1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/mmu_context.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/io_uring.h>
83 #include <uapi/linux/io_uring.h>
88 #define IORING_MAX_ENTRIES 32768
89 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
92 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
94 #define IORING_FILE_TABLE_SHIFT 9
95 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
96 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
97 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 u32 head ____cacheline_aligned_in_smp;
101 u32 tail ____cacheline_aligned_in_smp;
105 * This data is shared with the application through the mmap at offsets
106 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
108 * The offsets to the member fields are published through struct
109 * io_sqring_offsets when calling io_uring_setup.
113 * Head and tail offsets into the ring; the offsets need to be
114 * masked to get valid indices.
116 * The kernel controls head of the sq ring and the tail of the cq ring,
117 * and the application controls tail of the sq ring and the head of the
120 struct io_uring sq, cq;
122 * Bitmasks to apply to head and tail offsets (constant, equals
125 u32 sq_ring_mask, cq_ring_mask;
126 /* Ring sizes (constant, power of 2) */
127 u32 sq_ring_entries, cq_ring_entries;
129 * Number of invalid entries dropped by the kernel due to
130 * invalid index stored in array
132 * Written by the kernel, shouldn't be modified by the
133 * application (i.e. get number of "new events" by comparing to
136 * After a new SQ head value was read by the application this
137 * counter includes all submissions that were dropped reaching
138 * the new SQ head (and possibly more).
144 * Written by the kernel, shouldn't be modified by the
147 * The application needs a full memory barrier before checking
148 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
152 * Number of completion events lost because the queue was full;
153 * this should be avoided by the application by making sure
154 * there are not more requests pending than there is space in
155 * the completion queue.
157 * Written by the kernel, shouldn't be modified by the
158 * application (i.e. get number of "new events" by comparing to
161 * As completion events come in out of order this counter is not
162 * ordered with any other data.
166 * Ring buffer of completion events.
168 * The kernel writes completion events fresh every time they are
169 * produced, so the application is allowed to modify pending
172 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
175 struct io_mapped_ubuf {
178 struct bio_vec *bvec;
179 unsigned int nr_bvecs;
182 struct fixed_file_table {
186 struct fixed_file_data {
187 struct fixed_file_table *table;
188 struct io_ring_ctx *ctx;
190 struct percpu_ref refs;
191 struct llist_head put_llist;
192 struct work_struct ref_work;
193 struct completion done;
198 struct percpu_ref refs;
199 } ____cacheline_aligned_in_smp;
203 unsigned int compat: 1;
204 unsigned int account_mem: 1;
205 unsigned int cq_overflow_flushed: 1;
206 unsigned int drain_next: 1;
207 unsigned int eventfd_async: 1;
210 * Ring buffer of indices into array of io_uring_sqe, which is
211 * mmapped by the application using the IORING_OFF_SQES offset.
213 * This indirection could e.g. be used to assign fixed
214 * io_uring_sqe entries to operations and only submit them to
215 * the queue when needed.
217 * The kernel modifies neither the indices array nor the entries
221 unsigned cached_sq_head;
224 unsigned sq_thread_idle;
225 unsigned cached_sq_dropped;
226 atomic_t cached_cq_overflow;
227 unsigned long sq_check_overflow;
229 struct list_head defer_list;
230 struct list_head timeout_list;
231 struct list_head cq_overflow_list;
233 wait_queue_head_t inflight_wait;
234 struct io_uring_sqe *sq_sqes;
235 } ____cacheline_aligned_in_smp;
237 struct io_rings *rings;
241 struct task_struct *sqo_thread; /* if using sq thread polling */
242 struct mm_struct *sqo_mm;
243 wait_queue_head_t sqo_wait;
246 * If used, fixed file set. Writers must ensure that ->refs is dead,
247 * readers must ensure that ->refs is alive as long as the file* is
248 * used. Only updated through io_uring_register(2).
250 struct fixed_file_data *file_data;
251 unsigned nr_user_files;
253 struct file *ring_file;
255 /* if used, fixed mapped user buffers */
256 unsigned nr_user_bufs;
257 struct io_mapped_ubuf *user_bufs;
259 struct user_struct *user;
261 const struct cred *creds;
263 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
264 struct completion *completions;
266 /* if all else fails... */
267 struct io_kiocb *fallback_req;
269 #if defined(CONFIG_UNIX)
270 struct socket *ring_sock;
273 struct idr personality_idr;
276 unsigned cached_cq_tail;
279 atomic_t cq_timeouts;
280 unsigned long cq_check_overflow;
281 struct wait_queue_head cq_wait;
282 struct fasync_struct *cq_fasync;
283 struct eventfd_ctx *cq_ev_fd;
284 } ____cacheline_aligned_in_smp;
287 struct mutex uring_lock;
288 wait_queue_head_t wait;
289 } ____cacheline_aligned_in_smp;
292 spinlock_t completion_lock;
293 struct llist_head poll_llist;
296 * ->poll_list is protected by the ctx->uring_lock for
297 * io_uring instances that don't use IORING_SETUP_SQPOLL.
298 * For SQPOLL, only the single threaded io_sq_thread() will
299 * manipulate the list, hence no extra locking is needed there.
301 struct list_head poll_list;
302 struct hlist_head *cancel_hash;
303 unsigned cancel_hash_bits;
304 bool poll_multi_file;
306 spinlock_t inflight_lock;
307 struct list_head inflight_list;
308 } ____cacheline_aligned_in_smp;
312 * First field must be the file pointer in all the
313 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
315 struct io_poll_iocb {
318 struct wait_queue_head *head;
324 struct wait_queue_entry wait;
329 struct file *put_file;
333 struct io_timeout_data {
334 struct io_kiocb *req;
335 struct hrtimer timer;
336 struct timespec64 ts;
337 enum hrtimer_mode mode;
343 struct sockaddr __user *addr;
344 int __user *addr_len;
369 /* NOTE: kiocb has the file as the first member, so don't do it here */
377 struct sockaddr __user *addr;
384 struct user_msghdr __user *msg;
397 struct filename *filename;
398 struct statx __user *buffer;
402 struct io_files_update {
428 struct epoll_event event;
431 struct io_async_connect {
432 struct sockaddr_storage address;
435 struct io_async_msghdr {
436 struct iovec fast_iov[UIO_FASTIOV];
438 struct sockaddr __user *uaddr;
440 struct sockaddr_storage addr;
444 struct iovec fast_iov[UIO_FASTIOV];
450 struct io_async_ctx {
452 struct io_async_rw rw;
453 struct io_async_msghdr msg;
454 struct io_async_connect connect;
455 struct io_timeout_data timeout;
460 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
461 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
462 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
463 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
464 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
471 REQ_F_IOPOLL_COMPLETED_BIT,
472 REQ_F_LINK_TIMEOUT_BIT,
476 REQ_F_TIMEOUT_NOSEQ_BIT,
477 REQ_F_COMP_LOCKED_BIT,
478 REQ_F_NEED_CLEANUP_BIT,
484 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
485 /* drain existing IO first */
486 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
488 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
489 /* doesn't sever on completion < 0 */
490 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
492 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
494 /* already grabbed next link */
495 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
496 /* fail rest of links */
497 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
498 /* on inflight list */
499 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
500 /* read/write uses file position */
501 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
502 /* must not punt to workers */
503 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
504 /* polled IO has completed */
505 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
506 /* has linked timeout */
507 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
508 /* timeout request */
509 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
511 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
512 /* must be punted even for NONBLOCK */
513 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
514 /* no timeout sequence */
515 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
516 /* completion under lock */
517 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
519 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
520 /* in overflow list */
521 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
525 * NOTE! Each of the iocb union members has the file pointer
526 * as the first entry in their struct definition. So you can
527 * access the file pointer through any of the sub-structs,
528 * or directly as just 'ki_filp' in this struct.
534 struct io_poll_iocb poll;
535 struct io_accept accept;
537 struct io_cancel cancel;
538 struct io_timeout timeout;
539 struct io_connect connect;
540 struct io_sr_msg sr_msg;
542 struct io_close close;
543 struct io_files_update files_update;
544 struct io_fadvise fadvise;
545 struct io_madvise madvise;
546 struct io_epoll epoll;
549 struct io_async_ctx *io;
551 * llist_node is only used for poll deferred completions
553 struct llist_node llist_node;
555 bool needs_fixed_file;
558 struct io_ring_ctx *ctx;
560 struct list_head list;
561 struct hlist_node hash_node;
563 struct list_head link_list;
570 struct list_head inflight_entry;
572 struct io_wq_work work;
575 #define IO_PLUG_THRESHOLD 2
576 #define IO_IOPOLL_BATCH 8
578 struct io_submit_state {
579 struct blk_plug plug;
582 * io_kiocb alloc cache
584 void *reqs[IO_IOPOLL_BATCH];
585 unsigned int free_reqs;
588 * File reference cache
592 unsigned int has_refs;
593 unsigned int used_refs;
594 unsigned int ios_left;
598 /* needs req->io allocated for deferral/async */
599 unsigned async_ctx : 1;
600 /* needs current->mm setup, does mm access */
601 unsigned needs_mm : 1;
602 /* needs req->file assigned */
603 unsigned needs_file : 1;
604 /* needs req->file assigned IFF fd is >= 0 */
605 unsigned fd_non_neg : 1;
606 /* hash wq insertion if file is a regular file */
607 unsigned hash_reg_file : 1;
608 /* unbound wq insertion if file is a non-regular file */
609 unsigned unbound_nonreg_file : 1;
610 /* opcode is not supported by this kernel */
611 unsigned not_supported : 1;
612 /* needs file table */
613 unsigned file_table : 1;
615 unsigned needs_fs : 1;
618 static const struct io_op_def io_op_defs[] = {
619 [IORING_OP_NOP] = {},
620 [IORING_OP_READV] = {
624 .unbound_nonreg_file = 1,
626 [IORING_OP_WRITEV] = {
631 .unbound_nonreg_file = 1,
633 [IORING_OP_FSYNC] = {
636 [IORING_OP_READ_FIXED] = {
638 .unbound_nonreg_file = 1,
640 [IORING_OP_WRITE_FIXED] = {
643 .unbound_nonreg_file = 1,
645 [IORING_OP_POLL_ADD] = {
647 .unbound_nonreg_file = 1,
649 [IORING_OP_POLL_REMOVE] = {},
650 [IORING_OP_SYNC_FILE_RANGE] = {
653 [IORING_OP_SENDMSG] = {
657 .unbound_nonreg_file = 1,
660 [IORING_OP_RECVMSG] = {
664 .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] = {
698 [IORING_OP_CLOSE] = {
702 [IORING_OP_FILES_UPDATE] = {
706 [IORING_OP_STATX] = {
715 .unbound_nonreg_file = 1,
717 [IORING_OP_WRITE] = {
720 .unbound_nonreg_file = 1,
722 [IORING_OP_FADVISE] = {
725 [IORING_OP_MADVISE] = {
731 .unbound_nonreg_file = 1,
736 .unbound_nonreg_file = 1,
738 [IORING_OP_OPENAT2] = {
744 [IORING_OP_EPOLL_CTL] = {
745 .unbound_nonreg_file = 1,
750 static void io_wq_submit_work(struct io_wq_work **workptr);
751 static void io_cqring_fill_event(struct io_kiocb *req, long res);
752 static void io_put_req(struct io_kiocb *req);
753 static void __io_double_put_req(struct io_kiocb *req);
754 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
755 static void io_queue_linked_timeout(struct io_kiocb *req);
756 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
757 struct io_uring_files_update *ip,
759 static int io_grab_files(struct io_kiocb *req);
760 static void io_ring_file_ref_flush(struct fixed_file_data *data);
761 static void io_cleanup_req(struct io_kiocb *req);
763 static struct kmem_cache *req_cachep;
765 static const struct file_operations io_uring_fops;
767 struct sock *io_uring_get_socket(struct file *file)
769 #if defined(CONFIG_UNIX)
770 if (file->f_op == &io_uring_fops) {
771 struct io_ring_ctx *ctx = file->private_data;
773 return ctx->ring_sock->sk;
778 EXPORT_SYMBOL(io_uring_get_socket);
780 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
782 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
784 complete(&ctx->completions[0]);
787 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
789 struct io_ring_ctx *ctx;
792 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
796 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
797 if (!ctx->fallback_req)
800 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
801 if (!ctx->completions)
805 * Use 5 bits less than the max cq entries, that should give us around
806 * 32 entries per hash list if totally full and uniformly spread.
808 hash_bits = ilog2(p->cq_entries);
812 ctx->cancel_hash_bits = hash_bits;
813 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
815 if (!ctx->cancel_hash)
817 __hash_init(ctx->cancel_hash, 1U << hash_bits);
819 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
820 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
823 ctx->flags = p->flags;
824 init_waitqueue_head(&ctx->cq_wait);
825 INIT_LIST_HEAD(&ctx->cq_overflow_list);
826 init_completion(&ctx->completions[0]);
827 init_completion(&ctx->completions[1]);
828 idr_init(&ctx->personality_idr);
829 mutex_init(&ctx->uring_lock);
830 init_waitqueue_head(&ctx->wait);
831 spin_lock_init(&ctx->completion_lock);
832 init_llist_head(&ctx->poll_llist);
833 INIT_LIST_HEAD(&ctx->poll_list);
834 INIT_LIST_HEAD(&ctx->defer_list);
835 INIT_LIST_HEAD(&ctx->timeout_list);
836 init_waitqueue_head(&ctx->inflight_wait);
837 spin_lock_init(&ctx->inflight_lock);
838 INIT_LIST_HEAD(&ctx->inflight_list);
841 if (ctx->fallback_req)
842 kmem_cache_free(req_cachep, ctx->fallback_req);
843 kfree(ctx->completions);
844 kfree(ctx->cancel_hash);
849 static inline bool __req_need_defer(struct io_kiocb *req)
851 struct io_ring_ctx *ctx = req->ctx;
853 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
854 + atomic_read(&ctx->cached_cq_overflow);
857 static inline bool req_need_defer(struct io_kiocb *req)
859 if (unlikely(req->flags & REQ_F_IO_DRAIN))
860 return __req_need_defer(req);
865 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
867 struct io_kiocb *req;
869 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
870 if (req && !req_need_defer(req)) {
871 list_del_init(&req->list);
878 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
880 struct io_kiocb *req;
882 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
884 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
886 if (!__req_need_defer(req)) {
887 list_del_init(&req->list);
895 static void __io_commit_cqring(struct io_ring_ctx *ctx)
897 struct io_rings *rings = ctx->rings;
899 /* order cqe stores with ring update */
900 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
902 if (wq_has_sleeper(&ctx->cq_wait)) {
903 wake_up_interruptible(&ctx->cq_wait);
904 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
908 static inline void io_req_work_grab_env(struct io_kiocb *req,
909 const struct io_op_def *def)
911 if (!req->work.mm && def->needs_mm) {
913 req->work.mm = current->mm;
915 if (!req->work.creds)
916 req->work.creds = get_current_cred();
917 if (!req->work.fs && def->needs_fs) {
918 spin_lock(¤t->fs->lock);
919 if (!current->fs->in_exec) {
920 req->work.fs = current->fs;
921 req->work.fs->users++;
923 req->work.flags |= IO_WQ_WORK_CANCEL;
925 spin_unlock(¤t->fs->lock);
927 if (!req->work.task_pid)
928 req->work.task_pid = task_pid_vnr(current);
931 static inline void io_req_work_drop_env(struct io_kiocb *req)
934 mmdrop(req->work.mm);
937 if (req->work.creds) {
938 put_cred(req->work.creds);
939 req->work.creds = NULL;
942 struct fs_struct *fs = req->work.fs;
944 spin_lock(&req->work.fs->lock);
947 spin_unlock(&req->work.fs->lock);
953 static inline bool io_prep_async_work(struct io_kiocb *req,
954 struct io_kiocb **link)
956 const struct io_op_def *def = &io_op_defs[req->opcode];
957 bool do_hashed = false;
959 if (req->flags & REQ_F_ISREG) {
960 if (def->hash_reg_file)
963 if (def->unbound_nonreg_file)
964 req->work.flags |= IO_WQ_WORK_UNBOUND;
967 io_req_work_grab_env(req, def);
969 *link = io_prep_linked_timeout(req);
973 static inline void io_queue_async_work(struct io_kiocb *req)
975 struct io_ring_ctx *ctx = req->ctx;
976 struct io_kiocb *link;
979 do_hashed = io_prep_async_work(req, &link);
981 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
984 io_wq_enqueue(ctx->io_wq, &req->work);
986 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
987 file_inode(req->file));
991 io_queue_linked_timeout(link);
994 static void io_kill_timeout(struct io_kiocb *req)
998 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1000 atomic_inc(&req->ctx->cq_timeouts);
1001 list_del_init(&req->list);
1002 req->flags |= REQ_F_COMP_LOCKED;
1003 io_cqring_fill_event(req, 0);
1008 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1010 struct io_kiocb *req, *tmp;
1012 spin_lock_irq(&ctx->completion_lock);
1013 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1014 io_kill_timeout(req);
1015 spin_unlock_irq(&ctx->completion_lock);
1018 static void io_commit_cqring(struct io_ring_ctx *ctx)
1020 struct io_kiocb *req;
1022 while ((req = io_get_timeout_req(ctx)) != NULL)
1023 io_kill_timeout(req);
1025 __io_commit_cqring(ctx);
1027 while ((req = io_get_deferred_req(ctx)) != NULL)
1028 io_queue_async_work(req);
1031 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1033 struct io_rings *rings = ctx->rings;
1036 tail = ctx->cached_cq_tail;
1038 * writes to the cq entry need to come after reading head; the
1039 * control dependency is enough as we're using WRITE_ONCE to
1042 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1045 ctx->cached_cq_tail++;
1046 return &rings->cqes[tail & ctx->cq_mask];
1049 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1053 if (!ctx->eventfd_async)
1055 return io_wq_current_is_worker() || in_interrupt();
1058 static void __io_cqring_ev_posted(struct io_ring_ctx *ctx, bool trigger_ev)
1060 if (waitqueue_active(&ctx->wait))
1061 wake_up(&ctx->wait);
1062 if (waitqueue_active(&ctx->sqo_wait))
1063 wake_up(&ctx->sqo_wait);
1065 eventfd_signal(ctx->cq_ev_fd, 1);
1068 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1070 __io_cqring_ev_posted(ctx, io_should_trigger_evfd(ctx));
1073 /* Returns true if there are no backlogged entries after the flush */
1074 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1076 struct io_rings *rings = ctx->rings;
1077 struct io_uring_cqe *cqe;
1078 struct io_kiocb *req;
1079 unsigned long flags;
1083 if (list_empty_careful(&ctx->cq_overflow_list))
1085 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1086 rings->cq_ring_entries))
1090 spin_lock_irqsave(&ctx->completion_lock, flags);
1092 /* if force is set, the ring is going away. always drop after that */
1094 ctx->cq_overflow_flushed = 1;
1097 while (!list_empty(&ctx->cq_overflow_list)) {
1098 cqe = io_get_cqring(ctx);
1102 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1104 list_move(&req->list, &list);
1105 req->flags &= ~REQ_F_OVERFLOW;
1107 WRITE_ONCE(cqe->user_data, req->user_data);
1108 WRITE_ONCE(cqe->res, req->result);
1109 WRITE_ONCE(cqe->flags, 0);
1111 WRITE_ONCE(ctx->rings->cq_overflow,
1112 atomic_inc_return(&ctx->cached_cq_overflow));
1116 io_commit_cqring(ctx);
1118 clear_bit(0, &ctx->sq_check_overflow);
1119 clear_bit(0, &ctx->cq_check_overflow);
1121 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1122 io_cqring_ev_posted(ctx);
1124 while (!list_empty(&list)) {
1125 req = list_first_entry(&list, struct io_kiocb, list);
1126 list_del(&req->list);
1133 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1135 struct io_ring_ctx *ctx = req->ctx;
1136 struct io_uring_cqe *cqe;
1138 trace_io_uring_complete(ctx, req->user_data, res);
1141 * If we can't get a cq entry, userspace overflowed the
1142 * submission (by quite a lot). Increment the overflow count in
1145 cqe = io_get_cqring(ctx);
1147 WRITE_ONCE(cqe->user_data, req->user_data);
1148 WRITE_ONCE(cqe->res, res);
1149 WRITE_ONCE(cqe->flags, 0);
1150 } else if (ctx->cq_overflow_flushed) {
1151 WRITE_ONCE(ctx->rings->cq_overflow,
1152 atomic_inc_return(&ctx->cached_cq_overflow));
1154 if (list_empty(&ctx->cq_overflow_list)) {
1155 set_bit(0, &ctx->sq_check_overflow);
1156 set_bit(0, &ctx->cq_check_overflow);
1158 req->flags |= REQ_F_OVERFLOW;
1159 refcount_inc(&req->refs);
1161 list_add_tail(&req->list, &ctx->cq_overflow_list);
1165 static void io_cqring_add_event(struct io_kiocb *req, long res)
1167 struct io_ring_ctx *ctx = req->ctx;
1168 unsigned long flags;
1170 spin_lock_irqsave(&ctx->completion_lock, flags);
1171 io_cqring_fill_event(req, res);
1172 io_commit_cqring(ctx);
1173 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1175 io_cqring_ev_posted(ctx);
1178 static inline bool io_is_fallback_req(struct io_kiocb *req)
1180 return req == (struct io_kiocb *)
1181 ((unsigned long) req->ctx->fallback_req & ~1UL);
1184 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1186 struct io_kiocb *req;
1188 req = ctx->fallback_req;
1189 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
1195 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
1196 struct io_submit_state *state)
1198 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1199 struct io_kiocb *req;
1202 req = kmem_cache_alloc(req_cachep, gfp);
1205 } else if (!state->free_reqs) {
1209 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1210 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1213 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1214 * retry single alloc to be on the safe side.
1216 if (unlikely(ret <= 0)) {
1217 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1218 if (!state->reqs[0])
1222 state->free_reqs = ret - 1;
1223 req = state->reqs[ret - 1];
1226 req = state->reqs[state->free_reqs];
1234 /* one is dropped after submission, the other at completion */
1235 refcount_set(&req->refs, 2);
1237 INIT_IO_WORK(&req->work, io_wq_submit_work);
1240 req = io_get_fallback_req(ctx);
1243 percpu_ref_put(&ctx->refs);
1247 static void __io_req_do_free(struct io_kiocb *req)
1249 if (likely(!io_is_fallback_req(req)))
1250 kmem_cache_free(req_cachep, req);
1252 clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req);
1255 static void __io_req_aux_free(struct io_kiocb *req)
1257 struct io_ring_ctx *ctx = req->ctx;
1259 if (req->flags & REQ_F_NEED_CLEANUP)
1260 io_cleanup_req(req);
1264 if (req->flags & REQ_F_FIXED_FILE)
1265 percpu_ref_put(&ctx->file_data->refs);
1270 io_req_work_drop_env(req);
1273 static void __io_free_req(struct io_kiocb *req)
1275 __io_req_aux_free(req);
1277 if (req->flags & REQ_F_INFLIGHT) {
1278 struct io_ring_ctx *ctx = req->ctx;
1279 unsigned long flags;
1281 spin_lock_irqsave(&ctx->inflight_lock, flags);
1282 list_del(&req->inflight_entry);
1283 if (waitqueue_active(&ctx->inflight_wait))
1284 wake_up(&ctx->inflight_wait);
1285 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1288 percpu_ref_put(&req->ctx->refs);
1289 __io_req_do_free(req);
1293 void *reqs[IO_IOPOLL_BATCH];
1298 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1300 int fixed_refs = rb->to_free;
1304 if (rb->need_iter) {
1305 int i, inflight = 0;
1306 unsigned long flags;
1309 for (i = 0; i < rb->to_free; i++) {
1310 struct io_kiocb *req = rb->reqs[i];
1312 if (req->flags & REQ_F_FIXED_FILE) {
1316 if (req->flags & REQ_F_INFLIGHT)
1318 __io_req_aux_free(req);
1323 spin_lock_irqsave(&ctx->inflight_lock, flags);
1324 for (i = 0; i < rb->to_free; i++) {
1325 struct io_kiocb *req = rb->reqs[i];
1327 if (req->flags & REQ_F_INFLIGHT) {
1328 list_del(&req->inflight_entry);
1333 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1335 if (waitqueue_active(&ctx->inflight_wait))
1336 wake_up(&ctx->inflight_wait);
1339 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1341 percpu_ref_put_many(&ctx->file_data->refs, fixed_refs);
1342 percpu_ref_put_many(&ctx->refs, rb->to_free);
1343 rb->to_free = rb->need_iter = 0;
1346 static bool io_link_cancel_timeout(struct io_kiocb *req)
1348 struct io_ring_ctx *ctx = req->ctx;
1351 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1353 io_cqring_fill_event(req, -ECANCELED);
1354 io_commit_cqring(ctx);
1355 req->flags &= ~REQ_F_LINK;
1363 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1365 struct io_ring_ctx *ctx = req->ctx;
1366 bool wake_ev = false;
1368 /* Already got next link */
1369 if (req->flags & REQ_F_LINK_NEXT)
1373 * The list should never be empty when we are called here. But could
1374 * potentially happen if the chain is messed up, check to be on the
1377 while (!list_empty(&req->link_list)) {
1378 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1379 struct io_kiocb, link_list);
1381 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1382 (nxt->flags & REQ_F_TIMEOUT))) {
1383 list_del_init(&nxt->link_list);
1384 wake_ev |= io_link_cancel_timeout(nxt);
1385 req->flags &= ~REQ_F_LINK_TIMEOUT;
1389 list_del_init(&req->link_list);
1390 if (!list_empty(&nxt->link_list))
1391 nxt->flags |= REQ_F_LINK;
1396 req->flags |= REQ_F_LINK_NEXT;
1398 io_cqring_ev_posted(ctx);
1402 * Called if REQ_F_LINK is set, and we fail the head request
1404 static void io_fail_links(struct io_kiocb *req)
1406 struct io_ring_ctx *ctx = req->ctx;
1407 unsigned long flags;
1409 spin_lock_irqsave(&ctx->completion_lock, flags);
1411 while (!list_empty(&req->link_list)) {
1412 struct io_kiocb *link = list_first_entry(&req->link_list,
1413 struct io_kiocb, link_list);
1415 list_del_init(&link->link_list);
1416 trace_io_uring_fail_link(req, link);
1418 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1419 link->opcode == IORING_OP_LINK_TIMEOUT) {
1420 io_link_cancel_timeout(link);
1422 io_cqring_fill_event(link, -ECANCELED);
1423 __io_double_put_req(link);
1425 req->flags &= ~REQ_F_LINK_TIMEOUT;
1428 io_commit_cqring(ctx);
1429 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1430 io_cqring_ev_posted(ctx);
1433 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1435 if (likely(!(req->flags & REQ_F_LINK)))
1439 * If LINK is set, we have dependent requests in this chain. If we
1440 * didn't fail this request, queue the first one up, moving any other
1441 * dependencies to the next request. In case of failure, fail the rest
1444 if (req->flags & REQ_F_FAIL_LINK) {
1446 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1447 REQ_F_LINK_TIMEOUT) {
1448 struct io_ring_ctx *ctx = req->ctx;
1449 unsigned long flags;
1452 * If this is a timeout link, we could be racing with the
1453 * timeout timer. Grab the completion lock for this case to
1454 * protect against that.
1456 spin_lock_irqsave(&ctx->completion_lock, flags);
1457 io_req_link_next(req, nxt);
1458 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1460 io_req_link_next(req, nxt);
1464 static void io_free_req(struct io_kiocb *req)
1466 struct io_kiocb *nxt = NULL;
1468 io_req_find_next(req, &nxt);
1472 io_queue_async_work(nxt);
1476 * Drop reference to request, return next in chain (if there is one) if this
1477 * was the last reference to this request.
1479 __attribute__((nonnull))
1480 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1482 if (refcount_dec_and_test(&req->refs)) {
1483 io_req_find_next(req, nxtptr);
1488 static void io_put_req(struct io_kiocb *req)
1490 if (refcount_dec_and_test(&req->refs))
1495 * Must only be used if we don't need to care about links, usually from
1496 * within the completion handling itself.
1498 static void __io_double_put_req(struct io_kiocb *req)
1500 /* drop both submit and complete references */
1501 if (refcount_sub_and_test(2, &req->refs))
1505 static void io_double_put_req(struct io_kiocb *req)
1507 /* drop both submit and complete references */
1508 if (refcount_sub_and_test(2, &req->refs))
1512 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1514 struct io_rings *rings = ctx->rings;
1516 if (test_bit(0, &ctx->cq_check_overflow)) {
1518 * noflush == true is from the waitqueue handler, just ensure
1519 * we wake up the task, and the next invocation will flush the
1520 * entries. We cannot safely to it from here.
1522 if (noflush && !list_empty(&ctx->cq_overflow_list))
1525 io_cqring_overflow_flush(ctx, false);
1528 /* See comment at the top of this file */
1530 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1533 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1535 struct io_rings *rings = ctx->rings;
1537 /* make sure SQ entry isn't read before tail */
1538 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1541 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1543 if ((req->flags & REQ_F_LINK) || io_is_fallback_req(req))
1546 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1549 rb->reqs[rb->to_free++] = req;
1550 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1551 io_free_req_many(req->ctx, rb);
1556 * Find and free completed poll iocbs
1558 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1559 struct list_head *done)
1561 struct req_batch rb;
1562 struct io_kiocb *req;
1564 rb.to_free = rb.need_iter = 0;
1565 while (!list_empty(done)) {
1566 req = list_first_entry(done, struct io_kiocb, list);
1567 list_del(&req->list);
1569 io_cqring_fill_event(req, req->result);
1572 if (refcount_dec_and_test(&req->refs) &&
1573 !io_req_multi_free(&rb, req))
1577 io_commit_cqring(ctx);
1578 io_free_req_many(ctx, &rb);
1581 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1584 struct io_kiocb *req, *tmp;
1590 * Only spin for completions if we don't have multiple devices hanging
1591 * off our complete list, and we're under the requested amount.
1593 spin = !ctx->poll_multi_file && *nr_events < min;
1596 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1597 struct kiocb *kiocb = &req->rw.kiocb;
1600 * Move completed entries to our local list. If we find a
1601 * request that requires polling, break out and complete
1602 * the done list first, if we have entries there.
1604 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1605 list_move_tail(&req->list, &done);
1608 if (!list_empty(&done))
1611 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1620 if (!list_empty(&done))
1621 io_iopoll_complete(ctx, nr_events, &done);
1627 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1628 * non-spinning poll check - we'll still enter the driver poll loop, but only
1629 * as a non-spinning completion check.
1631 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1634 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1637 ret = io_do_iopoll(ctx, nr_events, min);
1640 if (!min || *nr_events >= min)
1648 * We can't just wait for polled events to come to us, we have to actively
1649 * find and complete them.
1651 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1653 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1656 mutex_lock(&ctx->uring_lock);
1657 while (!list_empty(&ctx->poll_list)) {
1658 unsigned int nr_events = 0;
1660 io_iopoll_getevents(ctx, &nr_events, 1);
1663 * Ensure we allow local-to-the-cpu processing to take place,
1664 * in this case we need to ensure that we reap all events.
1668 mutex_unlock(&ctx->uring_lock);
1671 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1674 int iters = 0, ret = 0;
1677 * We disallow the app entering submit/complete with polling, but we
1678 * still need to lock the ring to prevent racing with polled issue
1679 * that got punted to a workqueue.
1681 mutex_lock(&ctx->uring_lock);
1686 * Don't enter poll loop if we already have events pending.
1687 * If we do, we can potentially be spinning for commands that
1688 * already triggered a CQE (eg in error).
1690 if (io_cqring_events(ctx, false))
1694 * If a submit got punted to a workqueue, we can have the
1695 * application entering polling for a command before it gets
1696 * issued. That app will hold the uring_lock for the duration
1697 * of the poll right here, so we need to take a breather every
1698 * now and then to ensure that the issue has a chance to add
1699 * the poll to the issued list. Otherwise we can spin here
1700 * forever, while the workqueue is stuck trying to acquire the
1703 if (!(++iters & 7)) {
1704 mutex_unlock(&ctx->uring_lock);
1705 mutex_lock(&ctx->uring_lock);
1708 if (*nr_events < min)
1709 tmin = min - *nr_events;
1711 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1715 } while (min && !*nr_events && !need_resched());
1717 mutex_unlock(&ctx->uring_lock);
1721 static void kiocb_end_write(struct io_kiocb *req)
1724 * Tell lockdep we inherited freeze protection from submission
1727 if (req->flags & REQ_F_ISREG) {
1728 struct inode *inode = file_inode(req->file);
1730 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1732 file_end_write(req->file);
1735 static inline void req_set_fail_links(struct io_kiocb *req)
1737 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1738 req->flags |= REQ_F_FAIL_LINK;
1741 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1743 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1745 if (kiocb->ki_flags & IOCB_WRITE)
1746 kiocb_end_write(req);
1748 if (res != req->result)
1749 req_set_fail_links(req);
1750 io_cqring_add_event(req, res);
1753 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1755 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1757 io_complete_rw_common(kiocb, res);
1761 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1763 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1764 struct io_kiocb *nxt = NULL;
1766 io_complete_rw_common(kiocb, res);
1767 io_put_req_find_next(req, &nxt);
1772 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1774 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1776 if (kiocb->ki_flags & IOCB_WRITE)
1777 kiocb_end_write(req);
1779 if (res != req->result)
1780 req_set_fail_links(req);
1783 req->flags |= REQ_F_IOPOLL_COMPLETED;
1787 * After the iocb has been issued, it's safe to be found on the poll list.
1788 * Adding the kiocb to the list AFTER submission ensures that we don't
1789 * find it from a io_iopoll_getevents() thread before the issuer is done
1790 * accessing the kiocb cookie.
1792 static void io_iopoll_req_issued(struct io_kiocb *req)
1794 struct io_ring_ctx *ctx = req->ctx;
1797 * Track whether we have multiple files in our lists. This will impact
1798 * how we do polling eventually, not spinning if we're on potentially
1799 * different devices.
1801 if (list_empty(&ctx->poll_list)) {
1802 ctx->poll_multi_file = false;
1803 } else if (!ctx->poll_multi_file) {
1804 struct io_kiocb *list_req;
1806 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1808 if (list_req->file != req->file)
1809 ctx->poll_multi_file = true;
1813 * For fast devices, IO may have already completed. If it has, add
1814 * it to the front so we find it first.
1816 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1817 list_add(&req->list, &ctx->poll_list);
1819 list_add_tail(&req->list, &ctx->poll_list);
1821 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1822 wq_has_sleeper(&ctx->sqo_wait))
1823 wake_up(&ctx->sqo_wait);
1826 static void io_file_put(struct io_submit_state *state)
1829 int diff = state->has_refs - state->used_refs;
1832 fput_many(state->file, diff);
1838 * Get as many references to a file as we have IOs left in this submission,
1839 * assuming most submissions are for one file, or at least that each file
1840 * has more than one submission.
1842 static struct file *io_file_get(struct io_submit_state *state, int fd)
1848 if (state->fd == fd) {
1855 state->file = fget_many(fd, state->ios_left);
1860 state->has_refs = state->ios_left;
1861 state->used_refs = 1;
1867 * If we tracked the file through the SCM inflight mechanism, we could support
1868 * any file. For now, just ensure that anything potentially problematic is done
1871 static bool io_file_supports_async(struct file *file)
1873 umode_t mode = file_inode(file)->i_mode;
1875 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
1877 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1883 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1884 bool force_nonblock)
1886 struct io_ring_ctx *ctx = req->ctx;
1887 struct kiocb *kiocb = &req->rw.kiocb;
1891 if (S_ISREG(file_inode(req->file)->i_mode))
1892 req->flags |= REQ_F_ISREG;
1894 kiocb->ki_pos = READ_ONCE(sqe->off);
1895 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
1896 req->flags |= REQ_F_CUR_POS;
1897 kiocb->ki_pos = req->file->f_pos;
1899 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1900 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1901 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1905 ioprio = READ_ONCE(sqe->ioprio);
1907 ret = ioprio_check_cap(ioprio);
1911 kiocb->ki_ioprio = ioprio;
1913 kiocb->ki_ioprio = get_current_ioprio();
1915 /* don't allow async punt if RWF_NOWAIT was requested */
1916 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1917 (req->file->f_flags & O_NONBLOCK))
1918 req->flags |= REQ_F_NOWAIT;
1921 kiocb->ki_flags |= IOCB_NOWAIT;
1923 if (ctx->flags & IORING_SETUP_IOPOLL) {
1924 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1925 !kiocb->ki_filp->f_op->iopoll)
1928 kiocb->ki_flags |= IOCB_HIPRI;
1929 kiocb->ki_complete = io_complete_rw_iopoll;
1932 if (kiocb->ki_flags & IOCB_HIPRI)
1934 kiocb->ki_complete = io_complete_rw;
1937 req->rw.addr = READ_ONCE(sqe->addr);
1938 req->rw.len = READ_ONCE(sqe->len);
1939 /* we own ->private, reuse it for the buffer index */
1940 req->rw.kiocb.private = (void *) (unsigned long)
1941 READ_ONCE(sqe->buf_index);
1945 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1951 case -ERESTARTNOINTR:
1952 case -ERESTARTNOHAND:
1953 case -ERESTART_RESTARTBLOCK:
1955 * We can't just restart the syscall, since previously
1956 * submitted sqes may already be in progress. Just fail this
1962 kiocb->ki_complete(kiocb, ret, 0);
1966 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1969 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1971 if (req->flags & REQ_F_CUR_POS)
1972 req->file->f_pos = kiocb->ki_pos;
1973 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1974 *nxt = __io_complete_rw(kiocb, ret);
1976 io_rw_done(kiocb, ret);
1979 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
1980 struct iov_iter *iter)
1982 struct io_ring_ctx *ctx = req->ctx;
1983 size_t len = req->rw.len;
1984 struct io_mapped_ubuf *imu;
1985 unsigned index, buf_index;
1989 /* attempt to use fixed buffers without having provided iovecs */
1990 if (unlikely(!ctx->user_bufs))
1993 buf_index = (unsigned long) req->rw.kiocb.private;
1994 if (unlikely(buf_index >= ctx->nr_user_bufs))
1997 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1998 imu = &ctx->user_bufs[index];
1999 buf_addr = req->rw.addr;
2002 if (buf_addr + len < buf_addr)
2004 /* not inside the mapped region */
2005 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2009 * May not be a start of buffer, set size appropriately
2010 * and advance us to the beginning.
2012 offset = buf_addr - imu->ubuf;
2013 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2017 * Don't use iov_iter_advance() here, as it's really slow for
2018 * using the latter parts of a big fixed buffer - it iterates
2019 * over each segment manually. We can cheat a bit here, because
2022 * 1) it's a BVEC iter, we set it up
2023 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2024 * first and last bvec
2026 * So just find our index, and adjust the iterator afterwards.
2027 * If the offset is within the first bvec (or the whole first
2028 * bvec, just use iov_iter_advance(). This makes it easier
2029 * since we can just skip the first segment, which may not
2030 * be PAGE_SIZE aligned.
2032 const struct bio_vec *bvec = imu->bvec;
2034 if (offset <= bvec->bv_len) {
2035 iov_iter_advance(iter, offset);
2037 unsigned long seg_skip;
2039 /* skip first vec */
2040 offset -= bvec->bv_len;
2041 seg_skip = 1 + (offset >> PAGE_SHIFT);
2043 iter->bvec = bvec + seg_skip;
2044 iter->nr_segs -= seg_skip;
2045 iter->count -= bvec->bv_len + offset;
2046 iter->iov_offset = offset & ~PAGE_MASK;
2053 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2054 struct iovec **iovec, struct iov_iter *iter)
2056 void __user *buf = u64_to_user_ptr(req->rw.addr);
2057 size_t sqe_len = req->rw.len;
2060 opcode = req->opcode;
2061 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2063 return io_import_fixed(req, rw, iter);
2066 /* buffer index only valid with fixed read/write */
2067 if (req->rw.kiocb.private)
2070 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2072 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2074 return ret < 0 ? ret : sqe_len;
2078 struct io_async_rw *iorw = &req->io->rw;
2081 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2082 if (iorw->iov == iorw->fast_iov)
2087 #ifdef CONFIG_COMPAT
2088 if (req->ctx->compat)
2089 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2093 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2097 * For files that don't have ->read_iter() and ->write_iter(), handle them
2098 * by looping over ->read() or ->write() manually.
2100 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2101 struct iov_iter *iter)
2106 * Don't support polled IO through this interface, and we can't
2107 * support non-blocking either. For the latter, this just causes
2108 * the kiocb to be handled from an async context.
2110 if (kiocb->ki_flags & IOCB_HIPRI)
2112 if (kiocb->ki_flags & IOCB_NOWAIT)
2115 while (iov_iter_count(iter)) {
2119 if (!iov_iter_is_bvec(iter)) {
2120 iovec = iov_iter_iovec(iter);
2122 /* fixed buffers import bvec */
2123 iovec.iov_base = kmap(iter->bvec->bv_page)
2125 iovec.iov_len = min(iter->count,
2126 iter->bvec->bv_len - iter->iov_offset);
2130 nr = file->f_op->read(file, iovec.iov_base,
2131 iovec.iov_len, &kiocb->ki_pos);
2133 nr = file->f_op->write(file, iovec.iov_base,
2134 iovec.iov_len, &kiocb->ki_pos);
2137 if (iov_iter_is_bvec(iter))
2138 kunmap(iter->bvec->bv_page);
2146 if (nr != iovec.iov_len)
2148 iov_iter_advance(iter, nr);
2154 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2155 struct iovec *iovec, struct iovec *fast_iov,
2156 struct iov_iter *iter)
2158 req->io->rw.nr_segs = iter->nr_segs;
2159 req->io->rw.size = io_size;
2160 req->io->rw.iov = iovec;
2161 if (!req->io->rw.iov) {
2162 req->io->rw.iov = req->io->rw.fast_iov;
2163 memcpy(req->io->rw.iov, fast_iov,
2164 sizeof(struct iovec) * iter->nr_segs);
2166 req->flags |= REQ_F_NEED_CLEANUP;
2170 static int io_alloc_async_ctx(struct io_kiocb *req)
2172 if (!io_op_defs[req->opcode].async_ctx)
2174 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2175 return req->io == NULL;
2178 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2179 struct iovec *iovec, struct iovec *fast_iov,
2180 struct iov_iter *iter)
2182 if (!io_op_defs[req->opcode].async_ctx)
2185 if (io_alloc_async_ctx(req))
2188 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2193 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2194 bool force_nonblock)
2196 struct io_async_ctx *io;
2197 struct iov_iter iter;
2200 ret = io_prep_rw(req, sqe, force_nonblock);
2204 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2207 /* either don't need iovec imported or already have it */
2208 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2212 io->rw.iov = io->rw.fast_iov;
2214 ret = io_import_iovec(READ, req, &io->rw.iov, &iter);
2219 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2223 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
2224 bool force_nonblock)
2226 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2227 struct kiocb *kiocb = &req->rw.kiocb;
2228 struct iov_iter iter;
2230 ssize_t io_size, ret;
2232 ret = io_import_iovec(READ, req, &iovec, &iter);
2236 /* Ensure we clear previously set non-block flag */
2237 if (!force_nonblock)
2238 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2242 if (req->flags & REQ_F_LINK)
2243 req->result = io_size;
2246 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2247 * we know to async punt it even if it was opened O_NONBLOCK
2249 if (force_nonblock && !io_file_supports_async(req->file)) {
2250 req->flags |= REQ_F_MUST_PUNT;
2254 iov_count = iov_iter_count(&iter);
2255 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2259 if (req->file->f_op->read_iter)
2260 ret2 = call_read_iter(req->file, kiocb, &iter);
2262 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2264 /* Catch -EAGAIN return for forced non-blocking submission */
2265 if (!force_nonblock || ret2 != -EAGAIN) {
2266 kiocb_done(kiocb, ret2, nxt, req->in_async);
2269 ret = io_setup_async_rw(req, io_size, iovec,
2270 inline_vecs, &iter);
2278 req->flags &= ~REQ_F_NEED_CLEANUP;
2282 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2283 bool force_nonblock)
2285 struct io_async_ctx *io;
2286 struct iov_iter iter;
2289 ret = io_prep_rw(req, sqe, force_nonblock);
2293 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2296 /* either don't need iovec imported or already have it */
2297 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2301 io->rw.iov = io->rw.fast_iov;
2303 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter);
2308 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2312 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
2313 bool force_nonblock)
2315 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2316 struct kiocb *kiocb = &req->rw.kiocb;
2317 struct iov_iter iter;
2319 ssize_t ret, io_size;
2321 ret = io_import_iovec(WRITE, req, &iovec, &iter);
2325 /* Ensure we clear previously set non-block flag */
2326 if (!force_nonblock)
2327 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2331 if (req->flags & REQ_F_LINK)
2332 req->result = io_size;
2335 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2336 * we know to async punt it even if it was opened O_NONBLOCK
2338 if (force_nonblock && !io_file_supports_async(req->file)) {
2339 req->flags |= REQ_F_MUST_PUNT;
2343 /* file path doesn't support NOWAIT for non-direct_IO */
2344 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2345 (req->flags & REQ_F_ISREG))
2348 iov_count = iov_iter_count(&iter);
2349 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2354 * Open-code file_start_write here to grab freeze protection,
2355 * which will be released by another thread in
2356 * io_complete_rw(). Fool lockdep by telling it the lock got
2357 * released so that it doesn't complain about the held lock when
2358 * we return to userspace.
2360 if (req->flags & REQ_F_ISREG) {
2361 __sb_start_write(file_inode(req->file)->i_sb,
2362 SB_FREEZE_WRITE, true);
2363 __sb_writers_release(file_inode(req->file)->i_sb,
2366 kiocb->ki_flags |= IOCB_WRITE;
2368 if (req->file->f_op->write_iter)
2369 ret2 = call_write_iter(req->file, kiocb, &iter);
2371 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2373 * Raw bdev writes will -EOPNOTSUPP for IOCB_NOWAIT. Just
2374 * retry them without IOCB_NOWAIT.
2376 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2378 if (!force_nonblock || ret2 != -EAGAIN) {
2379 kiocb_done(kiocb, ret2, nxt, req->in_async);
2382 ret = io_setup_async_rw(req, io_size, iovec,
2383 inline_vecs, &iter);
2390 req->flags &= ~REQ_F_NEED_CLEANUP;
2396 * IORING_OP_NOP just posts a completion event, nothing else.
2398 static int io_nop(struct io_kiocb *req)
2400 struct io_ring_ctx *ctx = req->ctx;
2402 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2405 io_cqring_add_event(req, 0);
2410 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2412 struct io_ring_ctx *ctx = req->ctx;
2417 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2419 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2422 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2423 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2426 req->sync.off = READ_ONCE(sqe->off);
2427 req->sync.len = READ_ONCE(sqe->len);
2431 static bool io_req_cancelled(struct io_kiocb *req)
2433 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2434 req_set_fail_links(req);
2435 io_cqring_add_event(req, -ECANCELED);
2443 static void io_link_work_cb(struct io_wq_work **workptr)
2445 struct io_wq_work *work = *workptr;
2446 struct io_kiocb *link = work->data;
2448 io_queue_linked_timeout(link);
2449 work->func = io_wq_submit_work;
2452 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
2454 struct io_kiocb *link;
2456 io_prep_async_work(nxt, &link);
2457 *workptr = &nxt->work;
2459 nxt->work.flags |= IO_WQ_WORK_CB;
2460 nxt->work.func = io_link_work_cb;
2461 nxt->work.data = link;
2465 static void io_fsync_finish(struct io_wq_work **workptr)
2467 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2468 loff_t end = req->sync.off + req->sync.len;
2469 struct io_kiocb *nxt = NULL;
2472 if (io_req_cancelled(req))
2475 ret = vfs_fsync_range(req->file, req->sync.off,
2476 end > 0 ? end : LLONG_MAX,
2477 req->sync.flags & IORING_FSYNC_DATASYNC);
2479 req_set_fail_links(req);
2480 io_cqring_add_event(req, ret);
2481 io_put_req_find_next(req, &nxt);
2483 io_wq_assign_next(workptr, nxt);
2486 static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt,
2487 bool force_nonblock)
2489 struct io_wq_work *work, *old_work;
2491 /* fsync always requires a blocking context */
2492 if (force_nonblock) {
2494 req->work.func = io_fsync_finish;
2498 work = old_work = &req->work;
2499 io_fsync_finish(&work);
2500 if (work && work != old_work)
2501 *nxt = container_of(work, struct io_kiocb, work);
2505 static void io_fallocate_finish(struct io_wq_work **workptr)
2507 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2508 struct io_kiocb *nxt = NULL;
2511 if (io_req_cancelled(req))
2514 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2517 req_set_fail_links(req);
2518 io_cqring_add_event(req, ret);
2519 io_put_req_find_next(req, &nxt);
2521 io_wq_assign_next(workptr, nxt);
2524 static int io_fallocate_prep(struct io_kiocb *req,
2525 const struct io_uring_sqe *sqe)
2527 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2530 req->sync.off = READ_ONCE(sqe->off);
2531 req->sync.len = READ_ONCE(sqe->addr);
2532 req->sync.mode = READ_ONCE(sqe->len);
2536 static int io_fallocate(struct io_kiocb *req, struct io_kiocb **nxt,
2537 bool force_nonblock)
2539 struct io_wq_work *work, *old_work;
2541 /* fallocate always requiring blocking context */
2542 if (force_nonblock) {
2544 req->work.func = io_fallocate_finish;
2548 work = old_work = &req->work;
2549 io_fallocate_finish(&work);
2550 if (work && work != old_work)
2551 *nxt = container_of(work, struct io_kiocb, work);
2556 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2558 const char __user *fname;
2561 if (sqe->ioprio || sqe->buf_index)
2563 if (sqe->flags & IOSQE_FIXED_FILE)
2565 if (req->flags & REQ_F_NEED_CLEANUP)
2568 req->open.dfd = READ_ONCE(sqe->fd);
2569 req->open.how.mode = READ_ONCE(sqe->len);
2570 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2571 req->open.how.flags = READ_ONCE(sqe->open_flags);
2573 req->open.filename = getname(fname);
2574 if (IS_ERR(req->open.filename)) {
2575 ret = PTR_ERR(req->open.filename);
2576 req->open.filename = NULL;
2580 req->flags |= REQ_F_NEED_CLEANUP;
2584 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2586 struct open_how __user *how;
2587 const char __user *fname;
2591 if (sqe->ioprio || sqe->buf_index)
2593 if (sqe->flags & IOSQE_FIXED_FILE)
2595 if (req->flags & REQ_F_NEED_CLEANUP)
2598 req->open.dfd = READ_ONCE(sqe->fd);
2599 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2600 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2601 len = READ_ONCE(sqe->len);
2603 if (len < OPEN_HOW_SIZE_VER0)
2606 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2611 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2612 req->open.how.flags |= O_LARGEFILE;
2614 req->open.filename = getname(fname);
2615 if (IS_ERR(req->open.filename)) {
2616 ret = PTR_ERR(req->open.filename);
2617 req->open.filename = NULL;
2621 req->flags |= REQ_F_NEED_CLEANUP;
2625 static int io_openat2(struct io_kiocb *req, struct io_kiocb **nxt,
2626 bool force_nonblock)
2628 struct open_flags op;
2635 ret = build_open_flags(&req->open.how, &op);
2639 ret = get_unused_fd_flags(req->open.how.flags);
2643 file = do_filp_open(req->open.dfd, req->open.filename, &op);
2646 ret = PTR_ERR(file);
2648 fsnotify_open(file);
2649 fd_install(ret, file);
2652 putname(req->open.filename);
2653 req->flags &= ~REQ_F_NEED_CLEANUP;
2655 req_set_fail_links(req);
2656 io_cqring_add_event(req, ret);
2657 io_put_req_find_next(req, nxt);
2661 static int io_openat(struct io_kiocb *req, struct io_kiocb **nxt,
2662 bool force_nonblock)
2664 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
2665 return io_openat2(req, nxt, force_nonblock);
2668 static int io_epoll_ctl_prep(struct io_kiocb *req,
2669 const struct io_uring_sqe *sqe)
2671 #if defined(CONFIG_EPOLL)
2672 if (sqe->ioprio || sqe->buf_index)
2675 req->epoll.epfd = READ_ONCE(sqe->fd);
2676 req->epoll.op = READ_ONCE(sqe->len);
2677 req->epoll.fd = READ_ONCE(sqe->off);
2679 if (ep_op_has_event(req->epoll.op)) {
2680 struct epoll_event __user *ev;
2682 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
2683 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
2693 static int io_epoll_ctl(struct io_kiocb *req, struct io_kiocb **nxt,
2694 bool force_nonblock)
2696 #if defined(CONFIG_EPOLL)
2697 struct io_epoll *ie = &req->epoll;
2700 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
2701 if (force_nonblock && ret == -EAGAIN)
2705 req_set_fail_links(req);
2706 io_cqring_add_event(req, ret);
2707 io_put_req_find_next(req, nxt);
2714 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2716 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2717 if (sqe->ioprio || sqe->buf_index || sqe->off)
2720 req->madvise.addr = READ_ONCE(sqe->addr);
2721 req->madvise.len = READ_ONCE(sqe->len);
2722 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
2729 static int io_madvise(struct io_kiocb *req, struct io_kiocb **nxt,
2730 bool force_nonblock)
2732 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2733 struct io_madvise *ma = &req->madvise;
2739 ret = do_madvise(ma->addr, ma->len, ma->advice);
2741 req_set_fail_links(req);
2742 io_cqring_add_event(req, ret);
2743 io_put_req_find_next(req, nxt);
2750 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2752 if (sqe->ioprio || sqe->buf_index || sqe->addr)
2755 req->fadvise.offset = READ_ONCE(sqe->off);
2756 req->fadvise.len = READ_ONCE(sqe->len);
2757 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
2761 static int io_fadvise(struct io_kiocb *req, struct io_kiocb **nxt,
2762 bool force_nonblock)
2764 struct io_fadvise *fa = &req->fadvise;
2767 if (force_nonblock) {
2768 switch (fa->advice) {
2769 case POSIX_FADV_NORMAL:
2770 case POSIX_FADV_RANDOM:
2771 case POSIX_FADV_SEQUENTIAL:
2778 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
2780 req_set_fail_links(req);
2781 io_cqring_add_event(req, ret);
2782 io_put_req_find_next(req, nxt);
2786 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2788 const char __user *fname;
2789 unsigned lookup_flags;
2792 if (sqe->ioprio || sqe->buf_index)
2794 if (sqe->flags & IOSQE_FIXED_FILE)
2796 if (req->flags & REQ_F_NEED_CLEANUP)
2799 req->open.dfd = READ_ONCE(sqe->fd);
2800 req->open.mask = READ_ONCE(sqe->len);
2801 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2802 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2803 req->open.how.flags = READ_ONCE(sqe->statx_flags);
2805 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
2808 req->open.filename = getname_flags(fname, lookup_flags, NULL);
2809 if (IS_ERR(req->open.filename)) {
2810 ret = PTR_ERR(req->open.filename);
2811 req->open.filename = NULL;
2815 req->flags |= REQ_F_NEED_CLEANUP;
2819 static int io_statx(struct io_kiocb *req, struct io_kiocb **nxt,
2820 bool force_nonblock)
2822 struct io_open *ctx = &req->open;
2823 unsigned lookup_flags;
2831 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
2835 /* filename_lookup() drops it, keep a reference */
2836 ctx->filename->refcnt++;
2838 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
2843 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
2845 if (retry_estale(ret, lookup_flags)) {
2846 lookup_flags |= LOOKUP_REVAL;
2850 ret = cp_statx(&stat, ctx->buffer);
2852 putname(ctx->filename);
2853 req->flags &= ~REQ_F_NEED_CLEANUP;
2855 req_set_fail_links(req);
2856 io_cqring_add_event(req, ret);
2857 io_put_req_find_next(req, nxt);
2861 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2864 * If we queue this for async, it must not be cancellable. That would
2865 * leave the 'file' in an undeterminate state.
2867 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
2869 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
2870 sqe->rw_flags || sqe->buf_index)
2872 if (sqe->flags & IOSQE_FIXED_FILE)
2875 req->close.fd = READ_ONCE(sqe->fd);
2876 if (req->file->f_op == &io_uring_fops ||
2877 req->close.fd == req->ctx->ring_fd)
2883 /* only called when __close_fd_get_file() is done */
2884 static void __io_close_finish(struct io_kiocb *req, struct io_kiocb **nxt)
2888 ret = filp_close(req->close.put_file, req->work.files);
2890 req_set_fail_links(req);
2891 io_cqring_add_event(req, ret);
2892 fput(req->close.put_file);
2893 io_put_req_find_next(req, nxt);
2896 static void io_close_finish(struct io_wq_work **workptr)
2898 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2899 struct io_kiocb *nxt = NULL;
2901 /* not cancellable, don't do io_req_cancelled() */
2902 __io_close_finish(req, &nxt);
2904 io_wq_assign_next(workptr, nxt);
2907 static int io_close(struct io_kiocb *req, struct io_kiocb **nxt,
2908 bool force_nonblock)
2912 req->close.put_file = NULL;
2913 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
2917 /* if the file has a flush method, be safe and punt to async */
2918 if (req->close.put_file->f_op->flush && !io_wq_current_is_worker())
2922 * No ->flush(), safely close from here and just punt the
2923 * fput() to async context.
2925 __io_close_finish(req, nxt);
2928 req->work.func = io_close_finish;
2930 * Do manual async queue here to avoid grabbing files - we don't
2931 * need the files, and it'll cause io_close_finish() to close
2932 * the file again and cause a double CQE entry for this request
2934 io_queue_async_work(req);
2938 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2940 struct io_ring_ctx *ctx = req->ctx;
2945 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2947 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2950 req->sync.off = READ_ONCE(sqe->off);
2951 req->sync.len = READ_ONCE(sqe->len);
2952 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
2956 static void io_sync_file_range_finish(struct io_wq_work **workptr)
2958 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2959 struct io_kiocb *nxt = NULL;
2962 if (io_req_cancelled(req))
2965 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
2968 req_set_fail_links(req);
2969 io_cqring_add_event(req, ret);
2970 io_put_req_find_next(req, &nxt);
2972 io_wq_assign_next(workptr, nxt);
2975 static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt,
2976 bool force_nonblock)
2978 struct io_wq_work *work, *old_work;
2980 /* sync_file_range always requires a blocking context */
2981 if (force_nonblock) {
2983 req->work.func = io_sync_file_range_finish;
2987 work = old_work = &req->work;
2988 io_sync_file_range_finish(&work);
2989 if (work && work != old_work)
2990 *nxt = container_of(work, struct io_kiocb, work);
2994 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2996 #if defined(CONFIG_NET)
2997 struct io_sr_msg *sr = &req->sr_msg;
2998 struct io_async_ctx *io = req->io;
3001 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3002 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3003 sr->len = READ_ONCE(sqe->len);
3005 #ifdef CONFIG_COMPAT
3006 if (req->ctx->compat)
3007 sr->msg_flags |= MSG_CMSG_COMPAT;
3010 if (!io || req->opcode == IORING_OP_SEND)
3012 /* iovec is already imported */
3013 if (req->flags & REQ_F_NEED_CLEANUP)
3016 io->msg.iov = io->msg.fast_iov;
3017 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3020 req->flags |= REQ_F_NEED_CLEANUP;
3027 static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt,
3028 bool force_nonblock)
3030 #if defined(CONFIG_NET)
3031 struct io_async_msghdr *kmsg = NULL;
3032 struct socket *sock;
3035 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3038 sock = sock_from_file(req->file, &ret);
3040 struct io_async_ctx io;
3044 kmsg = &req->io->msg;
3045 kmsg->msg.msg_name = &req->io->msg.addr;
3046 /* if iov is set, it's allocated already */
3048 kmsg->iov = kmsg->fast_iov;
3049 kmsg->msg.msg_iter.iov = kmsg->iov;
3051 struct io_sr_msg *sr = &req->sr_msg;
3054 kmsg->msg.msg_name = &io.msg.addr;
3056 io.msg.iov = io.msg.fast_iov;
3057 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3058 sr->msg_flags, &io.msg.iov);
3063 flags = req->sr_msg.msg_flags;
3064 if (flags & MSG_DONTWAIT)
3065 req->flags |= REQ_F_NOWAIT;
3066 else if (force_nonblock)
3067 flags |= MSG_DONTWAIT;
3069 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3070 if (force_nonblock && ret == -EAGAIN) {
3073 if (io_alloc_async_ctx(req)) {
3074 if (kmsg->iov != kmsg->fast_iov)
3078 req->flags |= REQ_F_NEED_CLEANUP;
3079 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3082 if (ret == -ERESTARTSYS)
3086 if (kmsg && kmsg->iov != kmsg->fast_iov)
3088 req->flags &= ~REQ_F_NEED_CLEANUP;
3089 io_cqring_add_event(req, ret);
3091 req_set_fail_links(req);
3092 io_put_req_find_next(req, nxt);
3099 static int io_send(struct io_kiocb *req, struct io_kiocb **nxt,
3100 bool force_nonblock)
3102 #if defined(CONFIG_NET)
3103 struct socket *sock;
3106 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3109 sock = sock_from_file(req->file, &ret);
3111 struct io_sr_msg *sr = &req->sr_msg;
3116 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3121 msg.msg_name = NULL;
3122 msg.msg_control = NULL;
3123 msg.msg_controllen = 0;
3124 msg.msg_namelen = 0;
3126 flags = req->sr_msg.msg_flags;
3127 if (flags & MSG_DONTWAIT)
3128 req->flags |= REQ_F_NOWAIT;
3129 else if (force_nonblock)
3130 flags |= MSG_DONTWAIT;
3132 msg.msg_flags = flags;
3133 ret = sock_sendmsg(sock, &msg);
3134 if (force_nonblock && ret == -EAGAIN)
3136 if (ret == -ERESTARTSYS)
3140 io_cqring_add_event(req, ret);
3142 req_set_fail_links(req);
3143 io_put_req_find_next(req, nxt);
3150 static int io_recvmsg_prep(struct io_kiocb *req,
3151 const struct io_uring_sqe *sqe)
3153 #if defined(CONFIG_NET)
3154 struct io_sr_msg *sr = &req->sr_msg;
3155 struct io_async_ctx *io = req->io;
3158 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3159 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3160 sr->len = READ_ONCE(sqe->len);
3162 #ifdef CONFIG_COMPAT
3163 if (req->ctx->compat)
3164 sr->msg_flags |= MSG_CMSG_COMPAT;
3167 if (!io || req->opcode == IORING_OP_RECV)
3169 /* iovec is already imported */
3170 if (req->flags & REQ_F_NEED_CLEANUP)
3173 io->msg.iov = io->msg.fast_iov;
3174 ret = recvmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3175 &io->msg.uaddr, &io->msg.iov);
3177 req->flags |= REQ_F_NEED_CLEANUP;
3184 static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt,
3185 bool force_nonblock)
3187 #if defined(CONFIG_NET)
3188 struct io_async_msghdr *kmsg = NULL;
3189 struct socket *sock;
3192 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3195 sock = sock_from_file(req->file, &ret);
3197 struct io_async_ctx io;
3201 kmsg = &req->io->msg;
3202 kmsg->msg.msg_name = &req->io->msg.addr;
3203 /* if iov is set, it's allocated already */
3205 kmsg->iov = kmsg->fast_iov;
3206 kmsg->msg.msg_iter.iov = kmsg->iov;
3208 struct io_sr_msg *sr = &req->sr_msg;
3211 kmsg->msg.msg_name = &io.msg.addr;
3213 io.msg.iov = io.msg.fast_iov;
3214 ret = recvmsg_copy_msghdr(&io.msg.msg, sr->msg,
3215 sr->msg_flags, &io.msg.uaddr,
3221 flags = req->sr_msg.msg_flags;
3222 if (flags & MSG_DONTWAIT)
3223 req->flags |= REQ_F_NOWAIT;
3224 else if (force_nonblock)
3225 flags |= MSG_DONTWAIT;
3227 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3228 kmsg->uaddr, flags);
3229 if (force_nonblock && ret == -EAGAIN) {
3232 if (io_alloc_async_ctx(req)) {
3233 if (kmsg->iov != kmsg->fast_iov)
3237 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3238 req->flags |= REQ_F_NEED_CLEANUP;
3241 if (ret == -ERESTARTSYS)
3245 if (kmsg && kmsg->iov != kmsg->fast_iov)
3247 req->flags &= ~REQ_F_NEED_CLEANUP;
3248 io_cqring_add_event(req, ret);
3250 req_set_fail_links(req);
3251 io_put_req_find_next(req, nxt);
3258 static int io_recv(struct io_kiocb *req, struct io_kiocb **nxt,
3259 bool force_nonblock)
3261 #if defined(CONFIG_NET)
3262 struct socket *sock;
3265 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3268 sock = sock_from_file(req->file, &ret);
3270 struct io_sr_msg *sr = &req->sr_msg;
3275 ret = import_single_range(READ, sr->buf, sr->len, &iov,
3280 msg.msg_name = NULL;
3281 msg.msg_control = NULL;
3282 msg.msg_controllen = 0;
3283 msg.msg_namelen = 0;
3284 msg.msg_iocb = NULL;
3287 flags = req->sr_msg.msg_flags;
3288 if (flags & MSG_DONTWAIT)
3289 req->flags |= REQ_F_NOWAIT;
3290 else if (force_nonblock)
3291 flags |= MSG_DONTWAIT;
3293 ret = sock_recvmsg(sock, &msg, flags);
3294 if (force_nonblock && ret == -EAGAIN)
3296 if (ret == -ERESTARTSYS)
3300 io_cqring_add_event(req, ret);
3302 req_set_fail_links(req);
3303 io_put_req_find_next(req, nxt);
3311 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3313 #if defined(CONFIG_NET)
3314 struct io_accept *accept = &req->accept;
3316 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3318 if (sqe->ioprio || sqe->len || sqe->buf_index)
3321 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3322 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3323 accept->flags = READ_ONCE(sqe->accept_flags);
3330 #if defined(CONFIG_NET)
3331 static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3332 bool force_nonblock)
3334 struct io_accept *accept = &req->accept;
3335 unsigned file_flags;
3338 file_flags = force_nonblock ? O_NONBLOCK : 0;
3339 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3340 accept->addr_len, accept->flags);
3341 if (ret == -EAGAIN && force_nonblock)
3343 if (ret == -ERESTARTSYS)
3346 req_set_fail_links(req);
3347 io_cqring_add_event(req, ret);
3348 io_put_req_find_next(req, nxt);
3352 static void io_accept_finish(struct io_wq_work **workptr)
3354 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3355 struct io_kiocb *nxt = NULL;
3357 if (io_req_cancelled(req))
3359 __io_accept(req, &nxt, false);
3361 io_wq_assign_next(workptr, nxt);
3365 static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3366 bool force_nonblock)
3368 #if defined(CONFIG_NET)
3371 ret = __io_accept(req, nxt, force_nonblock);
3372 if (ret == -EAGAIN && force_nonblock) {
3373 req->work.func = io_accept_finish;
3383 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3385 #if defined(CONFIG_NET)
3386 struct io_connect *conn = &req->connect;
3387 struct io_async_ctx *io = req->io;
3389 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3391 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3394 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3395 conn->addr_len = READ_ONCE(sqe->addr2);
3400 return move_addr_to_kernel(conn->addr, conn->addr_len,
3401 &io->connect.address);
3407 static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt,
3408 bool force_nonblock)
3410 #if defined(CONFIG_NET)
3411 struct io_async_ctx __io, *io;
3412 unsigned file_flags;
3418 ret = move_addr_to_kernel(req->connect.addr,
3419 req->connect.addr_len,
3420 &__io.connect.address);
3426 file_flags = force_nonblock ? O_NONBLOCK : 0;
3428 ret = __sys_connect_file(req->file, &io->connect.address,
3429 req->connect.addr_len, file_flags);
3430 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
3433 if (io_alloc_async_ctx(req)) {
3437 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
3440 if (ret == -ERESTARTSYS)
3444 req_set_fail_links(req);
3445 io_cqring_add_event(req, ret);
3446 io_put_req_find_next(req, nxt);
3453 static void io_poll_remove_one(struct io_kiocb *req)
3455 struct io_poll_iocb *poll = &req->poll;
3457 spin_lock(&poll->head->lock);
3458 WRITE_ONCE(poll->canceled, true);
3459 if (!list_empty(&poll->wait.entry)) {
3460 list_del_init(&poll->wait.entry);
3461 io_queue_async_work(req);
3463 spin_unlock(&poll->head->lock);
3464 hash_del(&req->hash_node);
3467 static void io_poll_remove_all(struct io_ring_ctx *ctx)
3469 struct hlist_node *tmp;
3470 struct io_kiocb *req;
3473 spin_lock_irq(&ctx->completion_lock);
3474 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
3475 struct hlist_head *list;
3477 list = &ctx->cancel_hash[i];
3478 hlist_for_each_entry_safe(req, tmp, list, hash_node)
3479 io_poll_remove_one(req);
3481 spin_unlock_irq(&ctx->completion_lock);
3484 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
3486 struct hlist_head *list;
3487 struct io_kiocb *req;
3489 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
3490 hlist_for_each_entry(req, list, hash_node) {
3491 if (sqe_addr == req->user_data) {
3492 io_poll_remove_one(req);
3500 static int io_poll_remove_prep(struct io_kiocb *req,
3501 const struct io_uring_sqe *sqe)
3503 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3505 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3509 req->poll.addr = READ_ONCE(sqe->addr);
3514 * Find a running poll command that matches one specified in sqe->addr,
3515 * and remove it if found.
3517 static int io_poll_remove(struct io_kiocb *req)
3519 struct io_ring_ctx *ctx = req->ctx;
3523 addr = req->poll.addr;
3524 spin_lock_irq(&ctx->completion_lock);
3525 ret = io_poll_cancel(ctx, addr);
3526 spin_unlock_irq(&ctx->completion_lock);
3528 io_cqring_add_event(req, ret);
3530 req_set_fail_links(req);
3535 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
3537 struct io_ring_ctx *ctx = req->ctx;
3539 req->poll.done = true;
3541 io_cqring_fill_event(req, error);
3543 io_cqring_fill_event(req, mangle_poll(mask));
3544 io_commit_cqring(ctx);
3547 static void io_poll_complete_work(struct io_wq_work **workptr)
3549 struct io_wq_work *work = *workptr;
3550 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3551 struct io_poll_iocb *poll = &req->poll;
3552 struct poll_table_struct pt = { ._key = poll->events };
3553 struct io_ring_ctx *ctx = req->ctx;
3554 struct io_kiocb *nxt = NULL;
3558 if (work->flags & IO_WQ_WORK_CANCEL) {
3559 WRITE_ONCE(poll->canceled, true);
3561 } else if (READ_ONCE(poll->canceled)) {
3565 if (ret != -ECANCELED)
3566 mask = vfs_poll(poll->file, &pt) & poll->events;
3569 * Note that ->ki_cancel callers also delete iocb from active_reqs after
3570 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
3571 * synchronize with them. In the cancellation case the list_del_init
3572 * itself is not actually needed, but harmless so we keep it in to
3573 * avoid further branches in the fast path.
3575 spin_lock_irq(&ctx->completion_lock);
3576 if (!mask && ret != -ECANCELED) {
3577 add_wait_queue(poll->head, &poll->wait);
3578 spin_unlock_irq(&ctx->completion_lock);
3581 hash_del(&req->hash_node);
3582 io_poll_complete(req, mask, ret);
3583 spin_unlock_irq(&ctx->completion_lock);
3585 io_cqring_ev_posted(ctx);
3588 req_set_fail_links(req);
3589 io_put_req_find_next(req, &nxt);
3591 io_wq_assign_next(workptr, nxt);
3594 static void __io_poll_flush(struct io_ring_ctx *ctx, struct llist_node *nodes)
3596 struct io_kiocb *req, *tmp;
3597 struct req_batch rb;
3599 rb.to_free = rb.need_iter = 0;
3600 spin_lock_irq(&ctx->completion_lock);
3601 llist_for_each_entry_safe(req, tmp, nodes, llist_node) {
3602 hash_del(&req->hash_node);
3603 io_poll_complete(req, req->result, 0);
3605 if (refcount_dec_and_test(&req->refs) &&
3606 !io_req_multi_free(&rb, req)) {
3607 req->flags |= REQ_F_COMP_LOCKED;
3611 spin_unlock_irq(&ctx->completion_lock);
3613 io_cqring_ev_posted(ctx);
3614 io_free_req_many(ctx, &rb);
3617 static void io_poll_flush(struct io_wq_work **workptr)
3619 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3620 struct llist_node *nodes;
3622 nodes = llist_del_all(&req->ctx->poll_llist);
3624 __io_poll_flush(req->ctx, nodes);
3627 static void io_poll_trigger_evfd(struct io_wq_work **workptr)
3629 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3631 eventfd_signal(req->ctx->cq_ev_fd, 1);
3635 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
3638 struct io_poll_iocb *poll = wait->private;
3639 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
3640 struct io_ring_ctx *ctx = req->ctx;
3641 __poll_t mask = key_to_poll(key);
3643 /* for instances that support it check for an event match first: */
3644 if (mask && !(mask & poll->events))
3647 list_del_init(&poll->wait.entry);
3650 * Run completion inline if we can. We're using trylock here because
3651 * we are violating the completion_lock -> poll wq lock ordering.
3652 * If we have a link timeout we're going to need the completion_lock
3653 * for finalizing the request, mark us as having grabbed that already.
3656 unsigned long flags;
3658 if (llist_empty(&ctx->poll_llist) &&
3659 spin_trylock_irqsave(&ctx->completion_lock, flags)) {
3662 hash_del(&req->hash_node);
3663 io_poll_complete(req, mask, 0);
3665 trigger_ev = io_should_trigger_evfd(ctx);
3666 if (trigger_ev && eventfd_signal_count()) {
3668 req->work.func = io_poll_trigger_evfd;
3670 req->flags |= REQ_F_COMP_LOCKED;
3674 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3675 __io_cqring_ev_posted(ctx, trigger_ev);
3678 req->llist_node.next = NULL;
3679 /* if the list wasn't empty, we're done */
3680 if (!llist_add(&req->llist_node, &ctx->poll_llist))
3683 req->work.func = io_poll_flush;
3687 io_queue_async_work(req);
3692 struct io_poll_table {
3693 struct poll_table_struct pt;
3694 struct io_kiocb *req;
3698 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
3699 struct poll_table_struct *p)
3701 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
3703 if (unlikely(pt->req->poll.head)) {
3704 pt->error = -EINVAL;
3709 pt->req->poll.head = head;
3710 add_wait_queue(head, &pt->req->poll.wait);
3713 static void io_poll_req_insert(struct io_kiocb *req)
3715 struct io_ring_ctx *ctx = req->ctx;
3716 struct hlist_head *list;
3718 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
3719 hlist_add_head(&req->hash_node, list);
3722 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3724 struct io_poll_iocb *poll = &req->poll;
3727 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3729 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3734 events = READ_ONCE(sqe->poll_events);
3735 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
3739 static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt)
3741 struct io_poll_iocb *poll = &req->poll;
3742 struct io_ring_ctx *ctx = req->ctx;
3743 struct io_poll_table ipt;
3744 bool cancel = false;
3747 INIT_IO_WORK(&req->work, io_poll_complete_work);
3748 INIT_HLIST_NODE(&req->hash_node);
3752 poll->canceled = false;
3754 ipt.pt._qproc = io_poll_queue_proc;
3755 ipt.pt._key = poll->events;
3757 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
3759 /* initialized the list so that we can do list_empty checks */
3760 INIT_LIST_HEAD(&poll->wait.entry);
3761 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
3762 poll->wait.private = poll;
3764 INIT_LIST_HEAD(&req->list);
3766 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
3768 spin_lock_irq(&ctx->completion_lock);
3769 if (likely(poll->head)) {
3770 spin_lock(&poll->head->lock);
3771 if (unlikely(list_empty(&poll->wait.entry))) {
3777 if (mask || ipt.error)
3778 list_del_init(&poll->wait.entry);
3780 WRITE_ONCE(poll->canceled, true);
3781 else if (!poll->done) /* actually waiting for an event */
3782 io_poll_req_insert(req);
3783 spin_unlock(&poll->head->lock);
3785 if (mask) { /* no async, we'd stolen it */
3787 io_poll_complete(req, mask, 0);
3789 spin_unlock_irq(&ctx->completion_lock);
3792 io_cqring_ev_posted(ctx);
3793 io_put_req_find_next(req, nxt);
3798 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
3800 struct io_timeout_data *data = container_of(timer,
3801 struct io_timeout_data, timer);
3802 struct io_kiocb *req = data->req;
3803 struct io_ring_ctx *ctx = req->ctx;
3804 unsigned long flags;
3806 atomic_inc(&ctx->cq_timeouts);
3808 spin_lock_irqsave(&ctx->completion_lock, flags);
3810 * We could be racing with timeout deletion. If the list is empty,
3811 * then timeout lookup already found it and will be handling it.
3813 if (!list_empty(&req->list)) {
3814 struct io_kiocb *prev;
3817 * Adjust the reqs sequence before the current one because it
3818 * will consume a slot in the cq_ring and the cq_tail
3819 * pointer will be increased, otherwise other timeout reqs may
3820 * return in advance without waiting for enough wait_nr.
3823 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
3825 list_del_init(&req->list);
3828 io_cqring_fill_event(req, -ETIME);
3829 io_commit_cqring(ctx);
3830 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3832 io_cqring_ev_posted(ctx);
3833 req_set_fail_links(req);
3835 return HRTIMER_NORESTART;
3838 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
3840 struct io_kiocb *req;
3843 list_for_each_entry(req, &ctx->timeout_list, list) {
3844 if (user_data == req->user_data) {
3845 list_del_init(&req->list);
3854 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
3858 req_set_fail_links(req);
3859 io_cqring_fill_event(req, -ECANCELED);
3864 static int io_timeout_remove_prep(struct io_kiocb *req,
3865 const struct io_uring_sqe *sqe)
3867 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3869 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
3872 req->timeout.addr = READ_ONCE(sqe->addr);
3873 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
3874 if (req->timeout.flags)
3881 * Remove or update an existing timeout command
3883 static int io_timeout_remove(struct io_kiocb *req)
3885 struct io_ring_ctx *ctx = req->ctx;
3888 spin_lock_irq(&ctx->completion_lock);
3889 ret = io_timeout_cancel(ctx, req->timeout.addr);
3891 io_cqring_fill_event(req, ret);
3892 io_commit_cqring(ctx);
3893 spin_unlock_irq(&ctx->completion_lock);
3894 io_cqring_ev_posted(ctx);
3896 req_set_fail_links(req);
3901 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3902 bool is_timeout_link)
3904 struct io_timeout_data *data;
3907 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3909 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
3911 if (sqe->off && is_timeout_link)
3913 flags = READ_ONCE(sqe->timeout_flags);
3914 if (flags & ~IORING_TIMEOUT_ABS)
3917 req->timeout.count = READ_ONCE(sqe->off);
3919 if (!req->io && io_alloc_async_ctx(req))
3922 data = &req->io->timeout;
3924 req->flags |= REQ_F_TIMEOUT;
3926 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
3929 if (flags & IORING_TIMEOUT_ABS)
3930 data->mode = HRTIMER_MODE_ABS;
3932 data->mode = HRTIMER_MODE_REL;
3934 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
3938 static int io_timeout(struct io_kiocb *req)
3941 struct io_ring_ctx *ctx = req->ctx;
3942 struct io_timeout_data *data;
3943 struct list_head *entry;
3946 data = &req->io->timeout;
3949 * sqe->off holds how many events that need to occur for this
3950 * timeout event to be satisfied. If it isn't set, then this is
3951 * a pure timeout request, sequence isn't used.
3953 count = req->timeout.count;
3955 req->flags |= REQ_F_TIMEOUT_NOSEQ;
3956 spin_lock_irq(&ctx->completion_lock);
3957 entry = ctx->timeout_list.prev;
3961 req->sequence = ctx->cached_sq_head + count - 1;
3962 data->seq_offset = count;
3965 * Insertion sort, ensuring the first entry in the list is always
3966 * the one we need first.
3968 spin_lock_irq(&ctx->completion_lock);
3969 list_for_each_prev(entry, &ctx->timeout_list) {
3970 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
3971 unsigned nxt_sq_head;
3972 long long tmp, tmp_nxt;
3973 u32 nxt_offset = nxt->io->timeout.seq_offset;
3975 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
3979 * Since cached_sq_head + count - 1 can overflow, use type long
3982 tmp = (long long)ctx->cached_sq_head + count - 1;
3983 nxt_sq_head = nxt->sequence - nxt_offset + 1;
3984 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
3987 * cached_sq_head may overflow, and it will never overflow twice
3988 * once there is some timeout req still be valid.
3990 if (ctx->cached_sq_head < nxt_sq_head)
3997 * Sequence of reqs after the insert one and itself should
3998 * be adjusted because each timeout req consumes a slot.
4003 req->sequence -= span;
4005 list_add(&req->list, entry);
4006 data->timer.function = io_timeout_fn;
4007 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4008 spin_unlock_irq(&ctx->completion_lock);
4012 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4014 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4016 return req->user_data == (unsigned long) data;
4019 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4021 enum io_wq_cancel cancel_ret;
4024 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4025 switch (cancel_ret) {
4026 case IO_WQ_CANCEL_OK:
4029 case IO_WQ_CANCEL_RUNNING:
4032 case IO_WQ_CANCEL_NOTFOUND:
4040 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4041 struct io_kiocb *req, __u64 sqe_addr,
4042 struct io_kiocb **nxt, int success_ret)
4044 unsigned long flags;
4047 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4048 if (ret != -ENOENT) {
4049 spin_lock_irqsave(&ctx->completion_lock, flags);
4053 spin_lock_irqsave(&ctx->completion_lock, flags);
4054 ret = io_timeout_cancel(ctx, sqe_addr);
4057 ret = io_poll_cancel(ctx, sqe_addr);
4061 io_cqring_fill_event(req, ret);
4062 io_commit_cqring(ctx);
4063 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4064 io_cqring_ev_posted(ctx);
4067 req_set_fail_links(req);
4068 io_put_req_find_next(req, nxt);
4071 static int io_async_cancel_prep(struct io_kiocb *req,
4072 const struct io_uring_sqe *sqe)
4074 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4076 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4080 req->cancel.addr = READ_ONCE(sqe->addr);
4084 static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt)
4086 struct io_ring_ctx *ctx = req->ctx;
4088 io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0);
4092 static int io_files_update_prep(struct io_kiocb *req,
4093 const struct io_uring_sqe *sqe)
4095 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4098 req->files_update.offset = READ_ONCE(sqe->off);
4099 req->files_update.nr_args = READ_ONCE(sqe->len);
4100 if (!req->files_update.nr_args)
4102 req->files_update.arg = READ_ONCE(sqe->addr);
4106 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4108 struct io_ring_ctx *ctx = req->ctx;
4109 struct io_uring_files_update up;
4115 up.offset = req->files_update.offset;
4116 up.fds = req->files_update.arg;
4118 mutex_lock(&ctx->uring_lock);
4119 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4120 mutex_unlock(&ctx->uring_lock);
4123 req_set_fail_links(req);
4124 io_cqring_add_event(req, ret);
4129 static int io_req_defer_prep(struct io_kiocb *req,
4130 const struct io_uring_sqe *sqe)
4134 if (io_op_defs[req->opcode].file_table) {
4135 ret = io_grab_files(req);
4140 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4142 switch (req->opcode) {
4145 case IORING_OP_READV:
4146 case IORING_OP_READ_FIXED:
4147 case IORING_OP_READ:
4148 ret = io_read_prep(req, sqe, true);
4150 case IORING_OP_WRITEV:
4151 case IORING_OP_WRITE_FIXED:
4152 case IORING_OP_WRITE:
4153 ret = io_write_prep(req, sqe, true);
4155 case IORING_OP_POLL_ADD:
4156 ret = io_poll_add_prep(req, sqe);
4158 case IORING_OP_POLL_REMOVE:
4159 ret = io_poll_remove_prep(req, sqe);
4161 case IORING_OP_FSYNC:
4162 ret = io_prep_fsync(req, sqe);
4164 case IORING_OP_SYNC_FILE_RANGE:
4165 ret = io_prep_sfr(req, sqe);
4167 case IORING_OP_SENDMSG:
4168 case IORING_OP_SEND:
4169 ret = io_sendmsg_prep(req, sqe);
4171 case IORING_OP_RECVMSG:
4172 case IORING_OP_RECV:
4173 ret = io_recvmsg_prep(req, sqe);
4175 case IORING_OP_CONNECT:
4176 ret = io_connect_prep(req, sqe);
4178 case IORING_OP_TIMEOUT:
4179 ret = io_timeout_prep(req, sqe, false);
4181 case IORING_OP_TIMEOUT_REMOVE:
4182 ret = io_timeout_remove_prep(req, sqe);
4184 case IORING_OP_ASYNC_CANCEL:
4185 ret = io_async_cancel_prep(req, sqe);
4187 case IORING_OP_LINK_TIMEOUT:
4188 ret = io_timeout_prep(req, sqe, true);
4190 case IORING_OP_ACCEPT:
4191 ret = io_accept_prep(req, sqe);
4193 case IORING_OP_FALLOCATE:
4194 ret = io_fallocate_prep(req, sqe);
4196 case IORING_OP_OPENAT:
4197 ret = io_openat_prep(req, sqe);
4199 case IORING_OP_CLOSE:
4200 ret = io_close_prep(req, sqe);
4202 case IORING_OP_FILES_UPDATE:
4203 ret = io_files_update_prep(req, sqe);
4205 case IORING_OP_STATX:
4206 ret = io_statx_prep(req, sqe);
4208 case IORING_OP_FADVISE:
4209 ret = io_fadvise_prep(req, sqe);
4211 case IORING_OP_MADVISE:
4212 ret = io_madvise_prep(req, sqe);
4214 case IORING_OP_OPENAT2:
4215 ret = io_openat2_prep(req, sqe);
4217 case IORING_OP_EPOLL_CTL:
4218 ret = io_epoll_ctl_prep(req, sqe);
4221 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
4230 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4232 struct io_ring_ctx *ctx = req->ctx;
4235 /* Still need defer if there is pending req in defer list. */
4236 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
4239 if (!req->io && io_alloc_async_ctx(req))
4242 ret = io_req_defer_prep(req, sqe);
4246 spin_lock_irq(&ctx->completion_lock);
4247 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
4248 spin_unlock_irq(&ctx->completion_lock);
4252 trace_io_uring_defer(ctx, req, req->user_data);
4253 list_add_tail(&req->list, &ctx->defer_list);
4254 spin_unlock_irq(&ctx->completion_lock);
4255 return -EIOCBQUEUED;
4258 static void io_cleanup_req(struct io_kiocb *req)
4260 struct io_async_ctx *io = req->io;
4262 switch (req->opcode) {
4263 case IORING_OP_READV:
4264 case IORING_OP_READ_FIXED:
4265 case IORING_OP_READ:
4266 case IORING_OP_WRITEV:
4267 case IORING_OP_WRITE_FIXED:
4268 case IORING_OP_WRITE:
4269 if (io->rw.iov != io->rw.fast_iov)
4272 case IORING_OP_SENDMSG:
4273 case IORING_OP_RECVMSG:
4274 if (io->msg.iov != io->msg.fast_iov)
4277 case IORING_OP_OPENAT:
4278 case IORING_OP_OPENAT2:
4279 case IORING_OP_STATX:
4280 putname(req->open.filename);
4284 req->flags &= ~REQ_F_NEED_CLEANUP;
4287 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4288 struct io_kiocb **nxt, bool force_nonblock)
4290 struct io_ring_ctx *ctx = req->ctx;
4293 switch (req->opcode) {
4297 case IORING_OP_READV:
4298 case IORING_OP_READ_FIXED:
4299 case IORING_OP_READ:
4301 ret = io_read_prep(req, sqe, force_nonblock);
4305 ret = io_read(req, nxt, force_nonblock);
4307 case IORING_OP_WRITEV:
4308 case IORING_OP_WRITE_FIXED:
4309 case IORING_OP_WRITE:
4311 ret = io_write_prep(req, sqe, force_nonblock);
4315 ret = io_write(req, nxt, force_nonblock);
4317 case IORING_OP_FSYNC:
4319 ret = io_prep_fsync(req, sqe);
4323 ret = io_fsync(req, nxt, force_nonblock);
4325 case IORING_OP_POLL_ADD:
4327 ret = io_poll_add_prep(req, sqe);
4331 ret = io_poll_add(req, nxt);
4333 case IORING_OP_POLL_REMOVE:
4335 ret = io_poll_remove_prep(req, sqe);
4339 ret = io_poll_remove(req);
4341 case IORING_OP_SYNC_FILE_RANGE:
4343 ret = io_prep_sfr(req, sqe);
4347 ret = io_sync_file_range(req, nxt, force_nonblock);
4349 case IORING_OP_SENDMSG:
4350 case IORING_OP_SEND:
4352 ret = io_sendmsg_prep(req, sqe);
4356 if (req->opcode == IORING_OP_SENDMSG)
4357 ret = io_sendmsg(req, nxt, force_nonblock);
4359 ret = io_send(req, nxt, force_nonblock);
4361 case IORING_OP_RECVMSG:
4362 case IORING_OP_RECV:
4364 ret = io_recvmsg_prep(req, sqe);
4368 if (req->opcode == IORING_OP_RECVMSG)
4369 ret = io_recvmsg(req, nxt, force_nonblock);
4371 ret = io_recv(req, nxt, force_nonblock);
4373 case IORING_OP_TIMEOUT:
4375 ret = io_timeout_prep(req, sqe, false);
4379 ret = io_timeout(req);
4381 case IORING_OP_TIMEOUT_REMOVE:
4383 ret = io_timeout_remove_prep(req, sqe);
4387 ret = io_timeout_remove(req);
4389 case IORING_OP_ACCEPT:
4391 ret = io_accept_prep(req, sqe);
4395 ret = io_accept(req, nxt, force_nonblock);
4397 case IORING_OP_CONNECT:
4399 ret = io_connect_prep(req, sqe);
4403 ret = io_connect(req, nxt, force_nonblock);
4405 case IORING_OP_ASYNC_CANCEL:
4407 ret = io_async_cancel_prep(req, sqe);
4411 ret = io_async_cancel(req, nxt);
4413 case IORING_OP_FALLOCATE:
4415 ret = io_fallocate_prep(req, sqe);
4419 ret = io_fallocate(req, nxt, force_nonblock);
4421 case IORING_OP_OPENAT:
4423 ret = io_openat_prep(req, sqe);
4427 ret = io_openat(req, nxt, force_nonblock);
4429 case IORING_OP_CLOSE:
4431 ret = io_close_prep(req, sqe);
4435 ret = io_close(req, nxt, force_nonblock);
4437 case IORING_OP_FILES_UPDATE:
4439 ret = io_files_update_prep(req, sqe);
4443 ret = io_files_update(req, force_nonblock);
4445 case IORING_OP_STATX:
4447 ret = io_statx_prep(req, sqe);
4451 ret = io_statx(req, nxt, force_nonblock);
4453 case IORING_OP_FADVISE:
4455 ret = io_fadvise_prep(req, sqe);
4459 ret = io_fadvise(req, nxt, force_nonblock);
4461 case IORING_OP_MADVISE:
4463 ret = io_madvise_prep(req, sqe);
4467 ret = io_madvise(req, nxt, force_nonblock);
4469 case IORING_OP_OPENAT2:
4471 ret = io_openat2_prep(req, sqe);
4475 ret = io_openat2(req, nxt, force_nonblock);
4477 case IORING_OP_EPOLL_CTL:
4479 ret = io_epoll_ctl_prep(req, sqe);
4483 ret = io_epoll_ctl(req, nxt, force_nonblock);
4493 if (ctx->flags & IORING_SETUP_IOPOLL) {
4494 const bool in_async = io_wq_current_is_worker();
4496 if (req->result == -EAGAIN)
4499 /* workqueue context doesn't hold uring_lock, grab it now */
4501 mutex_lock(&ctx->uring_lock);
4503 io_iopoll_req_issued(req);
4506 mutex_unlock(&ctx->uring_lock);
4512 static void io_wq_submit_work(struct io_wq_work **workptr)
4514 struct io_wq_work *work = *workptr;
4515 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4516 struct io_kiocb *nxt = NULL;
4519 /* if NO_CANCEL is set, we must still run the work */
4520 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
4521 IO_WQ_WORK_CANCEL) {
4526 req->in_async = true;
4528 ret = io_issue_sqe(req, NULL, &nxt, false);
4530 * We can get EAGAIN for polled IO even though we're
4531 * forcing a sync submission from here, since we can't
4532 * wait for request slots on the block side.
4540 /* drop submission reference */
4544 req_set_fail_links(req);
4545 io_cqring_add_event(req, ret);
4549 /* if a dependent link is ready, pass it back */
4551 io_wq_assign_next(workptr, nxt);
4554 static int io_req_needs_file(struct io_kiocb *req, int fd)
4556 if (!io_op_defs[req->opcode].needs_file)
4558 if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg)
4563 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
4566 struct fixed_file_table *table;
4568 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
4569 return table->files[index & IORING_FILE_TABLE_MASK];;
4572 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
4573 const struct io_uring_sqe *sqe)
4575 struct io_ring_ctx *ctx = req->ctx;
4579 flags = READ_ONCE(sqe->flags);
4580 fd = READ_ONCE(sqe->fd);
4582 if (!io_req_needs_file(req, fd))
4585 if (flags & IOSQE_FIXED_FILE) {
4586 if (unlikely(!ctx->file_data ||
4587 (unsigned) fd >= ctx->nr_user_files))
4589 fd = array_index_nospec(fd, ctx->nr_user_files);
4590 req->file = io_file_from_index(ctx, fd);
4593 req->flags |= REQ_F_FIXED_FILE;
4594 percpu_ref_get(&ctx->file_data->refs);
4596 if (req->needs_fixed_file)
4598 trace_io_uring_file_get(ctx, fd);
4599 req->file = io_file_get(state, fd);
4600 if (unlikely(!req->file))
4607 static int io_grab_files(struct io_kiocb *req)
4610 struct io_ring_ctx *ctx = req->ctx;
4612 if (req->work.files)
4614 if (!ctx->ring_file)
4618 spin_lock_irq(&ctx->inflight_lock);
4620 * We use the f_ops->flush() handler to ensure that we can flush
4621 * out work accessing these files if the fd is closed. Check if
4622 * the fd has changed since we started down this path, and disallow
4623 * this operation if it has.
4625 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
4626 list_add(&req->inflight_entry, &ctx->inflight_list);
4627 req->flags |= REQ_F_INFLIGHT;
4628 req->work.files = current->files;
4631 spin_unlock_irq(&ctx->inflight_lock);
4637 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
4639 struct io_timeout_data *data = container_of(timer,
4640 struct io_timeout_data, timer);
4641 struct io_kiocb *req = data->req;
4642 struct io_ring_ctx *ctx = req->ctx;
4643 struct io_kiocb *prev = NULL;
4644 unsigned long flags;
4646 spin_lock_irqsave(&ctx->completion_lock, flags);
4649 * We don't expect the list to be empty, that will only happen if we
4650 * race with the completion of the linked work.
4652 if (!list_empty(&req->link_list)) {
4653 prev = list_entry(req->link_list.prev, struct io_kiocb,
4655 if (refcount_inc_not_zero(&prev->refs)) {
4656 list_del_init(&req->link_list);
4657 prev->flags &= ~REQ_F_LINK_TIMEOUT;
4662 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4665 req_set_fail_links(prev);
4666 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
4670 io_cqring_add_event(req, -ETIME);
4673 return HRTIMER_NORESTART;
4676 static void io_queue_linked_timeout(struct io_kiocb *req)
4678 struct io_ring_ctx *ctx = req->ctx;
4681 * If the list is now empty, then our linked request finished before
4682 * we got a chance to setup the timer
4684 spin_lock_irq(&ctx->completion_lock);
4685 if (!list_empty(&req->link_list)) {
4686 struct io_timeout_data *data = &req->io->timeout;
4688 data->timer.function = io_link_timeout_fn;
4689 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
4692 spin_unlock_irq(&ctx->completion_lock);
4694 /* drop submission reference */
4698 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
4700 struct io_kiocb *nxt;
4702 if (!(req->flags & REQ_F_LINK))
4705 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
4707 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
4710 req->flags |= REQ_F_LINK_TIMEOUT;
4714 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4716 struct io_kiocb *linked_timeout;
4717 struct io_kiocb *nxt = NULL;
4718 const struct cred *old_creds = NULL;
4722 linked_timeout = io_prep_linked_timeout(req);
4724 if (req->work.creds && req->work.creds != current_cred()) {
4726 revert_creds(old_creds);
4727 if (old_creds == req->work.creds)
4728 old_creds = NULL; /* restored original creds */
4730 old_creds = override_creds(req->work.creds);
4733 ret = io_issue_sqe(req, sqe, &nxt, true);
4736 * We async punt it if the file wasn't marked NOWAIT, or if the file
4737 * doesn't support non-blocking read/write attempts
4739 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
4740 (req->flags & REQ_F_MUST_PUNT))) {
4742 if (io_op_defs[req->opcode].file_table) {
4743 ret = io_grab_files(req);
4749 * Queued up for async execution, worker will release
4750 * submit reference when the iocb is actually submitted.
4752 io_queue_async_work(req);
4757 /* drop submission reference */
4758 io_put_req_find_next(req, &nxt);
4760 if (linked_timeout) {
4762 io_queue_linked_timeout(linked_timeout);
4764 io_put_req(linked_timeout);
4767 /* and drop final reference, if we failed */
4769 io_cqring_add_event(req, ret);
4770 req_set_fail_links(req);
4778 if (req->flags & REQ_F_FORCE_ASYNC)
4783 revert_creds(old_creds);
4786 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4790 ret = io_req_defer(req, sqe);
4792 if (ret != -EIOCBQUEUED) {
4794 io_cqring_add_event(req, ret);
4795 req_set_fail_links(req);
4796 io_double_put_req(req);
4798 } else if (req->flags & REQ_F_FORCE_ASYNC) {
4799 ret = io_req_defer_prep(req, sqe);
4800 if (unlikely(ret < 0))
4803 * Never try inline submit of IOSQE_ASYNC is set, go straight
4804 * to async execution.
4806 req->work.flags |= IO_WQ_WORK_CONCURRENT;
4807 io_queue_async_work(req);
4809 __io_queue_sqe(req, sqe);
4813 static inline void io_queue_link_head(struct io_kiocb *req)
4815 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
4816 io_cqring_add_event(req, -ECANCELED);
4817 io_double_put_req(req);
4819 io_queue_sqe(req, NULL);
4822 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
4823 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
4825 static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4826 struct io_submit_state *state, struct io_kiocb **link)
4828 struct io_ring_ctx *ctx = req->ctx;
4829 unsigned int sqe_flags;
4832 sqe_flags = READ_ONCE(sqe->flags);
4834 /* enforce forwards compatibility on users */
4835 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
4840 id = READ_ONCE(sqe->personality);
4842 req->work.creds = idr_find(&ctx->personality_idr, id);
4843 if (unlikely(!req->work.creds)) {
4847 get_cred(req->work.creds);
4850 /* same numerical values with corresponding REQ_F_*, safe to copy */
4851 req->flags |= sqe_flags & (IOSQE_IO_DRAIN|IOSQE_IO_HARDLINK|
4854 ret = io_req_set_file(state, req, sqe);
4855 if (unlikely(ret)) {
4857 io_cqring_add_event(req, ret);
4858 io_double_put_req(req);
4863 * If we already have a head request, queue this one for async
4864 * submittal once the head completes. If we don't have a head but
4865 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
4866 * submitted sync once the chain is complete. If none of those
4867 * conditions are true (normal request), then just queue it.
4870 struct io_kiocb *head = *link;
4873 * Taking sequential execution of a link, draining both sides
4874 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
4875 * requests in the link. So, it drains the head and the
4876 * next after the link request. The last one is done via
4877 * drain_next flag to persist the effect across calls.
4879 if (sqe_flags & IOSQE_IO_DRAIN) {
4880 head->flags |= REQ_F_IO_DRAIN;
4881 ctx->drain_next = 1;
4883 if (io_alloc_async_ctx(req)) {
4888 ret = io_req_defer_prep(req, sqe);
4890 /* fail even hard links since we don't submit */
4891 head->flags |= REQ_F_FAIL_LINK;
4894 trace_io_uring_link(ctx, req, head);
4895 list_add_tail(&req->link_list, &head->link_list);
4897 /* last request of a link, enqueue the link */
4898 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) {
4899 io_queue_link_head(head);
4903 if (unlikely(ctx->drain_next)) {
4904 req->flags |= REQ_F_IO_DRAIN;
4905 req->ctx->drain_next = 0;
4907 if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
4908 req->flags |= REQ_F_LINK;
4909 INIT_LIST_HEAD(&req->link_list);
4910 ret = io_req_defer_prep(req, sqe);
4912 req->flags |= REQ_F_FAIL_LINK;
4915 io_queue_sqe(req, sqe);
4923 * Batched submission is done, ensure local IO is flushed out.
4925 static void io_submit_state_end(struct io_submit_state *state)
4927 blk_finish_plug(&state->plug);
4929 if (state->free_reqs)
4930 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
4934 * Start submission side cache.
4936 static void io_submit_state_start(struct io_submit_state *state,
4937 unsigned int max_ios)
4939 blk_start_plug(&state->plug);
4940 state->free_reqs = 0;
4942 state->ios_left = max_ios;
4945 static void io_commit_sqring(struct io_ring_ctx *ctx)
4947 struct io_rings *rings = ctx->rings;
4950 * Ensure any loads from the SQEs are done at this point,
4951 * since once we write the new head, the application could
4952 * write new data to them.
4954 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
4958 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
4959 * that is mapped by userspace. This means that care needs to be taken to
4960 * ensure that reads are stable, as we cannot rely on userspace always
4961 * being a good citizen. If members of the sqe are validated and then later
4962 * used, it's important that those reads are done through READ_ONCE() to
4963 * prevent a re-load down the line.
4965 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
4966 const struct io_uring_sqe **sqe_ptr)
4968 u32 *sq_array = ctx->sq_array;
4972 * The cached sq head (or cq tail) serves two purposes:
4974 * 1) allows us to batch the cost of updating the user visible
4976 * 2) allows the kernel side to track the head on its own, even
4977 * though the application is the one updating it.
4979 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
4980 if (likely(head < ctx->sq_entries)) {
4982 * All io need record the previous position, if LINK vs DARIN,
4983 * it can be used to mark the position of the first IO in the
4986 req->sequence = ctx->cached_sq_head;
4987 *sqe_ptr = &ctx->sq_sqes[head];
4988 req->opcode = READ_ONCE((*sqe_ptr)->opcode);
4989 req->user_data = READ_ONCE((*sqe_ptr)->user_data);
4990 ctx->cached_sq_head++;
4994 /* drop invalid entries */
4995 ctx->cached_sq_head++;
4996 ctx->cached_sq_dropped++;
4997 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5001 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5002 struct file *ring_file, int ring_fd,
5003 struct mm_struct **mm, bool async)
5005 struct io_submit_state state, *statep = NULL;
5006 struct io_kiocb *link = NULL;
5007 int i, submitted = 0;
5008 bool mm_fault = false;
5010 /* if we have a backlog and couldn't flush it all, return BUSY */
5011 if (test_bit(0, &ctx->sq_check_overflow)) {
5012 if (!list_empty(&ctx->cq_overflow_list) &&
5013 !io_cqring_overflow_flush(ctx, false))
5017 /* make sure SQ entry isn't read before tail */
5018 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5020 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5023 if (nr > IO_PLUG_THRESHOLD) {
5024 io_submit_state_start(&state, nr);
5028 ctx->ring_fd = ring_fd;
5029 ctx->ring_file = ring_file;
5031 for (i = 0; i < nr; i++) {
5032 const struct io_uring_sqe *sqe;
5033 struct io_kiocb *req;
5036 req = io_get_req(ctx, statep);
5037 if (unlikely(!req)) {
5039 submitted = -EAGAIN;
5042 if (!io_get_sqring(ctx, req, &sqe)) {
5043 __io_req_do_free(req);
5047 /* will complete beyond this point, count as submitted */
5050 if (unlikely(req->opcode >= IORING_OP_LAST)) {
5053 io_cqring_add_event(req, err);
5054 io_double_put_req(req);
5058 if (io_op_defs[req->opcode].needs_mm && !*mm) {
5059 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
5060 if (unlikely(mm_fault)) {
5064 use_mm(ctx->sqo_mm);
5068 req->in_async = async;
5069 req->needs_fixed_file = async;
5070 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5072 if (!io_submit_sqe(req, sqe, statep, &link))
5076 if (unlikely(submitted != nr)) {
5077 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5079 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5082 io_queue_link_head(link);
5084 io_submit_state_end(&state);
5086 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5087 io_commit_sqring(ctx);
5092 static int io_sq_thread(void *data)
5094 struct io_ring_ctx *ctx = data;
5095 struct mm_struct *cur_mm = NULL;
5096 const struct cred *old_cred;
5097 mm_segment_t old_fs;
5099 unsigned long timeout;
5102 complete(&ctx->completions[1]);
5106 old_cred = override_creds(ctx->creds);
5108 timeout = jiffies + ctx->sq_thread_idle;
5109 while (!kthread_should_park()) {
5110 unsigned int to_submit;
5112 if (!list_empty(&ctx->poll_list)) {
5113 unsigned nr_events = 0;
5115 mutex_lock(&ctx->uring_lock);
5116 if (!list_empty(&ctx->poll_list))
5117 io_iopoll_getevents(ctx, &nr_events, 0);
5119 timeout = jiffies + ctx->sq_thread_idle;
5120 mutex_unlock(&ctx->uring_lock);
5123 to_submit = io_sqring_entries(ctx);
5126 * If submit got -EBUSY, flag us as needing the application
5127 * to enter the kernel to reap and flush events.
5129 if (!to_submit || ret == -EBUSY) {
5131 * Drop cur_mm before scheduling, we can't hold it for
5132 * long periods (or over schedule()). Do this before
5133 * adding ourselves to the waitqueue, as the unuse/drop
5143 * We're polling. If we're within the defined idle
5144 * period, then let us spin without work before going
5145 * to sleep. The exception is if we got EBUSY doing
5146 * more IO, we should wait for the application to
5147 * reap events and wake us up.
5149 if (!list_empty(&ctx->poll_list) ||
5150 (!time_after(jiffies, timeout) && ret != -EBUSY &&
5151 !percpu_ref_is_dying(&ctx->refs))) {
5156 prepare_to_wait(&ctx->sqo_wait, &wait,
5157 TASK_INTERRUPTIBLE);
5160 * While doing polled IO, before going to sleep, we need
5161 * to check if there are new reqs added to poll_list, it
5162 * is because reqs may have been punted to io worker and
5163 * will be added to poll_list later, hence check the
5166 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
5167 !list_empty_careful(&ctx->poll_list)) {
5168 finish_wait(&ctx->sqo_wait, &wait);
5172 /* Tell userspace we may need a wakeup call */
5173 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
5174 /* make sure to read SQ tail after writing flags */
5177 to_submit = io_sqring_entries(ctx);
5178 if (!to_submit || ret == -EBUSY) {
5179 if (kthread_should_park()) {
5180 finish_wait(&ctx->sqo_wait, &wait);
5183 if (signal_pending(current))
5184 flush_signals(current);
5186 finish_wait(&ctx->sqo_wait, &wait);
5188 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5191 finish_wait(&ctx->sqo_wait, &wait);
5193 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5196 mutex_lock(&ctx->uring_lock);
5197 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
5198 mutex_unlock(&ctx->uring_lock);
5199 timeout = jiffies + ctx->sq_thread_idle;
5207 revert_creds(old_cred);
5214 struct io_wait_queue {
5215 struct wait_queue_entry wq;
5216 struct io_ring_ctx *ctx;
5218 unsigned nr_timeouts;
5221 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
5223 struct io_ring_ctx *ctx = iowq->ctx;
5226 * Wake up if we have enough events, or if a timeout occurred since we
5227 * started waiting. For timeouts, we always want to return to userspace,
5228 * regardless of event count.
5230 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
5231 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
5234 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
5235 int wake_flags, void *key)
5237 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
5240 /* use noflush == true, as we can't safely rely on locking context */
5241 if (!io_should_wake(iowq, true))
5244 return autoremove_wake_function(curr, mode, wake_flags, key);
5248 * Wait until events become available, if we don't already have some. The
5249 * application must reap them itself, as they reside on the shared cq ring.
5251 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
5252 const sigset_t __user *sig, size_t sigsz)
5254 struct io_wait_queue iowq = {
5257 .func = io_wake_function,
5258 .entry = LIST_HEAD_INIT(iowq.wq.entry),
5261 .to_wait = min_events,
5263 struct io_rings *rings = ctx->rings;
5266 if (io_cqring_events(ctx, false) >= min_events)
5270 #ifdef CONFIG_COMPAT
5271 if (in_compat_syscall())
5272 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
5276 ret = set_user_sigmask(sig, sigsz);
5282 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
5283 trace_io_uring_cqring_wait(ctx, min_events);
5285 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
5286 TASK_INTERRUPTIBLE);
5287 if (io_should_wake(&iowq, false))
5290 if (signal_pending(current)) {
5295 finish_wait(&ctx->wait, &iowq.wq);
5297 restore_saved_sigmask_unless(ret == -EINTR);
5299 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
5302 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
5304 #if defined(CONFIG_UNIX)
5305 if (ctx->ring_sock) {
5306 struct sock *sock = ctx->ring_sock->sk;
5307 struct sk_buff *skb;
5309 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
5315 for (i = 0; i < ctx->nr_user_files; i++) {
5318 file = io_file_from_index(ctx, i);
5325 static void io_file_ref_kill(struct percpu_ref *ref)
5327 struct fixed_file_data *data;
5329 data = container_of(ref, struct fixed_file_data, refs);
5330 complete(&data->done);
5333 static void io_file_ref_exit_and_free(struct work_struct *work)
5335 struct fixed_file_data *data;
5337 data = container_of(work, struct fixed_file_data, ref_work);
5340 * Ensure any percpu-ref atomic switch callback has run, it could have
5341 * been in progress when the files were being unregistered. Once
5342 * that's done, we can safely exit and free the ref and containing
5346 percpu_ref_exit(&data->refs);
5350 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
5352 struct fixed_file_data *data = ctx->file_data;
5353 unsigned nr_tables, i;
5358 percpu_ref_kill_and_confirm(&data->refs, io_file_ref_kill);
5359 flush_work(&data->ref_work);
5360 wait_for_completion(&data->done);
5361 io_ring_file_ref_flush(data);
5363 __io_sqe_files_unregister(ctx);
5364 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
5365 for (i = 0; i < nr_tables; i++)
5366 kfree(data->table[i].files);
5368 INIT_WORK(&data->ref_work, io_file_ref_exit_and_free);
5369 queue_work(system_wq, &data->ref_work);
5370 ctx->file_data = NULL;
5371 ctx->nr_user_files = 0;
5375 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
5377 if (ctx->sqo_thread) {
5378 wait_for_completion(&ctx->completions[1]);
5380 * The park is a bit of a work-around, without it we get
5381 * warning spews on shutdown with SQPOLL set and affinity
5382 * set to a single CPU.
5384 kthread_park(ctx->sqo_thread);
5385 kthread_stop(ctx->sqo_thread);
5386 ctx->sqo_thread = NULL;
5390 static void io_finish_async(struct io_ring_ctx *ctx)
5392 io_sq_thread_stop(ctx);
5395 io_wq_destroy(ctx->io_wq);
5400 #if defined(CONFIG_UNIX)
5402 * Ensure the UNIX gc is aware of our file set, so we are certain that
5403 * the io_uring can be safely unregistered on process exit, even if we have
5404 * loops in the file referencing.
5406 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
5408 struct sock *sk = ctx->ring_sock->sk;
5409 struct scm_fp_list *fpl;
5410 struct sk_buff *skb;
5413 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
5414 unsigned long inflight = ctx->user->unix_inflight + nr;
5416 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
5420 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
5424 skb = alloc_skb(0, GFP_KERNEL);
5433 fpl->user = get_uid(ctx->user);
5434 for (i = 0; i < nr; i++) {
5435 struct file *file = io_file_from_index(ctx, i + offset);
5439 fpl->fp[nr_files] = get_file(file);
5440 unix_inflight(fpl->user, fpl->fp[nr_files]);
5445 fpl->max = SCM_MAX_FD;
5446 fpl->count = nr_files;
5447 UNIXCB(skb).fp = fpl;
5448 skb->destructor = unix_destruct_scm;
5449 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
5450 skb_queue_head(&sk->sk_receive_queue, skb);
5452 for (i = 0; i < nr_files; i++)
5463 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
5464 * causes regular reference counting to break down. We rely on the UNIX
5465 * garbage collection to take care of this problem for us.
5467 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5469 unsigned left, total;
5473 left = ctx->nr_user_files;
5475 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
5477 ret = __io_sqe_files_scm(ctx, this_files, total);
5481 total += this_files;
5487 while (total < ctx->nr_user_files) {
5488 struct file *file = io_file_from_index(ctx, total);
5498 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5504 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
5509 for (i = 0; i < nr_tables; i++) {
5510 struct fixed_file_table *table = &ctx->file_data->table[i];
5511 unsigned this_files;
5513 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
5514 table->files = kcalloc(this_files, sizeof(struct file *),
5518 nr_files -= this_files;
5524 for (i = 0; i < nr_tables; i++) {
5525 struct fixed_file_table *table = &ctx->file_data->table[i];
5526 kfree(table->files);
5531 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
5533 #if defined(CONFIG_UNIX)
5534 struct sock *sock = ctx->ring_sock->sk;
5535 struct sk_buff_head list, *head = &sock->sk_receive_queue;
5536 struct sk_buff *skb;
5539 __skb_queue_head_init(&list);
5542 * Find the skb that holds this file in its SCM_RIGHTS. When found,
5543 * remove this entry and rearrange the file array.
5545 skb = skb_dequeue(head);
5547 struct scm_fp_list *fp;
5549 fp = UNIXCB(skb).fp;
5550 for (i = 0; i < fp->count; i++) {
5553 if (fp->fp[i] != file)
5556 unix_notinflight(fp->user, fp->fp[i]);
5557 left = fp->count - 1 - i;
5559 memmove(&fp->fp[i], &fp->fp[i + 1],
5560 left * sizeof(struct file *));
5567 __skb_queue_tail(&list, skb);
5577 __skb_queue_tail(&list, skb);
5579 skb = skb_dequeue(head);
5582 if (skb_peek(&list)) {
5583 spin_lock_irq(&head->lock);
5584 while ((skb = __skb_dequeue(&list)) != NULL)
5585 __skb_queue_tail(head, skb);
5586 spin_unlock_irq(&head->lock);
5593 struct io_file_put {
5594 struct llist_node llist;
5596 struct completion *done;
5599 static void io_ring_file_ref_flush(struct fixed_file_data *data)
5601 struct io_file_put *pfile, *tmp;
5602 struct llist_node *node;
5604 while ((node = llist_del_all(&data->put_llist)) != NULL) {
5605 llist_for_each_entry_safe(pfile, tmp, node, llist) {
5606 io_ring_file_put(data->ctx, pfile->file);
5608 complete(pfile->done);
5615 static void io_ring_file_ref_switch(struct work_struct *work)
5617 struct fixed_file_data *data;
5619 data = container_of(work, struct fixed_file_data, ref_work);
5620 io_ring_file_ref_flush(data);
5621 percpu_ref_switch_to_percpu(&data->refs);
5624 static void io_file_data_ref_zero(struct percpu_ref *ref)
5626 struct fixed_file_data *data;
5628 data = container_of(ref, struct fixed_file_data, refs);
5631 * We can't safely switch from inside this context, punt to wq. If
5632 * the table ref is going away, the table is being unregistered.
5633 * Don't queue up the async work for that case, the caller will
5636 if (!percpu_ref_is_dying(&data->refs))
5637 queue_work(system_wq, &data->ref_work);
5640 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
5643 __s32 __user *fds = (__s32 __user *) arg;
5653 if (nr_args > IORING_MAX_FIXED_FILES)
5656 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
5657 if (!ctx->file_data)
5659 ctx->file_data->ctx = ctx;
5660 init_completion(&ctx->file_data->done);
5662 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
5663 ctx->file_data->table = kcalloc(nr_tables,
5664 sizeof(struct fixed_file_table),
5666 if (!ctx->file_data->table) {
5667 kfree(ctx->file_data);
5668 ctx->file_data = NULL;
5672 if (percpu_ref_init(&ctx->file_data->refs, io_file_data_ref_zero,
5673 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
5674 kfree(ctx->file_data->table);
5675 kfree(ctx->file_data);
5676 ctx->file_data = NULL;
5679 ctx->file_data->put_llist.first = NULL;
5680 INIT_WORK(&ctx->file_data->ref_work, io_ring_file_ref_switch);
5682 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
5683 percpu_ref_exit(&ctx->file_data->refs);
5684 kfree(ctx->file_data->table);
5685 kfree(ctx->file_data);
5686 ctx->file_data = NULL;
5690 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
5691 struct fixed_file_table *table;
5695 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
5697 /* allow sparse sets */
5703 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5704 index = i & IORING_FILE_TABLE_MASK;
5712 * Don't allow io_uring instances to be registered. If UNIX
5713 * isn't enabled, then this causes a reference cycle and this
5714 * instance can never get freed. If UNIX is enabled we'll
5715 * handle it just fine, but there's still no point in allowing
5716 * a ring fd as it doesn't support regular read/write anyway.
5718 if (file->f_op == &io_uring_fops) {
5723 table->files[index] = file;
5727 for (i = 0; i < ctx->nr_user_files; i++) {
5728 file = io_file_from_index(ctx, i);
5732 for (i = 0; i < nr_tables; i++)
5733 kfree(ctx->file_data->table[i].files);
5735 kfree(ctx->file_data->table);
5736 kfree(ctx->file_data);
5737 ctx->file_data = NULL;
5738 ctx->nr_user_files = 0;
5742 ret = io_sqe_files_scm(ctx);
5744 io_sqe_files_unregister(ctx);
5749 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
5752 #if defined(CONFIG_UNIX)
5753 struct sock *sock = ctx->ring_sock->sk;
5754 struct sk_buff_head *head = &sock->sk_receive_queue;
5755 struct sk_buff *skb;
5758 * See if we can merge this file into an existing skb SCM_RIGHTS
5759 * file set. If there's no room, fall back to allocating a new skb
5760 * and filling it in.
5762 spin_lock_irq(&head->lock);
5763 skb = skb_peek(head);
5765 struct scm_fp_list *fpl = UNIXCB(skb).fp;
5767 if (fpl->count < SCM_MAX_FD) {
5768 __skb_unlink(skb, head);
5769 spin_unlock_irq(&head->lock);
5770 fpl->fp[fpl->count] = get_file(file);
5771 unix_inflight(fpl->user, fpl->fp[fpl->count]);
5773 spin_lock_irq(&head->lock);
5774 __skb_queue_head(head, skb);
5779 spin_unlock_irq(&head->lock);
5786 return __io_sqe_files_scm(ctx, 1, index);
5792 static void io_atomic_switch(struct percpu_ref *ref)
5794 struct fixed_file_data *data;
5797 * Juggle reference to ensure we hit zero, if needed, so we can
5798 * switch back to percpu mode
5800 data = container_of(ref, struct fixed_file_data, refs);
5801 percpu_ref_put(&data->refs);
5802 percpu_ref_get(&data->refs);
5805 static bool io_queue_file_removal(struct fixed_file_data *data,
5808 struct io_file_put *pfile, pfile_stack;
5809 DECLARE_COMPLETION_ONSTACK(done);
5812 * If we fail allocating the struct we need for doing async reomval
5813 * of this file, just punt to sync and wait for it.
5815 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
5817 pfile = &pfile_stack;
5818 pfile->done = &done;
5822 llist_add(&pfile->llist, &data->put_llist);
5824 if (pfile == &pfile_stack) {
5825 percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch);
5826 wait_for_completion(&done);
5827 flush_work(&data->ref_work);
5834 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
5835 struct io_uring_files_update *up,
5838 struct fixed_file_data *data = ctx->file_data;
5839 bool ref_switch = false;
5845 if (check_add_overflow(up->offset, nr_args, &done))
5847 if (done > ctx->nr_user_files)
5851 fds = u64_to_user_ptr(up->fds);
5853 struct fixed_file_table *table;
5857 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
5861 i = array_index_nospec(up->offset, ctx->nr_user_files);
5862 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5863 index = i & IORING_FILE_TABLE_MASK;
5864 if (table->files[index]) {
5865 file = io_file_from_index(ctx, index);
5866 table->files[index] = NULL;
5867 if (io_queue_file_removal(data, file))
5877 * Don't allow io_uring instances to be registered. If
5878 * UNIX isn't enabled, then this causes a reference
5879 * cycle and this instance can never get freed. If UNIX
5880 * is enabled we'll handle it just fine, but there's
5881 * still no point in allowing a ring fd as it doesn't
5882 * support regular read/write anyway.
5884 if (file->f_op == &io_uring_fops) {
5889 table->files[index] = file;
5890 err = io_sqe_file_register(ctx, file, i);
5900 percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch);
5902 return done ? done : err;
5904 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
5907 struct io_uring_files_update up;
5909 if (!ctx->file_data)
5913 if (copy_from_user(&up, arg, sizeof(up)))
5918 return __io_sqe_files_update(ctx, &up, nr_args);
5921 static void io_put_work(struct io_wq_work *work)
5923 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5928 static void io_get_work(struct io_wq_work *work)
5930 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5932 refcount_inc(&req->refs);
5935 static int io_init_wq_offload(struct io_ring_ctx *ctx,
5936 struct io_uring_params *p)
5938 struct io_wq_data data;
5940 struct io_ring_ctx *ctx_attach;
5941 unsigned int concurrency;
5944 data.user = ctx->user;
5945 data.get_work = io_get_work;
5946 data.put_work = io_put_work;
5948 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
5949 /* Do QD, or 4 * CPUS, whatever is smallest */
5950 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
5952 ctx->io_wq = io_wq_create(concurrency, &data);
5953 if (IS_ERR(ctx->io_wq)) {
5954 ret = PTR_ERR(ctx->io_wq);
5960 f = fdget(p->wq_fd);
5964 if (f.file->f_op != &io_uring_fops) {
5969 ctx_attach = f.file->private_data;
5970 /* @io_wq is protected by holding the fd */
5971 if (!io_wq_get(ctx_attach->io_wq, &data)) {
5976 ctx->io_wq = ctx_attach->io_wq;
5982 static int io_sq_offload_start(struct io_ring_ctx *ctx,
5983 struct io_uring_params *p)
5987 init_waitqueue_head(&ctx->sqo_wait);
5988 mmgrab(current->mm);
5989 ctx->sqo_mm = current->mm;
5991 if (ctx->flags & IORING_SETUP_SQPOLL) {
5993 if (!capable(CAP_SYS_ADMIN))
5996 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
5997 if (!ctx->sq_thread_idle)
5998 ctx->sq_thread_idle = HZ;
6000 if (p->flags & IORING_SETUP_SQ_AFF) {
6001 int cpu = p->sq_thread_cpu;
6004 if (cpu >= nr_cpu_ids)
6006 if (!cpu_online(cpu))
6009 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6013 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6016 if (IS_ERR(ctx->sqo_thread)) {
6017 ret = PTR_ERR(ctx->sqo_thread);
6018 ctx->sqo_thread = NULL;
6021 wake_up_process(ctx->sqo_thread);
6022 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6023 /* Can't have SQ_AFF without SQPOLL */
6028 ret = io_init_wq_offload(ctx, p);
6034 io_finish_async(ctx);
6035 mmdrop(ctx->sqo_mm);
6040 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6042 atomic_long_sub(nr_pages, &user->locked_vm);
6045 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6047 unsigned long page_limit, cur_pages, new_pages;
6049 /* Don't allow more pages than we can safely lock */
6050 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6053 cur_pages = atomic_long_read(&user->locked_vm);
6054 new_pages = cur_pages + nr_pages;
6055 if (new_pages > page_limit)
6057 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6058 new_pages) != cur_pages);
6063 static void io_mem_free(void *ptr)
6070 page = virt_to_head_page(ptr);
6071 if (put_page_testzero(page))
6072 free_compound_page(page);
6075 static void *io_mem_alloc(size_t size)
6077 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6080 return (void *) __get_free_pages(gfp_flags, get_order(size));
6083 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6086 struct io_rings *rings;
6087 size_t off, sq_array_size;
6089 off = struct_size(rings, cqes, cq_entries);
6090 if (off == SIZE_MAX)
6094 off = ALIGN(off, SMP_CACHE_BYTES);
6099 sq_array_size = array_size(sizeof(u32), sq_entries);
6100 if (sq_array_size == SIZE_MAX)
6103 if (check_add_overflow(off, sq_array_size, &off))
6112 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6116 pages = (size_t)1 << get_order(
6117 rings_size(sq_entries, cq_entries, NULL));
6118 pages += (size_t)1 << get_order(
6119 array_size(sizeof(struct io_uring_sqe), sq_entries));
6124 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6128 if (!ctx->user_bufs)
6131 for (i = 0; i < ctx->nr_user_bufs; i++) {
6132 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6134 for (j = 0; j < imu->nr_bvecs; j++)
6135 unpin_user_page(imu->bvec[j].bv_page);
6137 if (ctx->account_mem)
6138 io_unaccount_mem(ctx->user, imu->nr_bvecs);
6143 kfree(ctx->user_bufs);
6144 ctx->user_bufs = NULL;
6145 ctx->nr_user_bufs = 0;
6149 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
6150 void __user *arg, unsigned index)
6152 struct iovec __user *src;
6154 #ifdef CONFIG_COMPAT
6156 struct compat_iovec __user *ciovs;
6157 struct compat_iovec ciov;
6159 ciovs = (struct compat_iovec __user *) arg;
6160 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
6163 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
6164 dst->iov_len = ciov.iov_len;
6168 src = (struct iovec __user *) arg;
6169 if (copy_from_user(dst, &src[index], sizeof(*dst)))
6174 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
6177 struct vm_area_struct **vmas = NULL;
6178 struct page **pages = NULL;
6179 int i, j, got_pages = 0;
6184 if (!nr_args || nr_args > UIO_MAXIOV)
6187 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
6189 if (!ctx->user_bufs)
6192 for (i = 0; i < nr_args; i++) {
6193 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6194 unsigned long off, start, end, ubuf;
6199 ret = io_copy_iov(ctx, &iov, arg, i);
6204 * Don't impose further limits on the size and buffer
6205 * constraints here, we'll -EINVAL later when IO is
6206 * submitted if they are wrong.
6209 if (!iov.iov_base || !iov.iov_len)
6212 /* arbitrary limit, but we need something */
6213 if (iov.iov_len > SZ_1G)
6216 ubuf = (unsigned long) iov.iov_base;
6217 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
6218 start = ubuf >> PAGE_SHIFT;
6219 nr_pages = end - start;
6221 if (ctx->account_mem) {
6222 ret = io_account_mem(ctx->user, nr_pages);
6228 if (!pages || nr_pages > got_pages) {
6231 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
6233 vmas = kvmalloc_array(nr_pages,
6234 sizeof(struct vm_area_struct *),
6236 if (!pages || !vmas) {
6238 if (ctx->account_mem)
6239 io_unaccount_mem(ctx->user, nr_pages);
6242 got_pages = nr_pages;
6245 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
6249 if (ctx->account_mem)
6250 io_unaccount_mem(ctx->user, nr_pages);
6255 down_read(¤t->mm->mmap_sem);
6256 pret = pin_user_pages(ubuf, nr_pages,
6257 FOLL_WRITE | FOLL_LONGTERM,
6259 if (pret == nr_pages) {
6260 /* don't support file backed memory */
6261 for (j = 0; j < nr_pages; j++) {
6262 struct vm_area_struct *vma = vmas[j];
6265 !is_file_hugepages(vma->vm_file)) {
6271 ret = pret < 0 ? pret : -EFAULT;
6273 up_read(¤t->mm->mmap_sem);
6276 * if we did partial map, or found file backed vmas,
6277 * release any pages we did get
6280 unpin_user_pages(pages, pret);
6281 if (ctx->account_mem)
6282 io_unaccount_mem(ctx->user, nr_pages);
6287 off = ubuf & ~PAGE_MASK;
6289 for (j = 0; j < nr_pages; j++) {
6292 vec_len = min_t(size_t, size, PAGE_SIZE - off);
6293 imu->bvec[j].bv_page = pages[j];
6294 imu->bvec[j].bv_len = vec_len;
6295 imu->bvec[j].bv_offset = off;
6299 /* store original address for later verification */
6301 imu->len = iov.iov_len;
6302 imu->nr_bvecs = nr_pages;
6304 ctx->nr_user_bufs++;
6312 io_sqe_buffer_unregister(ctx);
6316 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
6318 __s32 __user *fds = arg;
6324 if (copy_from_user(&fd, fds, sizeof(*fds)))
6327 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
6328 if (IS_ERR(ctx->cq_ev_fd)) {
6329 int ret = PTR_ERR(ctx->cq_ev_fd);
6330 ctx->cq_ev_fd = NULL;
6337 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
6339 if (ctx->cq_ev_fd) {
6340 eventfd_ctx_put(ctx->cq_ev_fd);
6341 ctx->cq_ev_fd = NULL;
6348 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
6350 io_finish_async(ctx);
6352 mmdrop(ctx->sqo_mm);
6354 io_iopoll_reap_events(ctx);
6355 io_sqe_buffer_unregister(ctx);
6356 io_sqe_files_unregister(ctx);
6357 io_eventfd_unregister(ctx);
6358 idr_destroy(&ctx->personality_idr);
6360 #if defined(CONFIG_UNIX)
6361 if (ctx->ring_sock) {
6362 ctx->ring_sock->file = NULL; /* so that iput() is called */
6363 sock_release(ctx->ring_sock);
6367 io_mem_free(ctx->rings);
6368 io_mem_free(ctx->sq_sqes);
6370 percpu_ref_exit(&ctx->refs);
6371 if (ctx->account_mem)
6372 io_unaccount_mem(ctx->user,
6373 ring_pages(ctx->sq_entries, ctx->cq_entries));
6374 free_uid(ctx->user);
6375 put_cred(ctx->creds);
6376 kfree(ctx->completions);
6377 kfree(ctx->cancel_hash);
6378 kmem_cache_free(req_cachep, ctx->fallback_req);
6382 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
6384 struct io_ring_ctx *ctx = file->private_data;
6387 poll_wait(file, &ctx->cq_wait, wait);
6389 * synchronizes with barrier from wq_has_sleeper call in
6393 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
6394 ctx->rings->sq_ring_entries)
6395 mask |= EPOLLOUT | EPOLLWRNORM;
6396 if (io_cqring_events(ctx, false))
6397 mask |= EPOLLIN | EPOLLRDNORM;
6402 static int io_uring_fasync(int fd, struct file *file, int on)
6404 struct io_ring_ctx *ctx = file->private_data;
6406 return fasync_helper(fd, file, on, &ctx->cq_fasync);
6409 static int io_remove_personalities(int id, void *p, void *data)
6411 struct io_ring_ctx *ctx = data;
6412 const struct cred *cred;
6414 cred = idr_remove(&ctx->personality_idr, id);
6420 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
6422 mutex_lock(&ctx->uring_lock);
6423 percpu_ref_kill(&ctx->refs);
6424 mutex_unlock(&ctx->uring_lock);
6427 * Wait for sq thread to idle, if we have one. It won't spin on new
6428 * work after we've killed the ctx ref above. This is important to do
6429 * before we cancel existing commands, as the thread could otherwise
6430 * be queueing new work post that. If that's work we need to cancel,
6431 * it could cause shutdown to hang.
6433 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
6436 io_kill_timeouts(ctx);
6437 io_poll_remove_all(ctx);
6440 io_wq_cancel_all(ctx->io_wq);
6442 io_iopoll_reap_events(ctx);
6443 /* if we failed setting up the ctx, we might not have any rings */
6445 io_cqring_overflow_flush(ctx, true);
6446 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
6447 wait_for_completion(&ctx->completions[0]);
6448 io_ring_ctx_free(ctx);
6451 static int io_uring_release(struct inode *inode, struct file *file)
6453 struct io_ring_ctx *ctx = file->private_data;
6455 file->private_data = NULL;
6456 io_ring_ctx_wait_and_kill(ctx);
6460 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
6461 struct files_struct *files)
6463 struct io_kiocb *req;
6466 while (!list_empty_careful(&ctx->inflight_list)) {
6467 struct io_kiocb *cancel_req = NULL;
6469 spin_lock_irq(&ctx->inflight_lock);
6470 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
6471 if (req->work.files != files)
6473 /* req is being completed, ignore */
6474 if (!refcount_inc_not_zero(&req->refs))
6480 prepare_to_wait(&ctx->inflight_wait, &wait,
6481 TASK_UNINTERRUPTIBLE);
6482 spin_unlock_irq(&ctx->inflight_lock);
6484 /* We need to keep going until we don't find a matching req */
6488 if (cancel_req->flags & REQ_F_OVERFLOW) {
6489 spin_lock_irq(&ctx->completion_lock);
6490 list_del(&cancel_req->list);
6491 cancel_req->flags &= ~REQ_F_OVERFLOW;
6492 if (list_empty(&ctx->cq_overflow_list)) {
6493 clear_bit(0, &ctx->sq_check_overflow);
6494 clear_bit(0, &ctx->cq_check_overflow);
6496 spin_unlock_irq(&ctx->completion_lock);
6498 WRITE_ONCE(ctx->rings->cq_overflow,
6499 atomic_inc_return(&ctx->cached_cq_overflow));
6502 * Put inflight ref and overflow ref. If that's
6503 * all we had, then we're done with this request.
6505 if (refcount_sub_and_test(2, &cancel_req->refs)) {
6506 io_put_req(cancel_req);
6511 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
6512 io_put_req(cancel_req);
6515 finish_wait(&ctx->inflight_wait, &wait);
6518 static int io_uring_flush(struct file *file, void *data)
6520 struct io_ring_ctx *ctx = file->private_data;
6522 io_uring_cancel_files(ctx, data);
6525 * If the task is going away, cancel work it may have pending
6527 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
6528 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
6533 static void *io_uring_validate_mmap_request(struct file *file,
6534 loff_t pgoff, size_t sz)
6536 struct io_ring_ctx *ctx = file->private_data;
6537 loff_t offset = pgoff << PAGE_SHIFT;
6542 case IORING_OFF_SQ_RING:
6543 case IORING_OFF_CQ_RING:
6546 case IORING_OFF_SQES:
6550 return ERR_PTR(-EINVAL);
6553 page = virt_to_head_page(ptr);
6554 if (sz > page_size(page))
6555 return ERR_PTR(-EINVAL);
6562 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6564 size_t sz = vma->vm_end - vma->vm_start;
6568 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
6570 return PTR_ERR(ptr);
6572 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
6573 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
6576 #else /* !CONFIG_MMU */
6578 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6580 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
6583 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
6585 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
6588 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
6589 unsigned long addr, unsigned long len,
6590 unsigned long pgoff, unsigned long flags)
6594 ptr = io_uring_validate_mmap_request(file, pgoff, len);
6596 return PTR_ERR(ptr);
6598 return (unsigned long) ptr;
6601 #endif /* !CONFIG_MMU */
6603 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
6604 u32, min_complete, u32, flags, const sigset_t __user *, sig,
6607 struct io_ring_ctx *ctx;
6612 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
6620 if (f.file->f_op != &io_uring_fops)
6624 ctx = f.file->private_data;
6625 if (!percpu_ref_tryget(&ctx->refs))
6629 * For SQ polling, the thread will do all submissions and completions.
6630 * Just return the requested submit count, and wake the thread if
6634 if (ctx->flags & IORING_SETUP_SQPOLL) {
6635 if (!list_empty_careful(&ctx->cq_overflow_list))
6636 io_cqring_overflow_flush(ctx, false);
6637 if (flags & IORING_ENTER_SQ_WAKEUP)
6638 wake_up(&ctx->sqo_wait);
6639 submitted = to_submit;
6640 } else if (to_submit) {
6641 struct mm_struct *cur_mm;
6643 mutex_lock(&ctx->uring_lock);
6644 /* already have mm, so io_submit_sqes() won't try to grab it */
6645 cur_mm = ctx->sqo_mm;
6646 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
6648 mutex_unlock(&ctx->uring_lock);
6650 if (submitted != to_submit)
6653 if (flags & IORING_ENTER_GETEVENTS) {
6654 unsigned nr_events = 0;
6656 min_complete = min(min_complete, ctx->cq_entries);
6658 if (ctx->flags & IORING_SETUP_IOPOLL) {
6659 ret = io_iopoll_check(ctx, &nr_events, min_complete);
6661 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
6666 percpu_ref_put(&ctx->refs);
6669 return submitted ? submitted : ret;
6672 #ifdef CONFIG_PROC_FS
6673 static int io_uring_show_cred(int id, void *p, void *data)
6675 const struct cred *cred = p;
6676 struct seq_file *m = data;
6677 struct user_namespace *uns = seq_user_ns(m);
6678 struct group_info *gi;
6683 seq_printf(m, "%5d\n", id);
6684 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
6685 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
6686 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
6687 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
6688 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
6689 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
6690 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
6691 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
6692 seq_puts(m, "\n\tGroups:\t");
6693 gi = cred->group_info;
6694 for (g = 0; g < gi->ngroups; g++) {
6695 seq_put_decimal_ull(m, g ? " " : "",
6696 from_kgid_munged(uns, gi->gid[g]));
6698 seq_puts(m, "\n\tCapEff:\t");
6699 cap = cred->cap_effective;
6700 CAP_FOR_EACH_U32(__capi)
6701 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
6706 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
6710 mutex_lock(&ctx->uring_lock);
6711 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
6712 for (i = 0; i < ctx->nr_user_files; i++) {
6713 struct fixed_file_table *table;
6716 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6717 f = table->files[i & IORING_FILE_TABLE_MASK];
6719 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
6721 seq_printf(m, "%5u: <none>\n", i);
6723 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
6724 for (i = 0; i < ctx->nr_user_bufs; i++) {
6725 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
6727 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
6728 (unsigned int) buf->len);
6730 if (!idr_is_empty(&ctx->personality_idr)) {
6731 seq_printf(m, "Personalities:\n");
6732 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
6734 mutex_unlock(&ctx->uring_lock);
6737 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
6739 struct io_ring_ctx *ctx = f->private_data;
6741 if (percpu_ref_tryget(&ctx->refs)) {
6742 __io_uring_show_fdinfo(ctx, m);
6743 percpu_ref_put(&ctx->refs);
6748 static const struct file_operations io_uring_fops = {
6749 .release = io_uring_release,
6750 .flush = io_uring_flush,
6751 .mmap = io_uring_mmap,
6753 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
6754 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
6756 .poll = io_uring_poll,
6757 .fasync = io_uring_fasync,
6758 #ifdef CONFIG_PROC_FS
6759 .show_fdinfo = io_uring_show_fdinfo,
6763 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
6764 struct io_uring_params *p)
6766 struct io_rings *rings;
6767 size_t size, sq_array_offset;
6769 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
6770 if (size == SIZE_MAX)
6773 rings = io_mem_alloc(size);
6778 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
6779 rings->sq_ring_mask = p->sq_entries - 1;
6780 rings->cq_ring_mask = p->cq_entries - 1;
6781 rings->sq_ring_entries = p->sq_entries;
6782 rings->cq_ring_entries = p->cq_entries;
6783 ctx->sq_mask = rings->sq_ring_mask;
6784 ctx->cq_mask = rings->cq_ring_mask;
6785 ctx->sq_entries = rings->sq_ring_entries;
6786 ctx->cq_entries = rings->cq_ring_entries;
6788 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
6789 if (size == SIZE_MAX) {
6790 io_mem_free(ctx->rings);
6795 ctx->sq_sqes = io_mem_alloc(size);
6796 if (!ctx->sq_sqes) {
6797 io_mem_free(ctx->rings);
6806 * Allocate an anonymous fd, this is what constitutes the application
6807 * visible backing of an io_uring instance. The application mmaps this
6808 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
6809 * we have to tie this fd to a socket for file garbage collection purposes.
6811 static int io_uring_get_fd(struct io_ring_ctx *ctx)
6816 #if defined(CONFIG_UNIX)
6817 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
6823 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
6827 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
6828 O_RDWR | O_CLOEXEC);
6831 ret = PTR_ERR(file);
6835 #if defined(CONFIG_UNIX)
6836 ctx->ring_sock->file = file;
6838 fd_install(ret, file);
6841 #if defined(CONFIG_UNIX)
6842 sock_release(ctx->ring_sock);
6843 ctx->ring_sock = NULL;
6848 static int io_uring_create(unsigned entries, struct io_uring_params *p)
6850 struct user_struct *user = NULL;
6851 struct io_ring_ctx *ctx;
6857 if (entries > IORING_MAX_ENTRIES) {
6858 if (!(p->flags & IORING_SETUP_CLAMP))
6860 entries = IORING_MAX_ENTRIES;
6864 * Use twice as many entries for the CQ ring. It's possible for the
6865 * application to drive a higher depth than the size of the SQ ring,
6866 * since the sqes are only used at submission time. This allows for
6867 * some flexibility in overcommitting a bit. If the application has
6868 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
6869 * of CQ ring entries manually.
6871 p->sq_entries = roundup_pow_of_two(entries);
6872 if (p->flags & IORING_SETUP_CQSIZE) {
6874 * If IORING_SETUP_CQSIZE is set, we do the same roundup
6875 * to a power-of-two, if it isn't already. We do NOT impose
6876 * any cq vs sq ring sizing.
6878 if (p->cq_entries < p->sq_entries)
6880 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
6881 if (!(p->flags & IORING_SETUP_CLAMP))
6883 p->cq_entries = IORING_MAX_CQ_ENTRIES;
6885 p->cq_entries = roundup_pow_of_two(p->cq_entries);
6887 p->cq_entries = 2 * p->sq_entries;
6890 user = get_uid(current_user());
6891 account_mem = !capable(CAP_IPC_LOCK);
6894 ret = io_account_mem(user,
6895 ring_pages(p->sq_entries, p->cq_entries));
6902 ctx = io_ring_ctx_alloc(p);
6905 io_unaccount_mem(user, ring_pages(p->sq_entries,
6910 ctx->compat = in_compat_syscall();
6911 ctx->account_mem = account_mem;
6913 ctx->creds = get_current_cred();
6915 ret = io_allocate_scq_urings(ctx, p);
6919 ret = io_sq_offload_start(ctx, p);
6923 memset(&p->sq_off, 0, sizeof(p->sq_off));
6924 p->sq_off.head = offsetof(struct io_rings, sq.head);
6925 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
6926 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
6927 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
6928 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
6929 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
6930 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
6932 memset(&p->cq_off, 0, sizeof(p->cq_off));
6933 p->cq_off.head = offsetof(struct io_rings, cq.head);
6934 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
6935 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
6936 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
6937 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
6938 p->cq_off.cqes = offsetof(struct io_rings, cqes);
6941 * Install ring fd as the very last thing, so we don't risk someone
6942 * having closed it before we finish setup
6944 ret = io_uring_get_fd(ctx);
6948 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
6949 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
6950 IORING_FEAT_CUR_PERSONALITY;
6951 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
6954 io_ring_ctx_wait_and_kill(ctx);
6959 * Sets up an aio uring context, and returns the fd. Applications asks for a
6960 * ring size, we return the actual sq/cq ring sizes (among other things) in the
6961 * params structure passed in.
6963 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
6965 struct io_uring_params p;
6969 if (copy_from_user(&p, params, sizeof(p)))
6971 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
6976 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
6977 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
6978 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
6981 ret = io_uring_create(entries, &p);
6985 if (copy_to_user(params, &p, sizeof(p)))
6991 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
6992 struct io_uring_params __user *, params)
6994 return io_uring_setup(entries, params);
6997 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
6999 struct io_uring_probe *p;
7003 size = struct_size(p, ops, nr_args);
7004 if (size == SIZE_MAX)
7006 p = kzalloc(size, GFP_KERNEL);
7011 if (copy_from_user(p, arg, size))
7014 if (memchr_inv(p, 0, size))
7017 p->last_op = IORING_OP_LAST - 1;
7018 if (nr_args > IORING_OP_LAST)
7019 nr_args = IORING_OP_LAST;
7021 for (i = 0; i < nr_args; i++) {
7023 if (!io_op_defs[i].not_supported)
7024 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7029 if (copy_to_user(arg, p, size))
7036 static int io_register_personality(struct io_ring_ctx *ctx)
7038 const struct cred *creds = get_current_cred();
7041 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7042 USHRT_MAX, GFP_KERNEL);
7048 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7050 const struct cred *old_creds;
7052 old_creds = idr_remove(&ctx->personality_idr, id);
7054 put_cred(old_creds);
7061 static bool io_register_op_must_quiesce(int op)
7064 case IORING_UNREGISTER_FILES:
7065 case IORING_REGISTER_FILES_UPDATE:
7066 case IORING_REGISTER_PROBE:
7067 case IORING_REGISTER_PERSONALITY:
7068 case IORING_UNREGISTER_PERSONALITY:
7075 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
7076 void __user *arg, unsigned nr_args)
7077 __releases(ctx->uring_lock)
7078 __acquires(ctx->uring_lock)
7083 * We're inside the ring mutex, if the ref is already dying, then
7084 * someone else killed the ctx or is already going through
7085 * io_uring_register().
7087 if (percpu_ref_is_dying(&ctx->refs))
7090 if (io_register_op_must_quiesce(opcode)) {
7091 percpu_ref_kill(&ctx->refs);
7094 * Drop uring mutex before waiting for references to exit. If
7095 * another thread is currently inside io_uring_enter() it might
7096 * need to grab the uring_lock to make progress. If we hold it
7097 * here across the drain wait, then we can deadlock. It's safe
7098 * to drop the mutex here, since no new references will come in
7099 * after we've killed the percpu ref.
7101 mutex_unlock(&ctx->uring_lock);
7102 ret = wait_for_completion_interruptible(&ctx->completions[0]);
7103 mutex_lock(&ctx->uring_lock);
7105 percpu_ref_resurrect(&ctx->refs);
7112 case IORING_REGISTER_BUFFERS:
7113 ret = io_sqe_buffer_register(ctx, arg, nr_args);
7115 case IORING_UNREGISTER_BUFFERS:
7119 ret = io_sqe_buffer_unregister(ctx);
7121 case IORING_REGISTER_FILES:
7122 ret = io_sqe_files_register(ctx, arg, nr_args);
7124 case IORING_UNREGISTER_FILES:
7128 ret = io_sqe_files_unregister(ctx);
7130 case IORING_REGISTER_FILES_UPDATE:
7131 ret = io_sqe_files_update(ctx, arg, nr_args);
7133 case IORING_REGISTER_EVENTFD:
7134 case IORING_REGISTER_EVENTFD_ASYNC:
7138 ret = io_eventfd_register(ctx, arg);
7141 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
7142 ctx->eventfd_async = 1;
7144 ctx->eventfd_async = 0;
7146 case IORING_UNREGISTER_EVENTFD:
7150 ret = io_eventfd_unregister(ctx);
7152 case IORING_REGISTER_PROBE:
7154 if (!arg || nr_args > 256)
7156 ret = io_probe(ctx, arg, nr_args);
7158 case IORING_REGISTER_PERSONALITY:
7162 ret = io_register_personality(ctx);
7164 case IORING_UNREGISTER_PERSONALITY:
7168 ret = io_unregister_personality(ctx, nr_args);
7175 if (io_register_op_must_quiesce(opcode)) {
7176 /* bring the ctx back to life */
7177 percpu_ref_reinit(&ctx->refs);
7179 reinit_completion(&ctx->completions[0]);
7184 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
7185 void __user *, arg, unsigned int, nr_args)
7187 struct io_ring_ctx *ctx;
7196 if (f.file->f_op != &io_uring_fops)
7199 ctx = f.file->private_data;
7201 mutex_lock(&ctx->uring_lock);
7202 ret = __io_uring_register(ctx, opcode, arg, nr_args);
7203 mutex_unlock(&ctx->uring_lock);
7204 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
7205 ctx->cq_ev_fd != NULL, ret);
7211 static int __init io_uring_init(void)
7213 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
7214 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
7215 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
7218 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
7219 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
7220 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
7221 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
7222 BUILD_BUG_SQE_ELEM(1, __u8, flags);
7223 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
7224 BUILD_BUG_SQE_ELEM(4, __s32, fd);
7225 BUILD_BUG_SQE_ELEM(8, __u64, off);
7226 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
7227 BUILD_BUG_SQE_ELEM(16, __u64, addr);
7228 BUILD_BUG_SQE_ELEM(24, __u32, len);
7229 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
7230 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
7231 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
7232 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
7233 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
7234 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
7235 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
7236 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
7237 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
7238 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
7239 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
7240 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
7241 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
7242 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
7243 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
7244 BUILD_BUG_SQE_ELEM(42, __u16, personality);
7246 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
7247 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
7250 __initcall(io_uring_init);