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>
79 #define CREATE_TRACE_POINTS
80 #include <trace/events/io_uring.h>
82 #include <uapi/linux/io_uring.h>
87 #define IORING_MAX_ENTRIES 32768
88 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
91 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
93 #define IORING_FILE_TABLE_SHIFT 9
94 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
95 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
96 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
99 u32 head ____cacheline_aligned_in_smp;
100 u32 tail ____cacheline_aligned_in_smp;
104 * This data is shared with the application through the mmap at offsets
105 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
107 * The offsets to the member fields are published through struct
108 * io_sqring_offsets when calling io_uring_setup.
112 * Head and tail offsets into the ring; the offsets need to be
113 * masked to get valid indices.
115 * The kernel controls head of the sq ring and the tail of the cq ring,
116 * and the application controls tail of the sq ring and the head of the
119 struct io_uring sq, cq;
121 * Bitmasks to apply to head and tail offsets (constant, equals
124 u32 sq_ring_mask, cq_ring_mask;
125 /* Ring sizes (constant, power of 2) */
126 u32 sq_ring_entries, cq_ring_entries;
128 * Number of invalid entries dropped by the kernel due to
129 * invalid index stored in array
131 * Written by the kernel, shouldn't be modified by the
132 * application (i.e. get number of "new events" by comparing to
135 * After a new SQ head value was read by the application this
136 * counter includes all submissions that were dropped reaching
137 * the new SQ head (and possibly more).
143 * Written by the kernel, shouldn't be modified by the
146 * The application needs a full memory barrier before checking
147 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
151 * Number of completion events lost because the queue was full;
152 * this should be avoided by the application by making sure
153 * there are not more requests pending than there is space in
154 * the completion queue.
156 * Written by the kernel, shouldn't be modified by the
157 * application (i.e. get number of "new events" by comparing to
160 * As completion events come in out of order this counter is not
161 * ordered with any other data.
165 * Ring buffer of completion events.
167 * The kernel writes completion events fresh every time they are
168 * produced, so the application is allowed to modify pending
171 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
174 struct io_mapped_ubuf {
177 struct bio_vec *bvec;
178 unsigned int nr_bvecs;
181 struct fixed_file_table {
189 struct fixed_file_data {
190 struct fixed_file_table *table;
191 struct io_ring_ctx *ctx;
193 struct percpu_ref refs;
194 struct llist_head put_llist;
196 struct work_struct ref_work;
197 struct completion done;
202 struct percpu_ref refs;
203 } ____cacheline_aligned_in_smp;
209 int cq_overflow_flushed: 1;
211 int eventfd_async: 1;
214 * Ring buffer of indices into array of io_uring_sqe, which is
215 * mmapped by the application using the IORING_OFF_SQES offset.
217 * This indirection could e.g. be used to assign fixed
218 * io_uring_sqe entries to operations and only submit them to
219 * the queue when needed.
221 * The kernel modifies neither the indices array nor the entries
225 unsigned cached_sq_head;
228 unsigned sq_thread_idle;
229 unsigned cached_sq_dropped;
230 atomic_t cached_cq_overflow;
231 unsigned long sq_check_overflow;
233 struct list_head defer_list;
234 struct list_head timeout_list;
235 struct list_head cq_overflow_list;
237 wait_queue_head_t inflight_wait;
238 struct io_uring_sqe *sq_sqes;
239 } ____cacheline_aligned_in_smp;
241 struct io_rings *rings;
245 struct task_struct *sqo_thread; /* if using sq thread polling */
246 struct mm_struct *sqo_mm;
247 wait_queue_head_t sqo_wait;
250 * If used, fixed file set. Writers must ensure that ->refs is dead,
251 * readers must ensure that ->refs is alive as long as the file* is
252 * used. Only updated through io_uring_register(2).
254 struct fixed_file_data *file_data;
255 unsigned nr_user_files;
257 struct file *ring_file;
259 /* if used, fixed mapped user buffers */
260 unsigned nr_user_bufs;
261 struct io_mapped_ubuf *user_bufs;
263 struct user_struct *user;
265 const struct cred *creds;
267 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
268 struct completion *completions;
270 /* if all else fails... */
271 struct io_kiocb *fallback_req;
273 #if defined(CONFIG_UNIX)
274 struct socket *ring_sock;
277 struct idr personality_idr;
280 unsigned cached_cq_tail;
283 atomic_t cq_timeouts;
284 unsigned long cq_check_overflow;
285 struct wait_queue_head cq_wait;
286 struct fasync_struct *cq_fasync;
287 struct eventfd_ctx *cq_ev_fd;
288 } ____cacheline_aligned_in_smp;
291 struct mutex uring_lock;
292 wait_queue_head_t wait;
293 } ____cacheline_aligned_in_smp;
296 spinlock_t completion_lock;
297 struct llist_head poll_llist;
300 * ->poll_list is protected by the ctx->uring_lock for
301 * io_uring instances that don't use IORING_SETUP_SQPOLL.
302 * For SQPOLL, only the single threaded io_sq_thread() will
303 * manipulate the list, hence no extra locking is needed there.
305 struct list_head poll_list;
306 struct hlist_head *cancel_hash;
307 unsigned cancel_hash_bits;
308 bool poll_multi_file;
310 spinlock_t inflight_lock;
311 struct list_head inflight_list;
312 } ____cacheline_aligned_in_smp;
316 * First field must be the file pointer in all the
317 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
319 struct io_poll_iocb {
322 struct wait_queue_head *head;
328 struct wait_queue_entry wait;
333 struct file *put_file;
337 struct io_timeout_data {
338 struct io_kiocb *req;
339 struct hrtimer timer;
340 struct timespec64 ts;
341 enum hrtimer_mode mode;
347 struct sockaddr __user *addr;
348 int __user *addr_len;
373 /* NOTE: kiocb has the file as the first member, so don't do it here */
381 struct sockaddr __user *addr;
388 struct user_msghdr __user *msg;
401 struct filename *filename;
402 struct statx __user *buffer;
406 struct io_files_update {
432 struct epoll_event event;
435 struct io_async_connect {
436 struct sockaddr_storage address;
439 struct io_async_msghdr {
440 struct iovec fast_iov[UIO_FASTIOV];
442 struct sockaddr __user *uaddr;
447 struct iovec fast_iov[UIO_FASTIOV];
453 struct io_async_open {
454 struct filename *filename;
457 struct io_async_ctx {
459 struct io_async_rw rw;
460 struct io_async_msghdr msg;
461 struct io_async_connect connect;
462 struct io_timeout_data timeout;
463 struct io_async_open open;
468 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
469 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
470 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
471 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
472 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
479 REQ_F_IOPOLL_COMPLETED_BIT,
480 REQ_F_LINK_TIMEOUT_BIT,
484 REQ_F_TIMEOUT_NOSEQ_BIT,
485 REQ_F_COMP_LOCKED_BIT,
490 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
491 /* drain existing IO first */
492 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
494 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
495 /* doesn't sever on completion < 0 */
496 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
498 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
500 /* already grabbed next link */
501 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
502 /* fail rest of links */
503 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
504 /* on inflight list */
505 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
506 /* read/write uses file position */
507 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
508 /* must not punt to workers */
509 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
510 /* polled IO has completed */
511 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
512 /* has linked timeout */
513 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
514 /* timeout request */
515 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
517 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
518 /* must be punted even for NONBLOCK */
519 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
520 /* no timeout sequence */
521 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
522 /* completion under lock */
523 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_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;
558 bool needs_fixed_file;
561 struct io_ring_ctx *ctx;
563 struct list_head list;
564 struct hlist_node hash_node;
566 struct list_head link_list;
573 struct list_head inflight_entry;
575 struct io_wq_work work;
578 #define IO_PLUG_THRESHOLD 2
579 #define IO_IOPOLL_BATCH 8
581 struct io_submit_state {
582 struct blk_plug plug;
585 * io_kiocb alloc cache
587 void *reqs[IO_IOPOLL_BATCH];
588 unsigned int free_reqs;
589 unsigned int cur_req;
592 * File reference cache
596 unsigned int has_refs;
597 unsigned int used_refs;
598 unsigned int ios_left;
602 /* needs req->io allocated for deferral/async */
603 unsigned async_ctx : 1;
604 /* needs current->mm setup, does mm access */
605 unsigned needs_mm : 1;
606 /* needs req->file assigned */
607 unsigned needs_file : 1;
608 /* needs req->file assigned IFF fd is >= 0 */
609 unsigned fd_non_neg : 1;
610 /* hash wq insertion if file is a regular file */
611 unsigned hash_reg_file : 1;
612 /* unbound wq insertion if file is a non-regular file */
613 unsigned unbound_nonreg_file : 1;
614 /* opcode is not supported by this kernel */
615 unsigned not_supported : 1;
616 /* needs file table */
617 unsigned file_table : 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,
661 [IORING_OP_RECVMSG] = {
665 .unbound_nonreg_file = 1,
667 [IORING_OP_TIMEOUT] = {
671 [IORING_OP_TIMEOUT_REMOVE] = {},
672 [IORING_OP_ACCEPT] = {
675 .unbound_nonreg_file = 1,
678 [IORING_OP_ASYNC_CANCEL] = {},
679 [IORING_OP_LINK_TIMEOUT] = {
683 [IORING_OP_CONNECT] = {
687 .unbound_nonreg_file = 1,
689 [IORING_OP_FALLOCATE] = {
692 [IORING_OP_OPENAT] = {
697 [IORING_OP_CLOSE] = {
701 [IORING_OP_FILES_UPDATE] = {
705 [IORING_OP_STATX] = {
713 .unbound_nonreg_file = 1,
715 [IORING_OP_WRITE] = {
718 .unbound_nonreg_file = 1,
720 [IORING_OP_FADVISE] = {
723 [IORING_OP_MADVISE] = {
729 .unbound_nonreg_file = 1,
734 .unbound_nonreg_file = 1,
736 [IORING_OP_OPENAT2] = {
741 [IORING_OP_EPOLL_CTL] = {
742 .unbound_nonreg_file = 1,
747 static void io_wq_submit_work(struct io_wq_work **workptr);
748 static void io_cqring_fill_event(struct io_kiocb *req, long res);
749 static void io_put_req(struct io_kiocb *req);
750 static void __io_double_put_req(struct io_kiocb *req);
751 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
752 static void io_queue_linked_timeout(struct io_kiocb *req);
753 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
754 struct io_uring_files_update *ip,
756 static int io_grab_files(struct io_kiocb *req);
758 static struct kmem_cache *req_cachep;
760 static const struct file_operations io_uring_fops;
762 struct sock *io_uring_get_socket(struct file *file)
764 #if defined(CONFIG_UNIX)
765 if (file->f_op == &io_uring_fops) {
766 struct io_ring_ctx *ctx = file->private_data;
768 return ctx->ring_sock->sk;
773 EXPORT_SYMBOL(io_uring_get_socket);
775 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
777 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
779 complete(&ctx->completions[0]);
782 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
784 struct io_ring_ctx *ctx;
787 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
791 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
792 if (!ctx->fallback_req)
795 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
796 if (!ctx->completions)
800 * Use 5 bits less than the max cq entries, that should give us around
801 * 32 entries per hash list if totally full and uniformly spread.
803 hash_bits = ilog2(p->cq_entries);
807 ctx->cancel_hash_bits = hash_bits;
808 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
810 if (!ctx->cancel_hash)
812 __hash_init(ctx->cancel_hash, 1U << hash_bits);
814 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
815 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
818 ctx->flags = p->flags;
819 init_waitqueue_head(&ctx->cq_wait);
820 INIT_LIST_HEAD(&ctx->cq_overflow_list);
821 init_completion(&ctx->completions[0]);
822 init_completion(&ctx->completions[1]);
823 idr_init(&ctx->personality_idr);
824 mutex_init(&ctx->uring_lock);
825 init_waitqueue_head(&ctx->wait);
826 spin_lock_init(&ctx->completion_lock);
827 init_llist_head(&ctx->poll_llist);
828 INIT_LIST_HEAD(&ctx->poll_list);
829 INIT_LIST_HEAD(&ctx->defer_list);
830 INIT_LIST_HEAD(&ctx->timeout_list);
831 init_waitqueue_head(&ctx->inflight_wait);
832 spin_lock_init(&ctx->inflight_lock);
833 INIT_LIST_HEAD(&ctx->inflight_list);
836 if (ctx->fallback_req)
837 kmem_cache_free(req_cachep, ctx->fallback_req);
838 kfree(ctx->completions);
839 kfree(ctx->cancel_hash);
844 static inline bool __req_need_defer(struct io_kiocb *req)
846 struct io_ring_ctx *ctx = req->ctx;
848 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
849 + atomic_read(&ctx->cached_cq_overflow);
852 static inline bool req_need_defer(struct io_kiocb *req)
854 if (unlikely(req->flags & REQ_F_IO_DRAIN))
855 return __req_need_defer(req);
860 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
862 struct io_kiocb *req;
864 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
865 if (req && !req_need_defer(req)) {
866 list_del_init(&req->list);
873 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
875 struct io_kiocb *req;
877 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
879 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
881 if (!__req_need_defer(req)) {
882 list_del_init(&req->list);
890 static void __io_commit_cqring(struct io_ring_ctx *ctx)
892 struct io_rings *rings = ctx->rings;
894 /* order cqe stores with ring update */
895 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
897 if (wq_has_sleeper(&ctx->cq_wait)) {
898 wake_up_interruptible(&ctx->cq_wait);
899 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
903 static inline void io_req_work_grab_env(struct io_kiocb *req,
904 const struct io_op_def *def)
906 if (!req->work.mm && def->needs_mm) {
908 req->work.mm = current->mm;
910 if (!req->work.creds)
911 req->work.creds = get_current_cred();
914 static inline void io_req_work_drop_env(struct io_kiocb *req)
917 mmdrop(req->work.mm);
920 if (req->work.creds) {
921 put_cred(req->work.creds);
922 req->work.creds = NULL;
926 static inline bool io_prep_async_work(struct io_kiocb *req,
927 struct io_kiocb **link)
929 const struct io_op_def *def = &io_op_defs[req->opcode];
930 bool do_hashed = false;
932 if (req->flags & REQ_F_ISREG) {
933 if (def->hash_reg_file)
936 if (def->unbound_nonreg_file)
937 req->work.flags |= IO_WQ_WORK_UNBOUND;
940 io_req_work_grab_env(req, def);
942 *link = io_prep_linked_timeout(req);
946 static inline void io_queue_async_work(struct io_kiocb *req)
948 struct io_ring_ctx *ctx = req->ctx;
949 struct io_kiocb *link;
952 do_hashed = io_prep_async_work(req, &link);
954 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
957 io_wq_enqueue(ctx->io_wq, &req->work);
959 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
960 file_inode(req->file));
964 io_queue_linked_timeout(link);
967 static void io_kill_timeout(struct io_kiocb *req)
971 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
973 atomic_inc(&req->ctx->cq_timeouts);
974 list_del_init(&req->list);
975 io_cqring_fill_event(req, 0);
980 static void io_kill_timeouts(struct io_ring_ctx *ctx)
982 struct io_kiocb *req, *tmp;
984 spin_lock_irq(&ctx->completion_lock);
985 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
986 io_kill_timeout(req);
987 spin_unlock_irq(&ctx->completion_lock);
990 static void io_commit_cqring(struct io_ring_ctx *ctx)
992 struct io_kiocb *req;
994 while ((req = io_get_timeout_req(ctx)) != NULL)
995 io_kill_timeout(req);
997 __io_commit_cqring(ctx);
999 while ((req = io_get_deferred_req(ctx)) != NULL)
1000 io_queue_async_work(req);
1003 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1005 struct io_rings *rings = ctx->rings;
1008 tail = ctx->cached_cq_tail;
1010 * writes to the cq entry need to come after reading head; the
1011 * control dependency is enough as we're using WRITE_ONCE to
1014 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1017 ctx->cached_cq_tail++;
1018 return &rings->cqes[tail & ctx->cq_mask];
1021 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1025 if (!ctx->eventfd_async)
1027 return io_wq_current_is_worker() || in_interrupt();
1030 static void __io_cqring_ev_posted(struct io_ring_ctx *ctx, bool trigger_ev)
1032 if (waitqueue_active(&ctx->wait))
1033 wake_up(&ctx->wait);
1034 if (waitqueue_active(&ctx->sqo_wait))
1035 wake_up(&ctx->sqo_wait);
1037 eventfd_signal(ctx->cq_ev_fd, 1);
1040 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1042 __io_cqring_ev_posted(ctx, io_should_trigger_evfd(ctx));
1045 /* Returns true if there are no backlogged entries after the flush */
1046 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1048 struct io_rings *rings = ctx->rings;
1049 struct io_uring_cqe *cqe;
1050 struct io_kiocb *req;
1051 unsigned long flags;
1055 if (list_empty_careful(&ctx->cq_overflow_list))
1057 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1058 rings->cq_ring_entries))
1062 spin_lock_irqsave(&ctx->completion_lock, flags);
1064 /* if force is set, the ring is going away. always drop after that */
1066 ctx->cq_overflow_flushed = 1;
1069 while (!list_empty(&ctx->cq_overflow_list)) {
1070 cqe = io_get_cqring(ctx);
1074 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1076 list_move(&req->list, &list);
1078 WRITE_ONCE(cqe->user_data, req->user_data);
1079 WRITE_ONCE(cqe->res, req->result);
1080 WRITE_ONCE(cqe->flags, 0);
1082 WRITE_ONCE(ctx->rings->cq_overflow,
1083 atomic_inc_return(&ctx->cached_cq_overflow));
1087 io_commit_cqring(ctx);
1089 clear_bit(0, &ctx->sq_check_overflow);
1090 clear_bit(0, &ctx->cq_check_overflow);
1092 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1093 io_cqring_ev_posted(ctx);
1095 while (!list_empty(&list)) {
1096 req = list_first_entry(&list, struct io_kiocb, list);
1097 list_del(&req->list);
1104 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1106 struct io_ring_ctx *ctx = req->ctx;
1107 struct io_uring_cqe *cqe;
1109 trace_io_uring_complete(ctx, req->user_data, res);
1112 * If we can't get a cq entry, userspace overflowed the
1113 * submission (by quite a lot). Increment the overflow count in
1116 cqe = io_get_cqring(ctx);
1118 WRITE_ONCE(cqe->user_data, req->user_data);
1119 WRITE_ONCE(cqe->res, res);
1120 WRITE_ONCE(cqe->flags, 0);
1121 } else if (ctx->cq_overflow_flushed) {
1122 WRITE_ONCE(ctx->rings->cq_overflow,
1123 atomic_inc_return(&ctx->cached_cq_overflow));
1125 if (list_empty(&ctx->cq_overflow_list)) {
1126 set_bit(0, &ctx->sq_check_overflow);
1127 set_bit(0, &ctx->cq_check_overflow);
1129 refcount_inc(&req->refs);
1131 list_add_tail(&req->list, &ctx->cq_overflow_list);
1135 static void io_cqring_add_event(struct io_kiocb *req, long res)
1137 struct io_ring_ctx *ctx = req->ctx;
1138 unsigned long flags;
1140 spin_lock_irqsave(&ctx->completion_lock, flags);
1141 io_cqring_fill_event(req, res);
1142 io_commit_cqring(ctx);
1143 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1145 io_cqring_ev_posted(ctx);
1148 static inline bool io_is_fallback_req(struct io_kiocb *req)
1150 return req == (struct io_kiocb *)
1151 ((unsigned long) req->ctx->fallback_req & ~1UL);
1154 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1156 struct io_kiocb *req;
1158 req = ctx->fallback_req;
1159 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
1165 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
1166 struct io_submit_state *state)
1168 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1169 struct io_kiocb *req;
1172 req = kmem_cache_alloc(req_cachep, gfp);
1175 } else if (!state->free_reqs) {
1179 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1180 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1183 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1184 * retry single alloc to be on the safe side.
1186 if (unlikely(ret <= 0)) {
1187 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1188 if (!state->reqs[0])
1192 state->free_reqs = ret - 1;
1194 req = state->reqs[0];
1196 req = state->reqs[state->cur_req];
1206 /* one is dropped after submission, the other at completion */
1207 refcount_set(&req->refs, 2);
1209 INIT_IO_WORK(&req->work, io_wq_submit_work);
1212 req = io_get_fallback_req(ctx);
1215 percpu_ref_put(&ctx->refs);
1219 static void __io_req_do_free(struct io_kiocb *req)
1221 if (likely(!io_is_fallback_req(req)))
1222 kmem_cache_free(req_cachep, req);
1224 clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req);
1227 static void __io_req_aux_free(struct io_kiocb *req)
1229 struct io_ring_ctx *ctx = req->ctx;
1233 if (req->flags & REQ_F_FIXED_FILE)
1234 percpu_ref_put(&ctx->file_data->refs);
1239 io_req_work_drop_env(req);
1242 static void __io_free_req(struct io_kiocb *req)
1244 __io_req_aux_free(req);
1246 if (req->flags & REQ_F_INFLIGHT) {
1247 struct io_ring_ctx *ctx = req->ctx;
1248 unsigned long flags;
1250 spin_lock_irqsave(&ctx->inflight_lock, flags);
1251 list_del(&req->inflight_entry);
1252 if (waitqueue_active(&ctx->inflight_wait))
1253 wake_up(&ctx->inflight_wait);
1254 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1257 percpu_ref_put(&req->ctx->refs);
1258 __io_req_do_free(req);
1262 void *reqs[IO_IOPOLL_BATCH];
1267 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1269 int fixed_refs = rb->to_free;
1273 if (rb->need_iter) {
1274 int i, inflight = 0;
1275 unsigned long flags;
1278 for (i = 0; i < rb->to_free; i++) {
1279 struct io_kiocb *req = rb->reqs[i];
1281 if (req->flags & REQ_F_FIXED_FILE) {
1285 if (req->flags & REQ_F_INFLIGHT)
1287 __io_req_aux_free(req);
1292 spin_lock_irqsave(&ctx->inflight_lock, flags);
1293 for (i = 0; i < rb->to_free; i++) {
1294 struct io_kiocb *req = rb->reqs[i];
1296 if (req->flags & REQ_F_INFLIGHT) {
1297 list_del(&req->inflight_entry);
1302 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1304 if (waitqueue_active(&ctx->inflight_wait))
1305 wake_up(&ctx->inflight_wait);
1308 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1310 percpu_ref_put_many(&ctx->file_data->refs, fixed_refs);
1311 percpu_ref_put_many(&ctx->refs, rb->to_free);
1312 rb->to_free = rb->need_iter = 0;
1315 static bool io_link_cancel_timeout(struct io_kiocb *req)
1317 struct io_ring_ctx *ctx = req->ctx;
1320 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1322 io_cqring_fill_event(req, -ECANCELED);
1323 io_commit_cqring(ctx);
1324 req->flags &= ~REQ_F_LINK;
1332 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1334 struct io_ring_ctx *ctx = req->ctx;
1335 bool wake_ev = false;
1337 /* Already got next link */
1338 if (req->flags & REQ_F_LINK_NEXT)
1342 * The list should never be empty when we are called here. But could
1343 * potentially happen if the chain is messed up, check to be on the
1346 while (!list_empty(&req->link_list)) {
1347 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1348 struct io_kiocb, link_list);
1350 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1351 (nxt->flags & REQ_F_TIMEOUT))) {
1352 list_del_init(&nxt->link_list);
1353 wake_ev |= io_link_cancel_timeout(nxt);
1354 req->flags &= ~REQ_F_LINK_TIMEOUT;
1358 list_del_init(&req->link_list);
1359 if (!list_empty(&nxt->link_list))
1360 nxt->flags |= REQ_F_LINK;
1365 req->flags |= REQ_F_LINK_NEXT;
1367 io_cqring_ev_posted(ctx);
1371 * Called if REQ_F_LINK is set, and we fail the head request
1373 static void io_fail_links(struct io_kiocb *req)
1375 struct io_ring_ctx *ctx = req->ctx;
1376 unsigned long flags;
1378 spin_lock_irqsave(&ctx->completion_lock, flags);
1380 while (!list_empty(&req->link_list)) {
1381 struct io_kiocb *link = list_first_entry(&req->link_list,
1382 struct io_kiocb, link_list);
1384 list_del_init(&link->link_list);
1385 trace_io_uring_fail_link(req, link);
1387 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1388 link->opcode == IORING_OP_LINK_TIMEOUT) {
1389 io_link_cancel_timeout(link);
1391 io_cqring_fill_event(link, -ECANCELED);
1392 __io_double_put_req(link);
1394 req->flags &= ~REQ_F_LINK_TIMEOUT;
1397 io_commit_cqring(ctx);
1398 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1399 io_cqring_ev_posted(ctx);
1402 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1404 if (likely(!(req->flags & REQ_F_LINK)))
1408 * If LINK is set, we have dependent requests in this chain. If we
1409 * didn't fail this request, queue the first one up, moving any other
1410 * dependencies to the next request. In case of failure, fail the rest
1413 if (req->flags & REQ_F_FAIL_LINK) {
1415 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1416 REQ_F_LINK_TIMEOUT) {
1417 struct io_ring_ctx *ctx = req->ctx;
1418 unsigned long flags;
1421 * If this is a timeout link, we could be racing with the
1422 * timeout timer. Grab the completion lock for this case to
1423 * protect against that.
1425 spin_lock_irqsave(&ctx->completion_lock, flags);
1426 io_req_link_next(req, nxt);
1427 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1429 io_req_link_next(req, nxt);
1433 static void io_free_req(struct io_kiocb *req)
1435 struct io_kiocb *nxt = NULL;
1437 io_req_find_next(req, &nxt);
1441 io_queue_async_work(nxt);
1445 * Drop reference to request, return next in chain (if there is one) if this
1446 * was the last reference to this request.
1448 __attribute__((nonnull))
1449 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1451 io_req_find_next(req, nxtptr);
1453 if (refcount_dec_and_test(&req->refs))
1457 static void io_put_req(struct io_kiocb *req)
1459 if (refcount_dec_and_test(&req->refs))
1464 * Must only be used if we don't need to care about links, usually from
1465 * within the completion handling itself.
1467 static void __io_double_put_req(struct io_kiocb *req)
1469 /* drop both submit and complete references */
1470 if (refcount_sub_and_test(2, &req->refs))
1474 static void io_double_put_req(struct io_kiocb *req)
1476 /* drop both submit and complete references */
1477 if (refcount_sub_and_test(2, &req->refs))
1481 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1483 struct io_rings *rings = ctx->rings;
1485 if (test_bit(0, &ctx->cq_check_overflow)) {
1487 * noflush == true is from the waitqueue handler, just ensure
1488 * we wake up the task, and the next invocation will flush the
1489 * entries. We cannot safely to it from here.
1491 if (noflush && !list_empty(&ctx->cq_overflow_list))
1494 io_cqring_overflow_flush(ctx, false);
1497 /* See comment at the top of this file */
1499 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1502 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1504 struct io_rings *rings = ctx->rings;
1506 /* make sure SQ entry isn't read before tail */
1507 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1510 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1512 if ((req->flags & REQ_F_LINK) || io_is_fallback_req(req))
1515 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1518 rb->reqs[rb->to_free++] = req;
1519 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1520 io_free_req_many(req->ctx, rb);
1525 * Find and free completed poll iocbs
1527 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1528 struct list_head *done)
1530 struct req_batch rb;
1531 struct io_kiocb *req;
1533 rb.to_free = rb.need_iter = 0;
1534 while (!list_empty(done)) {
1535 req = list_first_entry(done, struct io_kiocb, list);
1536 list_del(&req->list);
1538 io_cqring_fill_event(req, req->result);
1541 if (refcount_dec_and_test(&req->refs) &&
1542 !io_req_multi_free(&rb, req))
1546 io_commit_cqring(ctx);
1547 io_free_req_many(ctx, &rb);
1550 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1553 struct io_kiocb *req, *tmp;
1559 * Only spin for completions if we don't have multiple devices hanging
1560 * off our complete list, and we're under the requested amount.
1562 spin = !ctx->poll_multi_file && *nr_events < min;
1565 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1566 struct kiocb *kiocb = &req->rw.kiocb;
1569 * Move completed entries to our local list. If we find a
1570 * request that requires polling, break out and complete
1571 * the done list first, if we have entries there.
1573 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1574 list_move_tail(&req->list, &done);
1577 if (!list_empty(&done))
1580 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1589 if (!list_empty(&done))
1590 io_iopoll_complete(ctx, nr_events, &done);
1596 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1597 * non-spinning poll check - we'll still enter the driver poll loop, but only
1598 * as a non-spinning completion check.
1600 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1603 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1606 ret = io_do_iopoll(ctx, nr_events, min);
1609 if (!min || *nr_events >= min)
1617 * We can't just wait for polled events to come to us, we have to actively
1618 * find and complete them.
1620 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1622 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1625 mutex_lock(&ctx->uring_lock);
1626 while (!list_empty(&ctx->poll_list)) {
1627 unsigned int nr_events = 0;
1629 io_iopoll_getevents(ctx, &nr_events, 1);
1632 * Ensure we allow local-to-the-cpu processing to take place,
1633 * in this case we need to ensure that we reap all events.
1637 mutex_unlock(&ctx->uring_lock);
1640 static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1643 int iters = 0, ret = 0;
1649 * Don't enter poll loop if we already have events pending.
1650 * If we do, we can potentially be spinning for commands that
1651 * already triggered a CQE (eg in error).
1653 if (io_cqring_events(ctx, false))
1657 * If a submit got punted to a workqueue, we can have the
1658 * application entering polling for a command before it gets
1659 * issued. That app will hold the uring_lock for the duration
1660 * of the poll right here, so we need to take a breather every
1661 * now and then to ensure that the issue has a chance to add
1662 * the poll to the issued list. Otherwise we can spin here
1663 * forever, while the workqueue is stuck trying to acquire the
1666 if (!(++iters & 7)) {
1667 mutex_unlock(&ctx->uring_lock);
1668 mutex_lock(&ctx->uring_lock);
1671 if (*nr_events < min)
1672 tmin = min - *nr_events;
1674 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1678 } while (min && !*nr_events && !need_resched());
1683 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1689 * We disallow the app entering submit/complete with polling, but we
1690 * still need to lock the ring to prevent racing with polled issue
1691 * that got punted to a workqueue.
1693 mutex_lock(&ctx->uring_lock);
1694 ret = __io_iopoll_check(ctx, nr_events, min);
1695 mutex_unlock(&ctx->uring_lock);
1699 static void kiocb_end_write(struct io_kiocb *req)
1702 * Tell lockdep we inherited freeze protection from submission
1705 if (req->flags & REQ_F_ISREG) {
1706 struct inode *inode = file_inode(req->file);
1708 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1710 file_end_write(req->file);
1713 static inline void req_set_fail_links(struct io_kiocb *req)
1715 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1716 req->flags |= REQ_F_FAIL_LINK;
1719 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1721 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1723 if (kiocb->ki_flags & IOCB_WRITE)
1724 kiocb_end_write(req);
1726 if (res != req->result)
1727 req_set_fail_links(req);
1728 io_cqring_add_event(req, res);
1731 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1733 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1735 io_complete_rw_common(kiocb, res);
1739 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1741 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1742 struct io_kiocb *nxt = NULL;
1744 io_complete_rw_common(kiocb, res);
1745 io_put_req_find_next(req, &nxt);
1750 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1752 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1754 if (kiocb->ki_flags & IOCB_WRITE)
1755 kiocb_end_write(req);
1757 if (res != req->result)
1758 req_set_fail_links(req);
1761 req->flags |= REQ_F_IOPOLL_COMPLETED;
1765 * After the iocb has been issued, it's safe to be found on the poll list.
1766 * Adding the kiocb to the list AFTER submission ensures that we don't
1767 * find it from a io_iopoll_getevents() thread before the issuer is done
1768 * accessing the kiocb cookie.
1770 static void io_iopoll_req_issued(struct io_kiocb *req)
1772 struct io_ring_ctx *ctx = req->ctx;
1775 * Track whether we have multiple files in our lists. This will impact
1776 * how we do polling eventually, not spinning if we're on potentially
1777 * different devices.
1779 if (list_empty(&ctx->poll_list)) {
1780 ctx->poll_multi_file = false;
1781 } else if (!ctx->poll_multi_file) {
1782 struct io_kiocb *list_req;
1784 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1786 if (list_req->file != req->file)
1787 ctx->poll_multi_file = true;
1791 * For fast devices, IO may have already completed. If it has, add
1792 * it to the front so we find it first.
1794 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1795 list_add(&req->list, &ctx->poll_list);
1797 list_add_tail(&req->list, &ctx->poll_list);
1800 static void io_file_put(struct io_submit_state *state)
1803 int diff = state->has_refs - state->used_refs;
1806 fput_many(state->file, diff);
1812 * Get as many references to a file as we have IOs left in this submission,
1813 * assuming most submissions are for one file, or at least that each file
1814 * has more than one submission.
1816 static struct file *io_file_get(struct io_submit_state *state, int fd)
1822 if (state->fd == fd) {
1829 state->file = fget_many(fd, state->ios_left);
1834 state->has_refs = state->ios_left;
1835 state->used_refs = 1;
1841 * If we tracked the file through the SCM inflight mechanism, we could support
1842 * any file. For now, just ensure that anything potentially problematic is done
1845 static bool io_file_supports_async(struct file *file)
1847 umode_t mode = file_inode(file)->i_mode;
1849 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
1851 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1857 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1858 bool force_nonblock)
1860 struct io_ring_ctx *ctx = req->ctx;
1861 struct kiocb *kiocb = &req->rw.kiocb;
1865 if (S_ISREG(file_inode(req->file)->i_mode))
1866 req->flags |= REQ_F_ISREG;
1868 kiocb->ki_pos = READ_ONCE(sqe->off);
1869 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
1870 req->flags |= REQ_F_CUR_POS;
1871 kiocb->ki_pos = req->file->f_pos;
1873 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1874 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1876 ioprio = READ_ONCE(sqe->ioprio);
1878 ret = ioprio_check_cap(ioprio);
1882 kiocb->ki_ioprio = ioprio;
1884 kiocb->ki_ioprio = get_current_ioprio();
1886 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1890 /* don't allow async punt if RWF_NOWAIT was requested */
1891 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1892 (req->file->f_flags & O_NONBLOCK))
1893 req->flags |= REQ_F_NOWAIT;
1896 kiocb->ki_flags |= IOCB_NOWAIT;
1898 if (ctx->flags & IORING_SETUP_IOPOLL) {
1899 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1900 !kiocb->ki_filp->f_op->iopoll)
1903 kiocb->ki_flags |= IOCB_HIPRI;
1904 kiocb->ki_complete = io_complete_rw_iopoll;
1907 if (kiocb->ki_flags & IOCB_HIPRI)
1909 kiocb->ki_complete = io_complete_rw;
1912 req->rw.addr = READ_ONCE(sqe->addr);
1913 req->rw.len = READ_ONCE(sqe->len);
1914 /* we own ->private, reuse it for the buffer index */
1915 req->rw.kiocb.private = (void *) (unsigned long)
1916 READ_ONCE(sqe->buf_index);
1920 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1926 case -ERESTARTNOINTR:
1927 case -ERESTARTNOHAND:
1928 case -ERESTART_RESTARTBLOCK:
1930 * We can't just restart the syscall, since previously
1931 * submitted sqes may already be in progress. Just fail this
1937 kiocb->ki_complete(kiocb, ret, 0);
1941 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1944 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1946 if (req->flags & REQ_F_CUR_POS)
1947 req->file->f_pos = kiocb->ki_pos;
1948 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1949 *nxt = __io_complete_rw(kiocb, ret);
1951 io_rw_done(kiocb, ret);
1954 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
1955 struct iov_iter *iter)
1957 struct io_ring_ctx *ctx = req->ctx;
1958 size_t len = req->rw.len;
1959 struct io_mapped_ubuf *imu;
1960 unsigned index, buf_index;
1964 /* attempt to use fixed buffers without having provided iovecs */
1965 if (unlikely(!ctx->user_bufs))
1968 buf_index = (unsigned long) req->rw.kiocb.private;
1969 if (unlikely(buf_index >= ctx->nr_user_bufs))
1972 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1973 imu = &ctx->user_bufs[index];
1974 buf_addr = req->rw.addr;
1977 if (buf_addr + len < buf_addr)
1979 /* not inside the mapped region */
1980 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1984 * May not be a start of buffer, set size appropriately
1985 * and advance us to the beginning.
1987 offset = buf_addr - imu->ubuf;
1988 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1992 * Don't use iov_iter_advance() here, as it's really slow for
1993 * using the latter parts of a big fixed buffer - it iterates
1994 * over each segment manually. We can cheat a bit here, because
1997 * 1) it's a BVEC iter, we set it up
1998 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1999 * first and last bvec
2001 * So just find our index, and adjust the iterator afterwards.
2002 * If the offset is within the first bvec (or the whole first
2003 * bvec, just use iov_iter_advance(). This makes it easier
2004 * since we can just skip the first segment, which may not
2005 * be PAGE_SIZE aligned.
2007 const struct bio_vec *bvec = imu->bvec;
2009 if (offset <= bvec->bv_len) {
2010 iov_iter_advance(iter, offset);
2012 unsigned long seg_skip;
2014 /* skip first vec */
2015 offset -= bvec->bv_len;
2016 seg_skip = 1 + (offset >> PAGE_SHIFT);
2018 iter->bvec = bvec + seg_skip;
2019 iter->nr_segs -= seg_skip;
2020 iter->count -= bvec->bv_len + offset;
2021 iter->iov_offset = offset & ~PAGE_MASK;
2028 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2029 struct iovec **iovec, struct iov_iter *iter)
2031 void __user *buf = u64_to_user_ptr(req->rw.addr);
2032 size_t sqe_len = req->rw.len;
2035 opcode = req->opcode;
2036 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2038 return io_import_fixed(req, rw, iter);
2041 /* buffer index only valid with fixed read/write */
2042 if (req->rw.kiocb.private)
2045 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2047 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2053 struct io_async_rw *iorw = &req->io->rw;
2056 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2057 if (iorw->iov == iorw->fast_iov)
2065 #ifdef CONFIG_COMPAT
2066 if (req->ctx->compat)
2067 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2071 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2075 * For files that don't have ->read_iter() and ->write_iter(), handle them
2076 * by looping over ->read() or ->write() manually.
2078 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2079 struct iov_iter *iter)
2084 * Don't support polled IO through this interface, and we can't
2085 * support non-blocking either. For the latter, this just causes
2086 * the kiocb to be handled from an async context.
2088 if (kiocb->ki_flags & IOCB_HIPRI)
2090 if (kiocb->ki_flags & IOCB_NOWAIT)
2093 while (iov_iter_count(iter)) {
2097 if (!iov_iter_is_bvec(iter)) {
2098 iovec = iov_iter_iovec(iter);
2100 /* fixed buffers import bvec */
2101 iovec.iov_base = kmap(iter->bvec->bv_page)
2103 iovec.iov_len = min(iter->count,
2104 iter->bvec->bv_len - iter->iov_offset);
2108 nr = file->f_op->read(file, iovec.iov_base,
2109 iovec.iov_len, &kiocb->ki_pos);
2111 nr = file->f_op->write(file, iovec.iov_base,
2112 iovec.iov_len, &kiocb->ki_pos);
2115 if (iov_iter_is_bvec(iter))
2116 kunmap(iter->bvec->bv_page);
2124 if (nr != iovec.iov_len)
2126 iov_iter_advance(iter, nr);
2132 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2133 struct iovec *iovec, struct iovec *fast_iov,
2134 struct iov_iter *iter)
2136 req->io->rw.nr_segs = iter->nr_segs;
2137 req->io->rw.size = io_size;
2138 req->io->rw.iov = iovec;
2139 if (!req->io->rw.iov) {
2140 req->io->rw.iov = req->io->rw.fast_iov;
2141 memcpy(req->io->rw.iov, fast_iov,
2142 sizeof(struct iovec) * iter->nr_segs);
2146 static int io_alloc_async_ctx(struct io_kiocb *req)
2148 if (!io_op_defs[req->opcode].async_ctx)
2150 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2151 return req->io == NULL;
2154 static void io_rw_async(struct io_wq_work **workptr)
2156 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2157 struct iovec *iov = NULL;
2159 if (req->io->rw.iov != req->io->rw.fast_iov)
2160 iov = req->io->rw.iov;
2161 io_wq_submit_work(workptr);
2165 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2166 struct iovec *iovec, struct iovec *fast_iov,
2167 struct iov_iter *iter)
2169 if (!io_op_defs[req->opcode].async_ctx)
2172 if (io_alloc_async_ctx(req))
2175 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2177 req->work.func = io_rw_async;
2181 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2182 bool force_nonblock)
2184 struct io_async_ctx *io;
2185 struct iov_iter iter;
2188 ret = io_prep_rw(req, sqe, force_nonblock);
2192 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2199 io->rw.iov = io->rw.fast_iov;
2201 ret = io_import_iovec(READ, req, &io->rw.iov, &iter);
2206 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2210 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
2211 bool force_nonblock)
2213 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2214 struct kiocb *kiocb = &req->rw.kiocb;
2215 struct iov_iter iter;
2217 ssize_t io_size, ret;
2219 ret = io_import_iovec(READ, req, &iovec, &iter);
2223 /* Ensure we clear previously set non-block flag */
2224 if (!force_nonblock)
2225 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2229 if (req->flags & REQ_F_LINK)
2230 req->result = io_size;
2233 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2234 * we know to async punt it even if it was opened O_NONBLOCK
2236 if (force_nonblock && !io_file_supports_async(req->file)) {
2237 req->flags |= REQ_F_MUST_PUNT;
2241 iov_count = iov_iter_count(&iter);
2242 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2246 if (req->file->f_op->read_iter)
2247 ret2 = call_read_iter(req->file, kiocb, &iter);
2249 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2251 /* Catch -EAGAIN return for forced non-blocking submission */
2252 if (!force_nonblock || ret2 != -EAGAIN) {
2253 kiocb_done(kiocb, ret2, nxt, req->in_async);
2256 ret = io_setup_async_rw(req, io_size, iovec,
2257 inline_vecs, &iter);
2264 if (!io_wq_current_is_worker())
2269 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2270 bool force_nonblock)
2272 struct io_async_ctx *io;
2273 struct iov_iter iter;
2276 ret = io_prep_rw(req, sqe, force_nonblock);
2280 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2287 io->rw.iov = io->rw.fast_iov;
2289 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter);
2294 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2298 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
2299 bool force_nonblock)
2301 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2302 struct kiocb *kiocb = &req->rw.kiocb;
2303 struct iov_iter iter;
2305 ssize_t ret, io_size;
2307 ret = io_import_iovec(WRITE, req, &iovec, &iter);
2311 /* Ensure we clear previously set non-block flag */
2312 if (!force_nonblock)
2313 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2317 if (req->flags & REQ_F_LINK)
2318 req->result = io_size;
2321 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2322 * we know to async punt it even if it was opened O_NONBLOCK
2324 if (force_nonblock && !io_file_supports_async(req->file)) {
2325 req->flags |= REQ_F_MUST_PUNT;
2329 /* file path doesn't support NOWAIT for non-direct_IO */
2330 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2331 (req->flags & REQ_F_ISREG))
2334 iov_count = iov_iter_count(&iter);
2335 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2340 * Open-code file_start_write here to grab freeze protection,
2341 * which will be released by another thread in
2342 * io_complete_rw(). Fool lockdep by telling it the lock got
2343 * released so that it doesn't complain about the held lock when
2344 * we return to userspace.
2346 if (req->flags & REQ_F_ISREG) {
2347 __sb_start_write(file_inode(req->file)->i_sb,
2348 SB_FREEZE_WRITE, true);
2349 __sb_writers_release(file_inode(req->file)->i_sb,
2352 kiocb->ki_flags |= IOCB_WRITE;
2354 if (req->file->f_op->write_iter)
2355 ret2 = call_write_iter(req->file, kiocb, &iter);
2357 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2358 if (!force_nonblock || ret2 != -EAGAIN) {
2359 kiocb_done(kiocb, ret2, nxt, req->in_async);
2362 ret = io_setup_async_rw(req, io_size, iovec,
2363 inline_vecs, &iter);
2370 if (!io_wq_current_is_worker())
2376 * IORING_OP_NOP just posts a completion event, nothing else.
2378 static int io_nop(struct io_kiocb *req)
2380 struct io_ring_ctx *ctx = req->ctx;
2382 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2385 io_cqring_add_event(req, 0);
2390 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2392 struct io_ring_ctx *ctx = req->ctx;
2397 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2399 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2402 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2403 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2406 req->sync.off = READ_ONCE(sqe->off);
2407 req->sync.len = READ_ONCE(sqe->len);
2411 static bool io_req_cancelled(struct io_kiocb *req)
2413 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2414 req_set_fail_links(req);
2415 io_cqring_add_event(req, -ECANCELED);
2423 static void io_link_work_cb(struct io_wq_work **workptr)
2425 struct io_wq_work *work = *workptr;
2426 struct io_kiocb *link = work->data;
2428 io_queue_linked_timeout(link);
2429 work->func = io_wq_submit_work;
2432 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
2434 struct io_kiocb *link;
2436 io_prep_async_work(nxt, &link);
2437 *workptr = &nxt->work;
2439 nxt->work.flags |= IO_WQ_WORK_CB;
2440 nxt->work.func = io_link_work_cb;
2441 nxt->work.data = link;
2445 static void io_fsync_finish(struct io_wq_work **workptr)
2447 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2448 loff_t end = req->sync.off + req->sync.len;
2449 struct io_kiocb *nxt = NULL;
2452 if (io_req_cancelled(req))
2455 ret = vfs_fsync_range(req->file, req->sync.off,
2456 end > 0 ? end : LLONG_MAX,
2457 req->sync.flags & IORING_FSYNC_DATASYNC);
2459 req_set_fail_links(req);
2460 io_cqring_add_event(req, ret);
2461 io_put_req_find_next(req, &nxt);
2463 io_wq_assign_next(workptr, nxt);
2466 static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt,
2467 bool force_nonblock)
2469 struct io_wq_work *work, *old_work;
2471 /* fsync always requires a blocking context */
2472 if (force_nonblock) {
2474 req->work.func = io_fsync_finish;
2478 work = old_work = &req->work;
2479 io_fsync_finish(&work);
2480 if (work && work != old_work)
2481 *nxt = container_of(work, struct io_kiocb, work);
2485 static void io_fallocate_finish(struct io_wq_work **workptr)
2487 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2488 struct io_kiocb *nxt = NULL;
2491 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2494 req_set_fail_links(req);
2495 io_cqring_add_event(req, ret);
2496 io_put_req_find_next(req, &nxt);
2498 io_wq_assign_next(workptr, nxt);
2501 static int io_fallocate_prep(struct io_kiocb *req,
2502 const struct io_uring_sqe *sqe)
2504 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2507 req->sync.off = READ_ONCE(sqe->off);
2508 req->sync.len = READ_ONCE(sqe->addr);
2509 req->sync.mode = READ_ONCE(sqe->len);
2513 static int io_fallocate(struct io_kiocb *req, struct io_kiocb **nxt,
2514 bool force_nonblock)
2516 struct io_wq_work *work, *old_work;
2518 /* fallocate always requiring blocking context */
2519 if (force_nonblock) {
2521 req->work.func = io_fallocate_finish;
2525 work = old_work = &req->work;
2526 io_fallocate_finish(&work);
2527 if (work && work != old_work)
2528 *nxt = container_of(work, struct io_kiocb, work);
2533 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2535 const char __user *fname;
2538 if (sqe->ioprio || sqe->buf_index)
2541 req->open.dfd = READ_ONCE(sqe->fd);
2542 req->open.how.mode = READ_ONCE(sqe->len);
2543 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2544 req->open.how.flags = READ_ONCE(sqe->open_flags);
2546 req->open.filename = getname(fname);
2547 if (IS_ERR(req->open.filename)) {
2548 ret = PTR_ERR(req->open.filename);
2549 req->open.filename = NULL;
2556 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2558 struct open_how __user *how;
2559 const char __user *fname;
2563 if (sqe->ioprio || sqe->buf_index)
2566 req->open.dfd = READ_ONCE(sqe->fd);
2567 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2568 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2569 len = READ_ONCE(sqe->len);
2571 if (len < OPEN_HOW_SIZE_VER0)
2574 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2579 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2580 req->open.how.flags |= O_LARGEFILE;
2582 req->open.filename = getname(fname);
2583 if (IS_ERR(req->open.filename)) {
2584 ret = PTR_ERR(req->open.filename);
2585 req->open.filename = NULL;
2592 static int io_openat2(struct io_kiocb *req, struct io_kiocb **nxt,
2593 bool force_nonblock)
2595 struct open_flags op;
2602 ret = build_open_flags(&req->open.how, &op);
2606 ret = get_unused_fd_flags(req->open.how.flags);
2610 file = do_filp_open(req->open.dfd, req->open.filename, &op);
2613 ret = PTR_ERR(file);
2615 fsnotify_open(file);
2616 fd_install(ret, file);
2619 putname(req->open.filename);
2621 req_set_fail_links(req);
2622 io_cqring_add_event(req, ret);
2623 io_put_req_find_next(req, nxt);
2627 static int io_openat(struct io_kiocb *req, struct io_kiocb **nxt,
2628 bool force_nonblock)
2630 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
2631 return io_openat2(req, nxt, force_nonblock);
2634 static int io_epoll_ctl_prep(struct io_kiocb *req,
2635 const struct io_uring_sqe *sqe)
2637 #if defined(CONFIG_EPOLL)
2638 if (sqe->ioprio || sqe->buf_index)
2641 req->epoll.epfd = READ_ONCE(sqe->fd);
2642 req->epoll.op = READ_ONCE(sqe->len);
2643 req->epoll.fd = READ_ONCE(sqe->off);
2645 if (ep_op_has_event(req->epoll.op)) {
2646 struct epoll_event __user *ev;
2648 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
2649 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
2659 static int io_epoll_ctl(struct io_kiocb *req, struct io_kiocb **nxt,
2660 bool force_nonblock)
2662 #if defined(CONFIG_EPOLL)
2663 struct io_epoll *ie = &req->epoll;
2666 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
2667 if (force_nonblock && ret == -EAGAIN)
2671 req_set_fail_links(req);
2672 io_cqring_add_event(req, ret);
2673 io_put_req_find_next(req, nxt);
2680 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2682 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2683 if (sqe->ioprio || sqe->buf_index || sqe->off)
2686 req->madvise.addr = READ_ONCE(sqe->addr);
2687 req->madvise.len = READ_ONCE(sqe->len);
2688 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
2695 static int io_madvise(struct io_kiocb *req, struct io_kiocb **nxt,
2696 bool force_nonblock)
2698 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2699 struct io_madvise *ma = &req->madvise;
2705 ret = do_madvise(ma->addr, ma->len, ma->advice);
2707 req_set_fail_links(req);
2708 io_cqring_add_event(req, ret);
2709 io_put_req_find_next(req, nxt);
2716 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2718 if (sqe->ioprio || sqe->buf_index || sqe->addr)
2721 req->fadvise.offset = READ_ONCE(sqe->off);
2722 req->fadvise.len = READ_ONCE(sqe->len);
2723 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
2727 static int io_fadvise(struct io_kiocb *req, struct io_kiocb **nxt,
2728 bool force_nonblock)
2730 struct io_fadvise *fa = &req->fadvise;
2733 /* DONTNEED may block, others _should_ not */
2734 if (fa->advice == POSIX_FADV_DONTNEED && force_nonblock)
2737 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
2739 req_set_fail_links(req);
2740 io_cqring_add_event(req, ret);
2741 io_put_req_find_next(req, nxt);
2745 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2747 const char __user *fname;
2748 unsigned lookup_flags;
2751 if (sqe->ioprio || sqe->buf_index)
2754 req->open.dfd = READ_ONCE(sqe->fd);
2755 req->open.mask = READ_ONCE(sqe->len);
2756 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2757 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2758 req->open.how.flags = READ_ONCE(sqe->statx_flags);
2760 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
2763 req->open.filename = getname_flags(fname, lookup_flags, NULL);
2764 if (IS_ERR(req->open.filename)) {
2765 ret = PTR_ERR(req->open.filename);
2766 req->open.filename = NULL;
2773 static int io_statx(struct io_kiocb *req, struct io_kiocb **nxt,
2774 bool force_nonblock)
2776 struct io_open *ctx = &req->open;
2777 unsigned lookup_flags;
2785 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
2789 /* filename_lookup() drops it, keep a reference */
2790 ctx->filename->refcnt++;
2792 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
2797 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
2799 if (retry_estale(ret, lookup_flags)) {
2800 lookup_flags |= LOOKUP_REVAL;
2804 ret = cp_statx(&stat, ctx->buffer);
2806 putname(ctx->filename);
2808 req_set_fail_links(req);
2809 io_cqring_add_event(req, ret);
2810 io_put_req_find_next(req, nxt);
2814 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2817 * If we queue this for async, it must not be cancellable. That would
2818 * leave the 'file' in an undeterminate state.
2820 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
2822 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
2823 sqe->rw_flags || sqe->buf_index)
2825 if (sqe->flags & IOSQE_FIXED_FILE)
2828 req->close.fd = READ_ONCE(sqe->fd);
2829 if (req->file->f_op == &io_uring_fops ||
2830 req->close.fd == req->ctx->ring_fd)
2836 static void io_close_finish(struct io_wq_work **workptr)
2838 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2839 struct io_kiocb *nxt = NULL;
2841 /* Invoked with files, we need to do the close */
2842 if (req->work.files) {
2845 ret = filp_close(req->close.put_file, req->work.files);
2847 req_set_fail_links(req);
2849 io_cqring_add_event(req, ret);
2852 fput(req->close.put_file);
2854 /* we bypassed the re-issue, drop the submission reference */
2856 io_put_req_find_next(req, &nxt);
2858 io_wq_assign_next(workptr, nxt);
2861 static int io_close(struct io_kiocb *req, struct io_kiocb **nxt,
2862 bool force_nonblock)
2866 req->close.put_file = NULL;
2867 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
2871 /* if the file has a flush method, be safe and punt to async */
2872 if (req->close.put_file->f_op->flush && !io_wq_current_is_worker())
2876 * No ->flush(), safely close from here and just punt the
2877 * fput() to async context.
2879 ret = filp_close(req->close.put_file, current->files);
2882 req_set_fail_links(req);
2883 io_cqring_add_event(req, ret);
2885 if (io_wq_current_is_worker()) {
2886 struct io_wq_work *old_work, *work;
2888 old_work = work = &req->work;
2889 io_close_finish(&work);
2890 if (work && work != old_work)
2891 *nxt = container_of(work, struct io_kiocb, work);
2896 req->work.func = io_close_finish;
2900 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2902 struct io_ring_ctx *ctx = req->ctx;
2907 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2909 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2912 req->sync.off = READ_ONCE(sqe->off);
2913 req->sync.len = READ_ONCE(sqe->len);
2914 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
2918 static void io_sync_file_range_finish(struct io_wq_work **workptr)
2920 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2921 struct io_kiocb *nxt = NULL;
2924 if (io_req_cancelled(req))
2927 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
2930 req_set_fail_links(req);
2931 io_cqring_add_event(req, ret);
2932 io_put_req_find_next(req, &nxt);
2934 io_wq_assign_next(workptr, nxt);
2937 static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt,
2938 bool force_nonblock)
2940 struct io_wq_work *work, *old_work;
2942 /* sync_file_range always requires a blocking context */
2943 if (force_nonblock) {
2945 req->work.func = io_sync_file_range_finish;
2949 work = old_work = &req->work;
2950 io_sync_file_range_finish(&work);
2951 if (work && work != old_work)
2952 *nxt = container_of(work, struct io_kiocb, work);
2956 #if defined(CONFIG_NET)
2957 static void io_sendrecv_async(struct io_wq_work **workptr)
2959 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2960 struct iovec *iov = NULL;
2962 if (req->io->rw.iov != req->io->rw.fast_iov)
2963 iov = req->io->msg.iov;
2964 io_wq_submit_work(workptr);
2969 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2971 #if defined(CONFIG_NET)
2972 struct io_sr_msg *sr = &req->sr_msg;
2973 struct io_async_ctx *io = req->io;
2975 sr->msg_flags = READ_ONCE(sqe->msg_flags);
2976 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
2977 sr->len = READ_ONCE(sqe->len);
2979 if (!io || req->opcode == IORING_OP_SEND)
2982 io->msg.iov = io->msg.fast_iov;
2983 return sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
2990 static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt,
2991 bool force_nonblock)
2993 #if defined(CONFIG_NET)
2994 struct io_async_msghdr *kmsg = NULL;
2995 struct socket *sock;
2998 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3001 sock = sock_from_file(req->file, &ret);
3003 struct io_async_ctx io;
3004 struct sockaddr_storage addr;
3008 kmsg = &req->io->msg;
3009 kmsg->msg.msg_name = &addr;
3010 /* if iov is set, it's allocated already */
3012 kmsg->iov = kmsg->fast_iov;
3013 kmsg->msg.msg_iter.iov = kmsg->iov;
3015 struct io_sr_msg *sr = &req->sr_msg;
3018 kmsg->msg.msg_name = &addr;
3020 io.msg.iov = io.msg.fast_iov;
3021 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3022 sr->msg_flags, &io.msg.iov);
3027 flags = req->sr_msg.msg_flags;
3028 if (flags & MSG_DONTWAIT)
3029 req->flags |= REQ_F_NOWAIT;
3030 else if (force_nonblock)
3031 flags |= MSG_DONTWAIT;
3033 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3034 if (force_nonblock && ret == -EAGAIN) {
3037 if (io_alloc_async_ctx(req))
3039 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3040 req->work.func = io_sendrecv_async;
3043 if (ret == -ERESTARTSYS)
3047 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
3049 io_cqring_add_event(req, ret);
3051 req_set_fail_links(req);
3052 io_put_req_find_next(req, nxt);
3059 static int io_send(struct io_kiocb *req, struct io_kiocb **nxt,
3060 bool force_nonblock)
3062 #if defined(CONFIG_NET)
3063 struct socket *sock;
3066 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3069 sock = sock_from_file(req->file, &ret);
3071 struct io_sr_msg *sr = &req->sr_msg;
3076 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3081 msg.msg_name = NULL;
3082 msg.msg_control = NULL;
3083 msg.msg_controllen = 0;
3084 msg.msg_namelen = 0;
3086 flags = req->sr_msg.msg_flags;
3087 if (flags & MSG_DONTWAIT)
3088 req->flags |= REQ_F_NOWAIT;
3089 else if (force_nonblock)
3090 flags |= MSG_DONTWAIT;
3092 msg.msg_flags = flags;
3093 ret = sock_sendmsg(sock, &msg);
3094 if (force_nonblock && ret == -EAGAIN)
3096 if (ret == -ERESTARTSYS)
3100 io_cqring_add_event(req, ret);
3102 req_set_fail_links(req);
3103 io_put_req_find_next(req, nxt);
3110 static int io_recvmsg_prep(struct io_kiocb *req,
3111 const struct io_uring_sqe *sqe)
3113 #if defined(CONFIG_NET)
3114 struct io_sr_msg *sr = &req->sr_msg;
3115 struct io_async_ctx *io = req->io;
3117 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3118 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3119 sr->len = READ_ONCE(sqe->len);
3121 if (!io || req->opcode == IORING_OP_RECV)
3124 io->msg.iov = io->msg.fast_iov;
3125 return recvmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3126 &io->msg.uaddr, &io->msg.iov);
3132 static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt,
3133 bool force_nonblock)
3135 #if defined(CONFIG_NET)
3136 struct io_async_msghdr *kmsg = NULL;
3137 struct socket *sock;
3140 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3143 sock = sock_from_file(req->file, &ret);
3145 struct io_async_ctx io;
3146 struct sockaddr_storage addr;
3150 kmsg = &req->io->msg;
3151 kmsg->msg.msg_name = &addr;
3152 /* if iov is set, it's allocated already */
3154 kmsg->iov = kmsg->fast_iov;
3155 kmsg->msg.msg_iter.iov = kmsg->iov;
3157 struct io_sr_msg *sr = &req->sr_msg;
3160 kmsg->msg.msg_name = &addr;
3162 io.msg.iov = io.msg.fast_iov;
3163 ret = recvmsg_copy_msghdr(&io.msg.msg, sr->msg,
3164 sr->msg_flags, &io.msg.uaddr,
3170 flags = req->sr_msg.msg_flags;
3171 if (flags & MSG_DONTWAIT)
3172 req->flags |= REQ_F_NOWAIT;
3173 else if (force_nonblock)
3174 flags |= MSG_DONTWAIT;
3176 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3177 kmsg->uaddr, flags);
3178 if (force_nonblock && ret == -EAGAIN) {
3181 if (io_alloc_async_ctx(req))
3183 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3184 req->work.func = io_sendrecv_async;
3187 if (ret == -ERESTARTSYS)
3191 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
3193 io_cqring_add_event(req, ret);
3195 req_set_fail_links(req);
3196 io_put_req_find_next(req, nxt);
3203 static int io_recv(struct io_kiocb *req, struct io_kiocb **nxt,
3204 bool force_nonblock)
3206 #if defined(CONFIG_NET)
3207 struct socket *sock;
3210 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3213 sock = sock_from_file(req->file, &ret);
3215 struct io_sr_msg *sr = &req->sr_msg;
3220 ret = import_single_range(READ, sr->buf, sr->len, &iov,
3225 msg.msg_name = NULL;
3226 msg.msg_control = NULL;
3227 msg.msg_controllen = 0;
3228 msg.msg_namelen = 0;
3229 msg.msg_iocb = NULL;
3232 flags = req->sr_msg.msg_flags;
3233 if (flags & MSG_DONTWAIT)
3234 req->flags |= REQ_F_NOWAIT;
3235 else if (force_nonblock)
3236 flags |= MSG_DONTWAIT;
3238 ret = sock_recvmsg(sock, &msg, flags);
3239 if (force_nonblock && ret == -EAGAIN)
3241 if (ret == -ERESTARTSYS)
3245 io_cqring_add_event(req, ret);
3247 req_set_fail_links(req);
3248 io_put_req_find_next(req, nxt);
3256 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3258 #if defined(CONFIG_NET)
3259 struct io_accept *accept = &req->accept;
3261 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3263 if (sqe->ioprio || sqe->len || sqe->buf_index)
3266 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3267 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3268 accept->flags = READ_ONCE(sqe->accept_flags);
3275 #if defined(CONFIG_NET)
3276 static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3277 bool force_nonblock)
3279 struct io_accept *accept = &req->accept;
3280 unsigned file_flags;
3283 file_flags = force_nonblock ? O_NONBLOCK : 0;
3284 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3285 accept->addr_len, accept->flags);
3286 if (ret == -EAGAIN && force_nonblock)
3288 if (ret == -ERESTARTSYS)
3291 req_set_fail_links(req);
3292 io_cqring_add_event(req, ret);
3293 io_put_req_find_next(req, nxt);
3297 static void io_accept_finish(struct io_wq_work **workptr)
3299 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3300 struct io_kiocb *nxt = NULL;
3302 if (io_req_cancelled(req))
3304 __io_accept(req, &nxt, false);
3306 io_wq_assign_next(workptr, nxt);
3310 static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3311 bool force_nonblock)
3313 #if defined(CONFIG_NET)
3316 ret = __io_accept(req, nxt, force_nonblock);
3317 if (ret == -EAGAIN && force_nonblock) {
3318 req->work.func = io_accept_finish;
3328 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3330 #if defined(CONFIG_NET)
3331 struct io_connect *conn = &req->connect;
3332 struct io_async_ctx *io = req->io;
3334 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3336 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3339 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3340 conn->addr_len = READ_ONCE(sqe->addr2);
3345 return move_addr_to_kernel(conn->addr, conn->addr_len,
3346 &io->connect.address);
3352 static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt,
3353 bool force_nonblock)
3355 #if defined(CONFIG_NET)
3356 struct io_async_ctx __io, *io;
3357 unsigned file_flags;
3363 ret = move_addr_to_kernel(req->connect.addr,
3364 req->connect.addr_len,
3365 &__io.connect.address);
3371 file_flags = force_nonblock ? O_NONBLOCK : 0;
3373 ret = __sys_connect_file(req->file, &io->connect.address,
3374 req->connect.addr_len, file_flags);
3375 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
3378 if (io_alloc_async_ctx(req)) {
3382 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
3385 if (ret == -ERESTARTSYS)
3389 req_set_fail_links(req);
3390 io_cqring_add_event(req, ret);
3391 io_put_req_find_next(req, nxt);
3398 static void io_poll_remove_one(struct io_kiocb *req)
3400 struct io_poll_iocb *poll = &req->poll;
3402 spin_lock(&poll->head->lock);
3403 WRITE_ONCE(poll->canceled, true);
3404 if (!list_empty(&poll->wait.entry)) {
3405 list_del_init(&poll->wait.entry);
3406 io_queue_async_work(req);
3408 spin_unlock(&poll->head->lock);
3409 hash_del(&req->hash_node);
3412 static void io_poll_remove_all(struct io_ring_ctx *ctx)
3414 struct hlist_node *tmp;
3415 struct io_kiocb *req;
3418 spin_lock_irq(&ctx->completion_lock);
3419 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
3420 struct hlist_head *list;
3422 list = &ctx->cancel_hash[i];
3423 hlist_for_each_entry_safe(req, tmp, list, hash_node)
3424 io_poll_remove_one(req);
3426 spin_unlock_irq(&ctx->completion_lock);
3429 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
3431 struct hlist_head *list;
3432 struct io_kiocb *req;
3434 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
3435 hlist_for_each_entry(req, list, hash_node) {
3436 if (sqe_addr == req->user_data) {
3437 io_poll_remove_one(req);
3445 static int io_poll_remove_prep(struct io_kiocb *req,
3446 const struct io_uring_sqe *sqe)
3448 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3450 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3454 req->poll.addr = READ_ONCE(sqe->addr);
3459 * Find a running poll command that matches one specified in sqe->addr,
3460 * and remove it if found.
3462 static int io_poll_remove(struct io_kiocb *req)
3464 struct io_ring_ctx *ctx = req->ctx;
3468 addr = req->poll.addr;
3469 spin_lock_irq(&ctx->completion_lock);
3470 ret = io_poll_cancel(ctx, addr);
3471 spin_unlock_irq(&ctx->completion_lock);
3473 io_cqring_add_event(req, ret);
3475 req_set_fail_links(req);
3480 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
3482 struct io_ring_ctx *ctx = req->ctx;
3484 req->poll.done = true;
3486 io_cqring_fill_event(req, error);
3488 io_cqring_fill_event(req, mangle_poll(mask));
3489 io_commit_cqring(ctx);
3492 static void io_poll_complete_work(struct io_wq_work **workptr)
3494 struct io_wq_work *work = *workptr;
3495 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3496 struct io_poll_iocb *poll = &req->poll;
3497 struct poll_table_struct pt = { ._key = poll->events };
3498 struct io_ring_ctx *ctx = req->ctx;
3499 struct io_kiocb *nxt = NULL;
3503 if (work->flags & IO_WQ_WORK_CANCEL) {
3504 WRITE_ONCE(poll->canceled, true);
3506 } else if (READ_ONCE(poll->canceled)) {
3510 if (ret != -ECANCELED)
3511 mask = vfs_poll(poll->file, &pt) & poll->events;
3514 * Note that ->ki_cancel callers also delete iocb from active_reqs after
3515 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
3516 * synchronize with them. In the cancellation case the list_del_init
3517 * itself is not actually needed, but harmless so we keep it in to
3518 * avoid further branches in the fast path.
3520 spin_lock_irq(&ctx->completion_lock);
3521 if (!mask && ret != -ECANCELED) {
3522 add_wait_queue(poll->head, &poll->wait);
3523 spin_unlock_irq(&ctx->completion_lock);
3526 hash_del(&req->hash_node);
3527 io_poll_complete(req, mask, ret);
3528 spin_unlock_irq(&ctx->completion_lock);
3530 io_cqring_ev_posted(ctx);
3533 req_set_fail_links(req);
3534 io_put_req_find_next(req, &nxt);
3536 io_wq_assign_next(workptr, nxt);
3539 static void __io_poll_flush(struct io_ring_ctx *ctx, struct llist_node *nodes)
3541 struct io_kiocb *req, *tmp;
3542 struct req_batch rb;
3544 rb.to_free = rb.need_iter = 0;
3545 spin_lock_irq(&ctx->completion_lock);
3546 llist_for_each_entry_safe(req, tmp, nodes, llist_node) {
3547 hash_del(&req->hash_node);
3548 io_poll_complete(req, req->result, 0);
3550 if (refcount_dec_and_test(&req->refs) &&
3551 !io_req_multi_free(&rb, req)) {
3552 req->flags |= REQ_F_COMP_LOCKED;
3556 spin_unlock_irq(&ctx->completion_lock);
3558 io_cqring_ev_posted(ctx);
3559 io_free_req_many(ctx, &rb);
3562 static void io_poll_flush(struct io_wq_work **workptr)
3564 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3565 struct llist_node *nodes;
3567 nodes = llist_del_all(&req->ctx->poll_llist);
3569 __io_poll_flush(req->ctx, nodes);
3572 static void io_poll_trigger_evfd(struct io_wq_work **workptr)
3574 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3576 eventfd_signal(req->ctx->cq_ev_fd, 1);
3580 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
3583 struct io_poll_iocb *poll = wait->private;
3584 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
3585 struct io_ring_ctx *ctx = req->ctx;
3586 __poll_t mask = key_to_poll(key);
3588 /* for instances that support it check for an event match first: */
3589 if (mask && !(mask & poll->events))
3592 list_del_init(&poll->wait.entry);
3595 * Run completion inline if we can. We're using trylock here because
3596 * we are violating the completion_lock -> poll wq lock ordering.
3597 * If we have a link timeout we're going to need the completion_lock
3598 * for finalizing the request, mark us as having grabbed that already.
3601 unsigned long flags;
3603 if (llist_empty(&ctx->poll_llist) &&
3604 spin_trylock_irqsave(&ctx->completion_lock, flags)) {
3607 hash_del(&req->hash_node);
3608 io_poll_complete(req, mask, 0);
3610 trigger_ev = io_should_trigger_evfd(ctx);
3611 if (trigger_ev && eventfd_signal_count()) {
3613 req->work.func = io_poll_trigger_evfd;
3615 req->flags |= REQ_F_COMP_LOCKED;
3619 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3620 __io_cqring_ev_posted(ctx, trigger_ev);
3623 req->llist_node.next = NULL;
3624 /* if the list wasn't empty, we're done */
3625 if (!llist_add(&req->llist_node, &ctx->poll_llist))
3628 req->work.func = io_poll_flush;
3632 io_queue_async_work(req);
3637 struct io_poll_table {
3638 struct poll_table_struct pt;
3639 struct io_kiocb *req;
3643 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
3644 struct poll_table_struct *p)
3646 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
3648 if (unlikely(pt->req->poll.head)) {
3649 pt->error = -EINVAL;
3654 pt->req->poll.head = head;
3655 add_wait_queue(head, &pt->req->poll.wait);
3658 static void io_poll_req_insert(struct io_kiocb *req)
3660 struct io_ring_ctx *ctx = req->ctx;
3661 struct hlist_head *list;
3663 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
3664 hlist_add_head(&req->hash_node, list);
3667 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3669 struct io_poll_iocb *poll = &req->poll;
3672 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3674 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3679 events = READ_ONCE(sqe->poll_events);
3680 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
3684 static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt)
3686 struct io_poll_iocb *poll = &req->poll;
3687 struct io_ring_ctx *ctx = req->ctx;
3688 struct io_poll_table ipt;
3689 bool cancel = false;
3692 INIT_IO_WORK(&req->work, io_poll_complete_work);
3693 INIT_HLIST_NODE(&req->hash_node);
3697 poll->canceled = false;
3699 ipt.pt._qproc = io_poll_queue_proc;
3700 ipt.pt._key = poll->events;
3702 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
3704 /* initialized the list so that we can do list_empty checks */
3705 INIT_LIST_HEAD(&poll->wait.entry);
3706 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
3707 poll->wait.private = poll;
3709 INIT_LIST_HEAD(&req->list);
3711 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
3713 spin_lock_irq(&ctx->completion_lock);
3714 if (likely(poll->head)) {
3715 spin_lock(&poll->head->lock);
3716 if (unlikely(list_empty(&poll->wait.entry))) {
3722 if (mask || ipt.error)
3723 list_del_init(&poll->wait.entry);
3725 WRITE_ONCE(poll->canceled, true);
3726 else if (!poll->done) /* actually waiting for an event */
3727 io_poll_req_insert(req);
3728 spin_unlock(&poll->head->lock);
3730 if (mask) { /* no async, we'd stolen it */
3732 io_poll_complete(req, mask, 0);
3734 spin_unlock_irq(&ctx->completion_lock);
3737 io_cqring_ev_posted(ctx);
3738 io_put_req_find_next(req, nxt);
3743 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
3745 struct io_timeout_data *data = container_of(timer,
3746 struct io_timeout_data, timer);
3747 struct io_kiocb *req = data->req;
3748 struct io_ring_ctx *ctx = req->ctx;
3749 unsigned long flags;
3751 atomic_inc(&ctx->cq_timeouts);
3753 spin_lock_irqsave(&ctx->completion_lock, flags);
3755 * We could be racing with timeout deletion. If the list is empty,
3756 * then timeout lookup already found it and will be handling it.
3758 if (!list_empty(&req->list)) {
3759 struct io_kiocb *prev;
3762 * Adjust the reqs sequence before the current one because it
3763 * will consume a slot in the cq_ring and the cq_tail
3764 * pointer will be increased, otherwise other timeout reqs may
3765 * return in advance without waiting for enough wait_nr.
3768 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
3770 list_del_init(&req->list);
3773 io_cqring_fill_event(req, -ETIME);
3774 io_commit_cqring(ctx);
3775 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3777 io_cqring_ev_posted(ctx);
3778 req_set_fail_links(req);
3780 return HRTIMER_NORESTART;
3783 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
3785 struct io_kiocb *req;
3788 list_for_each_entry(req, &ctx->timeout_list, list) {
3789 if (user_data == req->user_data) {
3790 list_del_init(&req->list);
3799 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
3803 req_set_fail_links(req);
3804 io_cqring_fill_event(req, -ECANCELED);
3809 static int io_timeout_remove_prep(struct io_kiocb *req,
3810 const struct io_uring_sqe *sqe)
3812 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3814 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
3817 req->timeout.addr = READ_ONCE(sqe->addr);
3818 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
3819 if (req->timeout.flags)
3826 * Remove or update an existing timeout command
3828 static int io_timeout_remove(struct io_kiocb *req)
3830 struct io_ring_ctx *ctx = req->ctx;
3833 spin_lock_irq(&ctx->completion_lock);
3834 ret = io_timeout_cancel(ctx, req->timeout.addr);
3836 io_cqring_fill_event(req, ret);
3837 io_commit_cqring(ctx);
3838 spin_unlock_irq(&ctx->completion_lock);
3839 io_cqring_ev_posted(ctx);
3841 req_set_fail_links(req);
3846 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3847 bool is_timeout_link)
3849 struct io_timeout_data *data;
3852 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3854 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
3856 if (sqe->off && is_timeout_link)
3858 flags = READ_ONCE(sqe->timeout_flags);
3859 if (flags & ~IORING_TIMEOUT_ABS)
3862 req->timeout.count = READ_ONCE(sqe->off);
3864 if (!req->io && io_alloc_async_ctx(req))
3867 data = &req->io->timeout;
3869 req->flags |= REQ_F_TIMEOUT;
3871 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
3874 if (flags & IORING_TIMEOUT_ABS)
3875 data->mode = HRTIMER_MODE_ABS;
3877 data->mode = HRTIMER_MODE_REL;
3879 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
3883 static int io_timeout(struct io_kiocb *req)
3886 struct io_ring_ctx *ctx = req->ctx;
3887 struct io_timeout_data *data;
3888 struct list_head *entry;
3891 data = &req->io->timeout;
3894 * sqe->off holds how many events that need to occur for this
3895 * timeout event to be satisfied. If it isn't set, then this is
3896 * a pure timeout request, sequence isn't used.
3898 count = req->timeout.count;
3900 req->flags |= REQ_F_TIMEOUT_NOSEQ;
3901 spin_lock_irq(&ctx->completion_lock);
3902 entry = ctx->timeout_list.prev;
3906 req->sequence = ctx->cached_sq_head + count - 1;
3907 data->seq_offset = count;
3910 * Insertion sort, ensuring the first entry in the list is always
3911 * the one we need first.
3913 spin_lock_irq(&ctx->completion_lock);
3914 list_for_each_prev(entry, &ctx->timeout_list) {
3915 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
3916 unsigned nxt_sq_head;
3917 long long tmp, tmp_nxt;
3918 u32 nxt_offset = nxt->io->timeout.seq_offset;
3920 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
3924 * Since cached_sq_head + count - 1 can overflow, use type long
3927 tmp = (long long)ctx->cached_sq_head + count - 1;
3928 nxt_sq_head = nxt->sequence - nxt_offset + 1;
3929 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
3932 * cached_sq_head may overflow, and it will never overflow twice
3933 * once there is some timeout req still be valid.
3935 if (ctx->cached_sq_head < nxt_sq_head)
3942 * Sequence of reqs after the insert one and itself should
3943 * be adjusted because each timeout req consumes a slot.
3948 req->sequence -= span;
3950 list_add(&req->list, entry);
3951 data->timer.function = io_timeout_fn;
3952 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
3953 spin_unlock_irq(&ctx->completion_lock);
3957 static bool io_cancel_cb(struct io_wq_work *work, void *data)
3959 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3961 return req->user_data == (unsigned long) data;
3964 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
3966 enum io_wq_cancel cancel_ret;
3969 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
3970 switch (cancel_ret) {
3971 case IO_WQ_CANCEL_OK:
3974 case IO_WQ_CANCEL_RUNNING:
3977 case IO_WQ_CANCEL_NOTFOUND:
3985 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
3986 struct io_kiocb *req, __u64 sqe_addr,
3987 struct io_kiocb **nxt, int success_ret)
3989 unsigned long flags;
3992 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
3993 if (ret != -ENOENT) {
3994 spin_lock_irqsave(&ctx->completion_lock, flags);
3998 spin_lock_irqsave(&ctx->completion_lock, flags);
3999 ret = io_timeout_cancel(ctx, sqe_addr);
4002 ret = io_poll_cancel(ctx, sqe_addr);
4006 io_cqring_fill_event(req, ret);
4007 io_commit_cqring(ctx);
4008 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4009 io_cqring_ev_posted(ctx);
4012 req_set_fail_links(req);
4013 io_put_req_find_next(req, nxt);
4016 static int io_async_cancel_prep(struct io_kiocb *req,
4017 const struct io_uring_sqe *sqe)
4019 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4021 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4025 req->cancel.addr = READ_ONCE(sqe->addr);
4029 static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt)
4031 struct io_ring_ctx *ctx = req->ctx;
4033 io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0);
4037 static int io_files_update_prep(struct io_kiocb *req,
4038 const struct io_uring_sqe *sqe)
4040 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4043 req->files_update.offset = READ_ONCE(sqe->off);
4044 req->files_update.nr_args = READ_ONCE(sqe->len);
4045 if (!req->files_update.nr_args)
4047 req->files_update.arg = READ_ONCE(sqe->addr);
4051 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4053 struct io_ring_ctx *ctx = req->ctx;
4054 struct io_uring_files_update up;
4060 up.offset = req->files_update.offset;
4061 up.fds = req->files_update.arg;
4063 mutex_lock(&ctx->uring_lock);
4064 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4065 mutex_unlock(&ctx->uring_lock);
4068 req_set_fail_links(req);
4069 io_cqring_add_event(req, ret);
4074 static int io_req_defer_prep(struct io_kiocb *req,
4075 const struct io_uring_sqe *sqe)
4079 if (io_op_defs[req->opcode].file_table) {
4080 ret = io_grab_files(req);
4085 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4087 switch (req->opcode) {
4090 case IORING_OP_READV:
4091 case IORING_OP_READ_FIXED:
4092 case IORING_OP_READ:
4093 ret = io_read_prep(req, sqe, true);
4095 case IORING_OP_WRITEV:
4096 case IORING_OP_WRITE_FIXED:
4097 case IORING_OP_WRITE:
4098 ret = io_write_prep(req, sqe, true);
4100 case IORING_OP_POLL_ADD:
4101 ret = io_poll_add_prep(req, sqe);
4103 case IORING_OP_POLL_REMOVE:
4104 ret = io_poll_remove_prep(req, sqe);
4106 case IORING_OP_FSYNC:
4107 ret = io_prep_fsync(req, sqe);
4109 case IORING_OP_SYNC_FILE_RANGE:
4110 ret = io_prep_sfr(req, sqe);
4112 case IORING_OP_SENDMSG:
4113 case IORING_OP_SEND:
4114 ret = io_sendmsg_prep(req, sqe);
4116 case IORING_OP_RECVMSG:
4117 case IORING_OP_RECV:
4118 ret = io_recvmsg_prep(req, sqe);
4120 case IORING_OP_CONNECT:
4121 ret = io_connect_prep(req, sqe);
4123 case IORING_OP_TIMEOUT:
4124 ret = io_timeout_prep(req, sqe, false);
4126 case IORING_OP_TIMEOUT_REMOVE:
4127 ret = io_timeout_remove_prep(req, sqe);
4129 case IORING_OP_ASYNC_CANCEL:
4130 ret = io_async_cancel_prep(req, sqe);
4132 case IORING_OP_LINK_TIMEOUT:
4133 ret = io_timeout_prep(req, sqe, true);
4135 case IORING_OP_ACCEPT:
4136 ret = io_accept_prep(req, sqe);
4138 case IORING_OP_FALLOCATE:
4139 ret = io_fallocate_prep(req, sqe);
4141 case IORING_OP_OPENAT:
4142 ret = io_openat_prep(req, sqe);
4144 case IORING_OP_CLOSE:
4145 ret = io_close_prep(req, sqe);
4147 case IORING_OP_FILES_UPDATE:
4148 ret = io_files_update_prep(req, sqe);
4150 case IORING_OP_STATX:
4151 ret = io_statx_prep(req, sqe);
4153 case IORING_OP_FADVISE:
4154 ret = io_fadvise_prep(req, sqe);
4156 case IORING_OP_MADVISE:
4157 ret = io_madvise_prep(req, sqe);
4159 case IORING_OP_OPENAT2:
4160 ret = io_openat2_prep(req, sqe);
4162 case IORING_OP_EPOLL_CTL:
4163 ret = io_epoll_ctl_prep(req, sqe);
4166 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
4175 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4177 struct io_ring_ctx *ctx = req->ctx;
4180 /* Still need defer if there is pending req in defer list. */
4181 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
4184 if (!req->io && io_alloc_async_ctx(req))
4187 ret = io_req_defer_prep(req, sqe);
4191 spin_lock_irq(&ctx->completion_lock);
4192 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
4193 spin_unlock_irq(&ctx->completion_lock);
4197 trace_io_uring_defer(ctx, req, req->user_data);
4198 list_add_tail(&req->list, &ctx->defer_list);
4199 spin_unlock_irq(&ctx->completion_lock);
4200 return -EIOCBQUEUED;
4203 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4204 struct io_kiocb **nxt, bool force_nonblock)
4206 struct io_ring_ctx *ctx = req->ctx;
4209 switch (req->opcode) {
4213 case IORING_OP_READV:
4214 case IORING_OP_READ_FIXED:
4215 case IORING_OP_READ:
4217 ret = io_read_prep(req, sqe, force_nonblock);
4221 ret = io_read(req, nxt, force_nonblock);
4223 case IORING_OP_WRITEV:
4224 case IORING_OP_WRITE_FIXED:
4225 case IORING_OP_WRITE:
4227 ret = io_write_prep(req, sqe, force_nonblock);
4231 ret = io_write(req, nxt, force_nonblock);
4233 case IORING_OP_FSYNC:
4235 ret = io_prep_fsync(req, sqe);
4239 ret = io_fsync(req, nxt, force_nonblock);
4241 case IORING_OP_POLL_ADD:
4243 ret = io_poll_add_prep(req, sqe);
4247 ret = io_poll_add(req, nxt);
4249 case IORING_OP_POLL_REMOVE:
4251 ret = io_poll_remove_prep(req, sqe);
4255 ret = io_poll_remove(req);
4257 case IORING_OP_SYNC_FILE_RANGE:
4259 ret = io_prep_sfr(req, sqe);
4263 ret = io_sync_file_range(req, nxt, force_nonblock);
4265 case IORING_OP_SENDMSG:
4266 case IORING_OP_SEND:
4268 ret = io_sendmsg_prep(req, sqe);
4272 if (req->opcode == IORING_OP_SENDMSG)
4273 ret = io_sendmsg(req, nxt, force_nonblock);
4275 ret = io_send(req, nxt, force_nonblock);
4277 case IORING_OP_RECVMSG:
4278 case IORING_OP_RECV:
4280 ret = io_recvmsg_prep(req, sqe);
4284 if (req->opcode == IORING_OP_RECVMSG)
4285 ret = io_recvmsg(req, nxt, force_nonblock);
4287 ret = io_recv(req, nxt, force_nonblock);
4289 case IORING_OP_TIMEOUT:
4291 ret = io_timeout_prep(req, sqe, false);
4295 ret = io_timeout(req);
4297 case IORING_OP_TIMEOUT_REMOVE:
4299 ret = io_timeout_remove_prep(req, sqe);
4303 ret = io_timeout_remove(req);
4305 case IORING_OP_ACCEPT:
4307 ret = io_accept_prep(req, sqe);
4311 ret = io_accept(req, nxt, force_nonblock);
4313 case IORING_OP_CONNECT:
4315 ret = io_connect_prep(req, sqe);
4319 ret = io_connect(req, nxt, force_nonblock);
4321 case IORING_OP_ASYNC_CANCEL:
4323 ret = io_async_cancel_prep(req, sqe);
4327 ret = io_async_cancel(req, nxt);
4329 case IORING_OP_FALLOCATE:
4331 ret = io_fallocate_prep(req, sqe);
4335 ret = io_fallocate(req, nxt, force_nonblock);
4337 case IORING_OP_OPENAT:
4339 ret = io_openat_prep(req, sqe);
4343 ret = io_openat(req, nxt, force_nonblock);
4345 case IORING_OP_CLOSE:
4347 ret = io_close_prep(req, sqe);
4351 ret = io_close(req, nxt, force_nonblock);
4353 case IORING_OP_FILES_UPDATE:
4355 ret = io_files_update_prep(req, sqe);
4359 ret = io_files_update(req, force_nonblock);
4361 case IORING_OP_STATX:
4363 ret = io_statx_prep(req, sqe);
4367 ret = io_statx(req, nxt, force_nonblock);
4369 case IORING_OP_FADVISE:
4371 ret = io_fadvise_prep(req, sqe);
4375 ret = io_fadvise(req, nxt, force_nonblock);
4377 case IORING_OP_MADVISE:
4379 ret = io_madvise_prep(req, sqe);
4383 ret = io_madvise(req, nxt, force_nonblock);
4385 case IORING_OP_OPENAT2:
4387 ret = io_openat2_prep(req, sqe);
4391 ret = io_openat2(req, nxt, force_nonblock);
4393 case IORING_OP_EPOLL_CTL:
4395 ret = io_epoll_ctl_prep(req, sqe);
4399 ret = io_epoll_ctl(req, nxt, force_nonblock);
4409 if (ctx->flags & IORING_SETUP_IOPOLL) {
4410 const bool in_async = io_wq_current_is_worker();
4412 if (req->result == -EAGAIN)
4415 /* workqueue context doesn't hold uring_lock, grab it now */
4417 mutex_lock(&ctx->uring_lock);
4419 io_iopoll_req_issued(req);
4422 mutex_unlock(&ctx->uring_lock);
4428 static void io_wq_submit_work(struct io_wq_work **workptr)
4430 struct io_wq_work *work = *workptr;
4431 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4432 struct io_kiocb *nxt = NULL;
4435 /* if NO_CANCEL is set, we must still run the work */
4436 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
4437 IO_WQ_WORK_CANCEL) {
4442 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
4443 req->in_async = true;
4445 ret = io_issue_sqe(req, NULL, &nxt, false);
4447 * We can get EAGAIN for polled IO even though we're
4448 * forcing a sync submission from here, since we can't
4449 * wait for request slots on the block side.
4457 /* drop submission reference */
4461 req_set_fail_links(req);
4462 io_cqring_add_event(req, ret);
4466 /* if a dependent link is ready, pass it back */
4468 io_wq_assign_next(workptr, nxt);
4471 static int io_req_needs_file(struct io_kiocb *req, int fd)
4473 if (!io_op_defs[req->opcode].needs_file)
4475 if (fd == -1 && io_op_defs[req->opcode].fd_non_neg)
4480 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
4483 struct fixed_file_table *table;
4485 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
4486 return table->files[index & IORING_FILE_TABLE_MASK];;
4489 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
4490 const struct io_uring_sqe *sqe)
4492 struct io_ring_ctx *ctx = req->ctx;
4496 flags = READ_ONCE(sqe->flags);
4497 fd = READ_ONCE(sqe->fd);
4499 if (!io_req_needs_file(req, fd))
4502 if (flags & IOSQE_FIXED_FILE) {
4503 if (unlikely(!ctx->file_data ||
4504 (unsigned) fd >= ctx->nr_user_files))
4506 fd = array_index_nospec(fd, ctx->nr_user_files);
4507 req->file = io_file_from_index(ctx, fd);
4510 req->flags |= REQ_F_FIXED_FILE;
4511 percpu_ref_get(&ctx->file_data->refs);
4513 if (req->needs_fixed_file)
4515 trace_io_uring_file_get(ctx, fd);
4516 req->file = io_file_get(state, fd);
4517 if (unlikely(!req->file))
4524 static int io_grab_files(struct io_kiocb *req)
4527 struct io_ring_ctx *ctx = req->ctx;
4529 if (req->work.files)
4531 if (!ctx->ring_file)
4535 spin_lock_irq(&ctx->inflight_lock);
4537 * We use the f_ops->flush() handler to ensure that we can flush
4538 * out work accessing these files if the fd is closed. Check if
4539 * the fd has changed since we started down this path, and disallow
4540 * this operation if it has.
4542 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
4543 list_add(&req->inflight_entry, &ctx->inflight_list);
4544 req->flags |= REQ_F_INFLIGHT;
4545 req->work.files = current->files;
4548 spin_unlock_irq(&ctx->inflight_lock);
4554 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
4556 struct io_timeout_data *data = container_of(timer,
4557 struct io_timeout_data, timer);
4558 struct io_kiocb *req = data->req;
4559 struct io_ring_ctx *ctx = req->ctx;
4560 struct io_kiocb *prev = NULL;
4561 unsigned long flags;
4563 spin_lock_irqsave(&ctx->completion_lock, flags);
4566 * We don't expect the list to be empty, that will only happen if we
4567 * race with the completion of the linked work.
4569 if (!list_empty(&req->link_list)) {
4570 prev = list_entry(req->link_list.prev, struct io_kiocb,
4572 if (refcount_inc_not_zero(&prev->refs)) {
4573 list_del_init(&req->link_list);
4574 prev->flags &= ~REQ_F_LINK_TIMEOUT;
4579 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4582 req_set_fail_links(prev);
4583 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
4587 io_cqring_add_event(req, -ETIME);
4590 return HRTIMER_NORESTART;
4593 static void io_queue_linked_timeout(struct io_kiocb *req)
4595 struct io_ring_ctx *ctx = req->ctx;
4598 * If the list is now empty, then our linked request finished before
4599 * we got a chance to setup the timer
4601 spin_lock_irq(&ctx->completion_lock);
4602 if (!list_empty(&req->link_list)) {
4603 struct io_timeout_data *data = &req->io->timeout;
4605 data->timer.function = io_link_timeout_fn;
4606 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
4609 spin_unlock_irq(&ctx->completion_lock);
4611 /* drop submission reference */
4615 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
4617 struct io_kiocb *nxt;
4619 if (!(req->flags & REQ_F_LINK))
4622 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
4624 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
4627 req->flags |= REQ_F_LINK_TIMEOUT;
4631 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4633 struct io_kiocb *linked_timeout;
4634 struct io_kiocb *nxt = NULL;
4638 linked_timeout = io_prep_linked_timeout(req);
4640 ret = io_issue_sqe(req, sqe, &nxt, true);
4643 * We async punt it if the file wasn't marked NOWAIT, or if the file
4644 * doesn't support non-blocking read/write attempts
4646 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
4647 (req->flags & REQ_F_MUST_PUNT))) {
4649 if (io_op_defs[req->opcode].file_table) {
4650 ret = io_grab_files(req);
4656 * Queued up for async execution, worker will release
4657 * submit reference when the iocb is actually submitted.
4659 io_queue_async_work(req);
4664 /* drop submission reference */
4667 if (linked_timeout) {
4669 io_queue_linked_timeout(linked_timeout);
4671 io_put_req(linked_timeout);
4674 /* and drop final reference, if we failed */
4676 io_cqring_add_event(req, ret);
4677 req_set_fail_links(req);
4685 if (req->flags & REQ_F_FORCE_ASYNC)
4691 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4695 ret = io_req_defer(req, sqe);
4697 if (ret != -EIOCBQUEUED) {
4699 io_cqring_add_event(req, ret);
4700 req_set_fail_links(req);
4701 io_double_put_req(req);
4703 } else if (req->flags & REQ_F_FORCE_ASYNC) {
4704 ret = io_req_defer_prep(req, sqe);
4705 if (unlikely(ret < 0))
4708 * Never try inline submit of IOSQE_ASYNC is set, go straight
4709 * to async execution.
4711 req->work.flags |= IO_WQ_WORK_CONCURRENT;
4712 io_queue_async_work(req);
4714 __io_queue_sqe(req, sqe);
4718 static inline void io_queue_link_head(struct io_kiocb *req)
4720 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
4721 io_cqring_add_event(req, -ECANCELED);
4722 io_double_put_req(req);
4724 io_queue_sqe(req, NULL);
4727 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
4728 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
4730 static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4731 struct io_submit_state *state, struct io_kiocb **link)
4733 const struct cred *old_creds = NULL;
4734 struct io_ring_ctx *ctx = req->ctx;
4735 unsigned int sqe_flags;
4738 sqe_flags = READ_ONCE(sqe->flags);
4740 /* enforce forwards compatibility on users */
4741 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
4746 id = READ_ONCE(sqe->personality);
4748 const struct cred *personality_creds;
4750 personality_creds = idr_find(&ctx->personality_idr, id);
4751 if (unlikely(!personality_creds)) {
4755 old_creds = override_creds(personality_creds);
4758 /* same numerical values with corresponding REQ_F_*, safe to copy */
4759 req->flags |= sqe_flags & (IOSQE_IO_DRAIN|IOSQE_IO_HARDLINK|
4762 ret = io_req_set_file(state, req, sqe);
4763 if (unlikely(ret)) {
4765 io_cqring_add_event(req, ret);
4766 io_double_put_req(req);
4768 revert_creds(old_creds);
4773 * If we already have a head request, queue this one for async
4774 * submittal once the head completes. If we don't have a head but
4775 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
4776 * submitted sync once the chain is complete. If none of those
4777 * conditions are true (normal request), then just queue it.
4780 struct io_kiocb *head = *link;
4783 * Taking sequential execution of a link, draining both sides
4784 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
4785 * requests in the link. So, it drains the head and the
4786 * next after the link request. The last one is done via
4787 * drain_next flag to persist the effect across calls.
4789 if (sqe_flags & IOSQE_IO_DRAIN) {
4790 head->flags |= REQ_F_IO_DRAIN;
4791 ctx->drain_next = 1;
4793 if (io_alloc_async_ctx(req)) {
4798 ret = io_req_defer_prep(req, sqe);
4800 /* fail even hard links since we don't submit */
4801 head->flags |= REQ_F_FAIL_LINK;
4804 trace_io_uring_link(ctx, req, head);
4805 list_add_tail(&req->link_list, &head->link_list);
4807 /* last request of a link, enqueue the link */
4808 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) {
4809 io_queue_link_head(head);
4813 if (unlikely(ctx->drain_next)) {
4814 req->flags |= REQ_F_IO_DRAIN;
4815 req->ctx->drain_next = 0;
4817 if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
4818 req->flags |= REQ_F_LINK;
4819 INIT_LIST_HEAD(&req->link_list);
4820 ret = io_req_defer_prep(req, sqe);
4822 req->flags |= REQ_F_FAIL_LINK;
4825 io_queue_sqe(req, sqe);
4830 revert_creds(old_creds);
4835 * Batched submission is done, ensure local IO is flushed out.
4837 static void io_submit_state_end(struct io_submit_state *state)
4839 blk_finish_plug(&state->plug);
4841 if (state->free_reqs)
4842 kmem_cache_free_bulk(req_cachep, state->free_reqs,
4843 &state->reqs[state->cur_req]);
4847 * Start submission side cache.
4849 static void io_submit_state_start(struct io_submit_state *state,
4850 unsigned int max_ios)
4852 blk_start_plug(&state->plug);
4853 state->free_reqs = 0;
4855 state->ios_left = max_ios;
4858 static void io_commit_sqring(struct io_ring_ctx *ctx)
4860 struct io_rings *rings = ctx->rings;
4863 * Ensure any loads from the SQEs are done at this point,
4864 * since once we write the new head, the application could
4865 * write new data to them.
4867 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
4871 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
4872 * that is mapped by userspace. This means that care needs to be taken to
4873 * ensure that reads are stable, as we cannot rely on userspace always
4874 * being a good citizen. If members of the sqe are validated and then later
4875 * used, it's important that those reads are done through READ_ONCE() to
4876 * prevent a re-load down the line.
4878 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
4879 const struct io_uring_sqe **sqe_ptr)
4881 u32 *sq_array = ctx->sq_array;
4885 * The cached sq head (or cq tail) serves two purposes:
4887 * 1) allows us to batch the cost of updating the user visible
4889 * 2) allows the kernel side to track the head on its own, even
4890 * though the application is the one updating it.
4892 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
4893 if (likely(head < ctx->sq_entries)) {
4895 * All io need record the previous position, if LINK vs DARIN,
4896 * it can be used to mark the position of the first IO in the
4899 req->sequence = ctx->cached_sq_head;
4900 *sqe_ptr = &ctx->sq_sqes[head];
4901 req->opcode = READ_ONCE((*sqe_ptr)->opcode);
4902 req->user_data = READ_ONCE((*sqe_ptr)->user_data);
4903 ctx->cached_sq_head++;
4907 /* drop invalid entries */
4908 ctx->cached_sq_head++;
4909 ctx->cached_sq_dropped++;
4910 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
4914 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
4915 struct file *ring_file, int ring_fd,
4916 struct mm_struct **mm, bool async)
4918 struct io_submit_state state, *statep = NULL;
4919 struct io_kiocb *link = NULL;
4920 int i, submitted = 0;
4921 bool mm_fault = false;
4923 /* if we have a backlog and couldn't flush it all, return BUSY */
4924 if (test_bit(0, &ctx->sq_check_overflow)) {
4925 if (!list_empty(&ctx->cq_overflow_list) &&
4926 !io_cqring_overflow_flush(ctx, false))
4930 /* make sure SQ entry isn't read before tail */
4931 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
4933 if (!percpu_ref_tryget_many(&ctx->refs, nr))
4936 if (nr > IO_PLUG_THRESHOLD) {
4937 io_submit_state_start(&state, nr);
4941 ctx->ring_fd = ring_fd;
4942 ctx->ring_file = ring_file;
4944 for (i = 0; i < nr; i++) {
4945 const struct io_uring_sqe *sqe;
4946 struct io_kiocb *req;
4948 req = io_get_req(ctx, statep);
4949 if (unlikely(!req)) {
4951 submitted = -EAGAIN;
4954 if (!io_get_sqring(ctx, req, &sqe)) {
4955 __io_req_do_free(req);
4959 /* will complete beyond this point, count as submitted */
4962 if (unlikely(req->opcode >= IORING_OP_LAST)) {
4963 io_cqring_add_event(req, -EINVAL);
4964 io_double_put_req(req);
4968 if (io_op_defs[req->opcode].needs_mm && !*mm) {
4969 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
4971 use_mm(ctx->sqo_mm);
4976 req->has_user = *mm != NULL;
4977 req->in_async = async;
4978 req->needs_fixed_file = async;
4979 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
4981 if (!io_submit_sqe(req, sqe, statep, &link))
4985 if (unlikely(submitted != nr)) {
4986 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
4988 percpu_ref_put_many(&ctx->refs, nr - ref_used);
4991 io_queue_link_head(link);
4993 io_submit_state_end(&state);
4995 /* Commit SQ ring head once we've consumed and submitted all SQEs */
4996 io_commit_sqring(ctx);
5001 static int io_sq_thread(void *data)
5003 struct io_ring_ctx *ctx = data;
5004 struct mm_struct *cur_mm = NULL;
5005 const struct cred *old_cred;
5006 mm_segment_t old_fs;
5009 unsigned long timeout;
5012 complete(&ctx->completions[1]);
5016 old_cred = override_creds(ctx->creds);
5018 ret = timeout = inflight = 0;
5019 while (!kthread_should_park()) {
5020 unsigned int to_submit;
5023 unsigned nr_events = 0;
5025 if (ctx->flags & IORING_SETUP_IOPOLL) {
5027 * inflight is the count of the maximum possible
5028 * entries we submitted, but it can be smaller
5029 * if we dropped some of them. If we don't have
5030 * poll entries available, then we know that we
5031 * have nothing left to poll for. Reset the
5032 * inflight count to zero in that case.
5034 mutex_lock(&ctx->uring_lock);
5035 if (!list_empty(&ctx->poll_list))
5036 __io_iopoll_check(ctx, &nr_events, 0);
5039 mutex_unlock(&ctx->uring_lock);
5042 * Normal IO, just pretend everything completed.
5043 * We don't have to poll completions for that.
5045 nr_events = inflight;
5048 inflight -= nr_events;
5050 timeout = jiffies + ctx->sq_thread_idle;
5053 to_submit = io_sqring_entries(ctx);
5056 * If submit got -EBUSY, flag us as needing the application
5057 * to enter the kernel to reap and flush events.
5059 if (!to_submit || ret == -EBUSY) {
5061 * We're polling. If we're within the defined idle
5062 * period, then let us spin without work before going
5063 * to sleep. The exception is if we got EBUSY doing
5064 * more IO, we should wait for the application to
5065 * reap events and wake us up.
5068 (!time_after(jiffies, timeout) && ret != -EBUSY)) {
5074 * Drop cur_mm before scheduling, we can't hold it for
5075 * long periods (or over schedule()). Do this before
5076 * adding ourselves to the waitqueue, as the unuse/drop
5085 prepare_to_wait(&ctx->sqo_wait, &wait,
5086 TASK_INTERRUPTIBLE);
5088 /* Tell userspace we may need a wakeup call */
5089 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
5090 /* make sure to read SQ tail after writing flags */
5093 to_submit = io_sqring_entries(ctx);
5094 if (!to_submit || ret == -EBUSY) {
5095 if (kthread_should_park()) {
5096 finish_wait(&ctx->sqo_wait, &wait);
5099 if (signal_pending(current))
5100 flush_signals(current);
5102 finish_wait(&ctx->sqo_wait, &wait);
5104 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5107 finish_wait(&ctx->sqo_wait, &wait);
5109 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5112 mutex_lock(&ctx->uring_lock);
5113 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
5114 mutex_unlock(&ctx->uring_lock);
5124 revert_creds(old_cred);
5131 struct io_wait_queue {
5132 struct wait_queue_entry wq;
5133 struct io_ring_ctx *ctx;
5135 unsigned nr_timeouts;
5138 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
5140 struct io_ring_ctx *ctx = iowq->ctx;
5143 * Wake up if we have enough events, or if a timeout occurred since we
5144 * started waiting. For timeouts, we always want to return to userspace,
5145 * regardless of event count.
5147 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
5148 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
5151 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
5152 int wake_flags, void *key)
5154 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
5157 /* use noflush == true, as we can't safely rely on locking context */
5158 if (!io_should_wake(iowq, true))
5161 return autoremove_wake_function(curr, mode, wake_flags, key);
5165 * Wait until events become available, if we don't already have some. The
5166 * application must reap them itself, as they reside on the shared cq ring.
5168 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
5169 const sigset_t __user *sig, size_t sigsz)
5171 struct io_wait_queue iowq = {
5174 .func = io_wake_function,
5175 .entry = LIST_HEAD_INIT(iowq.wq.entry),
5178 .to_wait = min_events,
5180 struct io_rings *rings = ctx->rings;
5183 if (io_cqring_events(ctx, false) >= min_events)
5187 #ifdef CONFIG_COMPAT
5188 if (in_compat_syscall())
5189 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
5193 ret = set_user_sigmask(sig, sigsz);
5199 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
5200 trace_io_uring_cqring_wait(ctx, min_events);
5202 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
5203 TASK_INTERRUPTIBLE);
5204 if (io_should_wake(&iowq, false))
5207 if (signal_pending(current)) {
5212 finish_wait(&ctx->wait, &iowq.wq);
5214 restore_saved_sigmask_unless(ret == -EINTR);
5216 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
5219 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
5221 #if defined(CONFIG_UNIX)
5222 if (ctx->ring_sock) {
5223 struct sock *sock = ctx->ring_sock->sk;
5224 struct sk_buff *skb;
5226 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
5232 for (i = 0; i < ctx->nr_user_files; i++) {
5235 file = io_file_from_index(ctx, i);
5242 static void io_file_ref_kill(struct percpu_ref *ref)
5244 struct fixed_file_data *data;
5246 data = container_of(ref, struct fixed_file_data, refs);
5247 complete(&data->done);
5250 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
5252 struct fixed_file_data *data = ctx->file_data;
5253 unsigned nr_tables, i;
5258 /* protect against inflight atomic switch, which drops the ref */
5259 percpu_ref_get(&data->refs);
5260 /* wait for existing switches */
5261 flush_work(&data->ref_work);
5262 percpu_ref_kill_and_confirm(&data->refs, io_file_ref_kill);
5263 wait_for_completion(&data->done);
5264 percpu_ref_put(&data->refs);
5265 /* flush potential new switch */
5266 flush_work(&data->ref_work);
5267 percpu_ref_exit(&data->refs);
5269 __io_sqe_files_unregister(ctx);
5270 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
5271 for (i = 0; i < nr_tables; i++)
5272 kfree(data->table[i].files);
5275 ctx->file_data = NULL;
5276 ctx->nr_user_files = 0;
5280 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
5282 if (ctx->sqo_thread) {
5283 wait_for_completion(&ctx->completions[1]);
5285 * The park is a bit of a work-around, without it we get
5286 * warning spews on shutdown with SQPOLL set and affinity
5287 * set to a single CPU.
5289 kthread_park(ctx->sqo_thread);
5290 kthread_stop(ctx->sqo_thread);
5291 ctx->sqo_thread = NULL;
5295 static void io_finish_async(struct io_ring_ctx *ctx)
5297 io_sq_thread_stop(ctx);
5300 io_wq_destroy(ctx->io_wq);
5305 #if defined(CONFIG_UNIX)
5307 * Ensure the UNIX gc is aware of our file set, so we are certain that
5308 * the io_uring can be safely unregistered on process exit, even if we have
5309 * loops in the file referencing.
5311 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
5313 struct sock *sk = ctx->ring_sock->sk;
5314 struct scm_fp_list *fpl;
5315 struct sk_buff *skb;
5318 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
5319 unsigned long inflight = ctx->user->unix_inflight + nr;
5321 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
5325 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
5329 skb = alloc_skb(0, GFP_KERNEL);
5338 fpl->user = get_uid(ctx->user);
5339 for (i = 0; i < nr; i++) {
5340 struct file *file = io_file_from_index(ctx, i + offset);
5344 fpl->fp[nr_files] = get_file(file);
5345 unix_inflight(fpl->user, fpl->fp[nr_files]);
5350 fpl->max = SCM_MAX_FD;
5351 fpl->count = nr_files;
5352 UNIXCB(skb).fp = fpl;
5353 skb->destructor = unix_destruct_scm;
5354 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
5355 skb_queue_head(&sk->sk_receive_queue, skb);
5357 for (i = 0; i < nr_files; i++)
5368 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
5369 * causes regular reference counting to break down. We rely on the UNIX
5370 * garbage collection to take care of this problem for us.
5372 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5374 unsigned left, total;
5378 left = ctx->nr_user_files;
5380 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
5382 ret = __io_sqe_files_scm(ctx, this_files, total);
5386 total += this_files;
5392 while (total < ctx->nr_user_files) {
5393 struct file *file = io_file_from_index(ctx, total);
5403 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5409 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
5414 for (i = 0; i < nr_tables; i++) {
5415 struct fixed_file_table *table = &ctx->file_data->table[i];
5416 unsigned this_files;
5418 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
5419 table->files = kcalloc(this_files, sizeof(struct file *),
5423 nr_files -= this_files;
5429 for (i = 0; i < nr_tables; i++) {
5430 struct fixed_file_table *table = &ctx->file_data->table[i];
5431 kfree(table->files);
5436 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
5438 #if defined(CONFIG_UNIX)
5439 struct sock *sock = ctx->ring_sock->sk;
5440 struct sk_buff_head list, *head = &sock->sk_receive_queue;
5441 struct sk_buff *skb;
5444 __skb_queue_head_init(&list);
5447 * Find the skb that holds this file in its SCM_RIGHTS. When found,
5448 * remove this entry and rearrange the file array.
5450 skb = skb_dequeue(head);
5452 struct scm_fp_list *fp;
5454 fp = UNIXCB(skb).fp;
5455 for (i = 0; i < fp->count; i++) {
5458 if (fp->fp[i] != file)
5461 unix_notinflight(fp->user, fp->fp[i]);
5462 left = fp->count - 1 - i;
5464 memmove(&fp->fp[i], &fp->fp[i + 1],
5465 left * sizeof(struct file *));
5472 __skb_queue_tail(&list, skb);
5482 __skb_queue_tail(&list, skb);
5484 skb = skb_dequeue(head);
5487 if (skb_peek(&list)) {
5488 spin_lock_irq(&head->lock);
5489 while ((skb = __skb_dequeue(&list)) != NULL)
5490 __skb_queue_tail(head, skb);
5491 spin_unlock_irq(&head->lock);
5498 struct io_file_put {
5499 struct llist_node llist;
5501 struct completion *done;
5504 static void io_ring_file_ref_switch(struct work_struct *work)
5506 struct io_file_put *pfile, *tmp;
5507 struct fixed_file_data *data;
5508 struct llist_node *node;
5510 data = container_of(work, struct fixed_file_data, ref_work);
5512 while ((node = llist_del_all(&data->put_llist)) != NULL) {
5513 llist_for_each_entry_safe(pfile, tmp, node, llist) {
5514 io_ring_file_put(data->ctx, pfile->file);
5516 complete(pfile->done);
5522 percpu_ref_get(&data->refs);
5523 percpu_ref_switch_to_percpu(&data->refs);
5526 static void io_file_data_ref_zero(struct percpu_ref *ref)
5528 struct fixed_file_data *data;
5530 data = container_of(ref, struct fixed_file_data, refs);
5532 /* we can't safely switch from inside this context, punt to wq */
5533 queue_work(system_wq, &data->ref_work);
5536 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
5539 __s32 __user *fds = (__s32 __user *) arg;
5549 if (nr_args > IORING_MAX_FIXED_FILES)
5552 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
5553 if (!ctx->file_data)
5555 ctx->file_data->ctx = ctx;
5556 init_completion(&ctx->file_data->done);
5558 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
5559 ctx->file_data->table = kcalloc(nr_tables,
5560 sizeof(struct fixed_file_table),
5562 if (!ctx->file_data->table) {
5563 kfree(ctx->file_data);
5564 ctx->file_data = NULL;
5568 if (percpu_ref_init(&ctx->file_data->refs, io_file_data_ref_zero,
5569 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
5570 kfree(ctx->file_data->table);
5571 kfree(ctx->file_data);
5572 ctx->file_data = NULL;
5575 ctx->file_data->put_llist.first = NULL;
5576 INIT_WORK(&ctx->file_data->ref_work, io_ring_file_ref_switch);
5578 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
5579 percpu_ref_exit(&ctx->file_data->refs);
5580 kfree(ctx->file_data->table);
5581 kfree(ctx->file_data);
5582 ctx->file_data = NULL;
5586 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
5587 struct fixed_file_table *table;
5591 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
5593 /* allow sparse sets */
5599 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5600 index = i & IORING_FILE_TABLE_MASK;
5608 * Don't allow io_uring instances to be registered. If UNIX
5609 * isn't enabled, then this causes a reference cycle and this
5610 * instance can never get freed. If UNIX is enabled we'll
5611 * handle it just fine, but there's still no point in allowing
5612 * a ring fd as it doesn't support regular read/write anyway.
5614 if (file->f_op == &io_uring_fops) {
5619 table->files[index] = file;
5623 for (i = 0; i < ctx->nr_user_files; i++) {
5624 file = io_file_from_index(ctx, i);
5628 for (i = 0; i < nr_tables; i++)
5629 kfree(ctx->file_data->table[i].files);
5631 kfree(ctx->file_data->table);
5632 kfree(ctx->file_data);
5633 ctx->file_data = NULL;
5634 ctx->nr_user_files = 0;
5638 ret = io_sqe_files_scm(ctx);
5640 io_sqe_files_unregister(ctx);
5645 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
5648 #if defined(CONFIG_UNIX)
5649 struct sock *sock = ctx->ring_sock->sk;
5650 struct sk_buff_head *head = &sock->sk_receive_queue;
5651 struct sk_buff *skb;
5654 * See if we can merge this file into an existing skb SCM_RIGHTS
5655 * file set. If there's no room, fall back to allocating a new skb
5656 * and filling it in.
5658 spin_lock_irq(&head->lock);
5659 skb = skb_peek(head);
5661 struct scm_fp_list *fpl = UNIXCB(skb).fp;
5663 if (fpl->count < SCM_MAX_FD) {
5664 __skb_unlink(skb, head);
5665 spin_unlock_irq(&head->lock);
5666 fpl->fp[fpl->count] = get_file(file);
5667 unix_inflight(fpl->user, fpl->fp[fpl->count]);
5669 spin_lock_irq(&head->lock);
5670 __skb_queue_head(head, skb);
5675 spin_unlock_irq(&head->lock);
5682 return __io_sqe_files_scm(ctx, 1, index);
5688 static void io_atomic_switch(struct percpu_ref *ref)
5690 struct fixed_file_data *data;
5692 data = container_of(ref, struct fixed_file_data, refs);
5693 clear_bit(FFD_F_ATOMIC, &data->state);
5696 static bool io_queue_file_removal(struct fixed_file_data *data,
5699 struct io_file_put *pfile, pfile_stack;
5700 DECLARE_COMPLETION_ONSTACK(done);
5703 * If we fail allocating the struct we need for doing async reomval
5704 * of this file, just punt to sync and wait for it.
5706 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
5708 pfile = &pfile_stack;
5709 pfile->done = &done;
5713 llist_add(&pfile->llist, &data->put_llist);
5715 if (pfile == &pfile_stack) {
5716 if (!test_and_set_bit(FFD_F_ATOMIC, &data->state)) {
5717 percpu_ref_put(&data->refs);
5718 percpu_ref_switch_to_atomic(&data->refs,
5721 wait_for_completion(&done);
5722 flush_work(&data->ref_work);
5729 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
5730 struct io_uring_files_update *up,
5733 struct fixed_file_data *data = ctx->file_data;
5734 bool ref_switch = false;
5740 if (check_add_overflow(up->offset, nr_args, &done))
5742 if (done > ctx->nr_user_files)
5746 fds = u64_to_user_ptr(up->fds);
5748 struct fixed_file_table *table;
5752 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
5756 i = array_index_nospec(up->offset, ctx->nr_user_files);
5757 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5758 index = i & IORING_FILE_TABLE_MASK;
5759 if (table->files[index]) {
5760 file = io_file_from_index(ctx, index);
5761 table->files[index] = NULL;
5762 if (io_queue_file_removal(data, file))
5772 * Don't allow io_uring instances to be registered. If
5773 * UNIX isn't enabled, then this causes a reference
5774 * cycle and this instance can never get freed. If UNIX
5775 * is enabled we'll handle it just fine, but there's
5776 * still no point in allowing a ring fd as it doesn't
5777 * support regular read/write anyway.
5779 if (file->f_op == &io_uring_fops) {
5784 table->files[index] = file;
5785 err = io_sqe_file_register(ctx, file, i);
5794 if (ref_switch && !test_and_set_bit(FFD_F_ATOMIC, &data->state)) {
5795 percpu_ref_put(&data->refs);
5796 percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch);
5799 return done ? done : err;
5801 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
5804 struct io_uring_files_update up;
5806 if (!ctx->file_data)
5810 if (copy_from_user(&up, arg, sizeof(up)))
5815 return __io_sqe_files_update(ctx, &up, nr_args);
5818 static void io_put_work(struct io_wq_work *work)
5820 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5825 static void io_get_work(struct io_wq_work *work)
5827 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5829 refcount_inc(&req->refs);
5832 static int io_init_wq_offload(struct io_ring_ctx *ctx,
5833 struct io_uring_params *p)
5835 struct io_wq_data data;
5837 struct io_ring_ctx *ctx_attach;
5838 unsigned int concurrency;
5841 data.user = ctx->user;
5842 data.get_work = io_get_work;
5843 data.put_work = io_put_work;
5845 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
5846 /* Do QD, or 4 * CPUS, whatever is smallest */
5847 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
5849 ctx->io_wq = io_wq_create(concurrency, &data);
5850 if (IS_ERR(ctx->io_wq)) {
5851 ret = PTR_ERR(ctx->io_wq);
5857 f = fdget(p->wq_fd);
5861 if (f.file->f_op != &io_uring_fops) {
5866 ctx_attach = f.file->private_data;
5867 /* @io_wq is protected by holding the fd */
5868 if (!io_wq_get(ctx_attach->io_wq, &data)) {
5873 ctx->io_wq = ctx_attach->io_wq;
5879 static int io_sq_offload_start(struct io_ring_ctx *ctx,
5880 struct io_uring_params *p)
5884 init_waitqueue_head(&ctx->sqo_wait);
5885 mmgrab(current->mm);
5886 ctx->sqo_mm = current->mm;
5888 if (ctx->flags & IORING_SETUP_SQPOLL) {
5890 if (!capable(CAP_SYS_ADMIN))
5893 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
5894 if (!ctx->sq_thread_idle)
5895 ctx->sq_thread_idle = HZ;
5897 if (p->flags & IORING_SETUP_SQ_AFF) {
5898 int cpu = p->sq_thread_cpu;
5901 if (cpu >= nr_cpu_ids)
5903 if (!cpu_online(cpu))
5906 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
5910 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
5913 if (IS_ERR(ctx->sqo_thread)) {
5914 ret = PTR_ERR(ctx->sqo_thread);
5915 ctx->sqo_thread = NULL;
5918 wake_up_process(ctx->sqo_thread);
5919 } else if (p->flags & IORING_SETUP_SQ_AFF) {
5920 /* Can't have SQ_AFF without SQPOLL */
5925 ret = io_init_wq_offload(ctx, p);
5931 io_finish_async(ctx);
5932 mmdrop(ctx->sqo_mm);
5937 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
5939 atomic_long_sub(nr_pages, &user->locked_vm);
5942 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
5944 unsigned long page_limit, cur_pages, new_pages;
5946 /* Don't allow more pages than we can safely lock */
5947 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
5950 cur_pages = atomic_long_read(&user->locked_vm);
5951 new_pages = cur_pages + nr_pages;
5952 if (new_pages > page_limit)
5954 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
5955 new_pages) != cur_pages);
5960 static void io_mem_free(void *ptr)
5967 page = virt_to_head_page(ptr);
5968 if (put_page_testzero(page))
5969 free_compound_page(page);
5972 static void *io_mem_alloc(size_t size)
5974 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
5977 return (void *) __get_free_pages(gfp_flags, get_order(size));
5980 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
5983 struct io_rings *rings;
5984 size_t off, sq_array_size;
5986 off = struct_size(rings, cqes, cq_entries);
5987 if (off == SIZE_MAX)
5991 off = ALIGN(off, SMP_CACHE_BYTES);
5996 sq_array_size = array_size(sizeof(u32), sq_entries);
5997 if (sq_array_size == SIZE_MAX)
6000 if (check_add_overflow(off, sq_array_size, &off))
6009 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6013 pages = (size_t)1 << get_order(
6014 rings_size(sq_entries, cq_entries, NULL));
6015 pages += (size_t)1 << get_order(
6016 array_size(sizeof(struct io_uring_sqe), sq_entries));
6021 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6025 if (!ctx->user_bufs)
6028 for (i = 0; i < ctx->nr_user_bufs; i++) {
6029 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6031 for (j = 0; j < imu->nr_bvecs; j++)
6032 put_user_page(imu->bvec[j].bv_page);
6034 if (ctx->account_mem)
6035 io_unaccount_mem(ctx->user, imu->nr_bvecs);
6040 kfree(ctx->user_bufs);
6041 ctx->user_bufs = NULL;
6042 ctx->nr_user_bufs = 0;
6046 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
6047 void __user *arg, unsigned index)
6049 struct iovec __user *src;
6051 #ifdef CONFIG_COMPAT
6053 struct compat_iovec __user *ciovs;
6054 struct compat_iovec ciov;
6056 ciovs = (struct compat_iovec __user *) arg;
6057 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
6060 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
6061 dst->iov_len = ciov.iov_len;
6065 src = (struct iovec __user *) arg;
6066 if (copy_from_user(dst, &src[index], sizeof(*dst)))
6071 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
6074 struct vm_area_struct **vmas = NULL;
6075 struct page **pages = NULL;
6076 int i, j, got_pages = 0;
6081 if (!nr_args || nr_args > UIO_MAXIOV)
6084 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
6086 if (!ctx->user_bufs)
6089 for (i = 0; i < nr_args; i++) {
6090 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6091 unsigned long off, start, end, ubuf;
6096 ret = io_copy_iov(ctx, &iov, arg, i);
6101 * Don't impose further limits on the size and buffer
6102 * constraints here, we'll -EINVAL later when IO is
6103 * submitted if they are wrong.
6106 if (!iov.iov_base || !iov.iov_len)
6109 /* arbitrary limit, but we need something */
6110 if (iov.iov_len > SZ_1G)
6113 ubuf = (unsigned long) iov.iov_base;
6114 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
6115 start = ubuf >> PAGE_SHIFT;
6116 nr_pages = end - start;
6118 if (ctx->account_mem) {
6119 ret = io_account_mem(ctx->user, nr_pages);
6125 if (!pages || nr_pages > got_pages) {
6128 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
6130 vmas = kvmalloc_array(nr_pages,
6131 sizeof(struct vm_area_struct *),
6133 if (!pages || !vmas) {
6135 if (ctx->account_mem)
6136 io_unaccount_mem(ctx->user, nr_pages);
6139 got_pages = nr_pages;
6142 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
6146 if (ctx->account_mem)
6147 io_unaccount_mem(ctx->user, nr_pages);
6152 down_read(¤t->mm->mmap_sem);
6153 pret = get_user_pages(ubuf, nr_pages,
6154 FOLL_WRITE | FOLL_LONGTERM,
6156 if (pret == nr_pages) {
6157 /* don't support file backed memory */
6158 for (j = 0; j < nr_pages; j++) {
6159 struct vm_area_struct *vma = vmas[j];
6162 !is_file_hugepages(vma->vm_file)) {
6168 ret = pret < 0 ? pret : -EFAULT;
6170 up_read(¤t->mm->mmap_sem);
6173 * if we did partial map, or found file backed vmas,
6174 * release any pages we did get
6177 put_user_pages(pages, pret);
6178 if (ctx->account_mem)
6179 io_unaccount_mem(ctx->user, nr_pages);
6184 off = ubuf & ~PAGE_MASK;
6186 for (j = 0; j < nr_pages; j++) {
6189 vec_len = min_t(size_t, size, PAGE_SIZE - off);
6190 imu->bvec[j].bv_page = pages[j];
6191 imu->bvec[j].bv_len = vec_len;
6192 imu->bvec[j].bv_offset = off;
6196 /* store original address for later verification */
6198 imu->len = iov.iov_len;
6199 imu->nr_bvecs = nr_pages;
6201 ctx->nr_user_bufs++;
6209 io_sqe_buffer_unregister(ctx);
6213 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
6215 __s32 __user *fds = arg;
6221 if (copy_from_user(&fd, fds, sizeof(*fds)))
6224 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
6225 if (IS_ERR(ctx->cq_ev_fd)) {
6226 int ret = PTR_ERR(ctx->cq_ev_fd);
6227 ctx->cq_ev_fd = NULL;
6234 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
6236 if (ctx->cq_ev_fd) {
6237 eventfd_ctx_put(ctx->cq_ev_fd);
6238 ctx->cq_ev_fd = NULL;
6245 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
6247 io_finish_async(ctx);
6249 mmdrop(ctx->sqo_mm);
6251 io_iopoll_reap_events(ctx);
6252 io_sqe_buffer_unregister(ctx);
6253 io_sqe_files_unregister(ctx);
6254 io_eventfd_unregister(ctx);
6256 #if defined(CONFIG_UNIX)
6257 if (ctx->ring_sock) {
6258 ctx->ring_sock->file = NULL; /* so that iput() is called */
6259 sock_release(ctx->ring_sock);
6263 io_mem_free(ctx->rings);
6264 io_mem_free(ctx->sq_sqes);
6266 percpu_ref_exit(&ctx->refs);
6267 if (ctx->account_mem)
6268 io_unaccount_mem(ctx->user,
6269 ring_pages(ctx->sq_entries, ctx->cq_entries));
6270 free_uid(ctx->user);
6271 put_cred(ctx->creds);
6272 kfree(ctx->completions);
6273 kfree(ctx->cancel_hash);
6274 kmem_cache_free(req_cachep, ctx->fallback_req);
6278 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
6280 struct io_ring_ctx *ctx = file->private_data;
6283 poll_wait(file, &ctx->cq_wait, wait);
6285 * synchronizes with barrier from wq_has_sleeper call in
6289 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
6290 ctx->rings->sq_ring_entries)
6291 mask |= EPOLLOUT | EPOLLWRNORM;
6292 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
6293 mask |= EPOLLIN | EPOLLRDNORM;
6298 static int io_uring_fasync(int fd, struct file *file, int on)
6300 struct io_ring_ctx *ctx = file->private_data;
6302 return fasync_helper(fd, file, on, &ctx->cq_fasync);
6305 static int io_remove_personalities(int id, void *p, void *data)
6307 struct io_ring_ctx *ctx = data;
6308 const struct cred *cred;
6310 cred = idr_remove(&ctx->personality_idr, id);
6316 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
6318 mutex_lock(&ctx->uring_lock);
6319 percpu_ref_kill(&ctx->refs);
6320 mutex_unlock(&ctx->uring_lock);
6322 io_kill_timeouts(ctx);
6323 io_poll_remove_all(ctx);
6326 io_wq_cancel_all(ctx->io_wq);
6328 io_iopoll_reap_events(ctx);
6329 /* if we failed setting up the ctx, we might not have any rings */
6331 io_cqring_overflow_flush(ctx, true);
6332 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
6333 wait_for_completion(&ctx->completions[0]);
6334 io_ring_ctx_free(ctx);
6337 static int io_uring_release(struct inode *inode, struct file *file)
6339 struct io_ring_ctx *ctx = file->private_data;
6341 file->private_data = NULL;
6342 io_ring_ctx_wait_and_kill(ctx);
6346 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
6347 struct files_struct *files)
6349 struct io_kiocb *req;
6352 while (!list_empty_careful(&ctx->inflight_list)) {
6353 struct io_kiocb *cancel_req = NULL;
6355 spin_lock_irq(&ctx->inflight_lock);
6356 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
6357 if (req->work.files != files)
6359 /* req is being completed, ignore */
6360 if (!refcount_inc_not_zero(&req->refs))
6366 prepare_to_wait(&ctx->inflight_wait, &wait,
6367 TASK_UNINTERRUPTIBLE);
6368 spin_unlock_irq(&ctx->inflight_lock);
6370 /* We need to keep going until we don't find a matching req */
6374 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
6375 io_put_req(cancel_req);
6378 finish_wait(&ctx->inflight_wait, &wait);
6381 static int io_uring_flush(struct file *file, void *data)
6383 struct io_ring_ctx *ctx = file->private_data;
6385 io_uring_cancel_files(ctx, data);
6389 static void *io_uring_validate_mmap_request(struct file *file,
6390 loff_t pgoff, size_t sz)
6392 struct io_ring_ctx *ctx = file->private_data;
6393 loff_t offset = pgoff << PAGE_SHIFT;
6398 case IORING_OFF_SQ_RING:
6399 case IORING_OFF_CQ_RING:
6402 case IORING_OFF_SQES:
6406 return ERR_PTR(-EINVAL);
6409 page = virt_to_head_page(ptr);
6410 if (sz > page_size(page))
6411 return ERR_PTR(-EINVAL);
6418 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6420 size_t sz = vma->vm_end - vma->vm_start;
6424 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
6426 return PTR_ERR(ptr);
6428 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
6429 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
6432 #else /* !CONFIG_MMU */
6434 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6436 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
6439 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
6441 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
6444 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
6445 unsigned long addr, unsigned long len,
6446 unsigned long pgoff, unsigned long flags)
6450 ptr = io_uring_validate_mmap_request(file, pgoff, len);
6452 return PTR_ERR(ptr);
6454 return (unsigned long) ptr;
6457 #endif /* !CONFIG_MMU */
6459 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
6460 u32, min_complete, u32, flags, const sigset_t __user *, sig,
6463 struct io_ring_ctx *ctx;
6468 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
6476 if (f.file->f_op != &io_uring_fops)
6480 ctx = f.file->private_data;
6481 if (!percpu_ref_tryget(&ctx->refs))
6485 * For SQ polling, the thread will do all submissions and completions.
6486 * Just return the requested submit count, and wake the thread if
6490 if (ctx->flags & IORING_SETUP_SQPOLL) {
6491 if (!list_empty_careful(&ctx->cq_overflow_list))
6492 io_cqring_overflow_flush(ctx, false);
6493 if (flags & IORING_ENTER_SQ_WAKEUP)
6494 wake_up(&ctx->sqo_wait);
6495 submitted = to_submit;
6496 } else if (to_submit) {
6497 struct mm_struct *cur_mm;
6499 mutex_lock(&ctx->uring_lock);
6500 /* already have mm, so io_submit_sqes() won't try to grab it */
6501 cur_mm = ctx->sqo_mm;
6502 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
6504 mutex_unlock(&ctx->uring_lock);
6506 if (submitted != to_submit)
6509 if (flags & IORING_ENTER_GETEVENTS) {
6510 unsigned nr_events = 0;
6512 min_complete = min(min_complete, ctx->cq_entries);
6514 if (ctx->flags & IORING_SETUP_IOPOLL) {
6515 ret = io_iopoll_check(ctx, &nr_events, min_complete);
6517 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
6522 percpu_ref_put(&ctx->refs);
6525 return submitted ? submitted : ret;
6528 static int io_uring_show_cred(int id, void *p, void *data)
6530 const struct cred *cred = p;
6531 struct seq_file *m = data;
6532 struct user_namespace *uns = seq_user_ns(m);
6533 struct group_info *gi;
6538 seq_printf(m, "%5d\n", id);
6539 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
6540 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
6541 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
6542 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
6543 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
6544 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
6545 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
6546 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
6547 seq_puts(m, "\n\tGroups:\t");
6548 gi = cred->group_info;
6549 for (g = 0; g < gi->ngroups; g++) {
6550 seq_put_decimal_ull(m, g ? " " : "",
6551 from_kgid_munged(uns, gi->gid[g]));
6553 seq_puts(m, "\n\tCapEff:\t");
6554 cap = cred->cap_effective;
6555 CAP_FOR_EACH_U32(__capi)
6556 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
6561 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
6565 mutex_lock(&ctx->uring_lock);
6566 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
6567 for (i = 0; i < ctx->nr_user_files; i++) {
6568 struct fixed_file_table *table;
6571 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6572 f = table->files[i & IORING_FILE_TABLE_MASK];
6574 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
6576 seq_printf(m, "%5u: <none>\n", i);
6578 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
6579 for (i = 0; i < ctx->nr_user_bufs; i++) {
6580 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
6582 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
6583 (unsigned int) buf->len);
6585 if (!idr_is_empty(&ctx->personality_idr)) {
6586 seq_printf(m, "Personalities:\n");
6587 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
6589 mutex_unlock(&ctx->uring_lock);
6592 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
6594 struct io_ring_ctx *ctx = f->private_data;
6596 if (percpu_ref_tryget(&ctx->refs)) {
6597 __io_uring_show_fdinfo(ctx, m);
6598 percpu_ref_put(&ctx->refs);
6602 static const struct file_operations io_uring_fops = {
6603 .release = io_uring_release,
6604 .flush = io_uring_flush,
6605 .mmap = io_uring_mmap,
6607 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
6608 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
6610 .poll = io_uring_poll,
6611 .fasync = io_uring_fasync,
6612 .show_fdinfo = io_uring_show_fdinfo,
6615 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
6616 struct io_uring_params *p)
6618 struct io_rings *rings;
6619 size_t size, sq_array_offset;
6621 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
6622 if (size == SIZE_MAX)
6625 rings = io_mem_alloc(size);
6630 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
6631 rings->sq_ring_mask = p->sq_entries - 1;
6632 rings->cq_ring_mask = p->cq_entries - 1;
6633 rings->sq_ring_entries = p->sq_entries;
6634 rings->cq_ring_entries = p->cq_entries;
6635 ctx->sq_mask = rings->sq_ring_mask;
6636 ctx->cq_mask = rings->cq_ring_mask;
6637 ctx->sq_entries = rings->sq_ring_entries;
6638 ctx->cq_entries = rings->cq_ring_entries;
6640 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
6641 if (size == SIZE_MAX) {
6642 io_mem_free(ctx->rings);
6647 ctx->sq_sqes = io_mem_alloc(size);
6648 if (!ctx->sq_sqes) {
6649 io_mem_free(ctx->rings);
6658 * Allocate an anonymous fd, this is what constitutes the application
6659 * visible backing of an io_uring instance. The application mmaps this
6660 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
6661 * we have to tie this fd to a socket for file garbage collection purposes.
6663 static int io_uring_get_fd(struct io_ring_ctx *ctx)
6668 #if defined(CONFIG_UNIX)
6669 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
6675 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
6679 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
6680 O_RDWR | O_CLOEXEC);
6683 ret = PTR_ERR(file);
6687 #if defined(CONFIG_UNIX)
6688 ctx->ring_sock->file = file;
6690 fd_install(ret, file);
6693 #if defined(CONFIG_UNIX)
6694 sock_release(ctx->ring_sock);
6695 ctx->ring_sock = NULL;
6700 static int io_uring_create(unsigned entries, struct io_uring_params *p)
6702 struct user_struct *user = NULL;
6703 struct io_ring_ctx *ctx;
6709 if (entries > IORING_MAX_ENTRIES) {
6710 if (!(p->flags & IORING_SETUP_CLAMP))
6712 entries = IORING_MAX_ENTRIES;
6716 * Use twice as many entries for the CQ ring. It's possible for the
6717 * application to drive a higher depth than the size of the SQ ring,
6718 * since the sqes are only used at submission time. This allows for
6719 * some flexibility in overcommitting a bit. If the application has
6720 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
6721 * of CQ ring entries manually.
6723 p->sq_entries = roundup_pow_of_two(entries);
6724 if (p->flags & IORING_SETUP_CQSIZE) {
6726 * If IORING_SETUP_CQSIZE is set, we do the same roundup
6727 * to a power-of-two, if it isn't already. We do NOT impose
6728 * any cq vs sq ring sizing.
6730 if (p->cq_entries < p->sq_entries)
6732 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
6733 if (!(p->flags & IORING_SETUP_CLAMP))
6735 p->cq_entries = IORING_MAX_CQ_ENTRIES;
6737 p->cq_entries = roundup_pow_of_two(p->cq_entries);
6739 p->cq_entries = 2 * p->sq_entries;
6742 user = get_uid(current_user());
6743 account_mem = !capable(CAP_IPC_LOCK);
6746 ret = io_account_mem(user,
6747 ring_pages(p->sq_entries, p->cq_entries));
6754 ctx = io_ring_ctx_alloc(p);
6757 io_unaccount_mem(user, ring_pages(p->sq_entries,
6762 ctx->compat = in_compat_syscall();
6763 ctx->account_mem = account_mem;
6765 ctx->creds = get_current_cred();
6767 ret = io_allocate_scq_urings(ctx, p);
6771 ret = io_sq_offload_start(ctx, p);
6775 memset(&p->sq_off, 0, sizeof(p->sq_off));
6776 p->sq_off.head = offsetof(struct io_rings, sq.head);
6777 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
6778 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
6779 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
6780 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
6781 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
6782 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
6784 memset(&p->cq_off, 0, sizeof(p->cq_off));
6785 p->cq_off.head = offsetof(struct io_rings, cq.head);
6786 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
6787 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
6788 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
6789 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
6790 p->cq_off.cqes = offsetof(struct io_rings, cqes);
6793 * Install ring fd as the very last thing, so we don't risk someone
6794 * having closed it before we finish setup
6796 ret = io_uring_get_fd(ctx);
6800 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
6801 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
6802 IORING_FEAT_CUR_PERSONALITY;
6803 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
6806 io_ring_ctx_wait_and_kill(ctx);
6811 * Sets up an aio uring context, and returns the fd. Applications asks for a
6812 * ring size, we return the actual sq/cq ring sizes (among other things) in the
6813 * params structure passed in.
6815 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
6817 struct io_uring_params p;
6821 if (copy_from_user(&p, params, sizeof(p)))
6823 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
6828 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
6829 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
6830 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
6833 ret = io_uring_create(entries, &p);
6837 if (copy_to_user(params, &p, sizeof(p)))
6843 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
6844 struct io_uring_params __user *, params)
6846 return io_uring_setup(entries, params);
6849 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
6851 struct io_uring_probe *p;
6855 size = struct_size(p, ops, nr_args);
6856 if (size == SIZE_MAX)
6858 p = kzalloc(size, GFP_KERNEL);
6863 if (copy_from_user(p, arg, size))
6866 if (memchr_inv(p, 0, size))
6869 p->last_op = IORING_OP_LAST - 1;
6870 if (nr_args > IORING_OP_LAST)
6871 nr_args = IORING_OP_LAST;
6873 for (i = 0; i < nr_args; i++) {
6875 if (!io_op_defs[i].not_supported)
6876 p->ops[i].flags = IO_URING_OP_SUPPORTED;
6881 if (copy_to_user(arg, p, size))
6888 static int io_register_personality(struct io_ring_ctx *ctx)
6890 const struct cred *creds = get_current_cred();
6893 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
6894 USHRT_MAX, GFP_KERNEL);
6900 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
6902 const struct cred *old_creds;
6904 old_creds = idr_remove(&ctx->personality_idr, id);
6906 put_cred(old_creds);
6913 static bool io_register_op_must_quiesce(int op)
6916 case IORING_UNREGISTER_FILES:
6917 case IORING_REGISTER_FILES_UPDATE:
6918 case IORING_REGISTER_PROBE:
6919 case IORING_REGISTER_PERSONALITY:
6920 case IORING_UNREGISTER_PERSONALITY:
6927 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
6928 void __user *arg, unsigned nr_args)
6929 __releases(ctx->uring_lock)
6930 __acquires(ctx->uring_lock)
6935 * We're inside the ring mutex, if the ref is already dying, then
6936 * someone else killed the ctx or is already going through
6937 * io_uring_register().
6939 if (percpu_ref_is_dying(&ctx->refs))
6942 if (io_register_op_must_quiesce(opcode)) {
6943 percpu_ref_kill(&ctx->refs);
6946 * Drop uring mutex before waiting for references to exit. If
6947 * another thread is currently inside io_uring_enter() it might
6948 * need to grab the uring_lock to make progress. If we hold it
6949 * here across the drain wait, then we can deadlock. It's safe
6950 * to drop the mutex here, since no new references will come in
6951 * after we've killed the percpu ref.
6953 mutex_unlock(&ctx->uring_lock);
6954 ret = wait_for_completion_interruptible(&ctx->completions[0]);
6955 mutex_lock(&ctx->uring_lock);
6957 percpu_ref_resurrect(&ctx->refs);
6964 case IORING_REGISTER_BUFFERS:
6965 ret = io_sqe_buffer_register(ctx, arg, nr_args);
6967 case IORING_UNREGISTER_BUFFERS:
6971 ret = io_sqe_buffer_unregister(ctx);
6973 case IORING_REGISTER_FILES:
6974 ret = io_sqe_files_register(ctx, arg, nr_args);
6976 case IORING_UNREGISTER_FILES:
6980 ret = io_sqe_files_unregister(ctx);
6982 case IORING_REGISTER_FILES_UPDATE:
6983 ret = io_sqe_files_update(ctx, arg, nr_args);
6985 case IORING_REGISTER_EVENTFD:
6986 case IORING_REGISTER_EVENTFD_ASYNC:
6990 ret = io_eventfd_register(ctx, arg);
6993 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
6994 ctx->eventfd_async = 1;
6996 ctx->eventfd_async = 0;
6998 case IORING_UNREGISTER_EVENTFD:
7002 ret = io_eventfd_unregister(ctx);
7004 case IORING_REGISTER_PROBE:
7006 if (!arg || nr_args > 256)
7008 ret = io_probe(ctx, arg, nr_args);
7010 case IORING_REGISTER_PERSONALITY:
7014 ret = io_register_personality(ctx);
7016 case IORING_UNREGISTER_PERSONALITY:
7020 ret = io_unregister_personality(ctx, nr_args);
7027 if (io_register_op_must_quiesce(opcode)) {
7028 /* bring the ctx back to life */
7029 percpu_ref_reinit(&ctx->refs);
7031 reinit_completion(&ctx->completions[0]);
7036 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
7037 void __user *, arg, unsigned int, nr_args)
7039 struct io_ring_ctx *ctx;
7048 if (f.file->f_op != &io_uring_fops)
7051 ctx = f.file->private_data;
7053 mutex_lock(&ctx->uring_lock);
7054 ret = __io_uring_register(ctx, opcode, arg, nr_args);
7055 mutex_unlock(&ctx->uring_lock);
7056 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
7057 ctx->cq_ev_fd != NULL, ret);
7063 static int __init io_uring_init(void)
7065 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
7066 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
7067 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
7070 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
7071 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
7072 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
7073 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
7074 BUILD_BUG_SQE_ELEM(1, __u8, flags);
7075 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
7076 BUILD_BUG_SQE_ELEM(4, __s32, fd);
7077 BUILD_BUG_SQE_ELEM(8, __u64, off);
7078 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
7079 BUILD_BUG_SQE_ELEM(16, __u64, addr);
7080 BUILD_BUG_SQE_ELEM(24, __u32, len);
7081 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
7082 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
7083 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
7084 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
7085 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
7086 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
7087 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
7088 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
7089 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
7090 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
7091 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
7092 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
7093 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
7094 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
7095 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
7096 BUILD_BUG_SQE_ELEM(42, __u16, personality);
7098 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
7099 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
7102 __initcall(io_uring_init);