1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/mmu_context.h>
59 #include <linux/percpu.h>
60 #include <linux/slab.h>
61 #include <linux/kthread.h>
62 #include <linux/blkdev.h>
63 #include <linux/bvec.h>
64 #include <linux/net.h>
66 #include <net/af_unix.h>
68 #include <linux/anon_inodes.h>
69 #include <linux/sched/mm.h>
70 #include <linux/uaccess.h>
71 #include <linux/nospec.h>
72 #include <linux/sizes.h>
73 #include <linux/hugetlb.h>
74 #include <linux/highmem.h>
75 #include <linux/namei.h>
76 #include <linux/fsnotify.h>
77 #include <linux/fadvise.h>
78 #include <linux/eventpoll.h>
79 #include <linux/fs_struct.h>
80 #include <linux/splice.h>
81 #include <linux/task_work.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp;
104 u32 tail ____cacheline_aligned_in_smp;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq, cq;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask, cq_ring_mask;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries, cq_ring_entries;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
155 * Number of completion events lost because the queue was full;
156 * this should be avoided by the application by making sure
157 * there are not more requests pending than there is space in
158 * the completion queue.
160 * Written by the kernel, shouldn't be modified by the
161 * application (i.e. get number of "new events" by comparing to
164 * As completion events come in out of order this counter is not
165 * ordered with any other data.
169 * Ring buffer of completion events.
171 * The kernel writes completion events fresh every time they are
172 * produced, so the application is allowed to modify pending
175 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
178 struct io_mapped_ubuf {
181 struct bio_vec *bvec;
182 unsigned int nr_bvecs;
185 struct fixed_file_table {
189 struct fixed_file_data {
190 struct fixed_file_table *table;
191 struct io_ring_ctx *ctx;
193 struct percpu_ref refs;
194 struct llist_head put_llist;
195 struct work_struct ref_work;
196 struct completion done;
200 struct list_head list;
208 struct percpu_ref refs;
209 } ____cacheline_aligned_in_smp;
213 unsigned int compat: 1;
214 unsigned int account_mem: 1;
215 unsigned int cq_overflow_flushed: 1;
216 unsigned int drain_next: 1;
217 unsigned int eventfd_async: 1;
220 * Ring buffer of indices into array of io_uring_sqe, which is
221 * mmapped by the application using the IORING_OFF_SQES offset.
223 * This indirection could e.g. be used to assign fixed
224 * io_uring_sqe entries to operations and only submit them to
225 * the queue when needed.
227 * The kernel modifies neither the indices array nor the entries
231 unsigned cached_sq_head;
234 unsigned sq_thread_idle;
235 unsigned cached_sq_dropped;
236 atomic_t cached_cq_overflow;
237 unsigned long sq_check_overflow;
239 struct list_head defer_list;
240 struct list_head timeout_list;
241 struct list_head cq_overflow_list;
243 wait_queue_head_t inflight_wait;
244 struct io_uring_sqe *sq_sqes;
245 } ____cacheline_aligned_in_smp;
247 struct io_rings *rings;
251 struct task_struct *sqo_thread; /* if using sq thread polling */
252 struct mm_struct *sqo_mm;
253 wait_queue_head_t sqo_wait;
256 * If used, fixed file set. Writers must ensure that ->refs is dead,
257 * readers must ensure that ->refs is alive as long as the file* is
258 * used. Only updated through io_uring_register(2).
260 struct fixed_file_data *file_data;
261 unsigned nr_user_files;
263 struct file *ring_file;
265 /* if used, fixed mapped user buffers */
266 unsigned nr_user_bufs;
267 struct io_mapped_ubuf *user_bufs;
269 struct user_struct *user;
271 const struct cred *creds;
273 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
274 struct completion *completions;
276 /* if all else fails... */
277 struct io_kiocb *fallback_req;
279 #if defined(CONFIG_UNIX)
280 struct socket *ring_sock;
283 struct idr io_buffer_idr;
285 struct idr personality_idr;
288 unsigned cached_cq_tail;
291 atomic_t cq_timeouts;
292 unsigned long cq_check_overflow;
293 struct wait_queue_head cq_wait;
294 struct fasync_struct *cq_fasync;
295 struct eventfd_ctx *cq_ev_fd;
296 } ____cacheline_aligned_in_smp;
299 struct mutex uring_lock;
300 wait_queue_head_t wait;
301 } ____cacheline_aligned_in_smp;
304 spinlock_t completion_lock;
307 * ->poll_list is protected by the ctx->uring_lock for
308 * io_uring instances that don't use IORING_SETUP_SQPOLL.
309 * For SQPOLL, only the single threaded io_sq_thread() will
310 * manipulate the list, hence no extra locking is needed there.
312 struct list_head poll_list;
313 struct hlist_head *cancel_hash;
314 unsigned cancel_hash_bits;
315 bool poll_multi_file;
317 spinlock_t inflight_lock;
318 struct list_head inflight_list;
319 } ____cacheline_aligned_in_smp;
323 * First field must be the file pointer in all the
324 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
326 struct io_poll_iocb {
329 struct wait_queue_head *head;
335 struct wait_queue_entry wait;
340 struct file *put_file;
344 struct io_timeout_data {
345 struct io_kiocb *req;
346 struct hrtimer timer;
347 struct timespec64 ts;
348 enum hrtimer_mode mode;
354 struct sockaddr __user *addr;
355 int __user *addr_len;
380 /* NOTE: kiocb has the file as the first member, so don't do it here */
388 struct sockaddr __user *addr;
395 struct user_msghdr __user *msg;
401 struct io_buffer *kbuf;
410 struct filename *filename;
411 struct statx __user *buffer;
415 struct io_files_update {
441 struct epoll_event event;
445 struct file *file_out;
446 struct file *file_in;
453 struct io_provide_buf {
462 struct io_async_connect {
463 struct sockaddr_storage address;
466 struct io_async_msghdr {
467 struct iovec fast_iov[UIO_FASTIOV];
469 struct sockaddr __user *uaddr;
471 struct sockaddr_storage addr;
475 struct iovec fast_iov[UIO_FASTIOV];
481 struct io_async_ctx {
483 struct io_async_rw rw;
484 struct io_async_msghdr msg;
485 struct io_async_connect connect;
486 struct io_timeout_data timeout;
491 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
492 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
493 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
494 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
495 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
496 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
503 REQ_F_IOPOLL_COMPLETED_BIT,
504 REQ_F_LINK_TIMEOUT_BIT,
508 REQ_F_TIMEOUT_NOSEQ_BIT,
509 REQ_F_COMP_LOCKED_BIT,
510 REQ_F_NEED_CLEANUP_BIT,
513 REQ_F_BUFFER_SELECTED_BIT,
515 /* not a real bit, just to check we're not overflowing the space */
521 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
522 /* drain existing IO first */
523 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
525 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
526 /* doesn't sever on completion < 0 */
527 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
529 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
530 /* IOSQE_BUFFER_SELECT */
531 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
533 /* already grabbed next link */
534 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
535 /* fail rest of links */
536 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
537 /* on inflight list */
538 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
539 /* read/write uses file position */
540 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
541 /* must not punt to workers */
542 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
543 /* polled IO has completed */
544 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
545 /* has linked timeout */
546 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
547 /* timeout request */
548 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
550 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
551 /* must be punted even for NONBLOCK */
552 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
553 /* no timeout sequence */
554 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
555 /* completion under lock */
556 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
558 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
559 /* in overflow list */
560 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
561 /* already went through poll handler */
562 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
563 /* buffer already selected */
564 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
568 struct io_poll_iocb poll;
569 struct io_wq_work work;
573 * NOTE! Each of the iocb union members has the file pointer
574 * as the first entry in their struct definition. So you can
575 * access the file pointer through any of the sub-structs,
576 * or directly as just 'ki_filp' in this struct.
582 struct io_poll_iocb poll;
583 struct io_accept accept;
585 struct io_cancel cancel;
586 struct io_timeout timeout;
587 struct io_connect connect;
588 struct io_sr_msg sr_msg;
590 struct io_close close;
591 struct io_files_update files_update;
592 struct io_fadvise fadvise;
593 struct io_madvise madvise;
594 struct io_epoll epoll;
595 struct io_splice splice;
596 struct io_provide_buf pbuf;
599 struct io_async_ctx *io;
600 bool needs_fixed_file;
603 struct io_ring_ctx *ctx;
604 struct list_head list;
608 struct task_struct *task;
615 struct list_head link_list;
617 struct list_head inflight_entry;
621 * Only commands that never go async can use the below fields,
622 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
623 * async armed poll handlers for regular commands. The latter
624 * restore the work, if needed.
627 struct callback_head task_work;
628 struct hlist_node hash_node;
629 struct async_poll *apoll;
632 struct io_wq_work work;
636 #define IO_PLUG_THRESHOLD 2
637 #define IO_IOPOLL_BATCH 8
639 struct io_submit_state {
640 struct blk_plug plug;
643 * io_kiocb alloc cache
645 void *reqs[IO_IOPOLL_BATCH];
646 unsigned int free_reqs;
649 * File reference cache
653 unsigned int has_refs;
654 unsigned int used_refs;
655 unsigned int ios_left;
659 /* needs req->io allocated for deferral/async */
660 unsigned async_ctx : 1;
661 /* needs current->mm setup, does mm access */
662 unsigned needs_mm : 1;
663 /* needs req->file assigned */
664 unsigned needs_file : 1;
665 /* needs req->file assigned IFF fd is >= 0 */
666 unsigned fd_non_neg : 1;
667 /* hash wq insertion if file is a regular file */
668 unsigned hash_reg_file : 1;
669 /* unbound wq insertion if file is a non-regular file */
670 unsigned unbound_nonreg_file : 1;
671 /* opcode is not supported by this kernel */
672 unsigned not_supported : 1;
673 /* needs file table */
674 unsigned file_table : 1;
676 unsigned needs_fs : 1;
677 /* set if opcode supports polled "wait" */
679 unsigned pollout : 1;
680 /* op supports buffer selection */
681 unsigned buffer_select : 1;
684 static const struct io_op_def io_op_defs[] = {
685 [IORING_OP_NOP] = {},
686 [IORING_OP_READV] = {
690 .unbound_nonreg_file = 1,
694 [IORING_OP_WRITEV] = {
699 .unbound_nonreg_file = 1,
702 [IORING_OP_FSYNC] = {
705 [IORING_OP_READ_FIXED] = {
707 .unbound_nonreg_file = 1,
710 [IORING_OP_WRITE_FIXED] = {
713 .unbound_nonreg_file = 1,
716 [IORING_OP_POLL_ADD] = {
718 .unbound_nonreg_file = 1,
720 [IORING_OP_POLL_REMOVE] = {},
721 [IORING_OP_SYNC_FILE_RANGE] = {
724 [IORING_OP_SENDMSG] = {
728 .unbound_nonreg_file = 1,
732 [IORING_OP_RECVMSG] = {
736 .unbound_nonreg_file = 1,
741 [IORING_OP_TIMEOUT] = {
745 [IORING_OP_TIMEOUT_REMOVE] = {},
746 [IORING_OP_ACCEPT] = {
749 .unbound_nonreg_file = 1,
753 [IORING_OP_ASYNC_CANCEL] = {},
754 [IORING_OP_LINK_TIMEOUT] = {
758 [IORING_OP_CONNECT] = {
762 .unbound_nonreg_file = 1,
765 [IORING_OP_FALLOCATE] = {
768 [IORING_OP_OPENAT] = {
774 [IORING_OP_CLOSE] = {
778 [IORING_OP_FILES_UPDATE] = {
782 [IORING_OP_STATX] = {
791 .unbound_nonreg_file = 1,
795 [IORING_OP_WRITE] = {
798 .unbound_nonreg_file = 1,
801 [IORING_OP_FADVISE] = {
804 [IORING_OP_MADVISE] = {
810 .unbound_nonreg_file = 1,
816 .unbound_nonreg_file = 1,
820 [IORING_OP_OPENAT2] = {
826 [IORING_OP_EPOLL_CTL] = {
827 .unbound_nonreg_file = 1,
830 [IORING_OP_SPLICE] = {
833 .unbound_nonreg_file = 1,
835 [IORING_OP_PROVIDE_BUFFERS] = {},
836 [IORING_OP_REMOVE_BUFFERS] = {},
839 static void io_wq_submit_work(struct io_wq_work **workptr);
840 static void io_cqring_fill_event(struct io_kiocb *req, long res);
841 static void io_put_req(struct io_kiocb *req);
842 static void __io_double_put_req(struct io_kiocb *req);
843 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
844 static void io_queue_linked_timeout(struct io_kiocb *req);
845 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
846 struct io_uring_files_update *ip,
848 static int io_grab_files(struct io_kiocb *req);
849 static void io_ring_file_ref_flush(struct fixed_file_data *data);
850 static void io_cleanup_req(struct io_kiocb *req);
851 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
852 int fd, struct file **out_file, bool fixed);
853 static void __io_queue_sqe(struct io_kiocb *req,
854 const struct io_uring_sqe *sqe);
856 static struct kmem_cache *req_cachep;
858 static const struct file_operations io_uring_fops;
860 struct sock *io_uring_get_socket(struct file *file)
862 #if defined(CONFIG_UNIX)
863 if (file->f_op == &io_uring_fops) {
864 struct io_ring_ctx *ctx = file->private_data;
866 return ctx->ring_sock->sk;
871 EXPORT_SYMBOL(io_uring_get_socket);
873 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
875 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
877 complete(&ctx->completions[0]);
880 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
882 struct io_ring_ctx *ctx;
885 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
889 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
890 if (!ctx->fallback_req)
893 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
894 if (!ctx->completions)
898 * Use 5 bits less than the max cq entries, that should give us around
899 * 32 entries per hash list if totally full and uniformly spread.
901 hash_bits = ilog2(p->cq_entries);
905 ctx->cancel_hash_bits = hash_bits;
906 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
908 if (!ctx->cancel_hash)
910 __hash_init(ctx->cancel_hash, 1U << hash_bits);
912 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
913 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
916 ctx->flags = p->flags;
917 init_waitqueue_head(&ctx->cq_wait);
918 INIT_LIST_HEAD(&ctx->cq_overflow_list);
919 init_completion(&ctx->completions[0]);
920 init_completion(&ctx->completions[1]);
921 idr_init(&ctx->io_buffer_idr);
922 idr_init(&ctx->personality_idr);
923 mutex_init(&ctx->uring_lock);
924 init_waitqueue_head(&ctx->wait);
925 spin_lock_init(&ctx->completion_lock);
926 INIT_LIST_HEAD(&ctx->poll_list);
927 INIT_LIST_HEAD(&ctx->defer_list);
928 INIT_LIST_HEAD(&ctx->timeout_list);
929 init_waitqueue_head(&ctx->inflight_wait);
930 spin_lock_init(&ctx->inflight_lock);
931 INIT_LIST_HEAD(&ctx->inflight_list);
934 if (ctx->fallback_req)
935 kmem_cache_free(req_cachep, ctx->fallback_req);
936 kfree(ctx->completions);
937 kfree(ctx->cancel_hash);
942 static inline bool __req_need_defer(struct io_kiocb *req)
944 struct io_ring_ctx *ctx = req->ctx;
946 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
947 + atomic_read(&ctx->cached_cq_overflow);
950 static inline bool req_need_defer(struct io_kiocb *req)
952 if (unlikely(req->flags & REQ_F_IO_DRAIN))
953 return __req_need_defer(req);
958 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
960 struct io_kiocb *req;
962 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
963 if (req && !req_need_defer(req)) {
964 list_del_init(&req->list);
971 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
973 struct io_kiocb *req;
975 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
977 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
979 if (!__req_need_defer(req)) {
980 list_del_init(&req->list);
988 static void __io_commit_cqring(struct io_ring_ctx *ctx)
990 struct io_rings *rings = ctx->rings;
992 /* order cqe stores with ring update */
993 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
995 if (wq_has_sleeper(&ctx->cq_wait)) {
996 wake_up_interruptible(&ctx->cq_wait);
997 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1001 static inline void io_req_work_grab_env(struct io_kiocb *req,
1002 const struct io_op_def *def)
1004 if (!req->work.mm && def->needs_mm) {
1005 mmgrab(current->mm);
1006 req->work.mm = current->mm;
1008 if (!req->work.creds)
1009 req->work.creds = get_current_cred();
1010 if (!req->work.fs && def->needs_fs) {
1011 spin_lock(¤t->fs->lock);
1012 if (!current->fs->in_exec) {
1013 req->work.fs = current->fs;
1014 req->work.fs->users++;
1016 req->work.flags |= IO_WQ_WORK_CANCEL;
1018 spin_unlock(¤t->fs->lock);
1020 if (!req->work.task_pid)
1021 req->work.task_pid = task_pid_vnr(current);
1024 static inline void io_req_work_drop_env(struct io_kiocb *req)
1027 mmdrop(req->work.mm);
1028 req->work.mm = NULL;
1030 if (req->work.creds) {
1031 put_cred(req->work.creds);
1032 req->work.creds = NULL;
1035 struct fs_struct *fs = req->work.fs;
1037 spin_lock(&req->work.fs->lock);
1040 spin_unlock(&req->work.fs->lock);
1046 static inline void io_prep_async_work(struct io_kiocb *req,
1047 struct io_kiocb **link)
1049 const struct io_op_def *def = &io_op_defs[req->opcode];
1051 if (req->flags & REQ_F_ISREG) {
1052 if (def->hash_reg_file)
1053 io_wq_hash_work(&req->work, file_inode(req->file));
1055 if (def->unbound_nonreg_file)
1056 req->work.flags |= IO_WQ_WORK_UNBOUND;
1059 io_req_work_grab_env(req, def);
1061 *link = io_prep_linked_timeout(req);
1064 static inline void io_queue_async_work(struct io_kiocb *req)
1066 struct io_ring_ctx *ctx = req->ctx;
1067 struct io_kiocb *link;
1069 io_prep_async_work(req, &link);
1071 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1072 &req->work, req->flags);
1073 io_wq_enqueue(ctx->io_wq, &req->work);
1076 io_queue_linked_timeout(link);
1079 static void io_kill_timeout(struct io_kiocb *req)
1083 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1085 atomic_inc(&req->ctx->cq_timeouts);
1086 list_del_init(&req->list);
1087 io_cqring_fill_event(req, 0);
1092 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1094 struct io_kiocb *req, *tmp;
1096 spin_lock_irq(&ctx->completion_lock);
1097 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1098 io_kill_timeout(req);
1099 spin_unlock_irq(&ctx->completion_lock);
1102 static void io_commit_cqring(struct io_ring_ctx *ctx)
1104 struct io_kiocb *req;
1106 while ((req = io_get_timeout_req(ctx)) != NULL)
1107 io_kill_timeout(req);
1109 __io_commit_cqring(ctx);
1111 while ((req = io_get_deferred_req(ctx)) != NULL)
1112 io_queue_async_work(req);
1115 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1117 struct io_rings *rings = ctx->rings;
1120 tail = ctx->cached_cq_tail;
1122 * writes to the cq entry need to come after reading head; the
1123 * control dependency is enough as we're using WRITE_ONCE to
1126 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1129 ctx->cached_cq_tail++;
1130 return &rings->cqes[tail & ctx->cq_mask];
1133 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1137 if (!ctx->eventfd_async)
1139 return io_wq_current_is_worker();
1142 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1144 if (waitqueue_active(&ctx->wait))
1145 wake_up(&ctx->wait);
1146 if (waitqueue_active(&ctx->sqo_wait))
1147 wake_up(&ctx->sqo_wait);
1148 if (io_should_trigger_evfd(ctx))
1149 eventfd_signal(ctx->cq_ev_fd, 1);
1152 /* Returns true if there are no backlogged entries after the flush */
1153 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1155 struct io_rings *rings = ctx->rings;
1156 struct io_uring_cqe *cqe;
1157 struct io_kiocb *req;
1158 unsigned long flags;
1162 if (list_empty_careful(&ctx->cq_overflow_list))
1164 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1165 rings->cq_ring_entries))
1169 spin_lock_irqsave(&ctx->completion_lock, flags);
1171 /* if force is set, the ring is going away. always drop after that */
1173 ctx->cq_overflow_flushed = 1;
1176 while (!list_empty(&ctx->cq_overflow_list)) {
1177 cqe = io_get_cqring(ctx);
1181 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1183 list_move(&req->list, &list);
1184 req->flags &= ~REQ_F_OVERFLOW;
1186 WRITE_ONCE(cqe->user_data, req->user_data);
1187 WRITE_ONCE(cqe->res, req->result);
1188 WRITE_ONCE(cqe->flags, req->cflags);
1190 WRITE_ONCE(ctx->rings->cq_overflow,
1191 atomic_inc_return(&ctx->cached_cq_overflow));
1195 io_commit_cqring(ctx);
1197 clear_bit(0, &ctx->sq_check_overflow);
1198 clear_bit(0, &ctx->cq_check_overflow);
1200 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1201 io_cqring_ev_posted(ctx);
1203 while (!list_empty(&list)) {
1204 req = list_first_entry(&list, struct io_kiocb, list);
1205 list_del(&req->list);
1212 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1214 struct io_ring_ctx *ctx = req->ctx;
1215 struct io_uring_cqe *cqe;
1217 trace_io_uring_complete(ctx, req->user_data, res);
1220 * If we can't get a cq entry, userspace overflowed the
1221 * submission (by quite a lot). Increment the overflow count in
1224 cqe = io_get_cqring(ctx);
1226 WRITE_ONCE(cqe->user_data, req->user_data);
1227 WRITE_ONCE(cqe->res, res);
1228 WRITE_ONCE(cqe->flags, cflags);
1229 } else if (ctx->cq_overflow_flushed) {
1230 WRITE_ONCE(ctx->rings->cq_overflow,
1231 atomic_inc_return(&ctx->cached_cq_overflow));
1233 if (list_empty(&ctx->cq_overflow_list)) {
1234 set_bit(0, &ctx->sq_check_overflow);
1235 set_bit(0, &ctx->cq_check_overflow);
1237 req->flags |= REQ_F_OVERFLOW;
1238 refcount_inc(&req->refs);
1240 req->cflags = cflags;
1241 list_add_tail(&req->list, &ctx->cq_overflow_list);
1245 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1247 __io_cqring_fill_event(req, res, 0);
1250 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1252 struct io_ring_ctx *ctx = req->ctx;
1253 unsigned long flags;
1255 spin_lock_irqsave(&ctx->completion_lock, flags);
1256 __io_cqring_fill_event(req, res, cflags);
1257 io_commit_cqring(ctx);
1258 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1260 io_cqring_ev_posted(ctx);
1263 static void io_cqring_add_event(struct io_kiocb *req, long res)
1265 __io_cqring_add_event(req, res, 0);
1268 static inline bool io_is_fallback_req(struct io_kiocb *req)
1270 return req == (struct io_kiocb *)
1271 ((unsigned long) req->ctx->fallback_req & ~1UL);
1274 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1276 struct io_kiocb *req;
1278 req = ctx->fallback_req;
1279 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
1285 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
1286 struct io_submit_state *state)
1288 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1289 struct io_kiocb *req;
1292 req = kmem_cache_alloc(req_cachep, gfp);
1295 } else if (!state->free_reqs) {
1299 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1300 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1303 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1304 * retry single alloc to be on the safe side.
1306 if (unlikely(ret <= 0)) {
1307 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1308 if (!state->reqs[0])
1312 state->free_reqs = ret - 1;
1313 req = state->reqs[ret - 1];
1316 req = state->reqs[state->free_reqs];
1324 /* one is dropped after submission, the other at completion */
1325 refcount_set(&req->refs, 2);
1327 INIT_IO_WORK(&req->work, io_wq_submit_work);
1330 req = io_get_fallback_req(ctx);
1333 percpu_ref_put(&ctx->refs);
1337 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1341 percpu_ref_put(&req->ctx->file_data->refs);
1346 static void __io_req_do_free(struct io_kiocb *req)
1348 if (likely(!io_is_fallback_req(req)))
1349 kmem_cache_free(req_cachep, req);
1351 clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req);
1354 static void __io_req_aux_free(struct io_kiocb *req)
1356 if (req->flags & REQ_F_NEED_CLEANUP)
1357 io_cleanup_req(req);
1361 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1363 io_req_work_drop_env(req);
1366 static void __io_free_req(struct io_kiocb *req)
1368 __io_req_aux_free(req);
1370 if (req->flags & REQ_F_INFLIGHT) {
1371 struct io_ring_ctx *ctx = req->ctx;
1372 unsigned long flags;
1374 spin_lock_irqsave(&ctx->inflight_lock, flags);
1375 list_del(&req->inflight_entry);
1376 if (waitqueue_active(&ctx->inflight_wait))
1377 wake_up(&ctx->inflight_wait);
1378 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1381 percpu_ref_put(&req->ctx->refs);
1382 __io_req_do_free(req);
1386 void *reqs[IO_IOPOLL_BATCH];
1391 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1393 int fixed_refs = rb->to_free;
1397 if (rb->need_iter) {
1398 int i, inflight = 0;
1399 unsigned long flags;
1402 for (i = 0; i < rb->to_free; i++) {
1403 struct io_kiocb *req = rb->reqs[i];
1405 if (req->flags & REQ_F_FIXED_FILE) {
1409 if (req->flags & REQ_F_INFLIGHT)
1411 __io_req_aux_free(req);
1416 spin_lock_irqsave(&ctx->inflight_lock, flags);
1417 for (i = 0; i < rb->to_free; i++) {
1418 struct io_kiocb *req = rb->reqs[i];
1420 if (req->flags & REQ_F_INFLIGHT) {
1421 list_del(&req->inflight_entry);
1426 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1428 if (waitqueue_active(&ctx->inflight_wait))
1429 wake_up(&ctx->inflight_wait);
1432 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1434 percpu_ref_put_many(&ctx->file_data->refs, fixed_refs);
1435 percpu_ref_put_many(&ctx->refs, rb->to_free);
1436 rb->to_free = rb->need_iter = 0;
1439 static bool io_link_cancel_timeout(struct io_kiocb *req)
1441 struct io_ring_ctx *ctx = req->ctx;
1444 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1446 io_cqring_fill_event(req, -ECANCELED);
1447 io_commit_cqring(ctx);
1448 req->flags &= ~REQ_F_LINK;
1456 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1458 struct io_ring_ctx *ctx = req->ctx;
1459 bool wake_ev = false;
1461 /* Already got next link */
1462 if (req->flags & REQ_F_LINK_NEXT)
1466 * The list should never be empty when we are called here. But could
1467 * potentially happen if the chain is messed up, check to be on the
1470 while (!list_empty(&req->link_list)) {
1471 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1472 struct io_kiocb, link_list);
1474 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1475 (nxt->flags & REQ_F_TIMEOUT))) {
1476 list_del_init(&nxt->link_list);
1477 wake_ev |= io_link_cancel_timeout(nxt);
1478 req->flags &= ~REQ_F_LINK_TIMEOUT;
1482 list_del_init(&req->link_list);
1483 if (!list_empty(&nxt->link_list))
1484 nxt->flags |= REQ_F_LINK;
1489 req->flags |= REQ_F_LINK_NEXT;
1491 io_cqring_ev_posted(ctx);
1495 * Called if REQ_F_LINK is set, and we fail the head request
1497 static void io_fail_links(struct io_kiocb *req)
1499 struct io_ring_ctx *ctx = req->ctx;
1500 unsigned long flags;
1502 spin_lock_irqsave(&ctx->completion_lock, flags);
1504 while (!list_empty(&req->link_list)) {
1505 struct io_kiocb *link = list_first_entry(&req->link_list,
1506 struct io_kiocb, link_list);
1508 list_del_init(&link->link_list);
1509 trace_io_uring_fail_link(req, link);
1511 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1512 link->opcode == IORING_OP_LINK_TIMEOUT) {
1513 io_link_cancel_timeout(link);
1515 io_cqring_fill_event(link, -ECANCELED);
1516 __io_double_put_req(link);
1518 req->flags &= ~REQ_F_LINK_TIMEOUT;
1521 io_commit_cqring(ctx);
1522 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1523 io_cqring_ev_posted(ctx);
1526 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1528 if (likely(!(req->flags & REQ_F_LINK)))
1532 * If LINK is set, we have dependent requests in this chain. If we
1533 * didn't fail this request, queue the first one up, moving any other
1534 * dependencies to the next request. In case of failure, fail the rest
1537 if (req->flags & REQ_F_FAIL_LINK) {
1539 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1540 REQ_F_LINK_TIMEOUT) {
1541 struct io_ring_ctx *ctx = req->ctx;
1542 unsigned long flags;
1545 * If this is a timeout link, we could be racing with the
1546 * timeout timer. Grab the completion lock for this case to
1547 * protect against that.
1549 spin_lock_irqsave(&ctx->completion_lock, flags);
1550 io_req_link_next(req, nxt);
1551 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1553 io_req_link_next(req, nxt);
1557 static void io_free_req(struct io_kiocb *req)
1559 struct io_kiocb *nxt = NULL;
1561 io_req_find_next(req, &nxt);
1565 io_queue_async_work(nxt);
1568 static void io_link_work_cb(struct io_wq_work **workptr)
1570 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1571 struct io_kiocb *link;
1573 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1574 io_queue_linked_timeout(link);
1575 io_wq_submit_work(workptr);
1578 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1580 struct io_kiocb *link;
1581 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1583 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1584 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1586 *workptr = &nxt->work;
1587 link = io_prep_linked_timeout(nxt);
1589 nxt->work.func = io_link_work_cb;
1593 * Drop reference to request, return next in chain (if there is one) if this
1594 * was the last reference to this request.
1596 __attribute__((nonnull))
1597 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1599 if (refcount_dec_and_test(&req->refs)) {
1600 io_req_find_next(req, nxtptr);
1605 static void io_put_req(struct io_kiocb *req)
1607 if (refcount_dec_and_test(&req->refs))
1611 static void io_steal_work(struct io_kiocb *req,
1612 struct io_wq_work **workptr)
1615 * It's in an io-wq worker, so there always should be at least
1616 * one reference, which will be dropped in io_put_work() just
1617 * after the current handler returns.
1619 * It also means, that if the counter dropped to 1, then there is
1620 * no asynchronous users left, so it's safe to steal the next work.
1622 if (refcount_read(&req->refs) == 1) {
1623 struct io_kiocb *nxt = NULL;
1625 io_req_find_next(req, &nxt);
1627 io_wq_assign_next(workptr, nxt);
1632 * Must only be used if we don't need to care about links, usually from
1633 * within the completion handling itself.
1635 static void __io_double_put_req(struct io_kiocb *req)
1637 /* drop both submit and complete references */
1638 if (refcount_sub_and_test(2, &req->refs))
1642 static void io_double_put_req(struct io_kiocb *req)
1644 /* drop both submit and complete references */
1645 if (refcount_sub_and_test(2, &req->refs))
1649 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1651 struct io_rings *rings = ctx->rings;
1653 if (test_bit(0, &ctx->cq_check_overflow)) {
1655 * noflush == true is from the waitqueue handler, just ensure
1656 * we wake up the task, and the next invocation will flush the
1657 * entries. We cannot safely to it from here.
1659 if (noflush && !list_empty(&ctx->cq_overflow_list))
1662 io_cqring_overflow_flush(ctx, false);
1665 /* See comment at the top of this file */
1667 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1670 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1672 struct io_rings *rings = ctx->rings;
1674 /* make sure SQ entry isn't read before tail */
1675 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1678 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1680 if ((req->flags & REQ_F_LINK) || io_is_fallback_req(req))
1683 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1686 rb->reqs[rb->to_free++] = req;
1687 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1688 io_free_req_many(req->ctx, rb);
1692 static int io_put_kbuf(struct io_kiocb *req)
1694 struct io_buffer *kbuf;
1697 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1698 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1699 cflags |= IORING_CQE_F_BUFFER;
1706 * Find and free completed poll iocbs
1708 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1709 struct list_head *done)
1711 struct req_batch rb;
1712 struct io_kiocb *req;
1714 rb.to_free = rb.need_iter = 0;
1715 while (!list_empty(done)) {
1718 req = list_first_entry(done, struct io_kiocb, list);
1719 list_del(&req->list);
1721 if (req->flags & REQ_F_BUFFER_SELECTED)
1722 cflags = io_put_kbuf(req);
1724 __io_cqring_fill_event(req, req->result, cflags);
1727 if (refcount_dec_and_test(&req->refs) &&
1728 !io_req_multi_free(&rb, req))
1732 io_commit_cqring(ctx);
1733 if (ctx->flags & IORING_SETUP_SQPOLL)
1734 io_cqring_ev_posted(ctx);
1735 io_free_req_many(ctx, &rb);
1738 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1741 struct io_kiocb *req, *tmp;
1747 * Only spin for completions if we don't have multiple devices hanging
1748 * off our complete list, and we're under the requested amount.
1750 spin = !ctx->poll_multi_file && *nr_events < min;
1753 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1754 struct kiocb *kiocb = &req->rw.kiocb;
1757 * Move completed entries to our local list. If we find a
1758 * request that requires polling, break out and complete
1759 * the done list first, if we have entries there.
1761 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1762 list_move_tail(&req->list, &done);
1765 if (!list_empty(&done))
1768 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1777 if (!list_empty(&done))
1778 io_iopoll_complete(ctx, nr_events, &done);
1784 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1785 * non-spinning poll check - we'll still enter the driver poll loop, but only
1786 * as a non-spinning completion check.
1788 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1791 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1794 ret = io_do_iopoll(ctx, nr_events, min);
1797 if (!min || *nr_events >= min)
1805 * We can't just wait for polled events to come to us, we have to actively
1806 * find and complete them.
1808 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1810 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1813 mutex_lock(&ctx->uring_lock);
1814 while (!list_empty(&ctx->poll_list)) {
1815 unsigned int nr_events = 0;
1817 io_iopoll_getevents(ctx, &nr_events, 1);
1820 * Ensure we allow local-to-the-cpu processing to take place,
1821 * in this case we need to ensure that we reap all events.
1825 mutex_unlock(&ctx->uring_lock);
1828 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1831 int iters = 0, ret = 0;
1834 * We disallow the app entering submit/complete with polling, but we
1835 * still need to lock the ring to prevent racing with polled issue
1836 * that got punted to a workqueue.
1838 mutex_lock(&ctx->uring_lock);
1843 * Don't enter poll loop if we already have events pending.
1844 * If we do, we can potentially be spinning for commands that
1845 * already triggered a CQE (eg in error).
1847 if (io_cqring_events(ctx, false))
1851 * If a submit got punted to a workqueue, we can have the
1852 * application entering polling for a command before it gets
1853 * issued. That app will hold the uring_lock for the duration
1854 * of the poll right here, so we need to take a breather every
1855 * now and then to ensure that the issue has a chance to add
1856 * the poll to the issued list. Otherwise we can spin here
1857 * forever, while the workqueue is stuck trying to acquire the
1860 if (!(++iters & 7)) {
1861 mutex_unlock(&ctx->uring_lock);
1862 mutex_lock(&ctx->uring_lock);
1865 if (*nr_events < min)
1866 tmin = min - *nr_events;
1868 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1872 } while (min && !*nr_events && !need_resched());
1874 mutex_unlock(&ctx->uring_lock);
1878 static void kiocb_end_write(struct io_kiocb *req)
1881 * Tell lockdep we inherited freeze protection from submission
1884 if (req->flags & REQ_F_ISREG) {
1885 struct inode *inode = file_inode(req->file);
1887 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1889 file_end_write(req->file);
1892 static inline void req_set_fail_links(struct io_kiocb *req)
1894 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1895 req->flags |= REQ_F_FAIL_LINK;
1898 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1900 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1903 if (kiocb->ki_flags & IOCB_WRITE)
1904 kiocb_end_write(req);
1906 if (res != req->result)
1907 req_set_fail_links(req);
1908 if (req->flags & REQ_F_BUFFER_SELECTED)
1909 cflags = io_put_kbuf(req);
1910 __io_cqring_add_event(req, res, cflags);
1913 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1915 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1917 io_complete_rw_common(kiocb, res);
1921 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1923 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1925 if (kiocb->ki_flags & IOCB_WRITE)
1926 kiocb_end_write(req);
1928 if (res != req->result)
1929 req_set_fail_links(req);
1932 req->flags |= REQ_F_IOPOLL_COMPLETED;
1936 * After the iocb has been issued, it's safe to be found on the poll list.
1937 * Adding the kiocb to the list AFTER submission ensures that we don't
1938 * find it from a io_iopoll_getevents() thread before the issuer is done
1939 * accessing the kiocb cookie.
1941 static void io_iopoll_req_issued(struct io_kiocb *req)
1943 struct io_ring_ctx *ctx = req->ctx;
1946 * Track whether we have multiple files in our lists. This will impact
1947 * how we do polling eventually, not spinning if we're on potentially
1948 * different devices.
1950 if (list_empty(&ctx->poll_list)) {
1951 ctx->poll_multi_file = false;
1952 } else if (!ctx->poll_multi_file) {
1953 struct io_kiocb *list_req;
1955 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1957 if (list_req->file != req->file)
1958 ctx->poll_multi_file = true;
1962 * For fast devices, IO may have already completed. If it has, add
1963 * it to the front so we find it first.
1965 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1966 list_add(&req->list, &ctx->poll_list);
1968 list_add_tail(&req->list, &ctx->poll_list);
1970 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1971 wq_has_sleeper(&ctx->sqo_wait))
1972 wake_up(&ctx->sqo_wait);
1975 static void io_file_put(struct io_submit_state *state)
1978 int diff = state->has_refs - state->used_refs;
1981 fput_many(state->file, diff);
1987 * Get as many references to a file as we have IOs left in this submission,
1988 * assuming most submissions are for one file, or at least that each file
1989 * has more than one submission.
1991 static struct file *__io_file_get(struct io_submit_state *state, int fd)
1997 if (state->fd == fd) {
2004 state->file = fget_many(fd, state->ios_left);
2009 state->has_refs = state->ios_left;
2010 state->used_refs = 1;
2016 * If we tracked the file through the SCM inflight mechanism, we could support
2017 * any file. For now, just ensure that anything potentially problematic is done
2020 static bool io_file_supports_async(struct file *file)
2022 umode_t mode = file_inode(file)->i_mode;
2024 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2026 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2032 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2033 bool force_nonblock)
2035 struct io_ring_ctx *ctx = req->ctx;
2036 struct kiocb *kiocb = &req->rw.kiocb;
2040 if (S_ISREG(file_inode(req->file)->i_mode))
2041 req->flags |= REQ_F_ISREG;
2043 kiocb->ki_pos = READ_ONCE(sqe->off);
2044 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2045 req->flags |= REQ_F_CUR_POS;
2046 kiocb->ki_pos = req->file->f_pos;
2048 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2049 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2050 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2054 ioprio = READ_ONCE(sqe->ioprio);
2056 ret = ioprio_check_cap(ioprio);
2060 kiocb->ki_ioprio = ioprio;
2062 kiocb->ki_ioprio = get_current_ioprio();
2064 /* don't allow async punt if RWF_NOWAIT was requested */
2065 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2066 (req->file->f_flags & O_NONBLOCK))
2067 req->flags |= REQ_F_NOWAIT;
2070 kiocb->ki_flags |= IOCB_NOWAIT;
2072 if (ctx->flags & IORING_SETUP_IOPOLL) {
2073 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2074 !kiocb->ki_filp->f_op->iopoll)
2077 kiocb->ki_flags |= IOCB_HIPRI;
2078 kiocb->ki_complete = io_complete_rw_iopoll;
2081 if (kiocb->ki_flags & IOCB_HIPRI)
2083 kiocb->ki_complete = io_complete_rw;
2086 req->rw.addr = READ_ONCE(sqe->addr);
2087 req->rw.len = READ_ONCE(sqe->len);
2088 /* we own ->private, reuse it for the buffer index / buffer ID */
2089 req->rw.kiocb.private = (void *) (unsigned long)
2090 READ_ONCE(sqe->buf_index);
2094 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2100 case -ERESTARTNOINTR:
2101 case -ERESTARTNOHAND:
2102 case -ERESTART_RESTARTBLOCK:
2104 * We can't just restart the syscall, since previously
2105 * submitted sqes may already be in progress. Just fail this
2111 kiocb->ki_complete(kiocb, ret, 0);
2115 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2117 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2119 if (req->flags & REQ_F_CUR_POS)
2120 req->file->f_pos = kiocb->ki_pos;
2121 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2122 io_complete_rw(kiocb, ret, 0);
2124 io_rw_done(kiocb, ret);
2127 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2128 struct iov_iter *iter)
2130 struct io_ring_ctx *ctx = req->ctx;
2131 size_t len = req->rw.len;
2132 struct io_mapped_ubuf *imu;
2133 unsigned index, buf_index;
2137 /* attempt to use fixed buffers without having provided iovecs */
2138 if (unlikely(!ctx->user_bufs))
2141 buf_index = (unsigned long) req->rw.kiocb.private;
2142 if (unlikely(buf_index >= ctx->nr_user_bufs))
2145 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2146 imu = &ctx->user_bufs[index];
2147 buf_addr = req->rw.addr;
2150 if (buf_addr + len < buf_addr)
2152 /* not inside the mapped region */
2153 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2157 * May not be a start of buffer, set size appropriately
2158 * and advance us to the beginning.
2160 offset = buf_addr - imu->ubuf;
2161 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2165 * Don't use iov_iter_advance() here, as it's really slow for
2166 * using the latter parts of a big fixed buffer - it iterates
2167 * over each segment manually. We can cheat a bit here, because
2170 * 1) it's a BVEC iter, we set it up
2171 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2172 * first and last bvec
2174 * So just find our index, and adjust the iterator afterwards.
2175 * If the offset is within the first bvec (or the whole first
2176 * bvec, just use iov_iter_advance(). This makes it easier
2177 * since we can just skip the first segment, which may not
2178 * be PAGE_SIZE aligned.
2180 const struct bio_vec *bvec = imu->bvec;
2182 if (offset <= bvec->bv_len) {
2183 iov_iter_advance(iter, offset);
2185 unsigned long seg_skip;
2187 /* skip first vec */
2188 offset -= bvec->bv_len;
2189 seg_skip = 1 + (offset >> PAGE_SHIFT);
2191 iter->bvec = bvec + seg_skip;
2192 iter->nr_segs -= seg_skip;
2193 iter->count -= bvec->bv_len + offset;
2194 iter->iov_offset = offset & ~PAGE_MASK;
2201 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2204 mutex_unlock(&ctx->uring_lock);
2207 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2210 * "Normal" inline submissions always hold the uring_lock, since we
2211 * grab it from the system call. Same is true for the SQPOLL offload.
2212 * The only exception is when we've detached the request and issue it
2213 * from an async worker thread, grab the lock for that case.
2216 mutex_lock(&ctx->uring_lock);
2219 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2220 int bgid, struct io_buffer *kbuf,
2223 struct io_buffer *head;
2225 if (req->flags & REQ_F_BUFFER_SELECTED)
2228 io_ring_submit_lock(req->ctx, needs_lock);
2230 lockdep_assert_held(&req->ctx->uring_lock);
2232 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2234 if (!list_empty(&head->list)) {
2235 kbuf = list_last_entry(&head->list, struct io_buffer,
2237 list_del(&kbuf->list);
2240 idr_remove(&req->ctx->io_buffer_idr, bgid);
2242 if (*len > kbuf->len)
2245 kbuf = ERR_PTR(-ENOBUFS);
2248 io_ring_submit_unlock(req->ctx, needs_lock);
2253 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2256 struct io_buffer *kbuf;
2259 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2260 bgid = (int) (unsigned long) req->rw.kiocb.private;
2261 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2264 req->rw.addr = (u64) (unsigned long) kbuf;
2265 req->flags |= REQ_F_BUFFER_SELECTED;
2266 return u64_to_user_ptr(kbuf->addr);
2269 #ifdef CONFIG_COMPAT
2270 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2273 struct compat_iovec __user *uiov;
2274 compat_ssize_t clen;
2278 uiov = u64_to_user_ptr(req->rw.addr);
2279 if (!access_ok(uiov, sizeof(*uiov)))
2281 if (__get_user(clen, &uiov->iov_len))
2287 buf = io_rw_buffer_select(req, &len, needs_lock);
2289 return PTR_ERR(buf);
2290 iov[0].iov_base = buf;
2291 iov[0].iov_len = (compat_size_t) len;
2296 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2299 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2303 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2306 len = iov[0].iov_len;
2309 buf = io_rw_buffer_select(req, &len, needs_lock);
2311 return PTR_ERR(buf);
2312 iov[0].iov_base = buf;
2313 iov[0].iov_len = len;
2317 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2320 if (req->flags & REQ_F_BUFFER_SELECTED)
2324 else if (req->rw.len > 1)
2327 #ifdef CONFIG_COMPAT
2328 if (req->ctx->compat)
2329 return io_compat_import(req, iov, needs_lock);
2332 return __io_iov_buffer_select(req, iov, needs_lock);
2335 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2336 struct iovec **iovec, struct iov_iter *iter,
2339 void __user *buf = u64_to_user_ptr(req->rw.addr);
2340 size_t sqe_len = req->rw.len;
2344 opcode = req->opcode;
2345 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2347 return io_import_fixed(req, rw, iter);
2350 /* buffer index only valid with fixed read/write, or buffer select */
2351 if (req->rw.kiocb.private && !(req->flags & REQ_F_BUFFER_SELECT))
2354 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2355 if (req->flags & REQ_F_BUFFER_SELECT) {
2356 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2359 return PTR_ERR(buf);
2361 req->rw.len = sqe_len;
2364 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2366 return ret < 0 ? ret : sqe_len;
2370 struct io_async_rw *iorw = &req->io->rw;
2373 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2374 if (iorw->iov == iorw->fast_iov)
2379 if (req->flags & REQ_F_BUFFER_SELECT) {
2380 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2382 ret = (*iovec)->iov_len;
2383 iov_iter_init(iter, rw, *iovec, 1, ret);
2389 #ifdef CONFIG_COMPAT
2390 if (req->ctx->compat)
2391 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2395 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2399 * For files that don't have ->read_iter() and ->write_iter(), handle them
2400 * by looping over ->read() or ->write() manually.
2402 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2403 struct iov_iter *iter)
2408 * Don't support polled IO through this interface, and we can't
2409 * support non-blocking either. For the latter, this just causes
2410 * the kiocb to be handled from an async context.
2412 if (kiocb->ki_flags & IOCB_HIPRI)
2414 if (kiocb->ki_flags & IOCB_NOWAIT)
2417 while (iov_iter_count(iter)) {
2421 if (!iov_iter_is_bvec(iter)) {
2422 iovec = iov_iter_iovec(iter);
2424 /* fixed buffers import bvec */
2425 iovec.iov_base = kmap(iter->bvec->bv_page)
2427 iovec.iov_len = min(iter->count,
2428 iter->bvec->bv_len - iter->iov_offset);
2432 nr = file->f_op->read(file, iovec.iov_base,
2433 iovec.iov_len, &kiocb->ki_pos);
2435 nr = file->f_op->write(file, iovec.iov_base,
2436 iovec.iov_len, &kiocb->ki_pos);
2439 if (iov_iter_is_bvec(iter))
2440 kunmap(iter->bvec->bv_page);
2448 if (nr != iovec.iov_len)
2450 iov_iter_advance(iter, nr);
2456 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2457 struct iovec *iovec, struct iovec *fast_iov,
2458 struct iov_iter *iter)
2460 req->io->rw.nr_segs = iter->nr_segs;
2461 req->io->rw.size = io_size;
2462 req->io->rw.iov = iovec;
2463 if (!req->io->rw.iov) {
2464 req->io->rw.iov = req->io->rw.fast_iov;
2465 memcpy(req->io->rw.iov, fast_iov,
2466 sizeof(struct iovec) * iter->nr_segs);
2468 req->flags |= REQ_F_NEED_CLEANUP;
2472 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2474 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2475 return req->io == NULL;
2478 static int io_alloc_async_ctx(struct io_kiocb *req)
2480 if (!io_op_defs[req->opcode].async_ctx)
2483 return __io_alloc_async_ctx(req);
2486 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2487 struct iovec *iovec, struct iovec *fast_iov,
2488 struct iov_iter *iter)
2490 if (!io_op_defs[req->opcode].async_ctx)
2493 if (__io_alloc_async_ctx(req))
2496 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2501 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2502 bool force_nonblock)
2504 struct io_async_ctx *io;
2505 struct iov_iter iter;
2508 ret = io_prep_rw(req, sqe, force_nonblock);
2512 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2515 /* either don't need iovec imported or already have it */
2516 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2520 io->rw.iov = io->rw.fast_iov;
2522 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2527 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2531 static int io_read(struct io_kiocb *req, bool force_nonblock)
2533 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2534 struct kiocb *kiocb = &req->rw.kiocb;
2535 struct iov_iter iter;
2537 ssize_t io_size, ret;
2539 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2543 /* Ensure we clear previously set non-block flag */
2544 if (!force_nonblock)
2545 kiocb->ki_flags &= ~IOCB_NOWAIT;
2549 if (req->flags & REQ_F_LINK)
2550 req->result = io_size;
2553 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2554 * we know to async punt it even if it was opened O_NONBLOCK
2556 if (force_nonblock && !io_file_supports_async(req->file))
2559 iov_count = iov_iter_count(&iter);
2560 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2564 if (req->file->f_op->read_iter)
2565 ret2 = call_read_iter(req->file, kiocb, &iter);
2567 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2569 /* Catch -EAGAIN return for forced non-blocking submission */
2570 if (!force_nonblock || ret2 != -EAGAIN) {
2571 kiocb_done(kiocb, ret2);
2574 ret = io_setup_async_rw(req, io_size, iovec,
2575 inline_vecs, &iter);
2578 /* any defer here is final, must blocking retry */
2579 if (!(req->flags & REQ_F_NOWAIT))
2580 req->flags |= REQ_F_MUST_PUNT;
2586 req->flags &= ~REQ_F_NEED_CLEANUP;
2590 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2591 bool force_nonblock)
2593 struct io_async_ctx *io;
2594 struct iov_iter iter;
2597 ret = io_prep_rw(req, sqe, force_nonblock);
2601 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2604 req->fsize = rlimit(RLIMIT_FSIZE);
2606 /* either don't need iovec imported or already have it */
2607 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2611 io->rw.iov = io->rw.fast_iov;
2613 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2618 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2622 static int io_write(struct io_kiocb *req, bool force_nonblock)
2624 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2625 struct kiocb *kiocb = &req->rw.kiocb;
2626 struct iov_iter iter;
2628 ssize_t ret, io_size;
2630 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2634 /* Ensure we clear previously set non-block flag */
2635 if (!force_nonblock)
2636 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2640 if (req->flags & REQ_F_LINK)
2641 req->result = io_size;
2644 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2645 * we know to async punt it even if it was opened O_NONBLOCK
2647 if (force_nonblock && !io_file_supports_async(req->file))
2650 /* file path doesn't support NOWAIT for non-direct_IO */
2651 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2652 (req->flags & REQ_F_ISREG))
2655 iov_count = iov_iter_count(&iter);
2656 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2661 * Open-code file_start_write here to grab freeze protection,
2662 * which will be released by another thread in
2663 * io_complete_rw(). Fool lockdep by telling it the lock got
2664 * released so that it doesn't complain about the held lock when
2665 * we return to userspace.
2667 if (req->flags & REQ_F_ISREG) {
2668 __sb_start_write(file_inode(req->file)->i_sb,
2669 SB_FREEZE_WRITE, true);
2670 __sb_writers_release(file_inode(req->file)->i_sb,
2673 kiocb->ki_flags |= IOCB_WRITE;
2675 if (!force_nonblock)
2676 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2678 if (req->file->f_op->write_iter)
2679 ret2 = call_write_iter(req->file, kiocb, &iter);
2681 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2683 if (!force_nonblock)
2684 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2687 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2688 * retry them without IOCB_NOWAIT.
2690 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2692 if (!force_nonblock || ret2 != -EAGAIN) {
2693 kiocb_done(kiocb, ret2);
2696 ret = io_setup_async_rw(req, io_size, iovec,
2697 inline_vecs, &iter);
2700 /* any defer here is final, must blocking retry */
2701 req->flags |= REQ_F_MUST_PUNT;
2706 req->flags &= ~REQ_F_NEED_CLEANUP;
2711 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2713 struct io_splice* sp = &req->splice;
2714 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2717 if (req->flags & REQ_F_NEED_CLEANUP)
2721 sp->off_in = READ_ONCE(sqe->splice_off_in);
2722 sp->off_out = READ_ONCE(sqe->off);
2723 sp->len = READ_ONCE(sqe->len);
2724 sp->flags = READ_ONCE(sqe->splice_flags);
2726 if (unlikely(sp->flags & ~valid_flags))
2729 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2730 (sp->flags & SPLICE_F_FD_IN_FIXED));
2733 req->flags |= REQ_F_NEED_CLEANUP;
2735 if (!S_ISREG(file_inode(sp->file_in)->i_mode))
2736 req->work.flags |= IO_WQ_WORK_UNBOUND;
2741 static bool io_splice_punt(struct file *file)
2743 if (get_pipe_info(file))
2745 if (!io_file_supports_async(file))
2747 return !(file->f_mode & O_NONBLOCK);
2750 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2752 struct io_splice *sp = &req->splice;
2753 struct file *in = sp->file_in;
2754 struct file *out = sp->file_out;
2755 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2756 loff_t *poff_in, *poff_out;
2759 if (force_nonblock) {
2760 if (io_splice_punt(in) || io_splice_punt(out))
2762 flags |= SPLICE_F_NONBLOCK;
2765 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2766 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2767 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2768 if (force_nonblock && ret == -EAGAIN)
2771 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2772 req->flags &= ~REQ_F_NEED_CLEANUP;
2774 io_cqring_add_event(req, ret);
2776 req_set_fail_links(req);
2782 * IORING_OP_NOP just posts a completion event, nothing else.
2784 static int io_nop(struct io_kiocb *req)
2786 struct io_ring_ctx *ctx = req->ctx;
2788 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2791 io_cqring_add_event(req, 0);
2796 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2798 struct io_ring_ctx *ctx = req->ctx;
2803 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2805 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2808 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2809 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2812 req->sync.off = READ_ONCE(sqe->off);
2813 req->sync.len = READ_ONCE(sqe->len);
2817 static bool io_req_cancelled(struct io_kiocb *req)
2819 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2820 req_set_fail_links(req);
2821 io_cqring_add_event(req, -ECANCELED);
2829 static void __io_fsync(struct io_kiocb *req)
2831 loff_t end = req->sync.off + req->sync.len;
2834 ret = vfs_fsync_range(req->file, req->sync.off,
2835 end > 0 ? end : LLONG_MAX,
2836 req->sync.flags & IORING_FSYNC_DATASYNC);
2838 req_set_fail_links(req);
2839 io_cqring_add_event(req, ret);
2843 static void io_fsync_finish(struct io_wq_work **workptr)
2845 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2847 if (io_req_cancelled(req))
2850 io_steal_work(req, workptr);
2853 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2855 /* fsync always requires a blocking context */
2856 if (force_nonblock) {
2857 req->work.func = io_fsync_finish;
2864 static void __io_fallocate(struct io_kiocb *req)
2868 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2869 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2871 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2873 req_set_fail_links(req);
2874 io_cqring_add_event(req, ret);
2878 static void io_fallocate_finish(struct io_wq_work **workptr)
2880 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2882 if (io_req_cancelled(req))
2884 __io_fallocate(req);
2885 io_steal_work(req, workptr);
2888 static int io_fallocate_prep(struct io_kiocb *req,
2889 const struct io_uring_sqe *sqe)
2891 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2894 req->sync.off = READ_ONCE(sqe->off);
2895 req->sync.len = READ_ONCE(sqe->addr);
2896 req->sync.mode = READ_ONCE(sqe->len);
2897 req->fsize = rlimit(RLIMIT_FSIZE);
2901 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2903 /* fallocate always requiring blocking context */
2904 if (force_nonblock) {
2905 req->work.func = io_fallocate_finish;
2909 __io_fallocate(req);
2913 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2915 const char __user *fname;
2918 if (sqe->ioprio || sqe->buf_index)
2920 if (sqe->flags & IOSQE_FIXED_FILE)
2922 if (req->flags & REQ_F_NEED_CLEANUP)
2925 req->open.dfd = READ_ONCE(sqe->fd);
2926 req->open.how.mode = READ_ONCE(sqe->len);
2927 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2928 req->open.how.flags = READ_ONCE(sqe->open_flags);
2930 req->open.filename = getname(fname);
2931 if (IS_ERR(req->open.filename)) {
2932 ret = PTR_ERR(req->open.filename);
2933 req->open.filename = NULL;
2937 req->flags |= REQ_F_NEED_CLEANUP;
2941 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2943 struct open_how __user *how;
2944 const char __user *fname;
2948 if (sqe->ioprio || sqe->buf_index)
2950 if (sqe->flags & IOSQE_FIXED_FILE)
2952 if (req->flags & REQ_F_NEED_CLEANUP)
2955 req->open.dfd = READ_ONCE(sqe->fd);
2956 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2957 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2958 len = READ_ONCE(sqe->len);
2960 if (len < OPEN_HOW_SIZE_VER0)
2963 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2968 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2969 req->open.how.flags |= O_LARGEFILE;
2971 req->open.filename = getname(fname);
2972 if (IS_ERR(req->open.filename)) {
2973 ret = PTR_ERR(req->open.filename);
2974 req->open.filename = NULL;
2978 req->flags |= REQ_F_NEED_CLEANUP;
2982 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
2984 struct open_flags op;
2991 ret = build_open_flags(&req->open.how, &op);
2995 ret = get_unused_fd_flags(req->open.how.flags);
2999 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3002 ret = PTR_ERR(file);
3004 fsnotify_open(file);
3005 fd_install(ret, file);
3008 putname(req->open.filename);
3009 req->flags &= ~REQ_F_NEED_CLEANUP;
3011 req_set_fail_links(req);
3012 io_cqring_add_event(req, ret);
3017 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3019 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
3020 return io_openat2(req, force_nonblock);
3023 static int io_remove_buffers_prep(struct io_kiocb *req,
3024 const struct io_uring_sqe *sqe)
3026 struct io_provide_buf *p = &req->pbuf;
3029 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3032 tmp = READ_ONCE(sqe->fd);
3033 if (!tmp || tmp > USHRT_MAX)
3036 memset(p, 0, sizeof(*p));
3038 p->bgid = READ_ONCE(sqe->buf_group);
3042 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3043 int bgid, unsigned nbufs)
3047 /* shouldn't happen */
3051 /* the head kbuf is the list itself */
3052 while (!list_empty(&buf->list)) {
3053 struct io_buffer *nxt;
3055 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3056 list_del(&nxt->list);
3063 idr_remove(&ctx->io_buffer_idr, bgid);
3068 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3070 struct io_provide_buf *p = &req->pbuf;
3071 struct io_ring_ctx *ctx = req->ctx;
3072 struct io_buffer *head;
3075 io_ring_submit_lock(ctx, !force_nonblock);
3077 lockdep_assert_held(&ctx->uring_lock);
3080 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3082 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3084 io_ring_submit_lock(ctx, !force_nonblock);
3086 req_set_fail_links(req);
3087 io_cqring_add_event(req, ret);
3092 static int io_provide_buffers_prep(struct io_kiocb *req,
3093 const struct io_uring_sqe *sqe)
3095 struct io_provide_buf *p = &req->pbuf;
3098 if (sqe->ioprio || sqe->rw_flags)
3101 tmp = READ_ONCE(sqe->fd);
3102 if (!tmp || tmp > USHRT_MAX)
3105 p->addr = READ_ONCE(sqe->addr);
3106 p->len = READ_ONCE(sqe->len);
3108 if (!access_ok(u64_to_user_ptr(p->addr), p->len))
3111 p->bgid = READ_ONCE(sqe->buf_group);
3112 tmp = READ_ONCE(sqe->off);
3113 if (tmp > USHRT_MAX)
3119 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3121 struct io_buffer *buf;
3122 u64 addr = pbuf->addr;
3123 int i, bid = pbuf->bid;
3125 for (i = 0; i < pbuf->nbufs; i++) {
3126 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3131 buf->len = pbuf->len;
3136 INIT_LIST_HEAD(&buf->list);
3139 list_add_tail(&buf->list, &(*head)->list);
3143 return i ? i : -ENOMEM;
3146 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3148 struct io_provide_buf *p = &req->pbuf;
3149 struct io_ring_ctx *ctx = req->ctx;
3150 struct io_buffer *head, *list;
3153 io_ring_submit_lock(ctx, !force_nonblock);
3155 lockdep_assert_held(&ctx->uring_lock);
3157 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3159 ret = io_add_buffers(p, &head);
3164 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3167 __io_remove_buffers(ctx, head, p->bgid, -1U);
3172 io_ring_submit_unlock(ctx, !force_nonblock);
3174 req_set_fail_links(req);
3175 io_cqring_add_event(req, ret);
3180 static int io_epoll_ctl_prep(struct io_kiocb *req,
3181 const struct io_uring_sqe *sqe)
3183 #if defined(CONFIG_EPOLL)
3184 if (sqe->ioprio || sqe->buf_index)
3187 req->epoll.epfd = READ_ONCE(sqe->fd);
3188 req->epoll.op = READ_ONCE(sqe->len);
3189 req->epoll.fd = READ_ONCE(sqe->off);
3191 if (ep_op_has_event(req->epoll.op)) {
3192 struct epoll_event __user *ev;
3194 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3195 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3205 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3207 #if defined(CONFIG_EPOLL)
3208 struct io_epoll *ie = &req->epoll;
3211 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3212 if (force_nonblock && ret == -EAGAIN)
3216 req_set_fail_links(req);
3217 io_cqring_add_event(req, ret);
3225 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3227 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3228 if (sqe->ioprio || sqe->buf_index || sqe->off)
3231 req->madvise.addr = READ_ONCE(sqe->addr);
3232 req->madvise.len = READ_ONCE(sqe->len);
3233 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3240 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3242 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3243 struct io_madvise *ma = &req->madvise;
3249 ret = do_madvise(ma->addr, ma->len, ma->advice);
3251 req_set_fail_links(req);
3252 io_cqring_add_event(req, ret);
3260 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3262 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3265 req->fadvise.offset = READ_ONCE(sqe->off);
3266 req->fadvise.len = READ_ONCE(sqe->len);
3267 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3271 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3273 struct io_fadvise *fa = &req->fadvise;
3276 if (force_nonblock) {
3277 switch (fa->advice) {
3278 case POSIX_FADV_NORMAL:
3279 case POSIX_FADV_RANDOM:
3280 case POSIX_FADV_SEQUENTIAL:
3287 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3289 req_set_fail_links(req);
3290 io_cqring_add_event(req, ret);
3295 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3297 const char __user *fname;
3298 unsigned lookup_flags;
3301 if (sqe->ioprio || sqe->buf_index)
3303 if (sqe->flags & IOSQE_FIXED_FILE)
3305 if (req->flags & REQ_F_NEED_CLEANUP)
3308 req->open.dfd = READ_ONCE(sqe->fd);
3309 req->open.mask = READ_ONCE(sqe->len);
3310 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3311 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3312 req->open.how.flags = READ_ONCE(sqe->statx_flags);
3314 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
3317 req->open.filename = getname_flags(fname, lookup_flags, NULL);
3318 if (IS_ERR(req->open.filename)) {
3319 ret = PTR_ERR(req->open.filename);
3320 req->open.filename = NULL;
3324 req->flags |= REQ_F_NEED_CLEANUP;
3328 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3330 struct io_open *ctx = &req->open;
3331 unsigned lookup_flags;
3339 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
3343 /* filename_lookup() drops it, keep a reference */
3344 ctx->filename->refcnt++;
3346 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
3351 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
3353 if (retry_estale(ret, lookup_flags)) {
3354 lookup_flags |= LOOKUP_REVAL;
3358 ret = cp_statx(&stat, ctx->buffer);
3360 putname(ctx->filename);
3361 req->flags &= ~REQ_F_NEED_CLEANUP;
3363 req_set_fail_links(req);
3364 io_cqring_add_event(req, ret);
3369 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3372 * If we queue this for async, it must not be cancellable. That would
3373 * leave the 'file' in an undeterminate state.
3375 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3377 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3378 sqe->rw_flags || sqe->buf_index)
3380 if (sqe->flags & IOSQE_FIXED_FILE)
3383 req->close.fd = READ_ONCE(sqe->fd);
3384 if (req->file->f_op == &io_uring_fops ||
3385 req->close.fd == req->ctx->ring_fd)
3391 /* only called when __close_fd_get_file() is done */
3392 static void __io_close_finish(struct io_kiocb *req)
3396 ret = filp_close(req->close.put_file, req->work.files);
3398 req_set_fail_links(req);
3399 io_cqring_add_event(req, ret);
3400 fput(req->close.put_file);
3404 static void io_close_finish(struct io_wq_work **workptr)
3406 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3408 /* not cancellable, don't do io_req_cancelled() */
3409 __io_close_finish(req);
3410 io_steal_work(req, workptr);
3413 static int io_close(struct io_kiocb *req, bool force_nonblock)
3417 req->close.put_file = NULL;
3418 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
3422 /* if the file has a flush method, be safe and punt to async */
3423 if (req->close.put_file->f_op->flush && force_nonblock) {
3424 /* submission ref will be dropped, take it for async */
3425 refcount_inc(&req->refs);
3427 req->work.func = io_close_finish;
3429 * Do manual async queue here to avoid grabbing files - we don't
3430 * need the files, and it'll cause io_close_finish() to close
3431 * the file again and cause a double CQE entry for this request
3433 io_queue_async_work(req);
3438 * No ->flush(), safely close from here and just punt the
3439 * fput() to async context.
3441 __io_close_finish(req);
3445 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3447 struct io_ring_ctx *ctx = req->ctx;
3452 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3454 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3457 req->sync.off = READ_ONCE(sqe->off);
3458 req->sync.len = READ_ONCE(sqe->len);
3459 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3463 static void __io_sync_file_range(struct io_kiocb *req)
3467 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3470 req_set_fail_links(req);
3471 io_cqring_add_event(req, ret);
3476 static void io_sync_file_range_finish(struct io_wq_work **workptr)
3478 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3479 struct io_kiocb *nxt = NULL;
3481 if (io_req_cancelled(req))
3483 __io_sync_file_range(req);
3484 io_put_req(req); /* put submission ref */
3486 io_wq_assign_next(workptr, nxt);
3489 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3491 /* sync_file_range always requires a blocking context */
3492 if (force_nonblock) {
3493 req->work.func = io_sync_file_range_finish;
3497 __io_sync_file_range(req);
3501 #if defined(CONFIG_NET)
3502 static int io_setup_async_msg(struct io_kiocb *req,
3503 struct io_async_msghdr *kmsg)
3507 if (io_alloc_async_ctx(req)) {
3508 if (kmsg->iov != kmsg->fast_iov)
3512 req->flags |= REQ_F_NEED_CLEANUP;
3513 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3517 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3519 struct io_sr_msg *sr = &req->sr_msg;
3520 struct io_async_ctx *io = req->io;
3523 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3524 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3525 sr->len = READ_ONCE(sqe->len);
3527 #ifdef CONFIG_COMPAT
3528 if (req->ctx->compat)
3529 sr->msg_flags |= MSG_CMSG_COMPAT;
3532 if (!io || req->opcode == IORING_OP_SEND)
3534 /* iovec is already imported */
3535 if (req->flags & REQ_F_NEED_CLEANUP)
3538 io->msg.iov = io->msg.fast_iov;
3539 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3542 req->flags |= REQ_F_NEED_CLEANUP;
3546 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3548 struct io_async_msghdr *kmsg = NULL;
3549 struct socket *sock;
3552 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3555 sock = sock_from_file(req->file, &ret);
3557 struct io_async_ctx io;
3561 kmsg = &req->io->msg;
3562 kmsg->msg.msg_name = &req->io->msg.addr;
3563 /* if iov is set, it's allocated already */
3565 kmsg->iov = kmsg->fast_iov;
3566 kmsg->msg.msg_iter.iov = kmsg->iov;
3568 struct io_sr_msg *sr = &req->sr_msg;
3571 kmsg->msg.msg_name = &io.msg.addr;
3573 io.msg.iov = io.msg.fast_iov;
3574 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3575 sr->msg_flags, &io.msg.iov);
3580 flags = req->sr_msg.msg_flags;
3581 if (flags & MSG_DONTWAIT)
3582 req->flags |= REQ_F_NOWAIT;
3583 else if (force_nonblock)
3584 flags |= MSG_DONTWAIT;
3586 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3587 if (force_nonblock && ret == -EAGAIN)
3588 return io_setup_async_msg(req, kmsg);
3589 if (ret == -ERESTARTSYS)
3593 if (kmsg && kmsg->iov != kmsg->fast_iov)
3595 req->flags &= ~REQ_F_NEED_CLEANUP;
3596 io_cqring_add_event(req, ret);
3598 req_set_fail_links(req);
3603 static int io_send(struct io_kiocb *req, bool force_nonblock)
3605 struct socket *sock;
3608 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3611 sock = sock_from_file(req->file, &ret);
3613 struct io_sr_msg *sr = &req->sr_msg;
3618 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3623 msg.msg_name = NULL;
3624 msg.msg_control = NULL;
3625 msg.msg_controllen = 0;
3626 msg.msg_namelen = 0;
3628 flags = req->sr_msg.msg_flags;
3629 if (flags & MSG_DONTWAIT)
3630 req->flags |= REQ_F_NOWAIT;
3631 else if (force_nonblock)
3632 flags |= MSG_DONTWAIT;
3634 msg.msg_flags = flags;
3635 ret = sock_sendmsg(sock, &msg);
3636 if (force_nonblock && ret == -EAGAIN)
3638 if (ret == -ERESTARTSYS)
3642 io_cqring_add_event(req, ret);
3644 req_set_fail_links(req);
3649 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3651 struct io_sr_msg *sr = &req->sr_msg;
3652 struct iovec __user *uiov;
3656 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3661 if (req->flags & REQ_F_BUFFER_SELECT) {
3664 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3666 sr->len = io->msg.iov[0].iov_len;
3667 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3671 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3672 &io->msg.iov, &io->msg.msg.msg_iter);
3680 #ifdef CONFIG_COMPAT
3681 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3682 struct io_async_ctx *io)
3684 struct compat_msghdr __user *msg_compat;
3685 struct io_sr_msg *sr = &req->sr_msg;
3686 struct compat_iovec __user *uiov;
3691 msg_compat = (struct compat_msghdr __user *) sr->msg;
3692 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3697 uiov = compat_ptr(ptr);
3698 if (req->flags & REQ_F_BUFFER_SELECT) {
3699 compat_ssize_t clen;
3703 if (!access_ok(uiov, sizeof(*uiov)))
3705 if (__get_user(clen, &uiov->iov_len))
3709 sr->len = io->msg.iov[0].iov_len;
3712 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3714 &io->msg.msg.msg_iter);
3723 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3725 io->msg.iov = io->msg.fast_iov;
3727 #ifdef CONFIG_COMPAT
3728 if (req->ctx->compat)
3729 return __io_compat_recvmsg_copy_hdr(req, io);
3732 return __io_recvmsg_copy_hdr(req, io);
3735 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3736 int *cflags, bool needs_lock)
3738 struct io_sr_msg *sr = &req->sr_msg;
3739 struct io_buffer *kbuf;
3741 if (!(req->flags & REQ_F_BUFFER_SELECT))
3744 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3749 req->flags |= REQ_F_BUFFER_SELECTED;
3751 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3752 *cflags |= IORING_CQE_F_BUFFER;
3756 static int io_recvmsg_prep(struct io_kiocb *req,
3757 const struct io_uring_sqe *sqe)
3759 struct io_sr_msg *sr = &req->sr_msg;
3760 struct io_async_ctx *io = req->io;
3763 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3764 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3765 sr->len = READ_ONCE(sqe->len);
3766 sr->bgid = READ_ONCE(sqe->buf_group);
3768 #ifdef CONFIG_COMPAT
3769 if (req->ctx->compat)
3770 sr->msg_flags |= MSG_CMSG_COMPAT;
3773 if (!io || req->opcode == IORING_OP_RECV)
3775 /* iovec is already imported */
3776 if (req->flags & REQ_F_NEED_CLEANUP)
3779 ret = io_recvmsg_copy_hdr(req, io);
3781 req->flags |= REQ_F_NEED_CLEANUP;
3785 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3787 struct io_async_msghdr *kmsg = NULL;
3788 struct socket *sock;
3789 int ret, cflags = 0;
3791 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3794 sock = sock_from_file(req->file, &ret);
3796 struct io_buffer *kbuf;
3797 struct io_async_ctx io;
3801 kmsg = &req->io->msg;
3802 kmsg->msg.msg_name = &req->io->msg.addr;
3803 /* if iov is set, it's allocated already */
3805 kmsg->iov = kmsg->fast_iov;
3806 kmsg->msg.msg_iter.iov = kmsg->iov;
3809 kmsg->msg.msg_name = &io.msg.addr;
3811 ret = io_recvmsg_copy_hdr(req, &io);
3816 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3818 return PTR_ERR(kbuf);
3820 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3821 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3822 1, req->sr_msg.len);
3825 flags = req->sr_msg.msg_flags;
3826 if (flags & MSG_DONTWAIT)
3827 req->flags |= REQ_F_NOWAIT;
3828 else if (force_nonblock)
3829 flags |= MSG_DONTWAIT;
3831 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3832 kmsg->uaddr, flags);
3833 if (force_nonblock && ret == -EAGAIN)
3834 return io_setup_async_msg(req, kmsg);
3835 if (ret == -ERESTARTSYS)
3839 if (kmsg && kmsg->iov != kmsg->fast_iov)
3841 req->flags &= ~REQ_F_NEED_CLEANUP;
3842 __io_cqring_add_event(req, ret, cflags);
3844 req_set_fail_links(req);
3849 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3851 struct io_buffer *kbuf = NULL;
3852 struct socket *sock;
3853 int ret, cflags = 0;
3855 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3858 sock = sock_from_file(req->file, &ret);
3860 struct io_sr_msg *sr = &req->sr_msg;
3861 void __user *buf = sr->buf;
3866 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3868 return PTR_ERR(kbuf);
3870 buf = u64_to_user_ptr(kbuf->addr);
3872 ret = import_single_range(READ, buf, sr->len, &iov,
3879 req->flags |= REQ_F_NEED_CLEANUP;
3880 msg.msg_name = NULL;
3881 msg.msg_control = NULL;
3882 msg.msg_controllen = 0;
3883 msg.msg_namelen = 0;
3884 msg.msg_iocb = NULL;
3887 flags = req->sr_msg.msg_flags;
3888 if (flags & MSG_DONTWAIT)
3889 req->flags |= REQ_F_NOWAIT;
3890 else if (force_nonblock)
3891 flags |= MSG_DONTWAIT;
3893 ret = sock_recvmsg(sock, &msg, flags);
3894 if (force_nonblock && ret == -EAGAIN)
3896 if (ret == -ERESTARTSYS)
3901 req->flags &= ~REQ_F_NEED_CLEANUP;
3902 __io_cqring_add_event(req, ret, cflags);
3904 req_set_fail_links(req);
3909 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3911 struct io_accept *accept = &req->accept;
3913 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3915 if (sqe->ioprio || sqe->len || sqe->buf_index)
3918 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3919 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3920 accept->flags = READ_ONCE(sqe->accept_flags);
3924 static int __io_accept(struct io_kiocb *req, bool force_nonblock)
3926 struct io_accept *accept = &req->accept;
3927 unsigned file_flags;
3930 file_flags = force_nonblock ? O_NONBLOCK : 0;
3931 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3932 accept->addr_len, accept->flags);
3933 if (ret == -EAGAIN && force_nonblock)
3935 if (ret == -ERESTARTSYS)
3938 req_set_fail_links(req);
3939 io_cqring_add_event(req, ret);
3944 static void io_accept_finish(struct io_wq_work **workptr)
3946 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3948 if (io_req_cancelled(req))
3950 __io_accept(req, false);
3951 io_steal_work(req, workptr);
3954 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3958 ret = __io_accept(req, force_nonblock);
3959 if (ret == -EAGAIN && force_nonblock) {
3960 req->work.func = io_accept_finish;
3966 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3968 struct io_connect *conn = &req->connect;
3969 struct io_async_ctx *io = req->io;
3971 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3973 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3976 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3977 conn->addr_len = READ_ONCE(sqe->addr2);
3982 return move_addr_to_kernel(conn->addr, conn->addr_len,
3983 &io->connect.address);
3986 static int io_connect(struct io_kiocb *req, bool force_nonblock)
3988 struct io_async_ctx __io, *io;
3989 unsigned file_flags;
3995 ret = move_addr_to_kernel(req->connect.addr,
3996 req->connect.addr_len,
3997 &__io.connect.address);
4003 file_flags = force_nonblock ? O_NONBLOCK : 0;
4005 ret = __sys_connect_file(req->file, &io->connect.address,
4006 req->connect.addr_len, file_flags);
4007 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4010 if (io_alloc_async_ctx(req)) {
4014 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4017 if (ret == -ERESTARTSYS)
4021 req_set_fail_links(req);
4022 io_cqring_add_event(req, ret);
4026 #else /* !CONFIG_NET */
4027 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4032 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4037 static int io_send(struct io_kiocb *req, bool force_nonblock)
4042 static int io_recvmsg_prep(struct io_kiocb *req,
4043 const struct io_uring_sqe *sqe)
4048 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4053 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4058 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4063 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4068 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4073 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4077 #endif /* CONFIG_NET */
4079 struct io_poll_table {
4080 struct poll_table_struct pt;
4081 struct io_kiocb *req;
4085 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4086 struct wait_queue_head *head)
4088 if (unlikely(poll->head)) {
4089 pt->error = -EINVAL;
4095 add_wait_queue(head, &poll->wait);
4098 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4099 struct poll_table_struct *p)
4101 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4103 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4106 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4107 __poll_t mask, task_work_func_t func)
4109 struct task_struct *tsk;
4111 /* for instances that support it check for an event match first: */
4112 if (mask && !(mask & poll->events))
4115 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4117 list_del_init(&poll->wait.entry);
4121 init_task_work(&req->task_work, func);
4123 * If this fails, then the task is exiting. If that is the case, then
4124 * the exit check will ultimately cancel these work items. Hence we
4125 * don't need to check here and handle it specifically.
4127 task_work_add(tsk, &req->task_work, true);
4128 wake_up_process(tsk);
4132 static void io_async_task_func(struct callback_head *cb)
4134 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4135 struct async_poll *apoll = req->apoll;
4136 struct io_ring_ctx *ctx = req->ctx;
4138 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4140 WARN_ON_ONCE(!list_empty(&req->apoll->poll.wait.entry));
4142 if (hash_hashed(&req->hash_node)) {
4143 spin_lock_irq(&ctx->completion_lock);
4144 hash_del(&req->hash_node);
4145 spin_unlock_irq(&ctx->completion_lock);
4148 /* restore ->work in case we need to retry again */
4149 memcpy(&req->work, &apoll->work, sizeof(req->work));
4151 __set_current_state(TASK_RUNNING);
4152 mutex_lock(&ctx->uring_lock);
4153 __io_queue_sqe(req, NULL);
4154 mutex_unlock(&ctx->uring_lock);
4159 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4162 struct io_kiocb *req = wait->private;
4163 struct io_poll_iocb *poll = &req->apoll->poll;
4165 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4168 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4171 static void io_poll_req_insert(struct io_kiocb *req)
4173 struct io_ring_ctx *ctx = req->ctx;
4174 struct hlist_head *list;
4176 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4177 hlist_add_head(&req->hash_node, list);
4180 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4181 struct io_poll_iocb *poll,
4182 struct io_poll_table *ipt, __poll_t mask,
4183 wait_queue_func_t wake_func)
4184 __acquires(&ctx->completion_lock)
4186 struct io_ring_ctx *ctx = req->ctx;
4187 bool cancel = false;
4189 poll->file = req->file;
4191 poll->done = poll->canceled = false;
4192 poll->events = mask;
4194 ipt->pt._key = mask;
4196 ipt->error = -EINVAL;
4198 INIT_LIST_HEAD(&poll->wait.entry);
4199 init_waitqueue_func_entry(&poll->wait, wake_func);
4200 poll->wait.private = req;
4202 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4204 spin_lock_irq(&ctx->completion_lock);
4205 if (likely(poll->head)) {
4206 spin_lock(&poll->head->lock);
4207 if (unlikely(list_empty(&poll->wait.entry))) {
4213 if (mask || ipt->error)
4214 list_del_init(&poll->wait.entry);
4216 WRITE_ONCE(poll->canceled, true);
4217 else if (!poll->done) /* actually waiting for an event */
4218 io_poll_req_insert(req);
4219 spin_unlock(&poll->head->lock);
4225 static bool io_arm_poll_handler(struct io_kiocb *req)
4227 const struct io_op_def *def = &io_op_defs[req->opcode];
4228 struct io_ring_ctx *ctx = req->ctx;
4229 struct async_poll *apoll;
4230 struct io_poll_table ipt;
4233 if (!req->file || !file_can_poll(req->file))
4235 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4237 if (!def->pollin && !def->pollout)
4240 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4241 if (unlikely(!apoll))
4244 req->flags |= REQ_F_POLLED;
4245 memcpy(&apoll->work, &req->work, sizeof(req->work));
4248 * Don't need a reference here, as we're adding it to the task
4249 * task_works list. If the task exits, the list is pruned.
4251 req->task = current;
4253 INIT_HLIST_NODE(&req->hash_node);
4257 mask |= POLLIN | POLLRDNORM;
4259 mask |= POLLOUT | POLLWRNORM;
4260 mask |= POLLERR | POLLPRI;
4262 ipt.pt._qproc = io_async_queue_proc;
4264 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4268 apoll->poll.done = true;
4269 spin_unlock_irq(&ctx->completion_lock);
4270 memcpy(&req->work, &apoll->work, sizeof(req->work));
4274 spin_unlock_irq(&ctx->completion_lock);
4275 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4276 apoll->poll.events);
4280 static bool __io_poll_remove_one(struct io_kiocb *req,
4281 struct io_poll_iocb *poll)
4283 bool do_complete = false;
4285 spin_lock(&poll->head->lock);
4286 WRITE_ONCE(poll->canceled, true);
4287 if (!list_empty(&poll->wait.entry)) {
4288 list_del_init(&poll->wait.entry);
4291 spin_unlock(&poll->head->lock);
4295 static bool io_poll_remove_one(struct io_kiocb *req)
4299 if (req->opcode == IORING_OP_POLL_ADD) {
4300 do_complete = __io_poll_remove_one(req, &req->poll);
4302 /* non-poll requests have submit ref still */
4303 do_complete = __io_poll_remove_one(req, &req->apoll->poll);
4308 hash_del(&req->hash_node);
4311 io_cqring_fill_event(req, -ECANCELED);
4312 io_commit_cqring(req->ctx);
4313 req->flags |= REQ_F_COMP_LOCKED;
4320 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4322 struct hlist_node *tmp;
4323 struct io_kiocb *req;
4326 spin_lock_irq(&ctx->completion_lock);
4327 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4328 struct hlist_head *list;
4330 list = &ctx->cancel_hash[i];
4331 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4332 io_poll_remove_one(req);
4334 spin_unlock_irq(&ctx->completion_lock);
4336 io_cqring_ev_posted(ctx);
4339 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4341 struct hlist_head *list;
4342 struct io_kiocb *req;
4344 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4345 hlist_for_each_entry(req, list, hash_node) {
4346 if (sqe_addr != req->user_data)
4348 if (io_poll_remove_one(req))
4356 static int io_poll_remove_prep(struct io_kiocb *req,
4357 const struct io_uring_sqe *sqe)
4359 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4361 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4365 req->poll.addr = READ_ONCE(sqe->addr);
4370 * Find a running poll command that matches one specified in sqe->addr,
4371 * and remove it if found.
4373 static int io_poll_remove(struct io_kiocb *req)
4375 struct io_ring_ctx *ctx = req->ctx;
4379 addr = req->poll.addr;
4380 spin_lock_irq(&ctx->completion_lock);
4381 ret = io_poll_cancel(ctx, addr);
4382 spin_unlock_irq(&ctx->completion_lock);
4384 io_cqring_add_event(req, ret);
4386 req_set_fail_links(req);
4391 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4393 struct io_ring_ctx *ctx = req->ctx;
4395 req->poll.done = true;
4396 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4397 io_commit_cqring(ctx);
4400 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4402 struct io_ring_ctx *ctx = req->ctx;
4404 spin_lock_irq(&ctx->completion_lock);
4405 hash_del(&req->hash_node);
4406 io_poll_complete(req, req->result, 0);
4407 req->flags |= REQ_F_COMP_LOCKED;
4408 io_put_req_find_next(req, nxt);
4409 spin_unlock_irq(&ctx->completion_lock);
4411 io_cqring_ev_posted(ctx);
4414 static void io_poll_task_func(struct callback_head *cb)
4416 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4417 struct io_kiocb *nxt = NULL;
4419 io_poll_task_handler(req, &nxt);
4421 struct io_ring_ctx *ctx = nxt->ctx;
4423 mutex_lock(&ctx->uring_lock);
4424 __io_queue_sqe(nxt, NULL);
4425 mutex_unlock(&ctx->uring_lock);
4429 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4432 struct io_kiocb *req = wait->private;
4433 struct io_poll_iocb *poll = &req->poll;
4435 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4438 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4439 struct poll_table_struct *p)
4441 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4443 __io_queue_proc(&pt->req->poll, pt, head);
4446 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4448 struct io_poll_iocb *poll = &req->poll;
4451 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4453 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4458 events = READ_ONCE(sqe->poll_events);
4459 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4462 * Don't need a reference here, as we're adding it to the task
4463 * task_works list. If the task exits, the list is pruned.
4465 req->task = current;
4469 static int io_poll_add(struct io_kiocb *req)
4471 struct io_poll_iocb *poll = &req->poll;
4472 struct io_ring_ctx *ctx = req->ctx;
4473 struct io_poll_table ipt;
4476 INIT_HLIST_NODE(&req->hash_node);
4477 INIT_LIST_HEAD(&req->list);
4478 ipt.pt._qproc = io_poll_queue_proc;
4480 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4483 if (mask) { /* no async, we'd stolen it */
4485 io_poll_complete(req, mask, 0);
4487 spin_unlock_irq(&ctx->completion_lock);
4490 io_cqring_ev_posted(ctx);
4496 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4498 struct io_timeout_data *data = container_of(timer,
4499 struct io_timeout_data, timer);
4500 struct io_kiocb *req = data->req;
4501 struct io_ring_ctx *ctx = req->ctx;
4502 unsigned long flags;
4504 atomic_inc(&ctx->cq_timeouts);
4506 spin_lock_irqsave(&ctx->completion_lock, flags);
4508 * We could be racing with timeout deletion. If the list is empty,
4509 * then timeout lookup already found it and will be handling it.
4511 if (!list_empty(&req->list)) {
4512 struct io_kiocb *prev;
4515 * Adjust the reqs sequence before the current one because it
4516 * will consume a slot in the cq_ring and the cq_tail
4517 * pointer will be increased, otherwise other timeout reqs may
4518 * return in advance without waiting for enough wait_nr.
4521 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
4523 list_del_init(&req->list);
4526 io_cqring_fill_event(req, -ETIME);
4527 io_commit_cqring(ctx);
4528 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4530 io_cqring_ev_posted(ctx);
4531 req_set_fail_links(req);
4533 return HRTIMER_NORESTART;
4536 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4538 struct io_kiocb *req;
4541 list_for_each_entry(req, &ctx->timeout_list, list) {
4542 if (user_data == req->user_data) {
4543 list_del_init(&req->list);
4552 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4556 req_set_fail_links(req);
4557 io_cqring_fill_event(req, -ECANCELED);
4562 static int io_timeout_remove_prep(struct io_kiocb *req,
4563 const struct io_uring_sqe *sqe)
4565 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4567 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4570 req->timeout.addr = READ_ONCE(sqe->addr);
4571 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4572 if (req->timeout.flags)
4579 * Remove or update an existing timeout command
4581 static int io_timeout_remove(struct io_kiocb *req)
4583 struct io_ring_ctx *ctx = req->ctx;
4586 spin_lock_irq(&ctx->completion_lock);
4587 ret = io_timeout_cancel(ctx, req->timeout.addr);
4589 io_cqring_fill_event(req, ret);
4590 io_commit_cqring(ctx);
4591 spin_unlock_irq(&ctx->completion_lock);
4592 io_cqring_ev_posted(ctx);
4594 req_set_fail_links(req);
4599 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4600 bool is_timeout_link)
4602 struct io_timeout_data *data;
4605 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4607 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4609 if (sqe->off && is_timeout_link)
4611 flags = READ_ONCE(sqe->timeout_flags);
4612 if (flags & ~IORING_TIMEOUT_ABS)
4615 req->timeout.count = READ_ONCE(sqe->off);
4617 if (!req->io && io_alloc_async_ctx(req))
4620 data = &req->io->timeout;
4622 req->flags |= REQ_F_TIMEOUT;
4624 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4627 if (flags & IORING_TIMEOUT_ABS)
4628 data->mode = HRTIMER_MODE_ABS;
4630 data->mode = HRTIMER_MODE_REL;
4632 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4636 static int io_timeout(struct io_kiocb *req)
4639 struct io_ring_ctx *ctx = req->ctx;
4640 struct io_timeout_data *data;
4641 struct list_head *entry;
4644 data = &req->io->timeout;
4647 * sqe->off holds how many events that need to occur for this
4648 * timeout event to be satisfied. If it isn't set, then this is
4649 * a pure timeout request, sequence isn't used.
4651 count = req->timeout.count;
4653 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4654 spin_lock_irq(&ctx->completion_lock);
4655 entry = ctx->timeout_list.prev;
4659 req->sequence = ctx->cached_sq_head + count - 1;
4660 data->seq_offset = count;
4663 * Insertion sort, ensuring the first entry in the list is always
4664 * the one we need first.
4666 spin_lock_irq(&ctx->completion_lock);
4667 list_for_each_prev(entry, &ctx->timeout_list) {
4668 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4669 unsigned nxt_sq_head;
4670 long long tmp, tmp_nxt;
4671 u32 nxt_offset = nxt->io->timeout.seq_offset;
4673 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4677 * Since cached_sq_head + count - 1 can overflow, use type long
4680 tmp = (long long)ctx->cached_sq_head + count - 1;
4681 nxt_sq_head = nxt->sequence - nxt_offset + 1;
4682 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
4685 * cached_sq_head may overflow, and it will never overflow twice
4686 * once there is some timeout req still be valid.
4688 if (ctx->cached_sq_head < nxt_sq_head)
4695 * Sequence of reqs after the insert one and itself should
4696 * be adjusted because each timeout req consumes a slot.
4701 req->sequence -= span;
4703 list_add(&req->list, entry);
4704 data->timer.function = io_timeout_fn;
4705 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4706 spin_unlock_irq(&ctx->completion_lock);
4710 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4712 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4714 return req->user_data == (unsigned long) data;
4717 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4719 enum io_wq_cancel cancel_ret;
4722 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4723 switch (cancel_ret) {
4724 case IO_WQ_CANCEL_OK:
4727 case IO_WQ_CANCEL_RUNNING:
4730 case IO_WQ_CANCEL_NOTFOUND:
4738 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4739 struct io_kiocb *req, __u64 sqe_addr,
4742 unsigned long flags;
4745 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4746 if (ret != -ENOENT) {
4747 spin_lock_irqsave(&ctx->completion_lock, flags);
4751 spin_lock_irqsave(&ctx->completion_lock, flags);
4752 ret = io_timeout_cancel(ctx, sqe_addr);
4755 ret = io_poll_cancel(ctx, sqe_addr);
4759 io_cqring_fill_event(req, ret);
4760 io_commit_cqring(ctx);
4761 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4762 io_cqring_ev_posted(ctx);
4765 req_set_fail_links(req);
4769 static int io_async_cancel_prep(struct io_kiocb *req,
4770 const struct io_uring_sqe *sqe)
4772 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4774 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4778 req->cancel.addr = READ_ONCE(sqe->addr);
4782 static int io_async_cancel(struct io_kiocb *req)
4784 struct io_ring_ctx *ctx = req->ctx;
4786 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4790 static int io_files_update_prep(struct io_kiocb *req,
4791 const struct io_uring_sqe *sqe)
4793 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4796 req->files_update.offset = READ_ONCE(sqe->off);
4797 req->files_update.nr_args = READ_ONCE(sqe->len);
4798 if (!req->files_update.nr_args)
4800 req->files_update.arg = READ_ONCE(sqe->addr);
4804 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4806 struct io_ring_ctx *ctx = req->ctx;
4807 struct io_uring_files_update up;
4813 up.offset = req->files_update.offset;
4814 up.fds = req->files_update.arg;
4816 mutex_lock(&ctx->uring_lock);
4817 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4818 mutex_unlock(&ctx->uring_lock);
4821 req_set_fail_links(req);
4822 io_cqring_add_event(req, ret);
4827 static int io_req_defer_prep(struct io_kiocb *req,
4828 const struct io_uring_sqe *sqe)
4832 if (io_op_defs[req->opcode].file_table) {
4833 ret = io_grab_files(req);
4838 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4840 switch (req->opcode) {
4843 case IORING_OP_READV:
4844 case IORING_OP_READ_FIXED:
4845 case IORING_OP_READ:
4846 ret = io_read_prep(req, sqe, true);
4848 case IORING_OP_WRITEV:
4849 case IORING_OP_WRITE_FIXED:
4850 case IORING_OP_WRITE:
4851 ret = io_write_prep(req, sqe, true);
4853 case IORING_OP_POLL_ADD:
4854 ret = io_poll_add_prep(req, sqe);
4856 case IORING_OP_POLL_REMOVE:
4857 ret = io_poll_remove_prep(req, sqe);
4859 case IORING_OP_FSYNC:
4860 ret = io_prep_fsync(req, sqe);
4862 case IORING_OP_SYNC_FILE_RANGE:
4863 ret = io_prep_sfr(req, sqe);
4865 case IORING_OP_SENDMSG:
4866 case IORING_OP_SEND:
4867 ret = io_sendmsg_prep(req, sqe);
4869 case IORING_OP_RECVMSG:
4870 case IORING_OP_RECV:
4871 ret = io_recvmsg_prep(req, sqe);
4873 case IORING_OP_CONNECT:
4874 ret = io_connect_prep(req, sqe);
4876 case IORING_OP_TIMEOUT:
4877 ret = io_timeout_prep(req, sqe, false);
4879 case IORING_OP_TIMEOUT_REMOVE:
4880 ret = io_timeout_remove_prep(req, sqe);
4882 case IORING_OP_ASYNC_CANCEL:
4883 ret = io_async_cancel_prep(req, sqe);
4885 case IORING_OP_LINK_TIMEOUT:
4886 ret = io_timeout_prep(req, sqe, true);
4888 case IORING_OP_ACCEPT:
4889 ret = io_accept_prep(req, sqe);
4891 case IORING_OP_FALLOCATE:
4892 ret = io_fallocate_prep(req, sqe);
4894 case IORING_OP_OPENAT:
4895 ret = io_openat_prep(req, sqe);
4897 case IORING_OP_CLOSE:
4898 ret = io_close_prep(req, sqe);
4900 case IORING_OP_FILES_UPDATE:
4901 ret = io_files_update_prep(req, sqe);
4903 case IORING_OP_STATX:
4904 ret = io_statx_prep(req, sqe);
4906 case IORING_OP_FADVISE:
4907 ret = io_fadvise_prep(req, sqe);
4909 case IORING_OP_MADVISE:
4910 ret = io_madvise_prep(req, sqe);
4912 case IORING_OP_OPENAT2:
4913 ret = io_openat2_prep(req, sqe);
4915 case IORING_OP_EPOLL_CTL:
4916 ret = io_epoll_ctl_prep(req, sqe);
4918 case IORING_OP_SPLICE:
4919 ret = io_splice_prep(req, sqe);
4921 case IORING_OP_PROVIDE_BUFFERS:
4922 ret = io_provide_buffers_prep(req, sqe);
4924 case IORING_OP_REMOVE_BUFFERS:
4925 ret = io_remove_buffers_prep(req, sqe);
4928 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
4937 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4939 struct io_ring_ctx *ctx = req->ctx;
4942 /* Still need defer if there is pending req in defer list. */
4943 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
4946 if (!req->io && io_alloc_async_ctx(req))
4949 ret = io_req_defer_prep(req, sqe);
4953 spin_lock_irq(&ctx->completion_lock);
4954 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
4955 spin_unlock_irq(&ctx->completion_lock);
4959 trace_io_uring_defer(ctx, req, req->user_data);
4960 list_add_tail(&req->list, &ctx->defer_list);
4961 spin_unlock_irq(&ctx->completion_lock);
4962 return -EIOCBQUEUED;
4965 static void io_cleanup_req(struct io_kiocb *req)
4967 struct io_async_ctx *io = req->io;
4969 switch (req->opcode) {
4970 case IORING_OP_READV:
4971 case IORING_OP_READ_FIXED:
4972 case IORING_OP_READ:
4973 if (req->flags & REQ_F_BUFFER_SELECTED)
4974 kfree((void *)(unsigned long)req->rw.addr);
4976 case IORING_OP_WRITEV:
4977 case IORING_OP_WRITE_FIXED:
4978 case IORING_OP_WRITE:
4979 if (io->rw.iov != io->rw.fast_iov)
4982 case IORING_OP_RECVMSG:
4983 if (req->flags & REQ_F_BUFFER_SELECTED)
4984 kfree(req->sr_msg.kbuf);
4986 case IORING_OP_SENDMSG:
4987 if (io->msg.iov != io->msg.fast_iov)
4990 case IORING_OP_RECV:
4991 if (req->flags & REQ_F_BUFFER_SELECTED)
4992 kfree(req->sr_msg.kbuf);
4994 case IORING_OP_OPENAT:
4995 case IORING_OP_OPENAT2:
4996 case IORING_OP_STATX:
4997 putname(req->open.filename);
4999 case IORING_OP_SPLICE:
5000 io_put_file(req, req->splice.file_in,
5001 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5005 req->flags &= ~REQ_F_NEED_CLEANUP;
5008 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5009 bool force_nonblock)
5011 struct io_ring_ctx *ctx = req->ctx;
5014 switch (req->opcode) {
5018 case IORING_OP_READV:
5019 case IORING_OP_READ_FIXED:
5020 case IORING_OP_READ:
5022 ret = io_read_prep(req, sqe, force_nonblock);
5026 ret = io_read(req, force_nonblock);
5028 case IORING_OP_WRITEV:
5029 case IORING_OP_WRITE_FIXED:
5030 case IORING_OP_WRITE:
5032 ret = io_write_prep(req, sqe, force_nonblock);
5036 ret = io_write(req, force_nonblock);
5038 case IORING_OP_FSYNC:
5040 ret = io_prep_fsync(req, sqe);
5044 ret = io_fsync(req, force_nonblock);
5046 case IORING_OP_POLL_ADD:
5048 ret = io_poll_add_prep(req, sqe);
5052 ret = io_poll_add(req);
5054 case IORING_OP_POLL_REMOVE:
5056 ret = io_poll_remove_prep(req, sqe);
5060 ret = io_poll_remove(req);
5062 case IORING_OP_SYNC_FILE_RANGE:
5064 ret = io_prep_sfr(req, sqe);
5068 ret = io_sync_file_range(req, force_nonblock);
5070 case IORING_OP_SENDMSG:
5071 case IORING_OP_SEND:
5073 ret = io_sendmsg_prep(req, sqe);
5077 if (req->opcode == IORING_OP_SENDMSG)
5078 ret = io_sendmsg(req, force_nonblock);
5080 ret = io_send(req, force_nonblock);
5082 case IORING_OP_RECVMSG:
5083 case IORING_OP_RECV:
5085 ret = io_recvmsg_prep(req, sqe);
5089 if (req->opcode == IORING_OP_RECVMSG)
5090 ret = io_recvmsg(req, force_nonblock);
5092 ret = io_recv(req, force_nonblock);
5094 case IORING_OP_TIMEOUT:
5096 ret = io_timeout_prep(req, sqe, false);
5100 ret = io_timeout(req);
5102 case IORING_OP_TIMEOUT_REMOVE:
5104 ret = io_timeout_remove_prep(req, sqe);
5108 ret = io_timeout_remove(req);
5110 case IORING_OP_ACCEPT:
5112 ret = io_accept_prep(req, sqe);
5116 ret = io_accept(req, force_nonblock);
5118 case IORING_OP_CONNECT:
5120 ret = io_connect_prep(req, sqe);
5124 ret = io_connect(req, force_nonblock);
5126 case IORING_OP_ASYNC_CANCEL:
5128 ret = io_async_cancel_prep(req, sqe);
5132 ret = io_async_cancel(req);
5134 case IORING_OP_FALLOCATE:
5136 ret = io_fallocate_prep(req, sqe);
5140 ret = io_fallocate(req, force_nonblock);
5142 case IORING_OP_OPENAT:
5144 ret = io_openat_prep(req, sqe);
5148 ret = io_openat(req, force_nonblock);
5150 case IORING_OP_CLOSE:
5152 ret = io_close_prep(req, sqe);
5156 ret = io_close(req, force_nonblock);
5158 case IORING_OP_FILES_UPDATE:
5160 ret = io_files_update_prep(req, sqe);
5164 ret = io_files_update(req, force_nonblock);
5166 case IORING_OP_STATX:
5168 ret = io_statx_prep(req, sqe);
5172 ret = io_statx(req, force_nonblock);
5174 case IORING_OP_FADVISE:
5176 ret = io_fadvise_prep(req, sqe);
5180 ret = io_fadvise(req, force_nonblock);
5182 case IORING_OP_MADVISE:
5184 ret = io_madvise_prep(req, sqe);
5188 ret = io_madvise(req, force_nonblock);
5190 case IORING_OP_OPENAT2:
5192 ret = io_openat2_prep(req, sqe);
5196 ret = io_openat2(req, force_nonblock);
5198 case IORING_OP_EPOLL_CTL:
5200 ret = io_epoll_ctl_prep(req, sqe);
5204 ret = io_epoll_ctl(req, force_nonblock);
5206 case IORING_OP_SPLICE:
5208 ret = io_splice_prep(req, sqe);
5212 ret = io_splice(req, force_nonblock);
5214 case IORING_OP_PROVIDE_BUFFERS:
5216 ret = io_provide_buffers_prep(req, sqe);
5220 ret = io_provide_buffers(req, force_nonblock);
5222 case IORING_OP_REMOVE_BUFFERS:
5224 ret = io_remove_buffers_prep(req, sqe);
5228 ret = io_remove_buffers(req, force_nonblock);
5238 if (ctx->flags & IORING_SETUP_IOPOLL) {
5239 const bool in_async = io_wq_current_is_worker();
5241 if (req->result == -EAGAIN)
5244 /* workqueue context doesn't hold uring_lock, grab it now */
5246 mutex_lock(&ctx->uring_lock);
5248 io_iopoll_req_issued(req);
5251 mutex_unlock(&ctx->uring_lock);
5257 static void io_wq_submit_work(struct io_wq_work **workptr)
5259 struct io_wq_work *work = *workptr;
5260 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5263 /* if NO_CANCEL is set, we must still run the work */
5264 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5265 IO_WQ_WORK_CANCEL) {
5271 ret = io_issue_sqe(req, NULL, false);
5273 * We can get EAGAIN for polled IO even though we're
5274 * forcing a sync submission from here, since we can't
5275 * wait for request slots on the block side.
5284 req_set_fail_links(req);
5285 io_cqring_add_event(req, ret);
5289 io_steal_work(req, workptr);
5292 static int io_req_needs_file(struct io_kiocb *req, int fd)
5294 if (!io_op_defs[req->opcode].needs_file)
5296 if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg)
5301 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5304 struct fixed_file_table *table;
5306 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5307 return table->files[index & IORING_FILE_TABLE_MASK];;
5310 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5311 int fd, struct file **out_file, bool fixed)
5313 struct io_ring_ctx *ctx = req->ctx;
5317 if (unlikely(!ctx->file_data ||
5318 (unsigned) fd >= ctx->nr_user_files))
5320 fd = array_index_nospec(fd, ctx->nr_user_files);
5321 file = io_file_from_index(ctx, fd);
5324 percpu_ref_get(&ctx->file_data->refs);
5326 trace_io_uring_file_get(ctx, fd);
5327 file = __io_file_get(state, fd);
5328 if (unlikely(!file))
5336 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5337 const struct io_uring_sqe *sqe)
5343 flags = READ_ONCE(sqe->flags);
5344 fd = READ_ONCE(sqe->fd);
5346 if (!io_req_needs_file(req, fd))
5349 fixed = (flags & IOSQE_FIXED_FILE);
5350 if (unlikely(!fixed && req->needs_fixed_file))
5353 return io_file_get(state, req, fd, &req->file, fixed);
5356 static int io_grab_files(struct io_kiocb *req)
5359 struct io_ring_ctx *ctx = req->ctx;
5361 if (req->work.files)
5363 if (!ctx->ring_file)
5367 spin_lock_irq(&ctx->inflight_lock);
5369 * We use the f_ops->flush() handler to ensure that we can flush
5370 * out work accessing these files if the fd is closed. Check if
5371 * the fd has changed since we started down this path, and disallow
5372 * this operation if it has.
5374 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5375 list_add(&req->inflight_entry, &ctx->inflight_list);
5376 req->flags |= REQ_F_INFLIGHT;
5377 req->work.files = current->files;
5380 spin_unlock_irq(&ctx->inflight_lock);
5386 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5388 struct io_timeout_data *data = container_of(timer,
5389 struct io_timeout_data, timer);
5390 struct io_kiocb *req = data->req;
5391 struct io_ring_ctx *ctx = req->ctx;
5392 struct io_kiocb *prev = NULL;
5393 unsigned long flags;
5395 spin_lock_irqsave(&ctx->completion_lock, flags);
5398 * We don't expect the list to be empty, that will only happen if we
5399 * race with the completion of the linked work.
5401 if (!list_empty(&req->link_list)) {
5402 prev = list_entry(req->link_list.prev, struct io_kiocb,
5404 if (refcount_inc_not_zero(&prev->refs)) {
5405 list_del_init(&req->link_list);
5406 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5411 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5414 req_set_fail_links(prev);
5415 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5418 io_cqring_add_event(req, -ETIME);
5421 return HRTIMER_NORESTART;
5424 static void io_queue_linked_timeout(struct io_kiocb *req)
5426 struct io_ring_ctx *ctx = req->ctx;
5429 * If the list is now empty, then our linked request finished before
5430 * we got a chance to setup the timer
5432 spin_lock_irq(&ctx->completion_lock);
5433 if (!list_empty(&req->link_list)) {
5434 struct io_timeout_data *data = &req->io->timeout;
5436 data->timer.function = io_link_timeout_fn;
5437 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5440 spin_unlock_irq(&ctx->completion_lock);
5442 /* drop submission reference */
5446 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5448 struct io_kiocb *nxt;
5450 if (!(req->flags & REQ_F_LINK))
5452 /* for polled retry, if flag is set, we already went through here */
5453 if (req->flags & REQ_F_POLLED)
5456 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5458 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5461 req->flags |= REQ_F_LINK_TIMEOUT;
5465 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5467 struct io_kiocb *linked_timeout;
5468 struct io_kiocb *nxt;
5469 const struct cred *old_creds = NULL;
5473 linked_timeout = io_prep_linked_timeout(req);
5475 if (req->work.creds && req->work.creds != current_cred()) {
5477 revert_creds(old_creds);
5478 if (old_creds == req->work.creds)
5479 old_creds = NULL; /* restored original creds */
5481 old_creds = override_creds(req->work.creds);
5484 ret = io_issue_sqe(req, sqe, true);
5487 * We async punt it if the file wasn't marked NOWAIT, or if the file
5488 * doesn't support non-blocking read/write attempts
5490 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5491 (req->flags & REQ_F_MUST_PUNT))) {
5492 if (io_arm_poll_handler(req)) {
5494 io_queue_linked_timeout(linked_timeout);
5498 if (io_op_defs[req->opcode].file_table) {
5499 ret = io_grab_files(req);
5505 * Queued up for async execution, worker will release
5506 * submit reference when the iocb is actually submitted.
5508 io_queue_async_work(req);
5514 /* drop submission reference */
5515 io_put_req_find_next(req, &nxt);
5517 if (linked_timeout) {
5519 io_queue_linked_timeout(linked_timeout);
5521 io_put_req(linked_timeout);
5524 /* and drop final reference, if we failed */
5526 io_cqring_add_event(req, ret);
5527 req_set_fail_links(req);
5533 if (req->flags & REQ_F_FORCE_ASYNC)
5539 revert_creds(old_creds);
5542 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5546 ret = io_req_defer(req, sqe);
5548 if (ret != -EIOCBQUEUED) {
5550 io_cqring_add_event(req, ret);
5551 req_set_fail_links(req);
5552 io_double_put_req(req);
5554 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5555 ret = io_req_defer_prep(req, sqe);
5556 if (unlikely(ret < 0))
5559 * Never try inline submit of IOSQE_ASYNC is set, go straight
5560 * to async execution.
5562 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5563 io_queue_async_work(req);
5565 __io_queue_sqe(req, sqe);
5569 static inline void io_queue_link_head(struct io_kiocb *req)
5571 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5572 io_cqring_add_event(req, -ECANCELED);
5573 io_double_put_req(req);
5575 io_queue_sqe(req, NULL);
5578 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5579 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5580 IOSQE_BUFFER_SELECT)
5582 static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5583 struct io_submit_state *state, struct io_kiocb **link)
5585 struct io_ring_ctx *ctx = req->ctx;
5586 unsigned int sqe_flags;
5589 sqe_flags = READ_ONCE(sqe->flags);
5591 /* enforce forwards compatibility on users */
5592 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
5597 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5598 !io_op_defs[req->opcode].buffer_select) {
5603 id = READ_ONCE(sqe->personality);
5605 req->work.creds = idr_find(&ctx->personality_idr, id);
5606 if (unlikely(!req->work.creds)) {
5610 get_cred(req->work.creds);
5613 /* same numerical values with corresponding REQ_F_*, safe to copy */
5614 req->flags |= sqe_flags & (IOSQE_IO_DRAIN | IOSQE_IO_HARDLINK |
5615 IOSQE_ASYNC | IOSQE_FIXED_FILE |
5616 IOSQE_BUFFER_SELECT);
5618 ret = io_req_set_file(state, req, sqe);
5619 if (unlikely(ret)) {
5621 io_cqring_add_event(req, ret);
5622 io_double_put_req(req);
5627 * If we already have a head request, queue this one for async
5628 * submittal once the head completes. If we don't have a head but
5629 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5630 * submitted sync once the chain is complete. If none of those
5631 * conditions are true (normal request), then just queue it.
5634 struct io_kiocb *head = *link;
5637 * Taking sequential execution of a link, draining both sides
5638 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5639 * requests in the link. So, it drains the head and the
5640 * next after the link request. The last one is done via
5641 * drain_next flag to persist the effect across calls.
5643 if (sqe_flags & IOSQE_IO_DRAIN) {
5644 head->flags |= REQ_F_IO_DRAIN;
5645 ctx->drain_next = 1;
5647 if (io_alloc_async_ctx(req)) {
5652 ret = io_req_defer_prep(req, sqe);
5654 /* fail even hard links since we don't submit */
5655 head->flags |= REQ_F_FAIL_LINK;
5658 trace_io_uring_link(ctx, req, head);
5659 list_add_tail(&req->link_list, &head->link_list);
5661 /* last request of a link, enqueue the link */
5662 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) {
5663 io_queue_link_head(head);
5667 if (unlikely(ctx->drain_next)) {
5668 req->flags |= REQ_F_IO_DRAIN;
5669 req->ctx->drain_next = 0;
5671 if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
5672 req->flags |= REQ_F_LINK;
5673 INIT_LIST_HEAD(&req->link_list);
5674 ret = io_req_defer_prep(req, sqe);
5676 req->flags |= REQ_F_FAIL_LINK;
5679 io_queue_sqe(req, sqe);
5687 * Batched submission is done, ensure local IO is flushed out.
5689 static void io_submit_state_end(struct io_submit_state *state)
5691 blk_finish_plug(&state->plug);
5693 if (state->free_reqs)
5694 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5698 * Start submission side cache.
5700 static void io_submit_state_start(struct io_submit_state *state,
5701 unsigned int max_ios)
5703 blk_start_plug(&state->plug);
5704 state->free_reqs = 0;
5706 state->ios_left = max_ios;
5709 static void io_commit_sqring(struct io_ring_ctx *ctx)
5711 struct io_rings *rings = ctx->rings;
5714 * Ensure any loads from the SQEs are done at this point,
5715 * since once we write the new head, the application could
5716 * write new data to them.
5718 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5722 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5723 * that is mapped by userspace. This means that care needs to be taken to
5724 * ensure that reads are stable, as we cannot rely on userspace always
5725 * being a good citizen. If members of the sqe are validated and then later
5726 * used, it's important that those reads are done through READ_ONCE() to
5727 * prevent a re-load down the line.
5729 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
5730 const struct io_uring_sqe **sqe_ptr)
5732 u32 *sq_array = ctx->sq_array;
5736 * The cached sq head (or cq tail) serves two purposes:
5738 * 1) allows us to batch the cost of updating the user visible
5740 * 2) allows the kernel side to track the head on its own, even
5741 * though the application is the one updating it.
5743 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5744 if (likely(head < ctx->sq_entries)) {
5746 * All io need record the previous position, if LINK vs DARIN,
5747 * it can be used to mark the position of the first IO in the
5750 req->sequence = ctx->cached_sq_head;
5751 *sqe_ptr = &ctx->sq_sqes[head];
5752 req->opcode = READ_ONCE((*sqe_ptr)->opcode);
5753 req->user_data = READ_ONCE((*sqe_ptr)->user_data);
5754 ctx->cached_sq_head++;
5758 /* drop invalid entries */
5759 ctx->cached_sq_head++;
5760 ctx->cached_sq_dropped++;
5761 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5765 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5766 struct file *ring_file, int ring_fd,
5767 struct mm_struct **mm, bool async)
5769 struct io_submit_state state, *statep = NULL;
5770 struct io_kiocb *link = NULL;
5771 int i, submitted = 0;
5772 bool mm_fault = false;
5774 /* if we have a backlog and couldn't flush it all, return BUSY */
5775 if (test_bit(0, &ctx->sq_check_overflow)) {
5776 if (!list_empty(&ctx->cq_overflow_list) &&
5777 !io_cqring_overflow_flush(ctx, false))
5781 /* make sure SQ entry isn't read before tail */
5782 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5784 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5787 if (nr > IO_PLUG_THRESHOLD) {
5788 io_submit_state_start(&state, nr);
5792 ctx->ring_fd = ring_fd;
5793 ctx->ring_file = ring_file;
5795 for (i = 0; i < nr; i++) {
5796 const struct io_uring_sqe *sqe;
5797 struct io_kiocb *req;
5800 req = io_get_req(ctx, statep);
5801 if (unlikely(!req)) {
5803 submitted = -EAGAIN;
5806 if (!io_get_sqring(ctx, req, &sqe)) {
5807 __io_req_do_free(req);
5811 /* will complete beyond this point, count as submitted */
5814 if (unlikely(req->opcode >= IORING_OP_LAST)) {
5817 io_cqring_add_event(req, err);
5818 io_double_put_req(req);
5822 if (io_op_defs[req->opcode].needs_mm && !*mm) {
5823 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
5824 if (unlikely(mm_fault)) {
5828 use_mm(ctx->sqo_mm);
5832 req->needs_fixed_file = async;
5833 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5835 if (!io_submit_sqe(req, sqe, statep, &link))
5839 if (unlikely(submitted != nr)) {
5840 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5842 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5845 io_queue_link_head(link);
5847 io_submit_state_end(&state);
5849 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5850 io_commit_sqring(ctx);
5855 static int io_sq_thread(void *data)
5857 struct io_ring_ctx *ctx = data;
5858 struct mm_struct *cur_mm = NULL;
5859 const struct cred *old_cred;
5860 mm_segment_t old_fs;
5862 unsigned long timeout;
5865 complete(&ctx->completions[1]);
5869 old_cred = override_creds(ctx->creds);
5871 timeout = jiffies + ctx->sq_thread_idle;
5872 while (!kthread_should_park()) {
5873 unsigned int to_submit;
5875 if (!list_empty(&ctx->poll_list)) {
5876 unsigned nr_events = 0;
5878 mutex_lock(&ctx->uring_lock);
5879 if (!list_empty(&ctx->poll_list))
5880 io_iopoll_getevents(ctx, &nr_events, 0);
5882 timeout = jiffies + ctx->sq_thread_idle;
5883 mutex_unlock(&ctx->uring_lock);
5886 to_submit = io_sqring_entries(ctx);
5889 * If submit got -EBUSY, flag us as needing the application
5890 * to enter the kernel to reap and flush events.
5892 if (!to_submit || ret == -EBUSY) {
5894 * Drop cur_mm before scheduling, we can't hold it for
5895 * long periods (or over schedule()). Do this before
5896 * adding ourselves to the waitqueue, as the unuse/drop
5906 * We're polling. If we're within the defined idle
5907 * period, then let us spin without work before going
5908 * to sleep. The exception is if we got EBUSY doing
5909 * more IO, we should wait for the application to
5910 * reap events and wake us up.
5912 if (!list_empty(&ctx->poll_list) ||
5913 (!time_after(jiffies, timeout) && ret != -EBUSY &&
5914 !percpu_ref_is_dying(&ctx->refs))) {
5915 if (current->task_works)
5921 prepare_to_wait(&ctx->sqo_wait, &wait,
5922 TASK_INTERRUPTIBLE);
5925 * While doing polled IO, before going to sleep, we need
5926 * to check if there are new reqs added to poll_list, it
5927 * is because reqs may have been punted to io worker and
5928 * will be added to poll_list later, hence check the
5931 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
5932 !list_empty_careful(&ctx->poll_list)) {
5933 finish_wait(&ctx->sqo_wait, &wait);
5937 /* Tell userspace we may need a wakeup call */
5938 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
5939 /* make sure to read SQ tail after writing flags */
5942 to_submit = io_sqring_entries(ctx);
5943 if (!to_submit || ret == -EBUSY) {
5944 if (kthread_should_park()) {
5945 finish_wait(&ctx->sqo_wait, &wait);
5948 if (current->task_works) {
5952 if (signal_pending(current))
5953 flush_signals(current);
5955 finish_wait(&ctx->sqo_wait, &wait);
5957 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5960 finish_wait(&ctx->sqo_wait, &wait);
5962 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5965 mutex_lock(&ctx->uring_lock);
5966 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
5967 mutex_unlock(&ctx->uring_lock);
5968 timeout = jiffies + ctx->sq_thread_idle;
5971 if (current->task_works)
5979 revert_creds(old_cred);
5986 struct io_wait_queue {
5987 struct wait_queue_entry wq;
5988 struct io_ring_ctx *ctx;
5990 unsigned nr_timeouts;
5993 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
5995 struct io_ring_ctx *ctx = iowq->ctx;
5998 * Wake up if we have enough events, or if a timeout occurred since we
5999 * started waiting. For timeouts, we always want to return to userspace,
6000 * regardless of event count.
6002 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6003 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6006 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6007 int wake_flags, void *key)
6009 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6012 /* use noflush == true, as we can't safely rely on locking context */
6013 if (!io_should_wake(iowq, true))
6016 return autoremove_wake_function(curr, mode, wake_flags, key);
6020 * Wait until events become available, if we don't already have some. The
6021 * application must reap them itself, as they reside on the shared cq ring.
6023 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6024 const sigset_t __user *sig, size_t sigsz)
6026 struct io_wait_queue iowq = {
6029 .func = io_wake_function,
6030 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6033 .to_wait = min_events,
6035 struct io_rings *rings = ctx->rings;
6039 if (io_cqring_events(ctx, false) >= min_events)
6041 if (!current->task_works)
6047 #ifdef CONFIG_COMPAT
6048 if (in_compat_syscall())
6049 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6053 ret = set_user_sigmask(sig, sigsz);
6059 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6060 trace_io_uring_cqring_wait(ctx, min_events);
6062 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6063 TASK_INTERRUPTIBLE);
6064 if (current->task_works)
6066 if (io_should_wake(&iowq, false))
6069 if (signal_pending(current)) {
6074 finish_wait(&ctx->wait, &iowq.wq);
6076 restore_saved_sigmask_unless(ret == -EINTR);
6078 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6081 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6083 #if defined(CONFIG_UNIX)
6084 if (ctx->ring_sock) {
6085 struct sock *sock = ctx->ring_sock->sk;
6086 struct sk_buff *skb;
6088 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6094 for (i = 0; i < ctx->nr_user_files; i++) {
6097 file = io_file_from_index(ctx, i);
6104 static void io_file_ref_kill(struct percpu_ref *ref)
6106 struct fixed_file_data *data;
6108 data = container_of(ref, struct fixed_file_data, refs);
6109 complete(&data->done);
6112 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6114 struct fixed_file_data *data = ctx->file_data;
6115 unsigned nr_tables, i;
6120 percpu_ref_kill_and_confirm(&data->refs, io_file_ref_kill);
6121 flush_work(&data->ref_work);
6122 wait_for_completion(&data->done);
6123 io_ring_file_ref_flush(data);
6124 percpu_ref_exit(&data->refs);
6126 __io_sqe_files_unregister(ctx);
6127 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6128 for (i = 0; i < nr_tables; i++)
6129 kfree(data->table[i].files);
6132 ctx->file_data = NULL;
6133 ctx->nr_user_files = 0;
6137 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6139 if (ctx->sqo_thread) {
6140 wait_for_completion(&ctx->completions[1]);
6142 * The park is a bit of a work-around, without it we get
6143 * warning spews on shutdown with SQPOLL set and affinity
6144 * set to a single CPU.
6146 kthread_park(ctx->sqo_thread);
6147 kthread_stop(ctx->sqo_thread);
6148 ctx->sqo_thread = NULL;
6152 static void io_finish_async(struct io_ring_ctx *ctx)
6154 io_sq_thread_stop(ctx);
6157 io_wq_destroy(ctx->io_wq);
6162 #if defined(CONFIG_UNIX)
6164 * Ensure the UNIX gc is aware of our file set, so we are certain that
6165 * the io_uring can be safely unregistered on process exit, even if we have
6166 * loops in the file referencing.
6168 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6170 struct sock *sk = ctx->ring_sock->sk;
6171 struct scm_fp_list *fpl;
6172 struct sk_buff *skb;
6175 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
6176 unsigned long inflight = ctx->user->unix_inflight + nr;
6178 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
6182 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6186 skb = alloc_skb(0, GFP_KERNEL);
6195 fpl->user = get_uid(ctx->user);
6196 for (i = 0; i < nr; i++) {
6197 struct file *file = io_file_from_index(ctx, i + offset);
6201 fpl->fp[nr_files] = get_file(file);
6202 unix_inflight(fpl->user, fpl->fp[nr_files]);
6207 fpl->max = SCM_MAX_FD;
6208 fpl->count = nr_files;
6209 UNIXCB(skb).fp = fpl;
6210 skb->destructor = unix_destruct_scm;
6211 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6212 skb_queue_head(&sk->sk_receive_queue, skb);
6214 for (i = 0; i < nr_files; i++)
6225 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6226 * causes regular reference counting to break down. We rely on the UNIX
6227 * garbage collection to take care of this problem for us.
6229 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6231 unsigned left, total;
6235 left = ctx->nr_user_files;
6237 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6239 ret = __io_sqe_files_scm(ctx, this_files, total);
6243 total += this_files;
6249 while (total < ctx->nr_user_files) {
6250 struct file *file = io_file_from_index(ctx, total);
6260 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6266 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6271 for (i = 0; i < nr_tables; i++) {
6272 struct fixed_file_table *table = &ctx->file_data->table[i];
6273 unsigned this_files;
6275 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6276 table->files = kcalloc(this_files, sizeof(struct file *),
6280 nr_files -= this_files;
6286 for (i = 0; i < nr_tables; i++) {
6287 struct fixed_file_table *table = &ctx->file_data->table[i];
6288 kfree(table->files);
6293 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6295 #if defined(CONFIG_UNIX)
6296 struct sock *sock = ctx->ring_sock->sk;
6297 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6298 struct sk_buff *skb;
6301 __skb_queue_head_init(&list);
6304 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6305 * remove this entry and rearrange the file array.
6307 skb = skb_dequeue(head);
6309 struct scm_fp_list *fp;
6311 fp = UNIXCB(skb).fp;
6312 for (i = 0; i < fp->count; i++) {
6315 if (fp->fp[i] != file)
6318 unix_notinflight(fp->user, fp->fp[i]);
6319 left = fp->count - 1 - i;
6321 memmove(&fp->fp[i], &fp->fp[i + 1],
6322 left * sizeof(struct file *));
6329 __skb_queue_tail(&list, skb);
6339 __skb_queue_tail(&list, skb);
6341 skb = skb_dequeue(head);
6344 if (skb_peek(&list)) {
6345 spin_lock_irq(&head->lock);
6346 while ((skb = __skb_dequeue(&list)) != NULL)
6347 __skb_queue_tail(head, skb);
6348 spin_unlock_irq(&head->lock);
6355 struct io_file_put {
6356 struct llist_node llist;
6360 static void io_ring_file_ref_flush(struct fixed_file_data *data)
6362 struct io_file_put *pfile, *tmp;
6363 struct llist_node *node;
6365 while ((node = llist_del_all(&data->put_llist)) != NULL) {
6366 llist_for_each_entry_safe(pfile, tmp, node, llist) {
6367 io_ring_file_put(data->ctx, pfile->file);
6373 static void io_ring_file_ref_switch(struct work_struct *work)
6375 struct fixed_file_data *data;
6377 data = container_of(work, struct fixed_file_data, ref_work);
6378 io_ring_file_ref_flush(data);
6379 percpu_ref_switch_to_percpu(&data->refs);
6382 static void io_file_data_ref_zero(struct percpu_ref *ref)
6384 struct fixed_file_data *data;
6386 data = container_of(ref, struct fixed_file_data, refs);
6389 * We can't safely switch from inside this context, punt to wq. If
6390 * the table ref is going away, the table is being unregistered.
6391 * Don't queue up the async work for that case, the caller will
6394 if (!percpu_ref_is_dying(&data->refs))
6395 queue_work(system_wq, &data->ref_work);
6398 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6401 __s32 __user *fds = (__s32 __user *) arg;
6411 if (nr_args > IORING_MAX_FIXED_FILES)
6414 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6415 if (!ctx->file_data)
6417 ctx->file_data->ctx = ctx;
6418 init_completion(&ctx->file_data->done);
6420 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6421 ctx->file_data->table = kcalloc(nr_tables,
6422 sizeof(struct fixed_file_table),
6424 if (!ctx->file_data->table) {
6425 kfree(ctx->file_data);
6426 ctx->file_data = NULL;
6430 if (percpu_ref_init(&ctx->file_data->refs, io_file_data_ref_zero,
6431 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6432 kfree(ctx->file_data->table);
6433 kfree(ctx->file_data);
6434 ctx->file_data = NULL;
6437 ctx->file_data->put_llist.first = NULL;
6438 INIT_WORK(&ctx->file_data->ref_work, io_ring_file_ref_switch);
6440 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6441 percpu_ref_exit(&ctx->file_data->refs);
6442 kfree(ctx->file_data->table);
6443 kfree(ctx->file_data);
6444 ctx->file_data = NULL;
6448 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6449 struct fixed_file_table *table;
6453 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6455 /* allow sparse sets */
6461 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6462 index = i & IORING_FILE_TABLE_MASK;
6470 * Don't allow io_uring instances to be registered. If UNIX
6471 * isn't enabled, then this causes a reference cycle and this
6472 * instance can never get freed. If UNIX is enabled we'll
6473 * handle it just fine, but there's still no point in allowing
6474 * a ring fd as it doesn't support regular read/write anyway.
6476 if (file->f_op == &io_uring_fops) {
6481 table->files[index] = file;
6485 for (i = 0; i < ctx->nr_user_files; i++) {
6486 file = io_file_from_index(ctx, i);
6490 for (i = 0; i < nr_tables; i++)
6491 kfree(ctx->file_data->table[i].files);
6493 kfree(ctx->file_data->table);
6494 kfree(ctx->file_data);
6495 ctx->file_data = NULL;
6496 ctx->nr_user_files = 0;
6500 ret = io_sqe_files_scm(ctx);
6502 io_sqe_files_unregister(ctx);
6507 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6510 #if defined(CONFIG_UNIX)
6511 struct sock *sock = ctx->ring_sock->sk;
6512 struct sk_buff_head *head = &sock->sk_receive_queue;
6513 struct sk_buff *skb;
6516 * See if we can merge this file into an existing skb SCM_RIGHTS
6517 * file set. If there's no room, fall back to allocating a new skb
6518 * and filling it in.
6520 spin_lock_irq(&head->lock);
6521 skb = skb_peek(head);
6523 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6525 if (fpl->count < SCM_MAX_FD) {
6526 __skb_unlink(skb, head);
6527 spin_unlock_irq(&head->lock);
6528 fpl->fp[fpl->count] = get_file(file);
6529 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6531 spin_lock_irq(&head->lock);
6532 __skb_queue_head(head, skb);
6537 spin_unlock_irq(&head->lock);
6544 return __io_sqe_files_scm(ctx, 1, index);
6550 static void io_atomic_switch(struct percpu_ref *ref)
6552 struct fixed_file_data *data;
6555 * Juggle reference to ensure we hit zero, if needed, so we can
6556 * switch back to percpu mode
6558 data = container_of(ref, struct fixed_file_data, refs);
6559 percpu_ref_put(&data->refs);
6560 percpu_ref_get(&data->refs);
6563 static int io_queue_file_removal(struct fixed_file_data *data,
6566 struct io_file_put *pfile;
6568 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6573 llist_add(&pfile->llist, &data->put_llist);
6577 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6578 struct io_uring_files_update *up,
6581 struct fixed_file_data *data = ctx->file_data;
6582 bool ref_switch = false;
6588 if (check_add_overflow(up->offset, nr_args, &done))
6590 if (done > ctx->nr_user_files)
6594 fds = u64_to_user_ptr(up->fds);
6596 struct fixed_file_table *table;
6600 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6604 i = array_index_nospec(up->offset, ctx->nr_user_files);
6605 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6606 index = i & IORING_FILE_TABLE_MASK;
6607 if (table->files[index]) {
6608 file = io_file_from_index(ctx, index);
6609 err = io_queue_file_removal(data, file);
6612 table->files[index] = NULL;
6622 * Don't allow io_uring instances to be registered. If
6623 * UNIX isn't enabled, then this causes a reference
6624 * cycle and this instance can never get freed. If UNIX
6625 * is enabled we'll handle it just fine, but there's
6626 * still no point in allowing a ring fd as it doesn't
6627 * support regular read/write anyway.
6629 if (file->f_op == &io_uring_fops) {
6634 table->files[index] = file;
6635 err = io_sqe_file_register(ctx, file, i);
6645 percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch);
6647 return done ? done : err;
6649 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6652 struct io_uring_files_update up;
6654 if (!ctx->file_data)
6658 if (copy_from_user(&up, arg, sizeof(up)))
6663 return __io_sqe_files_update(ctx, &up, nr_args);
6666 static void io_free_work(struct io_wq_work *work)
6668 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6670 /* Consider that io_steal_work() relies on this ref */
6674 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6675 struct io_uring_params *p)
6677 struct io_wq_data data;
6679 struct io_ring_ctx *ctx_attach;
6680 unsigned int concurrency;
6683 data.user = ctx->user;
6684 data.free_work = io_free_work;
6686 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6687 /* Do QD, or 4 * CPUS, whatever is smallest */
6688 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6690 ctx->io_wq = io_wq_create(concurrency, &data);
6691 if (IS_ERR(ctx->io_wq)) {
6692 ret = PTR_ERR(ctx->io_wq);
6698 f = fdget(p->wq_fd);
6702 if (f.file->f_op != &io_uring_fops) {
6707 ctx_attach = f.file->private_data;
6708 /* @io_wq is protected by holding the fd */
6709 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6714 ctx->io_wq = ctx_attach->io_wq;
6720 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6721 struct io_uring_params *p)
6725 init_waitqueue_head(&ctx->sqo_wait);
6726 mmgrab(current->mm);
6727 ctx->sqo_mm = current->mm;
6729 if (ctx->flags & IORING_SETUP_SQPOLL) {
6731 if (!capable(CAP_SYS_ADMIN))
6734 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6735 if (!ctx->sq_thread_idle)
6736 ctx->sq_thread_idle = HZ;
6738 if (p->flags & IORING_SETUP_SQ_AFF) {
6739 int cpu = p->sq_thread_cpu;
6742 if (cpu >= nr_cpu_ids)
6744 if (!cpu_online(cpu))
6747 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6751 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6754 if (IS_ERR(ctx->sqo_thread)) {
6755 ret = PTR_ERR(ctx->sqo_thread);
6756 ctx->sqo_thread = NULL;
6759 wake_up_process(ctx->sqo_thread);
6760 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6761 /* Can't have SQ_AFF without SQPOLL */
6766 ret = io_init_wq_offload(ctx, p);
6772 io_finish_async(ctx);
6773 mmdrop(ctx->sqo_mm);
6778 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6780 atomic_long_sub(nr_pages, &user->locked_vm);
6783 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6785 unsigned long page_limit, cur_pages, new_pages;
6787 /* Don't allow more pages than we can safely lock */
6788 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6791 cur_pages = atomic_long_read(&user->locked_vm);
6792 new_pages = cur_pages + nr_pages;
6793 if (new_pages > page_limit)
6795 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6796 new_pages) != cur_pages);
6801 static void io_mem_free(void *ptr)
6808 page = virt_to_head_page(ptr);
6809 if (put_page_testzero(page))
6810 free_compound_page(page);
6813 static void *io_mem_alloc(size_t size)
6815 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6818 return (void *) __get_free_pages(gfp_flags, get_order(size));
6821 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6824 struct io_rings *rings;
6825 size_t off, sq_array_size;
6827 off = struct_size(rings, cqes, cq_entries);
6828 if (off == SIZE_MAX)
6832 off = ALIGN(off, SMP_CACHE_BYTES);
6837 sq_array_size = array_size(sizeof(u32), sq_entries);
6838 if (sq_array_size == SIZE_MAX)
6841 if (check_add_overflow(off, sq_array_size, &off))
6850 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6854 pages = (size_t)1 << get_order(
6855 rings_size(sq_entries, cq_entries, NULL));
6856 pages += (size_t)1 << get_order(
6857 array_size(sizeof(struct io_uring_sqe), sq_entries));
6862 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6866 if (!ctx->user_bufs)
6869 for (i = 0; i < ctx->nr_user_bufs; i++) {
6870 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6872 for (j = 0; j < imu->nr_bvecs; j++)
6873 unpin_user_page(imu->bvec[j].bv_page);
6875 if (ctx->account_mem)
6876 io_unaccount_mem(ctx->user, imu->nr_bvecs);
6881 kfree(ctx->user_bufs);
6882 ctx->user_bufs = NULL;
6883 ctx->nr_user_bufs = 0;
6887 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
6888 void __user *arg, unsigned index)
6890 struct iovec __user *src;
6892 #ifdef CONFIG_COMPAT
6894 struct compat_iovec __user *ciovs;
6895 struct compat_iovec ciov;
6897 ciovs = (struct compat_iovec __user *) arg;
6898 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
6901 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
6902 dst->iov_len = ciov.iov_len;
6906 src = (struct iovec __user *) arg;
6907 if (copy_from_user(dst, &src[index], sizeof(*dst)))
6912 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
6915 struct vm_area_struct **vmas = NULL;
6916 struct page **pages = NULL;
6917 int i, j, got_pages = 0;
6922 if (!nr_args || nr_args > UIO_MAXIOV)
6925 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
6927 if (!ctx->user_bufs)
6930 for (i = 0; i < nr_args; i++) {
6931 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6932 unsigned long off, start, end, ubuf;
6937 ret = io_copy_iov(ctx, &iov, arg, i);
6942 * Don't impose further limits on the size and buffer
6943 * constraints here, we'll -EINVAL later when IO is
6944 * submitted if they are wrong.
6947 if (!iov.iov_base || !iov.iov_len)
6950 /* arbitrary limit, but we need something */
6951 if (iov.iov_len > SZ_1G)
6954 ubuf = (unsigned long) iov.iov_base;
6955 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
6956 start = ubuf >> PAGE_SHIFT;
6957 nr_pages = end - start;
6959 if (ctx->account_mem) {
6960 ret = io_account_mem(ctx->user, nr_pages);
6966 if (!pages || nr_pages > got_pages) {
6969 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
6971 vmas = kvmalloc_array(nr_pages,
6972 sizeof(struct vm_area_struct *),
6974 if (!pages || !vmas) {
6976 if (ctx->account_mem)
6977 io_unaccount_mem(ctx->user, nr_pages);
6980 got_pages = nr_pages;
6983 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
6987 if (ctx->account_mem)
6988 io_unaccount_mem(ctx->user, nr_pages);
6993 down_read(¤t->mm->mmap_sem);
6994 pret = pin_user_pages(ubuf, nr_pages,
6995 FOLL_WRITE | FOLL_LONGTERM,
6997 if (pret == nr_pages) {
6998 /* don't support file backed memory */
6999 for (j = 0; j < nr_pages; j++) {
7000 struct vm_area_struct *vma = vmas[j];
7003 !is_file_hugepages(vma->vm_file)) {
7009 ret = pret < 0 ? pret : -EFAULT;
7011 up_read(¤t->mm->mmap_sem);
7014 * if we did partial map, or found file backed vmas,
7015 * release any pages we did get
7018 unpin_user_pages(pages, pret);
7019 if (ctx->account_mem)
7020 io_unaccount_mem(ctx->user, nr_pages);
7025 off = ubuf & ~PAGE_MASK;
7027 for (j = 0; j < nr_pages; j++) {
7030 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7031 imu->bvec[j].bv_page = pages[j];
7032 imu->bvec[j].bv_len = vec_len;
7033 imu->bvec[j].bv_offset = off;
7037 /* store original address for later verification */
7039 imu->len = iov.iov_len;
7040 imu->nr_bvecs = nr_pages;
7042 ctx->nr_user_bufs++;
7050 io_sqe_buffer_unregister(ctx);
7054 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7056 __s32 __user *fds = arg;
7062 if (copy_from_user(&fd, fds, sizeof(*fds)))
7065 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7066 if (IS_ERR(ctx->cq_ev_fd)) {
7067 int ret = PTR_ERR(ctx->cq_ev_fd);
7068 ctx->cq_ev_fd = NULL;
7075 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7077 if (ctx->cq_ev_fd) {
7078 eventfd_ctx_put(ctx->cq_ev_fd);
7079 ctx->cq_ev_fd = NULL;
7086 static int __io_destroy_buffers(int id, void *p, void *data)
7088 struct io_ring_ctx *ctx = data;
7089 struct io_buffer *buf = p;
7091 __io_remove_buffers(ctx, buf, id, -1U);
7095 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7097 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7098 idr_destroy(&ctx->io_buffer_idr);
7101 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7103 io_finish_async(ctx);
7105 mmdrop(ctx->sqo_mm);
7107 io_iopoll_reap_events(ctx);
7108 io_sqe_buffer_unregister(ctx);
7109 io_sqe_files_unregister(ctx);
7110 io_eventfd_unregister(ctx);
7111 io_destroy_buffers(ctx);
7112 idr_destroy(&ctx->personality_idr);
7114 #if defined(CONFIG_UNIX)
7115 if (ctx->ring_sock) {
7116 ctx->ring_sock->file = NULL; /* so that iput() is called */
7117 sock_release(ctx->ring_sock);
7121 io_mem_free(ctx->rings);
7122 io_mem_free(ctx->sq_sqes);
7124 percpu_ref_exit(&ctx->refs);
7125 if (ctx->account_mem)
7126 io_unaccount_mem(ctx->user,
7127 ring_pages(ctx->sq_entries, ctx->cq_entries));
7128 free_uid(ctx->user);
7129 put_cred(ctx->creds);
7130 kfree(ctx->completions);
7131 kfree(ctx->cancel_hash);
7132 kmem_cache_free(req_cachep, ctx->fallback_req);
7136 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7138 struct io_ring_ctx *ctx = file->private_data;
7141 poll_wait(file, &ctx->cq_wait, wait);
7143 * synchronizes with barrier from wq_has_sleeper call in
7147 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7148 ctx->rings->sq_ring_entries)
7149 mask |= EPOLLOUT | EPOLLWRNORM;
7150 if (io_cqring_events(ctx, false))
7151 mask |= EPOLLIN | EPOLLRDNORM;
7156 static int io_uring_fasync(int fd, struct file *file, int on)
7158 struct io_ring_ctx *ctx = file->private_data;
7160 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7163 static int io_remove_personalities(int id, void *p, void *data)
7165 struct io_ring_ctx *ctx = data;
7166 const struct cred *cred;
7168 cred = idr_remove(&ctx->personality_idr, id);
7174 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7176 mutex_lock(&ctx->uring_lock);
7177 percpu_ref_kill(&ctx->refs);
7178 mutex_unlock(&ctx->uring_lock);
7181 * Wait for sq thread to idle, if we have one. It won't spin on new
7182 * work after we've killed the ctx ref above. This is important to do
7183 * before we cancel existing commands, as the thread could otherwise
7184 * be queueing new work post that. If that's work we need to cancel,
7185 * it could cause shutdown to hang.
7187 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
7190 io_kill_timeouts(ctx);
7191 io_poll_remove_all(ctx);
7194 io_wq_cancel_all(ctx->io_wq);
7196 io_iopoll_reap_events(ctx);
7197 /* if we failed setting up the ctx, we might not have any rings */
7199 io_cqring_overflow_flush(ctx, true);
7200 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7201 wait_for_completion(&ctx->completions[0]);
7202 io_ring_ctx_free(ctx);
7205 static int io_uring_release(struct inode *inode, struct file *file)
7207 struct io_ring_ctx *ctx = file->private_data;
7209 file->private_data = NULL;
7210 io_ring_ctx_wait_and_kill(ctx);
7214 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7215 struct files_struct *files)
7217 struct io_kiocb *req;
7220 while (!list_empty_careful(&ctx->inflight_list)) {
7221 struct io_kiocb *cancel_req = NULL;
7223 spin_lock_irq(&ctx->inflight_lock);
7224 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7225 if (req->work.files != files)
7227 /* req is being completed, ignore */
7228 if (!refcount_inc_not_zero(&req->refs))
7234 prepare_to_wait(&ctx->inflight_wait, &wait,
7235 TASK_UNINTERRUPTIBLE);
7236 spin_unlock_irq(&ctx->inflight_lock);
7238 /* We need to keep going until we don't find a matching req */
7242 if (cancel_req->flags & REQ_F_OVERFLOW) {
7243 spin_lock_irq(&ctx->completion_lock);
7244 list_del(&cancel_req->list);
7245 cancel_req->flags &= ~REQ_F_OVERFLOW;
7246 if (list_empty(&ctx->cq_overflow_list)) {
7247 clear_bit(0, &ctx->sq_check_overflow);
7248 clear_bit(0, &ctx->cq_check_overflow);
7250 spin_unlock_irq(&ctx->completion_lock);
7252 WRITE_ONCE(ctx->rings->cq_overflow,
7253 atomic_inc_return(&ctx->cached_cq_overflow));
7256 * Put inflight ref and overflow ref. If that's
7257 * all we had, then we're done with this request.
7259 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7260 io_put_req(cancel_req);
7265 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7266 io_put_req(cancel_req);
7269 finish_wait(&ctx->inflight_wait, &wait);
7272 static int io_uring_flush(struct file *file, void *data)
7274 struct io_ring_ctx *ctx = file->private_data;
7276 io_uring_cancel_files(ctx, data);
7279 * If the task is going away, cancel work it may have pending
7281 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7282 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
7287 static void *io_uring_validate_mmap_request(struct file *file,
7288 loff_t pgoff, size_t sz)
7290 struct io_ring_ctx *ctx = file->private_data;
7291 loff_t offset = pgoff << PAGE_SHIFT;
7296 case IORING_OFF_SQ_RING:
7297 case IORING_OFF_CQ_RING:
7300 case IORING_OFF_SQES:
7304 return ERR_PTR(-EINVAL);
7307 page = virt_to_head_page(ptr);
7308 if (sz > page_size(page))
7309 return ERR_PTR(-EINVAL);
7316 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7318 size_t sz = vma->vm_end - vma->vm_start;
7322 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7324 return PTR_ERR(ptr);
7326 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7327 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7330 #else /* !CONFIG_MMU */
7332 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7334 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7337 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7339 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7342 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7343 unsigned long addr, unsigned long len,
7344 unsigned long pgoff, unsigned long flags)
7348 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7350 return PTR_ERR(ptr);
7352 return (unsigned long) ptr;
7355 #endif /* !CONFIG_MMU */
7357 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7358 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7361 struct io_ring_ctx *ctx;
7366 if (current->task_works)
7369 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7377 if (f.file->f_op != &io_uring_fops)
7381 ctx = f.file->private_data;
7382 if (!percpu_ref_tryget(&ctx->refs))
7386 * For SQ polling, the thread will do all submissions and completions.
7387 * Just return the requested submit count, and wake the thread if
7391 if (ctx->flags & IORING_SETUP_SQPOLL) {
7392 if (!list_empty_careful(&ctx->cq_overflow_list))
7393 io_cqring_overflow_flush(ctx, false);
7394 if (flags & IORING_ENTER_SQ_WAKEUP)
7395 wake_up(&ctx->sqo_wait);
7396 submitted = to_submit;
7397 } else if (to_submit) {
7398 struct mm_struct *cur_mm;
7400 mutex_lock(&ctx->uring_lock);
7401 /* already have mm, so io_submit_sqes() won't try to grab it */
7402 cur_mm = ctx->sqo_mm;
7403 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
7405 mutex_unlock(&ctx->uring_lock);
7407 if (submitted != to_submit)
7410 if (flags & IORING_ENTER_GETEVENTS) {
7411 unsigned nr_events = 0;
7413 min_complete = min(min_complete, ctx->cq_entries);
7416 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7417 * space applications don't need to do io completion events
7418 * polling again, they can rely on io_sq_thread to do polling
7419 * work, which can reduce cpu usage and uring_lock contention.
7421 if (ctx->flags & IORING_SETUP_IOPOLL &&
7422 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7423 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7425 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7430 percpu_ref_put(&ctx->refs);
7433 return submitted ? submitted : ret;
7436 #ifdef CONFIG_PROC_FS
7437 static int io_uring_show_cred(int id, void *p, void *data)
7439 const struct cred *cred = p;
7440 struct seq_file *m = data;
7441 struct user_namespace *uns = seq_user_ns(m);
7442 struct group_info *gi;
7447 seq_printf(m, "%5d\n", id);
7448 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7449 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7450 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7451 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7452 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7453 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7454 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7455 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7456 seq_puts(m, "\n\tGroups:\t");
7457 gi = cred->group_info;
7458 for (g = 0; g < gi->ngroups; g++) {
7459 seq_put_decimal_ull(m, g ? " " : "",
7460 from_kgid_munged(uns, gi->gid[g]));
7462 seq_puts(m, "\n\tCapEff:\t");
7463 cap = cred->cap_effective;
7464 CAP_FOR_EACH_U32(__capi)
7465 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7470 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7474 mutex_lock(&ctx->uring_lock);
7475 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7476 for (i = 0; i < ctx->nr_user_files; i++) {
7477 struct fixed_file_table *table;
7480 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7481 f = table->files[i & IORING_FILE_TABLE_MASK];
7483 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7485 seq_printf(m, "%5u: <none>\n", i);
7487 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7488 for (i = 0; i < ctx->nr_user_bufs; i++) {
7489 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7491 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7492 (unsigned int) buf->len);
7494 if (!idr_is_empty(&ctx->personality_idr)) {
7495 seq_printf(m, "Personalities:\n");
7496 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7498 seq_printf(m, "PollList:\n");
7499 spin_lock_irq(&ctx->completion_lock);
7500 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7501 struct hlist_head *list = &ctx->cancel_hash[i];
7502 struct io_kiocb *req;
7504 hlist_for_each_entry(req, list, hash_node)
7505 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7506 req->task->task_works != NULL);
7508 spin_unlock_irq(&ctx->completion_lock);
7509 mutex_unlock(&ctx->uring_lock);
7512 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7514 struct io_ring_ctx *ctx = f->private_data;
7516 if (percpu_ref_tryget(&ctx->refs)) {
7517 __io_uring_show_fdinfo(ctx, m);
7518 percpu_ref_put(&ctx->refs);
7523 static const struct file_operations io_uring_fops = {
7524 .release = io_uring_release,
7525 .flush = io_uring_flush,
7526 .mmap = io_uring_mmap,
7528 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7529 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7531 .poll = io_uring_poll,
7532 .fasync = io_uring_fasync,
7533 #ifdef CONFIG_PROC_FS
7534 .show_fdinfo = io_uring_show_fdinfo,
7538 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7539 struct io_uring_params *p)
7541 struct io_rings *rings;
7542 size_t size, sq_array_offset;
7544 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7545 if (size == SIZE_MAX)
7548 rings = io_mem_alloc(size);
7553 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7554 rings->sq_ring_mask = p->sq_entries - 1;
7555 rings->cq_ring_mask = p->cq_entries - 1;
7556 rings->sq_ring_entries = p->sq_entries;
7557 rings->cq_ring_entries = p->cq_entries;
7558 ctx->sq_mask = rings->sq_ring_mask;
7559 ctx->cq_mask = rings->cq_ring_mask;
7560 ctx->sq_entries = rings->sq_ring_entries;
7561 ctx->cq_entries = rings->cq_ring_entries;
7563 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7564 if (size == SIZE_MAX) {
7565 io_mem_free(ctx->rings);
7570 ctx->sq_sqes = io_mem_alloc(size);
7571 if (!ctx->sq_sqes) {
7572 io_mem_free(ctx->rings);
7581 * Allocate an anonymous fd, this is what constitutes the application
7582 * visible backing of an io_uring instance. The application mmaps this
7583 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7584 * we have to tie this fd to a socket for file garbage collection purposes.
7586 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7591 #if defined(CONFIG_UNIX)
7592 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7598 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7602 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7603 O_RDWR | O_CLOEXEC);
7606 ret = PTR_ERR(file);
7610 #if defined(CONFIG_UNIX)
7611 ctx->ring_sock->file = file;
7613 fd_install(ret, file);
7616 #if defined(CONFIG_UNIX)
7617 sock_release(ctx->ring_sock);
7618 ctx->ring_sock = NULL;
7623 static int io_uring_create(unsigned entries, struct io_uring_params *p)
7625 struct user_struct *user = NULL;
7626 struct io_ring_ctx *ctx;
7632 if (entries > IORING_MAX_ENTRIES) {
7633 if (!(p->flags & IORING_SETUP_CLAMP))
7635 entries = IORING_MAX_ENTRIES;
7639 * Use twice as many entries for the CQ ring. It's possible for the
7640 * application to drive a higher depth than the size of the SQ ring,
7641 * since the sqes are only used at submission time. This allows for
7642 * some flexibility in overcommitting a bit. If the application has
7643 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7644 * of CQ ring entries manually.
7646 p->sq_entries = roundup_pow_of_two(entries);
7647 if (p->flags & IORING_SETUP_CQSIZE) {
7649 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7650 * to a power-of-two, if it isn't already. We do NOT impose
7651 * any cq vs sq ring sizing.
7653 if (p->cq_entries < p->sq_entries)
7655 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7656 if (!(p->flags & IORING_SETUP_CLAMP))
7658 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7660 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7662 p->cq_entries = 2 * p->sq_entries;
7665 user = get_uid(current_user());
7666 account_mem = !capable(CAP_IPC_LOCK);
7669 ret = io_account_mem(user,
7670 ring_pages(p->sq_entries, p->cq_entries));
7677 ctx = io_ring_ctx_alloc(p);
7680 io_unaccount_mem(user, ring_pages(p->sq_entries,
7685 ctx->compat = in_compat_syscall();
7686 ctx->account_mem = account_mem;
7688 ctx->creds = get_current_cred();
7690 ret = io_allocate_scq_urings(ctx, p);
7694 ret = io_sq_offload_start(ctx, p);
7698 memset(&p->sq_off, 0, sizeof(p->sq_off));
7699 p->sq_off.head = offsetof(struct io_rings, sq.head);
7700 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7701 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7702 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7703 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7704 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7705 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7707 memset(&p->cq_off, 0, sizeof(p->cq_off));
7708 p->cq_off.head = offsetof(struct io_rings, cq.head);
7709 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7710 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7711 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7712 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7713 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7716 * Install ring fd as the very last thing, so we don't risk someone
7717 * having closed it before we finish setup
7719 ret = io_uring_get_fd(ctx);
7723 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7724 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7725 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7726 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7729 io_ring_ctx_wait_and_kill(ctx);
7734 * Sets up an aio uring context, and returns the fd. Applications asks for a
7735 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7736 * params structure passed in.
7738 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7740 struct io_uring_params p;
7744 if (copy_from_user(&p, params, sizeof(p)))
7746 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7751 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7752 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7753 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7756 ret = io_uring_create(entries, &p);
7760 if (copy_to_user(params, &p, sizeof(p)))
7766 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7767 struct io_uring_params __user *, params)
7769 return io_uring_setup(entries, params);
7772 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7774 struct io_uring_probe *p;
7778 size = struct_size(p, ops, nr_args);
7779 if (size == SIZE_MAX)
7781 p = kzalloc(size, GFP_KERNEL);
7786 if (copy_from_user(p, arg, size))
7789 if (memchr_inv(p, 0, size))
7792 p->last_op = IORING_OP_LAST - 1;
7793 if (nr_args > IORING_OP_LAST)
7794 nr_args = IORING_OP_LAST;
7796 for (i = 0; i < nr_args; i++) {
7798 if (!io_op_defs[i].not_supported)
7799 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7804 if (copy_to_user(arg, p, size))
7811 static int io_register_personality(struct io_ring_ctx *ctx)
7813 const struct cred *creds = get_current_cred();
7816 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7817 USHRT_MAX, GFP_KERNEL);
7823 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7825 const struct cred *old_creds;
7827 old_creds = idr_remove(&ctx->personality_idr, id);
7829 put_cred(old_creds);
7836 static bool io_register_op_must_quiesce(int op)
7839 case IORING_UNREGISTER_FILES:
7840 case IORING_REGISTER_FILES_UPDATE:
7841 case IORING_REGISTER_PROBE:
7842 case IORING_REGISTER_PERSONALITY:
7843 case IORING_UNREGISTER_PERSONALITY:
7850 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
7851 void __user *arg, unsigned nr_args)
7852 __releases(ctx->uring_lock)
7853 __acquires(ctx->uring_lock)
7858 * We're inside the ring mutex, if the ref is already dying, then
7859 * someone else killed the ctx or is already going through
7860 * io_uring_register().
7862 if (percpu_ref_is_dying(&ctx->refs))
7865 if (io_register_op_must_quiesce(opcode)) {
7866 percpu_ref_kill(&ctx->refs);
7869 * Drop uring mutex before waiting for references to exit. If
7870 * another thread is currently inside io_uring_enter() it might
7871 * need to grab the uring_lock to make progress. If we hold it
7872 * here across the drain wait, then we can deadlock. It's safe
7873 * to drop the mutex here, since no new references will come in
7874 * after we've killed the percpu ref.
7876 mutex_unlock(&ctx->uring_lock);
7877 ret = wait_for_completion_interruptible(&ctx->completions[0]);
7878 mutex_lock(&ctx->uring_lock);
7880 percpu_ref_resurrect(&ctx->refs);
7887 case IORING_REGISTER_BUFFERS:
7888 ret = io_sqe_buffer_register(ctx, arg, nr_args);
7890 case IORING_UNREGISTER_BUFFERS:
7894 ret = io_sqe_buffer_unregister(ctx);
7896 case IORING_REGISTER_FILES:
7897 ret = io_sqe_files_register(ctx, arg, nr_args);
7899 case IORING_UNREGISTER_FILES:
7903 ret = io_sqe_files_unregister(ctx);
7905 case IORING_REGISTER_FILES_UPDATE:
7906 ret = io_sqe_files_update(ctx, arg, nr_args);
7908 case IORING_REGISTER_EVENTFD:
7909 case IORING_REGISTER_EVENTFD_ASYNC:
7913 ret = io_eventfd_register(ctx, arg);
7916 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
7917 ctx->eventfd_async = 1;
7919 ctx->eventfd_async = 0;
7921 case IORING_UNREGISTER_EVENTFD:
7925 ret = io_eventfd_unregister(ctx);
7927 case IORING_REGISTER_PROBE:
7929 if (!arg || nr_args > 256)
7931 ret = io_probe(ctx, arg, nr_args);
7933 case IORING_REGISTER_PERSONALITY:
7937 ret = io_register_personality(ctx);
7939 case IORING_UNREGISTER_PERSONALITY:
7943 ret = io_unregister_personality(ctx, nr_args);
7950 if (io_register_op_must_quiesce(opcode)) {
7951 /* bring the ctx back to life */
7952 percpu_ref_reinit(&ctx->refs);
7954 reinit_completion(&ctx->completions[0]);
7959 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
7960 void __user *, arg, unsigned int, nr_args)
7962 struct io_ring_ctx *ctx;
7971 if (f.file->f_op != &io_uring_fops)
7974 ctx = f.file->private_data;
7976 mutex_lock(&ctx->uring_lock);
7977 ret = __io_uring_register(ctx, opcode, arg, nr_args);
7978 mutex_unlock(&ctx->uring_lock);
7979 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
7980 ctx->cq_ev_fd != NULL, ret);
7986 static int __init io_uring_init(void)
7988 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
7989 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
7990 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
7993 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
7994 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
7995 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
7996 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
7997 BUILD_BUG_SQE_ELEM(1, __u8, flags);
7998 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
7999 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8000 BUILD_BUG_SQE_ELEM(8, __u64, off);
8001 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8002 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8003 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8004 BUILD_BUG_SQE_ELEM(24, __u32, len);
8005 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8006 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8007 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8008 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8009 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8010 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8011 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8012 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8013 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8014 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8015 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8016 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8017 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8018 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8019 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8020 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8021 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8022 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8024 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8025 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8026 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8029 __initcall(io_uring_init);