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 *msg;
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;
508 struct callback_head task_work;
511 struct io_async_ctx {
513 struct io_async_rw rw;
514 struct io_async_msghdr msg;
515 struct io_async_connect connect;
516 struct io_timeout_data timeout;
521 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
522 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
523 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
524 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
525 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
526 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
533 REQ_F_LINK_TIMEOUT_BIT,
535 REQ_F_COMP_LOCKED_BIT,
536 REQ_F_NEED_CLEANUP_BIT,
539 REQ_F_BUFFER_SELECTED_BIT,
540 REQ_F_NO_FILE_TABLE_BIT,
541 REQ_F_QUEUE_TIMEOUT_BIT,
542 REQ_F_WORK_INITIALIZED_BIT,
543 REQ_F_TASK_PINNED_BIT,
545 /* not a real bit, just to check we're not overflowing the space */
551 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
552 /* drain existing IO first */
553 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
555 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
556 /* doesn't sever on completion < 0 */
557 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
559 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
560 /* IOSQE_BUFFER_SELECT */
561 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
564 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
565 /* fail rest of links */
566 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
567 /* on inflight list */
568 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
569 /* read/write uses file position */
570 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
571 /* must not punt to workers */
572 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
573 /* has linked timeout */
574 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
576 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
577 /* completion under lock */
578 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
580 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
581 /* in overflow list */
582 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
583 /* already went through poll handler */
584 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
585 /* buffer already selected */
586 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
587 /* doesn't need file table for this request */
588 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
589 /* needs to queue linked timeout */
590 REQ_F_QUEUE_TIMEOUT = BIT(REQ_F_QUEUE_TIMEOUT_BIT),
591 /* io_wq_work is initialized */
592 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
593 /* req->task is refcounted */
594 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
598 struct io_poll_iocb poll;
599 struct io_wq_work work;
603 * NOTE! Each of the iocb union members has the file pointer
604 * as the first entry in their struct definition. So you can
605 * access the file pointer through any of the sub-structs,
606 * or directly as just 'ki_filp' in this struct.
612 struct io_poll_iocb poll;
613 struct io_accept accept;
615 struct io_cancel cancel;
616 struct io_timeout timeout;
617 struct io_connect connect;
618 struct io_sr_msg sr_msg;
620 struct io_close close;
621 struct io_files_update files_update;
622 struct io_fadvise fadvise;
623 struct io_madvise madvise;
624 struct io_epoll epoll;
625 struct io_splice splice;
626 struct io_provide_buf pbuf;
627 struct io_statx statx;
630 struct io_async_ctx *io;
633 /* polled IO has completed */
638 struct io_ring_ctx *ctx;
639 struct list_head list;
642 struct task_struct *task;
648 struct list_head link_list;
650 struct list_head inflight_entry;
652 struct percpu_ref *fixed_file_refs;
656 * Only commands that never go async can use the below fields,
657 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
658 * async armed poll handlers for regular commands. The latter
659 * restore the work, if needed.
662 struct hlist_node hash_node;
663 struct async_poll *apoll;
665 struct io_wq_work work;
667 struct callback_head task_work;
670 #define IO_IOPOLL_BATCH 8
672 struct io_comp_state {
674 struct list_head list;
675 struct io_ring_ctx *ctx;
678 struct io_submit_state {
679 struct blk_plug plug;
682 * io_kiocb alloc cache
684 void *reqs[IO_IOPOLL_BATCH];
685 unsigned int free_reqs;
688 * Batch completion logic
690 struct io_comp_state comp;
693 * File reference cache
697 unsigned int has_refs;
698 unsigned int used_refs;
699 unsigned int ios_left;
703 /* needs req->io allocated for deferral/async */
704 unsigned async_ctx : 1;
705 /* needs current->mm setup, does mm access */
706 unsigned needs_mm : 1;
707 /* needs req->file assigned */
708 unsigned needs_file : 1;
709 /* don't fail if file grab fails */
710 unsigned needs_file_no_error : 1;
711 /* hash wq insertion if file is a regular file */
712 unsigned hash_reg_file : 1;
713 /* unbound wq insertion if file is a non-regular file */
714 unsigned unbound_nonreg_file : 1;
715 /* opcode is not supported by this kernel */
716 unsigned not_supported : 1;
717 /* needs file table */
718 unsigned file_table : 1;
720 unsigned needs_fs : 1;
721 /* set if opcode supports polled "wait" */
723 unsigned pollout : 1;
724 /* op supports buffer selection */
725 unsigned buffer_select : 1;
728 static const struct io_op_def io_op_defs[] = {
729 [IORING_OP_NOP] = {},
730 [IORING_OP_READV] = {
734 .unbound_nonreg_file = 1,
738 [IORING_OP_WRITEV] = {
743 .unbound_nonreg_file = 1,
746 [IORING_OP_FSYNC] = {
749 [IORING_OP_READ_FIXED] = {
751 .unbound_nonreg_file = 1,
754 [IORING_OP_WRITE_FIXED] = {
757 .unbound_nonreg_file = 1,
760 [IORING_OP_POLL_ADD] = {
762 .unbound_nonreg_file = 1,
764 [IORING_OP_POLL_REMOVE] = {},
765 [IORING_OP_SYNC_FILE_RANGE] = {
768 [IORING_OP_SENDMSG] = {
772 .unbound_nonreg_file = 1,
776 [IORING_OP_RECVMSG] = {
780 .unbound_nonreg_file = 1,
785 [IORING_OP_TIMEOUT] = {
789 [IORING_OP_TIMEOUT_REMOVE] = {},
790 [IORING_OP_ACCEPT] = {
793 .unbound_nonreg_file = 1,
797 [IORING_OP_ASYNC_CANCEL] = {},
798 [IORING_OP_LINK_TIMEOUT] = {
802 [IORING_OP_CONNECT] = {
806 .unbound_nonreg_file = 1,
809 [IORING_OP_FALLOCATE] = {
812 [IORING_OP_OPENAT] = {
816 [IORING_OP_CLOSE] = {
818 .needs_file_no_error = 1,
821 [IORING_OP_FILES_UPDATE] = {
825 [IORING_OP_STATX] = {
833 .unbound_nonreg_file = 1,
837 [IORING_OP_WRITE] = {
840 .unbound_nonreg_file = 1,
843 [IORING_OP_FADVISE] = {
846 [IORING_OP_MADVISE] = {
852 .unbound_nonreg_file = 1,
858 .unbound_nonreg_file = 1,
862 [IORING_OP_OPENAT2] = {
866 [IORING_OP_EPOLL_CTL] = {
867 .unbound_nonreg_file = 1,
870 [IORING_OP_SPLICE] = {
873 .unbound_nonreg_file = 1,
875 [IORING_OP_PROVIDE_BUFFERS] = {},
876 [IORING_OP_REMOVE_BUFFERS] = {},
880 .unbound_nonreg_file = 1,
884 enum io_mem_account {
889 static bool io_rw_reissue(struct io_kiocb *req, long res);
890 static void io_cqring_fill_event(struct io_kiocb *req, long res);
891 static void io_put_req(struct io_kiocb *req);
892 static void io_double_put_req(struct io_kiocb *req);
893 static void __io_double_put_req(struct io_kiocb *req);
894 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
895 static void io_queue_linked_timeout(struct io_kiocb *req);
896 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
897 struct io_uring_files_update *ip,
899 static int io_grab_files(struct io_kiocb *req);
900 static void io_complete_rw_common(struct kiocb *kiocb, long res,
901 struct io_comp_state *cs);
902 static void io_cleanup_req(struct io_kiocb *req);
903 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
904 int fd, struct file **out_file, bool fixed);
905 static void __io_queue_sqe(struct io_kiocb *req,
906 const struct io_uring_sqe *sqe,
907 struct io_comp_state *cs);
909 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
910 struct iovec **iovec, struct iov_iter *iter,
912 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
913 struct iovec *iovec, struct iovec *fast_iov,
914 struct iov_iter *iter);
916 static struct kmem_cache *req_cachep;
918 static const struct file_operations io_uring_fops;
920 struct sock *io_uring_get_socket(struct file *file)
922 #if defined(CONFIG_UNIX)
923 if (file->f_op == &io_uring_fops) {
924 struct io_ring_ctx *ctx = file->private_data;
926 return ctx->ring_sock->sk;
931 EXPORT_SYMBOL(io_uring_get_socket);
933 static void io_get_req_task(struct io_kiocb *req)
935 if (req->flags & REQ_F_TASK_PINNED)
937 get_task_struct(req->task);
938 req->flags |= REQ_F_TASK_PINNED;
941 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
942 static void __io_put_req_task(struct io_kiocb *req)
944 if (req->flags & REQ_F_TASK_PINNED)
945 put_task_struct(req->task);
948 static void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
950 struct mm_struct *mm = current->mm;
953 kthread_unuse_mm(mm);
958 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
961 if (unlikely(!ctx->sqo_mm || !mmget_not_zero(ctx->sqo_mm)))
963 kthread_use_mm(ctx->sqo_mm);
969 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
970 struct io_kiocb *req)
972 if (!io_op_defs[req->opcode].needs_mm)
974 return __io_sq_thread_acquire_mm(ctx);
977 static inline void req_set_fail_links(struct io_kiocb *req)
979 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
980 req->flags |= REQ_F_FAIL_LINK;
983 static void io_file_put_work(struct work_struct *work);
986 * Note: must call io_req_init_async() for the first time you
987 * touch any members of io_wq_work.
989 static inline void io_req_init_async(struct io_kiocb *req)
991 if (req->flags & REQ_F_WORK_INITIALIZED)
994 memset(&req->work, 0, sizeof(req->work));
995 req->flags |= REQ_F_WORK_INITIALIZED;
998 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1000 return ctx->flags & IORING_SETUP_SQPOLL;
1003 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1005 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1007 complete(&ctx->ref_comp);
1010 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1012 return !req->timeout.off;
1015 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1017 struct io_ring_ctx *ctx;
1020 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1024 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1025 if (!ctx->fallback_req)
1029 * Use 5 bits less than the max cq entries, that should give us around
1030 * 32 entries per hash list if totally full and uniformly spread.
1032 hash_bits = ilog2(p->cq_entries);
1036 ctx->cancel_hash_bits = hash_bits;
1037 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1039 if (!ctx->cancel_hash)
1041 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1043 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1044 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1047 ctx->flags = p->flags;
1048 init_waitqueue_head(&ctx->sqo_wait);
1049 init_waitqueue_head(&ctx->cq_wait);
1050 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1051 init_completion(&ctx->ref_comp);
1052 init_completion(&ctx->sq_thread_comp);
1053 idr_init(&ctx->io_buffer_idr);
1054 idr_init(&ctx->personality_idr);
1055 mutex_init(&ctx->uring_lock);
1056 init_waitqueue_head(&ctx->wait);
1057 spin_lock_init(&ctx->completion_lock);
1058 INIT_LIST_HEAD(&ctx->poll_list);
1059 INIT_LIST_HEAD(&ctx->defer_list);
1060 INIT_LIST_HEAD(&ctx->timeout_list);
1061 init_waitqueue_head(&ctx->inflight_wait);
1062 spin_lock_init(&ctx->inflight_lock);
1063 INIT_LIST_HEAD(&ctx->inflight_list);
1064 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1065 init_llist_head(&ctx->file_put_llist);
1068 if (ctx->fallback_req)
1069 kmem_cache_free(req_cachep, ctx->fallback_req);
1070 kfree(ctx->cancel_hash);
1075 static inline bool __req_need_defer(struct io_kiocb *req)
1077 struct io_ring_ctx *ctx = req->ctx;
1079 return req->sequence != ctx->cached_cq_tail
1080 + atomic_read(&ctx->cached_cq_overflow);
1083 static inline bool req_need_defer(struct io_kiocb *req)
1085 if (unlikely(req->flags & REQ_F_IO_DRAIN))
1086 return __req_need_defer(req);
1091 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1093 struct io_rings *rings = ctx->rings;
1095 /* order cqe stores with ring update */
1096 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1098 if (wq_has_sleeper(&ctx->cq_wait)) {
1099 wake_up_interruptible(&ctx->cq_wait);
1100 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1104 static void io_req_work_grab_env(struct io_kiocb *req)
1106 const struct io_op_def *def = &io_op_defs[req->opcode];
1108 io_req_init_async(req);
1110 if (!req->work.mm && def->needs_mm) {
1111 mmgrab(current->mm);
1112 req->work.mm = current->mm;
1114 if (!req->work.creds)
1115 req->work.creds = get_current_cred();
1116 if (!req->work.fs && def->needs_fs) {
1117 spin_lock(¤t->fs->lock);
1118 if (!current->fs->in_exec) {
1119 req->work.fs = current->fs;
1120 req->work.fs->users++;
1122 req->work.flags |= IO_WQ_WORK_CANCEL;
1124 spin_unlock(¤t->fs->lock);
1128 static inline void io_req_work_drop_env(struct io_kiocb *req)
1130 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1134 mmdrop(req->work.mm);
1135 req->work.mm = NULL;
1137 if (req->work.creds) {
1138 put_cred(req->work.creds);
1139 req->work.creds = NULL;
1142 struct fs_struct *fs = req->work.fs;
1144 spin_lock(&req->work.fs->lock);
1147 spin_unlock(&req->work.fs->lock);
1153 static void io_prep_async_work(struct io_kiocb *req)
1155 const struct io_op_def *def = &io_op_defs[req->opcode];
1157 if (req->flags & REQ_F_ISREG) {
1158 if (def->hash_reg_file)
1159 io_wq_hash_work(&req->work, file_inode(req->file));
1161 if (def->unbound_nonreg_file)
1162 req->work.flags |= IO_WQ_WORK_UNBOUND;
1165 io_req_work_grab_env(req);
1168 static void io_prep_async_link(struct io_kiocb *req)
1170 struct io_kiocb *cur;
1172 io_prep_async_work(req);
1173 if (req->flags & REQ_F_LINK_HEAD)
1174 list_for_each_entry(cur, &req->link_list, link_list)
1175 io_prep_async_work(cur);
1178 static void __io_queue_async_work(struct io_kiocb *req)
1180 struct io_ring_ctx *ctx = req->ctx;
1181 struct io_kiocb *link = io_prep_linked_timeout(req);
1183 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1184 &req->work, req->flags);
1185 io_wq_enqueue(ctx->io_wq, &req->work);
1188 io_queue_linked_timeout(link);
1191 static void io_queue_async_work(struct io_kiocb *req)
1193 /* init ->work of the whole link before punting */
1194 io_prep_async_link(req);
1195 __io_queue_async_work(req);
1198 static void io_kill_timeout(struct io_kiocb *req)
1202 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1204 atomic_inc(&req->ctx->cq_timeouts);
1205 list_del_init(&req->list);
1206 req->flags |= REQ_F_COMP_LOCKED;
1207 io_cqring_fill_event(req, 0);
1212 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1214 struct io_kiocb *req, *tmp;
1216 spin_lock_irq(&ctx->completion_lock);
1217 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1218 io_kill_timeout(req);
1219 spin_unlock_irq(&ctx->completion_lock);
1222 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1225 struct io_kiocb *req = list_first_entry(&ctx->defer_list,
1226 struct io_kiocb, list);
1228 if (req_need_defer(req))
1230 list_del_init(&req->list);
1231 /* punt-init is done before queueing for defer */
1232 __io_queue_async_work(req);
1233 } while (!list_empty(&ctx->defer_list));
1236 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1238 while (!list_empty(&ctx->timeout_list)) {
1239 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1240 struct io_kiocb, list);
1242 if (io_is_timeout_noseq(req))
1244 if (req->timeout.target_seq != ctx->cached_cq_tail
1245 - atomic_read(&ctx->cq_timeouts))
1248 list_del_init(&req->list);
1249 io_kill_timeout(req);
1253 static void io_commit_cqring(struct io_ring_ctx *ctx)
1255 io_flush_timeouts(ctx);
1256 __io_commit_cqring(ctx);
1258 if (unlikely(!list_empty(&ctx->defer_list)))
1259 __io_queue_deferred(ctx);
1262 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1264 struct io_rings *rings = ctx->rings;
1267 tail = ctx->cached_cq_tail;
1269 * writes to the cq entry need to come after reading head; the
1270 * control dependency is enough as we're using WRITE_ONCE to
1273 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1276 ctx->cached_cq_tail++;
1277 return &rings->cqes[tail & ctx->cq_mask];
1280 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1284 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1286 if (!ctx->eventfd_async)
1288 return io_wq_current_is_worker();
1291 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1293 if (waitqueue_active(&ctx->wait))
1294 wake_up(&ctx->wait);
1295 if (waitqueue_active(&ctx->sqo_wait))
1296 wake_up(&ctx->sqo_wait);
1297 if (io_should_trigger_evfd(ctx))
1298 eventfd_signal(ctx->cq_ev_fd, 1);
1301 /* Returns true if there are no backlogged entries after the flush */
1302 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1304 struct io_rings *rings = ctx->rings;
1305 struct io_uring_cqe *cqe;
1306 struct io_kiocb *req;
1307 unsigned long flags;
1311 if (list_empty_careful(&ctx->cq_overflow_list))
1313 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1314 rings->cq_ring_entries))
1318 spin_lock_irqsave(&ctx->completion_lock, flags);
1320 /* if force is set, the ring is going away. always drop after that */
1322 ctx->cq_overflow_flushed = 1;
1325 while (!list_empty(&ctx->cq_overflow_list)) {
1326 cqe = io_get_cqring(ctx);
1330 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1332 list_move(&req->list, &list);
1333 req->flags &= ~REQ_F_OVERFLOW;
1335 WRITE_ONCE(cqe->user_data, req->user_data);
1336 WRITE_ONCE(cqe->res, req->result);
1337 WRITE_ONCE(cqe->flags, req->cflags);
1339 WRITE_ONCE(ctx->rings->cq_overflow,
1340 atomic_inc_return(&ctx->cached_cq_overflow));
1344 io_commit_cqring(ctx);
1346 clear_bit(0, &ctx->sq_check_overflow);
1347 clear_bit(0, &ctx->cq_check_overflow);
1349 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1350 io_cqring_ev_posted(ctx);
1352 while (!list_empty(&list)) {
1353 req = list_first_entry(&list, struct io_kiocb, list);
1354 list_del(&req->list);
1361 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1363 struct io_ring_ctx *ctx = req->ctx;
1364 struct io_uring_cqe *cqe;
1366 trace_io_uring_complete(ctx, req->user_data, res);
1369 * If we can't get a cq entry, userspace overflowed the
1370 * submission (by quite a lot). Increment the overflow count in
1373 cqe = io_get_cqring(ctx);
1375 WRITE_ONCE(cqe->user_data, req->user_data);
1376 WRITE_ONCE(cqe->res, res);
1377 WRITE_ONCE(cqe->flags, cflags);
1378 } else if (ctx->cq_overflow_flushed) {
1379 WRITE_ONCE(ctx->rings->cq_overflow,
1380 atomic_inc_return(&ctx->cached_cq_overflow));
1382 if (list_empty(&ctx->cq_overflow_list)) {
1383 set_bit(0, &ctx->sq_check_overflow);
1384 set_bit(0, &ctx->cq_check_overflow);
1386 req->flags |= REQ_F_OVERFLOW;
1387 refcount_inc(&req->refs);
1389 req->cflags = cflags;
1390 list_add_tail(&req->list, &ctx->cq_overflow_list);
1394 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1396 __io_cqring_fill_event(req, res, 0);
1399 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1401 struct io_ring_ctx *ctx = req->ctx;
1402 unsigned long flags;
1404 spin_lock_irqsave(&ctx->completion_lock, flags);
1405 __io_cqring_fill_event(req, res, cflags);
1406 io_commit_cqring(ctx);
1407 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1409 io_cqring_ev_posted(ctx);
1412 static void io_submit_flush_completions(struct io_comp_state *cs)
1414 struct io_ring_ctx *ctx = cs->ctx;
1416 spin_lock_irq(&ctx->completion_lock);
1417 while (!list_empty(&cs->list)) {
1418 struct io_kiocb *req;
1420 req = list_first_entry(&cs->list, struct io_kiocb, list);
1421 list_del(&req->list);
1422 io_cqring_fill_event(req, req->result);
1423 if (!(req->flags & REQ_F_LINK_HEAD)) {
1424 req->flags |= REQ_F_COMP_LOCKED;
1427 spin_unlock_irq(&ctx->completion_lock);
1429 spin_lock_irq(&ctx->completion_lock);
1432 io_commit_cqring(ctx);
1433 spin_unlock_irq(&ctx->completion_lock);
1435 io_cqring_ev_posted(ctx);
1439 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1440 struct io_comp_state *cs)
1443 io_cqring_add_event(req, res, cflags);
1447 list_add_tail(&req->list, &cs->list);
1449 io_submit_flush_completions(cs);
1453 static void io_req_complete(struct io_kiocb *req, long res)
1455 __io_req_complete(req, res, 0, NULL);
1458 static inline bool io_is_fallback_req(struct io_kiocb *req)
1460 return req == (struct io_kiocb *)
1461 ((unsigned long) req->ctx->fallback_req & ~1UL);
1464 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1466 struct io_kiocb *req;
1468 req = ctx->fallback_req;
1469 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1475 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1476 struct io_submit_state *state)
1478 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1479 struct io_kiocb *req;
1481 if (!state->free_reqs) {
1485 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1486 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1489 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1490 * retry single alloc to be on the safe side.
1492 if (unlikely(ret <= 0)) {
1493 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1494 if (!state->reqs[0])
1498 state->free_reqs = ret - 1;
1499 req = state->reqs[ret - 1];
1502 req = state->reqs[state->free_reqs];
1507 return io_get_fallback_req(ctx);
1510 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1514 percpu_ref_put(req->fixed_file_refs);
1519 static void io_dismantle_req(struct io_kiocb *req)
1521 if (req->flags & REQ_F_NEED_CLEANUP)
1522 io_cleanup_req(req);
1526 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1527 __io_put_req_task(req);
1528 io_req_work_drop_env(req);
1530 if (req->flags & REQ_F_INFLIGHT) {
1531 struct io_ring_ctx *ctx = req->ctx;
1532 unsigned long flags;
1534 spin_lock_irqsave(&ctx->inflight_lock, flags);
1535 list_del(&req->inflight_entry);
1536 if (waitqueue_active(&ctx->inflight_wait))
1537 wake_up(&ctx->inflight_wait);
1538 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1542 static void __io_free_req(struct io_kiocb *req)
1544 struct io_ring_ctx *ctx;
1546 io_dismantle_req(req);
1548 if (likely(!io_is_fallback_req(req)))
1549 kmem_cache_free(req_cachep, req);
1551 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1552 percpu_ref_put(&ctx->refs);
1555 static bool io_link_cancel_timeout(struct io_kiocb *req)
1557 struct io_ring_ctx *ctx = req->ctx;
1560 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1562 io_cqring_fill_event(req, -ECANCELED);
1563 io_commit_cqring(ctx);
1564 req->flags &= ~REQ_F_LINK_HEAD;
1572 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1574 struct io_kiocb *link;
1577 if (list_empty(&req->link_list))
1579 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1580 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1583 list_del_init(&link->link_list);
1584 wake_ev = io_link_cancel_timeout(link);
1585 req->flags &= ~REQ_F_LINK_TIMEOUT;
1589 static void io_kill_linked_timeout(struct io_kiocb *req)
1591 struct io_ring_ctx *ctx = req->ctx;
1594 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1595 unsigned long flags;
1597 spin_lock_irqsave(&ctx->completion_lock, flags);
1598 wake_ev = __io_kill_linked_timeout(req);
1599 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1601 wake_ev = __io_kill_linked_timeout(req);
1605 io_cqring_ev_posted(ctx);
1608 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1610 struct io_kiocb *nxt;
1613 * The list should never be empty when we are called here. But could
1614 * potentially happen if the chain is messed up, check to be on the
1617 if (unlikely(list_empty(&req->link_list)))
1620 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1621 list_del_init(&req->link_list);
1622 if (!list_empty(&nxt->link_list))
1623 nxt->flags |= REQ_F_LINK_HEAD;
1628 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1630 static void __io_fail_links(struct io_kiocb *req)
1632 struct io_ring_ctx *ctx = req->ctx;
1634 while (!list_empty(&req->link_list)) {
1635 struct io_kiocb *link = list_first_entry(&req->link_list,
1636 struct io_kiocb, link_list);
1638 list_del_init(&link->link_list);
1639 trace_io_uring_fail_link(req, link);
1641 io_cqring_fill_event(link, -ECANCELED);
1642 __io_double_put_req(link);
1643 req->flags &= ~REQ_F_LINK_TIMEOUT;
1646 io_commit_cqring(ctx);
1647 io_cqring_ev_posted(ctx);
1650 static void io_fail_links(struct io_kiocb *req)
1652 struct io_ring_ctx *ctx = req->ctx;
1654 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1655 unsigned long flags;
1657 spin_lock_irqsave(&ctx->completion_lock, flags);
1658 __io_fail_links(req);
1659 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1661 __io_fail_links(req);
1664 io_cqring_ev_posted(ctx);
1667 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1669 req->flags &= ~REQ_F_LINK_HEAD;
1670 if (req->flags & REQ_F_LINK_TIMEOUT)
1671 io_kill_linked_timeout(req);
1674 * If LINK is set, we have dependent requests in this chain. If we
1675 * didn't fail this request, queue the first one up, moving any other
1676 * dependencies to the next request. In case of failure, fail the rest
1679 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1680 return io_req_link_next(req);
1685 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1687 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1689 return __io_req_find_next(req);
1692 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1694 struct io_ring_ctx *ctx = req->ctx;
1696 spin_lock_irq(&ctx->completion_lock);
1697 io_cqring_fill_event(req, error);
1698 io_commit_cqring(ctx);
1699 spin_unlock_irq(&ctx->completion_lock);
1701 io_cqring_ev_posted(ctx);
1702 req_set_fail_links(req);
1703 io_double_put_req(req);
1706 static void io_req_task_cancel(struct callback_head *cb)
1708 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1710 __io_req_task_cancel(req, -ECANCELED);
1713 static void __io_req_task_submit(struct io_kiocb *req)
1715 struct io_ring_ctx *ctx = req->ctx;
1717 if (!__io_sq_thread_acquire_mm(ctx)) {
1718 mutex_lock(&ctx->uring_lock);
1719 __io_queue_sqe(req, NULL, NULL);
1720 mutex_unlock(&ctx->uring_lock);
1722 __io_req_task_cancel(req, -EFAULT);
1726 static void io_req_task_submit(struct callback_head *cb)
1728 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1730 __io_req_task_submit(req);
1733 static void io_req_task_queue(struct io_kiocb *req)
1735 struct task_struct *tsk = req->task;
1738 init_task_work(&req->task_work, io_req_task_submit);
1740 ret = task_work_add(tsk, &req->task_work, true);
1741 if (unlikely(ret)) {
1742 init_task_work(&req->task_work, io_req_task_cancel);
1743 tsk = io_wq_get_task(req->ctx->io_wq);
1744 task_work_add(tsk, &req->task_work, true);
1746 wake_up_process(tsk);
1749 static void io_queue_next(struct io_kiocb *req)
1751 struct io_kiocb *nxt = io_req_find_next(req);
1754 io_req_task_queue(nxt);
1757 static void io_free_req(struct io_kiocb *req)
1764 void *reqs[IO_IOPOLL_BATCH];
1768 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1769 struct req_batch *rb)
1771 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1772 percpu_ref_put_many(&ctx->refs, rb->to_free);
1776 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1777 struct req_batch *rb)
1780 __io_req_free_batch_flush(ctx, rb);
1783 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1785 if (unlikely(io_is_fallback_req(req))) {
1789 if (req->flags & REQ_F_LINK_HEAD)
1792 io_dismantle_req(req);
1793 rb->reqs[rb->to_free++] = req;
1794 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1795 __io_req_free_batch_flush(req->ctx, rb);
1799 * Drop reference to request, return next in chain (if there is one) if this
1800 * was the last reference to this request.
1802 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1804 struct io_kiocb *nxt = NULL;
1806 if (refcount_dec_and_test(&req->refs)) {
1807 nxt = io_req_find_next(req);
1813 static void io_put_req(struct io_kiocb *req)
1815 if (refcount_dec_and_test(&req->refs))
1819 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1821 struct io_kiocb *timeout, *nxt = NULL;
1824 * A ref is owned by io-wq in which context we're. So, if that's the
1825 * last one, it's safe to steal next work. False negatives are Ok,
1826 * it just will be re-punted async in io_put_work()
1828 if (refcount_read(&req->refs) != 1)
1831 nxt = io_req_find_next(req);
1835 timeout = io_prep_linked_timeout(nxt);
1837 nxt->flags |= REQ_F_QUEUE_TIMEOUT;
1842 * Must only be used if we don't need to care about links, usually from
1843 * within the completion handling itself.
1845 static void __io_double_put_req(struct io_kiocb *req)
1847 /* drop both submit and complete references */
1848 if (refcount_sub_and_test(2, &req->refs))
1852 static void io_double_put_req(struct io_kiocb *req)
1854 /* drop both submit and complete references */
1855 if (refcount_sub_and_test(2, &req->refs))
1859 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1861 struct io_rings *rings = ctx->rings;
1863 if (test_bit(0, &ctx->cq_check_overflow)) {
1865 * noflush == true is from the waitqueue handler, just ensure
1866 * we wake up the task, and the next invocation will flush the
1867 * entries. We cannot safely to it from here.
1869 if (noflush && !list_empty(&ctx->cq_overflow_list))
1872 io_cqring_overflow_flush(ctx, false);
1875 /* See comment at the top of this file */
1877 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1880 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1882 struct io_rings *rings = ctx->rings;
1884 /* make sure SQ entry isn't read before tail */
1885 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1888 static int io_put_kbuf(struct io_kiocb *req)
1890 struct io_buffer *kbuf;
1893 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1894 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1895 cflags |= IORING_CQE_F_BUFFER;
1901 static inline bool io_run_task_work(void)
1903 if (current->task_works) {
1904 __set_current_state(TASK_RUNNING);
1912 static void io_iopoll_queue(struct list_head *again)
1914 struct io_kiocb *req;
1917 req = list_first_entry(again, struct io_kiocb, list);
1918 list_del(&req->list);
1919 if (!io_rw_reissue(req, -EAGAIN))
1920 io_complete_rw_common(&req->rw.kiocb, -EAGAIN, NULL);
1921 } while (!list_empty(again));
1925 * Find and free completed poll iocbs
1927 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1928 struct list_head *done)
1930 struct req_batch rb;
1931 struct io_kiocb *req;
1934 /* order with ->result store in io_complete_rw_iopoll() */
1938 while (!list_empty(done)) {
1941 req = list_first_entry(done, struct io_kiocb, list);
1942 if (READ_ONCE(req->result) == -EAGAIN) {
1943 req->iopoll_completed = 0;
1944 list_move_tail(&req->list, &again);
1947 list_del(&req->list);
1949 if (req->flags & REQ_F_BUFFER_SELECTED)
1950 cflags = io_put_kbuf(req);
1952 __io_cqring_fill_event(req, req->result, cflags);
1955 if (refcount_dec_and_test(&req->refs))
1956 io_req_free_batch(&rb, req);
1959 io_commit_cqring(ctx);
1960 if (ctx->flags & IORING_SETUP_SQPOLL)
1961 io_cqring_ev_posted(ctx);
1962 io_req_free_batch_finish(ctx, &rb);
1964 if (!list_empty(&again))
1965 io_iopoll_queue(&again);
1968 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1971 struct io_kiocb *req, *tmp;
1977 * Only spin for completions if we don't have multiple devices hanging
1978 * off our complete list, and we're under the requested amount.
1980 spin = !ctx->poll_multi_file && *nr_events < min;
1983 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1984 struct kiocb *kiocb = &req->rw.kiocb;
1987 * Move completed and retryable entries to our local lists.
1988 * If we find a request that requires polling, break out
1989 * and complete those lists first, if we have entries there.
1991 if (READ_ONCE(req->iopoll_completed)) {
1992 list_move_tail(&req->list, &done);
1995 if (!list_empty(&done))
1998 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2007 if (!list_empty(&done))
2008 io_iopoll_complete(ctx, nr_events, &done);
2014 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2015 * non-spinning poll check - we'll still enter the driver poll loop, but only
2016 * as a non-spinning completion check.
2018 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2021 while (!list_empty(&ctx->poll_list) && !need_resched()) {
2024 ret = io_do_iopoll(ctx, nr_events, min);
2027 if (!min || *nr_events >= min)
2035 * We can't just wait for polled events to come to us, we have to actively
2036 * find and complete them.
2038 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
2040 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2043 mutex_lock(&ctx->uring_lock);
2044 while (!list_empty(&ctx->poll_list)) {
2045 unsigned int nr_events = 0;
2047 io_iopoll_getevents(ctx, &nr_events, 1);
2050 * Ensure we allow local-to-the-cpu processing to take place,
2051 * in this case we need to ensure that we reap all events.
2055 mutex_unlock(&ctx->uring_lock);
2058 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
2061 int iters = 0, ret = 0;
2064 * We disallow the app entering submit/complete with polling, but we
2065 * still need to lock the ring to prevent racing with polled issue
2066 * that got punted to a workqueue.
2068 mutex_lock(&ctx->uring_lock);
2073 * Don't enter poll loop if we already have events pending.
2074 * If we do, we can potentially be spinning for commands that
2075 * already triggered a CQE (eg in error).
2077 if (io_cqring_events(ctx, false))
2081 * If a submit got punted to a workqueue, we can have the
2082 * application entering polling for a command before it gets
2083 * issued. That app will hold the uring_lock for the duration
2084 * of the poll right here, so we need to take a breather every
2085 * now and then to ensure that the issue has a chance to add
2086 * the poll to the issued list. Otherwise we can spin here
2087 * forever, while the workqueue is stuck trying to acquire the
2090 if (!(++iters & 7)) {
2091 mutex_unlock(&ctx->uring_lock);
2093 mutex_lock(&ctx->uring_lock);
2096 if (*nr_events < min)
2097 tmin = min - *nr_events;
2099 ret = io_iopoll_getevents(ctx, nr_events, tmin);
2103 } while (min && !*nr_events && !need_resched());
2105 mutex_unlock(&ctx->uring_lock);
2109 static void kiocb_end_write(struct io_kiocb *req)
2112 * Tell lockdep we inherited freeze protection from submission
2115 if (req->flags & REQ_F_ISREG) {
2116 struct inode *inode = file_inode(req->file);
2118 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2120 file_end_write(req->file);
2123 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2124 struct io_comp_state *cs)
2126 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2129 if (kiocb->ki_flags & IOCB_WRITE)
2130 kiocb_end_write(req);
2132 if (res != req->result)
2133 req_set_fail_links(req);
2134 if (req->flags & REQ_F_BUFFER_SELECTED)
2135 cflags = io_put_kbuf(req);
2136 __io_req_complete(req, res, cflags, cs);
2140 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2142 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2143 ssize_t ret = -ECANCELED;
2144 struct iov_iter iter;
2152 switch (req->opcode) {
2153 case IORING_OP_READV:
2154 case IORING_OP_READ_FIXED:
2155 case IORING_OP_READ:
2158 case IORING_OP_WRITEV:
2159 case IORING_OP_WRITE_FIXED:
2160 case IORING_OP_WRITE:
2164 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2169 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2172 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2177 req_set_fail_links(req);
2178 io_req_complete(req, ret);
2182 static void io_rw_resubmit(struct callback_head *cb)
2184 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2185 struct io_ring_ctx *ctx = req->ctx;
2188 err = io_sq_thread_acquire_mm(ctx, req);
2190 if (io_resubmit_prep(req, err)) {
2191 refcount_inc(&req->refs);
2192 io_queue_async_work(req);
2197 static bool io_rw_reissue(struct io_kiocb *req, long res)
2200 struct task_struct *tsk;
2203 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2207 init_task_work(&req->task_work, io_rw_resubmit);
2208 ret = task_work_add(tsk, &req->task_work, true);
2210 wake_up_process(tsk);
2217 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2218 struct io_comp_state *cs)
2220 if (!io_rw_reissue(req, res))
2221 io_complete_rw_common(&req->rw.kiocb, res, cs);
2224 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2226 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2228 __io_complete_rw(req, res, res2, NULL);
2231 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2233 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2235 if (kiocb->ki_flags & IOCB_WRITE)
2236 kiocb_end_write(req);
2238 if (res != -EAGAIN && res != req->result)
2239 req_set_fail_links(req);
2241 WRITE_ONCE(req->result, res);
2242 /* order with io_poll_complete() checking ->result */
2244 WRITE_ONCE(req->iopoll_completed, 1);
2248 * After the iocb has been issued, it's safe to be found on the poll list.
2249 * Adding the kiocb to the list AFTER submission ensures that we don't
2250 * find it from a io_iopoll_getevents() thread before the issuer is done
2251 * accessing the kiocb cookie.
2253 static void io_iopoll_req_issued(struct io_kiocb *req)
2255 struct io_ring_ctx *ctx = req->ctx;
2258 * Track whether we have multiple files in our lists. This will impact
2259 * how we do polling eventually, not spinning if we're on potentially
2260 * different devices.
2262 if (list_empty(&ctx->poll_list)) {
2263 ctx->poll_multi_file = false;
2264 } else if (!ctx->poll_multi_file) {
2265 struct io_kiocb *list_req;
2267 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
2269 if (list_req->file != req->file)
2270 ctx->poll_multi_file = true;
2274 * For fast devices, IO may have already completed. If it has, add
2275 * it to the front so we find it first.
2277 if (READ_ONCE(req->iopoll_completed))
2278 list_add(&req->list, &ctx->poll_list);
2280 list_add_tail(&req->list, &ctx->poll_list);
2282 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2283 wq_has_sleeper(&ctx->sqo_wait))
2284 wake_up(&ctx->sqo_wait);
2287 static void __io_state_file_put(struct io_submit_state *state)
2289 int diff = state->has_refs - state->used_refs;
2292 fput_many(state->file, diff);
2296 static inline void io_state_file_put(struct io_submit_state *state)
2299 __io_state_file_put(state);
2303 * Get as many references to a file as we have IOs left in this submission,
2304 * assuming most submissions are for one file, or at least that each file
2305 * has more than one submission.
2307 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2313 if (state->fd == fd) {
2318 __io_state_file_put(state);
2320 state->file = fget_many(fd, state->ios_left);
2325 state->has_refs = state->ios_left;
2326 state->used_refs = 1;
2331 static bool io_bdev_nowait(struct block_device *bdev)
2334 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2341 * If we tracked the file through the SCM inflight mechanism, we could support
2342 * any file. For now, just ensure that anything potentially problematic is done
2345 static bool io_file_supports_async(struct file *file, int rw)
2347 umode_t mode = file_inode(file)->i_mode;
2349 if (S_ISBLK(mode)) {
2350 if (io_bdev_nowait(file->f_inode->i_bdev))
2354 if (S_ISCHR(mode) || S_ISSOCK(mode))
2356 if (S_ISREG(mode)) {
2357 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2358 file->f_op != &io_uring_fops)
2363 /* any ->read/write should understand O_NONBLOCK */
2364 if (file->f_flags & O_NONBLOCK)
2367 if (!(file->f_mode & FMODE_NOWAIT))
2371 return file->f_op->read_iter != NULL;
2373 return file->f_op->write_iter != NULL;
2376 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2377 bool force_nonblock)
2379 struct io_ring_ctx *ctx = req->ctx;
2380 struct kiocb *kiocb = &req->rw.kiocb;
2384 if (S_ISREG(file_inode(req->file)->i_mode))
2385 req->flags |= REQ_F_ISREG;
2387 kiocb->ki_pos = READ_ONCE(sqe->off);
2388 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2389 req->flags |= REQ_F_CUR_POS;
2390 kiocb->ki_pos = req->file->f_pos;
2392 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2393 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2394 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2398 ioprio = READ_ONCE(sqe->ioprio);
2400 ret = ioprio_check_cap(ioprio);
2404 kiocb->ki_ioprio = ioprio;
2406 kiocb->ki_ioprio = get_current_ioprio();
2408 /* don't allow async punt if RWF_NOWAIT was requested */
2409 if (kiocb->ki_flags & IOCB_NOWAIT)
2410 req->flags |= REQ_F_NOWAIT;
2412 if (kiocb->ki_flags & IOCB_DIRECT)
2413 io_get_req_task(req);
2416 kiocb->ki_flags |= IOCB_NOWAIT;
2418 if (ctx->flags & IORING_SETUP_IOPOLL) {
2419 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2420 !kiocb->ki_filp->f_op->iopoll)
2423 kiocb->ki_flags |= IOCB_HIPRI;
2424 kiocb->ki_complete = io_complete_rw_iopoll;
2425 req->iopoll_completed = 0;
2426 io_get_req_task(req);
2428 if (kiocb->ki_flags & IOCB_HIPRI)
2430 kiocb->ki_complete = io_complete_rw;
2433 req->rw.addr = READ_ONCE(sqe->addr);
2434 req->rw.len = READ_ONCE(sqe->len);
2435 req->buf_index = READ_ONCE(sqe->buf_index);
2439 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2445 case -ERESTARTNOINTR:
2446 case -ERESTARTNOHAND:
2447 case -ERESTART_RESTARTBLOCK:
2449 * We can't just restart the syscall, since previously
2450 * submitted sqes may already be in progress. Just fail this
2456 kiocb->ki_complete(kiocb, ret, 0);
2460 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2461 struct io_comp_state *cs)
2463 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2465 if (req->flags & REQ_F_CUR_POS)
2466 req->file->f_pos = kiocb->ki_pos;
2467 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2468 __io_complete_rw(req, ret, 0, cs);
2470 io_rw_done(kiocb, ret);
2473 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2474 struct iov_iter *iter)
2476 struct io_ring_ctx *ctx = req->ctx;
2477 size_t len = req->rw.len;
2478 struct io_mapped_ubuf *imu;
2479 u16 index, buf_index;
2483 /* attempt to use fixed buffers without having provided iovecs */
2484 if (unlikely(!ctx->user_bufs))
2487 buf_index = req->buf_index;
2488 if (unlikely(buf_index >= ctx->nr_user_bufs))
2491 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2492 imu = &ctx->user_bufs[index];
2493 buf_addr = req->rw.addr;
2496 if (buf_addr + len < buf_addr)
2498 /* not inside the mapped region */
2499 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2503 * May not be a start of buffer, set size appropriately
2504 * and advance us to the beginning.
2506 offset = buf_addr - imu->ubuf;
2507 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2511 * Don't use iov_iter_advance() here, as it's really slow for
2512 * using the latter parts of a big fixed buffer - it iterates
2513 * over each segment manually. We can cheat a bit here, because
2516 * 1) it's a BVEC iter, we set it up
2517 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2518 * first and last bvec
2520 * So just find our index, and adjust the iterator afterwards.
2521 * If the offset is within the first bvec (or the whole first
2522 * bvec, just use iov_iter_advance(). This makes it easier
2523 * since we can just skip the first segment, which may not
2524 * be PAGE_SIZE aligned.
2526 const struct bio_vec *bvec = imu->bvec;
2528 if (offset <= bvec->bv_len) {
2529 iov_iter_advance(iter, offset);
2531 unsigned long seg_skip;
2533 /* skip first vec */
2534 offset -= bvec->bv_len;
2535 seg_skip = 1 + (offset >> PAGE_SHIFT);
2537 iter->bvec = bvec + seg_skip;
2538 iter->nr_segs -= seg_skip;
2539 iter->count -= bvec->bv_len + offset;
2540 iter->iov_offset = offset & ~PAGE_MASK;
2547 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2550 mutex_unlock(&ctx->uring_lock);
2553 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2556 * "Normal" inline submissions always hold the uring_lock, since we
2557 * grab it from the system call. Same is true for the SQPOLL offload.
2558 * The only exception is when we've detached the request and issue it
2559 * from an async worker thread, grab the lock for that case.
2562 mutex_lock(&ctx->uring_lock);
2565 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2566 int bgid, struct io_buffer *kbuf,
2569 struct io_buffer *head;
2571 if (req->flags & REQ_F_BUFFER_SELECTED)
2574 io_ring_submit_lock(req->ctx, needs_lock);
2576 lockdep_assert_held(&req->ctx->uring_lock);
2578 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2580 if (!list_empty(&head->list)) {
2581 kbuf = list_last_entry(&head->list, struct io_buffer,
2583 list_del(&kbuf->list);
2586 idr_remove(&req->ctx->io_buffer_idr, bgid);
2588 if (*len > kbuf->len)
2591 kbuf = ERR_PTR(-ENOBUFS);
2594 io_ring_submit_unlock(req->ctx, needs_lock);
2599 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2602 struct io_buffer *kbuf;
2605 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2606 bgid = req->buf_index;
2607 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2610 req->rw.addr = (u64) (unsigned long) kbuf;
2611 req->flags |= REQ_F_BUFFER_SELECTED;
2612 return u64_to_user_ptr(kbuf->addr);
2615 #ifdef CONFIG_COMPAT
2616 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2619 struct compat_iovec __user *uiov;
2620 compat_ssize_t clen;
2624 uiov = u64_to_user_ptr(req->rw.addr);
2625 if (!access_ok(uiov, sizeof(*uiov)))
2627 if (__get_user(clen, &uiov->iov_len))
2633 buf = io_rw_buffer_select(req, &len, needs_lock);
2635 return PTR_ERR(buf);
2636 iov[0].iov_base = buf;
2637 iov[0].iov_len = (compat_size_t) len;
2642 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2645 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2649 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2652 len = iov[0].iov_len;
2655 buf = io_rw_buffer_select(req, &len, needs_lock);
2657 return PTR_ERR(buf);
2658 iov[0].iov_base = buf;
2659 iov[0].iov_len = len;
2663 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2666 if (req->flags & REQ_F_BUFFER_SELECTED) {
2667 struct io_buffer *kbuf;
2669 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2670 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2671 iov[0].iov_len = kbuf->len;
2676 else if (req->rw.len > 1)
2679 #ifdef CONFIG_COMPAT
2680 if (req->ctx->compat)
2681 return io_compat_import(req, iov, needs_lock);
2684 return __io_iov_buffer_select(req, iov, needs_lock);
2687 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2688 struct iovec **iovec, struct iov_iter *iter,
2691 void __user *buf = u64_to_user_ptr(req->rw.addr);
2692 size_t sqe_len = req->rw.len;
2696 opcode = req->opcode;
2697 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2699 return io_import_fixed(req, rw, iter);
2702 /* buffer index only valid with fixed read/write, or buffer select */
2703 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2706 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2707 if (req->flags & REQ_F_BUFFER_SELECT) {
2708 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2711 return PTR_ERR(buf);
2713 req->rw.len = sqe_len;
2716 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2718 return ret < 0 ? ret : sqe_len;
2722 struct io_async_rw *iorw = &req->io->rw;
2725 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2726 if (iorw->iov == iorw->fast_iov)
2731 if (req->flags & REQ_F_BUFFER_SELECT) {
2732 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2734 ret = (*iovec)->iov_len;
2735 iov_iter_init(iter, rw, *iovec, 1, ret);
2741 #ifdef CONFIG_COMPAT
2742 if (req->ctx->compat)
2743 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2747 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2751 * For files that don't have ->read_iter() and ->write_iter(), handle them
2752 * by looping over ->read() or ->write() manually.
2754 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2755 struct iov_iter *iter)
2760 * Don't support polled IO through this interface, and we can't
2761 * support non-blocking either. For the latter, this just causes
2762 * the kiocb to be handled from an async context.
2764 if (kiocb->ki_flags & IOCB_HIPRI)
2766 if (kiocb->ki_flags & IOCB_NOWAIT)
2769 while (iov_iter_count(iter)) {
2773 if (!iov_iter_is_bvec(iter)) {
2774 iovec = iov_iter_iovec(iter);
2776 /* fixed buffers import bvec */
2777 iovec.iov_base = kmap(iter->bvec->bv_page)
2779 iovec.iov_len = min(iter->count,
2780 iter->bvec->bv_len - iter->iov_offset);
2784 nr = file->f_op->read(file, iovec.iov_base,
2785 iovec.iov_len, &kiocb->ki_pos);
2787 nr = file->f_op->write(file, iovec.iov_base,
2788 iovec.iov_len, &kiocb->ki_pos);
2791 if (iov_iter_is_bvec(iter))
2792 kunmap(iter->bvec->bv_page);
2800 if (nr != iovec.iov_len)
2802 iov_iter_advance(iter, nr);
2808 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2809 struct iovec *iovec, struct iovec *fast_iov,
2810 struct iov_iter *iter)
2812 req->io->rw.nr_segs = iter->nr_segs;
2813 req->io->rw.size = io_size;
2814 req->io->rw.iov = iovec;
2815 if (!req->io->rw.iov) {
2816 req->io->rw.iov = req->io->rw.fast_iov;
2817 if (req->io->rw.iov != fast_iov)
2818 memcpy(req->io->rw.iov, fast_iov,
2819 sizeof(struct iovec) * iter->nr_segs);
2821 req->flags |= REQ_F_NEED_CLEANUP;
2825 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2827 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2828 return req->io == NULL;
2831 static int io_alloc_async_ctx(struct io_kiocb *req)
2833 if (!io_op_defs[req->opcode].async_ctx)
2836 return __io_alloc_async_ctx(req);
2839 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2840 struct iovec *iovec, struct iovec *fast_iov,
2841 struct iov_iter *iter)
2843 if (!io_op_defs[req->opcode].async_ctx)
2846 if (__io_alloc_async_ctx(req))
2849 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2854 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2855 bool force_nonblock)
2857 struct io_async_ctx *io;
2858 struct iov_iter iter;
2861 ret = io_prep_rw(req, sqe, force_nonblock);
2865 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2868 /* either don't need iovec imported or already have it */
2869 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2873 io->rw.iov = io->rw.fast_iov;
2875 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2880 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2884 static void io_async_buf_cancel(struct callback_head *cb)
2886 struct io_async_rw *rw;
2887 struct io_kiocb *req;
2889 rw = container_of(cb, struct io_async_rw, task_work);
2890 req = rw->wpq.wait.private;
2891 __io_req_task_cancel(req, -ECANCELED);
2894 static void io_async_buf_retry(struct callback_head *cb)
2896 struct io_async_rw *rw;
2897 struct io_kiocb *req;
2899 rw = container_of(cb, struct io_async_rw, task_work);
2900 req = rw->wpq.wait.private;
2902 __io_req_task_submit(req);
2905 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2906 int sync, void *arg)
2908 struct wait_page_queue *wpq;
2909 struct io_kiocb *req = wait->private;
2910 struct io_async_rw *rw = &req->io->rw;
2911 struct wait_page_key *key = arg;
2912 struct task_struct *tsk;
2915 wpq = container_of(wait, struct wait_page_queue, wait);
2917 ret = wake_page_match(wpq, key);
2921 list_del_init(&wait->entry);
2923 init_task_work(&rw->task_work, io_async_buf_retry);
2924 /* submit ref gets dropped, acquire a new one */
2925 refcount_inc(&req->refs);
2927 ret = task_work_add(tsk, &rw->task_work, true);
2928 if (unlikely(ret)) {
2929 /* queue just for cancelation */
2930 init_task_work(&rw->task_work, io_async_buf_cancel);
2931 tsk = io_wq_get_task(req->ctx->io_wq);
2932 task_work_add(tsk, &rw->task_work, true);
2934 wake_up_process(tsk);
2938 static bool io_rw_should_retry(struct io_kiocb *req)
2940 struct kiocb *kiocb = &req->rw.kiocb;
2943 /* never retry for NOWAIT, we just complete with -EAGAIN */
2944 if (req->flags & REQ_F_NOWAIT)
2947 /* already tried, or we're doing O_DIRECT */
2948 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
2951 * just use poll if we can, and don't attempt if the fs doesn't
2952 * support callback based unlocks
2954 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
2958 * If request type doesn't require req->io to defer in general,
2959 * we need to allocate it here
2961 if (!req->io && __io_alloc_async_ctx(req))
2964 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
2965 io_async_buf_func, req);
2967 io_get_req_task(req);
2974 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
2976 if (req->file->f_op->read_iter)
2977 return call_read_iter(req->file, &req->rw.kiocb, iter);
2978 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
2981 static int io_read(struct io_kiocb *req, bool force_nonblock,
2982 struct io_comp_state *cs)
2984 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2985 struct kiocb *kiocb = &req->rw.kiocb;
2986 struct iov_iter iter;
2988 ssize_t io_size, ret;
2990 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2994 /* Ensure we clear previously set non-block flag */
2995 if (!force_nonblock)
2996 kiocb->ki_flags &= ~IOCB_NOWAIT;
2999 req->result = io_size;
3001 /* If the file doesn't support async, just async punt */
3002 if (force_nonblock && !io_file_supports_async(req->file, READ))
3005 iov_count = iov_iter_count(&iter);
3006 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3008 unsigned long nr_segs = iter.nr_segs;
3011 ret2 = io_iter_do_read(req, &iter);
3013 /* Catch -EAGAIN return for forced non-blocking submission */
3014 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3015 kiocb_done(kiocb, ret2, cs);
3017 iter.count = iov_count;
3018 iter.nr_segs = nr_segs;
3020 ret = io_setup_async_rw(req, io_size, iovec,
3021 inline_vecs, &iter);
3024 /* if we can retry, do so with the callbacks armed */
3025 if (io_rw_should_retry(req)) {
3026 ret2 = io_iter_do_read(req, &iter);
3027 if (ret2 == -EIOCBQUEUED) {
3029 } else if (ret2 != -EAGAIN) {
3030 kiocb_done(kiocb, ret2, cs);
3034 kiocb->ki_flags &= ~IOCB_WAITQ;
3039 if (!(req->flags & REQ_F_NEED_CLEANUP))
3044 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3045 bool force_nonblock)
3047 struct io_async_ctx *io;
3048 struct iov_iter iter;
3051 ret = io_prep_rw(req, sqe, force_nonblock);
3055 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3058 req->fsize = rlimit(RLIMIT_FSIZE);
3060 /* either don't need iovec imported or already have it */
3061 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3065 io->rw.iov = io->rw.fast_iov;
3067 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
3072 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
3076 static int io_write(struct io_kiocb *req, bool force_nonblock,
3077 struct io_comp_state *cs)
3079 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3080 struct kiocb *kiocb = &req->rw.kiocb;
3081 struct iov_iter iter;
3083 ssize_t ret, io_size;
3085 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3089 /* Ensure we clear previously set non-block flag */
3090 if (!force_nonblock)
3091 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3094 req->result = io_size;
3096 /* If the file doesn't support async, just async punt */
3097 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3100 /* file path doesn't support NOWAIT for non-direct_IO */
3101 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3102 (req->flags & REQ_F_ISREG))
3105 iov_count = iov_iter_count(&iter);
3106 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3108 unsigned long nr_segs = iter.nr_segs;
3112 * Open-code file_start_write here to grab freeze protection,
3113 * which will be released by another thread in
3114 * io_complete_rw(). Fool lockdep by telling it the lock got
3115 * released so that it doesn't complain about the held lock when
3116 * we return to userspace.
3118 if (req->flags & REQ_F_ISREG) {
3119 __sb_start_write(file_inode(req->file)->i_sb,
3120 SB_FREEZE_WRITE, true);
3121 __sb_writers_release(file_inode(req->file)->i_sb,
3124 kiocb->ki_flags |= IOCB_WRITE;
3126 if (!force_nonblock)
3127 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3129 if (req->file->f_op->write_iter)
3130 ret2 = call_write_iter(req->file, kiocb, &iter);
3132 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3134 if (!force_nonblock)
3135 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3138 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3139 * retry them without IOCB_NOWAIT.
3141 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3143 if (!force_nonblock || ret2 != -EAGAIN) {
3144 kiocb_done(kiocb, ret2, cs);
3146 iter.count = iov_count;
3147 iter.nr_segs = nr_segs;
3149 ret = io_setup_async_rw(req, io_size, iovec,
3150 inline_vecs, &iter);
3157 if (!(req->flags & REQ_F_NEED_CLEANUP))
3162 static int __io_splice_prep(struct io_kiocb *req,
3163 const struct io_uring_sqe *sqe)
3165 struct io_splice* sp = &req->splice;
3166 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3169 if (req->flags & REQ_F_NEED_CLEANUP)
3171 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3175 sp->len = READ_ONCE(sqe->len);
3176 sp->flags = READ_ONCE(sqe->splice_flags);
3178 if (unlikely(sp->flags & ~valid_flags))
3181 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3182 (sp->flags & SPLICE_F_FD_IN_FIXED));
3185 req->flags |= REQ_F_NEED_CLEANUP;
3187 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3189 * Splice operation will be punted aync, and here need to
3190 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3192 io_req_init_async(req);
3193 req->work.flags |= IO_WQ_WORK_UNBOUND;
3199 static int io_tee_prep(struct io_kiocb *req,
3200 const struct io_uring_sqe *sqe)
3202 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3204 return __io_splice_prep(req, sqe);
3207 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3209 struct io_splice *sp = &req->splice;
3210 struct file *in = sp->file_in;
3211 struct file *out = sp->file_out;
3212 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3218 ret = do_tee(in, out, sp->len, flags);
3220 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3221 req->flags &= ~REQ_F_NEED_CLEANUP;
3224 req_set_fail_links(req);
3225 io_req_complete(req, ret);
3229 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3231 struct io_splice* sp = &req->splice;
3233 sp->off_in = READ_ONCE(sqe->splice_off_in);
3234 sp->off_out = READ_ONCE(sqe->off);
3235 return __io_splice_prep(req, sqe);
3238 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3240 struct io_splice *sp = &req->splice;
3241 struct file *in = sp->file_in;
3242 struct file *out = sp->file_out;
3243 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3244 loff_t *poff_in, *poff_out;
3250 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3251 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3254 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3256 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3257 req->flags &= ~REQ_F_NEED_CLEANUP;
3260 req_set_fail_links(req);
3261 io_req_complete(req, ret);
3266 * IORING_OP_NOP just posts a completion event, nothing else.
3268 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3270 struct io_ring_ctx *ctx = req->ctx;
3272 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3275 __io_req_complete(req, 0, 0, cs);
3279 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3281 struct io_ring_ctx *ctx = req->ctx;
3286 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3288 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3291 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3292 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3295 req->sync.off = READ_ONCE(sqe->off);
3296 req->sync.len = READ_ONCE(sqe->len);
3300 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3302 loff_t end = req->sync.off + req->sync.len;
3305 /* fsync always requires a blocking context */
3309 ret = vfs_fsync_range(req->file, req->sync.off,
3310 end > 0 ? end : LLONG_MAX,
3311 req->sync.flags & IORING_FSYNC_DATASYNC);
3313 req_set_fail_links(req);
3314 io_req_complete(req, ret);
3318 static int io_fallocate_prep(struct io_kiocb *req,
3319 const struct io_uring_sqe *sqe)
3321 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3323 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3326 req->sync.off = READ_ONCE(sqe->off);
3327 req->sync.len = READ_ONCE(sqe->addr);
3328 req->sync.mode = READ_ONCE(sqe->len);
3329 req->fsize = rlimit(RLIMIT_FSIZE);
3333 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3337 /* fallocate always requiring blocking context */
3341 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3342 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3344 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3346 req_set_fail_links(req);
3347 io_req_complete(req, ret);
3351 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3353 const char __user *fname;
3356 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3358 if (unlikely(sqe->ioprio || sqe->buf_index))
3360 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3363 /* open.how should be already initialised */
3364 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3365 req->open.how.flags |= O_LARGEFILE;
3367 req->open.dfd = READ_ONCE(sqe->fd);
3368 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3369 req->open.filename = getname(fname);
3370 if (IS_ERR(req->open.filename)) {
3371 ret = PTR_ERR(req->open.filename);
3372 req->open.filename = NULL;
3375 req->open.nofile = rlimit(RLIMIT_NOFILE);
3376 req->flags |= REQ_F_NEED_CLEANUP;
3380 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3384 if (req->flags & REQ_F_NEED_CLEANUP)
3386 mode = READ_ONCE(sqe->len);
3387 flags = READ_ONCE(sqe->open_flags);
3388 req->open.how = build_open_how(flags, mode);
3389 return __io_openat_prep(req, sqe);
3392 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3394 struct open_how __user *how;
3398 if (req->flags & REQ_F_NEED_CLEANUP)
3400 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3401 len = READ_ONCE(sqe->len);
3402 if (len < OPEN_HOW_SIZE_VER0)
3405 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3410 return __io_openat_prep(req, sqe);
3413 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3415 struct open_flags op;
3422 ret = build_open_flags(&req->open.how, &op);
3426 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3430 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3433 ret = PTR_ERR(file);
3435 fsnotify_open(file);
3436 fd_install(ret, file);
3439 putname(req->open.filename);
3440 req->flags &= ~REQ_F_NEED_CLEANUP;
3442 req_set_fail_links(req);
3443 io_req_complete(req, ret);
3447 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3449 return io_openat2(req, force_nonblock);
3452 static int io_remove_buffers_prep(struct io_kiocb *req,
3453 const struct io_uring_sqe *sqe)
3455 struct io_provide_buf *p = &req->pbuf;
3458 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3461 tmp = READ_ONCE(sqe->fd);
3462 if (!tmp || tmp > USHRT_MAX)
3465 memset(p, 0, sizeof(*p));
3467 p->bgid = READ_ONCE(sqe->buf_group);
3471 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3472 int bgid, unsigned nbufs)
3476 /* shouldn't happen */
3480 /* the head kbuf is the list itself */
3481 while (!list_empty(&buf->list)) {
3482 struct io_buffer *nxt;
3484 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3485 list_del(&nxt->list);
3492 idr_remove(&ctx->io_buffer_idr, bgid);
3497 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3498 struct io_comp_state *cs)
3500 struct io_provide_buf *p = &req->pbuf;
3501 struct io_ring_ctx *ctx = req->ctx;
3502 struct io_buffer *head;
3505 io_ring_submit_lock(ctx, !force_nonblock);
3507 lockdep_assert_held(&ctx->uring_lock);
3510 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3512 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3514 io_ring_submit_lock(ctx, !force_nonblock);
3516 req_set_fail_links(req);
3517 __io_req_complete(req, ret, 0, cs);
3521 static int io_provide_buffers_prep(struct io_kiocb *req,
3522 const struct io_uring_sqe *sqe)
3524 struct io_provide_buf *p = &req->pbuf;
3527 if (sqe->ioprio || sqe->rw_flags)
3530 tmp = READ_ONCE(sqe->fd);
3531 if (!tmp || tmp > USHRT_MAX)
3534 p->addr = READ_ONCE(sqe->addr);
3535 p->len = READ_ONCE(sqe->len);
3537 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3540 p->bgid = READ_ONCE(sqe->buf_group);
3541 tmp = READ_ONCE(sqe->off);
3542 if (tmp > USHRT_MAX)
3548 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3550 struct io_buffer *buf;
3551 u64 addr = pbuf->addr;
3552 int i, bid = pbuf->bid;
3554 for (i = 0; i < pbuf->nbufs; i++) {
3555 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3560 buf->len = pbuf->len;
3565 INIT_LIST_HEAD(&buf->list);
3568 list_add_tail(&buf->list, &(*head)->list);
3572 return i ? i : -ENOMEM;
3575 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3576 struct io_comp_state *cs)
3578 struct io_provide_buf *p = &req->pbuf;
3579 struct io_ring_ctx *ctx = req->ctx;
3580 struct io_buffer *head, *list;
3583 io_ring_submit_lock(ctx, !force_nonblock);
3585 lockdep_assert_held(&ctx->uring_lock);
3587 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3589 ret = io_add_buffers(p, &head);
3594 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3597 __io_remove_buffers(ctx, head, p->bgid, -1U);
3602 io_ring_submit_unlock(ctx, !force_nonblock);
3604 req_set_fail_links(req);
3605 __io_req_complete(req, ret, 0, cs);
3609 static int io_epoll_ctl_prep(struct io_kiocb *req,
3610 const struct io_uring_sqe *sqe)
3612 #if defined(CONFIG_EPOLL)
3613 if (sqe->ioprio || sqe->buf_index)
3615 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3618 req->epoll.epfd = READ_ONCE(sqe->fd);
3619 req->epoll.op = READ_ONCE(sqe->len);
3620 req->epoll.fd = READ_ONCE(sqe->off);
3622 if (ep_op_has_event(req->epoll.op)) {
3623 struct epoll_event __user *ev;
3625 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3626 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3636 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3637 struct io_comp_state *cs)
3639 #if defined(CONFIG_EPOLL)
3640 struct io_epoll *ie = &req->epoll;
3643 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3644 if (force_nonblock && ret == -EAGAIN)
3648 req_set_fail_links(req);
3649 __io_req_complete(req, ret, 0, cs);
3656 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3658 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3659 if (sqe->ioprio || sqe->buf_index || sqe->off)
3661 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3664 req->madvise.addr = READ_ONCE(sqe->addr);
3665 req->madvise.len = READ_ONCE(sqe->len);
3666 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3673 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3675 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3676 struct io_madvise *ma = &req->madvise;
3682 ret = do_madvise(ma->addr, ma->len, ma->advice);
3684 req_set_fail_links(req);
3685 io_req_complete(req, ret);
3692 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3694 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3696 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3699 req->fadvise.offset = READ_ONCE(sqe->off);
3700 req->fadvise.len = READ_ONCE(sqe->len);
3701 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3705 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3707 struct io_fadvise *fa = &req->fadvise;
3710 if (force_nonblock) {
3711 switch (fa->advice) {
3712 case POSIX_FADV_NORMAL:
3713 case POSIX_FADV_RANDOM:
3714 case POSIX_FADV_SEQUENTIAL:
3721 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3723 req_set_fail_links(req);
3724 io_req_complete(req, ret);
3728 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3730 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3732 if (sqe->ioprio || sqe->buf_index)
3734 if (req->flags & REQ_F_FIXED_FILE)
3737 req->statx.dfd = READ_ONCE(sqe->fd);
3738 req->statx.mask = READ_ONCE(sqe->len);
3739 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3740 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3741 req->statx.flags = READ_ONCE(sqe->statx_flags);
3746 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3748 struct io_statx *ctx = &req->statx;
3751 if (force_nonblock) {
3752 /* only need file table for an actual valid fd */
3753 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3754 req->flags |= REQ_F_NO_FILE_TABLE;
3758 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3762 req_set_fail_links(req);
3763 io_req_complete(req, ret);
3767 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3770 * If we queue this for async, it must not be cancellable. That would
3771 * leave the 'file' in an undeterminate state, and here need to modify
3772 * io_wq_work.flags, so initialize io_wq_work firstly.
3774 io_req_init_async(req);
3775 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3777 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3779 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3780 sqe->rw_flags || sqe->buf_index)
3782 if (req->flags & REQ_F_FIXED_FILE)
3785 req->close.fd = READ_ONCE(sqe->fd);
3786 if ((req->file && req->file->f_op == &io_uring_fops) ||
3787 req->close.fd == req->ctx->ring_fd)
3790 req->close.put_file = NULL;
3794 static int io_close(struct io_kiocb *req, bool force_nonblock,
3795 struct io_comp_state *cs)
3797 struct io_close *close = &req->close;
3800 /* might be already done during nonblock submission */
3801 if (!close->put_file) {
3802 ret = __close_fd_get_file(close->fd, &close->put_file);
3804 return (ret == -ENOENT) ? -EBADF : ret;
3807 /* if the file has a flush method, be safe and punt to async */
3808 if (close->put_file->f_op->flush && force_nonblock) {
3809 /* was never set, but play safe */
3810 req->flags &= ~REQ_F_NOWAIT;
3811 /* avoid grabbing files - we don't need the files */
3812 req->flags |= REQ_F_NO_FILE_TABLE;
3816 /* No ->flush() or already async, safely close from here */
3817 ret = filp_close(close->put_file, req->work.files);
3819 req_set_fail_links(req);
3820 fput(close->put_file);
3821 close->put_file = NULL;
3822 __io_req_complete(req, ret, 0, cs);
3826 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3828 struct io_ring_ctx *ctx = req->ctx;
3833 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3835 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3838 req->sync.off = READ_ONCE(sqe->off);
3839 req->sync.len = READ_ONCE(sqe->len);
3840 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3844 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3848 /* sync_file_range always requires a blocking context */
3852 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3855 req_set_fail_links(req);
3856 io_req_complete(req, ret);
3860 #if defined(CONFIG_NET)
3861 static int io_setup_async_msg(struct io_kiocb *req,
3862 struct io_async_msghdr *kmsg)
3866 if (io_alloc_async_ctx(req)) {
3867 if (kmsg->iov != kmsg->fast_iov)
3871 req->flags |= REQ_F_NEED_CLEANUP;
3872 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3876 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3878 struct io_sr_msg *sr = &req->sr_msg;
3879 struct io_async_ctx *io = req->io;
3882 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3885 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3886 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3887 sr->len = READ_ONCE(sqe->len);
3889 #ifdef CONFIG_COMPAT
3890 if (req->ctx->compat)
3891 sr->msg_flags |= MSG_CMSG_COMPAT;
3894 if (!io || req->opcode == IORING_OP_SEND)
3896 /* iovec is already imported */
3897 if (req->flags & REQ_F_NEED_CLEANUP)
3900 io->msg.iov = io->msg.fast_iov;
3901 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3904 req->flags |= REQ_F_NEED_CLEANUP;
3908 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3909 struct io_comp_state *cs)
3911 struct io_async_msghdr *kmsg = NULL;
3912 struct socket *sock;
3915 sock = sock_from_file(req->file, &ret);
3917 struct io_async_ctx io;
3921 kmsg = &req->io->msg;
3922 kmsg->msg.msg_name = &req->io->msg.addr;
3923 /* if iov is set, it's allocated already */
3925 kmsg->iov = kmsg->fast_iov;
3926 kmsg->msg.msg_iter.iov = kmsg->iov;
3928 struct io_sr_msg *sr = &req->sr_msg;
3931 kmsg->msg.msg_name = &io.msg.addr;
3933 io.msg.iov = io.msg.fast_iov;
3934 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3935 sr->msg_flags, &io.msg.iov);
3940 flags = req->sr_msg.msg_flags;
3941 if (flags & MSG_DONTWAIT)
3942 req->flags |= REQ_F_NOWAIT;
3943 else if (force_nonblock)
3944 flags |= MSG_DONTWAIT;
3946 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3947 if (force_nonblock && ret == -EAGAIN)
3948 return io_setup_async_msg(req, kmsg);
3949 if (ret == -ERESTARTSYS)
3953 if (kmsg && kmsg->iov != kmsg->fast_iov)
3955 req->flags &= ~REQ_F_NEED_CLEANUP;
3957 req_set_fail_links(req);
3958 __io_req_complete(req, ret, 0, cs);
3962 static int io_send(struct io_kiocb *req, bool force_nonblock,
3963 struct io_comp_state *cs)
3965 struct socket *sock;
3968 sock = sock_from_file(req->file, &ret);
3970 struct io_sr_msg *sr = &req->sr_msg;
3975 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3980 msg.msg_name = NULL;
3981 msg.msg_control = NULL;
3982 msg.msg_controllen = 0;
3983 msg.msg_namelen = 0;
3985 flags = req->sr_msg.msg_flags;
3986 if (flags & MSG_DONTWAIT)
3987 req->flags |= REQ_F_NOWAIT;
3988 else if (force_nonblock)
3989 flags |= MSG_DONTWAIT;
3991 msg.msg_flags = flags;
3992 ret = sock_sendmsg(sock, &msg);
3993 if (force_nonblock && ret == -EAGAIN)
3995 if (ret == -ERESTARTSYS)
4000 req_set_fail_links(req);
4001 __io_req_complete(req, ret, 0, cs);
4005 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
4007 struct io_sr_msg *sr = &req->sr_msg;
4008 struct iovec __user *uiov;
4012 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
4017 if (req->flags & REQ_F_BUFFER_SELECT) {
4020 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
4022 sr->len = io->msg.iov[0].iov_len;
4023 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
4027 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4028 &io->msg.iov, &io->msg.msg.msg_iter);
4036 #ifdef CONFIG_COMPAT
4037 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4038 struct io_async_ctx *io)
4040 struct compat_msghdr __user *msg_compat;
4041 struct io_sr_msg *sr = &req->sr_msg;
4042 struct compat_iovec __user *uiov;
4047 msg_compat = (struct compat_msghdr __user *) sr->msg;
4048 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
4053 uiov = compat_ptr(ptr);
4054 if (req->flags & REQ_F_BUFFER_SELECT) {
4055 compat_ssize_t clen;
4059 if (!access_ok(uiov, sizeof(*uiov)))
4061 if (__get_user(clen, &uiov->iov_len))
4065 sr->len = io->msg.iov[0].iov_len;
4068 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4070 &io->msg.msg.msg_iter);
4079 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
4081 io->msg.iov = io->msg.fast_iov;
4083 #ifdef CONFIG_COMPAT
4084 if (req->ctx->compat)
4085 return __io_compat_recvmsg_copy_hdr(req, io);
4088 return __io_recvmsg_copy_hdr(req, io);
4091 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4092 int *cflags, bool needs_lock)
4094 struct io_sr_msg *sr = &req->sr_msg;
4095 struct io_buffer *kbuf;
4097 if (!(req->flags & REQ_F_BUFFER_SELECT))
4100 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4105 req->flags |= REQ_F_BUFFER_SELECTED;
4107 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
4108 *cflags |= IORING_CQE_F_BUFFER;
4112 static int io_recvmsg_prep(struct io_kiocb *req,
4113 const struct io_uring_sqe *sqe)
4115 struct io_sr_msg *sr = &req->sr_msg;
4116 struct io_async_ctx *io = req->io;
4119 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4122 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4123 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4124 sr->len = READ_ONCE(sqe->len);
4125 sr->bgid = READ_ONCE(sqe->buf_group);
4127 #ifdef CONFIG_COMPAT
4128 if (req->ctx->compat)
4129 sr->msg_flags |= MSG_CMSG_COMPAT;
4132 if (!io || req->opcode == IORING_OP_RECV)
4134 /* iovec is already imported */
4135 if (req->flags & REQ_F_NEED_CLEANUP)
4138 ret = io_recvmsg_copy_hdr(req, io);
4140 req->flags |= REQ_F_NEED_CLEANUP;
4144 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4145 struct io_comp_state *cs)
4147 struct io_async_msghdr *kmsg = NULL;
4148 struct socket *sock;
4149 int ret, cflags = 0;
4151 sock = sock_from_file(req->file, &ret);
4153 struct io_buffer *kbuf;
4154 struct io_async_ctx io;
4158 kmsg = &req->io->msg;
4159 kmsg->msg.msg_name = &req->io->msg.addr;
4160 /* if iov is set, it's allocated already */
4162 kmsg->iov = kmsg->fast_iov;
4163 kmsg->msg.msg_iter.iov = kmsg->iov;
4166 kmsg->msg.msg_name = &io.msg.addr;
4168 ret = io_recvmsg_copy_hdr(req, &io);
4173 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4175 return PTR_ERR(kbuf);
4177 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4178 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4179 1, req->sr_msg.len);
4182 flags = req->sr_msg.msg_flags;
4183 if (flags & MSG_DONTWAIT)
4184 req->flags |= REQ_F_NOWAIT;
4185 else if (force_nonblock)
4186 flags |= MSG_DONTWAIT;
4188 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
4189 kmsg->uaddr, flags);
4190 if (force_nonblock && ret == -EAGAIN)
4191 return io_setup_async_msg(req, kmsg);
4192 if (ret == -ERESTARTSYS)
4196 if (kmsg && kmsg->iov != kmsg->fast_iov)
4198 req->flags &= ~REQ_F_NEED_CLEANUP;
4200 req_set_fail_links(req);
4201 __io_req_complete(req, ret, cflags, cs);
4205 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4206 struct io_comp_state *cs)
4208 struct io_buffer *kbuf = NULL;
4209 struct socket *sock;
4210 int ret, cflags = 0;
4212 sock = sock_from_file(req->file, &ret);
4214 struct io_sr_msg *sr = &req->sr_msg;
4215 void __user *buf = sr->buf;
4220 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4222 return PTR_ERR(kbuf);
4224 buf = u64_to_user_ptr(kbuf->addr);
4226 ret = import_single_range(READ, buf, sr->len, &iov,
4233 req->flags |= REQ_F_NEED_CLEANUP;
4234 msg.msg_name = NULL;
4235 msg.msg_control = NULL;
4236 msg.msg_controllen = 0;
4237 msg.msg_namelen = 0;
4238 msg.msg_iocb = NULL;
4241 flags = req->sr_msg.msg_flags;
4242 if (flags & MSG_DONTWAIT)
4243 req->flags |= REQ_F_NOWAIT;
4244 else if (force_nonblock)
4245 flags |= MSG_DONTWAIT;
4247 ret = sock_recvmsg(sock, &msg, flags);
4248 if (force_nonblock && ret == -EAGAIN)
4250 if (ret == -ERESTARTSYS)
4255 req->flags &= ~REQ_F_NEED_CLEANUP;
4257 req_set_fail_links(req);
4258 __io_req_complete(req, ret, cflags, cs);
4262 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4264 struct io_accept *accept = &req->accept;
4266 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4268 if (sqe->ioprio || sqe->len || sqe->buf_index)
4271 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4272 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4273 accept->flags = READ_ONCE(sqe->accept_flags);
4274 accept->nofile = rlimit(RLIMIT_NOFILE);
4278 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4279 struct io_comp_state *cs)
4281 struct io_accept *accept = &req->accept;
4282 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4285 if (req->file->f_flags & O_NONBLOCK)
4286 req->flags |= REQ_F_NOWAIT;
4288 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4289 accept->addr_len, accept->flags,
4291 if (ret == -EAGAIN && force_nonblock)
4294 if (ret == -ERESTARTSYS)
4296 req_set_fail_links(req);
4298 __io_req_complete(req, ret, 0, cs);
4302 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4304 struct io_connect *conn = &req->connect;
4305 struct io_async_ctx *io = req->io;
4307 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4309 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4312 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4313 conn->addr_len = READ_ONCE(sqe->addr2);
4318 return move_addr_to_kernel(conn->addr, conn->addr_len,
4319 &io->connect.address);
4322 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4323 struct io_comp_state *cs)
4325 struct io_async_ctx __io, *io;
4326 unsigned file_flags;
4332 ret = move_addr_to_kernel(req->connect.addr,
4333 req->connect.addr_len,
4334 &__io.connect.address);
4340 file_flags = force_nonblock ? O_NONBLOCK : 0;
4342 ret = __sys_connect_file(req->file, &io->connect.address,
4343 req->connect.addr_len, file_flags);
4344 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4347 if (io_alloc_async_ctx(req)) {
4351 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4354 if (ret == -ERESTARTSYS)
4358 req_set_fail_links(req);
4359 __io_req_complete(req, ret, 0, cs);
4362 #else /* !CONFIG_NET */
4363 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4368 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4369 struct io_comp_state *cs)
4374 static int io_send(struct io_kiocb *req, bool force_nonblock,
4375 struct io_comp_state *cs)
4380 static int io_recvmsg_prep(struct io_kiocb *req,
4381 const struct io_uring_sqe *sqe)
4386 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4387 struct io_comp_state *cs)
4392 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4393 struct io_comp_state *cs)
4398 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4403 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4404 struct io_comp_state *cs)
4409 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4414 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4415 struct io_comp_state *cs)
4419 #endif /* CONFIG_NET */
4421 struct io_poll_table {
4422 struct poll_table_struct pt;
4423 struct io_kiocb *req;
4427 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
4429 struct task_struct *tsk = req->task;
4430 struct io_ring_ctx *ctx = req->ctx;
4431 int ret, notify = TWA_RESUME;
4434 * SQPOLL kernel thread doesn't need notification, just a wakeup.
4435 * If we're not using an eventfd, then TWA_RESUME is always fine,
4436 * as we won't have dependencies between request completions for
4437 * other kernel wait conditions.
4439 if (ctx->flags & IORING_SETUP_SQPOLL)
4441 else if (ctx->cq_ev_fd)
4442 notify = TWA_SIGNAL;
4444 ret = task_work_add(tsk, cb, notify);
4446 wake_up_process(tsk);
4450 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4451 __poll_t mask, task_work_func_t func)
4453 struct task_struct *tsk;
4456 /* for instances that support it check for an event match first: */
4457 if (mask && !(mask & poll->events))
4460 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4462 list_del_init(&poll->wait.entry);
4466 init_task_work(&req->task_work, func);
4468 * If this fails, then the task is exiting. When a task exits, the
4469 * work gets canceled, so just cancel this request as well instead
4470 * of executing it. We can't safely execute it anyway, as we may not
4471 * have the needed state needed for it anyway.
4473 ret = io_req_task_work_add(req, &req->task_work);
4474 if (unlikely(ret)) {
4475 WRITE_ONCE(poll->canceled, true);
4476 tsk = io_wq_get_task(req->ctx->io_wq);
4477 task_work_add(tsk, &req->task_work, 0);
4478 wake_up_process(tsk);
4483 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4484 __acquires(&req->ctx->completion_lock)
4486 struct io_ring_ctx *ctx = req->ctx;
4488 if (!req->result && !READ_ONCE(poll->canceled)) {
4489 struct poll_table_struct pt = { ._key = poll->events };
4491 req->result = vfs_poll(req->file, &pt) & poll->events;
4494 spin_lock_irq(&ctx->completion_lock);
4495 if (!req->result && !READ_ONCE(poll->canceled)) {
4496 add_wait_queue(poll->head, &poll->wait);
4503 static void io_poll_remove_double(struct io_kiocb *req)
4505 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4507 lockdep_assert_held(&req->ctx->completion_lock);
4509 if (poll && poll->head) {
4510 struct wait_queue_head *head = poll->head;
4512 spin_lock(&head->lock);
4513 list_del_init(&poll->wait.entry);
4514 if (poll->wait.private)
4515 refcount_dec(&req->refs);
4517 spin_unlock(&head->lock);
4521 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4523 struct io_ring_ctx *ctx = req->ctx;
4525 io_poll_remove_double(req);
4526 req->poll.done = true;
4527 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4528 io_commit_cqring(ctx);
4531 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4533 struct io_ring_ctx *ctx = req->ctx;
4535 if (io_poll_rewait(req, &req->poll)) {
4536 spin_unlock_irq(&ctx->completion_lock);
4540 hash_del(&req->hash_node);
4541 io_poll_complete(req, req->result, 0);
4542 req->flags |= REQ_F_COMP_LOCKED;
4543 *nxt = io_put_req_find_next(req);
4544 spin_unlock_irq(&ctx->completion_lock);
4546 io_cqring_ev_posted(ctx);
4549 static void io_poll_task_func(struct callback_head *cb)
4551 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4552 struct io_kiocb *nxt = NULL;
4554 io_poll_task_handler(req, &nxt);
4556 __io_req_task_submit(nxt);
4559 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4560 int sync, void *key)
4562 struct io_kiocb *req = wait->private;
4563 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4564 __poll_t mask = key_to_poll(key);
4566 /* for instances that support it check for an event match first: */
4567 if (mask && !(mask & poll->events))
4570 if (req->poll.head) {
4573 spin_lock(&req->poll.head->lock);
4574 done = list_empty(&req->poll.wait.entry);
4576 list_del_init(&req->poll.wait.entry);
4577 spin_unlock(&req->poll.head->lock);
4579 __io_async_wake(req, poll, mask, io_poll_task_func);
4581 refcount_dec(&req->refs);
4585 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4586 wait_queue_func_t wake_func)
4590 poll->canceled = false;
4591 poll->events = events;
4592 INIT_LIST_HEAD(&poll->wait.entry);
4593 init_waitqueue_func_entry(&poll->wait, wake_func);
4596 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4597 struct wait_queue_head *head)
4599 struct io_kiocb *req = pt->req;
4602 * If poll->head is already set, it's because the file being polled
4603 * uses multiple waitqueues for poll handling (eg one for read, one
4604 * for write). Setup a separate io_poll_iocb if this happens.
4606 if (unlikely(poll->head)) {
4607 /* already have a 2nd entry, fail a third attempt */
4609 pt->error = -EINVAL;
4612 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4614 pt->error = -ENOMEM;
4617 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4618 refcount_inc(&req->refs);
4619 poll->wait.private = req;
4620 req->io = (void *) poll;
4626 if (poll->events & EPOLLEXCLUSIVE)
4627 add_wait_queue_exclusive(head, &poll->wait);
4629 add_wait_queue(head, &poll->wait);
4632 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4633 struct poll_table_struct *p)
4635 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4637 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4640 static void io_async_task_func(struct callback_head *cb)
4642 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4643 struct async_poll *apoll = req->apoll;
4644 struct io_ring_ctx *ctx = req->ctx;
4646 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4648 if (io_poll_rewait(req, &apoll->poll)) {
4649 spin_unlock_irq(&ctx->completion_lock);
4653 /* If req is still hashed, it cannot have been canceled. Don't check. */
4654 if (hash_hashed(&req->hash_node))
4655 hash_del(&req->hash_node);
4657 spin_unlock_irq(&ctx->completion_lock);
4659 /* restore ->work in case we need to retry again */
4660 if (req->flags & REQ_F_WORK_INITIALIZED)
4661 memcpy(&req->work, &apoll->work, sizeof(req->work));
4664 if (!READ_ONCE(apoll->poll.canceled))
4665 __io_req_task_submit(req);
4667 __io_req_task_cancel(req, -ECANCELED);
4670 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4673 struct io_kiocb *req = wait->private;
4674 struct io_poll_iocb *poll = &req->apoll->poll;
4676 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4679 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4682 static void io_poll_req_insert(struct io_kiocb *req)
4684 struct io_ring_ctx *ctx = req->ctx;
4685 struct hlist_head *list;
4687 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4688 hlist_add_head(&req->hash_node, list);
4691 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4692 struct io_poll_iocb *poll,
4693 struct io_poll_table *ipt, __poll_t mask,
4694 wait_queue_func_t wake_func)
4695 __acquires(&ctx->completion_lock)
4697 struct io_ring_ctx *ctx = req->ctx;
4698 bool cancel = false;
4700 io_init_poll_iocb(poll, mask, wake_func);
4701 poll->file = req->file;
4702 poll->wait.private = req;
4704 ipt->pt._key = mask;
4706 ipt->error = -EINVAL;
4708 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4710 spin_lock_irq(&ctx->completion_lock);
4711 if (likely(poll->head)) {
4712 spin_lock(&poll->head->lock);
4713 if (unlikely(list_empty(&poll->wait.entry))) {
4719 if (mask || ipt->error)
4720 list_del_init(&poll->wait.entry);
4722 WRITE_ONCE(poll->canceled, true);
4723 else if (!poll->done) /* actually waiting for an event */
4724 io_poll_req_insert(req);
4725 spin_unlock(&poll->head->lock);
4731 static bool io_arm_poll_handler(struct io_kiocb *req)
4733 const struct io_op_def *def = &io_op_defs[req->opcode];
4734 struct io_ring_ctx *ctx = req->ctx;
4735 struct async_poll *apoll;
4736 struct io_poll_table ipt;
4740 if (!req->file || !file_can_poll(req->file))
4742 if (req->flags & REQ_F_POLLED)
4744 if (!def->pollin && !def->pollout)
4747 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4748 if (unlikely(!apoll))
4751 req->flags |= REQ_F_POLLED;
4752 if (req->flags & REQ_F_WORK_INITIALIZED)
4753 memcpy(&apoll->work, &req->work, sizeof(req->work));
4754 had_io = req->io != NULL;
4756 io_get_req_task(req);
4758 INIT_HLIST_NODE(&req->hash_node);
4762 mask |= POLLIN | POLLRDNORM;
4764 mask |= POLLOUT | POLLWRNORM;
4765 mask |= POLLERR | POLLPRI;
4767 ipt.pt._qproc = io_async_queue_proc;
4769 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4773 /* only remove double add if we did it here */
4775 io_poll_remove_double(req);
4776 spin_unlock_irq(&ctx->completion_lock);
4777 if (req->flags & REQ_F_WORK_INITIALIZED)
4778 memcpy(&req->work, &apoll->work, sizeof(req->work));
4782 spin_unlock_irq(&ctx->completion_lock);
4783 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4784 apoll->poll.events);
4788 static bool __io_poll_remove_one(struct io_kiocb *req,
4789 struct io_poll_iocb *poll)
4791 bool do_complete = false;
4793 spin_lock(&poll->head->lock);
4794 WRITE_ONCE(poll->canceled, true);
4795 if (!list_empty(&poll->wait.entry)) {
4796 list_del_init(&poll->wait.entry);
4799 spin_unlock(&poll->head->lock);
4800 hash_del(&req->hash_node);
4804 static bool io_poll_remove_one(struct io_kiocb *req)
4808 if (req->opcode == IORING_OP_POLL_ADD) {
4809 io_poll_remove_double(req);
4810 do_complete = __io_poll_remove_one(req, &req->poll);
4812 struct async_poll *apoll = req->apoll;
4814 /* non-poll requests have submit ref still */
4815 do_complete = __io_poll_remove_one(req, &apoll->poll);
4819 * restore ->work because we will call
4820 * io_req_work_drop_env below when dropping the
4823 if (req->flags & REQ_F_WORK_INITIALIZED)
4824 memcpy(&req->work, &apoll->work,
4831 io_cqring_fill_event(req, -ECANCELED);
4832 io_commit_cqring(req->ctx);
4833 req->flags |= REQ_F_COMP_LOCKED;
4840 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4842 struct hlist_node *tmp;
4843 struct io_kiocb *req;
4846 spin_lock_irq(&ctx->completion_lock);
4847 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4848 struct hlist_head *list;
4850 list = &ctx->cancel_hash[i];
4851 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4852 posted += io_poll_remove_one(req);
4854 spin_unlock_irq(&ctx->completion_lock);
4857 io_cqring_ev_posted(ctx);
4860 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4862 struct hlist_head *list;
4863 struct io_kiocb *req;
4865 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4866 hlist_for_each_entry(req, list, hash_node) {
4867 if (sqe_addr != req->user_data)
4869 if (io_poll_remove_one(req))
4877 static int io_poll_remove_prep(struct io_kiocb *req,
4878 const struct io_uring_sqe *sqe)
4880 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4882 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4886 req->poll.addr = READ_ONCE(sqe->addr);
4891 * Find a running poll command that matches one specified in sqe->addr,
4892 * and remove it if found.
4894 static int io_poll_remove(struct io_kiocb *req)
4896 struct io_ring_ctx *ctx = req->ctx;
4900 addr = req->poll.addr;
4901 spin_lock_irq(&ctx->completion_lock);
4902 ret = io_poll_cancel(ctx, addr);
4903 spin_unlock_irq(&ctx->completion_lock);
4906 req_set_fail_links(req);
4907 io_req_complete(req, ret);
4911 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4914 struct io_kiocb *req = wait->private;
4915 struct io_poll_iocb *poll = &req->poll;
4917 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4920 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4921 struct poll_table_struct *p)
4923 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4925 __io_queue_proc(&pt->req->poll, pt, head);
4928 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4930 struct io_poll_iocb *poll = &req->poll;
4933 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4935 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4940 events = READ_ONCE(sqe->poll32_events);
4942 events = swahw32(events);
4944 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4945 (events & EPOLLEXCLUSIVE);
4947 io_get_req_task(req);
4951 static int io_poll_add(struct io_kiocb *req)
4953 struct io_poll_iocb *poll = &req->poll;
4954 struct io_ring_ctx *ctx = req->ctx;
4955 struct io_poll_table ipt;
4958 INIT_HLIST_NODE(&req->hash_node);
4959 INIT_LIST_HEAD(&req->list);
4960 ipt.pt._qproc = io_poll_queue_proc;
4962 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4965 if (mask) { /* no async, we'd stolen it */
4967 io_poll_complete(req, mask, 0);
4969 spin_unlock_irq(&ctx->completion_lock);
4972 io_cqring_ev_posted(ctx);
4978 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4980 struct io_timeout_data *data = container_of(timer,
4981 struct io_timeout_data, timer);
4982 struct io_kiocb *req = data->req;
4983 struct io_ring_ctx *ctx = req->ctx;
4984 unsigned long flags;
4986 atomic_inc(&ctx->cq_timeouts);
4988 spin_lock_irqsave(&ctx->completion_lock, flags);
4990 * We could be racing with timeout deletion. If the list is empty,
4991 * then timeout lookup already found it and will be handling it.
4993 if (!list_empty(&req->list))
4994 list_del_init(&req->list);
4996 io_cqring_fill_event(req, -ETIME);
4997 io_commit_cqring(ctx);
4998 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5000 io_cqring_ev_posted(ctx);
5001 req_set_fail_links(req);
5003 return HRTIMER_NORESTART;
5006 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5008 struct io_kiocb *req;
5011 list_for_each_entry(req, &ctx->timeout_list, list) {
5012 if (user_data == req->user_data) {
5013 list_del_init(&req->list);
5022 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5026 req_set_fail_links(req);
5027 io_cqring_fill_event(req, -ECANCELED);
5032 static int io_timeout_remove_prep(struct io_kiocb *req,
5033 const struct io_uring_sqe *sqe)
5035 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5037 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
5040 req->timeout.addr = READ_ONCE(sqe->addr);
5041 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5042 if (req->timeout.flags)
5049 * Remove or update an existing timeout command
5051 static int io_timeout_remove(struct io_kiocb *req)
5053 struct io_ring_ctx *ctx = req->ctx;
5056 spin_lock_irq(&ctx->completion_lock);
5057 ret = io_timeout_cancel(ctx, req->timeout.addr);
5059 io_cqring_fill_event(req, ret);
5060 io_commit_cqring(ctx);
5061 spin_unlock_irq(&ctx->completion_lock);
5062 io_cqring_ev_posted(ctx);
5064 req_set_fail_links(req);
5069 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5070 bool is_timeout_link)
5072 struct io_timeout_data *data;
5074 u32 off = READ_ONCE(sqe->off);
5076 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5078 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5080 if (off && is_timeout_link)
5082 flags = READ_ONCE(sqe->timeout_flags);
5083 if (flags & ~IORING_TIMEOUT_ABS)
5086 req->timeout.off = off;
5088 if (!req->io && io_alloc_async_ctx(req))
5091 data = &req->io->timeout;
5094 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5097 if (flags & IORING_TIMEOUT_ABS)
5098 data->mode = HRTIMER_MODE_ABS;
5100 data->mode = HRTIMER_MODE_REL;
5102 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5106 static int io_timeout(struct io_kiocb *req)
5108 struct io_ring_ctx *ctx = req->ctx;
5109 struct io_timeout_data *data = &req->io->timeout;
5110 struct list_head *entry;
5111 u32 tail, off = req->timeout.off;
5113 spin_lock_irq(&ctx->completion_lock);
5116 * sqe->off holds how many events that need to occur for this
5117 * timeout event to be satisfied. If it isn't set, then this is
5118 * a pure timeout request, sequence isn't used.
5120 if (io_is_timeout_noseq(req)) {
5121 entry = ctx->timeout_list.prev;
5125 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5126 req->timeout.target_seq = tail + off;
5129 * Insertion sort, ensuring the first entry in the list is always
5130 * the one we need first.
5132 list_for_each_prev(entry, &ctx->timeout_list) {
5133 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
5135 if (io_is_timeout_noseq(nxt))
5137 /* nxt.seq is behind @tail, otherwise would've been completed */
5138 if (off >= nxt->timeout.target_seq - tail)
5142 list_add(&req->list, entry);
5143 data->timer.function = io_timeout_fn;
5144 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5145 spin_unlock_irq(&ctx->completion_lock);
5149 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5151 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5153 return req->user_data == (unsigned long) data;
5156 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5158 enum io_wq_cancel cancel_ret;
5161 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5162 switch (cancel_ret) {
5163 case IO_WQ_CANCEL_OK:
5166 case IO_WQ_CANCEL_RUNNING:
5169 case IO_WQ_CANCEL_NOTFOUND:
5177 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5178 struct io_kiocb *req, __u64 sqe_addr,
5181 unsigned long flags;
5184 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5185 if (ret != -ENOENT) {
5186 spin_lock_irqsave(&ctx->completion_lock, flags);
5190 spin_lock_irqsave(&ctx->completion_lock, flags);
5191 ret = io_timeout_cancel(ctx, sqe_addr);
5194 ret = io_poll_cancel(ctx, sqe_addr);
5198 io_cqring_fill_event(req, ret);
5199 io_commit_cqring(ctx);
5200 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5201 io_cqring_ev_posted(ctx);
5204 req_set_fail_links(req);
5208 static int io_async_cancel_prep(struct io_kiocb *req,
5209 const struct io_uring_sqe *sqe)
5211 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5213 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
5217 req->cancel.addr = READ_ONCE(sqe->addr);
5221 static int io_async_cancel(struct io_kiocb *req)
5223 struct io_ring_ctx *ctx = req->ctx;
5225 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5229 static int io_files_update_prep(struct io_kiocb *req,
5230 const struct io_uring_sqe *sqe)
5232 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
5235 req->files_update.offset = READ_ONCE(sqe->off);
5236 req->files_update.nr_args = READ_ONCE(sqe->len);
5237 if (!req->files_update.nr_args)
5239 req->files_update.arg = READ_ONCE(sqe->addr);
5243 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5244 struct io_comp_state *cs)
5246 struct io_ring_ctx *ctx = req->ctx;
5247 struct io_uring_files_update up;
5253 up.offset = req->files_update.offset;
5254 up.fds = req->files_update.arg;
5256 mutex_lock(&ctx->uring_lock);
5257 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5258 mutex_unlock(&ctx->uring_lock);
5261 req_set_fail_links(req);
5262 __io_req_complete(req, ret, 0, cs);
5266 static int io_req_defer_prep(struct io_kiocb *req,
5267 const struct io_uring_sqe *sqe)
5274 if (io_op_defs[req->opcode].file_table) {
5275 io_req_init_async(req);
5276 ret = io_grab_files(req);
5281 switch (req->opcode) {
5284 case IORING_OP_READV:
5285 case IORING_OP_READ_FIXED:
5286 case IORING_OP_READ:
5287 ret = io_read_prep(req, sqe, true);
5289 case IORING_OP_WRITEV:
5290 case IORING_OP_WRITE_FIXED:
5291 case IORING_OP_WRITE:
5292 ret = io_write_prep(req, sqe, true);
5294 case IORING_OP_POLL_ADD:
5295 ret = io_poll_add_prep(req, sqe);
5297 case IORING_OP_POLL_REMOVE:
5298 ret = io_poll_remove_prep(req, sqe);
5300 case IORING_OP_FSYNC:
5301 ret = io_prep_fsync(req, sqe);
5303 case IORING_OP_SYNC_FILE_RANGE:
5304 ret = io_prep_sfr(req, sqe);
5306 case IORING_OP_SENDMSG:
5307 case IORING_OP_SEND:
5308 ret = io_sendmsg_prep(req, sqe);
5310 case IORING_OP_RECVMSG:
5311 case IORING_OP_RECV:
5312 ret = io_recvmsg_prep(req, sqe);
5314 case IORING_OP_CONNECT:
5315 ret = io_connect_prep(req, sqe);
5317 case IORING_OP_TIMEOUT:
5318 ret = io_timeout_prep(req, sqe, false);
5320 case IORING_OP_TIMEOUT_REMOVE:
5321 ret = io_timeout_remove_prep(req, sqe);
5323 case IORING_OP_ASYNC_CANCEL:
5324 ret = io_async_cancel_prep(req, sqe);
5326 case IORING_OP_LINK_TIMEOUT:
5327 ret = io_timeout_prep(req, sqe, true);
5329 case IORING_OP_ACCEPT:
5330 ret = io_accept_prep(req, sqe);
5332 case IORING_OP_FALLOCATE:
5333 ret = io_fallocate_prep(req, sqe);
5335 case IORING_OP_OPENAT:
5336 ret = io_openat_prep(req, sqe);
5338 case IORING_OP_CLOSE:
5339 ret = io_close_prep(req, sqe);
5341 case IORING_OP_FILES_UPDATE:
5342 ret = io_files_update_prep(req, sqe);
5344 case IORING_OP_STATX:
5345 ret = io_statx_prep(req, sqe);
5347 case IORING_OP_FADVISE:
5348 ret = io_fadvise_prep(req, sqe);
5350 case IORING_OP_MADVISE:
5351 ret = io_madvise_prep(req, sqe);
5353 case IORING_OP_OPENAT2:
5354 ret = io_openat2_prep(req, sqe);
5356 case IORING_OP_EPOLL_CTL:
5357 ret = io_epoll_ctl_prep(req, sqe);
5359 case IORING_OP_SPLICE:
5360 ret = io_splice_prep(req, sqe);
5362 case IORING_OP_PROVIDE_BUFFERS:
5363 ret = io_provide_buffers_prep(req, sqe);
5365 case IORING_OP_REMOVE_BUFFERS:
5366 ret = io_remove_buffers_prep(req, sqe);
5369 ret = io_tee_prep(req, sqe);
5372 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5381 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5383 struct io_ring_ctx *ctx = req->ctx;
5386 /* Still need defer if there is pending req in defer list. */
5387 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5391 if (io_alloc_async_ctx(req))
5393 ret = io_req_defer_prep(req, sqe);
5397 io_prep_async_link(req);
5399 spin_lock_irq(&ctx->completion_lock);
5400 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5401 spin_unlock_irq(&ctx->completion_lock);
5405 trace_io_uring_defer(ctx, req, req->user_data);
5406 list_add_tail(&req->list, &ctx->defer_list);
5407 spin_unlock_irq(&ctx->completion_lock);
5408 return -EIOCBQUEUED;
5411 static void io_cleanup_req(struct io_kiocb *req)
5413 struct io_async_ctx *io = req->io;
5415 switch (req->opcode) {
5416 case IORING_OP_READV:
5417 case IORING_OP_READ_FIXED:
5418 case IORING_OP_READ:
5419 if (req->flags & REQ_F_BUFFER_SELECTED)
5420 kfree((void *)(unsigned long)req->rw.addr);
5422 case IORING_OP_WRITEV:
5423 case IORING_OP_WRITE_FIXED:
5424 case IORING_OP_WRITE:
5425 if (io->rw.iov != io->rw.fast_iov)
5428 case IORING_OP_RECVMSG:
5429 if (req->flags & REQ_F_BUFFER_SELECTED)
5430 kfree(req->sr_msg.kbuf);
5432 case IORING_OP_SENDMSG:
5433 if (io->msg.iov != io->msg.fast_iov)
5436 case IORING_OP_RECV:
5437 if (req->flags & REQ_F_BUFFER_SELECTED)
5438 kfree(req->sr_msg.kbuf);
5440 case IORING_OP_OPENAT:
5441 case IORING_OP_OPENAT2:
5443 case IORING_OP_SPLICE:
5445 io_put_file(req, req->splice.file_in,
5446 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5450 req->flags &= ~REQ_F_NEED_CLEANUP;
5453 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5454 bool force_nonblock, struct io_comp_state *cs)
5456 struct io_ring_ctx *ctx = req->ctx;
5459 switch (req->opcode) {
5461 ret = io_nop(req, cs);
5463 case IORING_OP_READV:
5464 case IORING_OP_READ_FIXED:
5465 case IORING_OP_READ:
5467 ret = io_read_prep(req, sqe, force_nonblock);
5471 ret = io_read(req, force_nonblock, cs);
5473 case IORING_OP_WRITEV:
5474 case IORING_OP_WRITE_FIXED:
5475 case IORING_OP_WRITE:
5477 ret = io_write_prep(req, sqe, force_nonblock);
5481 ret = io_write(req, force_nonblock, cs);
5483 case IORING_OP_FSYNC:
5485 ret = io_prep_fsync(req, sqe);
5489 ret = io_fsync(req, force_nonblock);
5491 case IORING_OP_POLL_ADD:
5493 ret = io_poll_add_prep(req, sqe);
5497 ret = io_poll_add(req);
5499 case IORING_OP_POLL_REMOVE:
5501 ret = io_poll_remove_prep(req, sqe);
5505 ret = io_poll_remove(req);
5507 case IORING_OP_SYNC_FILE_RANGE:
5509 ret = io_prep_sfr(req, sqe);
5513 ret = io_sync_file_range(req, force_nonblock);
5515 case IORING_OP_SENDMSG:
5516 case IORING_OP_SEND:
5518 ret = io_sendmsg_prep(req, sqe);
5522 if (req->opcode == IORING_OP_SENDMSG)
5523 ret = io_sendmsg(req, force_nonblock, cs);
5525 ret = io_send(req, force_nonblock, cs);
5527 case IORING_OP_RECVMSG:
5528 case IORING_OP_RECV:
5530 ret = io_recvmsg_prep(req, sqe);
5534 if (req->opcode == IORING_OP_RECVMSG)
5535 ret = io_recvmsg(req, force_nonblock, cs);
5537 ret = io_recv(req, force_nonblock, cs);
5539 case IORING_OP_TIMEOUT:
5541 ret = io_timeout_prep(req, sqe, false);
5545 ret = io_timeout(req);
5547 case IORING_OP_TIMEOUT_REMOVE:
5549 ret = io_timeout_remove_prep(req, sqe);
5553 ret = io_timeout_remove(req);
5555 case IORING_OP_ACCEPT:
5557 ret = io_accept_prep(req, sqe);
5561 ret = io_accept(req, force_nonblock, cs);
5563 case IORING_OP_CONNECT:
5565 ret = io_connect_prep(req, sqe);
5569 ret = io_connect(req, force_nonblock, cs);
5571 case IORING_OP_ASYNC_CANCEL:
5573 ret = io_async_cancel_prep(req, sqe);
5577 ret = io_async_cancel(req);
5579 case IORING_OP_FALLOCATE:
5581 ret = io_fallocate_prep(req, sqe);
5585 ret = io_fallocate(req, force_nonblock);
5587 case IORING_OP_OPENAT:
5589 ret = io_openat_prep(req, sqe);
5593 ret = io_openat(req, force_nonblock);
5595 case IORING_OP_CLOSE:
5597 ret = io_close_prep(req, sqe);
5601 ret = io_close(req, force_nonblock, cs);
5603 case IORING_OP_FILES_UPDATE:
5605 ret = io_files_update_prep(req, sqe);
5609 ret = io_files_update(req, force_nonblock, cs);
5611 case IORING_OP_STATX:
5613 ret = io_statx_prep(req, sqe);
5617 ret = io_statx(req, force_nonblock);
5619 case IORING_OP_FADVISE:
5621 ret = io_fadvise_prep(req, sqe);
5625 ret = io_fadvise(req, force_nonblock);
5627 case IORING_OP_MADVISE:
5629 ret = io_madvise_prep(req, sqe);
5633 ret = io_madvise(req, force_nonblock);
5635 case IORING_OP_OPENAT2:
5637 ret = io_openat2_prep(req, sqe);
5641 ret = io_openat2(req, force_nonblock);
5643 case IORING_OP_EPOLL_CTL:
5645 ret = io_epoll_ctl_prep(req, sqe);
5649 ret = io_epoll_ctl(req, force_nonblock, cs);
5651 case IORING_OP_SPLICE:
5653 ret = io_splice_prep(req, sqe);
5657 ret = io_splice(req, force_nonblock);
5659 case IORING_OP_PROVIDE_BUFFERS:
5661 ret = io_provide_buffers_prep(req, sqe);
5665 ret = io_provide_buffers(req, force_nonblock, cs);
5667 case IORING_OP_REMOVE_BUFFERS:
5669 ret = io_remove_buffers_prep(req, sqe);
5673 ret = io_remove_buffers(req, force_nonblock, cs);
5677 ret = io_tee_prep(req, sqe);
5681 ret = io_tee(req, force_nonblock);
5691 /* If the op doesn't have a file, we're not polling for it */
5692 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5693 const bool in_async = io_wq_current_is_worker();
5695 /* workqueue context doesn't hold uring_lock, grab it now */
5697 mutex_lock(&ctx->uring_lock);
5699 io_iopoll_req_issued(req);
5702 mutex_unlock(&ctx->uring_lock);
5708 static void io_arm_async_linked_timeout(struct io_kiocb *req)
5710 struct io_kiocb *link;
5712 /* link head's timeout is queued in io_queue_async_work() */
5713 if (!(req->flags & REQ_F_QUEUE_TIMEOUT))
5716 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
5717 io_queue_linked_timeout(link);
5720 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5722 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5725 io_arm_async_linked_timeout(req);
5727 /* if NO_CANCEL is set, we must still run the work */
5728 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5729 IO_WQ_WORK_CANCEL) {
5735 ret = io_issue_sqe(req, NULL, false, NULL);
5737 * We can get EAGAIN for polled IO even though we're
5738 * forcing a sync submission from here, since we can't
5739 * wait for request slots on the block side.
5748 req_set_fail_links(req);
5749 io_req_complete(req, ret);
5752 return io_steal_work(req);
5755 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5758 struct fixed_file_table *table;
5760 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5761 return table->files[index & IORING_FILE_TABLE_MASK];
5764 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5765 int fd, struct file **out_file, bool fixed)
5767 struct io_ring_ctx *ctx = req->ctx;
5771 if (unlikely(!ctx->file_data ||
5772 (unsigned) fd >= ctx->nr_user_files))
5774 fd = array_index_nospec(fd, ctx->nr_user_files);
5775 file = io_file_from_index(ctx, fd);
5777 req->fixed_file_refs = ctx->file_data->cur_refs;
5778 percpu_ref_get(req->fixed_file_refs);
5781 trace_io_uring_file_get(ctx, fd);
5782 file = __io_file_get(state, fd);
5785 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5792 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5797 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5798 if (unlikely(!fixed && io_async_submit(req->ctx)))
5801 return io_file_get(state, req, fd, &req->file, fixed);
5804 static int io_grab_files(struct io_kiocb *req)
5807 struct io_ring_ctx *ctx = req->ctx;
5809 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5811 if (!ctx->ring_file)
5815 spin_lock_irq(&ctx->inflight_lock);
5817 * We use the f_ops->flush() handler to ensure that we can flush
5818 * out work accessing these files if the fd is closed. Check if
5819 * the fd has changed since we started down this path, and disallow
5820 * this operation if it has.
5822 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5823 list_add(&req->inflight_entry, &ctx->inflight_list);
5824 req->flags |= REQ_F_INFLIGHT;
5825 req->work.files = current->files;
5828 spin_unlock_irq(&ctx->inflight_lock);
5834 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5836 struct io_timeout_data *data = container_of(timer,
5837 struct io_timeout_data, timer);
5838 struct io_kiocb *req = data->req;
5839 struct io_ring_ctx *ctx = req->ctx;
5840 struct io_kiocb *prev = NULL;
5841 unsigned long flags;
5843 spin_lock_irqsave(&ctx->completion_lock, flags);
5846 * We don't expect the list to be empty, that will only happen if we
5847 * race with the completion of the linked work.
5849 if (!list_empty(&req->link_list)) {
5850 prev = list_entry(req->link_list.prev, struct io_kiocb,
5852 if (refcount_inc_not_zero(&prev->refs)) {
5853 list_del_init(&req->link_list);
5854 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5859 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5862 req_set_fail_links(prev);
5863 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5866 io_req_complete(req, -ETIME);
5868 return HRTIMER_NORESTART;
5871 static void io_queue_linked_timeout(struct io_kiocb *req)
5873 struct io_ring_ctx *ctx = req->ctx;
5876 * If the list is now empty, then our linked request finished before
5877 * we got a chance to setup the timer
5879 spin_lock_irq(&ctx->completion_lock);
5880 if (!list_empty(&req->link_list)) {
5881 struct io_timeout_data *data = &req->io->timeout;
5883 data->timer.function = io_link_timeout_fn;
5884 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5887 spin_unlock_irq(&ctx->completion_lock);
5889 /* drop submission reference */
5893 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5895 struct io_kiocb *nxt;
5897 if (!(req->flags & REQ_F_LINK_HEAD))
5899 /* for polled retry, if flag is set, we already went through here */
5900 if (req->flags & REQ_F_POLLED)
5903 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5905 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5908 req->flags |= REQ_F_LINK_TIMEOUT;
5912 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5913 struct io_comp_state *cs)
5915 struct io_kiocb *linked_timeout;
5916 struct io_kiocb *nxt;
5917 const struct cred *old_creds = NULL;
5921 linked_timeout = io_prep_linked_timeout(req);
5923 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5924 req->work.creds != current_cred()) {
5926 revert_creds(old_creds);
5927 if (old_creds == req->work.creds)
5928 old_creds = NULL; /* restored original creds */
5930 old_creds = override_creds(req->work.creds);
5933 ret = io_issue_sqe(req, sqe, true, cs);
5936 * We async punt it if the file wasn't marked NOWAIT, or if the file
5937 * doesn't support non-blocking read/write attempts
5939 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5940 if (io_arm_poll_handler(req)) {
5942 io_queue_linked_timeout(linked_timeout);
5946 io_req_init_async(req);
5948 if (io_op_defs[req->opcode].file_table) {
5949 ret = io_grab_files(req);
5955 * Queued up for async execution, worker will release
5956 * submit reference when the iocb is actually submitted.
5958 io_queue_async_work(req);
5963 /* drop submission reference */
5964 nxt = io_put_req_find_next(req);
5966 if (linked_timeout) {
5968 io_queue_linked_timeout(linked_timeout);
5970 io_put_req(linked_timeout);
5973 /* and drop final reference, if we failed */
5975 req_set_fail_links(req);
5976 io_req_complete(req, ret);
5981 if (req->flags & REQ_F_FORCE_ASYNC)
5987 revert_creds(old_creds);
5990 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5991 struct io_comp_state *cs)
5995 ret = io_req_defer(req, sqe);
5997 if (ret != -EIOCBQUEUED) {
5999 req_set_fail_links(req);
6001 io_req_complete(req, ret);
6003 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6006 if (io_alloc_async_ctx(req))
6008 ret = io_req_defer_prep(req, sqe);
6009 if (unlikely(ret < 0))
6014 * Never try inline submit of IOSQE_ASYNC is set, go straight
6015 * to async execution.
6017 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6018 io_queue_async_work(req);
6020 __io_queue_sqe(req, sqe, cs);
6024 static inline void io_queue_link_head(struct io_kiocb *req,
6025 struct io_comp_state *cs)
6027 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6029 io_req_complete(req, -ECANCELED);
6031 io_queue_sqe(req, NULL, cs);
6034 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6035 struct io_kiocb **link, struct io_comp_state *cs)
6037 struct io_ring_ctx *ctx = req->ctx;
6041 * If we already have a head request, queue this one for async
6042 * submittal once the head completes. If we don't have a head but
6043 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6044 * submitted sync once the chain is complete. If none of those
6045 * conditions are true (normal request), then just queue it.
6048 struct io_kiocb *head = *link;
6051 * Taking sequential execution of a link, draining both sides
6052 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6053 * requests in the link. So, it drains the head and the
6054 * next after the link request. The last one is done via
6055 * drain_next flag to persist the effect across calls.
6057 if (req->flags & REQ_F_IO_DRAIN) {
6058 head->flags |= REQ_F_IO_DRAIN;
6059 ctx->drain_next = 1;
6061 if (io_alloc_async_ctx(req))
6064 ret = io_req_defer_prep(req, sqe);
6066 /* fail even hard links since we don't submit */
6067 head->flags |= REQ_F_FAIL_LINK;
6070 trace_io_uring_link(ctx, req, head);
6071 io_get_req_task(req);
6072 list_add_tail(&req->link_list, &head->link_list);
6074 /* last request of a link, enqueue the link */
6075 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6076 io_queue_link_head(head, cs);
6080 if (unlikely(ctx->drain_next)) {
6081 req->flags |= REQ_F_IO_DRAIN;
6082 ctx->drain_next = 0;
6084 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6085 req->flags |= REQ_F_LINK_HEAD;
6086 INIT_LIST_HEAD(&req->link_list);
6088 if (io_alloc_async_ctx(req))
6091 ret = io_req_defer_prep(req, sqe);
6093 req->flags |= REQ_F_FAIL_LINK;
6096 io_queue_sqe(req, sqe, cs);
6104 * Batched submission is done, ensure local IO is flushed out.
6106 static void io_submit_state_end(struct io_submit_state *state)
6108 if (!list_empty(&state->comp.list))
6109 io_submit_flush_completions(&state->comp);
6110 blk_finish_plug(&state->plug);
6111 io_state_file_put(state);
6112 if (state->free_reqs)
6113 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6117 * Start submission side cache.
6119 static void io_submit_state_start(struct io_submit_state *state,
6120 struct io_ring_ctx *ctx, unsigned int max_ios)
6122 blk_start_plug(&state->plug);
6124 state->plug.nowait = true;
6127 INIT_LIST_HEAD(&state->comp.list);
6128 state->comp.ctx = ctx;
6129 state->free_reqs = 0;
6131 state->ios_left = max_ios;
6134 static void io_commit_sqring(struct io_ring_ctx *ctx)
6136 struct io_rings *rings = ctx->rings;
6139 * Ensure any loads from the SQEs are done at this point,
6140 * since once we write the new head, the application could
6141 * write new data to them.
6143 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6147 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6148 * that is mapped by userspace. This means that care needs to be taken to
6149 * ensure that reads are stable, as we cannot rely on userspace always
6150 * being a good citizen. If members of the sqe are validated and then later
6151 * used, it's important that those reads are done through READ_ONCE() to
6152 * prevent a re-load down the line.
6154 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6156 u32 *sq_array = ctx->sq_array;
6160 * The cached sq head (or cq tail) serves two purposes:
6162 * 1) allows us to batch the cost of updating the user visible
6164 * 2) allows the kernel side to track the head on its own, even
6165 * though the application is the one updating it.
6167 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6168 if (likely(head < ctx->sq_entries))
6169 return &ctx->sq_sqes[head];
6171 /* drop invalid entries */
6172 ctx->cached_sq_dropped++;
6173 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6177 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6179 ctx->cached_sq_head++;
6182 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6183 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6184 IOSQE_BUFFER_SELECT)
6186 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6187 const struct io_uring_sqe *sqe,
6188 struct io_submit_state *state)
6190 unsigned int sqe_flags;
6194 * All io need record the previous position, if LINK vs DARIN,
6195 * it can be used to mark the position of the first IO in the
6198 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
6199 req->opcode = READ_ONCE(sqe->opcode);
6200 req->user_data = READ_ONCE(sqe->user_data);
6205 /* one is dropped after submission, the other at completion */
6206 refcount_set(&req->refs, 2);
6207 req->task = current;
6210 if (unlikely(req->opcode >= IORING_OP_LAST))
6213 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6216 sqe_flags = READ_ONCE(sqe->flags);
6217 /* enforce forwards compatibility on users */
6218 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6221 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6222 !io_op_defs[req->opcode].buffer_select)
6225 id = READ_ONCE(sqe->personality);
6227 io_req_init_async(req);
6228 req->work.creds = idr_find(&ctx->personality_idr, id);
6229 if (unlikely(!req->work.creds))
6231 get_cred(req->work.creds);
6234 /* same numerical values with corresponding REQ_F_*, safe to copy */
6235 req->flags |= sqe_flags;
6237 if (!io_op_defs[req->opcode].needs_file)
6240 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6243 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6244 struct file *ring_file, int ring_fd)
6246 struct io_submit_state state;
6247 struct io_kiocb *link = NULL;
6248 int i, submitted = 0;
6250 /* if we have a backlog and couldn't flush it all, return BUSY */
6251 if (test_bit(0, &ctx->sq_check_overflow)) {
6252 if (!list_empty(&ctx->cq_overflow_list) &&
6253 !io_cqring_overflow_flush(ctx, false))
6257 /* make sure SQ entry isn't read before tail */
6258 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6260 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6263 io_submit_state_start(&state, ctx, nr);
6265 ctx->ring_fd = ring_fd;
6266 ctx->ring_file = ring_file;
6268 for (i = 0; i < nr; i++) {
6269 const struct io_uring_sqe *sqe;
6270 struct io_kiocb *req;
6273 sqe = io_get_sqe(ctx);
6274 if (unlikely(!sqe)) {
6275 io_consume_sqe(ctx);
6278 req = io_alloc_req(ctx, &state);
6279 if (unlikely(!req)) {
6281 submitted = -EAGAIN;
6285 err = io_init_req(ctx, req, sqe, &state);
6286 io_consume_sqe(ctx);
6287 /* will complete beyond this point, count as submitted */
6290 if (unlikely(err)) {
6293 io_req_complete(req, err);
6297 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6298 true, io_async_submit(ctx));
6299 err = io_submit_sqe(req, sqe, &link, &state.comp);
6304 if (unlikely(submitted != nr)) {
6305 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6307 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6310 io_queue_link_head(link, &state.comp);
6311 io_submit_state_end(&state);
6313 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6314 io_commit_sqring(ctx);
6319 static int io_sq_thread(void *data)
6321 struct io_ring_ctx *ctx = data;
6322 const struct cred *old_cred;
6324 unsigned long timeout;
6327 complete(&ctx->sq_thread_comp);
6329 old_cred = override_creds(ctx->creds);
6331 timeout = jiffies + ctx->sq_thread_idle;
6332 while (!kthread_should_park()) {
6333 unsigned int to_submit;
6335 if (!list_empty(&ctx->poll_list)) {
6336 unsigned nr_events = 0;
6338 mutex_lock(&ctx->uring_lock);
6339 if (!list_empty(&ctx->poll_list))
6340 io_iopoll_getevents(ctx, &nr_events, 0);
6342 timeout = jiffies + ctx->sq_thread_idle;
6343 mutex_unlock(&ctx->uring_lock);
6346 to_submit = io_sqring_entries(ctx);
6349 * If submit got -EBUSY, flag us as needing the application
6350 * to enter the kernel to reap and flush events.
6352 if (!to_submit || ret == -EBUSY || need_resched()) {
6354 * Drop cur_mm before scheduling, we can't hold it for
6355 * long periods (or over schedule()). Do this before
6356 * adding ourselves to the waitqueue, as the unuse/drop
6359 io_sq_thread_drop_mm(ctx);
6362 * We're polling. If we're within the defined idle
6363 * period, then let us spin without work before going
6364 * to sleep. The exception is if we got EBUSY doing
6365 * more IO, we should wait for the application to
6366 * reap events and wake us up.
6368 if (!list_empty(&ctx->poll_list) || need_resched() ||
6369 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6370 !percpu_ref_is_dying(&ctx->refs))) {
6376 prepare_to_wait(&ctx->sqo_wait, &wait,
6377 TASK_INTERRUPTIBLE);
6380 * While doing polled IO, before going to sleep, we need
6381 * to check if there are new reqs added to poll_list, it
6382 * is because reqs may have been punted to io worker and
6383 * will be added to poll_list later, hence check the
6386 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6387 !list_empty_careful(&ctx->poll_list)) {
6388 finish_wait(&ctx->sqo_wait, &wait);
6392 /* Tell userspace we may need a wakeup call */
6393 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6394 /* make sure to read SQ tail after writing flags */
6397 to_submit = io_sqring_entries(ctx);
6398 if (!to_submit || ret == -EBUSY) {
6399 if (kthread_should_park()) {
6400 finish_wait(&ctx->sqo_wait, &wait);
6403 if (io_run_task_work()) {
6404 finish_wait(&ctx->sqo_wait, &wait);
6407 if (signal_pending(current))
6408 flush_signals(current);
6410 finish_wait(&ctx->sqo_wait, &wait);
6412 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6416 finish_wait(&ctx->sqo_wait, &wait);
6418 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6421 mutex_lock(&ctx->uring_lock);
6422 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6423 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6424 mutex_unlock(&ctx->uring_lock);
6425 timeout = jiffies + ctx->sq_thread_idle;
6430 io_sq_thread_drop_mm(ctx);
6431 revert_creds(old_cred);
6438 struct io_wait_queue {
6439 struct wait_queue_entry wq;
6440 struct io_ring_ctx *ctx;
6442 unsigned nr_timeouts;
6445 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6447 struct io_ring_ctx *ctx = iowq->ctx;
6450 * Wake up if we have enough events, or if a timeout occurred since we
6451 * started waiting. For timeouts, we always want to return to userspace,
6452 * regardless of event count.
6454 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6455 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6458 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6459 int wake_flags, void *key)
6461 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6464 /* use noflush == true, as we can't safely rely on locking context */
6465 if (!io_should_wake(iowq, true))
6468 return autoremove_wake_function(curr, mode, wake_flags, key);
6472 * Wait until events become available, if we don't already have some. The
6473 * application must reap them itself, as they reside on the shared cq ring.
6475 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6476 const sigset_t __user *sig, size_t sigsz)
6478 struct io_wait_queue iowq = {
6481 .func = io_wake_function,
6482 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6485 .to_wait = min_events,
6487 struct io_rings *rings = ctx->rings;
6491 if (io_cqring_events(ctx, false) >= min_events)
6493 if (!io_run_task_work())
6498 #ifdef CONFIG_COMPAT
6499 if (in_compat_syscall())
6500 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6504 ret = set_user_sigmask(sig, sigsz);
6510 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6511 trace_io_uring_cqring_wait(ctx, min_events);
6513 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6514 TASK_INTERRUPTIBLE);
6515 /* make sure we run task_work before checking for signals */
6516 if (io_run_task_work())
6518 if (signal_pending(current)) {
6519 if (current->jobctl & JOBCTL_TASK_WORK) {
6520 spin_lock_irq(¤t->sighand->siglock);
6521 current->jobctl &= ~JOBCTL_TASK_WORK;
6522 recalc_sigpending();
6523 spin_unlock_irq(¤t->sighand->siglock);
6529 if (io_should_wake(&iowq, false))
6533 finish_wait(&ctx->wait, &iowq.wq);
6535 restore_saved_sigmask_unless(ret == -EINTR);
6537 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6540 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6542 #if defined(CONFIG_UNIX)
6543 if (ctx->ring_sock) {
6544 struct sock *sock = ctx->ring_sock->sk;
6545 struct sk_buff *skb;
6547 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6553 for (i = 0; i < ctx->nr_user_files; i++) {
6556 file = io_file_from_index(ctx, i);
6563 static void io_file_ref_kill(struct percpu_ref *ref)
6565 struct fixed_file_data *data;
6567 data = container_of(ref, struct fixed_file_data, refs);
6568 complete(&data->done);
6571 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6573 struct fixed_file_data *data = ctx->file_data;
6574 struct fixed_file_ref_node *ref_node = NULL;
6575 unsigned nr_tables, i;
6580 spin_lock(&data->lock);
6581 if (!list_empty(&data->ref_list))
6582 ref_node = list_first_entry(&data->ref_list,
6583 struct fixed_file_ref_node, node);
6584 spin_unlock(&data->lock);
6586 percpu_ref_kill(&ref_node->refs);
6588 percpu_ref_kill(&data->refs);
6590 /* wait for all refs nodes to complete */
6591 flush_delayed_work(&ctx->file_put_work);
6592 wait_for_completion(&data->done);
6594 __io_sqe_files_unregister(ctx);
6595 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6596 for (i = 0; i < nr_tables; i++)
6597 kfree(data->table[i].files);
6599 percpu_ref_exit(&data->refs);
6601 ctx->file_data = NULL;
6602 ctx->nr_user_files = 0;
6606 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6608 if (ctx->sqo_thread) {
6609 wait_for_completion(&ctx->sq_thread_comp);
6611 * The park is a bit of a work-around, without it we get
6612 * warning spews on shutdown with SQPOLL set and affinity
6613 * set to a single CPU.
6615 kthread_park(ctx->sqo_thread);
6616 kthread_stop(ctx->sqo_thread);
6617 ctx->sqo_thread = NULL;
6621 static void io_finish_async(struct io_ring_ctx *ctx)
6623 io_sq_thread_stop(ctx);
6626 io_wq_destroy(ctx->io_wq);
6631 #if defined(CONFIG_UNIX)
6633 * Ensure the UNIX gc is aware of our file set, so we are certain that
6634 * the io_uring can be safely unregistered on process exit, even if we have
6635 * loops in the file referencing.
6637 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6639 struct sock *sk = ctx->ring_sock->sk;
6640 struct scm_fp_list *fpl;
6641 struct sk_buff *skb;
6644 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6648 skb = alloc_skb(0, GFP_KERNEL);
6657 fpl->user = get_uid(ctx->user);
6658 for (i = 0; i < nr; i++) {
6659 struct file *file = io_file_from_index(ctx, i + offset);
6663 fpl->fp[nr_files] = get_file(file);
6664 unix_inflight(fpl->user, fpl->fp[nr_files]);
6669 fpl->max = SCM_MAX_FD;
6670 fpl->count = nr_files;
6671 UNIXCB(skb).fp = fpl;
6672 skb->destructor = unix_destruct_scm;
6673 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6674 skb_queue_head(&sk->sk_receive_queue, skb);
6676 for (i = 0; i < nr_files; i++)
6687 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6688 * causes regular reference counting to break down. We rely on the UNIX
6689 * garbage collection to take care of this problem for us.
6691 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6693 unsigned left, total;
6697 left = ctx->nr_user_files;
6699 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6701 ret = __io_sqe_files_scm(ctx, this_files, total);
6705 total += this_files;
6711 while (total < ctx->nr_user_files) {
6712 struct file *file = io_file_from_index(ctx, total);
6722 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6728 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6733 for (i = 0; i < nr_tables; i++) {
6734 struct fixed_file_table *table = &ctx->file_data->table[i];
6735 unsigned this_files;
6737 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6738 table->files = kcalloc(this_files, sizeof(struct file *),
6742 nr_files -= this_files;
6748 for (i = 0; i < nr_tables; i++) {
6749 struct fixed_file_table *table = &ctx->file_data->table[i];
6750 kfree(table->files);
6755 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6757 #if defined(CONFIG_UNIX)
6758 struct sock *sock = ctx->ring_sock->sk;
6759 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6760 struct sk_buff *skb;
6763 __skb_queue_head_init(&list);
6766 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6767 * remove this entry and rearrange the file array.
6769 skb = skb_dequeue(head);
6771 struct scm_fp_list *fp;
6773 fp = UNIXCB(skb).fp;
6774 for (i = 0; i < fp->count; i++) {
6777 if (fp->fp[i] != file)
6780 unix_notinflight(fp->user, fp->fp[i]);
6781 left = fp->count - 1 - i;
6783 memmove(&fp->fp[i], &fp->fp[i + 1],
6784 left * sizeof(struct file *));
6791 __skb_queue_tail(&list, skb);
6801 __skb_queue_tail(&list, skb);
6803 skb = skb_dequeue(head);
6806 if (skb_peek(&list)) {
6807 spin_lock_irq(&head->lock);
6808 while ((skb = __skb_dequeue(&list)) != NULL)
6809 __skb_queue_tail(head, skb);
6810 spin_unlock_irq(&head->lock);
6817 struct io_file_put {
6818 struct list_head list;
6822 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6824 struct fixed_file_data *file_data = ref_node->file_data;
6825 struct io_ring_ctx *ctx = file_data->ctx;
6826 struct io_file_put *pfile, *tmp;
6828 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6829 list_del(&pfile->list);
6830 io_ring_file_put(ctx, pfile->file);
6834 spin_lock(&file_data->lock);
6835 list_del(&ref_node->node);
6836 spin_unlock(&file_data->lock);
6838 percpu_ref_exit(&ref_node->refs);
6840 percpu_ref_put(&file_data->refs);
6843 static void io_file_put_work(struct work_struct *work)
6845 struct io_ring_ctx *ctx;
6846 struct llist_node *node;
6848 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6849 node = llist_del_all(&ctx->file_put_llist);
6852 struct fixed_file_ref_node *ref_node;
6853 struct llist_node *next = node->next;
6855 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6856 __io_file_put_work(ref_node);
6861 static void io_file_data_ref_zero(struct percpu_ref *ref)
6863 struct fixed_file_ref_node *ref_node;
6864 struct io_ring_ctx *ctx;
6868 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6869 ctx = ref_node->file_data->ctx;
6871 if (percpu_ref_is_dying(&ctx->file_data->refs))
6874 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6876 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6878 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6881 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6882 struct io_ring_ctx *ctx)
6884 struct fixed_file_ref_node *ref_node;
6886 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6888 return ERR_PTR(-ENOMEM);
6890 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6893 return ERR_PTR(-ENOMEM);
6895 INIT_LIST_HEAD(&ref_node->node);
6896 INIT_LIST_HEAD(&ref_node->file_list);
6897 ref_node->file_data = ctx->file_data;
6901 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6903 percpu_ref_exit(&ref_node->refs);
6907 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6910 __s32 __user *fds = (__s32 __user *) arg;
6915 struct fixed_file_ref_node *ref_node;
6921 if (nr_args > IORING_MAX_FIXED_FILES)
6924 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6925 if (!ctx->file_data)
6927 ctx->file_data->ctx = ctx;
6928 init_completion(&ctx->file_data->done);
6929 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6930 spin_lock_init(&ctx->file_data->lock);
6932 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6933 ctx->file_data->table = kcalloc(nr_tables,
6934 sizeof(struct fixed_file_table),
6936 if (!ctx->file_data->table) {
6937 kfree(ctx->file_data);
6938 ctx->file_data = NULL;
6942 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6943 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6944 kfree(ctx->file_data->table);
6945 kfree(ctx->file_data);
6946 ctx->file_data = NULL;
6950 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6951 percpu_ref_exit(&ctx->file_data->refs);
6952 kfree(ctx->file_data->table);
6953 kfree(ctx->file_data);
6954 ctx->file_data = NULL;
6958 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6959 struct fixed_file_table *table;
6963 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6965 /* allow sparse sets */
6971 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6972 index = i & IORING_FILE_TABLE_MASK;
6980 * Don't allow io_uring instances to be registered. If UNIX
6981 * isn't enabled, then this causes a reference cycle and this
6982 * instance can never get freed. If UNIX is enabled we'll
6983 * handle it just fine, but there's still no point in allowing
6984 * a ring fd as it doesn't support regular read/write anyway.
6986 if (file->f_op == &io_uring_fops) {
6991 table->files[index] = file;
6995 for (i = 0; i < ctx->nr_user_files; i++) {
6996 file = io_file_from_index(ctx, i);
7000 for (i = 0; i < nr_tables; i++)
7001 kfree(ctx->file_data->table[i].files);
7003 kfree(ctx->file_data->table);
7004 kfree(ctx->file_data);
7005 ctx->file_data = NULL;
7006 ctx->nr_user_files = 0;
7010 ret = io_sqe_files_scm(ctx);
7012 io_sqe_files_unregister(ctx);
7016 ref_node = alloc_fixed_file_ref_node(ctx);
7017 if (IS_ERR(ref_node)) {
7018 io_sqe_files_unregister(ctx);
7019 return PTR_ERR(ref_node);
7022 ctx->file_data->cur_refs = &ref_node->refs;
7023 spin_lock(&ctx->file_data->lock);
7024 list_add(&ref_node->node, &ctx->file_data->ref_list);
7025 spin_unlock(&ctx->file_data->lock);
7026 percpu_ref_get(&ctx->file_data->refs);
7030 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7033 #if defined(CONFIG_UNIX)
7034 struct sock *sock = ctx->ring_sock->sk;
7035 struct sk_buff_head *head = &sock->sk_receive_queue;
7036 struct sk_buff *skb;
7039 * See if we can merge this file into an existing skb SCM_RIGHTS
7040 * file set. If there's no room, fall back to allocating a new skb
7041 * and filling it in.
7043 spin_lock_irq(&head->lock);
7044 skb = skb_peek(head);
7046 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7048 if (fpl->count < SCM_MAX_FD) {
7049 __skb_unlink(skb, head);
7050 spin_unlock_irq(&head->lock);
7051 fpl->fp[fpl->count] = get_file(file);
7052 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7054 spin_lock_irq(&head->lock);
7055 __skb_queue_head(head, skb);
7060 spin_unlock_irq(&head->lock);
7067 return __io_sqe_files_scm(ctx, 1, index);
7073 static int io_queue_file_removal(struct fixed_file_data *data,
7076 struct io_file_put *pfile;
7077 struct percpu_ref *refs = data->cur_refs;
7078 struct fixed_file_ref_node *ref_node;
7080 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7084 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7086 list_add(&pfile->list, &ref_node->file_list);
7091 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7092 struct io_uring_files_update *up,
7095 struct fixed_file_data *data = ctx->file_data;
7096 struct fixed_file_ref_node *ref_node;
7101 bool needs_switch = false;
7103 if (check_add_overflow(up->offset, nr_args, &done))
7105 if (done > ctx->nr_user_files)
7108 ref_node = alloc_fixed_file_ref_node(ctx);
7109 if (IS_ERR(ref_node))
7110 return PTR_ERR(ref_node);
7113 fds = u64_to_user_ptr(up->fds);
7115 struct fixed_file_table *table;
7119 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7123 i = array_index_nospec(up->offset, ctx->nr_user_files);
7124 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7125 index = i & IORING_FILE_TABLE_MASK;
7126 if (table->files[index]) {
7127 file = io_file_from_index(ctx, index);
7128 err = io_queue_file_removal(data, file);
7131 table->files[index] = NULL;
7132 needs_switch = true;
7141 * Don't allow io_uring instances to be registered. If
7142 * UNIX isn't enabled, then this causes a reference
7143 * cycle and this instance can never get freed. If UNIX
7144 * is enabled we'll handle it just fine, but there's
7145 * still no point in allowing a ring fd as it doesn't
7146 * support regular read/write anyway.
7148 if (file->f_op == &io_uring_fops) {
7153 table->files[index] = file;
7154 err = io_sqe_file_register(ctx, file, i);
7164 percpu_ref_kill(data->cur_refs);
7165 spin_lock(&data->lock);
7166 list_add(&ref_node->node, &data->ref_list);
7167 data->cur_refs = &ref_node->refs;
7168 spin_unlock(&data->lock);
7169 percpu_ref_get(&ctx->file_data->refs);
7171 destroy_fixed_file_ref_node(ref_node);
7173 return done ? done : err;
7176 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7179 struct io_uring_files_update up;
7181 if (!ctx->file_data)
7185 if (copy_from_user(&up, arg, sizeof(up)))
7190 return __io_sqe_files_update(ctx, &up, nr_args);
7193 static void io_free_work(struct io_wq_work *work)
7195 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7197 /* Consider that io_steal_work() relies on this ref */
7201 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7202 struct io_uring_params *p)
7204 struct io_wq_data data;
7206 struct io_ring_ctx *ctx_attach;
7207 unsigned int concurrency;
7210 data.user = ctx->user;
7211 data.free_work = io_free_work;
7212 data.do_work = io_wq_submit_work;
7214 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7215 /* Do QD, or 4 * CPUS, whatever is smallest */
7216 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7218 ctx->io_wq = io_wq_create(concurrency, &data);
7219 if (IS_ERR(ctx->io_wq)) {
7220 ret = PTR_ERR(ctx->io_wq);
7226 f = fdget(p->wq_fd);
7230 if (f.file->f_op != &io_uring_fops) {
7235 ctx_attach = f.file->private_data;
7236 /* @io_wq is protected by holding the fd */
7237 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7242 ctx->io_wq = ctx_attach->io_wq;
7248 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7249 struct io_uring_params *p)
7253 if (ctx->flags & IORING_SETUP_SQPOLL) {
7254 mmgrab(current->mm);
7255 ctx->sqo_mm = current->mm;
7258 if (!capable(CAP_SYS_ADMIN))
7261 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7262 if (!ctx->sq_thread_idle)
7263 ctx->sq_thread_idle = HZ;
7265 if (p->flags & IORING_SETUP_SQ_AFF) {
7266 int cpu = p->sq_thread_cpu;
7269 if (cpu >= nr_cpu_ids)
7271 if (!cpu_online(cpu))
7274 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7278 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7281 if (IS_ERR(ctx->sqo_thread)) {
7282 ret = PTR_ERR(ctx->sqo_thread);
7283 ctx->sqo_thread = NULL;
7286 wake_up_process(ctx->sqo_thread);
7287 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7288 /* Can't have SQ_AFF without SQPOLL */
7293 ret = io_init_wq_offload(ctx, p);
7299 io_finish_async(ctx);
7301 mmdrop(ctx->sqo_mm);
7307 static inline void __io_unaccount_mem(struct user_struct *user,
7308 unsigned long nr_pages)
7310 atomic_long_sub(nr_pages, &user->locked_vm);
7313 static inline int __io_account_mem(struct user_struct *user,
7314 unsigned long nr_pages)
7316 unsigned long page_limit, cur_pages, new_pages;
7318 /* Don't allow more pages than we can safely lock */
7319 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7322 cur_pages = atomic_long_read(&user->locked_vm);
7323 new_pages = cur_pages + nr_pages;
7324 if (new_pages > page_limit)
7326 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7327 new_pages) != cur_pages);
7332 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7333 enum io_mem_account acct)
7336 __io_unaccount_mem(ctx->user, nr_pages);
7339 if (acct == ACCT_LOCKED)
7340 ctx->sqo_mm->locked_vm -= nr_pages;
7341 else if (acct == ACCT_PINNED)
7342 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7346 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7347 enum io_mem_account acct)
7351 if (ctx->limit_mem) {
7352 ret = __io_account_mem(ctx->user, nr_pages);
7358 if (acct == ACCT_LOCKED)
7359 ctx->sqo_mm->locked_vm += nr_pages;
7360 else if (acct == ACCT_PINNED)
7361 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7367 static void io_mem_free(void *ptr)
7374 page = virt_to_head_page(ptr);
7375 if (put_page_testzero(page))
7376 free_compound_page(page);
7379 static void *io_mem_alloc(size_t size)
7381 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7384 return (void *) __get_free_pages(gfp_flags, get_order(size));
7387 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7390 struct io_rings *rings;
7391 size_t off, sq_array_size;
7393 off = struct_size(rings, cqes, cq_entries);
7394 if (off == SIZE_MAX)
7398 off = ALIGN(off, SMP_CACHE_BYTES);
7403 sq_array_size = array_size(sizeof(u32), sq_entries);
7404 if (sq_array_size == SIZE_MAX)
7407 if (check_add_overflow(off, sq_array_size, &off))
7416 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7420 pages = (size_t)1 << get_order(
7421 rings_size(sq_entries, cq_entries, NULL));
7422 pages += (size_t)1 << get_order(
7423 array_size(sizeof(struct io_uring_sqe), sq_entries));
7428 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7432 if (!ctx->user_bufs)
7435 for (i = 0; i < ctx->nr_user_bufs; i++) {
7436 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7438 for (j = 0; j < imu->nr_bvecs; j++)
7439 unpin_user_page(imu->bvec[j].bv_page);
7441 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7446 kfree(ctx->user_bufs);
7447 ctx->user_bufs = NULL;
7448 ctx->nr_user_bufs = 0;
7452 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7453 void __user *arg, unsigned index)
7455 struct iovec __user *src;
7457 #ifdef CONFIG_COMPAT
7459 struct compat_iovec __user *ciovs;
7460 struct compat_iovec ciov;
7462 ciovs = (struct compat_iovec __user *) arg;
7463 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7466 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7467 dst->iov_len = ciov.iov_len;
7471 src = (struct iovec __user *) arg;
7472 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7477 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7480 struct vm_area_struct **vmas = NULL;
7481 struct page **pages = NULL;
7482 int i, j, got_pages = 0;
7487 if (!nr_args || nr_args > UIO_MAXIOV)
7490 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7492 if (!ctx->user_bufs)
7495 for (i = 0; i < nr_args; i++) {
7496 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7497 unsigned long off, start, end, ubuf;
7502 ret = io_copy_iov(ctx, &iov, arg, i);
7507 * Don't impose further limits on the size and buffer
7508 * constraints here, we'll -EINVAL later when IO is
7509 * submitted if they are wrong.
7512 if (!iov.iov_base || !iov.iov_len)
7515 /* arbitrary limit, but we need something */
7516 if (iov.iov_len > SZ_1G)
7519 ubuf = (unsigned long) iov.iov_base;
7520 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7521 start = ubuf >> PAGE_SHIFT;
7522 nr_pages = end - start;
7524 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7529 if (!pages || nr_pages > got_pages) {
7532 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7534 vmas = kvmalloc_array(nr_pages,
7535 sizeof(struct vm_area_struct *),
7537 if (!pages || !vmas) {
7539 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7542 got_pages = nr_pages;
7545 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7549 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7554 mmap_read_lock(current->mm);
7555 pret = pin_user_pages(ubuf, nr_pages,
7556 FOLL_WRITE | FOLL_LONGTERM,
7558 if (pret == nr_pages) {
7559 /* don't support file backed memory */
7560 for (j = 0; j < nr_pages; j++) {
7561 struct vm_area_struct *vma = vmas[j];
7564 !is_file_hugepages(vma->vm_file)) {
7570 ret = pret < 0 ? pret : -EFAULT;
7572 mmap_read_unlock(current->mm);
7575 * if we did partial map, or found file backed vmas,
7576 * release any pages we did get
7579 unpin_user_pages(pages, pret);
7580 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7585 off = ubuf & ~PAGE_MASK;
7587 for (j = 0; j < nr_pages; j++) {
7590 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7591 imu->bvec[j].bv_page = pages[j];
7592 imu->bvec[j].bv_len = vec_len;
7593 imu->bvec[j].bv_offset = off;
7597 /* store original address for later verification */
7599 imu->len = iov.iov_len;
7600 imu->nr_bvecs = nr_pages;
7602 ctx->nr_user_bufs++;
7610 io_sqe_buffer_unregister(ctx);
7614 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7616 __s32 __user *fds = arg;
7622 if (copy_from_user(&fd, fds, sizeof(*fds)))
7625 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7626 if (IS_ERR(ctx->cq_ev_fd)) {
7627 int ret = PTR_ERR(ctx->cq_ev_fd);
7628 ctx->cq_ev_fd = NULL;
7635 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7637 if (ctx->cq_ev_fd) {
7638 eventfd_ctx_put(ctx->cq_ev_fd);
7639 ctx->cq_ev_fd = NULL;
7646 static int __io_destroy_buffers(int id, void *p, void *data)
7648 struct io_ring_ctx *ctx = data;
7649 struct io_buffer *buf = p;
7651 __io_remove_buffers(ctx, buf, id, -1U);
7655 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7657 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7658 idr_destroy(&ctx->io_buffer_idr);
7661 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7663 io_finish_async(ctx);
7665 mmdrop(ctx->sqo_mm);
7669 io_iopoll_reap_events(ctx);
7670 io_sqe_buffer_unregister(ctx);
7671 io_sqe_files_unregister(ctx);
7672 io_eventfd_unregister(ctx);
7673 io_destroy_buffers(ctx);
7674 idr_destroy(&ctx->personality_idr);
7676 #if defined(CONFIG_UNIX)
7677 if (ctx->ring_sock) {
7678 ctx->ring_sock->file = NULL; /* so that iput() is called */
7679 sock_release(ctx->ring_sock);
7683 io_mem_free(ctx->rings);
7684 io_mem_free(ctx->sq_sqes);
7686 percpu_ref_exit(&ctx->refs);
7687 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7689 free_uid(ctx->user);
7690 put_cred(ctx->creds);
7691 kfree(ctx->cancel_hash);
7692 kmem_cache_free(req_cachep, ctx->fallback_req);
7696 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7698 struct io_ring_ctx *ctx = file->private_data;
7701 poll_wait(file, &ctx->cq_wait, wait);
7703 * synchronizes with barrier from wq_has_sleeper call in
7707 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7708 ctx->rings->sq_ring_entries)
7709 mask |= EPOLLOUT | EPOLLWRNORM;
7710 if (io_cqring_events(ctx, false))
7711 mask |= EPOLLIN | EPOLLRDNORM;
7716 static int io_uring_fasync(int fd, struct file *file, int on)
7718 struct io_ring_ctx *ctx = file->private_data;
7720 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7723 static int io_remove_personalities(int id, void *p, void *data)
7725 struct io_ring_ctx *ctx = data;
7726 const struct cred *cred;
7728 cred = idr_remove(&ctx->personality_idr, id);
7734 static void io_ring_exit_work(struct work_struct *work)
7736 struct io_ring_ctx *ctx;
7738 ctx = container_of(work, struct io_ring_ctx, exit_work);
7740 io_cqring_overflow_flush(ctx, true);
7743 * If we're doing polled IO and end up having requests being
7744 * submitted async (out-of-line), then completions can come in while
7745 * we're waiting for refs to drop. We need to reap these manually,
7746 * as nobody else will be looking for them.
7748 while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20)) {
7749 io_iopoll_reap_events(ctx);
7751 io_cqring_overflow_flush(ctx, true);
7753 io_ring_ctx_free(ctx);
7756 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7758 mutex_lock(&ctx->uring_lock);
7759 percpu_ref_kill(&ctx->refs);
7760 mutex_unlock(&ctx->uring_lock);
7762 io_kill_timeouts(ctx);
7763 io_poll_remove_all(ctx);
7766 io_wq_cancel_all(ctx->io_wq);
7768 io_iopoll_reap_events(ctx);
7769 /* if we failed setting up the ctx, we might not have any rings */
7771 io_cqring_overflow_flush(ctx, true);
7772 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7773 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7774 queue_work(system_wq, &ctx->exit_work);
7777 static int io_uring_release(struct inode *inode, struct file *file)
7779 struct io_ring_ctx *ctx = file->private_data;
7781 file->private_data = NULL;
7782 io_ring_ctx_wait_and_kill(ctx);
7786 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7788 struct files_struct *files = data;
7790 return work->files == files;
7793 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7794 struct files_struct *files)
7796 if (list_empty_careful(&ctx->inflight_list))
7799 /* cancel all at once, should be faster than doing it one by one*/
7800 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7802 while (!list_empty_careful(&ctx->inflight_list)) {
7803 struct io_kiocb *cancel_req = NULL, *req;
7806 spin_lock_irq(&ctx->inflight_lock);
7807 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7808 if (req->work.files != files)
7810 /* req is being completed, ignore */
7811 if (!refcount_inc_not_zero(&req->refs))
7817 prepare_to_wait(&ctx->inflight_wait, &wait,
7818 TASK_UNINTERRUPTIBLE);
7819 spin_unlock_irq(&ctx->inflight_lock);
7821 /* We need to keep going until we don't find a matching req */
7825 if (cancel_req->flags & REQ_F_OVERFLOW) {
7826 spin_lock_irq(&ctx->completion_lock);
7827 list_del(&cancel_req->list);
7828 cancel_req->flags &= ~REQ_F_OVERFLOW;
7829 if (list_empty(&ctx->cq_overflow_list)) {
7830 clear_bit(0, &ctx->sq_check_overflow);
7831 clear_bit(0, &ctx->cq_check_overflow);
7833 spin_unlock_irq(&ctx->completion_lock);
7835 WRITE_ONCE(ctx->rings->cq_overflow,
7836 atomic_inc_return(&ctx->cached_cq_overflow));
7839 * Put inflight ref and overflow ref. If that's
7840 * all we had, then we're done with this request.
7842 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7843 io_free_req(cancel_req);
7844 finish_wait(&ctx->inflight_wait, &wait);
7848 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7849 io_put_req(cancel_req);
7853 finish_wait(&ctx->inflight_wait, &wait);
7857 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7859 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7860 struct task_struct *task = data;
7862 return req->task == task;
7865 static int io_uring_flush(struct file *file, void *data)
7867 struct io_ring_ctx *ctx = file->private_data;
7869 io_uring_cancel_files(ctx, data);
7872 * If the task is going away, cancel work it may have pending
7874 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7875 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7880 static void *io_uring_validate_mmap_request(struct file *file,
7881 loff_t pgoff, size_t sz)
7883 struct io_ring_ctx *ctx = file->private_data;
7884 loff_t offset = pgoff << PAGE_SHIFT;
7889 case IORING_OFF_SQ_RING:
7890 case IORING_OFF_CQ_RING:
7893 case IORING_OFF_SQES:
7897 return ERR_PTR(-EINVAL);
7900 page = virt_to_head_page(ptr);
7901 if (sz > page_size(page))
7902 return ERR_PTR(-EINVAL);
7909 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7911 size_t sz = vma->vm_end - vma->vm_start;
7915 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7917 return PTR_ERR(ptr);
7919 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7920 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7923 #else /* !CONFIG_MMU */
7925 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7927 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7930 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7932 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7935 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7936 unsigned long addr, unsigned long len,
7937 unsigned long pgoff, unsigned long flags)
7941 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7943 return PTR_ERR(ptr);
7945 return (unsigned long) ptr;
7948 #endif /* !CONFIG_MMU */
7950 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7951 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7954 struct io_ring_ctx *ctx;
7961 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7969 if (f.file->f_op != &io_uring_fops)
7973 ctx = f.file->private_data;
7974 if (!percpu_ref_tryget(&ctx->refs))
7978 * For SQ polling, the thread will do all submissions and completions.
7979 * Just return the requested submit count, and wake the thread if
7983 if (ctx->flags & IORING_SETUP_SQPOLL) {
7984 if (!list_empty_careful(&ctx->cq_overflow_list))
7985 io_cqring_overflow_flush(ctx, false);
7986 if (flags & IORING_ENTER_SQ_WAKEUP)
7987 wake_up(&ctx->sqo_wait);
7988 submitted = to_submit;
7989 } else if (to_submit) {
7990 mutex_lock(&ctx->uring_lock);
7991 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
7992 mutex_unlock(&ctx->uring_lock);
7994 if (submitted != to_submit)
7997 if (flags & IORING_ENTER_GETEVENTS) {
7998 unsigned nr_events = 0;
8000 min_complete = min(min_complete, ctx->cq_entries);
8003 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8004 * space applications don't need to do io completion events
8005 * polling again, they can rely on io_sq_thread to do polling
8006 * work, which can reduce cpu usage and uring_lock contention.
8008 if (ctx->flags & IORING_SETUP_IOPOLL &&
8009 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8010 ret = io_iopoll_check(ctx, &nr_events, min_complete);
8012 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8017 percpu_ref_put(&ctx->refs);
8020 return submitted ? submitted : ret;
8023 #ifdef CONFIG_PROC_FS
8024 static int io_uring_show_cred(int id, void *p, void *data)
8026 const struct cred *cred = p;
8027 struct seq_file *m = data;
8028 struct user_namespace *uns = seq_user_ns(m);
8029 struct group_info *gi;
8034 seq_printf(m, "%5d\n", id);
8035 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8036 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8037 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8038 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8039 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8040 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8041 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8042 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8043 seq_puts(m, "\n\tGroups:\t");
8044 gi = cred->group_info;
8045 for (g = 0; g < gi->ngroups; g++) {
8046 seq_put_decimal_ull(m, g ? " " : "",
8047 from_kgid_munged(uns, gi->gid[g]));
8049 seq_puts(m, "\n\tCapEff:\t");
8050 cap = cred->cap_effective;
8051 CAP_FOR_EACH_U32(__capi)
8052 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8057 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8061 mutex_lock(&ctx->uring_lock);
8062 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8063 for (i = 0; i < ctx->nr_user_files; i++) {
8064 struct fixed_file_table *table;
8067 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8068 f = table->files[i & IORING_FILE_TABLE_MASK];
8070 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8072 seq_printf(m, "%5u: <none>\n", i);
8074 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8075 for (i = 0; i < ctx->nr_user_bufs; i++) {
8076 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8078 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8079 (unsigned int) buf->len);
8081 if (!idr_is_empty(&ctx->personality_idr)) {
8082 seq_printf(m, "Personalities:\n");
8083 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8085 seq_printf(m, "PollList:\n");
8086 spin_lock_irq(&ctx->completion_lock);
8087 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8088 struct hlist_head *list = &ctx->cancel_hash[i];
8089 struct io_kiocb *req;
8091 hlist_for_each_entry(req, list, hash_node)
8092 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8093 req->task->task_works != NULL);
8095 spin_unlock_irq(&ctx->completion_lock);
8096 mutex_unlock(&ctx->uring_lock);
8099 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8101 struct io_ring_ctx *ctx = f->private_data;
8103 if (percpu_ref_tryget(&ctx->refs)) {
8104 __io_uring_show_fdinfo(ctx, m);
8105 percpu_ref_put(&ctx->refs);
8110 static const struct file_operations io_uring_fops = {
8111 .release = io_uring_release,
8112 .flush = io_uring_flush,
8113 .mmap = io_uring_mmap,
8115 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8116 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8118 .poll = io_uring_poll,
8119 .fasync = io_uring_fasync,
8120 #ifdef CONFIG_PROC_FS
8121 .show_fdinfo = io_uring_show_fdinfo,
8125 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8126 struct io_uring_params *p)
8128 struct io_rings *rings;
8129 size_t size, sq_array_offset;
8131 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8132 if (size == SIZE_MAX)
8135 rings = io_mem_alloc(size);
8140 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8141 rings->sq_ring_mask = p->sq_entries - 1;
8142 rings->cq_ring_mask = p->cq_entries - 1;
8143 rings->sq_ring_entries = p->sq_entries;
8144 rings->cq_ring_entries = p->cq_entries;
8145 ctx->sq_mask = rings->sq_ring_mask;
8146 ctx->cq_mask = rings->cq_ring_mask;
8147 ctx->sq_entries = rings->sq_ring_entries;
8148 ctx->cq_entries = rings->cq_ring_entries;
8150 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8151 if (size == SIZE_MAX) {
8152 io_mem_free(ctx->rings);
8157 ctx->sq_sqes = io_mem_alloc(size);
8158 if (!ctx->sq_sqes) {
8159 io_mem_free(ctx->rings);
8168 * Allocate an anonymous fd, this is what constitutes the application
8169 * visible backing of an io_uring instance. The application mmaps this
8170 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8171 * we have to tie this fd to a socket for file garbage collection purposes.
8173 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8178 #if defined(CONFIG_UNIX)
8179 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8185 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8189 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8190 O_RDWR | O_CLOEXEC);
8193 ret = PTR_ERR(file);
8197 #if defined(CONFIG_UNIX)
8198 ctx->ring_sock->file = file;
8200 fd_install(ret, file);
8203 #if defined(CONFIG_UNIX)
8204 sock_release(ctx->ring_sock);
8205 ctx->ring_sock = NULL;
8210 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8211 struct io_uring_params __user *params)
8213 struct user_struct *user = NULL;
8214 struct io_ring_ctx *ctx;
8220 if (entries > IORING_MAX_ENTRIES) {
8221 if (!(p->flags & IORING_SETUP_CLAMP))
8223 entries = IORING_MAX_ENTRIES;
8227 * Use twice as many entries for the CQ ring. It's possible for the
8228 * application to drive a higher depth than the size of the SQ ring,
8229 * since the sqes are only used at submission time. This allows for
8230 * some flexibility in overcommitting a bit. If the application has
8231 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8232 * of CQ ring entries manually.
8234 p->sq_entries = roundup_pow_of_two(entries);
8235 if (p->flags & IORING_SETUP_CQSIZE) {
8237 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8238 * to a power-of-two, if it isn't already. We do NOT impose
8239 * any cq vs sq ring sizing.
8241 if (p->cq_entries < p->sq_entries)
8243 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8244 if (!(p->flags & IORING_SETUP_CLAMP))
8246 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8248 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8250 p->cq_entries = 2 * p->sq_entries;
8253 user = get_uid(current_user());
8254 limit_mem = !capable(CAP_IPC_LOCK);
8257 ret = __io_account_mem(user,
8258 ring_pages(p->sq_entries, p->cq_entries));
8265 ctx = io_ring_ctx_alloc(p);
8268 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8273 ctx->compat = in_compat_syscall();
8275 ctx->creds = get_current_cred();
8277 ret = io_allocate_scq_urings(ctx, p);
8281 ret = io_sq_offload_start(ctx, p);
8285 memset(&p->sq_off, 0, sizeof(p->sq_off));
8286 p->sq_off.head = offsetof(struct io_rings, sq.head);
8287 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8288 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8289 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8290 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8291 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8292 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8294 memset(&p->cq_off, 0, sizeof(p->cq_off));
8295 p->cq_off.head = offsetof(struct io_rings, cq.head);
8296 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8297 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8298 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8299 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8300 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8301 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8303 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8304 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8305 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8306 IORING_FEAT_POLL_32BITS;
8308 if (copy_to_user(params, p, sizeof(*p))) {
8313 * Install ring fd as the very last thing, so we don't risk someone
8314 * having closed it before we finish setup
8316 ret = io_uring_get_fd(ctx);
8320 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8321 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8323 ctx->limit_mem = limit_mem;
8326 io_ring_ctx_wait_and_kill(ctx);
8331 * Sets up an aio uring context, and returns the fd. Applications asks for a
8332 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8333 * params structure passed in.
8335 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8337 struct io_uring_params p;
8340 if (copy_from_user(&p, params, sizeof(p)))
8342 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8347 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8348 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8349 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8352 return io_uring_create(entries, &p, params);
8355 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8356 struct io_uring_params __user *, params)
8358 return io_uring_setup(entries, params);
8361 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8363 struct io_uring_probe *p;
8367 size = struct_size(p, ops, nr_args);
8368 if (size == SIZE_MAX)
8370 p = kzalloc(size, GFP_KERNEL);
8375 if (copy_from_user(p, arg, size))
8378 if (memchr_inv(p, 0, size))
8381 p->last_op = IORING_OP_LAST - 1;
8382 if (nr_args > IORING_OP_LAST)
8383 nr_args = IORING_OP_LAST;
8385 for (i = 0; i < nr_args; i++) {
8387 if (!io_op_defs[i].not_supported)
8388 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8393 if (copy_to_user(arg, p, size))
8400 static int io_register_personality(struct io_ring_ctx *ctx)
8402 const struct cred *creds = get_current_cred();
8405 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8406 USHRT_MAX, GFP_KERNEL);
8412 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8414 const struct cred *old_creds;
8416 old_creds = idr_remove(&ctx->personality_idr, id);
8418 put_cred(old_creds);
8425 static bool io_register_op_must_quiesce(int op)
8428 case IORING_UNREGISTER_FILES:
8429 case IORING_REGISTER_FILES_UPDATE:
8430 case IORING_REGISTER_PROBE:
8431 case IORING_REGISTER_PERSONALITY:
8432 case IORING_UNREGISTER_PERSONALITY:
8439 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8440 void __user *arg, unsigned nr_args)
8441 __releases(ctx->uring_lock)
8442 __acquires(ctx->uring_lock)
8447 * We're inside the ring mutex, if the ref is already dying, then
8448 * someone else killed the ctx or is already going through
8449 * io_uring_register().
8451 if (percpu_ref_is_dying(&ctx->refs))
8454 if (io_register_op_must_quiesce(opcode)) {
8455 percpu_ref_kill(&ctx->refs);
8458 * Drop uring mutex before waiting for references to exit. If
8459 * another thread is currently inside io_uring_enter() it might
8460 * need to grab the uring_lock to make progress. If we hold it
8461 * here across the drain wait, then we can deadlock. It's safe
8462 * to drop the mutex here, since no new references will come in
8463 * after we've killed the percpu ref.
8465 mutex_unlock(&ctx->uring_lock);
8466 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8467 mutex_lock(&ctx->uring_lock);
8469 percpu_ref_resurrect(&ctx->refs);
8476 case IORING_REGISTER_BUFFERS:
8477 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8479 case IORING_UNREGISTER_BUFFERS:
8483 ret = io_sqe_buffer_unregister(ctx);
8485 case IORING_REGISTER_FILES:
8486 ret = io_sqe_files_register(ctx, arg, nr_args);
8488 case IORING_UNREGISTER_FILES:
8492 ret = io_sqe_files_unregister(ctx);
8494 case IORING_REGISTER_FILES_UPDATE:
8495 ret = io_sqe_files_update(ctx, arg, nr_args);
8497 case IORING_REGISTER_EVENTFD:
8498 case IORING_REGISTER_EVENTFD_ASYNC:
8502 ret = io_eventfd_register(ctx, arg);
8505 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8506 ctx->eventfd_async = 1;
8508 ctx->eventfd_async = 0;
8510 case IORING_UNREGISTER_EVENTFD:
8514 ret = io_eventfd_unregister(ctx);
8516 case IORING_REGISTER_PROBE:
8518 if (!arg || nr_args > 256)
8520 ret = io_probe(ctx, arg, nr_args);
8522 case IORING_REGISTER_PERSONALITY:
8526 ret = io_register_personality(ctx);
8528 case IORING_UNREGISTER_PERSONALITY:
8532 ret = io_unregister_personality(ctx, nr_args);
8539 if (io_register_op_must_quiesce(opcode)) {
8540 /* bring the ctx back to life */
8541 percpu_ref_reinit(&ctx->refs);
8543 reinit_completion(&ctx->ref_comp);
8548 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8549 void __user *, arg, unsigned int, nr_args)
8551 struct io_ring_ctx *ctx;
8560 if (f.file->f_op != &io_uring_fops)
8563 ctx = f.file->private_data;
8565 mutex_lock(&ctx->uring_lock);
8566 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8567 mutex_unlock(&ctx->uring_lock);
8568 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8569 ctx->cq_ev_fd != NULL, ret);
8575 static int __init io_uring_init(void)
8577 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8578 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8579 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8582 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8583 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8584 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8585 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8586 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8587 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8588 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8589 BUILD_BUG_SQE_ELEM(8, __u64, off);
8590 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8591 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8592 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8593 BUILD_BUG_SQE_ELEM(24, __u32, len);
8594 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8595 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8596 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8597 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8598 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8599 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8600 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8601 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8602 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8603 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8604 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8605 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8606 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8607 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8608 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8609 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8610 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8611 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8612 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8614 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8615 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8616 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8619 __initcall(io_uring_init);