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 <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp;
104 u32 tail ____cacheline_aligned_in_smp;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq, cq;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask, cq_ring_mask;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries, cq_ring_entries;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
157 * Written by the application, shouldn't be modified by the
162 * Number of completion events lost because the queue was full;
163 * this should be avoided by the application by making sure
164 * there are not more requests pending than there is space in
165 * the completion queue.
167 * Written by the kernel, shouldn't be modified by the
168 * application (i.e. get number of "new events" by comparing to
171 * As completion events come in out of order this counter is not
172 * ordered with any other data.
176 * Ring buffer of completion events.
178 * The kernel writes completion events fresh every time they are
179 * produced, so the application is allowed to modify pending
182 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
185 struct io_mapped_ubuf {
188 struct bio_vec *bvec;
189 unsigned int nr_bvecs;
192 struct fixed_file_table {
196 struct fixed_file_ref_node {
197 struct percpu_ref refs;
198 struct list_head node;
199 struct list_head file_list;
200 struct fixed_file_data *file_data;
201 struct llist_node llist;
204 struct fixed_file_data {
205 struct fixed_file_table *table;
206 struct io_ring_ctx *ctx;
208 struct percpu_ref *cur_refs;
209 struct percpu_ref refs;
210 struct completion done;
211 struct list_head ref_list;
216 struct list_head list;
224 struct percpu_ref refs;
225 } ____cacheline_aligned_in_smp;
229 unsigned int compat: 1;
230 unsigned int limit_mem: 1;
231 unsigned int cq_overflow_flushed: 1;
232 unsigned int drain_next: 1;
233 unsigned int eventfd_async: 1;
236 * Ring buffer of indices into array of io_uring_sqe, which is
237 * mmapped by the application using the IORING_OFF_SQES offset.
239 * This indirection could e.g. be used to assign fixed
240 * io_uring_sqe entries to operations and only submit them to
241 * the queue when needed.
243 * The kernel modifies neither the indices array nor the entries
247 unsigned cached_sq_head;
250 unsigned sq_thread_idle;
251 unsigned cached_sq_dropped;
252 atomic_t cached_cq_overflow;
253 unsigned long sq_check_overflow;
255 struct list_head defer_list;
256 struct list_head timeout_list;
257 struct list_head cq_overflow_list;
259 wait_queue_head_t inflight_wait;
260 struct io_uring_sqe *sq_sqes;
261 } ____cacheline_aligned_in_smp;
263 struct io_rings *rings;
267 struct task_struct *sqo_thread; /* if using sq thread polling */
268 struct mm_struct *sqo_mm;
269 wait_queue_head_t sqo_wait;
272 * If used, fixed file set. Writers must ensure that ->refs is dead,
273 * readers must ensure that ->refs is alive as long as the file* is
274 * used. Only updated through io_uring_register(2).
276 struct fixed_file_data *file_data;
277 unsigned nr_user_files;
279 struct file *ring_file;
281 /* if used, fixed mapped user buffers */
282 unsigned nr_user_bufs;
283 struct io_mapped_ubuf *user_bufs;
285 struct user_struct *user;
287 const struct cred *creds;
289 struct completion ref_comp;
290 struct completion sq_thread_comp;
292 /* if all else fails... */
293 struct io_kiocb *fallback_req;
295 #if defined(CONFIG_UNIX)
296 struct socket *ring_sock;
299 struct idr io_buffer_idr;
301 struct idr personality_idr;
304 unsigned cached_cq_tail;
307 atomic_t cq_timeouts;
308 unsigned long cq_check_overflow;
309 struct wait_queue_head cq_wait;
310 struct fasync_struct *cq_fasync;
311 struct eventfd_ctx *cq_ev_fd;
312 } ____cacheline_aligned_in_smp;
315 struct mutex uring_lock;
316 wait_queue_head_t wait;
317 } ____cacheline_aligned_in_smp;
320 spinlock_t completion_lock;
323 * ->poll_list is protected by the ctx->uring_lock for
324 * io_uring instances that don't use IORING_SETUP_SQPOLL.
325 * For SQPOLL, only the single threaded io_sq_thread() will
326 * manipulate the list, hence no extra locking is needed there.
328 struct list_head poll_list;
329 struct hlist_head *cancel_hash;
330 unsigned cancel_hash_bits;
331 bool poll_multi_file;
333 spinlock_t inflight_lock;
334 struct list_head inflight_list;
335 } ____cacheline_aligned_in_smp;
337 struct delayed_work file_put_work;
338 struct llist_head file_put_llist;
340 struct work_struct exit_work;
344 * First field must be the file pointer in all the
345 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
347 struct io_poll_iocb {
350 struct wait_queue_head *head;
356 struct wait_queue_entry wait;
361 struct file *put_file;
365 struct io_timeout_data {
366 struct io_kiocb *req;
367 struct hrtimer timer;
368 struct timespec64 ts;
369 enum hrtimer_mode mode;
374 struct sockaddr __user *addr;
375 int __user *addr_len;
377 unsigned long nofile;
402 /* NOTE: kiocb has the file as the first member, so don't do it here */
410 struct sockaddr __user *addr;
417 struct user_msghdr __user *msg;
423 struct io_buffer *kbuf;
429 struct filename *filename;
431 unsigned long nofile;
434 struct io_files_update {
460 struct epoll_event event;
464 struct file *file_out;
465 struct file *file_in;
472 struct io_provide_buf {
486 const char __user *filename;
487 struct statx __user *buffer;
490 struct io_async_connect {
491 struct sockaddr_storage address;
494 struct io_async_msghdr {
495 struct iovec fast_iov[UIO_FASTIOV];
497 struct sockaddr __user *uaddr;
499 struct sockaddr_storage addr;
503 struct iovec fast_iov[UIO_FASTIOV];
507 struct wait_page_queue wpq;
508 struct callback_head task_work;
511 struct io_async_ctx {
513 struct io_async_rw rw;
514 struct io_async_msghdr msg;
515 struct io_async_connect connect;
516 struct io_timeout_data timeout;
521 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
522 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
523 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
524 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
525 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
526 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
533 REQ_F_LINK_TIMEOUT_BIT,
536 REQ_F_TIMEOUT_NOSEQ_BIT,
537 REQ_F_COMP_LOCKED_BIT,
538 REQ_F_NEED_CLEANUP_BIT,
541 REQ_F_BUFFER_SELECTED_BIT,
542 REQ_F_NO_FILE_TABLE_BIT,
543 REQ_F_QUEUE_TIMEOUT_BIT,
544 REQ_F_WORK_INITIALIZED_BIT,
545 REQ_F_TASK_PINNED_BIT,
547 /* not a real bit, just to check we're not overflowing the space */
553 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
554 /* drain existing IO first */
555 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
557 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
558 /* doesn't sever on completion < 0 */
559 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
561 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
562 /* IOSQE_BUFFER_SELECT */
563 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
566 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
567 /* fail rest of links */
568 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
569 /* on inflight list */
570 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
571 /* read/write uses file position */
572 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
573 /* must not punt to workers */
574 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
575 /* has linked timeout */
576 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
577 /* timeout request */
578 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
580 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
581 /* no timeout sequence */
582 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
583 /* completion under lock */
584 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
586 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
587 /* in overflow list */
588 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
589 /* already went through poll handler */
590 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
591 /* buffer already selected */
592 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
593 /* doesn't need file table for this request */
594 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
595 /* needs to queue linked timeout */
596 REQ_F_QUEUE_TIMEOUT = BIT(REQ_F_QUEUE_TIMEOUT_BIT),
597 /* io_wq_work is initialized */
598 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
599 /* req->task is refcounted */
600 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
604 struct io_poll_iocb poll;
605 struct io_wq_work work;
609 * NOTE! Each of the iocb union members has the file pointer
610 * as the first entry in their struct definition. So you can
611 * access the file pointer through any of the sub-structs,
612 * or directly as just 'ki_filp' in this struct.
618 struct io_poll_iocb poll;
619 struct io_accept accept;
621 struct io_cancel cancel;
622 struct io_timeout timeout;
623 struct io_connect connect;
624 struct io_sr_msg sr_msg;
626 struct io_close close;
627 struct io_files_update files_update;
628 struct io_fadvise fadvise;
629 struct io_madvise madvise;
630 struct io_epoll epoll;
631 struct io_splice splice;
632 struct io_provide_buf pbuf;
633 struct io_statx statx;
636 struct io_async_ctx *io;
639 /* polled IO has completed */
644 struct io_ring_ctx *ctx;
645 struct list_head list;
648 struct task_struct *task;
654 struct list_head link_list;
656 struct list_head inflight_entry;
658 struct percpu_ref *fixed_file_refs;
662 * Only commands that never go async can use the below fields,
663 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
664 * async armed poll handlers for regular commands. The latter
665 * restore the work, if needed.
668 struct hlist_node hash_node;
669 struct async_poll *apoll;
671 struct io_wq_work work;
673 struct callback_head task_work;
676 #define IO_IOPOLL_BATCH 8
678 struct io_comp_state {
680 struct list_head list;
681 struct io_ring_ctx *ctx;
684 struct io_submit_state {
685 struct blk_plug plug;
688 * io_kiocb alloc cache
690 void *reqs[IO_IOPOLL_BATCH];
691 unsigned int free_reqs;
694 * Batch completion logic
696 struct io_comp_state comp;
699 * File reference cache
703 unsigned int has_refs;
704 unsigned int used_refs;
705 unsigned int ios_left;
709 /* needs req->io allocated for deferral/async */
710 unsigned async_ctx : 1;
711 /* needs current->mm setup, does mm access */
712 unsigned needs_mm : 1;
713 /* needs req->file assigned */
714 unsigned needs_file : 1;
715 /* don't fail if file grab fails */
716 unsigned needs_file_no_error : 1;
717 /* hash wq insertion if file is a regular file */
718 unsigned hash_reg_file : 1;
719 /* unbound wq insertion if file is a non-regular file */
720 unsigned unbound_nonreg_file : 1;
721 /* opcode is not supported by this kernel */
722 unsigned not_supported : 1;
723 /* needs file table */
724 unsigned file_table : 1;
726 unsigned needs_fs : 1;
727 /* set if opcode supports polled "wait" */
729 unsigned pollout : 1;
730 /* op supports buffer selection */
731 unsigned buffer_select : 1;
734 static const struct io_op_def io_op_defs[] = {
735 [IORING_OP_NOP] = {},
736 [IORING_OP_READV] = {
740 .unbound_nonreg_file = 1,
744 [IORING_OP_WRITEV] = {
749 .unbound_nonreg_file = 1,
752 [IORING_OP_FSYNC] = {
755 [IORING_OP_READ_FIXED] = {
757 .unbound_nonreg_file = 1,
760 [IORING_OP_WRITE_FIXED] = {
763 .unbound_nonreg_file = 1,
766 [IORING_OP_POLL_ADD] = {
768 .unbound_nonreg_file = 1,
770 [IORING_OP_POLL_REMOVE] = {},
771 [IORING_OP_SYNC_FILE_RANGE] = {
774 [IORING_OP_SENDMSG] = {
778 .unbound_nonreg_file = 1,
782 [IORING_OP_RECVMSG] = {
786 .unbound_nonreg_file = 1,
791 [IORING_OP_TIMEOUT] = {
795 [IORING_OP_TIMEOUT_REMOVE] = {},
796 [IORING_OP_ACCEPT] = {
799 .unbound_nonreg_file = 1,
803 [IORING_OP_ASYNC_CANCEL] = {},
804 [IORING_OP_LINK_TIMEOUT] = {
808 [IORING_OP_CONNECT] = {
812 .unbound_nonreg_file = 1,
815 [IORING_OP_FALLOCATE] = {
818 [IORING_OP_OPENAT] = {
822 [IORING_OP_CLOSE] = {
824 .needs_file_no_error = 1,
827 [IORING_OP_FILES_UPDATE] = {
831 [IORING_OP_STATX] = {
839 .unbound_nonreg_file = 1,
843 [IORING_OP_WRITE] = {
846 .unbound_nonreg_file = 1,
849 [IORING_OP_FADVISE] = {
852 [IORING_OP_MADVISE] = {
858 .unbound_nonreg_file = 1,
864 .unbound_nonreg_file = 1,
868 [IORING_OP_OPENAT2] = {
872 [IORING_OP_EPOLL_CTL] = {
873 .unbound_nonreg_file = 1,
876 [IORING_OP_SPLICE] = {
879 .unbound_nonreg_file = 1,
881 [IORING_OP_PROVIDE_BUFFERS] = {},
882 [IORING_OP_REMOVE_BUFFERS] = {},
886 .unbound_nonreg_file = 1,
890 enum io_mem_account {
895 static bool io_rw_reissue(struct io_kiocb *req, long res);
896 static void io_cqring_fill_event(struct io_kiocb *req, long res);
897 static void io_put_req(struct io_kiocb *req);
898 static void io_double_put_req(struct io_kiocb *req);
899 static void __io_double_put_req(struct io_kiocb *req);
900 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
901 static void io_queue_linked_timeout(struct io_kiocb *req);
902 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
903 struct io_uring_files_update *ip,
905 static int io_grab_files(struct io_kiocb *req);
906 static void io_complete_rw_common(struct kiocb *kiocb, long res,
907 struct io_comp_state *cs);
908 static void io_cleanup_req(struct io_kiocb *req);
909 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
910 int fd, struct file **out_file, bool fixed);
911 static void __io_queue_sqe(struct io_kiocb *req,
912 const struct io_uring_sqe *sqe,
913 struct io_comp_state *cs);
915 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
916 struct iovec **iovec, struct iov_iter *iter,
918 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
919 struct iovec *iovec, struct iovec *fast_iov,
920 struct iov_iter *iter);
922 static struct kmem_cache *req_cachep;
924 static const struct file_operations io_uring_fops;
926 struct sock *io_uring_get_socket(struct file *file)
928 #if defined(CONFIG_UNIX)
929 if (file->f_op == &io_uring_fops) {
930 struct io_ring_ctx *ctx = file->private_data;
932 return ctx->ring_sock->sk;
937 EXPORT_SYMBOL(io_uring_get_socket);
939 static void io_get_req_task(struct io_kiocb *req)
941 if (req->flags & REQ_F_TASK_PINNED)
943 get_task_struct(req->task);
944 req->flags |= REQ_F_TASK_PINNED;
947 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
948 static void __io_put_req_task(struct io_kiocb *req)
950 if (req->flags & REQ_F_TASK_PINNED)
951 put_task_struct(req->task);
954 static void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
956 struct mm_struct *mm = current->mm;
959 kthread_unuse_mm(mm);
964 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
967 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
969 kthread_use_mm(ctx->sqo_mm);
975 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
976 struct io_kiocb *req)
978 if (!io_op_defs[req->opcode].needs_mm)
980 return __io_sq_thread_acquire_mm(ctx);
983 static inline void req_set_fail_links(struct io_kiocb *req)
985 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
986 req->flags |= REQ_F_FAIL_LINK;
989 static void io_file_put_work(struct work_struct *work);
992 * Note: must call io_req_init_async() for the first time you
993 * touch any members of io_wq_work.
995 static inline void io_req_init_async(struct io_kiocb *req)
997 if (req->flags & REQ_F_WORK_INITIALIZED)
1000 memset(&req->work, 0, sizeof(req->work));
1001 req->flags |= REQ_F_WORK_INITIALIZED;
1004 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1006 return ctx->flags & IORING_SETUP_SQPOLL;
1009 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1011 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1013 complete(&ctx->ref_comp);
1016 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1018 struct io_ring_ctx *ctx;
1021 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1025 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1026 if (!ctx->fallback_req)
1030 * Use 5 bits less than the max cq entries, that should give us around
1031 * 32 entries per hash list if totally full and uniformly spread.
1033 hash_bits = ilog2(p->cq_entries);
1037 ctx->cancel_hash_bits = hash_bits;
1038 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1040 if (!ctx->cancel_hash)
1042 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1044 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1045 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1048 ctx->flags = p->flags;
1049 init_waitqueue_head(&ctx->sqo_wait);
1050 init_waitqueue_head(&ctx->cq_wait);
1051 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1052 init_completion(&ctx->ref_comp);
1053 init_completion(&ctx->sq_thread_comp);
1054 idr_init(&ctx->io_buffer_idr);
1055 idr_init(&ctx->personality_idr);
1056 mutex_init(&ctx->uring_lock);
1057 init_waitqueue_head(&ctx->wait);
1058 spin_lock_init(&ctx->completion_lock);
1059 INIT_LIST_HEAD(&ctx->poll_list);
1060 INIT_LIST_HEAD(&ctx->defer_list);
1061 INIT_LIST_HEAD(&ctx->timeout_list);
1062 init_waitqueue_head(&ctx->inflight_wait);
1063 spin_lock_init(&ctx->inflight_lock);
1064 INIT_LIST_HEAD(&ctx->inflight_list);
1065 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1066 init_llist_head(&ctx->file_put_llist);
1069 if (ctx->fallback_req)
1070 kmem_cache_free(req_cachep, ctx->fallback_req);
1071 kfree(ctx->cancel_hash);
1076 static inline bool __req_need_defer(struct io_kiocb *req)
1078 struct io_ring_ctx *ctx = req->ctx;
1080 return req->sequence != ctx->cached_cq_tail
1081 + atomic_read(&ctx->cached_cq_overflow);
1084 static inline bool req_need_defer(struct io_kiocb *req)
1086 if (unlikely(req->flags & REQ_F_IO_DRAIN))
1087 return __req_need_defer(req);
1092 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1094 struct io_rings *rings = ctx->rings;
1096 /* order cqe stores with ring update */
1097 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1099 if (wq_has_sleeper(&ctx->cq_wait)) {
1100 wake_up_interruptible(&ctx->cq_wait);
1101 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1105 static inline void io_req_work_grab_env(struct io_kiocb *req,
1106 const struct io_op_def *def)
1108 if (!req->work.mm && def->needs_mm) {
1109 mmgrab(current->mm);
1110 req->work.mm = current->mm;
1112 if (!req->work.creds)
1113 req->work.creds = get_current_cred();
1114 if (!req->work.fs && def->needs_fs) {
1115 spin_lock(¤t->fs->lock);
1116 if (!current->fs->in_exec) {
1117 req->work.fs = current->fs;
1118 req->work.fs->users++;
1120 req->work.flags |= IO_WQ_WORK_CANCEL;
1122 spin_unlock(¤t->fs->lock);
1126 static inline void io_req_work_drop_env(struct io_kiocb *req)
1128 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1132 mmdrop(req->work.mm);
1133 req->work.mm = NULL;
1135 if (req->work.creds) {
1136 put_cred(req->work.creds);
1137 req->work.creds = NULL;
1140 struct fs_struct *fs = req->work.fs;
1142 spin_lock(&req->work.fs->lock);
1145 spin_unlock(&req->work.fs->lock);
1151 static inline void io_prep_async_work(struct io_kiocb *req,
1152 struct io_kiocb **link)
1154 const struct io_op_def *def = &io_op_defs[req->opcode];
1156 if (req->flags & REQ_F_ISREG) {
1157 if (def->hash_reg_file)
1158 io_wq_hash_work(&req->work, file_inode(req->file));
1160 if (def->unbound_nonreg_file)
1161 req->work.flags |= IO_WQ_WORK_UNBOUND;
1164 io_req_init_async(req);
1165 io_req_work_grab_env(req, def);
1167 *link = io_prep_linked_timeout(req);
1170 static inline void io_queue_async_work(struct io_kiocb *req)
1172 struct io_ring_ctx *ctx = req->ctx;
1173 struct io_kiocb *link;
1175 io_prep_async_work(req, &link);
1177 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1178 &req->work, req->flags);
1179 io_wq_enqueue(ctx->io_wq, &req->work);
1182 io_queue_linked_timeout(link);
1185 static void io_kill_timeout(struct io_kiocb *req)
1189 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1191 atomic_inc(&req->ctx->cq_timeouts);
1192 list_del_init(&req->list);
1193 req->flags |= REQ_F_COMP_LOCKED;
1194 io_cqring_fill_event(req, 0);
1199 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1201 struct io_kiocb *req, *tmp;
1203 spin_lock_irq(&ctx->completion_lock);
1204 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1205 io_kill_timeout(req);
1206 spin_unlock_irq(&ctx->completion_lock);
1209 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1212 struct io_kiocb *req = list_first_entry(&ctx->defer_list,
1213 struct io_kiocb, list);
1215 if (req_need_defer(req))
1217 list_del_init(&req->list);
1218 io_queue_async_work(req);
1219 } while (!list_empty(&ctx->defer_list));
1222 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1224 while (!list_empty(&ctx->timeout_list)) {
1225 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1226 struct io_kiocb, list);
1228 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
1230 if (req->timeout.target_seq != ctx->cached_cq_tail
1231 - atomic_read(&ctx->cq_timeouts))
1234 list_del_init(&req->list);
1235 io_kill_timeout(req);
1239 static void io_commit_cqring(struct io_ring_ctx *ctx)
1241 io_flush_timeouts(ctx);
1242 __io_commit_cqring(ctx);
1244 if (unlikely(!list_empty(&ctx->defer_list)))
1245 __io_queue_deferred(ctx);
1248 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1250 struct io_rings *rings = ctx->rings;
1253 tail = ctx->cached_cq_tail;
1255 * writes to the cq entry need to come after reading head; the
1256 * control dependency is enough as we're using WRITE_ONCE to
1259 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1262 ctx->cached_cq_tail++;
1263 return &rings->cqes[tail & ctx->cq_mask];
1266 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1270 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1272 if (!ctx->eventfd_async)
1274 return io_wq_current_is_worker();
1277 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1279 if (waitqueue_active(&ctx->wait))
1280 wake_up(&ctx->wait);
1281 if (waitqueue_active(&ctx->sqo_wait))
1282 wake_up(&ctx->sqo_wait);
1283 if (io_should_trigger_evfd(ctx))
1284 eventfd_signal(ctx->cq_ev_fd, 1);
1287 /* Returns true if there are no backlogged entries after the flush */
1288 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1290 struct io_rings *rings = ctx->rings;
1291 struct io_uring_cqe *cqe;
1292 struct io_kiocb *req;
1293 unsigned long flags;
1297 if (list_empty_careful(&ctx->cq_overflow_list))
1299 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1300 rings->cq_ring_entries))
1304 spin_lock_irqsave(&ctx->completion_lock, flags);
1306 /* if force is set, the ring is going away. always drop after that */
1308 ctx->cq_overflow_flushed = 1;
1311 while (!list_empty(&ctx->cq_overflow_list)) {
1312 cqe = io_get_cqring(ctx);
1316 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1318 list_move(&req->list, &list);
1319 req->flags &= ~REQ_F_OVERFLOW;
1321 WRITE_ONCE(cqe->user_data, req->user_data);
1322 WRITE_ONCE(cqe->res, req->result);
1323 WRITE_ONCE(cqe->flags, req->cflags);
1325 WRITE_ONCE(ctx->rings->cq_overflow,
1326 atomic_inc_return(&ctx->cached_cq_overflow));
1330 io_commit_cqring(ctx);
1332 clear_bit(0, &ctx->sq_check_overflow);
1333 clear_bit(0, &ctx->cq_check_overflow);
1335 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1336 io_cqring_ev_posted(ctx);
1338 while (!list_empty(&list)) {
1339 req = list_first_entry(&list, struct io_kiocb, list);
1340 list_del(&req->list);
1347 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1349 struct io_ring_ctx *ctx = req->ctx;
1350 struct io_uring_cqe *cqe;
1352 trace_io_uring_complete(ctx, req->user_data, res);
1355 * If we can't get a cq entry, userspace overflowed the
1356 * submission (by quite a lot). Increment the overflow count in
1359 cqe = io_get_cqring(ctx);
1361 WRITE_ONCE(cqe->user_data, req->user_data);
1362 WRITE_ONCE(cqe->res, res);
1363 WRITE_ONCE(cqe->flags, cflags);
1364 } else if (ctx->cq_overflow_flushed) {
1365 WRITE_ONCE(ctx->rings->cq_overflow,
1366 atomic_inc_return(&ctx->cached_cq_overflow));
1368 if (list_empty(&ctx->cq_overflow_list)) {
1369 set_bit(0, &ctx->sq_check_overflow);
1370 set_bit(0, &ctx->cq_check_overflow);
1372 req->flags |= REQ_F_OVERFLOW;
1373 refcount_inc(&req->refs);
1375 req->cflags = cflags;
1376 list_add_tail(&req->list, &ctx->cq_overflow_list);
1380 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1382 __io_cqring_fill_event(req, res, 0);
1385 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1387 struct io_ring_ctx *ctx = req->ctx;
1388 unsigned long flags;
1390 spin_lock_irqsave(&ctx->completion_lock, flags);
1391 __io_cqring_fill_event(req, res, cflags);
1392 io_commit_cqring(ctx);
1393 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1395 io_cqring_ev_posted(ctx);
1398 static void io_submit_flush_completions(struct io_comp_state *cs)
1400 struct io_ring_ctx *ctx = cs->ctx;
1402 spin_lock_irq(&ctx->completion_lock);
1403 while (!list_empty(&cs->list)) {
1404 struct io_kiocb *req;
1406 req = list_first_entry(&cs->list, struct io_kiocb, list);
1407 list_del(&req->list);
1408 io_cqring_fill_event(req, req->result);
1409 if (!(req->flags & REQ_F_LINK_HEAD)) {
1410 req->flags |= REQ_F_COMP_LOCKED;
1413 spin_unlock_irq(&ctx->completion_lock);
1415 spin_lock_irq(&ctx->completion_lock);
1418 io_commit_cqring(ctx);
1419 spin_unlock_irq(&ctx->completion_lock);
1421 io_cqring_ev_posted(ctx);
1425 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1426 struct io_comp_state *cs)
1429 io_cqring_add_event(req, res, cflags);
1433 list_add_tail(&req->list, &cs->list);
1435 io_submit_flush_completions(cs);
1439 static void io_req_complete(struct io_kiocb *req, long res)
1441 __io_req_complete(req, res, 0, NULL);
1444 static inline bool io_is_fallback_req(struct io_kiocb *req)
1446 return req == (struct io_kiocb *)
1447 ((unsigned long) req->ctx->fallback_req & ~1UL);
1450 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1452 struct io_kiocb *req;
1454 req = ctx->fallback_req;
1455 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1461 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1462 struct io_submit_state *state)
1464 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1465 struct io_kiocb *req;
1467 if (!state->free_reqs) {
1471 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1472 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1475 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1476 * retry single alloc to be on the safe side.
1478 if (unlikely(ret <= 0)) {
1479 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1480 if (!state->reqs[0])
1484 state->free_reqs = ret - 1;
1485 req = state->reqs[ret - 1];
1488 req = state->reqs[state->free_reqs];
1493 return io_get_fallback_req(ctx);
1496 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1500 percpu_ref_put(req->fixed_file_refs);
1505 static void io_dismantle_req(struct io_kiocb *req)
1507 if (req->flags & REQ_F_NEED_CLEANUP)
1508 io_cleanup_req(req);
1512 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1513 __io_put_req_task(req);
1514 io_req_work_drop_env(req);
1516 if (req->flags & REQ_F_INFLIGHT) {
1517 struct io_ring_ctx *ctx = req->ctx;
1518 unsigned long flags;
1520 spin_lock_irqsave(&ctx->inflight_lock, flags);
1521 list_del(&req->inflight_entry);
1522 if (waitqueue_active(&ctx->inflight_wait))
1523 wake_up(&ctx->inflight_wait);
1524 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1528 static void __io_free_req(struct io_kiocb *req)
1530 io_dismantle_req(req);
1531 percpu_ref_put(&req->ctx->refs);
1532 if (likely(!io_is_fallback_req(req)))
1533 kmem_cache_free(req_cachep, req);
1535 clear_bit_unlock(0, (unsigned long *) &req->ctx->fallback_req);
1538 static bool io_link_cancel_timeout(struct io_kiocb *req)
1540 struct io_ring_ctx *ctx = req->ctx;
1543 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1545 io_cqring_fill_event(req, -ECANCELED);
1546 io_commit_cqring(ctx);
1547 req->flags &= ~REQ_F_LINK_HEAD;
1555 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1557 struct io_ring_ctx *ctx = req->ctx;
1558 bool wake_ev = false;
1561 * The list should never be empty when we are called here. But could
1562 * potentially happen if the chain is messed up, check to be on the
1565 while (!list_empty(&req->link_list)) {
1566 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1567 struct io_kiocb, link_list);
1569 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1570 (nxt->flags & REQ_F_TIMEOUT))) {
1571 list_del_init(&nxt->link_list);
1572 wake_ev |= io_link_cancel_timeout(nxt);
1573 req->flags &= ~REQ_F_LINK_TIMEOUT;
1577 list_del_init(&req->link_list);
1578 if (!list_empty(&nxt->link_list))
1579 nxt->flags |= REQ_F_LINK_HEAD;
1585 io_cqring_ev_posted(ctx);
1589 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1591 static void io_fail_links(struct io_kiocb *req)
1593 struct io_ring_ctx *ctx = req->ctx;
1594 unsigned long flags;
1596 spin_lock_irqsave(&ctx->completion_lock, flags);
1598 while (!list_empty(&req->link_list)) {
1599 struct io_kiocb *link = list_first_entry(&req->link_list,
1600 struct io_kiocb, link_list);
1602 list_del_init(&link->link_list);
1603 trace_io_uring_fail_link(req, link);
1605 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1606 link->opcode == IORING_OP_LINK_TIMEOUT) {
1607 io_link_cancel_timeout(link);
1609 io_cqring_fill_event(link, -ECANCELED);
1610 __io_double_put_req(link);
1612 req->flags &= ~REQ_F_LINK_TIMEOUT;
1615 io_commit_cqring(ctx);
1616 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1617 io_cqring_ev_posted(ctx);
1620 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1622 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1624 req->flags &= ~REQ_F_LINK_HEAD;
1627 * If LINK is set, we have dependent requests in this chain. If we
1628 * didn't fail this request, queue the first one up, moving any other
1629 * dependencies to the next request. In case of failure, fail the rest
1632 if (req->flags & REQ_F_FAIL_LINK) {
1634 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1635 REQ_F_LINK_TIMEOUT) {
1636 struct io_ring_ctx *ctx = req->ctx;
1637 unsigned long flags;
1640 * If this is a timeout link, we could be racing with the
1641 * timeout timer. Grab the completion lock for this case to
1642 * protect against that.
1644 spin_lock_irqsave(&ctx->completion_lock, flags);
1645 io_req_link_next(req, nxt);
1646 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1648 io_req_link_next(req, nxt);
1652 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1654 struct io_ring_ctx *ctx = req->ctx;
1656 spin_lock_irq(&ctx->completion_lock);
1657 io_cqring_fill_event(req, error);
1658 io_commit_cqring(ctx);
1659 spin_unlock_irq(&ctx->completion_lock);
1661 io_cqring_ev_posted(ctx);
1662 req_set_fail_links(req);
1663 io_double_put_req(req);
1666 static void io_req_task_cancel(struct callback_head *cb)
1668 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1670 __io_req_task_cancel(req, -ECANCELED);
1673 static void __io_req_task_submit(struct io_kiocb *req)
1675 struct io_ring_ctx *ctx = req->ctx;
1677 __set_current_state(TASK_RUNNING);
1678 if (!__io_sq_thread_acquire_mm(ctx)) {
1679 mutex_lock(&ctx->uring_lock);
1680 __io_queue_sqe(req, NULL, NULL);
1681 mutex_unlock(&ctx->uring_lock);
1683 __io_req_task_cancel(req, -EFAULT);
1687 static void io_req_task_submit(struct callback_head *cb)
1689 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1691 __io_req_task_submit(req);
1694 static void io_req_task_queue(struct io_kiocb *req)
1696 struct task_struct *tsk = req->task;
1699 init_task_work(&req->task_work, io_req_task_submit);
1701 ret = task_work_add(tsk, &req->task_work, true);
1702 if (unlikely(ret)) {
1703 init_task_work(&req->task_work, io_req_task_cancel);
1704 tsk = io_wq_get_task(req->ctx->io_wq);
1705 task_work_add(tsk, &req->task_work, true);
1707 wake_up_process(tsk);
1710 static void io_queue_next(struct io_kiocb *req)
1712 struct io_kiocb *nxt = NULL;
1714 io_req_find_next(req, &nxt);
1716 io_req_task_queue(nxt);
1719 static void io_free_req(struct io_kiocb *req)
1726 void *reqs[IO_IOPOLL_BATCH];
1730 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1731 struct req_batch *rb)
1733 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1734 percpu_ref_put_many(&ctx->refs, rb->to_free);
1738 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1739 struct req_batch *rb)
1742 __io_req_free_batch_flush(ctx, rb);
1745 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1747 if (unlikely(io_is_fallback_req(req))) {
1751 if (req->flags & REQ_F_LINK_HEAD)
1754 io_dismantle_req(req);
1755 rb->reqs[rb->to_free++] = req;
1756 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1757 __io_req_free_batch_flush(req->ctx, rb);
1761 * Drop reference to request, return next in chain (if there is one) if this
1762 * was the last reference to this request.
1764 __attribute__((nonnull))
1765 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1767 if (refcount_dec_and_test(&req->refs)) {
1768 io_req_find_next(req, nxtptr);
1773 static void io_put_req(struct io_kiocb *req)
1775 if (refcount_dec_and_test(&req->refs))
1779 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1781 struct io_kiocb *nxt = NULL;
1784 * A ref is owned by io-wq in which context we're. So, if that's the
1785 * last one, it's safe to steal next work. False negatives are Ok,
1786 * it just will be re-punted async in io_put_work()
1788 if (refcount_read(&req->refs) != 1)
1791 io_req_find_next(req, &nxt);
1795 if ((nxt->flags & REQ_F_ISREG) && io_op_defs[nxt->opcode].hash_reg_file)
1796 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1798 io_req_task_queue(nxt);
1800 * If we're going to return actual work, here should be timeout prep:
1802 * link = io_prep_linked_timeout(nxt);
1804 * nxt->flags |= REQ_F_QUEUE_TIMEOUT;
1810 * Must only be used if we don't need to care about links, usually from
1811 * within the completion handling itself.
1813 static void __io_double_put_req(struct io_kiocb *req)
1815 /* drop both submit and complete references */
1816 if (refcount_sub_and_test(2, &req->refs))
1820 static void io_double_put_req(struct io_kiocb *req)
1822 /* drop both submit and complete references */
1823 if (refcount_sub_and_test(2, &req->refs))
1827 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1829 struct io_rings *rings = ctx->rings;
1831 if (test_bit(0, &ctx->cq_check_overflow)) {
1833 * noflush == true is from the waitqueue handler, just ensure
1834 * we wake up the task, and the next invocation will flush the
1835 * entries. We cannot safely to it from here.
1837 if (noflush && !list_empty(&ctx->cq_overflow_list))
1840 io_cqring_overflow_flush(ctx, false);
1843 /* See comment at the top of this file */
1845 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1848 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1850 struct io_rings *rings = ctx->rings;
1852 /* make sure SQ entry isn't read before tail */
1853 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1856 static int io_put_kbuf(struct io_kiocb *req)
1858 struct io_buffer *kbuf;
1861 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1862 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1863 cflags |= IORING_CQE_F_BUFFER;
1869 static void io_iopoll_queue(struct list_head *again)
1871 struct io_kiocb *req;
1874 req = list_first_entry(again, struct io_kiocb, list);
1875 list_del(&req->list);
1877 /* should have ->mm unless io_uring is dying, kill reqs then */
1878 if (unlikely(!current->mm) || !io_rw_reissue(req, -EAGAIN))
1879 io_complete_rw_common(&req->rw.kiocb, -EAGAIN, NULL);
1880 } while (!list_empty(again));
1884 * Find and free completed poll iocbs
1886 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1887 struct list_head *done)
1889 struct req_batch rb;
1890 struct io_kiocb *req;
1893 /* order with ->result store in io_complete_rw_iopoll() */
1897 while (!list_empty(done)) {
1900 req = list_first_entry(done, struct io_kiocb, list);
1901 if (READ_ONCE(req->result) == -EAGAIN) {
1902 req->iopoll_completed = 0;
1903 list_move_tail(&req->list, &again);
1906 list_del(&req->list);
1908 if (req->flags & REQ_F_BUFFER_SELECTED)
1909 cflags = io_put_kbuf(req);
1911 __io_cqring_fill_event(req, req->result, cflags);
1914 if (refcount_dec_and_test(&req->refs))
1915 io_req_free_batch(&rb, req);
1918 io_commit_cqring(ctx);
1919 if (ctx->flags & IORING_SETUP_SQPOLL)
1920 io_cqring_ev_posted(ctx);
1921 io_req_free_batch_finish(ctx, &rb);
1923 if (!list_empty(&again))
1924 io_iopoll_queue(&again);
1927 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1930 struct io_kiocb *req, *tmp;
1936 * Only spin for completions if we don't have multiple devices hanging
1937 * off our complete list, and we're under the requested amount.
1939 spin = !ctx->poll_multi_file && *nr_events < min;
1942 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1943 struct kiocb *kiocb = &req->rw.kiocb;
1946 * Move completed and retryable entries to our local lists.
1947 * If we find a request that requires polling, break out
1948 * and complete those lists first, if we have entries there.
1950 if (READ_ONCE(req->iopoll_completed)) {
1951 list_move_tail(&req->list, &done);
1954 if (!list_empty(&done))
1957 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1966 if (!list_empty(&done))
1967 io_iopoll_complete(ctx, nr_events, &done);
1973 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1974 * non-spinning poll check - we'll still enter the driver poll loop, but only
1975 * as a non-spinning completion check.
1977 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1980 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1983 ret = io_do_iopoll(ctx, nr_events, min);
1986 if (!min || *nr_events >= min)
1994 * We can't just wait for polled events to come to us, we have to actively
1995 * find and complete them.
1997 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1999 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2002 mutex_lock(&ctx->uring_lock);
2003 while (!list_empty(&ctx->poll_list)) {
2004 unsigned int nr_events = 0;
2006 io_iopoll_getevents(ctx, &nr_events, 1);
2009 * Ensure we allow local-to-the-cpu processing to take place,
2010 * in this case we need to ensure that we reap all events.
2014 mutex_unlock(&ctx->uring_lock);
2017 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
2020 int iters = 0, ret = 0;
2023 * We disallow the app entering submit/complete with polling, but we
2024 * still need to lock the ring to prevent racing with polled issue
2025 * that got punted to a workqueue.
2027 mutex_lock(&ctx->uring_lock);
2032 * Don't enter poll loop if we already have events pending.
2033 * If we do, we can potentially be spinning for commands that
2034 * already triggered a CQE (eg in error).
2036 if (io_cqring_events(ctx, false))
2040 * If a submit got punted to a workqueue, we can have the
2041 * application entering polling for a command before it gets
2042 * issued. That app will hold the uring_lock for the duration
2043 * of the poll right here, so we need to take a breather every
2044 * now and then to ensure that the issue has a chance to add
2045 * the poll to the issued list. Otherwise we can spin here
2046 * forever, while the workqueue is stuck trying to acquire the
2049 if (!(++iters & 7)) {
2050 mutex_unlock(&ctx->uring_lock);
2051 if (current->task_works)
2053 mutex_lock(&ctx->uring_lock);
2056 if (*nr_events < min)
2057 tmin = min - *nr_events;
2059 ret = io_iopoll_getevents(ctx, nr_events, tmin);
2063 } while (min && !*nr_events && !need_resched());
2065 mutex_unlock(&ctx->uring_lock);
2069 static void kiocb_end_write(struct io_kiocb *req)
2072 * Tell lockdep we inherited freeze protection from submission
2075 if (req->flags & REQ_F_ISREG) {
2076 struct inode *inode = file_inode(req->file);
2078 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2080 file_end_write(req->file);
2083 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2084 struct io_comp_state *cs)
2086 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2089 if (kiocb->ki_flags & IOCB_WRITE)
2090 kiocb_end_write(req);
2092 if (res != req->result)
2093 req_set_fail_links(req);
2094 if (req->flags & REQ_F_BUFFER_SELECTED)
2095 cflags = io_put_kbuf(req);
2096 __io_req_complete(req, res, cflags, cs);
2100 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2102 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2103 ssize_t ret = -ECANCELED;
2104 struct iov_iter iter;
2112 switch (req->opcode) {
2113 case IORING_OP_READV:
2114 case IORING_OP_READ_FIXED:
2115 case IORING_OP_READ:
2118 case IORING_OP_WRITEV:
2119 case IORING_OP_WRITE_FIXED:
2120 case IORING_OP_WRITE:
2124 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2129 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2132 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2137 req_set_fail_links(req);
2138 io_req_complete(req, ret);
2142 static void io_rw_resubmit(struct callback_head *cb)
2144 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2145 struct io_ring_ctx *ctx = req->ctx;
2148 __set_current_state(TASK_RUNNING);
2150 err = io_sq_thread_acquire_mm(ctx, req);
2152 if (io_resubmit_prep(req, err)) {
2153 refcount_inc(&req->refs);
2154 io_queue_async_work(req);
2159 static bool io_rw_reissue(struct io_kiocb *req, long res)
2162 struct task_struct *tsk;
2165 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2169 init_task_work(&req->task_work, io_rw_resubmit);
2170 ret = task_work_add(tsk, &req->task_work, true);
2177 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2178 struct io_comp_state *cs)
2180 if (!io_rw_reissue(req, res))
2181 io_complete_rw_common(&req->rw.kiocb, res, cs);
2184 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2186 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2188 __io_complete_rw(req, res, res2, NULL);
2191 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2193 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2195 if (kiocb->ki_flags & IOCB_WRITE)
2196 kiocb_end_write(req);
2198 if (res != -EAGAIN && res != req->result)
2199 req_set_fail_links(req);
2201 WRITE_ONCE(req->result, res);
2202 /* order with io_poll_complete() checking ->result */
2204 WRITE_ONCE(req->iopoll_completed, 1);
2208 * After the iocb has been issued, it's safe to be found on the poll list.
2209 * Adding the kiocb to the list AFTER submission ensures that we don't
2210 * find it from a io_iopoll_getevents() thread before the issuer is done
2211 * accessing the kiocb cookie.
2213 static void io_iopoll_req_issued(struct io_kiocb *req)
2215 struct io_ring_ctx *ctx = req->ctx;
2218 * Track whether we have multiple files in our lists. This will impact
2219 * how we do polling eventually, not spinning if we're on potentially
2220 * different devices.
2222 if (list_empty(&ctx->poll_list)) {
2223 ctx->poll_multi_file = false;
2224 } else if (!ctx->poll_multi_file) {
2225 struct io_kiocb *list_req;
2227 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
2229 if (list_req->file != req->file)
2230 ctx->poll_multi_file = true;
2234 * For fast devices, IO may have already completed. If it has, add
2235 * it to the front so we find it first.
2237 if (READ_ONCE(req->iopoll_completed))
2238 list_add(&req->list, &ctx->poll_list);
2240 list_add_tail(&req->list, &ctx->poll_list);
2242 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2243 wq_has_sleeper(&ctx->sqo_wait))
2244 wake_up(&ctx->sqo_wait);
2247 static void __io_state_file_put(struct io_submit_state *state)
2249 int diff = state->has_refs - state->used_refs;
2252 fput_many(state->file, diff);
2256 static inline void io_state_file_put(struct io_submit_state *state)
2259 __io_state_file_put(state);
2263 * Get as many references to a file as we have IOs left in this submission,
2264 * assuming most submissions are for one file, or at least that each file
2265 * has more than one submission.
2267 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2273 if (state->fd == fd) {
2278 __io_state_file_put(state);
2280 state->file = fget_many(fd, state->ios_left);
2285 state->has_refs = state->ios_left;
2286 state->used_refs = 1;
2291 static bool io_bdev_nowait(struct block_device *bdev)
2294 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2301 * If we tracked the file through the SCM inflight mechanism, we could support
2302 * any file. For now, just ensure that anything potentially problematic is done
2305 static bool io_file_supports_async(struct file *file, int rw)
2307 umode_t mode = file_inode(file)->i_mode;
2309 if (S_ISBLK(mode)) {
2310 if (io_bdev_nowait(file->f_inode->i_bdev))
2314 if (S_ISCHR(mode) || S_ISSOCK(mode))
2316 if (S_ISREG(mode)) {
2317 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2318 file->f_op != &io_uring_fops)
2323 /* any ->read/write should understand O_NONBLOCK */
2324 if (file->f_flags & O_NONBLOCK)
2327 if (!(file->f_mode & FMODE_NOWAIT))
2331 return file->f_op->read_iter != NULL;
2333 return file->f_op->write_iter != NULL;
2336 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2337 bool force_nonblock)
2339 struct io_ring_ctx *ctx = req->ctx;
2340 struct kiocb *kiocb = &req->rw.kiocb;
2344 if (S_ISREG(file_inode(req->file)->i_mode))
2345 req->flags |= REQ_F_ISREG;
2347 kiocb->ki_pos = READ_ONCE(sqe->off);
2348 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2349 req->flags |= REQ_F_CUR_POS;
2350 kiocb->ki_pos = req->file->f_pos;
2352 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2353 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2354 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2358 ioprio = READ_ONCE(sqe->ioprio);
2360 ret = ioprio_check_cap(ioprio);
2364 kiocb->ki_ioprio = ioprio;
2366 kiocb->ki_ioprio = get_current_ioprio();
2368 /* don't allow async punt if RWF_NOWAIT was requested */
2369 if (kiocb->ki_flags & IOCB_NOWAIT)
2370 req->flags |= REQ_F_NOWAIT;
2372 if (kiocb->ki_flags & IOCB_DIRECT)
2373 io_get_req_task(req);
2376 kiocb->ki_flags |= IOCB_NOWAIT;
2378 if (ctx->flags & IORING_SETUP_IOPOLL) {
2379 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2380 !kiocb->ki_filp->f_op->iopoll)
2383 kiocb->ki_flags |= IOCB_HIPRI;
2384 kiocb->ki_complete = io_complete_rw_iopoll;
2385 req->iopoll_completed = 0;
2386 io_get_req_task(req);
2388 if (kiocb->ki_flags & IOCB_HIPRI)
2390 kiocb->ki_complete = io_complete_rw;
2393 req->rw.addr = READ_ONCE(sqe->addr);
2394 req->rw.len = READ_ONCE(sqe->len);
2395 req->buf_index = READ_ONCE(sqe->buf_index);
2399 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2405 case -ERESTARTNOINTR:
2406 case -ERESTARTNOHAND:
2407 case -ERESTART_RESTARTBLOCK:
2409 * We can't just restart the syscall, since previously
2410 * submitted sqes may already be in progress. Just fail this
2416 kiocb->ki_complete(kiocb, ret, 0);
2420 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2421 struct io_comp_state *cs)
2423 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2425 if (req->flags & REQ_F_CUR_POS)
2426 req->file->f_pos = kiocb->ki_pos;
2427 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2428 __io_complete_rw(req, ret, 0, cs);
2430 io_rw_done(kiocb, ret);
2433 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2434 struct iov_iter *iter)
2436 struct io_ring_ctx *ctx = req->ctx;
2437 size_t len = req->rw.len;
2438 struct io_mapped_ubuf *imu;
2439 u16 index, buf_index;
2443 /* attempt to use fixed buffers without having provided iovecs */
2444 if (unlikely(!ctx->user_bufs))
2447 buf_index = req->buf_index;
2448 if (unlikely(buf_index >= ctx->nr_user_bufs))
2451 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2452 imu = &ctx->user_bufs[index];
2453 buf_addr = req->rw.addr;
2456 if (buf_addr + len < buf_addr)
2458 /* not inside the mapped region */
2459 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2463 * May not be a start of buffer, set size appropriately
2464 * and advance us to the beginning.
2466 offset = buf_addr - imu->ubuf;
2467 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2471 * Don't use iov_iter_advance() here, as it's really slow for
2472 * using the latter parts of a big fixed buffer - it iterates
2473 * over each segment manually. We can cheat a bit here, because
2476 * 1) it's a BVEC iter, we set it up
2477 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2478 * first and last bvec
2480 * So just find our index, and adjust the iterator afterwards.
2481 * If the offset is within the first bvec (or the whole first
2482 * bvec, just use iov_iter_advance(). This makes it easier
2483 * since we can just skip the first segment, which may not
2484 * be PAGE_SIZE aligned.
2486 const struct bio_vec *bvec = imu->bvec;
2488 if (offset <= bvec->bv_len) {
2489 iov_iter_advance(iter, offset);
2491 unsigned long seg_skip;
2493 /* skip first vec */
2494 offset -= bvec->bv_len;
2495 seg_skip = 1 + (offset >> PAGE_SHIFT);
2497 iter->bvec = bvec + seg_skip;
2498 iter->nr_segs -= seg_skip;
2499 iter->count -= bvec->bv_len + offset;
2500 iter->iov_offset = offset & ~PAGE_MASK;
2507 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2510 mutex_unlock(&ctx->uring_lock);
2513 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2516 * "Normal" inline submissions always hold the uring_lock, since we
2517 * grab it from the system call. Same is true for the SQPOLL offload.
2518 * The only exception is when we've detached the request and issue it
2519 * from an async worker thread, grab the lock for that case.
2522 mutex_lock(&ctx->uring_lock);
2525 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2526 int bgid, struct io_buffer *kbuf,
2529 struct io_buffer *head;
2531 if (req->flags & REQ_F_BUFFER_SELECTED)
2534 io_ring_submit_lock(req->ctx, needs_lock);
2536 lockdep_assert_held(&req->ctx->uring_lock);
2538 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2540 if (!list_empty(&head->list)) {
2541 kbuf = list_last_entry(&head->list, struct io_buffer,
2543 list_del(&kbuf->list);
2546 idr_remove(&req->ctx->io_buffer_idr, bgid);
2548 if (*len > kbuf->len)
2551 kbuf = ERR_PTR(-ENOBUFS);
2554 io_ring_submit_unlock(req->ctx, needs_lock);
2559 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2562 struct io_buffer *kbuf;
2565 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2566 bgid = req->buf_index;
2567 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2570 req->rw.addr = (u64) (unsigned long) kbuf;
2571 req->flags |= REQ_F_BUFFER_SELECTED;
2572 return u64_to_user_ptr(kbuf->addr);
2575 #ifdef CONFIG_COMPAT
2576 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2579 struct compat_iovec __user *uiov;
2580 compat_ssize_t clen;
2584 uiov = u64_to_user_ptr(req->rw.addr);
2585 if (!access_ok(uiov, sizeof(*uiov)))
2587 if (__get_user(clen, &uiov->iov_len))
2593 buf = io_rw_buffer_select(req, &len, needs_lock);
2595 return PTR_ERR(buf);
2596 iov[0].iov_base = buf;
2597 iov[0].iov_len = (compat_size_t) len;
2602 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2605 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2609 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2612 len = iov[0].iov_len;
2615 buf = io_rw_buffer_select(req, &len, needs_lock);
2617 return PTR_ERR(buf);
2618 iov[0].iov_base = buf;
2619 iov[0].iov_len = len;
2623 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2626 if (req->flags & REQ_F_BUFFER_SELECTED) {
2627 struct io_buffer *kbuf;
2629 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2630 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2631 iov[0].iov_len = kbuf->len;
2636 else if (req->rw.len > 1)
2639 #ifdef CONFIG_COMPAT
2640 if (req->ctx->compat)
2641 return io_compat_import(req, iov, needs_lock);
2644 return __io_iov_buffer_select(req, iov, needs_lock);
2647 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2648 struct iovec **iovec, struct iov_iter *iter,
2651 void __user *buf = u64_to_user_ptr(req->rw.addr);
2652 size_t sqe_len = req->rw.len;
2656 opcode = req->opcode;
2657 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2659 return io_import_fixed(req, rw, iter);
2662 /* buffer index only valid with fixed read/write, or buffer select */
2663 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2666 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2667 if (req->flags & REQ_F_BUFFER_SELECT) {
2668 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2671 return PTR_ERR(buf);
2673 req->rw.len = sqe_len;
2676 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2678 return ret < 0 ? ret : sqe_len;
2682 struct io_async_rw *iorw = &req->io->rw;
2685 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2686 if (iorw->iov == iorw->fast_iov)
2691 if (req->flags & REQ_F_BUFFER_SELECT) {
2692 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2694 ret = (*iovec)->iov_len;
2695 iov_iter_init(iter, rw, *iovec, 1, ret);
2701 #ifdef CONFIG_COMPAT
2702 if (req->ctx->compat)
2703 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2707 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2711 * For files that don't have ->read_iter() and ->write_iter(), handle them
2712 * by looping over ->read() or ->write() manually.
2714 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2715 struct iov_iter *iter)
2720 * Don't support polled IO through this interface, and we can't
2721 * support non-blocking either. For the latter, this just causes
2722 * the kiocb to be handled from an async context.
2724 if (kiocb->ki_flags & IOCB_HIPRI)
2726 if (kiocb->ki_flags & IOCB_NOWAIT)
2729 while (iov_iter_count(iter)) {
2733 if (!iov_iter_is_bvec(iter)) {
2734 iovec = iov_iter_iovec(iter);
2736 /* fixed buffers import bvec */
2737 iovec.iov_base = kmap(iter->bvec->bv_page)
2739 iovec.iov_len = min(iter->count,
2740 iter->bvec->bv_len - iter->iov_offset);
2744 nr = file->f_op->read(file, iovec.iov_base,
2745 iovec.iov_len, &kiocb->ki_pos);
2747 nr = file->f_op->write(file, iovec.iov_base,
2748 iovec.iov_len, &kiocb->ki_pos);
2751 if (iov_iter_is_bvec(iter))
2752 kunmap(iter->bvec->bv_page);
2760 if (nr != iovec.iov_len)
2762 iov_iter_advance(iter, nr);
2768 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2769 struct iovec *iovec, struct iovec *fast_iov,
2770 struct iov_iter *iter)
2772 req->io->rw.nr_segs = iter->nr_segs;
2773 req->io->rw.size = io_size;
2774 req->io->rw.iov = iovec;
2775 if (!req->io->rw.iov) {
2776 req->io->rw.iov = req->io->rw.fast_iov;
2777 if (req->io->rw.iov != fast_iov)
2778 memcpy(req->io->rw.iov, fast_iov,
2779 sizeof(struct iovec) * iter->nr_segs);
2781 req->flags |= REQ_F_NEED_CLEANUP;
2785 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2787 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2788 return req->io == NULL;
2791 static int io_alloc_async_ctx(struct io_kiocb *req)
2793 if (!io_op_defs[req->opcode].async_ctx)
2796 return __io_alloc_async_ctx(req);
2799 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2800 struct iovec *iovec, struct iovec *fast_iov,
2801 struct iov_iter *iter)
2803 if (!io_op_defs[req->opcode].async_ctx)
2806 if (__io_alloc_async_ctx(req))
2809 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2814 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2815 bool force_nonblock)
2817 struct io_async_ctx *io;
2818 struct iov_iter iter;
2821 ret = io_prep_rw(req, sqe, force_nonblock);
2825 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2828 /* either don't need iovec imported or already have it */
2829 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2833 io->rw.iov = io->rw.fast_iov;
2835 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2840 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2844 static void io_async_buf_cancel(struct callback_head *cb)
2846 struct io_async_rw *rw;
2847 struct io_kiocb *req;
2849 rw = container_of(cb, struct io_async_rw, task_work);
2850 req = rw->wpq.wait.private;
2851 __io_req_task_cancel(req, -ECANCELED);
2854 static void io_async_buf_retry(struct callback_head *cb)
2856 struct io_async_rw *rw;
2857 struct io_kiocb *req;
2859 rw = container_of(cb, struct io_async_rw, task_work);
2860 req = rw->wpq.wait.private;
2862 __io_req_task_submit(req);
2865 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2866 int sync, void *arg)
2868 struct wait_page_queue *wpq;
2869 struct io_kiocb *req = wait->private;
2870 struct io_async_rw *rw = &req->io->rw;
2871 struct wait_page_key *key = arg;
2872 struct task_struct *tsk;
2875 wpq = container_of(wait, struct wait_page_queue, wait);
2877 ret = wake_page_match(wpq, key);
2881 list_del_init(&wait->entry);
2883 init_task_work(&rw->task_work, io_async_buf_retry);
2884 /* submit ref gets dropped, acquire a new one */
2885 refcount_inc(&req->refs);
2887 ret = task_work_add(tsk, &rw->task_work, true);
2888 if (unlikely(ret)) {
2889 /* queue just for cancelation */
2890 init_task_work(&rw->task_work, io_async_buf_cancel);
2891 tsk = io_wq_get_task(req->ctx->io_wq);
2892 task_work_add(tsk, &rw->task_work, true);
2894 wake_up_process(tsk);
2898 static bool io_rw_should_retry(struct io_kiocb *req)
2900 struct kiocb *kiocb = &req->rw.kiocb;
2903 /* never retry for NOWAIT, we just complete with -EAGAIN */
2904 if (req->flags & REQ_F_NOWAIT)
2907 /* already tried, or we're doing O_DIRECT */
2908 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
2911 * just use poll if we can, and don't attempt if the fs doesn't
2912 * support callback based unlocks
2914 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
2918 * If request type doesn't require req->io to defer in general,
2919 * we need to allocate it here
2921 if (!req->io && __io_alloc_async_ctx(req))
2924 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
2925 io_async_buf_func, req);
2927 io_get_req_task(req);
2934 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
2936 if (req->file->f_op->read_iter)
2937 return call_read_iter(req->file, &req->rw.kiocb, iter);
2938 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
2941 static int io_read(struct io_kiocb *req, bool force_nonblock,
2942 struct io_comp_state *cs)
2944 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2945 struct kiocb *kiocb = &req->rw.kiocb;
2946 struct iov_iter iter;
2948 ssize_t io_size, ret;
2950 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2954 /* Ensure we clear previously set non-block flag */
2955 if (!force_nonblock)
2956 kiocb->ki_flags &= ~IOCB_NOWAIT;
2959 req->result = io_size;
2961 /* If the file doesn't support async, just async punt */
2962 if (force_nonblock && !io_file_supports_async(req->file, READ))
2965 iov_count = iov_iter_count(&iter);
2966 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2968 unsigned long nr_segs = iter.nr_segs;
2971 ret2 = io_iter_do_read(req, &iter);
2973 /* Catch -EAGAIN return for forced non-blocking submission */
2974 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
2975 kiocb_done(kiocb, ret2, cs);
2977 iter.count = iov_count;
2978 iter.nr_segs = nr_segs;
2980 ret = io_setup_async_rw(req, io_size, iovec,
2981 inline_vecs, &iter);
2984 /* if we can retry, do so with the callbacks armed */
2985 if (io_rw_should_retry(req)) {
2986 ret2 = io_iter_do_read(req, &iter);
2987 if (ret2 == -EIOCBQUEUED) {
2989 } else if (ret2 != -EAGAIN) {
2990 kiocb_done(kiocb, ret2, cs);
2994 kiocb->ki_flags &= ~IOCB_WAITQ;
2999 if (!(req->flags & REQ_F_NEED_CLEANUP))
3004 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3005 bool force_nonblock)
3007 struct io_async_ctx *io;
3008 struct iov_iter iter;
3011 ret = io_prep_rw(req, sqe, force_nonblock);
3015 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3018 req->fsize = rlimit(RLIMIT_FSIZE);
3020 /* either don't need iovec imported or already have it */
3021 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3025 io->rw.iov = io->rw.fast_iov;
3027 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
3032 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
3036 static int io_write(struct io_kiocb *req, bool force_nonblock,
3037 struct io_comp_state *cs)
3039 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3040 struct kiocb *kiocb = &req->rw.kiocb;
3041 struct iov_iter iter;
3043 ssize_t ret, io_size;
3045 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3049 /* Ensure we clear previously set non-block flag */
3050 if (!force_nonblock)
3051 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3054 req->result = io_size;
3056 /* If the file doesn't support async, just async punt */
3057 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3060 /* file path doesn't support NOWAIT for non-direct_IO */
3061 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3062 (req->flags & REQ_F_ISREG))
3065 iov_count = iov_iter_count(&iter);
3066 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3068 unsigned long nr_segs = iter.nr_segs;
3072 * Open-code file_start_write here to grab freeze protection,
3073 * which will be released by another thread in
3074 * io_complete_rw(). Fool lockdep by telling it the lock got
3075 * released so that it doesn't complain about the held lock when
3076 * we return to userspace.
3078 if (req->flags & REQ_F_ISREG) {
3079 __sb_start_write(file_inode(req->file)->i_sb,
3080 SB_FREEZE_WRITE, true);
3081 __sb_writers_release(file_inode(req->file)->i_sb,
3084 kiocb->ki_flags |= IOCB_WRITE;
3086 if (!force_nonblock)
3087 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3089 if (req->file->f_op->write_iter)
3090 ret2 = call_write_iter(req->file, kiocb, &iter);
3092 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3094 if (!force_nonblock)
3095 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3098 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3099 * retry them without IOCB_NOWAIT.
3101 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3103 if (!force_nonblock || ret2 != -EAGAIN) {
3104 kiocb_done(kiocb, ret2, cs);
3106 iter.count = iov_count;
3107 iter.nr_segs = nr_segs;
3109 ret = io_setup_async_rw(req, io_size, iovec,
3110 inline_vecs, &iter);
3117 if (!(req->flags & REQ_F_NEED_CLEANUP))
3122 static int __io_splice_prep(struct io_kiocb *req,
3123 const struct io_uring_sqe *sqe)
3125 struct io_splice* sp = &req->splice;
3126 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3129 if (req->flags & REQ_F_NEED_CLEANUP)
3131 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3135 sp->len = READ_ONCE(sqe->len);
3136 sp->flags = READ_ONCE(sqe->splice_flags);
3138 if (unlikely(sp->flags & ~valid_flags))
3141 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3142 (sp->flags & SPLICE_F_FD_IN_FIXED));
3145 req->flags |= REQ_F_NEED_CLEANUP;
3147 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3149 * Splice operation will be punted aync, and here need to
3150 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3152 io_req_init_async(req);
3153 req->work.flags |= IO_WQ_WORK_UNBOUND;
3159 static int io_tee_prep(struct io_kiocb *req,
3160 const struct io_uring_sqe *sqe)
3162 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3164 return __io_splice_prep(req, sqe);
3167 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3169 struct io_splice *sp = &req->splice;
3170 struct file *in = sp->file_in;
3171 struct file *out = sp->file_out;
3172 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3178 ret = do_tee(in, out, sp->len, flags);
3180 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3181 req->flags &= ~REQ_F_NEED_CLEANUP;
3184 req_set_fail_links(req);
3185 io_req_complete(req, ret);
3189 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3191 struct io_splice* sp = &req->splice;
3193 sp->off_in = READ_ONCE(sqe->splice_off_in);
3194 sp->off_out = READ_ONCE(sqe->off);
3195 return __io_splice_prep(req, sqe);
3198 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3200 struct io_splice *sp = &req->splice;
3201 struct file *in = sp->file_in;
3202 struct file *out = sp->file_out;
3203 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3204 loff_t *poff_in, *poff_out;
3210 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3211 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3214 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3216 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3217 req->flags &= ~REQ_F_NEED_CLEANUP;
3220 req_set_fail_links(req);
3221 io_req_complete(req, ret);
3226 * IORING_OP_NOP just posts a completion event, nothing else.
3228 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3230 struct io_ring_ctx *ctx = req->ctx;
3232 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3235 __io_req_complete(req, 0, 0, cs);
3239 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3241 struct io_ring_ctx *ctx = req->ctx;
3246 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3248 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3251 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3252 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3255 req->sync.off = READ_ONCE(sqe->off);
3256 req->sync.len = READ_ONCE(sqe->len);
3260 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3262 loff_t end = req->sync.off + req->sync.len;
3265 /* fsync always requires a blocking context */
3269 ret = vfs_fsync_range(req->file, req->sync.off,
3270 end > 0 ? end : LLONG_MAX,
3271 req->sync.flags & IORING_FSYNC_DATASYNC);
3273 req_set_fail_links(req);
3274 io_req_complete(req, ret);
3278 static int io_fallocate_prep(struct io_kiocb *req,
3279 const struct io_uring_sqe *sqe)
3281 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3283 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3286 req->sync.off = READ_ONCE(sqe->off);
3287 req->sync.len = READ_ONCE(sqe->addr);
3288 req->sync.mode = READ_ONCE(sqe->len);
3289 req->fsize = rlimit(RLIMIT_FSIZE);
3293 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3297 /* fallocate always requiring blocking context */
3301 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3302 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3304 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3306 req_set_fail_links(req);
3307 io_req_complete(req, ret);
3311 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3313 const char __user *fname;
3316 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3318 if (unlikely(sqe->ioprio || sqe->buf_index))
3320 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3323 /* open.how should be already initialised */
3324 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3325 req->open.how.flags |= O_LARGEFILE;
3327 req->open.dfd = READ_ONCE(sqe->fd);
3328 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3329 req->open.filename = getname(fname);
3330 if (IS_ERR(req->open.filename)) {
3331 ret = PTR_ERR(req->open.filename);
3332 req->open.filename = NULL;
3335 req->open.nofile = rlimit(RLIMIT_NOFILE);
3336 req->flags |= REQ_F_NEED_CLEANUP;
3340 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3344 if (req->flags & REQ_F_NEED_CLEANUP)
3346 mode = READ_ONCE(sqe->len);
3347 flags = READ_ONCE(sqe->open_flags);
3348 req->open.how = build_open_how(flags, mode);
3349 return __io_openat_prep(req, sqe);
3352 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3354 struct open_how __user *how;
3358 if (req->flags & REQ_F_NEED_CLEANUP)
3360 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3361 len = READ_ONCE(sqe->len);
3362 if (len < OPEN_HOW_SIZE_VER0)
3365 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3370 return __io_openat_prep(req, sqe);
3373 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3375 struct open_flags op;
3382 ret = build_open_flags(&req->open.how, &op);
3386 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3390 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3393 ret = PTR_ERR(file);
3395 fsnotify_open(file);
3396 fd_install(ret, file);
3399 putname(req->open.filename);
3400 req->flags &= ~REQ_F_NEED_CLEANUP;
3402 req_set_fail_links(req);
3403 io_req_complete(req, ret);
3407 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3409 return io_openat2(req, force_nonblock);
3412 static int io_remove_buffers_prep(struct io_kiocb *req,
3413 const struct io_uring_sqe *sqe)
3415 struct io_provide_buf *p = &req->pbuf;
3418 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3421 tmp = READ_ONCE(sqe->fd);
3422 if (!tmp || tmp > USHRT_MAX)
3425 memset(p, 0, sizeof(*p));
3427 p->bgid = READ_ONCE(sqe->buf_group);
3431 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3432 int bgid, unsigned nbufs)
3436 /* shouldn't happen */
3440 /* the head kbuf is the list itself */
3441 while (!list_empty(&buf->list)) {
3442 struct io_buffer *nxt;
3444 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3445 list_del(&nxt->list);
3452 idr_remove(&ctx->io_buffer_idr, bgid);
3457 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3458 struct io_comp_state *cs)
3460 struct io_provide_buf *p = &req->pbuf;
3461 struct io_ring_ctx *ctx = req->ctx;
3462 struct io_buffer *head;
3465 io_ring_submit_lock(ctx, !force_nonblock);
3467 lockdep_assert_held(&ctx->uring_lock);
3470 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3472 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3474 io_ring_submit_lock(ctx, !force_nonblock);
3476 req_set_fail_links(req);
3477 __io_req_complete(req, ret, 0, cs);
3481 static int io_provide_buffers_prep(struct io_kiocb *req,
3482 const struct io_uring_sqe *sqe)
3484 struct io_provide_buf *p = &req->pbuf;
3487 if (sqe->ioprio || sqe->rw_flags)
3490 tmp = READ_ONCE(sqe->fd);
3491 if (!tmp || tmp > USHRT_MAX)
3494 p->addr = READ_ONCE(sqe->addr);
3495 p->len = READ_ONCE(sqe->len);
3497 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3500 p->bgid = READ_ONCE(sqe->buf_group);
3501 tmp = READ_ONCE(sqe->off);
3502 if (tmp > USHRT_MAX)
3508 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3510 struct io_buffer *buf;
3511 u64 addr = pbuf->addr;
3512 int i, bid = pbuf->bid;
3514 for (i = 0; i < pbuf->nbufs; i++) {
3515 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3520 buf->len = pbuf->len;
3525 INIT_LIST_HEAD(&buf->list);
3528 list_add_tail(&buf->list, &(*head)->list);
3532 return i ? i : -ENOMEM;
3535 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3536 struct io_comp_state *cs)
3538 struct io_provide_buf *p = &req->pbuf;
3539 struct io_ring_ctx *ctx = req->ctx;
3540 struct io_buffer *head, *list;
3543 io_ring_submit_lock(ctx, !force_nonblock);
3545 lockdep_assert_held(&ctx->uring_lock);
3547 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3549 ret = io_add_buffers(p, &head);
3554 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3557 __io_remove_buffers(ctx, head, p->bgid, -1U);
3562 io_ring_submit_unlock(ctx, !force_nonblock);
3564 req_set_fail_links(req);
3565 __io_req_complete(req, ret, 0, cs);
3569 static int io_epoll_ctl_prep(struct io_kiocb *req,
3570 const struct io_uring_sqe *sqe)
3572 #if defined(CONFIG_EPOLL)
3573 if (sqe->ioprio || sqe->buf_index)
3575 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3578 req->epoll.epfd = READ_ONCE(sqe->fd);
3579 req->epoll.op = READ_ONCE(sqe->len);
3580 req->epoll.fd = READ_ONCE(sqe->off);
3582 if (ep_op_has_event(req->epoll.op)) {
3583 struct epoll_event __user *ev;
3585 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3586 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3596 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3597 struct io_comp_state *cs)
3599 #if defined(CONFIG_EPOLL)
3600 struct io_epoll *ie = &req->epoll;
3603 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3604 if (force_nonblock && ret == -EAGAIN)
3608 req_set_fail_links(req);
3609 __io_req_complete(req, ret, 0, cs);
3616 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3618 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3619 if (sqe->ioprio || sqe->buf_index || sqe->off)
3621 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3624 req->madvise.addr = READ_ONCE(sqe->addr);
3625 req->madvise.len = READ_ONCE(sqe->len);
3626 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3633 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3635 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3636 struct io_madvise *ma = &req->madvise;
3642 ret = do_madvise(ma->addr, ma->len, ma->advice);
3644 req_set_fail_links(req);
3645 io_req_complete(req, ret);
3652 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3654 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3656 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3659 req->fadvise.offset = READ_ONCE(sqe->off);
3660 req->fadvise.len = READ_ONCE(sqe->len);
3661 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3665 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3667 struct io_fadvise *fa = &req->fadvise;
3670 if (force_nonblock) {
3671 switch (fa->advice) {
3672 case POSIX_FADV_NORMAL:
3673 case POSIX_FADV_RANDOM:
3674 case POSIX_FADV_SEQUENTIAL:
3681 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3683 req_set_fail_links(req);
3684 io_req_complete(req, ret);
3688 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3690 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3692 if (sqe->ioprio || sqe->buf_index)
3694 if (req->flags & REQ_F_FIXED_FILE)
3697 req->statx.dfd = READ_ONCE(sqe->fd);
3698 req->statx.mask = READ_ONCE(sqe->len);
3699 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3700 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3701 req->statx.flags = READ_ONCE(sqe->statx_flags);
3706 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3708 struct io_statx *ctx = &req->statx;
3711 if (force_nonblock) {
3712 /* only need file table for an actual valid fd */
3713 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3714 req->flags |= REQ_F_NO_FILE_TABLE;
3718 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3722 req_set_fail_links(req);
3723 io_req_complete(req, ret);
3727 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3730 * If we queue this for async, it must not be cancellable. That would
3731 * leave the 'file' in an undeterminate state, and here need to modify
3732 * io_wq_work.flags, so initialize io_wq_work firstly.
3734 io_req_init_async(req);
3735 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3737 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3739 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3740 sqe->rw_flags || sqe->buf_index)
3742 if (req->flags & REQ_F_FIXED_FILE)
3745 req->close.fd = READ_ONCE(sqe->fd);
3746 if ((req->file && req->file->f_op == &io_uring_fops) ||
3747 req->close.fd == req->ctx->ring_fd)
3750 req->close.put_file = NULL;
3754 static int io_close(struct io_kiocb *req, bool force_nonblock,
3755 struct io_comp_state *cs)
3757 struct io_close *close = &req->close;
3760 /* might be already done during nonblock submission */
3761 if (!close->put_file) {
3762 ret = __close_fd_get_file(close->fd, &close->put_file);
3764 return (ret == -ENOENT) ? -EBADF : ret;
3767 /* if the file has a flush method, be safe and punt to async */
3768 if (close->put_file->f_op->flush && force_nonblock) {
3769 /* was never set, but play safe */
3770 req->flags &= ~REQ_F_NOWAIT;
3771 /* avoid grabbing files - we don't need the files */
3772 req->flags |= REQ_F_NO_FILE_TABLE;
3776 /* No ->flush() or already async, safely close from here */
3777 ret = filp_close(close->put_file, req->work.files);
3779 req_set_fail_links(req);
3780 fput(close->put_file);
3781 close->put_file = NULL;
3782 __io_req_complete(req, ret, 0, cs);
3786 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3788 struct io_ring_ctx *ctx = req->ctx;
3793 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3795 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3798 req->sync.off = READ_ONCE(sqe->off);
3799 req->sync.len = READ_ONCE(sqe->len);
3800 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3804 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3808 /* sync_file_range always requires a blocking context */
3812 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3815 req_set_fail_links(req);
3816 io_req_complete(req, ret);
3820 #if defined(CONFIG_NET)
3821 static int io_setup_async_msg(struct io_kiocb *req,
3822 struct io_async_msghdr *kmsg)
3826 if (io_alloc_async_ctx(req)) {
3827 if (kmsg->iov != kmsg->fast_iov)
3831 req->flags |= REQ_F_NEED_CLEANUP;
3832 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3836 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3838 struct io_sr_msg *sr = &req->sr_msg;
3839 struct io_async_ctx *io = req->io;
3842 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3845 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3846 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3847 sr->len = READ_ONCE(sqe->len);
3849 #ifdef CONFIG_COMPAT
3850 if (req->ctx->compat)
3851 sr->msg_flags |= MSG_CMSG_COMPAT;
3854 if (!io || req->opcode == IORING_OP_SEND)
3856 /* iovec is already imported */
3857 if (req->flags & REQ_F_NEED_CLEANUP)
3860 io->msg.iov = io->msg.fast_iov;
3861 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3864 req->flags |= REQ_F_NEED_CLEANUP;
3868 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3869 struct io_comp_state *cs)
3871 struct io_async_msghdr *kmsg = NULL;
3872 struct socket *sock;
3875 sock = sock_from_file(req->file, &ret);
3877 struct io_async_ctx io;
3881 kmsg = &req->io->msg;
3882 kmsg->msg.msg_name = &req->io->msg.addr;
3883 /* if iov is set, it's allocated already */
3885 kmsg->iov = kmsg->fast_iov;
3886 kmsg->msg.msg_iter.iov = kmsg->iov;
3888 struct io_sr_msg *sr = &req->sr_msg;
3891 kmsg->msg.msg_name = &io.msg.addr;
3893 io.msg.iov = io.msg.fast_iov;
3894 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3895 sr->msg_flags, &io.msg.iov);
3900 flags = req->sr_msg.msg_flags;
3901 if (flags & MSG_DONTWAIT)
3902 req->flags |= REQ_F_NOWAIT;
3903 else if (force_nonblock)
3904 flags |= MSG_DONTWAIT;
3906 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3907 if (force_nonblock && ret == -EAGAIN)
3908 return io_setup_async_msg(req, kmsg);
3909 if (ret == -ERESTARTSYS)
3913 if (kmsg && kmsg->iov != kmsg->fast_iov)
3915 req->flags &= ~REQ_F_NEED_CLEANUP;
3917 req_set_fail_links(req);
3918 __io_req_complete(req, ret, 0, cs);
3922 static int io_send(struct io_kiocb *req, bool force_nonblock,
3923 struct io_comp_state *cs)
3925 struct socket *sock;
3928 sock = sock_from_file(req->file, &ret);
3930 struct io_sr_msg *sr = &req->sr_msg;
3935 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3940 msg.msg_name = NULL;
3941 msg.msg_control = NULL;
3942 msg.msg_controllen = 0;
3943 msg.msg_namelen = 0;
3945 flags = req->sr_msg.msg_flags;
3946 if (flags & MSG_DONTWAIT)
3947 req->flags |= REQ_F_NOWAIT;
3948 else if (force_nonblock)
3949 flags |= MSG_DONTWAIT;
3951 msg.msg_flags = flags;
3952 ret = sock_sendmsg(sock, &msg);
3953 if (force_nonblock && ret == -EAGAIN)
3955 if (ret == -ERESTARTSYS)
3960 req_set_fail_links(req);
3961 __io_req_complete(req, ret, 0, cs);
3965 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3967 struct io_sr_msg *sr = &req->sr_msg;
3968 struct iovec __user *uiov;
3972 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3977 if (req->flags & REQ_F_BUFFER_SELECT) {
3980 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3982 sr->len = io->msg.iov[0].iov_len;
3983 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3987 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3988 &io->msg.iov, &io->msg.msg.msg_iter);
3996 #ifdef CONFIG_COMPAT
3997 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3998 struct io_async_ctx *io)
4000 struct compat_msghdr __user *msg_compat;
4001 struct io_sr_msg *sr = &req->sr_msg;
4002 struct compat_iovec __user *uiov;
4007 msg_compat = (struct compat_msghdr __user *) sr->msg;
4008 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
4013 uiov = compat_ptr(ptr);
4014 if (req->flags & REQ_F_BUFFER_SELECT) {
4015 compat_ssize_t clen;
4019 if (!access_ok(uiov, sizeof(*uiov)))
4021 if (__get_user(clen, &uiov->iov_len))
4025 sr->len = io->msg.iov[0].iov_len;
4028 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4030 &io->msg.msg.msg_iter);
4039 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
4041 io->msg.iov = io->msg.fast_iov;
4043 #ifdef CONFIG_COMPAT
4044 if (req->ctx->compat)
4045 return __io_compat_recvmsg_copy_hdr(req, io);
4048 return __io_recvmsg_copy_hdr(req, io);
4051 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4052 int *cflags, bool needs_lock)
4054 struct io_sr_msg *sr = &req->sr_msg;
4055 struct io_buffer *kbuf;
4057 if (!(req->flags & REQ_F_BUFFER_SELECT))
4060 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4065 req->flags |= REQ_F_BUFFER_SELECTED;
4067 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
4068 *cflags |= IORING_CQE_F_BUFFER;
4072 static int io_recvmsg_prep(struct io_kiocb *req,
4073 const struct io_uring_sqe *sqe)
4075 struct io_sr_msg *sr = &req->sr_msg;
4076 struct io_async_ctx *io = req->io;
4079 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4082 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4083 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4084 sr->len = READ_ONCE(sqe->len);
4085 sr->bgid = READ_ONCE(sqe->buf_group);
4087 #ifdef CONFIG_COMPAT
4088 if (req->ctx->compat)
4089 sr->msg_flags |= MSG_CMSG_COMPAT;
4092 if (!io || req->opcode == IORING_OP_RECV)
4094 /* iovec is already imported */
4095 if (req->flags & REQ_F_NEED_CLEANUP)
4098 ret = io_recvmsg_copy_hdr(req, io);
4100 req->flags |= REQ_F_NEED_CLEANUP;
4104 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4105 struct io_comp_state *cs)
4107 struct io_async_msghdr *kmsg = NULL;
4108 struct socket *sock;
4109 int ret, cflags = 0;
4111 sock = sock_from_file(req->file, &ret);
4113 struct io_buffer *kbuf;
4114 struct io_async_ctx io;
4118 kmsg = &req->io->msg;
4119 kmsg->msg.msg_name = &req->io->msg.addr;
4120 /* if iov is set, it's allocated already */
4122 kmsg->iov = kmsg->fast_iov;
4123 kmsg->msg.msg_iter.iov = kmsg->iov;
4126 kmsg->msg.msg_name = &io.msg.addr;
4128 ret = io_recvmsg_copy_hdr(req, &io);
4133 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4135 return PTR_ERR(kbuf);
4137 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4138 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4139 1, req->sr_msg.len);
4142 flags = req->sr_msg.msg_flags;
4143 if (flags & MSG_DONTWAIT)
4144 req->flags |= REQ_F_NOWAIT;
4145 else if (force_nonblock)
4146 flags |= MSG_DONTWAIT;
4148 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
4149 kmsg->uaddr, flags);
4150 if (force_nonblock && ret == -EAGAIN)
4151 return io_setup_async_msg(req, kmsg);
4152 if (ret == -ERESTARTSYS)
4156 if (kmsg && kmsg->iov != kmsg->fast_iov)
4158 req->flags &= ~REQ_F_NEED_CLEANUP;
4160 req_set_fail_links(req);
4161 __io_req_complete(req, ret, cflags, cs);
4165 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4166 struct io_comp_state *cs)
4168 struct io_buffer *kbuf = NULL;
4169 struct socket *sock;
4170 int ret, cflags = 0;
4172 sock = sock_from_file(req->file, &ret);
4174 struct io_sr_msg *sr = &req->sr_msg;
4175 void __user *buf = sr->buf;
4180 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4182 return PTR_ERR(kbuf);
4184 buf = u64_to_user_ptr(kbuf->addr);
4186 ret = import_single_range(READ, buf, sr->len, &iov,
4193 req->flags |= REQ_F_NEED_CLEANUP;
4194 msg.msg_name = NULL;
4195 msg.msg_control = NULL;
4196 msg.msg_controllen = 0;
4197 msg.msg_namelen = 0;
4198 msg.msg_iocb = NULL;
4201 flags = req->sr_msg.msg_flags;
4202 if (flags & MSG_DONTWAIT)
4203 req->flags |= REQ_F_NOWAIT;
4204 else if (force_nonblock)
4205 flags |= MSG_DONTWAIT;
4207 ret = sock_recvmsg(sock, &msg, flags);
4208 if (force_nonblock && ret == -EAGAIN)
4210 if (ret == -ERESTARTSYS)
4215 req->flags &= ~REQ_F_NEED_CLEANUP;
4217 req_set_fail_links(req);
4218 __io_req_complete(req, ret, cflags, cs);
4222 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4224 struct io_accept *accept = &req->accept;
4226 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4228 if (sqe->ioprio || sqe->len || sqe->buf_index)
4231 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4232 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4233 accept->flags = READ_ONCE(sqe->accept_flags);
4234 accept->nofile = rlimit(RLIMIT_NOFILE);
4238 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4239 struct io_comp_state *cs)
4241 struct io_accept *accept = &req->accept;
4242 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4245 if (req->file->f_flags & O_NONBLOCK)
4246 req->flags |= REQ_F_NOWAIT;
4248 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4249 accept->addr_len, accept->flags,
4251 if (ret == -EAGAIN && force_nonblock)
4254 if (ret == -ERESTARTSYS)
4256 req_set_fail_links(req);
4258 __io_req_complete(req, ret, 0, cs);
4262 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4264 struct io_connect *conn = &req->connect;
4265 struct io_async_ctx *io = req->io;
4267 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4269 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4272 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4273 conn->addr_len = READ_ONCE(sqe->addr2);
4278 return move_addr_to_kernel(conn->addr, conn->addr_len,
4279 &io->connect.address);
4282 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4283 struct io_comp_state *cs)
4285 struct io_async_ctx __io, *io;
4286 unsigned file_flags;
4292 ret = move_addr_to_kernel(req->connect.addr,
4293 req->connect.addr_len,
4294 &__io.connect.address);
4300 file_flags = force_nonblock ? O_NONBLOCK : 0;
4302 ret = __sys_connect_file(req->file, &io->connect.address,
4303 req->connect.addr_len, file_flags);
4304 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4307 if (io_alloc_async_ctx(req)) {
4311 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4314 if (ret == -ERESTARTSYS)
4318 req_set_fail_links(req);
4319 __io_req_complete(req, ret, 0, cs);
4322 #else /* !CONFIG_NET */
4323 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4328 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4329 struct io_comp_state *cs)
4334 static int io_send(struct io_kiocb *req, bool force_nonblock,
4335 struct io_comp_state *cs)
4340 static int io_recvmsg_prep(struct io_kiocb *req,
4341 const struct io_uring_sqe *sqe)
4346 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4347 struct io_comp_state *cs)
4352 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4353 struct io_comp_state *cs)
4358 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4363 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4364 struct io_comp_state *cs)
4369 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4374 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4375 struct io_comp_state *cs)
4379 #endif /* CONFIG_NET */
4381 struct io_poll_table {
4382 struct poll_table_struct pt;
4383 struct io_kiocb *req;
4387 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4388 __poll_t mask, task_work_func_t func)
4390 struct task_struct *tsk;
4393 /* for instances that support it check for an event match first: */
4394 if (mask && !(mask & poll->events))
4397 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4399 list_del_init(&poll->wait.entry);
4403 init_task_work(&req->task_work, func);
4405 * If this fails, then the task is exiting. When a task exits, the
4406 * work gets canceled, so just cancel this request as well instead
4407 * of executing it. We can't safely execute it anyway, as we may not
4408 * have the needed state needed for it anyway.
4410 ret = task_work_add(tsk, &req->task_work, true);
4411 if (unlikely(ret)) {
4412 WRITE_ONCE(poll->canceled, true);
4413 tsk = io_wq_get_task(req->ctx->io_wq);
4414 task_work_add(tsk, &req->task_work, true);
4416 wake_up_process(tsk);
4420 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4421 __acquires(&req->ctx->completion_lock)
4423 struct io_ring_ctx *ctx = req->ctx;
4425 if (!req->result && !READ_ONCE(poll->canceled)) {
4426 struct poll_table_struct pt = { ._key = poll->events };
4428 req->result = vfs_poll(req->file, &pt) & poll->events;
4431 spin_lock_irq(&ctx->completion_lock);
4432 if (!req->result && !READ_ONCE(poll->canceled)) {
4433 add_wait_queue(poll->head, &poll->wait);
4440 static void io_poll_remove_double(struct io_kiocb *req)
4442 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4444 lockdep_assert_held(&req->ctx->completion_lock);
4446 if (poll && poll->head) {
4447 struct wait_queue_head *head = poll->head;
4449 spin_lock(&head->lock);
4450 list_del_init(&poll->wait.entry);
4451 if (poll->wait.private)
4452 refcount_dec(&req->refs);
4454 spin_unlock(&head->lock);
4458 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4460 struct io_ring_ctx *ctx = req->ctx;
4462 io_poll_remove_double(req);
4463 req->poll.done = true;
4464 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4465 io_commit_cqring(ctx);
4468 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4470 struct io_ring_ctx *ctx = req->ctx;
4472 if (io_poll_rewait(req, &req->poll)) {
4473 spin_unlock_irq(&ctx->completion_lock);
4477 hash_del(&req->hash_node);
4478 io_poll_complete(req, req->result, 0);
4479 req->flags |= REQ_F_COMP_LOCKED;
4480 io_put_req_find_next(req, nxt);
4481 spin_unlock_irq(&ctx->completion_lock);
4483 io_cqring_ev_posted(ctx);
4486 static void io_poll_task_func(struct callback_head *cb)
4488 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4489 struct io_kiocb *nxt = NULL;
4491 io_poll_task_handler(req, &nxt);
4493 struct io_ring_ctx *ctx = nxt->ctx;
4495 mutex_lock(&ctx->uring_lock);
4496 __io_queue_sqe(nxt, NULL, NULL);
4497 mutex_unlock(&ctx->uring_lock);
4501 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4502 int sync, void *key)
4504 struct io_kiocb *req = wait->private;
4505 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4506 __poll_t mask = key_to_poll(key);
4508 /* for instances that support it check for an event match first: */
4509 if (mask && !(mask & poll->events))
4512 if (req->poll.head) {
4515 spin_lock(&req->poll.head->lock);
4516 done = list_empty(&req->poll.wait.entry);
4518 list_del_init(&req->poll.wait.entry);
4519 spin_unlock(&req->poll.head->lock);
4521 __io_async_wake(req, poll, mask, io_poll_task_func);
4523 refcount_dec(&req->refs);
4527 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4528 wait_queue_func_t wake_func)
4532 poll->canceled = false;
4533 poll->events = events;
4534 INIT_LIST_HEAD(&poll->wait.entry);
4535 init_waitqueue_func_entry(&poll->wait, wake_func);
4538 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4539 struct wait_queue_head *head)
4541 struct io_kiocb *req = pt->req;
4544 * If poll->head is already set, it's because the file being polled
4545 * uses multiple waitqueues for poll handling (eg one for read, one
4546 * for write). Setup a separate io_poll_iocb if this happens.
4548 if (unlikely(poll->head)) {
4549 /* already have a 2nd entry, fail a third attempt */
4551 pt->error = -EINVAL;
4554 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4556 pt->error = -ENOMEM;
4559 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4560 refcount_inc(&req->refs);
4561 poll->wait.private = req;
4562 req->io = (void *) poll;
4568 if (poll->events & EPOLLEXCLUSIVE)
4569 add_wait_queue_exclusive(head, &poll->wait);
4571 add_wait_queue(head, &poll->wait);
4574 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4575 struct poll_table_struct *p)
4577 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4579 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4582 static void io_async_task_func(struct callback_head *cb)
4584 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4585 struct async_poll *apoll = req->apoll;
4586 struct io_ring_ctx *ctx = req->ctx;
4587 bool canceled = false;
4589 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4591 if (io_poll_rewait(req, &apoll->poll)) {
4592 spin_unlock_irq(&ctx->completion_lock);
4596 /* If req is still hashed, it cannot have been canceled. Don't check. */
4597 if (hash_hashed(&req->hash_node)) {
4598 hash_del(&req->hash_node);
4600 canceled = READ_ONCE(apoll->poll.canceled);
4602 io_cqring_fill_event(req, -ECANCELED);
4603 io_commit_cqring(ctx);
4607 spin_unlock_irq(&ctx->completion_lock);
4609 /* restore ->work in case we need to retry again */
4610 if (req->flags & REQ_F_WORK_INITIALIZED)
4611 memcpy(&req->work, &apoll->work, sizeof(req->work));
4615 __set_current_state(TASK_RUNNING);
4616 if (io_sq_thread_acquire_mm(ctx, req)) {
4617 io_cqring_add_event(req, -EFAULT, 0);
4620 mutex_lock(&ctx->uring_lock);
4621 __io_queue_sqe(req, NULL, NULL);
4622 mutex_unlock(&ctx->uring_lock);
4624 io_cqring_ev_posted(ctx);
4626 req_set_fail_links(req);
4627 io_double_put_req(req);
4631 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4634 struct io_kiocb *req = wait->private;
4635 struct io_poll_iocb *poll = &req->apoll->poll;
4637 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4640 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4643 static void io_poll_req_insert(struct io_kiocb *req)
4645 struct io_ring_ctx *ctx = req->ctx;
4646 struct hlist_head *list;
4648 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4649 hlist_add_head(&req->hash_node, list);
4652 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4653 struct io_poll_iocb *poll,
4654 struct io_poll_table *ipt, __poll_t mask,
4655 wait_queue_func_t wake_func)
4656 __acquires(&ctx->completion_lock)
4658 struct io_ring_ctx *ctx = req->ctx;
4659 bool cancel = false;
4661 io_init_poll_iocb(poll, mask, wake_func);
4662 poll->file = req->file;
4663 poll->wait.private = req;
4665 ipt->pt._key = mask;
4667 ipt->error = -EINVAL;
4669 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4671 spin_lock_irq(&ctx->completion_lock);
4672 if (likely(poll->head)) {
4673 spin_lock(&poll->head->lock);
4674 if (unlikely(list_empty(&poll->wait.entry))) {
4680 if (mask || ipt->error)
4681 list_del_init(&poll->wait.entry);
4683 WRITE_ONCE(poll->canceled, true);
4684 else if (!poll->done) /* actually waiting for an event */
4685 io_poll_req_insert(req);
4686 spin_unlock(&poll->head->lock);
4692 static bool io_arm_poll_handler(struct io_kiocb *req)
4694 const struct io_op_def *def = &io_op_defs[req->opcode];
4695 struct io_ring_ctx *ctx = req->ctx;
4696 struct async_poll *apoll;
4697 struct io_poll_table ipt;
4701 if (!req->file || !file_can_poll(req->file))
4703 if (req->flags & REQ_F_POLLED)
4705 if (!def->pollin && !def->pollout)
4708 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4709 if (unlikely(!apoll))
4712 req->flags |= REQ_F_POLLED;
4713 if (req->flags & REQ_F_WORK_INITIALIZED)
4714 memcpy(&apoll->work, &req->work, sizeof(req->work));
4715 had_io = req->io != NULL;
4717 io_get_req_task(req);
4719 INIT_HLIST_NODE(&req->hash_node);
4723 mask |= POLLIN | POLLRDNORM;
4725 mask |= POLLOUT | POLLWRNORM;
4726 mask |= POLLERR | POLLPRI;
4728 ipt.pt._qproc = io_async_queue_proc;
4730 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4734 /* only remove double add if we did it here */
4736 io_poll_remove_double(req);
4737 spin_unlock_irq(&ctx->completion_lock);
4738 if (req->flags & REQ_F_WORK_INITIALIZED)
4739 memcpy(&req->work, &apoll->work, sizeof(req->work));
4743 spin_unlock_irq(&ctx->completion_lock);
4744 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4745 apoll->poll.events);
4749 static bool __io_poll_remove_one(struct io_kiocb *req,
4750 struct io_poll_iocb *poll)
4752 bool do_complete = false;
4754 spin_lock(&poll->head->lock);
4755 WRITE_ONCE(poll->canceled, true);
4756 if (!list_empty(&poll->wait.entry)) {
4757 list_del_init(&poll->wait.entry);
4760 spin_unlock(&poll->head->lock);
4761 hash_del(&req->hash_node);
4765 static bool io_poll_remove_one(struct io_kiocb *req)
4769 if (req->opcode == IORING_OP_POLL_ADD) {
4770 io_poll_remove_double(req);
4771 do_complete = __io_poll_remove_one(req, &req->poll);
4773 struct async_poll *apoll = req->apoll;
4775 /* non-poll requests have submit ref still */
4776 do_complete = __io_poll_remove_one(req, &apoll->poll);
4780 * restore ->work because we will call
4781 * io_req_work_drop_env below when dropping the
4784 if (req->flags & REQ_F_WORK_INITIALIZED)
4785 memcpy(&req->work, &apoll->work,
4792 io_cqring_fill_event(req, -ECANCELED);
4793 io_commit_cqring(req->ctx);
4794 req->flags |= REQ_F_COMP_LOCKED;
4801 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4803 struct hlist_node *tmp;
4804 struct io_kiocb *req;
4807 spin_lock_irq(&ctx->completion_lock);
4808 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4809 struct hlist_head *list;
4811 list = &ctx->cancel_hash[i];
4812 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4813 posted += io_poll_remove_one(req);
4815 spin_unlock_irq(&ctx->completion_lock);
4818 io_cqring_ev_posted(ctx);
4821 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4823 struct hlist_head *list;
4824 struct io_kiocb *req;
4826 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4827 hlist_for_each_entry(req, list, hash_node) {
4828 if (sqe_addr != req->user_data)
4830 if (io_poll_remove_one(req))
4838 static int io_poll_remove_prep(struct io_kiocb *req,
4839 const struct io_uring_sqe *sqe)
4841 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4843 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4847 req->poll.addr = READ_ONCE(sqe->addr);
4852 * Find a running poll command that matches one specified in sqe->addr,
4853 * and remove it if found.
4855 static int io_poll_remove(struct io_kiocb *req)
4857 struct io_ring_ctx *ctx = req->ctx;
4861 addr = req->poll.addr;
4862 spin_lock_irq(&ctx->completion_lock);
4863 ret = io_poll_cancel(ctx, addr);
4864 spin_unlock_irq(&ctx->completion_lock);
4867 req_set_fail_links(req);
4868 io_req_complete(req, ret);
4872 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4875 struct io_kiocb *req = wait->private;
4876 struct io_poll_iocb *poll = &req->poll;
4878 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4881 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4882 struct poll_table_struct *p)
4884 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4886 __io_queue_proc(&pt->req->poll, pt, head);
4889 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4891 struct io_poll_iocb *poll = &req->poll;
4894 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4896 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4901 events = READ_ONCE(sqe->poll32_events);
4903 events = swahw32(events);
4905 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4906 (events & EPOLLEXCLUSIVE);
4908 io_get_req_task(req);
4912 static int io_poll_add(struct io_kiocb *req)
4914 struct io_poll_iocb *poll = &req->poll;
4915 struct io_ring_ctx *ctx = req->ctx;
4916 struct io_poll_table ipt;
4919 INIT_HLIST_NODE(&req->hash_node);
4920 INIT_LIST_HEAD(&req->list);
4921 ipt.pt._qproc = io_poll_queue_proc;
4923 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4926 if (mask) { /* no async, we'd stolen it */
4928 io_poll_complete(req, mask, 0);
4930 spin_unlock_irq(&ctx->completion_lock);
4933 io_cqring_ev_posted(ctx);
4939 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4941 struct io_timeout_data *data = container_of(timer,
4942 struct io_timeout_data, timer);
4943 struct io_kiocb *req = data->req;
4944 struct io_ring_ctx *ctx = req->ctx;
4945 unsigned long flags;
4947 atomic_inc(&ctx->cq_timeouts);
4949 spin_lock_irqsave(&ctx->completion_lock, flags);
4951 * We could be racing with timeout deletion. If the list is empty,
4952 * then timeout lookup already found it and will be handling it.
4954 if (!list_empty(&req->list))
4955 list_del_init(&req->list);
4957 io_cqring_fill_event(req, -ETIME);
4958 io_commit_cqring(ctx);
4959 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4961 io_cqring_ev_posted(ctx);
4962 req_set_fail_links(req);
4964 return HRTIMER_NORESTART;
4967 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4969 struct io_kiocb *req;
4972 list_for_each_entry(req, &ctx->timeout_list, list) {
4973 if (user_data == req->user_data) {
4974 list_del_init(&req->list);
4983 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4987 req_set_fail_links(req);
4988 io_cqring_fill_event(req, -ECANCELED);
4993 static int io_timeout_remove_prep(struct io_kiocb *req,
4994 const struct io_uring_sqe *sqe)
4996 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4998 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
5001 req->timeout.addr = READ_ONCE(sqe->addr);
5002 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5003 if (req->timeout.flags)
5010 * Remove or update an existing timeout command
5012 static int io_timeout_remove(struct io_kiocb *req)
5014 struct io_ring_ctx *ctx = req->ctx;
5017 spin_lock_irq(&ctx->completion_lock);
5018 ret = io_timeout_cancel(ctx, req->timeout.addr);
5020 io_cqring_fill_event(req, ret);
5021 io_commit_cqring(ctx);
5022 spin_unlock_irq(&ctx->completion_lock);
5023 io_cqring_ev_posted(ctx);
5025 req_set_fail_links(req);
5030 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5031 bool is_timeout_link)
5033 struct io_timeout_data *data;
5035 u32 off = READ_ONCE(sqe->off);
5037 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5039 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5041 if (off && is_timeout_link)
5043 flags = READ_ONCE(sqe->timeout_flags);
5044 if (flags & ~IORING_TIMEOUT_ABS)
5047 req->timeout.off = off;
5049 if (!req->io && io_alloc_async_ctx(req))
5052 data = &req->io->timeout;
5054 req->flags |= REQ_F_TIMEOUT;
5056 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5059 if (flags & IORING_TIMEOUT_ABS)
5060 data->mode = HRTIMER_MODE_ABS;
5062 data->mode = HRTIMER_MODE_REL;
5064 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5068 static int io_timeout(struct io_kiocb *req)
5070 struct io_ring_ctx *ctx = req->ctx;
5071 struct io_timeout_data *data = &req->io->timeout;
5072 struct list_head *entry;
5073 u32 tail, off = req->timeout.off;
5075 spin_lock_irq(&ctx->completion_lock);
5078 * sqe->off holds how many events that need to occur for this
5079 * timeout event to be satisfied. If it isn't set, then this is
5080 * a pure timeout request, sequence isn't used.
5083 req->flags |= REQ_F_TIMEOUT_NOSEQ;
5084 entry = ctx->timeout_list.prev;
5088 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5089 req->timeout.target_seq = tail + off;
5092 * Insertion sort, ensuring the first entry in the list is always
5093 * the one we need first.
5095 list_for_each_prev(entry, &ctx->timeout_list) {
5096 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
5098 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
5100 /* nxt.seq is behind @tail, otherwise would've been completed */
5101 if (off >= nxt->timeout.target_seq - tail)
5105 list_add(&req->list, entry);
5106 data->timer.function = io_timeout_fn;
5107 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5108 spin_unlock_irq(&ctx->completion_lock);
5112 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5114 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5116 return req->user_data == (unsigned long) data;
5119 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5121 enum io_wq_cancel cancel_ret;
5124 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5125 switch (cancel_ret) {
5126 case IO_WQ_CANCEL_OK:
5129 case IO_WQ_CANCEL_RUNNING:
5132 case IO_WQ_CANCEL_NOTFOUND:
5140 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5141 struct io_kiocb *req, __u64 sqe_addr,
5144 unsigned long flags;
5147 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5148 if (ret != -ENOENT) {
5149 spin_lock_irqsave(&ctx->completion_lock, flags);
5153 spin_lock_irqsave(&ctx->completion_lock, flags);
5154 ret = io_timeout_cancel(ctx, sqe_addr);
5157 ret = io_poll_cancel(ctx, sqe_addr);
5161 io_cqring_fill_event(req, ret);
5162 io_commit_cqring(ctx);
5163 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5164 io_cqring_ev_posted(ctx);
5167 req_set_fail_links(req);
5171 static int io_async_cancel_prep(struct io_kiocb *req,
5172 const struct io_uring_sqe *sqe)
5174 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5176 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
5180 req->cancel.addr = READ_ONCE(sqe->addr);
5184 static int io_async_cancel(struct io_kiocb *req)
5186 struct io_ring_ctx *ctx = req->ctx;
5188 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5192 static int io_files_update_prep(struct io_kiocb *req,
5193 const struct io_uring_sqe *sqe)
5195 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
5198 req->files_update.offset = READ_ONCE(sqe->off);
5199 req->files_update.nr_args = READ_ONCE(sqe->len);
5200 if (!req->files_update.nr_args)
5202 req->files_update.arg = READ_ONCE(sqe->addr);
5206 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5207 struct io_comp_state *cs)
5209 struct io_ring_ctx *ctx = req->ctx;
5210 struct io_uring_files_update up;
5216 up.offset = req->files_update.offset;
5217 up.fds = req->files_update.arg;
5219 mutex_lock(&ctx->uring_lock);
5220 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5221 mutex_unlock(&ctx->uring_lock);
5224 req_set_fail_links(req);
5225 __io_req_complete(req, ret, 0, cs);
5229 static int io_req_defer_prep(struct io_kiocb *req,
5230 const struct io_uring_sqe *sqe, bool for_async)
5237 if (io_op_defs[req->opcode].file_table) {
5238 io_req_init_async(req);
5239 ret = io_grab_files(req);
5244 if (for_async || (req->flags & REQ_F_WORK_INITIALIZED)) {
5245 io_req_init_async(req);
5246 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
5249 switch (req->opcode) {
5252 case IORING_OP_READV:
5253 case IORING_OP_READ_FIXED:
5254 case IORING_OP_READ:
5255 ret = io_read_prep(req, sqe, true);
5257 case IORING_OP_WRITEV:
5258 case IORING_OP_WRITE_FIXED:
5259 case IORING_OP_WRITE:
5260 ret = io_write_prep(req, sqe, true);
5262 case IORING_OP_POLL_ADD:
5263 ret = io_poll_add_prep(req, sqe);
5265 case IORING_OP_POLL_REMOVE:
5266 ret = io_poll_remove_prep(req, sqe);
5268 case IORING_OP_FSYNC:
5269 ret = io_prep_fsync(req, sqe);
5271 case IORING_OP_SYNC_FILE_RANGE:
5272 ret = io_prep_sfr(req, sqe);
5274 case IORING_OP_SENDMSG:
5275 case IORING_OP_SEND:
5276 ret = io_sendmsg_prep(req, sqe);
5278 case IORING_OP_RECVMSG:
5279 case IORING_OP_RECV:
5280 ret = io_recvmsg_prep(req, sqe);
5282 case IORING_OP_CONNECT:
5283 ret = io_connect_prep(req, sqe);
5285 case IORING_OP_TIMEOUT:
5286 ret = io_timeout_prep(req, sqe, false);
5288 case IORING_OP_TIMEOUT_REMOVE:
5289 ret = io_timeout_remove_prep(req, sqe);
5291 case IORING_OP_ASYNC_CANCEL:
5292 ret = io_async_cancel_prep(req, sqe);
5294 case IORING_OP_LINK_TIMEOUT:
5295 ret = io_timeout_prep(req, sqe, true);
5297 case IORING_OP_ACCEPT:
5298 ret = io_accept_prep(req, sqe);
5300 case IORING_OP_FALLOCATE:
5301 ret = io_fallocate_prep(req, sqe);
5303 case IORING_OP_OPENAT:
5304 ret = io_openat_prep(req, sqe);
5306 case IORING_OP_CLOSE:
5307 ret = io_close_prep(req, sqe);
5309 case IORING_OP_FILES_UPDATE:
5310 ret = io_files_update_prep(req, sqe);
5312 case IORING_OP_STATX:
5313 ret = io_statx_prep(req, sqe);
5315 case IORING_OP_FADVISE:
5316 ret = io_fadvise_prep(req, sqe);
5318 case IORING_OP_MADVISE:
5319 ret = io_madvise_prep(req, sqe);
5321 case IORING_OP_OPENAT2:
5322 ret = io_openat2_prep(req, sqe);
5324 case IORING_OP_EPOLL_CTL:
5325 ret = io_epoll_ctl_prep(req, sqe);
5327 case IORING_OP_SPLICE:
5328 ret = io_splice_prep(req, sqe);
5330 case IORING_OP_PROVIDE_BUFFERS:
5331 ret = io_provide_buffers_prep(req, sqe);
5333 case IORING_OP_REMOVE_BUFFERS:
5334 ret = io_remove_buffers_prep(req, sqe);
5337 ret = io_tee_prep(req, sqe);
5340 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5349 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5351 struct io_ring_ctx *ctx = req->ctx;
5354 /* Still need defer if there is pending req in defer list. */
5355 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5359 if (io_alloc_async_ctx(req))
5361 ret = io_req_defer_prep(req, sqe, true);
5366 spin_lock_irq(&ctx->completion_lock);
5367 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5368 spin_unlock_irq(&ctx->completion_lock);
5372 trace_io_uring_defer(ctx, req, req->user_data);
5373 list_add_tail(&req->list, &ctx->defer_list);
5374 spin_unlock_irq(&ctx->completion_lock);
5375 return -EIOCBQUEUED;
5378 static void io_cleanup_req(struct io_kiocb *req)
5380 struct io_async_ctx *io = req->io;
5382 switch (req->opcode) {
5383 case IORING_OP_READV:
5384 case IORING_OP_READ_FIXED:
5385 case IORING_OP_READ:
5386 if (req->flags & REQ_F_BUFFER_SELECTED)
5387 kfree((void *)(unsigned long)req->rw.addr);
5389 case IORING_OP_WRITEV:
5390 case IORING_OP_WRITE_FIXED:
5391 case IORING_OP_WRITE:
5392 if (io->rw.iov != io->rw.fast_iov)
5395 case IORING_OP_RECVMSG:
5396 if (req->flags & REQ_F_BUFFER_SELECTED)
5397 kfree(req->sr_msg.kbuf);
5399 case IORING_OP_SENDMSG:
5400 if (io->msg.iov != io->msg.fast_iov)
5403 case IORING_OP_RECV:
5404 if (req->flags & REQ_F_BUFFER_SELECTED)
5405 kfree(req->sr_msg.kbuf);
5407 case IORING_OP_OPENAT:
5408 case IORING_OP_OPENAT2:
5410 case IORING_OP_SPLICE:
5412 io_put_file(req, req->splice.file_in,
5413 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5417 req->flags &= ~REQ_F_NEED_CLEANUP;
5420 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5421 bool force_nonblock, struct io_comp_state *cs)
5423 struct io_ring_ctx *ctx = req->ctx;
5426 switch (req->opcode) {
5428 ret = io_nop(req, cs);
5430 case IORING_OP_READV:
5431 case IORING_OP_READ_FIXED:
5432 case IORING_OP_READ:
5434 ret = io_read_prep(req, sqe, force_nonblock);
5438 ret = io_read(req, force_nonblock, cs);
5440 case IORING_OP_WRITEV:
5441 case IORING_OP_WRITE_FIXED:
5442 case IORING_OP_WRITE:
5444 ret = io_write_prep(req, sqe, force_nonblock);
5448 ret = io_write(req, force_nonblock, cs);
5450 case IORING_OP_FSYNC:
5452 ret = io_prep_fsync(req, sqe);
5456 ret = io_fsync(req, force_nonblock);
5458 case IORING_OP_POLL_ADD:
5460 ret = io_poll_add_prep(req, sqe);
5464 ret = io_poll_add(req);
5466 case IORING_OP_POLL_REMOVE:
5468 ret = io_poll_remove_prep(req, sqe);
5472 ret = io_poll_remove(req);
5474 case IORING_OP_SYNC_FILE_RANGE:
5476 ret = io_prep_sfr(req, sqe);
5480 ret = io_sync_file_range(req, force_nonblock);
5482 case IORING_OP_SENDMSG:
5483 case IORING_OP_SEND:
5485 ret = io_sendmsg_prep(req, sqe);
5489 if (req->opcode == IORING_OP_SENDMSG)
5490 ret = io_sendmsg(req, force_nonblock, cs);
5492 ret = io_send(req, force_nonblock, cs);
5494 case IORING_OP_RECVMSG:
5495 case IORING_OP_RECV:
5497 ret = io_recvmsg_prep(req, sqe);
5501 if (req->opcode == IORING_OP_RECVMSG)
5502 ret = io_recvmsg(req, force_nonblock, cs);
5504 ret = io_recv(req, force_nonblock, cs);
5506 case IORING_OP_TIMEOUT:
5508 ret = io_timeout_prep(req, sqe, false);
5512 ret = io_timeout(req);
5514 case IORING_OP_TIMEOUT_REMOVE:
5516 ret = io_timeout_remove_prep(req, sqe);
5520 ret = io_timeout_remove(req);
5522 case IORING_OP_ACCEPT:
5524 ret = io_accept_prep(req, sqe);
5528 ret = io_accept(req, force_nonblock, cs);
5530 case IORING_OP_CONNECT:
5532 ret = io_connect_prep(req, sqe);
5536 ret = io_connect(req, force_nonblock, cs);
5538 case IORING_OP_ASYNC_CANCEL:
5540 ret = io_async_cancel_prep(req, sqe);
5544 ret = io_async_cancel(req);
5546 case IORING_OP_FALLOCATE:
5548 ret = io_fallocate_prep(req, sqe);
5552 ret = io_fallocate(req, force_nonblock);
5554 case IORING_OP_OPENAT:
5556 ret = io_openat_prep(req, sqe);
5560 ret = io_openat(req, force_nonblock);
5562 case IORING_OP_CLOSE:
5564 ret = io_close_prep(req, sqe);
5568 ret = io_close(req, force_nonblock, cs);
5570 case IORING_OP_FILES_UPDATE:
5572 ret = io_files_update_prep(req, sqe);
5576 ret = io_files_update(req, force_nonblock, cs);
5578 case IORING_OP_STATX:
5580 ret = io_statx_prep(req, sqe);
5584 ret = io_statx(req, force_nonblock);
5586 case IORING_OP_FADVISE:
5588 ret = io_fadvise_prep(req, sqe);
5592 ret = io_fadvise(req, force_nonblock);
5594 case IORING_OP_MADVISE:
5596 ret = io_madvise_prep(req, sqe);
5600 ret = io_madvise(req, force_nonblock);
5602 case IORING_OP_OPENAT2:
5604 ret = io_openat2_prep(req, sqe);
5608 ret = io_openat2(req, force_nonblock);
5610 case IORING_OP_EPOLL_CTL:
5612 ret = io_epoll_ctl_prep(req, sqe);
5616 ret = io_epoll_ctl(req, force_nonblock, cs);
5618 case IORING_OP_SPLICE:
5620 ret = io_splice_prep(req, sqe);
5624 ret = io_splice(req, force_nonblock);
5626 case IORING_OP_PROVIDE_BUFFERS:
5628 ret = io_provide_buffers_prep(req, sqe);
5632 ret = io_provide_buffers(req, force_nonblock, cs);
5634 case IORING_OP_REMOVE_BUFFERS:
5636 ret = io_remove_buffers_prep(req, sqe);
5640 ret = io_remove_buffers(req, force_nonblock, cs);
5644 ret = io_tee_prep(req, sqe);
5648 ret = io_tee(req, force_nonblock);
5658 /* If the op doesn't have a file, we're not polling for it */
5659 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5660 const bool in_async = io_wq_current_is_worker();
5662 /* workqueue context doesn't hold uring_lock, grab it now */
5664 mutex_lock(&ctx->uring_lock);
5666 io_iopoll_req_issued(req);
5669 mutex_unlock(&ctx->uring_lock);
5675 static void io_arm_async_linked_timeout(struct io_kiocb *req)
5677 struct io_kiocb *link;
5679 /* link head's timeout is queued in io_queue_async_work() */
5680 if (!(req->flags & REQ_F_QUEUE_TIMEOUT))
5683 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
5684 io_queue_linked_timeout(link);
5687 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5689 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5692 io_arm_async_linked_timeout(req);
5694 /* if NO_CANCEL is set, we must still run the work */
5695 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5696 IO_WQ_WORK_CANCEL) {
5702 ret = io_issue_sqe(req, NULL, false, NULL);
5704 * We can get EAGAIN for polled IO even though we're
5705 * forcing a sync submission from here, since we can't
5706 * wait for request slots on the block side.
5715 req_set_fail_links(req);
5716 io_req_complete(req, ret);
5719 return io_steal_work(req);
5722 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5725 struct fixed_file_table *table;
5727 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5728 return table->files[index & IORING_FILE_TABLE_MASK];
5731 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5732 int fd, struct file **out_file, bool fixed)
5734 struct io_ring_ctx *ctx = req->ctx;
5738 if (unlikely(!ctx->file_data ||
5739 (unsigned) fd >= ctx->nr_user_files))
5741 fd = array_index_nospec(fd, ctx->nr_user_files);
5742 file = io_file_from_index(ctx, fd);
5744 req->fixed_file_refs = ctx->file_data->cur_refs;
5745 percpu_ref_get(req->fixed_file_refs);
5748 trace_io_uring_file_get(ctx, fd);
5749 file = __io_file_get(state, fd);
5752 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5759 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5764 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5765 if (unlikely(!fixed && io_async_submit(req->ctx)))
5768 return io_file_get(state, req, fd, &req->file, fixed);
5771 static int io_grab_files(struct io_kiocb *req)
5774 struct io_ring_ctx *ctx = req->ctx;
5776 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5778 if (!ctx->ring_file)
5782 spin_lock_irq(&ctx->inflight_lock);
5784 * We use the f_ops->flush() handler to ensure that we can flush
5785 * out work accessing these files if the fd is closed. Check if
5786 * the fd has changed since we started down this path, and disallow
5787 * this operation if it has.
5789 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5790 list_add(&req->inflight_entry, &ctx->inflight_list);
5791 req->flags |= REQ_F_INFLIGHT;
5792 req->work.files = current->files;
5795 spin_unlock_irq(&ctx->inflight_lock);
5801 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5803 struct io_timeout_data *data = container_of(timer,
5804 struct io_timeout_data, timer);
5805 struct io_kiocb *req = data->req;
5806 struct io_ring_ctx *ctx = req->ctx;
5807 struct io_kiocb *prev = NULL;
5808 unsigned long flags;
5810 spin_lock_irqsave(&ctx->completion_lock, flags);
5813 * We don't expect the list to be empty, that will only happen if we
5814 * race with the completion of the linked work.
5816 if (!list_empty(&req->link_list)) {
5817 prev = list_entry(req->link_list.prev, struct io_kiocb,
5819 if (refcount_inc_not_zero(&prev->refs)) {
5820 list_del_init(&req->link_list);
5821 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5826 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5829 req_set_fail_links(prev);
5830 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5833 io_req_complete(req, -ETIME);
5835 return HRTIMER_NORESTART;
5838 static void io_queue_linked_timeout(struct io_kiocb *req)
5840 struct io_ring_ctx *ctx = req->ctx;
5843 * If the list is now empty, then our linked request finished before
5844 * we got a chance to setup the timer
5846 spin_lock_irq(&ctx->completion_lock);
5847 if (!list_empty(&req->link_list)) {
5848 struct io_timeout_data *data = &req->io->timeout;
5850 data->timer.function = io_link_timeout_fn;
5851 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5854 spin_unlock_irq(&ctx->completion_lock);
5856 /* drop submission reference */
5860 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5862 struct io_kiocb *nxt;
5864 if (!(req->flags & REQ_F_LINK_HEAD))
5866 /* for polled retry, if flag is set, we already went through here */
5867 if (req->flags & REQ_F_POLLED)
5870 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5872 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5875 req->flags |= REQ_F_LINK_TIMEOUT;
5879 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5880 struct io_comp_state *cs)
5882 struct io_kiocb *linked_timeout;
5883 struct io_kiocb *nxt;
5884 const struct cred *old_creds = NULL;
5888 linked_timeout = io_prep_linked_timeout(req);
5890 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5891 req->work.creds != current_cred()) {
5893 revert_creds(old_creds);
5894 if (old_creds == req->work.creds)
5895 old_creds = NULL; /* restored original creds */
5897 old_creds = override_creds(req->work.creds);
5900 ret = io_issue_sqe(req, sqe, true, cs);
5903 * We async punt it if the file wasn't marked NOWAIT, or if the file
5904 * doesn't support non-blocking read/write attempts
5906 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5907 if (io_arm_poll_handler(req)) {
5909 io_queue_linked_timeout(linked_timeout);
5913 io_req_init_async(req);
5915 if (io_op_defs[req->opcode].file_table) {
5916 ret = io_grab_files(req);
5922 * Queued up for async execution, worker will release
5923 * submit reference when the iocb is actually submitted.
5925 io_queue_async_work(req);
5931 /* drop submission reference */
5932 io_put_req_find_next(req, &nxt);
5934 if (linked_timeout) {
5936 io_queue_linked_timeout(linked_timeout);
5938 io_put_req(linked_timeout);
5941 /* and drop final reference, if we failed */
5943 req_set_fail_links(req);
5944 io_req_complete(req, ret);
5949 if (req->flags & REQ_F_FORCE_ASYNC)
5955 revert_creds(old_creds);
5958 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5959 struct io_comp_state *cs)
5963 ret = io_req_defer(req, sqe);
5965 if (ret != -EIOCBQUEUED) {
5967 req_set_fail_links(req);
5969 io_req_complete(req, ret);
5971 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5974 if (io_alloc_async_ctx(req))
5976 ret = io_req_defer_prep(req, sqe, true);
5977 if (unlikely(ret < 0))
5982 * Never try inline submit of IOSQE_ASYNC is set, go straight
5983 * to async execution.
5985 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5986 io_queue_async_work(req);
5988 __io_queue_sqe(req, sqe, cs);
5992 static inline void io_queue_link_head(struct io_kiocb *req,
5993 struct io_comp_state *cs)
5995 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5997 io_req_complete(req, -ECANCELED);
5999 io_queue_sqe(req, NULL, cs);
6002 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6003 struct io_kiocb **link, struct io_comp_state *cs)
6005 struct io_ring_ctx *ctx = req->ctx;
6009 * If we already have a head request, queue this one for async
6010 * submittal once the head completes. If we don't have a head but
6011 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6012 * submitted sync once the chain is complete. If none of those
6013 * conditions are true (normal request), then just queue it.
6016 struct io_kiocb *head = *link;
6019 * Taking sequential execution of a link, draining both sides
6020 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6021 * requests in the link. So, it drains the head and the
6022 * next after the link request. The last one is done via
6023 * drain_next flag to persist the effect across calls.
6025 if (req->flags & REQ_F_IO_DRAIN) {
6026 head->flags |= REQ_F_IO_DRAIN;
6027 ctx->drain_next = 1;
6029 if (io_alloc_async_ctx(req))
6032 ret = io_req_defer_prep(req, sqe, false);
6034 /* fail even hard links since we don't submit */
6035 head->flags |= REQ_F_FAIL_LINK;
6038 trace_io_uring_link(ctx, req, head);
6039 io_get_req_task(req);
6040 list_add_tail(&req->link_list, &head->link_list);
6042 /* last request of a link, enqueue the link */
6043 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6044 io_queue_link_head(head, cs);
6048 if (unlikely(ctx->drain_next)) {
6049 req->flags |= REQ_F_IO_DRAIN;
6050 ctx->drain_next = 0;
6052 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6053 req->flags |= REQ_F_LINK_HEAD;
6054 INIT_LIST_HEAD(&req->link_list);
6056 if (io_alloc_async_ctx(req))
6059 ret = io_req_defer_prep(req, sqe, false);
6061 req->flags |= REQ_F_FAIL_LINK;
6064 io_queue_sqe(req, sqe, cs);
6072 * Batched submission is done, ensure local IO is flushed out.
6074 static void io_submit_state_end(struct io_submit_state *state)
6076 if (!list_empty(&state->comp.list))
6077 io_submit_flush_completions(&state->comp);
6078 blk_finish_plug(&state->plug);
6079 io_state_file_put(state);
6080 if (state->free_reqs)
6081 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6085 * Start submission side cache.
6087 static void io_submit_state_start(struct io_submit_state *state,
6088 struct io_ring_ctx *ctx, unsigned int max_ios)
6090 blk_start_plug(&state->plug);
6092 state->plug.nowait = true;
6095 INIT_LIST_HEAD(&state->comp.list);
6096 state->comp.ctx = ctx;
6097 state->free_reqs = 0;
6099 state->ios_left = max_ios;
6102 static void io_commit_sqring(struct io_ring_ctx *ctx)
6104 struct io_rings *rings = ctx->rings;
6107 * Ensure any loads from the SQEs are done at this point,
6108 * since once we write the new head, the application could
6109 * write new data to them.
6111 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6115 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6116 * that is mapped by userspace. This means that care needs to be taken to
6117 * ensure that reads are stable, as we cannot rely on userspace always
6118 * being a good citizen. If members of the sqe are validated and then later
6119 * used, it's important that those reads are done through READ_ONCE() to
6120 * prevent a re-load down the line.
6122 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6124 u32 *sq_array = ctx->sq_array;
6128 * The cached sq head (or cq tail) serves two purposes:
6130 * 1) allows us to batch the cost of updating the user visible
6132 * 2) allows the kernel side to track the head on its own, even
6133 * though the application is the one updating it.
6135 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6136 if (likely(head < ctx->sq_entries))
6137 return &ctx->sq_sqes[head];
6139 /* drop invalid entries */
6140 ctx->cached_sq_dropped++;
6141 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6145 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6147 ctx->cached_sq_head++;
6150 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6151 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6152 IOSQE_BUFFER_SELECT)
6154 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6155 const struct io_uring_sqe *sqe,
6156 struct io_submit_state *state)
6158 unsigned int sqe_flags;
6162 * All io need record the previous position, if LINK vs DARIN,
6163 * it can be used to mark the position of the first IO in the
6166 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
6167 req->opcode = READ_ONCE(sqe->opcode);
6168 req->user_data = READ_ONCE(sqe->user_data);
6173 /* one is dropped after submission, the other at completion */
6174 refcount_set(&req->refs, 2);
6175 req->task = current;
6178 if (unlikely(req->opcode >= IORING_OP_LAST))
6181 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6184 sqe_flags = READ_ONCE(sqe->flags);
6185 /* enforce forwards compatibility on users */
6186 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6189 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6190 !io_op_defs[req->opcode].buffer_select)
6193 id = READ_ONCE(sqe->personality);
6195 io_req_init_async(req);
6196 req->work.creds = idr_find(&ctx->personality_idr, id);
6197 if (unlikely(!req->work.creds))
6199 get_cred(req->work.creds);
6202 /* same numerical values with corresponding REQ_F_*, safe to copy */
6203 req->flags |= sqe_flags;
6205 if (!io_op_defs[req->opcode].needs_file)
6208 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6211 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6212 struct file *ring_file, int ring_fd)
6214 struct io_submit_state state;
6215 struct io_kiocb *link = NULL;
6216 int i, submitted = 0;
6218 /* if we have a backlog and couldn't flush it all, return BUSY */
6219 if (test_bit(0, &ctx->sq_check_overflow)) {
6220 if (!list_empty(&ctx->cq_overflow_list) &&
6221 !io_cqring_overflow_flush(ctx, false))
6225 /* make sure SQ entry isn't read before tail */
6226 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6228 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6231 io_submit_state_start(&state, ctx, nr);
6233 ctx->ring_fd = ring_fd;
6234 ctx->ring_file = ring_file;
6236 for (i = 0; i < nr; i++) {
6237 const struct io_uring_sqe *sqe;
6238 struct io_kiocb *req;
6241 sqe = io_get_sqe(ctx);
6242 if (unlikely(!sqe)) {
6243 io_consume_sqe(ctx);
6246 req = io_alloc_req(ctx, &state);
6247 if (unlikely(!req)) {
6249 submitted = -EAGAIN;
6253 err = io_init_req(ctx, req, sqe, &state);
6254 io_consume_sqe(ctx);
6255 /* will complete beyond this point, count as submitted */
6258 if (unlikely(err)) {
6261 io_req_complete(req, err);
6265 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6266 true, io_async_submit(ctx));
6267 err = io_submit_sqe(req, sqe, &link, &state.comp);
6272 if (unlikely(submitted != nr)) {
6273 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6275 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6278 io_queue_link_head(link, &state.comp);
6279 io_submit_state_end(&state);
6281 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6282 io_commit_sqring(ctx);
6287 static int io_sq_thread(void *data)
6289 struct io_ring_ctx *ctx = data;
6290 const struct cred *old_cred;
6292 unsigned long timeout;
6295 complete(&ctx->sq_thread_comp);
6297 old_cred = override_creds(ctx->creds);
6299 timeout = jiffies + ctx->sq_thread_idle;
6300 while (!kthread_should_park()) {
6301 unsigned int to_submit;
6303 if (!list_empty(&ctx->poll_list)) {
6304 unsigned nr_events = 0;
6306 mutex_lock(&ctx->uring_lock);
6307 if (!list_empty(&ctx->poll_list))
6308 io_iopoll_getevents(ctx, &nr_events, 0);
6310 timeout = jiffies + ctx->sq_thread_idle;
6311 mutex_unlock(&ctx->uring_lock);
6314 to_submit = io_sqring_entries(ctx);
6317 * If submit got -EBUSY, flag us as needing the application
6318 * to enter the kernel to reap and flush events.
6320 if (!to_submit || ret == -EBUSY || need_resched()) {
6322 * Drop cur_mm before scheduling, we can't hold it for
6323 * long periods (or over schedule()). Do this before
6324 * adding ourselves to the waitqueue, as the unuse/drop
6327 io_sq_thread_drop_mm(ctx);
6330 * We're polling. If we're within the defined idle
6331 * period, then let us spin without work before going
6332 * to sleep. The exception is if we got EBUSY doing
6333 * more IO, we should wait for the application to
6334 * reap events and wake us up.
6336 if (!list_empty(&ctx->poll_list) || need_resched() ||
6337 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6338 !percpu_ref_is_dying(&ctx->refs))) {
6339 if (current->task_works)
6345 prepare_to_wait(&ctx->sqo_wait, &wait,
6346 TASK_INTERRUPTIBLE);
6349 * While doing polled IO, before going to sleep, we need
6350 * to check if there are new reqs added to poll_list, it
6351 * is because reqs may have been punted to io worker and
6352 * will be added to poll_list later, hence check the
6355 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6356 !list_empty_careful(&ctx->poll_list)) {
6357 finish_wait(&ctx->sqo_wait, &wait);
6361 /* Tell userspace we may need a wakeup call */
6362 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6363 /* make sure to read SQ tail after writing flags */
6366 to_submit = io_sqring_entries(ctx);
6367 if (!to_submit || ret == -EBUSY) {
6368 if (kthread_should_park()) {
6369 finish_wait(&ctx->sqo_wait, &wait);
6372 if (current->task_works) {
6374 finish_wait(&ctx->sqo_wait, &wait);
6377 if (signal_pending(current))
6378 flush_signals(current);
6380 finish_wait(&ctx->sqo_wait, &wait);
6382 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6386 finish_wait(&ctx->sqo_wait, &wait);
6388 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6391 mutex_lock(&ctx->uring_lock);
6392 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6393 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6394 mutex_unlock(&ctx->uring_lock);
6395 timeout = jiffies + ctx->sq_thread_idle;
6398 if (current->task_works)
6401 io_sq_thread_drop_mm(ctx);
6402 revert_creds(old_cred);
6409 struct io_wait_queue {
6410 struct wait_queue_entry wq;
6411 struct io_ring_ctx *ctx;
6413 unsigned nr_timeouts;
6416 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6418 struct io_ring_ctx *ctx = iowq->ctx;
6421 * Wake up if we have enough events, or if a timeout occurred since we
6422 * started waiting. For timeouts, we always want to return to userspace,
6423 * regardless of event count.
6425 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6426 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6429 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6430 int wake_flags, void *key)
6432 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6435 /* use noflush == true, as we can't safely rely on locking context */
6436 if (!io_should_wake(iowq, true))
6439 return autoremove_wake_function(curr, mode, wake_flags, key);
6443 * Wait until events become available, if we don't already have some. The
6444 * application must reap them itself, as they reside on the shared cq ring.
6446 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6447 const sigset_t __user *sig, size_t sigsz)
6449 struct io_wait_queue iowq = {
6452 .func = io_wake_function,
6453 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6456 .to_wait = min_events,
6458 struct io_rings *rings = ctx->rings;
6462 if (io_cqring_events(ctx, false) >= min_events)
6464 if (!current->task_works)
6470 #ifdef CONFIG_COMPAT
6471 if (in_compat_syscall())
6472 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6476 ret = set_user_sigmask(sig, sigsz);
6482 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6483 trace_io_uring_cqring_wait(ctx, min_events);
6485 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6486 TASK_INTERRUPTIBLE);
6487 if (current->task_works)
6489 if (io_should_wake(&iowq, false))
6492 if (signal_pending(current)) {
6497 finish_wait(&ctx->wait, &iowq.wq);
6499 restore_saved_sigmask_unless(ret == -EINTR);
6501 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6504 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6506 #if defined(CONFIG_UNIX)
6507 if (ctx->ring_sock) {
6508 struct sock *sock = ctx->ring_sock->sk;
6509 struct sk_buff *skb;
6511 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6517 for (i = 0; i < ctx->nr_user_files; i++) {
6520 file = io_file_from_index(ctx, i);
6527 static void io_file_ref_kill(struct percpu_ref *ref)
6529 struct fixed_file_data *data;
6531 data = container_of(ref, struct fixed_file_data, refs);
6532 complete(&data->done);
6535 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6537 struct fixed_file_data *data = ctx->file_data;
6538 struct fixed_file_ref_node *ref_node = NULL;
6539 unsigned nr_tables, i;
6544 spin_lock(&data->lock);
6545 if (!list_empty(&data->ref_list))
6546 ref_node = list_first_entry(&data->ref_list,
6547 struct fixed_file_ref_node, node);
6548 spin_unlock(&data->lock);
6550 percpu_ref_kill(&ref_node->refs);
6552 percpu_ref_kill(&data->refs);
6554 /* wait for all refs nodes to complete */
6555 flush_delayed_work(&ctx->file_put_work);
6556 wait_for_completion(&data->done);
6558 __io_sqe_files_unregister(ctx);
6559 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6560 for (i = 0; i < nr_tables; i++)
6561 kfree(data->table[i].files);
6563 percpu_ref_exit(&data->refs);
6565 ctx->file_data = NULL;
6566 ctx->nr_user_files = 0;
6570 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6572 if (ctx->sqo_thread) {
6573 wait_for_completion(&ctx->sq_thread_comp);
6575 * The park is a bit of a work-around, without it we get
6576 * warning spews on shutdown with SQPOLL set and affinity
6577 * set to a single CPU.
6579 kthread_park(ctx->sqo_thread);
6580 kthread_stop(ctx->sqo_thread);
6581 ctx->sqo_thread = NULL;
6585 static void io_finish_async(struct io_ring_ctx *ctx)
6587 io_sq_thread_stop(ctx);
6590 io_wq_destroy(ctx->io_wq);
6595 #if defined(CONFIG_UNIX)
6597 * Ensure the UNIX gc is aware of our file set, so we are certain that
6598 * the io_uring can be safely unregistered on process exit, even if we have
6599 * loops in the file referencing.
6601 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6603 struct sock *sk = ctx->ring_sock->sk;
6604 struct scm_fp_list *fpl;
6605 struct sk_buff *skb;
6608 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6612 skb = alloc_skb(0, GFP_KERNEL);
6621 fpl->user = get_uid(ctx->user);
6622 for (i = 0; i < nr; i++) {
6623 struct file *file = io_file_from_index(ctx, i + offset);
6627 fpl->fp[nr_files] = get_file(file);
6628 unix_inflight(fpl->user, fpl->fp[nr_files]);
6633 fpl->max = SCM_MAX_FD;
6634 fpl->count = nr_files;
6635 UNIXCB(skb).fp = fpl;
6636 skb->destructor = unix_destruct_scm;
6637 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6638 skb_queue_head(&sk->sk_receive_queue, skb);
6640 for (i = 0; i < nr_files; i++)
6651 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6652 * causes regular reference counting to break down. We rely on the UNIX
6653 * garbage collection to take care of this problem for us.
6655 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6657 unsigned left, total;
6661 left = ctx->nr_user_files;
6663 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6665 ret = __io_sqe_files_scm(ctx, this_files, total);
6669 total += this_files;
6675 while (total < ctx->nr_user_files) {
6676 struct file *file = io_file_from_index(ctx, total);
6686 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6692 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6697 for (i = 0; i < nr_tables; i++) {
6698 struct fixed_file_table *table = &ctx->file_data->table[i];
6699 unsigned this_files;
6701 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6702 table->files = kcalloc(this_files, sizeof(struct file *),
6706 nr_files -= this_files;
6712 for (i = 0; i < nr_tables; i++) {
6713 struct fixed_file_table *table = &ctx->file_data->table[i];
6714 kfree(table->files);
6719 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6721 #if defined(CONFIG_UNIX)
6722 struct sock *sock = ctx->ring_sock->sk;
6723 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6724 struct sk_buff *skb;
6727 __skb_queue_head_init(&list);
6730 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6731 * remove this entry and rearrange the file array.
6733 skb = skb_dequeue(head);
6735 struct scm_fp_list *fp;
6737 fp = UNIXCB(skb).fp;
6738 for (i = 0; i < fp->count; i++) {
6741 if (fp->fp[i] != file)
6744 unix_notinflight(fp->user, fp->fp[i]);
6745 left = fp->count - 1 - i;
6747 memmove(&fp->fp[i], &fp->fp[i + 1],
6748 left * sizeof(struct file *));
6755 __skb_queue_tail(&list, skb);
6765 __skb_queue_tail(&list, skb);
6767 skb = skb_dequeue(head);
6770 if (skb_peek(&list)) {
6771 spin_lock_irq(&head->lock);
6772 while ((skb = __skb_dequeue(&list)) != NULL)
6773 __skb_queue_tail(head, skb);
6774 spin_unlock_irq(&head->lock);
6781 struct io_file_put {
6782 struct list_head list;
6786 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6788 struct fixed_file_data *file_data = ref_node->file_data;
6789 struct io_ring_ctx *ctx = file_data->ctx;
6790 struct io_file_put *pfile, *tmp;
6792 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6793 list_del(&pfile->list);
6794 io_ring_file_put(ctx, pfile->file);
6798 spin_lock(&file_data->lock);
6799 list_del(&ref_node->node);
6800 spin_unlock(&file_data->lock);
6802 percpu_ref_exit(&ref_node->refs);
6804 percpu_ref_put(&file_data->refs);
6807 static void io_file_put_work(struct work_struct *work)
6809 struct io_ring_ctx *ctx;
6810 struct llist_node *node;
6812 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6813 node = llist_del_all(&ctx->file_put_llist);
6816 struct fixed_file_ref_node *ref_node;
6817 struct llist_node *next = node->next;
6819 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6820 __io_file_put_work(ref_node);
6825 static void io_file_data_ref_zero(struct percpu_ref *ref)
6827 struct fixed_file_ref_node *ref_node;
6828 struct io_ring_ctx *ctx;
6832 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6833 ctx = ref_node->file_data->ctx;
6835 if (percpu_ref_is_dying(&ctx->file_data->refs))
6838 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6840 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6842 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6845 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6846 struct io_ring_ctx *ctx)
6848 struct fixed_file_ref_node *ref_node;
6850 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6852 return ERR_PTR(-ENOMEM);
6854 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6857 return ERR_PTR(-ENOMEM);
6859 INIT_LIST_HEAD(&ref_node->node);
6860 INIT_LIST_HEAD(&ref_node->file_list);
6861 ref_node->file_data = ctx->file_data;
6865 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6867 percpu_ref_exit(&ref_node->refs);
6871 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6874 __s32 __user *fds = (__s32 __user *) arg;
6879 struct fixed_file_ref_node *ref_node;
6885 if (nr_args > IORING_MAX_FIXED_FILES)
6888 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6889 if (!ctx->file_data)
6891 ctx->file_data->ctx = ctx;
6892 init_completion(&ctx->file_data->done);
6893 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6894 spin_lock_init(&ctx->file_data->lock);
6896 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6897 ctx->file_data->table = kcalloc(nr_tables,
6898 sizeof(struct fixed_file_table),
6900 if (!ctx->file_data->table) {
6901 kfree(ctx->file_data);
6902 ctx->file_data = NULL;
6906 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6907 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6908 kfree(ctx->file_data->table);
6909 kfree(ctx->file_data);
6910 ctx->file_data = NULL;
6914 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6915 percpu_ref_exit(&ctx->file_data->refs);
6916 kfree(ctx->file_data->table);
6917 kfree(ctx->file_data);
6918 ctx->file_data = NULL;
6922 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6923 struct fixed_file_table *table;
6927 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6929 /* allow sparse sets */
6935 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6936 index = i & IORING_FILE_TABLE_MASK;
6944 * Don't allow io_uring instances to be registered. If UNIX
6945 * isn't enabled, then this causes a reference cycle and this
6946 * instance can never get freed. If UNIX is enabled we'll
6947 * handle it just fine, but there's still no point in allowing
6948 * a ring fd as it doesn't support regular read/write anyway.
6950 if (file->f_op == &io_uring_fops) {
6955 table->files[index] = file;
6959 for (i = 0; i < ctx->nr_user_files; i++) {
6960 file = io_file_from_index(ctx, i);
6964 for (i = 0; i < nr_tables; i++)
6965 kfree(ctx->file_data->table[i].files);
6967 kfree(ctx->file_data->table);
6968 kfree(ctx->file_data);
6969 ctx->file_data = NULL;
6970 ctx->nr_user_files = 0;
6974 ret = io_sqe_files_scm(ctx);
6976 io_sqe_files_unregister(ctx);
6980 ref_node = alloc_fixed_file_ref_node(ctx);
6981 if (IS_ERR(ref_node)) {
6982 io_sqe_files_unregister(ctx);
6983 return PTR_ERR(ref_node);
6986 ctx->file_data->cur_refs = &ref_node->refs;
6987 spin_lock(&ctx->file_data->lock);
6988 list_add(&ref_node->node, &ctx->file_data->ref_list);
6989 spin_unlock(&ctx->file_data->lock);
6990 percpu_ref_get(&ctx->file_data->refs);
6994 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6997 #if defined(CONFIG_UNIX)
6998 struct sock *sock = ctx->ring_sock->sk;
6999 struct sk_buff_head *head = &sock->sk_receive_queue;
7000 struct sk_buff *skb;
7003 * See if we can merge this file into an existing skb SCM_RIGHTS
7004 * file set. If there's no room, fall back to allocating a new skb
7005 * and filling it in.
7007 spin_lock_irq(&head->lock);
7008 skb = skb_peek(head);
7010 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7012 if (fpl->count < SCM_MAX_FD) {
7013 __skb_unlink(skb, head);
7014 spin_unlock_irq(&head->lock);
7015 fpl->fp[fpl->count] = get_file(file);
7016 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7018 spin_lock_irq(&head->lock);
7019 __skb_queue_head(head, skb);
7024 spin_unlock_irq(&head->lock);
7031 return __io_sqe_files_scm(ctx, 1, index);
7037 static int io_queue_file_removal(struct fixed_file_data *data,
7040 struct io_file_put *pfile;
7041 struct percpu_ref *refs = data->cur_refs;
7042 struct fixed_file_ref_node *ref_node;
7044 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7048 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7050 list_add(&pfile->list, &ref_node->file_list);
7055 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7056 struct io_uring_files_update *up,
7059 struct fixed_file_data *data = ctx->file_data;
7060 struct fixed_file_ref_node *ref_node;
7065 bool needs_switch = false;
7067 if (check_add_overflow(up->offset, nr_args, &done))
7069 if (done > ctx->nr_user_files)
7072 ref_node = alloc_fixed_file_ref_node(ctx);
7073 if (IS_ERR(ref_node))
7074 return PTR_ERR(ref_node);
7077 fds = u64_to_user_ptr(up->fds);
7079 struct fixed_file_table *table;
7083 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7087 i = array_index_nospec(up->offset, ctx->nr_user_files);
7088 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7089 index = i & IORING_FILE_TABLE_MASK;
7090 if (table->files[index]) {
7091 file = io_file_from_index(ctx, index);
7092 err = io_queue_file_removal(data, file);
7095 table->files[index] = NULL;
7096 needs_switch = true;
7105 * Don't allow io_uring instances to be registered. If
7106 * UNIX isn't enabled, then this causes a reference
7107 * cycle and this instance can never get freed. If UNIX
7108 * is enabled we'll handle it just fine, but there's
7109 * still no point in allowing a ring fd as it doesn't
7110 * support regular read/write anyway.
7112 if (file->f_op == &io_uring_fops) {
7117 table->files[index] = file;
7118 err = io_sqe_file_register(ctx, file, i);
7128 percpu_ref_kill(data->cur_refs);
7129 spin_lock(&data->lock);
7130 list_add(&ref_node->node, &data->ref_list);
7131 data->cur_refs = &ref_node->refs;
7132 spin_unlock(&data->lock);
7133 percpu_ref_get(&ctx->file_data->refs);
7135 destroy_fixed_file_ref_node(ref_node);
7137 return done ? done : err;
7140 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7143 struct io_uring_files_update up;
7145 if (!ctx->file_data)
7149 if (copy_from_user(&up, arg, sizeof(up)))
7154 return __io_sqe_files_update(ctx, &up, nr_args);
7157 static void io_free_work(struct io_wq_work *work)
7159 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7161 /* Consider that io_steal_work() relies on this ref */
7165 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7166 struct io_uring_params *p)
7168 struct io_wq_data data;
7170 struct io_ring_ctx *ctx_attach;
7171 unsigned int concurrency;
7174 data.user = ctx->user;
7175 data.free_work = io_free_work;
7176 data.do_work = io_wq_submit_work;
7178 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7179 /* Do QD, or 4 * CPUS, whatever is smallest */
7180 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7182 ctx->io_wq = io_wq_create(concurrency, &data);
7183 if (IS_ERR(ctx->io_wq)) {
7184 ret = PTR_ERR(ctx->io_wq);
7190 f = fdget(p->wq_fd);
7194 if (f.file->f_op != &io_uring_fops) {
7199 ctx_attach = f.file->private_data;
7200 /* @io_wq is protected by holding the fd */
7201 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7206 ctx->io_wq = ctx_attach->io_wq;
7212 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7213 struct io_uring_params *p)
7217 mmgrab(current->mm);
7218 ctx->sqo_mm = current->mm;
7220 if (ctx->flags & IORING_SETUP_SQPOLL) {
7222 if (!capable(CAP_SYS_ADMIN))
7225 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7226 if (!ctx->sq_thread_idle)
7227 ctx->sq_thread_idle = HZ;
7229 if (p->flags & IORING_SETUP_SQ_AFF) {
7230 int cpu = p->sq_thread_cpu;
7233 if (cpu >= nr_cpu_ids)
7235 if (!cpu_online(cpu))
7238 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7242 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7245 if (IS_ERR(ctx->sqo_thread)) {
7246 ret = PTR_ERR(ctx->sqo_thread);
7247 ctx->sqo_thread = NULL;
7250 wake_up_process(ctx->sqo_thread);
7251 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7252 /* Can't have SQ_AFF without SQPOLL */
7257 ret = io_init_wq_offload(ctx, p);
7263 io_finish_async(ctx);
7264 mmdrop(ctx->sqo_mm);
7269 static inline void __io_unaccount_mem(struct user_struct *user,
7270 unsigned long nr_pages)
7272 atomic_long_sub(nr_pages, &user->locked_vm);
7275 static inline int __io_account_mem(struct user_struct *user,
7276 unsigned long nr_pages)
7278 unsigned long page_limit, cur_pages, new_pages;
7280 /* Don't allow more pages than we can safely lock */
7281 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7284 cur_pages = atomic_long_read(&user->locked_vm);
7285 new_pages = cur_pages + nr_pages;
7286 if (new_pages > page_limit)
7288 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7289 new_pages) != cur_pages);
7294 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7295 enum io_mem_account acct)
7298 __io_unaccount_mem(ctx->user, nr_pages);
7301 if (acct == ACCT_LOCKED)
7302 ctx->sqo_mm->locked_vm -= nr_pages;
7303 else if (acct == ACCT_PINNED)
7304 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7308 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7309 enum io_mem_account acct)
7313 if (ctx->limit_mem) {
7314 ret = __io_account_mem(ctx->user, nr_pages);
7320 if (acct == ACCT_LOCKED)
7321 ctx->sqo_mm->locked_vm += nr_pages;
7322 else if (acct == ACCT_PINNED)
7323 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7329 static void io_mem_free(void *ptr)
7336 page = virt_to_head_page(ptr);
7337 if (put_page_testzero(page))
7338 free_compound_page(page);
7341 static void *io_mem_alloc(size_t size)
7343 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7346 return (void *) __get_free_pages(gfp_flags, get_order(size));
7349 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7352 struct io_rings *rings;
7353 size_t off, sq_array_size;
7355 off = struct_size(rings, cqes, cq_entries);
7356 if (off == SIZE_MAX)
7360 off = ALIGN(off, SMP_CACHE_BYTES);
7365 sq_array_size = array_size(sizeof(u32), sq_entries);
7366 if (sq_array_size == SIZE_MAX)
7369 if (check_add_overflow(off, sq_array_size, &off))
7378 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7382 pages = (size_t)1 << get_order(
7383 rings_size(sq_entries, cq_entries, NULL));
7384 pages += (size_t)1 << get_order(
7385 array_size(sizeof(struct io_uring_sqe), sq_entries));
7390 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7394 if (!ctx->user_bufs)
7397 for (i = 0; i < ctx->nr_user_bufs; i++) {
7398 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7400 for (j = 0; j < imu->nr_bvecs; j++)
7401 unpin_user_page(imu->bvec[j].bv_page);
7403 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7408 kfree(ctx->user_bufs);
7409 ctx->user_bufs = NULL;
7410 ctx->nr_user_bufs = 0;
7414 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7415 void __user *arg, unsigned index)
7417 struct iovec __user *src;
7419 #ifdef CONFIG_COMPAT
7421 struct compat_iovec __user *ciovs;
7422 struct compat_iovec ciov;
7424 ciovs = (struct compat_iovec __user *) arg;
7425 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7428 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7429 dst->iov_len = ciov.iov_len;
7433 src = (struct iovec __user *) arg;
7434 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7439 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7442 struct vm_area_struct **vmas = NULL;
7443 struct page **pages = NULL;
7444 int i, j, got_pages = 0;
7449 if (!nr_args || nr_args > UIO_MAXIOV)
7452 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7454 if (!ctx->user_bufs)
7457 for (i = 0; i < nr_args; i++) {
7458 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7459 unsigned long off, start, end, ubuf;
7464 ret = io_copy_iov(ctx, &iov, arg, i);
7469 * Don't impose further limits on the size and buffer
7470 * constraints here, we'll -EINVAL later when IO is
7471 * submitted if they are wrong.
7474 if (!iov.iov_base || !iov.iov_len)
7477 /* arbitrary limit, but we need something */
7478 if (iov.iov_len > SZ_1G)
7481 ubuf = (unsigned long) iov.iov_base;
7482 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7483 start = ubuf >> PAGE_SHIFT;
7484 nr_pages = end - start;
7486 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7491 if (!pages || nr_pages > got_pages) {
7494 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7496 vmas = kvmalloc_array(nr_pages,
7497 sizeof(struct vm_area_struct *),
7499 if (!pages || !vmas) {
7501 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7504 got_pages = nr_pages;
7507 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7511 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7516 mmap_read_lock(current->mm);
7517 pret = pin_user_pages(ubuf, nr_pages,
7518 FOLL_WRITE | FOLL_LONGTERM,
7520 if (pret == nr_pages) {
7521 /* don't support file backed memory */
7522 for (j = 0; j < nr_pages; j++) {
7523 struct vm_area_struct *vma = vmas[j];
7526 !is_file_hugepages(vma->vm_file)) {
7532 ret = pret < 0 ? pret : -EFAULT;
7534 mmap_read_unlock(current->mm);
7537 * if we did partial map, or found file backed vmas,
7538 * release any pages we did get
7541 unpin_user_pages(pages, pret);
7542 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7547 off = ubuf & ~PAGE_MASK;
7549 for (j = 0; j < nr_pages; j++) {
7552 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7553 imu->bvec[j].bv_page = pages[j];
7554 imu->bvec[j].bv_len = vec_len;
7555 imu->bvec[j].bv_offset = off;
7559 /* store original address for later verification */
7561 imu->len = iov.iov_len;
7562 imu->nr_bvecs = nr_pages;
7564 ctx->nr_user_bufs++;
7572 io_sqe_buffer_unregister(ctx);
7576 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7578 __s32 __user *fds = arg;
7584 if (copy_from_user(&fd, fds, sizeof(*fds)))
7587 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7588 if (IS_ERR(ctx->cq_ev_fd)) {
7589 int ret = PTR_ERR(ctx->cq_ev_fd);
7590 ctx->cq_ev_fd = NULL;
7597 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7599 if (ctx->cq_ev_fd) {
7600 eventfd_ctx_put(ctx->cq_ev_fd);
7601 ctx->cq_ev_fd = NULL;
7608 static int __io_destroy_buffers(int id, void *p, void *data)
7610 struct io_ring_ctx *ctx = data;
7611 struct io_buffer *buf = p;
7613 __io_remove_buffers(ctx, buf, id, -1U);
7617 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7619 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7620 idr_destroy(&ctx->io_buffer_idr);
7623 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7625 io_finish_async(ctx);
7627 mmdrop(ctx->sqo_mm);
7631 io_iopoll_reap_events(ctx);
7632 io_sqe_buffer_unregister(ctx);
7633 io_sqe_files_unregister(ctx);
7634 io_eventfd_unregister(ctx);
7635 io_destroy_buffers(ctx);
7636 idr_destroy(&ctx->personality_idr);
7638 #if defined(CONFIG_UNIX)
7639 if (ctx->ring_sock) {
7640 ctx->ring_sock->file = NULL; /* so that iput() is called */
7641 sock_release(ctx->ring_sock);
7645 io_mem_free(ctx->rings);
7646 io_mem_free(ctx->sq_sqes);
7648 percpu_ref_exit(&ctx->refs);
7649 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7651 free_uid(ctx->user);
7652 put_cred(ctx->creds);
7653 kfree(ctx->cancel_hash);
7654 kmem_cache_free(req_cachep, ctx->fallback_req);
7658 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7660 struct io_ring_ctx *ctx = file->private_data;
7663 poll_wait(file, &ctx->cq_wait, wait);
7665 * synchronizes with barrier from wq_has_sleeper call in
7669 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7670 ctx->rings->sq_ring_entries)
7671 mask |= EPOLLOUT | EPOLLWRNORM;
7672 if (io_cqring_events(ctx, false))
7673 mask |= EPOLLIN | EPOLLRDNORM;
7678 static int io_uring_fasync(int fd, struct file *file, int on)
7680 struct io_ring_ctx *ctx = file->private_data;
7682 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7685 static int io_remove_personalities(int id, void *p, void *data)
7687 struct io_ring_ctx *ctx = data;
7688 const struct cred *cred;
7690 cred = idr_remove(&ctx->personality_idr, id);
7696 static void io_ring_exit_work(struct work_struct *work)
7698 struct io_ring_ctx *ctx;
7700 ctx = container_of(work, struct io_ring_ctx, exit_work);
7702 io_cqring_overflow_flush(ctx, true);
7705 * If we're doing polled IO and end up having requests being
7706 * submitted async (out-of-line), then completions can come in while
7707 * we're waiting for refs to drop. We need to reap these manually,
7708 * as nobody else will be looking for them.
7710 while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20)) {
7711 io_iopoll_reap_events(ctx);
7713 io_cqring_overflow_flush(ctx, true);
7715 io_ring_ctx_free(ctx);
7718 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7720 mutex_lock(&ctx->uring_lock);
7721 percpu_ref_kill(&ctx->refs);
7722 mutex_unlock(&ctx->uring_lock);
7724 io_kill_timeouts(ctx);
7725 io_poll_remove_all(ctx);
7728 io_wq_cancel_all(ctx->io_wq);
7730 io_iopoll_reap_events(ctx);
7731 /* if we failed setting up the ctx, we might not have any rings */
7733 io_cqring_overflow_flush(ctx, true);
7734 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7735 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7736 queue_work(system_wq, &ctx->exit_work);
7739 static int io_uring_release(struct inode *inode, struct file *file)
7741 struct io_ring_ctx *ctx = file->private_data;
7743 file->private_data = NULL;
7744 io_ring_ctx_wait_and_kill(ctx);
7748 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7750 struct files_struct *files = data;
7752 return work->files == files;
7755 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7756 struct files_struct *files)
7758 if (list_empty_careful(&ctx->inflight_list))
7761 /* cancel all at once, should be faster than doing it one by one*/
7762 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7764 while (!list_empty_careful(&ctx->inflight_list)) {
7765 struct io_kiocb *cancel_req = NULL, *req;
7768 spin_lock_irq(&ctx->inflight_lock);
7769 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7770 if (req->work.files != files)
7772 /* req is being completed, ignore */
7773 if (!refcount_inc_not_zero(&req->refs))
7779 prepare_to_wait(&ctx->inflight_wait, &wait,
7780 TASK_UNINTERRUPTIBLE);
7781 spin_unlock_irq(&ctx->inflight_lock);
7783 /* We need to keep going until we don't find a matching req */
7787 if (cancel_req->flags & REQ_F_OVERFLOW) {
7788 spin_lock_irq(&ctx->completion_lock);
7789 list_del(&cancel_req->list);
7790 cancel_req->flags &= ~REQ_F_OVERFLOW;
7791 if (list_empty(&ctx->cq_overflow_list)) {
7792 clear_bit(0, &ctx->sq_check_overflow);
7793 clear_bit(0, &ctx->cq_check_overflow);
7795 spin_unlock_irq(&ctx->completion_lock);
7797 WRITE_ONCE(ctx->rings->cq_overflow,
7798 atomic_inc_return(&ctx->cached_cq_overflow));
7801 * Put inflight ref and overflow ref. If that's
7802 * all we had, then we're done with this request.
7804 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7805 io_free_req(cancel_req);
7806 finish_wait(&ctx->inflight_wait, &wait);
7810 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7811 io_put_req(cancel_req);
7815 finish_wait(&ctx->inflight_wait, &wait);
7819 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7821 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7822 struct task_struct *task = data;
7824 return req->task == task;
7827 static int io_uring_flush(struct file *file, void *data)
7829 struct io_ring_ctx *ctx = file->private_data;
7831 io_uring_cancel_files(ctx, data);
7834 * If the task is going away, cancel work it may have pending
7836 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7837 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7842 static void *io_uring_validate_mmap_request(struct file *file,
7843 loff_t pgoff, size_t sz)
7845 struct io_ring_ctx *ctx = file->private_data;
7846 loff_t offset = pgoff << PAGE_SHIFT;
7851 case IORING_OFF_SQ_RING:
7852 case IORING_OFF_CQ_RING:
7855 case IORING_OFF_SQES:
7859 return ERR_PTR(-EINVAL);
7862 page = virt_to_head_page(ptr);
7863 if (sz > page_size(page))
7864 return ERR_PTR(-EINVAL);
7871 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7873 size_t sz = vma->vm_end - vma->vm_start;
7877 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7879 return PTR_ERR(ptr);
7881 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7882 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7885 #else /* !CONFIG_MMU */
7887 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7889 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7892 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7894 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7897 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7898 unsigned long addr, unsigned long len,
7899 unsigned long pgoff, unsigned long flags)
7903 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7905 return PTR_ERR(ptr);
7907 return (unsigned long) ptr;
7910 #endif /* !CONFIG_MMU */
7912 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7913 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7916 struct io_ring_ctx *ctx;
7921 if (current->task_works)
7924 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7932 if (f.file->f_op != &io_uring_fops)
7936 ctx = f.file->private_data;
7937 if (!percpu_ref_tryget(&ctx->refs))
7941 * For SQ polling, the thread will do all submissions and completions.
7942 * Just return the requested submit count, and wake the thread if
7946 if (ctx->flags & IORING_SETUP_SQPOLL) {
7947 if (!list_empty_careful(&ctx->cq_overflow_list))
7948 io_cqring_overflow_flush(ctx, false);
7949 if (flags & IORING_ENTER_SQ_WAKEUP)
7950 wake_up(&ctx->sqo_wait);
7951 submitted = to_submit;
7952 } else if (to_submit) {
7953 mutex_lock(&ctx->uring_lock);
7954 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
7955 mutex_unlock(&ctx->uring_lock);
7957 if (submitted != to_submit)
7960 if (flags & IORING_ENTER_GETEVENTS) {
7961 unsigned nr_events = 0;
7963 min_complete = min(min_complete, ctx->cq_entries);
7966 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7967 * space applications don't need to do io completion events
7968 * polling again, they can rely on io_sq_thread to do polling
7969 * work, which can reduce cpu usage and uring_lock contention.
7971 if (ctx->flags & IORING_SETUP_IOPOLL &&
7972 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7973 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7975 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7980 percpu_ref_put(&ctx->refs);
7983 return submitted ? submitted : ret;
7986 #ifdef CONFIG_PROC_FS
7987 static int io_uring_show_cred(int id, void *p, void *data)
7989 const struct cred *cred = p;
7990 struct seq_file *m = data;
7991 struct user_namespace *uns = seq_user_ns(m);
7992 struct group_info *gi;
7997 seq_printf(m, "%5d\n", id);
7998 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7999 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8000 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8001 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8002 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8003 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8004 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8005 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8006 seq_puts(m, "\n\tGroups:\t");
8007 gi = cred->group_info;
8008 for (g = 0; g < gi->ngroups; g++) {
8009 seq_put_decimal_ull(m, g ? " " : "",
8010 from_kgid_munged(uns, gi->gid[g]));
8012 seq_puts(m, "\n\tCapEff:\t");
8013 cap = cred->cap_effective;
8014 CAP_FOR_EACH_U32(__capi)
8015 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8020 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8024 mutex_lock(&ctx->uring_lock);
8025 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8026 for (i = 0; i < ctx->nr_user_files; i++) {
8027 struct fixed_file_table *table;
8030 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8031 f = table->files[i & IORING_FILE_TABLE_MASK];
8033 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8035 seq_printf(m, "%5u: <none>\n", i);
8037 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8038 for (i = 0; i < ctx->nr_user_bufs; i++) {
8039 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8041 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8042 (unsigned int) buf->len);
8044 if (!idr_is_empty(&ctx->personality_idr)) {
8045 seq_printf(m, "Personalities:\n");
8046 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8048 seq_printf(m, "PollList:\n");
8049 spin_lock_irq(&ctx->completion_lock);
8050 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8051 struct hlist_head *list = &ctx->cancel_hash[i];
8052 struct io_kiocb *req;
8054 hlist_for_each_entry(req, list, hash_node)
8055 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8056 req->task->task_works != NULL);
8058 spin_unlock_irq(&ctx->completion_lock);
8059 mutex_unlock(&ctx->uring_lock);
8062 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8064 struct io_ring_ctx *ctx = f->private_data;
8066 if (percpu_ref_tryget(&ctx->refs)) {
8067 __io_uring_show_fdinfo(ctx, m);
8068 percpu_ref_put(&ctx->refs);
8073 static const struct file_operations io_uring_fops = {
8074 .release = io_uring_release,
8075 .flush = io_uring_flush,
8076 .mmap = io_uring_mmap,
8078 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8079 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8081 .poll = io_uring_poll,
8082 .fasync = io_uring_fasync,
8083 #ifdef CONFIG_PROC_FS
8084 .show_fdinfo = io_uring_show_fdinfo,
8088 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8089 struct io_uring_params *p)
8091 struct io_rings *rings;
8092 size_t size, sq_array_offset;
8094 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8095 if (size == SIZE_MAX)
8098 rings = io_mem_alloc(size);
8103 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8104 rings->sq_ring_mask = p->sq_entries - 1;
8105 rings->cq_ring_mask = p->cq_entries - 1;
8106 rings->sq_ring_entries = p->sq_entries;
8107 rings->cq_ring_entries = p->cq_entries;
8108 ctx->sq_mask = rings->sq_ring_mask;
8109 ctx->cq_mask = rings->cq_ring_mask;
8110 ctx->sq_entries = rings->sq_ring_entries;
8111 ctx->cq_entries = rings->cq_ring_entries;
8113 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8114 if (size == SIZE_MAX) {
8115 io_mem_free(ctx->rings);
8120 ctx->sq_sqes = io_mem_alloc(size);
8121 if (!ctx->sq_sqes) {
8122 io_mem_free(ctx->rings);
8131 * Allocate an anonymous fd, this is what constitutes the application
8132 * visible backing of an io_uring instance. The application mmaps this
8133 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8134 * we have to tie this fd to a socket for file garbage collection purposes.
8136 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8141 #if defined(CONFIG_UNIX)
8142 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8148 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8152 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8153 O_RDWR | O_CLOEXEC);
8156 ret = PTR_ERR(file);
8160 #if defined(CONFIG_UNIX)
8161 ctx->ring_sock->file = file;
8163 fd_install(ret, file);
8166 #if defined(CONFIG_UNIX)
8167 sock_release(ctx->ring_sock);
8168 ctx->ring_sock = NULL;
8173 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8174 struct io_uring_params __user *params)
8176 struct user_struct *user = NULL;
8177 struct io_ring_ctx *ctx;
8183 if (entries > IORING_MAX_ENTRIES) {
8184 if (!(p->flags & IORING_SETUP_CLAMP))
8186 entries = IORING_MAX_ENTRIES;
8190 * Use twice as many entries for the CQ ring. It's possible for the
8191 * application to drive a higher depth than the size of the SQ ring,
8192 * since the sqes are only used at submission time. This allows for
8193 * some flexibility in overcommitting a bit. If the application has
8194 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8195 * of CQ ring entries manually.
8197 p->sq_entries = roundup_pow_of_two(entries);
8198 if (p->flags & IORING_SETUP_CQSIZE) {
8200 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8201 * to a power-of-two, if it isn't already. We do NOT impose
8202 * any cq vs sq ring sizing.
8204 if (p->cq_entries < p->sq_entries)
8206 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8207 if (!(p->flags & IORING_SETUP_CLAMP))
8209 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8211 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8213 p->cq_entries = 2 * p->sq_entries;
8216 user = get_uid(current_user());
8217 limit_mem = !capable(CAP_IPC_LOCK);
8220 ret = __io_account_mem(user,
8221 ring_pages(p->sq_entries, p->cq_entries));
8228 ctx = io_ring_ctx_alloc(p);
8231 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8236 ctx->compat = in_compat_syscall();
8238 ctx->creds = get_current_cred();
8240 ret = io_allocate_scq_urings(ctx, p);
8244 ret = io_sq_offload_start(ctx, p);
8248 memset(&p->sq_off, 0, sizeof(p->sq_off));
8249 p->sq_off.head = offsetof(struct io_rings, sq.head);
8250 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8251 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8252 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8253 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8254 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8255 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8257 memset(&p->cq_off, 0, sizeof(p->cq_off));
8258 p->cq_off.head = offsetof(struct io_rings, cq.head);
8259 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8260 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8261 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8262 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8263 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8264 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8266 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8267 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8268 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8269 IORING_FEAT_POLL_32BITS;
8271 if (copy_to_user(params, p, sizeof(*p))) {
8276 * Install ring fd as the very last thing, so we don't risk someone
8277 * having closed it before we finish setup
8279 ret = io_uring_get_fd(ctx);
8283 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8284 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8286 ctx->limit_mem = limit_mem;
8289 io_ring_ctx_wait_and_kill(ctx);
8294 * Sets up an aio uring context, and returns the fd. Applications asks for a
8295 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8296 * params structure passed in.
8298 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8300 struct io_uring_params p;
8303 if (copy_from_user(&p, params, sizeof(p)))
8305 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8310 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8311 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8312 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8315 return io_uring_create(entries, &p, params);
8318 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8319 struct io_uring_params __user *, params)
8321 return io_uring_setup(entries, params);
8324 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8326 struct io_uring_probe *p;
8330 size = struct_size(p, ops, nr_args);
8331 if (size == SIZE_MAX)
8333 p = kzalloc(size, GFP_KERNEL);
8338 if (copy_from_user(p, arg, size))
8341 if (memchr_inv(p, 0, size))
8344 p->last_op = IORING_OP_LAST - 1;
8345 if (nr_args > IORING_OP_LAST)
8346 nr_args = IORING_OP_LAST;
8348 for (i = 0; i < nr_args; i++) {
8350 if (!io_op_defs[i].not_supported)
8351 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8356 if (copy_to_user(arg, p, size))
8363 static int io_register_personality(struct io_ring_ctx *ctx)
8365 const struct cred *creds = get_current_cred();
8368 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8369 USHRT_MAX, GFP_KERNEL);
8375 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8377 const struct cred *old_creds;
8379 old_creds = idr_remove(&ctx->personality_idr, id);
8381 put_cred(old_creds);
8388 static bool io_register_op_must_quiesce(int op)
8391 case IORING_UNREGISTER_FILES:
8392 case IORING_REGISTER_FILES_UPDATE:
8393 case IORING_REGISTER_PROBE:
8394 case IORING_REGISTER_PERSONALITY:
8395 case IORING_UNREGISTER_PERSONALITY:
8402 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8403 void __user *arg, unsigned nr_args)
8404 __releases(ctx->uring_lock)
8405 __acquires(ctx->uring_lock)
8410 * We're inside the ring mutex, if the ref is already dying, then
8411 * someone else killed the ctx or is already going through
8412 * io_uring_register().
8414 if (percpu_ref_is_dying(&ctx->refs))
8417 if (io_register_op_must_quiesce(opcode)) {
8418 percpu_ref_kill(&ctx->refs);
8421 * Drop uring mutex before waiting for references to exit. If
8422 * another thread is currently inside io_uring_enter() it might
8423 * need to grab the uring_lock to make progress. If we hold it
8424 * here across the drain wait, then we can deadlock. It's safe
8425 * to drop the mutex here, since no new references will come in
8426 * after we've killed the percpu ref.
8428 mutex_unlock(&ctx->uring_lock);
8429 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8430 mutex_lock(&ctx->uring_lock);
8432 percpu_ref_resurrect(&ctx->refs);
8439 case IORING_REGISTER_BUFFERS:
8440 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8442 case IORING_UNREGISTER_BUFFERS:
8446 ret = io_sqe_buffer_unregister(ctx);
8448 case IORING_REGISTER_FILES:
8449 ret = io_sqe_files_register(ctx, arg, nr_args);
8451 case IORING_UNREGISTER_FILES:
8455 ret = io_sqe_files_unregister(ctx);
8457 case IORING_REGISTER_FILES_UPDATE:
8458 ret = io_sqe_files_update(ctx, arg, nr_args);
8460 case IORING_REGISTER_EVENTFD:
8461 case IORING_REGISTER_EVENTFD_ASYNC:
8465 ret = io_eventfd_register(ctx, arg);
8468 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8469 ctx->eventfd_async = 1;
8471 ctx->eventfd_async = 0;
8473 case IORING_UNREGISTER_EVENTFD:
8477 ret = io_eventfd_unregister(ctx);
8479 case IORING_REGISTER_PROBE:
8481 if (!arg || nr_args > 256)
8483 ret = io_probe(ctx, arg, nr_args);
8485 case IORING_REGISTER_PERSONALITY:
8489 ret = io_register_personality(ctx);
8491 case IORING_UNREGISTER_PERSONALITY:
8495 ret = io_unregister_personality(ctx, nr_args);
8502 if (io_register_op_must_quiesce(opcode)) {
8503 /* bring the ctx back to life */
8504 percpu_ref_reinit(&ctx->refs);
8506 reinit_completion(&ctx->ref_comp);
8511 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8512 void __user *, arg, unsigned int, nr_args)
8514 struct io_ring_ctx *ctx;
8523 if (f.file->f_op != &io_uring_fops)
8526 ctx = f.file->private_data;
8528 mutex_lock(&ctx->uring_lock);
8529 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8530 mutex_unlock(&ctx->uring_lock);
8531 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8532 ctx->cq_ev_fd != NULL, ret);
8538 static int __init io_uring_init(void)
8540 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8541 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8542 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8545 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8546 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8547 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8548 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8549 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8550 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8551 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8552 BUILD_BUG_SQE_ELEM(8, __u64, off);
8553 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8554 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8555 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8556 BUILD_BUG_SQE_ELEM(24, __u32, len);
8557 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8558 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8559 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8560 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8561 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8562 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8563 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8564 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8565 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8566 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8567 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8568 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8569 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8570 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8571 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8572 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8573 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8574 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8575 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8577 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8578 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8579 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8582 __initcall(io_uring_init);