1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/mmu_context.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
79 #define CREATE_TRACE_POINTS
80 #include <trace/events/io_uring.h>
82 #include <uapi/linux/io_uring.h>
87 #define IORING_MAX_ENTRIES 32768
88 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
91 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
93 #define IORING_FILE_TABLE_SHIFT 9
94 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
95 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
96 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
99 u32 head ____cacheline_aligned_in_smp;
100 u32 tail ____cacheline_aligned_in_smp;
104 * This data is shared with the application through the mmap at offsets
105 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
107 * The offsets to the member fields are published through struct
108 * io_sqring_offsets when calling io_uring_setup.
112 * Head and tail offsets into the ring; the offsets need to be
113 * masked to get valid indices.
115 * The kernel controls head of the sq ring and the tail of the cq ring,
116 * and the application controls tail of the sq ring and the head of the
119 struct io_uring sq, cq;
121 * Bitmasks to apply to head and tail offsets (constant, equals
124 u32 sq_ring_mask, cq_ring_mask;
125 /* Ring sizes (constant, power of 2) */
126 u32 sq_ring_entries, cq_ring_entries;
128 * Number of invalid entries dropped by the kernel due to
129 * invalid index stored in array
131 * Written by the kernel, shouldn't be modified by the
132 * application (i.e. get number of "new events" by comparing to
135 * After a new SQ head value was read by the application this
136 * counter includes all submissions that were dropped reaching
137 * the new SQ head (and possibly more).
143 * Written by the kernel, shouldn't be modified by the
146 * The application needs a full memory barrier before checking
147 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
151 * Number of completion events lost because the queue was full;
152 * this should be avoided by the application by making sure
153 * there are not more requests pending than there is space in
154 * the completion queue.
156 * Written by the kernel, shouldn't be modified by the
157 * application (i.e. get number of "new events" by comparing to
160 * As completion events come in out of order this counter is not
161 * ordered with any other data.
165 * Ring buffer of completion events.
167 * The kernel writes completion events fresh every time they are
168 * produced, so the application is allowed to modify pending
171 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
174 struct io_mapped_ubuf {
177 struct bio_vec *bvec;
178 unsigned int nr_bvecs;
181 struct fixed_file_table {
189 struct fixed_file_data {
190 struct fixed_file_table *table;
191 struct io_ring_ctx *ctx;
193 struct percpu_ref refs;
194 struct llist_head put_llist;
196 struct work_struct ref_work;
197 struct completion done;
202 struct percpu_ref refs;
203 } ____cacheline_aligned_in_smp;
209 int cq_overflow_flushed: 1;
211 int eventfd_async: 1;
214 * Ring buffer of indices into array of io_uring_sqe, which is
215 * mmapped by the application using the IORING_OFF_SQES offset.
217 * This indirection could e.g. be used to assign fixed
218 * io_uring_sqe entries to operations and only submit them to
219 * the queue when needed.
221 * The kernel modifies neither the indices array nor the entries
225 unsigned cached_sq_head;
228 unsigned sq_thread_idle;
229 unsigned cached_sq_dropped;
230 atomic_t cached_cq_overflow;
231 unsigned long sq_check_overflow;
233 struct list_head defer_list;
234 struct list_head timeout_list;
235 struct list_head cq_overflow_list;
237 wait_queue_head_t inflight_wait;
238 struct io_uring_sqe *sq_sqes;
239 } ____cacheline_aligned_in_smp;
241 struct io_rings *rings;
245 struct task_struct *sqo_thread; /* if using sq thread polling */
246 struct mm_struct *sqo_mm;
247 wait_queue_head_t sqo_wait;
250 * If used, fixed file set. Writers must ensure that ->refs is dead,
251 * readers must ensure that ->refs is alive as long as the file* is
252 * used. Only updated through io_uring_register(2).
254 struct fixed_file_data *file_data;
255 unsigned nr_user_files;
257 struct file *ring_file;
259 /* if used, fixed mapped user buffers */
260 unsigned nr_user_bufs;
261 struct io_mapped_ubuf *user_bufs;
263 struct user_struct *user;
265 const struct cred *creds;
267 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
268 struct completion *completions;
270 /* if all else fails... */
271 struct io_kiocb *fallback_req;
273 #if defined(CONFIG_UNIX)
274 struct socket *ring_sock;
277 struct idr personality_idr;
280 unsigned cached_cq_tail;
283 atomic_t cq_timeouts;
284 unsigned long cq_check_overflow;
285 struct wait_queue_head cq_wait;
286 struct fasync_struct *cq_fasync;
287 struct eventfd_ctx *cq_ev_fd;
288 } ____cacheline_aligned_in_smp;
291 struct mutex uring_lock;
292 wait_queue_head_t wait;
293 } ____cacheline_aligned_in_smp;
296 spinlock_t completion_lock;
297 struct llist_head poll_llist;
300 * ->poll_list is protected by the ctx->uring_lock for
301 * io_uring instances that don't use IORING_SETUP_SQPOLL.
302 * For SQPOLL, only the single threaded io_sq_thread() will
303 * manipulate the list, hence no extra locking is needed there.
305 struct list_head poll_list;
306 struct hlist_head *cancel_hash;
307 unsigned cancel_hash_bits;
308 bool poll_multi_file;
310 spinlock_t inflight_lock;
311 struct list_head inflight_list;
312 } ____cacheline_aligned_in_smp;
316 * First field must be the file pointer in all the
317 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
319 struct io_poll_iocb {
322 struct wait_queue_head *head;
328 struct wait_queue_entry wait;
333 struct file *put_file;
337 struct io_timeout_data {
338 struct io_kiocb *req;
339 struct hrtimer timer;
340 struct timespec64 ts;
341 enum hrtimer_mode mode;
347 struct sockaddr __user *addr;
348 int __user *addr_len;
373 /* NOTE: kiocb has the file as the first member, so don't do it here */
381 struct sockaddr __user *addr;
388 struct user_msghdr __user *msg;
401 struct filename *filename;
402 struct statx __user *buffer;
406 struct io_files_update {
432 struct epoll_event event;
435 struct io_async_connect {
436 struct sockaddr_storage address;
439 struct io_async_msghdr {
440 struct iovec fast_iov[UIO_FASTIOV];
442 struct sockaddr __user *uaddr;
447 struct iovec fast_iov[UIO_FASTIOV];
453 struct io_async_open {
454 struct filename *filename;
457 struct io_async_ctx {
459 struct io_async_rw rw;
460 struct io_async_msghdr msg;
461 struct io_async_connect connect;
462 struct io_timeout_data timeout;
463 struct io_async_open open;
468 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
469 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
470 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
471 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
472 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
479 REQ_F_IOPOLL_COMPLETED_BIT,
480 REQ_F_LINK_TIMEOUT_BIT,
484 REQ_F_TIMEOUT_NOSEQ_BIT,
485 REQ_F_COMP_LOCKED_BIT,
490 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
491 /* drain existing IO first */
492 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
494 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
495 /* doesn't sever on completion < 0 */
496 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
498 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
500 /* already grabbed next link */
501 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
502 /* fail rest of links */
503 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
504 /* on inflight list */
505 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
506 /* read/write uses file position */
507 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
508 /* must not punt to workers */
509 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
510 /* polled IO has completed */
511 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
512 /* has linked timeout */
513 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
514 /* timeout request */
515 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
517 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
518 /* must be punted even for NONBLOCK */
519 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
520 /* no timeout sequence */
521 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
522 /* completion under lock */
523 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
527 * NOTE! Each of the iocb union members has the file pointer
528 * as the first entry in their struct definition. So you can
529 * access the file pointer through any of the sub-structs,
530 * or directly as just 'ki_filp' in this struct.
536 struct io_poll_iocb poll;
537 struct io_accept accept;
539 struct io_cancel cancel;
540 struct io_timeout timeout;
541 struct io_connect connect;
542 struct io_sr_msg sr_msg;
544 struct io_close close;
545 struct io_files_update files_update;
546 struct io_fadvise fadvise;
547 struct io_madvise madvise;
548 struct io_epoll epoll;
551 struct io_async_ctx *io;
553 * llist_node is only used for poll deferred completions
555 struct llist_node llist_node;
558 bool needs_fixed_file;
561 struct io_ring_ctx *ctx;
563 struct list_head list;
564 struct hlist_node hash_node;
566 struct list_head link_list;
573 struct list_head inflight_entry;
575 struct io_wq_work work;
578 #define IO_PLUG_THRESHOLD 2
579 #define IO_IOPOLL_BATCH 8
581 struct io_submit_state {
582 struct blk_plug plug;
585 * io_kiocb alloc cache
587 void *reqs[IO_IOPOLL_BATCH];
588 unsigned int free_reqs;
591 * File reference cache
595 unsigned int has_refs;
596 unsigned int used_refs;
597 unsigned int ios_left;
601 /* needs req->io allocated for deferral/async */
602 unsigned async_ctx : 1;
603 /* needs current->mm setup, does mm access */
604 unsigned needs_mm : 1;
605 /* needs req->file assigned */
606 unsigned needs_file : 1;
607 /* needs req->file assigned IFF fd is >= 0 */
608 unsigned fd_non_neg : 1;
609 /* hash wq insertion if file is a regular file */
610 unsigned hash_reg_file : 1;
611 /* unbound wq insertion if file is a non-regular file */
612 unsigned unbound_nonreg_file : 1;
613 /* opcode is not supported by this kernel */
614 unsigned not_supported : 1;
615 /* needs file table */
616 unsigned file_table : 1;
619 static const struct io_op_def io_op_defs[] = {
620 [IORING_OP_NOP] = {},
621 [IORING_OP_READV] = {
625 .unbound_nonreg_file = 1,
627 [IORING_OP_WRITEV] = {
632 .unbound_nonreg_file = 1,
634 [IORING_OP_FSYNC] = {
637 [IORING_OP_READ_FIXED] = {
639 .unbound_nonreg_file = 1,
641 [IORING_OP_WRITE_FIXED] = {
644 .unbound_nonreg_file = 1,
646 [IORING_OP_POLL_ADD] = {
648 .unbound_nonreg_file = 1,
650 [IORING_OP_POLL_REMOVE] = {},
651 [IORING_OP_SYNC_FILE_RANGE] = {
654 [IORING_OP_SENDMSG] = {
658 .unbound_nonreg_file = 1,
660 [IORING_OP_RECVMSG] = {
664 .unbound_nonreg_file = 1,
666 [IORING_OP_TIMEOUT] = {
670 [IORING_OP_TIMEOUT_REMOVE] = {},
671 [IORING_OP_ACCEPT] = {
674 .unbound_nonreg_file = 1,
677 [IORING_OP_ASYNC_CANCEL] = {},
678 [IORING_OP_LINK_TIMEOUT] = {
682 [IORING_OP_CONNECT] = {
686 .unbound_nonreg_file = 1,
688 [IORING_OP_FALLOCATE] = {
691 [IORING_OP_OPENAT] = {
696 [IORING_OP_CLOSE] = {
700 [IORING_OP_FILES_UPDATE] = {
704 [IORING_OP_STATX] = {
712 .unbound_nonreg_file = 1,
714 [IORING_OP_WRITE] = {
717 .unbound_nonreg_file = 1,
719 [IORING_OP_FADVISE] = {
722 [IORING_OP_MADVISE] = {
728 .unbound_nonreg_file = 1,
733 .unbound_nonreg_file = 1,
735 [IORING_OP_OPENAT2] = {
740 [IORING_OP_EPOLL_CTL] = {
741 .unbound_nonreg_file = 1,
746 static void io_wq_submit_work(struct io_wq_work **workptr);
747 static void io_cqring_fill_event(struct io_kiocb *req, long res);
748 static void io_put_req(struct io_kiocb *req);
749 static void __io_double_put_req(struct io_kiocb *req);
750 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
751 static void io_queue_linked_timeout(struct io_kiocb *req);
752 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
753 struct io_uring_files_update *ip,
755 static int io_grab_files(struct io_kiocb *req);
757 static struct kmem_cache *req_cachep;
759 static const struct file_operations io_uring_fops;
761 struct sock *io_uring_get_socket(struct file *file)
763 #if defined(CONFIG_UNIX)
764 if (file->f_op == &io_uring_fops) {
765 struct io_ring_ctx *ctx = file->private_data;
767 return ctx->ring_sock->sk;
772 EXPORT_SYMBOL(io_uring_get_socket);
774 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
776 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
778 complete(&ctx->completions[0]);
781 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
783 struct io_ring_ctx *ctx;
786 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
790 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
791 if (!ctx->fallback_req)
794 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
795 if (!ctx->completions)
799 * Use 5 bits less than the max cq entries, that should give us around
800 * 32 entries per hash list if totally full and uniformly spread.
802 hash_bits = ilog2(p->cq_entries);
806 ctx->cancel_hash_bits = hash_bits;
807 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
809 if (!ctx->cancel_hash)
811 __hash_init(ctx->cancel_hash, 1U << hash_bits);
813 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
814 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
817 ctx->flags = p->flags;
818 init_waitqueue_head(&ctx->cq_wait);
819 INIT_LIST_HEAD(&ctx->cq_overflow_list);
820 init_completion(&ctx->completions[0]);
821 init_completion(&ctx->completions[1]);
822 idr_init(&ctx->personality_idr);
823 mutex_init(&ctx->uring_lock);
824 init_waitqueue_head(&ctx->wait);
825 spin_lock_init(&ctx->completion_lock);
826 init_llist_head(&ctx->poll_llist);
827 INIT_LIST_HEAD(&ctx->poll_list);
828 INIT_LIST_HEAD(&ctx->defer_list);
829 INIT_LIST_HEAD(&ctx->timeout_list);
830 init_waitqueue_head(&ctx->inflight_wait);
831 spin_lock_init(&ctx->inflight_lock);
832 INIT_LIST_HEAD(&ctx->inflight_list);
835 if (ctx->fallback_req)
836 kmem_cache_free(req_cachep, ctx->fallback_req);
837 kfree(ctx->completions);
838 kfree(ctx->cancel_hash);
843 static inline bool __req_need_defer(struct io_kiocb *req)
845 struct io_ring_ctx *ctx = req->ctx;
847 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
848 + atomic_read(&ctx->cached_cq_overflow);
851 static inline bool req_need_defer(struct io_kiocb *req)
853 if (unlikely(req->flags & REQ_F_IO_DRAIN))
854 return __req_need_defer(req);
859 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
861 struct io_kiocb *req;
863 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
864 if (req && !req_need_defer(req)) {
865 list_del_init(&req->list);
872 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
874 struct io_kiocb *req;
876 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
878 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
880 if (!__req_need_defer(req)) {
881 list_del_init(&req->list);
889 static void __io_commit_cqring(struct io_ring_ctx *ctx)
891 struct io_rings *rings = ctx->rings;
893 /* order cqe stores with ring update */
894 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
896 if (wq_has_sleeper(&ctx->cq_wait)) {
897 wake_up_interruptible(&ctx->cq_wait);
898 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
902 static inline void io_req_work_grab_env(struct io_kiocb *req,
903 const struct io_op_def *def)
905 if (!req->work.mm && def->needs_mm) {
907 req->work.mm = current->mm;
909 if (!req->work.creds)
910 req->work.creds = get_current_cred();
913 static inline void io_req_work_drop_env(struct io_kiocb *req)
916 mmdrop(req->work.mm);
919 if (req->work.creds) {
920 put_cred(req->work.creds);
921 req->work.creds = NULL;
925 static inline bool io_prep_async_work(struct io_kiocb *req,
926 struct io_kiocb **link)
928 const struct io_op_def *def = &io_op_defs[req->opcode];
929 bool do_hashed = false;
931 if (req->flags & REQ_F_ISREG) {
932 if (def->hash_reg_file)
935 if (def->unbound_nonreg_file)
936 req->work.flags |= IO_WQ_WORK_UNBOUND;
939 io_req_work_grab_env(req, def);
941 *link = io_prep_linked_timeout(req);
945 static inline void io_queue_async_work(struct io_kiocb *req)
947 struct io_ring_ctx *ctx = req->ctx;
948 struct io_kiocb *link;
951 do_hashed = io_prep_async_work(req, &link);
953 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
956 io_wq_enqueue(ctx->io_wq, &req->work);
958 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
959 file_inode(req->file));
963 io_queue_linked_timeout(link);
966 static void io_kill_timeout(struct io_kiocb *req)
970 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
972 atomic_inc(&req->ctx->cq_timeouts);
973 list_del_init(&req->list);
974 io_cqring_fill_event(req, 0);
979 static void io_kill_timeouts(struct io_ring_ctx *ctx)
981 struct io_kiocb *req, *tmp;
983 spin_lock_irq(&ctx->completion_lock);
984 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
985 io_kill_timeout(req);
986 spin_unlock_irq(&ctx->completion_lock);
989 static void io_commit_cqring(struct io_ring_ctx *ctx)
991 struct io_kiocb *req;
993 while ((req = io_get_timeout_req(ctx)) != NULL)
994 io_kill_timeout(req);
996 __io_commit_cqring(ctx);
998 while ((req = io_get_deferred_req(ctx)) != NULL)
999 io_queue_async_work(req);
1002 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1004 struct io_rings *rings = ctx->rings;
1007 tail = ctx->cached_cq_tail;
1009 * writes to the cq entry need to come after reading head; the
1010 * control dependency is enough as we're using WRITE_ONCE to
1013 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1016 ctx->cached_cq_tail++;
1017 return &rings->cqes[tail & ctx->cq_mask];
1020 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1024 if (!ctx->eventfd_async)
1026 return io_wq_current_is_worker() || in_interrupt();
1029 static void __io_cqring_ev_posted(struct io_ring_ctx *ctx, bool trigger_ev)
1031 if (waitqueue_active(&ctx->wait))
1032 wake_up(&ctx->wait);
1033 if (waitqueue_active(&ctx->sqo_wait))
1034 wake_up(&ctx->sqo_wait);
1036 eventfd_signal(ctx->cq_ev_fd, 1);
1039 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1041 __io_cqring_ev_posted(ctx, io_should_trigger_evfd(ctx));
1044 /* Returns true if there are no backlogged entries after the flush */
1045 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1047 struct io_rings *rings = ctx->rings;
1048 struct io_uring_cqe *cqe;
1049 struct io_kiocb *req;
1050 unsigned long flags;
1054 if (list_empty_careful(&ctx->cq_overflow_list))
1056 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1057 rings->cq_ring_entries))
1061 spin_lock_irqsave(&ctx->completion_lock, flags);
1063 /* if force is set, the ring is going away. always drop after that */
1065 ctx->cq_overflow_flushed = 1;
1068 while (!list_empty(&ctx->cq_overflow_list)) {
1069 cqe = io_get_cqring(ctx);
1073 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1075 list_move(&req->list, &list);
1077 WRITE_ONCE(cqe->user_data, req->user_data);
1078 WRITE_ONCE(cqe->res, req->result);
1079 WRITE_ONCE(cqe->flags, 0);
1081 WRITE_ONCE(ctx->rings->cq_overflow,
1082 atomic_inc_return(&ctx->cached_cq_overflow));
1086 io_commit_cqring(ctx);
1088 clear_bit(0, &ctx->sq_check_overflow);
1089 clear_bit(0, &ctx->cq_check_overflow);
1091 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1092 io_cqring_ev_posted(ctx);
1094 while (!list_empty(&list)) {
1095 req = list_first_entry(&list, struct io_kiocb, list);
1096 list_del(&req->list);
1103 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1105 struct io_ring_ctx *ctx = req->ctx;
1106 struct io_uring_cqe *cqe;
1108 trace_io_uring_complete(ctx, req->user_data, res);
1111 * If we can't get a cq entry, userspace overflowed the
1112 * submission (by quite a lot). Increment the overflow count in
1115 cqe = io_get_cqring(ctx);
1117 WRITE_ONCE(cqe->user_data, req->user_data);
1118 WRITE_ONCE(cqe->res, res);
1119 WRITE_ONCE(cqe->flags, 0);
1120 } else if (ctx->cq_overflow_flushed) {
1121 WRITE_ONCE(ctx->rings->cq_overflow,
1122 atomic_inc_return(&ctx->cached_cq_overflow));
1124 if (list_empty(&ctx->cq_overflow_list)) {
1125 set_bit(0, &ctx->sq_check_overflow);
1126 set_bit(0, &ctx->cq_check_overflow);
1128 refcount_inc(&req->refs);
1130 list_add_tail(&req->list, &ctx->cq_overflow_list);
1134 static void io_cqring_add_event(struct io_kiocb *req, long res)
1136 struct io_ring_ctx *ctx = req->ctx;
1137 unsigned long flags;
1139 spin_lock_irqsave(&ctx->completion_lock, flags);
1140 io_cqring_fill_event(req, res);
1141 io_commit_cqring(ctx);
1142 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1144 io_cqring_ev_posted(ctx);
1147 static inline bool io_is_fallback_req(struct io_kiocb *req)
1149 return req == (struct io_kiocb *)
1150 ((unsigned long) req->ctx->fallback_req & ~1UL);
1153 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1155 struct io_kiocb *req;
1157 req = ctx->fallback_req;
1158 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
1164 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
1165 struct io_submit_state *state)
1167 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1168 struct io_kiocb *req;
1171 req = kmem_cache_alloc(req_cachep, gfp);
1174 } else if (!state->free_reqs) {
1178 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1179 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1182 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1183 * retry single alloc to be on the safe side.
1185 if (unlikely(ret <= 0)) {
1186 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1187 if (!state->reqs[0])
1191 state->free_reqs = ret - 1;
1192 req = state->reqs[ret - 1];
1195 req = state->reqs[state->free_reqs];
1203 /* one is dropped after submission, the other at completion */
1204 refcount_set(&req->refs, 2);
1206 INIT_IO_WORK(&req->work, io_wq_submit_work);
1209 req = io_get_fallback_req(ctx);
1212 percpu_ref_put(&ctx->refs);
1216 static void __io_req_do_free(struct io_kiocb *req)
1218 if (likely(!io_is_fallback_req(req)))
1219 kmem_cache_free(req_cachep, req);
1221 clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req);
1224 static void __io_req_aux_free(struct io_kiocb *req)
1226 struct io_ring_ctx *ctx = req->ctx;
1230 if (req->flags & REQ_F_FIXED_FILE)
1231 percpu_ref_put(&ctx->file_data->refs);
1236 io_req_work_drop_env(req);
1239 static void __io_free_req(struct io_kiocb *req)
1241 __io_req_aux_free(req);
1243 if (req->flags & REQ_F_INFLIGHT) {
1244 struct io_ring_ctx *ctx = req->ctx;
1245 unsigned long flags;
1247 spin_lock_irqsave(&ctx->inflight_lock, flags);
1248 list_del(&req->inflight_entry);
1249 if (waitqueue_active(&ctx->inflight_wait))
1250 wake_up(&ctx->inflight_wait);
1251 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1254 percpu_ref_put(&req->ctx->refs);
1255 __io_req_do_free(req);
1259 void *reqs[IO_IOPOLL_BATCH];
1264 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1266 int fixed_refs = rb->to_free;
1270 if (rb->need_iter) {
1271 int i, inflight = 0;
1272 unsigned long flags;
1275 for (i = 0; i < rb->to_free; i++) {
1276 struct io_kiocb *req = rb->reqs[i];
1278 if (req->flags & REQ_F_FIXED_FILE) {
1282 if (req->flags & REQ_F_INFLIGHT)
1284 __io_req_aux_free(req);
1289 spin_lock_irqsave(&ctx->inflight_lock, flags);
1290 for (i = 0; i < rb->to_free; i++) {
1291 struct io_kiocb *req = rb->reqs[i];
1293 if (req->flags & REQ_F_INFLIGHT) {
1294 list_del(&req->inflight_entry);
1299 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1301 if (waitqueue_active(&ctx->inflight_wait))
1302 wake_up(&ctx->inflight_wait);
1305 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1307 percpu_ref_put_many(&ctx->file_data->refs, fixed_refs);
1308 percpu_ref_put_many(&ctx->refs, rb->to_free);
1309 rb->to_free = rb->need_iter = 0;
1312 static bool io_link_cancel_timeout(struct io_kiocb *req)
1314 struct io_ring_ctx *ctx = req->ctx;
1317 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1319 io_cqring_fill_event(req, -ECANCELED);
1320 io_commit_cqring(ctx);
1321 req->flags &= ~REQ_F_LINK;
1329 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1331 struct io_ring_ctx *ctx = req->ctx;
1332 bool wake_ev = false;
1334 /* Already got next link */
1335 if (req->flags & REQ_F_LINK_NEXT)
1339 * The list should never be empty when we are called here. But could
1340 * potentially happen if the chain is messed up, check to be on the
1343 while (!list_empty(&req->link_list)) {
1344 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1345 struct io_kiocb, link_list);
1347 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1348 (nxt->flags & REQ_F_TIMEOUT))) {
1349 list_del_init(&nxt->link_list);
1350 wake_ev |= io_link_cancel_timeout(nxt);
1351 req->flags &= ~REQ_F_LINK_TIMEOUT;
1355 list_del_init(&req->link_list);
1356 if (!list_empty(&nxt->link_list))
1357 nxt->flags |= REQ_F_LINK;
1362 req->flags |= REQ_F_LINK_NEXT;
1364 io_cqring_ev_posted(ctx);
1368 * Called if REQ_F_LINK is set, and we fail the head request
1370 static void io_fail_links(struct io_kiocb *req)
1372 struct io_ring_ctx *ctx = req->ctx;
1373 unsigned long flags;
1375 spin_lock_irqsave(&ctx->completion_lock, flags);
1377 while (!list_empty(&req->link_list)) {
1378 struct io_kiocb *link = list_first_entry(&req->link_list,
1379 struct io_kiocb, link_list);
1381 list_del_init(&link->link_list);
1382 trace_io_uring_fail_link(req, link);
1384 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1385 link->opcode == IORING_OP_LINK_TIMEOUT) {
1386 io_link_cancel_timeout(link);
1388 io_cqring_fill_event(link, -ECANCELED);
1389 __io_double_put_req(link);
1391 req->flags &= ~REQ_F_LINK_TIMEOUT;
1394 io_commit_cqring(ctx);
1395 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1396 io_cqring_ev_posted(ctx);
1399 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1401 if (likely(!(req->flags & REQ_F_LINK)))
1405 * If LINK is set, we have dependent requests in this chain. If we
1406 * didn't fail this request, queue the first one up, moving any other
1407 * dependencies to the next request. In case of failure, fail the rest
1410 if (req->flags & REQ_F_FAIL_LINK) {
1412 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1413 REQ_F_LINK_TIMEOUT) {
1414 struct io_ring_ctx *ctx = req->ctx;
1415 unsigned long flags;
1418 * If this is a timeout link, we could be racing with the
1419 * timeout timer. Grab the completion lock for this case to
1420 * protect against that.
1422 spin_lock_irqsave(&ctx->completion_lock, flags);
1423 io_req_link_next(req, nxt);
1424 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1426 io_req_link_next(req, nxt);
1430 static void io_free_req(struct io_kiocb *req)
1432 struct io_kiocb *nxt = NULL;
1434 io_req_find_next(req, &nxt);
1438 io_queue_async_work(nxt);
1442 * Drop reference to request, return next in chain (if there is one) if this
1443 * was the last reference to this request.
1445 __attribute__((nonnull))
1446 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1448 io_req_find_next(req, nxtptr);
1450 if (refcount_dec_and_test(&req->refs))
1454 static void io_put_req(struct io_kiocb *req)
1456 if (refcount_dec_and_test(&req->refs))
1461 * Must only be used if we don't need to care about links, usually from
1462 * within the completion handling itself.
1464 static void __io_double_put_req(struct io_kiocb *req)
1466 /* drop both submit and complete references */
1467 if (refcount_sub_and_test(2, &req->refs))
1471 static void io_double_put_req(struct io_kiocb *req)
1473 /* drop both submit and complete references */
1474 if (refcount_sub_and_test(2, &req->refs))
1478 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1480 struct io_rings *rings = ctx->rings;
1482 if (test_bit(0, &ctx->cq_check_overflow)) {
1484 * noflush == true is from the waitqueue handler, just ensure
1485 * we wake up the task, and the next invocation will flush the
1486 * entries. We cannot safely to it from here.
1488 if (noflush && !list_empty(&ctx->cq_overflow_list))
1491 io_cqring_overflow_flush(ctx, false);
1494 /* See comment at the top of this file */
1496 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1499 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1501 struct io_rings *rings = ctx->rings;
1503 /* make sure SQ entry isn't read before tail */
1504 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1507 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1509 if ((req->flags & REQ_F_LINK) || io_is_fallback_req(req))
1512 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1515 rb->reqs[rb->to_free++] = req;
1516 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1517 io_free_req_many(req->ctx, rb);
1522 * Find and free completed poll iocbs
1524 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1525 struct list_head *done)
1527 struct req_batch rb;
1528 struct io_kiocb *req;
1530 rb.to_free = rb.need_iter = 0;
1531 while (!list_empty(done)) {
1532 req = list_first_entry(done, struct io_kiocb, list);
1533 list_del(&req->list);
1535 io_cqring_fill_event(req, req->result);
1538 if (refcount_dec_and_test(&req->refs) &&
1539 !io_req_multi_free(&rb, req))
1543 io_commit_cqring(ctx);
1544 io_free_req_many(ctx, &rb);
1547 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1550 struct io_kiocb *req, *tmp;
1556 * Only spin for completions if we don't have multiple devices hanging
1557 * off our complete list, and we're under the requested amount.
1559 spin = !ctx->poll_multi_file && *nr_events < min;
1562 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1563 struct kiocb *kiocb = &req->rw.kiocb;
1566 * Move completed entries to our local list. If we find a
1567 * request that requires polling, break out and complete
1568 * the done list first, if we have entries there.
1570 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1571 list_move_tail(&req->list, &done);
1574 if (!list_empty(&done))
1577 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1586 if (!list_empty(&done))
1587 io_iopoll_complete(ctx, nr_events, &done);
1593 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1594 * non-spinning poll check - we'll still enter the driver poll loop, but only
1595 * as a non-spinning completion check.
1597 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1600 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1603 ret = io_do_iopoll(ctx, nr_events, min);
1606 if (!min || *nr_events >= min)
1614 * We can't just wait for polled events to come to us, we have to actively
1615 * find and complete them.
1617 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1619 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1622 mutex_lock(&ctx->uring_lock);
1623 while (!list_empty(&ctx->poll_list)) {
1624 unsigned int nr_events = 0;
1626 io_iopoll_getevents(ctx, &nr_events, 1);
1629 * Ensure we allow local-to-the-cpu processing to take place,
1630 * in this case we need to ensure that we reap all events.
1634 mutex_unlock(&ctx->uring_lock);
1637 static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1640 int iters = 0, ret = 0;
1646 * Don't enter poll loop if we already have events pending.
1647 * If we do, we can potentially be spinning for commands that
1648 * already triggered a CQE (eg in error).
1650 if (io_cqring_events(ctx, false))
1654 * If a submit got punted to a workqueue, we can have the
1655 * application entering polling for a command before it gets
1656 * issued. That app will hold the uring_lock for the duration
1657 * of the poll right here, so we need to take a breather every
1658 * now and then to ensure that the issue has a chance to add
1659 * the poll to the issued list. Otherwise we can spin here
1660 * forever, while the workqueue is stuck trying to acquire the
1663 if (!(++iters & 7)) {
1664 mutex_unlock(&ctx->uring_lock);
1665 mutex_lock(&ctx->uring_lock);
1668 if (*nr_events < min)
1669 tmin = min - *nr_events;
1671 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1675 } while (min && !*nr_events && !need_resched());
1680 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1686 * We disallow the app entering submit/complete with polling, but we
1687 * still need to lock the ring to prevent racing with polled issue
1688 * that got punted to a workqueue.
1690 mutex_lock(&ctx->uring_lock);
1691 ret = __io_iopoll_check(ctx, nr_events, min);
1692 mutex_unlock(&ctx->uring_lock);
1696 static void kiocb_end_write(struct io_kiocb *req)
1699 * Tell lockdep we inherited freeze protection from submission
1702 if (req->flags & REQ_F_ISREG) {
1703 struct inode *inode = file_inode(req->file);
1705 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1707 file_end_write(req->file);
1710 static inline void req_set_fail_links(struct io_kiocb *req)
1712 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1713 req->flags |= REQ_F_FAIL_LINK;
1716 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1718 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1720 if (kiocb->ki_flags & IOCB_WRITE)
1721 kiocb_end_write(req);
1723 if (res != req->result)
1724 req_set_fail_links(req);
1725 io_cqring_add_event(req, res);
1728 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1730 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1732 io_complete_rw_common(kiocb, res);
1736 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1738 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1739 struct io_kiocb *nxt = NULL;
1741 io_complete_rw_common(kiocb, res);
1742 io_put_req_find_next(req, &nxt);
1747 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1749 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1751 if (kiocb->ki_flags & IOCB_WRITE)
1752 kiocb_end_write(req);
1754 if (res != req->result)
1755 req_set_fail_links(req);
1758 req->flags |= REQ_F_IOPOLL_COMPLETED;
1762 * After the iocb has been issued, it's safe to be found on the poll list.
1763 * Adding the kiocb to the list AFTER submission ensures that we don't
1764 * find it from a io_iopoll_getevents() thread before the issuer is done
1765 * accessing the kiocb cookie.
1767 static void io_iopoll_req_issued(struct io_kiocb *req)
1769 struct io_ring_ctx *ctx = req->ctx;
1772 * Track whether we have multiple files in our lists. This will impact
1773 * how we do polling eventually, not spinning if we're on potentially
1774 * different devices.
1776 if (list_empty(&ctx->poll_list)) {
1777 ctx->poll_multi_file = false;
1778 } else if (!ctx->poll_multi_file) {
1779 struct io_kiocb *list_req;
1781 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1783 if (list_req->file != req->file)
1784 ctx->poll_multi_file = true;
1788 * For fast devices, IO may have already completed. If it has, add
1789 * it to the front so we find it first.
1791 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1792 list_add(&req->list, &ctx->poll_list);
1794 list_add_tail(&req->list, &ctx->poll_list);
1797 static void io_file_put(struct io_submit_state *state)
1800 int diff = state->has_refs - state->used_refs;
1803 fput_many(state->file, diff);
1809 * Get as many references to a file as we have IOs left in this submission,
1810 * assuming most submissions are for one file, or at least that each file
1811 * has more than one submission.
1813 static struct file *io_file_get(struct io_submit_state *state, int fd)
1819 if (state->fd == fd) {
1826 state->file = fget_many(fd, state->ios_left);
1831 state->has_refs = state->ios_left;
1832 state->used_refs = 1;
1838 * If we tracked the file through the SCM inflight mechanism, we could support
1839 * any file. For now, just ensure that anything potentially problematic is done
1842 static bool io_file_supports_async(struct file *file)
1844 umode_t mode = file_inode(file)->i_mode;
1846 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
1848 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1854 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1855 bool force_nonblock)
1857 struct io_ring_ctx *ctx = req->ctx;
1858 struct kiocb *kiocb = &req->rw.kiocb;
1862 if (S_ISREG(file_inode(req->file)->i_mode))
1863 req->flags |= REQ_F_ISREG;
1865 kiocb->ki_pos = READ_ONCE(sqe->off);
1866 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
1867 req->flags |= REQ_F_CUR_POS;
1868 kiocb->ki_pos = req->file->f_pos;
1870 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1871 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1872 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1876 ioprio = READ_ONCE(sqe->ioprio);
1878 ret = ioprio_check_cap(ioprio);
1882 kiocb->ki_ioprio = ioprio;
1884 kiocb->ki_ioprio = get_current_ioprio();
1886 /* don't allow async punt if RWF_NOWAIT was requested */
1887 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1888 (req->file->f_flags & O_NONBLOCK))
1889 req->flags |= REQ_F_NOWAIT;
1892 kiocb->ki_flags |= IOCB_NOWAIT;
1894 if (ctx->flags & IORING_SETUP_IOPOLL) {
1895 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1896 !kiocb->ki_filp->f_op->iopoll)
1899 kiocb->ki_flags |= IOCB_HIPRI;
1900 kiocb->ki_complete = io_complete_rw_iopoll;
1903 if (kiocb->ki_flags & IOCB_HIPRI)
1905 kiocb->ki_complete = io_complete_rw;
1908 req->rw.addr = READ_ONCE(sqe->addr);
1909 req->rw.len = READ_ONCE(sqe->len);
1910 /* we own ->private, reuse it for the buffer index */
1911 req->rw.kiocb.private = (void *) (unsigned long)
1912 READ_ONCE(sqe->buf_index);
1916 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1922 case -ERESTARTNOINTR:
1923 case -ERESTARTNOHAND:
1924 case -ERESTART_RESTARTBLOCK:
1926 * We can't just restart the syscall, since previously
1927 * submitted sqes may already be in progress. Just fail this
1933 kiocb->ki_complete(kiocb, ret, 0);
1937 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1940 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1942 if (req->flags & REQ_F_CUR_POS)
1943 req->file->f_pos = kiocb->ki_pos;
1944 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1945 *nxt = __io_complete_rw(kiocb, ret);
1947 io_rw_done(kiocb, ret);
1950 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
1951 struct iov_iter *iter)
1953 struct io_ring_ctx *ctx = req->ctx;
1954 size_t len = req->rw.len;
1955 struct io_mapped_ubuf *imu;
1956 unsigned index, buf_index;
1960 /* attempt to use fixed buffers without having provided iovecs */
1961 if (unlikely(!ctx->user_bufs))
1964 buf_index = (unsigned long) req->rw.kiocb.private;
1965 if (unlikely(buf_index >= ctx->nr_user_bufs))
1968 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1969 imu = &ctx->user_bufs[index];
1970 buf_addr = req->rw.addr;
1973 if (buf_addr + len < buf_addr)
1975 /* not inside the mapped region */
1976 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1980 * May not be a start of buffer, set size appropriately
1981 * and advance us to the beginning.
1983 offset = buf_addr - imu->ubuf;
1984 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1988 * Don't use iov_iter_advance() here, as it's really slow for
1989 * using the latter parts of a big fixed buffer - it iterates
1990 * over each segment manually. We can cheat a bit here, because
1993 * 1) it's a BVEC iter, we set it up
1994 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1995 * first and last bvec
1997 * So just find our index, and adjust the iterator afterwards.
1998 * If the offset is within the first bvec (or the whole first
1999 * bvec, just use iov_iter_advance(). This makes it easier
2000 * since we can just skip the first segment, which may not
2001 * be PAGE_SIZE aligned.
2003 const struct bio_vec *bvec = imu->bvec;
2005 if (offset <= bvec->bv_len) {
2006 iov_iter_advance(iter, offset);
2008 unsigned long seg_skip;
2010 /* skip first vec */
2011 offset -= bvec->bv_len;
2012 seg_skip = 1 + (offset >> PAGE_SHIFT);
2014 iter->bvec = bvec + seg_skip;
2015 iter->nr_segs -= seg_skip;
2016 iter->count -= bvec->bv_len + offset;
2017 iter->iov_offset = offset & ~PAGE_MASK;
2024 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2025 struct iovec **iovec, struct iov_iter *iter)
2027 void __user *buf = u64_to_user_ptr(req->rw.addr);
2028 size_t sqe_len = req->rw.len;
2031 opcode = req->opcode;
2032 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2034 return io_import_fixed(req, rw, iter);
2037 /* buffer index only valid with fixed read/write */
2038 if (req->rw.kiocb.private)
2041 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2043 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2049 struct io_async_rw *iorw = &req->io->rw;
2052 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2053 if (iorw->iov == iorw->fast_iov)
2061 #ifdef CONFIG_COMPAT
2062 if (req->ctx->compat)
2063 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2067 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2071 * For files that don't have ->read_iter() and ->write_iter(), handle them
2072 * by looping over ->read() or ->write() manually.
2074 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2075 struct iov_iter *iter)
2080 * Don't support polled IO through this interface, and we can't
2081 * support non-blocking either. For the latter, this just causes
2082 * the kiocb to be handled from an async context.
2084 if (kiocb->ki_flags & IOCB_HIPRI)
2086 if (kiocb->ki_flags & IOCB_NOWAIT)
2089 while (iov_iter_count(iter)) {
2093 if (!iov_iter_is_bvec(iter)) {
2094 iovec = iov_iter_iovec(iter);
2096 /* fixed buffers import bvec */
2097 iovec.iov_base = kmap(iter->bvec->bv_page)
2099 iovec.iov_len = min(iter->count,
2100 iter->bvec->bv_len - iter->iov_offset);
2104 nr = file->f_op->read(file, iovec.iov_base,
2105 iovec.iov_len, &kiocb->ki_pos);
2107 nr = file->f_op->write(file, iovec.iov_base,
2108 iovec.iov_len, &kiocb->ki_pos);
2111 if (iov_iter_is_bvec(iter))
2112 kunmap(iter->bvec->bv_page);
2120 if (nr != iovec.iov_len)
2122 iov_iter_advance(iter, nr);
2128 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2129 struct iovec *iovec, struct iovec *fast_iov,
2130 struct iov_iter *iter)
2132 req->io->rw.nr_segs = iter->nr_segs;
2133 req->io->rw.size = io_size;
2134 req->io->rw.iov = iovec;
2135 if (!req->io->rw.iov) {
2136 req->io->rw.iov = req->io->rw.fast_iov;
2137 memcpy(req->io->rw.iov, fast_iov,
2138 sizeof(struct iovec) * iter->nr_segs);
2142 static int io_alloc_async_ctx(struct io_kiocb *req)
2144 if (!io_op_defs[req->opcode].async_ctx)
2146 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2147 return req->io == NULL;
2150 static void io_rw_async(struct io_wq_work **workptr)
2152 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2153 struct iovec *iov = NULL;
2155 if (req->io->rw.iov != req->io->rw.fast_iov)
2156 iov = req->io->rw.iov;
2157 io_wq_submit_work(workptr);
2161 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2162 struct iovec *iovec, struct iovec *fast_iov,
2163 struct iov_iter *iter)
2165 if (!io_op_defs[req->opcode].async_ctx)
2168 if (io_alloc_async_ctx(req))
2171 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2173 req->work.func = io_rw_async;
2177 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2178 bool force_nonblock)
2180 struct io_async_ctx *io;
2181 struct iov_iter iter;
2184 ret = io_prep_rw(req, sqe, force_nonblock);
2188 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2195 io->rw.iov = io->rw.fast_iov;
2197 ret = io_import_iovec(READ, req, &io->rw.iov, &iter);
2202 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2206 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
2207 bool force_nonblock)
2209 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2210 struct kiocb *kiocb = &req->rw.kiocb;
2211 struct iov_iter iter;
2213 ssize_t io_size, ret;
2215 ret = io_import_iovec(READ, req, &iovec, &iter);
2219 /* Ensure we clear previously set non-block flag */
2220 if (!force_nonblock)
2221 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2225 if (req->flags & REQ_F_LINK)
2226 req->result = io_size;
2229 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2230 * we know to async punt it even if it was opened O_NONBLOCK
2232 if (force_nonblock && !io_file_supports_async(req->file)) {
2233 req->flags |= REQ_F_MUST_PUNT;
2237 iov_count = iov_iter_count(&iter);
2238 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2242 if (req->file->f_op->read_iter)
2243 ret2 = call_read_iter(req->file, kiocb, &iter);
2245 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2247 /* Catch -EAGAIN return for forced non-blocking submission */
2248 if (!force_nonblock || ret2 != -EAGAIN) {
2249 kiocb_done(kiocb, ret2, nxt, req->in_async);
2252 ret = io_setup_async_rw(req, io_size, iovec,
2253 inline_vecs, &iter);
2260 if (!io_wq_current_is_worker())
2265 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2266 bool force_nonblock)
2268 struct io_async_ctx *io;
2269 struct iov_iter iter;
2272 ret = io_prep_rw(req, sqe, force_nonblock);
2276 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2283 io->rw.iov = io->rw.fast_iov;
2285 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter);
2290 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2294 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
2295 bool force_nonblock)
2297 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2298 struct kiocb *kiocb = &req->rw.kiocb;
2299 struct iov_iter iter;
2301 ssize_t ret, io_size;
2303 ret = io_import_iovec(WRITE, req, &iovec, &iter);
2307 /* Ensure we clear previously set non-block flag */
2308 if (!force_nonblock)
2309 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2313 if (req->flags & REQ_F_LINK)
2314 req->result = io_size;
2317 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2318 * we know to async punt it even if it was opened O_NONBLOCK
2320 if (force_nonblock && !io_file_supports_async(req->file)) {
2321 req->flags |= REQ_F_MUST_PUNT;
2325 /* file path doesn't support NOWAIT for non-direct_IO */
2326 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2327 (req->flags & REQ_F_ISREG))
2330 iov_count = iov_iter_count(&iter);
2331 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2336 * Open-code file_start_write here to grab freeze protection,
2337 * which will be released by another thread in
2338 * io_complete_rw(). Fool lockdep by telling it the lock got
2339 * released so that it doesn't complain about the held lock when
2340 * we return to userspace.
2342 if (req->flags & REQ_F_ISREG) {
2343 __sb_start_write(file_inode(req->file)->i_sb,
2344 SB_FREEZE_WRITE, true);
2345 __sb_writers_release(file_inode(req->file)->i_sb,
2348 kiocb->ki_flags |= IOCB_WRITE;
2350 if (req->file->f_op->write_iter)
2351 ret2 = call_write_iter(req->file, kiocb, &iter);
2353 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2354 if (!force_nonblock || ret2 != -EAGAIN) {
2355 kiocb_done(kiocb, ret2, nxt, req->in_async);
2358 ret = io_setup_async_rw(req, io_size, iovec,
2359 inline_vecs, &iter);
2366 if (!io_wq_current_is_worker())
2372 * IORING_OP_NOP just posts a completion event, nothing else.
2374 static int io_nop(struct io_kiocb *req)
2376 struct io_ring_ctx *ctx = req->ctx;
2378 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2381 io_cqring_add_event(req, 0);
2386 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2388 struct io_ring_ctx *ctx = req->ctx;
2393 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2395 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2398 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2399 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2402 req->sync.off = READ_ONCE(sqe->off);
2403 req->sync.len = READ_ONCE(sqe->len);
2407 static bool io_req_cancelled(struct io_kiocb *req)
2409 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2410 req_set_fail_links(req);
2411 io_cqring_add_event(req, -ECANCELED);
2419 static void io_link_work_cb(struct io_wq_work **workptr)
2421 struct io_wq_work *work = *workptr;
2422 struct io_kiocb *link = work->data;
2424 io_queue_linked_timeout(link);
2425 work->func = io_wq_submit_work;
2428 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
2430 struct io_kiocb *link;
2432 io_prep_async_work(nxt, &link);
2433 *workptr = &nxt->work;
2435 nxt->work.flags |= IO_WQ_WORK_CB;
2436 nxt->work.func = io_link_work_cb;
2437 nxt->work.data = link;
2441 static void io_fsync_finish(struct io_wq_work **workptr)
2443 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2444 loff_t end = req->sync.off + req->sync.len;
2445 struct io_kiocb *nxt = NULL;
2448 if (io_req_cancelled(req))
2451 ret = vfs_fsync_range(req->file, req->sync.off,
2452 end > 0 ? end : LLONG_MAX,
2453 req->sync.flags & IORING_FSYNC_DATASYNC);
2455 req_set_fail_links(req);
2456 io_cqring_add_event(req, ret);
2457 io_put_req_find_next(req, &nxt);
2459 io_wq_assign_next(workptr, nxt);
2462 static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt,
2463 bool force_nonblock)
2465 struct io_wq_work *work, *old_work;
2467 /* fsync always requires a blocking context */
2468 if (force_nonblock) {
2470 req->work.func = io_fsync_finish;
2474 work = old_work = &req->work;
2475 io_fsync_finish(&work);
2476 if (work && work != old_work)
2477 *nxt = container_of(work, struct io_kiocb, work);
2481 static void io_fallocate_finish(struct io_wq_work **workptr)
2483 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2484 struct io_kiocb *nxt = NULL;
2487 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2490 req_set_fail_links(req);
2491 io_cqring_add_event(req, ret);
2492 io_put_req_find_next(req, &nxt);
2494 io_wq_assign_next(workptr, nxt);
2497 static int io_fallocate_prep(struct io_kiocb *req,
2498 const struct io_uring_sqe *sqe)
2500 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2503 req->sync.off = READ_ONCE(sqe->off);
2504 req->sync.len = READ_ONCE(sqe->addr);
2505 req->sync.mode = READ_ONCE(sqe->len);
2509 static int io_fallocate(struct io_kiocb *req, struct io_kiocb **nxt,
2510 bool force_nonblock)
2512 struct io_wq_work *work, *old_work;
2514 /* fallocate always requiring blocking context */
2515 if (force_nonblock) {
2517 req->work.func = io_fallocate_finish;
2521 work = old_work = &req->work;
2522 io_fallocate_finish(&work);
2523 if (work && work != old_work)
2524 *nxt = container_of(work, struct io_kiocb, work);
2529 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2531 const char __user *fname;
2534 if (sqe->ioprio || sqe->buf_index)
2537 req->open.dfd = READ_ONCE(sqe->fd);
2538 req->open.how.mode = READ_ONCE(sqe->len);
2539 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2540 req->open.how.flags = READ_ONCE(sqe->open_flags);
2542 req->open.filename = getname(fname);
2543 if (IS_ERR(req->open.filename)) {
2544 ret = PTR_ERR(req->open.filename);
2545 req->open.filename = NULL;
2552 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2554 struct open_how __user *how;
2555 const char __user *fname;
2559 if (sqe->ioprio || sqe->buf_index)
2562 req->open.dfd = READ_ONCE(sqe->fd);
2563 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2564 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2565 len = READ_ONCE(sqe->len);
2567 if (len < OPEN_HOW_SIZE_VER0)
2570 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2575 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2576 req->open.how.flags |= O_LARGEFILE;
2578 req->open.filename = getname(fname);
2579 if (IS_ERR(req->open.filename)) {
2580 ret = PTR_ERR(req->open.filename);
2581 req->open.filename = NULL;
2588 static int io_openat2(struct io_kiocb *req, struct io_kiocb **nxt,
2589 bool force_nonblock)
2591 struct open_flags op;
2598 ret = build_open_flags(&req->open.how, &op);
2602 ret = get_unused_fd_flags(req->open.how.flags);
2606 file = do_filp_open(req->open.dfd, req->open.filename, &op);
2609 ret = PTR_ERR(file);
2611 fsnotify_open(file);
2612 fd_install(ret, file);
2615 putname(req->open.filename);
2617 req_set_fail_links(req);
2618 io_cqring_add_event(req, ret);
2619 io_put_req_find_next(req, nxt);
2623 static int io_openat(struct io_kiocb *req, struct io_kiocb **nxt,
2624 bool force_nonblock)
2626 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
2627 return io_openat2(req, nxt, force_nonblock);
2630 static int io_epoll_ctl_prep(struct io_kiocb *req,
2631 const struct io_uring_sqe *sqe)
2633 #if defined(CONFIG_EPOLL)
2634 if (sqe->ioprio || sqe->buf_index)
2637 req->epoll.epfd = READ_ONCE(sqe->fd);
2638 req->epoll.op = READ_ONCE(sqe->len);
2639 req->epoll.fd = READ_ONCE(sqe->off);
2641 if (ep_op_has_event(req->epoll.op)) {
2642 struct epoll_event __user *ev;
2644 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
2645 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
2655 static int io_epoll_ctl(struct io_kiocb *req, struct io_kiocb **nxt,
2656 bool force_nonblock)
2658 #if defined(CONFIG_EPOLL)
2659 struct io_epoll *ie = &req->epoll;
2662 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
2663 if (force_nonblock && ret == -EAGAIN)
2667 req_set_fail_links(req);
2668 io_cqring_add_event(req, ret);
2669 io_put_req_find_next(req, nxt);
2676 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2678 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2679 if (sqe->ioprio || sqe->buf_index || sqe->off)
2682 req->madvise.addr = READ_ONCE(sqe->addr);
2683 req->madvise.len = READ_ONCE(sqe->len);
2684 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
2691 static int io_madvise(struct io_kiocb *req, struct io_kiocb **nxt,
2692 bool force_nonblock)
2694 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
2695 struct io_madvise *ma = &req->madvise;
2701 ret = do_madvise(ma->addr, ma->len, ma->advice);
2703 req_set_fail_links(req);
2704 io_cqring_add_event(req, ret);
2705 io_put_req_find_next(req, nxt);
2712 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2714 if (sqe->ioprio || sqe->buf_index || sqe->addr)
2717 req->fadvise.offset = READ_ONCE(sqe->off);
2718 req->fadvise.len = READ_ONCE(sqe->len);
2719 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
2723 static int io_fadvise(struct io_kiocb *req, struct io_kiocb **nxt,
2724 bool force_nonblock)
2726 struct io_fadvise *fa = &req->fadvise;
2729 if (force_nonblock) {
2730 switch (fa->advice) {
2731 case POSIX_FADV_NORMAL:
2732 case POSIX_FADV_RANDOM:
2733 case POSIX_FADV_SEQUENTIAL:
2740 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
2742 req_set_fail_links(req);
2743 io_cqring_add_event(req, ret);
2744 io_put_req_find_next(req, nxt);
2748 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2750 const char __user *fname;
2751 unsigned lookup_flags;
2754 if (sqe->ioprio || sqe->buf_index)
2757 req->open.dfd = READ_ONCE(sqe->fd);
2758 req->open.mask = READ_ONCE(sqe->len);
2759 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2760 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2761 req->open.how.flags = READ_ONCE(sqe->statx_flags);
2763 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
2766 req->open.filename = getname_flags(fname, lookup_flags, NULL);
2767 if (IS_ERR(req->open.filename)) {
2768 ret = PTR_ERR(req->open.filename);
2769 req->open.filename = NULL;
2776 static int io_statx(struct io_kiocb *req, struct io_kiocb **nxt,
2777 bool force_nonblock)
2779 struct io_open *ctx = &req->open;
2780 unsigned lookup_flags;
2788 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
2792 /* filename_lookup() drops it, keep a reference */
2793 ctx->filename->refcnt++;
2795 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
2800 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
2802 if (retry_estale(ret, lookup_flags)) {
2803 lookup_flags |= LOOKUP_REVAL;
2807 ret = cp_statx(&stat, ctx->buffer);
2809 putname(ctx->filename);
2811 req_set_fail_links(req);
2812 io_cqring_add_event(req, ret);
2813 io_put_req_find_next(req, nxt);
2817 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2820 * If we queue this for async, it must not be cancellable. That would
2821 * leave the 'file' in an undeterminate state.
2823 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
2825 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
2826 sqe->rw_flags || sqe->buf_index)
2828 if (sqe->flags & IOSQE_FIXED_FILE)
2831 req->close.fd = READ_ONCE(sqe->fd);
2832 if (req->file->f_op == &io_uring_fops ||
2833 req->close.fd == req->ctx->ring_fd)
2839 static void io_close_finish(struct io_wq_work **workptr)
2841 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2842 struct io_kiocb *nxt = NULL;
2844 /* Invoked with files, we need to do the close */
2845 if (req->work.files) {
2848 ret = filp_close(req->close.put_file, req->work.files);
2850 req_set_fail_links(req);
2851 io_cqring_add_event(req, ret);
2854 fput(req->close.put_file);
2856 io_put_req_find_next(req, &nxt);
2858 io_wq_assign_next(workptr, nxt);
2861 static int io_close(struct io_kiocb *req, struct io_kiocb **nxt,
2862 bool force_nonblock)
2866 req->close.put_file = NULL;
2867 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
2871 /* if the file has a flush method, be safe and punt to async */
2872 if (req->close.put_file->f_op->flush && !io_wq_current_is_worker())
2876 * No ->flush(), safely close from here and just punt the
2877 * fput() to async context.
2879 ret = filp_close(req->close.put_file, current->files);
2882 req_set_fail_links(req);
2883 io_cqring_add_event(req, ret);
2885 if (io_wq_current_is_worker()) {
2886 struct io_wq_work *old_work, *work;
2888 old_work = work = &req->work;
2889 io_close_finish(&work);
2890 if (work && work != old_work)
2891 *nxt = container_of(work, struct io_kiocb, work);
2896 req->work.func = io_close_finish;
2898 * Do manual async queue here to avoid grabbing files - we don't
2899 * need the files, and it'll cause io_close_finish() to close
2900 * the file again and cause a double CQE entry for this request
2902 io_queue_async_work(req);
2906 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2908 struct io_ring_ctx *ctx = req->ctx;
2913 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2915 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2918 req->sync.off = READ_ONCE(sqe->off);
2919 req->sync.len = READ_ONCE(sqe->len);
2920 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
2924 static void io_sync_file_range_finish(struct io_wq_work **workptr)
2926 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2927 struct io_kiocb *nxt = NULL;
2930 if (io_req_cancelled(req))
2933 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
2936 req_set_fail_links(req);
2937 io_cqring_add_event(req, ret);
2938 io_put_req_find_next(req, &nxt);
2940 io_wq_assign_next(workptr, nxt);
2943 static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt,
2944 bool force_nonblock)
2946 struct io_wq_work *work, *old_work;
2948 /* sync_file_range always requires a blocking context */
2949 if (force_nonblock) {
2951 req->work.func = io_sync_file_range_finish;
2955 work = old_work = &req->work;
2956 io_sync_file_range_finish(&work);
2957 if (work && work != old_work)
2958 *nxt = container_of(work, struct io_kiocb, work);
2962 #if defined(CONFIG_NET)
2963 static void io_sendrecv_async(struct io_wq_work **workptr)
2965 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2966 struct iovec *iov = NULL;
2968 if (req->io->rw.iov != req->io->rw.fast_iov)
2969 iov = req->io->msg.iov;
2970 io_wq_submit_work(workptr);
2975 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2977 #if defined(CONFIG_NET)
2978 struct io_sr_msg *sr = &req->sr_msg;
2979 struct io_async_ctx *io = req->io;
2981 sr->msg_flags = READ_ONCE(sqe->msg_flags);
2982 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
2983 sr->len = READ_ONCE(sqe->len);
2985 if (!io || req->opcode == IORING_OP_SEND)
2988 io->msg.iov = io->msg.fast_iov;
2989 return sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
2996 static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt,
2997 bool force_nonblock)
2999 #if defined(CONFIG_NET)
3000 struct io_async_msghdr *kmsg = NULL;
3001 struct socket *sock;
3004 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3007 sock = sock_from_file(req->file, &ret);
3009 struct io_async_ctx io;
3010 struct sockaddr_storage addr;
3014 kmsg = &req->io->msg;
3015 kmsg->msg.msg_name = &addr;
3016 /* if iov is set, it's allocated already */
3018 kmsg->iov = kmsg->fast_iov;
3019 kmsg->msg.msg_iter.iov = kmsg->iov;
3021 struct io_sr_msg *sr = &req->sr_msg;
3024 kmsg->msg.msg_name = &addr;
3026 io.msg.iov = io.msg.fast_iov;
3027 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3028 sr->msg_flags, &io.msg.iov);
3033 flags = req->sr_msg.msg_flags;
3034 if (flags & MSG_DONTWAIT)
3035 req->flags |= REQ_F_NOWAIT;
3036 else if (force_nonblock)
3037 flags |= MSG_DONTWAIT;
3039 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3040 if (force_nonblock && ret == -EAGAIN) {
3043 if (io_alloc_async_ctx(req))
3045 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3046 req->work.func = io_sendrecv_async;
3049 if (ret == -ERESTARTSYS)
3053 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
3055 io_cqring_add_event(req, ret);
3057 req_set_fail_links(req);
3058 io_put_req_find_next(req, nxt);
3065 static int io_send(struct io_kiocb *req, struct io_kiocb **nxt,
3066 bool force_nonblock)
3068 #if defined(CONFIG_NET)
3069 struct socket *sock;
3072 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3075 sock = sock_from_file(req->file, &ret);
3077 struct io_sr_msg *sr = &req->sr_msg;
3082 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3087 msg.msg_name = NULL;
3088 msg.msg_control = NULL;
3089 msg.msg_controllen = 0;
3090 msg.msg_namelen = 0;
3092 flags = req->sr_msg.msg_flags;
3093 if (flags & MSG_DONTWAIT)
3094 req->flags |= REQ_F_NOWAIT;
3095 else if (force_nonblock)
3096 flags |= MSG_DONTWAIT;
3098 msg.msg_flags = flags;
3099 ret = sock_sendmsg(sock, &msg);
3100 if (force_nonblock && ret == -EAGAIN)
3102 if (ret == -ERESTARTSYS)
3106 io_cqring_add_event(req, ret);
3108 req_set_fail_links(req);
3109 io_put_req_find_next(req, nxt);
3116 static int io_recvmsg_prep(struct io_kiocb *req,
3117 const struct io_uring_sqe *sqe)
3119 #if defined(CONFIG_NET)
3120 struct io_sr_msg *sr = &req->sr_msg;
3121 struct io_async_ctx *io = req->io;
3123 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3124 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3125 sr->len = READ_ONCE(sqe->len);
3127 if (!io || req->opcode == IORING_OP_RECV)
3130 io->msg.iov = io->msg.fast_iov;
3131 return recvmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3132 &io->msg.uaddr, &io->msg.iov);
3138 static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt,
3139 bool force_nonblock)
3141 #if defined(CONFIG_NET)
3142 struct io_async_msghdr *kmsg = NULL;
3143 struct socket *sock;
3146 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3149 sock = sock_from_file(req->file, &ret);
3151 struct io_async_ctx io;
3152 struct sockaddr_storage addr;
3156 kmsg = &req->io->msg;
3157 kmsg->msg.msg_name = &addr;
3158 /* if iov is set, it's allocated already */
3160 kmsg->iov = kmsg->fast_iov;
3161 kmsg->msg.msg_iter.iov = kmsg->iov;
3163 struct io_sr_msg *sr = &req->sr_msg;
3166 kmsg->msg.msg_name = &addr;
3168 io.msg.iov = io.msg.fast_iov;
3169 ret = recvmsg_copy_msghdr(&io.msg.msg, sr->msg,
3170 sr->msg_flags, &io.msg.uaddr,
3176 flags = req->sr_msg.msg_flags;
3177 if (flags & MSG_DONTWAIT)
3178 req->flags |= REQ_F_NOWAIT;
3179 else if (force_nonblock)
3180 flags |= MSG_DONTWAIT;
3182 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3183 kmsg->uaddr, flags);
3184 if (force_nonblock && ret == -EAGAIN) {
3187 if (io_alloc_async_ctx(req))
3189 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
3190 req->work.func = io_sendrecv_async;
3193 if (ret == -ERESTARTSYS)
3197 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
3199 io_cqring_add_event(req, ret);
3201 req_set_fail_links(req);
3202 io_put_req_find_next(req, nxt);
3209 static int io_recv(struct io_kiocb *req, struct io_kiocb **nxt,
3210 bool force_nonblock)
3212 #if defined(CONFIG_NET)
3213 struct socket *sock;
3216 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3219 sock = sock_from_file(req->file, &ret);
3221 struct io_sr_msg *sr = &req->sr_msg;
3226 ret = import_single_range(READ, sr->buf, sr->len, &iov,
3231 msg.msg_name = NULL;
3232 msg.msg_control = NULL;
3233 msg.msg_controllen = 0;
3234 msg.msg_namelen = 0;
3235 msg.msg_iocb = NULL;
3238 flags = req->sr_msg.msg_flags;
3239 if (flags & MSG_DONTWAIT)
3240 req->flags |= REQ_F_NOWAIT;
3241 else if (force_nonblock)
3242 flags |= MSG_DONTWAIT;
3244 ret = sock_recvmsg(sock, &msg, flags);
3245 if (force_nonblock && ret == -EAGAIN)
3247 if (ret == -ERESTARTSYS)
3251 io_cqring_add_event(req, ret);
3253 req_set_fail_links(req);
3254 io_put_req_find_next(req, nxt);
3262 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3264 #if defined(CONFIG_NET)
3265 struct io_accept *accept = &req->accept;
3267 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3269 if (sqe->ioprio || sqe->len || sqe->buf_index)
3272 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3273 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3274 accept->flags = READ_ONCE(sqe->accept_flags);
3281 #if defined(CONFIG_NET)
3282 static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3283 bool force_nonblock)
3285 struct io_accept *accept = &req->accept;
3286 unsigned file_flags;
3289 file_flags = force_nonblock ? O_NONBLOCK : 0;
3290 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3291 accept->addr_len, accept->flags);
3292 if (ret == -EAGAIN && force_nonblock)
3294 if (ret == -ERESTARTSYS)
3297 req_set_fail_links(req);
3298 io_cqring_add_event(req, ret);
3299 io_put_req_find_next(req, nxt);
3303 static void io_accept_finish(struct io_wq_work **workptr)
3305 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3306 struct io_kiocb *nxt = NULL;
3308 if (io_req_cancelled(req))
3310 __io_accept(req, &nxt, false);
3312 io_wq_assign_next(workptr, nxt);
3316 static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
3317 bool force_nonblock)
3319 #if defined(CONFIG_NET)
3322 ret = __io_accept(req, nxt, force_nonblock);
3323 if (ret == -EAGAIN && force_nonblock) {
3324 req->work.func = io_accept_finish;
3334 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3336 #if defined(CONFIG_NET)
3337 struct io_connect *conn = &req->connect;
3338 struct io_async_ctx *io = req->io;
3340 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3342 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3345 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3346 conn->addr_len = READ_ONCE(sqe->addr2);
3351 return move_addr_to_kernel(conn->addr, conn->addr_len,
3352 &io->connect.address);
3358 static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt,
3359 bool force_nonblock)
3361 #if defined(CONFIG_NET)
3362 struct io_async_ctx __io, *io;
3363 unsigned file_flags;
3369 ret = move_addr_to_kernel(req->connect.addr,
3370 req->connect.addr_len,
3371 &__io.connect.address);
3377 file_flags = force_nonblock ? O_NONBLOCK : 0;
3379 ret = __sys_connect_file(req->file, &io->connect.address,
3380 req->connect.addr_len, file_flags);
3381 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
3384 if (io_alloc_async_ctx(req)) {
3388 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
3391 if (ret == -ERESTARTSYS)
3395 req_set_fail_links(req);
3396 io_cqring_add_event(req, ret);
3397 io_put_req_find_next(req, nxt);
3404 static void io_poll_remove_one(struct io_kiocb *req)
3406 struct io_poll_iocb *poll = &req->poll;
3408 spin_lock(&poll->head->lock);
3409 WRITE_ONCE(poll->canceled, true);
3410 if (!list_empty(&poll->wait.entry)) {
3411 list_del_init(&poll->wait.entry);
3412 io_queue_async_work(req);
3414 spin_unlock(&poll->head->lock);
3415 hash_del(&req->hash_node);
3418 static void io_poll_remove_all(struct io_ring_ctx *ctx)
3420 struct hlist_node *tmp;
3421 struct io_kiocb *req;
3424 spin_lock_irq(&ctx->completion_lock);
3425 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
3426 struct hlist_head *list;
3428 list = &ctx->cancel_hash[i];
3429 hlist_for_each_entry_safe(req, tmp, list, hash_node)
3430 io_poll_remove_one(req);
3432 spin_unlock_irq(&ctx->completion_lock);
3435 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
3437 struct hlist_head *list;
3438 struct io_kiocb *req;
3440 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
3441 hlist_for_each_entry(req, list, hash_node) {
3442 if (sqe_addr == req->user_data) {
3443 io_poll_remove_one(req);
3451 static int io_poll_remove_prep(struct io_kiocb *req,
3452 const struct io_uring_sqe *sqe)
3454 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3456 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3460 req->poll.addr = READ_ONCE(sqe->addr);
3465 * Find a running poll command that matches one specified in sqe->addr,
3466 * and remove it if found.
3468 static int io_poll_remove(struct io_kiocb *req)
3470 struct io_ring_ctx *ctx = req->ctx;
3474 addr = req->poll.addr;
3475 spin_lock_irq(&ctx->completion_lock);
3476 ret = io_poll_cancel(ctx, addr);
3477 spin_unlock_irq(&ctx->completion_lock);
3479 io_cqring_add_event(req, ret);
3481 req_set_fail_links(req);
3486 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
3488 struct io_ring_ctx *ctx = req->ctx;
3490 req->poll.done = true;
3492 io_cqring_fill_event(req, error);
3494 io_cqring_fill_event(req, mangle_poll(mask));
3495 io_commit_cqring(ctx);
3498 static void io_poll_complete_work(struct io_wq_work **workptr)
3500 struct io_wq_work *work = *workptr;
3501 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3502 struct io_poll_iocb *poll = &req->poll;
3503 struct poll_table_struct pt = { ._key = poll->events };
3504 struct io_ring_ctx *ctx = req->ctx;
3505 struct io_kiocb *nxt = NULL;
3509 if (work->flags & IO_WQ_WORK_CANCEL) {
3510 WRITE_ONCE(poll->canceled, true);
3512 } else if (READ_ONCE(poll->canceled)) {
3516 if (ret != -ECANCELED)
3517 mask = vfs_poll(poll->file, &pt) & poll->events;
3520 * Note that ->ki_cancel callers also delete iocb from active_reqs after
3521 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
3522 * synchronize with them. In the cancellation case the list_del_init
3523 * itself is not actually needed, but harmless so we keep it in to
3524 * avoid further branches in the fast path.
3526 spin_lock_irq(&ctx->completion_lock);
3527 if (!mask && ret != -ECANCELED) {
3528 add_wait_queue(poll->head, &poll->wait);
3529 spin_unlock_irq(&ctx->completion_lock);
3532 hash_del(&req->hash_node);
3533 io_poll_complete(req, mask, ret);
3534 spin_unlock_irq(&ctx->completion_lock);
3536 io_cqring_ev_posted(ctx);
3539 req_set_fail_links(req);
3540 io_put_req_find_next(req, &nxt);
3542 io_wq_assign_next(workptr, nxt);
3545 static void __io_poll_flush(struct io_ring_ctx *ctx, struct llist_node *nodes)
3547 struct io_kiocb *req, *tmp;
3548 struct req_batch rb;
3550 rb.to_free = rb.need_iter = 0;
3551 spin_lock_irq(&ctx->completion_lock);
3552 llist_for_each_entry_safe(req, tmp, nodes, llist_node) {
3553 hash_del(&req->hash_node);
3554 io_poll_complete(req, req->result, 0);
3556 if (refcount_dec_and_test(&req->refs) &&
3557 !io_req_multi_free(&rb, req)) {
3558 req->flags |= REQ_F_COMP_LOCKED;
3562 spin_unlock_irq(&ctx->completion_lock);
3564 io_cqring_ev_posted(ctx);
3565 io_free_req_many(ctx, &rb);
3568 static void io_poll_flush(struct io_wq_work **workptr)
3570 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3571 struct llist_node *nodes;
3573 nodes = llist_del_all(&req->ctx->poll_llist);
3575 __io_poll_flush(req->ctx, nodes);
3578 static void io_poll_trigger_evfd(struct io_wq_work **workptr)
3580 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3582 eventfd_signal(req->ctx->cq_ev_fd, 1);
3586 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
3589 struct io_poll_iocb *poll = wait->private;
3590 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
3591 struct io_ring_ctx *ctx = req->ctx;
3592 __poll_t mask = key_to_poll(key);
3594 /* for instances that support it check for an event match first: */
3595 if (mask && !(mask & poll->events))
3598 list_del_init(&poll->wait.entry);
3601 * Run completion inline if we can. We're using trylock here because
3602 * we are violating the completion_lock -> poll wq lock ordering.
3603 * If we have a link timeout we're going to need the completion_lock
3604 * for finalizing the request, mark us as having grabbed that already.
3607 unsigned long flags;
3609 if (llist_empty(&ctx->poll_llist) &&
3610 spin_trylock_irqsave(&ctx->completion_lock, flags)) {
3613 hash_del(&req->hash_node);
3614 io_poll_complete(req, mask, 0);
3616 trigger_ev = io_should_trigger_evfd(ctx);
3617 if (trigger_ev && eventfd_signal_count()) {
3619 req->work.func = io_poll_trigger_evfd;
3621 req->flags |= REQ_F_COMP_LOCKED;
3625 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3626 __io_cqring_ev_posted(ctx, trigger_ev);
3629 req->llist_node.next = NULL;
3630 /* if the list wasn't empty, we're done */
3631 if (!llist_add(&req->llist_node, &ctx->poll_llist))
3634 req->work.func = io_poll_flush;
3638 io_queue_async_work(req);
3643 struct io_poll_table {
3644 struct poll_table_struct pt;
3645 struct io_kiocb *req;
3649 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
3650 struct poll_table_struct *p)
3652 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
3654 if (unlikely(pt->req->poll.head)) {
3655 pt->error = -EINVAL;
3660 pt->req->poll.head = head;
3661 add_wait_queue(head, &pt->req->poll.wait);
3664 static void io_poll_req_insert(struct io_kiocb *req)
3666 struct io_ring_ctx *ctx = req->ctx;
3667 struct hlist_head *list;
3669 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
3670 hlist_add_head(&req->hash_node, list);
3673 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3675 struct io_poll_iocb *poll = &req->poll;
3678 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3680 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3685 events = READ_ONCE(sqe->poll_events);
3686 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
3690 static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt)
3692 struct io_poll_iocb *poll = &req->poll;
3693 struct io_ring_ctx *ctx = req->ctx;
3694 struct io_poll_table ipt;
3695 bool cancel = false;
3698 INIT_IO_WORK(&req->work, io_poll_complete_work);
3699 INIT_HLIST_NODE(&req->hash_node);
3703 poll->canceled = false;
3705 ipt.pt._qproc = io_poll_queue_proc;
3706 ipt.pt._key = poll->events;
3708 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
3710 /* initialized the list so that we can do list_empty checks */
3711 INIT_LIST_HEAD(&poll->wait.entry);
3712 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
3713 poll->wait.private = poll;
3715 INIT_LIST_HEAD(&req->list);
3717 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
3719 spin_lock_irq(&ctx->completion_lock);
3720 if (likely(poll->head)) {
3721 spin_lock(&poll->head->lock);
3722 if (unlikely(list_empty(&poll->wait.entry))) {
3728 if (mask || ipt.error)
3729 list_del_init(&poll->wait.entry);
3731 WRITE_ONCE(poll->canceled, true);
3732 else if (!poll->done) /* actually waiting for an event */
3733 io_poll_req_insert(req);
3734 spin_unlock(&poll->head->lock);
3736 if (mask) { /* no async, we'd stolen it */
3738 io_poll_complete(req, mask, 0);
3740 spin_unlock_irq(&ctx->completion_lock);
3743 io_cqring_ev_posted(ctx);
3744 io_put_req_find_next(req, nxt);
3749 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
3751 struct io_timeout_data *data = container_of(timer,
3752 struct io_timeout_data, timer);
3753 struct io_kiocb *req = data->req;
3754 struct io_ring_ctx *ctx = req->ctx;
3755 unsigned long flags;
3757 atomic_inc(&ctx->cq_timeouts);
3759 spin_lock_irqsave(&ctx->completion_lock, flags);
3761 * We could be racing with timeout deletion. If the list is empty,
3762 * then timeout lookup already found it and will be handling it.
3764 if (!list_empty(&req->list)) {
3765 struct io_kiocb *prev;
3768 * Adjust the reqs sequence before the current one because it
3769 * will consume a slot in the cq_ring and the cq_tail
3770 * pointer will be increased, otherwise other timeout reqs may
3771 * return in advance without waiting for enough wait_nr.
3774 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
3776 list_del_init(&req->list);
3779 io_cqring_fill_event(req, -ETIME);
3780 io_commit_cqring(ctx);
3781 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3783 io_cqring_ev_posted(ctx);
3784 req_set_fail_links(req);
3786 return HRTIMER_NORESTART;
3789 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
3791 struct io_kiocb *req;
3794 list_for_each_entry(req, &ctx->timeout_list, list) {
3795 if (user_data == req->user_data) {
3796 list_del_init(&req->list);
3805 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
3809 req_set_fail_links(req);
3810 io_cqring_fill_event(req, -ECANCELED);
3815 static int io_timeout_remove_prep(struct io_kiocb *req,
3816 const struct io_uring_sqe *sqe)
3818 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3820 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
3823 req->timeout.addr = READ_ONCE(sqe->addr);
3824 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
3825 if (req->timeout.flags)
3832 * Remove or update an existing timeout command
3834 static int io_timeout_remove(struct io_kiocb *req)
3836 struct io_ring_ctx *ctx = req->ctx;
3839 spin_lock_irq(&ctx->completion_lock);
3840 ret = io_timeout_cancel(ctx, req->timeout.addr);
3842 io_cqring_fill_event(req, ret);
3843 io_commit_cqring(ctx);
3844 spin_unlock_irq(&ctx->completion_lock);
3845 io_cqring_ev_posted(ctx);
3847 req_set_fail_links(req);
3852 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3853 bool is_timeout_link)
3855 struct io_timeout_data *data;
3858 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3860 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
3862 if (sqe->off && is_timeout_link)
3864 flags = READ_ONCE(sqe->timeout_flags);
3865 if (flags & ~IORING_TIMEOUT_ABS)
3868 req->timeout.count = READ_ONCE(sqe->off);
3870 if (!req->io && io_alloc_async_ctx(req))
3873 data = &req->io->timeout;
3875 req->flags |= REQ_F_TIMEOUT;
3877 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
3880 if (flags & IORING_TIMEOUT_ABS)
3881 data->mode = HRTIMER_MODE_ABS;
3883 data->mode = HRTIMER_MODE_REL;
3885 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
3889 static int io_timeout(struct io_kiocb *req)
3892 struct io_ring_ctx *ctx = req->ctx;
3893 struct io_timeout_data *data;
3894 struct list_head *entry;
3897 data = &req->io->timeout;
3900 * sqe->off holds how many events that need to occur for this
3901 * timeout event to be satisfied. If it isn't set, then this is
3902 * a pure timeout request, sequence isn't used.
3904 count = req->timeout.count;
3906 req->flags |= REQ_F_TIMEOUT_NOSEQ;
3907 spin_lock_irq(&ctx->completion_lock);
3908 entry = ctx->timeout_list.prev;
3912 req->sequence = ctx->cached_sq_head + count - 1;
3913 data->seq_offset = count;
3916 * Insertion sort, ensuring the first entry in the list is always
3917 * the one we need first.
3919 spin_lock_irq(&ctx->completion_lock);
3920 list_for_each_prev(entry, &ctx->timeout_list) {
3921 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
3922 unsigned nxt_sq_head;
3923 long long tmp, tmp_nxt;
3924 u32 nxt_offset = nxt->io->timeout.seq_offset;
3926 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
3930 * Since cached_sq_head + count - 1 can overflow, use type long
3933 tmp = (long long)ctx->cached_sq_head + count - 1;
3934 nxt_sq_head = nxt->sequence - nxt_offset + 1;
3935 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
3938 * cached_sq_head may overflow, and it will never overflow twice
3939 * once there is some timeout req still be valid.
3941 if (ctx->cached_sq_head < nxt_sq_head)
3948 * Sequence of reqs after the insert one and itself should
3949 * be adjusted because each timeout req consumes a slot.
3954 req->sequence -= span;
3956 list_add(&req->list, entry);
3957 data->timer.function = io_timeout_fn;
3958 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
3959 spin_unlock_irq(&ctx->completion_lock);
3963 static bool io_cancel_cb(struct io_wq_work *work, void *data)
3965 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3967 return req->user_data == (unsigned long) data;
3970 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
3972 enum io_wq_cancel cancel_ret;
3975 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
3976 switch (cancel_ret) {
3977 case IO_WQ_CANCEL_OK:
3980 case IO_WQ_CANCEL_RUNNING:
3983 case IO_WQ_CANCEL_NOTFOUND:
3991 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
3992 struct io_kiocb *req, __u64 sqe_addr,
3993 struct io_kiocb **nxt, int success_ret)
3995 unsigned long flags;
3998 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
3999 if (ret != -ENOENT) {
4000 spin_lock_irqsave(&ctx->completion_lock, flags);
4004 spin_lock_irqsave(&ctx->completion_lock, flags);
4005 ret = io_timeout_cancel(ctx, sqe_addr);
4008 ret = io_poll_cancel(ctx, sqe_addr);
4012 io_cqring_fill_event(req, ret);
4013 io_commit_cqring(ctx);
4014 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4015 io_cqring_ev_posted(ctx);
4018 req_set_fail_links(req);
4019 io_put_req_find_next(req, nxt);
4022 static int io_async_cancel_prep(struct io_kiocb *req,
4023 const struct io_uring_sqe *sqe)
4025 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4027 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4031 req->cancel.addr = READ_ONCE(sqe->addr);
4035 static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt)
4037 struct io_ring_ctx *ctx = req->ctx;
4039 io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0);
4043 static int io_files_update_prep(struct io_kiocb *req,
4044 const struct io_uring_sqe *sqe)
4046 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4049 req->files_update.offset = READ_ONCE(sqe->off);
4050 req->files_update.nr_args = READ_ONCE(sqe->len);
4051 if (!req->files_update.nr_args)
4053 req->files_update.arg = READ_ONCE(sqe->addr);
4057 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4059 struct io_ring_ctx *ctx = req->ctx;
4060 struct io_uring_files_update up;
4066 up.offset = req->files_update.offset;
4067 up.fds = req->files_update.arg;
4069 mutex_lock(&ctx->uring_lock);
4070 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4071 mutex_unlock(&ctx->uring_lock);
4074 req_set_fail_links(req);
4075 io_cqring_add_event(req, ret);
4080 static int io_req_defer_prep(struct io_kiocb *req,
4081 const struct io_uring_sqe *sqe)
4085 if (io_op_defs[req->opcode].file_table) {
4086 ret = io_grab_files(req);
4091 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4093 switch (req->opcode) {
4096 case IORING_OP_READV:
4097 case IORING_OP_READ_FIXED:
4098 case IORING_OP_READ:
4099 ret = io_read_prep(req, sqe, true);
4101 case IORING_OP_WRITEV:
4102 case IORING_OP_WRITE_FIXED:
4103 case IORING_OP_WRITE:
4104 ret = io_write_prep(req, sqe, true);
4106 case IORING_OP_POLL_ADD:
4107 ret = io_poll_add_prep(req, sqe);
4109 case IORING_OP_POLL_REMOVE:
4110 ret = io_poll_remove_prep(req, sqe);
4112 case IORING_OP_FSYNC:
4113 ret = io_prep_fsync(req, sqe);
4115 case IORING_OP_SYNC_FILE_RANGE:
4116 ret = io_prep_sfr(req, sqe);
4118 case IORING_OP_SENDMSG:
4119 case IORING_OP_SEND:
4120 ret = io_sendmsg_prep(req, sqe);
4122 case IORING_OP_RECVMSG:
4123 case IORING_OP_RECV:
4124 ret = io_recvmsg_prep(req, sqe);
4126 case IORING_OP_CONNECT:
4127 ret = io_connect_prep(req, sqe);
4129 case IORING_OP_TIMEOUT:
4130 ret = io_timeout_prep(req, sqe, false);
4132 case IORING_OP_TIMEOUT_REMOVE:
4133 ret = io_timeout_remove_prep(req, sqe);
4135 case IORING_OP_ASYNC_CANCEL:
4136 ret = io_async_cancel_prep(req, sqe);
4138 case IORING_OP_LINK_TIMEOUT:
4139 ret = io_timeout_prep(req, sqe, true);
4141 case IORING_OP_ACCEPT:
4142 ret = io_accept_prep(req, sqe);
4144 case IORING_OP_FALLOCATE:
4145 ret = io_fallocate_prep(req, sqe);
4147 case IORING_OP_OPENAT:
4148 ret = io_openat_prep(req, sqe);
4150 case IORING_OP_CLOSE:
4151 ret = io_close_prep(req, sqe);
4153 case IORING_OP_FILES_UPDATE:
4154 ret = io_files_update_prep(req, sqe);
4156 case IORING_OP_STATX:
4157 ret = io_statx_prep(req, sqe);
4159 case IORING_OP_FADVISE:
4160 ret = io_fadvise_prep(req, sqe);
4162 case IORING_OP_MADVISE:
4163 ret = io_madvise_prep(req, sqe);
4165 case IORING_OP_OPENAT2:
4166 ret = io_openat2_prep(req, sqe);
4168 case IORING_OP_EPOLL_CTL:
4169 ret = io_epoll_ctl_prep(req, sqe);
4172 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
4181 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4183 struct io_ring_ctx *ctx = req->ctx;
4186 /* Still need defer if there is pending req in defer list. */
4187 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
4190 if (!req->io && io_alloc_async_ctx(req))
4193 ret = io_req_defer_prep(req, sqe);
4197 spin_lock_irq(&ctx->completion_lock);
4198 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
4199 spin_unlock_irq(&ctx->completion_lock);
4203 trace_io_uring_defer(ctx, req, req->user_data);
4204 list_add_tail(&req->list, &ctx->defer_list);
4205 spin_unlock_irq(&ctx->completion_lock);
4206 return -EIOCBQUEUED;
4209 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4210 struct io_kiocb **nxt, bool force_nonblock)
4212 struct io_ring_ctx *ctx = req->ctx;
4215 switch (req->opcode) {
4219 case IORING_OP_READV:
4220 case IORING_OP_READ_FIXED:
4221 case IORING_OP_READ:
4223 ret = io_read_prep(req, sqe, force_nonblock);
4227 ret = io_read(req, nxt, force_nonblock);
4229 case IORING_OP_WRITEV:
4230 case IORING_OP_WRITE_FIXED:
4231 case IORING_OP_WRITE:
4233 ret = io_write_prep(req, sqe, force_nonblock);
4237 ret = io_write(req, nxt, force_nonblock);
4239 case IORING_OP_FSYNC:
4241 ret = io_prep_fsync(req, sqe);
4245 ret = io_fsync(req, nxt, force_nonblock);
4247 case IORING_OP_POLL_ADD:
4249 ret = io_poll_add_prep(req, sqe);
4253 ret = io_poll_add(req, nxt);
4255 case IORING_OP_POLL_REMOVE:
4257 ret = io_poll_remove_prep(req, sqe);
4261 ret = io_poll_remove(req);
4263 case IORING_OP_SYNC_FILE_RANGE:
4265 ret = io_prep_sfr(req, sqe);
4269 ret = io_sync_file_range(req, nxt, force_nonblock);
4271 case IORING_OP_SENDMSG:
4272 case IORING_OP_SEND:
4274 ret = io_sendmsg_prep(req, sqe);
4278 if (req->opcode == IORING_OP_SENDMSG)
4279 ret = io_sendmsg(req, nxt, force_nonblock);
4281 ret = io_send(req, nxt, force_nonblock);
4283 case IORING_OP_RECVMSG:
4284 case IORING_OP_RECV:
4286 ret = io_recvmsg_prep(req, sqe);
4290 if (req->opcode == IORING_OP_RECVMSG)
4291 ret = io_recvmsg(req, nxt, force_nonblock);
4293 ret = io_recv(req, nxt, force_nonblock);
4295 case IORING_OP_TIMEOUT:
4297 ret = io_timeout_prep(req, sqe, false);
4301 ret = io_timeout(req);
4303 case IORING_OP_TIMEOUT_REMOVE:
4305 ret = io_timeout_remove_prep(req, sqe);
4309 ret = io_timeout_remove(req);
4311 case IORING_OP_ACCEPT:
4313 ret = io_accept_prep(req, sqe);
4317 ret = io_accept(req, nxt, force_nonblock);
4319 case IORING_OP_CONNECT:
4321 ret = io_connect_prep(req, sqe);
4325 ret = io_connect(req, nxt, force_nonblock);
4327 case IORING_OP_ASYNC_CANCEL:
4329 ret = io_async_cancel_prep(req, sqe);
4333 ret = io_async_cancel(req, nxt);
4335 case IORING_OP_FALLOCATE:
4337 ret = io_fallocate_prep(req, sqe);
4341 ret = io_fallocate(req, nxt, force_nonblock);
4343 case IORING_OP_OPENAT:
4345 ret = io_openat_prep(req, sqe);
4349 ret = io_openat(req, nxt, force_nonblock);
4351 case IORING_OP_CLOSE:
4353 ret = io_close_prep(req, sqe);
4357 ret = io_close(req, nxt, force_nonblock);
4359 case IORING_OP_FILES_UPDATE:
4361 ret = io_files_update_prep(req, sqe);
4365 ret = io_files_update(req, force_nonblock);
4367 case IORING_OP_STATX:
4369 ret = io_statx_prep(req, sqe);
4373 ret = io_statx(req, nxt, force_nonblock);
4375 case IORING_OP_FADVISE:
4377 ret = io_fadvise_prep(req, sqe);
4381 ret = io_fadvise(req, nxt, force_nonblock);
4383 case IORING_OP_MADVISE:
4385 ret = io_madvise_prep(req, sqe);
4389 ret = io_madvise(req, nxt, force_nonblock);
4391 case IORING_OP_OPENAT2:
4393 ret = io_openat2_prep(req, sqe);
4397 ret = io_openat2(req, nxt, force_nonblock);
4399 case IORING_OP_EPOLL_CTL:
4401 ret = io_epoll_ctl_prep(req, sqe);
4405 ret = io_epoll_ctl(req, nxt, force_nonblock);
4415 if (ctx->flags & IORING_SETUP_IOPOLL) {
4416 const bool in_async = io_wq_current_is_worker();
4418 if (req->result == -EAGAIN)
4421 /* workqueue context doesn't hold uring_lock, grab it now */
4423 mutex_lock(&ctx->uring_lock);
4425 io_iopoll_req_issued(req);
4428 mutex_unlock(&ctx->uring_lock);
4434 static void io_wq_submit_work(struct io_wq_work **workptr)
4436 struct io_wq_work *work = *workptr;
4437 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4438 struct io_kiocb *nxt = NULL;
4441 /* if NO_CANCEL is set, we must still run the work */
4442 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
4443 IO_WQ_WORK_CANCEL) {
4448 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
4449 req->in_async = true;
4451 ret = io_issue_sqe(req, NULL, &nxt, false);
4453 * We can get EAGAIN for polled IO even though we're
4454 * forcing a sync submission from here, since we can't
4455 * wait for request slots on the block side.
4463 /* drop submission reference */
4467 req_set_fail_links(req);
4468 io_cqring_add_event(req, ret);
4472 /* if a dependent link is ready, pass it back */
4474 io_wq_assign_next(workptr, nxt);
4477 static int io_req_needs_file(struct io_kiocb *req, int fd)
4479 if (!io_op_defs[req->opcode].needs_file)
4481 if (fd == -1 && io_op_defs[req->opcode].fd_non_neg)
4486 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
4489 struct fixed_file_table *table;
4491 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
4492 return table->files[index & IORING_FILE_TABLE_MASK];;
4495 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
4496 const struct io_uring_sqe *sqe)
4498 struct io_ring_ctx *ctx = req->ctx;
4502 flags = READ_ONCE(sqe->flags);
4503 fd = READ_ONCE(sqe->fd);
4505 if (!io_req_needs_file(req, fd))
4508 if (flags & IOSQE_FIXED_FILE) {
4509 if (unlikely(!ctx->file_data ||
4510 (unsigned) fd >= ctx->nr_user_files))
4512 fd = array_index_nospec(fd, ctx->nr_user_files);
4513 req->file = io_file_from_index(ctx, fd);
4516 req->flags |= REQ_F_FIXED_FILE;
4517 percpu_ref_get(&ctx->file_data->refs);
4519 if (req->needs_fixed_file)
4521 trace_io_uring_file_get(ctx, fd);
4522 req->file = io_file_get(state, fd);
4523 if (unlikely(!req->file))
4530 static int io_grab_files(struct io_kiocb *req)
4533 struct io_ring_ctx *ctx = req->ctx;
4535 if (req->work.files)
4537 if (!ctx->ring_file)
4541 spin_lock_irq(&ctx->inflight_lock);
4543 * We use the f_ops->flush() handler to ensure that we can flush
4544 * out work accessing these files if the fd is closed. Check if
4545 * the fd has changed since we started down this path, and disallow
4546 * this operation if it has.
4548 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
4549 list_add(&req->inflight_entry, &ctx->inflight_list);
4550 req->flags |= REQ_F_INFLIGHT;
4551 req->work.files = current->files;
4554 spin_unlock_irq(&ctx->inflight_lock);
4560 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
4562 struct io_timeout_data *data = container_of(timer,
4563 struct io_timeout_data, timer);
4564 struct io_kiocb *req = data->req;
4565 struct io_ring_ctx *ctx = req->ctx;
4566 struct io_kiocb *prev = NULL;
4567 unsigned long flags;
4569 spin_lock_irqsave(&ctx->completion_lock, flags);
4572 * We don't expect the list to be empty, that will only happen if we
4573 * race with the completion of the linked work.
4575 if (!list_empty(&req->link_list)) {
4576 prev = list_entry(req->link_list.prev, struct io_kiocb,
4578 if (refcount_inc_not_zero(&prev->refs)) {
4579 list_del_init(&req->link_list);
4580 prev->flags &= ~REQ_F_LINK_TIMEOUT;
4585 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4588 req_set_fail_links(prev);
4589 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
4593 io_cqring_add_event(req, -ETIME);
4596 return HRTIMER_NORESTART;
4599 static void io_queue_linked_timeout(struct io_kiocb *req)
4601 struct io_ring_ctx *ctx = req->ctx;
4604 * If the list is now empty, then our linked request finished before
4605 * we got a chance to setup the timer
4607 spin_lock_irq(&ctx->completion_lock);
4608 if (!list_empty(&req->link_list)) {
4609 struct io_timeout_data *data = &req->io->timeout;
4611 data->timer.function = io_link_timeout_fn;
4612 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
4615 spin_unlock_irq(&ctx->completion_lock);
4617 /* drop submission reference */
4621 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
4623 struct io_kiocb *nxt;
4625 if (!(req->flags & REQ_F_LINK))
4628 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
4630 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
4633 req->flags |= REQ_F_LINK_TIMEOUT;
4637 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4639 struct io_kiocb *linked_timeout;
4640 struct io_kiocb *nxt = NULL;
4644 linked_timeout = io_prep_linked_timeout(req);
4646 ret = io_issue_sqe(req, sqe, &nxt, true);
4649 * We async punt it if the file wasn't marked NOWAIT, or if the file
4650 * doesn't support non-blocking read/write attempts
4652 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
4653 (req->flags & REQ_F_MUST_PUNT))) {
4655 if (io_op_defs[req->opcode].file_table) {
4656 ret = io_grab_files(req);
4662 * Queued up for async execution, worker will release
4663 * submit reference when the iocb is actually submitted.
4665 io_queue_async_work(req);
4670 /* drop submission reference */
4673 if (linked_timeout) {
4675 io_queue_linked_timeout(linked_timeout);
4677 io_put_req(linked_timeout);
4680 /* and drop final reference, if we failed */
4682 io_cqring_add_event(req, ret);
4683 req_set_fail_links(req);
4691 if (req->flags & REQ_F_FORCE_ASYNC)
4697 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4701 ret = io_req_defer(req, sqe);
4703 if (ret != -EIOCBQUEUED) {
4705 io_cqring_add_event(req, ret);
4706 req_set_fail_links(req);
4707 io_double_put_req(req);
4709 } else if (req->flags & REQ_F_FORCE_ASYNC) {
4710 ret = io_req_defer_prep(req, sqe);
4711 if (unlikely(ret < 0))
4714 * Never try inline submit of IOSQE_ASYNC is set, go straight
4715 * to async execution.
4717 req->work.flags |= IO_WQ_WORK_CONCURRENT;
4718 io_queue_async_work(req);
4720 __io_queue_sqe(req, sqe);
4724 static inline void io_queue_link_head(struct io_kiocb *req)
4726 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
4727 io_cqring_add_event(req, -ECANCELED);
4728 io_double_put_req(req);
4730 io_queue_sqe(req, NULL);
4733 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
4734 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
4736 static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4737 struct io_submit_state *state, struct io_kiocb **link)
4739 const struct cred *old_creds = NULL;
4740 struct io_ring_ctx *ctx = req->ctx;
4741 unsigned int sqe_flags;
4744 sqe_flags = READ_ONCE(sqe->flags);
4746 /* enforce forwards compatibility on users */
4747 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
4752 id = READ_ONCE(sqe->personality);
4754 const struct cred *personality_creds;
4756 personality_creds = idr_find(&ctx->personality_idr, id);
4757 if (unlikely(!personality_creds)) {
4761 old_creds = override_creds(personality_creds);
4764 /* same numerical values with corresponding REQ_F_*, safe to copy */
4765 req->flags |= sqe_flags & (IOSQE_IO_DRAIN|IOSQE_IO_HARDLINK|
4768 ret = io_req_set_file(state, req, sqe);
4769 if (unlikely(ret)) {
4771 io_cqring_add_event(req, ret);
4772 io_double_put_req(req);
4774 revert_creds(old_creds);
4779 * If we already have a head request, queue this one for async
4780 * submittal once the head completes. If we don't have a head but
4781 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
4782 * submitted sync once the chain is complete. If none of those
4783 * conditions are true (normal request), then just queue it.
4786 struct io_kiocb *head = *link;
4789 * Taking sequential execution of a link, draining both sides
4790 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
4791 * requests in the link. So, it drains the head and the
4792 * next after the link request. The last one is done via
4793 * drain_next flag to persist the effect across calls.
4795 if (sqe_flags & IOSQE_IO_DRAIN) {
4796 head->flags |= REQ_F_IO_DRAIN;
4797 ctx->drain_next = 1;
4799 if (io_alloc_async_ctx(req)) {
4804 ret = io_req_defer_prep(req, sqe);
4806 /* fail even hard links since we don't submit */
4807 head->flags |= REQ_F_FAIL_LINK;
4810 trace_io_uring_link(ctx, req, head);
4811 list_add_tail(&req->link_list, &head->link_list);
4813 /* last request of a link, enqueue the link */
4814 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) {
4815 io_queue_link_head(head);
4819 if (unlikely(ctx->drain_next)) {
4820 req->flags |= REQ_F_IO_DRAIN;
4821 req->ctx->drain_next = 0;
4823 if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
4824 req->flags |= REQ_F_LINK;
4825 INIT_LIST_HEAD(&req->link_list);
4826 ret = io_req_defer_prep(req, sqe);
4828 req->flags |= REQ_F_FAIL_LINK;
4831 io_queue_sqe(req, sqe);
4836 revert_creds(old_creds);
4841 * Batched submission is done, ensure local IO is flushed out.
4843 static void io_submit_state_end(struct io_submit_state *state)
4845 blk_finish_plug(&state->plug);
4847 if (state->free_reqs)
4848 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
4852 * Start submission side cache.
4854 static void io_submit_state_start(struct io_submit_state *state,
4855 unsigned int max_ios)
4857 blk_start_plug(&state->plug);
4858 state->free_reqs = 0;
4860 state->ios_left = max_ios;
4863 static void io_commit_sqring(struct io_ring_ctx *ctx)
4865 struct io_rings *rings = ctx->rings;
4868 * Ensure any loads from the SQEs are done at this point,
4869 * since once we write the new head, the application could
4870 * write new data to them.
4872 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
4876 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
4877 * that is mapped by userspace. This means that care needs to be taken to
4878 * ensure that reads are stable, as we cannot rely on userspace always
4879 * being a good citizen. If members of the sqe are validated and then later
4880 * used, it's important that those reads are done through READ_ONCE() to
4881 * prevent a re-load down the line.
4883 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
4884 const struct io_uring_sqe **sqe_ptr)
4886 u32 *sq_array = ctx->sq_array;
4890 * The cached sq head (or cq tail) serves two purposes:
4892 * 1) allows us to batch the cost of updating the user visible
4894 * 2) allows the kernel side to track the head on its own, even
4895 * though the application is the one updating it.
4897 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
4898 if (likely(head < ctx->sq_entries)) {
4900 * All io need record the previous position, if LINK vs DARIN,
4901 * it can be used to mark the position of the first IO in the
4904 req->sequence = ctx->cached_sq_head;
4905 *sqe_ptr = &ctx->sq_sqes[head];
4906 req->opcode = READ_ONCE((*sqe_ptr)->opcode);
4907 req->user_data = READ_ONCE((*sqe_ptr)->user_data);
4908 ctx->cached_sq_head++;
4912 /* drop invalid entries */
4913 ctx->cached_sq_head++;
4914 ctx->cached_sq_dropped++;
4915 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
4919 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
4920 struct file *ring_file, int ring_fd,
4921 struct mm_struct **mm, bool async)
4923 struct io_submit_state state, *statep = NULL;
4924 struct io_kiocb *link = NULL;
4925 int i, submitted = 0;
4926 bool mm_fault = false;
4928 /* if we have a backlog and couldn't flush it all, return BUSY */
4929 if (test_bit(0, &ctx->sq_check_overflow)) {
4930 if (!list_empty(&ctx->cq_overflow_list) &&
4931 !io_cqring_overflow_flush(ctx, false))
4935 /* make sure SQ entry isn't read before tail */
4936 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
4938 if (!percpu_ref_tryget_many(&ctx->refs, nr))
4941 if (nr > IO_PLUG_THRESHOLD) {
4942 io_submit_state_start(&state, nr);
4946 ctx->ring_fd = ring_fd;
4947 ctx->ring_file = ring_file;
4949 for (i = 0; i < nr; i++) {
4950 const struct io_uring_sqe *sqe;
4951 struct io_kiocb *req;
4953 req = io_get_req(ctx, statep);
4954 if (unlikely(!req)) {
4956 submitted = -EAGAIN;
4959 if (!io_get_sqring(ctx, req, &sqe)) {
4960 __io_req_do_free(req);
4964 /* will complete beyond this point, count as submitted */
4967 if (unlikely(req->opcode >= IORING_OP_LAST)) {
4968 io_cqring_add_event(req, -EINVAL);
4969 io_double_put_req(req);
4973 if (io_op_defs[req->opcode].needs_mm && !*mm) {
4974 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
4976 use_mm(ctx->sqo_mm);
4981 req->has_user = *mm != NULL;
4982 req->in_async = async;
4983 req->needs_fixed_file = async;
4984 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
4986 if (!io_submit_sqe(req, sqe, statep, &link))
4990 if (unlikely(submitted != nr)) {
4991 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
4993 percpu_ref_put_many(&ctx->refs, nr - ref_used);
4996 io_queue_link_head(link);
4998 io_submit_state_end(&state);
5000 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5001 io_commit_sqring(ctx);
5006 static int io_sq_thread(void *data)
5008 struct io_ring_ctx *ctx = data;
5009 struct mm_struct *cur_mm = NULL;
5010 const struct cred *old_cred;
5011 mm_segment_t old_fs;
5014 unsigned long timeout;
5017 complete(&ctx->completions[1]);
5021 old_cred = override_creds(ctx->creds);
5023 ret = timeout = inflight = 0;
5024 while (!kthread_should_park()) {
5025 unsigned int to_submit;
5028 unsigned nr_events = 0;
5030 if (ctx->flags & IORING_SETUP_IOPOLL) {
5032 * inflight is the count of the maximum possible
5033 * entries we submitted, but it can be smaller
5034 * if we dropped some of them. If we don't have
5035 * poll entries available, then we know that we
5036 * have nothing left to poll for. Reset the
5037 * inflight count to zero in that case.
5039 mutex_lock(&ctx->uring_lock);
5040 if (!list_empty(&ctx->poll_list))
5041 __io_iopoll_check(ctx, &nr_events, 0);
5044 mutex_unlock(&ctx->uring_lock);
5047 * Normal IO, just pretend everything completed.
5048 * We don't have to poll completions for that.
5050 nr_events = inflight;
5053 inflight -= nr_events;
5055 timeout = jiffies + ctx->sq_thread_idle;
5058 to_submit = io_sqring_entries(ctx);
5061 * If submit got -EBUSY, flag us as needing the application
5062 * to enter the kernel to reap and flush events.
5064 if (!to_submit || ret == -EBUSY) {
5066 * We're polling. If we're within the defined idle
5067 * period, then let us spin without work before going
5068 * to sleep. The exception is if we got EBUSY doing
5069 * more IO, we should wait for the application to
5070 * reap events and wake us up.
5073 (!time_after(jiffies, timeout) && ret != -EBUSY &&
5074 !percpu_ref_is_dying(&ctx->refs))) {
5080 * Drop cur_mm before scheduling, we can't hold it for
5081 * long periods (or over schedule()). Do this before
5082 * adding ourselves to the waitqueue, as the unuse/drop
5091 prepare_to_wait(&ctx->sqo_wait, &wait,
5092 TASK_INTERRUPTIBLE);
5094 /* Tell userspace we may need a wakeup call */
5095 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
5096 /* make sure to read SQ tail after writing flags */
5099 to_submit = io_sqring_entries(ctx);
5100 if (!to_submit || ret == -EBUSY) {
5101 if (kthread_should_park()) {
5102 finish_wait(&ctx->sqo_wait, &wait);
5105 if (signal_pending(current))
5106 flush_signals(current);
5108 finish_wait(&ctx->sqo_wait, &wait);
5110 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5113 finish_wait(&ctx->sqo_wait, &wait);
5115 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
5118 mutex_lock(&ctx->uring_lock);
5119 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
5120 mutex_unlock(&ctx->uring_lock);
5130 revert_creds(old_cred);
5137 struct io_wait_queue {
5138 struct wait_queue_entry wq;
5139 struct io_ring_ctx *ctx;
5141 unsigned nr_timeouts;
5144 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
5146 struct io_ring_ctx *ctx = iowq->ctx;
5149 * Wake up if we have enough events, or if a timeout occurred since we
5150 * started waiting. For timeouts, we always want to return to userspace,
5151 * regardless of event count.
5153 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
5154 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
5157 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
5158 int wake_flags, void *key)
5160 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
5163 /* use noflush == true, as we can't safely rely on locking context */
5164 if (!io_should_wake(iowq, true))
5167 return autoremove_wake_function(curr, mode, wake_flags, key);
5171 * Wait until events become available, if we don't already have some. The
5172 * application must reap them itself, as they reside on the shared cq ring.
5174 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
5175 const sigset_t __user *sig, size_t sigsz)
5177 struct io_wait_queue iowq = {
5180 .func = io_wake_function,
5181 .entry = LIST_HEAD_INIT(iowq.wq.entry),
5184 .to_wait = min_events,
5186 struct io_rings *rings = ctx->rings;
5189 if (io_cqring_events(ctx, false) >= min_events)
5193 #ifdef CONFIG_COMPAT
5194 if (in_compat_syscall())
5195 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
5199 ret = set_user_sigmask(sig, sigsz);
5205 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
5206 trace_io_uring_cqring_wait(ctx, min_events);
5208 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
5209 TASK_INTERRUPTIBLE);
5210 if (io_should_wake(&iowq, false))
5213 if (signal_pending(current)) {
5218 finish_wait(&ctx->wait, &iowq.wq);
5220 restore_saved_sigmask_unless(ret == -EINTR);
5222 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
5225 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
5227 #if defined(CONFIG_UNIX)
5228 if (ctx->ring_sock) {
5229 struct sock *sock = ctx->ring_sock->sk;
5230 struct sk_buff *skb;
5232 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
5238 for (i = 0; i < ctx->nr_user_files; i++) {
5241 file = io_file_from_index(ctx, i);
5248 static void io_file_ref_kill(struct percpu_ref *ref)
5250 struct fixed_file_data *data;
5252 data = container_of(ref, struct fixed_file_data, refs);
5253 complete(&data->done);
5256 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
5258 struct fixed_file_data *data = ctx->file_data;
5259 unsigned nr_tables, i;
5264 /* protect against inflight atomic switch, which drops the ref */
5265 percpu_ref_get(&data->refs);
5266 /* wait for existing switches */
5267 flush_work(&data->ref_work);
5268 percpu_ref_kill_and_confirm(&data->refs, io_file_ref_kill);
5269 wait_for_completion(&data->done);
5270 percpu_ref_put(&data->refs);
5271 /* flush potential new switch */
5272 flush_work(&data->ref_work);
5273 percpu_ref_exit(&data->refs);
5275 __io_sqe_files_unregister(ctx);
5276 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
5277 for (i = 0; i < nr_tables; i++)
5278 kfree(data->table[i].files);
5281 ctx->file_data = NULL;
5282 ctx->nr_user_files = 0;
5286 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
5288 if (ctx->sqo_thread) {
5289 wait_for_completion(&ctx->completions[1]);
5291 * The park is a bit of a work-around, without it we get
5292 * warning spews on shutdown with SQPOLL set and affinity
5293 * set to a single CPU.
5295 kthread_park(ctx->sqo_thread);
5296 kthread_stop(ctx->sqo_thread);
5297 ctx->sqo_thread = NULL;
5301 static void io_finish_async(struct io_ring_ctx *ctx)
5303 io_sq_thread_stop(ctx);
5306 io_wq_destroy(ctx->io_wq);
5311 #if defined(CONFIG_UNIX)
5313 * Ensure the UNIX gc is aware of our file set, so we are certain that
5314 * the io_uring can be safely unregistered on process exit, even if we have
5315 * loops in the file referencing.
5317 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
5319 struct sock *sk = ctx->ring_sock->sk;
5320 struct scm_fp_list *fpl;
5321 struct sk_buff *skb;
5324 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
5325 unsigned long inflight = ctx->user->unix_inflight + nr;
5327 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
5331 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
5335 skb = alloc_skb(0, GFP_KERNEL);
5344 fpl->user = get_uid(ctx->user);
5345 for (i = 0; i < nr; i++) {
5346 struct file *file = io_file_from_index(ctx, i + offset);
5350 fpl->fp[nr_files] = get_file(file);
5351 unix_inflight(fpl->user, fpl->fp[nr_files]);
5356 fpl->max = SCM_MAX_FD;
5357 fpl->count = nr_files;
5358 UNIXCB(skb).fp = fpl;
5359 skb->destructor = unix_destruct_scm;
5360 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
5361 skb_queue_head(&sk->sk_receive_queue, skb);
5363 for (i = 0; i < nr_files; i++)
5374 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
5375 * causes regular reference counting to break down. We rely on the UNIX
5376 * garbage collection to take care of this problem for us.
5378 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5380 unsigned left, total;
5384 left = ctx->nr_user_files;
5386 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
5388 ret = __io_sqe_files_scm(ctx, this_files, total);
5392 total += this_files;
5398 while (total < ctx->nr_user_files) {
5399 struct file *file = io_file_from_index(ctx, total);
5409 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
5415 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
5420 for (i = 0; i < nr_tables; i++) {
5421 struct fixed_file_table *table = &ctx->file_data->table[i];
5422 unsigned this_files;
5424 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
5425 table->files = kcalloc(this_files, sizeof(struct file *),
5429 nr_files -= this_files;
5435 for (i = 0; i < nr_tables; i++) {
5436 struct fixed_file_table *table = &ctx->file_data->table[i];
5437 kfree(table->files);
5442 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
5444 #if defined(CONFIG_UNIX)
5445 struct sock *sock = ctx->ring_sock->sk;
5446 struct sk_buff_head list, *head = &sock->sk_receive_queue;
5447 struct sk_buff *skb;
5450 __skb_queue_head_init(&list);
5453 * Find the skb that holds this file in its SCM_RIGHTS. When found,
5454 * remove this entry and rearrange the file array.
5456 skb = skb_dequeue(head);
5458 struct scm_fp_list *fp;
5460 fp = UNIXCB(skb).fp;
5461 for (i = 0; i < fp->count; i++) {
5464 if (fp->fp[i] != file)
5467 unix_notinflight(fp->user, fp->fp[i]);
5468 left = fp->count - 1 - i;
5470 memmove(&fp->fp[i], &fp->fp[i + 1],
5471 left * sizeof(struct file *));
5478 __skb_queue_tail(&list, skb);
5488 __skb_queue_tail(&list, skb);
5490 skb = skb_dequeue(head);
5493 if (skb_peek(&list)) {
5494 spin_lock_irq(&head->lock);
5495 while ((skb = __skb_dequeue(&list)) != NULL)
5496 __skb_queue_tail(head, skb);
5497 spin_unlock_irq(&head->lock);
5504 struct io_file_put {
5505 struct llist_node llist;
5507 struct completion *done;
5510 static void io_ring_file_ref_switch(struct work_struct *work)
5512 struct io_file_put *pfile, *tmp;
5513 struct fixed_file_data *data;
5514 struct llist_node *node;
5516 data = container_of(work, struct fixed_file_data, ref_work);
5518 while ((node = llist_del_all(&data->put_llist)) != NULL) {
5519 llist_for_each_entry_safe(pfile, tmp, node, llist) {
5520 io_ring_file_put(data->ctx, pfile->file);
5522 complete(pfile->done);
5528 percpu_ref_get(&data->refs);
5529 percpu_ref_switch_to_percpu(&data->refs);
5532 static void io_file_data_ref_zero(struct percpu_ref *ref)
5534 struct fixed_file_data *data;
5536 data = container_of(ref, struct fixed_file_data, refs);
5538 /* we can't safely switch from inside this context, punt to wq */
5539 queue_work(system_wq, &data->ref_work);
5542 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
5545 __s32 __user *fds = (__s32 __user *) arg;
5555 if (nr_args > IORING_MAX_FIXED_FILES)
5558 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
5559 if (!ctx->file_data)
5561 ctx->file_data->ctx = ctx;
5562 init_completion(&ctx->file_data->done);
5564 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
5565 ctx->file_data->table = kcalloc(nr_tables,
5566 sizeof(struct fixed_file_table),
5568 if (!ctx->file_data->table) {
5569 kfree(ctx->file_data);
5570 ctx->file_data = NULL;
5574 if (percpu_ref_init(&ctx->file_data->refs, io_file_data_ref_zero,
5575 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
5576 kfree(ctx->file_data->table);
5577 kfree(ctx->file_data);
5578 ctx->file_data = NULL;
5581 ctx->file_data->put_llist.first = NULL;
5582 INIT_WORK(&ctx->file_data->ref_work, io_ring_file_ref_switch);
5584 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
5585 percpu_ref_exit(&ctx->file_data->refs);
5586 kfree(ctx->file_data->table);
5587 kfree(ctx->file_data);
5588 ctx->file_data = NULL;
5592 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
5593 struct fixed_file_table *table;
5597 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
5599 /* allow sparse sets */
5605 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5606 index = i & IORING_FILE_TABLE_MASK;
5614 * Don't allow io_uring instances to be registered. If UNIX
5615 * isn't enabled, then this causes a reference cycle and this
5616 * instance can never get freed. If UNIX is enabled we'll
5617 * handle it just fine, but there's still no point in allowing
5618 * a ring fd as it doesn't support regular read/write anyway.
5620 if (file->f_op == &io_uring_fops) {
5625 table->files[index] = file;
5629 for (i = 0; i < ctx->nr_user_files; i++) {
5630 file = io_file_from_index(ctx, i);
5634 for (i = 0; i < nr_tables; i++)
5635 kfree(ctx->file_data->table[i].files);
5637 kfree(ctx->file_data->table);
5638 kfree(ctx->file_data);
5639 ctx->file_data = NULL;
5640 ctx->nr_user_files = 0;
5644 ret = io_sqe_files_scm(ctx);
5646 io_sqe_files_unregister(ctx);
5651 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
5654 #if defined(CONFIG_UNIX)
5655 struct sock *sock = ctx->ring_sock->sk;
5656 struct sk_buff_head *head = &sock->sk_receive_queue;
5657 struct sk_buff *skb;
5660 * See if we can merge this file into an existing skb SCM_RIGHTS
5661 * file set. If there's no room, fall back to allocating a new skb
5662 * and filling it in.
5664 spin_lock_irq(&head->lock);
5665 skb = skb_peek(head);
5667 struct scm_fp_list *fpl = UNIXCB(skb).fp;
5669 if (fpl->count < SCM_MAX_FD) {
5670 __skb_unlink(skb, head);
5671 spin_unlock_irq(&head->lock);
5672 fpl->fp[fpl->count] = get_file(file);
5673 unix_inflight(fpl->user, fpl->fp[fpl->count]);
5675 spin_lock_irq(&head->lock);
5676 __skb_queue_head(head, skb);
5681 spin_unlock_irq(&head->lock);
5688 return __io_sqe_files_scm(ctx, 1, index);
5694 static void io_atomic_switch(struct percpu_ref *ref)
5696 struct fixed_file_data *data;
5698 data = container_of(ref, struct fixed_file_data, refs);
5699 clear_bit(FFD_F_ATOMIC, &data->state);
5702 static bool io_queue_file_removal(struct fixed_file_data *data,
5705 struct io_file_put *pfile, pfile_stack;
5706 DECLARE_COMPLETION_ONSTACK(done);
5709 * If we fail allocating the struct we need for doing async reomval
5710 * of this file, just punt to sync and wait for it.
5712 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
5714 pfile = &pfile_stack;
5715 pfile->done = &done;
5719 llist_add(&pfile->llist, &data->put_llist);
5721 if (pfile == &pfile_stack) {
5722 if (!test_and_set_bit(FFD_F_ATOMIC, &data->state)) {
5723 percpu_ref_put(&data->refs);
5724 percpu_ref_switch_to_atomic(&data->refs,
5727 wait_for_completion(&done);
5728 flush_work(&data->ref_work);
5735 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
5736 struct io_uring_files_update *up,
5739 struct fixed_file_data *data = ctx->file_data;
5740 bool ref_switch = false;
5746 if (check_add_overflow(up->offset, nr_args, &done))
5748 if (done > ctx->nr_user_files)
5752 fds = u64_to_user_ptr(up->fds);
5754 struct fixed_file_table *table;
5758 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
5762 i = array_index_nospec(up->offset, ctx->nr_user_files);
5763 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
5764 index = i & IORING_FILE_TABLE_MASK;
5765 if (table->files[index]) {
5766 file = io_file_from_index(ctx, index);
5767 table->files[index] = NULL;
5768 if (io_queue_file_removal(data, file))
5778 * Don't allow io_uring instances to be registered. If
5779 * UNIX isn't enabled, then this causes a reference
5780 * cycle and this instance can never get freed. If UNIX
5781 * is enabled we'll handle it just fine, but there's
5782 * still no point in allowing a ring fd as it doesn't
5783 * support regular read/write anyway.
5785 if (file->f_op == &io_uring_fops) {
5790 table->files[index] = file;
5791 err = io_sqe_file_register(ctx, file, i);
5800 if (ref_switch && !test_and_set_bit(FFD_F_ATOMIC, &data->state)) {
5801 percpu_ref_put(&data->refs);
5802 percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch);
5805 return done ? done : err;
5807 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
5810 struct io_uring_files_update up;
5812 if (!ctx->file_data)
5816 if (copy_from_user(&up, arg, sizeof(up)))
5821 return __io_sqe_files_update(ctx, &up, nr_args);
5824 static void io_put_work(struct io_wq_work *work)
5826 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5831 static void io_get_work(struct io_wq_work *work)
5833 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5835 refcount_inc(&req->refs);
5838 static int io_init_wq_offload(struct io_ring_ctx *ctx,
5839 struct io_uring_params *p)
5841 struct io_wq_data data;
5843 struct io_ring_ctx *ctx_attach;
5844 unsigned int concurrency;
5847 data.user = ctx->user;
5848 data.get_work = io_get_work;
5849 data.put_work = io_put_work;
5851 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
5852 /* Do QD, or 4 * CPUS, whatever is smallest */
5853 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
5855 ctx->io_wq = io_wq_create(concurrency, &data);
5856 if (IS_ERR(ctx->io_wq)) {
5857 ret = PTR_ERR(ctx->io_wq);
5863 f = fdget(p->wq_fd);
5867 if (f.file->f_op != &io_uring_fops) {
5872 ctx_attach = f.file->private_data;
5873 /* @io_wq is protected by holding the fd */
5874 if (!io_wq_get(ctx_attach->io_wq, &data)) {
5879 ctx->io_wq = ctx_attach->io_wq;
5885 static int io_sq_offload_start(struct io_ring_ctx *ctx,
5886 struct io_uring_params *p)
5890 init_waitqueue_head(&ctx->sqo_wait);
5891 mmgrab(current->mm);
5892 ctx->sqo_mm = current->mm;
5894 if (ctx->flags & IORING_SETUP_SQPOLL) {
5896 if (!capable(CAP_SYS_ADMIN))
5899 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
5900 if (!ctx->sq_thread_idle)
5901 ctx->sq_thread_idle = HZ;
5903 if (p->flags & IORING_SETUP_SQ_AFF) {
5904 int cpu = p->sq_thread_cpu;
5907 if (cpu >= nr_cpu_ids)
5909 if (!cpu_online(cpu))
5912 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
5916 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
5919 if (IS_ERR(ctx->sqo_thread)) {
5920 ret = PTR_ERR(ctx->sqo_thread);
5921 ctx->sqo_thread = NULL;
5924 wake_up_process(ctx->sqo_thread);
5925 } else if (p->flags & IORING_SETUP_SQ_AFF) {
5926 /* Can't have SQ_AFF without SQPOLL */
5931 ret = io_init_wq_offload(ctx, p);
5937 io_finish_async(ctx);
5938 mmdrop(ctx->sqo_mm);
5943 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
5945 atomic_long_sub(nr_pages, &user->locked_vm);
5948 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
5950 unsigned long page_limit, cur_pages, new_pages;
5952 /* Don't allow more pages than we can safely lock */
5953 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
5956 cur_pages = atomic_long_read(&user->locked_vm);
5957 new_pages = cur_pages + nr_pages;
5958 if (new_pages > page_limit)
5960 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
5961 new_pages) != cur_pages);
5966 static void io_mem_free(void *ptr)
5973 page = virt_to_head_page(ptr);
5974 if (put_page_testzero(page))
5975 free_compound_page(page);
5978 static void *io_mem_alloc(size_t size)
5980 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
5983 return (void *) __get_free_pages(gfp_flags, get_order(size));
5986 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
5989 struct io_rings *rings;
5990 size_t off, sq_array_size;
5992 off = struct_size(rings, cqes, cq_entries);
5993 if (off == SIZE_MAX)
5997 off = ALIGN(off, SMP_CACHE_BYTES);
6002 sq_array_size = array_size(sizeof(u32), sq_entries);
6003 if (sq_array_size == SIZE_MAX)
6006 if (check_add_overflow(off, sq_array_size, &off))
6015 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6019 pages = (size_t)1 << get_order(
6020 rings_size(sq_entries, cq_entries, NULL));
6021 pages += (size_t)1 << get_order(
6022 array_size(sizeof(struct io_uring_sqe), sq_entries));
6027 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6031 if (!ctx->user_bufs)
6034 for (i = 0; i < ctx->nr_user_bufs; i++) {
6035 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6037 for (j = 0; j < imu->nr_bvecs; j++)
6038 put_user_page(imu->bvec[j].bv_page);
6040 if (ctx->account_mem)
6041 io_unaccount_mem(ctx->user, imu->nr_bvecs);
6046 kfree(ctx->user_bufs);
6047 ctx->user_bufs = NULL;
6048 ctx->nr_user_bufs = 0;
6052 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
6053 void __user *arg, unsigned index)
6055 struct iovec __user *src;
6057 #ifdef CONFIG_COMPAT
6059 struct compat_iovec __user *ciovs;
6060 struct compat_iovec ciov;
6062 ciovs = (struct compat_iovec __user *) arg;
6063 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
6066 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
6067 dst->iov_len = ciov.iov_len;
6071 src = (struct iovec __user *) arg;
6072 if (copy_from_user(dst, &src[index], sizeof(*dst)))
6077 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
6080 struct vm_area_struct **vmas = NULL;
6081 struct page **pages = NULL;
6082 int i, j, got_pages = 0;
6087 if (!nr_args || nr_args > UIO_MAXIOV)
6090 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
6092 if (!ctx->user_bufs)
6095 for (i = 0; i < nr_args; i++) {
6096 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6097 unsigned long off, start, end, ubuf;
6102 ret = io_copy_iov(ctx, &iov, arg, i);
6107 * Don't impose further limits on the size and buffer
6108 * constraints here, we'll -EINVAL later when IO is
6109 * submitted if they are wrong.
6112 if (!iov.iov_base || !iov.iov_len)
6115 /* arbitrary limit, but we need something */
6116 if (iov.iov_len > SZ_1G)
6119 ubuf = (unsigned long) iov.iov_base;
6120 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
6121 start = ubuf >> PAGE_SHIFT;
6122 nr_pages = end - start;
6124 if (ctx->account_mem) {
6125 ret = io_account_mem(ctx->user, nr_pages);
6131 if (!pages || nr_pages > got_pages) {
6134 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
6136 vmas = kvmalloc_array(nr_pages,
6137 sizeof(struct vm_area_struct *),
6139 if (!pages || !vmas) {
6141 if (ctx->account_mem)
6142 io_unaccount_mem(ctx->user, nr_pages);
6145 got_pages = nr_pages;
6148 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
6152 if (ctx->account_mem)
6153 io_unaccount_mem(ctx->user, nr_pages);
6158 down_read(¤t->mm->mmap_sem);
6159 pret = get_user_pages(ubuf, nr_pages,
6160 FOLL_WRITE | FOLL_LONGTERM,
6162 if (pret == nr_pages) {
6163 /* don't support file backed memory */
6164 for (j = 0; j < nr_pages; j++) {
6165 struct vm_area_struct *vma = vmas[j];
6168 !is_file_hugepages(vma->vm_file)) {
6174 ret = pret < 0 ? pret : -EFAULT;
6176 up_read(¤t->mm->mmap_sem);
6179 * if we did partial map, or found file backed vmas,
6180 * release any pages we did get
6183 put_user_pages(pages, pret);
6184 if (ctx->account_mem)
6185 io_unaccount_mem(ctx->user, nr_pages);
6190 off = ubuf & ~PAGE_MASK;
6192 for (j = 0; j < nr_pages; j++) {
6195 vec_len = min_t(size_t, size, PAGE_SIZE - off);
6196 imu->bvec[j].bv_page = pages[j];
6197 imu->bvec[j].bv_len = vec_len;
6198 imu->bvec[j].bv_offset = off;
6202 /* store original address for later verification */
6204 imu->len = iov.iov_len;
6205 imu->nr_bvecs = nr_pages;
6207 ctx->nr_user_bufs++;
6215 io_sqe_buffer_unregister(ctx);
6219 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
6221 __s32 __user *fds = arg;
6227 if (copy_from_user(&fd, fds, sizeof(*fds)))
6230 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
6231 if (IS_ERR(ctx->cq_ev_fd)) {
6232 int ret = PTR_ERR(ctx->cq_ev_fd);
6233 ctx->cq_ev_fd = NULL;
6240 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
6242 if (ctx->cq_ev_fd) {
6243 eventfd_ctx_put(ctx->cq_ev_fd);
6244 ctx->cq_ev_fd = NULL;
6251 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
6253 io_finish_async(ctx);
6255 mmdrop(ctx->sqo_mm);
6257 io_iopoll_reap_events(ctx);
6258 io_sqe_buffer_unregister(ctx);
6259 io_sqe_files_unregister(ctx);
6260 io_eventfd_unregister(ctx);
6262 #if defined(CONFIG_UNIX)
6263 if (ctx->ring_sock) {
6264 ctx->ring_sock->file = NULL; /* so that iput() is called */
6265 sock_release(ctx->ring_sock);
6269 io_mem_free(ctx->rings);
6270 io_mem_free(ctx->sq_sqes);
6272 percpu_ref_exit(&ctx->refs);
6273 if (ctx->account_mem)
6274 io_unaccount_mem(ctx->user,
6275 ring_pages(ctx->sq_entries, ctx->cq_entries));
6276 free_uid(ctx->user);
6277 put_cred(ctx->creds);
6278 kfree(ctx->completions);
6279 kfree(ctx->cancel_hash);
6280 kmem_cache_free(req_cachep, ctx->fallback_req);
6284 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
6286 struct io_ring_ctx *ctx = file->private_data;
6289 poll_wait(file, &ctx->cq_wait, wait);
6291 * synchronizes with barrier from wq_has_sleeper call in
6295 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
6296 ctx->rings->sq_ring_entries)
6297 mask |= EPOLLOUT | EPOLLWRNORM;
6298 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
6299 mask |= EPOLLIN | EPOLLRDNORM;
6304 static int io_uring_fasync(int fd, struct file *file, int on)
6306 struct io_ring_ctx *ctx = file->private_data;
6308 return fasync_helper(fd, file, on, &ctx->cq_fasync);
6311 static int io_remove_personalities(int id, void *p, void *data)
6313 struct io_ring_ctx *ctx = data;
6314 const struct cred *cred;
6316 cred = idr_remove(&ctx->personality_idr, id);
6322 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
6324 mutex_lock(&ctx->uring_lock);
6325 percpu_ref_kill(&ctx->refs);
6326 mutex_unlock(&ctx->uring_lock);
6329 * Wait for sq thread to idle, if we have one. It won't spin on new
6330 * work after we've killed the ctx ref above. This is important to do
6331 * before we cancel existing commands, as the thread could otherwise
6332 * be queueing new work post that. If that's work we need to cancel,
6333 * it could cause shutdown to hang.
6335 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
6338 io_kill_timeouts(ctx);
6339 io_poll_remove_all(ctx);
6342 io_wq_cancel_all(ctx->io_wq);
6344 io_iopoll_reap_events(ctx);
6345 /* if we failed setting up the ctx, we might not have any rings */
6347 io_cqring_overflow_flush(ctx, true);
6348 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
6349 wait_for_completion(&ctx->completions[0]);
6350 io_ring_ctx_free(ctx);
6353 static int io_uring_release(struct inode *inode, struct file *file)
6355 struct io_ring_ctx *ctx = file->private_data;
6357 file->private_data = NULL;
6358 io_ring_ctx_wait_and_kill(ctx);
6362 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
6363 struct files_struct *files)
6365 struct io_kiocb *req;
6368 while (!list_empty_careful(&ctx->inflight_list)) {
6369 struct io_kiocb *cancel_req = NULL;
6371 spin_lock_irq(&ctx->inflight_lock);
6372 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
6373 if (req->work.files != files)
6375 /* req is being completed, ignore */
6376 if (!refcount_inc_not_zero(&req->refs))
6382 prepare_to_wait(&ctx->inflight_wait, &wait,
6383 TASK_UNINTERRUPTIBLE);
6384 spin_unlock_irq(&ctx->inflight_lock);
6386 /* We need to keep going until we don't find a matching req */
6390 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
6391 io_put_req(cancel_req);
6394 finish_wait(&ctx->inflight_wait, &wait);
6397 static int io_uring_flush(struct file *file, void *data)
6399 struct io_ring_ctx *ctx = file->private_data;
6401 io_uring_cancel_files(ctx, data);
6405 static void *io_uring_validate_mmap_request(struct file *file,
6406 loff_t pgoff, size_t sz)
6408 struct io_ring_ctx *ctx = file->private_data;
6409 loff_t offset = pgoff << PAGE_SHIFT;
6414 case IORING_OFF_SQ_RING:
6415 case IORING_OFF_CQ_RING:
6418 case IORING_OFF_SQES:
6422 return ERR_PTR(-EINVAL);
6425 page = virt_to_head_page(ptr);
6426 if (sz > page_size(page))
6427 return ERR_PTR(-EINVAL);
6434 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6436 size_t sz = vma->vm_end - vma->vm_start;
6440 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
6442 return PTR_ERR(ptr);
6444 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
6445 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
6448 #else /* !CONFIG_MMU */
6450 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6452 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
6455 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
6457 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
6460 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
6461 unsigned long addr, unsigned long len,
6462 unsigned long pgoff, unsigned long flags)
6466 ptr = io_uring_validate_mmap_request(file, pgoff, len);
6468 return PTR_ERR(ptr);
6470 return (unsigned long) ptr;
6473 #endif /* !CONFIG_MMU */
6475 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
6476 u32, min_complete, u32, flags, const sigset_t __user *, sig,
6479 struct io_ring_ctx *ctx;
6484 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
6492 if (f.file->f_op != &io_uring_fops)
6496 ctx = f.file->private_data;
6497 if (!percpu_ref_tryget(&ctx->refs))
6501 * For SQ polling, the thread will do all submissions and completions.
6502 * Just return the requested submit count, and wake the thread if
6506 if (ctx->flags & IORING_SETUP_SQPOLL) {
6507 if (!list_empty_careful(&ctx->cq_overflow_list))
6508 io_cqring_overflow_flush(ctx, false);
6509 if (flags & IORING_ENTER_SQ_WAKEUP)
6510 wake_up(&ctx->sqo_wait);
6511 submitted = to_submit;
6512 } else if (to_submit) {
6513 struct mm_struct *cur_mm;
6515 mutex_lock(&ctx->uring_lock);
6516 /* already have mm, so io_submit_sqes() won't try to grab it */
6517 cur_mm = ctx->sqo_mm;
6518 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
6520 mutex_unlock(&ctx->uring_lock);
6522 if (submitted != to_submit)
6525 if (flags & IORING_ENTER_GETEVENTS) {
6526 unsigned nr_events = 0;
6528 min_complete = min(min_complete, ctx->cq_entries);
6530 if (ctx->flags & IORING_SETUP_IOPOLL) {
6531 ret = io_iopoll_check(ctx, &nr_events, min_complete);
6533 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
6538 percpu_ref_put(&ctx->refs);
6541 return submitted ? submitted : ret;
6544 static int io_uring_show_cred(int id, void *p, void *data)
6546 const struct cred *cred = p;
6547 struct seq_file *m = data;
6548 struct user_namespace *uns = seq_user_ns(m);
6549 struct group_info *gi;
6554 seq_printf(m, "%5d\n", id);
6555 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
6556 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
6557 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
6558 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
6559 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
6560 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
6561 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
6562 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
6563 seq_puts(m, "\n\tGroups:\t");
6564 gi = cred->group_info;
6565 for (g = 0; g < gi->ngroups; g++) {
6566 seq_put_decimal_ull(m, g ? " " : "",
6567 from_kgid_munged(uns, gi->gid[g]));
6569 seq_puts(m, "\n\tCapEff:\t");
6570 cap = cred->cap_effective;
6571 CAP_FOR_EACH_U32(__capi)
6572 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
6577 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
6581 mutex_lock(&ctx->uring_lock);
6582 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
6583 for (i = 0; i < ctx->nr_user_files; i++) {
6584 struct fixed_file_table *table;
6587 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6588 f = table->files[i & IORING_FILE_TABLE_MASK];
6590 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
6592 seq_printf(m, "%5u: <none>\n", i);
6594 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
6595 for (i = 0; i < ctx->nr_user_bufs; i++) {
6596 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
6598 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
6599 (unsigned int) buf->len);
6601 if (!idr_is_empty(&ctx->personality_idr)) {
6602 seq_printf(m, "Personalities:\n");
6603 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
6605 mutex_unlock(&ctx->uring_lock);
6608 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
6610 struct io_ring_ctx *ctx = f->private_data;
6612 if (percpu_ref_tryget(&ctx->refs)) {
6613 __io_uring_show_fdinfo(ctx, m);
6614 percpu_ref_put(&ctx->refs);
6618 static const struct file_operations io_uring_fops = {
6619 .release = io_uring_release,
6620 .flush = io_uring_flush,
6621 .mmap = io_uring_mmap,
6623 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
6624 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
6626 .poll = io_uring_poll,
6627 .fasync = io_uring_fasync,
6628 .show_fdinfo = io_uring_show_fdinfo,
6631 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
6632 struct io_uring_params *p)
6634 struct io_rings *rings;
6635 size_t size, sq_array_offset;
6637 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
6638 if (size == SIZE_MAX)
6641 rings = io_mem_alloc(size);
6646 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
6647 rings->sq_ring_mask = p->sq_entries - 1;
6648 rings->cq_ring_mask = p->cq_entries - 1;
6649 rings->sq_ring_entries = p->sq_entries;
6650 rings->cq_ring_entries = p->cq_entries;
6651 ctx->sq_mask = rings->sq_ring_mask;
6652 ctx->cq_mask = rings->cq_ring_mask;
6653 ctx->sq_entries = rings->sq_ring_entries;
6654 ctx->cq_entries = rings->cq_ring_entries;
6656 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
6657 if (size == SIZE_MAX) {
6658 io_mem_free(ctx->rings);
6663 ctx->sq_sqes = io_mem_alloc(size);
6664 if (!ctx->sq_sqes) {
6665 io_mem_free(ctx->rings);
6674 * Allocate an anonymous fd, this is what constitutes the application
6675 * visible backing of an io_uring instance. The application mmaps this
6676 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
6677 * we have to tie this fd to a socket for file garbage collection purposes.
6679 static int io_uring_get_fd(struct io_ring_ctx *ctx)
6684 #if defined(CONFIG_UNIX)
6685 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
6691 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
6695 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
6696 O_RDWR | O_CLOEXEC);
6699 ret = PTR_ERR(file);
6703 #if defined(CONFIG_UNIX)
6704 ctx->ring_sock->file = file;
6706 fd_install(ret, file);
6709 #if defined(CONFIG_UNIX)
6710 sock_release(ctx->ring_sock);
6711 ctx->ring_sock = NULL;
6716 static int io_uring_create(unsigned entries, struct io_uring_params *p)
6718 struct user_struct *user = NULL;
6719 struct io_ring_ctx *ctx;
6725 if (entries > IORING_MAX_ENTRIES) {
6726 if (!(p->flags & IORING_SETUP_CLAMP))
6728 entries = IORING_MAX_ENTRIES;
6732 * Use twice as many entries for the CQ ring. It's possible for the
6733 * application to drive a higher depth than the size of the SQ ring,
6734 * since the sqes are only used at submission time. This allows for
6735 * some flexibility in overcommitting a bit. If the application has
6736 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
6737 * of CQ ring entries manually.
6739 p->sq_entries = roundup_pow_of_two(entries);
6740 if (p->flags & IORING_SETUP_CQSIZE) {
6742 * If IORING_SETUP_CQSIZE is set, we do the same roundup
6743 * to a power-of-two, if it isn't already. We do NOT impose
6744 * any cq vs sq ring sizing.
6746 if (p->cq_entries < p->sq_entries)
6748 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
6749 if (!(p->flags & IORING_SETUP_CLAMP))
6751 p->cq_entries = IORING_MAX_CQ_ENTRIES;
6753 p->cq_entries = roundup_pow_of_two(p->cq_entries);
6755 p->cq_entries = 2 * p->sq_entries;
6758 user = get_uid(current_user());
6759 account_mem = !capable(CAP_IPC_LOCK);
6762 ret = io_account_mem(user,
6763 ring_pages(p->sq_entries, p->cq_entries));
6770 ctx = io_ring_ctx_alloc(p);
6773 io_unaccount_mem(user, ring_pages(p->sq_entries,
6778 ctx->compat = in_compat_syscall();
6779 ctx->account_mem = account_mem;
6781 ctx->creds = get_current_cred();
6783 ret = io_allocate_scq_urings(ctx, p);
6787 ret = io_sq_offload_start(ctx, p);
6791 memset(&p->sq_off, 0, sizeof(p->sq_off));
6792 p->sq_off.head = offsetof(struct io_rings, sq.head);
6793 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
6794 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
6795 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
6796 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
6797 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
6798 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
6800 memset(&p->cq_off, 0, sizeof(p->cq_off));
6801 p->cq_off.head = offsetof(struct io_rings, cq.head);
6802 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
6803 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
6804 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
6805 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
6806 p->cq_off.cqes = offsetof(struct io_rings, cqes);
6809 * Install ring fd as the very last thing, so we don't risk someone
6810 * having closed it before we finish setup
6812 ret = io_uring_get_fd(ctx);
6816 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
6817 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
6818 IORING_FEAT_CUR_PERSONALITY;
6819 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
6822 io_ring_ctx_wait_and_kill(ctx);
6827 * Sets up an aio uring context, and returns the fd. Applications asks for a
6828 * ring size, we return the actual sq/cq ring sizes (among other things) in the
6829 * params structure passed in.
6831 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
6833 struct io_uring_params p;
6837 if (copy_from_user(&p, params, sizeof(p)))
6839 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
6844 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
6845 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
6846 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
6849 ret = io_uring_create(entries, &p);
6853 if (copy_to_user(params, &p, sizeof(p)))
6859 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
6860 struct io_uring_params __user *, params)
6862 return io_uring_setup(entries, params);
6865 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
6867 struct io_uring_probe *p;
6871 size = struct_size(p, ops, nr_args);
6872 if (size == SIZE_MAX)
6874 p = kzalloc(size, GFP_KERNEL);
6879 if (copy_from_user(p, arg, size))
6882 if (memchr_inv(p, 0, size))
6885 p->last_op = IORING_OP_LAST - 1;
6886 if (nr_args > IORING_OP_LAST)
6887 nr_args = IORING_OP_LAST;
6889 for (i = 0; i < nr_args; i++) {
6891 if (!io_op_defs[i].not_supported)
6892 p->ops[i].flags = IO_URING_OP_SUPPORTED;
6897 if (copy_to_user(arg, p, size))
6904 static int io_register_personality(struct io_ring_ctx *ctx)
6906 const struct cred *creds = get_current_cred();
6909 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
6910 USHRT_MAX, GFP_KERNEL);
6916 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
6918 const struct cred *old_creds;
6920 old_creds = idr_remove(&ctx->personality_idr, id);
6922 put_cred(old_creds);
6929 static bool io_register_op_must_quiesce(int op)
6932 case IORING_UNREGISTER_FILES:
6933 case IORING_REGISTER_FILES_UPDATE:
6934 case IORING_REGISTER_PROBE:
6935 case IORING_REGISTER_PERSONALITY:
6936 case IORING_UNREGISTER_PERSONALITY:
6943 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
6944 void __user *arg, unsigned nr_args)
6945 __releases(ctx->uring_lock)
6946 __acquires(ctx->uring_lock)
6951 * We're inside the ring mutex, if the ref is already dying, then
6952 * someone else killed the ctx or is already going through
6953 * io_uring_register().
6955 if (percpu_ref_is_dying(&ctx->refs))
6958 if (io_register_op_must_quiesce(opcode)) {
6959 percpu_ref_kill(&ctx->refs);
6962 * Drop uring mutex before waiting for references to exit. If
6963 * another thread is currently inside io_uring_enter() it might
6964 * need to grab the uring_lock to make progress. If we hold it
6965 * here across the drain wait, then we can deadlock. It's safe
6966 * to drop the mutex here, since no new references will come in
6967 * after we've killed the percpu ref.
6969 mutex_unlock(&ctx->uring_lock);
6970 ret = wait_for_completion_interruptible(&ctx->completions[0]);
6971 mutex_lock(&ctx->uring_lock);
6973 percpu_ref_resurrect(&ctx->refs);
6980 case IORING_REGISTER_BUFFERS:
6981 ret = io_sqe_buffer_register(ctx, arg, nr_args);
6983 case IORING_UNREGISTER_BUFFERS:
6987 ret = io_sqe_buffer_unregister(ctx);
6989 case IORING_REGISTER_FILES:
6990 ret = io_sqe_files_register(ctx, arg, nr_args);
6992 case IORING_UNREGISTER_FILES:
6996 ret = io_sqe_files_unregister(ctx);
6998 case IORING_REGISTER_FILES_UPDATE:
6999 ret = io_sqe_files_update(ctx, arg, nr_args);
7001 case IORING_REGISTER_EVENTFD:
7002 case IORING_REGISTER_EVENTFD_ASYNC:
7006 ret = io_eventfd_register(ctx, arg);
7009 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
7010 ctx->eventfd_async = 1;
7012 ctx->eventfd_async = 0;
7014 case IORING_UNREGISTER_EVENTFD:
7018 ret = io_eventfd_unregister(ctx);
7020 case IORING_REGISTER_PROBE:
7022 if (!arg || nr_args > 256)
7024 ret = io_probe(ctx, arg, nr_args);
7026 case IORING_REGISTER_PERSONALITY:
7030 ret = io_register_personality(ctx);
7032 case IORING_UNREGISTER_PERSONALITY:
7036 ret = io_unregister_personality(ctx, nr_args);
7043 if (io_register_op_must_quiesce(opcode)) {
7044 /* bring the ctx back to life */
7045 percpu_ref_reinit(&ctx->refs);
7047 reinit_completion(&ctx->completions[0]);
7052 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
7053 void __user *, arg, unsigned int, nr_args)
7055 struct io_ring_ctx *ctx;
7064 if (f.file->f_op != &io_uring_fops)
7067 ctx = f.file->private_data;
7069 mutex_lock(&ctx->uring_lock);
7070 ret = __io_uring_register(ctx, opcode, arg, nr_args);
7071 mutex_unlock(&ctx->uring_lock);
7072 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
7073 ctx->cq_ev_fd != NULL, ret);
7079 static int __init io_uring_init(void)
7081 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
7082 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
7083 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
7086 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
7087 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
7088 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
7089 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
7090 BUILD_BUG_SQE_ELEM(1, __u8, flags);
7091 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
7092 BUILD_BUG_SQE_ELEM(4, __s32, fd);
7093 BUILD_BUG_SQE_ELEM(8, __u64, off);
7094 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
7095 BUILD_BUG_SQE_ELEM(16, __u64, addr);
7096 BUILD_BUG_SQE_ELEM(24, __u32, len);
7097 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
7098 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
7099 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
7100 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
7101 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
7102 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
7103 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
7104 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
7105 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
7106 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
7107 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
7108 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
7109 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
7110 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
7111 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
7112 BUILD_BUG_SQE_ELEM(42, __u16, personality);
7114 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
7115 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
7118 __initcall(io_uring_init);