1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/mmu_context.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/io_uring.h>
83 #include <uapi/linux/io_uring.h>
88 #define IORING_MAX_ENTRIES 32768
89 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
92 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
94 #define IORING_FILE_TABLE_SHIFT 9
95 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
96 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
97 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 u32 head ____cacheline_aligned_in_smp;
101 u32 tail ____cacheline_aligned_in_smp;
105 * This data is shared with the application through the mmap at offsets
106 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
108 * The offsets to the member fields are published through struct
109 * io_sqring_offsets when calling io_uring_setup.
113 * Head and tail offsets into the ring; the offsets need to be
114 * masked to get valid indices.
116 * The kernel controls head of the sq ring and the tail of the cq ring,
117 * and the application controls tail of the sq ring and the head of the
120 struct io_uring sq, cq;
122 * Bitmasks to apply to head and tail offsets (constant, equals
125 u32 sq_ring_mask, cq_ring_mask;
126 /* Ring sizes (constant, power of 2) */
127 u32 sq_ring_entries, cq_ring_entries;
129 * Number of invalid entries dropped by the kernel due to
130 * invalid index stored in array
132 * Written by the kernel, shouldn't be modified by the
133 * application (i.e. get number of "new events" by comparing to
136 * After a new SQ head value was read by the application this
137 * counter includes all submissions that were dropped reaching
138 * the new SQ head (and possibly more).
144 * Written by the kernel, shouldn't be modified by the
147 * The application needs a full memory barrier before checking
148 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
152 * Number of completion events lost because the queue was full;
153 * this should be avoided by the application by making sure
154 * there are not more requests pending than there is space in
155 * the completion queue.
157 * Written by the kernel, shouldn't be modified by the
158 * application (i.e. get number of "new events" by comparing to
161 * As completion events come in out of order this counter is not
162 * ordered with any other data.
166 * Ring buffer of completion events.
168 * The kernel writes completion events fresh every time they are
169 * produced, so the application is allowed to modify pending
172 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
175 struct io_mapped_ubuf {
178 struct bio_vec *bvec;
179 unsigned int nr_bvecs;
182 struct fixed_file_table {
186 struct fixed_file_data {
187 struct fixed_file_table *table;
188 struct io_ring_ctx *ctx;
190 struct percpu_ref refs;
191 struct llist_head put_llist;
192 struct work_struct ref_work;
193 struct completion done;
198 struct percpu_ref refs;
199 } ____cacheline_aligned_in_smp;
203 unsigned int compat: 1;
204 unsigned int account_mem: 1;
205 unsigned int cq_overflow_flushed: 1;
206 unsigned int drain_next: 1;
207 unsigned int eventfd_async: 1;
210 * Ring buffer of indices into array of io_uring_sqe, which is
211 * mmapped by the application using the IORING_OFF_SQES offset.
213 * This indirection could e.g. be used to assign fixed
214 * io_uring_sqe entries to operations and only submit them to
215 * the queue when needed.
217 * The kernel modifies neither the indices array nor the entries
221 unsigned cached_sq_head;
224 unsigned sq_thread_idle;
225 unsigned cached_sq_dropped;
226 atomic_t cached_cq_overflow;
227 unsigned long sq_check_overflow;
229 struct list_head defer_list;
230 struct list_head timeout_list;
231 struct list_head cq_overflow_list;
233 wait_queue_head_t inflight_wait;
234 struct io_uring_sqe *sq_sqes;
235 } ____cacheline_aligned_in_smp;
237 struct io_rings *rings;
241 struct task_struct *sqo_thread; /* if using sq thread polling */
242 struct mm_struct *sqo_mm;
243 wait_queue_head_t sqo_wait;
246 * If used, fixed file set. Writers must ensure that ->refs is dead,
247 * readers must ensure that ->refs is alive as long as the file* is
248 * used. Only updated through io_uring_register(2).
250 struct fixed_file_data *file_data;
251 unsigned nr_user_files;
253 struct file *ring_file;
255 /* if used, fixed mapped user buffers */
256 unsigned nr_user_bufs;
257 struct io_mapped_ubuf *user_bufs;
259 struct user_struct *user;
261 const struct cred *creds;
263 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
264 struct completion *completions;
266 /* if all else fails... */
267 struct io_kiocb *fallback_req;
269 #if defined(CONFIG_UNIX)
270 struct socket *ring_sock;
273 struct idr personality_idr;
276 unsigned cached_cq_tail;
279 atomic_t cq_timeouts;
280 unsigned long cq_check_overflow;
281 struct wait_queue_head cq_wait;
282 struct fasync_struct *cq_fasync;
283 struct eventfd_ctx *cq_ev_fd;
284 } ____cacheline_aligned_in_smp;
287 struct mutex uring_lock;
288 wait_queue_head_t wait;
289 } ____cacheline_aligned_in_smp;
292 spinlock_t completion_lock;
293 struct llist_head poll_llist;
296 * ->poll_list is protected by the ctx->uring_lock for
297 * io_uring instances that don't use IORING_SETUP_SQPOLL.
298 * For SQPOLL, only the single threaded io_sq_thread() will
299 * manipulate the list, hence no extra locking is needed there.
301 struct list_head poll_list;
302 struct hlist_head *cancel_hash;
303 unsigned cancel_hash_bits;
304 bool poll_multi_file;
306 spinlock_t inflight_lock;
307 struct list_head inflight_list;
308 } ____cacheline_aligned_in_smp;
312 * First field must be the file pointer in all the
313 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
315 struct io_poll_iocb {
318 struct wait_queue_head *head;
324 struct wait_queue_entry wait;
329 struct file *put_file;
333 struct io_timeout_data {
334 struct io_kiocb *req;
335 struct hrtimer timer;
336 struct timespec64 ts;
337 enum hrtimer_mode mode;
343 struct sockaddr __user *addr;
344 int __user *addr_len;
346 unsigned long nofile;
370 /* NOTE: kiocb has the file as the first member, so don't do it here */
378 struct sockaddr __user *addr;
385 struct user_msghdr __user *msg;
398 struct filename *filename;
399 struct statx __user *buffer;
401 unsigned long nofile;
404 struct io_files_update {
430 struct epoll_event event;
433 struct io_async_connect {
434 struct sockaddr_storage address;
437 struct io_async_msghdr {
438 struct iovec fast_iov[UIO_FASTIOV];
440 struct sockaddr __user *uaddr;
442 struct sockaddr_storage addr;
446 struct iovec fast_iov[UIO_FASTIOV];
452 struct io_async_ctx {
454 struct io_async_rw rw;
455 struct io_async_msghdr msg;
456 struct io_async_connect connect;
457 struct io_timeout_data timeout;
462 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
463 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
464 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
465 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
466 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
473 REQ_F_IOPOLL_COMPLETED_BIT,
474 REQ_F_LINK_TIMEOUT_BIT,
478 REQ_F_TIMEOUT_NOSEQ_BIT,
479 REQ_F_COMP_LOCKED_BIT,
480 REQ_F_NEED_CLEANUP_BIT,
486 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
487 /* drain existing IO first */
488 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
490 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
491 /* doesn't sever on completion < 0 */
492 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
494 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
496 /* already grabbed next link */
497 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
498 /* fail rest of links */
499 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
500 /* on inflight list */
501 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
502 /* read/write uses file position */
503 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
504 /* must not punt to workers */
505 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
506 /* polled IO has completed */
507 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
508 /* has linked timeout */
509 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
510 /* timeout request */
511 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
513 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
514 /* must be punted even for NONBLOCK */
515 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
516 /* no timeout sequence */
517 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
518 /* completion under lock */
519 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
521 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
522 /* in overflow list */
523 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
527 * NOTE! Each of the iocb union members has the file pointer
528 * as the first entry in their struct definition. So you can
529 * access the file pointer through any of the sub-structs,
530 * or directly as just 'ki_filp' in this struct.
536 struct io_poll_iocb poll;
537 struct io_accept accept;
539 struct io_cancel cancel;
540 struct io_timeout timeout;
541 struct io_connect connect;
542 struct io_sr_msg sr_msg;
544 struct io_close close;
545 struct io_files_update files_update;
546 struct io_fadvise fadvise;
547 struct io_madvise madvise;
548 struct io_epoll epoll;
551 struct io_async_ctx *io;
553 * llist_node is only used for poll deferred completions
555 struct llist_node llist_node;
557 bool needs_fixed_file;
560 struct io_ring_ctx *ctx;
562 struct list_head list;
563 struct hlist_node hash_node;
565 struct list_head link_list;
572 struct list_head inflight_entry;
574 struct io_wq_work work;
577 #define IO_PLUG_THRESHOLD 2
578 #define IO_IOPOLL_BATCH 8
580 struct io_submit_state {
581 struct blk_plug plug;
584 * io_kiocb alloc cache
586 void *reqs[IO_IOPOLL_BATCH];
587 unsigned int free_reqs;
590 * File reference cache
594 unsigned int has_refs;
595 unsigned int used_refs;
596 unsigned int ios_left;
600 /* needs req->io allocated for deferral/async */
601 unsigned async_ctx : 1;
602 /* needs current->mm setup, does mm access */
603 unsigned needs_mm : 1;
604 /* needs req->file assigned */
605 unsigned needs_file : 1;
606 /* needs req->file assigned IFF fd is >= 0 */
607 unsigned fd_non_neg : 1;
608 /* hash wq insertion if file is a regular file */
609 unsigned hash_reg_file : 1;
610 /* unbound wq insertion if file is a non-regular file */
611 unsigned unbound_nonreg_file : 1;
612 /* opcode is not supported by this kernel */
613 unsigned not_supported : 1;
614 /* needs file table */
615 unsigned file_table : 1;
617 unsigned needs_fs : 1;
620 static const struct io_op_def io_op_defs[] = {
621 [IORING_OP_NOP] = {},
622 [IORING_OP_READV] = {
626 .unbound_nonreg_file = 1,
628 [IORING_OP_WRITEV] = {
633 .unbound_nonreg_file = 1,
635 [IORING_OP_FSYNC] = {
638 [IORING_OP_READ_FIXED] = {
640 .unbound_nonreg_file = 1,
642 [IORING_OP_WRITE_FIXED] = {
645 .unbound_nonreg_file = 1,
647 [IORING_OP_POLL_ADD] = {
649 .unbound_nonreg_file = 1,
651 [IORING_OP_POLL_REMOVE] = {},
652 [IORING_OP_SYNC_FILE_RANGE] = {
655 [IORING_OP_SENDMSG] = {
659 .unbound_nonreg_file = 1,
662 [IORING_OP_RECVMSG] = {
666 .unbound_nonreg_file = 1,
669 [IORING_OP_TIMEOUT] = {
673 [IORING_OP_TIMEOUT_REMOVE] = {},
674 [IORING_OP_ACCEPT] = {
677 .unbound_nonreg_file = 1,
680 [IORING_OP_ASYNC_CANCEL] = {},
681 [IORING_OP_LINK_TIMEOUT] = {
685 [IORING_OP_CONNECT] = {
689 .unbound_nonreg_file = 1,
691 [IORING_OP_FALLOCATE] = {
694 [IORING_OP_OPENAT] = {
700 [IORING_OP_CLOSE] = {
704 [IORING_OP_FILES_UPDATE] = {
708 [IORING_OP_STATX] = {
717 .unbound_nonreg_file = 1,
719 [IORING_OP_WRITE] = {
722 .unbound_nonreg_file = 1,
724 [IORING_OP_FADVISE] = {
727 [IORING_OP_MADVISE] = {
733 .unbound_nonreg_file = 1,
738 .unbound_nonreg_file = 1,
740 [IORING_OP_OPENAT2] = {
746 [IORING_OP_EPOLL_CTL] = {
747 .unbound_nonreg_file = 1,
752 static void io_wq_submit_work(struct io_wq_work **workptr);
753 static void io_cqring_fill_event(struct io_kiocb *req, long res);
754 static void io_put_req(struct io_kiocb *req);
755 static void __io_double_put_req(struct io_kiocb *req);
756 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
757 static void io_queue_linked_timeout(struct io_kiocb *req);
758 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
759 struct io_uring_files_update *ip,
761 static int io_grab_files(struct io_kiocb *req);
762 static void io_ring_file_ref_flush(struct fixed_file_data *data);
763 static void io_cleanup_req(struct io_kiocb *req);
765 static struct kmem_cache *req_cachep;
767 static const struct file_operations io_uring_fops;
769 struct sock *io_uring_get_socket(struct file *file)
771 #if defined(CONFIG_UNIX)
772 if (file->f_op == &io_uring_fops) {
773 struct io_ring_ctx *ctx = file->private_data;
775 return ctx->ring_sock->sk;
780 EXPORT_SYMBOL(io_uring_get_socket);
782 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
784 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
786 complete(&ctx->completions[0]);
789 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
791 struct io_ring_ctx *ctx;
794 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
798 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
799 if (!ctx->fallback_req)
802 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
803 if (!ctx->completions)
807 * Use 5 bits less than the max cq entries, that should give us around
808 * 32 entries per hash list if totally full and uniformly spread.
810 hash_bits = ilog2(p->cq_entries);
814 ctx->cancel_hash_bits = hash_bits;
815 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
817 if (!ctx->cancel_hash)
819 __hash_init(ctx->cancel_hash, 1U << hash_bits);
821 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
822 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
825 ctx->flags = p->flags;
826 init_waitqueue_head(&ctx->cq_wait);
827 INIT_LIST_HEAD(&ctx->cq_overflow_list);
828 init_completion(&ctx->completions[0]);
829 init_completion(&ctx->completions[1]);
830 idr_init(&ctx->personality_idr);
831 mutex_init(&ctx->uring_lock);
832 init_waitqueue_head(&ctx->wait);
833 spin_lock_init(&ctx->completion_lock);
834 init_llist_head(&ctx->poll_llist);
835 INIT_LIST_HEAD(&ctx->poll_list);
836 INIT_LIST_HEAD(&ctx->defer_list);
837 INIT_LIST_HEAD(&ctx->timeout_list);
838 init_waitqueue_head(&ctx->inflight_wait);
839 spin_lock_init(&ctx->inflight_lock);
840 INIT_LIST_HEAD(&ctx->inflight_list);
843 if (ctx->fallback_req)
844 kmem_cache_free(req_cachep, ctx->fallback_req);
845 kfree(ctx->completions);
846 kfree(ctx->cancel_hash);
851 static inline bool __req_need_defer(struct io_kiocb *req)
853 struct io_ring_ctx *ctx = req->ctx;
855 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
856 + atomic_read(&ctx->cached_cq_overflow);
859 static inline bool req_need_defer(struct io_kiocb *req)
861 if (unlikely(req->flags & REQ_F_IO_DRAIN))
862 return __req_need_defer(req);
867 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
869 struct io_kiocb *req;
871 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
872 if (req && !req_need_defer(req)) {
873 list_del_init(&req->list);
880 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
882 struct io_kiocb *req;
884 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
886 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
888 if (!__req_need_defer(req)) {
889 list_del_init(&req->list);
897 static void __io_commit_cqring(struct io_ring_ctx *ctx)
899 struct io_rings *rings = ctx->rings;
901 /* order cqe stores with ring update */
902 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
904 if (wq_has_sleeper(&ctx->cq_wait)) {
905 wake_up_interruptible(&ctx->cq_wait);
906 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
910 static inline void io_req_work_grab_env(struct io_kiocb *req,
911 const struct io_op_def *def)
913 if (!req->work.mm && def->needs_mm) {
915 req->work.mm = current->mm;
917 if (!req->work.creds)
918 req->work.creds = get_current_cred();
919 if (!req->work.fs && def->needs_fs) {
920 spin_lock(¤t->fs->lock);
921 if (!current->fs->in_exec) {
922 req->work.fs = current->fs;
923 req->work.fs->users++;
925 req->work.flags |= IO_WQ_WORK_CANCEL;
927 spin_unlock(¤t->fs->lock);
929 if (!req->work.task_pid)
930 req->work.task_pid = task_pid_vnr(current);
933 static inline void io_req_work_drop_env(struct io_kiocb *req)
936 mmdrop(req->work.mm);
939 if (req->work.creds) {
940 put_cred(req->work.creds);
941 req->work.creds = NULL;
944 struct fs_struct *fs = req->work.fs;
946 spin_lock(&req->work.fs->lock);
949 spin_unlock(&req->work.fs->lock);
955 static inline bool io_prep_async_work(struct io_kiocb *req,
956 struct io_kiocb **link)
958 const struct io_op_def *def = &io_op_defs[req->opcode];
959 bool do_hashed = false;
961 if (req->flags & REQ_F_ISREG) {
962 if (def->hash_reg_file)
965 if (def->unbound_nonreg_file)
966 req->work.flags |= IO_WQ_WORK_UNBOUND;
969 io_req_work_grab_env(req, def);
971 *link = io_prep_linked_timeout(req);
975 static inline void io_queue_async_work(struct io_kiocb *req)
977 struct io_ring_ctx *ctx = req->ctx;
978 struct io_kiocb *link;
981 do_hashed = io_prep_async_work(req, &link);
983 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
986 io_wq_enqueue(ctx->io_wq, &req->work);
988 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
989 file_inode(req->file));
993 io_queue_linked_timeout(link);
996 static void io_kill_timeout(struct io_kiocb *req)
1000 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1002 atomic_inc(&req->ctx->cq_timeouts);
1003 list_del_init(&req->list);
1004 req->flags |= REQ_F_COMP_LOCKED;
1005 io_cqring_fill_event(req, 0);
1010 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1012 struct io_kiocb *req, *tmp;
1014 spin_lock_irq(&ctx->completion_lock);
1015 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1016 io_kill_timeout(req);
1017 spin_unlock_irq(&ctx->completion_lock);
1020 static void io_commit_cqring(struct io_ring_ctx *ctx)
1022 struct io_kiocb *req;
1024 while ((req = io_get_timeout_req(ctx)) != NULL)
1025 io_kill_timeout(req);
1027 __io_commit_cqring(ctx);
1029 while ((req = io_get_deferred_req(ctx)) != NULL)
1030 io_queue_async_work(req);
1033 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1035 struct io_rings *rings = ctx->rings;
1038 tail = ctx->cached_cq_tail;
1040 * writes to the cq entry need to come after reading head; the
1041 * control dependency is enough as we're using WRITE_ONCE to
1044 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1047 ctx->cached_cq_tail++;
1048 return &rings->cqes[tail & ctx->cq_mask];
1051 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1055 if (!ctx->eventfd_async)
1057 return io_wq_current_is_worker() || in_interrupt();
1060 static void __io_cqring_ev_posted(struct io_ring_ctx *ctx, bool trigger_ev)
1062 if (waitqueue_active(&ctx->wait))
1063 wake_up(&ctx->wait);
1064 if (waitqueue_active(&ctx->sqo_wait))
1065 wake_up(&ctx->sqo_wait);
1067 eventfd_signal(ctx->cq_ev_fd, 1);
1070 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1072 __io_cqring_ev_posted(ctx, io_should_trigger_evfd(ctx));
1075 /* Returns true if there are no backlogged entries after the flush */
1076 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1078 struct io_rings *rings = ctx->rings;
1079 struct io_uring_cqe *cqe;
1080 struct io_kiocb *req;
1081 unsigned long flags;
1085 if (list_empty_careful(&ctx->cq_overflow_list))
1087 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1088 rings->cq_ring_entries))
1092 spin_lock_irqsave(&ctx->completion_lock, flags);
1094 /* if force is set, the ring is going away. always drop after that */
1096 ctx->cq_overflow_flushed = 1;
1099 while (!list_empty(&ctx->cq_overflow_list)) {
1100 cqe = io_get_cqring(ctx);
1104 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1106 list_move(&req->list, &list);
1107 req->flags &= ~REQ_F_OVERFLOW;
1109 WRITE_ONCE(cqe->user_data, req->user_data);
1110 WRITE_ONCE(cqe->res, req->result);
1111 WRITE_ONCE(cqe->flags, 0);
1113 WRITE_ONCE(ctx->rings->cq_overflow,
1114 atomic_inc_return(&ctx->cached_cq_overflow));
1118 io_commit_cqring(ctx);
1120 clear_bit(0, &ctx->sq_check_overflow);
1121 clear_bit(0, &ctx->cq_check_overflow);
1123 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1124 io_cqring_ev_posted(ctx);
1126 while (!list_empty(&list)) {
1127 req = list_first_entry(&list, struct io_kiocb, list);
1128 list_del(&req->list);
1135 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1137 struct io_ring_ctx *ctx = req->ctx;
1138 struct io_uring_cqe *cqe;
1140 trace_io_uring_complete(ctx, req->user_data, res);
1143 * If we can't get a cq entry, userspace overflowed the
1144 * submission (by quite a lot). Increment the overflow count in
1147 cqe = io_get_cqring(ctx);
1149 WRITE_ONCE(cqe->user_data, req->user_data);
1150 WRITE_ONCE(cqe->res, res);
1151 WRITE_ONCE(cqe->flags, 0);
1152 } else if (ctx->cq_overflow_flushed) {
1153 WRITE_ONCE(ctx->rings->cq_overflow,
1154 atomic_inc_return(&ctx->cached_cq_overflow));
1156 if (list_empty(&ctx->cq_overflow_list)) {
1157 set_bit(0, &ctx->sq_check_overflow);
1158 set_bit(0, &ctx->cq_check_overflow);
1160 req->flags |= REQ_F_OVERFLOW;
1161 refcount_inc(&req->refs);
1163 list_add_tail(&req->list, &ctx->cq_overflow_list);
1167 static void io_cqring_add_event(struct io_kiocb *req, long res)
1169 struct io_ring_ctx *ctx = req->ctx;
1170 unsigned long flags;
1172 spin_lock_irqsave(&ctx->completion_lock, flags);
1173 io_cqring_fill_event(req, res);
1174 io_commit_cqring(ctx);
1175 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1177 io_cqring_ev_posted(ctx);
1180 static inline bool io_is_fallback_req(struct io_kiocb *req)
1182 return req == (struct io_kiocb *)
1183 ((unsigned long) req->ctx->fallback_req & ~1UL);
1186 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1188 struct io_kiocb *req;
1190 req = ctx->fallback_req;
1191 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
1197 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
1198 struct io_submit_state *state)
1200 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1201 struct io_kiocb *req;
1204 req = kmem_cache_alloc(req_cachep, gfp);
1207 } else if (!state->free_reqs) {
1211 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1212 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1215 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1216 * retry single alloc to be on the safe side.
1218 if (unlikely(ret <= 0)) {
1219 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1220 if (!state->reqs[0])
1224 state->free_reqs = ret - 1;
1225 req = state->reqs[ret - 1];
1228 req = state->reqs[state->free_reqs];
1236 /* one is dropped after submission, the other at completion */
1237 refcount_set(&req->refs, 2);
1239 INIT_IO_WORK(&req->work, io_wq_submit_work);
1242 req = io_get_fallback_req(ctx);
1245 percpu_ref_put(&ctx->refs);
1249 static void __io_req_do_free(struct io_kiocb *req)
1251 if (likely(!io_is_fallback_req(req)))
1252 kmem_cache_free(req_cachep, req);
1254 clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req);
1257 static void __io_req_aux_free(struct io_kiocb *req)
1259 struct io_ring_ctx *ctx = req->ctx;
1261 if (req->flags & REQ_F_NEED_CLEANUP)
1262 io_cleanup_req(req);
1266 if (req->flags & REQ_F_FIXED_FILE)
1267 percpu_ref_put(&ctx->file_data->refs);
1272 io_req_work_drop_env(req);
1275 static void __io_free_req(struct io_kiocb *req)
1277 __io_req_aux_free(req);
1279 if (req->flags & REQ_F_INFLIGHT) {
1280 struct io_ring_ctx *ctx = req->ctx;
1281 unsigned long flags;
1283 spin_lock_irqsave(&ctx->inflight_lock, flags);
1284 list_del(&req->inflight_entry);
1285 if (waitqueue_active(&ctx->inflight_wait))
1286 wake_up(&ctx->inflight_wait);
1287 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1290 percpu_ref_put(&req->ctx->refs);
1291 __io_req_do_free(req);
1295 void *reqs[IO_IOPOLL_BATCH];
1300 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1302 int fixed_refs = rb->to_free;
1306 if (rb->need_iter) {
1307 int i, inflight = 0;
1308 unsigned long flags;
1311 for (i = 0; i < rb->to_free; i++) {
1312 struct io_kiocb *req = rb->reqs[i];
1314 if (req->flags & REQ_F_FIXED_FILE) {
1318 if (req->flags & REQ_F_INFLIGHT)
1320 __io_req_aux_free(req);
1325 spin_lock_irqsave(&ctx->inflight_lock, flags);
1326 for (i = 0; i < rb->to_free; i++) {
1327 struct io_kiocb *req = rb->reqs[i];
1329 if (req->flags & REQ_F_INFLIGHT) {
1330 list_del(&req->inflight_entry);
1335 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1337 if (waitqueue_active(&ctx->inflight_wait))
1338 wake_up(&ctx->inflight_wait);
1341 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1343 percpu_ref_put_many(&ctx->file_data->refs, fixed_refs);
1344 percpu_ref_put_many(&ctx->refs, rb->to_free);
1345 rb->to_free = rb->need_iter = 0;
1348 static bool io_link_cancel_timeout(struct io_kiocb *req)
1350 struct io_ring_ctx *ctx = req->ctx;
1353 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1355 io_cqring_fill_event(req, -ECANCELED);
1356 io_commit_cqring(ctx);
1357 req->flags &= ~REQ_F_LINK;
1365 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1367 struct io_ring_ctx *ctx = req->ctx;
1368 bool wake_ev = false;
1370 /* Already got next link */
1371 if (req->flags & REQ_F_LINK_NEXT)
1375 * The list should never be empty when we are called here. But could
1376 * potentially happen if the chain is messed up, check to be on the
1379 while (!list_empty(&req->link_list)) {
1380 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1381 struct io_kiocb, link_list);
1383 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1384 (nxt->flags & REQ_F_TIMEOUT))) {
1385 list_del_init(&nxt->link_list);
1386 wake_ev |= io_link_cancel_timeout(nxt);
1387 req->flags &= ~REQ_F_LINK_TIMEOUT;
1391 list_del_init(&req->link_list);
1392 if (!list_empty(&nxt->link_list))
1393 nxt->flags |= REQ_F_LINK;
1398 req->flags |= REQ_F_LINK_NEXT;
1400 io_cqring_ev_posted(ctx);
1404 * Called if REQ_F_LINK is set, and we fail the head request
1406 static void io_fail_links(struct io_kiocb *req)
1408 struct io_ring_ctx *ctx = req->ctx;
1409 unsigned long flags;
1411 spin_lock_irqsave(&ctx->completion_lock, flags);
1413 while (!list_empty(&req->link_list)) {
1414 struct io_kiocb *link = list_first_entry(&req->link_list,
1415 struct io_kiocb, link_list);
1417 list_del_init(&link->link_list);
1418 trace_io_uring_fail_link(req, link);
1420 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1421 link->opcode == IORING_OP_LINK_TIMEOUT) {
1422 io_link_cancel_timeout(link);
1424 io_cqring_fill_event(link, -ECANCELED);
1425 __io_double_put_req(link);
1427 req->flags &= ~REQ_F_LINK_TIMEOUT;
1430 io_commit_cqring(ctx);
1431 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1432 io_cqring_ev_posted(ctx);
1435 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1437 if (likely(!(req->flags & REQ_F_LINK)))
1441 * If LINK is set, we have dependent requests in this chain. If we
1442 * didn't fail this request, queue the first one up, moving any other
1443 * dependencies to the next request. In case of failure, fail the rest
1446 if (req->flags & REQ_F_FAIL_LINK) {
1448 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1449 REQ_F_LINK_TIMEOUT) {
1450 struct io_ring_ctx *ctx = req->ctx;
1451 unsigned long flags;
1454 * If this is a timeout link, we could be racing with the
1455 * timeout timer. Grab the completion lock for this case to
1456 * protect against that.
1458 spin_lock_irqsave(&ctx->completion_lock, flags);
1459 io_req_link_next(req, nxt);
1460 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1462 io_req_link_next(req, nxt);
1466 static void io_free_req(struct io_kiocb *req)
1468 struct io_kiocb *nxt = NULL;
1470 io_req_find_next(req, &nxt);
1474 io_queue_async_work(nxt);
1478 * Drop reference to request, return next in chain (if there is one) if this
1479 * was the last reference to this request.
1481 __attribute__((nonnull))
1482 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1484 if (refcount_dec_and_test(&req->refs)) {
1485 io_req_find_next(req, nxtptr);
1490 static void io_put_req(struct io_kiocb *req)
1492 if (refcount_dec_and_test(&req->refs))
1497 * Must only be used if we don't need to care about links, usually from
1498 * within the completion handling itself.
1500 static void __io_double_put_req(struct io_kiocb *req)
1502 /* drop both submit and complete references */
1503 if (refcount_sub_and_test(2, &req->refs))
1507 static void io_double_put_req(struct io_kiocb *req)
1509 /* drop both submit and complete references */
1510 if (refcount_sub_and_test(2, &req->refs))
1514 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1516 struct io_rings *rings = ctx->rings;
1518 if (test_bit(0, &ctx->cq_check_overflow)) {
1520 * noflush == true is from the waitqueue handler, just ensure
1521 * we wake up the task, and the next invocation will flush the
1522 * entries. We cannot safely to it from here.
1524 if (noflush && !list_empty(&ctx->cq_overflow_list))
1527 io_cqring_overflow_flush(ctx, false);
1530 /* See comment at the top of this file */
1532 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1535 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1537 struct io_rings *rings = ctx->rings;
1539 /* make sure SQ entry isn't read before tail */
1540 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1543 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1545 if ((req->flags & REQ_F_LINK) || io_is_fallback_req(req))
1548 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1551 rb->reqs[rb->to_free++] = req;
1552 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1553 io_free_req_many(req->ctx, rb);
1558 * Find and free completed poll iocbs
1560 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1561 struct list_head *done)
1563 struct req_batch rb;
1564 struct io_kiocb *req;
1566 rb.to_free = rb.need_iter = 0;
1567 while (!list_empty(done)) {
1568 req = list_first_entry(done, struct io_kiocb, list);
1569 list_del(&req->list);
1571 io_cqring_fill_event(req, req->result);
1574 if (refcount_dec_and_test(&req->refs) &&
1575 !io_req_multi_free(&rb, req))
1579 io_commit_cqring(ctx);
1580 io_free_req_many(ctx, &rb);
1583 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1586 struct io_kiocb *req, *tmp;
1592 * Only spin for completions if we don't have multiple devices hanging
1593 * off our complete list, and we're under the requested amount.
1595 spin = !ctx->poll_multi_file && *nr_events < min;
1598 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1599 struct kiocb *kiocb = &req->rw.kiocb;
1602 * Move completed entries to our local list. If we find a
1603 * request that requires polling, break out and complete
1604 * the done list first, if we have entries there.
1606 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1607 list_move_tail(&req->list, &done);
1610 if (!list_empty(&done))
1613 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1622 if (!list_empty(&done))
1623 io_iopoll_complete(ctx, nr_events, &done);
1629 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1630 * non-spinning poll check - we'll still enter the driver poll loop, but only
1631 * as a non-spinning completion check.
1633 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1636 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1639 ret = io_do_iopoll(ctx, nr_events, min);
1642 if (!min || *nr_events >= min)
1650 * We can't just wait for polled events to come to us, we have to actively
1651 * find and complete them.
1653 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1655 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1658 mutex_lock(&ctx->uring_lock);
1659 while (!list_empty(&ctx->poll_list)) {
1660 unsigned int nr_events = 0;
1662 io_iopoll_getevents(ctx, &nr_events, 1);
1665 * Ensure we allow local-to-the-cpu processing to take place,
1666 * in this case we need to ensure that we reap all events.
1670 mutex_unlock(&ctx->uring_lock);
1673 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1676 int iters = 0, ret = 0;
1679 * We disallow the app entering submit/complete with polling, but we
1680 * still need to lock the ring to prevent racing with polled issue
1681 * that got punted to a workqueue.
1683 mutex_lock(&ctx->uring_lock);
1688 * Don't enter poll loop if we already have events pending.
1689 * If we do, we can potentially be spinning for commands that
1690 * already triggered a CQE (eg in error).
1692 if (io_cqring_events(ctx, false))
1696 * If a submit got punted to a workqueue, we can have the
1697 * application entering polling for a command before it gets
1698 * issued. That app will hold the uring_lock for the duration
1699 * of the poll right here, so we need to take a breather every
1700 * now and then to ensure that the issue has a chance to add
1701 * the poll to the issued list. Otherwise we can spin here
1702 * forever, while the workqueue is stuck trying to acquire the
1705 if (!(++iters & 7)) {
1706 mutex_unlock(&ctx->uring_lock);
1707 mutex_lock(&ctx->uring_lock);
1710 if (*nr_events < min)
1711 tmin = min - *nr_events;
1713 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1717 } while (min && !*nr_events && !need_resched());
1719 mutex_unlock(&ctx->uring_lock);
1723 static void kiocb_end_write(struct io_kiocb *req)
1726 * Tell lockdep we inherited freeze protection from submission
1729 if (req->flags & REQ_F_ISREG) {
1730 struct inode *inode = file_inode(req->file);
1732 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1734 file_end_write(req->file);
1737 static inline void req_set_fail_links(struct io_kiocb *req)
1739 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1740 req->flags |= REQ_F_FAIL_LINK;
1743 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1745 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1747 if (kiocb->ki_flags & IOCB_WRITE)
1748 kiocb_end_write(req);
1750 if (res != req->result)
1751 req_set_fail_links(req);
1752 io_cqring_add_event(req, res);
1755 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1757 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1759 io_complete_rw_common(kiocb, res);
1763 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1765 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1766 struct io_kiocb *nxt = NULL;
1768 io_complete_rw_common(kiocb, res);
1769 io_put_req_find_next(req, &nxt);
1774 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1776 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1778 if (kiocb->ki_flags & IOCB_WRITE)
1779 kiocb_end_write(req);
1781 if (res != req->result)
1782 req_set_fail_links(req);
1785 req->flags |= REQ_F_IOPOLL_COMPLETED;
1789 * After the iocb has been issued, it's safe to be found on the poll list.
1790 * Adding the kiocb to the list AFTER submission ensures that we don't
1791 * find it from a io_iopoll_getevents() thread before the issuer is done
1792 * accessing the kiocb cookie.
1794 static void io_iopoll_req_issued(struct io_kiocb *req)
1796 struct io_ring_ctx *ctx = req->ctx;
1799 * Track whether we have multiple files in our lists. This will impact
1800 * how we do polling eventually, not spinning if we're on potentially
1801 * different devices.
1803 if (list_empty(&ctx->poll_list)) {
1804 ctx->poll_multi_file = false;
1805 } else if (!ctx->poll_multi_file) {
1806 struct io_kiocb *list_req;
1808 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1810 if (list_req->file != req->file)
1811 ctx->poll_multi_file = true;
1815 * For fast devices, IO may have already completed. If it has, add
1816 * it to the front so we find it first.
1818 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1819 list_add(&req->list, &ctx->poll_list);
1821 list_add_tail(&req->list, &ctx->poll_list);
1823 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1824 wq_has_sleeper(&ctx->sqo_wait))
1825 wake_up(&ctx->sqo_wait);
1828 static void io_file_put(struct io_submit_state *state)
1831 int diff = state->has_refs - state->used_refs;
1834 fput_many(state->file, diff);
1840 * Get as many references to a file as we have IOs left in this submission,
1841 * assuming most submissions are for one file, or at least that each file
1842 * has more than one submission.
1844 static struct file *io_file_get(struct io_submit_state *state, int fd)
1850 if (state->fd == fd) {
1857 state->file = fget_many(fd, state->ios_left);
1862 state->has_refs = state->ios_left;
1863 state->used_refs = 1;
1869 * If we tracked the file through the SCM inflight mechanism, we could support
1870 * any file. For now, just ensure that anything potentially problematic is done
1873 static bool io_file_supports_async(struct file *file)
1875 umode_t mode = file_inode(file)->i_mode;
1877 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
1879 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1885 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1886 bool force_nonblock)
1888 struct io_ring_ctx *ctx = req->ctx;
1889 struct kiocb *kiocb = &req->rw.kiocb;
1893 if (S_ISREG(file_inode(req->file)->i_mode))
1894 req->flags |= REQ_F_ISREG;
1896 kiocb->ki_pos = READ_ONCE(sqe->off);
1897 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
1898 req->flags |= REQ_F_CUR_POS;
1899 kiocb->ki_pos = req->file->f_pos;
1901 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1902 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1903 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1907 ioprio = READ_ONCE(sqe->ioprio);
1909 ret = ioprio_check_cap(ioprio);
1913 kiocb->ki_ioprio = ioprio;
1915 kiocb->ki_ioprio = get_current_ioprio();
1917 /* don't allow async punt if RWF_NOWAIT was requested */
1918 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1919 (req->file->f_flags & O_NONBLOCK))
1920 req->flags |= REQ_F_NOWAIT;
1923 kiocb->ki_flags |= IOCB_NOWAIT;
1925 if (ctx->flags & IORING_SETUP_IOPOLL) {
1926 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1927 !kiocb->ki_filp->f_op->iopoll)
1930 kiocb->ki_flags |= IOCB_HIPRI;
1931 kiocb->ki_complete = io_complete_rw_iopoll;
1934 if (kiocb->ki_flags & IOCB_HIPRI)
1936 kiocb->ki_complete = io_complete_rw;
1939 req->rw.addr = READ_ONCE(sqe->addr);
1940 req->rw.len = READ_ONCE(sqe->len);
1941 /* we own ->private, reuse it for the buffer index */
1942 req->rw.kiocb.private = (void *) (unsigned long)
1943 READ_ONCE(sqe->buf_index);
1947 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1953 case -ERESTARTNOINTR:
1954 case -ERESTARTNOHAND:
1955 case -ERESTART_RESTARTBLOCK:
1957 * We can't just restart the syscall, since previously
1958 * submitted sqes may already be in progress. Just fail this
1964 kiocb->ki_complete(kiocb, ret, 0);
1968 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1971 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1973 if (req->flags & REQ_F_CUR_POS)
1974 req->file->f_pos = kiocb->ki_pos;
1975 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1976 *nxt = __io_complete_rw(kiocb, ret);
1978 io_rw_done(kiocb, ret);
1981 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
1982 struct iov_iter *iter)
1984 struct io_ring_ctx *ctx = req->ctx;
1985 size_t len = req->rw.len;
1986 struct io_mapped_ubuf *imu;
1987 unsigned index, buf_index;
1991 /* attempt to use fixed buffers without having provided iovecs */
1992 if (unlikely(!ctx->user_bufs))
1995 buf_index = (unsigned long) req->rw.kiocb.private;
1996 if (unlikely(buf_index >= ctx->nr_user_bufs))
1999 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2000 imu = &ctx->user_bufs[index];
2001 buf_addr = req->rw.addr;
2004 if (buf_addr + len < buf_addr)
2006 /* not inside the mapped region */
2007 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2011 * May not be a start of buffer, set size appropriately
2012 * and advance us to the beginning.
2014 offset = buf_addr - imu->ubuf;
2015 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2019 * Don't use iov_iter_advance() here, as it's really slow for
2020 * using the latter parts of a big fixed buffer - it iterates
2021 * over each segment manually. We can cheat a bit here, because
2024 * 1) it's a BVEC iter, we set it up
2025 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2026 * first and last bvec
2028 * So just find our index, and adjust the iterator afterwards.
2029 * If the offset is within the first bvec (or the whole first
2030 * bvec, just use iov_iter_advance(). This makes it easier
2031 * since we can just skip the first segment, which may not
2032 * be PAGE_SIZE aligned.
2034 const struct bio_vec *bvec = imu->bvec;
2036 if (offset <= bvec->bv_len) {
2037 iov_iter_advance(iter, offset);
2039 unsigned long seg_skip;
2041 /* skip first vec */
2042 offset -= bvec->bv_len;
2043 seg_skip = 1 + (offset >> PAGE_SHIFT);
2045 iter->bvec = bvec + seg_skip;
2046 iter->nr_segs -= seg_skip;
2047 iter->count -= bvec->bv_len + offset;
2048 iter->iov_offset = offset & ~PAGE_MASK;
2055 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2056 struct iovec **iovec, struct iov_iter *iter)
2058 void __user *buf = u64_to_user_ptr(req->rw.addr);
2059 size_t sqe_len = req->rw.len;
2062 opcode = req->opcode;
2063 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2065 return io_import_fixed(req, rw, iter);
2068 /* buffer index only valid with fixed read/write */
2069 if (req->rw.kiocb.private)
2072 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2074 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2076 return ret < 0 ? ret : sqe_len;
2080 struct io_async_rw *iorw = &req->io->rw;
2083 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2084 if (iorw->iov == iorw->fast_iov)
2089 #ifdef CONFIG_COMPAT
2090 if (req->ctx->compat)
2091 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2095 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2099 * For files that don't have ->read_iter() and ->write_iter(), handle them
2100 * by looping over ->read() or ->write() manually.
2102 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2103 struct iov_iter *iter)
2108 * Don't support polled IO through this interface, and we can't
2109 * support non-blocking either. For the latter, this just causes
2110 * the kiocb to be handled from an async context.
2112 if (kiocb->ki_flags & IOCB_HIPRI)
2114 if (kiocb->ki_flags & IOCB_NOWAIT)
2117 while (iov_iter_count(iter)) {
2121 if (!iov_iter_is_bvec(iter)) {
2122 iovec = iov_iter_iovec(iter);
2124 /* fixed buffers import bvec */
2125 iovec.iov_base = kmap(iter->bvec->bv_page)
2127 iovec.iov_len = min(iter->count,
2128 iter->bvec->bv_len - iter->iov_offset);
2132 nr = file->f_op->read(file, iovec.iov_base,
2133 iovec.iov_len, &kiocb->ki_pos);
2135 nr = file->f_op->write(file, iovec.iov_base,
2136 iovec.iov_len, &kiocb->ki_pos);
2139 if (iov_iter_is_bvec(iter))
2140 kunmap(iter->bvec->bv_page);
2148 if (nr != iovec.iov_len)
2150 iov_iter_advance(iter, nr);
2156 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2157 struct iovec *iovec, struct iovec *fast_iov,
2158 struct iov_iter *iter)
2160 req->io->rw.nr_segs = iter->nr_segs;
2161 req->io->rw.size = io_size;
2162 req->io->rw.iov = iovec;
2163 if (!req->io->rw.iov) {
2164 req->io->rw.iov = req->io->rw.fast_iov;
2165 memcpy(req->io->rw.iov, fast_iov,
2166 sizeof(struct iovec) * iter->nr_segs);
2168 req->flags |= REQ_F_NEED_CLEANUP;
2172 static int io_alloc_async_ctx(struct io_kiocb *req)
2174 if (!io_op_defs[req->opcode].async_ctx)
2176 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2177 return req->io == NULL;
2180 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2181 struct iovec *iovec, struct iovec *fast_iov,
2182 struct iov_iter *iter)
2184 if (!io_op_defs[req->opcode].async_ctx)
2187 if (io_alloc_async_ctx(req))
2190 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2195 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2196 bool force_nonblock)
2198 struct io_async_ctx *io;
2199 struct iov_iter iter;
2202 ret = io_prep_rw(req, sqe, force_nonblock);
2206 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2209 /* either don't need iovec imported or already have it */
2210 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2214 io->rw.iov = io->rw.fast_iov;
2216 ret = io_import_iovec(READ, req, &io->rw.iov, &iter);
2221 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2225 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
2226 bool force_nonblock)
2228 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2229 struct kiocb *kiocb = &req->rw.kiocb;
2230 struct iov_iter iter;
2232 ssize_t io_size, ret;
2234 ret = io_import_iovec(READ, req, &iovec, &iter);
2238 /* Ensure we clear previously set non-block flag */
2239 if (!force_nonblock)
2240 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2244 if (req->flags & REQ_F_LINK)
2245 req->result = io_size;
2248 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2249 * we know to async punt it even if it was opened O_NONBLOCK
2251 if (force_nonblock && !io_file_supports_async(req->file)) {
2252 req->flags |= REQ_F_MUST_PUNT;
2256 iov_count = iov_iter_count(&iter);
2257 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2261 if (req->file->f_op->read_iter)
2262 ret2 = call_read_iter(req->file, kiocb, &iter);
2264 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2266 /* Catch -EAGAIN return for forced non-blocking submission */
2267 if (!force_nonblock || ret2 != -EAGAIN) {
2268 kiocb_done(kiocb, ret2, nxt, req->in_async);
2271 ret = io_setup_async_rw(req, io_size, iovec,
2272 inline_vecs, &iter);
2280 req->flags &= ~REQ_F_NEED_CLEANUP;
2284 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2285 bool force_nonblock)
2287 struct io_async_ctx *io;
2288 struct iov_iter iter;
2291 ret = io_prep_rw(req, sqe, force_nonblock);
2295 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2298 /* either don't need iovec imported or already have it */
2299 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2303 io->rw.iov = io->rw.fast_iov;
2305 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter);
2310 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2314 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
2315 bool force_nonblock)
2317 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2318 struct kiocb *kiocb = &req->rw.kiocb;
2319 struct iov_iter iter;
2321 ssize_t ret, io_size;
2323 ret = io_import_iovec(WRITE, req, &iovec, &iter);
2327 /* Ensure we clear previously set non-block flag */
2328 if (!force_nonblock)
2329 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2333 if (req->flags & REQ_F_LINK)
2334 req->result = io_size;
2337 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2338 * we know to async punt it even if it was opened O_NONBLOCK
2340 if (force_nonblock && !io_file_supports_async(req->file)) {
2341 req->flags |= REQ_F_MUST_PUNT;
2345 /* file path doesn't support NOWAIT for non-direct_IO */
2346 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2347 (req->flags & REQ_F_ISREG))
2350 iov_count = iov_iter_count(&iter);
2351 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2356 * Open-code file_start_write here to grab freeze protection,
2357 * which will be released by another thread in
2358 * io_complete_rw(). Fool lockdep by telling it the lock got
2359 * released so that it doesn't complain about the held lock when
2360 * we return to userspace.
2362 if (req->flags & REQ_F_ISREG) {
2363 __sb_start_write(file_inode(req->file)->i_sb,
2364 SB_FREEZE_WRITE, true);
2365 __sb_writers_release(file_inode(req->file)->i_sb,
2368 kiocb->ki_flags |= IOCB_WRITE;
2370 if (req->file->f_op->write_iter)
2371 ret2 = call_write_iter(req->file, kiocb, &iter);
2373 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2375 * Raw bdev writes will -EOPNOTSUPP for IOCB_NOWAIT. Just
2376 * retry them without IOCB_NOWAIT.
2378 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2380 if (!force_nonblock || ret2 != -EAGAIN) {
2381 kiocb_done(kiocb, ret2, nxt, req->in_async);
2384 ret = io_setup_async_rw(req, io_size, iovec,
2385 inline_vecs, &iter);
2392 req->flags &= ~REQ_F_NEED_CLEANUP;
2398 * IORING_OP_NOP just posts a completion event, nothing else.
2400 static int io_nop(struct io_kiocb *req)
2402 struct io_ring_ctx *ctx = req->ctx;
2404 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2407 io_cqring_add_event(req, 0);
2412 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2414 struct io_ring_ctx *ctx = req->ctx;
2419 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2421 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2424 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2425 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2428 req->sync.off = READ_ONCE(sqe->off);
2429 req->sync.len = READ_ONCE(sqe->len);
2433 static bool io_req_cancelled(struct io_kiocb *req)
2435 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2436 req_set_fail_links(req);
2437 io_cqring_add_event(req, -ECANCELED);
2445 static void io_link_work_cb(struct io_wq_work **workptr)
2447 struct io_wq_work *work = *workptr;
2448 struct io_kiocb *link = work->data;
2450 io_queue_linked_timeout(link);
2451 work->func = io_wq_submit_work;
2454 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
2456 struct io_kiocb *link;
2458 io_prep_async_work(nxt, &link);
2459 *workptr = &nxt->work;
2461 nxt->work.flags |= IO_WQ_WORK_CB;
2462 nxt->work.func = io_link_work_cb;
2463 nxt->work.data = link;
2467 static void io_fsync_finish(struct io_wq_work **workptr)
2469 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2470 loff_t end = req->sync.off + req->sync.len;
2471 struct io_kiocb *nxt = NULL;
2474 if (io_req_cancelled(req))
2477 ret = vfs_fsync_range(req->file, req->sync.off,
2478 end > 0 ? end : LLONG_MAX,
2479 req->sync.flags & IORING_FSYNC_DATASYNC);
2481 req_set_fail_links(req);
2482 io_cqring_add_event(req, ret);
2483 io_put_req_find_next(req, &nxt);
2485 io_wq_assign_next(workptr, nxt);
2488 static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt,
2489 bool force_nonblock)
2491 struct io_wq_work *work, *old_work;
2493 /* fsync always requires a blocking context */
2494 if (force_nonblock) {
2496 req->work.func = io_fsync_finish;
2500 work = old_work = &req->work;
2501 io_fsync_finish(&work);
2502 if (work && work != old_work)
2503 *nxt = container_of(work, struct io_kiocb, work);
2507 static void io_fallocate_finish(struct io_wq_work **workptr)
2509 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2510 struct io_kiocb *nxt = NULL;
2513 if (io_req_cancelled(req))
2516 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2519 req_set_fail_links(req);
2520 io_cqring_add_event(req, ret);
2521 io_put_req_find_next(req, &nxt);
2523 io_wq_assign_next(workptr, nxt);
2526 static int io_fallocate_prep(struct io_kiocb *req,
2527 const struct io_uring_sqe *sqe)
2529 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2532 req->sync.off = READ_ONCE(sqe->off);
2533 req->sync.len = READ_ONCE(sqe->addr);
2534 req->sync.mode = READ_ONCE(sqe->len);
2538 static int io_fallocate(struct io_kiocb *req, struct io_kiocb **nxt,
2539 bool force_nonblock)
2541 struct io_wq_work *work, *old_work;
2543 /* fallocate always requiring blocking context */
2544 if (force_nonblock) {
2546 req->work.func = io_fallocate_finish;
2550 work = old_work = &req->work;
2551 io_fallocate_finish(&work);
2552 if (work && work != old_work)
2553 *nxt = container_of(work, struct io_kiocb, work);
2558 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2560 const char __user *fname;
2563 if (sqe->ioprio || sqe->buf_index)
2565 if (sqe->flags & IOSQE_FIXED_FILE)
2567 if (req->flags & REQ_F_NEED_CLEANUP)
2570 req->open.dfd = READ_ONCE(sqe->fd);
2571 req->open.how.mode = READ_ONCE(sqe->len);
2572 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2573 req->open.how.flags = READ_ONCE(sqe->open_flags);
2575 req->open.filename = getname(fname);
2576 if (IS_ERR(req->open.filename)) {
2577 ret = PTR_ERR(req->open.filename);
2578 req->open.filename = NULL;
2582 req->open.nofile = rlimit(RLIMIT_NOFILE);
2583 req->flags |= REQ_F_NEED_CLEANUP;
2587 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2589 struct open_how __user *how;
2590 const char __user *fname;
2594 if (sqe->ioprio || sqe->buf_index)
2596 if (sqe->flags & IOSQE_FIXED_FILE)
2598 if (req->flags & REQ_F_NEED_CLEANUP)
2601 req->open.dfd = READ_ONCE(sqe->fd);
2602 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2603 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2604 len = READ_ONCE(sqe->len);
2606 if (len < OPEN_HOW_SIZE_VER0)
2609 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2614 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2615 req->open.how.flags |= O_LARGEFILE;
2617 req->open.filename = getname(fname);
2618 if (IS_ERR(req->open.filename)) {
2619 ret = PTR_ERR(req->open.filename);
2620 req->open.filename = NULL;
2624 req->open.nofile = rlimit(RLIMIT_NOFILE);
2625 req->flags |= REQ_F_NEED_CLEANUP;
2629 static int io_openat2(struct io_kiocb *req, struct io_kiocb **nxt,
2630 bool force_nonblock)
2632 struct open_flags op;
2639 ret = build_open_flags(&req->open.how, &op);
2643 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
2647 file = do_filp_open(req->open.dfd, req->open.filename, &op);
2650 ret = PTR_ERR(file);
2652 fsnotify_open(file);
2653 fd_install(ret, file);
2656 putname(req->open.filename);
2657 req->flags &= ~REQ_F_NEED_CLEANUP;
2659 req_set_fail_links(req);
2660 io_cqring_add_event(req, ret);
2661 io_put_req_find_next(req, nxt);
2665 static int io_openat(struct io_kiocb *req, struct io_kiocb **nxt,
2666 bool force_nonblock)
2668 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
2669 return io_openat2(req, nxt, force_nonblock);
2672 static int io_epoll_ctl_prep(struct io_kiocb *req,
2673 const struct io_uring_sqe *sqe)
2675 #if defined(CONFIG_EPOLL)
2676 if (sqe->ioprio || sqe->buf_index)
2679 req->epoll.epfd = READ_ONCE(sqe->fd);
2680 req->epoll.op = READ_ONCE(sqe->len);
2681 req->epoll.fd = READ_ONCE(sqe->off);
2683 if (ep_op_has_event(req->epoll.op)) {
2684 struct epoll_event __user *ev;
2686 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
2687 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
2697 static int io_epoll_ctl(struct io_kiocb *req, struct io_kiocb **nxt,
2698 bool force_nonblock)
2700 #if defined(CONFIG_EPOLL)
2701 struct io_epoll *ie = &req->epoll;
2704 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
2705 if (force_nonblock && ret == -EAGAIN)
2709 req_set_fail_links(req);
2710 io_cqring_add_event(req, ret);
2711 io_put_req_find_next(req, nxt);
2718 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2720 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2721 if (sqe->ioprio || sqe->buf_index || sqe->off)
2724 req->madvise.addr = READ_ONCE(sqe->addr);
2725 req->madvise.len = READ_ONCE(sqe->len);
2726 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
2733 static int io_madvise(struct io_kiocb *req, struct io_kiocb **nxt,
2734 bool force_nonblock)
2736 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2737 struct io_madvise *ma = &req->madvise;
2743 ret = do_madvise(ma->addr, ma->len, ma->advice);
2745 req_set_fail_links(req);
2746 io_cqring_add_event(req, ret);
2747 io_put_req_find_next(req, nxt);
2754 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2756 if (sqe->ioprio || sqe->buf_index || sqe->addr)
2759 req->fadvise.offset = READ_ONCE(sqe->off);
2760 req->fadvise.len = READ_ONCE(sqe->len);
2761 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
2765 static int io_fadvise(struct io_kiocb *req, struct io_kiocb **nxt,
2766 bool force_nonblock)
2768 struct io_fadvise *fa = &req->fadvise;
2771 if (force_nonblock) {
2772 switch (fa->advice) {
2773 case POSIX_FADV_NORMAL:
2774 case POSIX_FADV_RANDOM:
2775 case POSIX_FADV_SEQUENTIAL:
2782 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
2784 req_set_fail_links(req);
2785 io_cqring_add_event(req, ret);
2786 io_put_req_find_next(req, nxt);
2790 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2792 const char __user *fname;
2793 unsigned lookup_flags;
2796 if (sqe->ioprio || sqe->buf_index)
2798 if (sqe->flags & IOSQE_FIXED_FILE)
2800 if (req->flags & REQ_F_NEED_CLEANUP)
2803 req->open.dfd = READ_ONCE(sqe->fd);
2804 req->open.mask = READ_ONCE(sqe->len);
2805 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2806 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2807 req->open.how.flags = READ_ONCE(sqe->statx_flags);
2809 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
2812 req->open.filename = getname_flags(fname, lookup_flags, NULL);
2813 if (IS_ERR(req->open.filename)) {
2814 ret = PTR_ERR(req->open.filename);
2815 req->open.filename = NULL;
2819 req->flags |= REQ_F_NEED_CLEANUP;
2823 static int io_statx(struct io_kiocb *req, struct io_kiocb **nxt,
2824 bool force_nonblock)
2826 struct io_open *ctx = &req->open;
2827 unsigned lookup_flags;
2835 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
2839 /* filename_lookup() drops it, keep a reference */
2840 ctx->filename->refcnt++;
2842 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
2847 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
2849 if (retry_estale(ret, lookup_flags)) {
2850 lookup_flags |= LOOKUP_REVAL;
2854 ret = cp_statx(&stat, ctx->buffer);
2856 putname(ctx->filename);
2857 req->flags &= ~REQ_F_NEED_CLEANUP;
2859 req_set_fail_links(req);
2860 io_cqring_add_event(req, ret);
2861 io_put_req_find_next(req, nxt);
2865 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2868 * If we queue this for async, it must not be cancellable. That would
2869 * leave the 'file' in an undeterminate state.
2871 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
2873 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
2874 sqe->rw_flags || sqe->buf_index)
2876 if (sqe->flags & IOSQE_FIXED_FILE)
2879 req->close.fd = READ_ONCE(sqe->fd);
2880 if (req->file->f_op == &io_uring_fops ||
2881 req->close.fd == req->ctx->ring_fd)
2887 /* only called when __close_fd_get_file() is done */
2888 static void __io_close_finish(struct io_kiocb *req, struct io_kiocb **nxt)
2892 ret = filp_close(req->close.put_file, req->work.files);
2894 req_set_fail_links(req);
2895 io_cqring_add_event(req, ret);
2896 fput(req->close.put_file);
2897 io_put_req_find_next(req, nxt);
2900 static void io_close_finish(struct io_wq_work **workptr)
2902 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2903 struct io_kiocb *nxt = NULL;
2905 /* not cancellable, don't do io_req_cancelled() */
2906 __io_close_finish(req, &nxt);
2908 io_wq_assign_next(workptr, nxt);
2911 static int io_close(struct io_kiocb *req, struct io_kiocb **nxt,
2912 bool force_nonblock)
2916 req->close.put_file = NULL;
2917 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
2921 /* if the file has a flush method, be safe and punt to async */
2922 if (req->close.put_file->f_op->flush && !io_wq_current_is_worker())
2926 * No ->flush(), safely close from here and just punt the
2927 * fput() to async context.
2929 __io_close_finish(req, nxt);
2932 req->work.func = io_close_finish;
2934 * Do manual async queue here to avoid grabbing files - we don't
2935 * need the files, and it'll cause io_close_finish() to close
2936 * the file again and cause a double CQE entry for this request
2938 io_queue_async_work(req);
2942 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2944 struct io_ring_ctx *ctx = req->ctx;
2949 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2951 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2954 req->sync.off = READ_ONCE(sqe->off);
2955 req->sync.len = READ_ONCE(sqe->len);
2956 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
2960 static void io_sync_file_range_finish(struct io_wq_work **workptr)
2962 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2963 struct io_kiocb *nxt = NULL;
2966 if (io_req_cancelled(req))
2969 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
2972 req_set_fail_links(req);
2973 io_cqring_add_event(req, ret);
2974 io_put_req_find_next(req, &nxt);
2976 io_wq_assign_next(workptr, nxt);
2979 static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt,
2980 bool force_nonblock)
2982 struct io_wq_work *work, *old_work;
2984 /* sync_file_range always requires a blocking context */
2985 if (force_nonblock) {
2987 req->work.func = io_sync_file_range_finish;
2991 work = old_work = &req->work;
2992 io_sync_file_range_finish(&work);
2993 if (work && work != old_work)
2994 *nxt = container_of(work, struct io_kiocb, work);
2998 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3000 #if defined(CONFIG_NET)
3001 struct io_sr_msg *sr = &req->sr_msg;
3002 struct io_async_ctx *io = req->io;
3005 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3006 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3007 sr->len = READ_ONCE(sqe->len);
3009 #ifdef CONFIG_COMPAT
3010 if (req->ctx->compat)
3011 sr->msg_flags |= MSG_CMSG_COMPAT;
3014 if (!io || req->opcode == IORING_OP_SEND)
3016 /* iovec is already imported */
3017 if (req->flags & REQ_F_NEED_CLEANUP)
3020 io->msg.iov = io->msg.fast_iov;
3021 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3024 req->flags |= REQ_F_NEED_CLEANUP;
3031 static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt,
3032 bool force_nonblock)
3034 #if defined(CONFIG_NET)
3035 struct io_async_msghdr *kmsg = NULL;
3036 struct socket *sock;
3039 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3042 sock = sock_from_file(req->file, &ret);
3044 struct io_async_ctx io;
3048 kmsg = &req->io->msg;
3049 kmsg->msg.msg_name = &req->io->msg.addr;
3050 /* if iov is set, it's allocated already */
3052 kmsg->iov = kmsg->fast_iov;
3053 kmsg->msg.msg_iter.iov = kmsg->iov;
3055 struct io_sr_msg *sr = &req->sr_msg;
3058 kmsg->msg.msg_name = &io.msg.addr;
3060 io.msg.iov = io.msg.fast_iov;
3061 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3062 sr->msg_flags, &io.msg.iov);
3067 flags = req->sr_msg.msg_flags;
3068 if (flags & MSG_DONTWAIT)
3069 req->flags |= REQ_F_NOWAIT;
3070 else if (force_nonblock)
3071 flags |= MSG_DONTWAIT;
3073 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3074 if (force_nonblock && ret == -EAGAIN) {
3077 if (io_alloc_async_ctx(req)) {
3078 if (kmsg->iov != kmsg->fast_iov)
3082 req->flags |= REQ_F_NEED_CLEANUP;
3083 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3086 if (ret == -ERESTARTSYS)
3090 if (kmsg && kmsg->iov != kmsg->fast_iov)
3092 req->flags &= ~REQ_F_NEED_CLEANUP;
3093 io_cqring_add_event(req, ret);
3095 req_set_fail_links(req);
3096 io_put_req_find_next(req, nxt);
3103 static int io_send(struct io_kiocb *req, struct io_kiocb **nxt,
3104 bool force_nonblock)
3106 #if defined(CONFIG_NET)
3107 struct socket *sock;
3110 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3113 sock = sock_from_file(req->file, &ret);
3115 struct io_sr_msg *sr = &req->sr_msg;
3120 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3125 msg.msg_name = NULL;
3126 msg.msg_control = NULL;
3127 msg.msg_controllen = 0;
3128 msg.msg_namelen = 0;
3130 flags = req->sr_msg.msg_flags;
3131 if (flags & MSG_DONTWAIT)
3132 req->flags |= REQ_F_NOWAIT;
3133 else if (force_nonblock)
3134 flags |= MSG_DONTWAIT;
3136 msg.msg_flags = flags;
3137 ret = sock_sendmsg(sock, &msg);
3138 if (force_nonblock && ret == -EAGAIN)
3140 if (ret == -ERESTARTSYS)
3144 io_cqring_add_event(req, ret);
3146 req_set_fail_links(req);
3147 io_put_req_find_next(req, nxt);
3154 static int io_recvmsg_prep(struct io_kiocb *req,
3155 const struct io_uring_sqe *sqe)
3157 #if defined(CONFIG_NET)
3158 struct io_sr_msg *sr = &req->sr_msg;
3159 struct io_async_ctx *io = req->io;
3162 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3163 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3164 sr->len = READ_ONCE(sqe->len);
3166 #ifdef CONFIG_COMPAT
3167 if (req->ctx->compat)
3168 sr->msg_flags |= MSG_CMSG_COMPAT;
3171 if (!io || req->opcode == IORING_OP_RECV)
3173 /* iovec is already imported */
3174 if (req->flags & REQ_F_NEED_CLEANUP)
3177 io->msg.iov = io->msg.fast_iov;
3178 ret = recvmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3179 &io->msg.uaddr, &io->msg.iov);
3181 req->flags |= REQ_F_NEED_CLEANUP;
3188 static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt,
3189 bool force_nonblock)
3191 #if defined(CONFIG_NET)
3192 struct io_async_msghdr *kmsg = NULL;
3193 struct socket *sock;
3196 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3199 sock = sock_from_file(req->file, &ret);
3201 struct io_async_ctx io;
3205 kmsg = &req->io->msg;
3206 kmsg->msg.msg_name = &req->io->msg.addr;
3207 /* if iov is set, it's allocated already */
3209 kmsg->iov = kmsg->fast_iov;
3210 kmsg->msg.msg_iter.iov = kmsg->iov;
3212 struct io_sr_msg *sr = &req->sr_msg;
3215 kmsg->msg.msg_name = &io.msg.addr;
3217 io.msg.iov = io.msg.fast_iov;
3218 ret = recvmsg_copy_msghdr(&io.msg.msg, sr->msg,
3219 sr->msg_flags, &io.msg.uaddr,
3225 flags = req->sr_msg.msg_flags;
3226 if (flags & MSG_DONTWAIT)
3227 req->flags |= REQ_F_NOWAIT;
3228 else if (force_nonblock)
3229 flags |= MSG_DONTWAIT;
3231 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3232 kmsg->uaddr, flags);
3233 if (force_nonblock && ret == -EAGAIN) {
3236 if (io_alloc_async_ctx(req)) {
3237 if (kmsg->iov != kmsg->fast_iov)
3241 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3242 req->flags |= REQ_F_NEED_CLEANUP;
3245 if (ret == -ERESTARTSYS)
3249 if (kmsg && kmsg->iov != kmsg->fast_iov)
3251 req->flags &= ~REQ_F_NEED_CLEANUP;
3252 io_cqring_add_event(req, ret);
3254 req_set_fail_links(req);
3255 io_put_req_find_next(req, nxt);
3262 static int io_recv(struct io_kiocb *req, struct io_kiocb **nxt,
3263 bool force_nonblock)
3265 #if defined(CONFIG_NET)
3266 struct socket *sock;
3269 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3272 sock = sock_from_file(req->file, &ret);
3274 struct io_sr_msg *sr = &req->sr_msg;
3279 ret = import_single_range(READ, sr->buf, sr->len, &iov,
3284 msg.msg_name = NULL;
3285 msg.msg_control = NULL;
3286 msg.msg_controllen = 0;
3287 msg.msg_namelen = 0;
3288 msg.msg_iocb = NULL;
3291 flags = req->sr_msg.msg_flags;
3292 if (flags & MSG_DONTWAIT)
3293 req->flags |= REQ_F_NOWAIT;
3294 else if (force_nonblock)
3295 flags |= MSG_DONTWAIT;
3297 ret = sock_recvmsg(sock, &msg, flags);
3298 if (force_nonblock && ret == -EAGAIN)
3300 if (ret == -ERESTARTSYS)
3304 io_cqring_add_event(req, ret);
3306 req_set_fail_links(req);
3307 io_put_req_find_next(req, nxt);
3315 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3317 #if defined(CONFIG_NET)
3318 struct io_accept *accept = &req->accept;
3320 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3322 if (sqe->ioprio || sqe->len || sqe->buf_index)
3325 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3326 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3327 accept->flags = READ_ONCE(sqe->accept_flags);
3328 accept->nofile = rlimit(RLIMIT_NOFILE);
3335 #if defined(CONFIG_NET)
3336 static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3337 bool force_nonblock)
3339 struct io_accept *accept = &req->accept;
3340 unsigned file_flags;
3343 file_flags = force_nonblock ? O_NONBLOCK : 0;
3344 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3345 accept->addr_len, accept->flags,
3347 if (ret == -EAGAIN && force_nonblock)
3349 if (ret == -ERESTARTSYS)
3352 req_set_fail_links(req);
3353 io_cqring_add_event(req, ret);
3354 io_put_req_find_next(req, nxt);
3358 static void io_accept_finish(struct io_wq_work **workptr)
3360 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3361 struct io_kiocb *nxt = NULL;
3363 if (io_req_cancelled(req))
3365 __io_accept(req, &nxt, false);
3367 io_wq_assign_next(workptr, nxt);
3371 static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3372 bool force_nonblock)
3374 #if defined(CONFIG_NET)
3377 ret = __io_accept(req, nxt, force_nonblock);
3378 if (ret == -EAGAIN && force_nonblock) {
3379 req->work.func = io_accept_finish;
3389 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3391 #if defined(CONFIG_NET)
3392 struct io_connect *conn = &req->connect;
3393 struct io_async_ctx *io = req->io;
3395 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3397 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3400 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3401 conn->addr_len = READ_ONCE(sqe->addr2);
3406 return move_addr_to_kernel(conn->addr, conn->addr_len,
3407 &io->connect.address);
3413 static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt,
3414 bool force_nonblock)
3416 #if defined(CONFIG_NET)
3417 struct io_async_ctx __io, *io;
3418 unsigned file_flags;
3424 ret = move_addr_to_kernel(req->connect.addr,
3425 req->connect.addr_len,
3426 &__io.connect.address);
3432 file_flags = force_nonblock ? O_NONBLOCK : 0;
3434 ret = __sys_connect_file(req->file, &io->connect.address,
3435 req->connect.addr_len, file_flags);
3436 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
3439 if (io_alloc_async_ctx(req)) {
3443 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
3446 if (ret == -ERESTARTSYS)
3450 req_set_fail_links(req);
3451 io_cqring_add_event(req, ret);
3452 io_put_req_find_next(req, nxt);
3459 static void io_poll_remove_one(struct io_kiocb *req)
3461 struct io_poll_iocb *poll = &req->poll;
3463 spin_lock(&poll->head->lock);
3464 WRITE_ONCE(poll->canceled, true);
3465 if (!list_empty(&poll->wait.entry)) {
3466 list_del_init(&poll->wait.entry);
3467 io_queue_async_work(req);
3469 spin_unlock(&poll->head->lock);
3470 hash_del(&req->hash_node);
3473 static void io_poll_remove_all(struct io_ring_ctx *ctx)
3475 struct hlist_node *tmp;
3476 struct io_kiocb *req;
3479 spin_lock_irq(&ctx->completion_lock);
3480 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
3481 struct hlist_head *list;
3483 list = &ctx->cancel_hash[i];
3484 hlist_for_each_entry_safe(req, tmp, list, hash_node)
3485 io_poll_remove_one(req);
3487 spin_unlock_irq(&ctx->completion_lock);
3490 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
3492 struct hlist_head *list;
3493 struct io_kiocb *req;
3495 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
3496 hlist_for_each_entry(req, list, hash_node) {
3497 if (sqe_addr == req->user_data) {
3498 io_poll_remove_one(req);
3506 static int io_poll_remove_prep(struct io_kiocb *req,
3507 const struct io_uring_sqe *sqe)
3509 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3511 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3515 req->poll.addr = READ_ONCE(sqe->addr);
3520 * Find a running poll command that matches one specified in sqe->addr,
3521 * and remove it if found.
3523 static int io_poll_remove(struct io_kiocb *req)
3525 struct io_ring_ctx *ctx = req->ctx;
3529 addr = req->poll.addr;
3530 spin_lock_irq(&ctx->completion_lock);
3531 ret = io_poll_cancel(ctx, addr);
3532 spin_unlock_irq(&ctx->completion_lock);
3534 io_cqring_add_event(req, ret);
3536 req_set_fail_links(req);
3541 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
3543 struct io_ring_ctx *ctx = req->ctx;
3545 req->poll.done = true;
3547 io_cqring_fill_event(req, error);
3549 io_cqring_fill_event(req, mangle_poll(mask));
3550 io_commit_cqring(ctx);
3553 static void io_poll_complete_work(struct io_wq_work **workptr)
3555 struct io_wq_work *work = *workptr;
3556 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3557 struct io_poll_iocb *poll = &req->poll;
3558 struct poll_table_struct pt = { ._key = poll->events };
3559 struct io_ring_ctx *ctx = req->ctx;
3560 struct io_kiocb *nxt = NULL;
3564 if (work->flags & IO_WQ_WORK_CANCEL) {
3565 WRITE_ONCE(poll->canceled, true);
3567 } else if (READ_ONCE(poll->canceled)) {
3571 if (ret != -ECANCELED)
3572 mask = vfs_poll(poll->file, &pt) & poll->events;
3575 * Note that ->ki_cancel callers also delete iocb from active_reqs after
3576 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
3577 * synchronize with them. In the cancellation case the list_del_init
3578 * itself is not actually needed, but harmless so we keep it in to
3579 * avoid further branches in the fast path.
3581 spin_lock_irq(&ctx->completion_lock);
3582 if (!mask && ret != -ECANCELED) {
3583 add_wait_queue(poll->head, &poll->wait);
3584 spin_unlock_irq(&ctx->completion_lock);
3587 hash_del(&req->hash_node);
3588 io_poll_complete(req, mask, ret);
3589 spin_unlock_irq(&ctx->completion_lock);
3591 io_cqring_ev_posted(ctx);
3594 req_set_fail_links(req);
3595 io_put_req_find_next(req, &nxt);
3597 io_wq_assign_next(workptr, nxt);
3600 static void __io_poll_flush(struct io_ring_ctx *ctx, struct llist_node *nodes)
3602 struct io_kiocb *req, *tmp;
3603 struct req_batch rb;
3605 rb.to_free = rb.need_iter = 0;
3606 spin_lock_irq(&ctx->completion_lock);
3607 llist_for_each_entry_safe(req, tmp, nodes, llist_node) {
3608 hash_del(&req->hash_node);
3609 io_poll_complete(req, req->result, 0);
3611 if (refcount_dec_and_test(&req->refs) &&
3612 !io_req_multi_free(&rb, req)) {
3613 req->flags |= REQ_F_COMP_LOCKED;
3617 spin_unlock_irq(&ctx->completion_lock);
3619 io_cqring_ev_posted(ctx);
3620 io_free_req_many(ctx, &rb);
3623 static void io_poll_flush(struct io_wq_work **workptr)
3625 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3626 struct llist_node *nodes;
3628 nodes = llist_del_all(&req->ctx->poll_llist);
3630 __io_poll_flush(req->ctx, nodes);
3633 static void io_poll_trigger_evfd(struct io_wq_work **workptr)
3635 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3637 eventfd_signal(req->ctx->cq_ev_fd, 1);
3641 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
3644 struct io_poll_iocb *poll = wait->private;
3645 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
3646 struct io_ring_ctx *ctx = req->ctx;
3647 __poll_t mask = key_to_poll(key);
3649 /* for instances that support it check for an event match first: */
3650 if (mask && !(mask & poll->events))
3653 list_del_init(&poll->wait.entry);
3656 * Run completion inline if we can. We're using trylock here because
3657 * we are violating the completion_lock -> poll wq lock ordering.
3658 * If we have a link timeout we're going to need the completion_lock
3659 * for finalizing the request, mark us as having grabbed that already.
3662 unsigned long flags;
3664 if (llist_empty(&ctx->poll_llist) &&
3665 spin_trylock_irqsave(&ctx->completion_lock, flags)) {
3668 hash_del(&req->hash_node);
3669 io_poll_complete(req, mask, 0);
3671 trigger_ev = io_should_trigger_evfd(ctx);
3672 if (trigger_ev && eventfd_signal_count()) {
3674 req->work.func = io_poll_trigger_evfd;
3676 req->flags |= REQ_F_COMP_LOCKED;
3680 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3681 __io_cqring_ev_posted(ctx, trigger_ev);
3684 req->llist_node.next = NULL;
3685 /* if the list wasn't empty, we're done */
3686 if (!llist_add(&req->llist_node, &ctx->poll_llist))
3689 req->work.func = io_poll_flush;
3693 io_queue_async_work(req);
3698 struct io_poll_table {
3699 struct poll_table_struct pt;
3700 struct io_kiocb *req;
3704 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
3705 struct poll_table_struct *p)
3707 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
3709 if (unlikely(pt->req->poll.head)) {
3710 pt->error = -EINVAL;
3715 pt->req->poll.head = head;
3716 add_wait_queue(head, &pt->req->poll.wait);
3719 static void io_poll_req_insert(struct io_kiocb *req)
3721 struct io_ring_ctx *ctx = req->ctx;
3722 struct hlist_head *list;
3724 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
3725 hlist_add_head(&req->hash_node, list);
3728 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3730 struct io_poll_iocb *poll = &req->poll;
3733 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3735 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3740 events = READ_ONCE(sqe->poll_events);
3741 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
3745 static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt)
3747 struct io_poll_iocb *poll = &req->poll;
3748 struct io_ring_ctx *ctx = req->ctx;
3749 struct io_poll_table ipt;
3750 bool cancel = false;
3753 INIT_IO_WORK(&req->work, io_poll_complete_work);
3754 INIT_HLIST_NODE(&req->hash_node);
3758 poll->canceled = false;
3760 ipt.pt._qproc = io_poll_queue_proc;
3761 ipt.pt._key = poll->events;
3763 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
3765 /* initialized the list so that we can do list_empty checks */
3766 INIT_LIST_HEAD(&poll->wait.entry);
3767 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
3768 poll->wait.private = poll;
3770 INIT_LIST_HEAD(&req->list);
3772 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
3774 spin_lock_irq(&ctx->completion_lock);
3775 if (likely(poll->head)) {
3776 spin_lock(&poll->head->lock);
3777 if (unlikely(list_empty(&poll->wait.entry))) {
3783 if (mask || ipt.error)
3784 list_del_init(&poll->wait.entry);
3786 WRITE_ONCE(poll->canceled, true);
3787 else if (!poll->done) /* actually waiting for an event */
3788 io_poll_req_insert(req);
3789 spin_unlock(&poll->head->lock);
3791 if (mask) { /* no async, we'd stolen it */
3793 io_poll_complete(req, mask, 0);
3795 spin_unlock_irq(&ctx->completion_lock);
3798 io_cqring_ev_posted(ctx);
3799 io_put_req_find_next(req, nxt);
3804 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
3806 struct io_timeout_data *data = container_of(timer,
3807 struct io_timeout_data, timer);
3808 struct io_kiocb *req = data->req;
3809 struct io_ring_ctx *ctx = req->ctx;
3810 unsigned long flags;
3812 atomic_inc(&ctx->cq_timeouts);
3814 spin_lock_irqsave(&ctx->completion_lock, flags);
3816 * We could be racing with timeout deletion. If the list is empty,
3817 * then timeout lookup already found it and will be handling it.
3819 if (!list_empty(&req->list)) {
3820 struct io_kiocb *prev;
3823 * Adjust the reqs sequence before the current one because it
3824 * will consume a slot in the cq_ring and the cq_tail
3825 * pointer will be increased, otherwise other timeout reqs may
3826 * return in advance without waiting for enough wait_nr.
3829 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
3831 list_del_init(&req->list);
3834 io_cqring_fill_event(req, -ETIME);
3835 io_commit_cqring(ctx);
3836 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3838 io_cqring_ev_posted(ctx);
3839 req_set_fail_links(req);
3841 return HRTIMER_NORESTART;
3844 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
3846 struct io_kiocb *req;
3849 list_for_each_entry(req, &ctx->timeout_list, list) {
3850 if (user_data == req->user_data) {
3851 list_del_init(&req->list);
3860 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
3864 req_set_fail_links(req);
3865 io_cqring_fill_event(req, -ECANCELED);
3870 static int io_timeout_remove_prep(struct io_kiocb *req,
3871 const struct io_uring_sqe *sqe)
3873 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3875 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
3878 req->timeout.addr = READ_ONCE(sqe->addr);
3879 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
3880 if (req->timeout.flags)
3887 * Remove or update an existing timeout command
3889 static int io_timeout_remove(struct io_kiocb *req)
3891 struct io_ring_ctx *ctx = req->ctx;
3894 spin_lock_irq(&ctx->completion_lock);
3895 ret = io_timeout_cancel(ctx, req->timeout.addr);
3897 io_cqring_fill_event(req, ret);
3898 io_commit_cqring(ctx);
3899 spin_unlock_irq(&ctx->completion_lock);
3900 io_cqring_ev_posted(ctx);
3902 req_set_fail_links(req);
3907 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3908 bool is_timeout_link)
3910 struct io_timeout_data *data;
3913 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3915 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
3917 if (sqe->off && is_timeout_link)
3919 flags = READ_ONCE(sqe->timeout_flags);
3920 if (flags & ~IORING_TIMEOUT_ABS)
3923 req->timeout.count = READ_ONCE(sqe->off);
3925 if (!req->io && io_alloc_async_ctx(req))
3928 data = &req->io->timeout;
3930 req->flags |= REQ_F_TIMEOUT;
3932 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
3935 if (flags & IORING_TIMEOUT_ABS)
3936 data->mode = HRTIMER_MODE_ABS;
3938 data->mode = HRTIMER_MODE_REL;
3940 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
3944 static int io_timeout(struct io_kiocb *req)
3947 struct io_ring_ctx *ctx = req->ctx;
3948 struct io_timeout_data *data;
3949 struct list_head *entry;
3952 data = &req->io->timeout;
3955 * sqe->off holds how many events that need to occur for this
3956 * timeout event to be satisfied. If it isn't set, then this is
3957 * a pure timeout request, sequence isn't used.
3959 count = req->timeout.count;
3961 req->flags |= REQ_F_TIMEOUT_NOSEQ;
3962 spin_lock_irq(&ctx->completion_lock);
3963 entry = ctx->timeout_list.prev;
3967 req->sequence = ctx->cached_sq_head + count - 1;
3968 data->seq_offset = count;
3971 * Insertion sort, ensuring the first entry in the list is always
3972 * the one we need first.
3974 spin_lock_irq(&ctx->completion_lock);
3975 list_for_each_prev(entry, &ctx->timeout_list) {
3976 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
3977 unsigned nxt_sq_head;
3978 long long tmp, tmp_nxt;
3979 u32 nxt_offset = nxt->io->timeout.seq_offset;
3981 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
3985 * Since cached_sq_head + count - 1 can overflow, use type long
3988 tmp = (long long)ctx->cached_sq_head + count - 1;
3989 nxt_sq_head = nxt->sequence - nxt_offset + 1;
3990 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
3993 * cached_sq_head may overflow, and it will never overflow twice
3994 * once there is some timeout req still be valid.
3996 if (ctx->cached_sq_head < nxt_sq_head)
4003 * Sequence of reqs after the insert one and itself should
4004 * be adjusted because each timeout req consumes a slot.
4009 req->sequence -= span;
4011 list_add(&req->list, entry);
4012 data->timer.function = io_timeout_fn;
4013 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4014 spin_unlock_irq(&ctx->completion_lock);
4018 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4020 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4022 return req->user_data == (unsigned long) data;
4025 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4027 enum io_wq_cancel cancel_ret;
4030 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4031 switch (cancel_ret) {
4032 case IO_WQ_CANCEL_OK:
4035 case IO_WQ_CANCEL_RUNNING:
4038 case IO_WQ_CANCEL_NOTFOUND:
4046 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4047 struct io_kiocb *req, __u64 sqe_addr,
4048 struct io_kiocb **nxt, int success_ret)
4050 unsigned long flags;
4053 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4054 if (ret != -ENOENT) {
4055 spin_lock_irqsave(&ctx->completion_lock, flags);
4059 spin_lock_irqsave(&ctx->completion_lock, flags);
4060 ret = io_timeout_cancel(ctx, sqe_addr);
4063 ret = io_poll_cancel(ctx, sqe_addr);
4067 io_cqring_fill_event(req, ret);
4068 io_commit_cqring(ctx);
4069 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4070 io_cqring_ev_posted(ctx);
4073 req_set_fail_links(req);
4074 io_put_req_find_next(req, nxt);
4077 static int io_async_cancel_prep(struct io_kiocb *req,
4078 const struct io_uring_sqe *sqe)
4080 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4082 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4086 req->cancel.addr = READ_ONCE(sqe->addr);
4090 static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt)
4092 struct io_ring_ctx *ctx = req->ctx;
4094 io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0);
4098 static int io_files_update_prep(struct io_kiocb *req,
4099 const struct io_uring_sqe *sqe)
4101 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4104 req->files_update.offset = READ_ONCE(sqe->off);
4105 req->files_update.nr_args = READ_ONCE(sqe->len);
4106 if (!req->files_update.nr_args)
4108 req->files_update.arg = READ_ONCE(sqe->addr);
4112 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4114 struct io_ring_ctx *ctx = req->ctx;
4115 struct io_uring_files_update up;
4121 up.offset = req->files_update.offset;
4122 up.fds = req->files_update.arg;
4124 mutex_lock(&ctx->uring_lock);
4125 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4126 mutex_unlock(&ctx->uring_lock);
4129 req_set_fail_links(req);
4130 io_cqring_add_event(req, ret);
4135 static int io_req_defer_prep(struct io_kiocb *req,
4136 const struct io_uring_sqe *sqe)
4143 if (io_op_defs[req->opcode].file_table) {
4144 ret = io_grab_files(req);
4149 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4151 switch (req->opcode) {
4154 case IORING_OP_READV:
4155 case IORING_OP_READ_FIXED:
4156 case IORING_OP_READ:
4157 ret = io_read_prep(req, sqe, true);
4159 case IORING_OP_WRITEV:
4160 case IORING_OP_WRITE_FIXED:
4161 case IORING_OP_WRITE:
4162 ret = io_write_prep(req, sqe, true);
4164 case IORING_OP_POLL_ADD:
4165 ret = io_poll_add_prep(req, sqe);
4167 case IORING_OP_POLL_REMOVE:
4168 ret = io_poll_remove_prep(req, sqe);
4170 case IORING_OP_FSYNC:
4171 ret = io_prep_fsync(req, sqe);
4173 case IORING_OP_SYNC_FILE_RANGE:
4174 ret = io_prep_sfr(req, sqe);
4176 case IORING_OP_SENDMSG:
4177 case IORING_OP_SEND:
4178 ret = io_sendmsg_prep(req, sqe);
4180 case IORING_OP_RECVMSG:
4181 case IORING_OP_RECV:
4182 ret = io_recvmsg_prep(req, sqe);
4184 case IORING_OP_CONNECT:
4185 ret = io_connect_prep(req, sqe);
4187 case IORING_OP_TIMEOUT:
4188 ret = io_timeout_prep(req, sqe, false);
4190 case IORING_OP_TIMEOUT_REMOVE:
4191 ret = io_timeout_remove_prep(req, sqe);
4193 case IORING_OP_ASYNC_CANCEL:
4194 ret = io_async_cancel_prep(req, sqe);
4196 case IORING_OP_LINK_TIMEOUT:
4197 ret = io_timeout_prep(req, sqe, true);
4199 case IORING_OP_ACCEPT:
4200 ret = io_accept_prep(req, sqe);
4202 case IORING_OP_FALLOCATE:
4203 ret = io_fallocate_prep(req, sqe);
4205 case IORING_OP_OPENAT:
4206 ret = io_openat_prep(req, sqe);
4208 case IORING_OP_CLOSE:
4209 ret = io_close_prep(req, sqe);
4211 case IORING_OP_FILES_UPDATE:
4212 ret = io_files_update_prep(req, sqe);
4214 case IORING_OP_STATX:
4215 ret = io_statx_prep(req, sqe);
4217 case IORING_OP_FADVISE:
4218 ret = io_fadvise_prep(req, sqe);
4220 case IORING_OP_MADVISE:
4221 ret = io_madvise_prep(req, sqe);
4223 case IORING_OP_OPENAT2:
4224 ret = io_openat2_prep(req, sqe);
4226 case IORING_OP_EPOLL_CTL:
4227 ret = io_epoll_ctl_prep(req, sqe);
4230 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
4239 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4241 struct io_ring_ctx *ctx = req->ctx;
4244 /* Still need defer if there is pending req in defer list. */
4245 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
4248 if (!req->io && io_alloc_async_ctx(req))
4251 ret = io_req_defer_prep(req, sqe);
4255 spin_lock_irq(&ctx->completion_lock);
4256 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
4257 spin_unlock_irq(&ctx->completion_lock);
4261 trace_io_uring_defer(ctx, req, req->user_data);
4262 list_add_tail(&req->list, &ctx->defer_list);
4263 spin_unlock_irq(&ctx->completion_lock);
4264 return -EIOCBQUEUED;
4267 static void io_cleanup_req(struct io_kiocb *req)
4269 struct io_async_ctx *io = req->io;
4271 switch (req->opcode) {
4272 case IORING_OP_READV:
4273 case IORING_OP_READ_FIXED:
4274 case IORING_OP_READ:
4275 case IORING_OP_WRITEV:
4276 case IORING_OP_WRITE_FIXED:
4277 case IORING_OP_WRITE:
4278 if (io->rw.iov != io->rw.fast_iov)
4281 case IORING_OP_SENDMSG:
4282 case IORING_OP_RECVMSG:
4283 if (io->msg.iov != io->msg.fast_iov)
4286 case IORING_OP_OPENAT:
4287 case IORING_OP_OPENAT2:
4288 case IORING_OP_STATX:
4289 putname(req->open.filename);
4293 req->flags &= ~REQ_F_NEED_CLEANUP;
4296 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4297 struct io_kiocb **nxt, bool force_nonblock)
4299 struct io_ring_ctx *ctx = req->ctx;
4302 switch (req->opcode) {
4306 case IORING_OP_READV:
4307 case IORING_OP_READ_FIXED:
4308 case IORING_OP_READ:
4310 ret = io_read_prep(req, sqe, force_nonblock);
4314 ret = io_read(req, nxt, force_nonblock);
4316 case IORING_OP_WRITEV:
4317 case IORING_OP_WRITE_FIXED:
4318 case IORING_OP_WRITE:
4320 ret = io_write_prep(req, sqe, force_nonblock);
4324 ret = io_write(req, nxt, force_nonblock);
4326 case IORING_OP_FSYNC:
4328 ret = io_prep_fsync(req, sqe);
4332 ret = io_fsync(req, nxt, force_nonblock);
4334 case IORING_OP_POLL_ADD:
4336 ret = io_poll_add_prep(req, sqe);
4340 ret = io_poll_add(req, nxt);
4342 case IORING_OP_POLL_REMOVE:
4344 ret = io_poll_remove_prep(req, sqe);
4348 ret = io_poll_remove(req);
4350 case IORING_OP_SYNC_FILE_RANGE:
4352 ret = io_prep_sfr(req, sqe);
4356 ret = io_sync_file_range(req, nxt, force_nonblock);
4358 case IORING_OP_SENDMSG:
4359 case IORING_OP_SEND:
4361 ret = io_sendmsg_prep(req, sqe);
4365 if (req->opcode == IORING_OP_SENDMSG)
4366 ret = io_sendmsg(req, nxt, force_nonblock);
4368 ret = io_send(req, nxt, force_nonblock);
4370 case IORING_OP_RECVMSG:
4371 case IORING_OP_RECV:
4373 ret = io_recvmsg_prep(req, sqe);
4377 if (req->opcode == IORING_OP_RECVMSG)
4378 ret = io_recvmsg(req, nxt, force_nonblock);
4380 ret = io_recv(req, nxt, force_nonblock);
4382 case IORING_OP_TIMEOUT:
4384 ret = io_timeout_prep(req, sqe, false);
4388 ret = io_timeout(req);
4390 case IORING_OP_TIMEOUT_REMOVE:
4392 ret = io_timeout_remove_prep(req, sqe);
4396 ret = io_timeout_remove(req);
4398 case IORING_OP_ACCEPT:
4400 ret = io_accept_prep(req, sqe);
4404 ret = io_accept(req, nxt, force_nonblock);
4406 case IORING_OP_CONNECT:
4408 ret = io_connect_prep(req, sqe);
4412 ret = io_connect(req, nxt, force_nonblock);
4414 case IORING_OP_ASYNC_CANCEL:
4416 ret = io_async_cancel_prep(req, sqe);
4420 ret = io_async_cancel(req, nxt);
4422 case IORING_OP_FALLOCATE:
4424 ret = io_fallocate_prep(req, sqe);
4428 ret = io_fallocate(req, nxt, force_nonblock);
4430 case IORING_OP_OPENAT:
4432 ret = io_openat_prep(req, sqe);
4436 ret = io_openat(req, nxt, force_nonblock);
4438 case IORING_OP_CLOSE:
4440 ret = io_close_prep(req, sqe);
4444 ret = io_close(req, nxt, force_nonblock);
4446 case IORING_OP_FILES_UPDATE:
4448 ret = io_files_update_prep(req, sqe);
4452 ret = io_files_update(req, force_nonblock);
4454 case IORING_OP_STATX:
4456 ret = io_statx_prep(req, sqe);
4460 ret = io_statx(req, nxt, force_nonblock);
4462 case IORING_OP_FADVISE:
4464 ret = io_fadvise_prep(req, sqe);
4468 ret = io_fadvise(req, nxt, force_nonblock);
4470 case IORING_OP_MADVISE:
4472 ret = io_madvise_prep(req, sqe);
4476 ret = io_madvise(req, nxt, force_nonblock);
4478 case IORING_OP_OPENAT2:
4480 ret = io_openat2_prep(req, sqe);
4484 ret = io_openat2(req, nxt, force_nonblock);
4486 case IORING_OP_EPOLL_CTL:
4488 ret = io_epoll_ctl_prep(req, sqe);
4492 ret = io_epoll_ctl(req, nxt, force_nonblock);
4502 if (ctx->flags & IORING_SETUP_IOPOLL) {
4503 const bool in_async = io_wq_current_is_worker();
4505 if (req->result == -EAGAIN)
4508 /* workqueue context doesn't hold uring_lock, grab it now */
4510 mutex_lock(&ctx->uring_lock);
4512 io_iopoll_req_issued(req);
4515 mutex_unlock(&ctx->uring_lock);
4521 static void io_wq_submit_work(struct io_wq_work **workptr)
4523 struct io_wq_work *work = *workptr;
4524 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4525 struct io_kiocb *nxt = NULL;
4528 /* if NO_CANCEL is set, we must still run the work */
4529 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
4530 IO_WQ_WORK_CANCEL) {
4535 req->in_async = true;
4537 ret = io_issue_sqe(req, NULL, &nxt, false);
4539 * We can get EAGAIN for polled IO even though we're
4540 * forcing a sync submission from here, since we can't
4541 * wait for request slots on the block side.
4549 /* drop submission reference */
4553 req_set_fail_links(req);
4554 io_cqring_add_event(req, ret);
4558 /* if a dependent link is ready, pass it back */
4560 io_wq_assign_next(workptr, nxt);
4563 static int io_req_needs_file(struct io_kiocb *req, int fd)
4565 if (!io_op_defs[req->opcode].needs_file)
4567 if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg)
4572 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
4575 struct fixed_file_table *table;
4577 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
4578 return table->files[index & IORING_FILE_TABLE_MASK];;
4581 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
4582 const struct io_uring_sqe *sqe)
4584 struct io_ring_ctx *ctx = req->ctx;
4588 flags = READ_ONCE(sqe->flags);
4589 fd = READ_ONCE(sqe->fd);
4591 if (!io_req_needs_file(req, fd))
4594 if (flags & IOSQE_FIXED_FILE) {
4595 if (unlikely(!ctx->file_data ||
4596 (unsigned) fd >= ctx->nr_user_files))
4598 fd = array_index_nospec(fd, ctx->nr_user_files);
4599 req->file = io_file_from_index(ctx, fd);
4602 req->flags |= REQ_F_FIXED_FILE;
4603 percpu_ref_get(&ctx->file_data->refs);
4605 if (req->needs_fixed_file)
4607 trace_io_uring_file_get(ctx, fd);
4608 req->file = io_file_get(state, fd);
4609 if (unlikely(!req->file))
4616 static int io_grab_files(struct io_kiocb *req)
4619 struct io_ring_ctx *ctx = req->ctx;
4621 if (req->work.files)
4623 if (!ctx->ring_file)
4627 spin_lock_irq(&ctx->inflight_lock);
4629 * We use the f_ops->flush() handler to ensure that we can flush
4630 * out work accessing these files if the fd is closed. Check if
4631 * the fd has changed since we started down this path, and disallow
4632 * this operation if it has.
4634 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
4635 list_add(&req->inflight_entry, &ctx->inflight_list);
4636 req->flags |= REQ_F_INFLIGHT;
4637 req->work.files = current->files;
4640 spin_unlock_irq(&ctx->inflight_lock);
4646 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
4648 struct io_timeout_data *data = container_of(timer,
4649 struct io_timeout_data, timer);
4650 struct io_kiocb *req = data->req;
4651 struct io_ring_ctx *ctx = req->ctx;
4652 struct io_kiocb *prev = NULL;
4653 unsigned long flags;
4655 spin_lock_irqsave(&ctx->completion_lock, flags);
4658 * We don't expect the list to be empty, that will only happen if we
4659 * race with the completion of the linked work.
4661 if (!list_empty(&req->link_list)) {
4662 prev = list_entry(req->link_list.prev, struct io_kiocb,
4664 if (refcount_inc_not_zero(&prev->refs)) {
4665 list_del_init(&req->link_list);
4666 prev->flags &= ~REQ_F_LINK_TIMEOUT;
4671 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4674 req_set_fail_links(prev);
4675 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
4679 io_cqring_add_event(req, -ETIME);
4682 return HRTIMER_NORESTART;
4685 static void io_queue_linked_timeout(struct io_kiocb *req)
4687 struct io_ring_ctx *ctx = req->ctx;
4690 * If the list is now empty, then our linked request finished before
4691 * we got a chance to setup the timer
4693 spin_lock_irq(&ctx->completion_lock);
4694 if (!list_empty(&req->link_list)) {
4695 struct io_timeout_data *data = &req->io->timeout;
4697 data->timer.function = io_link_timeout_fn;
4698 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
4701 spin_unlock_irq(&ctx->completion_lock);
4703 /* drop submission reference */
4707 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
4709 struct io_kiocb *nxt;
4711 if (!(req->flags & REQ_F_LINK))
4714 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
4716 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
4719 req->flags |= REQ_F_LINK_TIMEOUT;
4723 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4725 struct io_kiocb *linked_timeout;
4726 struct io_kiocb *nxt = NULL;
4727 const struct cred *old_creds = NULL;
4731 linked_timeout = io_prep_linked_timeout(req);
4733 if (req->work.creds && req->work.creds != current_cred()) {
4735 revert_creds(old_creds);
4736 if (old_creds == req->work.creds)
4737 old_creds = NULL; /* restored original creds */
4739 old_creds = override_creds(req->work.creds);
4742 ret = io_issue_sqe(req, sqe, &nxt, true);
4745 * We async punt it if the file wasn't marked NOWAIT, or if the file
4746 * doesn't support non-blocking read/write attempts
4748 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
4749 (req->flags & REQ_F_MUST_PUNT))) {
4751 if (io_op_defs[req->opcode].file_table) {
4752 ret = io_grab_files(req);
4758 * Queued up for async execution, worker will release
4759 * submit reference when the iocb is actually submitted.
4761 io_queue_async_work(req);
4766 /* drop submission reference */
4767 io_put_req_find_next(req, &nxt);
4769 if (linked_timeout) {
4771 io_queue_linked_timeout(linked_timeout);
4773 io_put_req(linked_timeout);
4776 /* and drop final reference, if we failed */
4778 io_cqring_add_event(req, ret);
4779 req_set_fail_links(req);
4787 if (req->flags & REQ_F_FORCE_ASYNC)
4792 revert_creds(old_creds);
4795 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4799 ret = io_req_defer(req, sqe);
4801 if (ret != -EIOCBQUEUED) {
4803 io_cqring_add_event(req, ret);
4804 req_set_fail_links(req);
4805 io_double_put_req(req);
4807 } else if (req->flags & REQ_F_FORCE_ASYNC) {
4808 ret = io_req_defer_prep(req, sqe);
4809 if (unlikely(ret < 0))
4812 * Never try inline submit of IOSQE_ASYNC is set, go straight
4813 * to async execution.
4815 req->work.flags |= IO_WQ_WORK_CONCURRENT;
4816 io_queue_async_work(req);
4818 __io_queue_sqe(req, sqe);
4822 static inline void io_queue_link_head(struct io_kiocb *req)
4824 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
4825 io_cqring_add_event(req, -ECANCELED);
4826 io_double_put_req(req);
4828 io_queue_sqe(req, NULL);
4831 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
4832 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
4834 static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4835 struct io_submit_state *state, struct io_kiocb **link)
4837 struct io_ring_ctx *ctx = req->ctx;
4838 unsigned int sqe_flags;
4841 sqe_flags = READ_ONCE(sqe->flags);
4843 /* enforce forwards compatibility on users */
4844 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
4849 id = READ_ONCE(sqe->personality);
4851 req->work.creds = idr_find(&ctx->personality_idr, id);
4852 if (unlikely(!req->work.creds)) {
4856 get_cred(req->work.creds);
4859 /* same numerical values with corresponding REQ_F_*, safe to copy */
4860 req->flags |= sqe_flags & (IOSQE_IO_DRAIN|IOSQE_IO_HARDLINK|
4863 ret = io_req_set_file(state, req, sqe);
4864 if (unlikely(ret)) {
4866 io_cqring_add_event(req, ret);
4867 io_double_put_req(req);
4872 * If we already have a head request, queue this one for async
4873 * submittal once the head completes. If we don't have a head but
4874 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
4875 * submitted sync once the chain is complete. If none of those
4876 * conditions are true (normal request), then just queue it.
4879 struct io_kiocb *head = *link;
4882 * Taking sequential execution of a link, draining both sides
4883 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
4884 * requests in the link. So, it drains the head and the
4885 * next after the link request. The last one is done via
4886 * drain_next flag to persist the effect across calls.
4888 if (sqe_flags & IOSQE_IO_DRAIN) {
4889 head->flags |= REQ_F_IO_DRAIN;
4890 ctx->drain_next = 1;
4892 if (io_alloc_async_ctx(req)) {
4897 ret = io_req_defer_prep(req, sqe);
4899 /* fail even hard links since we don't submit */
4900 head->flags |= REQ_F_FAIL_LINK;
4903 trace_io_uring_link(ctx, req, head);
4904 list_add_tail(&req->link_list, &head->link_list);
4906 /* last request of a link, enqueue the link */
4907 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) {
4908 io_queue_link_head(head);
4912 if (unlikely(ctx->drain_next)) {
4913 req->flags |= REQ_F_IO_DRAIN;
4914 req->ctx->drain_next = 0;
4916 if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
4917 req->flags |= REQ_F_LINK;
4918 INIT_LIST_HEAD(&req->link_list);
4920 if (io_alloc_async_ctx(req)) {
4924 ret = io_req_defer_prep(req, sqe);
4926 req->flags |= REQ_F_FAIL_LINK;
4929 io_queue_sqe(req, sqe);
4937 * Batched submission is done, ensure local IO is flushed out.
4939 static void io_submit_state_end(struct io_submit_state *state)
4941 blk_finish_plug(&state->plug);
4943 if (state->free_reqs)
4944 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
4948 * Start submission side cache.
4950 static void io_submit_state_start(struct io_submit_state *state,
4951 unsigned int max_ios)
4953 blk_start_plug(&state->plug);
4954 state->free_reqs = 0;
4956 state->ios_left = max_ios;
4959 static void io_commit_sqring(struct io_ring_ctx *ctx)
4961 struct io_rings *rings = ctx->rings;
4964 * Ensure any loads from the SQEs are done at this point,
4965 * since once we write the new head, the application could
4966 * write new data to them.
4968 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
4972 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
4973 * that is mapped by userspace. This means that care needs to be taken to
4974 * ensure that reads are stable, as we cannot rely on userspace always
4975 * being a good citizen. If members of the sqe are validated and then later
4976 * used, it's important that those reads are done through READ_ONCE() to
4977 * prevent a re-load down the line.
4979 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
4980 const struct io_uring_sqe **sqe_ptr)
4982 u32 *sq_array = ctx->sq_array;
4986 * The cached sq head (or cq tail) serves two purposes:
4988 * 1) allows us to batch the cost of updating the user visible
4990 * 2) allows the kernel side to track the head on its own, even
4991 * though the application is the one updating it.
4993 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
4994 if (likely(head < ctx->sq_entries)) {
4996 * All io need record the previous position, if LINK vs DARIN,
4997 * it can be used to mark the position of the first IO in the
5000 req->sequence = ctx->cached_sq_head;
5001 *sqe_ptr = &ctx->sq_sqes[head];
5002 req->opcode = READ_ONCE((*sqe_ptr)->opcode);
5003 req->user_data = READ_ONCE((*sqe_ptr)->user_data);
5004 ctx->cached_sq_head++;
5008 /* drop invalid entries */
5009 ctx->cached_sq_head++;
5010 ctx->cached_sq_dropped++;
5011 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5015 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5016 struct file *ring_file, int ring_fd,
5017 struct mm_struct **mm, bool async)
5019 struct io_submit_state state, *statep = NULL;
5020 struct io_kiocb *link = NULL;
5021 int i, submitted = 0;
5022 bool mm_fault = false;
5024 /* if we have a backlog and couldn't flush it all, return BUSY */
5025 if (test_bit(0, &ctx->sq_check_overflow)) {
5026 if (!list_empty(&ctx->cq_overflow_list) &&
5027 !io_cqring_overflow_flush(ctx, false))
5031 /* make sure SQ entry isn't read before tail */
5032 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5034 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5037 if (nr > IO_PLUG_THRESHOLD) {
5038 io_submit_state_start(&state, nr);
5042 ctx->ring_fd = ring_fd;
5043 ctx->ring_file = ring_file;
5045 for (i = 0; i < nr; i++) {
5046 const struct io_uring_sqe *sqe;
5047 struct io_kiocb *req;
5050 req = io_get_req(ctx, statep);
5051 if (unlikely(!req)) {
5053 submitted = -EAGAIN;
5056 if (!io_get_sqring(ctx, req, &sqe)) {
5057 __io_req_do_free(req);
5061 /* will complete beyond this point, count as submitted */
5064 if (unlikely(req->opcode >= IORING_OP_LAST)) {
5067 io_cqring_add_event(req, err);
5068 io_double_put_req(req);
5072 if (io_op_defs[req->opcode].needs_mm && !*mm) {
5073 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
5074 if (unlikely(mm_fault)) {
5078 use_mm(ctx->sqo_mm);
5082 req->in_async = async;
5083 req->needs_fixed_file = async;
5084 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5086 if (!io_submit_sqe(req, sqe, statep, &link))
5090 if (unlikely(submitted != nr)) {
5091 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5093 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5096 io_queue_link_head(link);
5098 io_submit_state_end(&state);
5100 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5101 io_commit_sqring(ctx);
5106 static int io_sq_thread(void *data)
5108 struct io_ring_ctx *ctx = data;
5109 struct mm_struct *cur_mm = NULL;
5110 const struct cred *old_cred;
5111 mm_segment_t old_fs;
5113 unsigned long timeout;
5116 complete(&ctx->completions[1]);
5120 old_cred = override_creds(ctx->creds);
5122 timeout = jiffies + ctx->sq_thread_idle;
5123 while (!kthread_should_park()) {
5124 unsigned int to_submit;
5126 if (!list_empty(&ctx->poll_list)) {
5127 unsigned nr_events = 0;
5129 mutex_lock(&ctx->uring_lock);
5130 if (!list_empty(&ctx->poll_list))
5131 io_iopoll_getevents(ctx, &nr_events, 0);
5133 timeout = jiffies + ctx->sq_thread_idle;
5134 mutex_unlock(&ctx->uring_lock);
5137 to_submit = io_sqring_entries(ctx);
5140 * If submit got -EBUSY, flag us as needing the application
5141 * to enter the kernel to reap and flush events.
5143 if (!to_submit || ret == -EBUSY) {
5145 * Drop cur_mm before scheduling, we can't hold it for
5146 * long periods (or over schedule()). Do this before
5147 * adding ourselves to the waitqueue, as the unuse/drop
5157 * We're polling. If we're within the defined idle
5158 * period, then let us spin without work before going
5159 * to sleep. The exception is if we got EBUSY doing
5160 * more IO, we should wait for the application to
5161 * reap events and wake us up.
5163 if (!list_empty(&ctx->poll_list) ||
5164 (!time_after(jiffies, timeout) && ret != -EBUSY &&
5165 !percpu_ref_is_dying(&ctx->refs))) {
5170 prepare_to_wait(&ctx->sqo_wait, &wait,
5171 TASK_INTERRUPTIBLE);
5174 * While doing polled IO, before going to sleep, we need
5175 * to check if there are new reqs added to poll_list, it
5176 * is because reqs may have been punted to io worker and
5177 * will be added to poll_list later, hence check the
5180 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
5181 !list_empty_careful(&ctx->poll_list)) {
5182 finish_wait(&ctx->sqo_wait, &wait);
5186 /* Tell userspace we may need a wakeup call */
5187 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
5188 /* make sure to read SQ tail after writing flags */
5191 to_submit = io_sqring_entries(ctx);
5192 if (!to_submit || ret == -EBUSY) {
5193 if (kthread_should_park()) {
5194 finish_wait(&ctx->sqo_wait, &wait);
5197 if (signal_pending(current))
5198 flush_signals(current);
5200 finish_wait(&ctx->sqo_wait, &wait);
5202 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5205 finish_wait(&ctx->sqo_wait, &wait);
5207 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5210 mutex_lock(&ctx->uring_lock);
5211 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
5212 mutex_unlock(&ctx->uring_lock);
5213 timeout = jiffies + ctx->sq_thread_idle;
5221 revert_creds(old_cred);
5228 struct io_wait_queue {
5229 struct wait_queue_entry wq;
5230 struct io_ring_ctx *ctx;
5232 unsigned nr_timeouts;
5235 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
5237 struct io_ring_ctx *ctx = iowq->ctx;
5240 * Wake up if we have enough events, or if a timeout occurred since we
5241 * started waiting. For timeouts, we always want to return to userspace,
5242 * regardless of event count.
5244 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
5245 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
5248 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
5249 int wake_flags, void *key)
5251 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
5254 /* use noflush == true, as we can't safely rely on locking context */
5255 if (!io_should_wake(iowq, true))
5258 return autoremove_wake_function(curr, mode, wake_flags, key);
5262 * Wait until events become available, if we don't already have some. The
5263 * application must reap them itself, as they reside on the shared cq ring.
5265 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
5266 const sigset_t __user *sig, size_t sigsz)
5268 struct io_wait_queue iowq = {
5271 .func = io_wake_function,
5272 .entry = LIST_HEAD_INIT(iowq.wq.entry),
5275 .to_wait = min_events,
5277 struct io_rings *rings = ctx->rings;
5280 if (io_cqring_events(ctx, false) >= min_events)
5284 #ifdef CONFIG_COMPAT
5285 if (in_compat_syscall())
5286 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
5290 ret = set_user_sigmask(sig, sigsz);
5296 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
5297 trace_io_uring_cqring_wait(ctx, min_events);
5299 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
5300 TASK_INTERRUPTIBLE);
5301 if (io_should_wake(&iowq, false))
5304 if (signal_pending(current)) {
5309 finish_wait(&ctx->wait, &iowq.wq);
5311 restore_saved_sigmask_unless(ret == -EINTR);
5313 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
5316 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
5318 #if defined(CONFIG_UNIX)
5319 if (ctx->ring_sock) {
5320 struct sock *sock = ctx->ring_sock->sk;
5321 struct sk_buff *skb;
5323 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
5329 for (i = 0; i < ctx->nr_user_files; i++) {
5332 file = io_file_from_index(ctx, i);
5339 static void io_file_ref_kill(struct percpu_ref *ref)
5341 struct fixed_file_data *data;
5343 data = container_of(ref, struct fixed_file_data, refs);
5344 complete(&data->done);
5347 static void io_file_ref_exit_and_free(struct work_struct *work)
5349 struct fixed_file_data *data;
5351 data = container_of(work, struct fixed_file_data, ref_work);
5354 * Ensure any percpu-ref atomic switch callback has run, it could have
5355 * been in progress when the files were being unregistered. Once
5356 * that's done, we can safely exit and free the ref and containing
5360 percpu_ref_exit(&data->refs);
5364 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
5366 struct fixed_file_data *data = ctx->file_data;
5367 unsigned nr_tables, i;
5372 percpu_ref_kill_and_confirm(&data->refs, io_file_ref_kill);
5373 flush_work(&data->ref_work);
5374 wait_for_completion(&data->done);
5375 io_ring_file_ref_flush(data);
5377 __io_sqe_files_unregister(ctx);
5378 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
5379 for (i = 0; i < nr_tables; i++)
5380 kfree(data->table[i].files);
5382 INIT_WORK(&data->ref_work, io_file_ref_exit_and_free);
5383 queue_work(system_wq, &data->ref_work);
5384 ctx->file_data = NULL;
5385 ctx->nr_user_files = 0;
5389 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
5391 if (ctx->sqo_thread) {
5392 wait_for_completion(&ctx->completions[1]);
5394 * The park is a bit of a work-around, without it we get
5395 * warning spews on shutdown with SQPOLL set and affinity
5396 * set to a single CPU.
5398 kthread_park(ctx->sqo_thread);
5399 kthread_stop(ctx->sqo_thread);
5400 ctx->sqo_thread = NULL;
5404 static void io_finish_async(struct io_ring_ctx *ctx)
5406 io_sq_thread_stop(ctx);
5409 io_wq_destroy(ctx->io_wq);
5414 #if defined(CONFIG_UNIX)
5416 * Ensure the UNIX gc is aware of our file set, so we are certain that
5417 * the io_uring can be safely unregistered on process exit, even if we have
5418 * loops in the file referencing.
5420 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
5422 struct sock *sk = ctx->ring_sock->sk;
5423 struct scm_fp_list *fpl;
5424 struct sk_buff *skb;
5427 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
5428 unsigned long inflight = ctx->user->unix_inflight + nr;
5430 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
5434 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
5438 skb = alloc_skb(0, GFP_KERNEL);
5447 fpl->user = get_uid(ctx->user);
5448 for (i = 0; i < nr; i++) {
5449 struct file *file = io_file_from_index(ctx, i + offset);
5453 fpl->fp[nr_files] = get_file(file);
5454 unix_inflight(fpl->user, fpl->fp[nr_files]);
5459 fpl->max = SCM_MAX_FD;
5460 fpl->count = nr_files;
5461 UNIXCB(skb).fp = fpl;
5462 skb->destructor = unix_destruct_scm;
5463 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
5464 skb_queue_head(&sk->sk_receive_queue, skb);
5466 for (i = 0; i < nr_files; i++)
5477 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
5478 * causes regular reference counting to break down. We rely on the UNIX
5479 * garbage collection to take care of this problem for us.
5481 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5483 unsigned left, total;
5487 left = ctx->nr_user_files;
5489 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
5491 ret = __io_sqe_files_scm(ctx, this_files, total);
5495 total += this_files;
5501 while (total < ctx->nr_user_files) {
5502 struct file *file = io_file_from_index(ctx, total);
5512 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5518 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
5523 for (i = 0; i < nr_tables; i++) {
5524 struct fixed_file_table *table = &ctx->file_data->table[i];
5525 unsigned this_files;
5527 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
5528 table->files = kcalloc(this_files, sizeof(struct file *),
5532 nr_files -= this_files;
5538 for (i = 0; i < nr_tables; i++) {
5539 struct fixed_file_table *table = &ctx->file_data->table[i];
5540 kfree(table->files);
5545 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
5547 #if defined(CONFIG_UNIX)
5548 struct sock *sock = ctx->ring_sock->sk;
5549 struct sk_buff_head list, *head = &sock->sk_receive_queue;
5550 struct sk_buff *skb;
5553 __skb_queue_head_init(&list);
5556 * Find the skb that holds this file in its SCM_RIGHTS. When found,
5557 * remove this entry and rearrange the file array.
5559 skb = skb_dequeue(head);
5561 struct scm_fp_list *fp;
5563 fp = UNIXCB(skb).fp;
5564 for (i = 0; i < fp->count; i++) {
5567 if (fp->fp[i] != file)
5570 unix_notinflight(fp->user, fp->fp[i]);
5571 left = fp->count - 1 - i;
5573 memmove(&fp->fp[i], &fp->fp[i + 1],
5574 left * sizeof(struct file *));
5581 __skb_queue_tail(&list, skb);
5591 __skb_queue_tail(&list, skb);
5593 skb = skb_dequeue(head);
5596 if (skb_peek(&list)) {
5597 spin_lock_irq(&head->lock);
5598 while ((skb = __skb_dequeue(&list)) != NULL)
5599 __skb_queue_tail(head, skb);
5600 spin_unlock_irq(&head->lock);
5607 struct io_file_put {
5608 struct llist_node llist;
5610 struct completion *done;
5613 static void io_ring_file_ref_flush(struct fixed_file_data *data)
5615 struct io_file_put *pfile, *tmp;
5616 struct llist_node *node;
5618 while ((node = llist_del_all(&data->put_llist)) != NULL) {
5619 llist_for_each_entry_safe(pfile, tmp, node, llist) {
5620 io_ring_file_put(data->ctx, pfile->file);
5622 complete(pfile->done);
5629 static void io_ring_file_ref_switch(struct work_struct *work)
5631 struct fixed_file_data *data;
5633 data = container_of(work, struct fixed_file_data, ref_work);
5634 io_ring_file_ref_flush(data);
5635 percpu_ref_switch_to_percpu(&data->refs);
5638 static void io_file_data_ref_zero(struct percpu_ref *ref)
5640 struct fixed_file_data *data;
5642 data = container_of(ref, struct fixed_file_data, refs);
5645 * We can't safely switch from inside this context, punt to wq. If
5646 * the table ref is going away, the table is being unregistered.
5647 * Don't queue up the async work for that case, the caller will
5650 if (!percpu_ref_is_dying(&data->refs))
5651 queue_work(system_wq, &data->ref_work);
5654 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
5657 __s32 __user *fds = (__s32 __user *) arg;
5667 if (nr_args > IORING_MAX_FIXED_FILES)
5670 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
5671 if (!ctx->file_data)
5673 ctx->file_data->ctx = ctx;
5674 init_completion(&ctx->file_data->done);
5676 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
5677 ctx->file_data->table = kcalloc(nr_tables,
5678 sizeof(struct fixed_file_table),
5680 if (!ctx->file_data->table) {
5681 kfree(ctx->file_data);
5682 ctx->file_data = NULL;
5686 if (percpu_ref_init(&ctx->file_data->refs, io_file_data_ref_zero,
5687 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
5688 kfree(ctx->file_data->table);
5689 kfree(ctx->file_data);
5690 ctx->file_data = NULL;
5693 ctx->file_data->put_llist.first = NULL;
5694 INIT_WORK(&ctx->file_data->ref_work, io_ring_file_ref_switch);
5696 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
5697 percpu_ref_exit(&ctx->file_data->refs);
5698 kfree(ctx->file_data->table);
5699 kfree(ctx->file_data);
5700 ctx->file_data = NULL;
5704 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
5705 struct fixed_file_table *table;
5709 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
5711 /* allow sparse sets */
5717 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5718 index = i & IORING_FILE_TABLE_MASK;
5726 * Don't allow io_uring instances to be registered. If UNIX
5727 * isn't enabled, then this causes a reference cycle and this
5728 * instance can never get freed. If UNIX is enabled we'll
5729 * handle it just fine, but there's still no point in allowing
5730 * a ring fd as it doesn't support regular read/write anyway.
5732 if (file->f_op == &io_uring_fops) {
5737 table->files[index] = file;
5741 for (i = 0; i < ctx->nr_user_files; i++) {
5742 file = io_file_from_index(ctx, i);
5746 for (i = 0; i < nr_tables; i++)
5747 kfree(ctx->file_data->table[i].files);
5749 kfree(ctx->file_data->table);
5750 kfree(ctx->file_data);
5751 ctx->file_data = NULL;
5752 ctx->nr_user_files = 0;
5756 ret = io_sqe_files_scm(ctx);
5758 io_sqe_files_unregister(ctx);
5763 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
5766 #if defined(CONFIG_UNIX)
5767 struct sock *sock = ctx->ring_sock->sk;
5768 struct sk_buff_head *head = &sock->sk_receive_queue;
5769 struct sk_buff *skb;
5772 * See if we can merge this file into an existing skb SCM_RIGHTS
5773 * file set. If there's no room, fall back to allocating a new skb
5774 * and filling it in.
5776 spin_lock_irq(&head->lock);
5777 skb = skb_peek(head);
5779 struct scm_fp_list *fpl = UNIXCB(skb).fp;
5781 if (fpl->count < SCM_MAX_FD) {
5782 __skb_unlink(skb, head);
5783 spin_unlock_irq(&head->lock);
5784 fpl->fp[fpl->count] = get_file(file);
5785 unix_inflight(fpl->user, fpl->fp[fpl->count]);
5787 spin_lock_irq(&head->lock);
5788 __skb_queue_head(head, skb);
5793 spin_unlock_irq(&head->lock);
5800 return __io_sqe_files_scm(ctx, 1, index);
5806 static void io_atomic_switch(struct percpu_ref *ref)
5808 struct fixed_file_data *data;
5811 * Juggle reference to ensure we hit zero, if needed, so we can
5812 * switch back to percpu mode
5814 data = container_of(ref, struct fixed_file_data, refs);
5815 percpu_ref_put(&data->refs);
5816 percpu_ref_get(&data->refs);
5819 static bool io_queue_file_removal(struct fixed_file_data *data,
5822 struct io_file_put *pfile, pfile_stack;
5823 DECLARE_COMPLETION_ONSTACK(done);
5826 * If we fail allocating the struct we need for doing async reomval
5827 * of this file, just punt to sync and wait for it.
5829 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
5831 pfile = &pfile_stack;
5832 pfile->done = &done;
5836 llist_add(&pfile->llist, &data->put_llist);
5838 if (pfile == &pfile_stack) {
5839 percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch);
5840 wait_for_completion(&done);
5841 flush_work(&data->ref_work);
5848 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
5849 struct io_uring_files_update *up,
5852 struct fixed_file_data *data = ctx->file_data;
5853 bool ref_switch = false;
5859 if (check_add_overflow(up->offset, nr_args, &done))
5861 if (done > ctx->nr_user_files)
5865 fds = u64_to_user_ptr(up->fds);
5867 struct fixed_file_table *table;
5871 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
5875 i = array_index_nospec(up->offset, ctx->nr_user_files);
5876 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5877 index = i & IORING_FILE_TABLE_MASK;
5878 if (table->files[index]) {
5879 file = io_file_from_index(ctx, index);
5880 table->files[index] = NULL;
5881 if (io_queue_file_removal(data, file))
5891 * Don't allow io_uring instances to be registered. If
5892 * UNIX isn't enabled, then this causes a reference
5893 * cycle and this instance can never get freed. If UNIX
5894 * is enabled we'll handle it just fine, but there's
5895 * still no point in allowing a ring fd as it doesn't
5896 * support regular read/write anyway.
5898 if (file->f_op == &io_uring_fops) {
5903 table->files[index] = file;
5904 err = io_sqe_file_register(ctx, file, i);
5914 percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch);
5916 return done ? done : err;
5918 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
5921 struct io_uring_files_update up;
5923 if (!ctx->file_data)
5927 if (copy_from_user(&up, arg, sizeof(up)))
5932 return __io_sqe_files_update(ctx, &up, nr_args);
5935 static void io_put_work(struct io_wq_work *work)
5937 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5942 static void io_get_work(struct io_wq_work *work)
5944 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5946 refcount_inc(&req->refs);
5949 static int io_init_wq_offload(struct io_ring_ctx *ctx,
5950 struct io_uring_params *p)
5952 struct io_wq_data data;
5954 struct io_ring_ctx *ctx_attach;
5955 unsigned int concurrency;
5958 data.user = ctx->user;
5959 data.get_work = io_get_work;
5960 data.put_work = io_put_work;
5962 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
5963 /* Do QD, or 4 * CPUS, whatever is smallest */
5964 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
5966 ctx->io_wq = io_wq_create(concurrency, &data);
5967 if (IS_ERR(ctx->io_wq)) {
5968 ret = PTR_ERR(ctx->io_wq);
5974 f = fdget(p->wq_fd);
5978 if (f.file->f_op != &io_uring_fops) {
5983 ctx_attach = f.file->private_data;
5984 /* @io_wq is protected by holding the fd */
5985 if (!io_wq_get(ctx_attach->io_wq, &data)) {
5990 ctx->io_wq = ctx_attach->io_wq;
5996 static int io_sq_offload_start(struct io_ring_ctx *ctx,
5997 struct io_uring_params *p)
6001 init_waitqueue_head(&ctx->sqo_wait);
6002 mmgrab(current->mm);
6003 ctx->sqo_mm = current->mm;
6005 if (ctx->flags & IORING_SETUP_SQPOLL) {
6007 if (!capable(CAP_SYS_ADMIN))
6010 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6011 if (!ctx->sq_thread_idle)
6012 ctx->sq_thread_idle = HZ;
6014 if (p->flags & IORING_SETUP_SQ_AFF) {
6015 int cpu = p->sq_thread_cpu;
6018 if (cpu >= nr_cpu_ids)
6020 if (!cpu_online(cpu))
6023 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6027 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6030 if (IS_ERR(ctx->sqo_thread)) {
6031 ret = PTR_ERR(ctx->sqo_thread);
6032 ctx->sqo_thread = NULL;
6035 wake_up_process(ctx->sqo_thread);
6036 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6037 /* Can't have SQ_AFF without SQPOLL */
6042 ret = io_init_wq_offload(ctx, p);
6048 io_finish_async(ctx);
6049 mmdrop(ctx->sqo_mm);
6054 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6056 atomic_long_sub(nr_pages, &user->locked_vm);
6059 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6061 unsigned long page_limit, cur_pages, new_pages;
6063 /* Don't allow more pages than we can safely lock */
6064 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6067 cur_pages = atomic_long_read(&user->locked_vm);
6068 new_pages = cur_pages + nr_pages;
6069 if (new_pages > page_limit)
6071 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6072 new_pages) != cur_pages);
6077 static void io_mem_free(void *ptr)
6084 page = virt_to_head_page(ptr);
6085 if (put_page_testzero(page))
6086 free_compound_page(page);
6089 static void *io_mem_alloc(size_t size)
6091 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6094 return (void *) __get_free_pages(gfp_flags, get_order(size));
6097 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6100 struct io_rings *rings;
6101 size_t off, sq_array_size;
6103 off = struct_size(rings, cqes, cq_entries);
6104 if (off == SIZE_MAX)
6108 off = ALIGN(off, SMP_CACHE_BYTES);
6113 sq_array_size = array_size(sizeof(u32), sq_entries);
6114 if (sq_array_size == SIZE_MAX)
6117 if (check_add_overflow(off, sq_array_size, &off))
6126 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6130 pages = (size_t)1 << get_order(
6131 rings_size(sq_entries, cq_entries, NULL));
6132 pages += (size_t)1 << get_order(
6133 array_size(sizeof(struct io_uring_sqe), sq_entries));
6138 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6142 if (!ctx->user_bufs)
6145 for (i = 0; i < ctx->nr_user_bufs; i++) {
6146 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6148 for (j = 0; j < imu->nr_bvecs; j++)
6149 unpin_user_page(imu->bvec[j].bv_page);
6151 if (ctx->account_mem)
6152 io_unaccount_mem(ctx->user, imu->nr_bvecs);
6157 kfree(ctx->user_bufs);
6158 ctx->user_bufs = NULL;
6159 ctx->nr_user_bufs = 0;
6163 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
6164 void __user *arg, unsigned index)
6166 struct iovec __user *src;
6168 #ifdef CONFIG_COMPAT
6170 struct compat_iovec __user *ciovs;
6171 struct compat_iovec ciov;
6173 ciovs = (struct compat_iovec __user *) arg;
6174 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
6177 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
6178 dst->iov_len = ciov.iov_len;
6182 src = (struct iovec __user *) arg;
6183 if (copy_from_user(dst, &src[index], sizeof(*dst)))
6188 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
6191 struct vm_area_struct **vmas = NULL;
6192 struct page **pages = NULL;
6193 int i, j, got_pages = 0;
6198 if (!nr_args || nr_args > UIO_MAXIOV)
6201 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
6203 if (!ctx->user_bufs)
6206 for (i = 0; i < nr_args; i++) {
6207 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6208 unsigned long off, start, end, ubuf;
6213 ret = io_copy_iov(ctx, &iov, arg, i);
6218 * Don't impose further limits on the size and buffer
6219 * constraints here, we'll -EINVAL later when IO is
6220 * submitted if they are wrong.
6223 if (!iov.iov_base || !iov.iov_len)
6226 /* arbitrary limit, but we need something */
6227 if (iov.iov_len > SZ_1G)
6230 ubuf = (unsigned long) iov.iov_base;
6231 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
6232 start = ubuf >> PAGE_SHIFT;
6233 nr_pages = end - start;
6235 if (ctx->account_mem) {
6236 ret = io_account_mem(ctx->user, nr_pages);
6242 if (!pages || nr_pages > got_pages) {
6245 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
6247 vmas = kvmalloc_array(nr_pages,
6248 sizeof(struct vm_area_struct *),
6250 if (!pages || !vmas) {
6252 if (ctx->account_mem)
6253 io_unaccount_mem(ctx->user, nr_pages);
6256 got_pages = nr_pages;
6259 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
6263 if (ctx->account_mem)
6264 io_unaccount_mem(ctx->user, nr_pages);
6269 down_read(¤t->mm->mmap_sem);
6270 pret = pin_user_pages(ubuf, nr_pages,
6271 FOLL_WRITE | FOLL_LONGTERM,
6273 if (pret == nr_pages) {
6274 /* don't support file backed memory */
6275 for (j = 0; j < nr_pages; j++) {
6276 struct vm_area_struct *vma = vmas[j];
6279 !is_file_hugepages(vma->vm_file)) {
6285 ret = pret < 0 ? pret : -EFAULT;
6287 up_read(¤t->mm->mmap_sem);
6290 * if we did partial map, or found file backed vmas,
6291 * release any pages we did get
6294 unpin_user_pages(pages, pret);
6295 if (ctx->account_mem)
6296 io_unaccount_mem(ctx->user, nr_pages);
6301 off = ubuf & ~PAGE_MASK;
6303 for (j = 0; j < nr_pages; j++) {
6306 vec_len = min_t(size_t, size, PAGE_SIZE - off);
6307 imu->bvec[j].bv_page = pages[j];
6308 imu->bvec[j].bv_len = vec_len;
6309 imu->bvec[j].bv_offset = off;
6313 /* store original address for later verification */
6315 imu->len = iov.iov_len;
6316 imu->nr_bvecs = nr_pages;
6318 ctx->nr_user_bufs++;
6326 io_sqe_buffer_unregister(ctx);
6330 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
6332 __s32 __user *fds = arg;
6338 if (copy_from_user(&fd, fds, sizeof(*fds)))
6341 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
6342 if (IS_ERR(ctx->cq_ev_fd)) {
6343 int ret = PTR_ERR(ctx->cq_ev_fd);
6344 ctx->cq_ev_fd = NULL;
6351 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
6353 if (ctx->cq_ev_fd) {
6354 eventfd_ctx_put(ctx->cq_ev_fd);
6355 ctx->cq_ev_fd = NULL;
6362 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
6364 io_finish_async(ctx);
6366 mmdrop(ctx->sqo_mm);
6368 io_iopoll_reap_events(ctx);
6369 io_sqe_buffer_unregister(ctx);
6370 io_sqe_files_unregister(ctx);
6371 io_eventfd_unregister(ctx);
6372 idr_destroy(&ctx->personality_idr);
6374 #if defined(CONFIG_UNIX)
6375 if (ctx->ring_sock) {
6376 ctx->ring_sock->file = NULL; /* so that iput() is called */
6377 sock_release(ctx->ring_sock);
6381 io_mem_free(ctx->rings);
6382 io_mem_free(ctx->sq_sqes);
6384 percpu_ref_exit(&ctx->refs);
6385 if (ctx->account_mem)
6386 io_unaccount_mem(ctx->user,
6387 ring_pages(ctx->sq_entries, ctx->cq_entries));
6388 free_uid(ctx->user);
6389 put_cred(ctx->creds);
6390 kfree(ctx->completions);
6391 kfree(ctx->cancel_hash);
6392 kmem_cache_free(req_cachep, ctx->fallback_req);
6396 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
6398 struct io_ring_ctx *ctx = file->private_data;
6401 poll_wait(file, &ctx->cq_wait, wait);
6403 * synchronizes with barrier from wq_has_sleeper call in
6407 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
6408 ctx->rings->sq_ring_entries)
6409 mask |= EPOLLOUT | EPOLLWRNORM;
6410 if (io_cqring_events(ctx, false))
6411 mask |= EPOLLIN | EPOLLRDNORM;
6416 static int io_uring_fasync(int fd, struct file *file, int on)
6418 struct io_ring_ctx *ctx = file->private_data;
6420 return fasync_helper(fd, file, on, &ctx->cq_fasync);
6423 static int io_remove_personalities(int id, void *p, void *data)
6425 struct io_ring_ctx *ctx = data;
6426 const struct cred *cred;
6428 cred = idr_remove(&ctx->personality_idr, id);
6434 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
6436 mutex_lock(&ctx->uring_lock);
6437 percpu_ref_kill(&ctx->refs);
6438 mutex_unlock(&ctx->uring_lock);
6441 * Wait for sq thread to idle, if we have one. It won't spin on new
6442 * work after we've killed the ctx ref above. This is important to do
6443 * before we cancel existing commands, as the thread could otherwise
6444 * be queueing new work post that. If that's work we need to cancel,
6445 * it could cause shutdown to hang.
6447 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
6450 io_kill_timeouts(ctx);
6451 io_poll_remove_all(ctx);
6454 io_wq_cancel_all(ctx->io_wq);
6456 io_iopoll_reap_events(ctx);
6457 /* if we failed setting up the ctx, we might not have any rings */
6459 io_cqring_overflow_flush(ctx, true);
6460 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
6461 wait_for_completion(&ctx->completions[0]);
6462 io_ring_ctx_free(ctx);
6465 static int io_uring_release(struct inode *inode, struct file *file)
6467 struct io_ring_ctx *ctx = file->private_data;
6469 file->private_data = NULL;
6470 io_ring_ctx_wait_and_kill(ctx);
6474 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
6475 struct files_struct *files)
6477 struct io_kiocb *req;
6480 while (!list_empty_careful(&ctx->inflight_list)) {
6481 struct io_kiocb *cancel_req = NULL;
6483 spin_lock_irq(&ctx->inflight_lock);
6484 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
6485 if (req->work.files != files)
6487 /* req is being completed, ignore */
6488 if (!refcount_inc_not_zero(&req->refs))
6494 prepare_to_wait(&ctx->inflight_wait, &wait,
6495 TASK_UNINTERRUPTIBLE);
6496 spin_unlock_irq(&ctx->inflight_lock);
6498 /* We need to keep going until we don't find a matching req */
6502 if (cancel_req->flags & REQ_F_OVERFLOW) {
6503 spin_lock_irq(&ctx->completion_lock);
6504 list_del(&cancel_req->list);
6505 cancel_req->flags &= ~REQ_F_OVERFLOW;
6506 if (list_empty(&ctx->cq_overflow_list)) {
6507 clear_bit(0, &ctx->sq_check_overflow);
6508 clear_bit(0, &ctx->cq_check_overflow);
6510 spin_unlock_irq(&ctx->completion_lock);
6512 WRITE_ONCE(ctx->rings->cq_overflow,
6513 atomic_inc_return(&ctx->cached_cq_overflow));
6516 * Put inflight ref and overflow ref. If that's
6517 * all we had, then we're done with this request.
6519 if (refcount_sub_and_test(2, &cancel_req->refs)) {
6520 io_put_req(cancel_req);
6525 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
6526 io_put_req(cancel_req);
6529 finish_wait(&ctx->inflight_wait, &wait);
6532 static int io_uring_flush(struct file *file, void *data)
6534 struct io_ring_ctx *ctx = file->private_data;
6536 io_uring_cancel_files(ctx, data);
6539 * If the task is going away, cancel work it may have pending
6541 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
6542 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
6547 static void *io_uring_validate_mmap_request(struct file *file,
6548 loff_t pgoff, size_t sz)
6550 struct io_ring_ctx *ctx = file->private_data;
6551 loff_t offset = pgoff << PAGE_SHIFT;
6556 case IORING_OFF_SQ_RING:
6557 case IORING_OFF_CQ_RING:
6560 case IORING_OFF_SQES:
6564 return ERR_PTR(-EINVAL);
6567 page = virt_to_head_page(ptr);
6568 if (sz > page_size(page))
6569 return ERR_PTR(-EINVAL);
6576 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6578 size_t sz = vma->vm_end - vma->vm_start;
6582 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
6584 return PTR_ERR(ptr);
6586 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
6587 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
6590 #else /* !CONFIG_MMU */
6592 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6594 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
6597 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
6599 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
6602 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
6603 unsigned long addr, unsigned long len,
6604 unsigned long pgoff, unsigned long flags)
6608 ptr = io_uring_validate_mmap_request(file, pgoff, len);
6610 return PTR_ERR(ptr);
6612 return (unsigned long) ptr;
6615 #endif /* !CONFIG_MMU */
6617 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
6618 u32, min_complete, u32, flags, const sigset_t __user *, sig,
6621 struct io_ring_ctx *ctx;
6626 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
6634 if (f.file->f_op != &io_uring_fops)
6638 ctx = f.file->private_data;
6639 if (!percpu_ref_tryget(&ctx->refs))
6643 * For SQ polling, the thread will do all submissions and completions.
6644 * Just return the requested submit count, and wake the thread if
6648 if (ctx->flags & IORING_SETUP_SQPOLL) {
6649 if (!list_empty_careful(&ctx->cq_overflow_list))
6650 io_cqring_overflow_flush(ctx, false);
6651 if (flags & IORING_ENTER_SQ_WAKEUP)
6652 wake_up(&ctx->sqo_wait);
6653 submitted = to_submit;
6654 } else if (to_submit) {
6655 struct mm_struct *cur_mm;
6657 mutex_lock(&ctx->uring_lock);
6658 /* already have mm, so io_submit_sqes() won't try to grab it */
6659 cur_mm = ctx->sqo_mm;
6660 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
6662 mutex_unlock(&ctx->uring_lock);
6664 if (submitted != to_submit)
6667 if (flags & IORING_ENTER_GETEVENTS) {
6668 unsigned nr_events = 0;
6670 min_complete = min(min_complete, ctx->cq_entries);
6672 if (ctx->flags & IORING_SETUP_IOPOLL) {
6673 ret = io_iopoll_check(ctx, &nr_events, min_complete);
6675 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
6680 percpu_ref_put(&ctx->refs);
6683 return submitted ? submitted : ret;
6686 #ifdef CONFIG_PROC_FS
6687 static int io_uring_show_cred(int id, void *p, void *data)
6689 const struct cred *cred = p;
6690 struct seq_file *m = data;
6691 struct user_namespace *uns = seq_user_ns(m);
6692 struct group_info *gi;
6697 seq_printf(m, "%5d\n", id);
6698 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
6699 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
6700 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
6701 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
6702 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
6703 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
6704 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
6705 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
6706 seq_puts(m, "\n\tGroups:\t");
6707 gi = cred->group_info;
6708 for (g = 0; g < gi->ngroups; g++) {
6709 seq_put_decimal_ull(m, g ? " " : "",
6710 from_kgid_munged(uns, gi->gid[g]));
6712 seq_puts(m, "\n\tCapEff:\t");
6713 cap = cred->cap_effective;
6714 CAP_FOR_EACH_U32(__capi)
6715 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
6720 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
6724 mutex_lock(&ctx->uring_lock);
6725 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
6726 for (i = 0; i < ctx->nr_user_files; i++) {
6727 struct fixed_file_table *table;
6730 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6731 f = table->files[i & IORING_FILE_TABLE_MASK];
6733 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
6735 seq_printf(m, "%5u: <none>\n", i);
6737 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
6738 for (i = 0; i < ctx->nr_user_bufs; i++) {
6739 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
6741 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
6742 (unsigned int) buf->len);
6744 if (!idr_is_empty(&ctx->personality_idr)) {
6745 seq_printf(m, "Personalities:\n");
6746 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
6748 mutex_unlock(&ctx->uring_lock);
6751 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
6753 struct io_ring_ctx *ctx = f->private_data;
6755 if (percpu_ref_tryget(&ctx->refs)) {
6756 __io_uring_show_fdinfo(ctx, m);
6757 percpu_ref_put(&ctx->refs);
6762 static const struct file_operations io_uring_fops = {
6763 .release = io_uring_release,
6764 .flush = io_uring_flush,
6765 .mmap = io_uring_mmap,
6767 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
6768 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
6770 .poll = io_uring_poll,
6771 .fasync = io_uring_fasync,
6772 #ifdef CONFIG_PROC_FS
6773 .show_fdinfo = io_uring_show_fdinfo,
6777 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
6778 struct io_uring_params *p)
6780 struct io_rings *rings;
6781 size_t size, sq_array_offset;
6783 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
6784 if (size == SIZE_MAX)
6787 rings = io_mem_alloc(size);
6792 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
6793 rings->sq_ring_mask = p->sq_entries - 1;
6794 rings->cq_ring_mask = p->cq_entries - 1;
6795 rings->sq_ring_entries = p->sq_entries;
6796 rings->cq_ring_entries = p->cq_entries;
6797 ctx->sq_mask = rings->sq_ring_mask;
6798 ctx->cq_mask = rings->cq_ring_mask;
6799 ctx->sq_entries = rings->sq_ring_entries;
6800 ctx->cq_entries = rings->cq_ring_entries;
6802 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
6803 if (size == SIZE_MAX) {
6804 io_mem_free(ctx->rings);
6809 ctx->sq_sqes = io_mem_alloc(size);
6810 if (!ctx->sq_sqes) {
6811 io_mem_free(ctx->rings);
6820 * Allocate an anonymous fd, this is what constitutes the application
6821 * visible backing of an io_uring instance. The application mmaps this
6822 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
6823 * we have to tie this fd to a socket for file garbage collection purposes.
6825 static int io_uring_get_fd(struct io_ring_ctx *ctx)
6830 #if defined(CONFIG_UNIX)
6831 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
6837 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
6841 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
6842 O_RDWR | O_CLOEXEC);
6845 ret = PTR_ERR(file);
6849 #if defined(CONFIG_UNIX)
6850 ctx->ring_sock->file = file;
6852 fd_install(ret, file);
6855 #if defined(CONFIG_UNIX)
6856 sock_release(ctx->ring_sock);
6857 ctx->ring_sock = NULL;
6862 static int io_uring_create(unsigned entries, struct io_uring_params *p)
6864 struct user_struct *user = NULL;
6865 struct io_ring_ctx *ctx;
6871 if (entries > IORING_MAX_ENTRIES) {
6872 if (!(p->flags & IORING_SETUP_CLAMP))
6874 entries = IORING_MAX_ENTRIES;
6878 * Use twice as many entries for the CQ ring. It's possible for the
6879 * application to drive a higher depth than the size of the SQ ring,
6880 * since the sqes are only used at submission time. This allows for
6881 * some flexibility in overcommitting a bit. If the application has
6882 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
6883 * of CQ ring entries manually.
6885 p->sq_entries = roundup_pow_of_two(entries);
6886 if (p->flags & IORING_SETUP_CQSIZE) {
6888 * If IORING_SETUP_CQSIZE is set, we do the same roundup
6889 * to a power-of-two, if it isn't already. We do NOT impose
6890 * any cq vs sq ring sizing.
6892 if (p->cq_entries < p->sq_entries)
6894 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
6895 if (!(p->flags & IORING_SETUP_CLAMP))
6897 p->cq_entries = IORING_MAX_CQ_ENTRIES;
6899 p->cq_entries = roundup_pow_of_two(p->cq_entries);
6901 p->cq_entries = 2 * p->sq_entries;
6904 user = get_uid(current_user());
6905 account_mem = !capable(CAP_IPC_LOCK);
6908 ret = io_account_mem(user,
6909 ring_pages(p->sq_entries, p->cq_entries));
6916 ctx = io_ring_ctx_alloc(p);
6919 io_unaccount_mem(user, ring_pages(p->sq_entries,
6924 ctx->compat = in_compat_syscall();
6925 ctx->account_mem = account_mem;
6927 ctx->creds = get_current_cred();
6929 ret = io_allocate_scq_urings(ctx, p);
6933 ret = io_sq_offload_start(ctx, p);
6937 memset(&p->sq_off, 0, sizeof(p->sq_off));
6938 p->sq_off.head = offsetof(struct io_rings, sq.head);
6939 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
6940 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
6941 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
6942 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
6943 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
6944 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
6946 memset(&p->cq_off, 0, sizeof(p->cq_off));
6947 p->cq_off.head = offsetof(struct io_rings, cq.head);
6948 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
6949 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
6950 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
6951 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
6952 p->cq_off.cqes = offsetof(struct io_rings, cqes);
6955 * Install ring fd as the very last thing, so we don't risk someone
6956 * having closed it before we finish setup
6958 ret = io_uring_get_fd(ctx);
6962 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
6963 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
6964 IORING_FEAT_CUR_PERSONALITY;
6965 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
6968 io_ring_ctx_wait_and_kill(ctx);
6973 * Sets up an aio uring context, and returns the fd. Applications asks for a
6974 * ring size, we return the actual sq/cq ring sizes (among other things) in the
6975 * params structure passed in.
6977 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
6979 struct io_uring_params p;
6983 if (copy_from_user(&p, params, sizeof(p)))
6985 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
6990 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
6991 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
6992 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
6995 ret = io_uring_create(entries, &p);
6999 if (copy_to_user(params, &p, sizeof(p)))
7005 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7006 struct io_uring_params __user *, params)
7008 return io_uring_setup(entries, params);
7011 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7013 struct io_uring_probe *p;
7017 size = struct_size(p, ops, nr_args);
7018 if (size == SIZE_MAX)
7020 p = kzalloc(size, GFP_KERNEL);
7025 if (copy_from_user(p, arg, size))
7028 if (memchr_inv(p, 0, size))
7031 p->last_op = IORING_OP_LAST - 1;
7032 if (nr_args > IORING_OP_LAST)
7033 nr_args = IORING_OP_LAST;
7035 for (i = 0; i < nr_args; i++) {
7037 if (!io_op_defs[i].not_supported)
7038 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7043 if (copy_to_user(arg, p, size))
7050 static int io_register_personality(struct io_ring_ctx *ctx)
7052 const struct cred *creds = get_current_cred();
7055 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7056 USHRT_MAX, GFP_KERNEL);
7062 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7064 const struct cred *old_creds;
7066 old_creds = idr_remove(&ctx->personality_idr, id);
7068 put_cred(old_creds);
7075 static bool io_register_op_must_quiesce(int op)
7078 case IORING_UNREGISTER_FILES:
7079 case IORING_REGISTER_FILES_UPDATE:
7080 case IORING_REGISTER_PROBE:
7081 case IORING_REGISTER_PERSONALITY:
7082 case IORING_UNREGISTER_PERSONALITY:
7089 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
7090 void __user *arg, unsigned nr_args)
7091 __releases(ctx->uring_lock)
7092 __acquires(ctx->uring_lock)
7097 * We're inside the ring mutex, if the ref is already dying, then
7098 * someone else killed the ctx or is already going through
7099 * io_uring_register().
7101 if (percpu_ref_is_dying(&ctx->refs))
7104 if (io_register_op_must_quiesce(opcode)) {
7105 percpu_ref_kill(&ctx->refs);
7108 * Drop uring mutex before waiting for references to exit. If
7109 * another thread is currently inside io_uring_enter() it might
7110 * need to grab the uring_lock to make progress. If we hold it
7111 * here across the drain wait, then we can deadlock. It's safe
7112 * to drop the mutex here, since no new references will come in
7113 * after we've killed the percpu ref.
7115 mutex_unlock(&ctx->uring_lock);
7116 ret = wait_for_completion_interruptible(&ctx->completions[0]);
7117 mutex_lock(&ctx->uring_lock);
7119 percpu_ref_resurrect(&ctx->refs);
7126 case IORING_REGISTER_BUFFERS:
7127 ret = io_sqe_buffer_register(ctx, arg, nr_args);
7129 case IORING_UNREGISTER_BUFFERS:
7133 ret = io_sqe_buffer_unregister(ctx);
7135 case IORING_REGISTER_FILES:
7136 ret = io_sqe_files_register(ctx, arg, nr_args);
7138 case IORING_UNREGISTER_FILES:
7142 ret = io_sqe_files_unregister(ctx);
7144 case IORING_REGISTER_FILES_UPDATE:
7145 ret = io_sqe_files_update(ctx, arg, nr_args);
7147 case IORING_REGISTER_EVENTFD:
7148 case IORING_REGISTER_EVENTFD_ASYNC:
7152 ret = io_eventfd_register(ctx, arg);
7155 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
7156 ctx->eventfd_async = 1;
7158 ctx->eventfd_async = 0;
7160 case IORING_UNREGISTER_EVENTFD:
7164 ret = io_eventfd_unregister(ctx);
7166 case IORING_REGISTER_PROBE:
7168 if (!arg || nr_args > 256)
7170 ret = io_probe(ctx, arg, nr_args);
7172 case IORING_REGISTER_PERSONALITY:
7176 ret = io_register_personality(ctx);
7178 case IORING_UNREGISTER_PERSONALITY:
7182 ret = io_unregister_personality(ctx, nr_args);
7189 if (io_register_op_must_quiesce(opcode)) {
7190 /* bring the ctx back to life */
7191 percpu_ref_reinit(&ctx->refs);
7193 reinit_completion(&ctx->completions[0]);
7198 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
7199 void __user *, arg, unsigned int, nr_args)
7201 struct io_ring_ctx *ctx;
7210 if (f.file->f_op != &io_uring_fops)
7213 ctx = f.file->private_data;
7215 mutex_lock(&ctx->uring_lock);
7216 ret = __io_uring_register(ctx, opcode, arg, nr_args);
7217 mutex_unlock(&ctx->uring_lock);
7218 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
7219 ctx->cq_ev_fd != NULL, ret);
7225 static int __init io_uring_init(void)
7227 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
7228 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
7229 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
7232 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
7233 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
7234 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
7235 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
7236 BUILD_BUG_SQE_ELEM(1, __u8, flags);
7237 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
7238 BUILD_BUG_SQE_ELEM(4, __s32, fd);
7239 BUILD_BUG_SQE_ELEM(8, __u64, off);
7240 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
7241 BUILD_BUG_SQE_ELEM(16, __u64, addr);
7242 BUILD_BUG_SQE_ELEM(24, __u32, len);
7243 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
7244 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
7245 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
7246 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
7247 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
7248 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
7249 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
7250 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
7251 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
7252 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
7253 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
7254 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
7255 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
7256 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
7257 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
7258 BUILD_BUG_SQE_ELEM(42, __u16, personality);
7260 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
7261 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
7264 __initcall(io_uring_init);