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(!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 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1671 req->flags &= ~REQ_F_LINK_HEAD;
1673 if (req->flags & REQ_F_LINK_TIMEOUT)
1674 io_kill_linked_timeout(req);
1677 * If LINK is set, we have dependent requests in this chain. If we
1678 * didn't fail this request, queue the first one up, moving any other
1679 * dependencies to the next request. In case of failure, fail the rest
1682 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1683 return io_req_link_next(req);
1688 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1690 struct io_ring_ctx *ctx = req->ctx;
1692 spin_lock_irq(&ctx->completion_lock);
1693 io_cqring_fill_event(req, error);
1694 io_commit_cqring(ctx);
1695 spin_unlock_irq(&ctx->completion_lock);
1697 io_cqring_ev_posted(ctx);
1698 req_set_fail_links(req);
1699 io_double_put_req(req);
1702 static void io_req_task_cancel(struct callback_head *cb)
1704 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1706 __io_req_task_cancel(req, -ECANCELED);
1709 static void __io_req_task_submit(struct io_kiocb *req)
1711 struct io_ring_ctx *ctx = req->ctx;
1713 __set_current_state(TASK_RUNNING);
1714 if (!__io_sq_thread_acquire_mm(ctx)) {
1715 mutex_lock(&ctx->uring_lock);
1716 __io_queue_sqe(req, NULL, NULL);
1717 mutex_unlock(&ctx->uring_lock);
1719 __io_req_task_cancel(req, -EFAULT);
1723 static void io_req_task_submit(struct callback_head *cb)
1725 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1727 __io_req_task_submit(req);
1730 static void io_req_task_queue(struct io_kiocb *req)
1732 struct task_struct *tsk = req->task;
1735 init_task_work(&req->task_work, io_req_task_submit);
1737 ret = task_work_add(tsk, &req->task_work, true);
1738 if (unlikely(ret)) {
1739 init_task_work(&req->task_work, io_req_task_cancel);
1740 tsk = io_wq_get_task(req->ctx->io_wq);
1741 task_work_add(tsk, &req->task_work, true);
1743 wake_up_process(tsk);
1746 static void io_queue_next(struct io_kiocb *req)
1748 struct io_kiocb *nxt = io_req_find_next(req);
1751 io_req_task_queue(nxt);
1754 static void io_free_req(struct io_kiocb *req)
1761 void *reqs[IO_IOPOLL_BATCH];
1765 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1766 struct req_batch *rb)
1768 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1769 percpu_ref_put_many(&ctx->refs, rb->to_free);
1773 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1774 struct req_batch *rb)
1777 __io_req_free_batch_flush(ctx, rb);
1780 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1782 if (unlikely(io_is_fallback_req(req))) {
1786 if (req->flags & REQ_F_LINK_HEAD)
1789 io_dismantle_req(req);
1790 rb->reqs[rb->to_free++] = req;
1791 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1792 __io_req_free_batch_flush(req->ctx, rb);
1796 * Drop reference to request, return next in chain (if there is one) if this
1797 * was the last reference to this request.
1799 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1801 struct io_kiocb *nxt = NULL;
1803 if (refcount_dec_and_test(&req->refs)) {
1804 nxt = io_req_find_next(req);
1810 static void io_put_req(struct io_kiocb *req)
1812 if (refcount_dec_and_test(&req->refs))
1816 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1818 struct io_kiocb *timeout, *nxt = NULL;
1821 * A ref is owned by io-wq in which context we're. So, if that's the
1822 * last one, it's safe to steal next work. False negatives are Ok,
1823 * it just will be re-punted async in io_put_work()
1825 if (refcount_read(&req->refs) != 1)
1828 nxt = io_req_find_next(req);
1832 timeout = io_prep_linked_timeout(nxt);
1834 nxt->flags |= REQ_F_QUEUE_TIMEOUT;
1839 * Must only be used if we don't need to care about links, usually from
1840 * within the completion handling itself.
1842 static void __io_double_put_req(struct io_kiocb *req)
1844 /* drop both submit and complete references */
1845 if (refcount_sub_and_test(2, &req->refs))
1849 static void io_double_put_req(struct io_kiocb *req)
1851 /* drop both submit and complete references */
1852 if (refcount_sub_and_test(2, &req->refs))
1856 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1858 struct io_rings *rings = ctx->rings;
1860 if (test_bit(0, &ctx->cq_check_overflow)) {
1862 * noflush == true is from the waitqueue handler, just ensure
1863 * we wake up the task, and the next invocation will flush the
1864 * entries. We cannot safely to it from here.
1866 if (noflush && !list_empty(&ctx->cq_overflow_list))
1869 io_cqring_overflow_flush(ctx, false);
1872 /* See comment at the top of this file */
1874 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1877 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1879 struct io_rings *rings = ctx->rings;
1881 /* make sure SQ entry isn't read before tail */
1882 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1885 static int io_put_kbuf(struct io_kiocb *req)
1887 struct io_buffer *kbuf;
1890 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1891 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1892 cflags |= IORING_CQE_F_BUFFER;
1898 static void io_iopoll_queue(struct list_head *again)
1900 struct io_kiocb *req;
1903 req = list_first_entry(again, struct io_kiocb, list);
1904 list_del(&req->list);
1905 if (!io_rw_reissue(req, -EAGAIN))
1906 io_complete_rw_common(&req->rw.kiocb, -EAGAIN, NULL);
1907 } while (!list_empty(again));
1911 * Find and free completed poll iocbs
1913 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1914 struct list_head *done)
1916 struct req_batch rb;
1917 struct io_kiocb *req;
1920 /* order with ->result store in io_complete_rw_iopoll() */
1924 while (!list_empty(done)) {
1927 req = list_first_entry(done, struct io_kiocb, list);
1928 if (READ_ONCE(req->result) == -EAGAIN) {
1929 req->iopoll_completed = 0;
1930 list_move_tail(&req->list, &again);
1933 list_del(&req->list);
1935 if (req->flags & REQ_F_BUFFER_SELECTED)
1936 cflags = io_put_kbuf(req);
1938 __io_cqring_fill_event(req, req->result, cflags);
1941 if (refcount_dec_and_test(&req->refs))
1942 io_req_free_batch(&rb, req);
1945 io_commit_cqring(ctx);
1946 if (ctx->flags & IORING_SETUP_SQPOLL)
1947 io_cqring_ev_posted(ctx);
1948 io_req_free_batch_finish(ctx, &rb);
1950 if (!list_empty(&again))
1951 io_iopoll_queue(&again);
1954 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1957 struct io_kiocb *req, *tmp;
1963 * Only spin for completions if we don't have multiple devices hanging
1964 * off our complete list, and we're under the requested amount.
1966 spin = !ctx->poll_multi_file && *nr_events < min;
1969 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1970 struct kiocb *kiocb = &req->rw.kiocb;
1973 * Move completed and retryable entries to our local lists.
1974 * If we find a request that requires polling, break out
1975 * and complete those lists first, if we have entries there.
1977 if (READ_ONCE(req->iopoll_completed)) {
1978 list_move_tail(&req->list, &done);
1981 if (!list_empty(&done))
1984 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1993 if (!list_empty(&done))
1994 io_iopoll_complete(ctx, nr_events, &done);
2000 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2001 * non-spinning poll check - we'll still enter the driver poll loop, but only
2002 * as a non-spinning completion check.
2004 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2007 while (!list_empty(&ctx->poll_list) && !need_resched()) {
2010 ret = io_do_iopoll(ctx, nr_events, min);
2013 if (!min || *nr_events >= min)
2021 * We can't just wait for polled events to come to us, we have to actively
2022 * find and complete them.
2024 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
2026 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2029 mutex_lock(&ctx->uring_lock);
2030 while (!list_empty(&ctx->poll_list)) {
2031 unsigned int nr_events = 0;
2033 io_iopoll_getevents(ctx, &nr_events, 1);
2036 * Ensure we allow local-to-the-cpu processing to take place,
2037 * in this case we need to ensure that we reap all events.
2041 mutex_unlock(&ctx->uring_lock);
2044 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
2047 int iters = 0, ret = 0;
2050 * We disallow the app entering submit/complete with polling, but we
2051 * still need to lock the ring to prevent racing with polled issue
2052 * that got punted to a workqueue.
2054 mutex_lock(&ctx->uring_lock);
2059 * Don't enter poll loop if we already have events pending.
2060 * If we do, we can potentially be spinning for commands that
2061 * already triggered a CQE (eg in error).
2063 if (io_cqring_events(ctx, false))
2067 * If a submit got punted to a workqueue, we can have the
2068 * application entering polling for a command before it gets
2069 * issued. That app will hold the uring_lock for the duration
2070 * of the poll right here, so we need to take a breather every
2071 * now and then to ensure that the issue has a chance to add
2072 * the poll to the issued list. Otherwise we can spin here
2073 * forever, while the workqueue is stuck trying to acquire the
2076 if (!(++iters & 7)) {
2077 mutex_unlock(&ctx->uring_lock);
2078 if (current->task_works)
2080 mutex_lock(&ctx->uring_lock);
2083 if (*nr_events < min)
2084 tmin = min - *nr_events;
2086 ret = io_iopoll_getevents(ctx, nr_events, tmin);
2090 } while (min && !*nr_events && !need_resched());
2092 mutex_unlock(&ctx->uring_lock);
2096 static void kiocb_end_write(struct io_kiocb *req)
2099 * Tell lockdep we inherited freeze protection from submission
2102 if (req->flags & REQ_F_ISREG) {
2103 struct inode *inode = file_inode(req->file);
2105 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2107 file_end_write(req->file);
2110 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2111 struct io_comp_state *cs)
2113 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2116 if (kiocb->ki_flags & IOCB_WRITE)
2117 kiocb_end_write(req);
2119 if (res != req->result)
2120 req_set_fail_links(req);
2121 if (req->flags & REQ_F_BUFFER_SELECTED)
2122 cflags = io_put_kbuf(req);
2123 __io_req_complete(req, res, cflags, cs);
2127 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2129 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2130 ssize_t ret = -ECANCELED;
2131 struct iov_iter iter;
2139 switch (req->opcode) {
2140 case IORING_OP_READV:
2141 case IORING_OP_READ_FIXED:
2142 case IORING_OP_READ:
2145 case IORING_OP_WRITEV:
2146 case IORING_OP_WRITE_FIXED:
2147 case IORING_OP_WRITE:
2151 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2156 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2159 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2164 req_set_fail_links(req);
2165 io_req_complete(req, ret);
2169 static void io_rw_resubmit(struct callback_head *cb)
2171 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2172 struct io_ring_ctx *ctx = req->ctx;
2175 __set_current_state(TASK_RUNNING);
2177 err = io_sq_thread_acquire_mm(ctx, req);
2179 if (io_resubmit_prep(req, err)) {
2180 refcount_inc(&req->refs);
2181 io_queue_async_work(req);
2186 static bool io_rw_reissue(struct io_kiocb *req, long res)
2189 struct task_struct *tsk;
2192 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2196 init_task_work(&req->task_work, io_rw_resubmit);
2197 ret = task_work_add(tsk, &req->task_work, true);
2199 wake_up_process(tsk);
2206 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2207 struct io_comp_state *cs)
2209 if (!io_rw_reissue(req, res))
2210 io_complete_rw_common(&req->rw.kiocb, res, cs);
2213 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2215 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2217 __io_complete_rw(req, res, res2, NULL);
2220 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2222 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2224 if (kiocb->ki_flags & IOCB_WRITE)
2225 kiocb_end_write(req);
2227 if (res != -EAGAIN && res != req->result)
2228 req_set_fail_links(req);
2230 WRITE_ONCE(req->result, res);
2231 /* order with io_poll_complete() checking ->result */
2233 WRITE_ONCE(req->iopoll_completed, 1);
2237 * After the iocb has been issued, it's safe to be found on the poll list.
2238 * Adding the kiocb to the list AFTER submission ensures that we don't
2239 * find it from a io_iopoll_getevents() thread before the issuer is done
2240 * accessing the kiocb cookie.
2242 static void io_iopoll_req_issued(struct io_kiocb *req)
2244 struct io_ring_ctx *ctx = req->ctx;
2247 * Track whether we have multiple files in our lists. This will impact
2248 * how we do polling eventually, not spinning if we're on potentially
2249 * different devices.
2251 if (list_empty(&ctx->poll_list)) {
2252 ctx->poll_multi_file = false;
2253 } else if (!ctx->poll_multi_file) {
2254 struct io_kiocb *list_req;
2256 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
2258 if (list_req->file != req->file)
2259 ctx->poll_multi_file = true;
2263 * For fast devices, IO may have already completed. If it has, add
2264 * it to the front so we find it first.
2266 if (READ_ONCE(req->iopoll_completed))
2267 list_add(&req->list, &ctx->poll_list);
2269 list_add_tail(&req->list, &ctx->poll_list);
2271 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2272 wq_has_sleeper(&ctx->sqo_wait))
2273 wake_up(&ctx->sqo_wait);
2276 static void __io_state_file_put(struct io_submit_state *state)
2278 int diff = state->has_refs - state->used_refs;
2281 fput_many(state->file, diff);
2285 static inline void io_state_file_put(struct io_submit_state *state)
2288 __io_state_file_put(state);
2292 * Get as many references to a file as we have IOs left in this submission,
2293 * assuming most submissions are for one file, or at least that each file
2294 * has more than one submission.
2296 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2302 if (state->fd == fd) {
2307 __io_state_file_put(state);
2309 state->file = fget_many(fd, state->ios_left);
2314 state->has_refs = state->ios_left;
2315 state->used_refs = 1;
2320 static bool io_bdev_nowait(struct block_device *bdev)
2323 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2330 * If we tracked the file through the SCM inflight mechanism, we could support
2331 * any file. For now, just ensure that anything potentially problematic is done
2334 static bool io_file_supports_async(struct file *file, int rw)
2336 umode_t mode = file_inode(file)->i_mode;
2338 if (S_ISBLK(mode)) {
2339 if (io_bdev_nowait(file->f_inode->i_bdev))
2343 if (S_ISCHR(mode) || S_ISSOCK(mode))
2345 if (S_ISREG(mode)) {
2346 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2347 file->f_op != &io_uring_fops)
2352 /* any ->read/write should understand O_NONBLOCK */
2353 if (file->f_flags & O_NONBLOCK)
2356 if (!(file->f_mode & FMODE_NOWAIT))
2360 return file->f_op->read_iter != NULL;
2362 return file->f_op->write_iter != NULL;
2365 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2366 bool force_nonblock)
2368 struct io_ring_ctx *ctx = req->ctx;
2369 struct kiocb *kiocb = &req->rw.kiocb;
2373 if (S_ISREG(file_inode(req->file)->i_mode))
2374 req->flags |= REQ_F_ISREG;
2376 kiocb->ki_pos = READ_ONCE(sqe->off);
2377 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2378 req->flags |= REQ_F_CUR_POS;
2379 kiocb->ki_pos = req->file->f_pos;
2381 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2382 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2383 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2387 ioprio = READ_ONCE(sqe->ioprio);
2389 ret = ioprio_check_cap(ioprio);
2393 kiocb->ki_ioprio = ioprio;
2395 kiocb->ki_ioprio = get_current_ioprio();
2397 /* don't allow async punt if RWF_NOWAIT was requested */
2398 if (kiocb->ki_flags & IOCB_NOWAIT)
2399 req->flags |= REQ_F_NOWAIT;
2401 if (kiocb->ki_flags & IOCB_DIRECT)
2402 io_get_req_task(req);
2405 kiocb->ki_flags |= IOCB_NOWAIT;
2407 if (ctx->flags & IORING_SETUP_IOPOLL) {
2408 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2409 !kiocb->ki_filp->f_op->iopoll)
2412 kiocb->ki_flags |= IOCB_HIPRI;
2413 kiocb->ki_complete = io_complete_rw_iopoll;
2414 req->iopoll_completed = 0;
2415 io_get_req_task(req);
2417 if (kiocb->ki_flags & IOCB_HIPRI)
2419 kiocb->ki_complete = io_complete_rw;
2422 req->rw.addr = READ_ONCE(sqe->addr);
2423 req->rw.len = READ_ONCE(sqe->len);
2424 req->buf_index = READ_ONCE(sqe->buf_index);
2428 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2434 case -ERESTARTNOINTR:
2435 case -ERESTARTNOHAND:
2436 case -ERESTART_RESTARTBLOCK:
2438 * We can't just restart the syscall, since previously
2439 * submitted sqes may already be in progress. Just fail this
2445 kiocb->ki_complete(kiocb, ret, 0);
2449 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2450 struct io_comp_state *cs)
2452 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2454 if (req->flags & REQ_F_CUR_POS)
2455 req->file->f_pos = kiocb->ki_pos;
2456 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2457 __io_complete_rw(req, ret, 0, cs);
2459 io_rw_done(kiocb, ret);
2462 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2463 struct iov_iter *iter)
2465 struct io_ring_ctx *ctx = req->ctx;
2466 size_t len = req->rw.len;
2467 struct io_mapped_ubuf *imu;
2468 u16 index, buf_index;
2472 /* attempt to use fixed buffers without having provided iovecs */
2473 if (unlikely(!ctx->user_bufs))
2476 buf_index = req->buf_index;
2477 if (unlikely(buf_index >= ctx->nr_user_bufs))
2480 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2481 imu = &ctx->user_bufs[index];
2482 buf_addr = req->rw.addr;
2485 if (buf_addr + len < buf_addr)
2487 /* not inside the mapped region */
2488 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2492 * May not be a start of buffer, set size appropriately
2493 * and advance us to the beginning.
2495 offset = buf_addr - imu->ubuf;
2496 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2500 * Don't use iov_iter_advance() here, as it's really slow for
2501 * using the latter parts of a big fixed buffer - it iterates
2502 * over each segment manually. We can cheat a bit here, because
2505 * 1) it's a BVEC iter, we set it up
2506 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2507 * first and last bvec
2509 * So just find our index, and adjust the iterator afterwards.
2510 * If the offset is within the first bvec (or the whole first
2511 * bvec, just use iov_iter_advance(). This makes it easier
2512 * since we can just skip the first segment, which may not
2513 * be PAGE_SIZE aligned.
2515 const struct bio_vec *bvec = imu->bvec;
2517 if (offset <= bvec->bv_len) {
2518 iov_iter_advance(iter, offset);
2520 unsigned long seg_skip;
2522 /* skip first vec */
2523 offset -= bvec->bv_len;
2524 seg_skip = 1 + (offset >> PAGE_SHIFT);
2526 iter->bvec = bvec + seg_skip;
2527 iter->nr_segs -= seg_skip;
2528 iter->count -= bvec->bv_len + offset;
2529 iter->iov_offset = offset & ~PAGE_MASK;
2536 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2539 mutex_unlock(&ctx->uring_lock);
2542 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2545 * "Normal" inline submissions always hold the uring_lock, since we
2546 * grab it from the system call. Same is true for the SQPOLL offload.
2547 * The only exception is when we've detached the request and issue it
2548 * from an async worker thread, grab the lock for that case.
2551 mutex_lock(&ctx->uring_lock);
2554 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2555 int bgid, struct io_buffer *kbuf,
2558 struct io_buffer *head;
2560 if (req->flags & REQ_F_BUFFER_SELECTED)
2563 io_ring_submit_lock(req->ctx, needs_lock);
2565 lockdep_assert_held(&req->ctx->uring_lock);
2567 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2569 if (!list_empty(&head->list)) {
2570 kbuf = list_last_entry(&head->list, struct io_buffer,
2572 list_del(&kbuf->list);
2575 idr_remove(&req->ctx->io_buffer_idr, bgid);
2577 if (*len > kbuf->len)
2580 kbuf = ERR_PTR(-ENOBUFS);
2583 io_ring_submit_unlock(req->ctx, needs_lock);
2588 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2591 struct io_buffer *kbuf;
2594 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2595 bgid = req->buf_index;
2596 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2599 req->rw.addr = (u64) (unsigned long) kbuf;
2600 req->flags |= REQ_F_BUFFER_SELECTED;
2601 return u64_to_user_ptr(kbuf->addr);
2604 #ifdef CONFIG_COMPAT
2605 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2608 struct compat_iovec __user *uiov;
2609 compat_ssize_t clen;
2613 uiov = u64_to_user_ptr(req->rw.addr);
2614 if (!access_ok(uiov, sizeof(*uiov)))
2616 if (__get_user(clen, &uiov->iov_len))
2622 buf = io_rw_buffer_select(req, &len, needs_lock);
2624 return PTR_ERR(buf);
2625 iov[0].iov_base = buf;
2626 iov[0].iov_len = (compat_size_t) len;
2631 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2634 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2638 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2641 len = iov[0].iov_len;
2644 buf = io_rw_buffer_select(req, &len, needs_lock);
2646 return PTR_ERR(buf);
2647 iov[0].iov_base = buf;
2648 iov[0].iov_len = len;
2652 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2655 if (req->flags & REQ_F_BUFFER_SELECTED) {
2656 struct io_buffer *kbuf;
2658 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2659 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2660 iov[0].iov_len = kbuf->len;
2665 else if (req->rw.len > 1)
2668 #ifdef CONFIG_COMPAT
2669 if (req->ctx->compat)
2670 return io_compat_import(req, iov, needs_lock);
2673 return __io_iov_buffer_select(req, iov, needs_lock);
2676 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2677 struct iovec **iovec, struct iov_iter *iter,
2680 void __user *buf = u64_to_user_ptr(req->rw.addr);
2681 size_t sqe_len = req->rw.len;
2685 opcode = req->opcode;
2686 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2688 return io_import_fixed(req, rw, iter);
2691 /* buffer index only valid with fixed read/write, or buffer select */
2692 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2695 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2696 if (req->flags & REQ_F_BUFFER_SELECT) {
2697 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2700 return PTR_ERR(buf);
2702 req->rw.len = sqe_len;
2705 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2707 return ret < 0 ? ret : sqe_len;
2711 struct io_async_rw *iorw = &req->io->rw;
2714 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2715 if (iorw->iov == iorw->fast_iov)
2720 if (req->flags & REQ_F_BUFFER_SELECT) {
2721 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2723 ret = (*iovec)->iov_len;
2724 iov_iter_init(iter, rw, *iovec, 1, ret);
2730 #ifdef CONFIG_COMPAT
2731 if (req->ctx->compat)
2732 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2736 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2740 * For files that don't have ->read_iter() and ->write_iter(), handle them
2741 * by looping over ->read() or ->write() manually.
2743 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2744 struct iov_iter *iter)
2749 * Don't support polled IO through this interface, and we can't
2750 * support non-blocking either. For the latter, this just causes
2751 * the kiocb to be handled from an async context.
2753 if (kiocb->ki_flags & IOCB_HIPRI)
2755 if (kiocb->ki_flags & IOCB_NOWAIT)
2758 while (iov_iter_count(iter)) {
2762 if (!iov_iter_is_bvec(iter)) {
2763 iovec = iov_iter_iovec(iter);
2765 /* fixed buffers import bvec */
2766 iovec.iov_base = kmap(iter->bvec->bv_page)
2768 iovec.iov_len = min(iter->count,
2769 iter->bvec->bv_len - iter->iov_offset);
2773 nr = file->f_op->read(file, iovec.iov_base,
2774 iovec.iov_len, &kiocb->ki_pos);
2776 nr = file->f_op->write(file, iovec.iov_base,
2777 iovec.iov_len, &kiocb->ki_pos);
2780 if (iov_iter_is_bvec(iter))
2781 kunmap(iter->bvec->bv_page);
2789 if (nr != iovec.iov_len)
2791 iov_iter_advance(iter, nr);
2797 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2798 struct iovec *iovec, struct iovec *fast_iov,
2799 struct iov_iter *iter)
2801 req->io->rw.nr_segs = iter->nr_segs;
2802 req->io->rw.size = io_size;
2803 req->io->rw.iov = iovec;
2804 if (!req->io->rw.iov) {
2805 req->io->rw.iov = req->io->rw.fast_iov;
2806 if (req->io->rw.iov != fast_iov)
2807 memcpy(req->io->rw.iov, fast_iov,
2808 sizeof(struct iovec) * iter->nr_segs);
2810 req->flags |= REQ_F_NEED_CLEANUP;
2814 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2816 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2817 return req->io == NULL;
2820 static int io_alloc_async_ctx(struct io_kiocb *req)
2822 if (!io_op_defs[req->opcode].async_ctx)
2825 return __io_alloc_async_ctx(req);
2828 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2829 struct iovec *iovec, struct iovec *fast_iov,
2830 struct iov_iter *iter)
2832 if (!io_op_defs[req->opcode].async_ctx)
2835 if (__io_alloc_async_ctx(req))
2838 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2843 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2844 bool force_nonblock)
2846 struct io_async_ctx *io;
2847 struct iov_iter iter;
2850 ret = io_prep_rw(req, sqe, force_nonblock);
2854 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2857 /* either don't need iovec imported or already have it */
2858 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2862 io->rw.iov = io->rw.fast_iov;
2864 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2869 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2873 static void io_async_buf_cancel(struct callback_head *cb)
2875 struct io_async_rw *rw;
2876 struct io_kiocb *req;
2878 rw = container_of(cb, struct io_async_rw, task_work);
2879 req = rw->wpq.wait.private;
2880 __io_req_task_cancel(req, -ECANCELED);
2883 static void io_async_buf_retry(struct callback_head *cb)
2885 struct io_async_rw *rw;
2886 struct io_kiocb *req;
2888 rw = container_of(cb, struct io_async_rw, task_work);
2889 req = rw->wpq.wait.private;
2891 __io_req_task_submit(req);
2894 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2895 int sync, void *arg)
2897 struct wait_page_queue *wpq;
2898 struct io_kiocb *req = wait->private;
2899 struct io_async_rw *rw = &req->io->rw;
2900 struct wait_page_key *key = arg;
2901 struct task_struct *tsk;
2904 wpq = container_of(wait, struct wait_page_queue, wait);
2906 ret = wake_page_match(wpq, key);
2910 list_del_init(&wait->entry);
2912 init_task_work(&rw->task_work, io_async_buf_retry);
2913 /* submit ref gets dropped, acquire a new one */
2914 refcount_inc(&req->refs);
2916 ret = task_work_add(tsk, &rw->task_work, true);
2917 if (unlikely(ret)) {
2918 /* queue just for cancelation */
2919 init_task_work(&rw->task_work, io_async_buf_cancel);
2920 tsk = io_wq_get_task(req->ctx->io_wq);
2921 task_work_add(tsk, &rw->task_work, true);
2923 wake_up_process(tsk);
2927 static bool io_rw_should_retry(struct io_kiocb *req)
2929 struct kiocb *kiocb = &req->rw.kiocb;
2932 /* never retry for NOWAIT, we just complete with -EAGAIN */
2933 if (req->flags & REQ_F_NOWAIT)
2936 /* already tried, or we're doing O_DIRECT */
2937 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
2940 * just use poll if we can, and don't attempt if the fs doesn't
2941 * support callback based unlocks
2943 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
2947 * If request type doesn't require req->io to defer in general,
2948 * we need to allocate it here
2950 if (!req->io && __io_alloc_async_ctx(req))
2953 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
2954 io_async_buf_func, req);
2956 io_get_req_task(req);
2963 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
2965 if (req->file->f_op->read_iter)
2966 return call_read_iter(req->file, &req->rw.kiocb, iter);
2967 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
2970 static int io_read(struct io_kiocb *req, bool force_nonblock,
2971 struct io_comp_state *cs)
2973 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2974 struct kiocb *kiocb = &req->rw.kiocb;
2975 struct iov_iter iter;
2977 ssize_t io_size, ret;
2979 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2983 /* Ensure we clear previously set non-block flag */
2984 if (!force_nonblock)
2985 kiocb->ki_flags &= ~IOCB_NOWAIT;
2988 req->result = io_size;
2990 /* If the file doesn't support async, just async punt */
2991 if (force_nonblock && !io_file_supports_async(req->file, READ))
2994 iov_count = iov_iter_count(&iter);
2995 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2997 unsigned long nr_segs = iter.nr_segs;
3000 ret2 = io_iter_do_read(req, &iter);
3002 /* Catch -EAGAIN return for forced non-blocking submission */
3003 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3004 kiocb_done(kiocb, ret2, cs);
3006 iter.count = iov_count;
3007 iter.nr_segs = nr_segs;
3009 ret = io_setup_async_rw(req, io_size, iovec,
3010 inline_vecs, &iter);
3013 /* if we can retry, do so with the callbacks armed */
3014 if (io_rw_should_retry(req)) {
3015 ret2 = io_iter_do_read(req, &iter);
3016 if (ret2 == -EIOCBQUEUED) {
3018 } else if (ret2 != -EAGAIN) {
3019 kiocb_done(kiocb, ret2, cs);
3023 kiocb->ki_flags &= ~IOCB_WAITQ;
3028 if (!(req->flags & REQ_F_NEED_CLEANUP))
3033 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3034 bool force_nonblock)
3036 struct io_async_ctx *io;
3037 struct iov_iter iter;
3040 ret = io_prep_rw(req, sqe, force_nonblock);
3044 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3047 req->fsize = rlimit(RLIMIT_FSIZE);
3049 /* either don't need iovec imported or already have it */
3050 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3054 io->rw.iov = io->rw.fast_iov;
3056 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
3061 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
3065 static int io_write(struct io_kiocb *req, bool force_nonblock,
3066 struct io_comp_state *cs)
3068 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3069 struct kiocb *kiocb = &req->rw.kiocb;
3070 struct iov_iter iter;
3072 ssize_t ret, io_size;
3074 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3078 /* Ensure we clear previously set non-block flag */
3079 if (!force_nonblock)
3080 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3083 req->result = io_size;
3085 /* If the file doesn't support async, just async punt */
3086 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3089 /* file path doesn't support NOWAIT for non-direct_IO */
3090 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3091 (req->flags & REQ_F_ISREG))
3094 iov_count = iov_iter_count(&iter);
3095 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3097 unsigned long nr_segs = iter.nr_segs;
3101 * Open-code file_start_write here to grab freeze protection,
3102 * which will be released by another thread in
3103 * io_complete_rw(). Fool lockdep by telling it the lock got
3104 * released so that it doesn't complain about the held lock when
3105 * we return to userspace.
3107 if (req->flags & REQ_F_ISREG) {
3108 __sb_start_write(file_inode(req->file)->i_sb,
3109 SB_FREEZE_WRITE, true);
3110 __sb_writers_release(file_inode(req->file)->i_sb,
3113 kiocb->ki_flags |= IOCB_WRITE;
3115 if (!force_nonblock)
3116 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3118 if (req->file->f_op->write_iter)
3119 ret2 = call_write_iter(req->file, kiocb, &iter);
3121 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3123 if (!force_nonblock)
3124 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3127 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3128 * retry them without IOCB_NOWAIT.
3130 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3132 if (!force_nonblock || ret2 != -EAGAIN) {
3133 kiocb_done(kiocb, ret2, cs);
3135 iter.count = iov_count;
3136 iter.nr_segs = nr_segs;
3138 ret = io_setup_async_rw(req, io_size, iovec,
3139 inline_vecs, &iter);
3146 if (!(req->flags & REQ_F_NEED_CLEANUP))
3151 static int __io_splice_prep(struct io_kiocb *req,
3152 const struct io_uring_sqe *sqe)
3154 struct io_splice* sp = &req->splice;
3155 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3158 if (req->flags & REQ_F_NEED_CLEANUP)
3160 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3164 sp->len = READ_ONCE(sqe->len);
3165 sp->flags = READ_ONCE(sqe->splice_flags);
3167 if (unlikely(sp->flags & ~valid_flags))
3170 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3171 (sp->flags & SPLICE_F_FD_IN_FIXED));
3174 req->flags |= REQ_F_NEED_CLEANUP;
3176 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3178 * Splice operation will be punted aync, and here need to
3179 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3181 io_req_init_async(req);
3182 req->work.flags |= IO_WQ_WORK_UNBOUND;
3188 static int io_tee_prep(struct io_kiocb *req,
3189 const struct io_uring_sqe *sqe)
3191 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3193 return __io_splice_prep(req, sqe);
3196 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3198 struct io_splice *sp = &req->splice;
3199 struct file *in = sp->file_in;
3200 struct file *out = sp->file_out;
3201 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3207 ret = do_tee(in, out, sp->len, flags);
3209 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3210 req->flags &= ~REQ_F_NEED_CLEANUP;
3213 req_set_fail_links(req);
3214 io_req_complete(req, ret);
3218 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3220 struct io_splice* sp = &req->splice;
3222 sp->off_in = READ_ONCE(sqe->splice_off_in);
3223 sp->off_out = READ_ONCE(sqe->off);
3224 return __io_splice_prep(req, sqe);
3227 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3229 struct io_splice *sp = &req->splice;
3230 struct file *in = sp->file_in;
3231 struct file *out = sp->file_out;
3232 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3233 loff_t *poff_in, *poff_out;
3239 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3240 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3243 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3245 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3246 req->flags &= ~REQ_F_NEED_CLEANUP;
3249 req_set_fail_links(req);
3250 io_req_complete(req, ret);
3255 * IORING_OP_NOP just posts a completion event, nothing else.
3257 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3259 struct io_ring_ctx *ctx = req->ctx;
3261 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3264 __io_req_complete(req, 0, 0, cs);
3268 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3270 struct io_ring_ctx *ctx = req->ctx;
3275 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3277 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3280 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3281 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3284 req->sync.off = READ_ONCE(sqe->off);
3285 req->sync.len = READ_ONCE(sqe->len);
3289 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3291 loff_t end = req->sync.off + req->sync.len;
3294 /* fsync always requires a blocking context */
3298 ret = vfs_fsync_range(req->file, req->sync.off,
3299 end > 0 ? end : LLONG_MAX,
3300 req->sync.flags & IORING_FSYNC_DATASYNC);
3302 req_set_fail_links(req);
3303 io_req_complete(req, ret);
3307 static int io_fallocate_prep(struct io_kiocb *req,
3308 const struct io_uring_sqe *sqe)
3310 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3312 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3315 req->sync.off = READ_ONCE(sqe->off);
3316 req->sync.len = READ_ONCE(sqe->addr);
3317 req->sync.mode = READ_ONCE(sqe->len);
3318 req->fsize = rlimit(RLIMIT_FSIZE);
3322 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3326 /* fallocate always requiring blocking context */
3330 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3331 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3333 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3335 req_set_fail_links(req);
3336 io_req_complete(req, ret);
3340 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3342 const char __user *fname;
3345 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3347 if (unlikely(sqe->ioprio || sqe->buf_index))
3349 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3352 /* open.how should be already initialised */
3353 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3354 req->open.how.flags |= O_LARGEFILE;
3356 req->open.dfd = READ_ONCE(sqe->fd);
3357 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3358 req->open.filename = getname(fname);
3359 if (IS_ERR(req->open.filename)) {
3360 ret = PTR_ERR(req->open.filename);
3361 req->open.filename = NULL;
3364 req->open.nofile = rlimit(RLIMIT_NOFILE);
3365 req->flags |= REQ_F_NEED_CLEANUP;
3369 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3373 if (req->flags & REQ_F_NEED_CLEANUP)
3375 mode = READ_ONCE(sqe->len);
3376 flags = READ_ONCE(sqe->open_flags);
3377 req->open.how = build_open_how(flags, mode);
3378 return __io_openat_prep(req, sqe);
3381 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3383 struct open_how __user *how;
3387 if (req->flags & REQ_F_NEED_CLEANUP)
3389 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3390 len = READ_ONCE(sqe->len);
3391 if (len < OPEN_HOW_SIZE_VER0)
3394 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3399 return __io_openat_prep(req, sqe);
3402 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3404 struct open_flags op;
3411 ret = build_open_flags(&req->open.how, &op);
3415 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3419 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3422 ret = PTR_ERR(file);
3424 fsnotify_open(file);
3425 fd_install(ret, file);
3428 putname(req->open.filename);
3429 req->flags &= ~REQ_F_NEED_CLEANUP;
3431 req_set_fail_links(req);
3432 io_req_complete(req, ret);
3436 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3438 return io_openat2(req, force_nonblock);
3441 static int io_remove_buffers_prep(struct io_kiocb *req,
3442 const struct io_uring_sqe *sqe)
3444 struct io_provide_buf *p = &req->pbuf;
3447 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3450 tmp = READ_ONCE(sqe->fd);
3451 if (!tmp || tmp > USHRT_MAX)
3454 memset(p, 0, sizeof(*p));
3456 p->bgid = READ_ONCE(sqe->buf_group);
3460 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3461 int bgid, unsigned nbufs)
3465 /* shouldn't happen */
3469 /* the head kbuf is the list itself */
3470 while (!list_empty(&buf->list)) {
3471 struct io_buffer *nxt;
3473 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3474 list_del(&nxt->list);
3481 idr_remove(&ctx->io_buffer_idr, bgid);
3486 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3487 struct io_comp_state *cs)
3489 struct io_provide_buf *p = &req->pbuf;
3490 struct io_ring_ctx *ctx = req->ctx;
3491 struct io_buffer *head;
3494 io_ring_submit_lock(ctx, !force_nonblock);
3496 lockdep_assert_held(&ctx->uring_lock);
3499 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3501 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3503 io_ring_submit_lock(ctx, !force_nonblock);
3505 req_set_fail_links(req);
3506 __io_req_complete(req, ret, 0, cs);
3510 static int io_provide_buffers_prep(struct io_kiocb *req,
3511 const struct io_uring_sqe *sqe)
3513 struct io_provide_buf *p = &req->pbuf;
3516 if (sqe->ioprio || sqe->rw_flags)
3519 tmp = READ_ONCE(sqe->fd);
3520 if (!tmp || tmp > USHRT_MAX)
3523 p->addr = READ_ONCE(sqe->addr);
3524 p->len = READ_ONCE(sqe->len);
3526 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3529 p->bgid = READ_ONCE(sqe->buf_group);
3530 tmp = READ_ONCE(sqe->off);
3531 if (tmp > USHRT_MAX)
3537 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3539 struct io_buffer *buf;
3540 u64 addr = pbuf->addr;
3541 int i, bid = pbuf->bid;
3543 for (i = 0; i < pbuf->nbufs; i++) {
3544 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3549 buf->len = pbuf->len;
3554 INIT_LIST_HEAD(&buf->list);
3557 list_add_tail(&buf->list, &(*head)->list);
3561 return i ? i : -ENOMEM;
3564 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3565 struct io_comp_state *cs)
3567 struct io_provide_buf *p = &req->pbuf;
3568 struct io_ring_ctx *ctx = req->ctx;
3569 struct io_buffer *head, *list;
3572 io_ring_submit_lock(ctx, !force_nonblock);
3574 lockdep_assert_held(&ctx->uring_lock);
3576 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3578 ret = io_add_buffers(p, &head);
3583 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3586 __io_remove_buffers(ctx, head, p->bgid, -1U);
3591 io_ring_submit_unlock(ctx, !force_nonblock);
3593 req_set_fail_links(req);
3594 __io_req_complete(req, ret, 0, cs);
3598 static int io_epoll_ctl_prep(struct io_kiocb *req,
3599 const struct io_uring_sqe *sqe)
3601 #if defined(CONFIG_EPOLL)
3602 if (sqe->ioprio || sqe->buf_index)
3604 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3607 req->epoll.epfd = READ_ONCE(sqe->fd);
3608 req->epoll.op = READ_ONCE(sqe->len);
3609 req->epoll.fd = READ_ONCE(sqe->off);
3611 if (ep_op_has_event(req->epoll.op)) {
3612 struct epoll_event __user *ev;
3614 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3615 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3625 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3626 struct io_comp_state *cs)
3628 #if defined(CONFIG_EPOLL)
3629 struct io_epoll *ie = &req->epoll;
3632 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3633 if (force_nonblock && ret == -EAGAIN)
3637 req_set_fail_links(req);
3638 __io_req_complete(req, ret, 0, cs);
3645 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3647 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3648 if (sqe->ioprio || sqe->buf_index || sqe->off)
3650 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3653 req->madvise.addr = READ_ONCE(sqe->addr);
3654 req->madvise.len = READ_ONCE(sqe->len);
3655 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3662 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3664 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3665 struct io_madvise *ma = &req->madvise;
3671 ret = do_madvise(ma->addr, ma->len, ma->advice);
3673 req_set_fail_links(req);
3674 io_req_complete(req, ret);
3681 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3683 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3685 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3688 req->fadvise.offset = READ_ONCE(sqe->off);
3689 req->fadvise.len = READ_ONCE(sqe->len);
3690 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3694 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3696 struct io_fadvise *fa = &req->fadvise;
3699 if (force_nonblock) {
3700 switch (fa->advice) {
3701 case POSIX_FADV_NORMAL:
3702 case POSIX_FADV_RANDOM:
3703 case POSIX_FADV_SEQUENTIAL:
3710 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3712 req_set_fail_links(req);
3713 io_req_complete(req, ret);
3717 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3719 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3721 if (sqe->ioprio || sqe->buf_index)
3723 if (req->flags & REQ_F_FIXED_FILE)
3726 req->statx.dfd = READ_ONCE(sqe->fd);
3727 req->statx.mask = READ_ONCE(sqe->len);
3728 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3729 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3730 req->statx.flags = READ_ONCE(sqe->statx_flags);
3735 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3737 struct io_statx *ctx = &req->statx;
3740 if (force_nonblock) {
3741 /* only need file table for an actual valid fd */
3742 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3743 req->flags |= REQ_F_NO_FILE_TABLE;
3747 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3751 req_set_fail_links(req);
3752 io_req_complete(req, ret);
3756 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3759 * If we queue this for async, it must not be cancellable. That would
3760 * leave the 'file' in an undeterminate state, and here need to modify
3761 * io_wq_work.flags, so initialize io_wq_work firstly.
3763 io_req_init_async(req);
3764 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3766 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3768 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3769 sqe->rw_flags || sqe->buf_index)
3771 if (req->flags & REQ_F_FIXED_FILE)
3774 req->close.fd = READ_ONCE(sqe->fd);
3775 if ((req->file && req->file->f_op == &io_uring_fops) ||
3776 req->close.fd == req->ctx->ring_fd)
3779 req->close.put_file = NULL;
3783 static int io_close(struct io_kiocb *req, bool force_nonblock,
3784 struct io_comp_state *cs)
3786 struct io_close *close = &req->close;
3789 /* might be already done during nonblock submission */
3790 if (!close->put_file) {
3791 ret = __close_fd_get_file(close->fd, &close->put_file);
3793 return (ret == -ENOENT) ? -EBADF : ret;
3796 /* if the file has a flush method, be safe and punt to async */
3797 if (close->put_file->f_op->flush && force_nonblock) {
3798 /* was never set, but play safe */
3799 req->flags &= ~REQ_F_NOWAIT;
3800 /* avoid grabbing files - we don't need the files */
3801 req->flags |= REQ_F_NO_FILE_TABLE;
3805 /* No ->flush() or already async, safely close from here */
3806 ret = filp_close(close->put_file, req->work.files);
3808 req_set_fail_links(req);
3809 fput(close->put_file);
3810 close->put_file = NULL;
3811 __io_req_complete(req, ret, 0, cs);
3815 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3817 struct io_ring_ctx *ctx = req->ctx;
3822 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3824 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3827 req->sync.off = READ_ONCE(sqe->off);
3828 req->sync.len = READ_ONCE(sqe->len);
3829 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3833 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3837 /* sync_file_range always requires a blocking context */
3841 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3844 req_set_fail_links(req);
3845 io_req_complete(req, ret);
3849 #if defined(CONFIG_NET)
3850 static int io_setup_async_msg(struct io_kiocb *req,
3851 struct io_async_msghdr *kmsg)
3855 if (io_alloc_async_ctx(req)) {
3856 if (kmsg->iov != kmsg->fast_iov)
3860 req->flags |= REQ_F_NEED_CLEANUP;
3861 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3865 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3867 struct io_sr_msg *sr = &req->sr_msg;
3868 struct io_async_ctx *io = req->io;
3871 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3874 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3875 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3876 sr->len = READ_ONCE(sqe->len);
3878 #ifdef CONFIG_COMPAT
3879 if (req->ctx->compat)
3880 sr->msg_flags |= MSG_CMSG_COMPAT;
3883 if (!io || req->opcode == IORING_OP_SEND)
3885 /* iovec is already imported */
3886 if (req->flags & REQ_F_NEED_CLEANUP)
3889 io->msg.iov = io->msg.fast_iov;
3890 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3893 req->flags |= REQ_F_NEED_CLEANUP;
3897 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3898 struct io_comp_state *cs)
3900 struct io_async_msghdr *kmsg = NULL;
3901 struct socket *sock;
3904 sock = sock_from_file(req->file, &ret);
3906 struct io_async_ctx io;
3910 kmsg = &req->io->msg;
3911 kmsg->msg.msg_name = &req->io->msg.addr;
3912 /* if iov is set, it's allocated already */
3914 kmsg->iov = kmsg->fast_iov;
3915 kmsg->msg.msg_iter.iov = kmsg->iov;
3917 struct io_sr_msg *sr = &req->sr_msg;
3920 kmsg->msg.msg_name = &io.msg.addr;
3922 io.msg.iov = io.msg.fast_iov;
3923 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3924 sr->msg_flags, &io.msg.iov);
3929 flags = req->sr_msg.msg_flags;
3930 if (flags & MSG_DONTWAIT)
3931 req->flags |= REQ_F_NOWAIT;
3932 else if (force_nonblock)
3933 flags |= MSG_DONTWAIT;
3935 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3936 if (force_nonblock && ret == -EAGAIN)
3937 return io_setup_async_msg(req, kmsg);
3938 if (ret == -ERESTARTSYS)
3942 if (kmsg && kmsg->iov != kmsg->fast_iov)
3944 req->flags &= ~REQ_F_NEED_CLEANUP;
3946 req_set_fail_links(req);
3947 __io_req_complete(req, ret, 0, cs);
3951 static int io_send(struct io_kiocb *req, bool force_nonblock,
3952 struct io_comp_state *cs)
3954 struct socket *sock;
3957 sock = sock_from_file(req->file, &ret);
3959 struct io_sr_msg *sr = &req->sr_msg;
3964 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3969 msg.msg_name = NULL;
3970 msg.msg_control = NULL;
3971 msg.msg_controllen = 0;
3972 msg.msg_namelen = 0;
3974 flags = req->sr_msg.msg_flags;
3975 if (flags & MSG_DONTWAIT)
3976 req->flags |= REQ_F_NOWAIT;
3977 else if (force_nonblock)
3978 flags |= MSG_DONTWAIT;
3980 msg.msg_flags = flags;
3981 ret = sock_sendmsg(sock, &msg);
3982 if (force_nonblock && ret == -EAGAIN)
3984 if (ret == -ERESTARTSYS)
3989 req_set_fail_links(req);
3990 __io_req_complete(req, ret, 0, cs);
3994 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3996 struct io_sr_msg *sr = &req->sr_msg;
3997 struct iovec __user *uiov;
4001 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
4006 if (req->flags & REQ_F_BUFFER_SELECT) {
4009 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
4011 sr->len = io->msg.iov[0].iov_len;
4012 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
4016 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4017 &io->msg.iov, &io->msg.msg.msg_iter);
4025 #ifdef CONFIG_COMPAT
4026 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4027 struct io_async_ctx *io)
4029 struct compat_msghdr __user *msg_compat;
4030 struct io_sr_msg *sr = &req->sr_msg;
4031 struct compat_iovec __user *uiov;
4036 msg_compat = (struct compat_msghdr __user *) sr->msg;
4037 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
4042 uiov = compat_ptr(ptr);
4043 if (req->flags & REQ_F_BUFFER_SELECT) {
4044 compat_ssize_t clen;
4048 if (!access_ok(uiov, sizeof(*uiov)))
4050 if (__get_user(clen, &uiov->iov_len))
4054 sr->len = io->msg.iov[0].iov_len;
4057 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4059 &io->msg.msg.msg_iter);
4068 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
4070 io->msg.iov = io->msg.fast_iov;
4072 #ifdef CONFIG_COMPAT
4073 if (req->ctx->compat)
4074 return __io_compat_recvmsg_copy_hdr(req, io);
4077 return __io_recvmsg_copy_hdr(req, io);
4080 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4081 int *cflags, bool needs_lock)
4083 struct io_sr_msg *sr = &req->sr_msg;
4084 struct io_buffer *kbuf;
4086 if (!(req->flags & REQ_F_BUFFER_SELECT))
4089 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4094 req->flags |= REQ_F_BUFFER_SELECTED;
4096 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
4097 *cflags |= IORING_CQE_F_BUFFER;
4101 static int io_recvmsg_prep(struct io_kiocb *req,
4102 const struct io_uring_sqe *sqe)
4104 struct io_sr_msg *sr = &req->sr_msg;
4105 struct io_async_ctx *io = req->io;
4108 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4111 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4112 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4113 sr->len = READ_ONCE(sqe->len);
4114 sr->bgid = READ_ONCE(sqe->buf_group);
4116 #ifdef CONFIG_COMPAT
4117 if (req->ctx->compat)
4118 sr->msg_flags |= MSG_CMSG_COMPAT;
4121 if (!io || req->opcode == IORING_OP_RECV)
4123 /* iovec is already imported */
4124 if (req->flags & REQ_F_NEED_CLEANUP)
4127 ret = io_recvmsg_copy_hdr(req, io);
4129 req->flags |= REQ_F_NEED_CLEANUP;
4133 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4134 struct io_comp_state *cs)
4136 struct io_async_msghdr *kmsg = NULL;
4137 struct socket *sock;
4138 int ret, cflags = 0;
4140 sock = sock_from_file(req->file, &ret);
4142 struct io_buffer *kbuf;
4143 struct io_async_ctx io;
4147 kmsg = &req->io->msg;
4148 kmsg->msg.msg_name = &req->io->msg.addr;
4149 /* if iov is set, it's allocated already */
4151 kmsg->iov = kmsg->fast_iov;
4152 kmsg->msg.msg_iter.iov = kmsg->iov;
4155 kmsg->msg.msg_name = &io.msg.addr;
4157 ret = io_recvmsg_copy_hdr(req, &io);
4162 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4164 return PTR_ERR(kbuf);
4166 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4167 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4168 1, req->sr_msg.len);
4171 flags = req->sr_msg.msg_flags;
4172 if (flags & MSG_DONTWAIT)
4173 req->flags |= REQ_F_NOWAIT;
4174 else if (force_nonblock)
4175 flags |= MSG_DONTWAIT;
4177 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
4178 kmsg->uaddr, flags);
4179 if (force_nonblock && ret == -EAGAIN)
4180 return io_setup_async_msg(req, kmsg);
4181 if (ret == -ERESTARTSYS)
4185 if (kmsg && kmsg->iov != kmsg->fast_iov)
4187 req->flags &= ~REQ_F_NEED_CLEANUP;
4189 req_set_fail_links(req);
4190 __io_req_complete(req, ret, cflags, cs);
4194 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4195 struct io_comp_state *cs)
4197 struct io_buffer *kbuf = NULL;
4198 struct socket *sock;
4199 int ret, cflags = 0;
4201 sock = sock_from_file(req->file, &ret);
4203 struct io_sr_msg *sr = &req->sr_msg;
4204 void __user *buf = sr->buf;
4209 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4211 return PTR_ERR(kbuf);
4213 buf = u64_to_user_ptr(kbuf->addr);
4215 ret = import_single_range(READ, buf, sr->len, &iov,
4222 req->flags |= REQ_F_NEED_CLEANUP;
4223 msg.msg_name = NULL;
4224 msg.msg_control = NULL;
4225 msg.msg_controllen = 0;
4226 msg.msg_namelen = 0;
4227 msg.msg_iocb = NULL;
4230 flags = req->sr_msg.msg_flags;
4231 if (flags & MSG_DONTWAIT)
4232 req->flags |= REQ_F_NOWAIT;
4233 else if (force_nonblock)
4234 flags |= MSG_DONTWAIT;
4236 ret = sock_recvmsg(sock, &msg, flags);
4237 if (force_nonblock && ret == -EAGAIN)
4239 if (ret == -ERESTARTSYS)
4244 req->flags &= ~REQ_F_NEED_CLEANUP;
4246 req_set_fail_links(req);
4247 __io_req_complete(req, ret, cflags, cs);
4251 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4253 struct io_accept *accept = &req->accept;
4255 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4257 if (sqe->ioprio || sqe->len || sqe->buf_index)
4260 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4261 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4262 accept->flags = READ_ONCE(sqe->accept_flags);
4263 accept->nofile = rlimit(RLIMIT_NOFILE);
4267 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4268 struct io_comp_state *cs)
4270 struct io_accept *accept = &req->accept;
4271 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4274 if (req->file->f_flags & O_NONBLOCK)
4275 req->flags |= REQ_F_NOWAIT;
4277 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4278 accept->addr_len, accept->flags,
4280 if (ret == -EAGAIN && force_nonblock)
4283 if (ret == -ERESTARTSYS)
4285 req_set_fail_links(req);
4287 __io_req_complete(req, ret, 0, cs);
4291 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4293 struct io_connect *conn = &req->connect;
4294 struct io_async_ctx *io = req->io;
4296 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4298 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4301 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4302 conn->addr_len = READ_ONCE(sqe->addr2);
4307 return move_addr_to_kernel(conn->addr, conn->addr_len,
4308 &io->connect.address);
4311 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4312 struct io_comp_state *cs)
4314 struct io_async_ctx __io, *io;
4315 unsigned file_flags;
4321 ret = move_addr_to_kernel(req->connect.addr,
4322 req->connect.addr_len,
4323 &__io.connect.address);
4329 file_flags = force_nonblock ? O_NONBLOCK : 0;
4331 ret = __sys_connect_file(req->file, &io->connect.address,
4332 req->connect.addr_len, file_flags);
4333 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4336 if (io_alloc_async_ctx(req)) {
4340 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4343 if (ret == -ERESTARTSYS)
4347 req_set_fail_links(req);
4348 __io_req_complete(req, ret, 0, cs);
4351 #else /* !CONFIG_NET */
4352 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4357 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4358 struct io_comp_state *cs)
4363 static int io_send(struct io_kiocb *req, bool force_nonblock,
4364 struct io_comp_state *cs)
4369 static int io_recvmsg_prep(struct io_kiocb *req,
4370 const struct io_uring_sqe *sqe)
4375 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4376 struct io_comp_state *cs)
4381 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4382 struct io_comp_state *cs)
4387 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4392 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4393 struct io_comp_state *cs)
4398 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4403 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4404 struct io_comp_state *cs)
4408 #endif /* CONFIG_NET */
4410 struct io_poll_table {
4411 struct poll_table_struct pt;
4412 struct io_kiocb *req;
4416 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4417 __poll_t mask, task_work_func_t func)
4419 struct task_struct *tsk;
4422 /* for instances that support it check for an event match first: */
4423 if (mask && !(mask & poll->events))
4426 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4428 list_del_init(&poll->wait.entry);
4432 init_task_work(&req->task_work, func);
4434 * If this fails, then the task is exiting. When a task exits, the
4435 * work gets canceled, so just cancel this request as well instead
4436 * of executing it. We can't safely execute it anyway, as we may not
4437 * have the needed state needed for it anyway.
4439 ret = task_work_add(tsk, &req->task_work, true);
4440 if (unlikely(ret)) {
4441 WRITE_ONCE(poll->canceled, true);
4442 tsk = io_wq_get_task(req->ctx->io_wq);
4443 task_work_add(tsk, &req->task_work, true);
4445 wake_up_process(tsk);
4449 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4450 __acquires(&req->ctx->completion_lock)
4452 struct io_ring_ctx *ctx = req->ctx;
4454 if (!req->result && !READ_ONCE(poll->canceled)) {
4455 struct poll_table_struct pt = { ._key = poll->events };
4457 req->result = vfs_poll(req->file, &pt) & poll->events;
4460 spin_lock_irq(&ctx->completion_lock);
4461 if (!req->result && !READ_ONCE(poll->canceled)) {
4462 add_wait_queue(poll->head, &poll->wait);
4469 static void io_poll_remove_double(struct io_kiocb *req)
4471 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4473 lockdep_assert_held(&req->ctx->completion_lock);
4475 if (poll && poll->head) {
4476 struct wait_queue_head *head = poll->head;
4478 spin_lock(&head->lock);
4479 list_del_init(&poll->wait.entry);
4480 if (poll->wait.private)
4481 refcount_dec(&req->refs);
4483 spin_unlock(&head->lock);
4487 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4489 struct io_ring_ctx *ctx = req->ctx;
4491 io_poll_remove_double(req);
4492 req->poll.done = true;
4493 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4494 io_commit_cqring(ctx);
4497 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4499 struct io_ring_ctx *ctx = req->ctx;
4501 if (io_poll_rewait(req, &req->poll)) {
4502 spin_unlock_irq(&ctx->completion_lock);
4506 hash_del(&req->hash_node);
4507 io_poll_complete(req, req->result, 0);
4508 req->flags |= REQ_F_COMP_LOCKED;
4509 *nxt = io_put_req_find_next(req);
4510 spin_unlock_irq(&ctx->completion_lock);
4512 io_cqring_ev_posted(ctx);
4515 static void io_poll_task_func(struct callback_head *cb)
4517 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4518 struct io_kiocb *nxt = NULL;
4520 io_poll_task_handler(req, &nxt);
4522 struct io_ring_ctx *ctx = nxt->ctx;
4524 mutex_lock(&ctx->uring_lock);
4525 __io_queue_sqe(nxt, NULL, NULL);
4526 mutex_unlock(&ctx->uring_lock);
4530 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4531 int sync, void *key)
4533 struct io_kiocb *req = wait->private;
4534 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4535 __poll_t mask = key_to_poll(key);
4537 /* for instances that support it check for an event match first: */
4538 if (mask && !(mask & poll->events))
4541 if (req->poll.head) {
4544 spin_lock(&req->poll.head->lock);
4545 done = list_empty(&req->poll.wait.entry);
4547 list_del_init(&req->poll.wait.entry);
4548 spin_unlock(&req->poll.head->lock);
4550 __io_async_wake(req, poll, mask, io_poll_task_func);
4552 refcount_dec(&req->refs);
4556 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4557 wait_queue_func_t wake_func)
4561 poll->canceled = false;
4562 poll->events = events;
4563 INIT_LIST_HEAD(&poll->wait.entry);
4564 init_waitqueue_func_entry(&poll->wait, wake_func);
4567 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4568 struct wait_queue_head *head)
4570 struct io_kiocb *req = pt->req;
4573 * If poll->head is already set, it's because the file being polled
4574 * uses multiple waitqueues for poll handling (eg one for read, one
4575 * for write). Setup a separate io_poll_iocb if this happens.
4577 if (unlikely(poll->head)) {
4578 /* already have a 2nd entry, fail a third attempt */
4580 pt->error = -EINVAL;
4583 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4585 pt->error = -ENOMEM;
4588 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4589 refcount_inc(&req->refs);
4590 poll->wait.private = req;
4591 req->io = (void *) poll;
4597 if (poll->events & EPOLLEXCLUSIVE)
4598 add_wait_queue_exclusive(head, &poll->wait);
4600 add_wait_queue(head, &poll->wait);
4603 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4604 struct poll_table_struct *p)
4606 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4608 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4611 static void io_async_task_func(struct callback_head *cb)
4613 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4614 struct async_poll *apoll = req->apoll;
4615 struct io_ring_ctx *ctx = req->ctx;
4616 bool canceled = false;
4618 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4620 if (io_poll_rewait(req, &apoll->poll)) {
4621 spin_unlock_irq(&ctx->completion_lock);
4625 /* If req is still hashed, it cannot have been canceled. Don't check. */
4626 if (hash_hashed(&req->hash_node)) {
4627 hash_del(&req->hash_node);
4629 canceled = READ_ONCE(apoll->poll.canceled);
4631 io_cqring_fill_event(req, -ECANCELED);
4632 io_commit_cqring(ctx);
4636 spin_unlock_irq(&ctx->completion_lock);
4638 /* restore ->work in case we need to retry again */
4639 if (req->flags & REQ_F_WORK_INITIALIZED)
4640 memcpy(&req->work, &apoll->work, sizeof(req->work));
4644 __set_current_state(TASK_RUNNING);
4645 if (io_sq_thread_acquire_mm(ctx, req)) {
4646 io_cqring_add_event(req, -EFAULT, 0);
4649 mutex_lock(&ctx->uring_lock);
4650 __io_queue_sqe(req, NULL, NULL);
4651 mutex_unlock(&ctx->uring_lock);
4653 io_cqring_ev_posted(ctx);
4655 req_set_fail_links(req);
4656 io_double_put_req(req);
4660 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4663 struct io_kiocb *req = wait->private;
4664 struct io_poll_iocb *poll = &req->apoll->poll;
4666 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4669 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4672 static void io_poll_req_insert(struct io_kiocb *req)
4674 struct io_ring_ctx *ctx = req->ctx;
4675 struct hlist_head *list;
4677 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4678 hlist_add_head(&req->hash_node, list);
4681 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4682 struct io_poll_iocb *poll,
4683 struct io_poll_table *ipt, __poll_t mask,
4684 wait_queue_func_t wake_func)
4685 __acquires(&ctx->completion_lock)
4687 struct io_ring_ctx *ctx = req->ctx;
4688 bool cancel = false;
4690 io_init_poll_iocb(poll, mask, wake_func);
4691 poll->file = req->file;
4692 poll->wait.private = req;
4694 ipt->pt._key = mask;
4696 ipt->error = -EINVAL;
4698 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4700 spin_lock_irq(&ctx->completion_lock);
4701 if (likely(poll->head)) {
4702 spin_lock(&poll->head->lock);
4703 if (unlikely(list_empty(&poll->wait.entry))) {
4709 if (mask || ipt->error)
4710 list_del_init(&poll->wait.entry);
4712 WRITE_ONCE(poll->canceled, true);
4713 else if (!poll->done) /* actually waiting for an event */
4714 io_poll_req_insert(req);
4715 spin_unlock(&poll->head->lock);
4721 static bool io_arm_poll_handler(struct io_kiocb *req)
4723 const struct io_op_def *def = &io_op_defs[req->opcode];
4724 struct io_ring_ctx *ctx = req->ctx;
4725 struct async_poll *apoll;
4726 struct io_poll_table ipt;
4730 if (!req->file || !file_can_poll(req->file))
4732 if (req->flags & REQ_F_POLLED)
4734 if (!def->pollin && !def->pollout)
4737 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4738 if (unlikely(!apoll))
4741 req->flags |= REQ_F_POLLED;
4742 if (req->flags & REQ_F_WORK_INITIALIZED)
4743 memcpy(&apoll->work, &req->work, sizeof(req->work));
4744 had_io = req->io != NULL;
4746 io_get_req_task(req);
4748 INIT_HLIST_NODE(&req->hash_node);
4752 mask |= POLLIN | POLLRDNORM;
4754 mask |= POLLOUT | POLLWRNORM;
4755 mask |= POLLERR | POLLPRI;
4757 ipt.pt._qproc = io_async_queue_proc;
4759 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4763 /* only remove double add if we did it here */
4765 io_poll_remove_double(req);
4766 spin_unlock_irq(&ctx->completion_lock);
4767 if (req->flags & REQ_F_WORK_INITIALIZED)
4768 memcpy(&req->work, &apoll->work, sizeof(req->work));
4772 spin_unlock_irq(&ctx->completion_lock);
4773 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4774 apoll->poll.events);
4778 static bool __io_poll_remove_one(struct io_kiocb *req,
4779 struct io_poll_iocb *poll)
4781 bool do_complete = false;
4783 spin_lock(&poll->head->lock);
4784 WRITE_ONCE(poll->canceled, true);
4785 if (!list_empty(&poll->wait.entry)) {
4786 list_del_init(&poll->wait.entry);
4789 spin_unlock(&poll->head->lock);
4790 hash_del(&req->hash_node);
4794 static bool io_poll_remove_one(struct io_kiocb *req)
4798 if (req->opcode == IORING_OP_POLL_ADD) {
4799 io_poll_remove_double(req);
4800 do_complete = __io_poll_remove_one(req, &req->poll);
4802 struct async_poll *apoll = req->apoll;
4804 /* non-poll requests have submit ref still */
4805 do_complete = __io_poll_remove_one(req, &apoll->poll);
4809 * restore ->work because we will call
4810 * io_req_work_drop_env below when dropping the
4813 if (req->flags & REQ_F_WORK_INITIALIZED)
4814 memcpy(&req->work, &apoll->work,
4821 io_cqring_fill_event(req, -ECANCELED);
4822 io_commit_cqring(req->ctx);
4823 req->flags |= REQ_F_COMP_LOCKED;
4830 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4832 struct hlist_node *tmp;
4833 struct io_kiocb *req;
4836 spin_lock_irq(&ctx->completion_lock);
4837 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4838 struct hlist_head *list;
4840 list = &ctx->cancel_hash[i];
4841 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4842 posted += io_poll_remove_one(req);
4844 spin_unlock_irq(&ctx->completion_lock);
4847 io_cqring_ev_posted(ctx);
4850 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4852 struct hlist_head *list;
4853 struct io_kiocb *req;
4855 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4856 hlist_for_each_entry(req, list, hash_node) {
4857 if (sqe_addr != req->user_data)
4859 if (io_poll_remove_one(req))
4867 static int io_poll_remove_prep(struct io_kiocb *req,
4868 const struct io_uring_sqe *sqe)
4870 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4872 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4876 req->poll.addr = READ_ONCE(sqe->addr);
4881 * Find a running poll command that matches one specified in sqe->addr,
4882 * and remove it if found.
4884 static int io_poll_remove(struct io_kiocb *req)
4886 struct io_ring_ctx *ctx = req->ctx;
4890 addr = req->poll.addr;
4891 spin_lock_irq(&ctx->completion_lock);
4892 ret = io_poll_cancel(ctx, addr);
4893 spin_unlock_irq(&ctx->completion_lock);
4896 req_set_fail_links(req);
4897 io_req_complete(req, ret);
4901 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4904 struct io_kiocb *req = wait->private;
4905 struct io_poll_iocb *poll = &req->poll;
4907 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4910 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4911 struct poll_table_struct *p)
4913 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4915 __io_queue_proc(&pt->req->poll, pt, head);
4918 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4920 struct io_poll_iocb *poll = &req->poll;
4923 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4925 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4930 events = READ_ONCE(sqe->poll32_events);
4932 events = swahw32(events);
4934 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4935 (events & EPOLLEXCLUSIVE);
4937 io_get_req_task(req);
4941 static int io_poll_add(struct io_kiocb *req)
4943 struct io_poll_iocb *poll = &req->poll;
4944 struct io_ring_ctx *ctx = req->ctx;
4945 struct io_poll_table ipt;
4948 INIT_HLIST_NODE(&req->hash_node);
4949 INIT_LIST_HEAD(&req->list);
4950 ipt.pt._qproc = io_poll_queue_proc;
4952 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4955 if (mask) { /* no async, we'd stolen it */
4957 io_poll_complete(req, mask, 0);
4959 spin_unlock_irq(&ctx->completion_lock);
4962 io_cqring_ev_posted(ctx);
4968 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4970 struct io_timeout_data *data = container_of(timer,
4971 struct io_timeout_data, timer);
4972 struct io_kiocb *req = data->req;
4973 struct io_ring_ctx *ctx = req->ctx;
4974 unsigned long flags;
4976 atomic_inc(&ctx->cq_timeouts);
4978 spin_lock_irqsave(&ctx->completion_lock, flags);
4980 * We could be racing with timeout deletion. If the list is empty,
4981 * then timeout lookup already found it and will be handling it.
4983 if (!list_empty(&req->list))
4984 list_del_init(&req->list);
4986 io_cqring_fill_event(req, -ETIME);
4987 io_commit_cqring(ctx);
4988 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4990 io_cqring_ev_posted(ctx);
4991 req_set_fail_links(req);
4993 return HRTIMER_NORESTART;
4996 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4998 struct io_kiocb *req;
5001 list_for_each_entry(req, &ctx->timeout_list, list) {
5002 if (user_data == req->user_data) {
5003 list_del_init(&req->list);
5012 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5016 req_set_fail_links(req);
5017 io_cqring_fill_event(req, -ECANCELED);
5022 static int io_timeout_remove_prep(struct io_kiocb *req,
5023 const struct io_uring_sqe *sqe)
5025 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5027 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
5030 req->timeout.addr = READ_ONCE(sqe->addr);
5031 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5032 if (req->timeout.flags)
5039 * Remove or update an existing timeout command
5041 static int io_timeout_remove(struct io_kiocb *req)
5043 struct io_ring_ctx *ctx = req->ctx;
5046 spin_lock_irq(&ctx->completion_lock);
5047 ret = io_timeout_cancel(ctx, req->timeout.addr);
5049 io_cqring_fill_event(req, ret);
5050 io_commit_cqring(ctx);
5051 spin_unlock_irq(&ctx->completion_lock);
5052 io_cqring_ev_posted(ctx);
5054 req_set_fail_links(req);
5059 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5060 bool is_timeout_link)
5062 struct io_timeout_data *data;
5064 u32 off = READ_ONCE(sqe->off);
5066 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5068 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5070 if (off && is_timeout_link)
5072 flags = READ_ONCE(sqe->timeout_flags);
5073 if (flags & ~IORING_TIMEOUT_ABS)
5076 req->timeout.off = off;
5078 if (!req->io && io_alloc_async_ctx(req))
5081 data = &req->io->timeout;
5084 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5087 if (flags & IORING_TIMEOUT_ABS)
5088 data->mode = HRTIMER_MODE_ABS;
5090 data->mode = HRTIMER_MODE_REL;
5092 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5096 static int io_timeout(struct io_kiocb *req)
5098 struct io_ring_ctx *ctx = req->ctx;
5099 struct io_timeout_data *data = &req->io->timeout;
5100 struct list_head *entry;
5101 u32 tail, off = req->timeout.off;
5103 spin_lock_irq(&ctx->completion_lock);
5106 * sqe->off holds how many events that need to occur for this
5107 * timeout event to be satisfied. If it isn't set, then this is
5108 * a pure timeout request, sequence isn't used.
5110 if (io_is_timeout_noseq(req)) {
5111 entry = ctx->timeout_list.prev;
5115 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5116 req->timeout.target_seq = tail + off;
5119 * Insertion sort, ensuring the first entry in the list is always
5120 * the one we need first.
5122 list_for_each_prev(entry, &ctx->timeout_list) {
5123 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
5125 if (io_is_timeout_noseq(nxt))
5127 /* nxt.seq is behind @tail, otherwise would've been completed */
5128 if (off >= nxt->timeout.target_seq - tail)
5132 list_add(&req->list, entry);
5133 data->timer.function = io_timeout_fn;
5134 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5135 spin_unlock_irq(&ctx->completion_lock);
5139 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5141 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5143 return req->user_data == (unsigned long) data;
5146 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5148 enum io_wq_cancel cancel_ret;
5151 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5152 switch (cancel_ret) {
5153 case IO_WQ_CANCEL_OK:
5156 case IO_WQ_CANCEL_RUNNING:
5159 case IO_WQ_CANCEL_NOTFOUND:
5167 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5168 struct io_kiocb *req, __u64 sqe_addr,
5171 unsigned long flags;
5174 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5175 if (ret != -ENOENT) {
5176 spin_lock_irqsave(&ctx->completion_lock, flags);
5180 spin_lock_irqsave(&ctx->completion_lock, flags);
5181 ret = io_timeout_cancel(ctx, sqe_addr);
5184 ret = io_poll_cancel(ctx, sqe_addr);
5188 io_cqring_fill_event(req, ret);
5189 io_commit_cqring(ctx);
5190 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5191 io_cqring_ev_posted(ctx);
5194 req_set_fail_links(req);
5198 static int io_async_cancel_prep(struct io_kiocb *req,
5199 const struct io_uring_sqe *sqe)
5201 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5203 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
5207 req->cancel.addr = READ_ONCE(sqe->addr);
5211 static int io_async_cancel(struct io_kiocb *req)
5213 struct io_ring_ctx *ctx = req->ctx;
5215 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5219 static int io_files_update_prep(struct io_kiocb *req,
5220 const struct io_uring_sqe *sqe)
5222 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
5225 req->files_update.offset = READ_ONCE(sqe->off);
5226 req->files_update.nr_args = READ_ONCE(sqe->len);
5227 if (!req->files_update.nr_args)
5229 req->files_update.arg = READ_ONCE(sqe->addr);
5233 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5234 struct io_comp_state *cs)
5236 struct io_ring_ctx *ctx = req->ctx;
5237 struct io_uring_files_update up;
5243 up.offset = req->files_update.offset;
5244 up.fds = req->files_update.arg;
5246 mutex_lock(&ctx->uring_lock);
5247 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5248 mutex_unlock(&ctx->uring_lock);
5251 req_set_fail_links(req);
5252 __io_req_complete(req, ret, 0, cs);
5256 static int io_req_defer_prep(struct io_kiocb *req,
5257 const struct io_uring_sqe *sqe)
5264 if (io_op_defs[req->opcode].file_table) {
5265 io_req_init_async(req);
5266 ret = io_grab_files(req);
5271 switch (req->opcode) {
5274 case IORING_OP_READV:
5275 case IORING_OP_READ_FIXED:
5276 case IORING_OP_READ:
5277 ret = io_read_prep(req, sqe, true);
5279 case IORING_OP_WRITEV:
5280 case IORING_OP_WRITE_FIXED:
5281 case IORING_OP_WRITE:
5282 ret = io_write_prep(req, sqe, true);
5284 case IORING_OP_POLL_ADD:
5285 ret = io_poll_add_prep(req, sqe);
5287 case IORING_OP_POLL_REMOVE:
5288 ret = io_poll_remove_prep(req, sqe);
5290 case IORING_OP_FSYNC:
5291 ret = io_prep_fsync(req, sqe);
5293 case IORING_OP_SYNC_FILE_RANGE:
5294 ret = io_prep_sfr(req, sqe);
5296 case IORING_OP_SENDMSG:
5297 case IORING_OP_SEND:
5298 ret = io_sendmsg_prep(req, sqe);
5300 case IORING_OP_RECVMSG:
5301 case IORING_OP_RECV:
5302 ret = io_recvmsg_prep(req, sqe);
5304 case IORING_OP_CONNECT:
5305 ret = io_connect_prep(req, sqe);
5307 case IORING_OP_TIMEOUT:
5308 ret = io_timeout_prep(req, sqe, false);
5310 case IORING_OP_TIMEOUT_REMOVE:
5311 ret = io_timeout_remove_prep(req, sqe);
5313 case IORING_OP_ASYNC_CANCEL:
5314 ret = io_async_cancel_prep(req, sqe);
5316 case IORING_OP_LINK_TIMEOUT:
5317 ret = io_timeout_prep(req, sqe, true);
5319 case IORING_OP_ACCEPT:
5320 ret = io_accept_prep(req, sqe);
5322 case IORING_OP_FALLOCATE:
5323 ret = io_fallocate_prep(req, sqe);
5325 case IORING_OP_OPENAT:
5326 ret = io_openat_prep(req, sqe);
5328 case IORING_OP_CLOSE:
5329 ret = io_close_prep(req, sqe);
5331 case IORING_OP_FILES_UPDATE:
5332 ret = io_files_update_prep(req, sqe);
5334 case IORING_OP_STATX:
5335 ret = io_statx_prep(req, sqe);
5337 case IORING_OP_FADVISE:
5338 ret = io_fadvise_prep(req, sqe);
5340 case IORING_OP_MADVISE:
5341 ret = io_madvise_prep(req, sqe);
5343 case IORING_OP_OPENAT2:
5344 ret = io_openat2_prep(req, sqe);
5346 case IORING_OP_EPOLL_CTL:
5347 ret = io_epoll_ctl_prep(req, sqe);
5349 case IORING_OP_SPLICE:
5350 ret = io_splice_prep(req, sqe);
5352 case IORING_OP_PROVIDE_BUFFERS:
5353 ret = io_provide_buffers_prep(req, sqe);
5355 case IORING_OP_REMOVE_BUFFERS:
5356 ret = io_remove_buffers_prep(req, sqe);
5359 ret = io_tee_prep(req, sqe);
5362 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5371 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5373 struct io_ring_ctx *ctx = req->ctx;
5376 /* Still need defer if there is pending req in defer list. */
5377 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5381 if (io_alloc_async_ctx(req))
5383 ret = io_req_defer_prep(req, sqe);
5387 io_prep_async_link(req);
5389 spin_lock_irq(&ctx->completion_lock);
5390 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5391 spin_unlock_irq(&ctx->completion_lock);
5395 trace_io_uring_defer(ctx, req, req->user_data);
5396 list_add_tail(&req->list, &ctx->defer_list);
5397 spin_unlock_irq(&ctx->completion_lock);
5398 return -EIOCBQUEUED;
5401 static void io_cleanup_req(struct io_kiocb *req)
5403 struct io_async_ctx *io = req->io;
5405 switch (req->opcode) {
5406 case IORING_OP_READV:
5407 case IORING_OP_READ_FIXED:
5408 case IORING_OP_READ:
5409 if (req->flags & REQ_F_BUFFER_SELECTED)
5410 kfree((void *)(unsigned long)req->rw.addr);
5412 case IORING_OP_WRITEV:
5413 case IORING_OP_WRITE_FIXED:
5414 case IORING_OP_WRITE:
5415 if (io->rw.iov != io->rw.fast_iov)
5418 case IORING_OP_RECVMSG:
5419 if (req->flags & REQ_F_BUFFER_SELECTED)
5420 kfree(req->sr_msg.kbuf);
5422 case IORING_OP_SENDMSG:
5423 if (io->msg.iov != io->msg.fast_iov)
5426 case IORING_OP_RECV:
5427 if (req->flags & REQ_F_BUFFER_SELECTED)
5428 kfree(req->sr_msg.kbuf);
5430 case IORING_OP_OPENAT:
5431 case IORING_OP_OPENAT2:
5433 case IORING_OP_SPLICE:
5435 io_put_file(req, req->splice.file_in,
5436 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5440 req->flags &= ~REQ_F_NEED_CLEANUP;
5443 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5444 bool force_nonblock, struct io_comp_state *cs)
5446 struct io_ring_ctx *ctx = req->ctx;
5449 switch (req->opcode) {
5451 ret = io_nop(req, cs);
5453 case IORING_OP_READV:
5454 case IORING_OP_READ_FIXED:
5455 case IORING_OP_READ:
5457 ret = io_read_prep(req, sqe, force_nonblock);
5461 ret = io_read(req, force_nonblock, cs);
5463 case IORING_OP_WRITEV:
5464 case IORING_OP_WRITE_FIXED:
5465 case IORING_OP_WRITE:
5467 ret = io_write_prep(req, sqe, force_nonblock);
5471 ret = io_write(req, force_nonblock, cs);
5473 case IORING_OP_FSYNC:
5475 ret = io_prep_fsync(req, sqe);
5479 ret = io_fsync(req, force_nonblock);
5481 case IORING_OP_POLL_ADD:
5483 ret = io_poll_add_prep(req, sqe);
5487 ret = io_poll_add(req);
5489 case IORING_OP_POLL_REMOVE:
5491 ret = io_poll_remove_prep(req, sqe);
5495 ret = io_poll_remove(req);
5497 case IORING_OP_SYNC_FILE_RANGE:
5499 ret = io_prep_sfr(req, sqe);
5503 ret = io_sync_file_range(req, force_nonblock);
5505 case IORING_OP_SENDMSG:
5506 case IORING_OP_SEND:
5508 ret = io_sendmsg_prep(req, sqe);
5512 if (req->opcode == IORING_OP_SENDMSG)
5513 ret = io_sendmsg(req, force_nonblock, cs);
5515 ret = io_send(req, force_nonblock, cs);
5517 case IORING_OP_RECVMSG:
5518 case IORING_OP_RECV:
5520 ret = io_recvmsg_prep(req, sqe);
5524 if (req->opcode == IORING_OP_RECVMSG)
5525 ret = io_recvmsg(req, force_nonblock, cs);
5527 ret = io_recv(req, force_nonblock, cs);
5529 case IORING_OP_TIMEOUT:
5531 ret = io_timeout_prep(req, sqe, false);
5535 ret = io_timeout(req);
5537 case IORING_OP_TIMEOUT_REMOVE:
5539 ret = io_timeout_remove_prep(req, sqe);
5543 ret = io_timeout_remove(req);
5545 case IORING_OP_ACCEPT:
5547 ret = io_accept_prep(req, sqe);
5551 ret = io_accept(req, force_nonblock, cs);
5553 case IORING_OP_CONNECT:
5555 ret = io_connect_prep(req, sqe);
5559 ret = io_connect(req, force_nonblock, cs);
5561 case IORING_OP_ASYNC_CANCEL:
5563 ret = io_async_cancel_prep(req, sqe);
5567 ret = io_async_cancel(req);
5569 case IORING_OP_FALLOCATE:
5571 ret = io_fallocate_prep(req, sqe);
5575 ret = io_fallocate(req, force_nonblock);
5577 case IORING_OP_OPENAT:
5579 ret = io_openat_prep(req, sqe);
5583 ret = io_openat(req, force_nonblock);
5585 case IORING_OP_CLOSE:
5587 ret = io_close_prep(req, sqe);
5591 ret = io_close(req, force_nonblock, cs);
5593 case IORING_OP_FILES_UPDATE:
5595 ret = io_files_update_prep(req, sqe);
5599 ret = io_files_update(req, force_nonblock, cs);
5601 case IORING_OP_STATX:
5603 ret = io_statx_prep(req, sqe);
5607 ret = io_statx(req, force_nonblock);
5609 case IORING_OP_FADVISE:
5611 ret = io_fadvise_prep(req, sqe);
5615 ret = io_fadvise(req, force_nonblock);
5617 case IORING_OP_MADVISE:
5619 ret = io_madvise_prep(req, sqe);
5623 ret = io_madvise(req, force_nonblock);
5625 case IORING_OP_OPENAT2:
5627 ret = io_openat2_prep(req, sqe);
5631 ret = io_openat2(req, force_nonblock);
5633 case IORING_OP_EPOLL_CTL:
5635 ret = io_epoll_ctl_prep(req, sqe);
5639 ret = io_epoll_ctl(req, force_nonblock, cs);
5641 case IORING_OP_SPLICE:
5643 ret = io_splice_prep(req, sqe);
5647 ret = io_splice(req, force_nonblock);
5649 case IORING_OP_PROVIDE_BUFFERS:
5651 ret = io_provide_buffers_prep(req, sqe);
5655 ret = io_provide_buffers(req, force_nonblock, cs);
5657 case IORING_OP_REMOVE_BUFFERS:
5659 ret = io_remove_buffers_prep(req, sqe);
5663 ret = io_remove_buffers(req, force_nonblock, cs);
5667 ret = io_tee_prep(req, sqe);
5671 ret = io_tee(req, force_nonblock);
5681 /* If the op doesn't have a file, we're not polling for it */
5682 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5683 const bool in_async = io_wq_current_is_worker();
5685 /* workqueue context doesn't hold uring_lock, grab it now */
5687 mutex_lock(&ctx->uring_lock);
5689 io_iopoll_req_issued(req);
5692 mutex_unlock(&ctx->uring_lock);
5698 static void io_arm_async_linked_timeout(struct io_kiocb *req)
5700 struct io_kiocb *link;
5702 /* link head's timeout is queued in io_queue_async_work() */
5703 if (!(req->flags & REQ_F_QUEUE_TIMEOUT))
5706 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
5707 io_queue_linked_timeout(link);
5710 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5712 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5715 io_arm_async_linked_timeout(req);
5717 /* if NO_CANCEL is set, we must still run the work */
5718 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5719 IO_WQ_WORK_CANCEL) {
5725 ret = io_issue_sqe(req, NULL, false, NULL);
5727 * We can get EAGAIN for polled IO even though we're
5728 * forcing a sync submission from here, since we can't
5729 * wait for request slots on the block side.
5738 req_set_fail_links(req);
5739 io_req_complete(req, ret);
5742 return io_steal_work(req);
5745 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5748 struct fixed_file_table *table;
5750 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5751 return table->files[index & IORING_FILE_TABLE_MASK];
5754 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5755 int fd, struct file **out_file, bool fixed)
5757 struct io_ring_ctx *ctx = req->ctx;
5761 if (unlikely(!ctx->file_data ||
5762 (unsigned) fd >= ctx->nr_user_files))
5764 fd = array_index_nospec(fd, ctx->nr_user_files);
5765 file = io_file_from_index(ctx, fd);
5767 req->fixed_file_refs = ctx->file_data->cur_refs;
5768 percpu_ref_get(req->fixed_file_refs);
5771 trace_io_uring_file_get(ctx, fd);
5772 file = __io_file_get(state, fd);
5775 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5782 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5787 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5788 if (unlikely(!fixed && io_async_submit(req->ctx)))
5791 return io_file_get(state, req, fd, &req->file, fixed);
5794 static int io_grab_files(struct io_kiocb *req)
5797 struct io_ring_ctx *ctx = req->ctx;
5799 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5801 if (!ctx->ring_file)
5805 spin_lock_irq(&ctx->inflight_lock);
5807 * We use the f_ops->flush() handler to ensure that we can flush
5808 * out work accessing these files if the fd is closed. Check if
5809 * the fd has changed since we started down this path, and disallow
5810 * this operation if it has.
5812 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5813 list_add(&req->inflight_entry, &ctx->inflight_list);
5814 req->flags |= REQ_F_INFLIGHT;
5815 req->work.files = current->files;
5818 spin_unlock_irq(&ctx->inflight_lock);
5824 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5826 struct io_timeout_data *data = container_of(timer,
5827 struct io_timeout_data, timer);
5828 struct io_kiocb *req = data->req;
5829 struct io_ring_ctx *ctx = req->ctx;
5830 struct io_kiocb *prev = NULL;
5831 unsigned long flags;
5833 spin_lock_irqsave(&ctx->completion_lock, flags);
5836 * We don't expect the list to be empty, that will only happen if we
5837 * race with the completion of the linked work.
5839 if (!list_empty(&req->link_list)) {
5840 prev = list_entry(req->link_list.prev, struct io_kiocb,
5842 if (refcount_inc_not_zero(&prev->refs)) {
5843 list_del_init(&req->link_list);
5844 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5849 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5852 req_set_fail_links(prev);
5853 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5856 io_req_complete(req, -ETIME);
5858 return HRTIMER_NORESTART;
5861 static void io_queue_linked_timeout(struct io_kiocb *req)
5863 struct io_ring_ctx *ctx = req->ctx;
5866 * If the list is now empty, then our linked request finished before
5867 * we got a chance to setup the timer
5869 spin_lock_irq(&ctx->completion_lock);
5870 if (!list_empty(&req->link_list)) {
5871 struct io_timeout_data *data = &req->io->timeout;
5873 data->timer.function = io_link_timeout_fn;
5874 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5877 spin_unlock_irq(&ctx->completion_lock);
5879 /* drop submission reference */
5883 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5885 struct io_kiocb *nxt;
5887 if (!(req->flags & REQ_F_LINK_HEAD))
5889 /* for polled retry, if flag is set, we already went through here */
5890 if (req->flags & REQ_F_POLLED)
5893 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5895 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5898 req->flags |= REQ_F_LINK_TIMEOUT;
5902 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5903 struct io_comp_state *cs)
5905 struct io_kiocb *linked_timeout;
5906 struct io_kiocb *nxt;
5907 const struct cred *old_creds = NULL;
5911 linked_timeout = io_prep_linked_timeout(req);
5913 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5914 req->work.creds != current_cred()) {
5916 revert_creds(old_creds);
5917 if (old_creds == req->work.creds)
5918 old_creds = NULL; /* restored original creds */
5920 old_creds = override_creds(req->work.creds);
5923 ret = io_issue_sqe(req, sqe, true, cs);
5926 * We async punt it if the file wasn't marked NOWAIT, or if the file
5927 * doesn't support non-blocking read/write attempts
5929 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5930 if (io_arm_poll_handler(req)) {
5932 io_queue_linked_timeout(linked_timeout);
5936 io_req_init_async(req);
5938 if (io_op_defs[req->opcode].file_table) {
5939 ret = io_grab_files(req);
5945 * Queued up for async execution, worker will release
5946 * submit reference when the iocb is actually submitted.
5948 io_queue_async_work(req);
5953 /* drop submission reference */
5954 nxt = io_put_req_find_next(req);
5956 if (linked_timeout) {
5958 io_queue_linked_timeout(linked_timeout);
5960 io_put_req(linked_timeout);
5963 /* and drop final reference, if we failed */
5965 req_set_fail_links(req);
5966 io_req_complete(req, ret);
5971 if (req->flags & REQ_F_FORCE_ASYNC)
5977 revert_creds(old_creds);
5980 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5981 struct io_comp_state *cs)
5985 ret = io_req_defer(req, sqe);
5987 if (ret != -EIOCBQUEUED) {
5989 req_set_fail_links(req);
5991 io_req_complete(req, ret);
5993 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5996 if (io_alloc_async_ctx(req))
5998 ret = io_req_defer_prep(req, sqe);
5999 if (unlikely(ret < 0))
6004 * Never try inline submit of IOSQE_ASYNC is set, go straight
6005 * to async execution.
6007 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6008 io_queue_async_work(req);
6010 __io_queue_sqe(req, sqe, cs);
6014 static inline void io_queue_link_head(struct io_kiocb *req,
6015 struct io_comp_state *cs)
6017 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6019 io_req_complete(req, -ECANCELED);
6021 io_queue_sqe(req, NULL, cs);
6024 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6025 struct io_kiocb **link, struct io_comp_state *cs)
6027 struct io_ring_ctx *ctx = req->ctx;
6031 * If we already have a head request, queue this one for async
6032 * submittal once the head completes. If we don't have a head but
6033 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6034 * submitted sync once the chain is complete. If none of those
6035 * conditions are true (normal request), then just queue it.
6038 struct io_kiocb *head = *link;
6041 * Taking sequential execution of a link, draining both sides
6042 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6043 * requests in the link. So, it drains the head and the
6044 * next after the link request. The last one is done via
6045 * drain_next flag to persist the effect across calls.
6047 if (req->flags & REQ_F_IO_DRAIN) {
6048 head->flags |= REQ_F_IO_DRAIN;
6049 ctx->drain_next = 1;
6051 if (io_alloc_async_ctx(req))
6054 ret = io_req_defer_prep(req, sqe);
6056 /* fail even hard links since we don't submit */
6057 head->flags |= REQ_F_FAIL_LINK;
6060 trace_io_uring_link(ctx, req, head);
6061 io_get_req_task(req);
6062 list_add_tail(&req->link_list, &head->link_list);
6064 /* last request of a link, enqueue the link */
6065 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6066 io_queue_link_head(head, cs);
6070 if (unlikely(ctx->drain_next)) {
6071 req->flags |= REQ_F_IO_DRAIN;
6072 ctx->drain_next = 0;
6074 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6075 req->flags |= REQ_F_LINK_HEAD;
6076 INIT_LIST_HEAD(&req->link_list);
6078 if (io_alloc_async_ctx(req))
6081 ret = io_req_defer_prep(req, sqe);
6083 req->flags |= REQ_F_FAIL_LINK;
6086 io_queue_sqe(req, sqe, cs);
6094 * Batched submission is done, ensure local IO is flushed out.
6096 static void io_submit_state_end(struct io_submit_state *state)
6098 if (!list_empty(&state->comp.list))
6099 io_submit_flush_completions(&state->comp);
6100 blk_finish_plug(&state->plug);
6101 io_state_file_put(state);
6102 if (state->free_reqs)
6103 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6107 * Start submission side cache.
6109 static void io_submit_state_start(struct io_submit_state *state,
6110 struct io_ring_ctx *ctx, unsigned int max_ios)
6112 blk_start_plug(&state->plug);
6114 state->plug.nowait = true;
6117 INIT_LIST_HEAD(&state->comp.list);
6118 state->comp.ctx = ctx;
6119 state->free_reqs = 0;
6121 state->ios_left = max_ios;
6124 static void io_commit_sqring(struct io_ring_ctx *ctx)
6126 struct io_rings *rings = ctx->rings;
6129 * Ensure any loads from the SQEs are done at this point,
6130 * since once we write the new head, the application could
6131 * write new data to them.
6133 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6137 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6138 * that is mapped by userspace. This means that care needs to be taken to
6139 * ensure that reads are stable, as we cannot rely on userspace always
6140 * being a good citizen. If members of the sqe are validated and then later
6141 * used, it's important that those reads are done through READ_ONCE() to
6142 * prevent a re-load down the line.
6144 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6146 u32 *sq_array = ctx->sq_array;
6150 * The cached sq head (or cq tail) serves two purposes:
6152 * 1) allows us to batch the cost of updating the user visible
6154 * 2) allows the kernel side to track the head on its own, even
6155 * though the application is the one updating it.
6157 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6158 if (likely(head < ctx->sq_entries))
6159 return &ctx->sq_sqes[head];
6161 /* drop invalid entries */
6162 ctx->cached_sq_dropped++;
6163 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6167 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6169 ctx->cached_sq_head++;
6172 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6173 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6174 IOSQE_BUFFER_SELECT)
6176 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6177 const struct io_uring_sqe *sqe,
6178 struct io_submit_state *state)
6180 unsigned int sqe_flags;
6184 * All io need record the previous position, if LINK vs DARIN,
6185 * it can be used to mark the position of the first IO in the
6188 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
6189 req->opcode = READ_ONCE(sqe->opcode);
6190 req->user_data = READ_ONCE(sqe->user_data);
6195 /* one is dropped after submission, the other at completion */
6196 refcount_set(&req->refs, 2);
6197 req->task = current;
6200 if (unlikely(req->opcode >= IORING_OP_LAST))
6203 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6206 sqe_flags = READ_ONCE(sqe->flags);
6207 /* enforce forwards compatibility on users */
6208 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6211 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6212 !io_op_defs[req->opcode].buffer_select)
6215 id = READ_ONCE(sqe->personality);
6217 io_req_init_async(req);
6218 req->work.creds = idr_find(&ctx->personality_idr, id);
6219 if (unlikely(!req->work.creds))
6221 get_cred(req->work.creds);
6224 /* same numerical values with corresponding REQ_F_*, safe to copy */
6225 req->flags |= sqe_flags;
6227 if (!io_op_defs[req->opcode].needs_file)
6230 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6233 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6234 struct file *ring_file, int ring_fd)
6236 struct io_submit_state state;
6237 struct io_kiocb *link = NULL;
6238 int i, submitted = 0;
6240 /* if we have a backlog and couldn't flush it all, return BUSY */
6241 if (test_bit(0, &ctx->sq_check_overflow)) {
6242 if (!list_empty(&ctx->cq_overflow_list) &&
6243 !io_cqring_overflow_flush(ctx, false))
6247 /* make sure SQ entry isn't read before tail */
6248 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6250 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6253 io_submit_state_start(&state, ctx, nr);
6255 ctx->ring_fd = ring_fd;
6256 ctx->ring_file = ring_file;
6258 for (i = 0; i < nr; i++) {
6259 const struct io_uring_sqe *sqe;
6260 struct io_kiocb *req;
6263 sqe = io_get_sqe(ctx);
6264 if (unlikely(!sqe)) {
6265 io_consume_sqe(ctx);
6268 req = io_alloc_req(ctx, &state);
6269 if (unlikely(!req)) {
6271 submitted = -EAGAIN;
6275 err = io_init_req(ctx, req, sqe, &state);
6276 io_consume_sqe(ctx);
6277 /* will complete beyond this point, count as submitted */
6280 if (unlikely(err)) {
6283 io_req_complete(req, err);
6287 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6288 true, io_async_submit(ctx));
6289 err = io_submit_sqe(req, sqe, &link, &state.comp);
6294 if (unlikely(submitted != nr)) {
6295 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6297 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6300 io_queue_link_head(link, &state.comp);
6301 io_submit_state_end(&state);
6303 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6304 io_commit_sqring(ctx);
6309 static int io_sq_thread(void *data)
6311 struct io_ring_ctx *ctx = data;
6312 const struct cred *old_cred;
6314 unsigned long timeout;
6317 complete(&ctx->sq_thread_comp);
6319 old_cred = override_creds(ctx->creds);
6321 timeout = jiffies + ctx->sq_thread_idle;
6322 while (!kthread_should_park()) {
6323 unsigned int to_submit;
6325 if (!list_empty(&ctx->poll_list)) {
6326 unsigned nr_events = 0;
6328 mutex_lock(&ctx->uring_lock);
6329 if (!list_empty(&ctx->poll_list))
6330 io_iopoll_getevents(ctx, &nr_events, 0);
6332 timeout = jiffies + ctx->sq_thread_idle;
6333 mutex_unlock(&ctx->uring_lock);
6336 to_submit = io_sqring_entries(ctx);
6339 * If submit got -EBUSY, flag us as needing the application
6340 * to enter the kernel to reap and flush events.
6342 if (!to_submit || ret == -EBUSY || need_resched()) {
6344 * Drop cur_mm before scheduling, we can't hold it for
6345 * long periods (or over schedule()). Do this before
6346 * adding ourselves to the waitqueue, as the unuse/drop
6349 io_sq_thread_drop_mm(ctx);
6352 * We're polling. If we're within the defined idle
6353 * period, then let us spin without work before going
6354 * to sleep. The exception is if we got EBUSY doing
6355 * more IO, we should wait for the application to
6356 * reap events and wake us up.
6358 if (!list_empty(&ctx->poll_list) || need_resched() ||
6359 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6360 !percpu_ref_is_dying(&ctx->refs))) {
6361 if (current->task_works)
6367 prepare_to_wait(&ctx->sqo_wait, &wait,
6368 TASK_INTERRUPTIBLE);
6371 * While doing polled IO, before going to sleep, we need
6372 * to check if there are new reqs added to poll_list, it
6373 * is because reqs may have been punted to io worker and
6374 * will be added to poll_list later, hence check the
6377 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6378 !list_empty_careful(&ctx->poll_list)) {
6379 finish_wait(&ctx->sqo_wait, &wait);
6383 /* Tell userspace we may need a wakeup call */
6384 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6385 /* make sure to read SQ tail after writing flags */
6388 to_submit = io_sqring_entries(ctx);
6389 if (!to_submit || ret == -EBUSY) {
6390 if (kthread_should_park()) {
6391 finish_wait(&ctx->sqo_wait, &wait);
6394 if (current->task_works) {
6396 finish_wait(&ctx->sqo_wait, &wait);
6399 if (signal_pending(current))
6400 flush_signals(current);
6402 finish_wait(&ctx->sqo_wait, &wait);
6404 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6408 finish_wait(&ctx->sqo_wait, &wait);
6410 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6413 mutex_lock(&ctx->uring_lock);
6414 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6415 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6416 mutex_unlock(&ctx->uring_lock);
6417 timeout = jiffies + ctx->sq_thread_idle;
6420 if (current->task_works)
6423 io_sq_thread_drop_mm(ctx);
6424 revert_creds(old_cred);
6431 struct io_wait_queue {
6432 struct wait_queue_entry wq;
6433 struct io_ring_ctx *ctx;
6435 unsigned nr_timeouts;
6438 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6440 struct io_ring_ctx *ctx = iowq->ctx;
6443 * Wake up if we have enough events, or if a timeout occurred since we
6444 * started waiting. For timeouts, we always want to return to userspace,
6445 * regardless of event count.
6447 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6448 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6451 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6452 int wake_flags, void *key)
6454 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6457 /* use noflush == true, as we can't safely rely on locking context */
6458 if (!io_should_wake(iowq, true))
6461 return autoremove_wake_function(curr, mode, wake_flags, key);
6465 * Wait until events become available, if we don't already have some. The
6466 * application must reap them itself, as they reside on the shared cq ring.
6468 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6469 const sigset_t __user *sig, size_t sigsz)
6471 struct io_wait_queue iowq = {
6474 .func = io_wake_function,
6475 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6478 .to_wait = min_events,
6480 struct io_rings *rings = ctx->rings;
6484 if (io_cqring_events(ctx, false) >= min_events)
6486 if (!current->task_works)
6492 #ifdef CONFIG_COMPAT
6493 if (in_compat_syscall())
6494 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6498 ret = set_user_sigmask(sig, sigsz);
6504 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6505 trace_io_uring_cqring_wait(ctx, min_events);
6507 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6508 TASK_INTERRUPTIBLE);
6509 if (current->task_works)
6511 if (io_should_wake(&iowq, false))
6514 if (signal_pending(current)) {
6519 finish_wait(&ctx->wait, &iowq.wq);
6521 restore_saved_sigmask_unless(ret == -EINTR);
6523 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6526 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6528 #if defined(CONFIG_UNIX)
6529 if (ctx->ring_sock) {
6530 struct sock *sock = ctx->ring_sock->sk;
6531 struct sk_buff *skb;
6533 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6539 for (i = 0; i < ctx->nr_user_files; i++) {
6542 file = io_file_from_index(ctx, i);
6549 static void io_file_ref_kill(struct percpu_ref *ref)
6551 struct fixed_file_data *data;
6553 data = container_of(ref, struct fixed_file_data, refs);
6554 complete(&data->done);
6557 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6559 struct fixed_file_data *data = ctx->file_data;
6560 struct fixed_file_ref_node *ref_node = NULL;
6561 unsigned nr_tables, i;
6566 spin_lock(&data->lock);
6567 if (!list_empty(&data->ref_list))
6568 ref_node = list_first_entry(&data->ref_list,
6569 struct fixed_file_ref_node, node);
6570 spin_unlock(&data->lock);
6572 percpu_ref_kill(&ref_node->refs);
6574 percpu_ref_kill(&data->refs);
6576 /* wait for all refs nodes to complete */
6577 flush_delayed_work(&ctx->file_put_work);
6578 wait_for_completion(&data->done);
6580 __io_sqe_files_unregister(ctx);
6581 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6582 for (i = 0; i < nr_tables; i++)
6583 kfree(data->table[i].files);
6585 percpu_ref_exit(&data->refs);
6587 ctx->file_data = NULL;
6588 ctx->nr_user_files = 0;
6592 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6594 if (ctx->sqo_thread) {
6595 wait_for_completion(&ctx->sq_thread_comp);
6597 * The park is a bit of a work-around, without it we get
6598 * warning spews on shutdown with SQPOLL set and affinity
6599 * set to a single CPU.
6601 kthread_park(ctx->sqo_thread);
6602 kthread_stop(ctx->sqo_thread);
6603 ctx->sqo_thread = NULL;
6607 static void io_finish_async(struct io_ring_ctx *ctx)
6609 io_sq_thread_stop(ctx);
6612 io_wq_destroy(ctx->io_wq);
6617 #if defined(CONFIG_UNIX)
6619 * Ensure the UNIX gc is aware of our file set, so we are certain that
6620 * the io_uring can be safely unregistered on process exit, even if we have
6621 * loops in the file referencing.
6623 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6625 struct sock *sk = ctx->ring_sock->sk;
6626 struct scm_fp_list *fpl;
6627 struct sk_buff *skb;
6630 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6634 skb = alloc_skb(0, GFP_KERNEL);
6643 fpl->user = get_uid(ctx->user);
6644 for (i = 0; i < nr; i++) {
6645 struct file *file = io_file_from_index(ctx, i + offset);
6649 fpl->fp[nr_files] = get_file(file);
6650 unix_inflight(fpl->user, fpl->fp[nr_files]);
6655 fpl->max = SCM_MAX_FD;
6656 fpl->count = nr_files;
6657 UNIXCB(skb).fp = fpl;
6658 skb->destructor = unix_destruct_scm;
6659 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6660 skb_queue_head(&sk->sk_receive_queue, skb);
6662 for (i = 0; i < nr_files; i++)
6673 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6674 * causes regular reference counting to break down. We rely on the UNIX
6675 * garbage collection to take care of this problem for us.
6677 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6679 unsigned left, total;
6683 left = ctx->nr_user_files;
6685 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6687 ret = __io_sqe_files_scm(ctx, this_files, total);
6691 total += this_files;
6697 while (total < ctx->nr_user_files) {
6698 struct file *file = io_file_from_index(ctx, total);
6708 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6714 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6719 for (i = 0; i < nr_tables; i++) {
6720 struct fixed_file_table *table = &ctx->file_data->table[i];
6721 unsigned this_files;
6723 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6724 table->files = kcalloc(this_files, sizeof(struct file *),
6728 nr_files -= this_files;
6734 for (i = 0; i < nr_tables; i++) {
6735 struct fixed_file_table *table = &ctx->file_data->table[i];
6736 kfree(table->files);
6741 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6743 #if defined(CONFIG_UNIX)
6744 struct sock *sock = ctx->ring_sock->sk;
6745 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6746 struct sk_buff *skb;
6749 __skb_queue_head_init(&list);
6752 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6753 * remove this entry and rearrange the file array.
6755 skb = skb_dequeue(head);
6757 struct scm_fp_list *fp;
6759 fp = UNIXCB(skb).fp;
6760 for (i = 0; i < fp->count; i++) {
6763 if (fp->fp[i] != file)
6766 unix_notinflight(fp->user, fp->fp[i]);
6767 left = fp->count - 1 - i;
6769 memmove(&fp->fp[i], &fp->fp[i + 1],
6770 left * sizeof(struct file *));
6777 __skb_queue_tail(&list, skb);
6787 __skb_queue_tail(&list, skb);
6789 skb = skb_dequeue(head);
6792 if (skb_peek(&list)) {
6793 spin_lock_irq(&head->lock);
6794 while ((skb = __skb_dequeue(&list)) != NULL)
6795 __skb_queue_tail(head, skb);
6796 spin_unlock_irq(&head->lock);
6803 struct io_file_put {
6804 struct list_head list;
6808 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6810 struct fixed_file_data *file_data = ref_node->file_data;
6811 struct io_ring_ctx *ctx = file_data->ctx;
6812 struct io_file_put *pfile, *tmp;
6814 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6815 list_del(&pfile->list);
6816 io_ring_file_put(ctx, pfile->file);
6820 spin_lock(&file_data->lock);
6821 list_del(&ref_node->node);
6822 spin_unlock(&file_data->lock);
6824 percpu_ref_exit(&ref_node->refs);
6826 percpu_ref_put(&file_data->refs);
6829 static void io_file_put_work(struct work_struct *work)
6831 struct io_ring_ctx *ctx;
6832 struct llist_node *node;
6834 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6835 node = llist_del_all(&ctx->file_put_llist);
6838 struct fixed_file_ref_node *ref_node;
6839 struct llist_node *next = node->next;
6841 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6842 __io_file_put_work(ref_node);
6847 static void io_file_data_ref_zero(struct percpu_ref *ref)
6849 struct fixed_file_ref_node *ref_node;
6850 struct io_ring_ctx *ctx;
6854 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6855 ctx = ref_node->file_data->ctx;
6857 if (percpu_ref_is_dying(&ctx->file_data->refs))
6860 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6862 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6864 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6867 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6868 struct io_ring_ctx *ctx)
6870 struct fixed_file_ref_node *ref_node;
6872 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6874 return ERR_PTR(-ENOMEM);
6876 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6879 return ERR_PTR(-ENOMEM);
6881 INIT_LIST_HEAD(&ref_node->node);
6882 INIT_LIST_HEAD(&ref_node->file_list);
6883 ref_node->file_data = ctx->file_data;
6887 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6889 percpu_ref_exit(&ref_node->refs);
6893 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6896 __s32 __user *fds = (__s32 __user *) arg;
6901 struct fixed_file_ref_node *ref_node;
6907 if (nr_args > IORING_MAX_FIXED_FILES)
6910 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6911 if (!ctx->file_data)
6913 ctx->file_data->ctx = ctx;
6914 init_completion(&ctx->file_data->done);
6915 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6916 spin_lock_init(&ctx->file_data->lock);
6918 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6919 ctx->file_data->table = kcalloc(nr_tables,
6920 sizeof(struct fixed_file_table),
6922 if (!ctx->file_data->table) {
6923 kfree(ctx->file_data);
6924 ctx->file_data = NULL;
6928 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6929 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6930 kfree(ctx->file_data->table);
6931 kfree(ctx->file_data);
6932 ctx->file_data = NULL;
6936 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6937 percpu_ref_exit(&ctx->file_data->refs);
6938 kfree(ctx->file_data->table);
6939 kfree(ctx->file_data);
6940 ctx->file_data = NULL;
6944 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6945 struct fixed_file_table *table;
6949 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6951 /* allow sparse sets */
6957 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6958 index = i & IORING_FILE_TABLE_MASK;
6966 * Don't allow io_uring instances to be registered. If UNIX
6967 * isn't enabled, then this causes a reference cycle and this
6968 * instance can never get freed. If UNIX is enabled we'll
6969 * handle it just fine, but there's still no point in allowing
6970 * a ring fd as it doesn't support regular read/write anyway.
6972 if (file->f_op == &io_uring_fops) {
6977 table->files[index] = file;
6981 for (i = 0; i < ctx->nr_user_files; i++) {
6982 file = io_file_from_index(ctx, i);
6986 for (i = 0; i < nr_tables; i++)
6987 kfree(ctx->file_data->table[i].files);
6989 kfree(ctx->file_data->table);
6990 kfree(ctx->file_data);
6991 ctx->file_data = NULL;
6992 ctx->nr_user_files = 0;
6996 ret = io_sqe_files_scm(ctx);
6998 io_sqe_files_unregister(ctx);
7002 ref_node = alloc_fixed_file_ref_node(ctx);
7003 if (IS_ERR(ref_node)) {
7004 io_sqe_files_unregister(ctx);
7005 return PTR_ERR(ref_node);
7008 ctx->file_data->cur_refs = &ref_node->refs;
7009 spin_lock(&ctx->file_data->lock);
7010 list_add(&ref_node->node, &ctx->file_data->ref_list);
7011 spin_unlock(&ctx->file_data->lock);
7012 percpu_ref_get(&ctx->file_data->refs);
7016 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7019 #if defined(CONFIG_UNIX)
7020 struct sock *sock = ctx->ring_sock->sk;
7021 struct sk_buff_head *head = &sock->sk_receive_queue;
7022 struct sk_buff *skb;
7025 * See if we can merge this file into an existing skb SCM_RIGHTS
7026 * file set. If there's no room, fall back to allocating a new skb
7027 * and filling it in.
7029 spin_lock_irq(&head->lock);
7030 skb = skb_peek(head);
7032 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7034 if (fpl->count < SCM_MAX_FD) {
7035 __skb_unlink(skb, head);
7036 spin_unlock_irq(&head->lock);
7037 fpl->fp[fpl->count] = get_file(file);
7038 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7040 spin_lock_irq(&head->lock);
7041 __skb_queue_head(head, skb);
7046 spin_unlock_irq(&head->lock);
7053 return __io_sqe_files_scm(ctx, 1, index);
7059 static int io_queue_file_removal(struct fixed_file_data *data,
7062 struct io_file_put *pfile;
7063 struct percpu_ref *refs = data->cur_refs;
7064 struct fixed_file_ref_node *ref_node;
7066 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7070 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7072 list_add(&pfile->list, &ref_node->file_list);
7077 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7078 struct io_uring_files_update *up,
7081 struct fixed_file_data *data = ctx->file_data;
7082 struct fixed_file_ref_node *ref_node;
7087 bool needs_switch = false;
7089 if (check_add_overflow(up->offset, nr_args, &done))
7091 if (done > ctx->nr_user_files)
7094 ref_node = alloc_fixed_file_ref_node(ctx);
7095 if (IS_ERR(ref_node))
7096 return PTR_ERR(ref_node);
7099 fds = u64_to_user_ptr(up->fds);
7101 struct fixed_file_table *table;
7105 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7109 i = array_index_nospec(up->offset, ctx->nr_user_files);
7110 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7111 index = i & IORING_FILE_TABLE_MASK;
7112 if (table->files[index]) {
7113 file = io_file_from_index(ctx, index);
7114 err = io_queue_file_removal(data, file);
7117 table->files[index] = NULL;
7118 needs_switch = true;
7127 * Don't allow io_uring instances to be registered. If
7128 * UNIX isn't enabled, then this causes a reference
7129 * cycle and this instance can never get freed. If UNIX
7130 * is enabled we'll handle it just fine, but there's
7131 * still no point in allowing a ring fd as it doesn't
7132 * support regular read/write anyway.
7134 if (file->f_op == &io_uring_fops) {
7139 table->files[index] = file;
7140 err = io_sqe_file_register(ctx, file, i);
7150 percpu_ref_kill(data->cur_refs);
7151 spin_lock(&data->lock);
7152 list_add(&ref_node->node, &data->ref_list);
7153 data->cur_refs = &ref_node->refs;
7154 spin_unlock(&data->lock);
7155 percpu_ref_get(&ctx->file_data->refs);
7157 destroy_fixed_file_ref_node(ref_node);
7159 return done ? done : err;
7162 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7165 struct io_uring_files_update up;
7167 if (!ctx->file_data)
7171 if (copy_from_user(&up, arg, sizeof(up)))
7176 return __io_sqe_files_update(ctx, &up, nr_args);
7179 static void io_free_work(struct io_wq_work *work)
7181 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7183 /* Consider that io_steal_work() relies on this ref */
7187 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7188 struct io_uring_params *p)
7190 struct io_wq_data data;
7192 struct io_ring_ctx *ctx_attach;
7193 unsigned int concurrency;
7196 data.user = ctx->user;
7197 data.free_work = io_free_work;
7198 data.do_work = io_wq_submit_work;
7200 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7201 /* Do QD, or 4 * CPUS, whatever is smallest */
7202 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7204 ctx->io_wq = io_wq_create(concurrency, &data);
7205 if (IS_ERR(ctx->io_wq)) {
7206 ret = PTR_ERR(ctx->io_wq);
7212 f = fdget(p->wq_fd);
7216 if (f.file->f_op != &io_uring_fops) {
7221 ctx_attach = f.file->private_data;
7222 /* @io_wq is protected by holding the fd */
7223 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7228 ctx->io_wq = ctx_attach->io_wq;
7234 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7235 struct io_uring_params *p)
7239 mmgrab(current->mm);
7240 ctx->sqo_mm = current->mm;
7242 if (ctx->flags & IORING_SETUP_SQPOLL) {
7244 if (!capable(CAP_SYS_ADMIN))
7247 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7248 if (!ctx->sq_thread_idle)
7249 ctx->sq_thread_idle = HZ;
7251 if (p->flags & IORING_SETUP_SQ_AFF) {
7252 int cpu = p->sq_thread_cpu;
7255 if (cpu >= nr_cpu_ids)
7257 if (!cpu_online(cpu))
7260 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7264 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7267 if (IS_ERR(ctx->sqo_thread)) {
7268 ret = PTR_ERR(ctx->sqo_thread);
7269 ctx->sqo_thread = NULL;
7272 wake_up_process(ctx->sqo_thread);
7273 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7274 /* Can't have SQ_AFF without SQPOLL */
7279 ret = io_init_wq_offload(ctx, p);
7285 io_finish_async(ctx);
7286 mmdrop(ctx->sqo_mm);
7291 static inline void __io_unaccount_mem(struct user_struct *user,
7292 unsigned long nr_pages)
7294 atomic_long_sub(nr_pages, &user->locked_vm);
7297 static inline int __io_account_mem(struct user_struct *user,
7298 unsigned long nr_pages)
7300 unsigned long page_limit, cur_pages, new_pages;
7302 /* Don't allow more pages than we can safely lock */
7303 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7306 cur_pages = atomic_long_read(&user->locked_vm);
7307 new_pages = cur_pages + nr_pages;
7308 if (new_pages > page_limit)
7310 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7311 new_pages) != cur_pages);
7316 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7317 enum io_mem_account acct)
7320 __io_unaccount_mem(ctx->user, nr_pages);
7323 if (acct == ACCT_LOCKED)
7324 ctx->sqo_mm->locked_vm -= nr_pages;
7325 else if (acct == ACCT_PINNED)
7326 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7330 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7331 enum io_mem_account acct)
7335 if (ctx->limit_mem) {
7336 ret = __io_account_mem(ctx->user, nr_pages);
7342 if (acct == ACCT_LOCKED)
7343 ctx->sqo_mm->locked_vm += nr_pages;
7344 else if (acct == ACCT_PINNED)
7345 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7351 static void io_mem_free(void *ptr)
7358 page = virt_to_head_page(ptr);
7359 if (put_page_testzero(page))
7360 free_compound_page(page);
7363 static void *io_mem_alloc(size_t size)
7365 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7368 return (void *) __get_free_pages(gfp_flags, get_order(size));
7371 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7374 struct io_rings *rings;
7375 size_t off, sq_array_size;
7377 off = struct_size(rings, cqes, cq_entries);
7378 if (off == SIZE_MAX)
7382 off = ALIGN(off, SMP_CACHE_BYTES);
7387 sq_array_size = array_size(sizeof(u32), sq_entries);
7388 if (sq_array_size == SIZE_MAX)
7391 if (check_add_overflow(off, sq_array_size, &off))
7400 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7404 pages = (size_t)1 << get_order(
7405 rings_size(sq_entries, cq_entries, NULL));
7406 pages += (size_t)1 << get_order(
7407 array_size(sizeof(struct io_uring_sqe), sq_entries));
7412 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7416 if (!ctx->user_bufs)
7419 for (i = 0; i < ctx->nr_user_bufs; i++) {
7420 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7422 for (j = 0; j < imu->nr_bvecs; j++)
7423 unpin_user_page(imu->bvec[j].bv_page);
7425 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7430 kfree(ctx->user_bufs);
7431 ctx->user_bufs = NULL;
7432 ctx->nr_user_bufs = 0;
7436 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7437 void __user *arg, unsigned index)
7439 struct iovec __user *src;
7441 #ifdef CONFIG_COMPAT
7443 struct compat_iovec __user *ciovs;
7444 struct compat_iovec ciov;
7446 ciovs = (struct compat_iovec __user *) arg;
7447 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7450 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7451 dst->iov_len = ciov.iov_len;
7455 src = (struct iovec __user *) arg;
7456 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7461 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7464 struct vm_area_struct **vmas = NULL;
7465 struct page **pages = NULL;
7466 int i, j, got_pages = 0;
7471 if (!nr_args || nr_args > UIO_MAXIOV)
7474 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7476 if (!ctx->user_bufs)
7479 for (i = 0; i < nr_args; i++) {
7480 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7481 unsigned long off, start, end, ubuf;
7486 ret = io_copy_iov(ctx, &iov, arg, i);
7491 * Don't impose further limits on the size and buffer
7492 * constraints here, we'll -EINVAL later when IO is
7493 * submitted if they are wrong.
7496 if (!iov.iov_base || !iov.iov_len)
7499 /* arbitrary limit, but we need something */
7500 if (iov.iov_len > SZ_1G)
7503 ubuf = (unsigned long) iov.iov_base;
7504 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7505 start = ubuf >> PAGE_SHIFT;
7506 nr_pages = end - start;
7508 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7513 if (!pages || nr_pages > got_pages) {
7516 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7518 vmas = kvmalloc_array(nr_pages,
7519 sizeof(struct vm_area_struct *),
7521 if (!pages || !vmas) {
7523 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7526 got_pages = nr_pages;
7529 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7533 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7538 mmap_read_lock(current->mm);
7539 pret = pin_user_pages(ubuf, nr_pages,
7540 FOLL_WRITE | FOLL_LONGTERM,
7542 if (pret == nr_pages) {
7543 /* don't support file backed memory */
7544 for (j = 0; j < nr_pages; j++) {
7545 struct vm_area_struct *vma = vmas[j];
7548 !is_file_hugepages(vma->vm_file)) {
7554 ret = pret < 0 ? pret : -EFAULT;
7556 mmap_read_unlock(current->mm);
7559 * if we did partial map, or found file backed vmas,
7560 * release any pages we did get
7563 unpin_user_pages(pages, pret);
7564 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7569 off = ubuf & ~PAGE_MASK;
7571 for (j = 0; j < nr_pages; j++) {
7574 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7575 imu->bvec[j].bv_page = pages[j];
7576 imu->bvec[j].bv_len = vec_len;
7577 imu->bvec[j].bv_offset = off;
7581 /* store original address for later verification */
7583 imu->len = iov.iov_len;
7584 imu->nr_bvecs = nr_pages;
7586 ctx->nr_user_bufs++;
7594 io_sqe_buffer_unregister(ctx);
7598 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7600 __s32 __user *fds = arg;
7606 if (copy_from_user(&fd, fds, sizeof(*fds)))
7609 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7610 if (IS_ERR(ctx->cq_ev_fd)) {
7611 int ret = PTR_ERR(ctx->cq_ev_fd);
7612 ctx->cq_ev_fd = NULL;
7619 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7621 if (ctx->cq_ev_fd) {
7622 eventfd_ctx_put(ctx->cq_ev_fd);
7623 ctx->cq_ev_fd = NULL;
7630 static int __io_destroy_buffers(int id, void *p, void *data)
7632 struct io_ring_ctx *ctx = data;
7633 struct io_buffer *buf = p;
7635 __io_remove_buffers(ctx, buf, id, -1U);
7639 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7641 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7642 idr_destroy(&ctx->io_buffer_idr);
7645 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7647 io_finish_async(ctx);
7649 mmdrop(ctx->sqo_mm);
7653 io_iopoll_reap_events(ctx);
7654 io_sqe_buffer_unregister(ctx);
7655 io_sqe_files_unregister(ctx);
7656 io_eventfd_unregister(ctx);
7657 io_destroy_buffers(ctx);
7658 idr_destroy(&ctx->personality_idr);
7660 #if defined(CONFIG_UNIX)
7661 if (ctx->ring_sock) {
7662 ctx->ring_sock->file = NULL; /* so that iput() is called */
7663 sock_release(ctx->ring_sock);
7667 io_mem_free(ctx->rings);
7668 io_mem_free(ctx->sq_sqes);
7670 percpu_ref_exit(&ctx->refs);
7671 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7673 free_uid(ctx->user);
7674 put_cred(ctx->creds);
7675 kfree(ctx->cancel_hash);
7676 kmem_cache_free(req_cachep, ctx->fallback_req);
7680 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7682 struct io_ring_ctx *ctx = file->private_data;
7685 poll_wait(file, &ctx->cq_wait, wait);
7687 * synchronizes with barrier from wq_has_sleeper call in
7691 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7692 ctx->rings->sq_ring_entries)
7693 mask |= EPOLLOUT | EPOLLWRNORM;
7694 if (io_cqring_events(ctx, false))
7695 mask |= EPOLLIN | EPOLLRDNORM;
7700 static int io_uring_fasync(int fd, struct file *file, int on)
7702 struct io_ring_ctx *ctx = file->private_data;
7704 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7707 static int io_remove_personalities(int id, void *p, void *data)
7709 struct io_ring_ctx *ctx = data;
7710 const struct cred *cred;
7712 cred = idr_remove(&ctx->personality_idr, id);
7718 static void io_ring_exit_work(struct work_struct *work)
7720 struct io_ring_ctx *ctx;
7722 ctx = container_of(work, struct io_ring_ctx, exit_work);
7724 io_cqring_overflow_flush(ctx, true);
7727 * If we're doing polled IO and end up having requests being
7728 * submitted async (out-of-line), then completions can come in while
7729 * we're waiting for refs to drop. We need to reap these manually,
7730 * as nobody else will be looking for them.
7732 while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20)) {
7733 io_iopoll_reap_events(ctx);
7735 io_cqring_overflow_flush(ctx, true);
7737 io_ring_ctx_free(ctx);
7740 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7742 mutex_lock(&ctx->uring_lock);
7743 percpu_ref_kill(&ctx->refs);
7744 mutex_unlock(&ctx->uring_lock);
7746 io_kill_timeouts(ctx);
7747 io_poll_remove_all(ctx);
7750 io_wq_cancel_all(ctx->io_wq);
7752 io_iopoll_reap_events(ctx);
7753 /* if we failed setting up the ctx, we might not have any rings */
7755 io_cqring_overflow_flush(ctx, true);
7756 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7757 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7758 queue_work(system_wq, &ctx->exit_work);
7761 static int io_uring_release(struct inode *inode, struct file *file)
7763 struct io_ring_ctx *ctx = file->private_data;
7765 file->private_data = NULL;
7766 io_ring_ctx_wait_and_kill(ctx);
7770 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7772 struct files_struct *files = data;
7774 return work->files == files;
7777 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7778 struct files_struct *files)
7780 if (list_empty_careful(&ctx->inflight_list))
7783 /* cancel all at once, should be faster than doing it one by one*/
7784 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7786 while (!list_empty_careful(&ctx->inflight_list)) {
7787 struct io_kiocb *cancel_req = NULL, *req;
7790 spin_lock_irq(&ctx->inflight_lock);
7791 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7792 if (req->work.files != files)
7794 /* req is being completed, ignore */
7795 if (!refcount_inc_not_zero(&req->refs))
7801 prepare_to_wait(&ctx->inflight_wait, &wait,
7802 TASK_UNINTERRUPTIBLE);
7803 spin_unlock_irq(&ctx->inflight_lock);
7805 /* We need to keep going until we don't find a matching req */
7809 if (cancel_req->flags & REQ_F_OVERFLOW) {
7810 spin_lock_irq(&ctx->completion_lock);
7811 list_del(&cancel_req->list);
7812 cancel_req->flags &= ~REQ_F_OVERFLOW;
7813 if (list_empty(&ctx->cq_overflow_list)) {
7814 clear_bit(0, &ctx->sq_check_overflow);
7815 clear_bit(0, &ctx->cq_check_overflow);
7817 spin_unlock_irq(&ctx->completion_lock);
7819 WRITE_ONCE(ctx->rings->cq_overflow,
7820 atomic_inc_return(&ctx->cached_cq_overflow));
7823 * Put inflight ref and overflow ref. If that's
7824 * all we had, then we're done with this request.
7826 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7827 io_free_req(cancel_req);
7828 finish_wait(&ctx->inflight_wait, &wait);
7832 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7833 io_put_req(cancel_req);
7837 finish_wait(&ctx->inflight_wait, &wait);
7841 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7843 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7844 struct task_struct *task = data;
7846 return req->task == task;
7849 static int io_uring_flush(struct file *file, void *data)
7851 struct io_ring_ctx *ctx = file->private_data;
7853 io_uring_cancel_files(ctx, data);
7856 * If the task is going away, cancel work it may have pending
7858 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7859 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7864 static void *io_uring_validate_mmap_request(struct file *file,
7865 loff_t pgoff, size_t sz)
7867 struct io_ring_ctx *ctx = file->private_data;
7868 loff_t offset = pgoff << PAGE_SHIFT;
7873 case IORING_OFF_SQ_RING:
7874 case IORING_OFF_CQ_RING:
7877 case IORING_OFF_SQES:
7881 return ERR_PTR(-EINVAL);
7884 page = virt_to_head_page(ptr);
7885 if (sz > page_size(page))
7886 return ERR_PTR(-EINVAL);
7893 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7895 size_t sz = vma->vm_end - vma->vm_start;
7899 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7901 return PTR_ERR(ptr);
7903 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7904 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7907 #else /* !CONFIG_MMU */
7909 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7911 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7914 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7916 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7919 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7920 unsigned long addr, unsigned long len,
7921 unsigned long pgoff, unsigned long flags)
7925 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7927 return PTR_ERR(ptr);
7929 return (unsigned long) ptr;
7932 #endif /* !CONFIG_MMU */
7934 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7935 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7938 struct io_ring_ctx *ctx;
7943 if (current->task_works)
7946 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7954 if (f.file->f_op != &io_uring_fops)
7958 ctx = f.file->private_data;
7959 if (!percpu_ref_tryget(&ctx->refs))
7963 * For SQ polling, the thread will do all submissions and completions.
7964 * Just return the requested submit count, and wake the thread if
7968 if (ctx->flags & IORING_SETUP_SQPOLL) {
7969 if (!list_empty_careful(&ctx->cq_overflow_list))
7970 io_cqring_overflow_flush(ctx, false);
7971 if (flags & IORING_ENTER_SQ_WAKEUP)
7972 wake_up(&ctx->sqo_wait);
7973 submitted = to_submit;
7974 } else if (to_submit) {
7975 mutex_lock(&ctx->uring_lock);
7976 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
7977 mutex_unlock(&ctx->uring_lock);
7979 if (submitted != to_submit)
7982 if (flags & IORING_ENTER_GETEVENTS) {
7983 unsigned nr_events = 0;
7985 min_complete = min(min_complete, ctx->cq_entries);
7988 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7989 * space applications don't need to do io completion events
7990 * polling again, they can rely on io_sq_thread to do polling
7991 * work, which can reduce cpu usage and uring_lock contention.
7993 if (ctx->flags & IORING_SETUP_IOPOLL &&
7994 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7995 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7997 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8002 percpu_ref_put(&ctx->refs);
8005 return submitted ? submitted : ret;
8008 #ifdef CONFIG_PROC_FS
8009 static int io_uring_show_cred(int id, void *p, void *data)
8011 const struct cred *cred = p;
8012 struct seq_file *m = data;
8013 struct user_namespace *uns = seq_user_ns(m);
8014 struct group_info *gi;
8019 seq_printf(m, "%5d\n", id);
8020 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8021 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8022 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8023 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8024 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8025 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8026 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8027 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8028 seq_puts(m, "\n\tGroups:\t");
8029 gi = cred->group_info;
8030 for (g = 0; g < gi->ngroups; g++) {
8031 seq_put_decimal_ull(m, g ? " " : "",
8032 from_kgid_munged(uns, gi->gid[g]));
8034 seq_puts(m, "\n\tCapEff:\t");
8035 cap = cred->cap_effective;
8036 CAP_FOR_EACH_U32(__capi)
8037 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8042 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8046 mutex_lock(&ctx->uring_lock);
8047 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8048 for (i = 0; i < ctx->nr_user_files; i++) {
8049 struct fixed_file_table *table;
8052 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8053 f = table->files[i & IORING_FILE_TABLE_MASK];
8055 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8057 seq_printf(m, "%5u: <none>\n", i);
8059 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8060 for (i = 0; i < ctx->nr_user_bufs; i++) {
8061 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8063 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8064 (unsigned int) buf->len);
8066 if (!idr_is_empty(&ctx->personality_idr)) {
8067 seq_printf(m, "Personalities:\n");
8068 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8070 seq_printf(m, "PollList:\n");
8071 spin_lock_irq(&ctx->completion_lock);
8072 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8073 struct hlist_head *list = &ctx->cancel_hash[i];
8074 struct io_kiocb *req;
8076 hlist_for_each_entry(req, list, hash_node)
8077 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8078 req->task->task_works != NULL);
8080 spin_unlock_irq(&ctx->completion_lock);
8081 mutex_unlock(&ctx->uring_lock);
8084 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8086 struct io_ring_ctx *ctx = f->private_data;
8088 if (percpu_ref_tryget(&ctx->refs)) {
8089 __io_uring_show_fdinfo(ctx, m);
8090 percpu_ref_put(&ctx->refs);
8095 static const struct file_operations io_uring_fops = {
8096 .release = io_uring_release,
8097 .flush = io_uring_flush,
8098 .mmap = io_uring_mmap,
8100 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8101 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8103 .poll = io_uring_poll,
8104 .fasync = io_uring_fasync,
8105 #ifdef CONFIG_PROC_FS
8106 .show_fdinfo = io_uring_show_fdinfo,
8110 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8111 struct io_uring_params *p)
8113 struct io_rings *rings;
8114 size_t size, sq_array_offset;
8116 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8117 if (size == SIZE_MAX)
8120 rings = io_mem_alloc(size);
8125 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8126 rings->sq_ring_mask = p->sq_entries - 1;
8127 rings->cq_ring_mask = p->cq_entries - 1;
8128 rings->sq_ring_entries = p->sq_entries;
8129 rings->cq_ring_entries = p->cq_entries;
8130 ctx->sq_mask = rings->sq_ring_mask;
8131 ctx->cq_mask = rings->cq_ring_mask;
8132 ctx->sq_entries = rings->sq_ring_entries;
8133 ctx->cq_entries = rings->cq_ring_entries;
8135 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8136 if (size == SIZE_MAX) {
8137 io_mem_free(ctx->rings);
8142 ctx->sq_sqes = io_mem_alloc(size);
8143 if (!ctx->sq_sqes) {
8144 io_mem_free(ctx->rings);
8153 * Allocate an anonymous fd, this is what constitutes the application
8154 * visible backing of an io_uring instance. The application mmaps this
8155 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8156 * we have to tie this fd to a socket for file garbage collection purposes.
8158 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8163 #if defined(CONFIG_UNIX)
8164 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8170 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8174 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8175 O_RDWR | O_CLOEXEC);
8178 ret = PTR_ERR(file);
8182 #if defined(CONFIG_UNIX)
8183 ctx->ring_sock->file = file;
8185 fd_install(ret, file);
8188 #if defined(CONFIG_UNIX)
8189 sock_release(ctx->ring_sock);
8190 ctx->ring_sock = NULL;
8195 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8196 struct io_uring_params __user *params)
8198 struct user_struct *user = NULL;
8199 struct io_ring_ctx *ctx;
8205 if (entries > IORING_MAX_ENTRIES) {
8206 if (!(p->flags & IORING_SETUP_CLAMP))
8208 entries = IORING_MAX_ENTRIES;
8212 * Use twice as many entries for the CQ ring. It's possible for the
8213 * application to drive a higher depth than the size of the SQ ring,
8214 * since the sqes are only used at submission time. This allows for
8215 * some flexibility in overcommitting a bit. If the application has
8216 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8217 * of CQ ring entries manually.
8219 p->sq_entries = roundup_pow_of_two(entries);
8220 if (p->flags & IORING_SETUP_CQSIZE) {
8222 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8223 * to a power-of-two, if it isn't already. We do NOT impose
8224 * any cq vs sq ring sizing.
8226 if (p->cq_entries < p->sq_entries)
8228 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8229 if (!(p->flags & IORING_SETUP_CLAMP))
8231 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8233 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8235 p->cq_entries = 2 * p->sq_entries;
8238 user = get_uid(current_user());
8239 limit_mem = !capable(CAP_IPC_LOCK);
8242 ret = __io_account_mem(user,
8243 ring_pages(p->sq_entries, p->cq_entries));
8250 ctx = io_ring_ctx_alloc(p);
8253 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8258 ctx->compat = in_compat_syscall();
8260 ctx->creds = get_current_cred();
8262 ret = io_allocate_scq_urings(ctx, p);
8266 ret = io_sq_offload_start(ctx, p);
8270 memset(&p->sq_off, 0, sizeof(p->sq_off));
8271 p->sq_off.head = offsetof(struct io_rings, sq.head);
8272 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8273 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8274 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8275 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8276 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8277 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8279 memset(&p->cq_off, 0, sizeof(p->cq_off));
8280 p->cq_off.head = offsetof(struct io_rings, cq.head);
8281 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8282 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8283 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8284 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8285 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8286 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8288 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8289 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8290 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8291 IORING_FEAT_POLL_32BITS;
8293 if (copy_to_user(params, p, sizeof(*p))) {
8298 * Install ring fd as the very last thing, so we don't risk someone
8299 * having closed it before we finish setup
8301 ret = io_uring_get_fd(ctx);
8305 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8306 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8308 ctx->limit_mem = limit_mem;
8311 io_ring_ctx_wait_and_kill(ctx);
8316 * Sets up an aio uring context, and returns the fd. Applications asks for a
8317 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8318 * params structure passed in.
8320 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8322 struct io_uring_params p;
8325 if (copy_from_user(&p, params, sizeof(p)))
8327 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8332 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8333 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8334 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8337 return io_uring_create(entries, &p, params);
8340 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8341 struct io_uring_params __user *, params)
8343 return io_uring_setup(entries, params);
8346 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8348 struct io_uring_probe *p;
8352 size = struct_size(p, ops, nr_args);
8353 if (size == SIZE_MAX)
8355 p = kzalloc(size, GFP_KERNEL);
8360 if (copy_from_user(p, arg, size))
8363 if (memchr_inv(p, 0, size))
8366 p->last_op = IORING_OP_LAST - 1;
8367 if (nr_args > IORING_OP_LAST)
8368 nr_args = IORING_OP_LAST;
8370 for (i = 0; i < nr_args; i++) {
8372 if (!io_op_defs[i].not_supported)
8373 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8378 if (copy_to_user(arg, p, size))
8385 static int io_register_personality(struct io_ring_ctx *ctx)
8387 const struct cred *creds = get_current_cred();
8390 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8391 USHRT_MAX, GFP_KERNEL);
8397 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8399 const struct cred *old_creds;
8401 old_creds = idr_remove(&ctx->personality_idr, id);
8403 put_cred(old_creds);
8410 static bool io_register_op_must_quiesce(int op)
8413 case IORING_UNREGISTER_FILES:
8414 case IORING_REGISTER_FILES_UPDATE:
8415 case IORING_REGISTER_PROBE:
8416 case IORING_REGISTER_PERSONALITY:
8417 case IORING_UNREGISTER_PERSONALITY:
8424 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8425 void __user *arg, unsigned nr_args)
8426 __releases(ctx->uring_lock)
8427 __acquires(ctx->uring_lock)
8432 * We're inside the ring mutex, if the ref is already dying, then
8433 * someone else killed the ctx or is already going through
8434 * io_uring_register().
8436 if (percpu_ref_is_dying(&ctx->refs))
8439 if (io_register_op_must_quiesce(opcode)) {
8440 percpu_ref_kill(&ctx->refs);
8443 * Drop uring mutex before waiting for references to exit. If
8444 * another thread is currently inside io_uring_enter() it might
8445 * need to grab the uring_lock to make progress. If we hold it
8446 * here across the drain wait, then we can deadlock. It's safe
8447 * to drop the mutex here, since no new references will come in
8448 * after we've killed the percpu ref.
8450 mutex_unlock(&ctx->uring_lock);
8451 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8452 mutex_lock(&ctx->uring_lock);
8454 percpu_ref_resurrect(&ctx->refs);
8461 case IORING_REGISTER_BUFFERS:
8462 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8464 case IORING_UNREGISTER_BUFFERS:
8468 ret = io_sqe_buffer_unregister(ctx);
8470 case IORING_REGISTER_FILES:
8471 ret = io_sqe_files_register(ctx, arg, nr_args);
8473 case IORING_UNREGISTER_FILES:
8477 ret = io_sqe_files_unregister(ctx);
8479 case IORING_REGISTER_FILES_UPDATE:
8480 ret = io_sqe_files_update(ctx, arg, nr_args);
8482 case IORING_REGISTER_EVENTFD:
8483 case IORING_REGISTER_EVENTFD_ASYNC:
8487 ret = io_eventfd_register(ctx, arg);
8490 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8491 ctx->eventfd_async = 1;
8493 ctx->eventfd_async = 0;
8495 case IORING_UNREGISTER_EVENTFD:
8499 ret = io_eventfd_unregister(ctx);
8501 case IORING_REGISTER_PROBE:
8503 if (!arg || nr_args > 256)
8505 ret = io_probe(ctx, arg, nr_args);
8507 case IORING_REGISTER_PERSONALITY:
8511 ret = io_register_personality(ctx);
8513 case IORING_UNREGISTER_PERSONALITY:
8517 ret = io_unregister_personality(ctx, nr_args);
8524 if (io_register_op_must_quiesce(opcode)) {
8525 /* bring the ctx back to life */
8526 percpu_ref_reinit(&ctx->refs);
8528 reinit_completion(&ctx->ref_comp);
8533 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8534 void __user *, arg, unsigned int, nr_args)
8536 struct io_ring_ctx *ctx;
8545 if (f.file->f_op != &io_uring_fops)
8548 ctx = f.file->private_data;
8550 mutex_lock(&ctx->uring_lock);
8551 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8552 mutex_unlock(&ctx->uring_lock);
8553 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8554 ctx->cq_ev_fd != NULL, ret);
8560 static int __init io_uring_init(void)
8562 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8563 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8564 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8567 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8568 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8569 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8570 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8571 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8572 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8573 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8574 BUILD_BUG_SQE_ELEM(8, __u64, off);
8575 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8576 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8577 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8578 BUILD_BUG_SQE_ELEM(24, __u32, len);
8579 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8580 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8581 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8582 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8583 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8584 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8585 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8586 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8587 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8588 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8589 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8590 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8591 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8592 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8593 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8594 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8595 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8596 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8597 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8599 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8600 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8601 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8604 __initcall(io_uring_init);