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>
82 #include <linux/io_uring.h>
84 #define CREATE_TRACE_POINTS
85 #include <trace/events/io_uring.h>
87 #include <uapi/linux/io_uring.h>
92 #define IORING_MAX_ENTRIES 32768
93 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
96 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
98 #define IORING_FILE_TABLE_SHIFT 9
99 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
100 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
101 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 u32 head ____cacheline_aligned_in_smp;
105 u32 tail ____cacheline_aligned_in_smp;
109 * This data is shared with the application through the mmap at offsets
110 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
112 * The offsets to the member fields are published through struct
113 * io_sqring_offsets when calling io_uring_setup.
117 * Head and tail offsets into the ring; the offsets need to be
118 * masked to get valid indices.
120 * The kernel controls head of the sq ring and the tail of the cq ring,
121 * and the application controls tail of the sq ring and the head of the
124 struct io_uring sq, cq;
126 * Bitmasks to apply to head and tail offsets (constant, equals
129 u32 sq_ring_mask, cq_ring_mask;
130 /* Ring sizes (constant, power of 2) */
131 u32 sq_ring_entries, cq_ring_entries;
133 * Number of invalid entries dropped by the kernel due to
134 * invalid index stored in array
136 * Written by the kernel, shouldn't be modified by the
137 * application (i.e. get number of "new events" by comparing to
140 * After a new SQ head value was read by the application this
141 * counter includes all submissions that were dropped reaching
142 * the new SQ head (and possibly more).
148 * Written by the kernel, shouldn't be modified by the
151 * The application needs a full memory barrier before checking
152 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
158 * Written by the application, shouldn't be modified by the
163 * Number of completion events lost because the queue was full;
164 * this should be avoided by the application by making sure
165 * there are not more requests pending than there is space in
166 * the completion queue.
168 * Written by the kernel, shouldn't be modified by the
169 * application (i.e. get number of "new events" by comparing to
172 * As completion events come in out of order this counter is not
173 * ordered with any other data.
177 * Ring buffer of completion events.
179 * The kernel writes completion events fresh every time they are
180 * produced, so the application is allowed to modify pending
183 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
186 struct io_mapped_ubuf {
189 struct bio_vec *bvec;
190 unsigned int nr_bvecs;
193 struct fixed_file_table {
197 struct fixed_file_ref_node {
198 struct percpu_ref refs;
199 struct list_head node;
200 struct list_head file_list;
201 struct fixed_file_data *file_data;
202 struct llist_node llist;
205 struct fixed_file_data {
206 struct fixed_file_table *table;
207 struct io_ring_ctx *ctx;
209 struct percpu_ref *cur_refs;
210 struct percpu_ref refs;
211 struct completion done;
212 struct list_head ref_list;
217 struct list_head list;
225 struct percpu_ref refs;
226 } ____cacheline_aligned_in_smp;
230 unsigned int compat: 1;
231 unsigned int limit_mem: 1;
232 unsigned int cq_overflow_flushed: 1;
233 unsigned int drain_next: 1;
234 unsigned int eventfd_async: 1;
237 * Ring buffer of indices into array of io_uring_sqe, which is
238 * mmapped by the application using the IORING_OFF_SQES offset.
240 * This indirection could e.g. be used to assign fixed
241 * io_uring_sqe entries to operations and only submit them to
242 * the queue when needed.
244 * The kernel modifies neither the indices array nor the entries
248 unsigned cached_sq_head;
251 unsigned sq_thread_idle;
252 unsigned cached_sq_dropped;
253 atomic_t cached_cq_overflow;
254 unsigned long sq_check_overflow;
256 struct list_head defer_list;
257 struct list_head timeout_list;
258 struct list_head cq_overflow_list;
260 wait_queue_head_t inflight_wait;
261 struct io_uring_sqe *sq_sqes;
262 } ____cacheline_aligned_in_smp;
264 struct io_rings *rings;
268 struct task_struct *sqo_thread; /* if using sq thread polling */
271 * For SQPOLL usage - we hold a reference to the parent task, so we
272 * have access to the ->files
274 struct task_struct *sqo_task;
276 /* Only used for accounting purposes */
277 struct mm_struct *mm_account;
279 wait_queue_head_t sqo_wait;
282 * If used, fixed file set. Writers must ensure that ->refs is dead,
283 * readers must ensure that ->refs is alive as long as the file* is
284 * used. Only updated through io_uring_register(2).
286 struct fixed_file_data *file_data;
287 unsigned nr_user_files;
289 /* if used, fixed mapped user buffers */
290 unsigned nr_user_bufs;
291 struct io_mapped_ubuf *user_bufs;
293 struct user_struct *user;
295 const struct cred *creds;
297 struct completion ref_comp;
298 struct completion sq_thread_comp;
300 /* if all else fails... */
301 struct io_kiocb *fallback_req;
303 #if defined(CONFIG_UNIX)
304 struct socket *ring_sock;
307 struct idr io_buffer_idr;
309 struct idr personality_idr;
312 unsigned cached_cq_tail;
315 atomic_t cq_timeouts;
316 unsigned long cq_check_overflow;
317 struct wait_queue_head cq_wait;
318 struct fasync_struct *cq_fasync;
319 struct eventfd_ctx *cq_ev_fd;
320 } ____cacheline_aligned_in_smp;
323 struct mutex uring_lock;
324 wait_queue_head_t wait;
325 } ____cacheline_aligned_in_smp;
328 spinlock_t completion_lock;
331 * ->iopoll_list is protected by the ctx->uring_lock for
332 * io_uring instances that don't use IORING_SETUP_SQPOLL.
333 * For SQPOLL, only the single threaded io_sq_thread() will
334 * manipulate the list, hence no extra locking is needed there.
336 struct list_head iopoll_list;
337 struct hlist_head *cancel_hash;
338 unsigned cancel_hash_bits;
339 bool poll_multi_file;
341 spinlock_t inflight_lock;
342 struct list_head inflight_list;
343 } ____cacheline_aligned_in_smp;
345 struct delayed_work file_put_work;
346 struct llist_head file_put_llist;
348 struct work_struct exit_work;
352 * First field must be the file pointer in all the
353 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
355 struct io_poll_iocb {
358 struct wait_queue_head *head;
364 struct wait_queue_entry wait;
369 struct file *put_file;
373 struct io_timeout_data {
374 struct io_kiocb *req;
375 struct hrtimer timer;
376 struct timespec64 ts;
377 enum hrtimer_mode mode;
382 struct sockaddr __user *addr;
383 int __user *addr_len;
385 unsigned long nofile;
407 struct list_head list;
411 /* NOTE: kiocb has the file as the first member, so don't do it here */
419 struct sockaddr __user *addr;
426 struct user_msghdr __user *umsg;
432 struct io_buffer *kbuf;
438 struct filename *filename;
440 unsigned long nofile;
443 struct io_files_update {
469 struct epoll_event event;
473 struct file *file_out;
474 struct file *file_in;
481 struct io_provide_buf {
495 const char __user *filename;
496 struct statx __user *buffer;
499 struct io_completion {
501 struct list_head list;
505 struct io_async_connect {
506 struct sockaddr_storage address;
509 struct io_async_msghdr {
510 struct iovec fast_iov[UIO_FASTIOV];
512 struct sockaddr __user *uaddr;
514 struct sockaddr_storage addr;
518 struct iovec fast_iov[UIO_FASTIOV];
519 const struct iovec *free_iovec;
520 struct iov_iter iter;
522 struct wait_page_queue wpq;
525 struct io_async_ctx {
527 struct io_async_rw rw;
528 struct io_async_msghdr msg;
529 struct io_async_connect connect;
530 struct io_timeout_data timeout;
535 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
536 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
537 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
538 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
539 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
540 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
547 REQ_F_LINK_TIMEOUT_BIT,
549 REQ_F_COMP_LOCKED_BIT,
550 REQ_F_NEED_CLEANUP_BIT,
552 REQ_F_BUFFER_SELECTED_BIT,
553 REQ_F_NO_FILE_TABLE_BIT,
554 REQ_F_WORK_INITIALIZED_BIT,
556 /* not a real bit, just to check we're not overflowing the space */
562 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
563 /* drain existing IO first */
564 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
566 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
567 /* doesn't sever on completion < 0 */
568 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
570 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
571 /* IOSQE_BUFFER_SELECT */
572 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
575 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
576 /* fail rest of links */
577 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
578 /* on inflight list */
579 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
580 /* read/write uses file position */
581 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
582 /* must not punt to workers */
583 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
584 /* has linked timeout */
585 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
587 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
588 /* completion under lock */
589 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
591 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
592 /* already went through poll handler */
593 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
594 /* buffer already selected */
595 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
596 /* doesn't need file table for this request */
597 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
598 /* io_wq_work is initialized */
599 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
603 struct io_poll_iocb poll;
604 struct io_poll_iocb *double_poll;
608 * NOTE! Each of the iocb union members has the file pointer
609 * as the first entry in their struct definition. So you can
610 * access the file pointer through any of the sub-structs,
611 * or directly as just 'ki_filp' in this struct.
617 struct io_poll_iocb poll;
618 struct io_accept accept;
620 struct io_cancel cancel;
621 struct io_timeout timeout;
622 struct io_connect connect;
623 struct io_sr_msg sr_msg;
625 struct io_close close;
626 struct io_files_update files_update;
627 struct io_fadvise fadvise;
628 struct io_madvise madvise;
629 struct io_epoll epoll;
630 struct io_splice splice;
631 struct io_provide_buf pbuf;
632 struct io_statx statx;
633 /* use only after cleaning per-op data, see io_clean_op() */
634 struct io_completion compl;
637 struct io_async_ctx *io;
639 /* polled IO has completed */
645 struct io_ring_ctx *ctx;
648 struct task_struct *task;
651 struct list_head link_list;
654 * 1. used with ctx->iopoll_list with reads/writes
655 * 2. to track reqs with ->files (see io_op_def::file_table)
657 struct list_head inflight_entry;
659 struct percpu_ref *fixed_file_refs;
660 struct callback_head task_work;
661 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
662 struct hlist_node hash_node;
663 struct async_poll *apoll;
664 struct io_wq_work work;
667 struct io_defer_entry {
668 struct list_head list;
669 struct io_kiocb *req;
673 #define IO_IOPOLL_BATCH 8
675 struct io_comp_state {
677 struct list_head list;
678 struct io_ring_ctx *ctx;
681 struct io_submit_state {
682 struct blk_plug plug;
685 * io_kiocb alloc cache
687 void *reqs[IO_IOPOLL_BATCH];
688 unsigned int free_reqs;
691 * Batch completion logic
693 struct io_comp_state comp;
696 * File reference cache
700 unsigned int has_refs;
701 unsigned int ios_left;
705 /* needs req->io allocated for deferral/async */
706 unsigned async_ctx : 1;
707 /* needs current->mm setup, does mm access */
708 unsigned needs_mm : 1;
709 /* needs req->file assigned */
710 unsigned needs_file : 1;
711 /* don't fail if file grab fails */
712 unsigned needs_file_no_error : 1;
713 /* hash wq insertion if file is a regular file */
714 unsigned hash_reg_file : 1;
715 /* unbound wq insertion if file is a non-regular file */
716 unsigned unbound_nonreg_file : 1;
717 /* opcode is not supported by this kernel */
718 unsigned not_supported : 1;
719 /* needs file table */
720 unsigned file_table : 1;
722 unsigned needs_fs : 1;
723 /* set if opcode supports polled "wait" */
725 unsigned pollout : 1;
726 /* op supports buffer selection */
727 unsigned buffer_select : 1;
728 unsigned needs_fsize : 1;
731 static const struct io_op_def io_op_defs[] = {
732 [IORING_OP_NOP] = {},
733 [IORING_OP_READV] = {
737 .unbound_nonreg_file = 1,
741 [IORING_OP_WRITEV] = {
746 .unbound_nonreg_file = 1,
750 [IORING_OP_FSYNC] = {
753 [IORING_OP_READ_FIXED] = {
755 .unbound_nonreg_file = 1,
758 [IORING_OP_WRITE_FIXED] = {
761 .unbound_nonreg_file = 1,
765 [IORING_OP_POLL_ADD] = {
767 .unbound_nonreg_file = 1,
769 [IORING_OP_POLL_REMOVE] = {},
770 [IORING_OP_SYNC_FILE_RANGE] = {
773 [IORING_OP_SENDMSG] = {
777 .unbound_nonreg_file = 1,
781 [IORING_OP_RECVMSG] = {
785 .unbound_nonreg_file = 1,
790 [IORING_OP_TIMEOUT] = {
794 [IORING_OP_TIMEOUT_REMOVE] = {},
795 [IORING_OP_ACCEPT] = {
798 .unbound_nonreg_file = 1,
802 [IORING_OP_ASYNC_CANCEL] = {},
803 [IORING_OP_LINK_TIMEOUT] = {
807 [IORING_OP_CONNECT] = {
811 .unbound_nonreg_file = 1,
814 [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,
850 [IORING_OP_FADVISE] = {
853 [IORING_OP_MADVISE] = {
859 .unbound_nonreg_file = 1,
865 .unbound_nonreg_file = 1,
869 [IORING_OP_OPENAT2] = {
873 [IORING_OP_EPOLL_CTL] = {
874 .unbound_nonreg_file = 1,
877 [IORING_OP_SPLICE] = {
880 .unbound_nonreg_file = 1,
882 [IORING_OP_PROVIDE_BUFFERS] = {},
883 [IORING_OP_REMOVE_BUFFERS] = {},
887 .unbound_nonreg_file = 1,
891 enum io_mem_account {
896 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
897 struct io_comp_state *cs);
898 static void io_cqring_fill_event(struct io_kiocb *req, long res);
899 static void io_put_req(struct io_kiocb *req);
900 static void io_double_put_req(struct io_kiocb *req);
901 static void __io_double_put_req(struct io_kiocb *req);
902 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
903 static void __io_queue_linked_timeout(struct io_kiocb *req);
904 static void io_queue_linked_timeout(struct io_kiocb *req);
905 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
906 struct io_uring_files_update *ip,
908 static int io_prep_work_files(struct io_kiocb *req);
909 static void __io_clean_op(struct io_kiocb *req);
910 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
911 int fd, struct file **out_file, bool fixed);
912 static void __io_queue_sqe(struct io_kiocb *req,
913 const struct io_uring_sqe *sqe,
914 struct io_comp_state *cs);
915 static void io_file_put_work(struct work_struct *work);
917 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
918 struct iovec **iovec, struct iov_iter *iter,
920 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
921 const struct iovec *fast_iov,
922 struct iov_iter *iter, bool force);
924 static struct kmem_cache *req_cachep;
926 static const struct file_operations io_uring_fops;
928 struct sock *io_uring_get_socket(struct file *file)
930 #if defined(CONFIG_UNIX)
931 if (file->f_op == &io_uring_fops) {
932 struct io_ring_ctx *ctx = file->private_data;
934 return ctx->ring_sock->sk;
939 EXPORT_SYMBOL(io_uring_get_socket);
941 static inline void io_clean_op(struct io_kiocb *req)
943 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
948 static void io_sq_thread_drop_mm(void)
950 struct mm_struct *mm = current->mm;
953 kthread_unuse_mm(mm);
958 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
961 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
962 !ctx->sqo_task->mm ||
963 !mmget_not_zero(ctx->sqo_task->mm)))
965 kthread_use_mm(ctx->sqo_task->mm);
971 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
972 struct io_kiocb *req)
974 if (!io_op_defs[req->opcode].needs_mm)
976 return __io_sq_thread_acquire_mm(ctx);
979 static inline void req_set_fail_links(struct io_kiocb *req)
981 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
982 req->flags |= REQ_F_FAIL_LINK;
986 * Note: must call io_req_init_async() for the first time you
987 * touch any members of io_wq_work.
989 static inline void io_req_init_async(struct io_kiocb *req)
991 if (req->flags & REQ_F_WORK_INITIALIZED)
994 memset(&req->work, 0, sizeof(req->work));
995 req->flags |= REQ_F_WORK_INITIALIZED;
998 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1000 return ctx->flags & IORING_SETUP_SQPOLL;
1003 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1005 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1007 complete(&ctx->ref_comp);
1010 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1012 return !req->timeout.off;
1015 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1017 struct io_ring_ctx *ctx;
1020 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1024 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1025 if (!ctx->fallback_req)
1029 * Use 5 bits less than the max cq entries, that should give us around
1030 * 32 entries per hash list if totally full and uniformly spread.
1032 hash_bits = ilog2(p->cq_entries);
1036 ctx->cancel_hash_bits = hash_bits;
1037 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1039 if (!ctx->cancel_hash)
1041 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1043 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1044 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1047 ctx->flags = p->flags;
1048 init_waitqueue_head(&ctx->sqo_wait);
1049 init_waitqueue_head(&ctx->cq_wait);
1050 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1051 init_completion(&ctx->ref_comp);
1052 init_completion(&ctx->sq_thread_comp);
1053 idr_init(&ctx->io_buffer_idr);
1054 idr_init(&ctx->personality_idr);
1055 mutex_init(&ctx->uring_lock);
1056 init_waitqueue_head(&ctx->wait);
1057 spin_lock_init(&ctx->completion_lock);
1058 INIT_LIST_HEAD(&ctx->iopoll_list);
1059 INIT_LIST_HEAD(&ctx->defer_list);
1060 INIT_LIST_HEAD(&ctx->timeout_list);
1061 init_waitqueue_head(&ctx->inflight_wait);
1062 spin_lock_init(&ctx->inflight_lock);
1063 INIT_LIST_HEAD(&ctx->inflight_list);
1064 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1065 init_llist_head(&ctx->file_put_llist);
1068 if (ctx->fallback_req)
1069 kmem_cache_free(req_cachep, ctx->fallback_req);
1070 kfree(ctx->cancel_hash);
1075 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1077 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1078 struct io_ring_ctx *ctx = req->ctx;
1080 return seq != ctx->cached_cq_tail
1081 + atomic_read(&ctx->cached_cq_overflow);
1087 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1089 struct io_rings *rings = ctx->rings;
1091 /* order cqe stores with ring update */
1092 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1094 if (wq_has_sleeper(&ctx->cq_wait)) {
1095 wake_up_interruptible(&ctx->cq_wait);
1096 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1101 * Returns true if we need to defer file table putting. This can only happen
1102 * from the error path with REQ_F_COMP_LOCKED set.
1104 static bool io_req_clean_work(struct io_kiocb *req)
1106 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1109 req->flags &= ~REQ_F_WORK_INITIALIZED;
1112 mmdrop(req->work.mm);
1113 req->work.mm = NULL;
1115 if (req->work.creds) {
1116 put_cred(req->work.creds);
1117 req->work.creds = NULL;
1120 struct fs_struct *fs = req->work.fs;
1122 if (req->flags & REQ_F_COMP_LOCKED)
1125 spin_lock(&req->work.fs->lock);
1128 spin_unlock(&req->work.fs->lock);
1131 req->work.fs = NULL;
1137 static void io_prep_async_work(struct io_kiocb *req)
1139 const struct io_op_def *def = &io_op_defs[req->opcode];
1141 io_req_init_async(req);
1143 if (req->flags & REQ_F_ISREG) {
1144 if (def->hash_reg_file || (req->ctx->flags & IORING_SETUP_IOPOLL))
1145 io_wq_hash_work(&req->work, file_inode(req->file));
1147 if (def->unbound_nonreg_file)
1148 req->work.flags |= IO_WQ_WORK_UNBOUND;
1150 if (!req->work.mm && def->needs_mm) {
1151 mmgrab(current->mm);
1152 req->work.mm = current->mm;
1154 if (!req->work.creds)
1155 req->work.creds = get_current_cred();
1156 if (!req->work.fs && def->needs_fs) {
1157 spin_lock(¤t->fs->lock);
1158 if (!current->fs->in_exec) {
1159 req->work.fs = current->fs;
1160 req->work.fs->users++;
1162 req->work.flags |= IO_WQ_WORK_CANCEL;
1164 spin_unlock(¤t->fs->lock);
1166 if (def->needs_fsize)
1167 req->work.fsize = rlimit(RLIMIT_FSIZE);
1169 req->work.fsize = RLIM_INFINITY;
1172 static void io_prep_async_link(struct io_kiocb *req)
1174 struct io_kiocb *cur;
1176 io_prep_async_work(req);
1177 if (req->flags & REQ_F_LINK_HEAD)
1178 list_for_each_entry(cur, &req->link_list, link_list)
1179 io_prep_async_work(cur);
1182 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1184 struct io_ring_ctx *ctx = req->ctx;
1185 struct io_kiocb *link = io_prep_linked_timeout(req);
1187 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1188 &req->work, req->flags);
1189 io_wq_enqueue(ctx->io_wq, &req->work);
1193 static void io_queue_async_work(struct io_kiocb *req)
1195 struct io_kiocb *link;
1197 /* init ->work of the whole link before punting */
1198 io_prep_async_link(req);
1199 link = __io_queue_async_work(req);
1202 io_queue_linked_timeout(link);
1205 static void io_kill_timeout(struct io_kiocb *req)
1209 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1211 atomic_set(&req->ctx->cq_timeouts,
1212 atomic_read(&req->ctx->cq_timeouts) + 1);
1213 list_del_init(&req->timeout.list);
1214 req->flags |= REQ_F_COMP_LOCKED;
1215 io_cqring_fill_event(req, 0);
1220 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1222 struct io_ring_ctx *ctx = req->ctx;
1224 if (!tsk || req->task == tsk)
1226 if ((ctx->flags & IORING_SETUP_SQPOLL) && req->task == ctx->sqo_thread)
1232 * Returns true if we found and killed one or more timeouts
1234 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1236 struct io_kiocb *req, *tmp;
1239 spin_lock_irq(&ctx->completion_lock);
1240 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1241 if (io_task_match(req, tsk)) {
1242 io_kill_timeout(req);
1246 spin_unlock_irq(&ctx->completion_lock);
1247 return canceled != 0;
1250 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1253 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1254 struct io_defer_entry, list);
1255 struct io_kiocb *link;
1257 if (req_need_defer(de->req, de->seq))
1259 list_del_init(&de->list);
1260 /* punt-init is done before queueing for defer */
1261 link = __io_queue_async_work(de->req);
1263 __io_queue_linked_timeout(link);
1264 /* drop submission reference */
1265 link->flags |= REQ_F_COMP_LOCKED;
1269 } while (!list_empty(&ctx->defer_list));
1272 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1274 while (!list_empty(&ctx->timeout_list)) {
1275 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1276 struct io_kiocb, timeout.list);
1278 if (io_is_timeout_noseq(req))
1280 if (req->timeout.target_seq != ctx->cached_cq_tail
1281 - atomic_read(&ctx->cq_timeouts))
1284 list_del_init(&req->timeout.list);
1285 io_kill_timeout(req);
1289 static void io_commit_cqring(struct io_ring_ctx *ctx)
1291 io_flush_timeouts(ctx);
1292 __io_commit_cqring(ctx);
1294 if (unlikely(!list_empty(&ctx->defer_list)))
1295 __io_queue_deferred(ctx);
1298 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1300 struct io_rings *rings = ctx->rings;
1303 tail = ctx->cached_cq_tail;
1305 * writes to the cq entry need to come after reading head; the
1306 * control dependency is enough as we're using WRITE_ONCE to
1309 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1312 ctx->cached_cq_tail++;
1313 return &rings->cqes[tail & ctx->cq_mask];
1316 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1320 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1322 if (!ctx->eventfd_async)
1324 return io_wq_current_is_worker();
1327 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1329 if (waitqueue_active(&ctx->wait))
1330 wake_up(&ctx->wait);
1331 if (waitqueue_active(&ctx->sqo_wait))
1332 wake_up(&ctx->sqo_wait);
1333 if (io_should_trigger_evfd(ctx))
1334 eventfd_signal(ctx->cq_ev_fd, 1);
1337 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1339 if (list_empty(&ctx->cq_overflow_list)) {
1340 clear_bit(0, &ctx->sq_check_overflow);
1341 clear_bit(0, &ctx->cq_check_overflow);
1342 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1346 static inline bool io_match_files(struct io_kiocb *req,
1347 struct files_struct *files)
1351 if (req->flags & REQ_F_WORK_INITIALIZED)
1352 return req->work.files == files;
1356 /* Returns true if there are no backlogged entries after the flush */
1357 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1358 struct task_struct *tsk,
1359 struct files_struct *files)
1361 struct io_rings *rings = ctx->rings;
1362 struct io_kiocb *req, *tmp;
1363 struct io_uring_cqe *cqe;
1364 unsigned long flags;
1368 if (list_empty_careful(&ctx->cq_overflow_list))
1370 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1371 rings->cq_ring_entries))
1375 spin_lock_irqsave(&ctx->completion_lock, flags);
1377 /* if force is set, the ring is going away. always drop after that */
1379 ctx->cq_overflow_flushed = 1;
1382 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1383 if (tsk && req->task != tsk)
1385 if (!io_match_files(req, files))
1388 cqe = io_get_cqring(ctx);
1392 list_move(&req->compl.list, &list);
1394 WRITE_ONCE(cqe->user_data, req->user_data);
1395 WRITE_ONCE(cqe->res, req->result);
1396 WRITE_ONCE(cqe->flags, req->compl.cflags);
1398 WRITE_ONCE(ctx->rings->cq_overflow,
1399 atomic_inc_return(&ctx->cached_cq_overflow));
1403 io_commit_cqring(ctx);
1404 io_cqring_mark_overflow(ctx);
1406 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1407 io_cqring_ev_posted(ctx);
1409 while (!list_empty(&list)) {
1410 req = list_first_entry(&list, struct io_kiocb, compl.list);
1411 list_del(&req->compl.list);
1418 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1420 struct io_ring_ctx *ctx = req->ctx;
1421 struct io_uring_cqe *cqe;
1423 trace_io_uring_complete(ctx, req->user_data, res);
1426 * If we can't get a cq entry, userspace overflowed the
1427 * submission (by quite a lot). Increment the overflow count in
1430 cqe = io_get_cqring(ctx);
1432 WRITE_ONCE(cqe->user_data, req->user_data);
1433 WRITE_ONCE(cqe->res, res);
1434 WRITE_ONCE(cqe->flags, cflags);
1435 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1437 * If we're in ring overflow flush mode, or in task cancel mode,
1438 * then we cannot store the request for later flushing, we need
1439 * to drop it on the floor.
1441 WRITE_ONCE(ctx->rings->cq_overflow,
1442 atomic_inc_return(&ctx->cached_cq_overflow));
1444 if (list_empty(&ctx->cq_overflow_list)) {
1445 set_bit(0, &ctx->sq_check_overflow);
1446 set_bit(0, &ctx->cq_check_overflow);
1447 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1451 req->compl.cflags = cflags;
1452 refcount_inc(&req->refs);
1453 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1457 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1459 __io_cqring_fill_event(req, res, 0);
1462 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1464 struct io_ring_ctx *ctx = req->ctx;
1465 unsigned long flags;
1467 spin_lock_irqsave(&ctx->completion_lock, flags);
1468 __io_cqring_fill_event(req, res, cflags);
1469 io_commit_cqring(ctx);
1470 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1472 io_cqring_ev_posted(ctx);
1475 static void io_submit_flush_completions(struct io_comp_state *cs)
1477 struct io_ring_ctx *ctx = cs->ctx;
1479 spin_lock_irq(&ctx->completion_lock);
1480 while (!list_empty(&cs->list)) {
1481 struct io_kiocb *req;
1483 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1484 list_del(&req->compl.list);
1485 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1486 if (!(req->flags & REQ_F_LINK_HEAD)) {
1487 req->flags |= REQ_F_COMP_LOCKED;
1490 spin_unlock_irq(&ctx->completion_lock);
1492 spin_lock_irq(&ctx->completion_lock);
1495 io_commit_cqring(ctx);
1496 spin_unlock_irq(&ctx->completion_lock);
1498 io_cqring_ev_posted(ctx);
1502 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1503 struct io_comp_state *cs)
1506 io_cqring_add_event(req, res, cflags);
1511 req->compl.cflags = cflags;
1512 list_add_tail(&req->compl.list, &cs->list);
1514 io_submit_flush_completions(cs);
1518 static void io_req_complete(struct io_kiocb *req, long res)
1520 __io_req_complete(req, res, 0, NULL);
1523 static inline bool io_is_fallback_req(struct io_kiocb *req)
1525 return req == (struct io_kiocb *)
1526 ((unsigned long) req->ctx->fallback_req & ~1UL);
1529 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1531 struct io_kiocb *req;
1533 req = ctx->fallback_req;
1534 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1540 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1541 struct io_submit_state *state)
1543 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1544 struct io_kiocb *req;
1546 if (!state->free_reqs) {
1550 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1551 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1554 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1555 * retry single alloc to be on the safe side.
1557 if (unlikely(ret <= 0)) {
1558 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1559 if (!state->reqs[0])
1563 state->free_reqs = ret - 1;
1564 req = state->reqs[ret - 1];
1567 req = state->reqs[state->free_reqs];
1572 return io_get_fallback_req(ctx);
1575 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1579 percpu_ref_put(req->fixed_file_refs);
1584 static bool io_dismantle_req(struct io_kiocb *req)
1591 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1593 return io_req_clean_work(req);
1596 static void __io_free_req_finish(struct io_kiocb *req)
1598 struct io_uring_task *tctx = req->task->io_uring;
1599 struct io_ring_ctx *ctx = req->ctx;
1601 atomic_long_inc(&tctx->req_complete);
1603 wake_up(&tctx->wait);
1604 put_task_struct(req->task);
1606 if (likely(!io_is_fallback_req(req)))
1607 kmem_cache_free(req_cachep, req);
1609 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1610 percpu_ref_put(&ctx->refs);
1613 static void io_req_task_file_table_put(struct callback_head *cb)
1615 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1616 struct fs_struct *fs = req->work.fs;
1618 spin_lock(&req->work.fs->lock);
1621 spin_unlock(&req->work.fs->lock);
1624 req->work.fs = NULL;
1625 __io_free_req_finish(req);
1628 static void __io_free_req(struct io_kiocb *req)
1630 if (!io_dismantle_req(req)) {
1631 __io_free_req_finish(req);
1635 init_task_work(&req->task_work, io_req_task_file_table_put);
1636 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1637 if (unlikely(ret)) {
1638 struct task_struct *tsk;
1640 tsk = io_wq_get_task(req->ctx->io_wq);
1641 task_work_add(tsk, &req->task_work, 0);
1646 static bool io_link_cancel_timeout(struct io_kiocb *req)
1648 struct io_ring_ctx *ctx = req->ctx;
1651 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1653 io_cqring_fill_event(req, -ECANCELED);
1654 io_commit_cqring(ctx);
1655 req->flags &= ~REQ_F_LINK_HEAD;
1663 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1665 struct io_kiocb *link;
1668 if (list_empty(&req->link_list))
1670 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1671 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1674 list_del_init(&link->link_list);
1675 link->flags |= REQ_F_COMP_LOCKED;
1676 wake_ev = io_link_cancel_timeout(link);
1677 req->flags &= ~REQ_F_LINK_TIMEOUT;
1681 static void io_kill_linked_timeout(struct io_kiocb *req)
1683 struct io_ring_ctx *ctx = req->ctx;
1686 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1687 unsigned long flags;
1689 spin_lock_irqsave(&ctx->completion_lock, flags);
1690 wake_ev = __io_kill_linked_timeout(req);
1691 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1693 wake_ev = __io_kill_linked_timeout(req);
1697 io_cqring_ev_posted(ctx);
1700 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1702 struct io_kiocb *nxt;
1705 * The list should never be empty when we are called here. But could
1706 * potentially happen if the chain is messed up, check to be on the
1709 if (unlikely(list_empty(&req->link_list)))
1712 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1713 list_del_init(&req->link_list);
1714 if (!list_empty(&nxt->link_list))
1715 nxt->flags |= REQ_F_LINK_HEAD;
1720 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1722 static void __io_fail_links(struct io_kiocb *req)
1724 struct io_ring_ctx *ctx = req->ctx;
1726 while (!list_empty(&req->link_list)) {
1727 struct io_kiocb *link = list_first_entry(&req->link_list,
1728 struct io_kiocb, link_list);
1730 list_del_init(&link->link_list);
1731 trace_io_uring_fail_link(req, link);
1733 io_cqring_fill_event(link, -ECANCELED);
1734 link->flags |= REQ_F_COMP_LOCKED;
1735 __io_double_put_req(link);
1736 req->flags &= ~REQ_F_LINK_TIMEOUT;
1739 io_commit_cqring(ctx);
1740 io_cqring_ev_posted(ctx);
1743 static void io_fail_links(struct io_kiocb *req)
1745 struct io_ring_ctx *ctx = req->ctx;
1747 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1748 unsigned long flags;
1750 spin_lock_irqsave(&ctx->completion_lock, flags);
1751 __io_fail_links(req);
1752 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1754 __io_fail_links(req);
1757 io_cqring_ev_posted(ctx);
1760 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1762 req->flags &= ~REQ_F_LINK_HEAD;
1763 if (req->flags & REQ_F_LINK_TIMEOUT)
1764 io_kill_linked_timeout(req);
1767 * If LINK is set, we have dependent requests in this chain. If we
1768 * didn't fail this request, queue the first one up, moving any other
1769 * dependencies to the next request. In case of failure, fail the rest
1772 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1773 return io_req_link_next(req);
1778 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1780 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1782 return __io_req_find_next(req);
1785 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb,
1788 struct task_struct *tsk = req->task;
1789 struct io_ring_ctx *ctx = req->ctx;
1792 if (tsk->flags & PF_EXITING)
1796 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1797 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1798 * processing task_work. There's no reliable way to tell if TWA_RESUME
1802 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1803 notify = TWA_SIGNAL;
1805 ret = task_work_add(tsk, cb, notify);
1807 wake_up_process(tsk);
1812 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1814 struct io_ring_ctx *ctx = req->ctx;
1816 spin_lock_irq(&ctx->completion_lock);
1817 io_cqring_fill_event(req, error);
1818 io_commit_cqring(ctx);
1819 spin_unlock_irq(&ctx->completion_lock);
1821 io_cqring_ev_posted(ctx);
1822 req_set_fail_links(req);
1823 io_double_put_req(req);
1826 static void io_req_task_cancel(struct callback_head *cb)
1828 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1829 struct io_ring_ctx *ctx = req->ctx;
1831 __io_req_task_cancel(req, -ECANCELED);
1832 percpu_ref_put(&ctx->refs);
1835 static void __io_req_task_submit(struct io_kiocb *req)
1837 struct io_ring_ctx *ctx = req->ctx;
1839 if (!__io_sq_thread_acquire_mm(ctx)) {
1840 mutex_lock(&ctx->uring_lock);
1841 __io_queue_sqe(req, NULL, NULL);
1842 mutex_unlock(&ctx->uring_lock);
1844 __io_req_task_cancel(req, -EFAULT);
1848 static void io_req_task_submit(struct callback_head *cb)
1850 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1851 struct io_ring_ctx *ctx = req->ctx;
1853 __io_req_task_submit(req);
1854 percpu_ref_put(&ctx->refs);
1857 static void io_req_task_queue(struct io_kiocb *req)
1861 init_task_work(&req->task_work, io_req_task_submit);
1862 percpu_ref_get(&req->ctx->refs);
1864 ret = io_req_task_work_add(req, &req->task_work, true);
1865 if (unlikely(ret)) {
1866 struct task_struct *tsk;
1868 init_task_work(&req->task_work, io_req_task_cancel);
1869 tsk = io_wq_get_task(req->ctx->io_wq);
1870 task_work_add(tsk, &req->task_work, 0);
1871 wake_up_process(tsk);
1875 static void io_queue_next(struct io_kiocb *req)
1877 struct io_kiocb *nxt = io_req_find_next(req);
1880 io_req_task_queue(nxt);
1883 static void io_free_req(struct io_kiocb *req)
1890 void *reqs[IO_IOPOLL_BATCH];
1893 struct task_struct *task;
1897 static inline void io_init_req_batch(struct req_batch *rb)
1904 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1905 struct req_batch *rb)
1907 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1908 percpu_ref_put_many(&ctx->refs, rb->to_free);
1912 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1913 struct req_batch *rb)
1916 __io_req_free_batch_flush(ctx, rb);
1918 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
1919 put_task_struct_many(rb->task, rb->task_refs);
1924 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1926 if (unlikely(io_is_fallback_req(req))) {
1930 if (req->flags & REQ_F_LINK_HEAD)
1933 if (req->task != rb->task) {
1935 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
1936 put_task_struct_many(rb->task, rb->task_refs);
1938 rb->task = req->task;
1943 WARN_ON_ONCE(io_dismantle_req(req));
1944 rb->reqs[rb->to_free++] = req;
1945 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1946 __io_req_free_batch_flush(req->ctx, rb);
1950 * Drop reference to request, return next in chain (if there is one) if this
1951 * was the last reference to this request.
1953 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1955 struct io_kiocb *nxt = NULL;
1957 if (refcount_dec_and_test(&req->refs)) {
1958 nxt = io_req_find_next(req);
1964 static void io_put_req(struct io_kiocb *req)
1966 if (refcount_dec_and_test(&req->refs))
1970 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1972 struct io_kiocb *nxt;
1975 * A ref is owned by io-wq in which context we're. So, if that's the
1976 * last one, it's safe to steal next work. False negatives are Ok,
1977 * it just will be re-punted async in io_put_work()
1979 if (refcount_read(&req->refs) != 1)
1982 nxt = io_req_find_next(req);
1983 return nxt ? &nxt->work : NULL;
1987 * Must only be used if we don't need to care about links, usually from
1988 * within the completion handling itself.
1990 static void __io_double_put_req(struct io_kiocb *req)
1992 /* drop both submit and complete references */
1993 if (refcount_sub_and_test(2, &req->refs))
1997 static void io_double_put_req(struct io_kiocb *req)
1999 /* drop both submit and complete references */
2000 if (refcount_sub_and_test(2, &req->refs))
2004 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2006 struct io_rings *rings = ctx->rings;
2008 if (test_bit(0, &ctx->cq_check_overflow)) {
2010 * noflush == true is from the waitqueue handler, just ensure
2011 * we wake up the task, and the next invocation will flush the
2012 * entries. We cannot safely to it from here.
2014 if (noflush && !list_empty(&ctx->cq_overflow_list))
2017 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2020 /* See comment at the top of this file */
2022 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2025 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2027 struct io_rings *rings = ctx->rings;
2029 /* make sure SQ entry isn't read before tail */
2030 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2033 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2035 unsigned int cflags;
2037 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2038 cflags |= IORING_CQE_F_BUFFER;
2039 req->flags &= ~REQ_F_BUFFER_SELECTED;
2044 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2046 struct io_buffer *kbuf;
2048 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2049 return io_put_kbuf(req, kbuf);
2052 static inline bool io_run_task_work(void)
2055 * Not safe to run on exiting task, and the task_work handling will
2056 * not add work to such a task.
2058 if (unlikely(current->flags & PF_EXITING))
2060 if (current->task_works) {
2061 __set_current_state(TASK_RUNNING);
2069 static void io_iopoll_queue(struct list_head *again)
2071 struct io_kiocb *req;
2074 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2075 list_del(&req->inflight_entry);
2076 __io_complete_rw(req, -EAGAIN, 0, NULL);
2077 } while (!list_empty(again));
2081 * Find and free completed poll iocbs
2083 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2084 struct list_head *done)
2086 struct req_batch rb;
2087 struct io_kiocb *req;
2090 /* order with ->result store in io_complete_rw_iopoll() */
2093 io_init_req_batch(&rb);
2094 while (!list_empty(done)) {
2097 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2098 if (READ_ONCE(req->result) == -EAGAIN) {
2100 req->iopoll_completed = 0;
2101 list_move_tail(&req->inflight_entry, &again);
2104 list_del(&req->inflight_entry);
2106 if (req->flags & REQ_F_BUFFER_SELECTED)
2107 cflags = io_put_rw_kbuf(req);
2109 __io_cqring_fill_event(req, req->result, cflags);
2112 if (refcount_dec_and_test(&req->refs))
2113 io_req_free_batch(&rb, req);
2116 io_commit_cqring(ctx);
2117 if (ctx->flags & IORING_SETUP_SQPOLL)
2118 io_cqring_ev_posted(ctx);
2119 io_req_free_batch_finish(ctx, &rb);
2121 if (!list_empty(&again))
2122 io_iopoll_queue(&again);
2125 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2128 struct io_kiocb *req, *tmp;
2134 * Only spin for completions if we don't have multiple devices hanging
2135 * off our complete list, and we're under the requested amount.
2137 spin = !ctx->poll_multi_file && *nr_events < min;
2140 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2141 struct kiocb *kiocb = &req->rw.kiocb;
2144 * Move completed and retryable entries to our local lists.
2145 * If we find a request that requires polling, break out
2146 * and complete those lists first, if we have entries there.
2148 if (READ_ONCE(req->iopoll_completed)) {
2149 list_move_tail(&req->inflight_entry, &done);
2152 if (!list_empty(&done))
2155 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2159 /* iopoll may have completed current req */
2160 if (READ_ONCE(req->iopoll_completed))
2161 list_move_tail(&req->inflight_entry, &done);
2168 if (!list_empty(&done))
2169 io_iopoll_complete(ctx, nr_events, &done);
2175 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2176 * non-spinning poll check - we'll still enter the driver poll loop, but only
2177 * as a non-spinning completion check.
2179 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2182 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2185 ret = io_do_iopoll(ctx, nr_events, min);
2188 if (*nr_events >= min)
2196 * We can't just wait for polled events to come to us, we have to actively
2197 * find and complete them.
2199 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2201 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2204 mutex_lock(&ctx->uring_lock);
2205 while (!list_empty(&ctx->iopoll_list)) {
2206 unsigned int nr_events = 0;
2208 io_do_iopoll(ctx, &nr_events, 0);
2210 /* let it sleep and repeat later if can't complete a request */
2214 * Ensure we allow local-to-the-cpu processing to take place,
2215 * in this case we need to ensure that we reap all events.
2216 * Also let task_work, etc. to progress by releasing the mutex
2218 if (need_resched()) {
2219 mutex_unlock(&ctx->uring_lock);
2221 mutex_lock(&ctx->uring_lock);
2224 mutex_unlock(&ctx->uring_lock);
2227 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2229 unsigned int nr_events = 0;
2230 int iters = 0, ret = 0;
2233 * We disallow the app entering submit/complete with polling, but we
2234 * still need to lock the ring to prevent racing with polled issue
2235 * that got punted to a workqueue.
2237 mutex_lock(&ctx->uring_lock);
2240 * Don't enter poll loop if we already have events pending.
2241 * If we do, we can potentially be spinning for commands that
2242 * already triggered a CQE (eg in error).
2244 if (io_cqring_events(ctx, false))
2248 * If a submit got punted to a workqueue, we can have the
2249 * application entering polling for a command before it gets
2250 * issued. That app will hold the uring_lock for the duration
2251 * of the poll right here, so we need to take a breather every
2252 * now and then to ensure that the issue has a chance to add
2253 * the poll to the issued list. Otherwise we can spin here
2254 * forever, while the workqueue is stuck trying to acquire the
2257 if (!(++iters & 7)) {
2258 mutex_unlock(&ctx->uring_lock);
2260 mutex_lock(&ctx->uring_lock);
2263 ret = io_iopoll_getevents(ctx, &nr_events, min);
2267 } while (min && !nr_events && !need_resched());
2269 mutex_unlock(&ctx->uring_lock);
2273 static void kiocb_end_write(struct io_kiocb *req)
2276 * Tell lockdep we inherited freeze protection from submission
2279 if (req->flags & REQ_F_ISREG) {
2280 struct inode *inode = file_inode(req->file);
2282 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2284 file_end_write(req->file);
2287 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2288 struct io_comp_state *cs)
2290 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2293 if (kiocb->ki_flags & IOCB_WRITE)
2294 kiocb_end_write(req);
2296 if (res != req->result)
2297 req_set_fail_links(req);
2298 if (req->flags & REQ_F_BUFFER_SELECTED)
2299 cflags = io_put_rw_kbuf(req);
2300 __io_req_complete(req, res, cflags, cs);
2304 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2306 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2307 ssize_t ret = -ECANCELED;
2308 struct iov_iter iter;
2316 switch (req->opcode) {
2317 case IORING_OP_READV:
2318 case IORING_OP_READ_FIXED:
2319 case IORING_OP_READ:
2322 case IORING_OP_WRITEV:
2323 case IORING_OP_WRITE_FIXED:
2324 case IORING_OP_WRITE:
2328 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2334 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2337 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2345 req_set_fail_links(req);
2346 io_req_complete(req, ret);
2351 static bool io_rw_reissue(struct io_kiocb *req, long res)
2354 umode_t mode = file_inode(req->file)->i_mode;
2357 if (!S_ISBLK(mode) && !S_ISREG(mode))
2359 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2362 ret = io_sq_thread_acquire_mm(req->ctx, req);
2364 if (io_resubmit_prep(req, ret)) {
2365 refcount_inc(&req->refs);
2366 io_queue_async_work(req);
2374 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2375 struct io_comp_state *cs)
2377 if (!io_rw_reissue(req, res))
2378 io_complete_rw_common(&req->rw.kiocb, res, cs);
2381 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2383 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2385 __io_complete_rw(req, res, res2, NULL);
2388 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2390 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2392 if (kiocb->ki_flags & IOCB_WRITE)
2393 kiocb_end_write(req);
2395 if (res != -EAGAIN && res != req->result)
2396 req_set_fail_links(req);
2398 WRITE_ONCE(req->result, res);
2399 /* order with io_poll_complete() checking ->result */
2401 WRITE_ONCE(req->iopoll_completed, 1);
2405 * After the iocb has been issued, it's safe to be found on the poll list.
2406 * Adding the kiocb to the list AFTER submission ensures that we don't
2407 * find it from a io_iopoll_getevents() thread before the issuer is done
2408 * accessing the kiocb cookie.
2410 static void io_iopoll_req_issued(struct io_kiocb *req)
2412 struct io_ring_ctx *ctx = req->ctx;
2415 * Track whether we have multiple files in our lists. This will impact
2416 * how we do polling eventually, not spinning if we're on potentially
2417 * different devices.
2419 if (list_empty(&ctx->iopoll_list)) {
2420 ctx->poll_multi_file = false;
2421 } else if (!ctx->poll_multi_file) {
2422 struct io_kiocb *list_req;
2424 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2426 if (list_req->file != req->file)
2427 ctx->poll_multi_file = true;
2431 * For fast devices, IO may have already completed. If it has, add
2432 * it to the front so we find it first.
2434 if (READ_ONCE(req->iopoll_completed))
2435 list_add(&req->inflight_entry, &ctx->iopoll_list);
2437 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2439 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2440 wq_has_sleeper(&ctx->sqo_wait))
2441 wake_up(&ctx->sqo_wait);
2444 static void __io_state_file_put(struct io_submit_state *state)
2446 if (state->has_refs)
2447 fput_many(state->file, state->has_refs);
2451 static inline void io_state_file_put(struct io_submit_state *state)
2454 __io_state_file_put(state);
2458 * Get as many references to a file as we have IOs left in this submission,
2459 * assuming most submissions are for one file, or at least that each file
2460 * has more than one submission.
2462 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2468 if (state->fd == fd) {
2473 __io_state_file_put(state);
2475 state->file = fget_many(fd, state->ios_left);
2481 state->has_refs = state->ios_left;
2485 static bool io_bdev_nowait(struct block_device *bdev)
2488 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2495 * If we tracked the file through the SCM inflight mechanism, we could support
2496 * any file. For now, just ensure that anything potentially problematic is done
2499 static bool io_file_supports_async(struct file *file, int rw)
2501 umode_t mode = file_inode(file)->i_mode;
2503 if (S_ISBLK(mode)) {
2504 if (io_bdev_nowait(file->f_inode->i_bdev))
2508 if (S_ISCHR(mode) || S_ISSOCK(mode))
2510 if (S_ISREG(mode)) {
2511 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2512 file->f_op != &io_uring_fops)
2517 /* any ->read/write should understand O_NONBLOCK */
2518 if (file->f_flags & O_NONBLOCK)
2521 if (!(file->f_mode & FMODE_NOWAIT))
2525 return file->f_op->read_iter != NULL;
2527 return file->f_op->write_iter != NULL;
2530 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2531 bool force_nonblock)
2533 struct io_ring_ctx *ctx = req->ctx;
2534 struct kiocb *kiocb = &req->rw.kiocb;
2538 if (S_ISREG(file_inode(req->file)->i_mode))
2539 req->flags |= REQ_F_ISREG;
2541 kiocb->ki_pos = READ_ONCE(sqe->off);
2542 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2543 req->flags |= REQ_F_CUR_POS;
2544 kiocb->ki_pos = req->file->f_pos;
2546 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2547 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2548 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2552 ioprio = READ_ONCE(sqe->ioprio);
2554 ret = ioprio_check_cap(ioprio);
2558 kiocb->ki_ioprio = ioprio;
2560 kiocb->ki_ioprio = get_current_ioprio();
2562 /* don't allow async punt if RWF_NOWAIT was requested */
2563 if (kiocb->ki_flags & IOCB_NOWAIT)
2564 req->flags |= REQ_F_NOWAIT;
2567 kiocb->ki_flags |= IOCB_NOWAIT;
2569 if (ctx->flags & IORING_SETUP_IOPOLL) {
2570 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2571 !kiocb->ki_filp->f_op->iopoll)
2574 kiocb->ki_flags |= IOCB_HIPRI;
2575 kiocb->ki_complete = io_complete_rw_iopoll;
2576 req->iopoll_completed = 0;
2578 if (kiocb->ki_flags & IOCB_HIPRI)
2580 kiocb->ki_complete = io_complete_rw;
2583 req->rw.addr = READ_ONCE(sqe->addr);
2584 req->rw.len = READ_ONCE(sqe->len);
2585 req->buf_index = READ_ONCE(sqe->buf_index);
2589 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2595 case -ERESTARTNOINTR:
2596 case -ERESTARTNOHAND:
2597 case -ERESTART_RESTARTBLOCK:
2599 * We can't just restart the syscall, since previously
2600 * submitted sqes may already be in progress. Just fail this
2606 kiocb->ki_complete(kiocb, ret, 0);
2610 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2611 struct io_comp_state *cs)
2613 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2615 /* add previously done IO, if any */
2616 if (req->io && req->io->rw.bytes_done > 0) {
2618 ret = req->io->rw.bytes_done;
2620 ret += req->io->rw.bytes_done;
2623 if (req->flags & REQ_F_CUR_POS)
2624 req->file->f_pos = kiocb->ki_pos;
2625 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2626 __io_complete_rw(req, ret, 0, cs);
2628 io_rw_done(kiocb, ret);
2631 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2632 struct iov_iter *iter)
2634 struct io_ring_ctx *ctx = req->ctx;
2635 size_t len = req->rw.len;
2636 struct io_mapped_ubuf *imu;
2637 u16 index, buf_index;
2641 /* attempt to use fixed buffers without having provided iovecs */
2642 if (unlikely(!ctx->user_bufs))
2645 buf_index = req->buf_index;
2646 if (unlikely(buf_index >= ctx->nr_user_bufs))
2649 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2650 imu = &ctx->user_bufs[index];
2651 buf_addr = req->rw.addr;
2654 if (buf_addr + len < buf_addr)
2656 /* not inside the mapped region */
2657 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2661 * May not be a start of buffer, set size appropriately
2662 * and advance us to the beginning.
2664 offset = buf_addr - imu->ubuf;
2665 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2669 * Don't use iov_iter_advance() here, as it's really slow for
2670 * using the latter parts of a big fixed buffer - it iterates
2671 * over each segment manually. We can cheat a bit here, because
2674 * 1) it's a BVEC iter, we set it up
2675 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2676 * first and last bvec
2678 * So just find our index, and adjust the iterator afterwards.
2679 * If the offset is within the first bvec (or the whole first
2680 * bvec, just use iov_iter_advance(). This makes it easier
2681 * since we can just skip the first segment, which may not
2682 * be PAGE_SIZE aligned.
2684 const struct bio_vec *bvec = imu->bvec;
2686 if (offset <= bvec->bv_len) {
2687 iov_iter_advance(iter, offset);
2689 unsigned long seg_skip;
2691 /* skip first vec */
2692 offset -= bvec->bv_len;
2693 seg_skip = 1 + (offset >> PAGE_SHIFT);
2695 iter->bvec = bvec + seg_skip;
2696 iter->nr_segs -= seg_skip;
2697 iter->count -= bvec->bv_len + offset;
2698 iter->iov_offset = offset & ~PAGE_MASK;
2705 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2708 mutex_unlock(&ctx->uring_lock);
2711 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2714 * "Normal" inline submissions always hold the uring_lock, since we
2715 * grab it from the system call. Same is true for the SQPOLL offload.
2716 * The only exception is when we've detached the request and issue it
2717 * from an async worker thread, grab the lock for that case.
2720 mutex_lock(&ctx->uring_lock);
2723 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2724 int bgid, struct io_buffer *kbuf,
2727 struct io_buffer *head;
2729 if (req->flags & REQ_F_BUFFER_SELECTED)
2732 io_ring_submit_lock(req->ctx, needs_lock);
2734 lockdep_assert_held(&req->ctx->uring_lock);
2736 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2738 if (!list_empty(&head->list)) {
2739 kbuf = list_last_entry(&head->list, struct io_buffer,
2741 list_del(&kbuf->list);
2744 idr_remove(&req->ctx->io_buffer_idr, bgid);
2746 if (*len > kbuf->len)
2749 kbuf = ERR_PTR(-ENOBUFS);
2752 io_ring_submit_unlock(req->ctx, needs_lock);
2757 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2760 struct io_buffer *kbuf;
2763 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2764 bgid = req->buf_index;
2765 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2768 req->rw.addr = (u64) (unsigned long) kbuf;
2769 req->flags |= REQ_F_BUFFER_SELECTED;
2770 return u64_to_user_ptr(kbuf->addr);
2773 #ifdef CONFIG_COMPAT
2774 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2777 struct compat_iovec __user *uiov;
2778 compat_ssize_t clen;
2782 uiov = u64_to_user_ptr(req->rw.addr);
2783 if (!access_ok(uiov, sizeof(*uiov)))
2785 if (__get_user(clen, &uiov->iov_len))
2791 buf = io_rw_buffer_select(req, &len, needs_lock);
2793 return PTR_ERR(buf);
2794 iov[0].iov_base = buf;
2795 iov[0].iov_len = (compat_size_t) len;
2800 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2803 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2807 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2810 len = iov[0].iov_len;
2813 buf = io_rw_buffer_select(req, &len, needs_lock);
2815 return PTR_ERR(buf);
2816 iov[0].iov_base = buf;
2817 iov[0].iov_len = len;
2821 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2824 if (req->flags & REQ_F_BUFFER_SELECTED) {
2825 struct io_buffer *kbuf;
2827 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2828 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2829 iov[0].iov_len = kbuf->len;
2834 else if (req->rw.len > 1)
2837 #ifdef CONFIG_COMPAT
2838 if (req->ctx->compat)
2839 return io_compat_import(req, iov, needs_lock);
2842 return __io_iov_buffer_select(req, iov, needs_lock);
2845 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2846 struct iovec **iovec, struct iov_iter *iter,
2849 void __user *buf = u64_to_user_ptr(req->rw.addr);
2850 size_t sqe_len = req->rw.len;
2854 opcode = req->opcode;
2855 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2857 return io_import_fixed(req, rw, iter);
2860 /* buffer index only valid with fixed read/write, or buffer select */
2861 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2864 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2865 if (req->flags & REQ_F_BUFFER_SELECT) {
2866 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2868 return PTR_ERR(buf);
2869 req->rw.len = sqe_len;
2872 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2874 return ret < 0 ? ret : sqe_len;
2877 if (req->flags & REQ_F_BUFFER_SELECT) {
2878 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2880 ret = (*iovec)->iov_len;
2881 iov_iter_init(iter, rw, *iovec, 1, ret);
2887 #ifdef CONFIG_COMPAT
2888 if (req->ctx->compat)
2889 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2893 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2896 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2897 struct iovec **iovec, struct iov_iter *iter,
2901 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
2903 return iov_iter_count(&req->io->rw.iter);
2906 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2908 return kiocb->ki_filp->f_mode & FMODE_STREAM ? NULL : &kiocb->ki_pos;
2912 * For files that don't have ->read_iter() and ->write_iter(), handle them
2913 * by looping over ->read() or ->write() manually.
2915 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2916 struct iov_iter *iter)
2921 * Don't support polled IO through this interface, and we can't
2922 * support non-blocking either. For the latter, this just causes
2923 * the kiocb to be handled from an async context.
2925 if (kiocb->ki_flags & IOCB_HIPRI)
2927 if (kiocb->ki_flags & IOCB_NOWAIT)
2930 while (iov_iter_count(iter)) {
2934 if (!iov_iter_is_bvec(iter)) {
2935 iovec = iov_iter_iovec(iter);
2937 /* fixed buffers import bvec */
2938 iovec.iov_base = kmap(iter->bvec->bv_page)
2940 iovec.iov_len = min(iter->count,
2941 iter->bvec->bv_len - iter->iov_offset);
2945 nr = file->f_op->read(file, iovec.iov_base,
2946 iovec.iov_len, io_kiocb_ppos(kiocb));
2948 nr = file->f_op->write(file, iovec.iov_base,
2949 iovec.iov_len, io_kiocb_ppos(kiocb));
2952 if (iov_iter_is_bvec(iter))
2953 kunmap(iter->bvec->bv_page);
2961 if (nr != iovec.iov_len)
2963 iov_iter_advance(iter, nr);
2969 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
2970 const struct iovec *fast_iov, struct iov_iter *iter)
2972 struct io_async_rw *rw = &req->io->rw;
2974 memcpy(&rw->iter, iter, sizeof(*iter));
2975 rw->free_iovec = NULL;
2977 /* can only be fixed buffers, no need to do anything */
2978 if (iter->type == ITER_BVEC)
2981 unsigned iov_off = 0;
2983 rw->iter.iov = rw->fast_iov;
2984 if (iter->iov != fast_iov) {
2985 iov_off = iter->iov - fast_iov;
2986 rw->iter.iov += iov_off;
2988 if (rw->fast_iov != fast_iov)
2989 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
2990 sizeof(struct iovec) * iter->nr_segs);
2992 rw->free_iovec = iovec;
2993 req->flags |= REQ_F_NEED_CLEANUP;
2997 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2999 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
3000 return req->io == NULL;
3003 static int io_alloc_async_ctx(struct io_kiocb *req)
3005 if (!io_op_defs[req->opcode].async_ctx)
3008 return __io_alloc_async_ctx(req);
3011 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3012 const struct iovec *fast_iov,
3013 struct iov_iter *iter, bool force)
3015 if (!force && !io_op_defs[req->opcode].async_ctx)
3018 if (__io_alloc_async_ctx(req))
3021 io_req_map_rw(req, iovec, fast_iov, iter);
3026 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
3027 bool force_nonblock)
3029 struct io_async_rw *iorw = &req->io->rw;
3033 iorw->iter.iov = iov = iorw->fast_iov;
3034 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, !force_nonblock);
3035 if (unlikely(ret < 0))
3038 iorw->iter.iov = iov;
3039 io_req_map_rw(req, iorw->iter.iov, iorw->fast_iov, &iorw->iter);
3043 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3044 bool force_nonblock)
3048 ret = io_prep_rw(req, sqe, force_nonblock);
3052 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3055 /* either don't need iovec imported or already have it */
3056 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3058 return io_rw_prep_async(req, READ, force_nonblock);
3062 * This is our waitqueue callback handler, registered through lock_page_async()
3063 * when we initially tried to do the IO with the iocb armed our waitqueue.
3064 * This gets called when the page is unlocked, and we generally expect that to
3065 * happen when the page IO is completed and the page is now uptodate. This will
3066 * queue a task_work based retry of the operation, attempting to copy the data
3067 * again. If the latter fails because the page was NOT uptodate, then we will
3068 * do a thread based blocking retry of the operation. That's the unexpected
3071 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3072 int sync, void *arg)
3074 struct wait_page_queue *wpq;
3075 struct io_kiocb *req = wait->private;
3076 struct wait_page_key *key = arg;
3079 wpq = container_of(wait, struct wait_page_queue, wait);
3081 if (!wake_page_match(wpq, key))
3084 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3085 list_del_init(&wait->entry);
3087 init_task_work(&req->task_work, io_req_task_submit);
3088 percpu_ref_get(&req->ctx->refs);
3090 /* submit ref gets dropped, acquire a new one */
3091 refcount_inc(&req->refs);
3092 ret = io_req_task_work_add(req, &req->task_work, true);
3093 if (unlikely(ret)) {
3094 struct task_struct *tsk;
3096 /* queue just for cancelation */
3097 init_task_work(&req->task_work, io_req_task_cancel);
3098 tsk = io_wq_get_task(req->ctx->io_wq);
3099 task_work_add(tsk, &req->task_work, 0);
3100 wake_up_process(tsk);
3106 * This controls whether a given IO request should be armed for async page
3107 * based retry. If we return false here, the request is handed to the async
3108 * worker threads for retry. If we're doing buffered reads on a regular file,
3109 * we prepare a private wait_page_queue entry and retry the operation. This
3110 * will either succeed because the page is now uptodate and unlocked, or it
3111 * will register a callback when the page is unlocked at IO completion. Through
3112 * that callback, io_uring uses task_work to setup a retry of the operation.
3113 * That retry will attempt the buffered read again. The retry will generally
3114 * succeed, or in rare cases where it fails, we then fall back to using the
3115 * async worker threads for a blocking retry.
3117 static bool io_rw_should_retry(struct io_kiocb *req)
3119 struct wait_page_queue *wait = &req->io->rw.wpq;
3120 struct kiocb *kiocb = &req->rw.kiocb;
3122 /* never retry for NOWAIT, we just complete with -EAGAIN */
3123 if (req->flags & REQ_F_NOWAIT)
3126 /* Only for buffered IO */
3127 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3131 * just use poll if we can, and don't attempt if the fs doesn't
3132 * support callback based unlocks
3134 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3137 wait->wait.func = io_async_buf_func;
3138 wait->wait.private = req;
3139 wait->wait.flags = 0;
3140 INIT_LIST_HEAD(&wait->wait.entry);
3141 kiocb->ki_flags |= IOCB_WAITQ;
3142 kiocb->ki_flags &= ~IOCB_NOWAIT;
3143 kiocb->ki_waitq = wait;
3147 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3149 if (req->file->f_op->read_iter)
3150 return call_read_iter(req->file, &req->rw.kiocb, iter);
3151 else if (req->file->f_op->read)
3152 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3157 static int io_read(struct io_kiocb *req, bool force_nonblock,
3158 struct io_comp_state *cs)
3160 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3161 struct kiocb *kiocb = &req->rw.kiocb;
3162 struct iov_iter __iter, *iter = &__iter;
3163 ssize_t io_size, ret, ret2;
3168 iter = &req->io->rw.iter;
3170 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3173 iov_count = iov_iter_count(iter);
3175 req->result = io_size;
3178 /* Ensure we clear previously set non-block flag */
3179 if (!force_nonblock)
3180 kiocb->ki_flags &= ~IOCB_NOWAIT;
3182 /* If the file doesn't support async, just async punt */
3183 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3187 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3191 ret = io_iter_do_read(req, iter);
3195 } else if (ret == -EIOCBQUEUED) {
3198 } else if (ret == -EAGAIN) {
3199 /* IOPOLL retry should happen for io-wq threads */
3200 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3202 /* no retry on NONBLOCK marked file */
3203 if (req->file->f_flags & O_NONBLOCK)
3205 /* some cases will consume bytes even on error returns */
3206 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3209 } else if (ret < 0) {
3210 /* make sure -ERESTARTSYS -> -EINTR is done */
3214 /* read it all, or we did blocking attempt. no retry. */
3215 if (!iov_iter_count(iter) || !force_nonblock ||
3216 (req->file->f_flags & O_NONBLOCK))
3221 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3228 /* it's copied and will be cleaned with ->io */
3230 /* now use our persistent iterator, if we aren't already */
3231 iter = &req->io->rw.iter;
3233 req->io->rw.bytes_done += ret;
3234 /* if we can retry, do so with the callbacks armed */
3235 if (!io_rw_should_retry(req)) {
3236 kiocb->ki_flags &= ~IOCB_WAITQ;
3241 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3242 * get -EIOCBQUEUED, then we'll get a notification when the desired
3243 * page gets unlocked. We can also get a partial read here, and if we
3244 * do, then just retry at the new offset.
3246 ret = io_iter_do_read(req, iter);
3247 if (ret == -EIOCBQUEUED) {
3250 } else if (ret > 0 && ret < io_size) {
3251 /* we got some bytes, but not all. retry. */
3255 kiocb_done(kiocb, ret, cs);
3258 /* it's reportedly faster than delegating the null check to kfree() */
3264 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3265 bool force_nonblock)
3269 ret = io_prep_rw(req, sqe, force_nonblock);
3273 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3276 /* either don't need iovec imported or already have it */
3277 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3279 return io_rw_prep_async(req, WRITE, force_nonblock);
3282 static int io_write(struct io_kiocb *req, bool force_nonblock,
3283 struct io_comp_state *cs)
3285 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3286 struct kiocb *kiocb = &req->rw.kiocb;
3287 struct iov_iter __iter, *iter = &__iter;
3289 ssize_t ret, ret2, io_size;
3292 iter = &req->io->rw.iter;
3294 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3297 iov_count = iov_iter_count(iter);
3299 req->result = io_size;
3301 /* Ensure we clear previously set non-block flag */
3302 if (!force_nonblock)
3303 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3305 /* If the file doesn't support async, just async punt */
3306 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3309 /* file path doesn't support NOWAIT for non-direct_IO */
3310 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3311 (req->flags & REQ_F_ISREG))
3314 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3319 * Open-code file_start_write here to grab freeze protection,
3320 * which will be released by another thread in
3321 * io_complete_rw(). Fool lockdep by telling it the lock got
3322 * released so that it doesn't complain about the held lock when
3323 * we return to userspace.
3325 if (req->flags & REQ_F_ISREG) {
3326 __sb_start_write(file_inode(req->file)->i_sb,
3327 SB_FREEZE_WRITE, true);
3328 __sb_writers_release(file_inode(req->file)->i_sb,
3331 kiocb->ki_flags |= IOCB_WRITE;
3333 if (req->file->f_op->write_iter)
3334 ret2 = call_write_iter(req->file, kiocb, iter);
3335 else if (req->file->f_op->write)
3336 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3341 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3342 * retry them without IOCB_NOWAIT.
3344 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3346 /* no retry on NONBLOCK marked file */
3347 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3349 if (!force_nonblock || ret2 != -EAGAIN) {
3350 /* IOPOLL retry should happen for io-wq threads */
3351 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3354 kiocb_done(kiocb, ret2, cs);
3357 /* some cases will consume bytes even on error returns */
3358 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3359 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3364 /* it's reportedly faster than delegating the null check to kfree() */
3370 static int __io_splice_prep(struct io_kiocb *req,
3371 const struct io_uring_sqe *sqe)
3373 struct io_splice* sp = &req->splice;
3374 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3377 if (req->flags & REQ_F_NEED_CLEANUP)
3379 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3383 sp->len = READ_ONCE(sqe->len);
3384 sp->flags = READ_ONCE(sqe->splice_flags);
3386 if (unlikely(sp->flags & ~valid_flags))
3389 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3390 (sp->flags & SPLICE_F_FD_IN_FIXED));
3393 req->flags |= REQ_F_NEED_CLEANUP;
3395 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3397 * Splice operation will be punted aync, and here need to
3398 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3400 io_req_init_async(req);
3401 req->work.flags |= IO_WQ_WORK_UNBOUND;
3407 static int io_tee_prep(struct io_kiocb *req,
3408 const struct io_uring_sqe *sqe)
3410 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3412 return __io_splice_prep(req, sqe);
3415 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3417 struct io_splice *sp = &req->splice;
3418 struct file *in = sp->file_in;
3419 struct file *out = sp->file_out;
3420 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3426 ret = do_tee(in, out, sp->len, flags);
3428 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3429 req->flags &= ~REQ_F_NEED_CLEANUP;
3432 req_set_fail_links(req);
3433 io_req_complete(req, ret);
3437 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3439 struct io_splice* sp = &req->splice;
3441 sp->off_in = READ_ONCE(sqe->splice_off_in);
3442 sp->off_out = READ_ONCE(sqe->off);
3443 return __io_splice_prep(req, sqe);
3446 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3448 struct io_splice *sp = &req->splice;
3449 struct file *in = sp->file_in;
3450 struct file *out = sp->file_out;
3451 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3452 loff_t *poff_in, *poff_out;
3458 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3459 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3462 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3464 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3465 req->flags &= ~REQ_F_NEED_CLEANUP;
3468 req_set_fail_links(req);
3469 io_req_complete(req, ret);
3474 * IORING_OP_NOP just posts a completion event, nothing else.
3476 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3478 struct io_ring_ctx *ctx = req->ctx;
3480 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3483 __io_req_complete(req, 0, 0, cs);
3487 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3489 struct io_ring_ctx *ctx = req->ctx;
3494 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3496 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3499 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3500 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3503 req->sync.off = READ_ONCE(sqe->off);
3504 req->sync.len = READ_ONCE(sqe->len);
3508 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3510 loff_t end = req->sync.off + req->sync.len;
3513 /* fsync always requires a blocking context */
3517 ret = vfs_fsync_range(req->file, req->sync.off,
3518 end > 0 ? end : LLONG_MAX,
3519 req->sync.flags & IORING_FSYNC_DATASYNC);
3521 req_set_fail_links(req);
3522 io_req_complete(req, ret);
3526 static int io_fallocate_prep(struct io_kiocb *req,
3527 const struct io_uring_sqe *sqe)
3529 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3531 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3534 req->sync.off = READ_ONCE(sqe->off);
3535 req->sync.len = READ_ONCE(sqe->addr);
3536 req->sync.mode = READ_ONCE(sqe->len);
3540 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3544 /* fallocate always requiring blocking context */
3547 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3550 req_set_fail_links(req);
3551 io_req_complete(req, ret);
3555 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3557 const char __user *fname;
3560 if (unlikely(sqe->ioprio || sqe->buf_index))
3562 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3565 /* open.how should be already initialised */
3566 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3567 req->open.how.flags |= O_LARGEFILE;
3569 req->open.dfd = READ_ONCE(sqe->fd);
3570 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3571 req->open.filename = getname(fname);
3572 if (IS_ERR(req->open.filename)) {
3573 ret = PTR_ERR(req->open.filename);
3574 req->open.filename = NULL;
3577 req->open.nofile = rlimit(RLIMIT_NOFILE);
3578 req->flags |= REQ_F_NEED_CLEANUP;
3582 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3586 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3588 if (req->flags & REQ_F_NEED_CLEANUP)
3590 mode = READ_ONCE(sqe->len);
3591 flags = READ_ONCE(sqe->open_flags);
3592 req->open.how = build_open_how(flags, mode);
3593 return __io_openat_prep(req, sqe);
3596 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3598 struct open_how __user *how;
3602 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3604 if (req->flags & REQ_F_NEED_CLEANUP)
3606 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3607 len = READ_ONCE(sqe->len);
3608 if (len < OPEN_HOW_SIZE_VER0)
3611 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3616 return __io_openat_prep(req, sqe);
3619 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3621 struct open_flags op;
3628 ret = build_open_flags(&req->open.how, &op);
3632 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3636 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3639 ret = PTR_ERR(file);
3641 fsnotify_open(file);
3642 fd_install(ret, file);
3645 putname(req->open.filename);
3646 req->flags &= ~REQ_F_NEED_CLEANUP;
3648 req_set_fail_links(req);
3649 io_req_complete(req, ret);
3653 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3655 return io_openat2(req, force_nonblock);
3658 static int io_remove_buffers_prep(struct io_kiocb *req,
3659 const struct io_uring_sqe *sqe)
3661 struct io_provide_buf *p = &req->pbuf;
3664 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3667 tmp = READ_ONCE(sqe->fd);
3668 if (!tmp || tmp > USHRT_MAX)
3671 memset(p, 0, sizeof(*p));
3673 p->bgid = READ_ONCE(sqe->buf_group);
3677 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3678 int bgid, unsigned nbufs)
3682 /* shouldn't happen */
3686 /* the head kbuf is the list itself */
3687 while (!list_empty(&buf->list)) {
3688 struct io_buffer *nxt;
3690 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3691 list_del(&nxt->list);
3698 idr_remove(&ctx->io_buffer_idr, bgid);
3703 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3704 struct io_comp_state *cs)
3706 struct io_provide_buf *p = &req->pbuf;
3707 struct io_ring_ctx *ctx = req->ctx;
3708 struct io_buffer *head;
3711 io_ring_submit_lock(ctx, !force_nonblock);
3713 lockdep_assert_held(&ctx->uring_lock);
3716 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3718 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3720 io_ring_submit_lock(ctx, !force_nonblock);
3722 req_set_fail_links(req);
3723 __io_req_complete(req, ret, 0, cs);
3727 static int io_provide_buffers_prep(struct io_kiocb *req,
3728 const struct io_uring_sqe *sqe)
3730 struct io_provide_buf *p = &req->pbuf;
3733 if (sqe->ioprio || sqe->rw_flags)
3736 tmp = READ_ONCE(sqe->fd);
3737 if (!tmp || tmp > USHRT_MAX)
3740 p->addr = READ_ONCE(sqe->addr);
3741 p->len = READ_ONCE(sqe->len);
3743 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3746 p->bgid = READ_ONCE(sqe->buf_group);
3747 tmp = READ_ONCE(sqe->off);
3748 if (tmp > USHRT_MAX)
3754 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3756 struct io_buffer *buf;
3757 u64 addr = pbuf->addr;
3758 int i, bid = pbuf->bid;
3760 for (i = 0; i < pbuf->nbufs; i++) {
3761 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3766 buf->len = pbuf->len;
3771 INIT_LIST_HEAD(&buf->list);
3774 list_add_tail(&buf->list, &(*head)->list);
3778 return i ? i : -ENOMEM;
3781 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3782 struct io_comp_state *cs)
3784 struct io_provide_buf *p = &req->pbuf;
3785 struct io_ring_ctx *ctx = req->ctx;
3786 struct io_buffer *head, *list;
3789 io_ring_submit_lock(ctx, !force_nonblock);
3791 lockdep_assert_held(&ctx->uring_lock);
3793 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3795 ret = io_add_buffers(p, &head);
3800 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3803 __io_remove_buffers(ctx, head, p->bgid, -1U);
3808 io_ring_submit_unlock(ctx, !force_nonblock);
3810 req_set_fail_links(req);
3811 __io_req_complete(req, ret, 0, cs);
3815 static int io_epoll_ctl_prep(struct io_kiocb *req,
3816 const struct io_uring_sqe *sqe)
3818 #if defined(CONFIG_EPOLL)
3819 if (sqe->ioprio || sqe->buf_index)
3821 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3824 req->epoll.epfd = READ_ONCE(sqe->fd);
3825 req->epoll.op = READ_ONCE(sqe->len);
3826 req->epoll.fd = READ_ONCE(sqe->off);
3828 if (ep_op_has_event(req->epoll.op)) {
3829 struct epoll_event __user *ev;
3831 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3832 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3842 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3843 struct io_comp_state *cs)
3845 #if defined(CONFIG_EPOLL)
3846 struct io_epoll *ie = &req->epoll;
3849 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3850 if (force_nonblock && ret == -EAGAIN)
3854 req_set_fail_links(req);
3855 __io_req_complete(req, ret, 0, cs);
3862 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3864 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3865 if (sqe->ioprio || sqe->buf_index || sqe->off)
3867 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3870 req->madvise.addr = READ_ONCE(sqe->addr);
3871 req->madvise.len = READ_ONCE(sqe->len);
3872 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3879 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3881 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3882 struct io_madvise *ma = &req->madvise;
3888 ret = do_madvise(ma->addr, ma->len, ma->advice);
3890 req_set_fail_links(req);
3891 io_req_complete(req, ret);
3898 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3900 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3902 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3905 req->fadvise.offset = READ_ONCE(sqe->off);
3906 req->fadvise.len = READ_ONCE(sqe->len);
3907 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3911 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3913 struct io_fadvise *fa = &req->fadvise;
3916 if (force_nonblock) {
3917 switch (fa->advice) {
3918 case POSIX_FADV_NORMAL:
3919 case POSIX_FADV_RANDOM:
3920 case POSIX_FADV_SEQUENTIAL:
3927 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3929 req_set_fail_links(req);
3930 io_req_complete(req, ret);
3934 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3936 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3938 if (sqe->ioprio || sqe->buf_index)
3940 if (req->flags & REQ_F_FIXED_FILE)
3943 req->statx.dfd = READ_ONCE(sqe->fd);
3944 req->statx.mask = READ_ONCE(sqe->len);
3945 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3946 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3947 req->statx.flags = READ_ONCE(sqe->statx_flags);
3952 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3954 struct io_statx *ctx = &req->statx;
3957 if (force_nonblock) {
3958 /* only need file table for an actual valid fd */
3959 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3960 req->flags |= REQ_F_NO_FILE_TABLE;
3964 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3968 req_set_fail_links(req);
3969 io_req_complete(req, ret);
3973 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3976 * If we queue this for async, it must not be cancellable. That would
3977 * leave the 'file' in an undeterminate state, and here need to modify
3978 * io_wq_work.flags, so initialize io_wq_work firstly.
3980 io_req_init_async(req);
3981 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3983 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3985 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3986 sqe->rw_flags || sqe->buf_index)
3988 if (req->flags & REQ_F_FIXED_FILE)
3991 req->close.fd = READ_ONCE(sqe->fd);
3992 if ((req->file && req->file->f_op == &io_uring_fops))
3995 req->close.put_file = NULL;
3999 static int io_close(struct io_kiocb *req, bool force_nonblock,
4000 struct io_comp_state *cs)
4002 struct io_close *close = &req->close;
4005 /* might be already done during nonblock submission */
4006 if (!close->put_file) {
4007 ret = __close_fd_get_file(close->fd, &close->put_file);
4009 return (ret == -ENOENT) ? -EBADF : ret;
4012 /* if the file has a flush method, be safe and punt to async */
4013 if (close->put_file->f_op->flush && force_nonblock) {
4014 /* was never set, but play safe */
4015 req->flags &= ~REQ_F_NOWAIT;
4016 /* avoid grabbing files - we don't need the files */
4017 req->flags |= REQ_F_NO_FILE_TABLE;
4021 /* No ->flush() or already async, safely close from here */
4022 ret = filp_close(close->put_file, req->work.files);
4024 req_set_fail_links(req);
4025 fput(close->put_file);
4026 close->put_file = NULL;
4027 __io_req_complete(req, ret, 0, cs);
4031 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4033 struct io_ring_ctx *ctx = req->ctx;
4038 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4040 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4043 req->sync.off = READ_ONCE(sqe->off);
4044 req->sync.len = READ_ONCE(sqe->len);
4045 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4049 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4053 /* sync_file_range always requires a blocking context */
4057 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4060 req_set_fail_links(req);
4061 io_req_complete(req, ret);
4065 #if defined(CONFIG_NET)
4066 static int io_setup_async_msg(struct io_kiocb *req,
4067 struct io_async_msghdr *kmsg)
4071 if (io_alloc_async_ctx(req)) {
4072 if (kmsg->iov != kmsg->fast_iov)
4076 req->flags |= REQ_F_NEED_CLEANUP;
4077 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
4081 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4082 struct io_async_msghdr *iomsg)
4084 iomsg->iov = iomsg->fast_iov;
4085 iomsg->msg.msg_name = &iomsg->addr;
4086 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4087 req->sr_msg.msg_flags, &iomsg->iov);
4090 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4092 struct io_sr_msg *sr = &req->sr_msg;
4093 struct io_async_ctx *io = req->io;
4096 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4099 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4100 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4101 sr->len = READ_ONCE(sqe->len);
4103 #ifdef CONFIG_COMPAT
4104 if (req->ctx->compat)
4105 sr->msg_flags |= MSG_CMSG_COMPAT;
4108 if (!io || req->opcode == IORING_OP_SEND)
4110 /* iovec is already imported */
4111 if (req->flags & REQ_F_NEED_CLEANUP)
4114 ret = io_sendmsg_copy_hdr(req, &io->msg);
4116 req->flags |= REQ_F_NEED_CLEANUP;
4120 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4121 struct io_comp_state *cs)
4123 struct io_async_msghdr iomsg, *kmsg;
4124 struct socket *sock;
4128 sock = sock_from_file(req->file, &ret);
4129 if (unlikely(!sock))
4133 kmsg = &req->io->msg;
4134 kmsg->msg.msg_name = &req->io->msg.addr;
4135 /* if iov is set, it's allocated already */
4137 kmsg->iov = kmsg->fast_iov;
4138 kmsg->msg.msg_iter.iov = kmsg->iov;
4140 ret = io_sendmsg_copy_hdr(req, &iomsg);
4146 flags = req->sr_msg.msg_flags;
4147 if (flags & MSG_DONTWAIT)
4148 req->flags |= REQ_F_NOWAIT;
4149 else if (force_nonblock)
4150 flags |= MSG_DONTWAIT;
4152 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4153 if (force_nonblock && ret == -EAGAIN)
4154 return io_setup_async_msg(req, kmsg);
4155 if (ret == -ERESTARTSYS)
4158 if (kmsg->iov != kmsg->fast_iov)
4160 req->flags &= ~REQ_F_NEED_CLEANUP;
4162 req_set_fail_links(req);
4163 __io_req_complete(req, ret, 0, cs);
4167 static int io_send(struct io_kiocb *req, bool force_nonblock,
4168 struct io_comp_state *cs)
4170 struct io_sr_msg *sr = &req->sr_msg;
4173 struct socket *sock;
4177 sock = sock_from_file(req->file, &ret);
4178 if (unlikely(!sock))
4181 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4185 msg.msg_name = NULL;
4186 msg.msg_control = NULL;
4187 msg.msg_controllen = 0;
4188 msg.msg_namelen = 0;
4190 flags = req->sr_msg.msg_flags;
4191 if (flags & MSG_DONTWAIT)
4192 req->flags |= REQ_F_NOWAIT;
4193 else if (force_nonblock)
4194 flags |= MSG_DONTWAIT;
4196 msg.msg_flags = flags;
4197 ret = sock_sendmsg(sock, &msg);
4198 if (force_nonblock && ret == -EAGAIN)
4200 if (ret == -ERESTARTSYS)
4204 req_set_fail_links(req);
4205 __io_req_complete(req, ret, 0, cs);
4209 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4210 struct io_async_msghdr *iomsg)
4212 struct io_sr_msg *sr = &req->sr_msg;
4213 struct iovec __user *uiov;
4217 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4218 &iomsg->uaddr, &uiov, &iov_len);
4222 if (req->flags & REQ_F_BUFFER_SELECT) {
4225 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4227 sr->len = iomsg->iov[0].iov_len;
4228 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4232 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4233 &iomsg->iov, &iomsg->msg.msg_iter);
4241 #ifdef CONFIG_COMPAT
4242 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4243 struct io_async_msghdr *iomsg)
4245 struct compat_msghdr __user *msg_compat;
4246 struct io_sr_msg *sr = &req->sr_msg;
4247 struct compat_iovec __user *uiov;
4252 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4253 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4258 uiov = compat_ptr(ptr);
4259 if (req->flags & REQ_F_BUFFER_SELECT) {
4260 compat_ssize_t clen;
4264 if (!access_ok(uiov, sizeof(*uiov)))
4266 if (__get_user(clen, &uiov->iov_len))
4270 sr->len = iomsg->iov[0].iov_len;
4273 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4275 &iomsg->msg.msg_iter);
4284 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4285 struct io_async_msghdr *iomsg)
4287 iomsg->msg.msg_name = &iomsg->addr;
4288 iomsg->iov = iomsg->fast_iov;
4290 #ifdef CONFIG_COMPAT
4291 if (req->ctx->compat)
4292 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4295 return __io_recvmsg_copy_hdr(req, iomsg);
4298 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4301 struct io_sr_msg *sr = &req->sr_msg;
4302 struct io_buffer *kbuf;
4304 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4309 req->flags |= REQ_F_BUFFER_SELECTED;
4313 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4315 return io_put_kbuf(req, req->sr_msg.kbuf);
4318 static int io_recvmsg_prep(struct io_kiocb *req,
4319 const struct io_uring_sqe *sqe)
4321 struct io_sr_msg *sr = &req->sr_msg;
4322 struct io_async_ctx *io = req->io;
4325 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4328 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4329 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4330 sr->len = READ_ONCE(sqe->len);
4331 sr->bgid = READ_ONCE(sqe->buf_group);
4333 #ifdef CONFIG_COMPAT
4334 if (req->ctx->compat)
4335 sr->msg_flags |= MSG_CMSG_COMPAT;
4338 if (!io || req->opcode == IORING_OP_RECV)
4340 /* iovec is already imported */
4341 if (req->flags & REQ_F_NEED_CLEANUP)
4344 ret = io_recvmsg_copy_hdr(req, &io->msg);
4346 req->flags |= REQ_F_NEED_CLEANUP;
4350 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4351 struct io_comp_state *cs)
4353 struct io_async_msghdr iomsg, *kmsg;
4354 struct socket *sock;
4355 struct io_buffer *kbuf;
4357 int ret, cflags = 0;
4359 sock = sock_from_file(req->file, &ret);
4360 if (unlikely(!sock))
4364 kmsg = &req->io->msg;
4365 kmsg->msg.msg_name = &req->io->msg.addr;
4366 /* if iov is set, it's allocated already */
4368 kmsg->iov = kmsg->fast_iov;
4369 kmsg->msg.msg_iter.iov = kmsg->iov;
4371 ret = io_recvmsg_copy_hdr(req, &iomsg);
4377 if (req->flags & REQ_F_BUFFER_SELECT) {
4378 kbuf = io_recv_buffer_select(req, !force_nonblock);
4380 return PTR_ERR(kbuf);
4381 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4382 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4383 1, req->sr_msg.len);
4386 flags = req->sr_msg.msg_flags;
4387 if (flags & MSG_DONTWAIT)
4388 req->flags |= REQ_F_NOWAIT;
4389 else if (force_nonblock)
4390 flags |= MSG_DONTWAIT;
4392 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4393 kmsg->uaddr, flags);
4394 if (force_nonblock && ret == -EAGAIN)
4395 return io_setup_async_msg(req, kmsg);
4396 if (ret == -ERESTARTSYS)
4399 if (req->flags & REQ_F_BUFFER_SELECTED)
4400 cflags = io_put_recv_kbuf(req);
4401 if (kmsg->iov != kmsg->fast_iov)
4403 req->flags &= ~REQ_F_NEED_CLEANUP;
4405 req_set_fail_links(req);
4406 __io_req_complete(req, ret, cflags, cs);
4410 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4411 struct io_comp_state *cs)
4413 struct io_buffer *kbuf;
4414 struct io_sr_msg *sr = &req->sr_msg;
4416 void __user *buf = sr->buf;
4417 struct socket *sock;
4420 int ret, cflags = 0;
4422 sock = sock_from_file(req->file, &ret);
4423 if (unlikely(!sock))
4426 if (req->flags & REQ_F_BUFFER_SELECT) {
4427 kbuf = io_recv_buffer_select(req, !force_nonblock);
4429 return PTR_ERR(kbuf);
4430 buf = u64_to_user_ptr(kbuf->addr);
4433 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4437 msg.msg_name = NULL;
4438 msg.msg_control = NULL;
4439 msg.msg_controllen = 0;
4440 msg.msg_namelen = 0;
4441 msg.msg_iocb = NULL;
4444 flags = req->sr_msg.msg_flags;
4445 if (flags & MSG_DONTWAIT)
4446 req->flags |= REQ_F_NOWAIT;
4447 else if (force_nonblock)
4448 flags |= MSG_DONTWAIT;
4450 ret = sock_recvmsg(sock, &msg, flags);
4451 if (force_nonblock && ret == -EAGAIN)
4453 if (ret == -ERESTARTSYS)
4456 if (req->flags & REQ_F_BUFFER_SELECTED)
4457 cflags = io_put_recv_kbuf(req);
4459 req_set_fail_links(req);
4460 __io_req_complete(req, ret, cflags, cs);
4464 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4466 struct io_accept *accept = &req->accept;
4468 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4470 if (sqe->ioprio || sqe->len || sqe->buf_index)
4473 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4474 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4475 accept->flags = READ_ONCE(sqe->accept_flags);
4476 accept->nofile = rlimit(RLIMIT_NOFILE);
4480 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4481 struct io_comp_state *cs)
4483 struct io_accept *accept = &req->accept;
4484 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4487 if (req->file->f_flags & O_NONBLOCK)
4488 req->flags |= REQ_F_NOWAIT;
4490 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4491 accept->addr_len, accept->flags,
4493 if (ret == -EAGAIN && force_nonblock)
4496 if (ret == -ERESTARTSYS)
4498 req_set_fail_links(req);
4500 __io_req_complete(req, ret, 0, cs);
4504 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4506 struct io_connect *conn = &req->connect;
4507 struct io_async_ctx *io = req->io;
4509 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4511 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4514 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4515 conn->addr_len = READ_ONCE(sqe->addr2);
4520 return move_addr_to_kernel(conn->addr, conn->addr_len,
4521 &io->connect.address);
4524 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4525 struct io_comp_state *cs)
4527 struct io_async_ctx __io, *io;
4528 unsigned file_flags;
4534 ret = move_addr_to_kernel(req->connect.addr,
4535 req->connect.addr_len,
4536 &__io.connect.address);
4542 file_flags = force_nonblock ? O_NONBLOCK : 0;
4544 ret = __sys_connect_file(req->file, &io->connect.address,
4545 req->connect.addr_len, file_flags);
4546 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4549 if (io_alloc_async_ctx(req)) {
4553 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4556 if (ret == -ERESTARTSYS)
4560 req_set_fail_links(req);
4561 __io_req_complete(req, ret, 0, cs);
4564 #else /* !CONFIG_NET */
4565 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4570 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4571 struct io_comp_state *cs)
4576 static int io_send(struct io_kiocb *req, bool force_nonblock,
4577 struct io_comp_state *cs)
4582 static int io_recvmsg_prep(struct io_kiocb *req,
4583 const struct io_uring_sqe *sqe)
4588 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4589 struct io_comp_state *cs)
4594 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4595 struct io_comp_state *cs)
4600 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4605 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4606 struct io_comp_state *cs)
4611 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4616 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4617 struct io_comp_state *cs)
4621 #endif /* CONFIG_NET */
4623 struct io_poll_table {
4624 struct poll_table_struct pt;
4625 struct io_kiocb *req;
4629 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4630 __poll_t mask, task_work_func_t func)
4635 /* for instances that support it check for an event match first: */
4636 if (mask && !(mask & poll->events))
4639 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4641 list_del_init(&poll->wait.entry);
4644 init_task_work(&req->task_work, func);
4645 percpu_ref_get(&req->ctx->refs);
4648 * If we using the signalfd wait_queue_head for this wakeup, then
4649 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4650 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4651 * either, as the normal wakeup will suffice.
4653 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4656 * If this fails, then the task is exiting. When a task exits, the
4657 * work gets canceled, so just cancel this request as well instead
4658 * of executing it. We can't safely execute it anyway, as we may not
4659 * have the needed state needed for it anyway.
4661 ret = io_req_task_work_add(req, &req->task_work, twa_signal_ok);
4662 if (unlikely(ret)) {
4663 struct task_struct *tsk;
4665 WRITE_ONCE(poll->canceled, true);
4666 tsk = io_wq_get_task(req->ctx->io_wq);
4667 task_work_add(tsk, &req->task_work, 0);
4668 wake_up_process(tsk);
4673 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4674 __acquires(&req->ctx->completion_lock)
4676 struct io_ring_ctx *ctx = req->ctx;
4678 if (!req->result && !READ_ONCE(poll->canceled)) {
4679 struct poll_table_struct pt = { ._key = poll->events };
4681 req->result = vfs_poll(req->file, &pt) & poll->events;
4684 spin_lock_irq(&ctx->completion_lock);
4685 if (!req->result && !READ_ONCE(poll->canceled)) {
4686 add_wait_queue(poll->head, &poll->wait);
4693 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4695 /* pure poll stashes this in ->io, poll driven retry elsewhere */
4696 if (req->opcode == IORING_OP_POLL_ADD)
4697 return (struct io_poll_iocb *) req->io;
4698 return req->apoll->double_poll;
4701 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4703 if (req->opcode == IORING_OP_POLL_ADD)
4705 return &req->apoll->poll;
4708 static void io_poll_remove_double(struct io_kiocb *req)
4710 struct io_poll_iocb *poll = io_poll_get_double(req);
4712 lockdep_assert_held(&req->ctx->completion_lock);
4714 if (poll && poll->head) {
4715 struct wait_queue_head *head = poll->head;
4717 spin_lock(&head->lock);
4718 list_del_init(&poll->wait.entry);
4719 if (poll->wait.private)
4720 refcount_dec(&req->refs);
4722 spin_unlock(&head->lock);
4726 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4728 struct io_ring_ctx *ctx = req->ctx;
4730 io_poll_remove_double(req);
4731 req->poll.done = true;
4732 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4733 io_commit_cqring(ctx);
4736 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4738 struct io_ring_ctx *ctx = req->ctx;
4740 if (io_poll_rewait(req, &req->poll)) {
4741 spin_unlock_irq(&ctx->completion_lock);
4745 hash_del(&req->hash_node);
4746 io_poll_complete(req, req->result, 0);
4747 req->flags |= REQ_F_COMP_LOCKED;
4748 *nxt = io_put_req_find_next(req);
4749 spin_unlock_irq(&ctx->completion_lock);
4751 io_cqring_ev_posted(ctx);
4754 static void io_poll_task_func(struct callback_head *cb)
4756 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4757 struct io_ring_ctx *ctx = req->ctx;
4758 struct io_kiocb *nxt = NULL;
4760 io_poll_task_handler(req, &nxt);
4762 __io_req_task_submit(nxt);
4763 percpu_ref_put(&ctx->refs);
4766 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4767 int sync, void *key)
4769 struct io_kiocb *req = wait->private;
4770 struct io_poll_iocb *poll = io_poll_get_single(req);
4771 __poll_t mask = key_to_poll(key);
4773 /* for instances that support it check for an event match first: */
4774 if (mask && !(mask & poll->events))
4777 list_del_init(&wait->entry);
4779 if (poll && poll->head) {
4782 spin_lock(&poll->head->lock);
4783 done = list_empty(&poll->wait.entry);
4785 list_del_init(&poll->wait.entry);
4786 /* make sure double remove sees this as being gone */
4787 wait->private = NULL;
4788 spin_unlock(&poll->head->lock);
4790 __io_async_wake(req, poll, mask, io_poll_task_func);
4792 refcount_dec(&req->refs);
4796 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4797 wait_queue_func_t wake_func)
4801 poll->canceled = false;
4802 poll->events = events;
4803 INIT_LIST_HEAD(&poll->wait.entry);
4804 init_waitqueue_func_entry(&poll->wait, wake_func);
4807 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4808 struct wait_queue_head *head,
4809 struct io_poll_iocb **poll_ptr)
4811 struct io_kiocb *req = pt->req;
4814 * If poll->head is already set, it's because the file being polled
4815 * uses multiple waitqueues for poll handling (eg one for read, one
4816 * for write). Setup a separate io_poll_iocb if this happens.
4818 if (unlikely(poll->head)) {
4819 /* already have a 2nd entry, fail a third attempt */
4821 pt->error = -EINVAL;
4824 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4826 pt->error = -ENOMEM;
4829 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4830 refcount_inc(&req->refs);
4831 poll->wait.private = req;
4838 if (poll->events & EPOLLEXCLUSIVE)
4839 add_wait_queue_exclusive(head, &poll->wait);
4841 add_wait_queue(head, &poll->wait);
4844 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4845 struct poll_table_struct *p)
4847 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4848 struct async_poll *apoll = pt->req->apoll;
4850 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4853 static void io_async_task_func(struct callback_head *cb)
4855 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4856 struct async_poll *apoll = req->apoll;
4857 struct io_ring_ctx *ctx = req->ctx;
4859 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4861 if (io_poll_rewait(req, &apoll->poll)) {
4862 spin_unlock_irq(&ctx->completion_lock);
4863 percpu_ref_put(&ctx->refs);
4867 /* If req is still hashed, it cannot have been canceled. Don't check. */
4868 if (hash_hashed(&req->hash_node))
4869 hash_del(&req->hash_node);
4871 io_poll_remove_double(req);
4872 spin_unlock_irq(&ctx->completion_lock);
4874 if (!READ_ONCE(apoll->poll.canceled))
4875 __io_req_task_submit(req);
4877 __io_req_task_cancel(req, -ECANCELED);
4879 percpu_ref_put(&ctx->refs);
4880 kfree(apoll->double_poll);
4884 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4887 struct io_kiocb *req = wait->private;
4888 struct io_poll_iocb *poll = &req->apoll->poll;
4890 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4893 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4896 static void io_poll_req_insert(struct io_kiocb *req)
4898 struct io_ring_ctx *ctx = req->ctx;
4899 struct hlist_head *list;
4901 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4902 hlist_add_head(&req->hash_node, list);
4905 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4906 struct io_poll_iocb *poll,
4907 struct io_poll_table *ipt, __poll_t mask,
4908 wait_queue_func_t wake_func)
4909 __acquires(&ctx->completion_lock)
4911 struct io_ring_ctx *ctx = req->ctx;
4912 bool cancel = false;
4914 io_init_poll_iocb(poll, mask, wake_func);
4915 poll->file = req->file;
4916 poll->wait.private = req;
4918 ipt->pt._key = mask;
4920 ipt->error = -EINVAL;
4922 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4924 spin_lock_irq(&ctx->completion_lock);
4925 if (likely(poll->head)) {
4926 spin_lock(&poll->head->lock);
4927 if (unlikely(list_empty(&poll->wait.entry))) {
4933 if (mask || ipt->error)
4934 list_del_init(&poll->wait.entry);
4936 WRITE_ONCE(poll->canceled, true);
4937 else if (!poll->done) /* actually waiting for an event */
4938 io_poll_req_insert(req);
4939 spin_unlock(&poll->head->lock);
4945 static bool io_arm_poll_handler(struct io_kiocb *req)
4947 const struct io_op_def *def = &io_op_defs[req->opcode];
4948 struct io_ring_ctx *ctx = req->ctx;
4949 struct async_poll *apoll;
4950 struct io_poll_table ipt;
4954 if (!req->file || !file_can_poll(req->file))
4956 if (req->flags & REQ_F_POLLED)
4960 else if (def->pollout)
4964 /* if we can't nonblock try, then no point in arming a poll handler */
4965 if (!io_file_supports_async(req->file, rw))
4968 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4969 if (unlikely(!apoll))
4971 apoll->double_poll = NULL;
4973 req->flags |= REQ_F_POLLED;
4975 INIT_HLIST_NODE(&req->hash_node);
4979 mask |= POLLIN | POLLRDNORM;
4981 mask |= POLLOUT | POLLWRNORM;
4982 mask |= POLLERR | POLLPRI;
4984 ipt.pt._qproc = io_async_queue_proc;
4986 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4988 if (ret || ipt.error) {
4989 io_poll_remove_double(req);
4990 spin_unlock_irq(&ctx->completion_lock);
4991 kfree(apoll->double_poll);
4995 spin_unlock_irq(&ctx->completion_lock);
4996 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4997 apoll->poll.events);
5001 static bool __io_poll_remove_one(struct io_kiocb *req,
5002 struct io_poll_iocb *poll)
5004 bool do_complete = false;
5006 spin_lock(&poll->head->lock);
5007 WRITE_ONCE(poll->canceled, true);
5008 if (!list_empty(&poll->wait.entry)) {
5009 list_del_init(&poll->wait.entry);
5012 spin_unlock(&poll->head->lock);
5013 hash_del(&req->hash_node);
5017 static bool io_poll_remove_one(struct io_kiocb *req)
5021 io_poll_remove_double(req);
5023 if (req->opcode == IORING_OP_POLL_ADD) {
5024 do_complete = __io_poll_remove_one(req, &req->poll);
5026 struct async_poll *apoll = req->apoll;
5028 /* non-poll requests have submit ref still */
5029 do_complete = __io_poll_remove_one(req, &apoll->poll);
5032 kfree(apoll->double_poll);
5038 io_cqring_fill_event(req, -ECANCELED);
5039 io_commit_cqring(req->ctx);
5040 req->flags |= REQ_F_COMP_LOCKED;
5041 req_set_fail_links(req);
5049 * Returns true if we found and killed one or more poll requests
5051 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5053 struct hlist_node *tmp;
5054 struct io_kiocb *req;
5057 spin_lock_irq(&ctx->completion_lock);
5058 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5059 struct hlist_head *list;
5061 list = &ctx->cancel_hash[i];
5062 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5063 if (io_task_match(req, tsk))
5064 posted += io_poll_remove_one(req);
5067 spin_unlock_irq(&ctx->completion_lock);
5070 io_cqring_ev_posted(ctx);
5075 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5077 struct hlist_head *list;
5078 struct io_kiocb *req;
5080 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5081 hlist_for_each_entry(req, list, hash_node) {
5082 if (sqe_addr != req->user_data)
5084 if (io_poll_remove_one(req))
5092 static int io_poll_remove_prep(struct io_kiocb *req,
5093 const struct io_uring_sqe *sqe)
5095 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5097 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5101 req->poll.addr = READ_ONCE(sqe->addr);
5106 * Find a running poll command that matches one specified in sqe->addr,
5107 * and remove it if found.
5109 static int io_poll_remove(struct io_kiocb *req)
5111 struct io_ring_ctx *ctx = req->ctx;
5115 addr = req->poll.addr;
5116 spin_lock_irq(&ctx->completion_lock);
5117 ret = io_poll_cancel(ctx, addr);
5118 spin_unlock_irq(&ctx->completion_lock);
5121 req_set_fail_links(req);
5122 io_req_complete(req, ret);
5126 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5129 struct io_kiocb *req = wait->private;
5130 struct io_poll_iocb *poll = &req->poll;
5132 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5135 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5136 struct poll_table_struct *p)
5138 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5140 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
5143 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5145 struct io_poll_iocb *poll = &req->poll;
5148 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5150 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5155 events = READ_ONCE(sqe->poll32_events);
5157 events = swahw32(events);
5159 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5160 (events & EPOLLEXCLUSIVE);
5164 static int io_poll_add(struct io_kiocb *req)
5166 struct io_poll_iocb *poll = &req->poll;
5167 struct io_ring_ctx *ctx = req->ctx;
5168 struct io_poll_table ipt;
5171 INIT_HLIST_NODE(&req->hash_node);
5172 ipt.pt._qproc = io_poll_queue_proc;
5174 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5177 if (mask) { /* no async, we'd stolen it */
5179 io_poll_complete(req, mask, 0);
5181 spin_unlock_irq(&ctx->completion_lock);
5184 io_cqring_ev_posted(ctx);
5190 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5192 struct io_timeout_data *data = container_of(timer,
5193 struct io_timeout_data, timer);
5194 struct io_kiocb *req = data->req;
5195 struct io_ring_ctx *ctx = req->ctx;
5196 unsigned long flags;
5198 spin_lock_irqsave(&ctx->completion_lock, flags);
5199 atomic_set(&req->ctx->cq_timeouts,
5200 atomic_read(&req->ctx->cq_timeouts) + 1);
5203 * We could be racing with timeout deletion. If the list is empty,
5204 * then timeout lookup already found it and will be handling it.
5206 if (!list_empty(&req->timeout.list))
5207 list_del_init(&req->timeout.list);
5209 io_cqring_fill_event(req, -ETIME);
5210 io_commit_cqring(ctx);
5211 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5213 io_cqring_ev_posted(ctx);
5214 req_set_fail_links(req);
5216 return HRTIMER_NORESTART;
5219 static int __io_timeout_cancel(struct io_kiocb *req)
5223 list_del_init(&req->timeout.list);
5225 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5229 req_set_fail_links(req);
5230 req->flags |= REQ_F_COMP_LOCKED;
5231 io_cqring_fill_event(req, -ECANCELED);
5236 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5238 struct io_kiocb *req;
5241 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5242 if (user_data == req->user_data) {
5251 return __io_timeout_cancel(req);
5254 static int io_timeout_remove_prep(struct io_kiocb *req,
5255 const struct io_uring_sqe *sqe)
5257 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5259 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5261 if (sqe->ioprio || sqe->buf_index || sqe->len)
5264 req->timeout.addr = READ_ONCE(sqe->addr);
5265 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5266 if (req->timeout.flags)
5273 * Remove or update an existing timeout command
5275 static int io_timeout_remove(struct io_kiocb *req)
5277 struct io_ring_ctx *ctx = req->ctx;
5280 spin_lock_irq(&ctx->completion_lock);
5281 ret = io_timeout_cancel(ctx, req->timeout.addr);
5283 io_cqring_fill_event(req, ret);
5284 io_commit_cqring(ctx);
5285 spin_unlock_irq(&ctx->completion_lock);
5286 io_cqring_ev_posted(ctx);
5288 req_set_fail_links(req);
5293 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5294 bool is_timeout_link)
5296 struct io_timeout_data *data;
5298 u32 off = READ_ONCE(sqe->off);
5300 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5302 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5304 if (off && is_timeout_link)
5306 flags = READ_ONCE(sqe->timeout_flags);
5307 if (flags & ~IORING_TIMEOUT_ABS)
5310 req->timeout.off = off;
5312 if (!req->io && io_alloc_async_ctx(req))
5315 data = &req->io->timeout;
5318 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5321 if (flags & IORING_TIMEOUT_ABS)
5322 data->mode = HRTIMER_MODE_ABS;
5324 data->mode = HRTIMER_MODE_REL;
5326 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5330 static int io_timeout(struct io_kiocb *req)
5332 struct io_ring_ctx *ctx = req->ctx;
5333 struct io_timeout_data *data = &req->io->timeout;
5334 struct list_head *entry;
5335 u32 tail, off = req->timeout.off;
5337 spin_lock_irq(&ctx->completion_lock);
5340 * sqe->off holds how many events that need to occur for this
5341 * timeout event to be satisfied. If it isn't set, then this is
5342 * a pure timeout request, sequence isn't used.
5344 if (io_is_timeout_noseq(req)) {
5345 entry = ctx->timeout_list.prev;
5349 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5350 req->timeout.target_seq = tail + off;
5353 * Insertion sort, ensuring the first entry in the list is always
5354 * the one we need first.
5356 list_for_each_prev(entry, &ctx->timeout_list) {
5357 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5360 if (io_is_timeout_noseq(nxt))
5362 /* nxt.seq is behind @tail, otherwise would've been completed */
5363 if (off >= nxt->timeout.target_seq - tail)
5367 list_add(&req->timeout.list, entry);
5368 data->timer.function = io_timeout_fn;
5369 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5370 spin_unlock_irq(&ctx->completion_lock);
5374 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5376 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5378 return req->user_data == (unsigned long) data;
5381 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5383 enum io_wq_cancel cancel_ret;
5386 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5387 switch (cancel_ret) {
5388 case IO_WQ_CANCEL_OK:
5391 case IO_WQ_CANCEL_RUNNING:
5394 case IO_WQ_CANCEL_NOTFOUND:
5402 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5403 struct io_kiocb *req, __u64 sqe_addr,
5406 unsigned long flags;
5409 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5410 if (ret != -ENOENT) {
5411 spin_lock_irqsave(&ctx->completion_lock, flags);
5415 spin_lock_irqsave(&ctx->completion_lock, flags);
5416 ret = io_timeout_cancel(ctx, sqe_addr);
5419 ret = io_poll_cancel(ctx, sqe_addr);
5423 io_cqring_fill_event(req, ret);
5424 io_commit_cqring(ctx);
5425 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5426 io_cqring_ev_posted(ctx);
5429 req_set_fail_links(req);
5433 static int io_async_cancel_prep(struct io_kiocb *req,
5434 const struct io_uring_sqe *sqe)
5436 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5438 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5440 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5443 req->cancel.addr = READ_ONCE(sqe->addr);
5447 static int io_async_cancel(struct io_kiocb *req)
5449 struct io_ring_ctx *ctx = req->ctx;
5451 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5455 static int io_files_update_prep(struct io_kiocb *req,
5456 const struct io_uring_sqe *sqe)
5458 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5460 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5462 if (sqe->ioprio || sqe->rw_flags)
5465 req->files_update.offset = READ_ONCE(sqe->off);
5466 req->files_update.nr_args = READ_ONCE(sqe->len);
5467 if (!req->files_update.nr_args)
5469 req->files_update.arg = READ_ONCE(sqe->addr);
5473 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5474 struct io_comp_state *cs)
5476 struct io_ring_ctx *ctx = req->ctx;
5477 struct io_uring_files_update up;
5483 up.offset = req->files_update.offset;
5484 up.fds = req->files_update.arg;
5486 mutex_lock(&ctx->uring_lock);
5487 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5488 mutex_unlock(&ctx->uring_lock);
5491 req_set_fail_links(req);
5492 __io_req_complete(req, ret, 0, cs);
5496 static int io_req_defer_prep(struct io_kiocb *req,
5497 const struct io_uring_sqe *sqe)
5504 if (io_alloc_async_ctx(req))
5506 ret = io_prep_work_files(req);
5510 io_prep_async_work(req);
5512 switch (req->opcode) {
5515 case IORING_OP_READV:
5516 case IORING_OP_READ_FIXED:
5517 case IORING_OP_READ:
5518 ret = io_read_prep(req, sqe, true);
5520 case IORING_OP_WRITEV:
5521 case IORING_OP_WRITE_FIXED:
5522 case IORING_OP_WRITE:
5523 ret = io_write_prep(req, sqe, true);
5525 case IORING_OP_POLL_ADD:
5526 ret = io_poll_add_prep(req, sqe);
5528 case IORING_OP_POLL_REMOVE:
5529 ret = io_poll_remove_prep(req, sqe);
5531 case IORING_OP_FSYNC:
5532 ret = io_prep_fsync(req, sqe);
5534 case IORING_OP_SYNC_FILE_RANGE:
5535 ret = io_prep_sfr(req, sqe);
5537 case IORING_OP_SENDMSG:
5538 case IORING_OP_SEND:
5539 ret = io_sendmsg_prep(req, sqe);
5541 case IORING_OP_RECVMSG:
5542 case IORING_OP_RECV:
5543 ret = io_recvmsg_prep(req, sqe);
5545 case IORING_OP_CONNECT:
5546 ret = io_connect_prep(req, sqe);
5548 case IORING_OP_TIMEOUT:
5549 ret = io_timeout_prep(req, sqe, false);
5551 case IORING_OP_TIMEOUT_REMOVE:
5552 ret = io_timeout_remove_prep(req, sqe);
5554 case IORING_OP_ASYNC_CANCEL:
5555 ret = io_async_cancel_prep(req, sqe);
5557 case IORING_OP_LINK_TIMEOUT:
5558 ret = io_timeout_prep(req, sqe, true);
5560 case IORING_OP_ACCEPT:
5561 ret = io_accept_prep(req, sqe);
5563 case IORING_OP_FALLOCATE:
5564 ret = io_fallocate_prep(req, sqe);
5566 case IORING_OP_OPENAT:
5567 ret = io_openat_prep(req, sqe);
5569 case IORING_OP_CLOSE:
5570 ret = io_close_prep(req, sqe);
5572 case IORING_OP_FILES_UPDATE:
5573 ret = io_files_update_prep(req, sqe);
5575 case IORING_OP_STATX:
5576 ret = io_statx_prep(req, sqe);
5578 case IORING_OP_FADVISE:
5579 ret = io_fadvise_prep(req, sqe);
5581 case IORING_OP_MADVISE:
5582 ret = io_madvise_prep(req, sqe);
5584 case IORING_OP_OPENAT2:
5585 ret = io_openat2_prep(req, sqe);
5587 case IORING_OP_EPOLL_CTL:
5588 ret = io_epoll_ctl_prep(req, sqe);
5590 case IORING_OP_SPLICE:
5591 ret = io_splice_prep(req, sqe);
5593 case IORING_OP_PROVIDE_BUFFERS:
5594 ret = io_provide_buffers_prep(req, sqe);
5596 case IORING_OP_REMOVE_BUFFERS:
5597 ret = io_remove_buffers_prep(req, sqe);
5600 ret = io_tee_prep(req, sqe);
5603 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5612 static u32 io_get_sequence(struct io_kiocb *req)
5614 struct io_kiocb *pos;
5615 struct io_ring_ctx *ctx = req->ctx;
5616 u32 total_submitted, nr_reqs = 1;
5618 if (req->flags & REQ_F_LINK_HEAD)
5619 list_for_each_entry(pos, &req->link_list, link_list)
5622 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5623 return total_submitted - nr_reqs;
5626 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5628 struct io_ring_ctx *ctx = req->ctx;
5629 struct io_defer_entry *de;
5633 /* Still need defer if there is pending req in defer list. */
5634 if (likely(list_empty_careful(&ctx->defer_list) &&
5635 !(req->flags & REQ_F_IO_DRAIN)))
5638 seq = io_get_sequence(req);
5639 /* Still a chance to pass the sequence check */
5640 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5644 ret = io_req_defer_prep(req, sqe);
5648 io_prep_async_link(req);
5649 de = kmalloc(sizeof(*de), GFP_KERNEL);
5653 spin_lock_irq(&ctx->completion_lock);
5654 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5655 spin_unlock_irq(&ctx->completion_lock);
5657 io_queue_async_work(req);
5658 return -EIOCBQUEUED;
5661 trace_io_uring_defer(ctx, req, req->user_data);
5664 list_add_tail(&de->list, &ctx->defer_list);
5665 spin_unlock_irq(&ctx->completion_lock);
5666 return -EIOCBQUEUED;
5669 static void io_req_drop_files(struct io_kiocb *req)
5671 struct io_ring_ctx *ctx = req->ctx;
5672 unsigned long flags;
5674 spin_lock_irqsave(&ctx->inflight_lock, flags);
5675 list_del(&req->inflight_entry);
5676 if (waitqueue_active(&ctx->inflight_wait))
5677 wake_up(&ctx->inflight_wait);
5678 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5679 req->flags &= ~REQ_F_INFLIGHT;
5680 put_files_struct(req->work.files);
5681 put_nsproxy(req->work.nsproxy);
5682 req->work.files = NULL;
5685 static void __io_clean_op(struct io_kiocb *req)
5687 struct io_async_ctx *io = req->io;
5689 if (req->flags & REQ_F_BUFFER_SELECTED) {
5690 switch (req->opcode) {
5691 case IORING_OP_READV:
5692 case IORING_OP_READ_FIXED:
5693 case IORING_OP_READ:
5694 kfree((void *)(unsigned long)req->rw.addr);
5696 case IORING_OP_RECVMSG:
5697 case IORING_OP_RECV:
5698 kfree(req->sr_msg.kbuf);
5701 req->flags &= ~REQ_F_BUFFER_SELECTED;
5704 if (req->flags & REQ_F_NEED_CLEANUP) {
5705 switch (req->opcode) {
5706 case IORING_OP_READV:
5707 case IORING_OP_READ_FIXED:
5708 case IORING_OP_READ:
5709 case IORING_OP_WRITEV:
5710 case IORING_OP_WRITE_FIXED:
5711 case IORING_OP_WRITE:
5712 if (io->rw.free_iovec)
5713 kfree(io->rw.free_iovec);
5715 case IORING_OP_RECVMSG:
5716 case IORING_OP_SENDMSG:
5717 if (io->msg.iov != io->msg.fast_iov)
5720 case IORING_OP_SPLICE:
5722 io_put_file(req, req->splice.file_in,
5723 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5725 case IORING_OP_OPENAT:
5726 case IORING_OP_OPENAT2:
5727 if (req->open.filename)
5728 putname(req->open.filename);
5731 req->flags &= ~REQ_F_NEED_CLEANUP;
5734 if (req->flags & REQ_F_INFLIGHT)
5735 io_req_drop_files(req);
5738 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5739 bool force_nonblock, struct io_comp_state *cs)
5741 struct io_ring_ctx *ctx = req->ctx;
5744 switch (req->opcode) {
5746 ret = io_nop(req, cs);
5748 case IORING_OP_READV:
5749 case IORING_OP_READ_FIXED:
5750 case IORING_OP_READ:
5752 ret = io_read_prep(req, sqe, force_nonblock);
5756 ret = io_read(req, force_nonblock, cs);
5758 case IORING_OP_WRITEV:
5759 case IORING_OP_WRITE_FIXED:
5760 case IORING_OP_WRITE:
5762 ret = io_write_prep(req, sqe, force_nonblock);
5766 ret = io_write(req, force_nonblock, cs);
5768 case IORING_OP_FSYNC:
5770 ret = io_prep_fsync(req, sqe);
5774 ret = io_fsync(req, force_nonblock);
5776 case IORING_OP_POLL_ADD:
5778 ret = io_poll_add_prep(req, sqe);
5782 ret = io_poll_add(req);
5784 case IORING_OP_POLL_REMOVE:
5786 ret = io_poll_remove_prep(req, sqe);
5790 ret = io_poll_remove(req);
5792 case IORING_OP_SYNC_FILE_RANGE:
5794 ret = io_prep_sfr(req, sqe);
5798 ret = io_sync_file_range(req, force_nonblock);
5800 case IORING_OP_SENDMSG:
5801 case IORING_OP_SEND:
5803 ret = io_sendmsg_prep(req, sqe);
5807 if (req->opcode == IORING_OP_SENDMSG)
5808 ret = io_sendmsg(req, force_nonblock, cs);
5810 ret = io_send(req, force_nonblock, cs);
5812 case IORING_OP_RECVMSG:
5813 case IORING_OP_RECV:
5815 ret = io_recvmsg_prep(req, sqe);
5819 if (req->opcode == IORING_OP_RECVMSG)
5820 ret = io_recvmsg(req, force_nonblock, cs);
5822 ret = io_recv(req, force_nonblock, cs);
5824 case IORING_OP_TIMEOUT:
5826 ret = io_timeout_prep(req, sqe, false);
5830 ret = io_timeout(req);
5832 case IORING_OP_TIMEOUT_REMOVE:
5834 ret = io_timeout_remove_prep(req, sqe);
5838 ret = io_timeout_remove(req);
5840 case IORING_OP_ACCEPT:
5842 ret = io_accept_prep(req, sqe);
5846 ret = io_accept(req, force_nonblock, cs);
5848 case IORING_OP_CONNECT:
5850 ret = io_connect_prep(req, sqe);
5854 ret = io_connect(req, force_nonblock, cs);
5856 case IORING_OP_ASYNC_CANCEL:
5858 ret = io_async_cancel_prep(req, sqe);
5862 ret = io_async_cancel(req);
5864 case IORING_OP_FALLOCATE:
5866 ret = io_fallocate_prep(req, sqe);
5870 ret = io_fallocate(req, force_nonblock);
5872 case IORING_OP_OPENAT:
5874 ret = io_openat_prep(req, sqe);
5878 ret = io_openat(req, force_nonblock);
5880 case IORING_OP_CLOSE:
5882 ret = io_close_prep(req, sqe);
5886 ret = io_close(req, force_nonblock, cs);
5888 case IORING_OP_FILES_UPDATE:
5890 ret = io_files_update_prep(req, sqe);
5894 ret = io_files_update(req, force_nonblock, cs);
5896 case IORING_OP_STATX:
5898 ret = io_statx_prep(req, sqe);
5902 ret = io_statx(req, force_nonblock);
5904 case IORING_OP_FADVISE:
5906 ret = io_fadvise_prep(req, sqe);
5910 ret = io_fadvise(req, force_nonblock);
5912 case IORING_OP_MADVISE:
5914 ret = io_madvise_prep(req, sqe);
5918 ret = io_madvise(req, force_nonblock);
5920 case IORING_OP_OPENAT2:
5922 ret = io_openat2_prep(req, sqe);
5926 ret = io_openat2(req, force_nonblock);
5928 case IORING_OP_EPOLL_CTL:
5930 ret = io_epoll_ctl_prep(req, sqe);
5934 ret = io_epoll_ctl(req, force_nonblock, cs);
5936 case IORING_OP_SPLICE:
5938 ret = io_splice_prep(req, sqe);
5942 ret = io_splice(req, force_nonblock);
5944 case IORING_OP_PROVIDE_BUFFERS:
5946 ret = io_provide_buffers_prep(req, sqe);
5950 ret = io_provide_buffers(req, force_nonblock, cs);
5952 case IORING_OP_REMOVE_BUFFERS:
5954 ret = io_remove_buffers_prep(req, sqe);
5958 ret = io_remove_buffers(req, force_nonblock, cs);
5962 ret = io_tee_prep(req, sqe);
5966 ret = io_tee(req, force_nonblock);
5976 /* If the op doesn't have a file, we're not polling for it */
5977 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5978 const bool in_async = io_wq_current_is_worker();
5980 /* workqueue context doesn't hold uring_lock, grab it now */
5982 mutex_lock(&ctx->uring_lock);
5984 io_iopoll_req_issued(req);
5987 mutex_unlock(&ctx->uring_lock);
5993 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5995 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5996 struct io_kiocb *timeout;
5999 timeout = io_prep_linked_timeout(req);
6001 io_queue_linked_timeout(timeout);
6003 /* if NO_CANCEL is set, we must still run the work */
6004 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6005 IO_WQ_WORK_CANCEL) {
6011 ret = io_issue_sqe(req, NULL, false, NULL);
6013 * We can get EAGAIN for polled IO even though we're
6014 * forcing a sync submission from here, since we can't
6015 * wait for request slots on the block side.
6024 req_set_fail_links(req);
6025 io_req_complete(req, ret);
6028 return io_steal_work(req);
6031 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6034 struct fixed_file_table *table;
6036 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6037 return table->files[index & IORING_FILE_TABLE_MASK];
6040 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
6041 int fd, struct file **out_file, bool fixed)
6043 struct io_ring_ctx *ctx = req->ctx;
6047 if (unlikely(!ctx->file_data ||
6048 (unsigned) fd >= ctx->nr_user_files))
6050 fd = array_index_nospec(fd, ctx->nr_user_files);
6051 file = io_file_from_index(ctx, fd);
6053 req->fixed_file_refs = ctx->file_data->cur_refs;
6054 percpu_ref_get(req->fixed_file_refs);
6057 trace_io_uring_file_get(ctx, fd);
6058 file = __io_file_get(state, fd);
6061 if (file || io_op_defs[req->opcode].needs_file_no_error) {
6068 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6073 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6074 if (unlikely(!fixed && io_async_submit(req->ctx)))
6077 return io_file_get(state, req, fd, &req->file, fixed);
6080 static int io_grab_files(struct io_kiocb *req)
6082 struct io_ring_ctx *ctx = req->ctx;
6084 io_req_init_async(req);
6086 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
6089 req->work.files = get_files_struct(current);
6090 get_nsproxy(current->nsproxy);
6091 req->work.nsproxy = current->nsproxy;
6092 req->flags |= REQ_F_INFLIGHT;
6094 spin_lock_irq(&ctx->inflight_lock);
6095 list_add(&req->inflight_entry, &ctx->inflight_list);
6096 spin_unlock_irq(&ctx->inflight_lock);
6100 static inline int io_prep_work_files(struct io_kiocb *req)
6102 if (!io_op_defs[req->opcode].file_table)
6104 return io_grab_files(req);
6107 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6109 struct io_timeout_data *data = container_of(timer,
6110 struct io_timeout_data, timer);
6111 struct io_kiocb *req = data->req;
6112 struct io_ring_ctx *ctx = req->ctx;
6113 struct io_kiocb *prev = NULL;
6114 unsigned long flags;
6116 spin_lock_irqsave(&ctx->completion_lock, flags);
6119 * We don't expect the list to be empty, that will only happen if we
6120 * race with the completion of the linked work.
6122 if (!list_empty(&req->link_list)) {
6123 prev = list_entry(req->link_list.prev, struct io_kiocb,
6125 if (refcount_inc_not_zero(&prev->refs)) {
6126 list_del_init(&req->link_list);
6127 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6132 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6135 req_set_fail_links(prev);
6136 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6139 io_req_complete(req, -ETIME);
6141 return HRTIMER_NORESTART;
6144 static void __io_queue_linked_timeout(struct io_kiocb *req)
6147 * If the list is now empty, then our linked request finished before
6148 * we got a chance to setup the timer
6150 if (!list_empty(&req->link_list)) {
6151 struct io_timeout_data *data = &req->io->timeout;
6153 data->timer.function = io_link_timeout_fn;
6154 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6159 static void io_queue_linked_timeout(struct io_kiocb *req)
6161 struct io_ring_ctx *ctx = req->ctx;
6163 spin_lock_irq(&ctx->completion_lock);
6164 __io_queue_linked_timeout(req);
6165 spin_unlock_irq(&ctx->completion_lock);
6167 /* drop submission reference */
6171 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6173 struct io_kiocb *nxt;
6175 if (!(req->flags & REQ_F_LINK_HEAD))
6177 if (req->flags & REQ_F_LINK_TIMEOUT)
6180 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6182 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6185 req->flags |= REQ_F_LINK_TIMEOUT;
6189 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6190 struct io_comp_state *cs)
6192 struct io_kiocb *linked_timeout;
6193 struct io_kiocb *nxt;
6194 const struct cred *old_creds = NULL;
6198 linked_timeout = io_prep_linked_timeout(req);
6200 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6201 req->work.creds != current_cred()) {
6203 revert_creds(old_creds);
6204 if (old_creds == req->work.creds)
6205 old_creds = NULL; /* restored original creds */
6207 old_creds = override_creds(req->work.creds);
6210 ret = io_issue_sqe(req, sqe, true, cs);
6213 * We async punt it if the file wasn't marked NOWAIT, or if the file
6214 * doesn't support non-blocking read/write attempts
6216 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6217 if (!io_arm_poll_handler(req)) {
6219 ret = io_prep_work_files(req);
6223 * Queued up for async execution, worker will release
6224 * submit reference when the iocb is actually submitted.
6226 io_queue_async_work(req);
6230 io_queue_linked_timeout(linked_timeout);
6234 if (unlikely(ret)) {
6236 /* un-prep timeout, so it'll be killed as any other linked */
6237 req->flags &= ~REQ_F_LINK_TIMEOUT;
6238 req_set_fail_links(req);
6240 io_req_complete(req, ret);
6244 /* drop submission reference */
6245 nxt = io_put_req_find_next(req);
6247 io_queue_linked_timeout(linked_timeout);
6252 if (req->flags & REQ_F_FORCE_ASYNC)
6258 revert_creds(old_creds);
6261 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6262 struct io_comp_state *cs)
6266 ret = io_req_defer(req, sqe);
6268 if (ret != -EIOCBQUEUED) {
6270 req_set_fail_links(req);
6272 io_req_complete(req, ret);
6274 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6276 ret = io_req_defer_prep(req, sqe);
6282 * Never try inline submit of IOSQE_ASYNC is set, go straight
6283 * to async execution.
6285 io_req_init_async(req);
6286 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6287 io_queue_async_work(req);
6289 __io_queue_sqe(req, sqe, cs);
6293 static inline void io_queue_link_head(struct io_kiocb *req,
6294 struct io_comp_state *cs)
6296 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6298 io_req_complete(req, -ECANCELED);
6300 io_queue_sqe(req, NULL, cs);
6303 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6304 struct io_kiocb **link, struct io_comp_state *cs)
6306 struct io_ring_ctx *ctx = req->ctx;
6310 * If we already have a head request, queue this one for async
6311 * submittal once the head completes. If we don't have a head but
6312 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6313 * submitted sync once the chain is complete. If none of those
6314 * conditions are true (normal request), then just queue it.
6317 struct io_kiocb *head = *link;
6320 * Taking sequential execution of a link, draining both sides
6321 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6322 * requests in the link. So, it drains the head and the
6323 * next after the link request. The last one is done via
6324 * drain_next flag to persist the effect across calls.
6326 if (req->flags & REQ_F_IO_DRAIN) {
6327 head->flags |= REQ_F_IO_DRAIN;
6328 ctx->drain_next = 1;
6330 ret = io_req_defer_prep(req, sqe);
6331 if (unlikely(ret)) {
6332 /* fail even hard links since we don't submit */
6333 head->flags |= REQ_F_FAIL_LINK;
6336 trace_io_uring_link(ctx, req, head);
6337 list_add_tail(&req->link_list, &head->link_list);
6339 /* last request of a link, enqueue the link */
6340 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6341 io_queue_link_head(head, cs);
6345 if (unlikely(ctx->drain_next)) {
6346 req->flags |= REQ_F_IO_DRAIN;
6347 ctx->drain_next = 0;
6349 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6350 req->flags |= REQ_F_LINK_HEAD;
6351 INIT_LIST_HEAD(&req->link_list);
6353 ret = io_req_defer_prep(req, sqe);
6355 req->flags |= REQ_F_FAIL_LINK;
6358 io_queue_sqe(req, sqe, cs);
6366 * Batched submission is done, ensure local IO is flushed out.
6368 static void io_submit_state_end(struct io_submit_state *state)
6370 if (!list_empty(&state->comp.list))
6371 io_submit_flush_completions(&state->comp);
6372 blk_finish_plug(&state->plug);
6373 io_state_file_put(state);
6374 if (state->free_reqs)
6375 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6379 * Start submission side cache.
6381 static void io_submit_state_start(struct io_submit_state *state,
6382 struct io_ring_ctx *ctx, unsigned int max_ios)
6384 blk_start_plug(&state->plug);
6386 INIT_LIST_HEAD(&state->comp.list);
6387 state->comp.ctx = ctx;
6388 state->free_reqs = 0;
6390 state->ios_left = max_ios;
6393 static void io_commit_sqring(struct io_ring_ctx *ctx)
6395 struct io_rings *rings = ctx->rings;
6398 * Ensure any loads from the SQEs are done at this point,
6399 * since once we write the new head, the application could
6400 * write new data to them.
6402 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6406 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6407 * that is mapped by userspace. This means that care needs to be taken to
6408 * ensure that reads are stable, as we cannot rely on userspace always
6409 * being a good citizen. If members of the sqe are validated and then later
6410 * used, it's important that those reads are done through READ_ONCE() to
6411 * prevent a re-load down the line.
6413 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6415 u32 *sq_array = ctx->sq_array;
6419 * The cached sq head (or cq tail) serves two purposes:
6421 * 1) allows us to batch the cost of updating the user visible
6423 * 2) allows the kernel side to track the head on its own, even
6424 * though the application is the one updating it.
6426 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6427 if (likely(head < ctx->sq_entries))
6428 return &ctx->sq_sqes[head];
6430 /* drop invalid entries */
6431 ctx->cached_sq_dropped++;
6432 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6436 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6438 ctx->cached_sq_head++;
6441 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6442 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6443 IOSQE_BUFFER_SELECT)
6445 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6446 const struct io_uring_sqe *sqe,
6447 struct io_submit_state *state)
6449 unsigned int sqe_flags;
6452 req->opcode = READ_ONCE(sqe->opcode);
6453 req->user_data = READ_ONCE(sqe->user_data);
6458 /* one is dropped after submission, the other at completion */
6459 refcount_set(&req->refs, 2);
6460 req->task = current;
6461 get_task_struct(req->task);
6462 atomic_long_inc(&req->task->io_uring->req_issue);
6465 if (unlikely(req->opcode >= IORING_OP_LAST))
6468 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6471 sqe_flags = READ_ONCE(sqe->flags);
6472 /* enforce forwards compatibility on users */
6473 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6476 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6477 !io_op_defs[req->opcode].buffer_select)
6480 id = READ_ONCE(sqe->personality);
6482 io_req_init_async(req);
6483 req->work.creds = idr_find(&ctx->personality_idr, id);
6484 if (unlikely(!req->work.creds))
6486 get_cred(req->work.creds);
6489 /* same numerical values with corresponding REQ_F_*, safe to copy */
6490 req->flags |= sqe_flags;
6492 if (!io_op_defs[req->opcode].needs_file)
6495 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6498 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6500 struct io_submit_state state;
6501 struct io_kiocb *link = NULL;
6502 int i, submitted = 0;
6504 /* if we have a backlog and couldn't flush it all, return BUSY */
6505 if (test_bit(0, &ctx->sq_check_overflow)) {
6506 if (!list_empty(&ctx->cq_overflow_list) &&
6507 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6511 /* make sure SQ entry isn't read before tail */
6512 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6514 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6517 io_submit_state_start(&state, ctx, nr);
6519 for (i = 0; i < nr; i++) {
6520 const struct io_uring_sqe *sqe;
6521 struct io_kiocb *req;
6524 sqe = io_get_sqe(ctx);
6525 if (unlikely(!sqe)) {
6526 io_consume_sqe(ctx);
6529 req = io_alloc_req(ctx, &state);
6530 if (unlikely(!req)) {
6532 submitted = -EAGAIN;
6536 err = io_init_req(ctx, req, sqe, &state);
6537 io_consume_sqe(ctx);
6538 /* will complete beyond this point, count as submitted */
6541 if (unlikely(err)) {
6544 io_req_complete(req, err);
6548 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6549 true, io_async_submit(ctx));
6550 err = io_submit_sqe(req, sqe, &link, &state.comp);
6555 if (unlikely(submitted != nr)) {
6556 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6558 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6561 io_queue_link_head(link, &state.comp);
6562 io_submit_state_end(&state);
6564 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6565 io_commit_sqring(ctx);
6570 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6572 /* Tell userspace we may need a wakeup call */
6573 spin_lock_irq(&ctx->completion_lock);
6574 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6575 spin_unlock_irq(&ctx->completion_lock);
6578 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6580 spin_lock_irq(&ctx->completion_lock);
6581 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6582 spin_unlock_irq(&ctx->completion_lock);
6585 static int io_sq_thread(void *data)
6587 struct io_ring_ctx *ctx = data;
6588 const struct cred *old_cred;
6590 unsigned long timeout;
6593 complete(&ctx->sq_thread_comp);
6595 old_cred = override_creds(ctx->creds);
6597 timeout = jiffies + ctx->sq_thread_idle;
6598 while (!kthread_should_park()) {
6599 unsigned int to_submit;
6601 if (!list_empty(&ctx->iopoll_list)) {
6602 unsigned nr_events = 0;
6604 mutex_lock(&ctx->uring_lock);
6605 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6606 io_do_iopoll(ctx, &nr_events, 0);
6608 timeout = jiffies + ctx->sq_thread_idle;
6609 mutex_unlock(&ctx->uring_lock);
6612 to_submit = io_sqring_entries(ctx);
6615 * If submit got -EBUSY, flag us as needing the application
6616 * to enter the kernel to reap and flush events.
6618 if (!to_submit || ret == -EBUSY || need_resched()) {
6620 * Drop cur_mm before scheduling, we can't hold it for
6621 * long periods (or over schedule()). Do this before
6622 * adding ourselves to the waitqueue, as the unuse/drop
6625 io_sq_thread_drop_mm();
6628 * We're polling. If we're within the defined idle
6629 * period, then let us spin without work before going
6630 * to sleep. The exception is if we got EBUSY doing
6631 * more IO, we should wait for the application to
6632 * reap events and wake us up.
6634 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6635 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6636 !percpu_ref_is_dying(&ctx->refs))) {
6642 prepare_to_wait(&ctx->sqo_wait, &wait,
6643 TASK_INTERRUPTIBLE);
6646 * While doing polled IO, before going to sleep, we need
6647 * to check if there are new reqs added to iopoll_list,
6648 * it is because reqs may have been punted to io worker
6649 * and will be added to iopoll_list later, hence check
6650 * the iopoll_list again.
6652 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6653 !list_empty_careful(&ctx->iopoll_list)) {
6654 finish_wait(&ctx->sqo_wait, &wait);
6658 io_ring_set_wakeup_flag(ctx);
6660 to_submit = io_sqring_entries(ctx);
6661 if (!to_submit || ret == -EBUSY) {
6662 if (kthread_should_park()) {
6663 finish_wait(&ctx->sqo_wait, &wait);
6666 if (io_run_task_work()) {
6667 finish_wait(&ctx->sqo_wait, &wait);
6668 io_ring_clear_wakeup_flag(ctx);
6671 if (signal_pending(current))
6672 flush_signals(current);
6674 finish_wait(&ctx->sqo_wait, &wait);
6676 io_ring_clear_wakeup_flag(ctx);
6680 finish_wait(&ctx->sqo_wait, &wait);
6682 io_ring_clear_wakeup_flag(ctx);
6685 mutex_lock(&ctx->uring_lock);
6686 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6687 ret = io_submit_sqes(ctx, to_submit);
6688 mutex_unlock(&ctx->uring_lock);
6689 timeout = jiffies + ctx->sq_thread_idle;
6694 io_sq_thread_drop_mm();
6695 revert_creds(old_cred);
6702 struct io_wait_queue {
6703 struct wait_queue_entry wq;
6704 struct io_ring_ctx *ctx;
6706 unsigned nr_timeouts;
6709 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6711 struct io_ring_ctx *ctx = iowq->ctx;
6714 * Wake up if we have enough events, or if a timeout occurred since we
6715 * started waiting. For timeouts, we always want to return to userspace,
6716 * regardless of event count.
6718 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6719 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6722 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6723 int wake_flags, void *key)
6725 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6728 /* use noflush == true, as we can't safely rely on locking context */
6729 if (!io_should_wake(iowq, true))
6732 return autoremove_wake_function(curr, mode, wake_flags, key);
6736 * Wait until events become available, if we don't already have some. The
6737 * application must reap them itself, as they reside on the shared cq ring.
6739 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6740 const sigset_t __user *sig, size_t sigsz)
6742 struct io_wait_queue iowq = {
6745 .func = io_wake_function,
6746 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6749 .to_wait = min_events,
6751 struct io_rings *rings = ctx->rings;
6755 if (io_cqring_events(ctx, false) >= min_events)
6757 if (!io_run_task_work())
6762 #ifdef CONFIG_COMPAT
6763 if (in_compat_syscall())
6764 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6768 ret = set_user_sigmask(sig, sigsz);
6774 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6775 trace_io_uring_cqring_wait(ctx, min_events);
6777 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6778 TASK_INTERRUPTIBLE);
6779 /* make sure we run task_work before checking for signals */
6780 if (io_run_task_work())
6782 if (signal_pending(current)) {
6783 if (current->jobctl & JOBCTL_TASK_WORK) {
6784 spin_lock_irq(¤t->sighand->siglock);
6785 current->jobctl &= ~JOBCTL_TASK_WORK;
6786 recalc_sigpending();
6787 spin_unlock_irq(¤t->sighand->siglock);
6793 if (io_should_wake(&iowq, false))
6797 finish_wait(&ctx->wait, &iowq.wq);
6799 restore_saved_sigmask_unless(ret == -EINTR);
6801 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6804 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6806 #if defined(CONFIG_UNIX)
6807 if (ctx->ring_sock) {
6808 struct sock *sock = ctx->ring_sock->sk;
6809 struct sk_buff *skb;
6811 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6817 for (i = 0; i < ctx->nr_user_files; i++) {
6820 file = io_file_from_index(ctx, i);
6827 static void io_file_ref_kill(struct percpu_ref *ref)
6829 struct fixed_file_data *data;
6831 data = container_of(ref, struct fixed_file_data, refs);
6832 complete(&data->done);
6835 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6837 struct fixed_file_data *data = ctx->file_data;
6838 struct fixed_file_ref_node *ref_node = NULL;
6839 unsigned nr_tables, i;
6844 spin_lock(&data->lock);
6845 if (!list_empty(&data->ref_list))
6846 ref_node = list_first_entry(&data->ref_list,
6847 struct fixed_file_ref_node, node);
6848 spin_unlock(&data->lock);
6850 percpu_ref_kill(&ref_node->refs);
6852 percpu_ref_kill(&data->refs);
6854 /* wait for all refs nodes to complete */
6855 flush_delayed_work(&ctx->file_put_work);
6856 wait_for_completion(&data->done);
6858 __io_sqe_files_unregister(ctx);
6859 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6860 for (i = 0; i < nr_tables; i++)
6861 kfree(data->table[i].files);
6863 percpu_ref_exit(&data->refs);
6865 ctx->file_data = NULL;
6866 ctx->nr_user_files = 0;
6870 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6872 if (ctx->sqo_thread) {
6873 wait_for_completion(&ctx->sq_thread_comp);
6875 * The park is a bit of a work-around, without it we get
6876 * warning spews on shutdown with SQPOLL set and affinity
6877 * set to a single CPU.
6879 kthread_park(ctx->sqo_thread);
6880 kthread_stop(ctx->sqo_thread);
6881 ctx->sqo_thread = NULL;
6885 static void io_finish_async(struct io_ring_ctx *ctx)
6887 io_sq_thread_stop(ctx);
6890 io_wq_destroy(ctx->io_wq);
6895 #if defined(CONFIG_UNIX)
6897 * Ensure the UNIX gc is aware of our file set, so we are certain that
6898 * the io_uring can be safely unregistered on process exit, even if we have
6899 * loops in the file referencing.
6901 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6903 struct sock *sk = ctx->ring_sock->sk;
6904 struct scm_fp_list *fpl;
6905 struct sk_buff *skb;
6908 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6912 skb = alloc_skb(0, GFP_KERNEL);
6921 fpl->user = get_uid(ctx->user);
6922 for (i = 0; i < nr; i++) {
6923 struct file *file = io_file_from_index(ctx, i + offset);
6927 fpl->fp[nr_files] = get_file(file);
6928 unix_inflight(fpl->user, fpl->fp[nr_files]);
6933 fpl->max = SCM_MAX_FD;
6934 fpl->count = nr_files;
6935 UNIXCB(skb).fp = fpl;
6936 skb->destructor = unix_destruct_scm;
6937 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6938 skb_queue_head(&sk->sk_receive_queue, skb);
6940 for (i = 0; i < nr_files; i++)
6951 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6952 * causes regular reference counting to break down. We rely on the UNIX
6953 * garbage collection to take care of this problem for us.
6955 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6957 unsigned left, total;
6961 left = ctx->nr_user_files;
6963 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6965 ret = __io_sqe_files_scm(ctx, this_files, total);
6969 total += this_files;
6975 while (total < ctx->nr_user_files) {
6976 struct file *file = io_file_from_index(ctx, total);
6986 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6992 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6997 for (i = 0; i < nr_tables; i++) {
6998 struct fixed_file_table *table = &ctx->file_data->table[i];
6999 unsigned this_files;
7001 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7002 table->files = kcalloc(this_files, sizeof(struct file *),
7006 nr_files -= this_files;
7012 for (i = 0; i < nr_tables; i++) {
7013 struct fixed_file_table *table = &ctx->file_data->table[i];
7014 kfree(table->files);
7019 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7021 #if defined(CONFIG_UNIX)
7022 struct sock *sock = ctx->ring_sock->sk;
7023 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7024 struct sk_buff *skb;
7027 __skb_queue_head_init(&list);
7030 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7031 * remove this entry and rearrange the file array.
7033 skb = skb_dequeue(head);
7035 struct scm_fp_list *fp;
7037 fp = UNIXCB(skb).fp;
7038 for (i = 0; i < fp->count; i++) {
7041 if (fp->fp[i] != file)
7044 unix_notinflight(fp->user, fp->fp[i]);
7045 left = fp->count - 1 - i;
7047 memmove(&fp->fp[i], &fp->fp[i + 1],
7048 left * sizeof(struct file *));
7055 __skb_queue_tail(&list, skb);
7065 __skb_queue_tail(&list, skb);
7067 skb = skb_dequeue(head);
7070 if (skb_peek(&list)) {
7071 spin_lock_irq(&head->lock);
7072 while ((skb = __skb_dequeue(&list)) != NULL)
7073 __skb_queue_tail(head, skb);
7074 spin_unlock_irq(&head->lock);
7081 struct io_file_put {
7082 struct list_head list;
7086 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7088 struct fixed_file_data *file_data = ref_node->file_data;
7089 struct io_ring_ctx *ctx = file_data->ctx;
7090 struct io_file_put *pfile, *tmp;
7092 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7093 list_del(&pfile->list);
7094 io_ring_file_put(ctx, pfile->file);
7098 spin_lock(&file_data->lock);
7099 list_del(&ref_node->node);
7100 spin_unlock(&file_data->lock);
7102 percpu_ref_exit(&ref_node->refs);
7104 percpu_ref_put(&file_data->refs);
7107 static void io_file_put_work(struct work_struct *work)
7109 struct io_ring_ctx *ctx;
7110 struct llist_node *node;
7112 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7113 node = llist_del_all(&ctx->file_put_llist);
7116 struct fixed_file_ref_node *ref_node;
7117 struct llist_node *next = node->next;
7119 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7120 __io_file_put_work(ref_node);
7125 static void io_file_data_ref_zero(struct percpu_ref *ref)
7127 struct fixed_file_ref_node *ref_node;
7128 struct io_ring_ctx *ctx;
7132 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7133 ctx = ref_node->file_data->ctx;
7135 if (percpu_ref_is_dying(&ctx->file_data->refs))
7138 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7140 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7142 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7145 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7146 struct io_ring_ctx *ctx)
7148 struct fixed_file_ref_node *ref_node;
7150 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7152 return ERR_PTR(-ENOMEM);
7154 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7157 return ERR_PTR(-ENOMEM);
7159 INIT_LIST_HEAD(&ref_node->node);
7160 INIT_LIST_HEAD(&ref_node->file_list);
7161 ref_node->file_data = ctx->file_data;
7165 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7167 percpu_ref_exit(&ref_node->refs);
7171 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7174 __s32 __user *fds = (__s32 __user *) arg;
7179 struct fixed_file_ref_node *ref_node;
7185 if (nr_args > IORING_MAX_FIXED_FILES)
7188 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7189 if (!ctx->file_data)
7191 ctx->file_data->ctx = ctx;
7192 init_completion(&ctx->file_data->done);
7193 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7194 spin_lock_init(&ctx->file_data->lock);
7196 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7197 ctx->file_data->table = kcalloc(nr_tables,
7198 sizeof(struct fixed_file_table),
7200 if (!ctx->file_data->table) {
7201 kfree(ctx->file_data);
7202 ctx->file_data = NULL;
7206 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7207 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7208 kfree(ctx->file_data->table);
7209 kfree(ctx->file_data);
7210 ctx->file_data = NULL;
7214 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7215 percpu_ref_exit(&ctx->file_data->refs);
7216 kfree(ctx->file_data->table);
7217 kfree(ctx->file_data);
7218 ctx->file_data = NULL;
7222 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7223 struct fixed_file_table *table;
7227 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7229 /* allow sparse sets */
7235 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7236 index = i & IORING_FILE_TABLE_MASK;
7244 * Don't allow io_uring instances to be registered. If UNIX
7245 * isn't enabled, then this causes a reference cycle and this
7246 * instance can never get freed. If UNIX is enabled we'll
7247 * handle it just fine, but there's still no point in allowing
7248 * a ring fd as it doesn't support regular read/write anyway.
7250 if (file->f_op == &io_uring_fops) {
7255 table->files[index] = file;
7259 for (i = 0; i < ctx->nr_user_files; i++) {
7260 file = io_file_from_index(ctx, i);
7264 for (i = 0; i < nr_tables; i++)
7265 kfree(ctx->file_data->table[i].files);
7267 percpu_ref_exit(&ctx->file_data->refs);
7268 kfree(ctx->file_data->table);
7269 kfree(ctx->file_data);
7270 ctx->file_data = NULL;
7271 ctx->nr_user_files = 0;
7275 ret = io_sqe_files_scm(ctx);
7277 io_sqe_files_unregister(ctx);
7281 ref_node = alloc_fixed_file_ref_node(ctx);
7282 if (IS_ERR(ref_node)) {
7283 io_sqe_files_unregister(ctx);
7284 return PTR_ERR(ref_node);
7287 ctx->file_data->cur_refs = &ref_node->refs;
7288 spin_lock(&ctx->file_data->lock);
7289 list_add(&ref_node->node, &ctx->file_data->ref_list);
7290 spin_unlock(&ctx->file_data->lock);
7291 percpu_ref_get(&ctx->file_data->refs);
7295 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7298 #if defined(CONFIG_UNIX)
7299 struct sock *sock = ctx->ring_sock->sk;
7300 struct sk_buff_head *head = &sock->sk_receive_queue;
7301 struct sk_buff *skb;
7304 * See if we can merge this file into an existing skb SCM_RIGHTS
7305 * file set. If there's no room, fall back to allocating a new skb
7306 * and filling it in.
7308 spin_lock_irq(&head->lock);
7309 skb = skb_peek(head);
7311 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7313 if (fpl->count < SCM_MAX_FD) {
7314 __skb_unlink(skb, head);
7315 spin_unlock_irq(&head->lock);
7316 fpl->fp[fpl->count] = get_file(file);
7317 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7319 spin_lock_irq(&head->lock);
7320 __skb_queue_head(head, skb);
7325 spin_unlock_irq(&head->lock);
7332 return __io_sqe_files_scm(ctx, 1, index);
7338 static int io_queue_file_removal(struct fixed_file_data *data,
7341 struct io_file_put *pfile;
7342 struct percpu_ref *refs = data->cur_refs;
7343 struct fixed_file_ref_node *ref_node;
7345 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7349 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7351 list_add(&pfile->list, &ref_node->file_list);
7356 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7357 struct io_uring_files_update *up,
7360 struct fixed_file_data *data = ctx->file_data;
7361 struct fixed_file_ref_node *ref_node;
7366 bool needs_switch = false;
7368 if (check_add_overflow(up->offset, nr_args, &done))
7370 if (done > ctx->nr_user_files)
7373 ref_node = alloc_fixed_file_ref_node(ctx);
7374 if (IS_ERR(ref_node))
7375 return PTR_ERR(ref_node);
7378 fds = u64_to_user_ptr(up->fds);
7380 struct fixed_file_table *table;
7384 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7388 i = array_index_nospec(up->offset, ctx->nr_user_files);
7389 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7390 index = i & IORING_FILE_TABLE_MASK;
7391 if (table->files[index]) {
7392 file = table->files[index];
7393 err = io_queue_file_removal(data, file);
7396 table->files[index] = NULL;
7397 needs_switch = true;
7406 * Don't allow io_uring instances to be registered. If
7407 * UNIX isn't enabled, then this causes a reference
7408 * cycle and this instance can never get freed. If UNIX
7409 * is enabled we'll handle it just fine, but there's
7410 * still no point in allowing a ring fd as it doesn't
7411 * support regular read/write anyway.
7413 if (file->f_op == &io_uring_fops) {
7418 table->files[index] = file;
7419 err = io_sqe_file_register(ctx, file, i);
7421 table->files[index] = NULL;
7432 percpu_ref_kill(data->cur_refs);
7433 spin_lock(&data->lock);
7434 list_add(&ref_node->node, &data->ref_list);
7435 data->cur_refs = &ref_node->refs;
7436 spin_unlock(&data->lock);
7437 percpu_ref_get(&ctx->file_data->refs);
7439 destroy_fixed_file_ref_node(ref_node);
7441 return done ? done : err;
7444 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7447 struct io_uring_files_update up;
7449 if (!ctx->file_data)
7453 if (copy_from_user(&up, arg, sizeof(up)))
7458 return __io_sqe_files_update(ctx, &up, nr_args);
7461 static void io_free_work(struct io_wq_work *work)
7463 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7465 /* Consider that io_steal_work() relies on this ref */
7469 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7470 struct io_uring_params *p)
7472 struct io_wq_data data;
7474 struct io_ring_ctx *ctx_attach;
7475 unsigned int concurrency;
7478 data.user = ctx->user;
7479 data.free_work = io_free_work;
7480 data.do_work = io_wq_submit_work;
7482 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7483 /* Do QD, or 4 * CPUS, whatever is smallest */
7484 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7486 ctx->io_wq = io_wq_create(concurrency, &data);
7487 if (IS_ERR(ctx->io_wq)) {
7488 ret = PTR_ERR(ctx->io_wq);
7494 f = fdget(p->wq_fd);
7498 if (f.file->f_op != &io_uring_fops) {
7503 ctx_attach = f.file->private_data;
7504 /* @io_wq is protected by holding the fd */
7505 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7510 ctx->io_wq = ctx_attach->io_wq;
7516 static int io_uring_alloc_task_context(struct task_struct *task)
7518 struct io_uring_task *tctx;
7520 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7521 if (unlikely(!tctx))
7525 init_waitqueue_head(&tctx->wait);
7528 atomic_long_set(&tctx->req_issue, 0);
7529 atomic_long_set(&tctx->req_complete, 0);
7530 task->io_uring = tctx;
7534 void __io_uring_free(struct task_struct *tsk)
7536 struct io_uring_task *tctx = tsk->io_uring;
7538 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7539 xa_destroy(&tctx->xa);
7541 tsk->io_uring = NULL;
7544 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7545 struct io_uring_params *p)
7549 if (ctx->flags & IORING_SETUP_SQPOLL) {
7551 if (!capable(CAP_SYS_ADMIN))
7554 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7555 if (!ctx->sq_thread_idle)
7556 ctx->sq_thread_idle = HZ;
7558 if (p->flags & IORING_SETUP_SQ_AFF) {
7559 int cpu = p->sq_thread_cpu;
7562 if (cpu >= nr_cpu_ids)
7564 if (!cpu_online(cpu))
7567 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7571 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7574 if (IS_ERR(ctx->sqo_thread)) {
7575 ret = PTR_ERR(ctx->sqo_thread);
7576 ctx->sqo_thread = NULL;
7579 ret = io_uring_alloc_task_context(ctx->sqo_thread);
7582 wake_up_process(ctx->sqo_thread);
7583 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7584 /* Can't have SQ_AFF without SQPOLL */
7589 ret = io_init_wq_offload(ctx, p);
7595 io_finish_async(ctx);
7599 static inline void __io_unaccount_mem(struct user_struct *user,
7600 unsigned long nr_pages)
7602 atomic_long_sub(nr_pages, &user->locked_vm);
7605 static inline int __io_account_mem(struct user_struct *user,
7606 unsigned long nr_pages)
7608 unsigned long page_limit, cur_pages, new_pages;
7610 /* Don't allow more pages than we can safely lock */
7611 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7614 cur_pages = atomic_long_read(&user->locked_vm);
7615 new_pages = cur_pages + nr_pages;
7616 if (new_pages > page_limit)
7618 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7619 new_pages) != cur_pages);
7624 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7625 enum io_mem_account acct)
7628 __io_unaccount_mem(ctx->user, nr_pages);
7630 if (ctx->mm_account) {
7631 if (acct == ACCT_LOCKED)
7632 ctx->mm_account->locked_vm -= nr_pages;
7633 else if (acct == ACCT_PINNED)
7634 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7638 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7639 enum io_mem_account acct)
7643 if (ctx->limit_mem) {
7644 ret = __io_account_mem(ctx->user, nr_pages);
7649 if (ctx->mm_account) {
7650 if (acct == ACCT_LOCKED)
7651 ctx->mm_account->locked_vm += nr_pages;
7652 else if (acct == ACCT_PINNED)
7653 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7659 static void io_mem_free(void *ptr)
7666 page = virt_to_head_page(ptr);
7667 if (put_page_testzero(page))
7668 free_compound_page(page);
7671 static void *io_mem_alloc(size_t size)
7673 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7676 return (void *) __get_free_pages(gfp_flags, get_order(size));
7679 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7682 struct io_rings *rings;
7683 size_t off, sq_array_size;
7685 off = struct_size(rings, cqes, cq_entries);
7686 if (off == SIZE_MAX)
7690 off = ALIGN(off, SMP_CACHE_BYTES);
7698 sq_array_size = array_size(sizeof(u32), sq_entries);
7699 if (sq_array_size == SIZE_MAX)
7702 if (check_add_overflow(off, sq_array_size, &off))
7708 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7712 pages = (size_t)1 << get_order(
7713 rings_size(sq_entries, cq_entries, NULL));
7714 pages += (size_t)1 << get_order(
7715 array_size(sizeof(struct io_uring_sqe), sq_entries));
7720 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7724 if (!ctx->user_bufs)
7727 for (i = 0; i < ctx->nr_user_bufs; i++) {
7728 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7730 for (j = 0; j < imu->nr_bvecs; j++)
7731 unpin_user_page(imu->bvec[j].bv_page);
7733 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7738 kfree(ctx->user_bufs);
7739 ctx->user_bufs = NULL;
7740 ctx->nr_user_bufs = 0;
7744 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7745 void __user *arg, unsigned index)
7747 struct iovec __user *src;
7749 #ifdef CONFIG_COMPAT
7751 struct compat_iovec __user *ciovs;
7752 struct compat_iovec ciov;
7754 ciovs = (struct compat_iovec __user *) arg;
7755 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7758 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7759 dst->iov_len = ciov.iov_len;
7763 src = (struct iovec __user *) arg;
7764 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7769 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7772 struct vm_area_struct **vmas = NULL;
7773 struct page **pages = NULL;
7774 int i, j, got_pages = 0;
7779 if (!nr_args || nr_args > UIO_MAXIOV)
7782 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7784 if (!ctx->user_bufs)
7787 for (i = 0; i < nr_args; i++) {
7788 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7789 unsigned long off, start, end, ubuf;
7794 ret = io_copy_iov(ctx, &iov, arg, i);
7799 * Don't impose further limits on the size and buffer
7800 * constraints here, we'll -EINVAL later when IO is
7801 * submitted if they are wrong.
7804 if (!iov.iov_base || !iov.iov_len)
7807 /* arbitrary limit, but we need something */
7808 if (iov.iov_len > SZ_1G)
7811 ubuf = (unsigned long) iov.iov_base;
7812 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7813 start = ubuf >> PAGE_SHIFT;
7814 nr_pages = end - start;
7816 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7821 if (!pages || nr_pages > got_pages) {
7824 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7826 vmas = kvmalloc_array(nr_pages,
7827 sizeof(struct vm_area_struct *),
7829 if (!pages || !vmas) {
7831 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7834 got_pages = nr_pages;
7837 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7841 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7846 mmap_read_lock(current->mm);
7847 pret = pin_user_pages(ubuf, nr_pages,
7848 FOLL_WRITE | FOLL_LONGTERM,
7850 if (pret == nr_pages) {
7851 /* don't support file backed memory */
7852 for (j = 0; j < nr_pages; j++) {
7853 struct vm_area_struct *vma = vmas[j];
7856 !is_file_hugepages(vma->vm_file)) {
7862 ret = pret < 0 ? pret : -EFAULT;
7864 mmap_read_unlock(current->mm);
7867 * if we did partial map, or found file backed vmas,
7868 * release any pages we did get
7871 unpin_user_pages(pages, pret);
7872 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7877 off = ubuf & ~PAGE_MASK;
7879 for (j = 0; j < nr_pages; j++) {
7882 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7883 imu->bvec[j].bv_page = pages[j];
7884 imu->bvec[j].bv_len = vec_len;
7885 imu->bvec[j].bv_offset = off;
7889 /* store original address for later verification */
7891 imu->len = iov.iov_len;
7892 imu->nr_bvecs = nr_pages;
7894 ctx->nr_user_bufs++;
7902 io_sqe_buffer_unregister(ctx);
7906 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7908 __s32 __user *fds = arg;
7914 if (copy_from_user(&fd, fds, sizeof(*fds)))
7917 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7918 if (IS_ERR(ctx->cq_ev_fd)) {
7919 int ret = PTR_ERR(ctx->cq_ev_fd);
7920 ctx->cq_ev_fd = NULL;
7927 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7929 if (ctx->cq_ev_fd) {
7930 eventfd_ctx_put(ctx->cq_ev_fd);
7931 ctx->cq_ev_fd = NULL;
7938 static int __io_destroy_buffers(int id, void *p, void *data)
7940 struct io_ring_ctx *ctx = data;
7941 struct io_buffer *buf = p;
7943 __io_remove_buffers(ctx, buf, id, -1U);
7947 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7949 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7950 idr_destroy(&ctx->io_buffer_idr);
7953 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7955 io_finish_async(ctx);
7956 io_sqe_buffer_unregister(ctx);
7958 if (ctx->sqo_task) {
7959 put_task_struct(ctx->sqo_task);
7960 ctx->sqo_task = NULL;
7961 mmdrop(ctx->mm_account);
7962 ctx->mm_account = NULL;
7965 io_sqe_files_unregister(ctx);
7966 io_eventfd_unregister(ctx);
7967 io_destroy_buffers(ctx);
7968 idr_destroy(&ctx->personality_idr);
7970 #if defined(CONFIG_UNIX)
7971 if (ctx->ring_sock) {
7972 ctx->ring_sock->file = NULL; /* so that iput() is called */
7973 sock_release(ctx->ring_sock);
7977 io_mem_free(ctx->rings);
7978 io_mem_free(ctx->sq_sqes);
7980 percpu_ref_exit(&ctx->refs);
7981 free_uid(ctx->user);
7982 put_cred(ctx->creds);
7983 kfree(ctx->cancel_hash);
7984 kmem_cache_free(req_cachep, ctx->fallback_req);
7988 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7990 struct io_ring_ctx *ctx = file->private_data;
7993 poll_wait(file, &ctx->cq_wait, wait);
7995 * synchronizes with barrier from wq_has_sleeper call in
7999 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
8000 ctx->rings->sq_ring_entries)
8001 mask |= EPOLLOUT | EPOLLWRNORM;
8002 if (io_cqring_events(ctx, false))
8003 mask |= EPOLLIN | EPOLLRDNORM;
8008 static int io_uring_fasync(int fd, struct file *file, int on)
8010 struct io_ring_ctx *ctx = file->private_data;
8012 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8015 static int io_remove_personalities(int id, void *p, void *data)
8017 struct io_ring_ctx *ctx = data;
8018 const struct cred *cred;
8020 cred = idr_remove(&ctx->personality_idr, id);
8026 static void io_ring_exit_work(struct work_struct *work)
8028 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8032 * If we're doing polled IO and end up having requests being
8033 * submitted async (out-of-line), then completions can come in while
8034 * we're waiting for refs to drop. We need to reap these manually,
8035 * as nobody else will be looking for them.
8039 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8040 io_iopoll_try_reap_events(ctx);
8041 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8042 io_ring_ctx_free(ctx);
8045 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8047 mutex_lock(&ctx->uring_lock);
8048 percpu_ref_kill(&ctx->refs);
8049 mutex_unlock(&ctx->uring_lock);
8051 io_kill_timeouts(ctx, NULL);
8052 io_poll_remove_all(ctx, NULL);
8055 io_wq_cancel_all(ctx->io_wq);
8057 /* if we failed setting up the ctx, we might not have any rings */
8059 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8060 io_iopoll_try_reap_events(ctx);
8061 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8064 * Do this upfront, so we won't have a grace period where the ring
8065 * is closed but resources aren't reaped yet. This can cause
8066 * spurious failure in setting up a new ring.
8068 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8071 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8073 * Use system_unbound_wq to avoid spawning tons of event kworkers
8074 * if we're exiting a ton of rings at the same time. It just adds
8075 * noise and overhead, there's no discernable change in runtime
8076 * over using system_wq.
8078 queue_work(system_unbound_wq, &ctx->exit_work);
8081 static int io_uring_release(struct inode *inode, struct file *file)
8083 struct io_ring_ctx *ctx = file->private_data;
8085 file->private_data = NULL;
8086 io_ring_ctx_wait_and_kill(ctx);
8090 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8092 struct files_struct *files = data;
8094 return !files || work->files == files;
8098 * Returns true if 'preq' is the link parent of 'req'
8100 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8102 struct io_kiocb *link;
8104 if (!(preq->flags & REQ_F_LINK_HEAD))
8107 list_for_each_entry(link, &preq->link_list, link_list) {
8115 static bool io_match_link_files(struct io_kiocb *req,
8116 struct files_struct *files)
8118 struct io_kiocb *link;
8120 if (io_match_files(req, files))
8122 if (req->flags & REQ_F_LINK_HEAD) {
8123 list_for_each_entry(link, &req->link_list, link_list) {
8124 if (io_match_files(link, files))
8132 * We're looking to cancel 'req' because it's holding on to our files, but
8133 * 'req' could be a link to another request. See if it is, and cancel that
8134 * parent request if so.
8136 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8138 struct hlist_node *tmp;
8139 struct io_kiocb *preq;
8143 spin_lock_irq(&ctx->completion_lock);
8144 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8145 struct hlist_head *list;
8147 list = &ctx->cancel_hash[i];
8148 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8149 found = io_match_link(preq, req);
8151 io_poll_remove_one(preq);
8156 spin_unlock_irq(&ctx->completion_lock);
8160 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8161 struct io_kiocb *req)
8163 struct io_kiocb *preq;
8166 spin_lock_irq(&ctx->completion_lock);
8167 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8168 found = io_match_link(preq, req);
8170 __io_timeout_cancel(preq);
8174 spin_unlock_irq(&ctx->completion_lock);
8178 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8180 return io_match_link(container_of(work, struct io_kiocb, work), data);
8183 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8185 enum io_wq_cancel cret;
8187 /* cancel this particular work, if it's running */
8188 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8189 if (cret != IO_WQ_CANCEL_NOTFOUND)
8192 /* find links that hold this pending, cancel those */
8193 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8194 if (cret != IO_WQ_CANCEL_NOTFOUND)
8197 /* if we have a poll link holding this pending, cancel that */
8198 if (io_poll_remove_link(ctx, req))
8201 /* final option, timeout link is holding this req pending */
8202 io_timeout_remove_link(ctx, req);
8205 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8206 struct files_struct *files)
8208 struct io_defer_entry *de = NULL;
8211 spin_lock_irq(&ctx->completion_lock);
8212 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8213 if (io_match_link_files(de->req, files)) {
8214 list_cut_position(&list, &ctx->defer_list, &de->list);
8218 spin_unlock_irq(&ctx->completion_lock);
8220 while (!list_empty(&list)) {
8221 de = list_first_entry(&list, struct io_defer_entry, list);
8222 list_del_init(&de->list);
8223 req_set_fail_links(de->req);
8224 io_put_req(de->req);
8225 io_req_complete(de->req, -ECANCELED);
8231 * Returns true if we found and killed one or more files pinning requests
8233 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8234 struct files_struct *files)
8236 if (list_empty_careful(&ctx->inflight_list))
8239 io_cancel_defer_files(ctx, files);
8240 /* cancel all at once, should be faster than doing it one by one*/
8241 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8243 while (!list_empty_careful(&ctx->inflight_list)) {
8244 struct io_kiocb *cancel_req = NULL, *req;
8247 spin_lock_irq(&ctx->inflight_lock);
8248 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8249 if (files && req->work.files != files)
8251 /* req is being completed, ignore */
8252 if (!refcount_inc_not_zero(&req->refs))
8258 prepare_to_wait(&ctx->inflight_wait, &wait,
8259 TASK_UNINTERRUPTIBLE);
8260 spin_unlock_irq(&ctx->inflight_lock);
8262 /* We need to keep going until we don't find a matching req */
8265 /* cancel this request, or head link requests */
8266 io_attempt_cancel(ctx, cancel_req);
8267 io_put_req(cancel_req);
8268 /* cancellations _may_ trigger task work */
8271 finish_wait(&ctx->inflight_wait, &wait);
8277 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8279 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8280 struct task_struct *task = data;
8282 return io_task_match(req, task);
8285 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8286 struct task_struct *task,
8287 struct files_struct *files)
8291 ret = io_uring_cancel_files(ctx, files);
8293 enum io_wq_cancel cret;
8295 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8296 if (cret != IO_WQ_CANCEL_NOTFOUND)
8299 /* SQPOLL thread does its own polling */
8300 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8301 while (!list_empty_careful(&ctx->iopoll_list)) {
8302 io_iopoll_try_reap_events(ctx);
8307 ret |= io_poll_remove_all(ctx, task);
8308 ret |= io_kill_timeouts(ctx, task);
8315 * We need to iteratively cancel requests, in case a request has dependent
8316 * hard links. These persist even for failure of cancelations, hence keep
8317 * looping until none are found.
8319 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8320 struct files_struct *files)
8322 struct task_struct *task = current;
8324 if (ctx->flags & IORING_SETUP_SQPOLL)
8325 task = ctx->sqo_thread;
8327 io_cqring_overflow_flush(ctx, true, task, files);
8329 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8336 * Note that this task has used io_uring. We use it for cancelation purposes.
8338 static int io_uring_add_task_file(struct file *file)
8340 if (unlikely(!current->io_uring)) {
8343 ret = io_uring_alloc_task_context(current);
8347 if (current->io_uring->last != file) {
8348 XA_STATE(xas, ¤t->io_uring->xa, (unsigned long) file);
8352 old = xas_load(&xas);
8356 xas_store(&xas, file);
8360 current->io_uring->last = file;
8367 * Remove this io_uring_file -> task mapping.
8369 static void io_uring_del_task_file(struct file *file)
8371 struct io_uring_task *tctx = current->io_uring;
8372 XA_STATE(xas, &tctx->xa, (unsigned long) file);
8374 if (tctx->last == file)
8378 file = xas_store(&xas, NULL);
8385 static void __io_uring_attempt_task_drop(struct file *file)
8387 XA_STATE(xas, ¤t->io_uring->xa, (unsigned long) file);
8391 old = xas_load(&xas);
8395 io_uring_del_task_file(file);
8399 * Drop task note for this file if we're the only ones that hold it after
8402 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8404 if (!current->io_uring)
8407 * fput() is pending, will be 2 if the only other ref is our potential
8408 * task file note. If the task is exiting, drop regardless of count.
8410 if (!exiting && atomic_long_read(&file->f_count) != 2)
8413 __io_uring_attempt_task_drop(file);
8416 void __io_uring_files_cancel(struct files_struct *files)
8418 struct io_uring_task *tctx = current->io_uring;
8419 XA_STATE(xas, &tctx->xa, 0);
8421 /* make sure overflow events are dropped */
8422 tctx->in_idle = true;
8425 struct io_ring_ctx *ctx;
8429 file = xas_next_entry(&xas, ULONG_MAX);
8435 ctx = file->private_data;
8437 io_uring_cancel_task_requests(ctx, files);
8439 io_uring_del_task_file(file);
8443 static inline bool io_uring_task_idle(struct io_uring_task *tctx)
8445 return atomic_long_read(&tctx->req_issue) ==
8446 atomic_long_read(&tctx->req_complete);
8450 * Find any io_uring fd that this task has registered or done IO on, and cancel
8453 void __io_uring_task_cancel(void)
8455 struct io_uring_task *tctx = current->io_uring;
8459 /* make sure overflow events are dropped */
8460 tctx->in_idle = true;
8462 while (!io_uring_task_idle(tctx)) {
8463 /* read completions before cancelations */
8464 completions = atomic_long_read(&tctx->req_complete);
8465 __io_uring_files_cancel(NULL);
8467 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8470 * If we've seen completions, retry. This avoids a race where
8471 * a completion comes in before we did prepare_to_wait().
8473 if (completions != atomic_long_read(&tctx->req_complete))
8475 if (io_uring_task_idle(tctx))
8480 finish_wait(&tctx->wait, &wait);
8481 tctx->in_idle = false;
8484 static int io_uring_flush(struct file *file, void *data)
8486 struct io_ring_ctx *ctx = file->private_data;
8489 * If the task is going away, cancel work it may have pending
8491 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8494 io_uring_cancel_task_requests(ctx, data);
8495 io_uring_attempt_task_drop(file, !data);
8499 static void *io_uring_validate_mmap_request(struct file *file,
8500 loff_t pgoff, size_t sz)
8502 struct io_ring_ctx *ctx = file->private_data;
8503 loff_t offset = pgoff << PAGE_SHIFT;
8508 case IORING_OFF_SQ_RING:
8509 case IORING_OFF_CQ_RING:
8512 case IORING_OFF_SQES:
8516 return ERR_PTR(-EINVAL);
8519 page = virt_to_head_page(ptr);
8520 if (sz > page_size(page))
8521 return ERR_PTR(-EINVAL);
8528 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8530 size_t sz = vma->vm_end - vma->vm_start;
8534 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8536 return PTR_ERR(ptr);
8538 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8539 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8542 #else /* !CONFIG_MMU */
8544 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8546 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8549 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8551 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8554 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8555 unsigned long addr, unsigned long len,
8556 unsigned long pgoff, unsigned long flags)
8560 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8562 return PTR_ERR(ptr);
8564 return (unsigned long) ptr;
8567 #endif /* !CONFIG_MMU */
8569 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8570 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8573 struct io_ring_ctx *ctx;
8580 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8588 if (f.file->f_op != &io_uring_fops)
8592 ctx = f.file->private_data;
8593 if (!percpu_ref_tryget(&ctx->refs))
8597 * For SQ polling, the thread will do all submissions and completions.
8598 * Just return the requested submit count, and wake the thread if
8602 if (ctx->flags & IORING_SETUP_SQPOLL) {
8603 if (!list_empty_careful(&ctx->cq_overflow_list))
8604 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8605 if (flags & IORING_ENTER_SQ_WAKEUP)
8606 wake_up(&ctx->sqo_wait);
8607 submitted = to_submit;
8608 } else if (to_submit) {
8609 ret = io_uring_add_task_file(f.file);
8612 mutex_lock(&ctx->uring_lock);
8613 submitted = io_submit_sqes(ctx, to_submit);
8614 mutex_unlock(&ctx->uring_lock);
8616 if (submitted != to_submit)
8619 if (flags & IORING_ENTER_GETEVENTS) {
8620 min_complete = min(min_complete, ctx->cq_entries);
8623 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8624 * space applications don't need to do io completion events
8625 * polling again, they can rely on io_sq_thread to do polling
8626 * work, which can reduce cpu usage and uring_lock contention.
8628 if (ctx->flags & IORING_SETUP_IOPOLL &&
8629 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8630 ret = io_iopoll_check(ctx, min_complete);
8632 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8637 percpu_ref_put(&ctx->refs);
8640 return submitted ? submitted : ret;
8643 #ifdef CONFIG_PROC_FS
8644 static int io_uring_show_cred(int id, void *p, void *data)
8646 const struct cred *cred = p;
8647 struct seq_file *m = data;
8648 struct user_namespace *uns = seq_user_ns(m);
8649 struct group_info *gi;
8654 seq_printf(m, "%5d\n", id);
8655 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8656 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8657 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8658 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8659 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8660 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8661 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8662 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8663 seq_puts(m, "\n\tGroups:\t");
8664 gi = cred->group_info;
8665 for (g = 0; g < gi->ngroups; g++) {
8666 seq_put_decimal_ull(m, g ? " " : "",
8667 from_kgid_munged(uns, gi->gid[g]));
8669 seq_puts(m, "\n\tCapEff:\t");
8670 cap = cred->cap_effective;
8671 CAP_FOR_EACH_U32(__capi)
8672 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8677 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8683 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8684 * since fdinfo case grabs it in the opposite direction of normal use
8685 * cases. If we fail to get the lock, we just don't iterate any
8686 * structures that could be going away outside the io_uring mutex.
8688 has_lock = mutex_trylock(&ctx->uring_lock);
8690 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8691 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8692 struct fixed_file_table *table;
8695 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8696 f = table->files[i & IORING_FILE_TABLE_MASK];
8698 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8700 seq_printf(m, "%5u: <none>\n", i);
8702 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8703 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8704 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8706 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8707 (unsigned int) buf->len);
8709 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8710 seq_printf(m, "Personalities:\n");
8711 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8713 seq_printf(m, "PollList:\n");
8714 spin_lock_irq(&ctx->completion_lock);
8715 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8716 struct hlist_head *list = &ctx->cancel_hash[i];
8717 struct io_kiocb *req;
8719 hlist_for_each_entry(req, list, hash_node)
8720 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8721 req->task->task_works != NULL);
8723 spin_unlock_irq(&ctx->completion_lock);
8725 mutex_unlock(&ctx->uring_lock);
8728 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8730 struct io_ring_ctx *ctx = f->private_data;
8732 if (percpu_ref_tryget(&ctx->refs)) {
8733 __io_uring_show_fdinfo(ctx, m);
8734 percpu_ref_put(&ctx->refs);
8739 static const struct file_operations io_uring_fops = {
8740 .release = io_uring_release,
8741 .flush = io_uring_flush,
8742 .mmap = io_uring_mmap,
8744 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8745 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8747 .poll = io_uring_poll,
8748 .fasync = io_uring_fasync,
8749 #ifdef CONFIG_PROC_FS
8750 .show_fdinfo = io_uring_show_fdinfo,
8754 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8755 struct io_uring_params *p)
8757 struct io_rings *rings;
8758 size_t size, sq_array_offset;
8760 /* make sure these are sane, as we already accounted them */
8761 ctx->sq_entries = p->sq_entries;
8762 ctx->cq_entries = p->cq_entries;
8764 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8765 if (size == SIZE_MAX)
8768 rings = io_mem_alloc(size);
8773 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8774 rings->sq_ring_mask = p->sq_entries - 1;
8775 rings->cq_ring_mask = p->cq_entries - 1;
8776 rings->sq_ring_entries = p->sq_entries;
8777 rings->cq_ring_entries = p->cq_entries;
8778 ctx->sq_mask = rings->sq_ring_mask;
8779 ctx->cq_mask = rings->cq_ring_mask;
8781 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8782 if (size == SIZE_MAX) {
8783 io_mem_free(ctx->rings);
8788 ctx->sq_sqes = io_mem_alloc(size);
8789 if (!ctx->sq_sqes) {
8790 io_mem_free(ctx->rings);
8799 * Allocate an anonymous fd, this is what constitutes the application
8800 * visible backing of an io_uring instance. The application mmaps this
8801 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8802 * we have to tie this fd to a socket for file garbage collection purposes.
8804 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8809 #if defined(CONFIG_UNIX)
8810 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8816 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8820 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8821 O_RDWR | O_CLOEXEC);
8825 ret = PTR_ERR(file);
8829 #if defined(CONFIG_UNIX)
8830 ctx->ring_sock->file = file;
8832 if (unlikely(io_uring_add_task_file(file))) {
8833 file = ERR_PTR(-ENOMEM);
8836 fd_install(ret, file);
8839 #if defined(CONFIG_UNIX)
8840 sock_release(ctx->ring_sock);
8841 ctx->ring_sock = NULL;
8846 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8847 struct io_uring_params __user *params)
8849 struct user_struct *user = NULL;
8850 struct io_ring_ctx *ctx;
8856 if (entries > IORING_MAX_ENTRIES) {
8857 if (!(p->flags & IORING_SETUP_CLAMP))
8859 entries = IORING_MAX_ENTRIES;
8863 * Use twice as many entries for the CQ ring. It's possible for the
8864 * application to drive a higher depth than the size of the SQ ring,
8865 * since the sqes are only used at submission time. This allows for
8866 * some flexibility in overcommitting a bit. If the application has
8867 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8868 * of CQ ring entries manually.
8870 p->sq_entries = roundup_pow_of_two(entries);
8871 if (p->flags & IORING_SETUP_CQSIZE) {
8873 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8874 * to a power-of-two, if it isn't already. We do NOT impose
8875 * any cq vs sq ring sizing.
8877 if (p->cq_entries < p->sq_entries)
8879 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8880 if (!(p->flags & IORING_SETUP_CLAMP))
8882 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8884 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8886 p->cq_entries = 2 * p->sq_entries;
8889 user = get_uid(current_user());
8890 limit_mem = !capable(CAP_IPC_LOCK);
8893 ret = __io_account_mem(user,
8894 ring_pages(p->sq_entries, p->cq_entries));
8901 ctx = io_ring_ctx_alloc(p);
8904 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8909 ctx->compat = in_compat_syscall();
8911 ctx->creds = get_current_cred();
8913 ctx->sqo_task = get_task_struct(current);
8916 * This is just grabbed for accounting purposes. When a process exits,
8917 * the mm is exited and dropped before the files, hence we need to hang
8918 * on to this mm purely for the purposes of being able to unaccount
8919 * memory (locked/pinned vm). It's not used for anything else.
8921 mmgrab(current->mm);
8922 ctx->mm_account = current->mm;
8925 * Account memory _before_ installing the file descriptor. Once
8926 * the descriptor is installed, it can get closed at any time. Also
8927 * do this before hitting the general error path, as ring freeing
8928 * will un-account as well.
8930 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8932 ctx->limit_mem = limit_mem;
8934 ret = io_allocate_scq_urings(ctx, p);
8938 ret = io_sq_offload_start(ctx, p);
8942 memset(&p->sq_off, 0, sizeof(p->sq_off));
8943 p->sq_off.head = offsetof(struct io_rings, sq.head);
8944 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8945 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8946 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8947 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8948 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8949 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8951 memset(&p->cq_off, 0, sizeof(p->cq_off));
8952 p->cq_off.head = offsetof(struct io_rings, cq.head);
8953 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8954 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8955 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8956 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8957 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8958 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8960 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8961 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8962 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8963 IORING_FEAT_POLL_32BITS;
8965 if (copy_to_user(params, p, sizeof(*p))) {
8971 * Install ring fd as the very last thing, so we don't risk someone
8972 * having closed it before we finish setup
8974 ret = io_uring_get_fd(ctx);
8978 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8981 io_ring_ctx_wait_and_kill(ctx);
8986 * Sets up an aio uring context, and returns the fd. Applications asks for a
8987 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8988 * params structure passed in.
8990 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8992 struct io_uring_params p;
8995 if (copy_from_user(&p, params, sizeof(p)))
8997 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9002 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9003 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9004 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
9007 return io_uring_create(entries, &p, params);
9010 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9011 struct io_uring_params __user *, params)
9013 return io_uring_setup(entries, params);
9016 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9018 struct io_uring_probe *p;
9022 size = struct_size(p, ops, nr_args);
9023 if (size == SIZE_MAX)
9025 p = kzalloc(size, GFP_KERNEL);
9030 if (copy_from_user(p, arg, size))
9033 if (memchr_inv(p, 0, size))
9036 p->last_op = IORING_OP_LAST - 1;
9037 if (nr_args > IORING_OP_LAST)
9038 nr_args = IORING_OP_LAST;
9040 for (i = 0; i < nr_args; i++) {
9042 if (!io_op_defs[i].not_supported)
9043 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9048 if (copy_to_user(arg, p, size))
9055 static int io_register_personality(struct io_ring_ctx *ctx)
9057 const struct cred *creds = get_current_cred();
9060 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9061 USHRT_MAX, GFP_KERNEL);
9067 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9069 const struct cred *old_creds;
9071 old_creds = idr_remove(&ctx->personality_idr, id);
9073 put_cred(old_creds);
9080 static bool io_register_op_must_quiesce(int op)
9083 case IORING_UNREGISTER_FILES:
9084 case IORING_REGISTER_FILES_UPDATE:
9085 case IORING_REGISTER_PROBE:
9086 case IORING_REGISTER_PERSONALITY:
9087 case IORING_UNREGISTER_PERSONALITY:
9094 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9095 void __user *arg, unsigned nr_args)
9096 __releases(ctx->uring_lock)
9097 __acquires(ctx->uring_lock)
9102 * We're inside the ring mutex, if the ref is already dying, then
9103 * someone else killed the ctx or is already going through
9104 * io_uring_register().
9106 if (percpu_ref_is_dying(&ctx->refs))
9109 if (io_register_op_must_quiesce(opcode)) {
9110 percpu_ref_kill(&ctx->refs);
9113 * Drop uring mutex before waiting for references to exit. If
9114 * another thread is currently inside io_uring_enter() it might
9115 * need to grab the uring_lock to make progress. If we hold it
9116 * here across the drain wait, then we can deadlock. It's safe
9117 * to drop the mutex here, since no new references will come in
9118 * after we've killed the percpu ref.
9120 mutex_unlock(&ctx->uring_lock);
9121 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9122 mutex_lock(&ctx->uring_lock);
9124 percpu_ref_resurrect(&ctx->refs);
9131 case IORING_REGISTER_BUFFERS:
9132 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9134 case IORING_UNREGISTER_BUFFERS:
9138 ret = io_sqe_buffer_unregister(ctx);
9140 case IORING_REGISTER_FILES:
9141 ret = io_sqe_files_register(ctx, arg, nr_args);
9143 case IORING_UNREGISTER_FILES:
9147 ret = io_sqe_files_unregister(ctx);
9149 case IORING_REGISTER_FILES_UPDATE:
9150 ret = io_sqe_files_update(ctx, arg, nr_args);
9152 case IORING_REGISTER_EVENTFD:
9153 case IORING_REGISTER_EVENTFD_ASYNC:
9157 ret = io_eventfd_register(ctx, arg);
9160 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9161 ctx->eventfd_async = 1;
9163 ctx->eventfd_async = 0;
9165 case IORING_UNREGISTER_EVENTFD:
9169 ret = io_eventfd_unregister(ctx);
9171 case IORING_REGISTER_PROBE:
9173 if (!arg || nr_args > 256)
9175 ret = io_probe(ctx, arg, nr_args);
9177 case IORING_REGISTER_PERSONALITY:
9181 ret = io_register_personality(ctx);
9183 case IORING_UNREGISTER_PERSONALITY:
9187 ret = io_unregister_personality(ctx, nr_args);
9194 if (io_register_op_must_quiesce(opcode)) {
9195 /* bring the ctx back to life */
9196 percpu_ref_reinit(&ctx->refs);
9198 reinit_completion(&ctx->ref_comp);
9203 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9204 void __user *, arg, unsigned int, nr_args)
9206 struct io_ring_ctx *ctx;
9215 if (f.file->f_op != &io_uring_fops)
9218 ctx = f.file->private_data;
9220 mutex_lock(&ctx->uring_lock);
9221 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9222 mutex_unlock(&ctx->uring_lock);
9223 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9224 ctx->cq_ev_fd != NULL, ret);
9230 static int __init io_uring_init(void)
9232 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9233 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9234 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9237 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9238 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9239 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9240 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9241 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9242 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9243 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9244 BUILD_BUG_SQE_ELEM(8, __u64, off);
9245 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9246 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9247 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9248 BUILD_BUG_SQE_ELEM(24, __u32, len);
9249 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9250 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9251 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9252 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9253 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9254 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9255 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9256 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9257 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9258 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9259 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9260 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9261 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9262 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9263 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9264 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9265 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9266 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9267 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9269 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9270 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9271 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9274 __initcall(io_uring_init);