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;
1868 if (S_ISREG(file_inode(req->file)->i_mode))
1869 req->flags |= REQ_F_ISREG;
1871 kiocb->ki_pos = READ_ONCE(sqe->off);
1872 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
1873 req->flags |= REQ_F_CUR_POS;
1874 kiocb->ki_pos = req->file->f_pos;
1876 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1877 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1879 ioprio = READ_ONCE(sqe->ioprio);
1881 ret = ioprio_check_cap(ioprio);
1885 kiocb->ki_ioprio = ioprio;
1887 kiocb->ki_ioprio = get_current_ioprio();
1889 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1893 /* don't allow async punt if RWF_NOWAIT was requested */
1894 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1895 (req->file->f_flags & O_NONBLOCK))
1896 req->flags |= REQ_F_NOWAIT;
1899 kiocb->ki_flags |= IOCB_NOWAIT;
1901 if (ctx->flags & IORING_SETUP_IOPOLL) {
1902 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1903 !kiocb->ki_filp->f_op->iopoll)
1906 kiocb->ki_flags |= IOCB_HIPRI;
1907 kiocb->ki_complete = io_complete_rw_iopoll;
1910 if (kiocb->ki_flags & IOCB_HIPRI)
1912 kiocb->ki_complete = io_complete_rw;
1915 req->rw.addr = READ_ONCE(sqe->addr);
1916 req->rw.len = READ_ONCE(sqe->len);
1917 /* we own ->private, reuse it for the buffer index */
1918 req->rw.kiocb.private = (void *) (unsigned long)
1919 READ_ONCE(sqe->buf_index);
1923 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1929 case -ERESTARTNOINTR:
1930 case -ERESTARTNOHAND:
1931 case -ERESTART_RESTARTBLOCK:
1933 * We can't just restart the syscall, since previously
1934 * submitted sqes may already be in progress. Just fail this
1940 kiocb->ki_complete(kiocb, ret, 0);
1944 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1947 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1949 if (req->flags & REQ_F_CUR_POS)
1950 req->file->f_pos = kiocb->ki_pos;
1951 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1952 *nxt = __io_complete_rw(kiocb, ret);
1954 io_rw_done(kiocb, ret);
1957 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
1958 struct iov_iter *iter)
1960 struct io_ring_ctx *ctx = req->ctx;
1961 size_t len = req->rw.len;
1962 struct io_mapped_ubuf *imu;
1963 unsigned index, buf_index;
1967 /* attempt to use fixed buffers without having provided iovecs */
1968 if (unlikely(!ctx->user_bufs))
1971 buf_index = (unsigned long) req->rw.kiocb.private;
1972 if (unlikely(buf_index >= ctx->nr_user_bufs))
1975 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1976 imu = &ctx->user_bufs[index];
1977 buf_addr = req->rw.addr;
1980 if (buf_addr + len < buf_addr)
1982 /* not inside the mapped region */
1983 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1987 * May not be a start of buffer, set size appropriately
1988 * and advance us to the beginning.
1990 offset = buf_addr - imu->ubuf;
1991 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1995 * Don't use iov_iter_advance() here, as it's really slow for
1996 * using the latter parts of a big fixed buffer - it iterates
1997 * over each segment manually. We can cheat a bit here, because
2000 * 1) it's a BVEC iter, we set it up
2001 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2002 * first and last bvec
2004 * So just find our index, and adjust the iterator afterwards.
2005 * If the offset is within the first bvec (or the whole first
2006 * bvec, just use iov_iter_advance(). This makes it easier
2007 * since we can just skip the first segment, which may not
2008 * be PAGE_SIZE aligned.
2010 const struct bio_vec *bvec = imu->bvec;
2012 if (offset <= bvec->bv_len) {
2013 iov_iter_advance(iter, offset);
2015 unsigned long seg_skip;
2017 /* skip first vec */
2018 offset -= bvec->bv_len;
2019 seg_skip = 1 + (offset >> PAGE_SHIFT);
2021 iter->bvec = bvec + seg_skip;
2022 iter->nr_segs -= seg_skip;
2023 iter->count -= bvec->bv_len + offset;
2024 iter->iov_offset = offset & ~PAGE_MASK;
2031 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2032 struct iovec **iovec, struct iov_iter *iter)
2034 void __user *buf = u64_to_user_ptr(req->rw.addr);
2035 size_t sqe_len = req->rw.len;
2038 opcode = req->opcode;
2039 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2041 return io_import_fixed(req, rw, iter);
2044 /* buffer index only valid with fixed read/write */
2045 if (req->rw.kiocb.private)
2048 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2050 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2056 struct io_async_rw *iorw = &req->io->rw;
2059 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2060 if (iorw->iov == iorw->fast_iov)
2068 #ifdef CONFIG_COMPAT
2069 if (req->ctx->compat)
2070 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2074 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2078 * For files that don't have ->read_iter() and ->write_iter(), handle them
2079 * by looping over ->read() or ->write() manually.
2081 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2082 struct iov_iter *iter)
2087 * Don't support polled IO through this interface, and we can't
2088 * support non-blocking either. For the latter, this just causes
2089 * the kiocb to be handled from an async context.
2091 if (kiocb->ki_flags & IOCB_HIPRI)
2093 if (kiocb->ki_flags & IOCB_NOWAIT)
2096 while (iov_iter_count(iter)) {
2100 if (!iov_iter_is_bvec(iter)) {
2101 iovec = iov_iter_iovec(iter);
2103 /* fixed buffers import bvec */
2104 iovec.iov_base = kmap(iter->bvec->bv_page)
2106 iovec.iov_len = min(iter->count,
2107 iter->bvec->bv_len - iter->iov_offset);
2111 nr = file->f_op->read(file, iovec.iov_base,
2112 iovec.iov_len, &kiocb->ki_pos);
2114 nr = file->f_op->write(file, iovec.iov_base,
2115 iovec.iov_len, &kiocb->ki_pos);
2118 if (iov_iter_is_bvec(iter))
2119 kunmap(iter->bvec->bv_page);
2127 if (nr != iovec.iov_len)
2129 iov_iter_advance(iter, nr);
2135 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2136 struct iovec *iovec, struct iovec *fast_iov,
2137 struct iov_iter *iter)
2139 req->io->rw.nr_segs = iter->nr_segs;
2140 req->io->rw.size = io_size;
2141 req->io->rw.iov = iovec;
2142 if (!req->io->rw.iov) {
2143 req->io->rw.iov = req->io->rw.fast_iov;
2144 memcpy(req->io->rw.iov, fast_iov,
2145 sizeof(struct iovec) * iter->nr_segs);
2149 static int io_alloc_async_ctx(struct io_kiocb *req)
2151 if (!io_op_defs[req->opcode].async_ctx)
2153 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2154 return req->io == NULL;
2157 static void io_rw_async(struct io_wq_work **workptr)
2159 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2160 struct iovec *iov = NULL;
2162 if (req->io->rw.iov != req->io->rw.fast_iov)
2163 iov = req->io->rw.iov;
2164 io_wq_submit_work(workptr);
2168 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2169 struct iovec *iovec, struct iovec *fast_iov,
2170 struct iov_iter *iter)
2172 if (!io_op_defs[req->opcode].async_ctx)
2175 if (io_alloc_async_ctx(req))
2178 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2180 req->work.func = io_rw_async;
2184 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2185 bool force_nonblock)
2187 struct io_async_ctx *io;
2188 struct iov_iter iter;
2191 ret = io_prep_rw(req, sqe, force_nonblock);
2195 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2202 io->rw.iov = io->rw.fast_iov;
2204 ret = io_import_iovec(READ, req, &io->rw.iov, &iter);
2209 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2213 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
2214 bool force_nonblock)
2216 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2217 struct kiocb *kiocb = &req->rw.kiocb;
2218 struct iov_iter iter;
2220 ssize_t io_size, ret;
2222 ret = io_import_iovec(READ, req, &iovec, &iter);
2226 /* Ensure we clear previously set non-block flag */
2227 if (!force_nonblock)
2228 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2232 if (req->flags & REQ_F_LINK)
2233 req->result = io_size;
2236 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2237 * we know to async punt it even if it was opened O_NONBLOCK
2239 if (force_nonblock && !io_file_supports_async(req->file)) {
2240 req->flags |= REQ_F_MUST_PUNT;
2244 iov_count = iov_iter_count(&iter);
2245 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2249 if (req->file->f_op->read_iter)
2250 ret2 = call_read_iter(req->file, kiocb, &iter);
2252 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2254 /* Catch -EAGAIN return for forced non-blocking submission */
2255 if (!force_nonblock || ret2 != -EAGAIN) {
2256 kiocb_done(kiocb, ret2, nxt, req->in_async);
2259 ret = io_setup_async_rw(req, io_size, iovec,
2260 inline_vecs, &iter);
2267 if (!io_wq_current_is_worker())
2272 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2273 bool force_nonblock)
2275 struct io_async_ctx *io;
2276 struct iov_iter iter;
2279 ret = io_prep_rw(req, sqe, force_nonblock);
2283 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2290 io->rw.iov = io->rw.fast_iov;
2292 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter);
2297 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2301 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
2302 bool force_nonblock)
2304 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2305 struct kiocb *kiocb = &req->rw.kiocb;
2306 struct iov_iter iter;
2308 ssize_t ret, io_size;
2310 ret = io_import_iovec(WRITE, req, &iovec, &iter);
2314 /* Ensure we clear previously set non-block flag */
2315 if (!force_nonblock)
2316 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2320 if (req->flags & REQ_F_LINK)
2321 req->result = io_size;
2324 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2325 * we know to async punt it even if it was opened O_NONBLOCK
2327 if (force_nonblock && !io_file_supports_async(req->file)) {
2328 req->flags |= REQ_F_MUST_PUNT;
2332 /* file path doesn't support NOWAIT for non-direct_IO */
2333 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2334 (req->flags & REQ_F_ISREG))
2337 iov_count = iov_iter_count(&iter);
2338 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2343 * Open-code file_start_write here to grab freeze protection,
2344 * which will be released by another thread in
2345 * io_complete_rw(). Fool lockdep by telling it the lock got
2346 * released so that it doesn't complain about the held lock when
2347 * we return to userspace.
2349 if (req->flags & REQ_F_ISREG) {
2350 __sb_start_write(file_inode(req->file)->i_sb,
2351 SB_FREEZE_WRITE, true);
2352 __sb_writers_release(file_inode(req->file)->i_sb,
2355 kiocb->ki_flags |= IOCB_WRITE;
2357 if (req->file->f_op->write_iter)
2358 ret2 = call_write_iter(req->file, kiocb, &iter);
2360 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2361 if (!force_nonblock || ret2 != -EAGAIN) {
2362 kiocb_done(kiocb, ret2, nxt, req->in_async);
2365 ret = io_setup_async_rw(req, io_size, iovec,
2366 inline_vecs, &iter);
2373 if (!io_wq_current_is_worker())
2379 * IORING_OP_NOP just posts a completion event, nothing else.
2381 static int io_nop(struct io_kiocb *req)
2383 struct io_ring_ctx *ctx = req->ctx;
2385 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2388 io_cqring_add_event(req, 0);
2393 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2395 struct io_ring_ctx *ctx = req->ctx;
2400 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2402 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2405 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2406 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2409 req->sync.off = READ_ONCE(sqe->off);
2410 req->sync.len = READ_ONCE(sqe->len);
2414 static bool io_req_cancelled(struct io_kiocb *req)
2416 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2417 req_set_fail_links(req);
2418 io_cqring_add_event(req, -ECANCELED);
2426 static void io_link_work_cb(struct io_wq_work **workptr)
2428 struct io_wq_work *work = *workptr;
2429 struct io_kiocb *link = work->data;
2431 io_queue_linked_timeout(link);
2432 work->func = io_wq_submit_work;
2435 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
2437 struct io_kiocb *link;
2439 io_prep_async_work(nxt, &link);
2440 *workptr = &nxt->work;
2442 nxt->work.flags |= IO_WQ_WORK_CB;
2443 nxt->work.func = io_link_work_cb;
2444 nxt->work.data = link;
2448 static void io_fsync_finish(struct io_wq_work **workptr)
2450 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2451 loff_t end = req->sync.off + req->sync.len;
2452 struct io_kiocb *nxt = NULL;
2455 if (io_req_cancelled(req))
2458 ret = vfs_fsync_range(req->file, req->sync.off,
2459 end > 0 ? end : LLONG_MAX,
2460 req->sync.flags & IORING_FSYNC_DATASYNC);
2462 req_set_fail_links(req);
2463 io_cqring_add_event(req, ret);
2464 io_put_req_find_next(req, &nxt);
2466 io_wq_assign_next(workptr, nxt);
2469 static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt,
2470 bool force_nonblock)
2472 struct io_wq_work *work, *old_work;
2474 /* fsync always requires a blocking context */
2475 if (force_nonblock) {
2477 req->work.func = io_fsync_finish;
2481 work = old_work = &req->work;
2482 io_fsync_finish(&work);
2483 if (work && work != old_work)
2484 *nxt = container_of(work, struct io_kiocb, work);
2488 static void io_fallocate_finish(struct io_wq_work **workptr)
2490 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2491 struct io_kiocb *nxt = NULL;
2494 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2497 req_set_fail_links(req);
2498 io_cqring_add_event(req, ret);
2499 io_put_req_find_next(req, &nxt);
2501 io_wq_assign_next(workptr, nxt);
2504 static int io_fallocate_prep(struct io_kiocb *req,
2505 const struct io_uring_sqe *sqe)
2507 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2510 req->sync.off = READ_ONCE(sqe->off);
2511 req->sync.len = READ_ONCE(sqe->addr);
2512 req->sync.mode = READ_ONCE(sqe->len);
2516 static int io_fallocate(struct io_kiocb *req, struct io_kiocb **nxt,
2517 bool force_nonblock)
2519 struct io_wq_work *work, *old_work;
2521 /* fallocate always requiring blocking context */
2522 if (force_nonblock) {
2524 req->work.func = io_fallocate_finish;
2528 work = old_work = &req->work;
2529 io_fallocate_finish(&work);
2530 if (work && work != old_work)
2531 *nxt = container_of(work, struct io_kiocb, work);
2536 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2538 const char __user *fname;
2541 if (sqe->ioprio || sqe->buf_index)
2544 req->open.dfd = READ_ONCE(sqe->fd);
2545 req->open.how.mode = READ_ONCE(sqe->len);
2546 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2547 req->open.how.flags = READ_ONCE(sqe->open_flags);
2549 req->open.filename = getname(fname);
2550 if (IS_ERR(req->open.filename)) {
2551 ret = PTR_ERR(req->open.filename);
2552 req->open.filename = NULL;
2559 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2561 struct open_how __user *how;
2562 const char __user *fname;
2566 if (sqe->ioprio || sqe->buf_index)
2569 req->open.dfd = READ_ONCE(sqe->fd);
2570 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2571 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2572 len = READ_ONCE(sqe->len);
2574 if (len < OPEN_HOW_SIZE_VER0)
2577 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2582 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2583 req->open.how.flags |= O_LARGEFILE;
2585 req->open.filename = getname(fname);
2586 if (IS_ERR(req->open.filename)) {
2587 ret = PTR_ERR(req->open.filename);
2588 req->open.filename = NULL;
2595 static int io_openat2(struct io_kiocb *req, struct io_kiocb **nxt,
2596 bool force_nonblock)
2598 struct open_flags op;
2605 ret = build_open_flags(&req->open.how, &op);
2609 ret = get_unused_fd_flags(req->open.how.flags);
2613 file = do_filp_open(req->open.dfd, req->open.filename, &op);
2616 ret = PTR_ERR(file);
2618 fsnotify_open(file);
2619 fd_install(ret, file);
2622 putname(req->open.filename);
2624 req_set_fail_links(req);
2625 io_cqring_add_event(req, ret);
2626 io_put_req_find_next(req, nxt);
2630 static int io_openat(struct io_kiocb *req, struct io_kiocb **nxt,
2631 bool force_nonblock)
2633 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
2634 return io_openat2(req, nxt, force_nonblock);
2637 static int io_epoll_ctl_prep(struct io_kiocb *req,
2638 const struct io_uring_sqe *sqe)
2640 #if defined(CONFIG_EPOLL)
2641 if (sqe->ioprio || sqe->buf_index)
2644 req->epoll.epfd = READ_ONCE(sqe->fd);
2645 req->epoll.op = READ_ONCE(sqe->len);
2646 req->epoll.fd = READ_ONCE(sqe->off);
2648 if (ep_op_has_event(req->epoll.op)) {
2649 struct epoll_event __user *ev;
2651 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
2652 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
2662 static int io_epoll_ctl(struct io_kiocb *req, struct io_kiocb **nxt,
2663 bool force_nonblock)
2665 #if defined(CONFIG_EPOLL)
2666 struct io_epoll *ie = &req->epoll;
2669 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
2670 if (force_nonblock && ret == -EAGAIN)
2674 req_set_fail_links(req);
2675 io_cqring_add_event(req, ret);
2676 io_put_req_find_next(req, nxt);
2683 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2685 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2686 if (sqe->ioprio || sqe->buf_index || sqe->off)
2689 req->madvise.addr = READ_ONCE(sqe->addr);
2690 req->madvise.len = READ_ONCE(sqe->len);
2691 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
2698 static int io_madvise(struct io_kiocb *req, struct io_kiocb **nxt,
2699 bool force_nonblock)
2701 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2702 struct io_madvise *ma = &req->madvise;
2708 ret = do_madvise(ma->addr, ma->len, ma->advice);
2710 req_set_fail_links(req);
2711 io_cqring_add_event(req, ret);
2712 io_put_req_find_next(req, nxt);
2719 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2721 if (sqe->ioprio || sqe->buf_index || sqe->addr)
2724 req->fadvise.offset = READ_ONCE(sqe->off);
2725 req->fadvise.len = READ_ONCE(sqe->len);
2726 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
2730 static int io_fadvise(struct io_kiocb *req, struct io_kiocb **nxt,
2731 bool force_nonblock)
2733 struct io_fadvise *fa = &req->fadvise;
2736 /* DONTNEED may block, others _should_ not */
2737 if (fa->advice == POSIX_FADV_DONTNEED && force_nonblock)
2740 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
2742 req_set_fail_links(req);
2743 io_cqring_add_event(req, ret);
2744 io_put_req_find_next(req, nxt);
2748 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2750 const char __user *fname;
2751 unsigned lookup_flags;
2754 if (sqe->ioprio || sqe->buf_index)
2757 req->open.dfd = READ_ONCE(sqe->fd);
2758 req->open.mask = READ_ONCE(sqe->len);
2759 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2760 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2761 req->open.how.flags = READ_ONCE(sqe->statx_flags);
2763 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
2766 req->open.filename = getname_flags(fname, lookup_flags, NULL);
2767 if (IS_ERR(req->open.filename)) {
2768 ret = PTR_ERR(req->open.filename);
2769 req->open.filename = NULL;
2776 static int io_statx(struct io_kiocb *req, struct io_kiocb **nxt,
2777 bool force_nonblock)
2779 struct io_open *ctx = &req->open;
2780 unsigned lookup_flags;
2788 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
2792 /* filename_lookup() drops it, keep a reference */
2793 ctx->filename->refcnt++;
2795 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
2800 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
2802 if (retry_estale(ret, lookup_flags)) {
2803 lookup_flags |= LOOKUP_REVAL;
2807 ret = cp_statx(&stat, ctx->buffer);
2809 putname(ctx->filename);
2811 req_set_fail_links(req);
2812 io_cqring_add_event(req, ret);
2813 io_put_req_find_next(req, nxt);
2817 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2820 * If we queue this for async, it must not be cancellable. That would
2821 * leave the 'file' in an undeterminate state.
2823 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
2825 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
2826 sqe->rw_flags || sqe->buf_index)
2828 if (sqe->flags & IOSQE_FIXED_FILE)
2831 req->close.fd = READ_ONCE(sqe->fd);
2832 if (req->file->f_op == &io_uring_fops ||
2833 req->close.fd == req->ctx->ring_fd)
2839 static void io_close_finish(struct io_wq_work **workptr)
2841 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2842 struct io_kiocb *nxt = NULL;
2844 /* Invoked with files, we need to do the close */
2845 if (req->work.files) {
2848 ret = filp_close(req->close.put_file, req->work.files);
2850 req_set_fail_links(req);
2852 io_cqring_add_event(req, ret);
2855 fput(req->close.put_file);
2857 /* we bypassed the re-issue, drop the submission reference */
2859 io_put_req_find_next(req, &nxt);
2861 io_wq_assign_next(workptr, nxt);
2864 static int io_close(struct io_kiocb *req, struct io_kiocb **nxt,
2865 bool force_nonblock)
2869 req->close.put_file = NULL;
2870 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
2874 /* if the file has a flush method, be safe and punt to async */
2875 if (req->close.put_file->f_op->flush && !io_wq_current_is_worker())
2879 * No ->flush(), safely close from here and just punt the
2880 * fput() to async context.
2882 ret = filp_close(req->close.put_file, current->files);
2885 req_set_fail_links(req);
2886 io_cqring_add_event(req, ret);
2888 if (io_wq_current_is_worker()) {
2889 struct io_wq_work *old_work, *work;
2891 old_work = work = &req->work;
2892 io_close_finish(&work);
2893 if (work && work != old_work)
2894 *nxt = container_of(work, struct io_kiocb, work);
2899 req->work.func = io_close_finish;
2903 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2905 struct io_ring_ctx *ctx = req->ctx;
2910 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2912 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2915 req->sync.off = READ_ONCE(sqe->off);
2916 req->sync.len = READ_ONCE(sqe->len);
2917 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
2921 static void io_sync_file_range_finish(struct io_wq_work **workptr)
2923 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2924 struct io_kiocb *nxt = NULL;
2927 if (io_req_cancelled(req))
2930 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
2933 req_set_fail_links(req);
2934 io_cqring_add_event(req, ret);
2935 io_put_req_find_next(req, &nxt);
2937 io_wq_assign_next(workptr, nxt);
2940 static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt,
2941 bool force_nonblock)
2943 struct io_wq_work *work, *old_work;
2945 /* sync_file_range always requires a blocking context */
2946 if (force_nonblock) {
2948 req->work.func = io_sync_file_range_finish;
2952 work = old_work = &req->work;
2953 io_sync_file_range_finish(&work);
2954 if (work && work != old_work)
2955 *nxt = container_of(work, struct io_kiocb, work);
2959 #if defined(CONFIG_NET)
2960 static void io_sendrecv_async(struct io_wq_work **workptr)
2962 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2963 struct iovec *iov = NULL;
2965 if (req->io->rw.iov != req->io->rw.fast_iov)
2966 iov = req->io->msg.iov;
2967 io_wq_submit_work(workptr);
2972 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2974 #if defined(CONFIG_NET)
2975 struct io_sr_msg *sr = &req->sr_msg;
2976 struct io_async_ctx *io = req->io;
2978 sr->msg_flags = READ_ONCE(sqe->msg_flags);
2979 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
2980 sr->len = READ_ONCE(sqe->len);
2982 if (!io || req->opcode == IORING_OP_SEND)
2985 io->msg.iov = io->msg.fast_iov;
2986 return sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
2993 static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt,
2994 bool force_nonblock)
2996 #if defined(CONFIG_NET)
2997 struct io_async_msghdr *kmsg = NULL;
2998 struct socket *sock;
3001 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3004 sock = sock_from_file(req->file, &ret);
3006 struct io_async_ctx io;
3007 struct sockaddr_storage addr;
3011 kmsg = &req->io->msg;
3012 kmsg->msg.msg_name = &addr;
3013 /* if iov is set, it's allocated already */
3015 kmsg->iov = kmsg->fast_iov;
3016 kmsg->msg.msg_iter.iov = kmsg->iov;
3018 struct io_sr_msg *sr = &req->sr_msg;
3021 kmsg->msg.msg_name = &addr;
3023 io.msg.iov = io.msg.fast_iov;
3024 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3025 sr->msg_flags, &io.msg.iov);
3030 flags = req->sr_msg.msg_flags;
3031 if (flags & MSG_DONTWAIT)
3032 req->flags |= REQ_F_NOWAIT;
3033 else if (force_nonblock)
3034 flags |= MSG_DONTWAIT;
3036 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3037 if (force_nonblock && ret == -EAGAIN) {
3040 if (io_alloc_async_ctx(req))
3042 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3043 req->work.func = io_sendrecv_async;
3046 if (ret == -ERESTARTSYS)
3050 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
3052 io_cqring_add_event(req, ret);
3054 req_set_fail_links(req);
3055 io_put_req_find_next(req, nxt);
3062 static int io_send(struct io_kiocb *req, struct io_kiocb **nxt,
3063 bool force_nonblock)
3065 #if defined(CONFIG_NET)
3066 struct socket *sock;
3069 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3072 sock = sock_from_file(req->file, &ret);
3074 struct io_sr_msg *sr = &req->sr_msg;
3079 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3084 msg.msg_name = NULL;
3085 msg.msg_control = NULL;
3086 msg.msg_controllen = 0;
3087 msg.msg_namelen = 0;
3089 flags = req->sr_msg.msg_flags;
3090 if (flags & MSG_DONTWAIT)
3091 req->flags |= REQ_F_NOWAIT;
3092 else if (force_nonblock)
3093 flags |= MSG_DONTWAIT;
3095 msg.msg_flags = flags;
3096 ret = sock_sendmsg(sock, &msg);
3097 if (force_nonblock && ret == -EAGAIN)
3099 if (ret == -ERESTARTSYS)
3103 io_cqring_add_event(req, ret);
3105 req_set_fail_links(req);
3106 io_put_req_find_next(req, nxt);
3113 static int io_recvmsg_prep(struct io_kiocb *req,
3114 const struct io_uring_sqe *sqe)
3116 #if defined(CONFIG_NET)
3117 struct io_sr_msg *sr = &req->sr_msg;
3118 struct io_async_ctx *io = req->io;
3120 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3121 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3122 sr->len = READ_ONCE(sqe->len);
3124 if (!io || req->opcode == IORING_OP_RECV)
3127 io->msg.iov = io->msg.fast_iov;
3128 return recvmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3129 &io->msg.uaddr, &io->msg.iov);
3135 static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt,
3136 bool force_nonblock)
3138 #if defined(CONFIG_NET)
3139 struct io_async_msghdr *kmsg = NULL;
3140 struct socket *sock;
3143 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3146 sock = sock_from_file(req->file, &ret);
3148 struct io_async_ctx io;
3149 struct sockaddr_storage addr;
3153 kmsg = &req->io->msg;
3154 kmsg->msg.msg_name = &addr;
3155 /* if iov is set, it's allocated already */
3157 kmsg->iov = kmsg->fast_iov;
3158 kmsg->msg.msg_iter.iov = kmsg->iov;
3160 struct io_sr_msg *sr = &req->sr_msg;
3163 kmsg->msg.msg_name = &addr;
3165 io.msg.iov = io.msg.fast_iov;
3166 ret = recvmsg_copy_msghdr(&io.msg.msg, sr->msg,
3167 sr->msg_flags, &io.msg.uaddr,
3173 flags = req->sr_msg.msg_flags;
3174 if (flags & MSG_DONTWAIT)
3175 req->flags |= REQ_F_NOWAIT;
3176 else if (force_nonblock)
3177 flags |= MSG_DONTWAIT;
3179 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3180 kmsg->uaddr, flags);
3181 if (force_nonblock && ret == -EAGAIN) {
3184 if (io_alloc_async_ctx(req))
3186 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3187 req->work.func = io_sendrecv_async;
3190 if (ret == -ERESTARTSYS)
3194 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
3196 io_cqring_add_event(req, ret);
3198 req_set_fail_links(req);
3199 io_put_req_find_next(req, nxt);
3206 static int io_recv(struct io_kiocb *req, struct io_kiocb **nxt,
3207 bool force_nonblock)
3209 #if defined(CONFIG_NET)
3210 struct socket *sock;
3213 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3216 sock = sock_from_file(req->file, &ret);
3218 struct io_sr_msg *sr = &req->sr_msg;
3223 ret = import_single_range(READ, sr->buf, sr->len, &iov,
3228 msg.msg_name = NULL;
3229 msg.msg_control = NULL;
3230 msg.msg_controllen = 0;
3231 msg.msg_namelen = 0;
3232 msg.msg_iocb = NULL;
3235 flags = req->sr_msg.msg_flags;
3236 if (flags & MSG_DONTWAIT)
3237 req->flags |= REQ_F_NOWAIT;
3238 else if (force_nonblock)
3239 flags |= MSG_DONTWAIT;
3241 ret = sock_recvmsg(sock, &msg, flags);
3242 if (force_nonblock && ret == -EAGAIN)
3244 if (ret == -ERESTARTSYS)
3248 io_cqring_add_event(req, ret);
3250 req_set_fail_links(req);
3251 io_put_req_find_next(req, nxt);
3259 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3261 #if defined(CONFIG_NET)
3262 struct io_accept *accept = &req->accept;
3264 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3266 if (sqe->ioprio || sqe->len || sqe->buf_index)
3269 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3270 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3271 accept->flags = READ_ONCE(sqe->accept_flags);
3278 #if defined(CONFIG_NET)
3279 static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3280 bool force_nonblock)
3282 struct io_accept *accept = &req->accept;
3283 unsigned file_flags;
3286 file_flags = force_nonblock ? O_NONBLOCK : 0;
3287 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3288 accept->addr_len, accept->flags);
3289 if (ret == -EAGAIN && force_nonblock)
3291 if (ret == -ERESTARTSYS)
3294 req_set_fail_links(req);
3295 io_cqring_add_event(req, ret);
3296 io_put_req_find_next(req, nxt);
3300 static void io_accept_finish(struct io_wq_work **workptr)
3302 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3303 struct io_kiocb *nxt = NULL;
3305 if (io_req_cancelled(req))
3307 __io_accept(req, &nxt, false);
3309 io_wq_assign_next(workptr, nxt);
3313 static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3314 bool force_nonblock)
3316 #if defined(CONFIG_NET)
3319 ret = __io_accept(req, nxt, force_nonblock);
3320 if (ret == -EAGAIN && force_nonblock) {
3321 req->work.func = io_accept_finish;
3331 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3333 #if defined(CONFIG_NET)
3334 struct io_connect *conn = &req->connect;
3335 struct io_async_ctx *io = req->io;
3337 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3339 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3342 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3343 conn->addr_len = READ_ONCE(sqe->addr2);
3348 return move_addr_to_kernel(conn->addr, conn->addr_len,
3349 &io->connect.address);
3355 static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt,
3356 bool force_nonblock)
3358 #if defined(CONFIG_NET)
3359 struct io_async_ctx __io, *io;
3360 unsigned file_flags;
3366 ret = move_addr_to_kernel(req->connect.addr,
3367 req->connect.addr_len,
3368 &__io.connect.address);
3374 file_flags = force_nonblock ? O_NONBLOCK : 0;
3376 ret = __sys_connect_file(req->file, &io->connect.address,
3377 req->connect.addr_len, file_flags);
3378 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
3381 if (io_alloc_async_ctx(req)) {
3385 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
3388 if (ret == -ERESTARTSYS)
3392 req_set_fail_links(req);
3393 io_cqring_add_event(req, ret);
3394 io_put_req_find_next(req, nxt);
3401 static void io_poll_remove_one(struct io_kiocb *req)
3403 struct io_poll_iocb *poll = &req->poll;
3405 spin_lock(&poll->head->lock);
3406 WRITE_ONCE(poll->canceled, true);
3407 if (!list_empty(&poll->wait.entry)) {
3408 list_del_init(&poll->wait.entry);
3409 io_queue_async_work(req);
3411 spin_unlock(&poll->head->lock);
3412 hash_del(&req->hash_node);
3415 static void io_poll_remove_all(struct io_ring_ctx *ctx)
3417 struct hlist_node *tmp;
3418 struct io_kiocb *req;
3421 spin_lock_irq(&ctx->completion_lock);
3422 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
3423 struct hlist_head *list;
3425 list = &ctx->cancel_hash[i];
3426 hlist_for_each_entry_safe(req, tmp, list, hash_node)
3427 io_poll_remove_one(req);
3429 spin_unlock_irq(&ctx->completion_lock);
3432 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
3434 struct hlist_head *list;
3435 struct io_kiocb *req;
3437 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
3438 hlist_for_each_entry(req, list, hash_node) {
3439 if (sqe_addr == req->user_data) {
3440 io_poll_remove_one(req);
3448 static int io_poll_remove_prep(struct io_kiocb *req,
3449 const struct io_uring_sqe *sqe)
3451 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3453 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3457 req->poll.addr = READ_ONCE(sqe->addr);
3462 * Find a running poll command that matches one specified in sqe->addr,
3463 * and remove it if found.
3465 static int io_poll_remove(struct io_kiocb *req)
3467 struct io_ring_ctx *ctx = req->ctx;
3471 addr = req->poll.addr;
3472 spin_lock_irq(&ctx->completion_lock);
3473 ret = io_poll_cancel(ctx, addr);
3474 spin_unlock_irq(&ctx->completion_lock);
3476 io_cqring_add_event(req, ret);
3478 req_set_fail_links(req);
3483 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
3485 struct io_ring_ctx *ctx = req->ctx;
3487 req->poll.done = true;
3489 io_cqring_fill_event(req, error);
3491 io_cqring_fill_event(req, mangle_poll(mask));
3492 io_commit_cqring(ctx);
3495 static void io_poll_complete_work(struct io_wq_work **workptr)
3497 struct io_wq_work *work = *workptr;
3498 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3499 struct io_poll_iocb *poll = &req->poll;
3500 struct poll_table_struct pt = { ._key = poll->events };
3501 struct io_ring_ctx *ctx = req->ctx;
3502 struct io_kiocb *nxt = NULL;
3506 if (work->flags & IO_WQ_WORK_CANCEL) {
3507 WRITE_ONCE(poll->canceled, true);
3509 } else if (READ_ONCE(poll->canceled)) {
3513 if (ret != -ECANCELED)
3514 mask = vfs_poll(poll->file, &pt) & poll->events;
3517 * Note that ->ki_cancel callers also delete iocb from active_reqs after
3518 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
3519 * synchronize with them. In the cancellation case the list_del_init
3520 * itself is not actually needed, but harmless so we keep it in to
3521 * avoid further branches in the fast path.
3523 spin_lock_irq(&ctx->completion_lock);
3524 if (!mask && ret != -ECANCELED) {
3525 add_wait_queue(poll->head, &poll->wait);
3526 spin_unlock_irq(&ctx->completion_lock);
3529 hash_del(&req->hash_node);
3530 io_poll_complete(req, mask, ret);
3531 spin_unlock_irq(&ctx->completion_lock);
3533 io_cqring_ev_posted(ctx);
3536 req_set_fail_links(req);
3537 io_put_req_find_next(req, &nxt);
3539 io_wq_assign_next(workptr, nxt);
3542 static void __io_poll_flush(struct io_ring_ctx *ctx, struct llist_node *nodes)
3544 struct io_kiocb *req, *tmp;
3545 struct req_batch rb;
3547 rb.to_free = rb.need_iter = 0;
3548 spin_lock_irq(&ctx->completion_lock);
3549 llist_for_each_entry_safe(req, tmp, nodes, llist_node) {
3550 hash_del(&req->hash_node);
3551 io_poll_complete(req, req->result, 0);
3553 if (refcount_dec_and_test(&req->refs) &&
3554 !io_req_multi_free(&rb, req)) {
3555 req->flags |= REQ_F_COMP_LOCKED;
3559 spin_unlock_irq(&ctx->completion_lock);
3561 io_cqring_ev_posted(ctx);
3562 io_free_req_many(ctx, &rb);
3565 static void io_poll_flush(struct io_wq_work **workptr)
3567 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3568 struct llist_node *nodes;
3570 nodes = llist_del_all(&req->ctx->poll_llist);
3572 __io_poll_flush(req->ctx, nodes);
3575 static void io_poll_trigger_evfd(struct io_wq_work **workptr)
3577 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3579 eventfd_signal(req->ctx->cq_ev_fd, 1);
3583 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
3586 struct io_poll_iocb *poll = wait->private;
3587 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
3588 struct io_ring_ctx *ctx = req->ctx;
3589 __poll_t mask = key_to_poll(key);
3591 /* for instances that support it check for an event match first: */
3592 if (mask && !(mask & poll->events))
3595 list_del_init(&poll->wait.entry);
3598 * Run completion inline if we can. We're using trylock here because
3599 * we are violating the completion_lock -> poll wq lock ordering.
3600 * If we have a link timeout we're going to need the completion_lock
3601 * for finalizing the request, mark us as having grabbed that already.
3604 unsigned long flags;
3606 if (llist_empty(&ctx->poll_llist) &&
3607 spin_trylock_irqsave(&ctx->completion_lock, flags)) {
3610 hash_del(&req->hash_node);
3611 io_poll_complete(req, mask, 0);
3613 trigger_ev = io_should_trigger_evfd(ctx);
3614 if (trigger_ev && eventfd_signal_count()) {
3616 req->work.func = io_poll_trigger_evfd;
3618 req->flags |= REQ_F_COMP_LOCKED;
3622 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3623 __io_cqring_ev_posted(ctx, trigger_ev);
3626 req->llist_node.next = NULL;
3627 /* if the list wasn't empty, we're done */
3628 if (!llist_add(&req->llist_node, &ctx->poll_llist))
3631 req->work.func = io_poll_flush;
3635 io_queue_async_work(req);
3640 struct io_poll_table {
3641 struct poll_table_struct pt;
3642 struct io_kiocb *req;
3646 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
3647 struct poll_table_struct *p)
3649 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
3651 if (unlikely(pt->req->poll.head)) {
3652 pt->error = -EINVAL;
3657 pt->req->poll.head = head;
3658 add_wait_queue(head, &pt->req->poll.wait);
3661 static void io_poll_req_insert(struct io_kiocb *req)
3663 struct io_ring_ctx *ctx = req->ctx;
3664 struct hlist_head *list;
3666 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
3667 hlist_add_head(&req->hash_node, list);
3670 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3672 struct io_poll_iocb *poll = &req->poll;
3675 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3677 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3682 events = READ_ONCE(sqe->poll_events);
3683 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
3687 static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt)
3689 struct io_poll_iocb *poll = &req->poll;
3690 struct io_ring_ctx *ctx = req->ctx;
3691 struct io_poll_table ipt;
3692 bool cancel = false;
3695 INIT_IO_WORK(&req->work, io_poll_complete_work);
3696 INIT_HLIST_NODE(&req->hash_node);
3700 poll->canceled = false;
3702 ipt.pt._qproc = io_poll_queue_proc;
3703 ipt.pt._key = poll->events;
3705 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
3707 /* initialized the list so that we can do list_empty checks */
3708 INIT_LIST_HEAD(&poll->wait.entry);
3709 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
3710 poll->wait.private = poll;
3712 INIT_LIST_HEAD(&req->list);
3714 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
3716 spin_lock_irq(&ctx->completion_lock);
3717 if (likely(poll->head)) {
3718 spin_lock(&poll->head->lock);
3719 if (unlikely(list_empty(&poll->wait.entry))) {
3725 if (mask || ipt.error)
3726 list_del_init(&poll->wait.entry);
3728 WRITE_ONCE(poll->canceled, true);
3729 else if (!poll->done) /* actually waiting for an event */
3730 io_poll_req_insert(req);
3731 spin_unlock(&poll->head->lock);
3733 if (mask) { /* no async, we'd stolen it */
3735 io_poll_complete(req, mask, 0);
3737 spin_unlock_irq(&ctx->completion_lock);
3740 io_cqring_ev_posted(ctx);
3741 io_put_req_find_next(req, nxt);
3746 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
3748 struct io_timeout_data *data = container_of(timer,
3749 struct io_timeout_data, timer);
3750 struct io_kiocb *req = data->req;
3751 struct io_ring_ctx *ctx = req->ctx;
3752 unsigned long flags;
3754 atomic_inc(&ctx->cq_timeouts);
3756 spin_lock_irqsave(&ctx->completion_lock, flags);
3758 * We could be racing with timeout deletion. If the list is empty,
3759 * then timeout lookup already found it and will be handling it.
3761 if (!list_empty(&req->list)) {
3762 struct io_kiocb *prev;
3765 * Adjust the reqs sequence before the current one because it
3766 * will consume a slot in the cq_ring and the cq_tail
3767 * pointer will be increased, otherwise other timeout reqs may
3768 * return in advance without waiting for enough wait_nr.
3771 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
3773 list_del_init(&req->list);
3776 io_cqring_fill_event(req, -ETIME);
3777 io_commit_cqring(ctx);
3778 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3780 io_cqring_ev_posted(ctx);
3781 req_set_fail_links(req);
3783 return HRTIMER_NORESTART;
3786 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
3788 struct io_kiocb *req;
3791 list_for_each_entry(req, &ctx->timeout_list, list) {
3792 if (user_data == req->user_data) {
3793 list_del_init(&req->list);
3802 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
3806 req_set_fail_links(req);
3807 io_cqring_fill_event(req, -ECANCELED);
3812 static int io_timeout_remove_prep(struct io_kiocb *req,
3813 const struct io_uring_sqe *sqe)
3815 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3817 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
3820 req->timeout.addr = READ_ONCE(sqe->addr);
3821 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
3822 if (req->timeout.flags)
3829 * Remove or update an existing timeout command
3831 static int io_timeout_remove(struct io_kiocb *req)
3833 struct io_ring_ctx *ctx = req->ctx;
3836 spin_lock_irq(&ctx->completion_lock);
3837 ret = io_timeout_cancel(ctx, req->timeout.addr);
3839 io_cqring_fill_event(req, ret);
3840 io_commit_cqring(ctx);
3841 spin_unlock_irq(&ctx->completion_lock);
3842 io_cqring_ev_posted(ctx);
3844 req_set_fail_links(req);
3849 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3850 bool is_timeout_link)
3852 struct io_timeout_data *data;
3855 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3857 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
3859 if (sqe->off && is_timeout_link)
3861 flags = READ_ONCE(sqe->timeout_flags);
3862 if (flags & ~IORING_TIMEOUT_ABS)
3865 req->timeout.count = READ_ONCE(sqe->off);
3867 if (!req->io && io_alloc_async_ctx(req))
3870 data = &req->io->timeout;
3872 req->flags |= REQ_F_TIMEOUT;
3874 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
3877 if (flags & IORING_TIMEOUT_ABS)
3878 data->mode = HRTIMER_MODE_ABS;
3880 data->mode = HRTIMER_MODE_REL;
3882 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
3886 static int io_timeout(struct io_kiocb *req)
3889 struct io_ring_ctx *ctx = req->ctx;
3890 struct io_timeout_data *data;
3891 struct list_head *entry;
3894 data = &req->io->timeout;
3897 * sqe->off holds how many events that need to occur for this
3898 * timeout event to be satisfied. If it isn't set, then this is
3899 * a pure timeout request, sequence isn't used.
3901 count = req->timeout.count;
3903 req->flags |= REQ_F_TIMEOUT_NOSEQ;
3904 spin_lock_irq(&ctx->completion_lock);
3905 entry = ctx->timeout_list.prev;
3909 req->sequence = ctx->cached_sq_head + count - 1;
3910 data->seq_offset = count;
3913 * Insertion sort, ensuring the first entry in the list is always
3914 * the one we need first.
3916 spin_lock_irq(&ctx->completion_lock);
3917 list_for_each_prev(entry, &ctx->timeout_list) {
3918 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
3919 unsigned nxt_sq_head;
3920 long long tmp, tmp_nxt;
3921 u32 nxt_offset = nxt->io->timeout.seq_offset;
3923 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
3927 * Since cached_sq_head + count - 1 can overflow, use type long
3930 tmp = (long long)ctx->cached_sq_head + count - 1;
3931 nxt_sq_head = nxt->sequence - nxt_offset + 1;
3932 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
3935 * cached_sq_head may overflow, and it will never overflow twice
3936 * once there is some timeout req still be valid.
3938 if (ctx->cached_sq_head < nxt_sq_head)
3945 * Sequence of reqs after the insert one and itself should
3946 * be adjusted because each timeout req consumes a slot.
3951 req->sequence -= span;
3953 list_add(&req->list, entry);
3954 data->timer.function = io_timeout_fn;
3955 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
3956 spin_unlock_irq(&ctx->completion_lock);
3960 static bool io_cancel_cb(struct io_wq_work *work, void *data)
3962 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3964 return req->user_data == (unsigned long) data;
3967 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
3969 enum io_wq_cancel cancel_ret;
3972 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
3973 switch (cancel_ret) {
3974 case IO_WQ_CANCEL_OK:
3977 case IO_WQ_CANCEL_RUNNING:
3980 case IO_WQ_CANCEL_NOTFOUND:
3988 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
3989 struct io_kiocb *req, __u64 sqe_addr,
3990 struct io_kiocb **nxt, int success_ret)
3992 unsigned long flags;
3995 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
3996 if (ret != -ENOENT) {
3997 spin_lock_irqsave(&ctx->completion_lock, flags);
4001 spin_lock_irqsave(&ctx->completion_lock, flags);
4002 ret = io_timeout_cancel(ctx, sqe_addr);
4005 ret = io_poll_cancel(ctx, sqe_addr);
4009 io_cqring_fill_event(req, ret);
4010 io_commit_cqring(ctx);
4011 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4012 io_cqring_ev_posted(ctx);
4015 req_set_fail_links(req);
4016 io_put_req_find_next(req, nxt);
4019 static int io_async_cancel_prep(struct io_kiocb *req,
4020 const struct io_uring_sqe *sqe)
4022 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4024 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4028 req->cancel.addr = READ_ONCE(sqe->addr);
4032 static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt)
4034 struct io_ring_ctx *ctx = req->ctx;
4036 io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0);
4040 static int io_files_update_prep(struct io_kiocb *req,
4041 const struct io_uring_sqe *sqe)
4043 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4046 req->files_update.offset = READ_ONCE(sqe->off);
4047 req->files_update.nr_args = READ_ONCE(sqe->len);
4048 if (!req->files_update.nr_args)
4050 req->files_update.arg = READ_ONCE(sqe->addr);
4054 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4056 struct io_ring_ctx *ctx = req->ctx;
4057 struct io_uring_files_update up;
4063 up.offset = req->files_update.offset;
4064 up.fds = req->files_update.arg;
4066 mutex_lock(&ctx->uring_lock);
4067 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4068 mutex_unlock(&ctx->uring_lock);
4071 req_set_fail_links(req);
4072 io_cqring_add_event(req, ret);
4077 static int io_req_defer_prep(struct io_kiocb *req,
4078 const struct io_uring_sqe *sqe)
4082 if (io_op_defs[req->opcode].file_table) {
4083 ret = io_grab_files(req);
4088 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4090 switch (req->opcode) {
4093 case IORING_OP_READV:
4094 case IORING_OP_READ_FIXED:
4095 case IORING_OP_READ:
4096 ret = io_read_prep(req, sqe, true);
4098 case IORING_OP_WRITEV:
4099 case IORING_OP_WRITE_FIXED:
4100 case IORING_OP_WRITE:
4101 ret = io_write_prep(req, sqe, true);
4103 case IORING_OP_POLL_ADD:
4104 ret = io_poll_add_prep(req, sqe);
4106 case IORING_OP_POLL_REMOVE:
4107 ret = io_poll_remove_prep(req, sqe);
4109 case IORING_OP_FSYNC:
4110 ret = io_prep_fsync(req, sqe);
4112 case IORING_OP_SYNC_FILE_RANGE:
4113 ret = io_prep_sfr(req, sqe);
4115 case IORING_OP_SENDMSG:
4116 case IORING_OP_SEND:
4117 ret = io_sendmsg_prep(req, sqe);
4119 case IORING_OP_RECVMSG:
4120 case IORING_OP_RECV:
4121 ret = io_recvmsg_prep(req, sqe);
4123 case IORING_OP_CONNECT:
4124 ret = io_connect_prep(req, sqe);
4126 case IORING_OP_TIMEOUT:
4127 ret = io_timeout_prep(req, sqe, false);
4129 case IORING_OP_TIMEOUT_REMOVE:
4130 ret = io_timeout_remove_prep(req, sqe);
4132 case IORING_OP_ASYNC_CANCEL:
4133 ret = io_async_cancel_prep(req, sqe);
4135 case IORING_OP_LINK_TIMEOUT:
4136 ret = io_timeout_prep(req, sqe, true);
4138 case IORING_OP_ACCEPT:
4139 ret = io_accept_prep(req, sqe);
4141 case IORING_OP_FALLOCATE:
4142 ret = io_fallocate_prep(req, sqe);
4144 case IORING_OP_OPENAT:
4145 ret = io_openat_prep(req, sqe);
4147 case IORING_OP_CLOSE:
4148 ret = io_close_prep(req, sqe);
4150 case IORING_OP_FILES_UPDATE:
4151 ret = io_files_update_prep(req, sqe);
4153 case IORING_OP_STATX:
4154 ret = io_statx_prep(req, sqe);
4156 case IORING_OP_FADVISE:
4157 ret = io_fadvise_prep(req, sqe);
4159 case IORING_OP_MADVISE:
4160 ret = io_madvise_prep(req, sqe);
4162 case IORING_OP_OPENAT2:
4163 ret = io_openat2_prep(req, sqe);
4165 case IORING_OP_EPOLL_CTL:
4166 ret = io_epoll_ctl_prep(req, sqe);
4169 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
4178 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4180 struct io_ring_ctx *ctx = req->ctx;
4183 /* Still need defer if there is pending req in defer list. */
4184 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
4187 if (!req->io && io_alloc_async_ctx(req))
4190 ret = io_req_defer_prep(req, sqe);
4194 spin_lock_irq(&ctx->completion_lock);
4195 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
4196 spin_unlock_irq(&ctx->completion_lock);
4200 trace_io_uring_defer(ctx, req, req->user_data);
4201 list_add_tail(&req->list, &ctx->defer_list);
4202 spin_unlock_irq(&ctx->completion_lock);
4203 return -EIOCBQUEUED;
4206 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4207 struct io_kiocb **nxt, bool force_nonblock)
4209 struct io_ring_ctx *ctx = req->ctx;
4212 switch (req->opcode) {
4216 case IORING_OP_READV:
4217 case IORING_OP_READ_FIXED:
4218 case IORING_OP_READ:
4220 ret = io_read_prep(req, sqe, force_nonblock);
4224 ret = io_read(req, nxt, force_nonblock);
4226 case IORING_OP_WRITEV:
4227 case IORING_OP_WRITE_FIXED:
4228 case IORING_OP_WRITE:
4230 ret = io_write_prep(req, sqe, force_nonblock);
4234 ret = io_write(req, nxt, force_nonblock);
4236 case IORING_OP_FSYNC:
4238 ret = io_prep_fsync(req, sqe);
4242 ret = io_fsync(req, nxt, force_nonblock);
4244 case IORING_OP_POLL_ADD:
4246 ret = io_poll_add_prep(req, sqe);
4250 ret = io_poll_add(req, nxt);
4252 case IORING_OP_POLL_REMOVE:
4254 ret = io_poll_remove_prep(req, sqe);
4258 ret = io_poll_remove(req);
4260 case IORING_OP_SYNC_FILE_RANGE:
4262 ret = io_prep_sfr(req, sqe);
4266 ret = io_sync_file_range(req, nxt, force_nonblock);
4268 case IORING_OP_SENDMSG:
4269 case IORING_OP_SEND:
4271 ret = io_sendmsg_prep(req, sqe);
4275 if (req->opcode == IORING_OP_SENDMSG)
4276 ret = io_sendmsg(req, nxt, force_nonblock);
4278 ret = io_send(req, nxt, force_nonblock);
4280 case IORING_OP_RECVMSG:
4281 case IORING_OP_RECV:
4283 ret = io_recvmsg_prep(req, sqe);
4287 if (req->opcode == IORING_OP_RECVMSG)
4288 ret = io_recvmsg(req, nxt, force_nonblock);
4290 ret = io_recv(req, nxt, force_nonblock);
4292 case IORING_OP_TIMEOUT:
4294 ret = io_timeout_prep(req, sqe, false);
4298 ret = io_timeout(req);
4300 case IORING_OP_TIMEOUT_REMOVE:
4302 ret = io_timeout_remove_prep(req, sqe);
4306 ret = io_timeout_remove(req);
4308 case IORING_OP_ACCEPT:
4310 ret = io_accept_prep(req, sqe);
4314 ret = io_accept(req, nxt, force_nonblock);
4316 case IORING_OP_CONNECT:
4318 ret = io_connect_prep(req, sqe);
4322 ret = io_connect(req, nxt, force_nonblock);
4324 case IORING_OP_ASYNC_CANCEL:
4326 ret = io_async_cancel_prep(req, sqe);
4330 ret = io_async_cancel(req, nxt);
4332 case IORING_OP_FALLOCATE:
4334 ret = io_fallocate_prep(req, sqe);
4338 ret = io_fallocate(req, nxt, force_nonblock);
4340 case IORING_OP_OPENAT:
4342 ret = io_openat_prep(req, sqe);
4346 ret = io_openat(req, nxt, force_nonblock);
4348 case IORING_OP_CLOSE:
4350 ret = io_close_prep(req, sqe);
4354 ret = io_close(req, nxt, force_nonblock);
4356 case IORING_OP_FILES_UPDATE:
4358 ret = io_files_update_prep(req, sqe);
4362 ret = io_files_update(req, force_nonblock);
4364 case IORING_OP_STATX:
4366 ret = io_statx_prep(req, sqe);
4370 ret = io_statx(req, nxt, force_nonblock);
4372 case IORING_OP_FADVISE:
4374 ret = io_fadvise_prep(req, sqe);
4378 ret = io_fadvise(req, nxt, force_nonblock);
4380 case IORING_OP_MADVISE:
4382 ret = io_madvise_prep(req, sqe);
4386 ret = io_madvise(req, nxt, force_nonblock);
4388 case IORING_OP_OPENAT2:
4390 ret = io_openat2_prep(req, sqe);
4394 ret = io_openat2(req, nxt, force_nonblock);
4396 case IORING_OP_EPOLL_CTL:
4398 ret = io_epoll_ctl_prep(req, sqe);
4402 ret = io_epoll_ctl(req, nxt, force_nonblock);
4412 if (ctx->flags & IORING_SETUP_IOPOLL) {
4413 const bool in_async = io_wq_current_is_worker();
4415 if (req->result == -EAGAIN)
4418 /* workqueue context doesn't hold uring_lock, grab it now */
4420 mutex_lock(&ctx->uring_lock);
4422 io_iopoll_req_issued(req);
4425 mutex_unlock(&ctx->uring_lock);
4431 static void io_wq_submit_work(struct io_wq_work **workptr)
4433 struct io_wq_work *work = *workptr;
4434 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4435 struct io_kiocb *nxt = NULL;
4438 /* if NO_CANCEL is set, we must still run the work */
4439 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
4440 IO_WQ_WORK_CANCEL) {
4445 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
4446 req->in_async = true;
4448 ret = io_issue_sqe(req, NULL, &nxt, false);
4450 * We can get EAGAIN for polled IO even though we're
4451 * forcing a sync submission from here, since we can't
4452 * wait for request slots on the block side.
4460 /* drop submission reference */
4464 req_set_fail_links(req);
4465 io_cqring_add_event(req, ret);
4469 /* if a dependent link is ready, pass it back */
4471 io_wq_assign_next(workptr, nxt);
4474 static int io_req_needs_file(struct io_kiocb *req, int fd)
4476 if (!io_op_defs[req->opcode].needs_file)
4478 if (fd == -1 && io_op_defs[req->opcode].fd_non_neg)
4483 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
4486 struct fixed_file_table *table;
4488 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
4489 return table->files[index & IORING_FILE_TABLE_MASK];;
4492 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
4493 const struct io_uring_sqe *sqe)
4495 struct io_ring_ctx *ctx = req->ctx;
4499 flags = READ_ONCE(sqe->flags);
4500 fd = READ_ONCE(sqe->fd);
4502 if (!io_req_needs_file(req, fd))
4505 if (flags & IOSQE_FIXED_FILE) {
4506 if (unlikely(!ctx->file_data ||
4507 (unsigned) fd >= ctx->nr_user_files))
4509 fd = array_index_nospec(fd, ctx->nr_user_files);
4510 req->file = io_file_from_index(ctx, fd);
4513 req->flags |= REQ_F_FIXED_FILE;
4514 percpu_ref_get(&ctx->file_data->refs);
4516 if (req->needs_fixed_file)
4518 trace_io_uring_file_get(ctx, fd);
4519 req->file = io_file_get(state, fd);
4520 if (unlikely(!req->file))
4527 static int io_grab_files(struct io_kiocb *req)
4530 struct io_ring_ctx *ctx = req->ctx;
4532 if (req->work.files)
4534 if (!ctx->ring_file)
4538 spin_lock_irq(&ctx->inflight_lock);
4540 * We use the f_ops->flush() handler to ensure that we can flush
4541 * out work accessing these files if the fd is closed. Check if
4542 * the fd has changed since we started down this path, and disallow
4543 * this operation if it has.
4545 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
4546 list_add(&req->inflight_entry, &ctx->inflight_list);
4547 req->flags |= REQ_F_INFLIGHT;
4548 req->work.files = current->files;
4551 spin_unlock_irq(&ctx->inflight_lock);
4557 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
4559 struct io_timeout_data *data = container_of(timer,
4560 struct io_timeout_data, timer);
4561 struct io_kiocb *req = data->req;
4562 struct io_ring_ctx *ctx = req->ctx;
4563 struct io_kiocb *prev = NULL;
4564 unsigned long flags;
4566 spin_lock_irqsave(&ctx->completion_lock, flags);
4569 * We don't expect the list to be empty, that will only happen if we
4570 * race with the completion of the linked work.
4572 if (!list_empty(&req->link_list)) {
4573 prev = list_entry(req->link_list.prev, struct io_kiocb,
4575 if (refcount_inc_not_zero(&prev->refs)) {
4576 list_del_init(&req->link_list);
4577 prev->flags &= ~REQ_F_LINK_TIMEOUT;
4582 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4585 req_set_fail_links(prev);
4586 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
4590 io_cqring_add_event(req, -ETIME);
4593 return HRTIMER_NORESTART;
4596 static void io_queue_linked_timeout(struct io_kiocb *req)
4598 struct io_ring_ctx *ctx = req->ctx;
4601 * If the list is now empty, then our linked request finished before
4602 * we got a chance to setup the timer
4604 spin_lock_irq(&ctx->completion_lock);
4605 if (!list_empty(&req->link_list)) {
4606 struct io_timeout_data *data = &req->io->timeout;
4608 data->timer.function = io_link_timeout_fn;
4609 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
4612 spin_unlock_irq(&ctx->completion_lock);
4614 /* drop submission reference */
4618 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
4620 struct io_kiocb *nxt;
4622 if (!(req->flags & REQ_F_LINK))
4625 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
4627 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
4630 req->flags |= REQ_F_LINK_TIMEOUT;
4634 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4636 struct io_kiocb *linked_timeout;
4637 struct io_kiocb *nxt = NULL;
4641 linked_timeout = io_prep_linked_timeout(req);
4643 ret = io_issue_sqe(req, sqe, &nxt, true);
4646 * We async punt it if the file wasn't marked NOWAIT, or if the file
4647 * doesn't support non-blocking read/write attempts
4649 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
4650 (req->flags & REQ_F_MUST_PUNT))) {
4652 if (io_op_defs[req->opcode].file_table) {
4653 ret = io_grab_files(req);
4659 * Queued up for async execution, worker will release
4660 * submit reference when the iocb is actually submitted.
4662 io_queue_async_work(req);
4667 /* drop submission reference */
4670 if (linked_timeout) {
4672 io_queue_linked_timeout(linked_timeout);
4674 io_put_req(linked_timeout);
4677 /* and drop final reference, if we failed */
4679 io_cqring_add_event(req, ret);
4680 req_set_fail_links(req);
4688 if (req->flags & REQ_F_FORCE_ASYNC)
4694 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4698 ret = io_req_defer(req, sqe);
4700 if (ret != -EIOCBQUEUED) {
4702 io_cqring_add_event(req, ret);
4703 req_set_fail_links(req);
4704 io_double_put_req(req);
4706 } else if (req->flags & REQ_F_FORCE_ASYNC) {
4707 ret = io_req_defer_prep(req, sqe);
4708 if (unlikely(ret < 0))
4711 * Never try inline submit of IOSQE_ASYNC is set, go straight
4712 * to async execution.
4714 req->work.flags |= IO_WQ_WORK_CONCURRENT;
4715 io_queue_async_work(req);
4717 __io_queue_sqe(req, sqe);
4721 static inline void io_queue_link_head(struct io_kiocb *req)
4723 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
4724 io_cqring_add_event(req, -ECANCELED);
4725 io_double_put_req(req);
4727 io_queue_sqe(req, NULL);
4730 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
4731 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
4733 static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4734 struct io_submit_state *state, struct io_kiocb **link)
4736 const struct cred *old_creds = NULL;
4737 struct io_ring_ctx *ctx = req->ctx;
4738 unsigned int sqe_flags;
4741 sqe_flags = READ_ONCE(sqe->flags);
4743 /* enforce forwards compatibility on users */
4744 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
4749 id = READ_ONCE(sqe->personality);
4751 const struct cred *personality_creds;
4753 personality_creds = idr_find(&ctx->personality_idr, id);
4754 if (unlikely(!personality_creds)) {
4758 old_creds = override_creds(personality_creds);
4761 /* same numerical values with corresponding REQ_F_*, safe to copy */
4762 req->flags |= sqe_flags & (IOSQE_IO_DRAIN|IOSQE_IO_HARDLINK|
4765 ret = io_req_set_file(state, req, sqe);
4766 if (unlikely(ret)) {
4768 io_cqring_add_event(req, ret);
4769 io_double_put_req(req);
4771 revert_creds(old_creds);
4776 * If we already have a head request, queue this one for async
4777 * submittal once the head completes. If we don't have a head but
4778 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
4779 * submitted sync once the chain is complete. If none of those
4780 * conditions are true (normal request), then just queue it.
4783 struct io_kiocb *head = *link;
4786 * Taking sequential execution of a link, draining both sides
4787 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
4788 * requests in the link. So, it drains the head and the
4789 * next after the link request. The last one is done via
4790 * drain_next flag to persist the effect across calls.
4792 if (sqe_flags & IOSQE_IO_DRAIN) {
4793 head->flags |= REQ_F_IO_DRAIN;
4794 ctx->drain_next = 1;
4796 if (io_alloc_async_ctx(req)) {
4801 ret = io_req_defer_prep(req, sqe);
4803 /* fail even hard links since we don't submit */
4804 head->flags |= REQ_F_FAIL_LINK;
4807 trace_io_uring_link(ctx, req, head);
4808 list_add_tail(&req->link_list, &head->link_list);
4810 /* last request of a link, enqueue the link */
4811 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) {
4812 io_queue_link_head(head);
4816 if (unlikely(ctx->drain_next)) {
4817 req->flags |= REQ_F_IO_DRAIN;
4818 req->ctx->drain_next = 0;
4820 if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
4821 req->flags |= REQ_F_LINK;
4822 INIT_LIST_HEAD(&req->link_list);
4823 ret = io_req_defer_prep(req, sqe);
4825 req->flags |= REQ_F_FAIL_LINK;
4828 io_queue_sqe(req, sqe);
4833 revert_creds(old_creds);
4838 * Batched submission is done, ensure local IO is flushed out.
4840 static void io_submit_state_end(struct io_submit_state *state)
4842 blk_finish_plug(&state->plug);
4844 if (state->free_reqs)
4845 kmem_cache_free_bulk(req_cachep, state->free_reqs,
4846 &state->reqs[state->cur_req]);
4850 * Start submission side cache.
4852 static void io_submit_state_start(struct io_submit_state *state,
4853 unsigned int max_ios)
4855 blk_start_plug(&state->plug);
4856 state->free_reqs = 0;
4858 state->ios_left = max_ios;
4861 static void io_commit_sqring(struct io_ring_ctx *ctx)
4863 struct io_rings *rings = ctx->rings;
4866 * Ensure any loads from the SQEs are done at this point,
4867 * since once we write the new head, the application could
4868 * write new data to them.
4870 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
4874 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
4875 * that is mapped by userspace. This means that care needs to be taken to
4876 * ensure that reads are stable, as we cannot rely on userspace always
4877 * being a good citizen. If members of the sqe are validated and then later
4878 * used, it's important that those reads are done through READ_ONCE() to
4879 * prevent a re-load down the line.
4881 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
4882 const struct io_uring_sqe **sqe_ptr)
4884 u32 *sq_array = ctx->sq_array;
4888 * The cached sq head (or cq tail) serves two purposes:
4890 * 1) allows us to batch the cost of updating the user visible
4892 * 2) allows the kernel side to track the head on its own, even
4893 * though the application is the one updating it.
4895 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
4896 if (likely(head < ctx->sq_entries)) {
4898 * All io need record the previous position, if LINK vs DARIN,
4899 * it can be used to mark the position of the first IO in the
4902 req->sequence = ctx->cached_sq_head;
4903 *sqe_ptr = &ctx->sq_sqes[head];
4904 req->opcode = READ_ONCE((*sqe_ptr)->opcode);
4905 req->user_data = READ_ONCE((*sqe_ptr)->user_data);
4906 ctx->cached_sq_head++;
4910 /* drop invalid entries */
4911 ctx->cached_sq_head++;
4912 ctx->cached_sq_dropped++;
4913 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
4917 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
4918 struct file *ring_file, int ring_fd,
4919 struct mm_struct **mm, bool async)
4921 struct io_submit_state state, *statep = NULL;
4922 struct io_kiocb *link = NULL;
4923 int i, submitted = 0;
4924 bool mm_fault = false;
4926 /* if we have a backlog and couldn't flush it all, return BUSY */
4927 if (test_bit(0, &ctx->sq_check_overflow)) {
4928 if (!list_empty(&ctx->cq_overflow_list) &&
4929 !io_cqring_overflow_flush(ctx, false))
4933 /* make sure SQ entry isn't read before tail */
4934 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
4936 if (!percpu_ref_tryget_many(&ctx->refs, nr))
4939 if (nr > IO_PLUG_THRESHOLD) {
4940 io_submit_state_start(&state, nr);
4944 ctx->ring_fd = ring_fd;
4945 ctx->ring_file = ring_file;
4947 for (i = 0; i < nr; i++) {
4948 const struct io_uring_sqe *sqe;
4949 struct io_kiocb *req;
4951 req = io_get_req(ctx, statep);
4952 if (unlikely(!req)) {
4954 submitted = -EAGAIN;
4957 if (!io_get_sqring(ctx, req, &sqe)) {
4958 __io_req_do_free(req);
4962 /* will complete beyond this point, count as submitted */
4965 if (unlikely(req->opcode >= IORING_OP_LAST)) {
4966 io_cqring_add_event(req, -EINVAL);
4967 io_double_put_req(req);
4971 if (io_op_defs[req->opcode].needs_mm && !*mm) {
4972 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
4974 use_mm(ctx->sqo_mm);
4979 req->has_user = *mm != NULL;
4980 req->in_async = async;
4981 req->needs_fixed_file = async;
4982 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
4984 if (!io_submit_sqe(req, sqe, statep, &link))
4988 if (unlikely(submitted != nr)) {
4989 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
4991 percpu_ref_put_many(&ctx->refs, nr - ref_used);
4994 io_queue_link_head(link);
4996 io_submit_state_end(&state);
4998 /* Commit SQ ring head once we've consumed and submitted all SQEs */
4999 io_commit_sqring(ctx);
5004 static int io_sq_thread(void *data)
5006 struct io_ring_ctx *ctx = data;
5007 struct mm_struct *cur_mm = NULL;
5008 const struct cred *old_cred;
5009 mm_segment_t old_fs;
5012 unsigned long timeout;
5015 complete(&ctx->completions[1]);
5019 old_cred = override_creds(ctx->creds);
5021 ret = timeout = inflight = 0;
5022 while (!kthread_should_park()) {
5023 unsigned int to_submit;
5026 unsigned nr_events = 0;
5028 if (ctx->flags & IORING_SETUP_IOPOLL) {
5030 * inflight is the count of the maximum possible
5031 * entries we submitted, but it can be smaller
5032 * if we dropped some of them. If we don't have
5033 * poll entries available, then we know that we
5034 * have nothing left to poll for. Reset the
5035 * inflight count to zero in that case.
5037 mutex_lock(&ctx->uring_lock);
5038 if (!list_empty(&ctx->poll_list))
5039 __io_iopoll_check(ctx, &nr_events, 0);
5042 mutex_unlock(&ctx->uring_lock);
5045 * Normal IO, just pretend everything completed.
5046 * We don't have to poll completions for that.
5048 nr_events = inflight;
5051 inflight -= nr_events;
5053 timeout = jiffies + ctx->sq_thread_idle;
5056 to_submit = io_sqring_entries(ctx);
5059 * If submit got -EBUSY, flag us as needing the application
5060 * to enter the kernel to reap and flush events.
5062 if (!to_submit || ret == -EBUSY) {
5064 * We're polling. If we're within the defined idle
5065 * period, then let us spin without work before going
5066 * to sleep. The exception is if we got EBUSY doing
5067 * more IO, we should wait for the application to
5068 * reap events and wake us up.
5071 (!time_after(jiffies, timeout) && ret != -EBUSY)) {
5077 * Drop cur_mm before scheduling, we can't hold it for
5078 * long periods (or over schedule()). Do this before
5079 * adding ourselves to the waitqueue, as the unuse/drop
5088 prepare_to_wait(&ctx->sqo_wait, &wait,
5089 TASK_INTERRUPTIBLE);
5091 /* Tell userspace we may need a wakeup call */
5092 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
5093 /* make sure to read SQ tail after writing flags */
5096 to_submit = io_sqring_entries(ctx);
5097 if (!to_submit || ret == -EBUSY) {
5098 if (kthread_should_park()) {
5099 finish_wait(&ctx->sqo_wait, &wait);
5102 if (signal_pending(current))
5103 flush_signals(current);
5105 finish_wait(&ctx->sqo_wait, &wait);
5107 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5110 finish_wait(&ctx->sqo_wait, &wait);
5112 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5115 mutex_lock(&ctx->uring_lock);
5116 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
5117 mutex_unlock(&ctx->uring_lock);
5127 revert_creds(old_cred);
5134 struct io_wait_queue {
5135 struct wait_queue_entry wq;
5136 struct io_ring_ctx *ctx;
5138 unsigned nr_timeouts;
5141 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
5143 struct io_ring_ctx *ctx = iowq->ctx;
5146 * Wake up if we have enough events, or if a timeout occurred since we
5147 * started waiting. For timeouts, we always want to return to userspace,
5148 * regardless of event count.
5150 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
5151 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
5154 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
5155 int wake_flags, void *key)
5157 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
5160 /* use noflush == true, as we can't safely rely on locking context */
5161 if (!io_should_wake(iowq, true))
5164 return autoremove_wake_function(curr, mode, wake_flags, key);
5168 * Wait until events become available, if we don't already have some. The
5169 * application must reap them itself, as they reside on the shared cq ring.
5171 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
5172 const sigset_t __user *sig, size_t sigsz)
5174 struct io_wait_queue iowq = {
5177 .func = io_wake_function,
5178 .entry = LIST_HEAD_INIT(iowq.wq.entry),
5181 .to_wait = min_events,
5183 struct io_rings *rings = ctx->rings;
5186 if (io_cqring_events(ctx, false) >= min_events)
5190 #ifdef CONFIG_COMPAT
5191 if (in_compat_syscall())
5192 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
5196 ret = set_user_sigmask(sig, sigsz);
5202 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
5203 trace_io_uring_cqring_wait(ctx, min_events);
5205 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
5206 TASK_INTERRUPTIBLE);
5207 if (io_should_wake(&iowq, false))
5210 if (signal_pending(current)) {
5215 finish_wait(&ctx->wait, &iowq.wq);
5217 restore_saved_sigmask_unless(ret == -EINTR);
5219 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
5222 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
5224 #if defined(CONFIG_UNIX)
5225 if (ctx->ring_sock) {
5226 struct sock *sock = ctx->ring_sock->sk;
5227 struct sk_buff *skb;
5229 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
5235 for (i = 0; i < ctx->nr_user_files; i++) {
5238 file = io_file_from_index(ctx, i);
5245 static void io_file_ref_kill(struct percpu_ref *ref)
5247 struct fixed_file_data *data;
5249 data = container_of(ref, struct fixed_file_data, refs);
5250 complete(&data->done);
5253 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
5255 struct fixed_file_data *data = ctx->file_data;
5256 unsigned nr_tables, i;
5261 /* protect against inflight atomic switch, which drops the ref */
5262 percpu_ref_get(&data->refs);
5263 /* wait for existing switches */
5264 flush_work(&data->ref_work);
5265 percpu_ref_kill_and_confirm(&data->refs, io_file_ref_kill);
5266 wait_for_completion(&data->done);
5267 percpu_ref_put(&data->refs);
5268 /* flush potential new switch */
5269 flush_work(&data->ref_work);
5270 percpu_ref_exit(&data->refs);
5272 __io_sqe_files_unregister(ctx);
5273 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
5274 for (i = 0; i < nr_tables; i++)
5275 kfree(data->table[i].files);
5278 ctx->file_data = NULL;
5279 ctx->nr_user_files = 0;
5283 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
5285 if (ctx->sqo_thread) {
5286 wait_for_completion(&ctx->completions[1]);
5288 * The park is a bit of a work-around, without it we get
5289 * warning spews on shutdown with SQPOLL set and affinity
5290 * set to a single CPU.
5292 kthread_park(ctx->sqo_thread);
5293 kthread_stop(ctx->sqo_thread);
5294 ctx->sqo_thread = NULL;
5298 static void io_finish_async(struct io_ring_ctx *ctx)
5300 io_sq_thread_stop(ctx);
5303 io_wq_destroy(ctx->io_wq);
5308 #if defined(CONFIG_UNIX)
5310 * Ensure the UNIX gc is aware of our file set, so we are certain that
5311 * the io_uring can be safely unregistered on process exit, even if we have
5312 * loops in the file referencing.
5314 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
5316 struct sock *sk = ctx->ring_sock->sk;
5317 struct scm_fp_list *fpl;
5318 struct sk_buff *skb;
5321 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
5322 unsigned long inflight = ctx->user->unix_inflight + nr;
5324 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
5328 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
5332 skb = alloc_skb(0, GFP_KERNEL);
5341 fpl->user = get_uid(ctx->user);
5342 for (i = 0; i < nr; i++) {
5343 struct file *file = io_file_from_index(ctx, i + offset);
5347 fpl->fp[nr_files] = get_file(file);
5348 unix_inflight(fpl->user, fpl->fp[nr_files]);
5353 fpl->max = SCM_MAX_FD;
5354 fpl->count = nr_files;
5355 UNIXCB(skb).fp = fpl;
5356 skb->destructor = unix_destruct_scm;
5357 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
5358 skb_queue_head(&sk->sk_receive_queue, skb);
5360 for (i = 0; i < nr_files; i++)
5371 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
5372 * causes regular reference counting to break down. We rely on the UNIX
5373 * garbage collection to take care of this problem for us.
5375 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5377 unsigned left, total;
5381 left = ctx->nr_user_files;
5383 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
5385 ret = __io_sqe_files_scm(ctx, this_files, total);
5389 total += this_files;
5395 while (total < ctx->nr_user_files) {
5396 struct file *file = io_file_from_index(ctx, total);
5406 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5412 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
5417 for (i = 0; i < nr_tables; i++) {
5418 struct fixed_file_table *table = &ctx->file_data->table[i];
5419 unsigned this_files;
5421 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
5422 table->files = kcalloc(this_files, sizeof(struct file *),
5426 nr_files -= this_files;
5432 for (i = 0; i < nr_tables; i++) {
5433 struct fixed_file_table *table = &ctx->file_data->table[i];
5434 kfree(table->files);
5439 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
5441 #if defined(CONFIG_UNIX)
5442 struct sock *sock = ctx->ring_sock->sk;
5443 struct sk_buff_head list, *head = &sock->sk_receive_queue;
5444 struct sk_buff *skb;
5447 __skb_queue_head_init(&list);
5450 * Find the skb that holds this file in its SCM_RIGHTS. When found,
5451 * remove this entry and rearrange the file array.
5453 skb = skb_dequeue(head);
5455 struct scm_fp_list *fp;
5457 fp = UNIXCB(skb).fp;
5458 for (i = 0; i < fp->count; i++) {
5461 if (fp->fp[i] != file)
5464 unix_notinflight(fp->user, fp->fp[i]);
5465 left = fp->count - 1 - i;
5467 memmove(&fp->fp[i], &fp->fp[i + 1],
5468 left * sizeof(struct file *));
5475 __skb_queue_tail(&list, skb);
5485 __skb_queue_tail(&list, skb);
5487 skb = skb_dequeue(head);
5490 if (skb_peek(&list)) {
5491 spin_lock_irq(&head->lock);
5492 while ((skb = __skb_dequeue(&list)) != NULL)
5493 __skb_queue_tail(head, skb);
5494 spin_unlock_irq(&head->lock);
5501 struct io_file_put {
5502 struct llist_node llist;
5504 struct completion *done;
5507 static void io_ring_file_ref_switch(struct work_struct *work)
5509 struct io_file_put *pfile, *tmp;
5510 struct fixed_file_data *data;
5511 struct llist_node *node;
5513 data = container_of(work, struct fixed_file_data, ref_work);
5515 while ((node = llist_del_all(&data->put_llist)) != NULL) {
5516 llist_for_each_entry_safe(pfile, tmp, node, llist) {
5517 io_ring_file_put(data->ctx, pfile->file);
5519 complete(pfile->done);
5525 percpu_ref_get(&data->refs);
5526 percpu_ref_switch_to_percpu(&data->refs);
5529 static void io_file_data_ref_zero(struct percpu_ref *ref)
5531 struct fixed_file_data *data;
5533 data = container_of(ref, struct fixed_file_data, refs);
5535 /* we can't safely switch from inside this context, punt to wq */
5536 queue_work(system_wq, &data->ref_work);
5539 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
5542 __s32 __user *fds = (__s32 __user *) arg;
5552 if (nr_args > IORING_MAX_FIXED_FILES)
5555 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
5556 if (!ctx->file_data)
5558 ctx->file_data->ctx = ctx;
5559 init_completion(&ctx->file_data->done);
5561 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
5562 ctx->file_data->table = kcalloc(nr_tables,
5563 sizeof(struct fixed_file_table),
5565 if (!ctx->file_data->table) {
5566 kfree(ctx->file_data);
5567 ctx->file_data = NULL;
5571 if (percpu_ref_init(&ctx->file_data->refs, io_file_data_ref_zero,
5572 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
5573 kfree(ctx->file_data->table);
5574 kfree(ctx->file_data);
5575 ctx->file_data = NULL;
5578 ctx->file_data->put_llist.first = NULL;
5579 INIT_WORK(&ctx->file_data->ref_work, io_ring_file_ref_switch);
5581 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
5582 percpu_ref_exit(&ctx->file_data->refs);
5583 kfree(ctx->file_data->table);
5584 kfree(ctx->file_data);
5585 ctx->file_data = NULL;
5589 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
5590 struct fixed_file_table *table;
5594 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
5596 /* allow sparse sets */
5602 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5603 index = i & IORING_FILE_TABLE_MASK;
5611 * Don't allow io_uring instances to be registered. If UNIX
5612 * isn't enabled, then this causes a reference cycle and this
5613 * instance can never get freed. If UNIX is enabled we'll
5614 * handle it just fine, but there's still no point in allowing
5615 * a ring fd as it doesn't support regular read/write anyway.
5617 if (file->f_op == &io_uring_fops) {
5622 table->files[index] = file;
5626 for (i = 0; i < ctx->nr_user_files; i++) {
5627 file = io_file_from_index(ctx, i);
5631 for (i = 0; i < nr_tables; i++)
5632 kfree(ctx->file_data->table[i].files);
5634 kfree(ctx->file_data->table);
5635 kfree(ctx->file_data);
5636 ctx->file_data = NULL;
5637 ctx->nr_user_files = 0;
5641 ret = io_sqe_files_scm(ctx);
5643 io_sqe_files_unregister(ctx);
5648 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
5651 #if defined(CONFIG_UNIX)
5652 struct sock *sock = ctx->ring_sock->sk;
5653 struct sk_buff_head *head = &sock->sk_receive_queue;
5654 struct sk_buff *skb;
5657 * See if we can merge this file into an existing skb SCM_RIGHTS
5658 * file set. If there's no room, fall back to allocating a new skb
5659 * and filling it in.
5661 spin_lock_irq(&head->lock);
5662 skb = skb_peek(head);
5664 struct scm_fp_list *fpl = UNIXCB(skb).fp;
5666 if (fpl->count < SCM_MAX_FD) {
5667 __skb_unlink(skb, head);
5668 spin_unlock_irq(&head->lock);
5669 fpl->fp[fpl->count] = get_file(file);
5670 unix_inflight(fpl->user, fpl->fp[fpl->count]);
5672 spin_lock_irq(&head->lock);
5673 __skb_queue_head(head, skb);
5678 spin_unlock_irq(&head->lock);
5685 return __io_sqe_files_scm(ctx, 1, index);
5691 static void io_atomic_switch(struct percpu_ref *ref)
5693 struct fixed_file_data *data;
5695 data = container_of(ref, struct fixed_file_data, refs);
5696 clear_bit(FFD_F_ATOMIC, &data->state);
5699 static bool io_queue_file_removal(struct fixed_file_data *data,
5702 struct io_file_put *pfile, pfile_stack;
5703 DECLARE_COMPLETION_ONSTACK(done);
5706 * If we fail allocating the struct we need for doing async reomval
5707 * of this file, just punt to sync and wait for it.
5709 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
5711 pfile = &pfile_stack;
5712 pfile->done = &done;
5716 llist_add(&pfile->llist, &data->put_llist);
5718 if (pfile == &pfile_stack) {
5719 if (!test_and_set_bit(FFD_F_ATOMIC, &data->state)) {
5720 percpu_ref_put(&data->refs);
5721 percpu_ref_switch_to_atomic(&data->refs,
5724 wait_for_completion(&done);
5725 flush_work(&data->ref_work);
5732 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
5733 struct io_uring_files_update *up,
5736 struct fixed_file_data *data = ctx->file_data;
5737 bool ref_switch = false;
5743 if (check_add_overflow(up->offset, nr_args, &done))
5745 if (done > ctx->nr_user_files)
5749 fds = u64_to_user_ptr(up->fds);
5751 struct fixed_file_table *table;
5755 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
5759 i = array_index_nospec(up->offset, ctx->nr_user_files);
5760 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5761 index = i & IORING_FILE_TABLE_MASK;
5762 if (table->files[index]) {
5763 file = io_file_from_index(ctx, index);
5764 table->files[index] = NULL;
5765 if (io_queue_file_removal(data, file))
5775 * Don't allow io_uring instances to be registered. If
5776 * UNIX isn't enabled, then this causes a reference
5777 * cycle and this instance can never get freed. If UNIX
5778 * is enabled we'll handle it just fine, but there's
5779 * still no point in allowing a ring fd as it doesn't
5780 * support regular read/write anyway.
5782 if (file->f_op == &io_uring_fops) {
5787 table->files[index] = file;
5788 err = io_sqe_file_register(ctx, file, i);
5797 if (ref_switch && !test_and_set_bit(FFD_F_ATOMIC, &data->state)) {
5798 percpu_ref_put(&data->refs);
5799 percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch);
5802 return done ? done : err;
5804 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
5807 struct io_uring_files_update up;
5809 if (!ctx->file_data)
5813 if (copy_from_user(&up, arg, sizeof(up)))
5818 return __io_sqe_files_update(ctx, &up, nr_args);
5821 static void io_put_work(struct io_wq_work *work)
5823 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5828 static void io_get_work(struct io_wq_work *work)
5830 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5832 refcount_inc(&req->refs);
5835 static int io_init_wq_offload(struct io_ring_ctx *ctx,
5836 struct io_uring_params *p)
5838 struct io_wq_data data;
5840 struct io_ring_ctx *ctx_attach;
5841 unsigned int concurrency;
5844 data.user = ctx->user;
5845 data.get_work = io_get_work;
5846 data.put_work = io_put_work;
5848 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
5849 /* Do QD, or 4 * CPUS, whatever is smallest */
5850 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
5852 ctx->io_wq = io_wq_create(concurrency, &data);
5853 if (IS_ERR(ctx->io_wq)) {
5854 ret = PTR_ERR(ctx->io_wq);
5860 f = fdget(p->wq_fd);
5864 if (f.file->f_op != &io_uring_fops) {
5869 ctx_attach = f.file->private_data;
5870 /* @io_wq is protected by holding the fd */
5871 if (!io_wq_get(ctx_attach->io_wq, &data)) {
5876 ctx->io_wq = ctx_attach->io_wq;
5882 static int io_sq_offload_start(struct io_ring_ctx *ctx,
5883 struct io_uring_params *p)
5887 init_waitqueue_head(&ctx->sqo_wait);
5888 mmgrab(current->mm);
5889 ctx->sqo_mm = current->mm;
5891 if (ctx->flags & IORING_SETUP_SQPOLL) {
5893 if (!capable(CAP_SYS_ADMIN))
5896 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
5897 if (!ctx->sq_thread_idle)
5898 ctx->sq_thread_idle = HZ;
5900 if (p->flags & IORING_SETUP_SQ_AFF) {
5901 int cpu = p->sq_thread_cpu;
5904 if (cpu >= nr_cpu_ids)
5906 if (!cpu_online(cpu))
5909 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
5913 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
5916 if (IS_ERR(ctx->sqo_thread)) {
5917 ret = PTR_ERR(ctx->sqo_thread);
5918 ctx->sqo_thread = NULL;
5921 wake_up_process(ctx->sqo_thread);
5922 } else if (p->flags & IORING_SETUP_SQ_AFF) {
5923 /* Can't have SQ_AFF without SQPOLL */
5928 ret = io_init_wq_offload(ctx, p);
5934 io_finish_async(ctx);
5935 mmdrop(ctx->sqo_mm);
5940 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
5942 atomic_long_sub(nr_pages, &user->locked_vm);
5945 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
5947 unsigned long page_limit, cur_pages, new_pages;
5949 /* Don't allow more pages than we can safely lock */
5950 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
5953 cur_pages = atomic_long_read(&user->locked_vm);
5954 new_pages = cur_pages + nr_pages;
5955 if (new_pages > page_limit)
5957 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
5958 new_pages) != cur_pages);
5963 static void io_mem_free(void *ptr)
5970 page = virt_to_head_page(ptr);
5971 if (put_page_testzero(page))
5972 free_compound_page(page);
5975 static void *io_mem_alloc(size_t size)
5977 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
5980 return (void *) __get_free_pages(gfp_flags, get_order(size));
5983 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
5986 struct io_rings *rings;
5987 size_t off, sq_array_size;
5989 off = struct_size(rings, cqes, cq_entries);
5990 if (off == SIZE_MAX)
5994 off = ALIGN(off, SMP_CACHE_BYTES);
5999 sq_array_size = array_size(sizeof(u32), sq_entries);
6000 if (sq_array_size == SIZE_MAX)
6003 if (check_add_overflow(off, sq_array_size, &off))
6012 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6016 pages = (size_t)1 << get_order(
6017 rings_size(sq_entries, cq_entries, NULL));
6018 pages += (size_t)1 << get_order(
6019 array_size(sizeof(struct io_uring_sqe), sq_entries));
6024 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6028 if (!ctx->user_bufs)
6031 for (i = 0; i < ctx->nr_user_bufs; i++) {
6032 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6034 for (j = 0; j < imu->nr_bvecs; j++)
6035 put_user_page(imu->bvec[j].bv_page);
6037 if (ctx->account_mem)
6038 io_unaccount_mem(ctx->user, imu->nr_bvecs);
6043 kfree(ctx->user_bufs);
6044 ctx->user_bufs = NULL;
6045 ctx->nr_user_bufs = 0;
6049 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
6050 void __user *arg, unsigned index)
6052 struct iovec __user *src;
6054 #ifdef CONFIG_COMPAT
6056 struct compat_iovec __user *ciovs;
6057 struct compat_iovec ciov;
6059 ciovs = (struct compat_iovec __user *) arg;
6060 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
6063 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
6064 dst->iov_len = ciov.iov_len;
6068 src = (struct iovec __user *) arg;
6069 if (copy_from_user(dst, &src[index], sizeof(*dst)))
6074 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
6077 struct vm_area_struct **vmas = NULL;
6078 struct page **pages = NULL;
6079 int i, j, got_pages = 0;
6084 if (!nr_args || nr_args > UIO_MAXIOV)
6087 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
6089 if (!ctx->user_bufs)
6092 for (i = 0; i < nr_args; i++) {
6093 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6094 unsigned long off, start, end, ubuf;
6099 ret = io_copy_iov(ctx, &iov, arg, i);
6104 * Don't impose further limits on the size and buffer
6105 * constraints here, we'll -EINVAL later when IO is
6106 * submitted if they are wrong.
6109 if (!iov.iov_base || !iov.iov_len)
6112 /* arbitrary limit, but we need something */
6113 if (iov.iov_len > SZ_1G)
6116 ubuf = (unsigned long) iov.iov_base;
6117 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
6118 start = ubuf >> PAGE_SHIFT;
6119 nr_pages = end - start;
6121 if (ctx->account_mem) {
6122 ret = io_account_mem(ctx->user, nr_pages);
6128 if (!pages || nr_pages > got_pages) {
6131 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
6133 vmas = kvmalloc_array(nr_pages,
6134 sizeof(struct vm_area_struct *),
6136 if (!pages || !vmas) {
6138 if (ctx->account_mem)
6139 io_unaccount_mem(ctx->user, nr_pages);
6142 got_pages = nr_pages;
6145 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
6149 if (ctx->account_mem)
6150 io_unaccount_mem(ctx->user, nr_pages);
6155 down_read(¤t->mm->mmap_sem);
6156 pret = get_user_pages(ubuf, nr_pages,
6157 FOLL_WRITE | FOLL_LONGTERM,
6159 if (pret == nr_pages) {
6160 /* don't support file backed memory */
6161 for (j = 0; j < nr_pages; j++) {
6162 struct vm_area_struct *vma = vmas[j];
6165 !is_file_hugepages(vma->vm_file)) {
6171 ret = pret < 0 ? pret : -EFAULT;
6173 up_read(¤t->mm->mmap_sem);
6176 * if we did partial map, or found file backed vmas,
6177 * release any pages we did get
6180 put_user_pages(pages, pret);
6181 if (ctx->account_mem)
6182 io_unaccount_mem(ctx->user, nr_pages);
6187 off = ubuf & ~PAGE_MASK;
6189 for (j = 0; j < nr_pages; j++) {
6192 vec_len = min_t(size_t, size, PAGE_SIZE - off);
6193 imu->bvec[j].bv_page = pages[j];
6194 imu->bvec[j].bv_len = vec_len;
6195 imu->bvec[j].bv_offset = off;
6199 /* store original address for later verification */
6201 imu->len = iov.iov_len;
6202 imu->nr_bvecs = nr_pages;
6204 ctx->nr_user_bufs++;
6212 io_sqe_buffer_unregister(ctx);
6216 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
6218 __s32 __user *fds = arg;
6224 if (copy_from_user(&fd, fds, sizeof(*fds)))
6227 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
6228 if (IS_ERR(ctx->cq_ev_fd)) {
6229 int ret = PTR_ERR(ctx->cq_ev_fd);
6230 ctx->cq_ev_fd = NULL;
6237 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
6239 if (ctx->cq_ev_fd) {
6240 eventfd_ctx_put(ctx->cq_ev_fd);
6241 ctx->cq_ev_fd = NULL;
6248 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
6250 io_finish_async(ctx);
6252 mmdrop(ctx->sqo_mm);
6254 io_iopoll_reap_events(ctx);
6255 io_sqe_buffer_unregister(ctx);
6256 io_sqe_files_unregister(ctx);
6257 io_eventfd_unregister(ctx);
6259 #if defined(CONFIG_UNIX)
6260 if (ctx->ring_sock) {
6261 ctx->ring_sock->file = NULL; /* so that iput() is called */
6262 sock_release(ctx->ring_sock);
6266 io_mem_free(ctx->rings);
6267 io_mem_free(ctx->sq_sqes);
6269 percpu_ref_exit(&ctx->refs);
6270 if (ctx->account_mem)
6271 io_unaccount_mem(ctx->user,
6272 ring_pages(ctx->sq_entries, ctx->cq_entries));
6273 free_uid(ctx->user);
6274 put_cred(ctx->creds);
6275 kfree(ctx->completions);
6276 kfree(ctx->cancel_hash);
6277 kmem_cache_free(req_cachep, ctx->fallback_req);
6281 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
6283 struct io_ring_ctx *ctx = file->private_data;
6286 poll_wait(file, &ctx->cq_wait, wait);
6288 * synchronizes with barrier from wq_has_sleeper call in
6292 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
6293 ctx->rings->sq_ring_entries)
6294 mask |= EPOLLOUT | EPOLLWRNORM;
6295 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
6296 mask |= EPOLLIN | EPOLLRDNORM;
6301 static int io_uring_fasync(int fd, struct file *file, int on)
6303 struct io_ring_ctx *ctx = file->private_data;
6305 return fasync_helper(fd, file, on, &ctx->cq_fasync);
6308 static int io_remove_personalities(int id, void *p, void *data)
6310 struct io_ring_ctx *ctx = data;
6311 const struct cred *cred;
6313 cred = idr_remove(&ctx->personality_idr, id);
6319 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
6321 mutex_lock(&ctx->uring_lock);
6322 percpu_ref_kill(&ctx->refs);
6323 mutex_unlock(&ctx->uring_lock);
6325 io_kill_timeouts(ctx);
6326 io_poll_remove_all(ctx);
6329 io_wq_cancel_all(ctx->io_wq);
6331 io_iopoll_reap_events(ctx);
6332 /* if we failed setting up the ctx, we might not have any rings */
6334 io_cqring_overflow_flush(ctx, true);
6335 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
6336 wait_for_completion(&ctx->completions[0]);
6337 io_ring_ctx_free(ctx);
6340 static int io_uring_release(struct inode *inode, struct file *file)
6342 struct io_ring_ctx *ctx = file->private_data;
6344 file->private_data = NULL;
6345 io_ring_ctx_wait_and_kill(ctx);
6349 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
6350 struct files_struct *files)
6352 struct io_kiocb *req;
6355 while (!list_empty_careful(&ctx->inflight_list)) {
6356 struct io_kiocb *cancel_req = NULL;
6358 spin_lock_irq(&ctx->inflight_lock);
6359 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
6360 if (req->work.files != files)
6362 /* req is being completed, ignore */
6363 if (!refcount_inc_not_zero(&req->refs))
6369 prepare_to_wait(&ctx->inflight_wait, &wait,
6370 TASK_UNINTERRUPTIBLE);
6371 spin_unlock_irq(&ctx->inflight_lock);
6373 /* We need to keep going until we don't find a matching req */
6377 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
6378 io_put_req(cancel_req);
6381 finish_wait(&ctx->inflight_wait, &wait);
6384 static int io_uring_flush(struct file *file, void *data)
6386 struct io_ring_ctx *ctx = file->private_data;
6388 io_uring_cancel_files(ctx, data);
6392 static void *io_uring_validate_mmap_request(struct file *file,
6393 loff_t pgoff, size_t sz)
6395 struct io_ring_ctx *ctx = file->private_data;
6396 loff_t offset = pgoff << PAGE_SHIFT;
6401 case IORING_OFF_SQ_RING:
6402 case IORING_OFF_CQ_RING:
6405 case IORING_OFF_SQES:
6409 return ERR_PTR(-EINVAL);
6412 page = virt_to_head_page(ptr);
6413 if (sz > page_size(page))
6414 return ERR_PTR(-EINVAL);
6421 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6423 size_t sz = vma->vm_end - vma->vm_start;
6427 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
6429 return PTR_ERR(ptr);
6431 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
6432 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
6435 #else /* !CONFIG_MMU */
6437 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6439 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
6442 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
6444 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
6447 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
6448 unsigned long addr, unsigned long len,
6449 unsigned long pgoff, unsigned long flags)
6453 ptr = io_uring_validate_mmap_request(file, pgoff, len);
6455 return PTR_ERR(ptr);
6457 return (unsigned long) ptr;
6460 #endif /* !CONFIG_MMU */
6462 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
6463 u32, min_complete, u32, flags, const sigset_t __user *, sig,
6466 struct io_ring_ctx *ctx;
6471 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
6479 if (f.file->f_op != &io_uring_fops)
6483 ctx = f.file->private_data;
6484 if (!percpu_ref_tryget(&ctx->refs))
6488 * For SQ polling, the thread will do all submissions and completions.
6489 * Just return the requested submit count, and wake the thread if
6493 if (ctx->flags & IORING_SETUP_SQPOLL) {
6494 if (!list_empty_careful(&ctx->cq_overflow_list))
6495 io_cqring_overflow_flush(ctx, false);
6496 if (flags & IORING_ENTER_SQ_WAKEUP)
6497 wake_up(&ctx->sqo_wait);
6498 submitted = to_submit;
6499 } else if (to_submit) {
6500 struct mm_struct *cur_mm;
6502 mutex_lock(&ctx->uring_lock);
6503 /* already have mm, so io_submit_sqes() won't try to grab it */
6504 cur_mm = ctx->sqo_mm;
6505 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
6507 mutex_unlock(&ctx->uring_lock);
6509 if (submitted != to_submit)
6512 if (flags & IORING_ENTER_GETEVENTS) {
6513 unsigned nr_events = 0;
6515 min_complete = min(min_complete, ctx->cq_entries);
6517 if (ctx->flags & IORING_SETUP_IOPOLL) {
6518 ret = io_iopoll_check(ctx, &nr_events, min_complete);
6520 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
6525 percpu_ref_put(&ctx->refs);
6528 return submitted ? submitted : ret;
6531 static int io_uring_show_cred(int id, void *p, void *data)
6533 const struct cred *cred = p;
6534 struct seq_file *m = data;
6535 struct user_namespace *uns = seq_user_ns(m);
6536 struct group_info *gi;
6541 seq_printf(m, "%5d\n", id);
6542 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
6543 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
6544 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
6545 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
6546 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
6547 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
6548 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
6549 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
6550 seq_puts(m, "\n\tGroups:\t");
6551 gi = cred->group_info;
6552 for (g = 0; g < gi->ngroups; g++) {
6553 seq_put_decimal_ull(m, g ? " " : "",
6554 from_kgid_munged(uns, gi->gid[g]));
6556 seq_puts(m, "\n\tCapEff:\t");
6557 cap = cred->cap_effective;
6558 CAP_FOR_EACH_U32(__capi)
6559 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
6564 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
6568 mutex_lock(&ctx->uring_lock);
6569 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
6570 for (i = 0; i < ctx->nr_user_files; i++) {
6571 struct fixed_file_table *table;
6574 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6575 f = table->files[i & IORING_FILE_TABLE_MASK];
6577 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
6579 seq_printf(m, "%5u: <none>\n", i);
6581 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
6582 for (i = 0; i < ctx->nr_user_bufs; i++) {
6583 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
6585 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
6586 (unsigned int) buf->len);
6588 if (!idr_is_empty(&ctx->personality_idr)) {
6589 seq_printf(m, "Personalities:\n");
6590 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
6592 mutex_unlock(&ctx->uring_lock);
6595 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
6597 struct io_ring_ctx *ctx = f->private_data;
6599 if (percpu_ref_tryget(&ctx->refs)) {
6600 __io_uring_show_fdinfo(ctx, m);
6601 percpu_ref_put(&ctx->refs);
6605 static const struct file_operations io_uring_fops = {
6606 .release = io_uring_release,
6607 .flush = io_uring_flush,
6608 .mmap = io_uring_mmap,
6610 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
6611 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
6613 .poll = io_uring_poll,
6614 .fasync = io_uring_fasync,
6615 .show_fdinfo = io_uring_show_fdinfo,
6618 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
6619 struct io_uring_params *p)
6621 struct io_rings *rings;
6622 size_t size, sq_array_offset;
6624 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
6625 if (size == SIZE_MAX)
6628 rings = io_mem_alloc(size);
6633 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
6634 rings->sq_ring_mask = p->sq_entries - 1;
6635 rings->cq_ring_mask = p->cq_entries - 1;
6636 rings->sq_ring_entries = p->sq_entries;
6637 rings->cq_ring_entries = p->cq_entries;
6638 ctx->sq_mask = rings->sq_ring_mask;
6639 ctx->cq_mask = rings->cq_ring_mask;
6640 ctx->sq_entries = rings->sq_ring_entries;
6641 ctx->cq_entries = rings->cq_ring_entries;
6643 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
6644 if (size == SIZE_MAX) {
6645 io_mem_free(ctx->rings);
6650 ctx->sq_sqes = io_mem_alloc(size);
6651 if (!ctx->sq_sqes) {
6652 io_mem_free(ctx->rings);
6661 * Allocate an anonymous fd, this is what constitutes the application
6662 * visible backing of an io_uring instance. The application mmaps this
6663 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
6664 * we have to tie this fd to a socket for file garbage collection purposes.
6666 static int io_uring_get_fd(struct io_ring_ctx *ctx)
6671 #if defined(CONFIG_UNIX)
6672 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
6678 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
6682 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
6683 O_RDWR | O_CLOEXEC);
6686 ret = PTR_ERR(file);
6690 #if defined(CONFIG_UNIX)
6691 ctx->ring_sock->file = file;
6693 fd_install(ret, file);
6696 #if defined(CONFIG_UNIX)
6697 sock_release(ctx->ring_sock);
6698 ctx->ring_sock = NULL;
6703 static int io_uring_create(unsigned entries, struct io_uring_params *p)
6705 struct user_struct *user = NULL;
6706 struct io_ring_ctx *ctx;
6712 if (entries > IORING_MAX_ENTRIES) {
6713 if (!(p->flags & IORING_SETUP_CLAMP))
6715 entries = IORING_MAX_ENTRIES;
6719 * Use twice as many entries for the CQ ring. It's possible for the
6720 * application to drive a higher depth than the size of the SQ ring,
6721 * since the sqes are only used at submission time. This allows for
6722 * some flexibility in overcommitting a bit. If the application has
6723 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
6724 * of CQ ring entries manually.
6726 p->sq_entries = roundup_pow_of_two(entries);
6727 if (p->flags & IORING_SETUP_CQSIZE) {
6729 * If IORING_SETUP_CQSIZE is set, we do the same roundup
6730 * to a power-of-two, if it isn't already. We do NOT impose
6731 * any cq vs sq ring sizing.
6733 if (p->cq_entries < p->sq_entries)
6735 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
6736 if (!(p->flags & IORING_SETUP_CLAMP))
6738 p->cq_entries = IORING_MAX_CQ_ENTRIES;
6740 p->cq_entries = roundup_pow_of_two(p->cq_entries);
6742 p->cq_entries = 2 * p->sq_entries;
6745 user = get_uid(current_user());
6746 account_mem = !capable(CAP_IPC_LOCK);
6749 ret = io_account_mem(user,
6750 ring_pages(p->sq_entries, p->cq_entries));
6757 ctx = io_ring_ctx_alloc(p);
6760 io_unaccount_mem(user, ring_pages(p->sq_entries,
6765 ctx->compat = in_compat_syscall();
6766 ctx->account_mem = account_mem;
6768 ctx->creds = get_current_cred();
6770 ret = io_allocate_scq_urings(ctx, p);
6774 ret = io_sq_offload_start(ctx, p);
6778 memset(&p->sq_off, 0, sizeof(p->sq_off));
6779 p->sq_off.head = offsetof(struct io_rings, sq.head);
6780 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
6781 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
6782 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
6783 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
6784 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
6785 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
6787 memset(&p->cq_off, 0, sizeof(p->cq_off));
6788 p->cq_off.head = offsetof(struct io_rings, cq.head);
6789 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
6790 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
6791 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
6792 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
6793 p->cq_off.cqes = offsetof(struct io_rings, cqes);
6796 * Install ring fd as the very last thing, so we don't risk someone
6797 * having closed it before we finish setup
6799 ret = io_uring_get_fd(ctx);
6803 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
6804 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
6805 IORING_FEAT_CUR_PERSONALITY;
6806 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
6809 io_ring_ctx_wait_and_kill(ctx);
6814 * Sets up an aio uring context, and returns the fd. Applications asks for a
6815 * ring size, we return the actual sq/cq ring sizes (among other things) in the
6816 * params structure passed in.
6818 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
6820 struct io_uring_params p;
6824 if (copy_from_user(&p, params, sizeof(p)))
6826 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
6831 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
6832 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
6833 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
6836 ret = io_uring_create(entries, &p);
6840 if (copy_to_user(params, &p, sizeof(p)))
6846 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
6847 struct io_uring_params __user *, params)
6849 return io_uring_setup(entries, params);
6852 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
6854 struct io_uring_probe *p;
6858 size = struct_size(p, ops, nr_args);
6859 if (size == SIZE_MAX)
6861 p = kzalloc(size, GFP_KERNEL);
6866 if (copy_from_user(p, arg, size))
6869 if (memchr_inv(p, 0, size))
6872 p->last_op = IORING_OP_LAST - 1;
6873 if (nr_args > IORING_OP_LAST)
6874 nr_args = IORING_OP_LAST;
6876 for (i = 0; i < nr_args; i++) {
6878 if (!io_op_defs[i].not_supported)
6879 p->ops[i].flags = IO_URING_OP_SUPPORTED;
6884 if (copy_to_user(arg, p, size))
6891 static int io_register_personality(struct io_ring_ctx *ctx)
6893 const struct cred *creds = get_current_cred();
6896 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
6897 USHRT_MAX, GFP_KERNEL);
6903 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
6905 const struct cred *old_creds;
6907 old_creds = idr_remove(&ctx->personality_idr, id);
6909 put_cred(old_creds);
6916 static bool io_register_op_must_quiesce(int op)
6919 case IORING_UNREGISTER_FILES:
6920 case IORING_REGISTER_FILES_UPDATE:
6921 case IORING_REGISTER_PROBE:
6922 case IORING_REGISTER_PERSONALITY:
6923 case IORING_UNREGISTER_PERSONALITY:
6930 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
6931 void __user *arg, unsigned nr_args)
6932 __releases(ctx->uring_lock)
6933 __acquires(ctx->uring_lock)
6938 * We're inside the ring mutex, if the ref is already dying, then
6939 * someone else killed the ctx or is already going through
6940 * io_uring_register().
6942 if (percpu_ref_is_dying(&ctx->refs))
6945 if (io_register_op_must_quiesce(opcode)) {
6946 percpu_ref_kill(&ctx->refs);
6949 * Drop uring mutex before waiting for references to exit. If
6950 * another thread is currently inside io_uring_enter() it might
6951 * need to grab the uring_lock to make progress. If we hold it
6952 * here across the drain wait, then we can deadlock. It's safe
6953 * to drop the mutex here, since no new references will come in
6954 * after we've killed the percpu ref.
6956 mutex_unlock(&ctx->uring_lock);
6957 ret = wait_for_completion_interruptible(&ctx->completions[0]);
6958 mutex_lock(&ctx->uring_lock);
6960 percpu_ref_resurrect(&ctx->refs);
6967 case IORING_REGISTER_BUFFERS:
6968 ret = io_sqe_buffer_register(ctx, arg, nr_args);
6970 case IORING_UNREGISTER_BUFFERS:
6974 ret = io_sqe_buffer_unregister(ctx);
6976 case IORING_REGISTER_FILES:
6977 ret = io_sqe_files_register(ctx, arg, nr_args);
6979 case IORING_UNREGISTER_FILES:
6983 ret = io_sqe_files_unregister(ctx);
6985 case IORING_REGISTER_FILES_UPDATE:
6986 ret = io_sqe_files_update(ctx, arg, nr_args);
6988 case IORING_REGISTER_EVENTFD:
6989 case IORING_REGISTER_EVENTFD_ASYNC:
6993 ret = io_eventfd_register(ctx, arg);
6996 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
6997 ctx->eventfd_async = 1;
6999 ctx->eventfd_async = 0;
7001 case IORING_UNREGISTER_EVENTFD:
7005 ret = io_eventfd_unregister(ctx);
7007 case IORING_REGISTER_PROBE:
7009 if (!arg || nr_args > 256)
7011 ret = io_probe(ctx, arg, nr_args);
7013 case IORING_REGISTER_PERSONALITY:
7017 ret = io_register_personality(ctx);
7019 case IORING_UNREGISTER_PERSONALITY:
7023 ret = io_unregister_personality(ctx, nr_args);
7030 if (io_register_op_must_quiesce(opcode)) {
7031 /* bring the ctx back to life */
7032 percpu_ref_reinit(&ctx->refs);
7034 reinit_completion(&ctx->completions[0]);
7039 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
7040 void __user *, arg, unsigned int, nr_args)
7042 struct io_ring_ctx *ctx;
7051 if (f.file->f_op != &io_uring_fops)
7054 ctx = f.file->private_data;
7056 mutex_lock(&ctx->uring_lock);
7057 ret = __io_uring_register(ctx, opcode, arg, nr_args);
7058 mutex_unlock(&ctx->uring_lock);
7059 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
7060 ctx->cq_ev_fd != NULL, ret);
7066 static int __init io_uring_init(void)
7068 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
7069 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
7070 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
7073 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
7074 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
7075 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
7076 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
7077 BUILD_BUG_SQE_ELEM(1, __u8, flags);
7078 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
7079 BUILD_BUG_SQE_ELEM(4, __s32, fd);
7080 BUILD_BUG_SQE_ELEM(8, __u64, off);
7081 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
7082 BUILD_BUG_SQE_ELEM(16, __u64, addr);
7083 BUILD_BUG_SQE_ELEM(24, __u32, len);
7084 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
7085 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
7086 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
7087 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
7088 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
7089 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
7090 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
7091 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
7092 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
7093 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
7094 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
7095 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
7096 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
7097 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
7098 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
7099 BUILD_BUG_SQE_ELEM(42, __u16, personality);
7101 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
7102 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
7105 __initcall(io_uring_init);