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)
1023 if (!ctx->eventfd_async)
1025 return io_wq_current_is_worker() || in_interrupt();
1028 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1030 if (waitqueue_active(&ctx->wait))
1031 wake_up(&ctx->wait);
1032 if (waitqueue_active(&ctx->sqo_wait))
1033 wake_up(&ctx->sqo_wait);
1034 if (ctx->cq_ev_fd && io_should_trigger_evfd(ctx))
1035 eventfd_signal(ctx->cq_ev_fd, 1);
1038 /* Returns true if there are no backlogged entries after the flush */
1039 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1041 struct io_rings *rings = ctx->rings;
1042 struct io_uring_cqe *cqe;
1043 struct io_kiocb *req;
1044 unsigned long flags;
1048 if (list_empty_careful(&ctx->cq_overflow_list))
1050 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1051 rings->cq_ring_entries))
1055 spin_lock_irqsave(&ctx->completion_lock, flags);
1057 /* if force is set, the ring is going away. always drop after that */
1059 ctx->cq_overflow_flushed = 1;
1062 while (!list_empty(&ctx->cq_overflow_list)) {
1063 cqe = io_get_cqring(ctx);
1067 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1069 list_move(&req->list, &list);
1071 WRITE_ONCE(cqe->user_data, req->user_data);
1072 WRITE_ONCE(cqe->res, req->result);
1073 WRITE_ONCE(cqe->flags, 0);
1075 WRITE_ONCE(ctx->rings->cq_overflow,
1076 atomic_inc_return(&ctx->cached_cq_overflow));
1080 io_commit_cqring(ctx);
1082 clear_bit(0, &ctx->sq_check_overflow);
1083 clear_bit(0, &ctx->cq_check_overflow);
1085 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1086 io_cqring_ev_posted(ctx);
1088 while (!list_empty(&list)) {
1089 req = list_first_entry(&list, struct io_kiocb, list);
1090 list_del(&req->list);
1097 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1099 struct io_ring_ctx *ctx = req->ctx;
1100 struct io_uring_cqe *cqe;
1102 trace_io_uring_complete(ctx, req->user_data, res);
1105 * If we can't get a cq entry, userspace overflowed the
1106 * submission (by quite a lot). Increment the overflow count in
1109 cqe = io_get_cqring(ctx);
1111 WRITE_ONCE(cqe->user_data, req->user_data);
1112 WRITE_ONCE(cqe->res, res);
1113 WRITE_ONCE(cqe->flags, 0);
1114 } else if (ctx->cq_overflow_flushed) {
1115 WRITE_ONCE(ctx->rings->cq_overflow,
1116 atomic_inc_return(&ctx->cached_cq_overflow));
1118 if (list_empty(&ctx->cq_overflow_list)) {
1119 set_bit(0, &ctx->sq_check_overflow);
1120 set_bit(0, &ctx->cq_check_overflow);
1122 refcount_inc(&req->refs);
1124 list_add_tail(&req->list, &ctx->cq_overflow_list);
1128 static void io_cqring_add_event(struct io_kiocb *req, long res)
1130 struct io_ring_ctx *ctx = req->ctx;
1131 unsigned long flags;
1133 spin_lock_irqsave(&ctx->completion_lock, flags);
1134 io_cqring_fill_event(req, res);
1135 io_commit_cqring(ctx);
1136 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1138 io_cqring_ev_posted(ctx);
1141 static inline bool io_is_fallback_req(struct io_kiocb *req)
1143 return req == (struct io_kiocb *)
1144 ((unsigned long) req->ctx->fallback_req & ~1UL);
1147 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1149 struct io_kiocb *req;
1151 req = ctx->fallback_req;
1152 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
1158 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
1159 struct io_submit_state *state)
1161 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1162 struct io_kiocb *req;
1165 req = kmem_cache_alloc(req_cachep, gfp);
1168 } else if (!state->free_reqs) {
1172 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1173 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1176 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1177 * retry single alloc to be on the safe side.
1179 if (unlikely(ret <= 0)) {
1180 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1181 if (!state->reqs[0])
1185 state->free_reqs = ret - 1;
1187 req = state->reqs[0];
1189 req = state->reqs[state->cur_req];
1199 /* one is dropped after submission, the other at completion */
1200 refcount_set(&req->refs, 2);
1202 INIT_IO_WORK(&req->work, io_wq_submit_work);
1205 req = io_get_fallback_req(ctx);
1208 percpu_ref_put(&ctx->refs);
1212 static void __io_req_do_free(struct io_kiocb *req)
1214 if (likely(!io_is_fallback_req(req)))
1215 kmem_cache_free(req_cachep, req);
1217 clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req);
1220 static void __io_req_aux_free(struct io_kiocb *req)
1222 struct io_ring_ctx *ctx = req->ctx;
1226 if (req->flags & REQ_F_FIXED_FILE)
1227 percpu_ref_put(&ctx->file_data->refs);
1232 io_req_work_drop_env(req);
1235 static void __io_free_req(struct io_kiocb *req)
1237 __io_req_aux_free(req);
1239 if (req->flags & REQ_F_INFLIGHT) {
1240 struct io_ring_ctx *ctx = req->ctx;
1241 unsigned long flags;
1243 spin_lock_irqsave(&ctx->inflight_lock, flags);
1244 list_del(&req->inflight_entry);
1245 if (waitqueue_active(&ctx->inflight_wait))
1246 wake_up(&ctx->inflight_wait);
1247 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1250 percpu_ref_put(&req->ctx->refs);
1251 __io_req_do_free(req);
1255 void *reqs[IO_IOPOLL_BATCH];
1260 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1262 int fixed_refs = rb->to_free;
1266 if (rb->need_iter) {
1267 int i, inflight = 0;
1268 unsigned long flags;
1271 for (i = 0; i < rb->to_free; i++) {
1272 struct io_kiocb *req = rb->reqs[i];
1274 if (req->flags & REQ_F_FIXED_FILE) {
1278 if (req->flags & REQ_F_INFLIGHT)
1280 __io_req_aux_free(req);
1285 spin_lock_irqsave(&ctx->inflight_lock, flags);
1286 for (i = 0; i < rb->to_free; i++) {
1287 struct io_kiocb *req = rb->reqs[i];
1289 if (req->flags & REQ_F_INFLIGHT) {
1290 list_del(&req->inflight_entry);
1295 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1297 if (waitqueue_active(&ctx->inflight_wait))
1298 wake_up(&ctx->inflight_wait);
1301 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1303 percpu_ref_put_many(&ctx->file_data->refs, fixed_refs);
1304 percpu_ref_put_many(&ctx->refs, rb->to_free);
1305 rb->to_free = rb->need_iter = 0;
1308 static bool io_link_cancel_timeout(struct io_kiocb *req)
1310 struct io_ring_ctx *ctx = req->ctx;
1313 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1315 io_cqring_fill_event(req, -ECANCELED);
1316 io_commit_cqring(ctx);
1317 req->flags &= ~REQ_F_LINK;
1325 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1327 struct io_ring_ctx *ctx = req->ctx;
1328 bool wake_ev = false;
1330 /* Already got next link */
1331 if (req->flags & REQ_F_LINK_NEXT)
1335 * The list should never be empty when we are called here. But could
1336 * potentially happen if the chain is messed up, check to be on the
1339 while (!list_empty(&req->link_list)) {
1340 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1341 struct io_kiocb, link_list);
1343 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1344 (nxt->flags & REQ_F_TIMEOUT))) {
1345 list_del_init(&nxt->link_list);
1346 wake_ev |= io_link_cancel_timeout(nxt);
1347 req->flags &= ~REQ_F_LINK_TIMEOUT;
1351 list_del_init(&req->link_list);
1352 if (!list_empty(&nxt->link_list))
1353 nxt->flags |= REQ_F_LINK;
1358 req->flags |= REQ_F_LINK_NEXT;
1360 io_cqring_ev_posted(ctx);
1364 * Called if REQ_F_LINK is set, and we fail the head request
1366 static void io_fail_links(struct io_kiocb *req)
1368 struct io_ring_ctx *ctx = req->ctx;
1369 unsigned long flags;
1371 spin_lock_irqsave(&ctx->completion_lock, flags);
1373 while (!list_empty(&req->link_list)) {
1374 struct io_kiocb *link = list_first_entry(&req->link_list,
1375 struct io_kiocb, link_list);
1377 list_del_init(&link->link_list);
1378 trace_io_uring_fail_link(req, link);
1380 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1381 link->opcode == IORING_OP_LINK_TIMEOUT) {
1382 io_link_cancel_timeout(link);
1384 io_cqring_fill_event(link, -ECANCELED);
1385 __io_double_put_req(link);
1387 req->flags &= ~REQ_F_LINK_TIMEOUT;
1390 io_commit_cqring(ctx);
1391 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1392 io_cqring_ev_posted(ctx);
1395 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1397 if (likely(!(req->flags & REQ_F_LINK)))
1401 * If LINK is set, we have dependent requests in this chain. If we
1402 * didn't fail this request, queue the first one up, moving any other
1403 * dependencies to the next request. In case of failure, fail the rest
1406 if (req->flags & REQ_F_FAIL_LINK) {
1408 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1409 REQ_F_LINK_TIMEOUT) {
1410 struct io_ring_ctx *ctx = req->ctx;
1411 unsigned long flags;
1414 * If this is a timeout link, we could be racing with the
1415 * timeout timer. Grab the completion lock for this case to
1416 * protect against that.
1418 spin_lock_irqsave(&ctx->completion_lock, flags);
1419 io_req_link_next(req, nxt);
1420 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1422 io_req_link_next(req, nxt);
1426 static void io_free_req(struct io_kiocb *req)
1428 struct io_kiocb *nxt = NULL;
1430 io_req_find_next(req, &nxt);
1434 io_queue_async_work(nxt);
1438 * Drop reference to request, return next in chain (if there is one) if this
1439 * was the last reference to this request.
1441 __attribute__((nonnull))
1442 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1444 io_req_find_next(req, nxtptr);
1446 if (refcount_dec_and_test(&req->refs))
1450 static void io_put_req(struct io_kiocb *req)
1452 if (refcount_dec_and_test(&req->refs))
1457 * Must only be used if we don't need to care about links, usually from
1458 * within the completion handling itself.
1460 static void __io_double_put_req(struct io_kiocb *req)
1462 /* drop both submit and complete references */
1463 if (refcount_sub_and_test(2, &req->refs))
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 unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1476 struct io_rings *rings = ctx->rings;
1478 if (test_bit(0, &ctx->cq_check_overflow)) {
1480 * noflush == true is from the waitqueue handler, just ensure
1481 * we wake up the task, and the next invocation will flush the
1482 * entries. We cannot safely to it from here.
1484 if (noflush && !list_empty(&ctx->cq_overflow_list))
1487 io_cqring_overflow_flush(ctx, false);
1490 /* See comment at the top of this file */
1492 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1495 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1497 struct io_rings *rings = ctx->rings;
1499 /* make sure SQ entry isn't read before tail */
1500 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1503 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1505 if ((req->flags & REQ_F_LINK) || io_is_fallback_req(req))
1508 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1511 rb->reqs[rb->to_free++] = req;
1512 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1513 io_free_req_many(req->ctx, rb);
1518 * Find and free completed poll iocbs
1520 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1521 struct list_head *done)
1523 struct req_batch rb;
1524 struct io_kiocb *req;
1526 rb.to_free = rb.need_iter = 0;
1527 while (!list_empty(done)) {
1528 req = list_first_entry(done, struct io_kiocb, list);
1529 list_del(&req->list);
1531 io_cqring_fill_event(req, req->result);
1534 if (refcount_dec_and_test(&req->refs) &&
1535 !io_req_multi_free(&rb, req))
1539 io_commit_cqring(ctx);
1540 io_free_req_many(ctx, &rb);
1543 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1546 struct io_kiocb *req, *tmp;
1552 * Only spin for completions if we don't have multiple devices hanging
1553 * off our complete list, and we're under the requested amount.
1555 spin = !ctx->poll_multi_file && *nr_events < min;
1558 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1559 struct kiocb *kiocb = &req->rw.kiocb;
1562 * Move completed entries to our local list. If we find a
1563 * request that requires polling, break out and complete
1564 * the done list first, if we have entries there.
1566 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1567 list_move_tail(&req->list, &done);
1570 if (!list_empty(&done))
1573 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1582 if (!list_empty(&done))
1583 io_iopoll_complete(ctx, nr_events, &done);
1589 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1590 * non-spinning poll check - we'll still enter the driver poll loop, but only
1591 * as a non-spinning completion check.
1593 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1596 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1599 ret = io_do_iopoll(ctx, nr_events, min);
1602 if (!min || *nr_events >= min)
1610 * We can't just wait for polled events to come to us, we have to actively
1611 * find and complete them.
1613 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1615 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1618 mutex_lock(&ctx->uring_lock);
1619 while (!list_empty(&ctx->poll_list)) {
1620 unsigned int nr_events = 0;
1622 io_iopoll_getevents(ctx, &nr_events, 1);
1625 * Ensure we allow local-to-the-cpu processing to take place,
1626 * in this case we need to ensure that we reap all events.
1630 mutex_unlock(&ctx->uring_lock);
1633 static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1636 int iters = 0, ret = 0;
1642 * Don't enter poll loop if we already have events pending.
1643 * If we do, we can potentially be spinning for commands that
1644 * already triggered a CQE (eg in error).
1646 if (io_cqring_events(ctx, false))
1650 * If a submit got punted to a workqueue, we can have the
1651 * application entering polling for a command before it gets
1652 * issued. That app will hold the uring_lock for the duration
1653 * of the poll right here, so we need to take a breather every
1654 * now and then to ensure that the issue has a chance to add
1655 * the poll to the issued list. Otherwise we can spin here
1656 * forever, while the workqueue is stuck trying to acquire the
1659 if (!(++iters & 7)) {
1660 mutex_unlock(&ctx->uring_lock);
1661 mutex_lock(&ctx->uring_lock);
1664 if (*nr_events < min)
1665 tmin = min - *nr_events;
1667 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1671 } while (min && !*nr_events && !need_resched());
1676 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1682 * We disallow the app entering submit/complete with polling, but we
1683 * still need to lock the ring to prevent racing with polled issue
1684 * that got punted to a workqueue.
1686 mutex_lock(&ctx->uring_lock);
1687 ret = __io_iopoll_check(ctx, nr_events, min);
1688 mutex_unlock(&ctx->uring_lock);
1692 static void kiocb_end_write(struct io_kiocb *req)
1695 * Tell lockdep we inherited freeze protection from submission
1698 if (req->flags & REQ_F_ISREG) {
1699 struct inode *inode = file_inode(req->file);
1701 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1703 file_end_write(req->file);
1706 static inline void req_set_fail_links(struct io_kiocb *req)
1708 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1709 req->flags |= REQ_F_FAIL_LINK;
1712 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1714 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1716 if (kiocb->ki_flags & IOCB_WRITE)
1717 kiocb_end_write(req);
1719 if (res != req->result)
1720 req_set_fail_links(req);
1721 io_cqring_add_event(req, res);
1724 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1726 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1728 io_complete_rw_common(kiocb, res);
1732 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1734 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1735 struct io_kiocb *nxt = NULL;
1737 io_complete_rw_common(kiocb, res);
1738 io_put_req_find_next(req, &nxt);
1743 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1745 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1747 if (kiocb->ki_flags & IOCB_WRITE)
1748 kiocb_end_write(req);
1750 if (res != req->result)
1751 req_set_fail_links(req);
1754 req->flags |= REQ_F_IOPOLL_COMPLETED;
1758 * After the iocb has been issued, it's safe to be found on the poll list.
1759 * Adding the kiocb to the list AFTER submission ensures that we don't
1760 * find it from a io_iopoll_getevents() thread before the issuer is done
1761 * accessing the kiocb cookie.
1763 static void io_iopoll_req_issued(struct io_kiocb *req)
1765 struct io_ring_ctx *ctx = req->ctx;
1768 * Track whether we have multiple files in our lists. This will impact
1769 * how we do polling eventually, not spinning if we're on potentially
1770 * different devices.
1772 if (list_empty(&ctx->poll_list)) {
1773 ctx->poll_multi_file = false;
1774 } else if (!ctx->poll_multi_file) {
1775 struct io_kiocb *list_req;
1777 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1779 if (list_req->file != req->file)
1780 ctx->poll_multi_file = true;
1784 * For fast devices, IO may have already completed. If it has, add
1785 * it to the front so we find it first.
1787 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1788 list_add(&req->list, &ctx->poll_list);
1790 list_add_tail(&req->list, &ctx->poll_list);
1793 static void io_file_put(struct io_submit_state *state)
1796 int diff = state->has_refs - state->used_refs;
1799 fput_many(state->file, diff);
1805 * Get as many references to a file as we have IOs left in this submission,
1806 * assuming most submissions are for one file, or at least that each file
1807 * has more than one submission.
1809 static struct file *io_file_get(struct io_submit_state *state, int fd)
1815 if (state->fd == fd) {
1822 state->file = fget_many(fd, state->ios_left);
1827 state->has_refs = state->ios_left;
1828 state->used_refs = 1;
1834 * If we tracked the file through the SCM inflight mechanism, we could support
1835 * any file. For now, just ensure that anything potentially problematic is done
1838 static bool io_file_supports_async(struct file *file)
1840 umode_t mode = file_inode(file)->i_mode;
1842 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
1844 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1850 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1851 bool force_nonblock)
1853 struct io_ring_ctx *ctx = req->ctx;
1854 struct kiocb *kiocb = &req->rw.kiocb;
1861 if (S_ISREG(file_inode(req->file)->i_mode))
1862 req->flags |= REQ_F_ISREG;
1864 kiocb->ki_pos = READ_ONCE(sqe->off);
1865 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
1866 req->flags |= REQ_F_CUR_POS;
1867 kiocb->ki_pos = req->file->f_pos;
1869 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1870 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1872 ioprio = READ_ONCE(sqe->ioprio);
1874 ret = ioprio_check_cap(ioprio);
1878 kiocb->ki_ioprio = ioprio;
1880 kiocb->ki_ioprio = get_current_ioprio();
1882 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1886 /* don't allow async punt if RWF_NOWAIT was requested */
1887 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1888 (req->file->f_flags & O_NONBLOCK))
1889 req->flags |= REQ_F_NOWAIT;
1892 kiocb->ki_flags |= IOCB_NOWAIT;
1894 if (ctx->flags & IORING_SETUP_IOPOLL) {
1895 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1896 !kiocb->ki_filp->f_op->iopoll)
1899 kiocb->ki_flags |= IOCB_HIPRI;
1900 kiocb->ki_complete = io_complete_rw_iopoll;
1903 if (kiocb->ki_flags & IOCB_HIPRI)
1905 kiocb->ki_complete = io_complete_rw;
1908 req->rw.addr = READ_ONCE(sqe->addr);
1909 req->rw.len = READ_ONCE(sqe->len);
1910 /* we own ->private, reuse it for the buffer index */
1911 req->rw.kiocb.private = (void *) (unsigned long)
1912 READ_ONCE(sqe->buf_index);
1916 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1922 case -ERESTARTNOINTR:
1923 case -ERESTARTNOHAND:
1924 case -ERESTART_RESTARTBLOCK:
1926 * We can't just restart the syscall, since previously
1927 * submitted sqes may already be in progress. Just fail this
1933 kiocb->ki_complete(kiocb, ret, 0);
1937 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1940 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1942 if (req->flags & REQ_F_CUR_POS)
1943 req->file->f_pos = kiocb->ki_pos;
1944 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1945 *nxt = __io_complete_rw(kiocb, ret);
1947 io_rw_done(kiocb, ret);
1950 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
1951 struct iov_iter *iter)
1953 struct io_ring_ctx *ctx = req->ctx;
1954 size_t len = req->rw.len;
1955 struct io_mapped_ubuf *imu;
1956 unsigned index, buf_index;
1960 /* attempt to use fixed buffers without having provided iovecs */
1961 if (unlikely(!ctx->user_bufs))
1964 buf_index = (unsigned long) req->rw.kiocb.private;
1965 if (unlikely(buf_index >= ctx->nr_user_bufs))
1968 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1969 imu = &ctx->user_bufs[index];
1970 buf_addr = req->rw.addr;
1973 if (buf_addr + len < buf_addr)
1975 /* not inside the mapped region */
1976 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1980 * May not be a start of buffer, set size appropriately
1981 * and advance us to the beginning.
1983 offset = buf_addr - imu->ubuf;
1984 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1988 * Don't use iov_iter_advance() here, as it's really slow for
1989 * using the latter parts of a big fixed buffer - it iterates
1990 * over each segment manually. We can cheat a bit here, because
1993 * 1) it's a BVEC iter, we set it up
1994 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1995 * first and last bvec
1997 * So just find our index, and adjust the iterator afterwards.
1998 * If the offset is within the first bvec (or the whole first
1999 * bvec, just use iov_iter_advance(). This makes it easier
2000 * since we can just skip the first segment, which may not
2001 * be PAGE_SIZE aligned.
2003 const struct bio_vec *bvec = imu->bvec;
2005 if (offset <= bvec->bv_len) {
2006 iov_iter_advance(iter, offset);
2008 unsigned long seg_skip;
2010 /* skip first vec */
2011 offset -= bvec->bv_len;
2012 seg_skip = 1 + (offset >> PAGE_SHIFT);
2014 iter->bvec = bvec + seg_skip;
2015 iter->nr_segs -= seg_skip;
2016 iter->count -= bvec->bv_len + offset;
2017 iter->iov_offset = offset & ~PAGE_MASK;
2024 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2025 struct iovec **iovec, struct iov_iter *iter)
2027 void __user *buf = u64_to_user_ptr(req->rw.addr);
2028 size_t sqe_len = req->rw.len;
2031 opcode = req->opcode;
2032 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2034 return io_import_fixed(req, rw, iter);
2037 /* buffer index only valid with fixed read/write */
2038 if (req->rw.kiocb.private)
2041 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2043 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2049 struct io_async_rw *iorw = &req->io->rw;
2052 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2053 if (iorw->iov == iorw->fast_iov)
2061 #ifdef CONFIG_COMPAT
2062 if (req->ctx->compat)
2063 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2067 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2071 * For files that don't have ->read_iter() and ->write_iter(), handle them
2072 * by looping over ->read() or ->write() manually.
2074 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2075 struct iov_iter *iter)
2080 * Don't support polled IO through this interface, and we can't
2081 * support non-blocking either. For the latter, this just causes
2082 * the kiocb to be handled from an async context.
2084 if (kiocb->ki_flags & IOCB_HIPRI)
2086 if (kiocb->ki_flags & IOCB_NOWAIT)
2089 while (iov_iter_count(iter)) {
2093 if (!iov_iter_is_bvec(iter)) {
2094 iovec = iov_iter_iovec(iter);
2096 /* fixed buffers import bvec */
2097 iovec.iov_base = kmap(iter->bvec->bv_page)
2099 iovec.iov_len = min(iter->count,
2100 iter->bvec->bv_len - iter->iov_offset);
2104 nr = file->f_op->read(file, iovec.iov_base,
2105 iovec.iov_len, &kiocb->ki_pos);
2107 nr = file->f_op->write(file, iovec.iov_base,
2108 iovec.iov_len, &kiocb->ki_pos);
2111 if (iov_iter_is_bvec(iter))
2112 kunmap(iter->bvec->bv_page);
2120 if (nr != iovec.iov_len)
2122 iov_iter_advance(iter, nr);
2128 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2129 struct iovec *iovec, struct iovec *fast_iov,
2130 struct iov_iter *iter)
2132 req->io->rw.nr_segs = iter->nr_segs;
2133 req->io->rw.size = io_size;
2134 req->io->rw.iov = iovec;
2135 if (!req->io->rw.iov) {
2136 req->io->rw.iov = req->io->rw.fast_iov;
2137 memcpy(req->io->rw.iov, fast_iov,
2138 sizeof(struct iovec) * iter->nr_segs);
2142 static int io_alloc_async_ctx(struct io_kiocb *req)
2144 if (!io_op_defs[req->opcode].async_ctx)
2146 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2147 return req->io == NULL;
2150 static void io_rw_async(struct io_wq_work **workptr)
2152 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2153 struct iovec *iov = NULL;
2155 if (req->io->rw.iov != req->io->rw.fast_iov)
2156 iov = req->io->rw.iov;
2157 io_wq_submit_work(workptr);
2161 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2162 struct iovec *iovec, struct iovec *fast_iov,
2163 struct iov_iter *iter)
2165 if (!io_op_defs[req->opcode].async_ctx)
2167 if (!req->io && io_alloc_async_ctx(req))
2170 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2171 req->work.func = io_rw_async;
2175 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2176 bool force_nonblock)
2178 struct io_async_ctx *io;
2179 struct iov_iter iter;
2182 ret = io_prep_rw(req, sqe, force_nonblock);
2186 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2193 io->rw.iov = io->rw.fast_iov;
2195 ret = io_import_iovec(READ, req, &io->rw.iov, &iter);
2200 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2204 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
2205 bool force_nonblock)
2207 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2208 struct kiocb *kiocb = &req->rw.kiocb;
2209 struct iov_iter iter;
2211 ssize_t io_size, ret;
2213 ret = io_import_iovec(READ, req, &iovec, &iter);
2217 /* Ensure we clear previously set non-block flag */
2218 if (!force_nonblock)
2219 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2223 if (req->flags & REQ_F_LINK)
2224 req->result = io_size;
2227 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2228 * we know to async punt it even if it was opened O_NONBLOCK
2230 if (force_nonblock && !io_file_supports_async(req->file)) {
2231 req->flags |= REQ_F_MUST_PUNT;
2235 iov_count = iov_iter_count(&iter);
2236 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2240 if (req->file->f_op->read_iter)
2241 ret2 = call_read_iter(req->file, kiocb, &iter);
2243 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2245 /* Catch -EAGAIN return for forced non-blocking submission */
2246 if (!force_nonblock || ret2 != -EAGAIN) {
2247 kiocb_done(kiocb, ret2, nxt, req->in_async);
2250 ret = io_setup_async_rw(req, io_size, iovec,
2251 inline_vecs, &iter);
2258 if (!io_wq_current_is_worker())
2263 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2264 bool force_nonblock)
2266 struct io_async_ctx *io;
2267 struct iov_iter iter;
2270 ret = io_prep_rw(req, sqe, force_nonblock);
2274 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2281 io->rw.iov = io->rw.fast_iov;
2283 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter);
2288 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2292 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
2293 bool force_nonblock)
2295 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2296 struct kiocb *kiocb = &req->rw.kiocb;
2297 struct iov_iter iter;
2299 ssize_t ret, io_size;
2301 ret = io_import_iovec(WRITE, req, &iovec, &iter);
2305 /* Ensure we clear previously set non-block flag */
2306 if (!force_nonblock)
2307 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2311 if (req->flags & REQ_F_LINK)
2312 req->result = io_size;
2315 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2316 * we know to async punt it even if it was opened O_NONBLOCK
2318 if (force_nonblock && !io_file_supports_async(req->file)) {
2319 req->flags |= REQ_F_MUST_PUNT;
2323 /* file path doesn't support NOWAIT for non-direct_IO */
2324 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2325 (req->flags & REQ_F_ISREG))
2328 iov_count = iov_iter_count(&iter);
2329 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2334 * Open-code file_start_write here to grab freeze protection,
2335 * which will be released by another thread in
2336 * io_complete_rw(). Fool lockdep by telling it the lock got
2337 * released so that it doesn't complain about the held lock when
2338 * we return to userspace.
2340 if (req->flags & REQ_F_ISREG) {
2341 __sb_start_write(file_inode(req->file)->i_sb,
2342 SB_FREEZE_WRITE, true);
2343 __sb_writers_release(file_inode(req->file)->i_sb,
2346 kiocb->ki_flags |= IOCB_WRITE;
2348 if (req->file->f_op->write_iter)
2349 ret2 = call_write_iter(req->file, kiocb, &iter);
2351 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2352 if (!force_nonblock || ret2 != -EAGAIN) {
2353 kiocb_done(kiocb, ret2, nxt, req->in_async);
2356 ret = io_setup_async_rw(req, io_size, iovec,
2357 inline_vecs, &iter);
2364 if (!io_wq_current_is_worker())
2370 * IORING_OP_NOP just posts a completion event, nothing else.
2372 static int io_nop(struct io_kiocb *req)
2374 struct io_ring_ctx *ctx = req->ctx;
2376 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2379 io_cqring_add_event(req, 0);
2384 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2386 struct io_ring_ctx *ctx = req->ctx;
2391 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2393 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2396 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2397 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2400 req->sync.off = READ_ONCE(sqe->off);
2401 req->sync.len = READ_ONCE(sqe->len);
2405 static bool io_req_cancelled(struct io_kiocb *req)
2407 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2408 req_set_fail_links(req);
2409 io_cqring_add_event(req, -ECANCELED);
2417 static void io_link_work_cb(struct io_wq_work **workptr)
2419 struct io_wq_work *work = *workptr;
2420 struct io_kiocb *link = work->data;
2422 io_queue_linked_timeout(link);
2423 work->func = io_wq_submit_work;
2426 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
2428 struct io_kiocb *link;
2430 io_prep_async_work(nxt, &link);
2431 *workptr = &nxt->work;
2433 nxt->work.flags |= IO_WQ_WORK_CB;
2434 nxt->work.func = io_link_work_cb;
2435 nxt->work.data = link;
2439 static void io_fsync_finish(struct io_wq_work **workptr)
2441 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2442 loff_t end = req->sync.off + req->sync.len;
2443 struct io_kiocb *nxt = NULL;
2446 if (io_req_cancelled(req))
2449 ret = vfs_fsync_range(req->file, req->sync.off,
2450 end > 0 ? end : LLONG_MAX,
2451 req->sync.flags & IORING_FSYNC_DATASYNC);
2453 req_set_fail_links(req);
2454 io_cqring_add_event(req, ret);
2455 io_put_req_find_next(req, &nxt);
2457 io_wq_assign_next(workptr, nxt);
2460 static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt,
2461 bool force_nonblock)
2463 struct io_wq_work *work, *old_work;
2465 /* fsync always requires a blocking context */
2466 if (force_nonblock) {
2468 req->work.func = io_fsync_finish;
2472 work = old_work = &req->work;
2473 io_fsync_finish(&work);
2474 if (work && work != old_work)
2475 *nxt = container_of(work, struct io_kiocb, work);
2479 static void io_fallocate_finish(struct io_wq_work **workptr)
2481 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2482 struct io_kiocb *nxt = NULL;
2485 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2488 req_set_fail_links(req);
2489 io_cqring_add_event(req, ret);
2490 io_put_req_find_next(req, &nxt);
2492 io_wq_assign_next(workptr, nxt);
2495 static int io_fallocate_prep(struct io_kiocb *req,
2496 const struct io_uring_sqe *sqe)
2498 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2501 req->sync.off = READ_ONCE(sqe->off);
2502 req->sync.len = READ_ONCE(sqe->addr);
2503 req->sync.mode = READ_ONCE(sqe->len);
2507 static int io_fallocate(struct io_kiocb *req, struct io_kiocb **nxt,
2508 bool force_nonblock)
2510 struct io_wq_work *work, *old_work;
2512 /* fallocate always requiring blocking context */
2513 if (force_nonblock) {
2515 req->work.func = io_fallocate_finish;
2519 work = old_work = &req->work;
2520 io_fallocate_finish(&work);
2521 if (work && work != old_work)
2522 *nxt = container_of(work, struct io_kiocb, work);
2527 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2529 const char __user *fname;
2532 if (sqe->ioprio || sqe->buf_index)
2535 req->open.dfd = READ_ONCE(sqe->fd);
2536 req->open.how.mode = READ_ONCE(sqe->len);
2537 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2538 req->open.how.flags = READ_ONCE(sqe->open_flags);
2540 req->open.filename = getname(fname);
2541 if (IS_ERR(req->open.filename)) {
2542 ret = PTR_ERR(req->open.filename);
2543 req->open.filename = NULL;
2550 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2552 struct open_how __user *how;
2553 const char __user *fname;
2557 if (sqe->ioprio || sqe->buf_index)
2560 req->open.dfd = READ_ONCE(sqe->fd);
2561 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2562 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2563 len = READ_ONCE(sqe->len);
2565 if (len < OPEN_HOW_SIZE_VER0)
2568 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2573 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2574 req->open.how.flags |= O_LARGEFILE;
2576 req->open.filename = getname(fname);
2577 if (IS_ERR(req->open.filename)) {
2578 ret = PTR_ERR(req->open.filename);
2579 req->open.filename = NULL;
2586 static int io_openat2(struct io_kiocb *req, struct io_kiocb **nxt,
2587 bool force_nonblock)
2589 struct open_flags op;
2596 ret = build_open_flags(&req->open.how, &op);
2600 ret = get_unused_fd_flags(req->open.how.flags);
2604 file = do_filp_open(req->open.dfd, req->open.filename, &op);
2607 ret = PTR_ERR(file);
2609 fsnotify_open(file);
2610 fd_install(ret, file);
2613 putname(req->open.filename);
2615 req_set_fail_links(req);
2616 io_cqring_add_event(req, ret);
2617 io_put_req_find_next(req, nxt);
2621 static int io_openat(struct io_kiocb *req, struct io_kiocb **nxt,
2622 bool force_nonblock)
2624 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
2625 return io_openat2(req, nxt, force_nonblock);
2628 static int io_epoll_ctl_prep(struct io_kiocb *req,
2629 const struct io_uring_sqe *sqe)
2631 #if defined(CONFIG_EPOLL)
2632 if (sqe->ioprio || sqe->buf_index)
2635 req->epoll.epfd = READ_ONCE(sqe->fd);
2636 req->epoll.op = READ_ONCE(sqe->len);
2637 req->epoll.fd = READ_ONCE(sqe->off);
2639 if (ep_op_has_event(req->epoll.op)) {
2640 struct epoll_event __user *ev;
2642 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
2643 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
2653 static int io_epoll_ctl(struct io_kiocb *req, struct io_kiocb **nxt,
2654 bool force_nonblock)
2656 #if defined(CONFIG_EPOLL)
2657 struct io_epoll *ie = &req->epoll;
2660 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
2661 if (force_nonblock && ret == -EAGAIN)
2665 req_set_fail_links(req);
2666 io_cqring_add_event(req, ret);
2667 io_put_req_find_next(req, nxt);
2674 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2676 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2677 if (sqe->ioprio || sqe->buf_index || sqe->off)
2680 req->madvise.addr = READ_ONCE(sqe->addr);
2681 req->madvise.len = READ_ONCE(sqe->len);
2682 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
2689 static int io_madvise(struct io_kiocb *req, struct io_kiocb **nxt,
2690 bool force_nonblock)
2692 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2693 struct io_madvise *ma = &req->madvise;
2699 ret = do_madvise(ma->addr, ma->len, ma->advice);
2701 req_set_fail_links(req);
2702 io_cqring_add_event(req, ret);
2703 io_put_req_find_next(req, nxt);
2710 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2712 if (sqe->ioprio || sqe->buf_index || sqe->addr)
2715 req->fadvise.offset = READ_ONCE(sqe->off);
2716 req->fadvise.len = READ_ONCE(sqe->len);
2717 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
2721 static int io_fadvise(struct io_kiocb *req, struct io_kiocb **nxt,
2722 bool force_nonblock)
2724 struct io_fadvise *fa = &req->fadvise;
2727 /* DONTNEED may block, others _should_ not */
2728 if (fa->advice == POSIX_FADV_DONTNEED && force_nonblock)
2731 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
2733 req_set_fail_links(req);
2734 io_cqring_add_event(req, ret);
2735 io_put_req_find_next(req, nxt);
2739 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2741 const char __user *fname;
2742 unsigned lookup_flags;
2745 if (sqe->ioprio || sqe->buf_index)
2748 req->open.dfd = READ_ONCE(sqe->fd);
2749 req->open.mask = READ_ONCE(sqe->len);
2750 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2751 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2752 req->open.how.flags = READ_ONCE(sqe->statx_flags);
2754 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
2757 req->open.filename = getname_flags(fname, lookup_flags, NULL);
2758 if (IS_ERR(req->open.filename)) {
2759 ret = PTR_ERR(req->open.filename);
2760 req->open.filename = NULL;
2767 static int io_statx(struct io_kiocb *req, struct io_kiocb **nxt,
2768 bool force_nonblock)
2770 struct io_open *ctx = &req->open;
2771 unsigned lookup_flags;
2779 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
2783 /* filename_lookup() drops it, keep a reference */
2784 ctx->filename->refcnt++;
2786 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
2791 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
2793 if (retry_estale(ret, lookup_flags)) {
2794 lookup_flags |= LOOKUP_REVAL;
2798 ret = cp_statx(&stat, ctx->buffer);
2800 putname(ctx->filename);
2802 req_set_fail_links(req);
2803 io_cqring_add_event(req, ret);
2804 io_put_req_find_next(req, nxt);
2808 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2811 * If we queue this for async, it must not be cancellable. That would
2812 * leave the 'file' in an undeterminate state.
2814 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
2816 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
2817 sqe->rw_flags || sqe->buf_index)
2819 if (sqe->flags & IOSQE_FIXED_FILE)
2822 req->close.fd = READ_ONCE(sqe->fd);
2823 if (req->file->f_op == &io_uring_fops ||
2824 req->close.fd == req->ctx->ring_fd)
2830 static void io_close_finish(struct io_wq_work **workptr)
2832 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2833 struct io_kiocb *nxt = NULL;
2835 /* Invoked with files, we need to do the close */
2836 if (req->work.files) {
2839 ret = filp_close(req->close.put_file, req->work.files);
2841 req_set_fail_links(req);
2843 io_cqring_add_event(req, ret);
2846 fput(req->close.put_file);
2848 /* we bypassed the re-issue, drop the submission reference */
2850 io_put_req_find_next(req, &nxt);
2852 io_wq_assign_next(workptr, nxt);
2855 static int io_close(struct io_kiocb *req, struct io_kiocb **nxt,
2856 bool force_nonblock)
2860 req->close.put_file = NULL;
2861 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
2865 /* if the file has a flush method, be safe and punt to async */
2866 if (req->close.put_file->f_op->flush && !io_wq_current_is_worker())
2870 * No ->flush(), safely close from here and just punt the
2871 * fput() to async context.
2873 ret = filp_close(req->close.put_file, current->files);
2876 req_set_fail_links(req);
2877 io_cqring_add_event(req, ret);
2879 if (io_wq_current_is_worker()) {
2880 struct io_wq_work *old_work, *work;
2882 old_work = work = &req->work;
2883 io_close_finish(&work);
2884 if (work && work != old_work)
2885 *nxt = container_of(work, struct io_kiocb, work);
2890 req->work.func = io_close_finish;
2894 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2896 struct io_ring_ctx *ctx = req->ctx;
2901 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2903 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2906 req->sync.off = READ_ONCE(sqe->off);
2907 req->sync.len = READ_ONCE(sqe->len);
2908 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
2912 static void io_sync_file_range_finish(struct io_wq_work **workptr)
2914 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2915 struct io_kiocb *nxt = NULL;
2918 if (io_req_cancelled(req))
2921 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
2924 req_set_fail_links(req);
2925 io_cqring_add_event(req, ret);
2926 io_put_req_find_next(req, &nxt);
2928 io_wq_assign_next(workptr, nxt);
2931 static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt,
2932 bool force_nonblock)
2934 struct io_wq_work *work, *old_work;
2936 /* sync_file_range always requires a blocking context */
2937 if (force_nonblock) {
2939 req->work.func = io_sync_file_range_finish;
2943 work = old_work = &req->work;
2944 io_sync_file_range_finish(&work);
2945 if (work && work != old_work)
2946 *nxt = container_of(work, struct io_kiocb, work);
2950 #if defined(CONFIG_NET)
2951 static void io_sendrecv_async(struct io_wq_work **workptr)
2953 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2954 struct iovec *iov = NULL;
2956 if (req->io->rw.iov != req->io->rw.fast_iov)
2957 iov = req->io->msg.iov;
2958 io_wq_submit_work(workptr);
2963 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2965 #if defined(CONFIG_NET)
2966 struct io_sr_msg *sr = &req->sr_msg;
2967 struct io_async_ctx *io = req->io;
2969 sr->msg_flags = READ_ONCE(sqe->msg_flags);
2970 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
2971 sr->len = READ_ONCE(sqe->len);
2973 if (!io || req->opcode == IORING_OP_SEND)
2976 io->msg.iov = io->msg.fast_iov;
2977 return sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
2984 static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt,
2985 bool force_nonblock)
2987 #if defined(CONFIG_NET)
2988 struct io_async_msghdr *kmsg = NULL;
2989 struct socket *sock;
2992 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2995 sock = sock_from_file(req->file, &ret);
2997 struct io_async_ctx io;
2998 struct sockaddr_storage addr;
3002 kmsg = &req->io->msg;
3003 kmsg->msg.msg_name = &addr;
3004 /* if iov is set, it's allocated already */
3006 kmsg->iov = kmsg->fast_iov;
3007 kmsg->msg.msg_iter.iov = kmsg->iov;
3009 struct io_sr_msg *sr = &req->sr_msg;
3012 kmsg->msg.msg_name = &addr;
3014 io.msg.iov = io.msg.fast_iov;
3015 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3016 sr->msg_flags, &io.msg.iov);
3021 flags = req->sr_msg.msg_flags;
3022 if (flags & MSG_DONTWAIT)
3023 req->flags |= REQ_F_NOWAIT;
3024 else if (force_nonblock)
3025 flags |= MSG_DONTWAIT;
3027 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3028 if (force_nonblock && ret == -EAGAIN) {
3031 if (io_alloc_async_ctx(req))
3033 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3034 req->work.func = io_sendrecv_async;
3037 if (ret == -ERESTARTSYS)
3041 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
3043 io_cqring_add_event(req, ret);
3045 req_set_fail_links(req);
3046 io_put_req_find_next(req, nxt);
3053 static int io_send(struct io_kiocb *req, struct io_kiocb **nxt,
3054 bool force_nonblock)
3056 #if defined(CONFIG_NET)
3057 struct socket *sock;
3060 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3063 sock = sock_from_file(req->file, &ret);
3065 struct io_sr_msg *sr = &req->sr_msg;
3070 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3075 msg.msg_name = NULL;
3076 msg.msg_control = NULL;
3077 msg.msg_controllen = 0;
3078 msg.msg_namelen = 0;
3080 flags = req->sr_msg.msg_flags;
3081 if (flags & MSG_DONTWAIT)
3082 req->flags |= REQ_F_NOWAIT;
3083 else if (force_nonblock)
3084 flags |= MSG_DONTWAIT;
3086 ret = __sys_sendmsg_sock(sock, &msg, flags);
3087 if (force_nonblock && ret == -EAGAIN)
3089 if (ret == -ERESTARTSYS)
3093 io_cqring_add_event(req, ret);
3095 req_set_fail_links(req);
3096 io_put_req_find_next(req, nxt);
3103 static int io_recvmsg_prep(struct io_kiocb *req,
3104 const struct io_uring_sqe *sqe)
3106 #if defined(CONFIG_NET)
3107 struct io_sr_msg *sr = &req->sr_msg;
3108 struct io_async_ctx *io = req->io;
3110 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3111 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3113 if (!io || req->opcode == IORING_OP_RECV)
3116 io->msg.iov = io->msg.fast_iov;
3117 return recvmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3118 &io->msg.uaddr, &io->msg.iov);
3124 static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt,
3125 bool force_nonblock)
3127 #if defined(CONFIG_NET)
3128 struct io_async_msghdr *kmsg = NULL;
3129 struct socket *sock;
3132 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3135 sock = sock_from_file(req->file, &ret);
3137 struct io_async_ctx io;
3138 struct sockaddr_storage addr;
3142 kmsg = &req->io->msg;
3143 kmsg->msg.msg_name = &addr;
3144 /* if iov is set, it's allocated already */
3146 kmsg->iov = kmsg->fast_iov;
3147 kmsg->msg.msg_iter.iov = kmsg->iov;
3149 struct io_sr_msg *sr = &req->sr_msg;
3152 kmsg->msg.msg_name = &addr;
3154 io.msg.iov = io.msg.fast_iov;
3155 ret = recvmsg_copy_msghdr(&io.msg.msg, sr->msg,
3156 sr->msg_flags, &io.msg.uaddr,
3162 flags = req->sr_msg.msg_flags;
3163 if (flags & MSG_DONTWAIT)
3164 req->flags |= REQ_F_NOWAIT;
3165 else if (force_nonblock)
3166 flags |= MSG_DONTWAIT;
3168 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3169 kmsg->uaddr, flags);
3170 if (force_nonblock && ret == -EAGAIN) {
3173 if (io_alloc_async_ctx(req))
3175 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3176 req->work.func = io_sendrecv_async;
3179 if (ret == -ERESTARTSYS)
3183 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
3185 io_cqring_add_event(req, ret);
3187 req_set_fail_links(req);
3188 io_put_req_find_next(req, nxt);
3195 static int io_recv(struct io_kiocb *req, struct io_kiocb **nxt,
3196 bool force_nonblock)
3198 #if defined(CONFIG_NET)
3199 struct socket *sock;
3202 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3205 sock = sock_from_file(req->file, &ret);
3207 struct io_sr_msg *sr = &req->sr_msg;
3212 ret = import_single_range(READ, sr->buf, sr->len, &iov,
3217 msg.msg_name = NULL;
3218 msg.msg_control = NULL;
3219 msg.msg_controllen = 0;
3220 msg.msg_namelen = 0;
3221 msg.msg_iocb = NULL;
3224 flags = req->sr_msg.msg_flags;
3225 if (flags & MSG_DONTWAIT)
3226 req->flags |= REQ_F_NOWAIT;
3227 else if (force_nonblock)
3228 flags |= MSG_DONTWAIT;
3230 ret = __sys_recvmsg_sock(sock, &msg, NULL, NULL, flags);
3231 if (force_nonblock && ret == -EAGAIN)
3233 if (ret == -ERESTARTSYS)
3237 io_cqring_add_event(req, ret);
3239 req_set_fail_links(req);
3240 io_put_req_find_next(req, nxt);
3248 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3250 #if defined(CONFIG_NET)
3251 struct io_accept *accept = &req->accept;
3253 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3255 if (sqe->ioprio || sqe->len || sqe->buf_index)
3258 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3259 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3260 accept->flags = READ_ONCE(sqe->accept_flags);
3267 #if defined(CONFIG_NET)
3268 static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3269 bool force_nonblock)
3271 struct io_accept *accept = &req->accept;
3272 unsigned file_flags;
3275 file_flags = force_nonblock ? O_NONBLOCK : 0;
3276 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3277 accept->addr_len, accept->flags);
3278 if (ret == -EAGAIN && force_nonblock)
3280 if (ret == -ERESTARTSYS)
3283 req_set_fail_links(req);
3284 io_cqring_add_event(req, ret);
3285 io_put_req_find_next(req, nxt);
3289 static void io_accept_finish(struct io_wq_work **workptr)
3291 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3292 struct io_kiocb *nxt = NULL;
3294 if (io_req_cancelled(req))
3296 __io_accept(req, &nxt, false);
3298 io_wq_assign_next(workptr, nxt);
3302 static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3303 bool force_nonblock)
3305 #if defined(CONFIG_NET)
3308 ret = __io_accept(req, nxt, force_nonblock);
3309 if (ret == -EAGAIN && force_nonblock) {
3310 req->work.func = io_accept_finish;
3320 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3322 #if defined(CONFIG_NET)
3323 struct io_connect *conn = &req->connect;
3324 struct io_async_ctx *io = req->io;
3326 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3328 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3331 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3332 conn->addr_len = READ_ONCE(sqe->addr2);
3337 return move_addr_to_kernel(conn->addr, conn->addr_len,
3338 &io->connect.address);
3344 static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt,
3345 bool force_nonblock)
3347 #if defined(CONFIG_NET)
3348 struct io_async_ctx __io, *io;
3349 unsigned file_flags;
3355 ret = move_addr_to_kernel(req->connect.addr,
3356 req->connect.addr_len,
3357 &__io.connect.address);
3363 file_flags = force_nonblock ? O_NONBLOCK : 0;
3365 ret = __sys_connect_file(req->file, &io->connect.address,
3366 req->connect.addr_len, file_flags);
3367 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
3370 if (io_alloc_async_ctx(req)) {
3374 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
3377 if (ret == -ERESTARTSYS)
3381 req_set_fail_links(req);
3382 io_cqring_add_event(req, ret);
3383 io_put_req_find_next(req, nxt);
3390 static void io_poll_remove_one(struct io_kiocb *req)
3392 struct io_poll_iocb *poll = &req->poll;
3394 spin_lock(&poll->head->lock);
3395 WRITE_ONCE(poll->canceled, true);
3396 if (!list_empty(&poll->wait.entry)) {
3397 list_del_init(&poll->wait.entry);
3398 io_queue_async_work(req);
3400 spin_unlock(&poll->head->lock);
3401 hash_del(&req->hash_node);
3404 static void io_poll_remove_all(struct io_ring_ctx *ctx)
3406 struct hlist_node *tmp;
3407 struct io_kiocb *req;
3410 spin_lock_irq(&ctx->completion_lock);
3411 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
3412 struct hlist_head *list;
3414 list = &ctx->cancel_hash[i];
3415 hlist_for_each_entry_safe(req, tmp, list, hash_node)
3416 io_poll_remove_one(req);
3418 spin_unlock_irq(&ctx->completion_lock);
3421 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
3423 struct hlist_head *list;
3424 struct io_kiocb *req;
3426 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
3427 hlist_for_each_entry(req, list, hash_node) {
3428 if (sqe_addr == req->user_data) {
3429 io_poll_remove_one(req);
3437 static int io_poll_remove_prep(struct io_kiocb *req,
3438 const struct io_uring_sqe *sqe)
3440 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3442 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3446 req->poll.addr = READ_ONCE(sqe->addr);
3451 * Find a running poll command that matches one specified in sqe->addr,
3452 * and remove it if found.
3454 static int io_poll_remove(struct io_kiocb *req)
3456 struct io_ring_ctx *ctx = req->ctx;
3460 addr = req->poll.addr;
3461 spin_lock_irq(&ctx->completion_lock);
3462 ret = io_poll_cancel(ctx, addr);
3463 spin_unlock_irq(&ctx->completion_lock);
3465 io_cqring_add_event(req, ret);
3467 req_set_fail_links(req);
3472 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
3474 struct io_ring_ctx *ctx = req->ctx;
3476 req->poll.done = true;
3478 io_cqring_fill_event(req, error);
3480 io_cqring_fill_event(req, mangle_poll(mask));
3481 io_commit_cqring(ctx);
3484 static void io_poll_complete_work(struct io_wq_work **workptr)
3486 struct io_wq_work *work = *workptr;
3487 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3488 struct io_poll_iocb *poll = &req->poll;
3489 struct poll_table_struct pt = { ._key = poll->events };
3490 struct io_ring_ctx *ctx = req->ctx;
3491 struct io_kiocb *nxt = NULL;
3495 if (work->flags & IO_WQ_WORK_CANCEL) {
3496 WRITE_ONCE(poll->canceled, true);
3498 } else if (READ_ONCE(poll->canceled)) {
3502 if (ret != -ECANCELED)
3503 mask = vfs_poll(poll->file, &pt) & poll->events;
3506 * Note that ->ki_cancel callers also delete iocb from active_reqs after
3507 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
3508 * synchronize with them. In the cancellation case the list_del_init
3509 * itself is not actually needed, but harmless so we keep it in to
3510 * avoid further branches in the fast path.
3512 spin_lock_irq(&ctx->completion_lock);
3513 if (!mask && ret != -ECANCELED) {
3514 add_wait_queue(poll->head, &poll->wait);
3515 spin_unlock_irq(&ctx->completion_lock);
3518 hash_del(&req->hash_node);
3519 io_poll_complete(req, mask, ret);
3520 spin_unlock_irq(&ctx->completion_lock);
3522 io_cqring_ev_posted(ctx);
3525 req_set_fail_links(req);
3526 io_put_req_find_next(req, &nxt);
3528 io_wq_assign_next(workptr, nxt);
3531 static void __io_poll_flush(struct io_ring_ctx *ctx, struct llist_node *nodes)
3533 struct io_kiocb *req, *tmp;
3534 struct req_batch rb;
3536 rb.to_free = rb.need_iter = 0;
3537 spin_lock_irq(&ctx->completion_lock);
3538 llist_for_each_entry_safe(req, tmp, nodes, llist_node) {
3539 hash_del(&req->hash_node);
3540 io_poll_complete(req, req->result, 0);
3542 if (refcount_dec_and_test(&req->refs) &&
3543 !io_req_multi_free(&rb, req)) {
3544 req->flags |= REQ_F_COMP_LOCKED;
3548 spin_unlock_irq(&ctx->completion_lock);
3550 io_cqring_ev_posted(ctx);
3551 io_free_req_many(ctx, &rb);
3554 static void io_poll_flush(struct io_wq_work **workptr)
3556 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3557 struct llist_node *nodes;
3559 nodes = llist_del_all(&req->ctx->poll_llist);
3561 __io_poll_flush(req->ctx, nodes);
3564 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
3567 struct io_poll_iocb *poll = wait->private;
3568 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
3569 struct io_ring_ctx *ctx = req->ctx;
3570 __poll_t mask = key_to_poll(key);
3572 /* for instances that support it check for an event match first: */
3573 if (mask && !(mask & poll->events))
3576 list_del_init(&poll->wait.entry);
3579 * Run completion inline if we can. We're using trylock here because
3580 * we are violating the completion_lock -> poll wq lock ordering.
3581 * If we have a link timeout we're going to need the completion_lock
3582 * for finalizing the request, mark us as having grabbed that already.
3585 unsigned long flags;
3587 if (llist_empty(&ctx->poll_llist) &&
3588 spin_trylock_irqsave(&ctx->completion_lock, flags)) {
3589 hash_del(&req->hash_node);
3590 io_poll_complete(req, mask, 0);
3591 req->flags |= REQ_F_COMP_LOCKED;
3593 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3595 io_cqring_ev_posted(ctx);
3599 req->llist_node.next = NULL;
3600 /* if the list wasn't empty, we're done */
3601 if (!llist_add(&req->llist_node, &ctx->poll_llist))
3604 req->work.func = io_poll_flush;
3608 io_queue_async_work(req);
3613 struct io_poll_table {
3614 struct poll_table_struct pt;
3615 struct io_kiocb *req;
3619 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
3620 struct poll_table_struct *p)
3622 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
3624 if (unlikely(pt->req->poll.head)) {
3625 pt->error = -EINVAL;
3630 pt->req->poll.head = head;
3631 add_wait_queue(head, &pt->req->poll.wait);
3634 static void io_poll_req_insert(struct io_kiocb *req)
3636 struct io_ring_ctx *ctx = req->ctx;
3637 struct hlist_head *list;
3639 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
3640 hlist_add_head(&req->hash_node, list);
3643 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3645 struct io_poll_iocb *poll = &req->poll;
3648 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3650 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3655 events = READ_ONCE(sqe->poll_events);
3656 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
3660 static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt)
3662 struct io_poll_iocb *poll = &req->poll;
3663 struct io_ring_ctx *ctx = req->ctx;
3664 struct io_poll_table ipt;
3665 bool cancel = false;
3668 INIT_IO_WORK(&req->work, io_poll_complete_work);
3669 INIT_HLIST_NODE(&req->hash_node);
3673 poll->canceled = false;
3675 ipt.pt._qproc = io_poll_queue_proc;
3676 ipt.pt._key = poll->events;
3678 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
3680 /* initialized the list so that we can do list_empty checks */
3681 INIT_LIST_HEAD(&poll->wait.entry);
3682 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
3683 poll->wait.private = poll;
3685 INIT_LIST_HEAD(&req->list);
3687 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
3689 spin_lock_irq(&ctx->completion_lock);
3690 if (likely(poll->head)) {
3691 spin_lock(&poll->head->lock);
3692 if (unlikely(list_empty(&poll->wait.entry))) {
3698 if (mask || ipt.error)
3699 list_del_init(&poll->wait.entry);
3701 WRITE_ONCE(poll->canceled, true);
3702 else if (!poll->done) /* actually waiting for an event */
3703 io_poll_req_insert(req);
3704 spin_unlock(&poll->head->lock);
3706 if (mask) { /* no async, we'd stolen it */
3708 io_poll_complete(req, mask, 0);
3710 spin_unlock_irq(&ctx->completion_lock);
3713 io_cqring_ev_posted(ctx);
3714 io_put_req_find_next(req, nxt);
3719 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
3721 struct io_timeout_data *data = container_of(timer,
3722 struct io_timeout_data, timer);
3723 struct io_kiocb *req = data->req;
3724 struct io_ring_ctx *ctx = req->ctx;
3725 unsigned long flags;
3727 atomic_inc(&ctx->cq_timeouts);
3729 spin_lock_irqsave(&ctx->completion_lock, flags);
3731 * We could be racing with timeout deletion. If the list is empty,
3732 * then timeout lookup already found it and will be handling it.
3734 if (!list_empty(&req->list)) {
3735 struct io_kiocb *prev;
3738 * Adjust the reqs sequence before the current one because it
3739 * will consume a slot in the cq_ring and the cq_tail
3740 * pointer will be increased, otherwise other timeout reqs may
3741 * return in advance without waiting for enough wait_nr.
3744 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
3746 list_del_init(&req->list);
3749 io_cqring_fill_event(req, -ETIME);
3750 io_commit_cqring(ctx);
3751 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3753 io_cqring_ev_posted(ctx);
3754 req_set_fail_links(req);
3756 return HRTIMER_NORESTART;
3759 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
3761 struct io_kiocb *req;
3764 list_for_each_entry(req, &ctx->timeout_list, list) {
3765 if (user_data == req->user_data) {
3766 list_del_init(&req->list);
3775 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
3779 req_set_fail_links(req);
3780 io_cqring_fill_event(req, -ECANCELED);
3785 static int io_timeout_remove_prep(struct io_kiocb *req,
3786 const struct io_uring_sqe *sqe)
3788 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3790 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
3793 req->timeout.addr = READ_ONCE(sqe->addr);
3794 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
3795 if (req->timeout.flags)
3802 * Remove or update an existing timeout command
3804 static int io_timeout_remove(struct io_kiocb *req)
3806 struct io_ring_ctx *ctx = req->ctx;
3809 spin_lock_irq(&ctx->completion_lock);
3810 ret = io_timeout_cancel(ctx, req->timeout.addr);
3812 io_cqring_fill_event(req, ret);
3813 io_commit_cqring(ctx);
3814 spin_unlock_irq(&ctx->completion_lock);
3815 io_cqring_ev_posted(ctx);
3817 req_set_fail_links(req);
3822 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3823 bool is_timeout_link)
3825 struct io_timeout_data *data;
3828 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3830 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
3832 if (sqe->off && is_timeout_link)
3834 flags = READ_ONCE(sqe->timeout_flags);
3835 if (flags & ~IORING_TIMEOUT_ABS)
3838 req->timeout.count = READ_ONCE(sqe->off);
3840 if (!req->io && io_alloc_async_ctx(req))
3843 data = &req->io->timeout;
3845 req->flags |= REQ_F_TIMEOUT;
3847 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
3850 if (flags & IORING_TIMEOUT_ABS)
3851 data->mode = HRTIMER_MODE_ABS;
3853 data->mode = HRTIMER_MODE_REL;
3855 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
3859 static int io_timeout(struct io_kiocb *req)
3862 struct io_ring_ctx *ctx = req->ctx;
3863 struct io_timeout_data *data;
3864 struct list_head *entry;
3867 data = &req->io->timeout;
3870 * sqe->off holds how many events that need to occur for this
3871 * timeout event to be satisfied. If it isn't set, then this is
3872 * a pure timeout request, sequence isn't used.
3874 count = req->timeout.count;
3876 req->flags |= REQ_F_TIMEOUT_NOSEQ;
3877 spin_lock_irq(&ctx->completion_lock);
3878 entry = ctx->timeout_list.prev;
3882 req->sequence = ctx->cached_sq_head + count - 1;
3883 data->seq_offset = count;
3886 * Insertion sort, ensuring the first entry in the list is always
3887 * the one we need first.
3889 spin_lock_irq(&ctx->completion_lock);
3890 list_for_each_prev(entry, &ctx->timeout_list) {
3891 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
3892 unsigned nxt_sq_head;
3893 long long tmp, tmp_nxt;
3894 u32 nxt_offset = nxt->io->timeout.seq_offset;
3896 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
3900 * Since cached_sq_head + count - 1 can overflow, use type long
3903 tmp = (long long)ctx->cached_sq_head + count - 1;
3904 nxt_sq_head = nxt->sequence - nxt_offset + 1;
3905 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
3908 * cached_sq_head may overflow, and it will never overflow twice
3909 * once there is some timeout req still be valid.
3911 if (ctx->cached_sq_head < nxt_sq_head)
3918 * Sequence of reqs after the insert one and itself should
3919 * be adjusted because each timeout req consumes a slot.
3924 req->sequence -= span;
3926 list_add(&req->list, entry);
3927 data->timer.function = io_timeout_fn;
3928 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
3929 spin_unlock_irq(&ctx->completion_lock);
3933 static bool io_cancel_cb(struct io_wq_work *work, void *data)
3935 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3937 return req->user_data == (unsigned long) data;
3940 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
3942 enum io_wq_cancel cancel_ret;
3945 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
3946 switch (cancel_ret) {
3947 case IO_WQ_CANCEL_OK:
3950 case IO_WQ_CANCEL_RUNNING:
3953 case IO_WQ_CANCEL_NOTFOUND:
3961 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
3962 struct io_kiocb *req, __u64 sqe_addr,
3963 struct io_kiocb **nxt, int success_ret)
3965 unsigned long flags;
3968 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
3969 if (ret != -ENOENT) {
3970 spin_lock_irqsave(&ctx->completion_lock, flags);
3974 spin_lock_irqsave(&ctx->completion_lock, flags);
3975 ret = io_timeout_cancel(ctx, sqe_addr);
3978 ret = io_poll_cancel(ctx, sqe_addr);
3982 io_cqring_fill_event(req, ret);
3983 io_commit_cqring(ctx);
3984 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3985 io_cqring_ev_posted(ctx);
3988 req_set_fail_links(req);
3989 io_put_req_find_next(req, nxt);
3992 static int io_async_cancel_prep(struct io_kiocb *req,
3993 const struct io_uring_sqe *sqe)
3995 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3997 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4001 req->cancel.addr = READ_ONCE(sqe->addr);
4005 static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt)
4007 struct io_ring_ctx *ctx = req->ctx;
4009 io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0);
4013 static int io_files_update_prep(struct io_kiocb *req,
4014 const struct io_uring_sqe *sqe)
4016 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4019 req->files_update.offset = READ_ONCE(sqe->off);
4020 req->files_update.nr_args = READ_ONCE(sqe->len);
4021 if (!req->files_update.nr_args)
4023 req->files_update.arg = READ_ONCE(sqe->addr);
4027 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4029 struct io_ring_ctx *ctx = req->ctx;
4030 struct io_uring_files_update up;
4036 up.offset = req->files_update.offset;
4037 up.fds = req->files_update.arg;
4039 mutex_lock(&ctx->uring_lock);
4040 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4041 mutex_unlock(&ctx->uring_lock);
4044 req_set_fail_links(req);
4045 io_cqring_add_event(req, ret);
4050 static int io_req_defer_prep(struct io_kiocb *req,
4051 const struct io_uring_sqe *sqe)
4055 if (io_op_defs[req->opcode].file_table) {
4056 ret = io_grab_files(req);
4061 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4063 switch (req->opcode) {
4066 case IORING_OP_READV:
4067 case IORING_OP_READ_FIXED:
4068 case IORING_OP_READ:
4069 ret = io_read_prep(req, sqe, true);
4071 case IORING_OP_WRITEV:
4072 case IORING_OP_WRITE_FIXED:
4073 case IORING_OP_WRITE:
4074 ret = io_write_prep(req, sqe, true);
4076 case IORING_OP_POLL_ADD:
4077 ret = io_poll_add_prep(req, sqe);
4079 case IORING_OP_POLL_REMOVE:
4080 ret = io_poll_remove_prep(req, sqe);
4082 case IORING_OP_FSYNC:
4083 ret = io_prep_fsync(req, sqe);
4085 case IORING_OP_SYNC_FILE_RANGE:
4086 ret = io_prep_sfr(req, sqe);
4088 case IORING_OP_SENDMSG:
4089 case IORING_OP_SEND:
4090 ret = io_sendmsg_prep(req, sqe);
4092 case IORING_OP_RECVMSG:
4093 case IORING_OP_RECV:
4094 ret = io_recvmsg_prep(req, sqe);
4096 case IORING_OP_CONNECT:
4097 ret = io_connect_prep(req, sqe);
4099 case IORING_OP_TIMEOUT:
4100 ret = io_timeout_prep(req, sqe, false);
4102 case IORING_OP_TIMEOUT_REMOVE:
4103 ret = io_timeout_remove_prep(req, sqe);
4105 case IORING_OP_ASYNC_CANCEL:
4106 ret = io_async_cancel_prep(req, sqe);
4108 case IORING_OP_LINK_TIMEOUT:
4109 ret = io_timeout_prep(req, sqe, true);
4111 case IORING_OP_ACCEPT:
4112 ret = io_accept_prep(req, sqe);
4114 case IORING_OP_FALLOCATE:
4115 ret = io_fallocate_prep(req, sqe);
4117 case IORING_OP_OPENAT:
4118 ret = io_openat_prep(req, sqe);
4120 case IORING_OP_CLOSE:
4121 ret = io_close_prep(req, sqe);
4123 case IORING_OP_FILES_UPDATE:
4124 ret = io_files_update_prep(req, sqe);
4126 case IORING_OP_STATX:
4127 ret = io_statx_prep(req, sqe);
4129 case IORING_OP_FADVISE:
4130 ret = io_fadvise_prep(req, sqe);
4132 case IORING_OP_MADVISE:
4133 ret = io_madvise_prep(req, sqe);
4135 case IORING_OP_OPENAT2:
4136 ret = io_openat2_prep(req, sqe);
4138 case IORING_OP_EPOLL_CTL:
4139 ret = io_epoll_ctl_prep(req, sqe);
4142 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
4151 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4153 struct io_ring_ctx *ctx = req->ctx;
4156 /* Still need defer if there is pending req in defer list. */
4157 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
4160 if (!req->io && io_alloc_async_ctx(req))
4163 ret = io_req_defer_prep(req, sqe);
4167 spin_lock_irq(&ctx->completion_lock);
4168 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
4169 spin_unlock_irq(&ctx->completion_lock);
4173 trace_io_uring_defer(ctx, req, req->user_data);
4174 list_add_tail(&req->list, &ctx->defer_list);
4175 spin_unlock_irq(&ctx->completion_lock);
4176 return -EIOCBQUEUED;
4179 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4180 struct io_kiocb **nxt, bool force_nonblock)
4182 struct io_ring_ctx *ctx = req->ctx;
4185 switch (req->opcode) {
4189 case IORING_OP_READV:
4190 case IORING_OP_READ_FIXED:
4191 case IORING_OP_READ:
4193 ret = io_read_prep(req, sqe, force_nonblock);
4197 ret = io_read(req, nxt, force_nonblock);
4199 case IORING_OP_WRITEV:
4200 case IORING_OP_WRITE_FIXED:
4201 case IORING_OP_WRITE:
4203 ret = io_write_prep(req, sqe, force_nonblock);
4207 ret = io_write(req, nxt, force_nonblock);
4209 case IORING_OP_FSYNC:
4211 ret = io_prep_fsync(req, sqe);
4215 ret = io_fsync(req, nxt, force_nonblock);
4217 case IORING_OP_POLL_ADD:
4219 ret = io_poll_add_prep(req, sqe);
4223 ret = io_poll_add(req, nxt);
4225 case IORING_OP_POLL_REMOVE:
4227 ret = io_poll_remove_prep(req, sqe);
4231 ret = io_poll_remove(req);
4233 case IORING_OP_SYNC_FILE_RANGE:
4235 ret = io_prep_sfr(req, sqe);
4239 ret = io_sync_file_range(req, nxt, force_nonblock);
4241 case IORING_OP_SENDMSG:
4242 case IORING_OP_SEND:
4244 ret = io_sendmsg_prep(req, sqe);
4248 if (req->opcode == IORING_OP_SENDMSG)
4249 ret = io_sendmsg(req, nxt, force_nonblock);
4251 ret = io_send(req, nxt, force_nonblock);
4253 case IORING_OP_RECVMSG:
4254 case IORING_OP_RECV:
4256 ret = io_recvmsg_prep(req, sqe);
4260 if (req->opcode == IORING_OP_RECVMSG)
4261 ret = io_recvmsg(req, nxt, force_nonblock);
4263 ret = io_recv(req, nxt, force_nonblock);
4265 case IORING_OP_TIMEOUT:
4267 ret = io_timeout_prep(req, sqe, false);
4271 ret = io_timeout(req);
4273 case IORING_OP_TIMEOUT_REMOVE:
4275 ret = io_timeout_remove_prep(req, sqe);
4279 ret = io_timeout_remove(req);
4281 case IORING_OP_ACCEPT:
4283 ret = io_accept_prep(req, sqe);
4287 ret = io_accept(req, nxt, force_nonblock);
4289 case IORING_OP_CONNECT:
4291 ret = io_connect_prep(req, sqe);
4295 ret = io_connect(req, nxt, force_nonblock);
4297 case IORING_OP_ASYNC_CANCEL:
4299 ret = io_async_cancel_prep(req, sqe);
4303 ret = io_async_cancel(req, nxt);
4305 case IORING_OP_FALLOCATE:
4307 ret = io_fallocate_prep(req, sqe);
4311 ret = io_fallocate(req, nxt, force_nonblock);
4313 case IORING_OP_OPENAT:
4315 ret = io_openat_prep(req, sqe);
4319 ret = io_openat(req, nxt, force_nonblock);
4321 case IORING_OP_CLOSE:
4323 ret = io_close_prep(req, sqe);
4327 ret = io_close(req, nxt, force_nonblock);
4329 case IORING_OP_FILES_UPDATE:
4331 ret = io_files_update_prep(req, sqe);
4335 ret = io_files_update(req, force_nonblock);
4337 case IORING_OP_STATX:
4339 ret = io_statx_prep(req, sqe);
4343 ret = io_statx(req, nxt, force_nonblock);
4345 case IORING_OP_FADVISE:
4347 ret = io_fadvise_prep(req, sqe);
4351 ret = io_fadvise(req, nxt, force_nonblock);
4353 case IORING_OP_MADVISE:
4355 ret = io_madvise_prep(req, sqe);
4359 ret = io_madvise(req, nxt, force_nonblock);
4361 case IORING_OP_OPENAT2:
4363 ret = io_openat2_prep(req, sqe);
4367 ret = io_openat2(req, nxt, force_nonblock);
4369 case IORING_OP_EPOLL_CTL:
4371 ret = io_epoll_ctl_prep(req, sqe);
4375 ret = io_epoll_ctl(req, nxt, force_nonblock);
4385 if (ctx->flags & IORING_SETUP_IOPOLL) {
4386 const bool in_async = io_wq_current_is_worker();
4388 if (req->result == -EAGAIN)
4391 /* workqueue context doesn't hold uring_lock, grab it now */
4393 mutex_lock(&ctx->uring_lock);
4395 io_iopoll_req_issued(req);
4398 mutex_unlock(&ctx->uring_lock);
4404 static void io_wq_submit_work(struct io_wq_work **workptr)
4406 struct io_wq_work *work = *workptr;
4407 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4408 struct io_kiocb *nxt = NULL;
4411 /* if NO_CANCEL is set, we must still run the work */
4412 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
4413 IO_WQ_WORK_CANCEL) {
4418 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
4419 req->in_async = true;
4421 ret = io_issue_sqe(req, NULL, &nxt, false);
4423 * We can get EAGAIN for polled IO even though we're
4424 * forcing a sync submission from here, since we can't
4425 * wait for request slots on the block side.
4433 /* drop submission reference */
4437 req_set_fail_links(req);
4438 io_cqring_add_event(req, ret);
4442 /* if a dependent link is ready, pass it back */
4444 io_wq_assign_next(workptr, nxt);
4447 static int io_req_needs_file(struct io_kiocb *req, int fd)
4449 if (!io_op_defs[req->opcode].needs_file)
4451 if (fd == -1 && io_op_defs[req->opcode].fd_non_neg)
4456 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
4459 struct fixed_file_table *table;
4461 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
4462 return table->files[index & IORING_FILE_TABLE_MASK];;
4465 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
4466 const struct io_uring_sqe *sqe)
4468 struct io_ring_ctx *ctx = req->ctx;
4472 flags = READ_ONCE(sqe->flags);
4473 fd = READ_ONCE(sqe->fd);
4475 if (!io_req_needs_file(req, fd))
4478 if (flags & IOSQE_FIXED_FILE) {
4479 if (unlikely(!ctx->file_data ||
4480 (unsigned) fd >= ctx->nr_user_files))
4482 fd = array_index_nospec(fd, ctx->nr_user_files);
4483 req->file = io_file_from_index(ctx, fd);
4486 req->flags |= REQ_F_FIXED_FILE;
4487 percpu_ref_get(&ctx->file_data->refs);
4489 if (req->needs_fixed_file)
4491 trace_io_uring_file_get(ctx, fd);
4492 req->file = io_file_get(state, fd);
4493 if (unlikely(!req->file))
4500 static int io_grab_files(struct io_kiocb *req)
4503 struct io_ring_ctx *ctx = req->ctx;
4505 if (req->work.files)
4507 if (!ctx->ring_file)
4511 spin_lock_irq(&ctx->inflight_lock);
4513 * We use the f_ops->flush() handler to ensure that we can flush
4514 * out work accessing these files if the fd is closed. Check if
4515 * the fd has changed since we started down this path, and disallow
4516 * this operation if it has.
4518 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
4519 list_add(&req->inflight_entry, &ctx->inflight_list);
4520 req->flags |= REQ_F_INFLIGHT;
4521 req->work.files = current->files;
4524 spin_unlock_irq(&ctx->inflight_lock);
4530 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
4532 struct io_timeout_data *data = container_of(timer,
4533 struct io_timeout_data, timer);
4534 struct io_kiocb *req = data->req;
4535 struct io_ring_ctx *ctx = req->ctx;
4536 struct io_kiocb *prev = NULL;
4537 unsigned long flags;
4539 spin_lock_irqsave(&ctx->completion_lock, flags);
4542 * We don't expect the list to be empty, that will only happen if we
4543 * race with the completion of the linked work.
4545 if (!list_empty(&req->link_list)) {
4546 prev = list_entry(req->link_list.prev, struct io_kiocb,
4548 if (refcount_inc_not_zero(&prev->refs)) {
4549 list_del_init(&req->link_list);
4550 prev->flags &= ~REQ_F_LINK_TIMEOUT;
4555 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4558 req_set_fail_links(prev);
4559 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
4563 io_cqring_add_event(req, -ETIME);
4566 return HRTIMER_NORESTART;
4569 static void io_queue_linked_timeout(struct io_kiocb *req)
4571 struct io_ring_ctx *ctx = req->ctx;
4574 * If the list is now empty, then our linked request finished before
4575 * we got a chance to setup the timer
4577 spin_lock_irq(&ctx->completion_lock);
4578 if (!list_empty(&req->link_list)) {
4579 struct io_timeout_data *data = &req->io->timeout;
4581 data->timer.function = io_link_timeout_fn;
4582 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
4585 spin_unlock_irq(&ctx->completion_lock);
4587 /* drop submission reference */
4591 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
4593 struct io_kiocb *nxt;
4595 if (!(req->flags & REQ_F_LINK))
4598 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
4600 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
4603 req->flags |= REQ_F_LINK_TIMEOUT;
4607 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4609 struct io_kiocb *linked_timeout;
4610 struct io_kiocb *nxt = NULL;
4614 linked_timeout = io_prep_linked_timeout(req);
4616 ret = io_issue_sqe(req, sqe, &nxt, true);
4619 * We async punt it if the file wasn't marked NOWAIT, or if the file
4620 * doesn't support non-blocking read/write attempts
4622 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
4623 (req->flags & REQ_F_MUST_PUNT))) {
4625 if (io_op_defs[req->opcode].file_table) {
4626 ret = io_grab_files(req);
4632 * Queued up for async execution, worker will release
4633 * submit reference when the iocb is actually submitted.
4635 io_queue_async_work(req);
4640 /* drop submission reference */
4643 if (linked_timeout) {
4645 io_queue_linked_timeout(linked_timeout);
4647 io_put_req(linked_timeout);
4650 /* and drop final reference, if we failed */
4652 io_cqring_add_event(req, ret);
4653 req_set_fail_links(req);
4661 if (req->flags & REQ_F_FORCE_ASYNC)
4667 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4671 ret = io_req_defer(req, sqe);
4673 if (ret != -EIOCBQUEUED) {
4675 io_cqring_add_event(req, ret);
4676 req_set_fail_links(req);
4677 io_double_put_req(req);
4679 } else if (req->flags & REQ_F_FORCE_ASYNC) {
4680 ret = io_req_defer_prep(req, sqe);
4681 if (unlikely(ret < 0))
4684 * Never try inline submit of IOSQE_ASYNC is set, go straight
4685 * to async execution.
4687 req->work.flags |= IO_WQ_WORK_CONCURRENT;
4688 io_queue_async_work(req);
4690 __io_queue_sqe(req, sqe);
4694 static inline void io_queue_link_head(struct io_kiocb *req)
4696 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
4697 io_cqring_add_event(req, -ECANCELED);
4698 io_double_put_req(req);
4700 io_queue_sqe(req, NULL);
4703 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
4704 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
4706 static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4707 struct io_submit_state *state, struct io_kiocb **link)
4709 const struct cred *old_creds = NULL;
4710 struct io_ring_ctx *ctx = req->ctx;
4711 unsigned int sqe_flags;
4714 sqe_flags = READ_ONCE(sqe->flags);
4716 /* enforce forwards compatibility on users */
4717 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
4722 id = READ_ONCE(sqe->personality);
4724 const struct cred *personality_creds;
4726 personality_creds = idr_find(&ctx->personality_idr, id);
4727 if (unlikely(!personality_creds)) {
4731 old_creds = override_creds(personality_creds);
4734 /* same numerical values with corresponding REQ_F_*, safe to copy */
4735 req->flags |= sqe_flags & (IOSQE_IO_DRAIN|IOSQE_IO_HARDLINK|
4738 ret = io_req_set_file(state, req, sqe);
4739 if (unlikely(ret)) {
4741 io_cqring_add_event(req, ret);
4742 io_double_put_req(req);
4744 revert_creds(old_creds);
4749 * If we already have a head request, queue this one for async
4750 * submittal once the head completes. If we don't have a head but
4751 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
4752 * submitted sync once the chain is complete. If none of those
4753 * conditions are true (normal request), then just queue it.
4756 struct io_kiocb *head = *link;
4759 * Taking sequential execution of a link, draining both sides
4760 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
4761 * requests in the link. So, it drains the head and the
4762 * next after the link request. The last one is done via
4763 * drain_next flag to persist the effect across calls.
4765 if (sqe_flags & IOSQE_IO_DRAIN) {
4766 head->flags |= REQ_F_IO_DRAIN;
4767 ctx->drain_next = 1;
4769 if (io_alloc_async_ctx(req)) {
4774 ret = io_req_defer_prep(req, sqe);
4776 /* fail even hard links since we don't submit */
4777 head->flags |= REQ_F_FAIL_LINK;
4780 trace_io_uring_link(ctx, req, head);
4781 list_add_tail(&req->link_list, &head->link_list);
4783 /* last request of a link, enqueue the link */
4784 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) {
4785 io_queue_link_head(head);
4789 if (unlikely(ctx->drain_next)) {
4790 req->flags |= REQ_F_IO_DRAIN;
4791 req->ctx->drain_next = 0;
4793 if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
4794 req->flags |= REQ_F_LINK;
4795 INIT_LIST_HEAD(&req->link_list);
4796 ret = io_req_defer_prep(req, sqe);
4798 req->flags |= REQ_F_FAIL_LINK;
4801 io_queue_sqe(req, sqe);
4806 revert_creds(old_creds);
4811 * Batched submission is done, ensure local IO is flushed out.
4813 static void io_submit_state_end(struct io_submit_state *state)
4815 blk_finish_plug(&state->plug);
4817 if (state->free_reqs)
4818 kmem_cache_free_bulk(req_cachep, state->free_reqs,
4819 &state->reqs[state->cur_req]);
4823 * Start submission side cache.
4825 static void io_submit_state_start(struct io_submit_state *state,
4826 unsigned int max_ios)
4828 blk_start_plug(&state->plug);
4829 state->free_reqs = 0;
4831 state->ios_left = max_ios;
4834 static void io_commit_sqring(struct io_ring_ctx *ctx)
4836 struct io_rings *rings = ctx->rings;
4839 * Ensure any loads from the SQEs are done at this point,
4840 * since once we write the new head, the application could
4841 * write new data to them.
4843 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
4847 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
4848 * that is mapped by userspace. This means that care needs to be taken to
4849 * ensure that reads are stable, as we cannot rely on userspace always
4850 * being a good citizen. If members of the sqe are validated and then later
4851 * used, it's important that those reads are done through READ_ONCE() to
4852 * prevent a re-load down the line.
4854 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
4855 const struct io_uring_sqe **sqe_ptr)
4857 u32 *sq_array = ctx->sq_array;
4861 * The cached sq head (or cq tail) serves two purposes:
4863 * 1) allows us to batch the cost of updating the user visible
4865 * 2) allows the kernel side to track the head on its own, even
4866 * though the application is the one updating it.
4868 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
4869 if (likely(head < ctx->sq_entries)) {
4871 * All io need record the previous position, if LINK vs DARIN,
4872 * it can be used to mark the position of the first IO in the
4875 req->sequence = ctx->cached_sq_head;
4876 *sqe_ptr = &ctx->sq_sqes[head];
4877 req->opcode = READ_ONCE((*sqe_ptr)->opcode);
4878 req->user_data = READ_ONCE((*sqe_ptr)->user_data);
4879 ctx->cached_sq_head++;
4883 /* drop invalid entries */
4884 ctx->cached_sq_head++;
4885 ctx->cached_sq_dropped++;
4886 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
4890 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
4891 struct file *ring_file, int ring_fd,
4892 struct mm_struct **mm, bool async)
4894 struct io_submit_state state, *statep = NULL;
4895 struct io_kiocb *link = NULL;
4896 int i, submitted = 0;
4897 bool mm_fault = false;
4899 /* if we have a backlog and couldn't flush it all, return BUSY */
4900 if (test_bit(0, &ctx->sq_check_overflow)) {
4901 if (!list_empty(&ctx->cq_overflow_list) &&
4902 !io_cqring_overflow_flush(ctx, false))
4906 /* make sure SQ entry isn't read before tail */
4907 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
4909 if (!percpu_ref_tryget_many(&ctx->refs, nr))
4912 if (nr > IO_PLUG_THRESHOLD) {
4913 io_submit_state_start(&state, nr);
4917 ctx->ring_fd = ring_fd;
4918 ctx->ring_file = ring_file;
4920 for (i = 0; i < nr; i++) {
4921 const struct io_uring_sqe *sqe;
4922 struct io_kiocb *req;
4924 req = io_get_req(ctx, statep);
4925 if (unlikely(!req)) {
4927 submitted = -EAGAIN;
4930 if (!io_get_sqring(ctx, req, &sqe)) {
4931 __io_req_do_free(req);
4935 /* will complete beyond this point, count as submitted */
4938 if (unlikely(req->opcode >= IORING_OP_LAST)) {
4939 io_cqring_add_event(req, -EINVAL);
4940 io_double_put_req(req);
4944 if (io_op_defs[req->opcode].needs_mm && !*mm) {
4945 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
4947 use_mm(ctx->sqo_mm);
4952 req->has_user = *mm != NULL;
4953 req->in_async = async;
4954 req->needs_fixed_file = async;
4955 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
4957 if (!io_submit_sqe(req, sqe, statep, &link))
4961 if (unlikely(submitted != nr)) {
4962 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
4964 percpu_ref_put_many(&ctx->refs, nr - ref_used);
4967 io_queue_link_head(link);
4969 io_submit_state_end(&state);
4971 /* Commit SQ ring head once we've consumed and submitted all SQEs */
4972 io_commit_sqring(ctx);
4977 static int io_sq_thread(void *data)
4979 struct io_ring_ctx *ctx = data;
4980 struct mm_struct *cur_mm = NULL;
4981 const struct cred *old_cred;
4982 mm_segment_t old_fs;
4985 unsigned long timeout;
4988 complete(&ctx->completions[1]);
4992 old_cred = override_creds(ctx->creds);
4994 ret = timeout = inflight = 0;
4995 while (!kthread_should_park()) {
4996 unsigned int to_submit;
4999 unsigned nr_events = 0;
5001 if (ctx->flags & IORING_SETUP_IOPOLL) {
5003 * inflight is the count of the maximum possible
5004 * entries we submitted, but it can be smaller
5005 * if we dropped some of them. If we don't have
5006 * poll entries available, then we know that we
5007 * have nothing left to poll for. Reset the
5008 * inflight count to zero in that case.
5010 mutex_lock(&ctx->uring_lock);
5011 if (!list_empty(&ctx->poll_list))
5012 __io_iopoll_check(ctx, &nr_events, 0);
5015 mutex_unlock(&ctx->uring_lock);
5018 * Normal IO, just pretend everything completed.
5019 * We don't have to poll completions for that.
5021 nr_events = inflight;
5024 inflight -= nr_events;
5026 timeout = jiffies + ctx->sq_thread_idle;
5029 to_submit = io_sqring_entries(ctx);
5032 * If submit got -EBUSY, flag us as needing the application
5033 * to enter the kernel to reap and flush events.
5035 if (!to_submit || ret == -EBUSY) {
5037 * We're polling. If we're within the defined idle
5038 * period, then let us spin without work before going
5039 * to sleep. The exception is if we got EBUSY doing
5040 * more IO, we should wait for the application to
5041 * reap events and wake us up.
5044 (!time_after(jiffies, timeout) && ret != -EBUSY)) {
5050 * Drop cur_mm before scheduling, we can't hold it for
5051 * long periods (or over schedule()). Do this before
5052 * adding ourselves to the waitqueue, as the unuse/drop
5061 prepare_to_wait(&ctx->sqo_wait, &wait,
5062 TASK_INTERRUPTIBLE);
5064 /* Tell userspace we may need a wakeup call */
5065 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
5066 /* make sure to read SQ tail after writing flags */
5069 to_submit = io_sqring_entries(ctx);
5070 if (!to_submit || ret == -EBUSY) {
5071 if (kthread_should_park()) {
5072 finish_wait(&ctx->sqo_wait, &wait);
5075 if (signal_pending(current))
5076 flush_signals(current);
5078 finish_wait(&ctx->sqo_wait, &wait);
5080 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5083 finish_wait(&ctx->sqo_wait, &wait);
5085 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5088 mutex_lock(&ctx->uring_lock);
5089 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
5090 mutex_unlock(&ctx->uring_lock);
5100 revert_creds(old_cred);
5107 struct io_wait_queue {
5108 struct wait_queue_entry wq;
5109 struct io_ring_ctx *ctx;
5111 unsigned nr_timeouts;
5114 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
5116 struct io_ring_ctx *ctx = iowq->ctx;
5119 * Wake up if we have enough events, or if a timeout occurred since we
5120 * started waiting. For timeouts, we always want to return to userspace,
5121 * regardless of event count.
5123 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
5124 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
5127 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
5128 int wake_flags, void *key)
5130 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
5133 /* use noflush == true, as we can't safely rely on locking context */
5134 if (!io_should_wake(iowq, true))
5137 return autoremove_wake_function(curr, mode, wake_flags, key);
5141 * Wait until events become available, if we don't already have some. The
5142 * application must reap them itself, as they reside on the shared cq ring.
5144 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
5145 const sigset_t __user *sig, size_t sigsz)
5147 struct io_wait_queue iowq = {
5150 .func = io_wake_function,
5151 .entry = LIST_HEAD_INIT(iowq.wq.entry),
5154 .to_wait = min_events,
5156 struct io_rings *rings = ctx->rings;
5159 if (io_cqring_events(ctx, false) >= min_events)
5163 #ifdef CONFIG_COMPAT
5164 if (in_compat_syscall())
5165 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
5169 ret = set_user_sigmask(sig, sigsz);
5175 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
5176 trace_io_uring_cqring_wait(ctx, min_events);
5178 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
5179 TASK_INTERRUPTIBLE);
5180 if (io_should_wake(&iowq, false))
5183 if (signal_pending(current)) {
5188 finish_wait(&ctx->wait, &iowq.wq);
5190 restore_saved_sigmask_unless(ret == -EINTR);
5192 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
5195 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
5197 #if defined(CONFIG_UNIX)
5198 if (ctx->ring_sock) {
5199 struct sock *sock = ctx->ring_sock->sk;
5200 struct sk_buff *skb;
5202 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
5208 for (i = 0; i < ctx->nr_user_files; i++) {
5211 file = io_file_from_index(ctx, i);
5218 static void io_file_ref_kill(struct percpu_ref *ref)
5220 struct fixed_file_data *data;
5222 data = container_of(ref, struct fixed_file_data, refs);
5223 complete(&data->done);
5226 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
5228 struct fixed_file_data *data = ctx->file_data;
5229 unsigned nr_tables, i;
5234 /* protect against inflight atomic switch, which drops the ref */
5235 percpu_ref_get(&data->refs);
5236 /* wait for existing switches */
5237 flush_work(&data->ref_work);
5238 percpu_ref_kill_and_confirm(&data->refs, io_file_ref_kill);
5239 wait_for_completion(&data->done);
5240 percpu_ref_put(&data->refs);
5241 /* flush potential new switch */
5242 flush_work(&data->ref_work);
5243 percpu_ref_exit(&data->refs);
5245 __io_sqe_files_unregister(ctx);
5246 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
5247 for (i = 0; i < nr_tables; i++)
5248 kfree(data->table[i].files);
5251 ctx->file_data = NULL;
5252 ctx->nr_user_files = 0;
5256 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
5258 if (ctx->sqo_thread) {
5259 wait_for_completion(&ctx->completions[1]);
5261 * The park is a bit of a work-around, without it we get
5262 * warning spews on shutdown with SQPOLL set and affinity
5263 * set to a single CPU.
5265 kthread_park(ctx->sqo_thread);
5266 kthread_stop(ctx->sqo_thread);
5267 ctx->sqo_thread = NULL;
5271 static void io_finish_async(struct io_ring_ctx *ctx)
5273 io_sq_thread_stop(ctx);
5276 io_wq_destroy(ctx->io_wq);
5281 #if defined(CONFIG_UNIX)
5283 * Ensure the UNIX gc is aware of our file set, so we are certain that
5284 * the io_uring can be safely unregistered on process exit, even if we have
5285 * loops in the file referencing.
5287 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
5289 struct sock *sk = ctx->ring_sock->sk;
5290 struct scm_fp_list *fpl;
5291 struct sk_buff *skb;
5294 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
5295 unsigned long inflight = ctx->user->unix_inflight + nr;
5297 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
5301 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
5305 skb = alloc_skb(0, GFP_KERNEL);
5314 fpl->user = get_uid(ctx->user);
5315 for (i = 0; i < nr; i++) {
5316 struct file *file = io_file_from_index(ctx, i + offset);
5320 fpl->fp[nr_files] = get_file(file);
5321 unix_inflight(fpl->user, fpl->fp[nr_files]);
5326 fpl->max = SCM_MAX_FD;
5327 fpl->count = nr_files;
5328 UNIXCB(skb).fp = fpl;
5329 skb->destructor = unix_destruct_scm;
5330 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
5331 skb_queue_head(&sk->sk_receive_queue, skb);
5333 for (i = 0; i < nr_files; i++)
5344 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
5345 * causes regular reference counting to break down. We rely on the UNIX
5346 * garbage collection to take care of this problem for us.
5348 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5350 unsigned left, total;
5354 left = ctx->nr_user_files;
5356 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
5358 ret = __io_sqe_files_scm(ctx, this_files, total);
5362 total += this_files;
5368 while (total < ctx->nr_user_files) {
5369 struct file *file = io_file_from_index(ctx, total);
5379 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5385 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
5390 for (i = 0; i < nr_tables; i++) {
5391 struct fixed_file_table *table = &ctx->file_data->table[i];
5392 unsigned this_files;
5394 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
5395 table->files = kcalloc(this_files, sizeof(struct file *),
5399 nr_files -= this_files;
5405 for (i = 0; i < nr_tables; i++) {
5406 struct fixed_file_table *table = &ctx->file_data->table[i];
5407 kfree(table->files);
5412 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
5414 #if defined(CONFIG_UNIX)
5415 struct sock *sock = ctx->ring_sock->sk;
5416 struct sk_buff_head list, *head = &sock->sk_receive_queue;
5417 struct sk_buff *skb;
5420 __skb_queue_head_init(&list);
5423 * Find the skb that holds this file in its SCM_RIGHTS. When found,
5424 * remove this entry and rearrange the file array.
5426 skb = skb_dequeue(head);
5428 struct scm_fp_list *fp;
5430 fp = UNIXCB(skb).fp;
5431 for (i = 0; i < fp->count; i++) {
5434 if (fp->fp[i] != file)
5437 unix_notinflight(fp->user, fp->fp[i]);
5438 left = fp->count - 1 - i;
5440 memmove(&fp->fp[i], &fp->fp[i + 1],
5441 left * sizeof(struct file *));
5448 __skb_queue_tail(&list, skb);
5458 __skb_queue_tail(&list, skb);
5460 skb = skb_dequeue(head);
5463 if (skb_peek(&list)) {
5464 spin_lock_irq(&head->lock);
5465 while ((skb = __skb_dequeue(&list)) != NULL)
5466 __skb_queue_tail(head, skb);
5467 spin_unlock_irq(&head->lock);
5474 struct io_file_put {
5475 struct llist_node llist;
5477 struct completion *done;
5480 static void io_ring_file_ref_switch(struct work_struct *work)
5482 struct io_file_put *pfile, *tmp;
5483 struct fixed_file_data *data;
5484 struct llist_node *node;
5486 data = container_of(work, struct fixed_file_data, ref_work);
5488 while ((node = llist_del_all(&data->put_llist)) != NULL) {
5489 llist_for_each_entry_safe(pfile, tmp, node, llist) {
5490 io_ring_file_put(data->ctx, pfile->file);
5492 complete(pfile->done);
5498 percpu_ref_get(&data->refs);
5499 percpu_ref_switch_to_percpu(&data->refs);
5502 static void io_file_data_ref_zero(struct percpu_ref *ref)
5504 struct fixed_file_data *data;
5506 data = container_of(ref, struct fixed_file_data, refs);
5508 /* we can't safely switch from inside this context, punt to wq */
5509 queue_work(system_wq, &data->ref_work);
5512 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
5515 __s32 __user *fds = (__s32 __user *) arg;
5525 if (nr_args > IORING_MAX_FIXED_FILES)
5528 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
5529 if (!ctx->file_data)
5531 ctx->file_data->ctx = ctx;
5532 init_completion(&ctx->file_data->done);
5534 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
5535 ctx->file_data->table = kcalloc(nr_tables,
5536 sizeof(struct fixed_file_table),
5538 if (!ctx->file_data->table) {
5539 kfree(ctx->file_data);
5540 ctx->file_data = NULL;
5544 if (percpu_ref_init(&ctx->file_data->refs, io_file_data_ref_zero,
5545 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
5546 kfree(ctx->file_data->table);
5547 kfree(ctx->file_data);
5548 ctx->file_data = NULL;
5551 ctx->file_data->put_llist.first = NULL;
5552 INIT_WORK(&ctx->file_data->ref_work, io_ring_file_ref_switch);
5554 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
5555 percpu_ref_exit(&ctx->file_data->refs);
5556 kfree(ctx->file_data->table);
5557 kfree(ctx->file_data);
5558 ctx->file_data = NULL;
5562 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
5563 struct fixed_file_table *table;
5567 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
5569 /* allow sparse sets */
5575 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5576 index = i & IORING_FILE_TABLE_MASK;
5584 * Don't allow io_uring instances to be registered. If UNIX
5585 * isn't enabled, then this causes a reference cycle and this
5586 * instance can never get freed. If UNIX is enabled we'll
5587 * handle it just fine, but there's still no point in allowing
5588 * a ring fd as it doesn't support regular read/write anyway.
5590 if (file->f_op == &io_uring_fops) {
5595 table->files[index] = file;
5599 for (i = 0; i < ctx->nr_user_files; i++) {
5600 file = io_file_from_index(ctx, i);
5604 for (i = 0; i < nr_tables; i++)
5605 kfree(ctx->file_data->table[i].files);
5607 kfree(ctx->file_data->table);
5608 kfree(ctx->file_data);
5609 ctx->file_data = NULL;
5610 ctx->nr_user_files = 0;
5614 ret = io_sqe_files_scm(ctx);
5616 io_sqe_files_unregister(ctx);
5621 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
5624 #if defined(CONFIG_UNIX)
5625 struct sock *sock = ctx->ring_sock->sk;
5626 struct sk_buff_head *head = &sock->sk_receive_queue;
5627 struct sk_buff *skb;
5630 * See if we can merge this file into an existing skb SCM_RIGHTS
5631 * file set. If there's no room, fall back to allocating a new skb
5632 * and filling it in.
5634 spin_lock_irq(&head->lock);
5635 skb = skb_peek(head);
5637 struct scm_fp_list *fpl = UNIXCB(skb).fp;
5639 if (fpl->count < SCM_MAX_FD) {
5640 __skb_unlink(skb, head);
5641 spin_unlock_irq(&head->lock);
5642 fpl->fp[fpl->count] = get_file(file);
5643 unix_inflight(fpl->user, fpl->fp[fpl->count]);
5645 spin_lock_irq(&head->lock);
5646 __skb_queue_head(head, skb);
5651 spin_unlock_irq(&head->lock);
5658 return __io_sqe_files_scm(ctx, 1, index);
5664 static void io_atomic_switch(struct percpu_ref *ref)
5666 struct fixed_file_data *data;
5668 data = container_of(ref, struct fixed_file_data, refs);
5669 clear_bit(FFD_F_ATOMIC, &data->state);
5672 static bool io_queue_file_removal(struct fixed_file_data *data,
5675 struct io_file_put *pfile, pfile_stack;
5676 DECLARE_COMPLETION_ONSTACK(done);
5679 * If we fail allocating the struct we need for doing async reomval
5680 * of this file, just punt to sync and wait for it.
5682 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
5684 pfile = &pfile_stack;
5685 pfile->done = &done;
5689 llist_add(&pfile->llist, &data->put_llist);
5691 if (pfile == &pfile_stack) {
5692 if (!test_and_set_bit(FFD_F_ATOMIC, &data->state)) {
5693 percpu_ref_put(&data->refs);
5694 percpu_ref_switch_to_atomic(&data->refs,
5697 wait_for_completion(&done);
5698 flush_work(&data->ref_work);
5705 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
5706 struct io_uring_files_update *up,
5709 struct fixed_file_data *data = ctx->file_data;
5710 bool ref_switch = false;
5716 if (check_add_overflow(up->offset, nr_args, &done))
5718 if (done > ctx->nr_user_files)
5722 fds = u64_to_user_ptr(up->fds);
5724 struct fixed_file_table *table;
5728 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
5732 i = array_index_nospec(up->offset, ctx->nr_user_files);
5733 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5734 index = i & IORING_FILE_TABLE_MASK;
5735 if (table->files[index]) {
5736 file = io_file_from_index(ctx, index);
5737 table->files[index] = NULL;
5738 if (io_queue_file_removal(data, file))
5748 * Don't allow io_uring instances to be registered. If
5749 * UNIX isn't enabled, then this causes a reference
5750 * cycle and this instance can never get freed. If UNIX
5751 * is enabled we'll handle it just fine, but there's
5752 * still no point in allowing a ring fd as it doesn't
5753 * support regular read/write anyway.
5755 if (file->f_op == &io_uring_fops) {
5760 table->files[index] = file;
5761 err = io_sqe_file_register(ctx, file, i);
5770 if (ref_switch && !test_and_set_bit(FFD_F_ATOMIC, &data->state)) {
5771 percpu_ref_put(&data->refs);
5772 percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch);
5775 return done ? done : err;
5777 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
5780 struct io_uring_files_update up;
5782 if (!ctx->file_data)
5786 if (copy_from_user(&up, arg, sizeof(up)))
5791 return __io_sqe_files_update(ctx, &up, nr_args);
5794 static void io_put_work(struct io_wq_work *work)
5796 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5801 static void io_get_work(struct io_wq_work *work)
5803 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5805 refcount_inc(&req->refs);
5808 static int io_init_wq_offload(struct io_ring_ctx *ctx,
5809 struct io_uring_params *p)
5811 struct io_wq_data data;
5813 struct io_ring_ctx *ctx_attach;
5814 unsigned int concurrency;
5817 data.user = ctx->user;
5818 data.get_work = io_get_work;
5819 data.put_work = io_put_work;
5821 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
5822 /* Do QD, or 4 * CPUS, whatever is smallest */
5823 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
5825 ctx->io_wq = io_wq_create(concurrency, &data);
5826 if (IS_ERR(ctx->io_wq)) {
5827 ret = PTR_ERR(ctx->io_wq);
5833 f = fdget(p->wq_fd);
5837 if (f.file->f_op != &io_uring_fops) {
5842 ctx_attach = f.file->private_data;
5843 /* @io_wq is protected by holding the fd */
5844 if (!io_wq_get(ctx_attach->io_wq, &data)) {
5849 ctx->io_wq = ctx_attach->io_wq;
5855 static int io_sq_offload_start(struct io_ring_ctx *ctx,
5856 struct io_uring_params *p)
5860 init_waitqueue_head(&ctx->sqo_wait);
5861 mmgrab(current->mm);
5862 ctx->sqo_mm = current->mm;
5864 if (ctx->flags & IORING_SETUP_SQPOLL) {
5866 if (!capable(CAP_SYS_ADMIN))
5869 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
5870 if (!ctx->sq_thread_idle)
5871 ctx->sq_thread_idle = HZ;
5873 if (p->flags & IORING_SETUP_SQ_AFF) {
5874 int cpu = p->sq_thread_cpu;
5877 if (cpu >= nr_cpu_ids)
5879 if (!cpu_online(cpu))
5882 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
5886 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
5889 if (IS_ERR(ctx->sqo_thread)) {
5890 ret = PTR_ERR(ctx->sqo_thread);
5891 ctx->sqo_thread = NULL;
5894 wake_up_process(ctx->sqo_thread);
5895 } else if (p->flags & IORING_SETUP_SQ_AFF) {
5896 /* Can't have SQ_AFF without SQPOLL */
5901 ret = io_init_wq_offload(ctx, p);
5907 io_finish_async(ctx);
5908 mmdrop(ctx->sqo_mm);
5913 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
5915 atomic_long_sub(nr_pages, &user->locked_vm);
5918 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
5920 unsigned long page_limit, cur_pages, new_pages;
5922 /* Don't allow more pages than we can safely lock */
5923 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
5926 cur_pages = atomic_long_read(&user->locked_vm);
5927 new_pages = cur_pages + nr_pages;
5928 if (new_pages > page_limit)
5930 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
5931 new_pages) != cur_pages);
5936 static void io_mem_free(void *ptr)
5943 page = virt_to_head_page(ptr);
5944 if (put_page_testzero(page))
5945 free_compound_page(page);
5948 static void *io_mem_alloc(size_t size)
5950 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
5953 return (void *) __get_free_pages(gfp_flags, get_order(size));
5956 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
5959 struct io_rings *rings;
5960 size_t off, sq_array_size;
5962 off = struct_size(rings, cqes, cq_entries);
5963 if (off == SIZE_MAX)
5967 off = ALIGN(off, SMP_CACHE_BYTES);
5972 sq_array_size = array_size(sizeof(u32), sq_entries);
5973 if (sq_array_size == SIZE_MAX)
5976 if (check_add_overflow(off, sq_array_size, &off))
5985 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
5989 pages = (size_t)1 << get_order(
5990 rings_size(sq_entries, cq_entries, NULL));
5991 pages += (size_t)1 << get_order(
5992 array_size(sizeof(struct io_uring_sqe), sq_entries));
5997 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6001 if (!ctx->user_bufs)
6004 for (i = 0; i < ctx->nr_user_bufs; i++) {
6005 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6007 for (j = 0; j < imu->nr_bvecs; j++)
6008 unpin_user_page(imu->bvec[j].bv_page);
6010 if (ctx->account_mem)
6011 io_unaccount_mem(ctx->user, imu->nr_bvecs);
6016 kfree(ctx->user_bufs);
6017 ctx->user_bufs = NULL;
6018 ctx->nr_user_bufs = 0;
6022 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
6023 void __user *arg, unsigned index)
6025 struct iovec __user *src;
6027 #ifdef CONFIG_COMPAT
6029 struct compat_iovec __user *ciovs;
6030 struct compat_iovec ciov;
6032 ciovs = (struct compat_iovec __user *) arg;
6033 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
6036 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
6037 dst->iov_len = ciov.iov_len;
6041 src = (struct iovec __user *) arg;
6042 if (copy_from_user(dst, &src[index], sizeof(*dst)))
6047 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
6050 struct vm_area_struct **vmas = NULL;
6051 struct page **pages = NULL;
6052 int i, j, got_pages = 0;
6057 if (!nr_args || nr_args > UIO_MAXIOV)
6060 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
6062 if (!ctx->user_bufs)
6065 for (i = 0; i < nr_args; i++) {
6066 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6067 unsigned long off, start, end, ubuf;
6072 ret = io_copy_iov(ctx, &iov, arg, i);
6077 * Don't impose further limits on the size and buffer
6078 * constraints here, we'll -EINVAL later when IO is
6079 * submitted if they are wrong.
6082 if (!iov.iov_base || !iov.iov_len)
6085 /* arbitrary limit, but we need something */
6086 if (iov.iov_len > SZ_1G)
6089 ubuf = (unsigned long) iov.iov_base;
6090 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
6091 start = ubuf >> PAGE_SHIFT;
6092 nr_pages = end - start;
6094 if (ctx->account_mem) {
6095 ret = io_account_mem(ctx->user, nr_pages);
6101 if (!pages || nr_pages > got_pages) {
6104 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
6106 vmas = kvmalloc_array(nr_pages,
6107 sizeof(struct vm_area_struct *),
6109 if (!pages || !vmas) {
6111 if (ctx->account_mem)
6112 io_unaccount_mem(ctx->user, nr_pages);
6115 got_pages = nr_pages;
6118 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
6122 if (ctx->account_mem)
6123 io_unaccount_mem(ctx->user, nr_pages);
6128 down_read(¤t->mm->mmap_sem);
6129 pret = pin_user_pages(ubuf, nr_pages,
6130 FOLL_WRITE | FOLL_LONGTERM,
6132 if (pret == nr_pages) {
6133 /* don't support file backed memory */
6134 for (j = 0; j < nr_pages; j++) {
6135 struct vm_area_struct *vma = vmas[j];
6138 !is_file_hugepages(vma->vm_file)) {
6144 ret = pret < 0 ? pret : -EFAULT;
6146 up_read(¤t->mm->mmap_sem);
6149 * if we did partial map, or found file backed vmas,
6150 * release any pages we did get
6153 unpin_user_pages(pages, pret);
6154 if (ctx->account_mem)
6155 io_unaccount_mem(ctx->user, nr_pages);
6160 off = ubuf & ~PAGE_MASK;
6162 for (j = 0; j < nr_pages; j++) {
6165 vec_len = min_t(size_t, size, PAGE_SIZE - off);
6166 imu->bvec[j].bv_page = pages[j];
6167 imu->bvec[j].bv_len = vec_len;
6168 imu->bvec[j].bv_offset = off;
6172 /* store original address for later verification */
6174 imu->len = iov.iov_len;
6175 imu->nr_bvecs = nr_pages;
6177 ctx->nr_user_bufs++;
6185 io_sqe_buffer_unregister(ctx);
6189 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
6191 __s32 __user *fds = arg;
6197 if (copy_from_user(&fd, fds, sizeof(*fds)))
6200 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
6201 if (IS_ERR(ctx->cq_ev_fd)) {
6202 int ret = PTR_ERR(ctx->cq_ev_fd);
6203 ctx->cq_ev_fd = NULL;
6210 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
6212 if (ctx->cq_ev_fd) {
6213 eventfd_ctx_put(ctx->cq_ev_fd);
6214 ctx->cq_ev_fd = NULL;
6221 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
6223 io_finish_async(ctx);
6225 mmdrop(ctx->sqo_mm);
6227 io_iopoll_reap_events(ctx);
6228 io_sqe_buffer_unregister(ctx);
6229 io_sqe_files_unregister(ctx);
6230 io_eventfd_unregister(ctx);
6232 #if defined(CONFIG_UNIX)
6233 if (ctx->ring_sock) {
6234 ctx->ring_sock->file = NULL; /* so that iput() is called */
6235 sock_release(ctx->ring_sock);
6239 io_mem_free(ctx->rings);
6240 io_mem_free(ctx->sq_sqes);
6242 percpu_ref_exit(&ctx->refs);
6243 if (ctx->account_mem)
6244 io_unaccount_mem(ctx->user,
6245 ring_pages(ctx->sq_entries, ctx->cq_entries));
6246 free_uid(ctx->user);
6247 put_cred(ctx->creds);
6248 kfree(ctx->completions);
6249 kfree(ctx->cancel_hash);
6250 kmem_cache_free(req_cachep, ctx->fallback_req);
6254 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
6256 struct io_ring_ctx *ctx = file->private_data;
6259 poll_wait(file, &ctx->cq_wait, wait);
6261 * synchronizes with barrier from wq_has_sleeper call in
6265 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
6266 ctx->rings->sq_ring_entries)
6267 mask |= EPOLLOUT | EPOLLWRNORM;
6268 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
6269 mask |= EPOLLIN | EPOLLRDNORM;
6274 static int io_uring_fasync(int fd, struct file *file, int on)
6276 struct io_ring_ctx *ctx = file->private_data;
6278 return fasync_helper(fd, file, on, &ctx->cq_fasync);
6281 static int io_remove_personalities(int id, void *p, void *data)
6283 struct io_ring_ctx *ctx = data;
6284 const struct cred *cred;
6286 cred = idr_remove(&ctx->personality_idr, id);
6292 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
6294 mutex_lock(&ctx->uring_lock);
6295 percpu_ref_kill(&ctx->refs);
6296 mutex_unlock(&ctx->uring_lock);
6298 io_kill_timeouts(ctx);
6299 io_poll_remove_all(ctx);
6302 io_wq_cancel_all(ctx->io_wq);
6304 io_iopoll_reap_events(ctx);
6305 /* if we failed setting up the ctx, we might not have any rings */
6307 io_cqring_overflow_flush(ctx, true);
6308 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
6309 wait_for_completion(&ctx->completions[0]);
6310 io_ring_ctx_free(ctx);
6313 static int io_uring_release(struct inode *inode, struct file *file)
6315 struct io_ring_ctx *ctx = file->private_data;
6317 file->private_data = NULL;
6318 io_ring_ctx_wait_and_kill(ctx);
6322 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
6323 struct files_struct *files)
6325 struct io_kiocb *req;
6328 while (!list_empty_careful(&ctx->inflight_list)) {
6329 struct io_kiocb *cancel_req = NULL;
6331 spin_lock_irq(&ctx->inflight_lock);
6332 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
6333 if (req->work.files != files)
6335 /* req is being completed, ignore */
6336 if (!refcount_inc_not_zero(&req->refs))
6342 prepare_to_wait(&ctx->inflight_wait, &wait,
6343 TASK_UNINTERRUPTIBLE);
6344 spin_unlock_irq(&ctx->inflight_lock);
6346 /* We need to keep going until we don't find a matching req */
6350 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
6351 io_put_req(cancel_req);
6354 finish_wait(&ctx->inflight_wait, &wait);
6357 static int io_uring_flush(struct file *file, void *data)
6359 struct io_ring_ctx *ctx = file->private_data;
6361 io_uring_cancel_files(ctx, data);
6365 static void *io_uring_validate_mmap_request(struct file *file,
6366 loff_t pgoff, size_t sz)
6368 struct io_ring_ctx *ctx = file->private_data;
6369 loff_t offset = pgoff << PAGE_SHIFT;
6374 case IORING_OFF_SQ_RING:
6375 case IORING_OFF_CQ_RING:
6378 case IORING_OFF_SQES:
6382 return ERR_PTR(-EINVAL);
6385 page = virt_to_head_page(ptr);
6386 if (sz > page_size(page))
6387 return ERR_PTR(-EINVAL);
6394 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6396 size_t sz = vma->vm_end - vma->vm_start;
6400 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
6402 return PTR_ERR(ptr);
6404 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
6405 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
6408 #else /* !CONFIG_MMU */
6410 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6412 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
6415 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
6417 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
6420 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
6421 unsigned long addr, unsigned long len,
6422 unsigned long pgoff, unsigned long flags)
6426 ptr = io_uring_validate_mmap_request(file, pgoff, len);
6428 return PTR_ERR(ptr);
6430 return (unsigned long) ptr;
6433 #endif /* !CONFIG_MMU */
6435 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
6436 u32, min_complete, u32, flags, const sigset_t __user *, sig,
6439 struct io_ring_ctx *ctx;
6444 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
6452 if (f.file->f_op != &io_uring_fops)
6456 ctx = f.file->private_data;
6457 if (!percpu_ref_tryget(&ctx->refs))
6461 * For SQ polling, the thread will do all submissions and completions.
6462 * Just return the requested submit count, and wake the thread if
6466 if (ctx->flags & IORING_SETUP_SQPOLL) {
6467 if (!list_empty_careful(&ctx->cq_overflow_list))
6468 io_cqring_overflow_flush(ctx, false);
6469 if (flags & IORING_ENTER_SQ_WAKEUP)
6470 wake_up(&ctx->sqo_wait);
6471 submitted = to_submit;
6472 } else if (to_submit) {
6473 struct mm_struct *cur_mm;
6475 mutex_lock(&ctx->uring_lock);
6476 /* already have mm, so io_submit_sqes() won't try to grab it */
6477 cur_mm = ctx->sqo_mm;
6478 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
6480 mutex_unlock(&ctx->uring_lock);
6482 if (submitted != to_submit)
6485 if (flags & IORING_ENTER_GETEVENTS) {
6486 unsigned nr_events = 0;
6488 min_complete = min(min_complete, ctx->cq_entries);
6490 if (ctx->flags & IORING_SETUP_IOPOLL) {
6491 ret = io_iopoll_check(ctx, &nr_events, min_complete);
6493 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
6498 percpu_ref_put(&ctx->refs);
6501 return submitted ? submitted : ret;
6504 static const struct file_operations io_uring_fops = {
6505 .release = io_uring_release,
6506 .flush = io_uring_flush,
6507 .mmap = io_uring_mmap,
6509 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
6510 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
6512 .poll = io_uring_poll,
6513 .fasync = io_uring_fasync,
6516 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
6517 struct io_uring_params *p)
6519 struct io_rings *rings;
6520 size_t size, sq_array_offset;
6522 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
6523 if (size == SIZE_MAX)
6526 rings = io_mem_alloc(size);
6531 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
6532 rings->sq_ring_mask = p->sq_entries - 1;
6533 rings->cq_ring_mask = p->cq_entries - 1;
6534 rings->sq_ring_entries = p->sq_entries;
6535 rings->cq_ring_entries = p->cq_entries;
6536 ctx->sq_mask = rings->sq_ring_mask;
6537 ctx->cq_mask = rings->cq_ring_mask;
6538 ctx->sq_entries = rings->sq_ring_entries;
6539 ctx->cq_entries = rings->cq_ring_entries;
6541 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
6542 if (size == SIZE_MAX) {
6543 io_mem_free(ctx->rings);
6548 ctx->sq_sqes = io_mem_alloc(size);
6549 if (!ctx->sq_sqes) {
6550 io_mem_free(ctx->rings);
6559 * Allocate an anonymous fd, this is what constitutes the application
6560 * visible backing of an io_uring instance. The application mmaps this
6561 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
6562 * we have to tie this fd to a socket for file garbage collection purposes.
6564 static int io_uring_get_fd(struct io_ring_ctx *ctx)
6569 #if defined(CONFIG_UNIX)
6570 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
6576 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
6580 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
6581 O_RDWR | O_CLOEXEC);
6584 ret = PTR_ERR(file);
6588 #if defined(CONFIG_UNIX)
6589 ctx->ring_sock->file = file;
6591 fd_install(ret, file);
6594 #if defined(CONFIG_UNIX)
6595 sock_release(ctx->ring_sock);
6596 ctx->ring_sock = NULL;
6601 static int io_uring_create(unsigned entries, struct io_uring_params *p)
6603 struct user_struct *user = NULL;
6604 struct io_ring_ctx *ctx;
6610 if (entries > IORING_MAX_ENTRIES) {
6611 if (!(p->flags & IORING_SETUP_CLAMP))
6613 entries = IORING_MAX_ENTRIES;
6617 * Use twice as many entries for the CQ ring. It's possible for the
6618 * application to drive a higher depth than the size of the SQ ring,
6619 * since the sqes are only used at submission time. This allows for
6620 * some flexibility in overcommitting a bit. If the application has
6621 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
6622 * of CQ ring entries manually.
6624 p->sq_entries = roundup_pow_of_two(entries);
6625 if (p->flags & IORING_SETUP_CQSIZE) {
6627 * If IORING_SETUP_CQSIZE is set, we do the same roundup
6628 * to a power-of-two, if it isn't already. We do NOT impose
6629 * any cq vs sq ring sizing.
6631 if (p->cq_entries < p->sq_entries)
6633 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
6634 if (!(p->flags & IORING_SETUP_CLAMP))
6636 p->cq_entries = IORING_MAX_CQ_ENTRIES;
6638 p->cq_entries = roundup_pow_of_two(p->cq_entries);
6640 p->cq_entries = 2 * p->sq_entries;
6643 user = get_uid(current_user());
6644 account_mem = !capable(CAP_IPC_LOCK);
6647 ret = io_account_mem(user,
6648 ring_pages(p->sq_entries, p->cq_entries));
6655 ctx = io_ring_ctx_alloc(p);
6658 io_unaccount_mem(user, ring_pages(p->sq_entries,
6663 ctx->compat = in_compat_syscall();
6664 ctx->account_mem = account_mem;
6666 ctx->creds = get_current_cred();
6668 ret = io_allocate_scq_urings(ctx, p);
6672 ret = io_sq_offload_start(ctx, p);
6676 memset(&p->sq_off, 0, sizeof(p->sq_off));
6677 p->sq_off.head = offsetof(struct io_rings, sq.head);
6678 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
6679 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
6680 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
6681 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
6682 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
6683 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
6685 memset(&p->cq_off, 0, sizeof(p->cq_off));
6686 p->cq_off.head = offsetof(struct io_rings, cq.head);
6687 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
6688 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
6689 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
6690 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
6691 p->cq_off.cqes = offsetof(struct io_rings, cqes);
6694 * Install ring fd as the very last thing, so we don't risk someone
6695 * having closed it before we finish setup
6697 ret = io_uring_get_fd(ctx);
6701 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
6702 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
6703 IORING_FEAT_CUR_PERSONALITY;
6704 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
6707 io_ring_ctx_wait_and_kill(ctx);
6712 * Sets up an aio uring context, and returns the fd. Applications asks for a
6713 * ring size, we return the actual sq/cq ring sizes (among other things) in the
6714 * params structure passed in.
6716 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
6718 struct io_uring_params p;
6722 if (copy_from_user(&p, params, sizeof(p)))
6724 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
6729 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
6730 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
6731 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
6734 ret = io_uring_create(entries, &p);
6738 if (copy_to_user(params, &p, sizeof(p)))
6744 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
6745 struct io_uring_params __user *, params)
6747 return io_uring_setup(entries, params);
6750 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
6752 struct io_uring_probe *p;
6756 size = struct_size(p, ops, nr_args);
6757 if (size == SIZE_MAX)
6759 p = kzalloc(size, GFP_KERNEL);
6764 if (copy_from_user(p, arg, size))
6767 if (memchr_inv(p, 0, size))
6770 p->last_op = IORING_OP_LAST - 1;
6771 if (nr_args > IORING_OP_LAST)
6772 nr_args = IORING_OP_LAST;
6774 for (i = 0; i < nr_args; i++) {
6776 if (!io_op_defs[i].not_supported)
6777 p->ops[i].flags = IO_URING_OP_SUPPORTED;
6782 if (copy_to_user(arg, p, size))
6789 static int io_register_personality(struct io_ring_ctx *ctx)
6791 const struct cred *creds = get_current_cred();
6794 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
6795 USHRT_MAX, GFP_KERNEL);
6801 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
6803 const struct cred *old_creds;
6805 old_creds = idr_remove(&ctx->personality_idr, id);
6807 put_cred(old_creds);
6814 static bool io_register_op_must_quiesce(int op)
6817 case IORING_UNREGISTER_FILES:
6818 case IORING_REGISTER_FILES_UPDATE:
6819 case IORING_REGISTER_PROBE:
6820 case IORING_REGISTER_PERSONALITY:
6821 case IORING_UNREGISTER_PERSONALITY:
6828 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
6829 void __user *arg, unsigned nr_args)
6830 __releases(ctx->uring_lock)
6831 __acquires(ctx->uring_lock)
6836 * We're inside the ring mutex, if the ref is already dying, then
6837 * someone else killed the ctx or is already going through
6838 * io_uring_register().
6840 if (percpu_ref_is_dying(&ctx->refs))
6843 if (io_register_op_must_quiesce(opcode)) {
6844 percpu_ref_kill(&ctx->refs);
6847 * Drop uring mutex before waiting for references to exit. If
6848 * another thread is currently inside io_uring_enter() it might
6849 * need to grab the uring_lock to make progress. If we hold it
6850 * here across the drain wait, then we can deadlock. It's safe
6851 * to drop the mutex here, since no new references will come in
6852 * after we've killed the percpu ref.
6854 mutex_unlock(&ctx->uring_lock);
6855 ret = wait_for_completion_interruptible(&ctx->completions[0]);
6856 mutex_lock(&ctx->uring_lock);
6858 percpu_ref_resurrect(&ctx->refs);
6865 case IORING_REGISTER_BUFFERS:
6866 ret = io_sqe_buffer_register(ctx, arg, nr_args);
6868 case IORING_UNREGISTER_BUFFERS:
6872 ret = io_sqe_buffer_unregister(ctx);
6874 case IORING_REGISTER_FILES:
6875 ret = io_sqe_files_register(ctx, arg, nr_args);
6877 case IORING_UNREGISTER_FILES:
6881 ret = io_sqe_files_unregister(ctx);
6883 case IORING_REGISTER_FILES_UPDATE:
6884 ret = io_sqe_files_update(ctx, arg, nr_args);
6886 case IORING_REGISTER_EVENTFD:
6887 case IORING_REGISTER_EVENTFD_ASYNC:
6891 ret = io_eventfd_register(ctx, arg);
6894 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
6895 ctx->eventfd_async = 1;
6897 ctx->eventfd_async = 0;
6899 case IORING_UNREGISTER_EVENTFD:
6903 ret = io_eventfd_unregister(ctx);
6905 case IORING_REGISTER_PROBE:
6907 if (!arg || nr_args > 256)
6909 ret = io_probe(ctx, arg, nr_args);
6911 case IORING_REGISTER_PERSONALITY:
6915 ret = io_register_personality(ctx);
6917 case IORING_UNREGISTER_PERSONALITY:
6921 ret = io_unregister_personality(ctx, nr_args);
6928 if (io_register_op_must_quiesce(opcode)) {
6929 /* bring the ctx back to life */
6930 percpu_ref_reinit(&ctx->refs);
6932 reinit_completion(&ctx->completions[0]);
6937 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
6938 void __user *, arg, unsigned int, nr_args)
6940 struct io_ring_ctx *ctx;
6949 if (f.file->f_op != &io_uring_fops)
6952 ctx = f.file->private_data;
6954 mutex_lock(&ctx->uring_lock);
6955 ret = __io_uring_register(ctx, opcode, arg, nr_args);
6956 mutex_unlock(&ctx->uring_lock);
6957 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
6958 ctx->cq_ev_fd != NULL, ret);
6964 static int __init io_uring_init(void)
6966 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
6967 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
6970 __initcall(io_uring_init);