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 <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/mmu_context.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>
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/io_uring.h>
83 #include <uapi/linux/io_uring.h>
88 #define IORING_MAX_ENTRIES 32768
89 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
92 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
94 #define IORING_FILE_TABLE_SHIFT 9
95 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
96 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
97 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 u32 head ____cacheline_aligned_in_smp;
101 u32 tail ____cacheline_aligned_in_smp;
105 * This data is shared with the application through the mmap at offsets
106 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
108 * The offsets to the member fields are published through struct
109 * io_sqring_offsets when calling io_uring_setup.
113 * Head and tail offsets into the ring; the offsets need to be
114 * masked to get valid indices.
116 * The kernel controls head of the sq ring and the tail of the cq ring,
117 * and the application controls tail of the sq ring and the head of the
120 struct io_uring sq, cq;
122 * Bitmasks to apply to head and tail offsets (constant, equals
125 u32 sq_ring_mask, cq_ring_mask;
126 /* Ring sizes (constant, power of 2) */
127 u32 sq_ring_entries, cq_ring_entries;
129 * Number of invalid entries dropped by the kernel due to
130 * invalid index stored in array
132 * Written by the kernel, shouldn't be modified by the
133 * application (i.e. get number of "new events" by comparing to
136 * After a new SQ head value was read by the application this
137 * counter includes all submissions that were dropped reaching
138 * the new SQ head (and possibly more).
144 * Written by the kernel, shouldn't be modified by the
147 * The application needs a full memory barrier before checking
148 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
152 * Number of completion events lost because the queue was full;
153 * this should be avoided by the application by making sure
154 * there are not more requests pending than there is space in
155 * the completion queue.
157 * Written by the kernel, shouldn't be modified by the
158 * application (i.e. get number of "new events" by comparing to
161 * As completion events come in out of order this counter is not
162 * ordered with any other data.
166 * Ring buffer of completion events.
168 * The kernel writes completion events fresh every time they are
169 * produced, so the application is allowed to modify pending
172 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
175 struct io_mapped_ubuf {
178 struct bio_vec *bvec;
179 unsigned int nr_bvecs;
182 struct fixed_file_table {
190 struct fixed_file_data {
191 struct fixed_file_table *table;
192 struct io_ring_ctx *ctx;
194 struct percpu_ref refs;
195 struct llist_head put_llist;
197 struct work_struct ref_work;
198 struct completion done;
203 struct percpu_ref refs;
204 } ____cacheline_aligned_in_smp;
208 unsigned int compat: 1;
209 unsigned int account_mem: 1;
210 unsigned int cq_overflow_flushed: 1;
211 unsigned int drain_next: 1;
212 unsigned int eventfd_async: 1;
215 * Ring buffer of indices into array of io_uring_sqe, which is
216 * mmapped by the application using the IORING_OFF_SQES offset.
218 * This indirection could e.g. be used to assign fixed
219 * io_uring_sqe entries to operations and only submit them to
220 * the queue when needed.
222 * The kernel modifies neither the indices array nor the entries
226 unsigned cached_sq_head;
229 unsigned sq_thread_idle;
230 unsigned cached_sq_dropped;
231 atomic_t cached_cq_overflow;
232 unsigned long sq_check_overflow;
234 struct list_head defer_list;
235 struct list_head timeout_list;
236 struct list_head cq_overflow_list;
238 wait_queue_head_t inflight_wait;
239 struct io_uring_sqe *sq_sqes;
240 } ____cacheline_aligned_in_smp;
242 struct io_rings *rings;
246 struct task_struct *sqo_thread; /* if using sq thread polling */
247 struct mm_struct *sqo_mm;
248 wait_queue_head_t sqo_wait;
251 * If used, fixed file set. Writers must ensure that ->refs is dead,
252 * readers must ensure that ->refs is alive as long as the file* is
253 * used. Only updated through io_uring_register(2).
255 struct fixed_file_data *file_data;
256 unsigned nr_user_files;
258 struct file *ring_file;
260 /* if used, fixed mapped user buffers */
261 unsigned nr_user_bufs;
262 struct io_mapped_ubuf *user_bufs;
264 struct user_struct *user;
266 const struct cred *creds;
268 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
269 struct completion *completions;
271 /* if all else fails... */
272 struct io_kiocb *fallback_req;
274 #if defined(CONFIG_UNIX)
275 struct socket *ring_sock;
278 struct idr personality_idr;
281 unsigned cached_cq_tail;
284 atomic_t cq_timeouts;
285 unsigned long cq_check_overflow;
286 struct wait_queue_head cq_wait;
287 struct fasync_struct *cq_fasync;
288 struct eventfd_ctx *cq_ev_fd;
289 } ____cacheline_aligned_in_smp;
292 struct mutex uring_lock;
293 wait_queue_head_t wait;
294 } ____cacheline_aligned_in_smp;
297 spinlock_t completion_lock;
298 struct llist_head poll_llist;
301 * ->poll_list is protected by the ctx->uring_lock for
302 * io_uring instances that don't use IORING_SETUP_SQPOLL.
303 * For SQPOLL, only the single threaded io_sq_thread() will
304 * manipulate the list, hence no extra locking is needed there.
306 struct list_head poll_list;
307 struct hlist_head *cancel_hash;
308 unsigned cancel_hash_bits;
309 bool poll_multi_file;
311 spinlock_t inflight_lock;
312 struct list_head inflight_list;
313 } ____cacheline_aligned_in_smp;
317 * First field must be the file pointer in all the
318 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
320 struct io_poll_iocb {
323 struct wait_queue_head *head;
329 struct wait_queue_entry wait;
334 struct file *put_file;
338 struct io_timeout_data {
339 struct io_kiocb *req;
340 struct hrtimer timer;
341 struct timespec64 ts;
342 enum hrtimer_mode mode;
348 struct sockaddr __user *addr;
349 int __user *addr_len;
374 /* NOTE: kiocb has the file as the first member, so don't do it here */
382 struct sockaddr __user *addr;
389 struct user_msghdr __user *msg;
402 struct filename *filename;
403 struct statx __user *buffer;
407 struct io_files_update {
433 struct epoll_event event;
436 struct io_async_connect {
437 struct sockaddr_storage address;
440 struct io_async_msghdr {
441 struct iovec fast_iov[UIO_FASTIOV];
443 struct sockaddr __user *uaddr;
448 struct iovec fast_iov[UIO_FASTIOV];
454 struct io_async_ctx {
456 struct io_async_rw rw;
457 struct io_async_msghdr msg;
458 struct io_async_connect connect;
459 struct io_timeout_data timeout;
464 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
465 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
466 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
467 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
468 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
475 REQ_F_IOPOLL_COMPLETED_BIT,
476 REQ_F_LINK_TIMEOUT_BIT,
480 REQ_F_TIMEOUT_NOSEQ_BIT,
481 REQ_F_COMP_LOCKED_BIT,
482 REQ_F_NEED_CLEANUP_BIT,
487 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
488 /* drain existing IO first */
489 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
491 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
492 /* doesn't sever on completion < 0 */
493 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
495 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
497 /* already grabbed next link */
498 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
499 /* fail rest of links */
500 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
501 /* on inflight list */
502 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
503 /* read/write uses file position */
504 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
505 /* must not punt to workers */
506 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
507 /* polled IO has completed */
508 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
509 /* has linked timeout */
510 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
511 /* timeout request */
512 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
514 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
515 /* must be punted even for NONBLOCK */
516 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
517 /* no timeout sequence */
518 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
519 /* completion under lock */
520 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
522 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
526 * NOTE! Each of the iocb union members has the file pointer
527 * as the first entry in their struct definition. So you can
528 * access the file pointer through any of the sub-structs,
529 * or directly as just 'ki_filp' in this struct.
535 struct io_poll_iocb poll;
536 struct io_accept accept;
538 struct io_cancel cancel;
539 struct io_timeout timeout;
540 struct io_connect connect;
541 struct io_sr_msg sr_msg;
543 struct io_close close;
544 struct io_files_update files_update;
545 struct io_fadvise fadvise;
546 struct io_madvise madvise;
547 struct io_epoll epoll;
550 struct io_async_ctx *io;
552 * llist_node is only used for poll deferred completions
554 struct llist_node llist_node;
556 bool needs_fixed_file;
559 struct io_ring_ctx *ctx;
561 struct list_head list;
562 struct hlist_node hash_node;
564 struct list_head link_list;
571 struct list_head inflight_entry;
573 struct io_wq_work work;
576 #define IO_PLUG_THRESHOLD 2
577 #define IO_IOPOLL_BATCH 8
579 struct io_submit_state {
580 struct blk_plug plug;
583 * io_kiocb alloc cache
585 void *reqs[IO_IOPOLL_BATCH];
586 unsigned int free_reqs;
589 * File reference cache
593 unsigned int has_refs;
594 unsigned int used_refs;
595 unsigned int ios_left;
599 /* needs req->io allocated for deferral/async */
600 unsigned async_ctx : 1;
601 /* needs current->mm setup, does mm access */
602 unsigned needs_mm : 1;
603 /* needs req->file assigned */
604 unsigned needs_file : 1;
605 /* needs req->file assigned IFF fd is >= 0 */
606 unsigned fd_non_neg : 1;
607 /* hash wq insertion if file is a regular file */
608 unsigned hash_reg_file : 1;
609 /* unbound wq insertion if file is a non-regular file */
610 unsigned unbound_nonreg_file : 1;
611 /* opcode is not supported by this kernel */
612 unsigned not_supported : 1;
613 /* needs file table */
614 unsigned file_table : 1;
616 unsigned needs_fs : 1;
619 static const struct io_op_def io_op_defs[] = {
620 [IORING_OP_NOP] = {},
621 [IORING_OP_READV] = {
625 .unbound_nonreg_file = 1,
627 [IORING_OP_WRITEV] = {
632 .unbound_nonreg_file = 1,
634 [IORING_OP_FSYNC] = {
637 [IORING_OP_READ_FIXED] = {
639 .unbound_nonreg_file = 1,
641 [IORING_OP_WRITE_FIXED] = {
644 .unbound_nonreg_file = 1,
646 [IORING_OP_POLL_ADD] = {
648 .unbound_nonreg_file = 1,
650 [IORING_OP_POLL_REMOVE] = {},
651 [IORING_OP_SYNC_FILE_RANGE] = {
654 [IORING_OP_SENDMSG] = {
658 .unbound_nonreg_file = 1,
661 [IORING_OP_RECVMSG] = {
665 .unbound_nonreg_file = 1,
668 [IORING_OP_TIMEOUT] = {
672 [IORING_OP_TIMEOUT_REMOVE] = {},
673 [IORING_OP_ACCEPT] = {
676 .unbound_nonreg_file = 1,
679 [IORING_OP_ASYNC_CANCEL] = {},
680 [IORING_OP_LINK_TIMEOUT] = {
684 [IORING_OP_CONNECT] = {
688 .unbound_nonreg_file = 1,
690 [IORING_OP_FALLOCATE] = {
693 [IORING_OP_OPENAT] = {
699 [IORING_OP_CLOSE] = {
703 [IORING_OP_FILES_UPDATE] = {
707 [IORING_OP_STATX] = {
716 .unbound_nonreg_file = 1,
718 [IORING_OP_WRITE] = {
721 .unbound_nonreg_file = 1,
723 [IORING_OP_FADVISE] = {
726 [IORING_OP_MADVISE] = {
732 .unbound_nonreg_file = 1,
737 .unbound_nonreg_file = 1,
739 [IORING_OP_OPENAT2] = {
745 [IORING_OP_EPOLL_CTL] = {
746 .unbound_nonreg_file = 1,
751 static void io_wq_submit_work(struct io_wq_work **workptr);
752 static void io_cqring_fill_event(struct io_kiocb *req, long res);
753 static void io_put_req(struct io_kiocb *req);
754 static void __io_double_put_req(struct io_kiocb *req);
755 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
756 static void io_queue_linked_timeout(struct io_kiocb *req);
757 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
758 struct io_uring_files_update *ip,
760 static int io_grab_files(struct io_kiocb *req);
761 static void io_ring_file_ref_flush(struct fixed_file_data *data);
762 static void io_cleanup_req(struct io_kiocb *req);
764 static struct kmem_cache *req_cachep;
766 static const struct file_operations io_uring_fops;
768 struct sock *io_uring_get_socket(struct file *file)
770 #if defined(CONFIG_UNIX)
771 if (file->f_op == &io_uring_fops) {
772 struct io_ring_ctx *ctx = file->private_data;
774 return ctx->ring_sock->sk;
779 EXPORT_SYMBOL(io_uring_get_socket);
781 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
783 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
785 complete(&ctx->completions[0]);
788 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
790 struct io_ring_ctx *ctx;
793 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
797 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
798 if (!ctx->fallback_req)
801 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
802 if (!ctx->completions)
806 * Use 5 bits less than the max cq entries, that should give us around
807 * 32 entries per hash list if totally full and uniformly spread.
809 hash_bits = ilog2(p->cq_entries);
813 ctx->cancel_hash_bits = hash_bits;
814 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
816 if (!ctx->cancel_hash)
818 __hash_init(ctx->cancel_hash, 1U << hash_bits);
820 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
821 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
824 ctx->flags = p->flags;
825 init_waitqueue_head(&ctx->cq_wait);
826 INIT_LIST_HEAD(&ctx->cq_overflow_list);
827 init_completion(&ctx->completions[0]);
828 init_completion(&ctx->completions[1]);
829 idr_init(&ctx->personality_idr);
830 mutex_init(&ctx->uring_lock);
831 init_waitqueue_head(&ctx->wait);
832 spin_lock_init(&ctx->completion_lock);
833 init_llist_head(&ctx->poll_llist);
834 INIT_LIST_HEAD(&ctx->poll_list);
835 INIT_LIST_HEAD(&ctx->defer_list);
836 INIT_LIST_HEAD(&ctx->timeout_list);
837 init_waitqueue_head(&ctx->inflight_wait);
838 spin_lock_init(&ctx->inflight_lock);
839 INIT_LIST_HEAD(&ctx->inflight_list);
842 if (ctx->fallback_req)
843 kmem_cache_free(req_cachep, ctx->fallback_req);
844 kfree(ctx->completions);
845 kfree(ctx->cancel_hash);
850 static inline bool __req_need_defer(struct io_kiocb *req)
852 struct io_ring_ctx *ctx = req->ctx;
854 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
855 + atomic_read(&ctx->cached_cq_overflow);
858 static inline bool req_need_defer(struct io_kiocb *req)
860 if (unlikely(req->flags & REQ_F_IO_DRAIN))
861 return __req_need_defer(req);
866 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
868 struct io_kiocb *req;
870 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
871 if (req && !req_need_defer(req)) {
872 list_del_init(&req->list);
879 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
881 struct io_kiocb *req;
883 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
885 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
887 if (!__req_need_defer(req)) {
888 list_del_init(&req->list);
896 static void __io_commit_cqring(struct io_ring_ctx *ctx)
898 struct io_rings *rings = ctx->rings;
900 /* order cqe stores with ring update */
901 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
903 if (wq_has_sleeper(&ctx->cq_wait)) {
904 wake_up_interruptible(&ctx->cq_wait);
905 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
909 static inline void io_req_work_grab_env(struct io_kiocb *req,
910 const struct io_op_def *def)
912 if (!req->work.mm && def->needs_mm) {
914 req->work.mm = current->mm;
916 if (!req->work.creds)
917 req->work.creds = get_current_cred();
918 if (!req->work.fs && def->needs_fs) {
919 spin_lock(¤t->fs->lock);
920 if (!current->fs->in_exec) {
921 req->work.fs = current->fs;
922 req->work.fs->users++;
924 req->work.flags |= IO_WQ_WORK_CANCEL;
926 spin_unlock(¤t->fs->lock);
930 static inline void io_req_work_drop_env(struct io_kiocb *req)
933 mmdrop(req->work.mm);
936 if (req->work.creds) {
937 put_cred(req->work.creds);
938 req->work.creds = NULL;
941 struct fs_struct *fs = req->work.fs;
943 spin_lock(&req->work.fs->lock);
946 spin_unlock(&req->work.fs->lock);
952 static inline bool io_prep_async_work(struct io_kiocb *req,
953 struct io_kiocb **link)
955 const struct io_op_def *def = &io_op_defs[req->opcode];
956 bool do_hashed = false;
958 if (req->flags & REQ_F_ISREG) {
959 if (def->hash_reg_file)
962 if (def->unbound_nonreg_file)
963 req->work.flags |= IO_WQ_WORK_UNBOUND;
966 io_req_work_grab_env(req, def);
968 *link = io_prep_linked_timeout(req);
972 static inline void io_queue_async_work(struct io_kiocb *req)
974 struct io_ring_ctx *ctx = req->ctx;
975 struct io_kiocb *link;
978 do_hashed = io_prep_async_work(req, &link);
980 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
983 io_wq_enqueue(ctx->io_wq, &req->work);
985 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
986 file_inode(req->file));
990 io_queue_linked_timeout(link);
993 static void io_kill_timeout(struct io_kiocb *req)
997 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
999 atomic_inc(&req->ctx->cq_timeouts);
1000 list_del_init(&req->list);
1001 io_cqring_fill_event(req, 0);
1006 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1008 struct io_kiocb *req, *tmp;
1010 spin_lock_irq(&ctx->completion_lock);
1011 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1012 io_kill_timeout(req);
1013 spin_unlock_irq(&ctx->completion_lock);
1016 static void io_commit_cqring(struct io_ring_ctx *ctx)
1018 struct io_kiocb *req;
1020 while ((req = io_get_timeout_req(ctx)) != NULL)
1021 io_kill_timeout(req);
1023 __io_commit_cqring(ctx);
1025 while ((req = io_get_deferred_req(ctx)) != NULL)
1026 io_queue_async_work(req);
1029 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1031 struct io_rings *rings = ctx->rings;
1034 tail = ctx->cached_cq_tail;
1036 * writes to the cq entry need to come after reading head; the
1037 * control dependency is enough as we're using WRITE_ONCE to
1040 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1043 ctx->cached_cq_tail++;
1044 return &rings->cqes[tail & ctx->cq_mask];
1047 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1051 if (!ctx->eventfd_async)
1053 return io_wq_current_is_worker() || in_interrupt();
1056 static void __io_cqring_ev_posted(struct io_ring_ctx *ctx, bool trigger_ev)
1058 if (waitqueue_active(&ctx->wait))
1059 wake_up(&ctx->wait);
1060 if (waitqueue_active(&ctx->sqo_wait))
1061 wake_up(&ctx->sqo_wait);
1063 eventfd_signal(ctx->cq_ev_fd, 1);
1066 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1068 __io_cqring_ev_posted(ctx, io_should_trigger_evfd(ctx));
1071 /* Returns true if there are no backlogged entries after the flush */
1072 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1074 struct io_rings *rings = ctx->rings;
1075 struct io_uring_cqe *cqe;
1076 struct io_kiocb *req;
1077 unsigned long flags;
1081 if (list_empty_careful(&ctx->cq_overflow_list))
1083 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1084 rings->cq_ring_entries))
1088 spin_lock_irqsave(&ctx->completion_lock, flags);
1090 /* if force is set, the ring is going away. always drop after that */
1092 ctx->cq_overflow_flushed = 1;
1095 while (!list_empty(&ctx->cq_overflow_list)) {
1096 cqe = io_get_cqring(ctx);
1100 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1102 list_move(&req->list, &list);
1104 WRITE_ONCE(cqe->user_data, req->user_data);
1105 WRITE_ONCE(cqe->res, req->result);
1106 WRITE_ONCE(cqe->flags, 0);
1108 WRITE_ONCE(ctx->rings->cq_overflow,
1109 atomic_inc_return(&ctx->cached_cq_overflow));
1113 io_commit_cqring(ctx);
1115 clear_bit(0, &ctx->sq_check_overflow);
1116 clear_bit(0, &ctx->cq_check_overflow);
1118 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1119 io_cqring_ev_posted(ctx);
1121 while (!list_empty(&list)) {
1122 req = list_first_entry(&list, struct io_kiocb, list);
1123 list_del(&req->list);
1130 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1132 struct io_ring_ctx *ctx = req->ctx;
1133 struct io_uring_cqe *cqe;
1135 trace_io_uring_complete(ctx, req->user_data, res);
1138 * If we can't get a cq entry, userspace overflowed the
1139 * submission (by quite a lot). Increment the overflow count in
1142 cqe = io_get_cqring(ctx);
1144 WRITE_ONCE(cqe->user_data, req->user_data);
1145 WRITE_ONCE(cqe->res, res);
1146 WRITE_ONCE(cqe->flags, 0);
1147 } else if (ctx->cq_overflow_flushed) {
1148 WRITE_ONCE(ctx->rings->cq_overflow,
1149 atomic_inc_return(&ctx->cached_cq_overflow));
1151 if (list_empty(&ctx->cq_overflow_list)) {
1152 set_bit(0, &ctx->sq_check_overflow);
1153 set_bit(0, &ctx->cq_check_overflow);
1155 refcount_inc(&req->refs);
1157 list_add_tail(&req->list, &ctx->cq_overflow_list);
1161 static void io_cqring_add_event(struct io_kiocb *req, long res)
1163 struct io_ring_ctx *ctx = req->ctx;
1164 unsigned long flags;
1166 spin_lock_irqsave(&ctx->completion_lock, flags);
1167 io_cqring_fill_event(req, res);
1168 io_commit_cqring(ctx);
1169 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1171 io_cqring_ev_posted(ctx);
1174 static inline bool io_is_fallback_req(struct io_kiocb *req)
1176 return req == (struct io_kiocb *)
1177 ((unsigned long) req->ctx->fallback_req & ~1UL);
1180 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1182 struct io_kiocb *req;
1184 req = ctx->fallback_req;
1185 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
1191 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
1192 struct io_submit_state *state)
1194 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1195 struct io_kiocb *req;
1198 req = kmem_cache_alloc(req_cachep, gfp);
1201 } else if (!state->free_reqs) {
1205 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1206 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1209 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1210 * retry single alloc to be on the safe side.
1212 if (unlikely(ret <= 0)) {
1213 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1214 if (!state->reqs[0])
1218 state->free_reqs = ret - 1;
1219 req = state->reqs[ret - 1];
1222 req = state->reqs[state->free_reqs];
1230 /* one is dropped after submission, the other at completion */
1231 refcount_set(&req->refs, 2);
1233 INIT_IO_WORK(&req->work, io_wq_submit_work);
1236 req = io_get_fallback_req(ctx);
1239 percpu_ref_put(&ctx->refs);
1243 static void __io_req_do_free(struct io_kiocb *req)
1245 if (likely(!io_is_fallback_req(req)))
1246 kmem_cache_free(req_cachep, req);
1248 clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req);
1251 static void __io_req_aux_free(struct io_kiocb *req)
1253 struct io_ring_ctx *ctx = req->ctx;
1257 if (req->flags & REQ_F_FIXED_FILE)
1258 percpu_ref_put(&ctx->file_data->refs);
1263 io_req_work_drop_env(req);
1266 static void __io_free_req(struct io_kiocb *req)
1268 __io_req_aux_free(req);
1270 if (req->flags & REQ_F_NEED_CLEANUP)
1271 io_cleanup_req(req);
1273 if (req->flags & REQ_F_INFLIGHT) {
1274 struct io_ring_ctx *ctx = req->ctx;
1275 unsigned long flags;
1277 spin_lock_irqsave(&ctx->inflight_lock, flags);
1278 list_del(&req->inflight_entry);
1279 if (waitqueue_active(&ctx->inflight_wait))
1280 wake_up(&ctx->inflight_wait);
1281 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1284 percpu_ref_put(&req->ctx->refs);
1285 __io_req_do_free(req);
1289 void *reqs[IO_IOPOLL_BATCH];
1294 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1296 int fixed_refs = rb->to_free;
1300 if (rb->need_iter) {
1301 int i, inflight = 0;
1302 unsigned long flags;
1305 for (i = 0; i < rb->to_free; i++) {
1306 struct io_kiocb *req = rb->reqs[i];
1308 if (req->flags & REQ_F_FIXED_FILE) {
1312 if (req->flags & REQ_F_INFLIGHT)
1314 __io_req_aux_free(req);
1319 spin_lock_irqsave(&ctx->inflight_lock, flags);
1320 for (i = 0; i < rb->to_free; i++) {
1321 struct io_kiocb *req = rb->reqs[i];
1323 if (req->flags & REQ_F_INFLIGHT) {
1324 list_del(&req->inflight_entry);
1329 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1331 if (waitqueue_active(&ctx->inflight_wait))
1332 wake_up(&ctx->inflight_wait);
1335 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1337 percpu_ref_put_many(&ctx->file_data->refs, fixed_refs);
1338 percpu_ref_put_many(&ctx->refs, rb->to_free);
1339 rb->to_free = rb->need_iter = 0;
1342 static bool io_link_cancel_timeout(struct io_kiocb *req)
1344 struct io_ring_ctx *ctx = req->ctx;
1347 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1349 io_cqring_fill_event(req, -ECANCELED);
1350 io_commit_cqring(ctx);
1351 req->flags &= ~REQ_F_LINK;
1359 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1361 struct io_ring_ctx *ctx = req->ctx;
1362 bool wake_ev = false;
1364 /* Already got next link */
1365 if (req->flags & REQ_F_LINK_NEXT)
1369 * The list should never be empty when we are called here. But could
1370 * potentially happen if the chain is messed up, check to be on the
1373 while (!list_empty(&req->link_list)) {
1374 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1375 struct io_kiocb, link_list);
1377 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1378 (nxt->flags & REQ_F_TIMEOUT))) {
1379 list_del_init(&nxt->link_list);
1380 wake_ev |= io_link_cancel_timeout(nxt);
1381 req->flags &= ~REQ_F_LINK_TIMEOUT;
1385 list_del_init(&req->link_list);
1386 if (!list_empty(&nxt->link_list))
1387 nxt->flags |= REQ_F_LINK;
1392 req->flags |= REQ_F_LINK_NEXT;
1394 io_cqring_ev_posted(ctx);
1398 * Called if REQ_F_LINK is set, and we fail the head request
1400 static void io_fail_links(struct io_kiocb *req)
1402 struct io_ring_ctx *ctx = req->ctx;
1403 unsigned long flags;
1405 spin_lock_irqsave(&ctx->completion_lock, flags);
1407 while (!list_empty(&req->link_list)) {
1408 struct io_kiocb *link = list_first_entry(&req->link_list,
1409 struct io_kiocb, link_list);
1411 list_del_init(&link->link_list);
1412 trace_io_uring_fail_link(req, link);
1414 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1415 link->opcode == IORING_OP_LINK_TIMEOUT) {
1416 io_link_cancel_timeout(link);
1418 io_cqring_fill_event(link, -ECANCELED);
1419 __io_double_put_req(link);
1421 req->flags &= ~REQ_F_LINK_TIMEOUT;
1424 io_commit_cqring(ctx);
1425 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1426 io_cqring_ev_posted(ctx);
1429 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1431 if (likely(!(req->flags & REQ_F_LINK)))
1435 * If LINK is set, we have dependent requests in this chain. If we
1436 * didn't fail this request, queue the first one up, moving any other
1437 * dependencies to the next request. In case of failure, fail the rest
1440 if (req->flags & REQ_F_FAIL_LINK) {
1442 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1443 REQ_F_LINK_TIMEOUT) {
1444 struct io_ring_ctx *ctx = req->ctx;
1445 unsigned long flags;
1448 * If this is a timeout link, we could be racing with the
1449 * timeout timer. Grab the completion lock for this case to
1450 * protect against that.
1452 spin_lock_irqsave(&ctx->completion_lock, flags);
1453 io_req_link_next(req, nxt);
1454 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1456 io_req_link_next(req, nxt);
1460 static void io_free_req(struct io_kiocb *req)
1462 struct io_kiocb *nxt = NULL;
1464 io_req_find_next(req, &nxt);
1468 io_queue_async_work(nxt);
1472 * Drop reference to request, return next in chain (if there is one) if this
1473 * was the last reference to this request.
1475 __attribute__((nonnull))
1476 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1478 io_req_find_next(req, nxtptr);
1480 if (refcount_dec_and_test(&req->refs))
1484 static void io_put_req(struct io_kiocb *req)
1486 if (refcount_dec_and_test(&req->refs))
1491 * Must only be used if we don't need to care about links, usually from
1492 * within the completion handling itself.
1494 static void __io_double_put_req(struct io_kiocb *req)
1496 /* drop both submit and complete references */
1497 if (refcount_sub_and_test(2, &req->refs))
1501 static void io_double_put_req(struct io_kiocb *req)
1503 /* drop both submit and complete references */
1504 if (refcount_sub_and_test(2, &req->refs))
1508 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1510 struct io_rings *rings = ctx->rings;
1512 if (test_bit(0, &ctx->cq_check_overflow)) {
1514 * noflush == true is from the waitqueue handler, just ensure
1515 * we wake up the task, and the next invocation will flush the
1516 * entries. We cannot safely to it from here.
1518 if (noflush && !list_empty(&ctx->cq_overflow_list))
1521 io_cqring_overflow_flush(ctx, false);
1524 /* See comment at the top of this file */
1526 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1529 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1531 struct io_rings *rings = ctx->rings;
1533 /* make sure SQ entry isn't read before tail */
1534 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1537 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1539 if ((req->flags & REQ_F_LINK) || io_is_fallback_req(req))
1542 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1545 rb->reqs[rb->to_free++] = req;
1546 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1547 io_free_req_many(req->ctx, rb);
1552 * Find and free completed poll iocbs
1554 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1555 struct list_head *done)
1557 struct req_batch rb;
1558 struct io_kiocb *req;
1560 rb.to_free = rb.need_iter = 0;
1561 while (!list_empty(done)) {
1562 req = list_first_entry(done, struct io_kiocb, list);
1563 list_del(&req->list);
1565 io_cqring_fill_event(req, req->result);
1568 if (refcount_dec_and_test(&req->refs) &&
1569 !io_req_multi_free(&rb, req))
1573 io_commit_cqring(ctx);
1574 io_free_req_many(ctx, &rb);
1577 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1580 struct io_kiocb *req, *tmp;
1586 * Only spin for completions if we don't have multiple devices hanging
1587 * off our complete list, and we're under the requested amount.
1589 spin = !ctx->poll_multi_file && *nr_events < min;
1592 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1593 struct kiocb *kiocb = &req->rw.kiocb;
1596 * Move completed entries to our local list. If we find a
1597 * request that requires polling, break out and complete
1598 * the done list first, if we have entries there.
1600 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1601 list_move_tail(&req->list, &done);
1604 if (!list_empty(&done))
1607 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1616 if (!list_empty(&done))
1617 io_iopoll_complete(ctx, nr_events, &done);
1623 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1624 * non-spinning poll check - we'll still enter the driver poll loop, but only
1625 * as a non-spinning completion check.
1627 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1630 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1633 ret = io_do_iopoll(ctx, nr_events, min);
1636 if (!min || *nr_events >= min)
1644 * We can't just wait for polled events to come to us, we have to actively
1645 * find and complete them.
1647 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1649 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1652 mutex_lock(&ctx->uring_lock);
1653 while (!list_empty(&ctx->poll_list)) {
1654 unsigned int nr_events = 0;
1656 io_iopoll_getevents(ctx, &nr_events, 1);
1659 * Ensure we allow local-to-the-cpu processing to take place,
1660 * in this case we need to ensure that we reap all events.
1664 mutex_unlock(&ctx->uring_lock);
1667 static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1670 int iters = 0, ret = 0;
1676 * Don't enter poll loop if we already have events pending.
1677 * If we do, we can potentially be spinning for commands that
1678 * already triggered a CQE (eg in error).
1680 if (io_cqring_events(ctx, false))
1684 * If a submit got punted to a workqueue, we can have the
1685 * application entering polling for a command before it gets
1686 * issued. That app will hold the uring_lock for the duration
1687 * of the poll right here, so we need to take a breather every
1688 * now and then to ensure that the issue has a chance to add
1689 * the poll to the issued list. Otherwise we can spin here
1690 * forever, while the workqueue is stuck trying to acquire the
1693 if (!(++iters & 7)) {
1694 mutex_unlock(&ctx->uring_lock);
1695 mutex_lock(&ctx->uring_lock);
1698 if (*nr_events < min)
1699 tmin = min - *nr_events;
1701 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1705 } while (min && !*nr_events && !need_resched());
1710 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1716 * We disallow the app entering submit/complete with polling, but we
1717 * still need to lock the ring to prevent racing with polled issue
1718 * that got punted to a workqueue.
1720 mutex_lock(&ctx->uring_lock);
1721 ret = __io_iopoll_check(ctx, nr_events, min);
1722 mutex_unlock(&ctx->uring_lock);
1726 static void kiocb_end_write(struct io_kiocb *req)
1729 * Tell lockdep we inherited freeze protection from submission
1732 if (req->flags & REQ_F_ISREG) {
1733 struct inode *inode = file_inode(req->file);
1735 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1737 file_end_write(req->file);
1740 static inline void req_set_fail_links(struct io_kiocb *req)
1742 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1743 req->flags |= REQ_F_FAIL_LINK;
1746 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1748 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1750 if (kiocb->ki_flags & IOCB_WRITE)
1751 kiocb_end_write(req);
1753 if (res != req->result)
1754 req_set_fail_links(req);
1755 io_cqring_add_event(req, res);
1758 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1760 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1762 io_complete_rw_common(kiocb, res);
1766 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1768 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1769 struct io_kiocb *nxt = NULL;
1771 io_complete_rw_common(kiocb, res);
1772 io_put_req_find_next(req, &nxt);
1777 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1779 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1781 if (kiocb->ki_flags & IOCB_WRITE)
1782 kiocb_end_write(req);
1784 if (res != req->result)
1785 req_set_fail_links(req);
1788 req->flags |= REQ_F_IOPOLL_COMPLETED;
1792 * After the iocb has been issued, it's safe to be found on the poll list.
1793 * Adding the kiocb to the list AFTER submission ensures that we don't
1794 * find it from a io_iopoll_getevents() thread before the issuer is done
1795 * accessing the kiocb cookie.
1797 static void io_iopoll_req_issued(struct io_kiocb *req)
1799 struct io_ring_ctx *ctx = req->ctx;
1802 * Track whether we have multiple files in our lists. This will impact
1803 * how we do polling eventually, not spinning if we're on potentially
1804 * different devices.
1806 if (list_empty(&ctx->poll_list)) {
1807 ctx->poll_multi_file = false;
1808 } else if (!ctx->poll_multi_file) {
1809 struct io_kiocb *list_req;
1811 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1813 if (list_req->file != req->file)
1814 ctx->poll_multi_file = true;
1818 * For fast devices, IO may have already completed. If it has, add
1819 * it to the front so we find it first.
1821 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1822 list_add(&req->list, &ctx->poll_list);
1824 list_add_tail(&req->list, &ctx->poll_list);
1827 static void io_file_put(struct io_submit_state *state)
1830 int diff = state->has_refs - state->used_refs;
1833 fput_many(state->file, diff);
1839 * Get as many references to a file as we have IOs left in this submission,
1840 * assuming most submissions are for one file, or at least that each file
1841 * has more than one submission.
1843 static struct file *io_file_get(struct io_submit_state *state, int fd)
1849 if (state->fd == fd) {
1856 state->file = fget_many(fd, state->ios_left);
1861 state->has_refs = state->ios_left;
1862 state->used_refs = 1;
1868 * If we tracked the file through the SCM inflight mechanism, we could support
1869 * any file. For now, just ensure that anything potentially problematic is done
1872 static bool io_file_supports_async(struct file *file)
1874 umode_t mode = file_inode(file)->i_mode;
1876 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
1878 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1884 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1885 bool force_nonblock)
1887 struct io_ring_ctx *ctx = req->ctx;
1888 struct kiocb *kiocb = &req->rw.kiocb;
1892 if (S_ISREG(file_inode(req->file)->i_mode))
1893 req->flags |= REQ_F_ISREG;
1895 kiocb->ki_pos = READ_ONCE(sqe->off);
1896 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
1897 req->flags |= REQ_F_CUR_POS;
1898 kiocb->ki_pos = req->file->f_pos;
1900 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1901 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1902 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1906 ioprio = READ_ONCE(sqe->ioprio);
1908 ret = ioprio_check_cap(ioprio);
1912 kiocb->ki_ioprio = ioprio;
1914 kiocb->ki_ioprio = get_current_ioprio();
1916 /* don't allow async punt if RWF_NOWAIT was requested */
1917 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1918 (req->file->f_flags & O_NONBLOCK))
1919 req->flags |= REQ_F_NOWAIT;
1922 kiocb->ki_flags |= IOCB_NOWAIT;
1924 if (ctx->flags & IORING_SETUP_IOPOLL) {
1925 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1926 !kiocb->ki_filp->f_op->iopoll)
1929 kiocb->ki_flags |= IOCB_HIPRI;
1930 kiocb->ki_complete = io_complete_rw_iopoll;
1933 if (kiocb->ki_flags & IOCB_HIPRI)
1935 kiocb->ki_complete = io_complete_rw;
1938 req->rw.addr = READ_ONCE(sqe->addr);
1939 req->rw.len = READ_ONCE(sqe->len);
1940 /* we own ->private, reuse it for the buffer index */
1941 req->rw.kiocb.private = (void *) (unsigned long)
1942 READ_ONCE(sqe->buf_index);
1946 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1952 case -ERESTARTNOINTR:
1953 case -ERESTARTNOHAND:
1954 case -ERESTART_RESTARTBLOCK:
1956 * We can't just restart the syscall, since previously
1957 * submitted sqes may already be in progress. Just fail this
1963 kiocb->ki_complete(kiocb, ret, 0);
1967 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1970 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1972 if (req->flags & REQ_F_CUR_POS)
1973 req->file->f_pos = kiocb->ki_pos;
1974 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1975 *nxt = __io_complete_rw(kiocb, ret);
1977 io_rw_done(kiocb, ret);
1980 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
1981 struct iov_iter *iter)
1983 struct io_ring_ctx *ctx = req->ctx;
1984 size_t len = req->rw.len;
1985 struct io_mapped_ubuf *imu;
1986 unsigned index, buf_index;
1990 /* attempt to use fixed buffers without having provided iovecs */
1991 if (unlikely(!ctx->user_bufs))
1994 buf_index = (unsigned long) req->rw.kiocb.private;
1995 if (unlikely(buf_index >= ctx->nr_user_bufs))
1998 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1999 imu = &ctx->user_bufs[index];
2000 buf_addr = req->rw.addr;
2003 if (buf_addr + len < buf_addr)
2005 /* not inside the mapped region */
2006 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2010 * May not be a start of buffer, set size appropriately
2011 * and advance us to the beginning.
2013 offset = buf_addr - imu->ubuf;
2014 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2018 * Don't use iov_iter_advance() here, as it's really slow for
2019 * using the latter parts of a big fixed buffer - it iterates
2020 * over each segment manually. We can cheat a bit here, because
2023 * 1) it's a BVEC iter, we set it up
2024 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2025 * first and last bvec
2027 * So just find our index, and adjust the iterator afterwards.
2028 * If the offset is within the first bvec (or the whole first
2029 * bvec, just use iov_iter_advance(). This makes it easier
2030 * since we can just skip the first segment, which may not
2031 * be PAGE_SIZE aligned.
2033 const struct bio_vec *bvec = imu->bvec;
2035 if (offset <= bvec->bv_len) {
2036 iov_iter_advance(iter, offset);
2038 unsigned long seg_skip;
2040 /* skip first vec */
2041 offset -= bvec->bv_len;
2042 seg_skip = 1 + (offset >> PAGE_SHIFT);
2044 iter->bvec = bvec + seg_skip;
2045 iter->nr_segs -= seg_skip;
2046 iter->count -= bvec->bv_len + offset;
2047 iter->iov_offset = offset & ~PAGE_MASK;
2054 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2055 struct iovec **iovec, struct iov_iter *iter)
2057 void __user *buf = u64_to_user_ptr(req->rw.addr);
2058 size_t sqe_len = req->rw.len;
2061 opcode = req->opcode;
2062 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2064 return io_import_fixed(req, rw, iter);
2067 /* buffer index only valid with fixed read/write */
2068 if (req->rw.kiocb.private)
2071 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2073 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2079 struct io_async_rw *iorw = &req->io->rw;
2082 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2083 if (iorw->iov == iorw->fast_iov)
2088 #ifdef CONFIG_COMPAT
2089 if (req->ctx->compat)
2090 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2094 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2098 * For files that don't have ->read_iter() and ->write_iter(), handle them
2099 * by looping over ->read() or ->write() manually.
2101 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2102 struct iov_iter *iter)
2107 * Don't support polled IO through this interface, and we can't
2108 * support non-blocking either. For the latter, this just causes
2109 * the kiocb to be handled from an async context.
2111 if (kiocb->ki_flags & IOCB_HIPRI)
2113 if (kiocb->ki_flags & IOCB_NOWAIT)
2116 while (iov_iter_count(iter)) {
2120 if (!iov_iter_is_bvec(iter)) {
2121 iovec = iov_iter_iovec(iter);
2123 /* fixed buffers import bvec */
2124 iovec.iov_base = kmap(iter->bvec->bv_page)
2126 iovec.iov_len = min(iter->count,
2127 iter->bvec->bv_len - iter->iov_offset);
2131 nr = file->f_op->read(file, iovec.iov_base,
2132 iovec.iov_len, &kiocb->ki_pos);
2134 nr = file->f_op->write(file, iovec.iov_base,
2135 iovec.iov_len, &kiocb->ki_pos);
2138 if (iov_iter_is_bvec(iter))
2139 kunmap(iter->bvec->bv_page);
2147 if (nr != iovec.iov_len)
2149 iov_iter_advance(iter, nr);
2155 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2156 struct iovec *iovec, struct iovec *fast_iov,
2157 struct iov_iter *iter)
2159 req->io->rw.nr_segs = iter->nr_segs;
2160 req->io->rw.size = io_size;
2161 req->io->rw.iov = iovec;
2162 if (!req->io->rw.iov) {
2163 req->io->rw.iov = req->io->rw.fast_iov;
2164 memcpy(req->io->rw.iov, fast_iov,
2165 sizeof(struct iovec) * iter->nr_segs);
2167 req->flags |= REQ_F_NEED_CLEANUP;
2171 static int io_alloc_async_ctx(struct io_kiocb *req)
2173 if (!io_op_defs[req->opcode].async_ctx)
2175 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2176 return req->io == NULL;
2179 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2180 struct iovec *iovec, struct iovec *fast_iov,
2181 struct iov_iter *iter)
2183 if (!io_op_defs[req->opcode].async_ctx)
2186 if (io_alloc_async_ctx(req))
2189 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2194 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2195 bool force_nonblock)
2197 struct io_async_ctx *io;
2198 struct iov_iter iter;
2201 ret = io_prep_rw(req, sqe, force_nonblock);
2205 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2208 /* either don't need iovec imported or already have it */
2209 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2213 io->rw.iov = io->rw.fast_iov;
2215 ret = io_import_iovec(READ, req, &io->rw.iov, &iter);
2220 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2224 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
2225 bool force_nonblock)
2227 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2228 struct kiocb *kiocb = &req->rw.kiocb;
2229 struct iov_iter iter;
2231 ssize_t io_size, ret;
2233 ret = io_import_iovec(READ, req, &iovec, &iter);
2237 /* Ensure we clear previously set non-block flag */
2238 if (!force_nonblock)
2239 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2243 if (req->flags & REQ_F_LINK)
2244 req->result = io_size;
2247 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2248 * we know to async punt it even if it was opened O_NONBLOCK
2250 if (force_nonblock && !io_file_supports_async(req->file)) {
2251 req->flags |= REQ_F_MUST_PUNT;
2255 iov_count = iov_iter_count(&iter);
2256 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2260 if (req->file->f_op->read_iter)
2261 ret2 = call_read_iter(req->file, kiocb, &iter);
2263 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2265 /* Catch -EAGAIN return for forced non-blocking submission */
2266 if (!force_nonblock || ret2 != -EAGAIN) {
2267 kiocb_done(kiocb, ret2, nxt, req->in_async);
2270 ret = io_setup_async_rw(req, io_size, iovec,
2271 inline_vecs, &iter);
2279 req->flags &= ~REQ_F_NEED_CLEANUP;
2283 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2284 bool force_nonblock)
2286 struct io_async_ctx *io;
2287 struct iov_iter iter;
2290 ret = io_prep_rw(req, sqe, force_nonblock);
2294 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2297 /* either don't need iovec imported or already have it */
2298 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2302 io->rw.iov = io->rw.fast_iov;
2304 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter);
2309 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2313 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
2314 bool force_nonblock)
2316 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2317 struct kiocb *kiocb = &req->rw.kiocb;
2318 struct iov_iter iter;
2320 ssize_t ret, io_size;
2322 ret = io_import_iovec(WRITE, req, &iovec, &iter);
2326 /* Ensure we clear previously set non-block flag */
2327 if (!force_nonblock)
2328 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2332 if (req->flags & REQ_F_LINK)
2333 req->result = io_size;
2336 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2337 * we know to async punt it even if it was opened O_NONBLOCK
2339 if (force_nonblock && !io_file_supports_async(req->file)) {
2340 req->flags |= REQ_F_MUST_PUNT;
2344 /* file path doesn't support NOWAIT for non-direct_IO */
2345 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2346 (req->flags & REQ_F_ISREG))
2349 iov_count = iov_iter_count(&iter);
2350 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2355 * Open-code file_start_write here to grab freeze protection,
2356 * which will be released by another thread in
2357 * io_complete_rw(). Fool lockdep by telling it the lock got
2358 * released so that it doesn't complain about the held lock when
2359 * we return to userspace.
2361 if (req->flags & REQ_F_ISREG) {
2362 __sb_start_write(file_inode(req->file)->i_sb,
2363 SB_FREEZE_WRITE, true);
2364 __sb_writers_release(file_inode(req->file)->i_sb,
2367 kiocb->ki_flags |= IOCB_WRITE;
2369 if (req->file->f_op->write_iter)
2370 ret2 = call_write_iter(req->file, kiocb, &iter);
2372 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2374 * Raw bdev writes will -EOPNOTSUPP for IOCB_NOWAIT. Just
2375 * retry them without IOCB_NOWAIT.
2377 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2379 if (!force_nonblock || ret2 != -EAGAIN) {
2380 kiocb_done(kiocb, ret2, nxt, req->in_async);
2383 ret = io_setup_async_rw(req, io_size, iovec,
2384 inline_vecs, &iter);
2391 req->flags &= ~REQ_F_NEED_CLEANUP;
2397 * IORING_OP_NOP just posts a completion event, nothing else.
2399 static int io_nop(struct io_kiocb *req)
2401 struct io_ring_ctx *ctx = req->ctx;
2403 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2406 io_cqring_add_event(req, 0);
2411 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2413 struct io_ring_ctx *ctx = req->ctx;
2418 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2420 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2423 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2424 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2427 req->sync.off = READ_ONCE(sqe->off);
2428 req->sync.len = READ_ONCE(sqe->len);
2432 static bool io_req_cancelled(struct io_kiocb *req)
2434 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2435 req_set_fail_links(req);
2436 io_cqring_add_event(req, -ECANCELED);
2444 static void io_link_work_cb(struct io_wq_work **workptr)
2446 struct io_wq_work *work = *workptr;
2447 struct io_kiocb *link = work->data;
2449 io_queue_linked_timeout(link);
2450 work->func = io_wq_submit_work;
2453 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
2455 struct io_kiocb *link;
2457 io_prep_async_work(nxt, &link);
2458 *workptr = &nxt->work;
2460 nxt->work.flags |= IO_WQ_WORK_CB;
2461 nxt->work.func = io_link_work_cb;
2462 nxt->work.data = link;
2466 static void io_fsync_finish(struct io_wq_work **workptr)
2468 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2469 loff_t end = req->sync.off + req->sync.len;
2470 struct io_kiocb *nxt = NULL;
2473 if (io_req_cancelled(req))
2476 ret = vfs_fsync_range(req->file, req->sync.off,
2477 end > 0 ? end : LLONG_MAX,
2478 req->sync.flags & IORING_FSYNC_DATASYNC);
2480 req_set_fail_links(req);
2481 io_cqring_add_event(req, ret);
2482 io_put_req_find_next(req, &nxt);
2484 io_wq_assign_next(workptr, nxt);
2487 static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt,
2488 bool force_nonblock)
2490 struct io_wq_work *work, *old_work;
2492 /* fsync always requires a blocking context */
2493 if (force_nonblock) {
2495 req->work.func = io_fsync_finish;
2499 work = old_work = &req->work;
2500 io_fsync_finish(&work);
2501 if (work && work != old_work)
2502 *nxt = container_of(work, struct io_kiocb, work);
2506 static void io_fallocate_finish(struct io_wq_work **workptr)
2508 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2509 struct io_kiocb *nxt = NULL;
2512 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2515 req_set_fail_links(req);
2516 io_cqring_add_event(req, ret);
2517 io_put_req_find_next(req, &nxt);
2519 io_wq_assign_next(workptr, nxt);
2522 static int io_fallocate_prep(struct io_kiocb *req,
2523 const struct io_uring_sqe *sqe)
2525 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2528 req->sync.off = READ_ONCE(sqe->off);
2529 req->sync.len = READ_ONCE(sqe->addr);
2530 req->sync.mode = READ_ONCE(sqe->len);
2534 static int io_fallocate(struct io_kiocb *req, struct io_kiocb **nxt,
2535 bool force_nonblock)
2537 struct io_wq_work *work, *old_work;
2539 /* fallocate always requiring blocking context */
2540 if (force_nonblock) {
2542 req->work.func = io_fallocate_finish;
2546 work = old_work = &req->work;
2547 io_fallocate_finish(&work);
2548 if (work && work != old_work)
2549 *nxt = container_of(work, struct io_kiocb, work);
2554 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2556 const char __user *fname;
2559 if (sqe->ioprio || sqe->buf_index)
2561 if (sqe->flags & IOSQE_FIXED_FILE)
2563 if (req->flags & REQ_F_NEED_CLEANUP)
2566 req->open.dfd = READ_ONCE(sqe->fd);
2567 req->open.how.mode = READ_ONCE(sqe->len);
2568 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2569 req->open.how.flags = READ_ONCE(sqe->open_flags);
2571 req->open.filename = getname(fname);
2572 if (IS_ERR(req->open.filename)) {
2573 ret = PTR_ERR(req->open.filename);
2574 req->open.filename = NULL;
2578 req->flags |= REQ_F_NEED_CLEANUP;
2582 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2584 struct open_how __user *how;
2585 const char __user *fname;
2589 if (sqe->ioprio || sqe->buf_index)
2591 if (sqe->flags & IOSQE_FIXED_FILE)
2593 if (req->flags & REQ_F_NEED_CLEANUP)
2596 req->open.dfd = READ_ONCE(sqe->fd);
2597 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2598 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2599 len = READ_ONCE(sqe->len);
2601 if (len < OPEN_HOW_SIZE_VER0)
2604 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2609 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2610 req->open.how.flags |= O_LARGEFILE;
2612 req->open.filename = getname(fname);
2613 if (IS_ERR(req->open.filename)) {
2614 ret = PTR_ERR(req->open.filename);
2615 req->open.filename = NULL;
2619 req->flags |= REQ_F_NEED_CLEANUP;
2623 static int io_openat2(struct io_kiocb *req, struct io_kiocb **nxt,
2624 bool force_nonblock)
2626 struct open_flags op;
2633 ret = build_open_flags(&req->open.how, &op);
2637 ret = get_unused_fd_flags(req->open.how.flags);
2641 file = do_filp_open(req->open.dfd, req->open.filename, &op);
2644 ret = PTR_ERR(file);
2646 fsnotify_open(file);
2647 fd_install(ret, file);
2650 putname(req->open.filename);
2651 req->flags &= ~REQ_F_NEED_CLEANUP;
2653 req_set_fail_links(req);
2654 io_cqring_add_event(req, ret);
2655 io_put_req_find_next(req, nxt);
2659 static int io_openat(struct io_kiocb *req, struct io_kiocb **nxt,
2660 bool force_nonblock)
2662 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
2663 return io_openat2(req, nxt, force_nonblock);
2666 static int io_epoll_ctl_prep(struct io_kiocb *req,
2667 const struct io_uring_sqe *sqe)
2669 #if defined(CONFIG_EPOLL)
2670 if (sqe->ioprio || sqe->buf_index)
2673 req->epoll.epfd = READ_ONCE(sqe->fd);
2674 req->epoll.op = READ_ONCE(sqe->len);
2675 req->epoll.fd = READ_ONCE(sqe->off);
2677 if (ep_op_has_event(req->epoll.op)) {
2678 struct epoll_event __user *ev;
2680 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
2681 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
2691 static int io_epoll_ctl(struct io_kiocb *req, struct io_kiocb **nxt,
2692 bool force_nonblock)
2694 #if defined(CONFIG_EPOLL)
2695 struct io_epoll *ie = &req->epoll;
2698 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
2699 if (force_nonblock && ret == -EAGAIN)
2703 req_set_fail_links(req);
2704 io_cqring_add_event(req, ret);
2705 io_put_req_find_next(req, nxt);
2712 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2714 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2715 if (sqe->ioprio || sqe->buf_index || sqe->off)
2718 req->madvise.addr = READ_ONCE(sqe->addr);
2719 req->madvise.len = READ_ONCE(sqe->len);
2720 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
2727 static int io_madvise(struct io_kiocb *req, struct io_kiocb **nxt,
2728 bool force_nonblock)
2730 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2731 struct io_madvise *ma = &req->madvise;
2737 ret = do_madvise(ma->addr, ma->len, ma->advice);
2739 req_set_fail_links(req);
2740 io_cqring_add_event(req, ret);
2741 io_put_req_find_next(req, nxt);
2748 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2750 if (sqe->ioprio || sqe->buf_index || sqe->addr)
2753 req->fadvise.offset = READ_ONCE(sqe->off);
2754 req->fadvise.len = READ_ONCE(sqe->len);
2755 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
2759 static int io_fadvise(struct io_kiocb *req, struct io_kiocb **nxt,
2760 bool force_nonblock)
2762 struct io_fadvise *fa = &req->fadvise;
2765 if (force_nonblock) {
2766 switch (fa->advice) {
2767 case POSIX_FADV_NORMAL:
2768 case POSIX_FADV_RANDOM:
2769 case POSIX_FADV_SEQUENTIAL:
2776 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
2778 req_set_fail_links(req);
2779 io_cqring_add_event(req, ret);
2780 io_put_req_find_next(req, nxt);
2784 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2786 const char __user *fname;
2787 unsigned lookup_flags;
2790 if (sqe->ioprio || sqe->buf_index)
2792 if (sqe->flags & IOSQE_FIXED_FILE)
2794 if (req->flags & REQ_F_NEED_CLEANUP)
2797 req->open.dfd = READ_ONCE(sqe->fd);
2798 req->open.mask = READ_ONCE(sqe->len);
2799 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2800 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2801 req->open.how.flags = READ_ONCE(sqe->statx_flags);
2803 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
2806 req->open.filename = getname_flags(fname, lookup_flags, NULL);
2807 if (IS_ERR(req->open.filename)) {
2808 ret = PTR_ERR(req->open.filename);
2809 req->open.filename = NULL;
2813 req->flags |= REQ_F_NEED_CLEANUP;
2817 static int io_statx(struct io_kiocb *req, struct io_kiocb **nxt,
2818 bool force_nonblock)
2820 struct io_open *ctx = &req->open;
2821 unsigned lookup_flags;
2829 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
2833 /* filename_lookup() drops it, keep a reference */
2834 ctx->filename->refcnt++;
2836 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
2841 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
2843 if (retry_estale(ret, lookup_flags)) {
2844 lookup_flags |= LOOKUP_REVAL;
2848 ret = cp_statx(&stat, ctx->buffer);
2850 putname(ctx->filename);
2851 req->flags &= ~REQ_F_NEED_CLEANUP;
2853 req_set_fail_links(req);
2854 io_cqring_add_event(req, ret);
2855 io_put_req_find_next(req, nxt);
2859 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2862 * If we queue this for async, it must not be cancellable. That would
2863 * leave the 'file' in an undeterminate state.
2865 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
2867 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
2868 sqe->rw_flags || sqe->buf_index)
2870 if (sqe->flags & IOSQE_FIXED_FILE)
2873 req->close.fd = READ_ONCE(sqe->fd);
2874 if (req->file->f_op == &io_uring_fops ||
2875 req->close.fd == req->ctx->ring_fd)
2881 /* only called when __close_fd_get_file() is done */
2882 static void __io_close_finish(struct io_kiocb *req, struct io_kiocb **nxt)
2886 ret = filp_close(req->close.put_file, req->work.files);
2888 req_set_fail_links(req);
2889 io_cqring_add_event(req, ret);
2890 fput(req->close.put_file);
2891 io_put_req_find_next(req, nxt);
2894 static void io_close_finish(struct io_wq_work **workptr)
2896 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2897 struct io_kiocb *nxt = NULL;
2899 __io_close_finish(req, &nxt);
2901 io_wq_assign_next(workptr, nxt);
2904 static int io_close(struct io_kiocb *req, struct io_kiocb **nxt,
2905 bool force_nonblock)
2909 req->close.put_file = NULL;
2910 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
2914 /* if the file has a flush method, be safe and punt to async */
2915 if (req->close.put_file->f_op->flush && !io_wq_current_is_worker())
2919 * No ->flush(), safely close from here and just punt the
2920 * fput() to async context.
2922 __io_close_finish(req, nxt);
2925 req->work.func = io_close_finish;
2927 * Do manual async queue here to avoid grabbing files - we don't
2928 * need the files, and it'll cause io_close_finish() to close
2929 * the file again and cause a double CQE entry for this request
2931 io_queue_async_work(req);
2935 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2937 struct io_ring_ctx *ctx = req->ctx;
2942 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2944 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2947 req->sync.off = READ_ONCE(sqe->off);
2948 req->sync.len = READ_ONCE(sqe->len);
2949 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
2953 static void io_sync_file_range_finish(struct io_wq_work **workptr)
2955 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2956 struct io_kiocb *nxt = NULL;
2959 if (io_req_cancelled(req))
2962 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
2965 req_set_fail_links(req);
2966 io_cqring_add_event(req, ret);
2967 io_put_req_find_next(req, &nxt);
2969 io_wq_assign_next(workptr, nxt);
2972 static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt,
2973 bool force_nonblock)
2975 struct io_wq_work *work, *old_work;
2977 /* sync_file_range always requires a blocking context */
2978 if (force_nonblock) {
2980 req->work.func = io_sync_file_range_finish;
2984 work = old_work = &req->work;
2985 io_sync_file_range_finish(&work);
2986 if (work && work != old_work)
2987 *nxt = container_of(work, struct io_kiocb, work);
2991 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2993 #if defined(CONFIG_NET)
2994 struct io_sr_msg *sr = &req->sr_msg;
2995 struct io_async_ctx *io = req->io;
2998 sr->msg_flags = READ_ONCE(sqe->msg_flags);
2999 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3000 sr->len = READ_ONCE(sqe->len);
3002 if (!io || req->opcode == IORING_OP_SEND)
3004 /* iovec is already imported */
3005 if (req->flags & REQ_F_NEED_CLEANUP)
3008 io->msg.iov = io->msg.fast_iov;
3009 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3012 req->flags |= REQ_F_NEED_CLEANUP;
3019 static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt,
3020 bool force_nonblock)
3022 #if defined(CONFIG_NET)
3023 struct io_async_msghdr *kmsg = NULL;
3024 struct socket *sock;
3027 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3030 sock = sock_from_file(req->file, &ret);
3032 struct io_async_ctx io;
3033 struct sockaddr_storage addr;
3037 kmsg = &req->io->msg;
3038 kmsg->msg.msg_name = &addr;
3039 /* if iov is set, it's allocated already */
3041 kmsg->iov = kmsg->fast_iov;
3042 kmsg->msg.msg_iter.iov = kmsg->iov;
3044 struct io_sr_msg *sr = &req->sr_msg;
3047 kmsg->msg.msg_name = &addr;
3049 io.msg.iov = io.msg.fast_iov;
3050 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3051 sr->msg_flags, &io.msg.iov);
3056 flags = req->sr_msg.msg_flags;
3057 if (flags & MSG_DONTWAIT)
3058 req->flags |= REQ_F_NOWAIT;
3059 else if (force_nonblock)
3060 flags |= MSG_DONTWAIT;
3062 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3063 if (force_nonblock && ret == -EAGAIN) {
3066 if (io_alloc_async_ctx(req)) {
3067 if (kmsg && kmsg->iov != kmsg->fast_iov)
3071 req->flags |= REQ_F_NEED_CLEANUP;
3072 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3075 if (ret == -ERESTARTSYS)
3079 if (kmsg && kmsg->iov != kmsg->fast_iov)
3081 req->flags &= ~REQ_F_NEED_CLEANUP;
3082 io_cqring_add_event(req, ret);
3084 req_set_fail_links(req);
3085 io_put_req_find_next(req, nxt);
3092 static int io_send(struct io_kiocb *req, struct io_kiocb **nxt,
3093 bool force_nonblock)
3095 #if defined(CONFIG_NET)
3096 struct socket *sock;
3099 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3102 sock = sock_from_file(req->file, &ret);
3104 struct io_sr_msg *sr = &req->sr_msg;
3109 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3114 msg.msg_name = NULL;
3115 msg.msg_control = NULL;
3116 msg.msg_controllen = 0;
3117 msg.msg_namelen = 0;
3119 flags = req->sr_msg.msg_flags;
3120 if (flags & MSG_DONTWAIT)
3121 req->flags |= REQ_F_NOWAIT;
3122 else if (force_nonblock)
3123 flags |= MSG_DONTWAIT;
3125 msg.msg_flags = flags;
3126 ret = sock_sendmsg(sock, &msg);
3127 if (force_nonblock && ret == -EAGAIN)
3129 if (ret == -ERESTARTSYS)
3133 io_cqring_add_event(req, ret);
3135 req_set_fail_links(req);
3136 io_put_req_find_next(req, nxt);
3143 static int io_recvmsg_prep(struct io_kiocb *req,
3144 const struct io_uring_sqe *sqe)
3146 #if defined(CONFIG_NET)
3147 struct io_sr_msg *sr = &req->sr_msg;
3148 struct io_async_ctx *io = req->io;
3151 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3152 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3153 sr->len = READ_ONCE(sqe->len);
3155 if (!io || req->opcode == IORING_OP_RECV)
3157 /* iovec is already imported */
3158 if (req->flags & REQ_F_NEED_CLEANUP)
3161 io->msg.iov = io->msg.fast_iov;
3162 ret = recvmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3163 &io->msg.uaddr, &io->msg.iov);
3165 req->flags |= REQ_F_NEED_CLEANUP;
3172 static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt,
3173 bool force_nonblock)
3175 #if defined(CONFIG_NET)
3176 struct io_async_msghdr *kmsg = NULL;
3177 struct socket *sock;
3180 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3183 sock = sock_from_file(req->file, &ret);
3185 struct io_async_ctx io;
3186 struct sockaddr_storage addr;
3190 kmsg = &req->io->msg;
3191 kmsg->msg.msg_name = &addr;
3192 /* if iov is set, it's allocated already */
3194 kmsg->iov = kmsg->fast_iov;
3195 kmsg->msg.msg_iter.iov = kmsg->iov;
3197 struct io_sr_msg *sr = &req->sr_msg;
3200 kmsg->msg.msg_name = &addr;
3202 io.msg.iov = io.msg.fast_iov;
3203 ret = recvmsg_copy_msghdr(&io.msg.msg, sr->msg,
3204 sr->msg_flags, &io.msg.uaddr,
3210 flags = req->sr_msg.msg_flags;
3211 if (flags & MSG_DONTWAIT)
3212 req->flags |= REQ_F_NOWAIT;
3213 else if (force_nonblock)
3214 flags |= MSG_DONTWAIT;
3216 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3217 kmsg->uaddr, flags);
3218 if (force_nonblock && ret == -EAGAIN) {
3221 if (io_alloc_async_ctx(req)) {
3222 if (kmsg && kmsg->iov != kmsg->fast_iov)
3226 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3227 req->flags |= REQ_F_NEED_CLEANUP;
3230 if (ret == -ERESTARTSYS)
3234 if (kmsg && kmsg->iov != kmsg->fast_iov)
3236 req->flags &= ~REQ_F_NEED_CLEANUP;
3237 io_cqring_add_event(req, ret);
3239 req_set_fail_links(req);
3240 io_put_req_find_next(req, nxt);
3247 static int io_recv(struct io_kiocb *req, struct io_kiocb **nxt,
3248 bool force_nonblock)
3250 #if defined(CONFIG_NET)
3251 struct socket *sock;
3254 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3257 sock = sock_from_file(req->file, &ret);
3259 struct io_sr_msg *sr = &req->sr_msg;
3264 ret = import_single_range(READ, sr->buf, sr->len, &iov,
3269 msg.msg_name = NULL;
3270 msg.msg_control = NULL;
3271 msg.msg_controllen = 0;
3272 msg.msg_namelen = 0;
3273 msg.msg_iocb = NULL;
3276 flags = req->sr_msg.msg_flags;
3277 if (flags & MSG_DONTWAIT)
3278 req->flags |= REQ_F_NOWAIT;
3279 else if (force_nonblock)
3280 flags |= MSG_DONTWAIT;
3282 ret = sock_recvmsg(sock, &msg, flags);
3283 if (force_nonblock && ret == -EAGAIN)
3285 if (ret == -ERESTARTSYS)
3289 io_cqring_add_event(req, ret);
3291 req_set_fail_links(req);
3292 io_put_req_find_next(req, nxt);
3300 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3302 #if defined(CONFIG_NET)
3303 struct io_accept *accept = &req->accept;
3305 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3307 if (sqe->ioprio || sqe->len || sqe->buf_index)
3310 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3311 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3312 accept->flags = READ_ONCE(sqe->accept_flags);
3319 #if defined(CONFIG_NET)
3320 static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3321 bool force_nonblock)
3323 struct io_accept *accept = &req->accept;
3324 unsigned file_flags;
3327 file_flags = force_nonblock ? O_NONBLOCK : 0;
3328 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3329 accept->addr_len, accept->flags);
3330 if (ret == -EAGAIN && force_nonblock)
3332 if (ret == -ERESTARTSYS)
3335 req_set_fail_links(req);
3336 io_cqring_add_event(req, ret);
3337 io_put_req_find_next(req, nxt);
3341 static void io_accept_finish(struct io_wq_work **workptr)
3343 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3344 struct io_kiocb *nxt = NULL;
3346 if (io_req_cancelled(req))
3348 __io_accept(req, &nxt, false);
3350 io_wq_assign_next(workptr, nxt);
3354 static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3355 bool force_nonblock)
3357 #if defined(CONFIG_NET)
3360 ret = __io_accept(req, nxt, force_nonblock);
3361 if (ret == -EAGAIN && force_nonblock) {
3362 req->work.func = io_accept_finish;
3372 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3374 #if defined(CONFIG_NET)
3375 struct io_connect *conn = &req->connect;
3376 struct io_async_ctx *io = req->io;
3378 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3380 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3383 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3384 conn->addr_len = READ_ONCE(sqe->addr2);
3389 return move_addr_to_kernel(conn->addr, conn->addr_len,
3390 &io->connect.address);
3396 static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt,
3397 bool force_nonblock)
3399 #if defined(CONFIG_NET)
3400 struct io_async_ctx __io, *io;
3401 unsigned file_flags;
3407 ret = move_addr_to_kernel(req->connect.addr,
3408 req->connect.addr_len,
3409 &__io.connect.address);
3415 file_flags = force_nonblock ? O_NONBLOCK : 0;
3417 ret = __sys_connect_file(req->file, &io->connect.address,
3418 req->connect.addr_len, file_flags);
3419 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
3422 if (io_alloc_async_ctx(req)) {
3426 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
3429 if (ret == -ERESTARTSYS)
3433 req_set_fail_links(req);
3434 io_cqring_add_event(req, ret);
3435 io_put_req_find_next(req, nxt);
3442 static void io_poll_remove_one(struct io_kiocb *req)
3444 struct io_poll_iocb *poll = &req->poll;
3446 spin_lock(&poll->head->lock);
3447 WRITE_ONCE(poll->canceled, true);
3448 if (!list_empty(&poll->wait.entry)) {
3449 list_del_init(&poll->wait.entry);
3450 io_queue_async_work(req);
3452 spin_unlock(&poll->head->lock);
3453 hash_del(&req->hash_node);
3456 static void io_poll_remove_all(struct io_ring_ctx *ctx)
3458 struct hlist_node *tmp;
3459 struct io_kiocb *req;
3462 spin_lock_irq(&ctx->completion_lock);
3463 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
3464 struct hlist_head *list;
3466 list = &ctx->cancel_hash[i];
3467 hlist_for_each_entry_safe(req, tmp, list, hash_node)
3468 io_poll_remove_one(req);
3470 spin_unlock_irq(&ctx->completion_lock);
3473 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
3475 struct hlist_head *list;
3476 struct io_kiocb *req;
3478 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
3479 hlist_for_each_entry(req, list, hash_node) {
3480 if (sqe_addr == req->user_data) {
3481 io_poll_remove_one(req);
3489 static int io_poll_remove_prep(struct io_kiocb *req,
3490 const struct io_uring_sqe *sqe)
3492 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3494 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3498 req->poll.addr = READ_ONCE(sqe->addr);
3503 * Find a running poll command that matches one specified in sqe->addr,
3504 * and remove it if found.
3506 static int io_poll_remove(struct io_kiocb *req)
3508 struct io_ring_ctx *ctx = req->ctx;
3512 addr = req->poll.addr;
3513 spin_lock_irq(&ctx->completion_lock);
3514 ret = io_poll_cancel(ctx, addr);
3515 spin_unlock_irq(&ctx->completion_lock);
3517 io_cqring_add_event(req, ret);
3519 req_set_fail_links(req);
3524 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
3526 struct io_ring_ctx *ctx = req->ctx;
3528 req->poll.done = true;
3530 io_cqring_fill_event(req, error);
3532 io_cqring_fill_event(req, mangle_poll(mask));
3533 io_commit_cqring(ctx);
3536 static void io_poll_complete_work(struct io_wq_work **workptr)
3538 struct io_wq_work *work = *workptr;
3539 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3540 struct io_poll_iocb *poll = &req->poll;
3541 struct poll_table_struct pt = { ._key = poll->events };
3542 struct io_ring_ctx *ctx = req->ctx;
3543 struct io_kiocb *nxt = NULL;
3547 if (work->flags & IO_WQ_WORK_CANCEL) {
3548 WRITE_ONCE(poll->canceled, true);
3550 } else if (READ_ONCE(poll->canceled)) {
3554 if (ret != -ECANCELED)
3555 mask = vfs_poll(poll->file, &pt) & poll->events;
3558 * Note that ->ki_cancel callers also delete iocb from active_reqs after
3559 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
3560 * synchronize with them. In the cancellation case the list_del_init
3561 * itself is not actually needed, but harmless so we keep it in to
3562 * avoid further branches in the fast path.
3564 spin_lock_irq(&ctx->completion_lock);
3565 if (!mask && ret != -ECANCELED) {
3566 add_wait_queue(poll->head, &poll->wait);
3567 spin_unlock_irq(&ctx->completion_lock);
3570 hash_del(&req->hash_node);
3571 io_poll_complete(req, mask, ret);
3572 spin_unlock_irq(&ctx->completion_lock);
3574 io_cqring_ev_posted(ctx);
3577 req_set_fail_links(req);
3578 io_put_req_find_next(req, &nxt);
3580 io_wq_assign_next(workptr, nxt);
3583 static void __io_poll_flush(struct io_ring_ctx *ctx, struct llist_node *nodes)
3585 struct io_kiocb *req, *tmp;
3586 struct req_batch rb;
3588 rb.to_free = rb.need_iter = 0;
3589 spin_lock_irq(&ctx->completion_lock);
3590 llist_for_each_entry_safe(req, tmp, nodes, llist_node) {
3591 hash_del(&req->hash_node);
3592 io_poll_complete(req, req->result, 0);
3594 if (refcount_dec_and_test(&req->refs) &&
3595 !io_req_multi_free(&rb, req)) {
3596 req->flags |= REQ_F_COMP_LOCKED;
3600 spin_unlock_irq(&ctx->completion_lock);
3602 io_cqring_ev_posted(ctx);
3603 io_free_req_many(ctx, &rb);
3606 static void io_poll_flush(struct io_wq_work **workptr)
3608 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3609 struct llist_node *nodes;
3611 nodes = llist_del_all(&req->ctx->poll_llist);
3613 __io_poll_flush(req->ctx, nodes);
3616 static void io_poll_trigger_evfd(struct io_wq_work **workptr)
3618 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3620 eventfd_signal(req->ctx->cq_ev_fd, 1);
3624 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
3627 struct io_poll_iocb *poll = wait->private;
3628 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
3629 struct io_ring_ctx *ctx = req->ctx;
3630 __poll_t mask = key_to_poll(key);
3632 /* for instances that support it check for an event match first: */
3633 if (mask && !(mask & poll->events))
3636 list_del_init(&poll->wait.entry);
3639 * Run completion inline if we can. We're using trylock here because
3640 * we are violating the completion_lock -> poll wq lock ordering.
3641 * If we have a link timeout we're going to need the completion_lock
3642 * for finalizing the request, mark us as having grabbed that already.
3645 unsigned long flags;
3647 if (llist_empty(&ctx->poll_llist) &&
3648 spin_trylock_irqsave(&ctx->completion_lock, flags)) {
3651 hash_del(&req->hash_node);
3652 io_poll_complete(req, mask, 0);
3654 trigger_ev = io_should_trigger_evfd(ctx);
3655 if (trigger_ev && eventfd_signal_count()) {
3657 req->work.func = io_poll_trigger_evfd;
3659 req->flags |= REQ_F_COMP_LOCKED;
3663 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3664 __io_cqring_ev_posted(ctx, trigger_ev);
3667 req->llist_node.next = NULL;
3668 /* if the list wasn't empty, we're done */
3669 if (!llist_add(&req->llist_node, &ctx->poll_llist))
3672 req->work.func = io_poll_flush;
3676 io_queue_async_work(req);
3681 struct io_poll_table {
3682 struct poll_table_struct pt;
3683 struct io_kiocb *req;
3687 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
3688 struct poll_table_struct *p)
3690 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
3692 if (unlikely(pt->req->poll.head)) {
3693 pt->error = -EINVAL;
3698 pt->req->poll.head = head;
3699 add_wait_queue(head, &pt->req->poll.wait);
3702 static void io_poll_req_insert(struct io_kiocb *req)
3704 struct io_ring_ctx *ctx = req->ctx;
3705 struct hlist_head *list;
3707 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
3708 hlist_add_head(&req->hash_node, list);
3711 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3713 struct io_poll_iocb *poll = &req->poll;
3716 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3718 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3723 events = READ_ONCE(sqe->poll_events);
3724 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
3728 static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt)
3730 struct io_poll_iocb *poll = &req->poll;
3731 struct io_ring_ctx *ctx = req->ctx;
3732 struct io_poll_table ipt;
3733 bool cancel = false;
3736 INIT_IO_WORK(&req->work, io_poll_complete_work);
3737 INIT_HLIST_NODE(&req->hash_node);
3741 poll->canceled = false;
3743 ipt.pt._qproc = io_poll_queue_proc;
3744 ipt.pt._key = poll->events;
3746 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
3748 /* initialized the list so that we can do list_empty checks */
3749 INIT_LIST_HEAD(&poll->wait.entry);
3750 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
3751 poll->wait.private = poll;
3753 INIT_LIST_HEAD(&req->list);
3755 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
3757 spin_lock_irq(&ctx->completion_lock);
3758 if (likely(poll->head)) {
3759 spin_lock(&poll->head->lock);
3760 if (unlikely(list_empty(&poll->wait.entry))) {
3766 if (mask || ipt.error)
3767 list_del_init(&poll->wait.entry);
3769 WRITE_ONCE(poll->canceled, true);
3770 else if (!poll->done) /* actually waiting for an event */
3771 io_poll_req_insert(req);
3772 spin_unlock(&poll->head->lock);
3774 if (mask) { /* no async, we'd stolen it */
3776 io_poll_complete(req, mask, 0);
3778 spin_unlock_irq(&ctx->completion_lock);
3781 io_cqring_ev_posted(ctx);
3782 io_put_req_find_next(req, nxt);
3787 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
3789 struct io_timeout_data *data = container_of(timer,
3790 struct io_timeout_data, timer);
3791 struct io_kiocb *req = data->req;
3792 struct io_ring_ctx *ctx = req->ctx;
3793 unsigned long flags;
3795 atomic_inc(&ctx->cq_timeouts);
3797 spin_lock_irqsave(&ctx->completion_lock, flags);
3799 * We could be racing with timeout deletion. If the list is empty,
3800 * then timeout lookup already found it and will be handling it.
3802 if (!list_empty(&req->list)) {
3803 struct io_kiocb *prev;
3806 * Adjust the reqs sequence before the current one because it
3807 * will consume a slot in the cq_ring and the cq_tail
3808 * pointer will be increased, otherwise other timeout reqs may
3809 * return in advance without waiting for enough wait_nr.
3812 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
3814 list_del_init(&req->list);
3817 io_cqring_fill_event(req, -ETIME);
3818 io_commit_cqring(ctx);
3819 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3821 io_cqring_ev_posted(ctx);
3822 req_set_fail_links(req);
3824 return HRTIMER_NORESTART;
3827 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
3829 struct io_kiocb *req;
3832 list_for_each_entry(req, &ctx->timeout_list, list) {
3833 if (user_data == req->user_data) {
3834 list_del_init(&req->list);
3843 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
3847 req_set_fail_links(req);
3848 io_cqring_fill_event(req, -ECANCELED);
3853 static int io_timeout_remove_prep(struct io_kiocb *req,
3854 const struct io_uring_sqe *sqe)
3856 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3858 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
3861 req->timeout.addr = READ_ONCE(sqe->addr);
3862 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
3863 if (req->timeout.flags)
3870 * Remove or update an existing timeout command
3872 static int io_timeout_remove(struct io_kiocb *req)
3874 struct io_ring_ctx *ctx = req->ctx;
3877 spin_lock_irq(&ctx->completion_lock);
3878 ret = io_timeout_cancel(ctx, req->timeout.addr);
3880 io_cqring_fill_event(req, ret);
3881 io_commit_cqring(ctx);
3882 spin_unlock_irq(&ctx->completion_lock);
3883 io_cqring_ev_posted(ctx);
3885 req_set_fail_links(req);
3890 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3891 bool is_timeout_link)
3893 struct io_timeout_data *data;
3896 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3898 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
3900 if (sqe->off && is_timeout_link)
3902 flags = READ_ONCE(sqe->timeout_flags);
3903 if (flags & ~IORING_TIMEOUT_ABS)
3906 req->timeout.count = READ_ONCE(sqe->off);
3908 if (!req->io && io_alloc_async_ctx(req))
3911 data = &req->io->timeout;
3913 req->flags |= REQ_F_TIMEOUT;
3915 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
3918 if (flags & IORING_TIMEOUT_ABS)
3919 data->mode = HRTIMER_MODE_ABS;
3921 data->mode = HRTIMER_MODE_REL;
3923 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
3927 static int io_timeout(struct io_kiocb *req)
3930 struct io_ring_ctx *ctx = req->ctx;
3931 struct io_timeout_data *data;
3932 struct list_head *entry;
3935 data = &req->io->timeout;
3938 * sqe->off holds how many events that need to occur for this
3939 * timeout event to be satisfied. If it isn't set, then this is
3940 * a pure timeout request, sequence isn't used.
3942 count = req->timeout.count;
3944 req->flags |= REQ_F_TIMEOUT_NOSEQ;
3945 spin_lock_irq(&ctx->completion_lock);
3946 entry = ctx->timeout_list.prev;
3950 req->sequence = ctx->cached_sq_head + count - 1;
3951 data->seq_offset = count;
3954 * Insertion sort, ensuring the first entry in the list is always
3955 * the one we need first.
3957 spin_lock_irq(&ctx->completion_lock);
3958 list_for_each_prev(entry, &ctx->timeout_list) {
3959 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
3960 unsigned nxt_sq_head;
3961 long long tmp, tmp_nxt;
3962 u32 nxt_offset = nxt->io->timeout.seq_offset;
3964 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
3968 * Since cached_sq_head + count - 1 can overflow, use type long
3971 tmp = (long long)ctx->cached_sq_head + count - 1;
3972 nxt_sq_head = nxt->sequence - nxt_offset + 1;
3973 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
3976 * cached_sq_head may overflow, and it will never overflow twice
3977 * once there is some timeout req still be valid.
3979 if (ctx->cached_sq_head < nxt_sq_head)
3986 * Sequence of reqs after the insert one and itself should
3987 * be adjusted because each timeout req consumes a slot.
3992 req->sequence -= span;
3994 list_add(&req->list, entry);
3995 data->timer.function = io_timeout_fn;
3996 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
3997 spin_unlock_irq(&ctx->completion_lock);
4001 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4003 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4005 return req->user_data == (unsigned long) data;
4008 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4010 enum io_wq_cancel cancel_ret;
4013 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4014 switch (cancel_ret) {
4015 case IO_WQ_CANCEL_OK:
4018 case IO_WQ_CANCEL_RUNNING:
4021 case IO_WQ_CANCEL_NOTFOUND:
4029 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4030 struct io_kiocb *req, __u64 sqe_addr,
4031 struct io_kiocb **nxt, int success_ret)
4033 unsigned long flags;
4036 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4037 if (ret != -ENOENT) {
4038 spin_lock_irqsave(&ctx->completion_lock, flags);
4042 spin_lock_irqsave(&ctx->completion_lock, flags);
4043 ret = io_timeout_cancel(ctx, sqe_addr);
4046 ret = io_poll_cancel(ctx, sqe_addr);
4050 io_cqring_fill_event(req, ret);
4051 io_commit_cqring(ctx);
4052 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4053 io_cqring_ev_posted(ctx);
4056 req_set_fail_links(req);
4057 io_put_req_find_next(req, nxt);
4060 static int io_async_cancel_prep(struct io_kiocb *req,
4061 const struct io_uring_sqe *sqe)
4063 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4065 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4069 req->cancel.addr = READ_ONCE(sqe->addr);
4073 static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt)
4075 struct io_ring_ctx *ctx = req->ctx;
4077 io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0);
4081 static int io_files_update_prep(struct io_kiocb *req,
4082 const struct io_uring_sqe *sqe)
4084 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4087 req->files_update.offset = READ_ONCE(sqe->off);
4088 req->files_update.nr_args = READ_ONCE(sqe->len);
4089 if (!req->files_update.nr_args)
4091 req->files_update.arg = READ_ONCE(sqe->addr);
4095 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4097 struct io_ring_ctx *ctx = req->ctx;
4098 struct io_uring_files_update up;
4104 up.offset = req->files_update.offset;
4105 up.fds = req->files_update.arg;
4107 mutex_lock(&ctx->uring_lock);
4108 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4109 mutex_unlock(&ctx->uring_lock);
4112 req_set_fail_links(req);
4113 io_cqring_add_event(req, ret);
4118 static int io_req_defer_prep(struct io_kiocb *req,
4119 const struct io_uring_sqe *sqe)
4123 if (io_op_defs[req->opcode].file_table) {
4124 ret = io_grab_files(req);
4129 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4131 switch (req->opcode) {
4134 case IORING_OP_READV:
4135 case IORING_OP_READ_FIXED:
4136 case IORING_OP_READ:
4137 ret = io_read_prep(req, sqe, true);
4139 case IORING_OP_WRITEV:
4140 case IORING_OP_WRITE_FIXED:
4141 case IORING_OP_WRITE:
4142 ret = io_write_prep(req, sqe, true);
4144 case IORING_OP_POLL_ADD:
4145 ret = io_poll_add_prep(req, sqe);
4147 case IORING_OP_POLL_REMOVE:
4148 ret = io_poll_remove_prep(req, sqe);
4150 case IORING_OP_FSYNC:
4151 ret = io_prep_fsync(req, sqe);
4153 case IORING_OP_SYNC_FILE_RANGE:
4154 ret = io_prep_sfr(req, sqe);
4156 case IORING_OP_SENDMSG:
4157 case IORING_OP_SEND:
4158 ret = io_sendmsg_prep(req, sqe);
4160 case IORING_OP_RECVMSG:
4161 case IORING_OP_RECV:
4162 ret = io_recvmsg_prep(req, sqe);
4164 case IORING_OP_CONNECT:
4165 ret = io_connect_prep(req, sqe);
4167 case IORING_OP_TIMEOUT:
4168 ret = io_timeout_prep(req, sqe, false);
4170 case IORING_OP_TIMEOUT_REMOVE:
4171 ret = io_timeout_remove_prep(req, sqe);
4173 case IORING_OP_ASYNC_CANCEL:
4174 ret = io_async_cancel_prep(req, sqe);
4176 case IORING_OP_LINK_TIMEOUT:
4177 ret = io_timeout_prep(req, sqe, true);
4179 case IORING_OP_ACCEPT:
4180 ret = io_accept_prep(req, sqe);
4182 case IORING_OP_FALLOCATE:
4183 ret = io_fallocate_prep(req, sqe);
4185 case IORING_OP_OPENAT:
4186 ret = io_openat_prep(req, sqe);
4188 case IORING_OP_CLOSE:
4189 ret = io_close_prep(req, sqe);
4191 case IORING_OP_FILES_UPDATE:
4192 ret = io_files_update_prep(req, sqe);
4194 case IORING_OP_STATX:
4195 ret = io_statx_prep(req, sqe);
4197 case IORING_OP_FADVISE:
4198 ret = io_fadvise_prep(req, sqe);
4200 case IORING_OP_MADVISE:
4201 ret = io_madvise_prep(req, sqe);
4203 case IORING_OP_OPENAT2:
4204 ret = io_openat2_prep(req, sqe);
4206 case IORING_OP_EPOLL_CTL:
4207 ret = io_epoll_ctl_prep(req, sqe);
4210 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
4219 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4221 struct io_ring_ctx *ctx = req->ctx;
4224 /* Still need defer if there is pending req in defer list. */
4225 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
4228 if (!req->io && io_alloc_async_ctx(req))
4231 ret = io_req_defer_prep(req, sqe);
4235 spin_lock_irq(&ctx->completion_lock);
4236 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
4237 spin_unlock_irq(&ctx->completion_lock);
4241 trace_io_uring_defer(ctx, req, req->user_data);
4242 list_add_tail(&req->list, &ctx->defer_list);
4243 spin_unlock_irq(&ctx->completion_lock);
4244 return -EIOCBQUEUED;
4247 static void io_cleanup_req(struct io_kiocb *req)
4249 struct io_async_ctx *io = req->io;
4251 switch (req->opcode) {
4252 case IORING_OP_READV:
4253 case IORING_OP_READ_FIXED:
4254 case IORING_OP_READ:
4255 case IORING_OP_WRITEV:
4256 case IORING_OP_WRITE_FIXED:
4257 case IORING_OP_WRITE:
4258 if (io->rw.iov != io->rw.fast_iov)
4261 case IORING_OP_SENDMSG:
4262 case IORING_OP_RECVMSG:
4263 if (io->msg.iov != io->msg.fast_iov)
4266 case IORING_OP_OPENAT:
4267 case IORING_OP_OPENAT2:
4268 case IORING_OP_STATX:
4269 putname(req->open.filename);
4273 req->flags &= ~REQ_F_NEED_CLEANUP;
4276 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4277 struct io_kiocb **nxt, bool force_nonblock)
4279 struct io_ring_ctx *ctx = req->ctx;
4282 switch (req->opcode) {
4286 case IORING_OP_READV:
4287 case IORING_OP_READ_FIXED:
4288 case IORING_OP_READ:
4290 ret = io_read_prep(req, sqe, force_nonblock);
4294 ret = io_read(req, nxt, force_nonblock);
4296 case IORING_OP_WRITEV:
4297 case IORING_OP_WRITE_FIXED:
4298 case IORING_OP_WRITE:
4300 ret = io_write_prep(req, sqe, force_nonblock);
4304 ret = io_write(req, nxt, force_nonblock);
4306 case IORING_OP_FSYNC:
4308 ret = io_prep_fsync(req, sqe);
4312 ret = io_fsync(req, nxt, force_nonblock);
4314 case IORING_OP_POLL_ADD:
4316 ret = io_poll_add_prep(req, sqe);
4320 ret = io_poll_add(req, nxt);
4322 case IORING_OP_POLL_REMOVE:
4324 ret = io_poll_remove_prep(req, sqe);
4328 ret = io_poll_remove(req);
4330 case IORING_OP_SYNC_FILE_RANGE:
4332 ret = io_prep_sfr(req, sqe);
4336 ret = io_sync_file_range(req, nxt, force_nonblock);
4338 case IORING_OP_SENDMSG:
4339 case IORING_OP_SEND:
4341 ret = io_sendmsg_prep(req, sqe);
4345 if (req->opcode == IORING_OP_SENDMSG)
4346 ret = io_sendmsg(req, nxt, force_nonblock);
4348 ret = io_send(req, nxt, force_nonblock);
4350 case IORING_OP_RECVMSG:
4351 case IORING_OP_RECV:
4353 ret = io_recvmsg_prep(req, sqe);
4357 if (req->opcode == IORING_OP_RECVMSG)
4358 ret = io_recvmsg(req, nxt, force_nonblock);
4360 ret = io_recv(req, nxt, force_nonblock);
4362 case IORING_OP_TIMEOUT:
4364 ret = io_timeout_prep(req, sqe, false);
4368 ret = io_timeout(req);
4370 case IORING_OP_TIMEOUT_REMOVE:
4372 ret = io_timeout_remove_prep(req, sqe);
4376 ret = io_timeout_remove(req);
4378 case IORING_OP_ACCEPT:
4380 ret = io_accept_prep(req, sqe);
4384 ret = io_accept(req, nxt, force_nonblock);
4386 case IORING_OP_CONNECT:
4388 ret = io_connect_prep(req, sqe);
4392 ret = io_connect(req, nxt, force_nonblock);
4394 case IORING_OP_ASYNC_CANCEL:
4396 ret = io_async_cancel_prep(req, sqe);
4400 ret = io_async_cancel(req, nxt);
4402 case IORING_OP_FALLOCATE:
4404 ret = io_fallocate_prep(req, sqe);
4408 ret = io_fallocate(req, nxt, force_nonblock);
4410 case IORING_OP_OPENAT:
4412 ret = io_openat_prep(req, sqe);
4416 ret = io_openat(req, nxt, force_nonblock);
4418 case IORING_OP_CLOSE:
4420 ret = io_close_prep(req, sqe);
4424 ret = io_close(req, nxt, force_nonblock);
4426 case IORING_OP_FILES_UPDATE:
4428 ret = io_files_update_prep(req, sqe);
4432 ret = io_files_update(req, force_nonblock);
4434 case IORING_OP_STATX:
4436 ret = io_statx_prep(req, sqe);
4440 ret = io_statx(req, nxt, force_nonblock);
4442 case IORING_OP_FADVISE:
4444 ret = io_fadvise_prep(req, sqe);
4448 ret = io_fadvise(req, nxt, force_nonblock);
4450 case IORING_OP_MADVISE:
4452 ret = io_madvise_prep(req, sqe);
4456 ret = io_madvise(req, nxt, force_nonblock);
4458 case IORING_OP_OPENAT2:
4460 ret = io_openat2_prep(req, sqe);
4464 ret = io_openat2(req, nxt, force_nonblock);
4466 case IORING_OP_EPOLL_CTL:
4468 ret = io_epoll_ctl_prep(req, sqe);
4472 ret = io_epoll_ctl(req, nxt, force_nonblock);
4482 if (ctx->flags & IORING_SETUP_IOPOLL) {
4483 const bool in_async = io_wq_current_is_worker();
4485 if (req->result == -EAGAIN)
4488 /* workqueue context doesn't hold uring_lock, grab it now */
4490 mutex_lock(&ctx->uring_lock);
4492 io_iopoll_req_issued(req);
4495 mutex_unlock(&ctx->uring_lock);
4501 static void io_wq_submit_work(struct io_wq_work **workptr)
4503 struct io_wq_work *work = *workptr;
4504 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4505 struct io_kiocb *nxt = NULL;
4508 /* if NO_CANCEL is set, we must still run the work */
4509 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
4510 IO_WQ_WORK_CANCEL) {
4515 req->in_async = true;
4517 ret = io_issue_sqe(req, NULL, &nxt, false);
4519 * We can get EAGAIN for polled IO even though we're
4520 * forcing a sync submission from here, since we can't
4521 * wait for request slots on the block side.
4529 /* drop submission reference */
4533 req_set_fail_links(req);
4534 io_cqring_add_event(req, ret);
4538 /* if a dependent link is ready, pass it back */
4540 io_wq_assign_next(workptr, nxt);
4543 static int io_req_needs_file(struct io_kiocb *req, int fd)
4545 if (!io_op_defs[req->opcode].needs_file)
4547 if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg)
4552 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
4555 struct fixed_file_table *table;
4557 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
4558 return table->files[index & IORING_FILE_TABLE_MASK];;
4561 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
4562 const struct io_uring_sqe *sqe)
4564 struct io_ring_ctx *ctx = req->ctx;
4568 flags = READ_ONCE(sqe->flags);
4569 fd = READ_ONCE(sqe->fd);
4571 if (!io_req_needs_file(req, fd))
4574 if (flags & IOSQE_FIXED_FILE) {
4575 if (unlikely(!ctx->file_data ||
4576 (unsigned) fd >= ctx->nr_user_files))
4578 fd = array_index_nospec(fd, ctx->nr_user_files);
4579 req->file = io_file_from_index(ctx, fd);
4582 req->flags |= REQ_F_FIXED_FILE;
4583 percpu_ref_get(&ctx->file_data->refs);
4585 if (req->needs_fixed_file)
4587 trace_io_uring_file_get(ctx, fd);
4588 req->file = io_file_get(state, fd);
4589 if (unlikely(!req->file))
4596 static int io_grab_files(struct io_kiocb *req)
4599 struct io_ring_ctx *ctx = req->ctx;
4601 if (req->work.files)
4603 if (!ctx->ring_file)
4607 spin_lock_irq(&ctx->inflight_lock);
4609 * We use the f_ops->flush() handler to ensure that we can flush
4610 * out work accessing these files if the fd is closed. Check if
4611 * the fd has changed since we started down this path, and disallow
4612 * this operation if it has.
4614 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
4615 list_add(&req->inflight_entry, &ctx->inflight_list);
4616 req->flags |= REQ_F_INFLIGHT;
4617 req->work.files = current->files;
4620 spin_unlock_irq(&ctx->inflight_lock);
4626 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
4628 struct io_timeout_data *data = container_of(timer,
4629 struct io_timeout_data, timer);
4630 struct io_kiocb *req = data->req;
4631 struct io_ring_ctx *ctx = req->ctx;
4632 struct io_kiocb *prev = NULL;
4633 unsigned long flags;
4635 spin_lock_irqsave(&ctx->completion_lock, flags);
4638 * We don't expect the list to be empty, that will only happen if we
4639 * race with the completion of the linked work.
4641 if (!list_empty(&req->link_list)) {
4642 prev = list_entry(req->link_list.prev, struct io_kiocb,
4644 if (refcount_inc_not_zero(&prev->refs)) {
4645 list_del_init(&req->link_list);
4646 prev->flags &= ~REQ_F_LINK_TIMEOUT;
4651 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4654 req_set_fail_links(prev);
4655 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
4659 io_cqring_add_event(req, -ETIME);
4662 return HRTIMER_NORESTART;
4665 static void io_queue_linked_timeout(struct io_kiocb *req)
4667 struct io_ring_ctx *ctx = req->ctx;
4670 * If the list is now empty, then our linked request finished before
4671 * we got a chance to setup the timer
4673 spin_lock_irq(&ctx->completion_lock);
4674 if (!list_empty(&req->link_list)) {
4675 struct io_timeout_data *data = &req->io->timeout;
4677 data->timer.function = io_link_timeout_fn;
4678 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
4681 spin_unlock_irq(&ctx->completion_lock);
4683 /* drop submission reference */
4687 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
4689 struct io_kiocb *nxt;
4691 if (!(req->flags & REQ_F_LINK))
4694 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
4696 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
4699 req->flags |= REQ_F_LINK_TIMEOUT;
4703 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4705 struct io_kiocb *linked_timeout;
4706 struct io_kiocb *nxt = NULL;
4710 linked_timeout = io_prep_linked_timeout(req);
4712 ret = io_issue_sqe(req, sqe, &nxt, true);
4715 * We async punt it if the file wasn't marked NOWAIT, or if the file
4716 * doesn't support non-blocking read/write attempts
4718 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
4719 (req->flags & REQ_F_MUST_PUNT))) {
4721 if (io_op_defs[req->opcode].file_table) {
4722 ret = io_grab_files(req);
4728 * Queued up for async execution, worker will release
4729 * submit reference when the iocb is actually submitted.
4731 io_queue_async_work(req);
4736 /* drop submission reference */
4739 if (linked_timeout) {
4741 io_queue_linked_timeout(linked_timeout);
4743 io_put_req(linked_timeout);
4746 /* and drop final reference, if we failed */
4748 io_cqring_add_event(req, ret);
4749 req_set_fail_links(req);
4757 if (req->flags & REQ_F_FORCE_ASYNC)
4763 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4767 ret = io_req_defer(req, sqe);
4769 if (ret != -EIOCBQUEUED) {
4771 io_cqring_add_event(req, ret);
4772 req_set_fail_links(req);
4773 io_double_put_req(req);
4775 } else if (req->flags & REQ_F_FORCE_ASYNC) {
4776 ret = io_req_defer_prep(req, sqe);
4777 if (unlikely(ret < 0))
4780 * Never try inline submit of IOSQE_ASYNC is set, go straight
4781 * to async execution.
4783 req->work.flags |= IO_WQ_WORK_CONCURRENT;
4784 io_queue_async_work(req);
4786 __io_queue_sqe(req, sqe);
4790 static inline void io_queue_link_head(struct io_kiocb *req)
4792 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
4793 io_cqring_add_event(req, -ECANCELED);
4794 io_double_put_req(req);
4796 io_queue_sqe(req, NULL);
4799 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
4800 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
4802 static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4803 struct io_submit_state *state, struct io_kiocb **link)
4805 const struct cred *old_creds = NULL;
4806 struct io_ring_ctx *ctx = req->ctx;
4807 unsigned int sqe_flags;
4810 sqe_flags = READ_ONCE(sqe->flags);
4812 /* enforce forwards compatibility on users */
4813 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
4818 id = READ_ONCE(sqe->personality);
4820 const struct cred *personality_creds;
4822 personality_creds = idr_find(&ctx->personality_idr, id);
4823 if (unlikely(!personality_creds)) {
4827 old_creds = override_creds(personality_creds);
4830 /* same numerical values with corresponding REQ_F_*, safe to copy */
4831 req->flags |= sqe_flags & (IOSQE_IO_DRAIN|IOSQE_IO_HARDLINK|
4834 ret = io_req_set_file(state, req, sqe);
4835 if (unlikely(ret)) {
4837 io_cqring_add_event(req, ret);
4838 io_double_put_req(req);
4840 revert_creds(old_creds);
4845 * If we already have a head request, queue this one for async
4846 * submittal once the head completes. If we don't have a head but
4847 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
4848 * submitted sync once the chain is complete. If none of those
4849 * conditions are true (normal request), then just queue it.
4852 struct io_kiocb *head = *link;
4855 * Taking sequential execution of a link, draining both sides
4856 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
4857 * requests in the link. So, it drains the head and the
4858 * next after the link request. The last one is done via
4859 * drain_next flag to persist the effect across calls.
4861 if (sqe_flags & IOSQE_IO_DRAIN) {
4862 head->flags |= REQ_F_IO_DRAIN;
4863 ctx->drain_next = 1;
4865 if (io_alloc_async_ctx(req)) {
4870 ret = io_req_defer_prep(req, sqe);
4872 /* fail even hard links since we don't submit */
4873 head->flags |= REQ_F_FAIL_LINK;
4876 trace_io_uring_link(ctx, req, head);
4877 list_add_tail(&req->link_list, &head->link_list);
4879 /* last request of a link, enqueue the link */
4880 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) {
4881 io_queue_link_head(head);
4885 if (unlikely(ctx->drain_next)) {
4886 req->flags |= REQ_F_IO_DRAIN;
4887 req->ctx->drain_next = 0;
4889 if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
4890 req->flags |= REQ_F_LINK;
4891 INIT_LIST_HEAD(&req->link_list);
4892 ret = io_req_defer_prep(req, sqe);
4894 req->flags |= REQ_F_FAIL_LINK;
4897 io_queue_sqe(req, sqe);
4902 revert_creds(old_creds);
4907 * Batched submission is done, ensure local IO is flushed out.
4909 static void io_submit_state_end(struct io_submit_state *state)
4911 blk_finish_plug(&state->plug);
4913 if (state->free_reqs)
4914 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
4918 * Start submission side cache.
4920 static void io_submit_state_start(struct io_submit_state *state,
4921 unsigned int max_ios)
4923 blk_start_plug(&state->plug);
4924 state->free_reqs = 0;
4926 state->ios_left = max_ios;
4929 static void io_commit_sqring(struct io_ring_ctx *ctx)
4931 struct io_rings *rings = ctx->rings;
4934 * Ensure any loads from the SQEs are done at this point,
4935 * since once we write the new head, the application could
4936 * write new data to them.
4938 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
4942 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
4943 * that is mapped by userspace. This means that care needs to be taken to
4944 * ensure that reads are stable, as we cannot rely on userspace always
4945 * being a good citizen. If members of the sqe are validated and then later
4946 * used, it's important that those reads are done through READ_ONCE() to
4947 * prevent a re-load down the line.
4949 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
4950 const struct io_uring_sqe **sqe_ptr)
4952 u32 *sq_array = ctx->sq_array;
4956 * The cached sq head (or cq tail) serves two purposes:
4958 * 1) allows us to batch the cost of updating the user visible
4960 * 2) allows the kernel side to track the head on its own, even
4961 * though the application is the one updating it.
4963 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
4964 if (likely(head < ctx->sq_entries)) {
4966 * All io need record the previous position, if LINK vs DARIN,
4967 * it can be used to mark the position of the first IO in the
4970 req->sequence = ctx->cached_sq_head;
4971 *sqe_ptr = &ctx->sq_sqes[head];
4972 req->opcode = READ_ONCE((*sqe_ptr)->opcode);
4973 req->user_data = READ_ONCE((*sqe_ptr)->user_data);
4974 ctx->cached_sq_head++;
4978 /* drop invalid entries */
4979 ctx->cached_sq_head++;
4980 ctx->cached_sq_dropped++;
4981 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
4985 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
4986 struct file *ring_file, int ring_fd,
4987 struct mm_struct **mm, bool async)
4989 struct io_submit_state state, *statep = NULL;
4990 struct io_kiocb *link = NULL;
4991 int i, submitted = 0;
4992 bool mm_fault = false;
4994 /* if we have a backlog and couldn't flush it all, return BUSY */
4995 if (test_bit(0, &ctx->sq_check_overflow)) {
4996 if (!list_empty(&ctx->cq_overflow_list) &&
4997 !io_cqring_overflow_flush(ctx, false))
5001 /* make sure SQ entry isn't read before tail */
5002 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5004 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5007 if (nr > IO_PLUG_THRESHOLD) {
5008 io_submit_state_start(&state, nr);
5012 ctx->ring_fd = ring_fd;
5013 ctx->ring_file = ring_file;
5015 for (i = 0; i < nr; i++) {
5016 const struct io_uring_sqe *sqe;
5017 struct io_kiocb *req;
5020 req = io_get_req(ctx, statep);
5021 if (unlikely(!req)) {
5023 submitted = -EAGAIN;
5026 if (!io_get_sqring(ctx, req, &sqe)) {
5027 __io_req_do_free(req);
5031 /* will complete beyond this point, count as submitted */
5034 if (unlikely(req->opcode >= IORING_OP_LAST)) {
5037 io_cqring_add_event(req, err);
5038 io_double_put_req(req);
5042 if (io_op_defs[req->opcode].needs_mm && !*mm) {
5043 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
5044 if (unlikely(mm_fault)) {
5048 use_mm(ctx->sqo_mm);
5052 req->in_async = async;
5053 req->needs_fixed_file = async;
5054 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5056 if (!io_submit_sqe(req, sqe, statep, &link))
5060 if (unlikely(submitted != nr)) {
5061 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5063 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5066 io_queue_link_head(link);
5068 io_submit_state_end(&state);
5070 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5071 io_commit_sqring(ctx);
5076 static int io_sq_thread(void *data)
5078 struct io_ring_ctx *ctx = data;
5079 struct mm_struct *cur_mm = NULL;
5080 const struct cred *old_cred;
5081 mm_segment_t old_fs;
5084 unsigned long timeout;
5087 complete(&ctx->completions[1]);
5091 old_cred = override_creds(ctx->creds);
5093 ret = timeout = inflight = 0;
5094 while (!kthread_should_park()) {
5095 unsigned int to_submit;
5098 unsigned nr_events = 0;
5100 if (ctx->flags & IORING_SETUP_IOPOLL) {
5102 * inflight is the count of the maximum possible
5103 * entries we submitted, but it can be smaller
5104 * if we dropped some of them. If we don't have
5105 * poll entries available, then we know that we
5106 * have nothing left to poll for. Reset the
5107 * inflight count to zero in that case.
5109 mutex_lock(&ctx->uring_lock);
5110 if (!list_empty(&ctx->poll_list))
5111 __io_iopoll_check(ctx, &nr_events, 0);
5114 mutex_unlock(&ctx->uring_lock);
5117 * Normal IO, just pretend everything completed.
5118 * We don't have to poll completions for that.
5120 nr_events = inflight;
5123 inflight -= nr_events;
5125 timeout = jiffies + ctx->sq_thread_idle;
5128 to_submit = io_sqring_entries(ctx);
5131 * If submit got -EBUSY, flag us as needing the application
5132 * to enter the kernel to reap and flush events.
5134 if (!to_submit || ret == -EBUSY) {
5136 * We're polling. If we're within the defined idle
5137 * period, then let us spin without work before going
5138 * to sleep. The exception is if we got EBUSY doing
5139 * more IO, we should wait for the application to
5140 * reap events and wake us up.
5143 (!time_after(jiffies, timeout) && ret != -EBUSY &&
5144 !percpu_ref_is_dying(&ctx->refs))) {
5150 * Drop cur_mm before scheduling, we can't hold it for
5151 * long periods (or over schedule()). Do this before
5152 * adding ourselves to the waitqueue, as the unuse/drop
5161 prepare_to_wait(&ctx->sqo_wait, &wait,
5162 TASK_INTERRUPTIBLE);
5164 /* Tell userspace we may need a wakeup call */
5165 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
5166 /* make sure to read SQ tail after writing flags */
5169 to_submit = io_sqring_entries(ctx);
5170 if (!to_submit || ret == -EBUSY) {
5171 if (kthread_should_park()) {
5172 finish_wait(&ctx->sqo_wait, &wait);
5175 if (signal_pending(current))
5176 flush_signals(current);
5178 finish_wait(&ctx->sqo_wait, &wait);
5180 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5183 finish_wait(&ctx->sqo_wait, &wait);
5185 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5188 mutex_lock(&ctx->uring_lock);
5189 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
5190 mutex_unlock(&ctx->uring_lock);
5200 revert_creds(old_cred);
5207 struct io_wait_queue {
5208 struct wait_queue_entry wq;
5209 struct io_ring_ctx *ctx;
5211 unsigned nr_timeouts;
5214 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
5216 struct io_ring_ctx *ctx = iowq->ctx;
5219 * Wake up if we have enough events, or if a timeout occurred since we
5220 * started waiting. For timeouts, we always want to return to userspace,
5221 * regardless of event count.
5223 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
5224 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
5227 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
5228 int wake_flags, void *key)
5230 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
5233 /* use noflush == true, as we can't safely rely on locking context */
5234 if (!io_should_wake(iowq, true))
5237 return autoremove_wake_function(curr, mode, wake_flags, key);
5241 * Wait until events become available, if we don't already have some. The
5242 * application must reap them itself, as they reside on the shared cq ring.
5244 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
5245 const sigset_t __user *sig, size_t sigsz)
5247 struct io_wait_queue iowq = {
5250 .func = io_wake_function,
5251 .entry = LIST_HEAD_INIT(iowq.wq.entry),
5254 .to_wait = min_events,
5256 struct io_rings *rings = ctx->rings;
5259 if (io_cqring_events(ctx, false) >= min_events)
5263 #ifdef CONFIG_COMPAT
5264 if (in_compat_syscall())
5265 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
5269 ret = set_user_sigmask(sig, sigsz);
5275 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
5276 trace_io_uring_cqring_wait(ctx, min_events);
5278 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
5279 TASK_INTERRUPTIBLE);
5280 if (io_should_wake(&iowq, false))
5283 if (signal_pending(current)) {
5288 finish_wait(&ctx->wait, &iowq.wq);
5290 restore_saved_sigmask_unless(ret == -EINTR);
5292 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
5295 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
5297 #if defined(CONFIG_UNIX)
5298 if (ctx->ring_sock) {
5299 struct sock *sock = ctx->ring_sock->sk;
5300 struct sk_buff *skb;
5302 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
5308 for (i = 0; i < ctx->nr_user_files; i++) {
5311 file = io_file_from_index(ctx, i);
5318 static void io_file_ref_kill(struct percpu_ref *ref)
5320 struct fixed_file_data *data;
5322 data = container_of(ref, struct fixed_file_data, refs);
5323 complete(&data->done);
5326 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
5328 struct fixed_file_data *data = ctx->file_data;
5329 unsigned nr_tables, i;
5334 percpu_ref_kill_and_confirm(&data->refs, io_file_ref_kill);
5335 flush_work(&data->ref_work);
5336 wait_for_completion(&data->done);
5337 io_ring_file_ref_flush(data);
5338 percpu_ref_exit(&data->refs);
5340 __io_sqe_files_unregister(ctx);
5341 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
5342 for (i = 0; i < nr_tables; i++)
5343 kfree(data->table[i].files);
5346 ctx->file_data = NULL;
5347 ctx->nr_user_files = 0;
5351 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
5353 if (ctx->sqo_thread) {
5354 wait_for_completion(&ctx->completions[1]);
5356 * The park is a bit of a work-around, without it we get
5357 * warning spews on shutdown with SQPOLL set and affinity
5358 * set to a single CPU.
5360 kthread_park(ctx->sqo_thread);
5361 kthread_stop(ctx->sqo_thread);
5362 ctx->sqo_thread = NULL;
5366 static void io_finish_async(struct io_ring_ctx *ctx)
5368 io_sq_thread_stop(ctx);
5371 io_wq_destroy(ctx->io_wq);
5376 #if defined(CONFIG_UNIX)
5378 * Ensure the UNIX gc is aware of our file set, so we are certain that
5379 * the io_uring can be safely unregistered on process exit, even if we have
5380 * loops in the file referencing.
5382 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
5384 struct sock *sk = ctx->ring_sock->sk;
5385 struct scm_fp_list *fpl;
5386 struct sk_buff *skb;
5389 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
5390 unsigned long inflight = ctx->user->unix_inflight + nr;
5392 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
5396 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
5400 skb = alloc_skb(0, GFP_KERNEL);
5409 fpl->user = get_uid(ctx->user);
5410 for (i = 0; i < nr; i++) {
5411 struct file *file = io_file_from_index(ctx, i + offset);
5415 fpl->fp[nr_files] = get_file(file);
5416 unix_inflight(fpl->user, fpl->fp[nr_files]);
5421 fpl->max = SCM_MAX_FD;
5422 fpl->count = nr_files;
5423 UNIXCB(skb).fp = fpl;
5424 skb->destructor = unix_destruct_scm;
5425 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
5426 skb_queue_head(&sk->sk_receive_queue, skb);
5428 for (i = 0; i < nr_files; i++)
5439 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
5440 * causes regular reference counting to break down. We rely on the UNIX
5441 * garbage collection to take care of this problem for us.
5443 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5445 unsigned left, total;
5449 left = ctx->nr_user_files;
5451 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
5453 ret = __io_sqe_files_scm(ctx, this_files, total);
5457 total += this_files;
5463 while (total < ctx->nr_user_files) {
5464 struct file *file = io_file_from_index(ctx, total);
5474 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5480 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
5485 for (i = 0; i < nr_tables; i++) {
5486 struct fixed_file_table *table = &ctx->file_data->table[i];
5487 unsigned this_files;
5489 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
5490 table->files = kcalloc(this_files, sizeof(struct file *),
5494 nr_files -= this_files;
5500 for (i = 0; i < nr_tables; i++) {
5501 struct fixed_file_table *table = &ctx->file_data->table[i];
5502 kfree(table->files);
5507 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
5509 #if defined(CONFIG_UNIX)
5510 struct sock *sock = ctx->ring_sock->sk;
5511 struct sk_buff_head list, *head = &sock->sk_receive_queue;
5512 struct sk_buff *skb;
5515 __skb_queue_head_init(&list);
5518 * Find the skb that holds this file in its SCM_RIGHTS. When found,
5519 * remove this entry and rearrange the file array.
5521 skb = skb_dequeue(head);
5523 struct scm_fp_list *fp;
5525 fp = UNIXCB(skb).fp;
5526 for (i = 0; i < fp->count; i++) {
5529 if (fp->fp[i] != file)
5532 unix_notinflight(fp->user, fp->fp[i]);
5533 left = fp->count - 1 - i;
5535 memmove(&fp->fp[i], &fp->fp[i + 1],
5536 left * sizeof(struct file *));
5543 __skb_queue_tail(&list, skb);
5553 __skb_queue_tail(&list, skb);
5555 skb = skb_dequeue(head);
5558 if (skb_peek(&list)) {
5559 spin_lock_irq(&head->lock);
5560 while ((skb = __skb_dequeue(&list)) != NULL)
5561 __skb_queue_tail(head, skb);
5562 spin_unlock_irq(&head->lock);
5569 struct io_file_put {
5570 struct llist_node llist;
5572 struct completion *done;
5575 static void io_ring_file_ref_flush(struct fixed_file_data *data)
5577 struct io_file_put *pfile, *tmp;
5578 struct llist_node *node;
5580 while ((node = llist_del_all(&data->put_llist)) != NULL) {
5581 llist_for_each_entry_safe(pfile, tmp, node, llist) {
5582 io_ring_file_put(data->ctx, pfile->file);
5584 complete(pfile->done);
5591 static void io_ring_file_ref_switch(struct work_struct *work)
5593 struct fixed_file_data *data;
5595 data = container_of(work, struct fixed_file_data, ref_work);
5596 io_ring_file_ref_flush(data);
5597 percpu_ref_get(&data->refs);
5598 percpu_ref_switch_to_percpu(&data->refs);
5601 static void io_file_data_ref_zero(struct percpu_ref *ref)
5603 struct fixed_file_data *data;
5605 data = container_of(ref, struct fixed_file_data, refs);
5608 * We can't safely switch from inside this context, punt to wq. If
5609 * the table ref is going away, the table is being unregistered.
5610 * Don't queue up the async work for that case, the caller will
5613 if (!percpu_ref_is_dying(&data->refs))
5614 queue_work(system_wq, &data->ref_work);
5617 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
5620 __s32 __user *fds = (__s32 __user *) arg;
5630 if (nr_args > IORING_MAX_FIXED_FILES)
5633 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
5634 if (!ctx->file_data)
5636 ctx->file_data->ctx = ctx;
5637 init_completion(&ctx->file_data->done);
5639 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
5640 ctx->file_data->table = kcalloc(nr_tables,
5641 sizeof(struct fixed_file_table),
5643 if (!ctx->file_data->table) {
5644 kfree(ctx->file_data);
5645 ctx->file_data = NULL;
5649 if (percpu_ref_init(&ctx->file_data->refs, io_file_data_ref_zero,
5650 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
5651 kfree(ctx->file_data->table);
5652 kfree(ctx->file_data);
5653 ctx->file_data = NULL;
5656 ctx->file_data->put_llist.first = NULL;
5657 INIT_WORK(&ctx->file_data->ref_work, io_ring_file_ref_switch);
5659 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
5660 percpu_ref_exit(&ctx->file_data->refs);
5661 kfree(ctx->file_data->table);
5662 kfree(ctx->file_data);
5663 ctx->file_data = NULL;
5667 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
5668 struct fixed_file_table *table;
5672 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
5674 /* allow sparse sets */
5680 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5681 index = i & IORING_FILE_TABLE_MASK;
5689 * Don't allow io_uring instances to be registered. If UNIX
5690 * isn't enabled, then this causes a reference cycle and this
5691 * instance can never get freed. If UNIX is enabled we'll
5692 * handle it just fine, but there's still no point in allowing
5693 * a ring fd as it doesn't support regular read/write anyway.
5695 if (file->f_op == &io_uring_fops) {
5700 table->files[index] = file;
5704 for (i = 0; i < ctx->nr_user_files; i++) {
5705 file = io_file_from_index(ctx, i);
5709 for (i = 0; i < nr_tables; i++)
5710 kfree(ctx->file_data->table[i].files);
5712 kfree(ctx->file_data->table);
5713 kfree(ctx->file_data);
5714 ctx->file_data = NULL;
5715 ctx->nr_user_files = 0;
5719 ret = io_sqe_files_scm(ctx);
5721 io_sqe_files_unregister(ctx);
5726 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
5729 #if defined(CONFIG_UNIX)
5730 struct sock *sock = ctx->ring_sock->sk;
5731 struct sk_buff_head *head = &sock->sk_receive_queue;
5732 struct sk_buff *skb;
5735 * See if we can merge this file into an existing skb SCM_RIGHTS
5736 * file set. If there's no room, fall back to allocating a new skb
5737 * and filling it in.
5739 spin_lock_irq(&head->lock);
5740 skb = skb_peek(head);
5742 struct scm_fp_list *fpl = UNIXCB(skb).fp;
5744 if (fpl->count < SCM_MAX_FD) {
5745 __skb_unlink(skb, head);
5746 spin_unlock_irq(&head->lock);
5747 fpl->fp[fpl->count] = get_file(file);
5748 unix_inflight(fpl->user, fpl->fp[fpl->count]);
5750 spin_lock_irq(&head->lock);
5751 __skb_queue_head(head, skb);
5756 spin_unlock_irq(&head->lock);
5763 return __io_sqe_files_scm(ctx, 1, index);
5769 static void io_atomic_switch(struct percpu_ref *ref)
5771 struct fixed_file_data *data;
5773 data = container_of(ref, struct fixed_file_data, refs);
5774 clear_bit(FFD_F_ATOMIC, &data->state);
5777 static bool io_queue_file_removal(struct fixed_file_data *data,
5780 struct io_file_put *pfile, pfile_stack;
5781 DECLARE_COMPLETION_ONSTACK(done);
5784 * If we fail allocating the struct we need for doing async reomval
5785 * of this file, just punt to sync and wait for it.
5787 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
5789 pfile = &pfile_stack;
5790 pfile->done = &done;
5794 llist_add(&pfile->llist, &data->put_llist);
5796 if (pfile == &pfile_stack) {
5797 if (!test_and_set_bit(FFD_F_ATOMIC, &data->state)) {
5798 percpu_ref_put(&data->refs);
5799 percpu_ref_switch_to_atomic(&data->refs,
5802 wait_for_completion(&done);
5803 flush_work(&data->ref_work);
5810 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
5811 struct io_uring_files_update *up,
5814 struct fixed_file_data *data = ctx->file_data;
5815 bool ref_switch = false;
5821 if (check_add_overflow(up->offset, nr_args, &done))
5823 if (done > ctx->nr_user_files)
5827 fds = u64_to_user_ptr(up->fds);
5829 struct fixed_file_table *table;
5833 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
5837 i = array_index_nospec(up->offset, ctx->nr_user_files);
5838 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5839 index = i & IORING_FILE_TABLE_MASK;
5840 if (table->files[index]) {
5841 file = io_file_from_index(ctx, index);
5842 table->files[index] = NULL;
5843 if (io_queue_file_removal(data, file))
5853 * Don't allow io_uring instances to be registered. If
5854 * UNIX isn't enabled, then this causes a reference
5855 * cycle and this instance can never get freed. If UNIX
5856 * is enabled we'll handle it just fine, but there's
5857 * still no point in allowing a ring fd as it doesn't
5858 * support regular read/write anyway.
5860 if (file->f_op == &io_uring_fops) {
5865 table->files[index] = file;
5866 err = io_sqe_file_register(ctx, file, i);
5875 if (ref_switch && !test_and_set_bit(FFD_F_ATOMIC, &data->state)) {
5876 percpu_ref_put(&data->refs);
5877 percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch);
5880 return done ? done : err;
5882 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
5885 struct io_uring_files_update up;
5887 if (!ctx->file_data)
5891 if (copy_from_user(&up, arg, sizeof(up)))
5896 return __io_sqe_files_update(ctx, &up, nr_args);
5899 static void io_put_work(struct io_wq_work *work)
5901 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5906 static void io_get_work(struct io_wq_work *work)
5908 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5910 refcount_inc(&req->refs);
5913 static int io_init_wq_offload(struct io_ring_ctx *ctx,
5914 struct io_uring_params *p)
5916 struct io_wq_data data;
5918 struct io_ring_ctx *ctx_attach;
5919 unsigned int concurrency;
5922 data.user = ctx->user;
5923 data.get_work = io_get_work;
5924 data.put_work = io_put_work;
5926 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
5927 /* Do QD, or 4 * CPUS, whatever is smallest */
5928 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
5930 ctx->io_wq = io_wq_create(concurrency, &data);
5931 if (IS_ERR(ctx->io_wq)) {
5932 ret = PTR_ERR(ctx->io_wq);
5938 f = fdget(p->wq_fd);
5942 if (f.file->f_op != &io_uring_fops) {
5947 ctx_attach = f.file->private_data;
5948 /* @io_wq is protected by holding the fd */
5949 if (!io_wq_get(ctx_attach->io_wq, &data)) {
5954 ctx->io_wq = ctx_attach->io_wq;
5960 static int io_sq_offload_start(struct io_ring_ctx *ctx,
5961 struct io_uring_params *p)
5965 init_waitqueue_head(&ctx->sqo_wait);
5966 mmgrab(current->mm);
5967 ctx->sqo_mm = current->mm;
5969 if (ctx->flags & IORING_SETUP_SQPOLL) {
5971 if (!capable(CAP_SYS_ADMIN))
5974 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
5975 if (!ctx->sq_thread_idle)
5976 ctx->sq_thread_idle = HZ;
5978 if (p->flags & IORING_SETUP_SQ_AFF) {
5979 int cpu = p->sq_thread_cpu;
5982 if (cpu >= nr_cpu_ids)
5984 if (!cpu_online(cpu))
5987 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
5991 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
5994 if (IS_ERR(ctx->sqo_thread)) {
5995 ret = PTR_ERR(ctx->sqo_thread);
5996 ctx->sqo_thread = NULL;
5999 wake_up_process(ctx->sqo_thread);
6000 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6001 /* Can't have SQ_AFF without SQPOLL */
6006 ret = io_init_wq_offload(ctx, p);
6012 io_finish_async(ctx);
6013 mmdrop(ctx->sqo_mm);
6018 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6020 atomic_long_sub(nr_pages, &user->locked_vm);
6023 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6025 unsigned long page_limit, cur_pages, new_pages;
6027 /* Don't allow more pages than we can safely lock */
6028 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6031 cur_pages = atomic_long_read(&user->locked_vm);
6032 new_pages = cur_pages + nr_pages;
6033 if (new_pages > page_limit)
6035 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6036 new_pages) != cur_pages);
6041 static void io_mem_free(void *ptr)
6048 page = virt_to_head_page(ptr);
6049 if (put_page_testzero(page))
6050 free_compound_page(page);
6053 static void *io_mem_alloc(size_t size)
6055 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6058 return (void *) __get_free_pages(gfp_flags, get_order(size));
6061 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6064 struct io_rings *rings;
6065 size_t off, sq_array_size;
6067 off = struct_size(rings, cqes, cq_entries);
6068 if (off == SIZE_MAX)
6072 off = ALIGN(off, SMP_CACHE_BYTES);
6077 sq_array_size = array_size(sizeof(u32), sq_entries);
6078 if (sq_array_size == SIZE_MAX)
6081 if (check_add_overflow(off, sq_array_size, &off))
6090 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6094 pages = (size_t)1 << get_order(
6095 rings_size(sq_entries, cq_entries, NULL));
6096 pages += (size_t)1 << get_order(
6097 array_size(sizeof(struct io_uring_sqe), sq_entries));
6102 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6106 if (!ctx->user_bufs)
6109 for (i = 0; i < ctx->nr_user_bufs; i++) {
6110 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6112 for (j = 0; j < imu->nr_bvecs; j++)
6113 unpin_user_page(imu->bvec[j].bv_page);
6115 if (ctx->account_mem)
6116 io_unaccount_mem(ctx->user, imu->nr_bvecs);
6121 kfree(ctx->user_bufs);
6122 ctx->user_bufs = NULL;
6123 ctx->nr_user_bufs = 0;
6127 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
6128 void __user *arg, unsigned index)
6130 struct iovec __user *src;
6132 #ifdef CONFIG_COMPAT
6134 struct compat_iovec __user *ciovs;
6135 struct compat_iovec ciov;
6137 ciovs = (struct compat_iovec __user *) arg;
6138 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
6141 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
6142 dst->iov_len = ciov.iov_len;
6146 src = (struct iovec __user *) arg;
6147 if (copy_from_user(dst, &src[index], sizeof(*dst)))
6152 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
6155 struct vm_area_struct **vmas = NULL;
6156 struct page **pages = NULL;
6157 int i, j, got_pages = 0;
6162 if (!nr_args || nr_args > UIO_MAXIOV)
6165 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
6167 if (!ctx->user_bufs)
6170 for (i = 0; i < nr_args; i++) {
6171 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6172 unsigned long off, start, end, ubuf;
6177 ret = io_copy_iov(ctx, &iov, arg, i);
6182 * Don't impose further limits on the size and buffer
6183 * constraints here, we'll -EINVAL later when IO is
6184 * submitted if they are wrong.
6187 if (!iov.iov_base || !iov.iov_len)
6190 /* arbitrary limit, but we need something */
6191 if (iov.iov_len > SZ_1G)
6194 ubuf = (unsigned long) iov.iov_base;
6195 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
6196 start = ubuf >> PAGE_SHIFT;
6197 nr_pages = end - start;
6199 if (ctx->account_mem) {
6200 ret = io_account_mem(ctx->user, nr_pages);
6206 if (!pages || nr_pages > got_pages) {
6209 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
6211 vmas = kvmalloc_array(nr_pages,
6212 sizeof(struct vm_area_struct *),
6214 if (!pages || !vmas) {
6216 if (ctx->account_mem)
6217 io_unaccount_mem(ctx->user, nr_pages);
6220 got_pages = nr_pages;
6223 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
6227 if (ctx->account_mem)
6228 io_unaccount_mem(ctx->user, nr_pages);
6233 down_read(¤t->mm->mmap_sem);
6234 pret = pin_user_pages(ubuf, nr_pages,
6235 FOLL_WRITE | FOLL_LONGTERM,
6237 if (pret == nr_pages) {
6238 /* don't support file backed memory */
6239 for (j = 0; j < nr_pages; j++) {
6240 struct vm_area_struct *vma = vmas[j];
6243 !is_file_hugepages(vma->vm_file)) {
6249 ret = pret < 0 ? pret : -EFAULT;
6251 up_read(¤t->mm->mmap_sem);
6254 * if we did partial map, or found file backed vmas,
6255 * release any pages we did get
6258 unpin_user_pages(pages, pret);
6259 if (ctx->account_mem)
6260 io_unaccount_mem(ctx->user, nr_pages);
6265 off = ubuf & ~PAGE_MASK;
6267 for (j = 0; j < nr_pages; j++) {
6270 vec_len = min_t(size_t, size, PAGE_SIZE - off);
6271 imu->bvec[j].bv_page = pages[j];
6272 imu->bvec[j].bv_len = vec_len;
6273 imu->bvec[j].bv_offset = off;
6277 /* store original address for later verification */
6279 imu->len = iov.iov_len;
6280 imu->nr_bvecs = nr_pages;
6282 ctx->nr_user_bufs++;
6290 io_sqe_buffer_unregister(ctx);
6294 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
6296 __s32 __user *fds = arg;
6302 if (copy_from_user(&fd, fds, sizeof(*fds)))
6305 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
6306 if (IS_ERR(ctx->cq_ev_fd)) {
6307 int ret = PTR_ERR(ctx->cq_ev_fd);
6308 ctx->cq_ev_fd = NULL;
6315 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
6317 if (ctx->cq_ev_fd) {
6318 eventfd_ctx_put(ctx->cq_ev_fd);
6319 ctx->cq_ev_fd = NULL;
6326 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
6328 io_finish_async(ctx);
6330 mmdrop(ctx->sqo_mm);
6332 io_iopoll_reap_events(ctx);
6333 io_sqe_buffer_unregister(ctx);
6334 io_sqe_files_unregister(ctx);
6335 io_eventfd_unregister(ctx);
6337 #if defined(CONFIG_UNIX)
6338 if (ctx->ring_sock) {
6339 ctx->ring_sock->file = NULL; /* so that iput() is called */
6340 sock_release(ctx->ring_sock);
6344 io_mem_free(ctx->rings);
6345 io_mem_free(ctx->sq_sqes);
6347 percpu_ref_exit(&ctx->refs);
6348 if (ctx->account_mem)
6349 io_unaccount_mem(ctx->user,
6350 ring_pages(ctx->sq_entries, ctx->cq_entries));
6351 free_uid(ctx->user);
6352 put_cred(ctx->creds);
6353 kfree(ctx->completions);
6354 kfree(ctx->cancel_hash);
6355 kmem_cache_free(req_cachep, ctx->fallback_req);
6359 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
6361 struct io_ring_ctx *ctx = file->private_data;
6364 poll_wait(file, &ctx->cq_wait, wait);
6366 * synchronizes with barrier from wq_has_sleeper call in
6370 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
6371 ctx->rings->sq_ring_entries)
6372 mask |= EPOLLOUT | EPOLLWRNORM;
6373 if (io_cqring_events(ctx, false))
6374 mask |= EPOLLIN | EPOLLRDNORM;
6379 static int io_uring_fasync(int fd, struct file *file, int on)
6381 struct io_ring_ctx *ctx = file->private_data;
6383 return fasync_helper(fd, file, on, &ctx->cq_fasync);
6386 static int io_remove_personalities(int id, void *p, void *data)
6388 struct io_ring_ctx *ctx = data;
6389 const struct cred *cred;
6391 cred = idr_remove(&ctx->personality_idr, id);
6397 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
6399 mutex_lock(&ctx->uring_lock);
6400 percpu_ref_kill(&ctx->refs);
6401 mutex_unlock(&ctx->uring_lock);
6404 * Wait for sq thread to idle, if we have one. It won't spin on new
6405 * work after we've killed the ctx ref above. This is important to do
6406 * before we cancel existing commands, as the thread could otherwise
6407 * be queueing new work post that. If that's work we need to cancel,
6408 * it could cause shutdown to hang.
6410 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
6413 io_kill_timeouts(ctx);
6414 io_poll_remove_all(ctx);
6417 io_wq_cancel_all(ctx->io_wq);
6419 io_iopoll_reap_events(ctx);
6420 /* if we failed setting up the ctx, we might not have any rings */
6422 io_cqring_overflow_flush(ctx, true);
6423 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
6424 wait_for_completion(&ctx->completions[0]);
6425 io_ring_ctx_free(ctx);
6428 static int io_uring_release(struct inode *inode, struct file *file)
6430 struct io_ring_ctx *ctx = file->private_data;
6432 file->private_data = NULL;
6433 io_ring_ctx_wait_and_kill(ctx);
6437 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
6438 struct files_struct *files)
6440 struct io_kiocb *req;
6443 while (!list_empty_careful(&ctx->inflight_list)) {
6444 struct io_kiocb *cancel_req = NULL;
6446 spin_lock_irq(&ctx->inflight_lock);
6447 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
6448 if (req->work.files != files)
6450 /* req is being completed, ignore */
6451 if (!refcount_inc_not_zero(&req->refs))
6457 prepare_to_wait(&ctx->inflight_wait, &wait,
6458 TASK_UNINTERRUPTIBLE);
6459 spin_unlock_irq(&ctx->inflight_lock);
6461 /* We need to keep going until we don't find a matching req */
6465 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
6466 io_put_req(cancel_req);
6469 finish_wait(&ctx->inflight_wait, &wait);
6472 static int io_uring_flush(struct file *file, void *data)
6474 struct io_ring_ctx *ctx = file->private_data;
6476 io_uring_cancel_files(ctx, data);
6480 static void *io_uring_validate_mmap_request(struct file *file,
6481 loff_t pgoff, size_t sz)
6483 struct io_ring_ctx *ctx = file->private_data;
6484 loff_t offset = pgoff << PAGE_SHIFT;
6489 case IORING_OFF_SQ_RING:
6490 case IORING_OFF_CQ_RING:
6493 case IORING_OFF_SQES:
6497 return ERR_PTR(-EINVAL);
6500 page = virt_to_head_page(ptr);
6501 if (sz > page_size(page))
6502 return ERR_PTR(-EINVAL);
6509 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6511 size_t sz = vma->vm_end - vma->vm_start;
6515 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
6517 return PTR_ERR(ptr);
6519 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
6520 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
6523 #else /* !CONFIG_MMU */
6525 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6527 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
6530 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
6532 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
6535 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
6536 unsigned long addr, unsigned long len,
6537 unsigned long pgoff, unsigned long flags)
6541 ptr = io_uring_validate_mmap_request(file, pgoff, len);
6543 return PTR_ERR(ptr);
6545 return (unsigned long) ptr;
6548 #endif /* !CONFIG_MMU */
6550 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
6551 u32, min_complete, u32, flags, const sigset_t __user *, sig,
6554 struct io_ring_ctx *ctx;
6559 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
6567 if (f.file->f_op != &io_uring_fops)
6571 ctx = f.file->private_data;
6572 if (!percpu_ref_tryget(&ctx->refs))
6576 * For SQ polling, the thread will do all submissions and completions.
6577 * Just return the requested submit count, and wake the thread if
6581 if (ctx->flags & IORING_SETUP_SQPOLL) {
6582 if (!list_empty_careful(&ctx->cq_overflow_list))
6583 io_cqring_overflow_flush(ctx, false);
6584 if (flags & IORING_ENTER_SQ_WAKEUP)
6585 wake_up(&ctx->sqo_wait);
6586 submitted = to_submit;
6587 } else if (to_submit) {
6588 struct mm_struct *cur_mm;
6590 mutex_lock(&ctx->uring_lock);
6591 /* already have mm, so io_submit_sqes() won't try to grab it */
6592 cur_mm = ctx->sqo_mm;
6593 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
6595 mutex_unlock(&ctx->uring_lock);
6597 if (submitted != to_submit)
6600 if (flags & IORING_ENTER_GETEVENTS) {
6601 unsigned nr_events = 0;
6603 min_complete = min(min_complete, ctx->cq_entries);
6605 if (ctx->flags & IORING_SETUP_IOPOLL) {
6606 ret = io_iopoll_check(ctx, &nr_events, min_complete);
6608 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
6613 percpu_ref_put(&ctx->refs);
6616 return submitted ? submitted : ret;
6619 static int io_uring_show_cred(int id, void *p, void *data)
6621 const struct cred *cred = p;
6622 struct seq_file *m = data;
6623 struct user_namespace *uns = seq_user_ns(m);
6624 struct group_info *gi;
6629 seq_printf(m, "%5d\n", id);
6630 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
6631 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
6632 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
6633 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
6634 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
6635 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
6636 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
6637 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
6638 seq_puts(m, "\n\tGroups:\t");
6639 gi = cred->group_info;
6640 for (g = 0; g < gi->ngroups; g++) {
6641 seq_put_decimal_ull(m, g ? " " : "",
6642 from_kgid_munged(uns, gi->gid[g]));
6644 seq_puts(m, "\n\tCapEff:\t");
6645 cap = cred->cap_effective;
6646 CAP_FOR_EACH_U32(__capi)
6647 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
6652 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
6656 mutex_lock(&ctx->uring_lock);
6657 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
6658 for (i = 0; i < ctx->nr_user_files; i++) {
6659 struct fixed_file_table *table;
6662 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6663 f = table->files[i & IORING_FILE_TABLE_MASK];
6665 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
6667 seq_printf(m, "%5u: <none>\n", i);
6669 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
6670 for (i = 0; i < ctx->nr_user_bufs; i++) {
6671 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
6673 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
6674 (unsigned int) buf->len);
6676 if (!idr_is_empty(&ctx->personality_idr)) {
6677 seq_printf(m, "Personalities:\n");
6678 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
6680 mutex_unlock(&ctx->uring_lock);
6683 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
6685 struct io_ring_ctx *ctx = f->private_data;
6687 if (percpu_ref_tryget(&ctx->refs)) {
6688 __io_uring_show_fdinfo(ctx, m);
6689 percpu_ref_put(&ctx->refs);
6693 static const struct file_operations io_uring_fops = {
6694 .release = io_uring_release,
6695 .flush = io_uring_flush,
6696 .mmap = io_uring_mmap,
6698 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
6699 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
6701 .poll = io_uring_poll,
6702 .fasync = io_uring_fasync,
6703 .show_fdinfo = io_uring_show_fdinfo,
6706 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
6707 struct io_uring_params *p)
6709 struct io_rings *rings;
6710 size_t size, sq_array_offset;
6712 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
6713 if (size == SIZE_MAX)
6716 rings = io_mem_alloc(size);
6721 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
6722 rings->sq_ring_mask = p->sq_entries - 1;
6723 rings->cq_ring_mask = p->cq_entries - 1;
6724 rings->sq_ring_entries = p->sq_entries;
6725 rings->cq_ring_entries = p->cq_entries;
6726 ctx->sq_mask = rings->sq_ring_mask;
6727 ctx->cq_mask = rings->cq_ring_mask;
6728 ctx->sq_entries = rings->sq_ring_entries;
6729 ctx->cq_entries = rings->cq_ring_entries;
6731 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
6732 if (size == SIZE_MAX) {
6733 io_mem_free(ctx->rings);
6738 ctx->sq_sqes = io_mem_alloc(size);
6739 if (!ctx->sq_sqes) {
6740 io_mem_free(ctx->rings);
6749 * Allocate an anonymous fd, this is what constitutes the application
6750 * visible backing of an io_uring instance. The application mmaps this
6751 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
6752 * we have to tie this fd to a socket for file garbage collection purposes.
6754 static int io_uring_get_fd(struct io_ring_ctx *ctx)
6759 #if defined(CONFIG_UNIX)
6760 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
6766 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
6770 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
6771 O_RDWR | O_CLOEXEC);
6774 ret = PTR_ERR(file);
6778 #if defined(CONFIG_UNIX)
6779 ctx->ring_sock->file = file;
6781 fd_install(ret, file);
6784 #if defined(CONFIG_UNIX)
6785 sock_release(ctx->ring_sock);
6786 ctx->ring_sock = NULL;
6791 static int io_uring_create(unsigned entries, struct io_uring_params *p)
6793 struct user_struct *user = NULL;
6794 struct io_ring_ctx *ctx;
6800 if (entries > IORING_MAX_ENTRIES) {
6801 if (!(p->flags & IORING_SETUP_CLAMP))
6803 entries = IORING_MAX_ENTRIES;
6807 * Use twice as many entries for the CQ ring. It's possible for the
6808 * application to drive a higher depth than the size of the SQ ring,
6809 * since the sqes are only used at submission time. This allows for
6810 * some flexibility in overcommitting a bit. If the application has
6811 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
6812 * of CQ ring entries manually.
6814 p->sq_entries = roundup_pow_of_two(entries);
6815 if (p->flags & IORING_SETUP_CQSIZE) {
6817 * If IORING_SETUP_CQSIZE is set, we do the same roundup
6818 * to a power-of-two, if it isn't already. We do NOT impose
6819 * any cq vs sq ring sizing.
6821 if (p->cq_entries < p->sq_entries)
6823 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
6824 if (!(p->flags & IORING_SETUP_CLAMP))
6826 p->cq_entries = IORING_MAX_CQ_ENTRIES;
6828 p->cq_entries = roundup_pow_of_two(p->cq_entries);
6830 p->cq_entries = 2 * p->sq_entries;
6833 user = get_uid(current_user());
6834 account_mem = !capable(CAP_IPC_LOCK);
6837 ret = io_account_mem(user,
6838 ring_pages(p->sq_entries, p->cq_entries));
6845 ctx = io_ring_ctx_alloc(p);
6848 io_unaccount_mem(user, ring_pages(p->sq_entries,
6853 ctx->compat = in_compat_syscall();
6854 ctx->account_mem = account_mem;
6856 ctx->creds = get_current_cred();
6858 ret = io_allocate_scq_urings(ctx, p);
6862 ret = io_sq_offload_start(ctx, p);
6866 memset(&p->sq_off, 0, sizeof(p->sq_off));
6867 p->sq_off.head = offsetof(struct io_rings, sq.head);
6868 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
6869 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
6870 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
6871 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
6872 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
6873 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
6875 memset(&p->cq_off, 0, sizeof(p->cq_off));
6876 p->cq_off.head = offsetof(struct io_rings, cq.head);
6877 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
6878 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
6879 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
6880 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
6881 p->cq_off.cqes = offsetof(struct io_rings, cqes);
6884 * Install ring fd as the very last thing, so we don't risk someone
6885 * having closed it before we finish setup
6887 ret = io_uring_get_fd(ctx);
6891 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
6892 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
6893 IORING_FEAT_CUR_PERSONALITY;
6894 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
6897 io_ring_ctx_wait_and_kill(ctx);
6902 * Sets up an aio uring context, and returns the fd. Applications asks for a
6903 * ring size, we return the actual sq/cq ring sizes (among other things) in the
6904 * params structure passed in.
6906 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
6908 struct io_uring_params p;
6912 if (copy_from_user(&p, params, sizeof(p)))
6914 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
6919 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
6920 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
6921 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
6924 ret = io_uring_create(entries, &p);
6928 if (copy_to_user(params, &p, sizeof(p)))
6934 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
6935 struct io_uring_params __user *, params)
6937 return io_uring_setup(entries, params);
6940 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
6942 struct io_uring_probe *p;
6946 size = struct_size(p, ops, nr_args);
6947 if (size == SIZE_MAX)
6949 p = kzalloc(size, GFP_KERNEL);
6954 if (copy_from_user(p, arg, size))
6957 if (memchr_inv(p, 0, size))
6960 p->last_op = IORING_OP_LAST - 1;
6961 if (nr_args > IORING_OP_LAST)
6962 nr_args = IORING_OP_LAST;
6964 for (i = 0; i < nr_args; i++) {
6966 if (!io_op_defs[i].not_supported)
6967 p->ops[i].flags = IO_URING_OP_SUPPORTED;
6972 if (copy_to_user(arg, p, size))
6979 static int io_register_personality(struct io_ring_ctx *ctx)
6981 const struct cred *creds = get_current_cred();
6984 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
6985 USHRT_MAX, GFP_KERNEL);
6991 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
6993 const struct cred *old_creds;
6995 old_creds = idr_remove(&ctx->personality_idr, id);
6997 put_cred(old_creds);
7004 static bool io_register_op_must_quiesce(int op)
7007 case IORING_UNREGISTER_FILES:
7008 case IORING_REGISTER_FILES_UPDATE:
7009 case IORING_REGISTER_PROBE:
7010 case IORING_REGISTER_PERSONALITY:
7011 case IORING_UNREGISTER_PERSONALITY:
7018 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
7019 void __user *arg, unsigned nr_args)
7020 __releases(ctx->uring_lock)
7021 __acquires(ctx->uring_lock)
7026 * We're inside the ring mutex, if the ref is already dying, then
7027 * someone else killed the ctx or is already going through
7028 * io_uring_register().
7030 if (percpu_ref_is_dying(&ctx->refs))
7033 if (io_register_op_must_quiesce(opcode)) {
7034 percpu_ref_kill(&ctx->refs);
7037 * Drop uring mutex before waiting for references to exit. If
7038 * another thread is currently inside io_uring_enter() it might
7039 * need to grab the uring_lock to make progress. If we hold it
7040 * here across the drain wait, then we can deadlock. It's safe
7041 * to drop the mutex here, since no new references will come in
7042 * after we've killed the percpu ref.
7044 mutex_unlock(&ctx->uring_lock);
7045 ret = wait_for_completion_interruptible(&ctx->completions[0]);
7046 mutex_lock(&ctx->uring_lock);
7048 percpu_ref_resurrect(&ctx->refs);
7055 case IORING_REGISTER_BUFFERS:
7056 ret = io_sqe_buffer_register(ctx, arg, nr_args);
7058 case IORING_UNREGISTER_BUFFERS:
7062 ret = io_sqe_buffer_unregister(ctx);
7064 case IORING_REGISTER_FILES:
7065 ret = io_sqe_files_register(ctx, arg, nr_args);
7067 case IORING_UNREGISTER_FILES:
7071 ret = io_sqe_files_unregister(ctx);
7073 case IORING_REGISTER_FILES_UPDATE:
7074 ret = io_sqe_files_update(ctx, arg, nr_args);
7076 case IORING_REGISTER_EVENTFD:
7077 case IORING_REGISTER_EVENTFD_ASYNC:
7081 ret = io_eventfd_register(ctx, arg);
7084 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
7085 ctx->eventfd_async = 1;
7087 ctx->eventfd_async = 0;
7089 case IORING_UNREGISTER_EVENTFD:
7093 ret = io_eventfd_unregister(ctx);
7095 case IORING_REGISTER_PROBE:
7097 if (!arg || nr_args > 256)
7099 ret = io_probe(ctx, arg, nr_args);
7101 case IORING_REGISTER_PERSONALITY:
7105 ret = io_register_personality(ctx);
7107 case IORING_UNREGISTER_PERSONALITY:
7111 ret = io_unregister_personality(ctx, nr_args);
7118 if (io_register_op_must_quiesce(opcode)) {
7119 /* bring the ctx back to life */
7120 percpu_ref_reinit(&ctx->refs);
7122 reinit_completion(&ctx->completions[0]);
7127 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
7128 void __user *, arg, unsigned int, nr_args)
7130 struct io_ring_ctx *ctx;
7139 if (f.file->f_op != &io_uring_fops)
7142 ctx = f.file->private_data;
7144 mutex_lock(&ctx->uring_lock);
7145 ret = __io_uring_register(ctx, opcode, arg, nr_args);
7146 mutex_unlock(&ctx->uring_lock);
7147 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
7148 ctx->cq_ev_fd != NULL, ret);
7154 static int __init io_uring_init(void)
7156 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
7157 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
7158 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
7161 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
7162 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
7163 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
7164 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
7165 BUILD_BUG_SQE_ELEM(1, __u8, flags);
7166 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
7167 BUILD_BUG_SQE_ELEM(4, __s32, fd);
7168 BUILD_BUG_SQE_ELEM(8, __u64, off);
7169 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
7170 BUILD_BUG_SQE_ELEM(16, __u64, addr);
7171 BUILD_BUG_SQE_ELEM(24, __u32, len);
7172 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
7173 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
7174 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
7175 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
7176 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
7177 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
7178 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
7179 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
7180 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
7181 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
7182 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
7183 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
7184 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
7185 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
7186 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
7187 BUILD_BUG_SQE_ELEM(42, __u16, personality);
7189 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
7190 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
7193 __initcall(io_uring_init);