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)
102 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
103 IORING_REGISTER_LAST + IORING_OP_LAST)
106 u32 head ____cacheline_aligned_in_smp;
107 u32 tail ____cacheline_aligned_in_smp;
111 * This data is shared with the application through the mmap at offsets
112 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
114 * The offsets to the member fields are published through struct
115 * io_sqring_offsets when calling io_uring_setup.
119 * Head and tail offsets into the ring; the offsets need to be
120 * masked to get valid indices.
122 * The kernel controls head of the sq ring and the tail of the cq ring,
123 * and the application controls tail of the sq ring and the head of the
126 struct io_uring sq, cq;
128 * Bitmasks to apply to head and tail offsets (constant, equals
131 u32 sq_ring_mask, cq_ring_mask;
132 /* Ring sizes (constant, power of 2) */
133 u32 sq_ring_entries, cq_ring_entries;
135 * Number of invalid entries dropped by the kernel due to
136 * invalid index stored in array
138 * Written by the kernel, shouldn't be modified by the
139 * application (i.e. get number of "new events" by comparing to
142 * After a new SQ head value was read by the application this
143 * counter includes all submissions that were dropped reaching
144 * the new SQ head (and possibly more).
150 * Written by the kernel, shouldn't be modified by the
153 * The application needs a full memory barrier before checking
154 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
160 * Written by the application, shouldn't be modified by the
165 * Number of completion events lost because the queue was full;
166 * this should be avoided by the application by making sure
167 * there are not more requests pending than there is space in
168 * the completion queue.
170 * Written by the kernel, shouldn't be modified by the
171 * application (i.e. get number of "new events" by comparing to
174 * As completion events come in out of order this counter is not
175 * ordered with any other data.
179 * Ring buffer of completion events.
181 * The kernel writes completion events fresh every time they are
182 * produced, so the application is allowed to modify pending
185 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
188 struct io_mapped_ubuf {
191 struct bio_vec *bvec;
192 unsigned int nr_bvecs;
195 struct fixed_file_table {
199 struct fixed_file_ref_node {
200 struct percpu_ref refs;
201 struct list_head node;
202 struct list_head file_list;
203 struct fixed_file_data *file_data;
204 struct llist_node llist;
207 struct fixed_file_data {
208 struct fixed_file_table *table;
209 struct io_ring_ctx *ctx;
211 struct percpu_ref *cur_refs;
212 struct percpu_ref refs;
213 struct completion done;
214 struct list_head ref_list;
219 struct list_head list;
225 struct io_restriction {
226 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
227 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
228 u8 sqe_flags_allowed;
229 u8 sqe_flags_required;
235 struct percpu_ref refs;
236 } ____cacheline_aligned_in_smp;
240 unsigned int compat: 1;
241 unsigned int limit_mem: 1;
242 unsigned int cq_overflow_flushed: 1;
243 unsigned int drain_next: 1;
244 unsigned int eventfd_async: 1;
245 unsigned int restricted: 1;
248 * Ring buffer of indices into array of io_uring_sqe, which is
249 * mmapped by the application using the IORING_OFF_SQES offset.
251 * This indirection could e.g. be used to assign fixed
252 * io_uring_sqe entries to operations and only submit them to
253 * the queue when needed.
255 * The kernel modifies neither the indices array nor the entries
259 unsigned cached_sq_head;
262 unsigned sq_thread_idle;
263 unsigned cached_sq_dropped;
264 atomic_t cached_cq_overflow;
265 unsigned long sq_check_overflow;
267 struct list_head defer_list;
268 struct list_head timeout_list;
269 struct list_head cq_overflow_list;
271 wait_queue_head_t inflight_wait;
272 struct io_uring_sqe *sq_sqes;
273 } ____cacheline_aligned_in_smp;
275 struct io_rings *rings;
279 struct task_struct *sqo_thread; /* if using sq thread polling */
282 * For SQPOLL usage - we hold a reference to the parent task, so we
283 * have access to the ->files
285 struct task_struct *sqo_task;
287 /* Only used for accounting purposes */
288 struct mm_struct *mm_account;
290 struct wait_queue_head *sqo_wait;
291 struct wait_queue_head __sqo_wait;
292 struct wait_queue_entry sqo_wait_entry;
295 * If used, fixed file set. Writers must ensure that ->refs is dead,
296 * readers must ensure that ->refs is alive as long as the file* is
297 * used. Only updated through io_uring_register(2).
299 struct fixed_file_data *file_data;
300 unsigned nr_user_files;
302 /* if used, fixed mapped user buffers */
303 unsigned nr_user_bufs;
304 struct io_mapped_ubuf *user_bufs;
306 struct user_struct *user;
308 const struct cred *creds;
310 struct completion ref_comp;
311 struct completion sq_thread_comp;
313 /* if all else fails... */
314 struct io_kiocb *fallback_req;
316 #if defined(CONFIG_UNIX)
317 struct socket *ring_sock;
320 struct idr io_buffer_idr;
322 struct idr personality_idr;
325 unsigned cached_cq_tail;
328 atomic_t cq_timeouts;
329 unsigned long cq_check_overflow;
330 struct wait_queue_head cq_wait;
331 struct fasync_struct *cq_fasync;
332 struct eventfd_ctx *cq_ev_fd;
333 } ____cacheline_aligned_in_smp;
336 struct mutex uring_lock;
337 wait_queue_head_t wait;
338 } ____cacheline_aligned_in_smp;
341 spinlock_t completion_lock;
344 * ->iopoll_list is protected by the ctx->uring_lock for
345 * io_uring instances that don't use IORING_SETUP_SQPOLL.
346 * For SQPOLL, only the single threaded io_sq_thread() will
347 * manipulate the list, hence no extra locking is needed there.
349 struct list_head iopoll_list;
350 struct hlist_head *cancel_hash;
351 unsigned cancel_hash_bits;
352 bool poll_multi_file;
354 spinlock_t inflight_lock;
355 struct list_head inflight_list;
356 } ____cacheline_aligned_in_smp;
358 struct delayed_work file_put_work;
359 struct llist_head file_put_llist;
361 struct work_struct exit_work;
362 struct io_restriction restrictions;
366 * First field must be the file pointer in all the
367 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
369 struct io_poll_iocb {
372 struct wait_queue_head *head;
378 struct wait_queue_entry wait;
383 struct file *put_file;
387 struct io_timeout_data {
388 struct io_kiocb *req;
389 struct hrtimer timer;
390 struct timespec64 ts;
391 enum hrtimer_mode mode;
396 struct sockaddr __user *addr;
397 int __user *addr_len;
399 unsigned long nofile;
421 struct list_head list;
425 /* NOTE: kiocb has the file as the first member, so don't do it here */
433 struct sockaddr __user *addr;
440 struct user_msghdr __user *umsg;
446 struct io_buffer *kbuf;
452 struct filename *filename;
454 unsigned long nofile;
457 struct io_files_update {
483 struct epoll_event event;
487 struct file *file_out;
488 struct file *file_in;
495 struct io_provide_buf {
509 const char __user *filename;
510 struct statx __user *buffer;
513 struct io_completion {
515 struct list_head list;
519 struct io_async_connect {
520 struct sockaddr_storage address;
523 struct io_async_msghdr {
524 struct iovec fast_iov[UIO_FASTIOV];
526 struct sockaddr __user *uaddr;
528 struct sockaddr_storage addr;
532 struct iovec fast_iov[UIO_FASTIOV];
533 const struct iovec *free_iovec;
534 struct iov_iter iter;
536 struct wait_page_queue wpq;
539 struct io_async_ctx {
541 struct io_async_rw rw;
542 struct io_async_msghdr msg;
543 struct io_async_connect connect;
544 struct io_timeout_data timeout;
549 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
550 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
551 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
552 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
553 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
554 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
561 REQ_F_LINK_TIMEOUT_BIT,
563 REQ_F_COMP_LOCKED_BIT,
564 REQ_F_NEED_CLEANUP_BIT,
566 REQ_F_BUFFER_SELECTED_BIT,
567 REQ_F_NO_FILE_TABLE_BIT,
568 REQ_F_WORK_INITIALIZED_BIT,
570 /* not a real bit, just to check we're not overflowing the space */
576 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
577 /* drain existing IO first */
578 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
580 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
581 /* doesn't sever on completion < 0 */
582 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
584 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
585 /* IOSQE_BUFFER_SELECT */
586 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
589 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
590 /* fail rest of links */
591 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
592 /* on inflight list */
593 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
594 /* read/write uses file position */
595 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
596 /* must not punt to workers */
597 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
598 /* has linked timeout */
599 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
601 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
602 /* completion under lock */
603 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
605 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
606 /* already went through poll handler */
607 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
608 /* buffer already selected */
609 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
610 /* doesn't need file table for this request */
611 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
612 /* io_wq_work is initialized */
613 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
617 struct io_poll_iocb poll;
618 struct io_poll_iocb *double_poll;
622 * NOTE! Each of the iocb union members has the file pointer
623 * as the first entry in their struct definition. So you can
624 * access the file pointer through any of the sub-structs,
625 * or directly as just 'ki_filp' in this struct.
631 struct io_poll_iocb poll;
632 struct io_accept accept;
634 struct io_cancel cancel;
635 struct io_timeout timeout;
636 struct io_connect connect;
637 struct io_sr_msg sr_msg;
639 struct io_close close;
640 struct io_files_update files_update;
641 struct io_fadvise fadvise;
642 struct io_madvise madvise;
643 struct io_epoll epoll;
644 struct io_splice splice;
645 struct io_provide_buf pbuf;
646 struct io_statx statx;
647 /* use only after cleaning per-op data, see io_clean_op() */
648 struct io_completion compl;
651 struct io_async_ctx *io;
653 /* polled IO has completed */
659 struct io_ring_ctx *ctx;
662 struct task_struct *task;
665 struct list_head link_list;
668 * 1. used with ctx->iopoll_list with reads/writes
669 * 2. to track reqs with ->files (see io_op_def::file_table)
671 struct list_head inflight_entry;
673 struct percpu_ref *fixed_file_refs;
674 struct callback_head task_work;
675 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
676 struct hlist_node hash_node;
677 struct async_poll *apoll;
678 struct io_wq_work work;
681 struct io_defer_entry {
682 struct list_head list;
683 struct io_kiocb *req;
687 #define IO_IOPOLL_BATCH 8
689 struct io_comp_state {
691 struct list_head list;
692 struct io_ring_ctx *ctx;
695 struct io_submit_state {
696 struct blk_plug plug;
699 * io_kiocb alloc cache
701 void *reqs[IO_IOPOLL_BATCH];
702 unsigned int free_reqs;
705 * Batch completion logic
707 struct io_comp_state comp;
710 * File reference cache
714 unsigned int has_refs;
715 unsigned int ios_left;
719 /* needs req->io allocated for deferral/async */
720 unsigned async_ctx : 1;
721 /* needs current->mm setup, does mm access */
722 unsigned needs_mm : 1;
723 /* needs req->file assigned */
724 unsigned needs_file : 1;
725 /* don't fail if file grab fails */
726 unsigned needs_file_no_error : 1;
727 /* hash wq insertion if file is a regular file */
728 unsigned hash_reg_file : 1;
729 /* unbound wq insertion if file is a non-regular file */
730 unsigned unbound_nonreg_file : 1;
731 /* opcode is not supported by this kernel */
732 unsigned not_supported : 1;
733 /* needs file table */
734 unsigned file_table : 1;
736 unsigned needs_fs : 1;
737 /* set if opcode supports polled "wait" */
739 unsigned pollout : 1;
740 /* op supports buffer selection */
741 unsigned buffer_select : 1;
742 unsigned needs_fsize : 1;
745 static const struct io_op_def io_op_defs[] = {
746 [IORING_OP_NOP] = {},
747 [IORING_OP_READV] = {
751 .unbound_nonreg_file = 1,
755 [IORING_OP_WRITEV] = {
760 .unbound_nonreg_file = 1,
764 [IORING_OP_FSYNC] = {
767 [IORING_OP_READ_FIXED] = {
769 .unbound_nonreg_file = 1,
772 [IORING_OP_WRITE_FIXED] = {
775 .unbound_nonreg_file = 1,
779 [IORING_OP_POLL_ADD] = {
781 .unbound_nonreg_file = 1,
783 [IORING_OP_POLL_REMOVE] = {},
784 [IORING_OP_SYNC_FILE_RANGE] = {
787 [IORING_OP_SENDMSG] = {
791 .unbound_nonreg_file = 1,
795 [IORING_OP_RECVMSG] = {
799 .unbound_nonreg_file = 1,
804 [IORING_OP_TIMEOUT] = {
808 [IORING_OP_TIMEOUT_REMOVE] = {},
809 [IORING_OP_ACCEPT] = {
812 .unbound_nonreg_file = 1,
816 [IORING_OP_ASYNC_CANCEL] = {},
817 [IORING_OP_LINK_TIMEOUT] = {
821 [IORING_OP_CONNECT] = {
825 .unbound_nonreg_file = 1,
828 [IORING_OP_FALLOCATE] = {
832 [IORING_OP_OPENAT] = {
836 [IORING_OP_CLOSE] = {
838 .needs_file_no_error = 1,
841 [IORING_OP_FILES_UPDATE] = {
845 [IORING_OP_STATX] = {
853 .unbound_nonreg_file = 1,
857 [IORING_OP_WRITE] = {
860 .unbound_nonreg_file = 1,
864 [IORING_OP_FADVISE] = {
867 [IORING_OP_MADVISE] = {
873 .unbound_nonreg_file = 1,
879 .unbound_nonreg_file = 1,
883 [IORING_OP_OPENAT2] = {
887 [IORING_OP_EPOLL_CTL] = {
888 .unbound_nonreg_file = 1,
891 [IORING_OP_SPLICE] = {
894 .unbound_nonreg_file = 1,
896 [IORING_OP_PROVIDE_BUFFERS] = {},
897 [IORING_OP_REMOVE_BUFFERS] = {},
901 .unbound_nonreg_file = 1,
905 enum io_mem_account {
910 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
911 struct io_comp_state *cs);
912 static void io_cqring_fill_event(struct io_kiocb *req, long res);
913 static void io_put_req(struct io_kiocb *req);
914 static void io_double_put_req(struct io_kiocb *req);
915 static void __io_double_put_req(struct io_kiocb *req);
916 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
917 static void __io_queue_linked_timeout(struct io_kiocb *req);
918 static void io_queue_linked_timeout(struct io_kiocb *req);
919 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
920 struct io_uring_files_update *ip,
922 static int io_prep_work_files(struct io_kiocb *req);
923 static void __io_clean_op(struct io_kiocb *req);
924 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
925 int fd, struct file **out_file, bool fixed);
926 static void __io_queue_sqe(struct io_kiocb *req,
927 const struct io_uring_sqe *sqe,
928 struct io_comp_state *cs);
929 static void io_file_put_work(struct work_struct *work);
931 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
932 struct iovec **iovec, struct iov_iter *iter,
934 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
935 const struct iovec *fast_iov,
936 struct iov_iter *iter, bool force);
938 static struct kmem_cache *req_cachep;
940 static const struct file_operations io_uring_fops;
942 struct sock *io_uring_get_socket(struct file *file)
944 #if defined(CONFIG_UNIX)
945 if (file->f_op == &io_uring_fops) {
946 struct io_ring_ctx *ctx = file->private_data;
948 return ctx->ring_sock->sk;
953 EXPORT_SYMBOL(io_uring_get_socket);
955 static inline void io_clean_op(struct io_kiocb *req)
957 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
962 static void io_sq_thread_drop_mm(void)
964 struct mm_struct *mm = current->mm;
967 kthread_unuse_mm(mm);
972 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
975 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
976 !ctx->sqo_task->mm ||
977 !mmget_not_zero(ctx->sqo_task->mm)))
979 kthread_use_mm(ctx->sqo_task->mm);
985 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
986 struct io_kiocb *req)
988 if (!io_op_defs[req->opcode].needs_mm)
990 return __io_sq_thread_acquire_mm(ctx);
993 static inline void req_set_fail_links(struct io_kiocb *req)
995 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
996 req->flags |= REQ_F_FAIL_LINK;
1000 * Note: must call io_req_init_async() for the first time you
1001 * touch any members of io_wq_work.
1003 static inline void io_req_init_async(struct io_kiocb *req)
1005 if (req->flags & REQ_F_WORK_INITIALIZED)
1008 memset(&req->work, 0, sizeof(req->work));
1009 req->flags |= REQ_F_WORK_INITIALIZED;
1012 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1014 return ctx->flags & IORING_SETUP_SQPOLL;
1017 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1019 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1021 complete(&ctx->ref_comp);
1024 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1026 return !req->timeout.off;
1029 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1031 struct io_ring_ctx *ctx;
1034 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1038 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1039 if (!ctx->fallback_req)
1043 * Use 5 bits less than the max cq entries, that should give us around
1044 * 32 entries per hash list if totally full and uniformly spread.
1046 hash_bits = ilog2(p->cq_entries);
1050 ctx->cancel_hash_bits = hash_bits;
1051 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1053 if (!ctx->cancel_hash)
1055 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1057 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1058 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1061 ctx->flags = p->flags;
1062 init_waitqueue_head(&ctx->__sqo_wait);
1063 ctx->sqo_wait = &ctx->__sqo_wait;
1064 init_waitqueue_head(&ctx->cq_wait);
1065 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1066 init_completion(&ctx->ref_comp);
1067 init_completion(&ctx->sq_thread_comp);
1068 idr_init(&ctx->io_buffer_idr);
1069 idr_init(&ctx->personality_idr);
1070 mutex_init(&ctx->uring_lock);
1071 init_waitqueue_head(&ctx->wait);
1072 spin_lock_init(&ctx->completion_lock);
1073 INIT_LIST_HEAD(&ctx->iopoll_list);
1074 INIT_LIST_HEAD(&ctx->defer_list);
1075 INIT_LIST_HEAD(&ctx->timeout_list);
1076 init_waitqueue_head(&ctx->inflight_wait);
1077 spin_lock_init(&ctx->inflight_lock);
1078 INIT_LIST_HEAD(&ctx->inflight_list);
1079 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1080 init_llist_head(&ctx->file_put_llist);
1083 if (ctx->fallback_req)
1084 kmem_cache_free(req_cachep, ctx->fallback_req);
1085 kfree(ctx->cancel_hash);
1090 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1092 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1093 struct io_ring_ctx *ctx = req->ctx;
1095 return seq != ctx->cached_cq_tail
1096 + atomic_read(&ctx->cached_cq_overflow);
1102 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1104 struct io_rings *rings = ctx->rings;
1106 /* order cqe stores with ring update */
1107 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1109 if (wq_has_sleeper(&ctx->cq_wait)) {
1110 wake_up_interruptible(&ctx->cq_wait);
1111 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1116 * Returns true if we need to defer file table putting. This can only happen
1117 * from the error path with REQ_F_COMP_LOCKED set.
1119 static bool io_req_clean_work(struct io_kiocb *req)
1121 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1124 req->flags &= ~REQ_F_WORK_INITIALIZED;
1127 mmdrop(req->work.mm);
1128 req->work.mm = NULL;
1130 if (req->work.creds) {
1131 put_cred(req->work.creds);
1132 req->work.creds = NULL;
1135 struct fs_struct *fs = req->work.fs;
1137 if (req->flags & REQ_F_COMP_LOCKED)
1140 spin_lock(&req->work.fs->lock);
1143 spin_unlock(&req->work.fs->lock);
1146 req->work.fs = NULL;
1152 static void io_prep_async_work(struct io_kiocb *req)
1154 const struct io_op_def *def = &io_op_defs[req->opcode];
1156 io_req_init_async(req);
1158 if (req->flags & REQ_F_ISREG) {
1159 if (def->hash_reg_file || (req->ctx->flags & IORING_SETUP_IOPOLL))
1160 io_wq_hash_work(&req->work, file_inode(req->file));
1162 if (def->unbound_nonreg_file)
1163 req->work.flags |= IO_WQ_WORK_UNBOUND;
1165 if (!req->work.mm && def->needs_mm) {
1166 mmgrab(current->mm);
1167 req->work.mm = current->mm;
1169 if (!req->work.creds)
1170 req->work.creds = get_current_cred();
1171 if (!req->work.fs && def->needs_fs) {
1172 spin_lock(¤t->fs->lock);
1173 if (!current->fs->in_exec) {
1174 req->work.fs = current->fs;
1175 req->work.fs->users++;
1177 req->work.flags |= IO_WQ_WORK_CANCEL;
1179 spin_unlock(¤t->fs->lock);
1181 if (def->needs_fsize)
1182 req->work.fsize = rlimit(RLIMIT_FSIZE);
1184 req->work.fsize = RLIM_INFINITY;
1187 static void io_prep_async_link(struct io_kiocb *req)
1189 struct io_kiocb *cur;
1191 io_prep_async_work(req);
1192 if (req->flags & REQ_F_LINK_HEAD)
1193 list_for_each_entry(cur, &req->link_list, link_list)
1194 io_prep_async_work(cur);
1197 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1199 struct io_ring_ctx *ctx = req->ctx;
1200 struct io_kiocb *link = io_prep_linked_timeout(req);
1202 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1203 &req->work, req->flags);
1204 io_wq_enqueue(ctx->io_wq, &req->work);
1208 static void io_queue_async_work(struct io_kiocb *req)
1210 struct io_kiocb *link;
1212 /* init ->work of the whole link before punting */
1213 io_prep_async_link(req);
1214 link = __io_queue_async_work(req);
1217 io_queue_linked_timeout(link);
1220 static void io_kill_timeout(struct io_kiocb *req)
1224 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1226 atomic_set(&req->ctx->cq_timeouts,
1227 atomic_read(&req->ctx->cq_timeouts) + 1);
1228 list_del_init(&req->timeout.list);
1229 req->flags |= REQ_F_COMP_LOCKED;
1230 io_cqring_fill_event(req, 0);
1235 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1237 struct io_ring_ctx *ctx = req->ctx;
1239 if (!tsk || req->task == tsk)
1241 if ((ctx->flags & IORING_SETUP_SQPOLL) && req->task == ctx->sqo_thread)
1247 * Returns true if we found and killed one or more timeouts
1249 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1251 struct io_kiocb *req, *tmp;
1254 spin_lock_irq(&ctx->completion_lock);
1255 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1256 if (io_task_match(req, tsk)) {
1257 io_kill_timeout(req);
1261 spin_unlock_irq(&ctx->completion_lock);
1262 return canceled != 0;
1265 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1268 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1269 struct io_defer_entry, list);
1270 struct io_kiocb *link;
1272 if (req_need_defer(de->req, de->seq))
1274 list_del_init(&de->list);
1275 /* punt-init is done before queueing for defer */
1276 link = __io_queue_async_work(de->req);
1278 __io_queue_linked_timeout(link);
1279 /* drop submission reference */
1280 link->flags |= REQ_F_COMP_LOCKED;
1284 } while (!list_empty(&ctx->defer_list));
1287 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1289 while (!list_empty(&ctx->timeout_list)) {
1290 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1291 struct io_kiocb, timeout.list);
1293 if (io_is_timeout_noseq(req))
1295 if (req->timeout.target_seq != ctx->cached_cq_tail
1296 - atomic_read(&ctx->cq_timeouts))
1299 list_del_init(&req->timeout.list);
1300 io_kill_timeout(req);
1304 static void io_commit_cqring(struct io_ring_ctx *ctx)
1306 io_flush_timeouts(ctx);
1307 __io_commit_cqring(ctx);
1309 if (unlikely(!list_empty(&ctx->defer_list)))
1310 __io_queue_deferred(ctx);
1313 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1315 struct io_rings *rings = ctx->rings;
1318 tail = ctx->cached_cq_tail;
1320 * writes to the cq entry need to come after reading head; the
1321 * control dependency is enough as we're using WRITE_ONCE to
1324 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1327 ctx->cached_cq_tail++;
1328 return &rings->cqes[tail & ctx->cq_mask];
1331 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1335 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1337 if (!ctx->eventfd_async)
1339 return io_wq_current_is_worker();
1342 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1344 if (waitqueue_active(&ctx->wait))
1345 wake_up(&ctx->wait);
1346 if (waitqueue_active(ctx->sqo_wait))
1347 wake_up(ctx->sqo_wait);
1348 if (io_should_trigger_evfd(ctx))
1349 eventfd_signal(ctx->cq_ev_fd, 1);
1352 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1354 if (list_empty(&ctx->cq_overflow_list)) {
1355 clear_bit(0, &ctx->sq_check_overflow);
1356 clear_bit(0, &ctx->cq_check_overflow);
1357 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1361 static inline bool io_match_files(struct io_kiocb *req,
1362 struct files_struct *files)
1366 if (req->flags & REQ_F_WORK_INITIALIZED)
1367 return req->work.files == files;
1371 /* Returns true if there are no backlogged entries after the flush */
1372 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1373 struct task_struct *tsk,
1374 struct files_struct *files)
1376 struct io_rings *rings = ctx->rings;
1377 struct io_kiocb *req, *tmp;
1378 struct io_uring_cqe *cqe;
1379 unsigned long flags;
1383 if (list_empty_careful(&ctx->cq_overflow_list))
1385 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1386 rings->cq_ring_entries))
1390 spin_lock_irqsave(&ctx->completion_lock, flags);
1392 /* if force is set, the ring is going away. always drop after that */
1394 ctx->cq_overflow_flushed = 1;
1397 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1398 if (tsk && req->task != tsk)
1400 if (!io_match_files(req, files))
1403 cqe = io_get_cqring(ctx);
1407 list_move(&req->compl.list, &list);
1409 WRITE_ONCE(cqe->user_data, req->user_data);
1410 WRITE_ONCE(cqe->res, req->result);
1411 WRITE_ONCE(cqe->flags, req->compl.cflags);
1413 WRITE_ONCE(ctx->rings->cq_overflow,
1414 atomic_inc_return(&ctx->cached_cq_overflow));
1418 io_commit_cqring(ctx);
1419 io_cqring_mark_overflow(ctx);
1421 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1422 io_cqring_ev_posted(ctx);
1424 while (!list_empty(&list)) {
1425 req = list_first_entry(&list, struct io_kiocb, compl.list);
1426 list_del(&req->compl.list);
1433 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1435 struct io_ring_ctx *ctx = req->ctx;
1436 struct io_uring_cqe *cqe;
1438 trace_io_uring_complete(ctx, req->user_data, res);
1441 * If we can't get a cq entry, userspace overflowed the
1442 * submission (by quite a lot). Increment the overflow count in
1445 cqe = io_get_cqring(ctx);
1447 WRITE_ONCE(cqe->user_data, req->user_data);
1448 WRITE_ONCE(cqe->res, res);
1449 WRITE_ONCE(cqe->flags, cflags);
1450 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1452 * If we're in ring overflow flush mode, or in task cancel mode,
1453 * then we cannot store the request for later flushing, we need
1454 * to drop it on the floor.
1456 WRITE_ONCE(ctx->rings->cq_overflow,
1457 atomic_inc_return(&ctx->cached_cq_overflow));
1459 if (list_empty(&ctx->cq_overflow_list)) {
1460 set_bit(0, &ctx->sq_check_overflow);
1461 set_bit(0, &ctx->cq_check_overflow);
1462 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1466 req->compl.cflags = cflags;
1467 refcount_inc(&req->refs);
1468 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1472 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1474 __io_cqring_fill_event(req, res, 0);
1477 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1479 struct io_ring_ctx *ctx = req->ctx;
1480 unsigned long flags;
1482 spin_lock_irqsave(&ctx->completion_lock, flags);
1483 __io_cqring_fill_event(req, res, cflags);
1484 io_commit_cqring(ctx);
1485 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1487 io_cqring_ev_posted(ctx);
1490 static void io_submit_flush_completions(struct io_comp_state *cs)
1492 struct io_ring_ctx *ctx = cs->ctx;
1494 spin_lock_irq(&ctx->completion_lock);
1495 while (!list_empty(&cs->list)) {
1496 struct io_kiocb *req;
1498 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1499 list_del(&req->compl.list);
1500 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1501 if (!(req->flags & REQ_F_LINK_HEAD)) {
1502 req->flags |= REQ_F_COMP_LOCKED;
1505 spin_unlock_irq(&ctx->completion_lock);
1507 spin_lock_irq(&ctx->completion_lock);
1510 io_commit_cqring(ctx);
1511 spin_unlock_irq(&ctx->completion_lock);
1513 io_cqring_ev_posted(ctx);
1517 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1518 struct io_comp_state *cs)
1521 io_cqring_add_event(req, res, cflags);
1526 req->compl.cflags = cflags;
1527 list_add_tail(&req->compl.list, &cs->list);
1529 io_submit_flush_completions(cs);
1533 static void io_req_complete(struct io_kiocb *req, long res)
1535 __io_req_complete(req, res, 0, NULL);
1538 static inline bool io_is_fallback_req(struct io_kiocb *req)
1540 return req == (struct io_kiocb *)
1541 ((unsigned long) req->ctx->fallback_req & ~1UL);
1544 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1546 struct io_kiocb *req;
1548 req = ctx->fallback_req;
1549 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1555 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1556 struct io_submit_state *state)
1558 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1559 struct io_kiocb *req;
1561 if (!state->free_reqs) {
1565 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1566 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1569 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1570 * retry single alloc to be on the safe side.
1572 if (unlikely(ret <= 0)) {
1573 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1574 if (!state->reqs[0])
1578 state->free_reqs = ret - 1;
1579 req = state->reqs[ret - 1];
1582 req = state->reqs[state->free_reqs];
1587 return io_get_fallback_req(ctx);
1590 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1594 percpu_ref_put(req->fixed_file_refs);
1599 static bool io_dismantle_req(struct io_kiocb *req)
1606 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1608 return io_req_clean_work(req);
1611 static void __io_free_req_finish(struct io_kiocb *req)
1613 struct io_uring_task *tctx = req->task->io_uring;
1614 struct io_ring_ctx *ctx = req->ctx;
1616 atomic_long_inc(&tctx->req_complete);
1618 wake_up(&tctx->wait);
1619 put_task_struct(req->task);
1621 if (likely(!io_is_fallback_req(req)))
1622 kmem_cache_free(req_cachep, req);
1624 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1625 percpu_ref_put(&ctx->refs);
1628 static void io_req_task_file_table_put(struct callback_head *cb)
1630 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1631 struct fs_struct *fs = req->work.fs;
1633 spin_lock(&req->work.fs->lock);
1636 spin_unlock(&req->work.fs->lock);
1639 req->work.fs = NULL;
1640 __io_free_req_finish(req);
1643 static void __io_free_req(struct io_kiocb *req)
1645 if (!io_dismantle_req(req)) {
1646 __io_free_req_finish(req);
1650 init_task_work(&req->task_work, io_req_task_file_table_put);
1651 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1652 if (unlikely(ret)) {
1653 struct task_struct *tsk;
1655 tsk = io_wq_get_task(req->ctx->io_wq);
1656 task_work_add(tsk, &req->task_work, 0);
1661 static bool io_link_cancel_timeout(struct io_kiocb *req)
1663 struct io_ring_ctx *ctx = req->ctx;
1666 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1668 io_cqring_fill_event(req, -ECANCELED);
1669 io_commit_cqring(ctx);
1670 req->flags &= ~REQ_F_LINK_HEAD;
1678 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1680 struct io_kiocb *link;
1683 if (list_empty(&req->link_list))
1685 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1686 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1689 list_del_init(&link->link_list);
1690 link->flags |= REQ_F_COMP_LOCKED;
1691 wake_ev = io_link_cancel_timeout(link);
1692 req->flags &= ~REQ_F_LINK_TIMEOUT;
1696 static void io_kill_linked_timeout(struct io_kiocb *req)
1698 struct io_ring_ctx *ctx = req->ctx;
1701 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1702 unsigned long flags;
1704 spin_lock_irqsave(&ctx->completion_lock, flags);
1705 wake_ev = __io_kill_linked_timeout(req);
1706 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1708 wake_ev = __io_kill_linked_timeout(req);
1712 io_cqring_ev_posted(ctx);
1715 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1717 struct io_kiocb *nxt;
1720 * The list should never be empty when we are called here. But could
1721 * potentially happen if the chain is messed up, check to be on the
1724 if (unlikely(list_empty(&req->link_list)))
1727 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1728 list_del_init(&req->link_list);
1729 if (!list_empty(&nxt->link_list))
1730 nxt->flags |= REQ_F_LINK_HEAD;
1735 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1737 static void __io_fail_links(struct io_kiocb *req)
1739 struct io_ring_ctx *ctx = req->ctx;
1741 while (!list_empty(&req->link_list)) {
1742 struct io_kiocb *link = list_first_entry(&req->link_list,
1743 struct io_kiocb, link_list);
1745 list_del_init(&link->link_list);
1746 trace_io_uring_fail_link(req, link);
1748 io_cqring_fill_event(link, -ECANCELED);
1749 link->flags |= REQ_F_COMP_LOCKED;
1750 __io_double_put_req(link);
1751 req->flags &= ~REQ_F_LINK_TIMEOUT;
1754 io_commit_cqring(ctx);
1755 io_cqring_ev_posted(ctx);
1758 static void io_fail_links(struct io_kiocb *req)
1760 struct io_ring_ctx *ctx = req->ctx;
1762 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1763 unsigned long flags;
1765 spin_lock_irqsave(&ctx->completion_lock, flags);
1766 __io_fail_links(req);
1767 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1769 __io_fail_links(req);
1772 io_cqring_ev_posted(ctx);
1775 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1777 req->flags &= ~REQ_F_LINK_HEAD;
1778 if (req->flags & REQ_F_LINK_TIMEOUT)
1779 io_kill_linked_timeout(req);
1782 * If LINK is set, we have dependent requests in this chain. If we
1783 * didn't fail this request, queue the first one up, moving any other
1784 * dependencies to the next request. In case of failure, fail the rest
1787 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1788 return io_req_link_next(req);
1793 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1795 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1797 return __io_req_find_next(req);
1800 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb,
1803 struct task_struct *tsk = req->task;
1804 struct io_ring_ctx *ctx = req->ctx;
1807 if (tsk->flags & PF_EXITING)
1811 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1812 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1813 * processing task_work. There's no reliable way to tell if TWA_RESUME
1817 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1818 notify = TWA_SIGNAL;
1820 ret = task_work_add(tsk, cb, notify);
1822 wake_up_process(tsk);
1827 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1829 struct io_ring_ctx *ctx = req->ctx;
1831 spin_lock_irq(&ctx->completion_lock);
1832 io_cqring_fill_event(req, error);
1833 io_commit_cqring(ctx);
1834 spin_unlock_irq(&ctx->completion_lock);
1836 io_cqring_ev_posted(ctx);
1837 req_set_fail_links(req);
1838 io_double_put_req(req);
1841 static void io_req_task_cancel(struct callback_head *cb)
1843 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1844 struct io_ring_ctx *ctx = req->ctx;
1846 __io_req_task_cancel(req, -ECANCELED);
1847 percpu_ref_put(&ctx->refs);
1850 static void __io_req_task_submit(struct io_kiocb *req)
1852 struct io_ring_ctx *ctx = req->ctx;
1854 if (!__io_sq_thread_acquire_mm(ctx)) {
1855 mutex_lock(&ctx->uring_lock);
1856 __io_queue_sqe(req, NULL, NULL);
1857 mutex_unlock(&ctx->uring_lock);
1859 __io_req_task_cancel(req, -EFAULT);
1863 static void io_req_task_submit(struct callback_head *cb)
1865 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1866 struct io_ring_ctx *ctx = req->ctx;
1868 __io_req_task_submit(req);
1869 percpu_ref_put(&ctx->refs);
1872 static void io_req_task_queue(struct io_kiocb *req)
1876 init_task_work(&req->task_work, io_req_task_submit);
1877 percpu_ref_get(&req->ctx->refs);
1879 ret = io_req_task_work_add(req, &req->task_work, true);
1880 if (unlikely(ret)) {
1881 struct task_struct *tsk;
1883 init_task_work(&req->task_work, io_req_task_cancel);
1884 tsk = io_wq_get_task(req->ctx->io_wq);
1885 task_work_add(tsk, &req->task_work, 0);
1886 wake_up_process(tsk);
1890 static void io_queue_next(struct io_kiocb *req)
1892 struct io_kiocb *nxt = io_req_find_next(req);
1895 io_req_task_queue(nxt);
1898 static void io_free_req(struct io_kiocb *req)
1905 void *reqs[IO_IOPOLL_BATCH];
1908 struct task_struct *task;
1912 static inline void io_init_req_batch(struct req_batch *rb)
1919 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1920 struct req_batch *rb)
1922 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1923 percpu_ref_put_many(&ctx->refs, rb->to_free);
1927 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1928 struct req_batch *rb)
1931 __io_req_free_batch_flush(ctx, rb);
1933 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
1934 put_task_struct_many(rb->task, rb->task_refs);
1939 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1941 if (unlikely(io_is_fallback_req(req))) {
1945 if (req->flags & REQ_F_LINK_HEAD)
1948 if (req->task != rb->task) {
1950 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
1951 put_task_struct_many(rb->task, rb->task_refs);
1953 rb->task = req->task;
1958 WARN_ON_ONCE(io_dismantle_req(req));
1959 rb->reqs[rb->to_free++] = req;
1960 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1961 __io_req_free_batch_flush(req->ctx, rb);
1965 * Drop reference to request, return next in chain (if there is one) if this
1966 * was the last reference to this request.
1968 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1970 struct io_kiocb *nxt = NULL;
1972 if (refcount_dec_and_test(&req->refs)) {
1973 nxt = io_req_find_next(req);
1979 static void io_put_req(struct io_kiocb *req)
1981 if (refcount_dec_and_test(&req->refs))
1985 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1987 struct io_kiocb *nxt;
1990 * A ref is owned by io-wq in which context we're. So, if that's the
1991 * last one, it's safe to steal next work. False negatives are Ok,
1992 * it just will be re-punted async in io_put_work()
1994 if (refcount_read(&req->refs) != 1)
1997 nxt = io_req_find_next(req);
1998 return nxt ? &nxt->work : NULL;
2002 * Must only be used if we don't need to care about links, usually from
2003 * within the completion handling itself.
2005 static void __io_double_put_req(struct io_kiocb *req)
2007 /* drop both submit and complete references */
2008 if (refcount_sub_and_test(2, &req->refs))
2012 static void io_double_put_req(struct io_kiocb *req)
2014 /* drop both submit and complete references */
2015 if (refcount_sub_and_test(2, &req->refs))
2019 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2021 struct io_rings *rings = ctx->rings;
2023 if (test_bit(0, &ctx->cq_check_overflow)) {
2025 * noflush == true is from the waitqueue handler, just ensure
2026 * we wake up the task, and the next invocation will flush the
2027 * entries. We cannot safely to it from here.
2029 if (noflush && !list_empty(&ctx->cq_overflow_list))
2032 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2035 /* See comment at the top of this file */
2037 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2040 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2042 struct io_rings *rings = ctx->rings;
2044 /* make sure SQ entry isn't read before tail */
2045 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2048 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2050 unsigned int cflags;
2052 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2053 cflags |= IORING_CQE_F_BUFFER;
2054 req->flags &= ~REQ_F_BUFFER_SELECTED;
2059 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2061 struct io_buffer *kbuf;
2063 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2064 return io_put_kbuf(req, kbuf);
2067 static inline bool io_run_task_work(void)
2070 * Not safe to run on exiting task, and the task_work handling will
2071 * not add work to such a task.
2073 if (unlikely(current->flags & PF_EXITING))
2075 if (current->task_works) {
2076 __set_current_state(TASK_RUNNING);
2084 static void io_iopoll_queue(struct list_head *again)
2086 struct io_kiocb *req;
2089 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2090 list_del(&req->inflight_entry);
2091 __io_complete_rw(req, -EAGAIN, 0, NULL);
2092 } while (!list_empty(again));
2096 * Find and free completed poll iocbs
2098 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2099 struct list_head *done)
2101 struct req_batch rb;
2102 struct io_kiocb *req;
2105 /* order with ->result store in io_complete_rw_iopoll() */
2108 io_init_req_batch(&rb);
2109 while (!list_empty(done)) {
2112 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2113 if (READ_ONCE(req->result) == -EAGAIN) {
2115 req->iopoll_completed = 0;
2116 list_move_tail(&req->inflight_entry, &again);
2119 list_del(&req->inflight_entry);
2121 if (req->flags & REQ_F_BUFFER_SELECTED)
2122 cflags = io_put_rw_kbuf(req);
2124 __io_cqring_fill_event(req, req->result, cflags);
2127 if (refcount_dec_and_test(&req->refs))
2128 io_req_free_batch(&rb, req);
2131 io_commit_cqring(ctx);
2132 if (ctx->flags & IORING_SETUP_SQPOLL)
2133 io_cqring_ev_posted(ctx);
2134 io_req_free_batch_finish(ctx, &rb);
2136 if (!list_empty(&again))
2137 io_iopoll_queue(&again);
2140 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2143 struct io_kiocb *req, *tmp;
2149 * Only spin for completions if we don't have multiple devices hanging
2150 * off our complete list, and we're under the requested amount.
2152 spin = !ctx->poll_multi_file && *nr_events < min;
2155 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2156 struct kiocb *kiocb = &req->rw.kiocb;
2159 * Move completed and retryable entries to our local lists.
2160 * If we find a request that requires polling, break out
2161 * and complete those lists first, if we have entries there.
2163 if (READ_ONCE(req->iopoll_completed)) {
2164 list_move_tail(&req->inflight_entry, &done);
2167 if (!list_empty(&done))
2170 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2174 /* iopoll may have completed current req */
2175 if (READ_ONCE(req->iopoll_completed))
2176 list_move_tail(&req->inflight_entry, &done);
2183 if (!list_empty(&done))
2184 io_iopoll_complete(ctx, nr_events, &done);
2190 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2191 * non-spinning poll check - we'll still enter the driver poll loop, but only
2192 * as a non-spinning completion check.
2194 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2197 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2200 ret = io_do_iopoll(ctx, nr_events, min);
2203 if (*nr_events >= min)
2211 * We can't just wait for polled events to come to us, we have to actively
2212 * find and complete them.
2214 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2216 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2219 mutex_lock(&ctx->uring_lock);
2220 while (!list_empty(&ctx->iopoll_list)) {
2221 unsigned int nr_events = 0;
2223 io_do_iopoll(ctx, &nr_events, 0);
2225 /* let it sleep and repeat later if can't complete a request */
2229 * Ensure we allow local-to-the-cpu processing to take place,
2230 * in this case we need to ensure that we reap all events.
2231 * Also let task_work, etc. to progress by releasing the mutex
2233 if (need_resched()) {
2234 mutex_unlock(&ctx->uring_lock);
2236 mutex_lock(&ctx->uring_lock);
2239 mutex_unlock(&ctx->uring_lock);
2242 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2244 unsigned int nr_events = 0;
2245 int iters = 0, ret = 0;
2248 * We disallow the app entering submit/complete with polling, but we
2249 * still need to lock the ring to prevent racing with polled issue
2250 * that got punted to a workqueue.
2252 mutex_lock(&ctx->uring_lock);
2255 * Don't enter poll loop if we already have events pending.
2256 * If we do, we can potentially be spinning for commands that
2257 * already triggered a CQE (eg in error).
2259 if (io_cqring_events(ctx, false))
2263 * If a submit got punted to a workqueue, we can have the
2264 * application entering polling for a command before it gets
2265 * issued. That app will hold the uring_lock for the duration
2266 * of the poll right here, so we need to take a breather every
2267 * now and then to ensure that the issue has a chance to add
2268 * the poll to the issued list. Otherwise we can spin here
2269 * forever, while the workqueue is stuck trying to acquire the
2272 if (!(++iters & 7)) {
2273 mutex_unlock(&ctx->uring_lock);
2275 mutex_lock(&ctx->uring_lock);
2278 ret = io_iopoll_getevents(ctx, &nr_events, min);
2282 } while (min && !nr_events && !need_resched());
2284 mutex_unlock(&ctx->uring_lock);
2288 static void kiocb_end_write(struct io_kiocb *req)
2291 * Tell lockdep we inherited freeze protection from submission
2294 if (req->flags & REQ_F_ISREG) {
2295 struct inode *inode = file_inode(req->file);
2297 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2299 file_end_write(req->file);
2302 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2303 struct io_comp_state *cs)
2305 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2308 if (kiocb->ki_flags & IOCB_WRITE)
2309 kiocb_end_write(req);
2311 if (res != req->result)
2312 req_set_fail_links(req);
2313 if (req->flags & REQ_F_BUFFER_SELECTED)
2314 cflags = io_put_rw_kbuf(req);
2315 __io_req_complete(req, res, cflags, cs);
2319 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2321 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2322 ssize_t ret = -ECANCELED;
2323 struct iov_iter iter;
2331 switch (req->opcode) {
2332 case IORING_OP_READV:
2333 case IORING_OP_READ_FIXED:
2334 case IORING_OP_READ:
2337 case IORING_OP_WRITEV:
2338 case IORING_OP_WRITE_FIXED:
2339 case IORING_OP_WRITE:
2343 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2349 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2352 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2360 req_set_fail_links(req);
2361 io_req_complete(req, ret);
2366 static bool io_rw_reissue(struct io_kiocb *req, long res)
2369 umode_t mode = file_inode(req->file)->i_mode;
2372 if (!S_ISBLK(mode) && !S_ISREG(mode))
2374 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2377 ret = io_sq_thread_acquire_mm(req->ctx, req);
2379 if (io_resubmit_prep(req, ret)) {
2380 refcount_inc(&req->refs);
2381 io_queue_async_work(req);
2389 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2390 struct io_comp_state *cs)
2392 if (!io_rw_reissue(req, res))
2393 io_complete_rw_common(&req->rw.kiocb, res, cs);
2396 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2398 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2400 __io_complete_rw(req, res, res2, NULL);
2403 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2405 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2407 if (kiocb->ki_flags & IOCB_WRITE)
2408 kiocb_end_write(req);
2410 if (res != -EAGAIN && res != req->result)
2411 req_set_fail_links(req);
2413 WRITE_ONCE(req->result, res);
2414 /* order with io_poll_complete() checking ->result */
2416 WRITE_ONCE(req->iopoll_completed, 1);
2420 * After the iocb has been issued, it's safe to be found on the poll list.
2421 * Adding the kiocb to the list AFTER submission ensures that we don't
2422 * find it from a io_iopoll_getevents() thread before the issuer is done
2423 * accessing the kiocb cookie.
2425 static void io_iopoll_req_issued(struct io_kiocb *req)
2427 struct io_ring_ctx *ctx = req->ctx;
2430 * Track whether we have multiple files in our lists. This will impact
2431 * how we do polling eventually, not spinning if we're on potentially
2432 * different devices.
2434 if (list_empty(&ctx->iopoll_list)) {
2435 ctx->poll_multi_file = false;
2436 } else if (!ctx->poll_multi_file) {
2437 struct io_kiocb *list_req;
2439 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2441 if (list_req->file != req->file)
2442 ctx->poll_multi_file = true;
2446 * For fast devices, IO may have already completed. If it has, add
2447 * it to the front so we find it first.
2449 if (READ_ONCE(req->iopoll_completed))
2450 list_add(&req->inflight_entry, &ctx->iopoll_list);
2452 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2454 if ((ctx->flags & IORING_SETUP_SQPOLL) && wq_has_sleeper(ctx->sqo_wait))
2455 wake_up(ctx->sqo_wait);
2458 static void __io_state_file_put(struct io_submit_state *state)
2460 if (state->has_refs)
2461 fput_many(state->file, state->has_refs);
2465 static inline void io_state_file_put(struct io_submit_state *state)
2468 __io_state_file_put(state);
2472 * Get as many references to a file as we have IOs left in this submission,
2473 * assuming most submissions are for one file, or at least that each file
2474 * has more than one submission.
2476 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2482 if (state->fd == fd) {
2487 __io_state_file_put(state);
2489 state->file = fget_many(fd, state->ios_left);
2495 state->has_refs = state->ios_left;
2499 static bool io_bdev_nowait(struct block_device *bdev)
2502 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2509 * If we tracked the file through the SCM inflight mechanism, we could support
2510 * any file. For now, just ensure that anything potentially problematic is done
2513 static bool io_file_supports_async(struct file *file, int rw)
2515 umode_t mode = file_inode(file)->i_mode;
2517 if (S_ISBLK(mode)) {
2518 if (io_bdev_nowait(file->f_inode->i_bdev))
2522 if (S_ISCHR(mode) || S_ISSOCK(mode))
2524 if (S_ISREG(mode)) {
2525 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2526 file->f_op != &io_uring_fops)
2531 /* any ->read/write should understand O_NONBLOCK */
2532 if (file->f_flags & O_NONBLOCK)
2535 if (!(file->f_mode & FMODE_NOWAIT))
2539 return file->f_op->read_iter != NULL;
2541 return file->f_op->write_iter != NULL;
2544 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2545 bool force_nonblock)
2547 struct io_ring_ctx *ctx = req->ctx;
2548 struct kiocb *kiocb = &req->rw.kiocb;
2552 if (S_ISREG(file_inode(req->file)->i_mode))
2553 req->flags |= REQ_F_ISREG;
2555 kiocb->ki_pos = READ_ONCE(sqe->off);
2556 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2557 req->flags |= REQ_F_CUR_POS;
2558 kiocb->ki_pos = req->file->f_pos;
2560 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2561 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2562 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2566 ioprio = READ_ONCE(sqe->ioprio);
2568 ret = ioprio_check_cap(ioprio);
2572 kiocb->ki_ioprio = ioprio;
2574 kiocb->ki_ioprio = get_current_ioprio();
2576 /* don't allow async punt if RWF_NOWAIT was requested */
2577 if (kiocb->ki_flags & IOCB_NOWAIT)
2578 req->flags |= REQ_F_NOWAIT;
2581 kiocb->ki_flags |= IOCB_NOWAIT;
2583 if (ctx->flags & IORING_SETUP_IOPOLL) {
2584 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2585 !kiocb->ki_filp->f_op->iopoll)
2588 kiocb->ki_flags |= IOCB_HIPRI;
2589 kiocb->ki_complete = io_complete_rw_iopoll;
2590 req->iopoll_completed = 0;
2592 if (kiocb->ki_flags & IOCB_HIPRI)
2594 kiocb->ki_complete = io_complete_rw;
2597 req->rw.addr = READ_ONCE(sqe->addr);
2598 req->rw.len = READ_ONCE(sqe->len);
2599 req->buf_index = READ_ONCE(sqe->buf_index);
2603 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2609 case -ERESTARTNOINTR:
2610 case -ERESTARTNOHAND:
2611 case -ERESTART_RESTARTBLOCK:
2613 * We can't just restart the syscall, since previously
2614 * submitted sqes may already be in progress. Just fail this
2620 kiocb->ki_complete(kiocb, ret, 0);
2624 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2625 struct io_comp_state *cs)
2627 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2629 /* add previously done IO, if any */
2630 if (req->io && req->io->rw.bytes_done > 0) {
2632 ret = req->io->rw.bytes_done;
2634 ret += req->io->rw.bytes_done;
2637 if (req->flags & REQ_F_CUR_POS)
2638 req->file->f_pos = kiocb->ki_pos;
2639 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2640 __io_complete_rw(req, ret, 0, cs);
2642 io_rw_done(kiocb, ret);
2645 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2646 struct iov_iter *iter)
2648 struct io_ring_ctx *ctx = req->ctx;
2649 size_t len = req->rw.len;
2650 struct io_mapped_ubuf *imu;
2651 u16 index, buf_index;
2655 /* attempt to use fixed buffers without having provided iovecs */
2656 if (unlikely(!ctx->user_bufs))
2659 buf_index = req->buf_index;
2660 if (unlikely(buf_index >= ctx->nr_user_bufs))
2663 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2664 imu = &ctx->user_bufs[index];
2665 buf_addr = req->rw.addr;
2668 if (buf_addr + len < buf_addr)
2670 /* not inside the mapped region */
2671 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2675 * May not be a start of buffer, set size appropriately
2676 * and advance us to the beginning.
2678 offset = buf_addr - imu->ubuf;
2679 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2683 * Don't use iov_iter_advance() here, as it's really slow for
2684 * using the latter parts of a big fixed buffer - it iterates
2685 * over each segment manually. We can cheat a bit here, because
2688 * 1) it's a BVEC iter, we set it up
2689 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2690 * first and last bvec
2692 * So just find our index, and adjust the iterator afterwards.
2693 * If the offset is within the first bvec (or the whole first
2694 * bvec, just use iov_iter_advance(). This makes it easier
2695 * since we can just skip the first segment, which may not
2696 * be PAGE_SIZE aligned.
2698 const struct bio_vec *bvec = imu->bvec;
2700 if (offset <= bvec->bv_len) {
2701 iov_iter_advance(iter, offset);
2703 unsigned long seg_skip;
2705 /* skip first vec */
2706 offset -= bvec->bv_len;
2707 seg_skip = 1 + (offset >> PAGE_SHIFT);
2709 iter->bvec = bvec + seg_skip;
2710 iter->nr_segs -= seg_skip;
2711 iter->count -= bvec->bv_len + offset;
2712 iter->iov_offset = offset & ~PAGE_MASK;
2719 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2722 mutex_unlock(&ctx->uring_lock);
2725 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2728 * "Normal" inline submissions always hold the uring_lock, since we
2729 * grab it from the system call. Same is true for the SQPOLL offload.
2730 * The only exception is when we've detached the request and issue it
2731 * from an async worker thread, grab the lock for that case.
2734 mutex_lock(&ctx->uring_lock);
2737 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2738 int bgid, struct io_buffer *kbuf,
2741 struct io_buffer *head;
2743 if (req->flags & REQ_F_BUFFER_SELECTED)
2746 io_ring_submit_lock(req->ctx, needs_lock);
2748 lockdep_assert_held(&req->ctx->uring_lock);
2750 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2752 if (!list_empty(&head->list)) {
2753 kbuf = list_last_entry(&head->list, struct io_buffer,
2755 list_del(&kbuf->list);
2758 idr_remove(&req->ctx->io_buffer_idr, bgid);
2760 if (*len > kbuf->len)
2763 kbuf = ERR_PTR(-ENOBUFS);
2766 io_ring_submit_unlock(req->ctx, needs_lock);
2771 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2774 struct io_buffer *kbuf;
2777 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2778 bgid = req->buf_index;
2779 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2782 req->rw.addr = (u64) (unsigned long) kbuf;
2783 req->flags |= REQ_F_BUFFER_SELECTED;
2784 return u64_to_user_ptr(kbuf->addr);
2787 #ifdef CONFIG_COMPAT
2788 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2791 struct compat_iovec __user *uiov;
2792 compat_ssize_t clen;
2796 uiov = u64_to_user_ptr(req->rw.addr);
2797 if (!access_ok(uiov, sizeof(*uiov)))
2799 if (__get_user(clen, &uiov->iov_len))
2805 buf = io_rw_buffer_select(req, &len, needs_lock);
2807 return PTR_ERR(buf);
2808 iov[0].iov_base = buf;
2809 iov[0].iov_len = (compat_size_t) len;
2814 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2817 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2821 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2824 len = iov[0].iov_len;
2827 buf = io_rw_buffer_select(req, &len, needs_lock);
2829 return PTR_ERR(buf);
2830 iov[0].iov_base = buf;
2831 iov[0].iov_len = len;
2835 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2838 if (req->flags & REQ_F_BUFFER_SELECTED) {
2839 struct io_buffer *kbuf;
2841 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2842 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2843 iov[0].iov_len = kbuf->len;
2848 else if (req->rw.len > 1)
2851 #ifdef CONFIG_COMPAT
2852 if (req->ctx->compat)
2853 return io_compat_import(req, iov, needs_lock);
2856 return __io_iov_buffer_select(req, iov, needs_lock);
2859 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2860 struct iovec **iovec, struct iov_iter *iter,
2863 void __user *buf = u64_to_user_ptr(req->rw.addr);
2864 size_t sqe_len = req->rw.len;
2868 opcode = req->opcode;
2869 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2871 return io_import_fixed(req, rw, iter);
2874 /* buffer index only valid with fixed read/write, or buffer select */
2875 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2878 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2879 if (req->flags & REQ_F_BUFFER_SELECT) {
2880 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2882 return PTR_ERR(buf);
2883 req->rw.len = sqe_len;
2886 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2888 return ret < 0 ? ret : sqe_len;
2891 if (req->flags & REQ_F_BUFFER_SELECT) {
2892 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2894 ret = (*iovec)->iov_len;
2895 iov_iter_init(iter, rw, *iovec, 1, ret);
2901 #ifdef CONFIG_COMPAT
2902 if (req->ctx->compat)
2903 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2907 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2910 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2911 struct iovec **iovec, struct iov_iter *iter,
2915 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
2917 return iov_iter_count(&req->io->rw.iter);
2920 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2922 return kiocb->ki_filp->f_mode & FMODE_STREAM ? NULL : &kiocb->ki_pos;
2926 * For files that don't have ->read_iter() and ->write_iter(), handle them
2927 * by looping over ->read() or ->write() manually.
2929 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2930 struct iov_iter *iter)
2935 * Don't support polled IO through this interface, and we can't
2936 * support non-blocking either. For the latter, this just causes
2937 * the kiocb to be handled from an async context.
2939 if (kiocb->ki_flags & IOCB_HIPRI)
2941 if (kiocb->ki_flags & IOCB_NOWAIT)
2944 while (iov_iter_count(iter)) {
2948 if (!iov_iter_is_bvec(iter)) {
2949 iovec = iov_iter_iovec(iter);
2951 /* fixed buffers import bvec */
2952 iovec.iov_base = kmap(iter->bvec->bv_page)
2954 iovec.iov_len = min(iter->count,
2955 iter->bvec->bv_len - iter->iov_offset);
2959 nr = file->f_op->read(file, iovec.iov_base,
2960 iovec.iov_len, io_kiocb_ppos(kiocb));
2962 nr = file->f_op->write(file, iovec.iov_base,
2963 iovec.iov_len, io_kiocb_ppos(kiocb));
2966 if (iov_iter_is_bvec(iter))
2967 kunmap(iter->bvec->bv_page);
2975 if (nr != iovec.iov_len)
2977 iov_iter_advance(iter, nr);
2983 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
2984 const struct iovec *fast_iov, struct iov_iter *iter)
2986 struct io_async_rw *rw = &req->io->rw;
2988 memcpy(&rw->iter, iter, sizeof(*iter));
2989 rw->free_iovec = NULL;
2991 /* can only be fixed buffers, no need to do anything */
2992 if (iter->type == ITER_BVEC)
2995 unsigned iov_off = 0;
2997 rw->iter.iov = rw->fast_iov;
2998 if (iter->iov != fast_iov) {
2999 iov_off = iter->iov - fast_iov;
3000 rw->iter.iov += iov_off;
3002 if (rw->fast_iov != fast_iov)
3003 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3004 sizeof(struct iovec) * iter->nr_segs);
3006 rw->free_iovec = iovec;
3007 req->flags |= REQ_F_NEED_CLEANUP;
3011 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
3013 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
3014 return req->io == NULL;
3017 static int io_alloc_async_ctx(struct io_kiocb *req)
3019 if (!io_op_defs[req->opcode].async_ctx)
3022 return __io_alloc_async_ctx(req);
3025 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3026 const struct iovec *fast_iov,
3027 struct iov_iter *iter, bool force)
3029 if (!force && !io_op_defs[req->opcode].async_ctx)
3032 if (__io_alloc_async_ctx(req))
3035 io_req_map_rw(req, iovec, fast_iov, iter);
3040 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
3041 bool force_nonblock)
3043 struct io_async_rw *iorw = &req->io->rw;
3047 iorw->iter.iov = iov = iorw->fast_iov;
3048 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, !force_nonblock);
3049 if (unlikely(ret < 0))
3052 iorw->iter.iov = iov;
3053 io_req_map_rw(req, iorw->iter.iov, iorw->fast_iov, &iorw->iter);
3057 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3058 bool force_nonblock)
3062 ret = io_prep_rw(req, sqe, force_nonblock);
3066 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3069 /* either don't need iovec imported or already have it */
3070 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3072 return io_rw_prep_async(req, READ, force_nonblock);
3076 * This is our waitqueue callback handler, registered through lock_page_async()
3077 * when we initially tried to do the IO with the iocb armed our waitqueue.
3078 * This gets called when the page is unlocked, and we generally expect that to
3079 * happen when the page IO is completed and the page is now uptodate. This will
3080 * queue a task_work based retry of the operation, attempting to copy the data
3081 * again. If the latter fails because the page was NOT uptodate, then we will
3082 * do a thread based blocking retry of the operation. That's the unexpected
3085 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3086 int sync, void *arg)
3088 struct wait_page_queue *wpq;
3089 struct io_kiocb *req = wait->private;
3090 struct wait_page_key *key = arg;
3093 wpq = container_of(wait, struct wait_page_queue, wait);
3095 if (!wake_page_match(wpq, key))
3098 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3099 list_del_init(&wait->entry);
3101 init_task_work(&req->task_work, io_req_task_submit);
3102 percpu_ref_get(&req->ctx->refs);
3104 /* submit ref gets dropped, acquire a new one */
3105 refcount_inc(&req->refs);
3106 ret = io_req_task_work_add(req, &req->task_work, true);
3107 if (unlikely(ret)) {
3108 struct task_struct *tsk;
3110 /* queue just for cancelation */
3111 init_task_work(&req->task_work, io_req_task_cancel);
3112 tsk = io_wq_get_task(req->ctx->io_wq);
3113 task_work_add(tsk, &req->task_work, 0);
3114 wake_up_process(tsk);
3120 * This controls whether a given IO request should be armed for async page
3121 * based retry. If we return false here, the request is handed to the async
3122 * worker threads for retry. If we're doing buffered reads on a regular file,
3123 * we prepare a private wait_page_queue entry and retry the operation. This
3124 * will either succeed because the page is now uptodate and unlocked, or it
3125 * will register a callback when the page is unlocked at IO completion. Through
3126 * that callback, io_uring uses task_work to setup a retry of the operation.
3127 * That retry will attempt the buffered read again. The retry will generally
3128 * succeed, or in rare cases where it fails, we then fall back to using the
3129 * async worker threads for a blocking retry.
3131 static bool io_rw_should_retry(struct io_kiocb *req)
3133 struct wait_page_queue *wait = &req->io->rw.wpq;
3134 struct kiocb *kiocb = &req->rw.kiocb;
3136 /* never retry for NOWAIT, we just complete with -EAGAIN */
3137 if (req->flags & REQ_F_NOWAIT)
3140 /* Only for buffered IO */
3141 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3145 * just use poll if we can, and don't attempt if the fs doesn't
3146 * support callback based unlocks
3148 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3151 wait->wait.func = io_async_buf_func;
3152 wait->wait.private = req;
3153 wait->wait.flags = 0;
3154 INIT_LIST_HEAD(&wait->wait.entry);
3155 kiocb->ki_flags |= IOCB_WAITQ;
3156 kiocb->ki_flags &= ~IOCB_NOWAIT;
3157 kiocb->ki_waitq = wait;
3161 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3163 if (req->file->f_op->read_iter)
3164 return call_read_iter(req->file, &req->rw.kiocb, iter);
3165 else if (req->file->f_op->read)
3166 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3171 static int io_read(struct io_kiocb *req, bool force_nonblock,
3172 struct io_comp_state *cs)
3174 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3175 struct kiocb *kiocb = &req->rw.kiocb;
3176 struct iov_iter __iter, *iter = &__iter;
3177 ssize_t io_size, ret, ret2;
3182 iter = &req->io->rw.iter;
3184 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3187 iov_count = iov_iter_count(iter);
3189 req->result = io_size;
3192 /* Ensure we clear previously set non-block flag */
3193 if (!force_nonblock)
3194 kiocb->ki_flags &= ~IOCB_NOWAIT;
3196 /* If the file doesn't support async, just async punt */
3197 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3201 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3205 ret = io_iter_do_read(req, iter);
3209 } else if (ret == -EIOCBQUEUED) {
3212 } else if (ret == -EAGAIN) {
3213 /* IOPOLL retry should happen for io-wq threads */
3214 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3216 /* no retry on NONBLOCK marked file */
3217 if (req->file->f_flags & O_NONBLOCK)
3219 /* some cases will consume bytes even on error returns */
3220 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3223 } else if (ret < 0) {
3224 /* make sure -ERESTARTSYS -> -EINTR is done */
3228 /* read it all, or we did blocking attempt. no retry. */
3229 if (!iov_iter_count(iter) || !force_nonblock ||
3230 (req->file->f_flags & O_NONBLOCK))
3235 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3242 /* it's copied and will be cleaned with ->io */
3244 /* now use our persistent iterator, if we aren't already */
3245 iter = &req->io->rw.iter;
3247 req->io->rw.bytes_done += ret;
3248 /* if we can retry, do so with the callbacks armed */
3249 if (!io_rw_should_retry(req)) {
3250 kiocb->ki_flags &= ~IOCB_WAITQ;
3255 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3256 * get -EIOCBQUEUED, then we'll get a notification when the desired
3257 * page gets unlocked. We can also get a partial read here, and if we
3258 * do, then just retry at the new offset.
3260 ret = io_iter_do_read(req, iter);
3261 if (ret == -EIOCBQUEUED) {
3264 } else if (ret > 0 && ret < io_size) {
3265 /* we got some bytes, but not all. retry. */
3269 kiocb_done(kiocb, ret, cs);
3272 /* it's reportedly faster than delegating the null check to kfree() */
3278 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3279 bool force_nonblock)
3283 ret = io_prep_rw(req, sqe, force_nonblock);
3287 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3290 /* either don't need iovec imported or already have it */
3291 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3293 return io_rw_prep_async(req, WRITE, force_nonblock);
3296 static int io_write(struct io_kiocb *req, bool force_nonblock,
3297 struct io_comp_state *cs)
3299 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3300 struct kiocb *kiocb = &req->rw.kiocb;
3301 struct iov_iter __iter, *iter = &__iter;
3303 ssize_t ret, ret2, io_size;
3306 iter = &req->io->rw.iter;
3308 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3311 iov_count = iov_iter_count(iter);
3313 req->result = io_size;
3315 /* Ensure we clear previously set non-block flag */
3316 if (!force_nonblock)
3317 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3319 /* If the file doesn't support async, just async punt */
3320 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3323 /* file path doesn't support NOWAIT for non-direct_IO */
3324 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3325 (req->flags & REQ_F_ISREG))
3328 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3333 * Open-code file_start_write here to grab freeze protection,
3334 * which will be released by another thread in
3335 * io_complete_rw(). Fool lockdep by telling it the lock got
3336 * released so that it doesn't complain about the held lock when
3337 * we return to userspace.
3339 if (req->flags & REQ_F_ISREG) {
3340 __sb_start_write(file_inode(req->file)->i_sb,
3341 SB_FREEZE_WRITE, true);
3342 __sb_writers_release(file_inode(req->file)->i_sb,
3345 kiocb->ki_flags |= IOCB_WRITE;
3347 if (req->file->f_op->write_iter)
3348 ret2 = call_write_iter(req->file, kiocb, iter);
3349 else if (req->file->f_op->write)
3350 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3355 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3356 * retry them without IOCB_NOWAIT.
3358 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3360 /* no retry on NONBLOCK marked file */
3361 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3363 if (!force_nonblock || ret2 != -EAGAIN) {
3364 /* IOPOLL retry should happen for io-wq threads */
3365 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3368 kiocb_done(kiocb, ret2, cs);
3371 /* some cases will consume bytes even on error returns */
3372 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3373 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3378 /* it's reportedly faster than delegating the null check to kfree() */
3384 static int __io_splice_prep(struct io_kiocb *req,
3385 const struct io_uring_sqe *sqe)
3387 struct io_splice* sp = &req->splice;
3388 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3391 if (req->flags & REQ_F_NEED_CLEANUP)
3393 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3397 sp->len = READ_ONCE(sqe->len);
3398 sp->flags = READ_ONCE(sqe->splice_flags);
3400 if (unlikely(sp->flags & ~valid_flags))
3403 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3404 (sp->flags & SPLICE_F_FD_IN_FIXED));
3407 req->flags |= REQ_F_NEED_CLEANUP;
3409 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3411 * Splice operation will be punted aync, and here need to
3412 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3414 io_req_init_async(req);
3415 req->work.flags |= IO_WQ_WORK_UNBOUND;
3421 static int io_tee_prep(struct io_kiocb *req,
3422 const struct io_uring_sqe *sqe)
3424 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3426 return __io_splice_prep(req, sqe);
3429 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3431 struct io_splice *sp = &req->splice;
3432 struct file *in = sp->file_in;
3433 struct file *out = sp->file_out;
3434 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3440 ret = do_tee(in, out, sp->len, flags);
3442 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3443 req->flags &= ~REQ_F_NEED_CLEANUP;
3446 req_set_fail_links(req);
3447 io_req_complete(req, ret);
3451 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3453 struct io_splice* sp = &req->splice;
3455 sp->off_in = READ_ONCE(sqe->splice_off_in);
3456 sp->off_out = READ_ONCE(sqe->off);
3457 return __io_splice_prep(req, sqe);
3460 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3462 struct io_splice *sp = &req->splice;
3463 struct file *in = sp->file_in;
3464 struct file *out = sp->file_out;
3465 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3466 loff_t *poff_in, *poff_out;
3472 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3473 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3476 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3478 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3479 req->flags &= ~REQ_F_NEED_CLEANUP;
3482 req_set_fail_links(req);
3483 io_req_complete(req, ret);
3488 * IORING_OP_NOP just posts a completion event, nothing else.
3490 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3492 struct io_ring_ctx *ctx = req->ctx;
3494 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3497 __io_req_complete(req, 0, 0, cs);
3501 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3503 struct io_ring_ctx *ctx = req->ctx;
3508 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3510 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3513 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3514 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3517 req->sync.off = READ_ONCE(sqe->off);
3518 req->sync.len = READ_ONCE(sqe->len);
3522 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3524 loff_t end = req->sync.off + req->sync.len;
3527 /* fsync always requires a blocking context */
3531 ret = vfs_fsync_range(req->file, req->sync.off,
3532 end > 0 ? end : LLONG_MAX,
3533 req->sync.flags & IORING_FSYNC_DATASYNC);
3535 req_set_fail_links(req);
3536 io_req_complete(req, ret);
3540 static int io_fallocate_prep(struct io_kiocb *req,
3541 const struct io_uring_sqe *sqe)
3543 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3545 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3548 req->sync.off = READ_ONCE(sqe->off);
3549 req->sync.len = READ_ONCE(sqe->addr);
3550 req->sync.mode = READ_ONCE(sqe->len);
3554 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3558 /* fallocate always requiring blocking context */
3561 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3564 req_set_fail_links(req);
3565 io_req_complete(req, ret);
3569 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3571 const char __user *fname;
3574 if (unlikely(sqe->ioprio || sqe->buf_index))
3576 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3579 /* open.how should be already initialised */
3580 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3581 req->open.how.flags |= O_LARGEFILE;
3583 req->open.dfd = READ_ONCE(sqe->fd);
3584 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3585 req->open.filename = getname(fname);
3586 if (IS_ERR(req->open.filename)) {
3587 ret = PTR_ERR(req->open.filename);
3588 req->open.filename = NULL;
3591 req->open.nofile = rlimit(RLIMIT_NOFILE);
3592 req->flags |= REQ_F_NEED_CLEANUP;
3596 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3600 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3602 if (req->flags & REQ_F_NEED_CLEANUP)
3604 mode = READ_ONCE(sqe->len);
3605 flags = READ_ONCE(sqe->open_flags);
3606 req->open.how = build_open_how(flags, mode);
3607 return __io_openat_prep(req, sqe);
3610 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3612 struct open_how __user *how;
3616 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3618 if (req->flags & REQ_F_NEED_CLEANUP)
3620 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3621 len = READ_ONCE(sqe->len);
3622 if (len < OPEN_HOW_SIZE_VER0)
3625 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3630 return __io_openat_prep(req, sqe);
3633 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3635 struct open_flags op;
3642 ret = build_open_flags(&req->open.how, &op);
3646 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3650 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3653 ret = PTR_ERR(file);
3655 fsnotify_open(file);
3656 fd_install(ret, file);
3659 putname(req->open.filename);
3660 req->flags &= ~REQ_F_NEED_CLEANUP;
3662 req_set_fail_links(req);
3663 io_req_complete(req, ret);
3667 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3669 return io_openat2(req, force_nonblock);
3672 static int io_remove_buffers_prep(struct io_kiocb *req,
3673 const struct io_uring_sqe *sqe)
3675 struct io_provide_buf *p = &req->pbuf;
3678 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3681 tmp = READ_ONCE(sqe->fd);
3682 if (!tmp || tmp > USHRT_MAX)
3685 memset(p, 0, sizeof(*p));
3687 p->bgid = READ_ONCE(sqe->buf_group);
3691 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3692 int bgid, unsigned nbufs)
3696 /* shouldn't happen */
3700 /* the head kbuf is the list itself */
3701 while (!list_empty(&buf->list)) {
3702 struct io_buffer *nxt;
3704 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3705 list_del(&nxt->list);
3712 idr_remove(&ctx->io_buffer_idr, bgid);
3717 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3718 struct io_comp_state *cs)
3720 struct io_provide_buf *p = &req->pbuf;
3721 struct io_ring_ctx *ctx = req->ctx;
3722 struct io_buffer *head;
3725 io_ring_submit_lock(ctx, !force_nonblock);
3727 lockdep_assert_held(&ctx->uring_lock);
3730 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3732 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3734 io_ring_submit_lock(ctx, !force_nonblock);
3736 req_set_fail_links(req);
3737 __io_req_complete(req, ret, 0, cs);
3741 static int io_provide_buffers_prep(struct io_kiocb *req,
3742 const struct io_uring_sqe *sqe)
3744 struct io_provide_buf *p = &req->pbuf;
3747 if (sqe->ioprio || sqe->rw_flags)
3750 tmp = READ_ONCE(sqe->fd);
3751 if (!tmp || tmp > USHRT_MAX)
3754 p->addr = READ_ONCE(sqe->addr);
3755 p->len = READ_ONCE(sqe->len);
3757 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3760 p->bgid = READ_ONCE(sqe->buf_group);
3761 tmp = READ_ONCE(sqe->off);
3762 if (tmp > USHRT_MAX)
3768 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3770 struct io_buffer *buf;
3771 u64 addr = pbuf->addr;
3772 int i, bid = pbuf->bid;
3774 for (i = 0; i < pbuf->nbufs; i++) {
3775 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3780 buf->len = pbuf->len;
3785 INIT_LIST_HEAD(&buf->list);
3788 list_add_tail(&buf->list, &(*head)->list);
3792 return i ? i : -ENOMEM;
3795 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3796 struct io_comp_state *cs)
3798 struct io_provide_buf *p = &req->pbuf;
3799 struct io_ring_ctx *ctx = req->ctx;
3800 struct io_buffer *head, *list;
3803 io_ring_submit_lock(ctx, !force_nonblock);
3805 lockdep_assert_held(&ctx->uring_lock);
3807 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3809 ret = io_add_buffers(p, &head);
3814 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3817 __io_remove_buffers(ctx, head, p->bgid, -1U);
3822 io_ring_submit_unlock(ctx, !force_nonblock);
3824 req_set_fail_links(req);
3825 __io_req_complete(req, ret, 0, cs);
3829 static int io_epoll_ctl_prep(struct io_kiocb *req,
3830 const struct io_uring_sqe *sqe)
3832 #if defined(CONFIG_EPOLL)
3833 if (sqe->ioprio || sqe->buf_index)
3835 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3838 req->epoll.epfd = READ_ONCE(sqe->fd);
3839 req->epoll.op = READ_ONCE(sqe->len);
3840 req->epoll.fd = READ_ONCE(sqe->off);
3842 if (ep_op_has_event(req->epoll.op)) {
3843 struct epoll_event __user *ev;
3845 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3846 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3856 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3857 struct io_comp_state *cs)
3859 #if defined(CONFIG_EPOLL)
3860 struct io_epoll *ie = &req->epoll;
3863 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3864 if (force_nonblock && ret == -EAGAIN)
3868 req_set_fail_links(req);
3869 __io_req_complete(req, ret, 0, cs);
3876 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3878 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3879 if (sqe->ioprio || sqe->buf_index || sqe->off)
3881 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3884 req->madvise.addr = READ_ONCE(sqe->addr);
3885 req->madvise.len = READ_ONCE(sqe->len);
3886 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3893 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3895 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3896 struct io_madvise *ma = &req->madvise;
3902 ret = do_madvise(ma->addr, ma->len, ma->advice);
3904 req_set_fail_links(req);
3905 io_req_complete(req, ret);
3912 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3914 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3916 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3919 req->fadvise.offset = READ_ONCE(sqe->off);
3920 req->fadvise.len = READ_ONCE(sqe->len);
3921 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3925 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3927 struct io_fadvise *fa = &req->fadvise;
3930 if (force_nonblock) {
3931 switch (fa->advice) {
3932 case POSIX_FADV_NORMAL:
3933 case POSIX_FADV_RANDOM:
3934 case POSIX_FADV_SEQUENTIAL:
3941 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3943 req_set_fail_links(req);
3944 io_req_complete(req, ret);
3948 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3950 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3952 if (sqe->ioprio || sqe->buf_index)
3954 if (req->flags & REQ_F_FIXED_FILE)
3957 req->statx.dfd = READ_ONCE(sqe->fd);
3958 req->statx.mask = READ_ONCE(sqe->len);
3959 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3960 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3961 req->statx.flags = READ_ONCE(sqe->statx_flags);
3966 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3968 struct io_statx *ctx = &req->statx;
3971 if (force_nonblock) {
3972 /* only need file table for an actual valid fd */
3973 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3974 req->flags |= REQ_F_NO_FILE_TABLE;
3978 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3982 req_set_fail_links(req);
3983 io_req_complete(req, ret);
3987 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3990 * If we queue this for async, it must not be cancellable. That would
3991 * leave the 'file' in an undeterminate state, and here need to modify
3992 * io_wq_work.flags, so initialize io_wq_work firstly.
3994 io_req_init_async(req);
3995 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3997 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3999 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4000 sqe->rw_flags || sqe->buf_index)
4002 if (req->flags & REQ_F_FIXED_FILE)
4005 req->close.fd = READ_ONCE(sqe->fd);
4006 if ((req->file && req->file->f_op == &io_uring_fops))
4009 req->close.put_file = NULL;
4013 static int io_close(struct io_kiocb *req, bool force_nonblock,
4014 struct io_comp_state *cs)
4016 struct io_close *close = &req->close;
4019 /* might be already done during nonblock submission */
4020 if (!close->put_file) {
4021 ret = __close_fd_get_file(close->fd, &close->put_file);
4023 return (ret == -ENOENT) ? -EBADF : ret;
4026 /* if the file has a flush method, be safe and punt to async */
4027 if (close->put_file->f_op->flush && force_nonblock) {
4028 /* was never set, but play safe */
4029 req->flags &= ~REQ_F_NOWAIT;
4030 /* avoid grabbing files - we don't need the files */
4031 req->flags |= REQ_F_NO_FILE_TABLE;
4035 /* No ->flush() or already async, safely close from here */
4036 ret = filp_close(close->put_file, req->work.files);
4038 req_set_fail_links(req);
4039 fput(close->put_file);
4040 close->put_file = NULL;
4041 __io_req_complete(req, ret, 0, cs);
4045 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4047 struct io_ring_ctx *ctx = req->ctx;
4052 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4054 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4057 req->sync.off = READ_ONCE(sqe->off);
4058 req->sync.len = READ_ONCE(sqe->len);
4059 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4063 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4067 /* sync_file_range always requires a blocking context */
4071 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4074 req_set_fail_links(req);
4075 io_req_complete(req, ret);
4079 #if defined(CONFIG_NET)
4080 static int io_setup_async_msg(struct io_kiocb *req,
4081 struct io_async_msghdr *kmsg)
4085 if (io_alloc_async_ctx(req)) {
4086 if (kmsg->iov != kmsg->fast_iov)
4090 req->flags |= REQ_F_NEED_CLEANUP;
4091 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
4095 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4096 struct io_async_msghdr *iomsg)
4098 iomsg->iov = iomsg->fast_iov;
4099 iomsg->msg.msg_name = &iomsg->addr;
4100 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4101 req->sr_msg.msg_flags, &iomsg->iov);
4104 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4106 struct io_sr_msg *sr = &req->sr_msg;
4107 struct io_async_ctx *io = req->io;
4110 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4113 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4114 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4115 sr->len = READ_ONCE(sqe->len);
4117 #ifdef CONFIG_COMPAT
4118 if (req->ctx->compat)
4119 sr->msg_flags |= MSG_CMSG_COMPAT;
4122 if (!io || req->opcode == IORING_OP_SEND)
4124 /* iovec is already imported */
4125 if (req->flags & REQ_F_NEED_CLEANUP)
4128 ret = io_sendmsg_copy_hdr(req, &io->msg);
4130 req->flags |= REQ_F_NEED_CLEANUP;
4134 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4135 struct io_comp_state *cs)
4137 struct io_async_msghdr iomsg, *kmsg;
4138 struct socket *sock;
4142 sock = sock_from_file(req->file, &ret);
4143 if (unlikely(!sock))
4147 kmsg = &req->io->msg;
4148 kmsg->msg.msg_name = &req->io->msg.addr;
4149 /* if iov is set, it's allocated already */
4151 kmsg->iov = kmsg->fast_iov;
4152 kmsg->msg.msg_iter.iov = kmsg->iov;
4154 ret = io_sendmsg_copy_hdr(req, &iomsg);
4160 flags = req->sr_msg.msg_flags;
4161 if (flags & MSG_DONTWAIT)
4162 req->flags |= REQ_F_NOWAIT;
4163 else if (force_nonblock)
4164 flags |= MSG_DONTWAIT;
4166 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4167 if (force_nonblock && ret == -EAGAIN)
4168 return io_setup_async_msg(req, kmsg);
4169 if (ret == -ERESTARTSYS)
4172 if (kmsg->iov != kmsg->fast_iov)
4174 req->flags &= ~REQ_F_NEED_CLEANUP;
4176 req_set_fail_links(req);
4177 __io_req_complete(req, ret, 0, cs);
4181 static int io_send(struct io_kiocb *req, bool force_nonblock,
4182 struct io_comp_state *cs)
4184 struct io_sr_msg *sr = &req->sr_msg;
4187 struct socket *sock;
4191 sock = sock_from_file(req->file, &ret);
4192 if (unlikely(!sock))
4195 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4199 msg.msg_name = NULL;
4200 msg.msg_control = NULL;
4201 msg.msg_controllen = 0;
4202 msg.msg_namelen = 0;
4204 flags = req->sr_msg.msg_flags;
4205 if (flags & MSG_DONTWAIT)
4206 req->flags |= REQ_F_NOWAIT;
4207 else if (force_nonblock)
4208 flags |= MSG_DONTWAIT;
4210 msg.msg_flags = flags;
4211 ret = sock_sendmsg(sock, &msg);
4212 if (force_nonblock && ret == -EAGAIN)
4214 if (ret == -ERESTARTSYS)
4218 req_set_fail_links(req);
4219 __io_req_complete(req, ret, 0, cs);
4223 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4224 struct io_async_msghdr *iomsg)
4226 struct io_sr_msg *sr = &req->sr_msg;
4227 struct iovec __user *uiov;
4231 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4232 &iomsg->uaddr, &uiov, &iov_len);
4236 if (req->flags & REQ_F_BUFFER_SELECT) {
4239 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4241 sr->len = iomsg->iov[0].iov_len;
4242 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4246 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4247 &iomsg->iov, &iomsg->msg.msg_iter);
4255 #ifdef CONFIG_COMPAT
4256 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4257 struct io_async_msghdr *iomsg)
4259 struct compat_msghdr __user *msg_compat;
4260 struct io_sr_msg *sr = &req->sr_msg;
4261 struct compat_iovec __user *uiov;
4266 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4267 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4272 uiov = compat_ptr(ptr);
4273 if (req->flags & REQ_F_BUFFER_SELECT) {
4274 compat_ssize_t clen;
4278 if (!access_ok(uiov, sizeof(*uiov)))
4280 if (__get_user(clen, &uiov->iov_len))
4284 sr->len = iomsg->iov[0].iov_len;
4287 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4289 &iomsg->msg.msg_iter);
4298 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4299 struct io_async_msghdr *iomsg)
4301 iomsg->msg.msg_name = &iomsg->addr;
4302 iomsg->iov = iomsg->fast_iov;
4304 #ifdef CONFIG_COMPAT
4305 if (req->ctx->compat)
4306 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4309 return __io_recvmsg_copy_hdr(req, iomsg);
4312 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4315 struct io_sr_msg *sr = &req->sr_msg;
4316 struct io_buffer *kbuf;
4318 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4323 req->flags |= REQ_F_BUFFER_SELECTED;
4327 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4329 return io_put_kbuf(req, req->sr_msg.kbuf);
4332 static int io_recvmsg_prep(struct io_kiocb *req,
4333 const struct io_uring_sqe *sqe)
4335 struct io_sr_msg *sr = &req->sr_msg;
4336 struct io_async_ctx *io = req->io;
4339 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4342 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4343 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4344 sr->len = READ_ONCE(sqe->len);
4345 sr->bgid = READ_ONCE(sqe->buf_group);
4347 #ifdef CONFIG_COMPAT
4348 if (req->ctx->compat)
4349 sr->msg_flags |= MSG_CMSG_COMPAT;
4352 if (!io || req->opcode == IORING_OP_RECV)
4354 /* iovec is already imported */
4355 if (req->flags & REQ_F_NEED_CLEANUP)
4358 ret = io_recvmsg_copy_hdr(req, &io->msg);
4360 req->flags |= REQ_F_NEED_CLEANUP;
4364 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4365 struct io_comp_state *cs)
4367 struct io_async_msghdr iomsg, *kmsg;
4368 struct socket *sock;
4369 struct io_buffer *kbuf;
4371 int ret, cflags = 0;
4373 sock = sock_from_file(req->file, &ret);
4374 if (unlikely(!sock))
4378 kmsg = &req->io->msg;
4379 kmsg->msg.msg_name = &req->io->msg.addr;
4380 /* if iov is set, it's allocated already */
4382 kmsg->iov = kmsg->fast_iov;
4383 kmsg->msg.msg_iter.iov = kmsg->iov;
4385 ret = io_recvmsg_copy_hdr(req, &iomsg);
4391 if (req->flags & REQ_F_BUFFER_SELECT) {
4392 kbuf = io_recv_buffer_select(req, !force_nonblock);
4394 return PTR_ERR(kbuf);
4395 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4396 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4397 1, req->sr_msg.len);
4400 flags = req->sr_msg.msg_flags;
4401 if (flags & MSG_DONTWAIT)
4402 req->flags |= REQ_F_NOWAIT;
4403 else if (force_nonblock)
4404 flags |= MSG_DONTWAIT;
4406 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4407 kmsg->uaddr, flags);
4408 if (force_nonblock && ret == -EAGAIN)
4409 return io_setup_async_msg(req, kmsg);
4410 if (ret == -ERESTARTSYS)
4413 if (req->flags & REQ_F_BUFFER_SELECTED)
4414 cflags = io_put_recv_kbuf(req);
4415 if (kmsg->iov != kmsg->fast_iov)
4417 req->flags &= ~REQ_F_NEED_CLEANUP;
4419 req_set_fail_links(req);
4420 __io_req_complete(req, ret, cflags, cs);
4424 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4425 struct io_comp_state *cs)
4427 struct io_buffer *kbuf;
4428 struct io_sr_msg *sr = &req->sr_msg;
4430 void __user *buf = sr->buf;
4431 struct socket *sock;
4434 int ret, cflags = 0;
4436 sock = sock_from_file(req->file, &ret);
4437 if (unlikely(!sock))
4440 if (req->flags & REQ_F_BUFFER_SELECT) {
4441 kbuf = io_recv_buffer_select(req, !force_nonblock);
4443 return PTR_ERR(kbuf);
4444 buf = u64_to_user_ptr(kbuf->addr);
4447 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4451 msg.msg_name = NULL;
4452 msg.msg_control = NULL;
4453 msg.msg_controllen = 0;
4454 msg.msg_namelen = 0;
4455 msg.msg_iocb = NULL;
4458 flags = req->sr_msg.msg_flags;
4459 if (flags & MSG_DONTWAIT)
4460 req->flags |= REQ_F_NOWAIT;
4461 else if (force_nonblock)
4462 flags |= MSG_DONTWAIT;
4464 ret = sock_recvmsg(sock, &msg, flags);
4465 if (force_nonblock && ret == -EAGAIN)
4467 if (ret == -ERESTARTSYS)
4470 if (req->flags & REQ_F_BUFFER_SELECTED)
4471 cflags = io_put_recv_kbuf(req);
4473 req_set_fail_links(req);
4474 __io_req_complete(req, ret, cflags, cs);
4478 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4480 struct io_accept *accept = &req->accept;
4482 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4484 if (sqe->ioprio || sqe->len || sqe->buf_index)
4487 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4488 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4489 accept->flags = READ_ONCE(sqe->accept_flags);
4490 accept->nofile = rlimit(RLIMIT_NOFILE);
4494 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4495 struct io_comp_state *cs)
4497 struct io_accept *accept = &req->accept;
4498 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4501 if (req->file->f_flags & O_NONBLOCK)
4502 req->flags |= REQ_F_NOWAIT;
4504 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4505 accept->addr_len, accept->flags,
4507 if (ret == -EAGAIN && force_nonblock)
4510 if (ret == -ERESTARTSYS)
4512 req_set_fail_links(req);
4514 __io_req_complete(req, ret, 0, cs);
4518 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4520 struct io_connect *conn = &req->connect;
4521 struct io_async_ctx *io = req->io;
4523 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4525 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4528 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4529 conn->addr_len = READ_ONCE(sqe->addr2);
4534 return move_addr_to_kernel(conn->addr, conn->addr_len,
4535 &io->connect.address);
4538 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4539 struct io_comp_state *cs)
4541 struct io_async_ctx __io, *io;
4542 unsigned file_flags;
4548 ret = move_addr_to_kernel(req->connect.addr,
4549 req->connect.addr_len,
4550 &__io.connect.address);
4556 file_flags = force_nonblock ? O_NONBLOCK : 0;
4558 ret = __sys_connect_file(req->file, &io->connect.address,
4559 req->connect.addr_len, file_flags);
4560 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4563 if (io_alloc_async_ctx(req)) {
4567 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4570 if (ret == -ERESTARTSYS)
4574 req_set_fail_links(req);
4575 __io_req_complete(req, ret, 0, cs);
4578 #else /* !CONFIG_NET */
4579 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4584 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4585 struct io_comp_state *cs)
4590 static int io_send(struct io_kiocb *req, bool force_nonblock,
4591 struct io_comp_state *cs)
4596 static int io_recvmsg_prep(struct io_kiocb *req,
4597 const struct io_uring_sqe *sqe)
4602 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4603 struct io_comp_state *cs)
4608 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4609 struct io_comp_state *cs)
4614 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4619 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4620 struct io_comp_state *cs)
4625 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4630 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4631 struct io_comp_state *cs)
4635 #endif /* CONFIG_NET */
4637 struct io_poll_table {
4638 struct poll_table_struct pt;
4639 struct io_kiocb *req;
4643 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4644 __poll_t mask, task_work_func_t func)
4649 /* for instances that support it check for an event match first: */
4650 if (mask && !(mask & poll->events))
4653 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4655 list_del_init(&poll->wait.entry);
4658 init_task_work(&req->task_work, func);
4659 percpu_ref_get(&req->ctx->refs);
4662 * If we using the signalfd wait_queue_head for this wakeup, then
4663 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4664 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4665 * either, as the normal wakeup will suffice.
4667 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4670 * If this fails, then the task is exiting. When a task exits, the
4671 * work gets canceled, so just cancel this request as well instead
4672 * of executing it. We can't safely execute it anyway, as we may not
4673 * have the needed state needed for it anyway.
4675 ret = io_req_task_work_add(req, &req->task_work, twa_signal_ok);
4676 if (unlikely(ret)) {
4677 struct task_struct *tsk;
4679 WRITE_ONCE(poll->canceled, true);
4680 tsk = io_wq_get_task(req->ctx->io_wq);
4681 task_work_add(tsk, &req->task_work, 0);
4682 wake_up_process(tsk);
4687 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4688 __acquires(&req->ctx->completion_lock)
4690 struct io_ring_ctx *ctx = req->ctx;
4692 if (!req->result && !READ_ONCE(poll->canceled)) {
4693 struct poll_table_struct pt = { ._key = poll->events };
4695 req->result = vfs_poll(req->file, &pt) & poll->events;
4698 spin_lock_irq(&ctx->completion_lock);
4699 if (!req->result && !READ_ONCE(poll->canceled)) {
4700 add_wait_queue(poll->head, &poll->wait);
4707 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4709 /* pure poll stashes this in ->io, poll driven retry elsewhere */
4710 if (req->opcode == IORING_OP_POLL_ADD)
4711 return (struct io_poll_iocb *) req->io;
4712 return req->apoll->double_poll;
4715 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4717 if (req->opcode == IORING_OP_POLL_ADD)
4719 return &req->apoll->poll;
4722 static void io_poll_remove_double(struct io_kiocb *req)
4724 struct io_poll_iocb *poll = io_poll_get_double(req);
4726 lockdep_assert_held(&req->ctx->completion_lock);
4728 if (poll && poll->head) {
4729 struct wait_queue_head *head = poll->head;
4731 spin_lock(&head->lock);
4732 list_del_init(&poll->wait.entry);
4733 if (poll->wait.private)
4734 refcount_dec(&req->refs);
4736 spin_unlock(&head->lock);
4740 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4742 struct io_ring_ctx *ctx = req->ctx;
4744 io_poll_remove_double(req);
4745 req->poll.done = true;
4746 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4747 io_commit_cqring(ctx);
4750 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4752 struct io_ring_ctx *ctx = req->ctx;
4754 if (io_poll_rewait(req, &req->poll)) {
4755 spin_unlock_irq(&ctx->completion_lock);
4759 hash_del(&req->hash_node);
4760 io_poll_complete(req, req->result, 0);
4761 req->flags |= REQ_F_COMP_LOCKED;
4762 *nxt = io_put_req_find_next(req);
4763 spin_unlock_irq(&ctx->completion_lock);
4765 io_cqring_ev_posted(ctx);
4768 static void io_poll_task_func(struct callback_head *cb)
4770 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4771 struct io_ring_ctx *ctx = req->ctx;
4772 struct io_kiocb *nxt = NULL;
4774 io_poll_task_handler(req, &nxt);
4776 __io_req_task_submit(nxt);
4777 percpu_ref_put(&ctx->refs);
4780 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4781 int sync, void *key)
4783 struct io_kiocb *req = wait->private;
4784 struct io_poll_iocb *poll = io_poll_get_single(req);
4785 __poll_t mask = key_to_poll(key);
4787 /* for instances that support it check for an event match first: */
4788 if (mask && !(mask & poll->events))
4791 list_del_init(&wait->entry);
4793 if (poll && poll->head) {
4796 spin_lock(&poll->head->lock);
4797 done = list_empty(&poll->wait.entry);
4799 list_del_init(&poll->wait.entry);
4800 /* make sure double remove sees this as being gone */
4801 wait->private = NULL;
4802 spin_unlock(&poll->head->lock);
4804 __io_async_wake(req, poll, mask, io_poll_task_func);
4806 refcount_dec(&req->refs);
4810 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4811 wait_queue_func_t wake_func)
4815 poll->canceled = false;
4816 poll->events = events;
4817 INIT_LIST_HEAD(&poll->wait.entry);
4818 init_waitqueue_func_entry(&poll->wait, wake_func);
4821 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4822 struct wait_queue_head *head,
4823 struct io_poll_iocb **poll_ptr)
4825 struct io_kiocb *req = pt->req;
4828 * If poll->head is already set, it's because the file being polled
4829 * uses multiple waitqueues for poll handling (eg one for read, one
4830 * for write). Setup a separate io_poll_iocb if this happens.
4832 if (unlikely(poll->head)) {
4833 /* already have a 2nd entry, fail a third attempt */
4835 pt->error = -EINVAL;
4838 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4840 pt->error = -ENOMEM;
4843 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4844 refcount_inc(&req->refs);
4845 poll->wait.private = req;
4852 if (poll->events & EPOLLEXCLUSIVE)
4853 add_wait_queue_exclusive(head, &poll->wait);
4855 add_wait_queue(head, &poll->wait);
4858 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4859 struct poll_table_struct *p)
4861 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4862 struct async_poll *apoll = pt->req->apoll;
4864 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4867 static void io_async_task_func(struct callback_head *cb)
4869 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4870 struct async_poll *apoll = req->apoll;
4871 struct io_ring_ctx *ctx = req->ctx;
4873 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4875 if (io_poll_rewait(req, &apoll->poll)) {
4876 spin_unlock_irq(&ctx->completion_lock);
4877 percpu_ref_put(&ctx->refs);
4881 /* If req is still hashed, it cannot have been canceled. Don't check. */
4882 if (hash_hashed(&req->hash_node))
4883 hash_del(&req->hash_node);
4885 io_poll_remove_double(req);
4886 spin_unlock_irq(&ctx->completion_lock);
4888 if (!READ_ONCE(apoll->poll.canceled))
4889 __io_req_task_submit(req);
4891 __io_req_task_cancel(req, -ECANCELED);
4893 percpu_ref_put(&ctx->refs);
4894 kfree(apoll->double_poll);
4898 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4901 struct io_kiocb *req = wait->private;
4902 struct io_poll_iocb *poll = &req->apoll->poll;
4904 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4907 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4910 static void io_poll_req_insert(struct io_kiocb *req)
4912 struct io_ring_ctx *ctx = req->ctx;
4913 struct hlist_head *list;
4915 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4916 hlist_add_head(&req->hash_node, list);
4919 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4920 struct io_poll_iocb *poll,
4921 struct io_poll_table *ipt, __poll_t mask,
4922 wait_queue_func_t wake_func)
4923 __acquires(&ctx->completion_lock)
4925 struct io_ring_ctx *ctx = req->ctx;
4926 bool cancel = false;
4928 io_init_poll_iocb(poll, mask, wake_func);
4929 poll->file = req->file;
4930 poll->wait.private = req;
4932 ipt->pt._key = mask;
4934 ipt->error = -EINVAL;
4936 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4938 spin_lock_irq(&ctx->completion_lock);
4939 if (likely(poll->head)) {
4940 spin_lock(&poll->head->lock);
4941 if (unlikely(list_empty(&poll->wait.entry))) {
4947 if (mask || ipt->error)
4948 list_del_init(&poll->wait.entry);
4950 WRITE_ONCE(poll->canceled, true);
4951 else if (!poll->done) /* actually waiting for an event */
4952 io_poll_req_insert(req);
4953 spin_unlock(&poll->head->lock);
4959 static bool io_arm_poll_handler(struct io_kiocb *req)
4961 const struct io_op_def *def = &io_op_defs[req->opcode];
4962 struct io_ring_ctx *ctx = req->ctx;
4963 struct async_poll *apoll;
4964 struct io_poll_table ipt;
4968 if (!req->file || !file_can_poll(req->file))
4970 if (req->flags & REQ_F_POLLED)
4974 else if (def->pollout)
4978 /* if we can't nonblock try, then no point in arming a poll handler */
4979 if (!io_file_supports_async(req->file, rw))
4982 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4983 if (unlikely(!apoll))
4985 apoll->double_poll = NULL;
4987 req->flags |= REQ_F_POLLED;
4989 INIT_HLIST_NODE(&req->hash_node);
4993 mask |= POLLIN | POLLRDNORM;
4995 mask |= POLLOUT | POLLWRNORM;
4996 mask |= POLLERR | POLLPRI;
4998 ipt.pt._qproc = io_async_queue_proc;
5000 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5002 if (ret || ipt.error) {
5003 io_poll_remove_double(req);
5004 spin_unlock_irq(&ctx->completion_lock);
5005 kfree(apoll->double_poll);
5009 spin_unlock_irq(&ctx->completion_lock);
5010 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5011 apoll->poll.events);
5015 static bool __io_poll_remove_one(struct io_kiocb *req,
5016 struct io_poll_iocb *poll)
5018 bool do_complete = false;
5020 spin_lock(&poll->head->lock);
5021 WRITE_ONCE(poll->canceled, true);
5022 if (!list_empty(&poll->wait.entry)) {
5023 list_del_init(&poll->wait.entry);
5026 spin_unlock(&poll->head->lock);
5027 hash_del(&req->hash_node);
5031 static bool io_poll_remove_one(struct io_kiocb *req)
5035 io_poll_remove_double(req);
5037 if (req->opcode == IORING_OP_POLL_ADD) {
5038 do_complete = __io_poll_remove_one(req, &req->poll);
5040 struct async_poll *apoll = req->apoll;
5042 /* non-poll requests have submit ref still */
5043 do_complete = __io_poll_remove_one(req, &apoll->poll);
5046 kfree(apoll->double_poll);
5052 io_cqring_fill_event(req, -ECANCELED);
5053 io_commit_cqring(req->ctx);
5054 req->flags |= REQ_F_COMP_LOCKED;
5055 req_set_fail_links(req);
5063 * Returns true if we found and killed one or more poll requests
5065 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5067 struct hlist_node *tmp;
5068 struct io_kiocb *req;
5071 spin_lock_irq(&ctx->completion_lock);
5072 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5073 struct hlist_head *list;
5075 list = &ctx->cancel_hash[i];
5076 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5077 if (io_task_match(req, tsk))
5078 posted += io_poll_remove_one(req);
5081 spin_unlock_irq(&ctx->completion_lock);
5084 io_cqring_ev_posted(ctx);
5089 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5091 struct hlist_head *list;
5092 struct io_kiocb *req;
5094 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5095 hlist_for_each_entry(req, list, hash_node) {
5096 if (sqe_addr != req->user_data)
5098 if (io_poll_remove_one(req))
5106 static int io_poll_remove_prep(struct io_kiocb *req,
5107 const struct io_uring_sqe *sqe)
5109 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5111 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5115 req->poll.addr = READ_ONCE(sqe->addr);
5120 * Find a running poll command that matches one specified in sqe->addr,
5121 * and remove it if found.
5123 static int io_poll_remove(struct io_kiocb *req)
5125 struct io_ring_ctx *ctx = req->ctx;
5129 addr = req->poll.addr;
5130 spin_lock_irq(&ctx->completion_lock);
5131 ret = io_poll_cancel(ctx, addr);
5132 spin_unlock_irq(&ctx->completion_lock);
5135 req_set_fail_links(req);
5136 io_req_complete(req, ret);
5140 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5143 struct io_kiocb *req = wait->private;
5144 struct io_poll_iocb *poll = &req->poll;
5146 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5149 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5150 struct poll_table_struct *p)
5152 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5154 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
5157 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5159 struct io_poll_iocb *poll = &req->poll;
5162 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5164 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5169 events = READ_ONCE(sqe->poll32_events);
5171 events = swahw32(events);
5173 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5174 (events & EPOLLEXCLUSIVE);
5178 static int io_poll_add(struct io_kiocb *req)
5180 struct io_poll_iocb *poll = &req->poll;
5181 struct io_ring_ctx *ctx = req->ctx;
5182 struct io_poll_table ipt;
5185 INIT_HLIST_NODE(&req->hash_node);
5186 ipt.pt._qproc = io_poll_queue_proc;
5188 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5191 if (mask) { /* no async, we'd stolen it */
5193 io_poll_complete(req, mask, 0);
5195 spin_unlock_irq(&ctx->completion_lock);
5198 io_cqring_ev_posted(ctx);
5204 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5206 struct io_timeout_data *data = container_of(timer,
5207 struct io_timeout_data, timer);
5208 struct io_kiocb *req = data->req;
5209 struct io_ring_ctx *ctx = req->ctx;
5210 unsigned long flags;
5212 spin_lock_irqsave(&ctx->completion_lock, flags);
5213 atomic_set(&req->ctx->cq_timeouts,
5214 atomic_read(&req->ctx->cq_timeouts) + 1);
5217 * We could be racing with timeout deletion. If the list is empty,
5218 * then timeout lookup already found it and will be handling it.
5220 if (!list_empty(&req->timeout.list))
5221 list_del_init(&req->timeout.list);
5223 io_cqring_fill_event(req, -ETIME);
5224 io_commit_cqring(ctx);
5225 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5227 io_cqring_ev_posted(ctx);
5228 req_set_fail_links(req);
5230 return HRTIMER_NORESTART;
5233 static int __io_timeout_cancel(struct io_kiocb *req)
5237 list_del_init(&req->timeout.list);
5239 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5243 req_set_fail_links(req);
5244 req->flags |= REQ_F_COMP_LOCKED;
5245 io_cqring_fill_event(req, -ECANCELED);
5250 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5252 struct io_kiocb *req;
5255 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5256 if (user_data == req->user_data) {
5265 return __io_timeout_cancel(req);
5268 static int io_timeout_remove_prep(struct io_kiocb *req,
5269 const struct io_uring_sqe *sqe)
5271 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5273 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5275 if (sqe->ioprio || sqe->buf_index || sqe->len)
5278 req->timeout.addr = READ_ONCE(sqe->addr);
5279 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5280 if (req->timeout.flags)
5287 * Remove or update an existing timeout command
5289 static int io_timeout_remove(struct io_kiocb *req)
5291 struct io_ring_ctx *ctx = req->ctx;
5294 spin_lock_irq(&ctx->completion_lock);
5295 ret = io_timeout_cancel(ctx, req->timeout.addr);
5297 io_cqring_fill_event(req, ret);
5298 io_commit_cqring(ctx);
5299 spin_unlock_irq(&ctx->completion_lock);
5300 io_cqring_ev_posted(ctx);
5302 req_set_fail_links(req);
5307 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5308 bool is_timeout_link)
5310 struct io_timeout_data *data;
5312 u32 off = READ_ONCE(sqe->off);
5314 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5316 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5318 if (off && is_timeout_link)
5320 flags = READ_ONCE(sqe->timeout_flags);
5321 if (flags & ~IORING_TIMEOUT_ABS)
5324 req->timeout.off = off;
5326 if (!req->io && io_alloc_async_ctx(req))
5329 data = &req->io->timeout;
5332 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5335 if (flags & IORING_TIMEOUT_ABS)
5336 data->mode = HRTIMER_MODE_ABS;
5338 data->mode = HRTIMER_MODE_REL;
5340 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5344 static int io_timeout(struct io_kiocb *req)
5346 struct io_ring_ctx *ctx = req->ctx;
5347 struct io_timeout_data *data = &req->io->timeout;
5348 struct list_head *entry;
5349 u32 tail, off = req->timeout.off;
5351 spin_lock_irq(&ctx->completion_lock);
5354 * sqe->off holds how many events that need to occur for this
5355 * timeout event to be satisfied. If it isn't set, then this is
5356 * a pure timeout request, sequence isn't used.
5358 if (io_is_timeout_noseq(req)) {
5359 entry = ctx->timeout_list.prev;
5363 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5364 req->timeout.target_seq = tail + off;
5367 * Insertion sort, ensuring the first entry in the list is always
5368 * the one we need first.
5370 list_for_each_prev(entry, &ctx->timeout_list) {
5371 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5374 if (io_is_timeout_noseq(nxt))
5376 /* nxt.seq is behind @tail, otherwise would've been completed */
5377 if (off >= nxt->timeout.target_seq - tail)
5381 list_add(&req->timeout.list, entry);
5382 data->timer.function = io_timeout_fn;
5383 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5384 spin_unlock_irq(&ctx->completion_lock);
5388 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5390 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5392 return req->user_data == (unsigned long) data;
5395 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5397 enum io_wq_cancel cancel_ret;
5400 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5401 switch (cancel_ret) {
5402 case IO_WQ_CANCEL_OK:
5405 case IO_WQ_CANCEL_RUNNING:
5408 case IO_WQ_CANCEL_NOTFOUND:
5416 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5417 struct io_kiocb *req, __u64 sqe_addr,
5420 unsigned long flags;
5423 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5424 if (ret != -ENOENT) {
5425 spin_lock_irqsave(&ctx->completion_lock, flags);
5429 spin_lock_irqsave(&ctx->completion_lock, flags);
5430 ret = io_timeout_cancel(ctx, sqe_addr);
5433 ret = io_poll_cancel(ctx, sqe_addr);
5437 io_cqring_fill_event(req, ret);
5438 io_commit_cqring(ctx);
5439 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5440 io_cqring_ev_posted(ctx);
5443 req_set_fail_links(req);
5447 static int io_async_cancel_prep(struct io_kiocb *req,
5448 const struct io_uring_sqe *sqe)
5450 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5452 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5454 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5457 req->cancel.addr = READ_ONCE(sqe->addr);
5461 static int io_async_cancel(struct io_kiocb *req)
5463 struct io_ring_ctx *ctx = req->ctx;
5465 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5469 static int io_files_update_prep(struct io_kiocb *req,
5470 const struct io_uring_sqe *sqe)
5472 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5474 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5476 if (sqe->ioprio || sqe->rw_flags)
5479 req->files_update.offset = READ_ONCE(sqe->off);
5480 req->files_update.nr_args = READ_ONCE(sqe->len);
5481 if (!req->files_update.nr_args)
5483 req->files_update.arg = READ_ONCE(sqe->addr);
5487 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5488 struct io_comp_state *cs)
5490 struct io_ring_ctx *ctx = req->ctx;
5491 struct io_uring_files_update up;
5497 up.offset = req->files_update.offset;
5498 up.fds = req->files_update.arg;
5500 mutex_lock(&ctx->uring_lock);
5501 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5502 mutex_unlock(&ctx->uring_lock);
5505 req_set_fail_links(req);
5506 __io_req_complete(req, ret, 0, cs);
5510 static int io_req_defer_prep(struct io_kiocb *req,
5511 const struct io_uring_sqe *sqe)
5518 if (io_alloc_async_ctx(req))
5520 ret = io_prep_work_files(req);
5524 io_prep_async_work(req);
5526 switch (req->opcode) {
5529 case IORING_OP_READV:
5530 case IORING_OP_READ_FIXED:
5531 case IORING_OP_READ:
5532 ret = io_read_prep(req, sqe, true);
5534 case IORING_OP_WRITEV:
5535 case IORING_OP_WRITE_FIXED:
5536 case IORING_OP_WRITE:
5537 ret = io_write_prep(req, sqe, true);
5539 case IORING_OP_POLL_ADD:
5540 ret = io_poll_add_prep(req, sqe);
5542 case IORING_OP_POLL_REMOVE:
5543 ret = io_poll_remove_prep(req, sqe);
5545 case IORING_OP_FSYNC:
5546 ret = io_prep_fsync(req, sqe);
5548 case IORING_OP_SYNC_FILE_RANGE:
5549 ret = io_prep_sfr(req, sqe);
5551 case IORING_OP_SENDMSG:
5552 case IORING_OP_SEND:
5553 ret = io_sendmsg_prep(req, sqe);
5555 case IORING_OP_RECVMSG:
5556 case IORING_OP_RECV:
5557 ret = io_recvmsg_prep(req, sqe);
5559 case IORING_OP_CONNECT:
5560 ret = io_connect_prep(req, sqe);
5562 case IORING_OP_TIMEOUT:
5563 ret = io_timeout_prep(req, sqe, false);
5565 case IORING_OP_TIMEOUT_REMOVE:
5566 ret = io_timeout_remove_prep(req, sqe);
5568 case IORING_OP_ASYNC_CANCEL:
5569 ret = io_async_cancel_prep(req, sqe);
5571 case IORING_OP_LINK_TIMEOUT:
5572 ret = io_timeout_prep(req, sqe, true);
5574 case IORING_OP_ACCEPT:
5575 ret = io_accept_prep(req, sqe);
5577 case IORING_OP_FALLOCATE:
5578 ret = io_fallocate_prep(req, sqe);
5580 case IORING_OP_OPENAT:
5581 ret = io_openat_prep(req, sqe);
5583 case IORING_OP_CLOSE:
5584 ret = io_close_prep(req, sqe);
5586 case IORING_OP_FILES_UPDATE:
5587 ret = io_files_update_prep(req, sqe);
5589 case IORING_OP_STATX:
5590 ret = io_statx_prep(req, sqe);
5592 case IORING_OP_FADVISE:
5593 ret = io_fadvise_prep(req, sqe);
5595 case IORING_OP_MADVISE:
5596 ret = io_madvise_prep(req, sqe);
5598 case IORING_OP_OPENAT2:
5599 ret = io_openat2_prep(req, sqe);
5601 case IORING_OP_EPOLL_CTL:
5602 ret = io_epoll_ctl_prep(req, sqe);
5604 case IORING_OP_SPLICE:
5605 ret = io_splice_prep(req, sqe);
5607 case IORING_OP_PROVIDE_BUFFERS:
5608 ret = io_provide_buffers_prep(req, sqe);
5610 case IORING_OP_REMOVE_BUFFERS:
5611 ret = io_remove_buffers_prep(req, sqe);
5614 ret = io_tee_prep(req, sqe);
5617 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5626 static u32 io_get_sequence(struct io_kiocb *req)
5628 struct io_kiocb *pos;
5629 struct io_ring_ctx *ctx = req->ctx;
5630 u32 total_submitted, nr_reqs = 1;
5632 if (req->flags & REQ_F_LINK_HEAD)
5633 list_for_each_entry(pos, &req->link_list, link_list)
5636 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5637 return total_submitted - nr_reqs;
5640 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5642 struct io_ring_ctx *ctx = req->ctx;
5643 struct io_defer_entry *de;
5647 /* Still need defer if there is pending req in defer list. */
5648 if (likely(list_empty_careful(&ctx->defer_list) &&
5649 !(req->flags & REQ_F_IO_DRAIN)))
5652 seq = io_get_sequence(req);
5653 /* Still a chance to pass the sequence check */
5654 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5658 ret = io_req_defer_prep(req, sqe);
5662 io_prep_async_link(req);
5663 de = kmalloc(sizeof(*de), GFP_KERNEL);
5667 spin_lock_irq(&ctx->completion_lock);
5668 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5669 spin_unlock_irq(&ctx->completion_lock);
5671 io_queue_async_work(req);
5672 return -EIOCBQUEUED;
5675 trace_io_uring_defer(ctx, req, req->user_data);
5678 list_add_tail(&de->list, &ctx->defer_list);
5679 spin_unlock_irq(&ctx->completion_lock);
5680 return -EIOCBQUEUED;
5683 static void io_req_drop_files(struct io_kiocb *req)
5685 struct io_ring_ctx *ctx = req->ctx;
5686 unsigned long flags;
5688 spin_lock_irqsave(&ctx->inflight_lock, flags);
5689 list_del(&req->inflight_entry);
5690 if (waitqueue_active(&ctx->inflight_wait))
5691 wake_up(&ctx->inflight_wait);
5692 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5693 req->flags &= ~REQ_F_INFLIGHT;
5694 put_files_struct(req->work.files);
5695 put_nsproxy(req->work.nsproxy);
5696 req->work.files = NULL;
5699 static void __io_clean_op(struct io_kiocb *req)
5701 struct io_async_ctx *io = req->io;
5703 if (req->flags & REQ_F_BUFFER_SELECTED) {
5704 switch (req->opcode) {
5705 case IORING_OP_READV:
5706 case IORING_OP_READ_FIXED:
5707 case IORING_OP_READ:
5708 kfree((void *)(unsigned long)req->rw.addr);
5710 case IORING_OP_RECVMSG:
5711 case IORING_OP_RECV:
5712 kfree(req->sr_msg.kbuf);
5715 req->flags &= ~REQ_F_BUFFER_SELECTED;
5718 if (req->flags & REQ_F_NEED_CLEANUP) {
5719 switch (req->opcode) {
5720 case IORING_OP_READV:
5721 case IORING_OP_READ_FIXED:
5722 case IORING_OP_READ:
5723 case IORING_OP_WRITEV:
5724 case IORING_OP_WRITE_FIXED:
5725 case IORING_OP_WRITE:
5726 if (io->rw.free_iovec)
5727 kfree(io->rw.free_iovec);
5729 case IORING_OP_RECVMSG:
5730 case IORING_OP_SENDMSG:
5731 if (io->msg.iov != io->msg.fast_iov)
5734 case IORING_OP_SPLICE:
5736 io_put_file(req, req->splice.file_in,
5737 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5739 case IORING_OP_OPENAT:
5740 case IORING_OP_OPENAT2:
5741 if (req->open.filename)
5742 putname(req->open.filename);
5745 req->flags &= ~REQ_F_NEED_CLEANUP;
5748 if (req->flags & REQ_F_INFLIGHT)
5749 io_req_drop_files(req);
5752 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5753 bool force_nonblock, struct io_comp_state *cs)
5755 struct io_ring_ctx *ctx = req->ctx;
5758 switch (req->opcode) {
5760 ret = io_nop(req, cs);
5762 case IORING_OP_READV:
5763 case IORING_OP_READ_FIXED:
5764 case IORING_OP_READ:
5766 ret = io_read_prep(req, sqe, force_nonblock);
5770 ret = io_read(req, force_nonblock, cs);
5772 case IORING_OP_WRITEV:
5773 case IORING_OP_WRITE_FIXED:
5774 case IORING_OP_WRITE:
5776 ret = io_write_prep(req, sqe, force_nonblock);
5780 ret = io_write(req, force_nonblock, cs);
5782 case IORING_OP_FSYNC:
5784 ret = io_prep_fsync(req, sqe);
5788 ret = io_fsync(req, force_nonblock);
5790 case IORING_OP_POLL_ADD:
5792 ret = io_poll_add_prep(req, sqe);
5796 ret = io_poll_add(req);
5798 case IORING_OP_POLL_REMOVE:
5800 ret = io_poll_remove_prep(req, sqe);
5804 ret = io_poll_remove(req);
5806 case IORING_OP_SYNC_FILE_RANGE:
5808 ret = io_prep_sfr(req, sqe);
5812 ret = io_sync_file_range(req, force_nonblock);
5814 case IORING_OP_SENDMSG:
5815 case IORING_OP_SEND:
5817 ret = io_sendmsg_prep(req, sqe);
5821 if (req->opcode == IORING_OP_SENDMSG)
5822 ret = io_sendmsg(req, force_nonblock, cs);
5824 ret = io_send(req, force_nonblock, cs);
5826 case IORING_OP_RECVMSG:
5827 case IORING_OP_RECV:
5829 ret = io_recvmsg_prep(req, sqe);
5833 if (req->opcode == IORING_OP_RECVMSG)
5834 ret = io_recvmsg(req, force_nonblock, cs);
5836 ret = io_recv(req, force_nonblock, cs);
5838 case IORING_OP_TIMEOUT:
5840 ret = io_timeout_prep(req, sqe, false);
5844 ret = io_timeout(req);
5846 case IORING_OP_TIMEOUT_REMOVE:
5848 ret = io_timeout_remove_prep(req, sqe);
5852 ret = io_timeout_remove(req);
5854 case IORING_OP_ACCEPT:
5856 ret = io_accept_prep(req, sqe);
5860 ret = io_accept(req, force_nonblock, cs);
5862 case IORING_OP_CONNECT:
5864 ret = io_connect_prep(req, sqe);
5868 ret = io_connect(req, force_nonblock, cs);
5870 case IORING_OP_ASYNC_CANCEL:
5872 ret = io_async_cancel_prep(req, sqe);
5876 ret = io_async_cancel(req);
5878 case IORING_OP_FALLOCATE:
5880 ret = io_fallocate_prep(req, sqe);
5884 ret = io_fallocate(req, force_nonblock);
5886 case IORING_OP_OPENAT:
5888 ret = io_openat_prep(req, sqe);
5892 ret = io_openat(req, force_nonblock);
5894 case IORING_OP_CLOSE:
5896 ret = io_close_prep(req, sqe);
5900 ret = io_close(req, force_nonblock, cs);
5902 case IORING_OP_FILES_UPDATE:
5904 ret = io_files_update_prep(req, sqe);
5908 ret = io_files_update(req, force_nonblock, cs);
5910 case IORING_OP_STATX:
5912 ret = io_statx_prep(req, sqe);
5916 ret = io_statx(req, force_nonblock);
5918 case IORING_OP_FADVISE:
5920 ret = io_fadvise_prep(req, sqe);
5924 ret = io_fadvise(req, force_nonblock);
5926 case IORING_OP_MADVISE:
5928 ret = io_madvise_prep(req, sqe);
5932 ret = io_madvise(req, force_nonblock);
5934 case IORING_OP_OPENAT2:
5936 ret = io_openat2_prep(req, sqe);
5940 ret = io_openat2(req, force_nonblock);
5942 case IORING_OP_EPOLL_CTL:
5944 ret = io_epoll_ctl_prep(req, sqe);
5948 ret = io_epoll_ctl(req, force_nonblock, cs);
5950 case IORING_OP_SPLICE:
5952 ret = io_splice_prep(req, sqe);
5956 ret = io_splice(req, force_nonblock);
5958 case IORING_OP_PROVIDE_BUFFERS:
5960 ret = io_provide_buffers_prep(req, sqe);
5964 ret = io_provide_buffers(req, force_nonblock, cs);
5966 case IORING_OP_REMOVE_BUFFERS:
5968 ret = io_remove_buffers_prep(req, sqe);
5972 ret = io_remove_buffers(req, force_nonblock, cs);
5976 ret = io_tee_prep(req, sqe);
5980 ret = io_tee(req, force_nonblock);
5990 /* If the op doesn't have a file, we're not polling for it */
5991 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5992 const bool in_async = io_wq_current_is_worker();
5994 /* workqueue context doesn't hold uring_lock, grab it now */
5996 mutex_lock(&ctx->uring_lock);
5998 io_iopoll_req_issued(req);
6001 mutex_unlock(&ctx->uring_lock);
6007 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6009 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6010 struct io_kiocb *timeout;
6013 timeout = io_prep_linked_timeout(req);
6015 io_queue_linked_timeout(timeout);
6017 /* if NO_CANCEL is set, we must still run the work */
6018 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6019 IO_WQ_WORK_CANCEL) {
6025 ret = io_issue_sqe(req, NULL, false, NULL);
6027 * We can get EAGAIN for polled IO even though we're
6028 * forcing a sync submission from here, since we can't
6029 * wait for request slots on the block side.
6038 req_set_fail_links(req);
6039 io_req_complete(req, ret);
6042 return io_steal_work(req);
6045 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6048 struct fixed_file_table *table;
6050 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6051 return table->files[index & IORING_FILE_TABLE_MASK];
6054 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
6055 int fd, struct file **out_file, bool fixed)
6057 struct io_ring_ctx *ctx = req->ctx;
6061 if (unlikely(!ctx->file_data ||
6062 (unsigned) fd >= ctx->nr_user_files))
6064 fd = array_index_nospec(fd, ctx->nr_user_files);
6065 file = io_file_from_index(ctx, fd);
6067 req->fixed_file_refs = ctx->file_data->cur_refs;
6068 percpu_ref_get(req->fixed_file_refs);
6071 trace_io_uring_file_get(ctx, fd);
6072 file = __io_file_get(state, fd);
6075 if (file || io_op_defs[req->opcode].needs_file_no_error) {
6082 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6087 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6088 if (unlikely(!fixed && io_async_submit(req->ctx)))
6091 return io_file_get(state, req, fd, &req->file, fixed);
6094 static int io_grab_files(struct io_kiocb *req)
6096 struct io_ring_ctx *ctx = req->ctx;
6098 io_req_init_async(req);
6100 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
6103 req->work.files = get_files_struct(current);
6104 get_nsproxy(current->nsproxy);
6105 req->work.nsproxy = current->nsproxy;
6106 req->flags |= REQ_F_INFLIGHT;
6108 spin_lock_irq(&ctx->inflight_lock);
6109 list_add(&req->inflight_entry, &ctx->inflight_list);
6110 spin_unlock_irq(&ctx->inflight_lock);
6114 static inline int io_prep_work_files(struct io_kiocb *req)
6116 if (!io_op_defs[req->opcode].file_table)
6118 return io_grab_files(req);
6121 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6123 struct io_timeout_data *data = container_of(timer,
6124 struct io_timeout_data, timer);
6125 struct io_kiocb *req = data->req;
6126 struct io_ring_ctx *ctx = req->ctx;
6127 struct io_kiocb *prev = NULL;
6128 unsigned long flags;
6130 spin_lock_irqsave(&ctx->completion_lock, flags);
6133 * We don't expect the list to be empty, that will only happen if we
6134 * race with the completion of the linked work.
6136 if (!list_empty(&req->link_list)) {
6137 prev = list_entry(req->link_list.prev, struct io_kiocb,
6139 if (refcount_inc_not_zero(&prev->refs)) {
6140 list_del_init(&req->link_list);
6141 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6146 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6149 req_set_fail_links(prev);
6150 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6153 io_req_complete(req, -ETIME);
6155 return HRTIMER_NORESTART;
6158 static void __io_queue_linked_timeout(struct io_kiocb *req)
6161 * If the list is now empty, then our linked request finished before
6162 * we got a chance to setup the timer
6164 if (!list_empty(&req->link_list)) {
6165 struct io_timeout_data *data = &req->io->timeout;
6167 data->timer.function = io_link_timeout_fn;
6168 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6173 static void io_queue_linked_timeout(struct io_kiocb *req)
6175 struct io_ring_ctx *ctx = req->ctx;
6177 spin_lock_irq(&ctx->completion_lock);
6178 __io_queue_linked_timeout(req);
6179 spin_unlock_irq(&ctx->completion_lock);
6181 /* drop submission reference */
6185 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6187 struct io_kiocb *nxt;
6189 if (!(req->flags & REQ_F_LINK_HEAD))
6191 if (req->flags & REQ_F_LINK_TIMEOUT)
6194 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6196 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6199 req->flags |= REQ_F_LINK_TIMEOUT;
6203 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6204 struct io_comp_state *cs)
6206 struct io_kiocb *linked_timeout;
6207 struct io_kiocb *nxt;
6208 const struct cred *old_creds = NULL;
6212 linked_timeout = io_prep_linked_timeout(req);
6214 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6215 req->work.creds != current_cred()) {
6217 revert_creds(old_creds);
6218 if (old_creds == req->work.creds)
6219 old_creds = NULL; /* restored original creds */
6221 old_creds = override_creds(req->work.creds);
6224 ret = io_issue_sqe(req, sqe, true, cs);
6227 * We async punt it if the file wasn't marked NOWAIT, or if the file
6228 * doesn't support non-blocking read/write attempts
6230 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6231 if (!io_arm_poll_handler(req)) {
6233 ret = io_prep_work_files(req);
6237 * Queued up for async execution, worker will release
6238 * submit reference when the iocb is actually submitted.
6240 io_queue_async_work(req);
6244 io_queue_linked_timeout(linked_timeout);
6248 if (unlikely(ret)) {
6250 /* un-prep timeout, so it'll be killed as any other linked */
6251 req->flags &= ~REQ_F_LINK_TIMEOUT;
6252 req_set_fail_links(req);
6254 io_req_complete(req, ret);
6258 /* drop submission reference */
6259 nxt = io_put_req_find_next(req);
6261 io_queue_linked_timeout(linked_timeout);
6266 if (req->flags & REQ_F_FORCE_ASYNC)
6272 revert_creds(old_creds);
6275 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6276 struct io_comp_state *cs)
6280 ret = io_req_defer(req, sqe);
6282 if (ret != -EIOCBQUEUED) {
6284 req_set_fail_links(req);
6286 io_req_complete(req, ret);
6288 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6290 ret = io_req_defer_prep(req, sqe);
6296 * Never try inline submit of IOSQE_ASYNC is set, go straight
6297 * to async execution.
6299 io_req_init_async(req);
6300 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6301 io_queue_async_work(req);
6303 __io_queue_sqe(req, sqe, cs);
6307 static inline void io_queue_link_head(struct io_kiocb *req,
6308 struct io_comp_state *cs)
6310 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6312 io_req_complete(req, -ECANCELED);
6314 io_queue_sqe(req, NULL, cs);
6317 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6318 struct io_kiocb **link, struct io_comp_state *cs)
6320 struct io_ring_ctx *ctx = req->ctx;
6324 * If we already have a head request, queue this one for async
6325 * submittal once the head completes. If we don't have a head but
6326 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6327 * submitted sync once the chain is complete. If none of those
6328 * conditions are true (normal request), then just queue it.
6331 struct io_kiocb *head = *link;
6334 * Taking sequential execution of a link, draining both sides
6335 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6336 * requests in the link. So, it drains the head and the
6337 * next after the link request. The last one is done via
6338 * drain_next flag to persist the effect across calls.
6340 if (req->flags & REQ_F_IO_DRAIN) {
6341 head->flags |= REQ_F_IO_DRAIN;
6342 ctx->drain_next = 1;
6344 ret = io_req_defer_prep(req, sqe);
6345 if (unlikely(ret)) {
6346 /* fail even hard links since we don't submit */
6347 head->flags |= REQ_F_FAIL_LINK;
6350 trace_io_uring_link(ctx, req, head);
6351 list_add_tail(&req->link_list, &head->link_list);
6353 /* last request of a link, enqueue the link */
6354 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6355 io_queue_link_head(head, cs);
6359 if (unlikely(ctx->drain_next)) {
6360 req->flags |= REQ_F_IO_DRAIN;
6361 ctx->drain_next = 0;
6363 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6364 req->flags |= REQ_F_LINK_HEAD;
6365 INIT_LIST_HEAD(&req->link_list);
6367 ret = io_req_defer_prep(req, sqe);
6369 req->flags |= REQ_F_FAIL_LINK;
6372 io_queue_sqe(req, sqe, cs);
6380 * Batched submission is done, ensure local IO is flushed out.
6382 static void io_submit_state_end(struct io_submit_state *state)
6384 if (!list_empty(&state->comp.list))
6385 io_submit_flush_completions(&state->comp);
6386 blk_finish_plug(&state->plug);
6387 io_state_file_put(state);
6388 if (state->free_reqs)
6389 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6393 * Start submission side cache.
6395 static void io_submit_state_start(struct io_submit_state *state,
6396 struct io_ring_ctx *ctx, unsigned int max_ios)
6398 blk_start_plug(&state->plug);
6400 INIT_LIST_HEAD(&state->comp.list);
6401 state->comp.ctx = ctx;
6402 state->free_reqs = 0;
6404 state->ios_left = max_ios;
6407 static void io_commit_sqring(struct io_ring_ctx *ctx)
6409 struct io_rings *rings = ctx->rings;
6412 * Ensure any loads from the SQEs are done at this point,
6413 * since once we write the new head, the application could
6414 * write new data to them.
6416 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6420 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6421 * that is mapped by userspace. This means that care needs to be taken to
6422 * ensure that reads are stable, as we cannot rely on userspace always
6423 * being a good citizen. If members of the sqe are validated and then later
6424 * used, it's important that those reads are done through READ_ONCE() to
6425 * prevent a re-load down the line.
6427 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6429 u32 *sq_array = ctx->sq_array;
6433 * The cached sq head (or cq tail) serves two purposes:
6435 * 1) allows us to batch the cost of updating the user visible
6437 * 2) allows the kernel side to track the head on its own, even
6438 * though the application is the one updating it.
6440 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6441 if (likely(head < ctx->sq_entries))
6442 return &ctx->sq_sqes[head];
6444 /* drop invalid entries */
6445 ctx->cached_sq_dropped++;
6446 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6450 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6452 ctx->cached_sq_head++;
6456 * Check SQE restrictions (opcode and flags).
6458 * Returns 'true' if SQE is allowed, 'false' otherwise.
6460 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6461 struct io_kiocb *req,
6462 unsigned int sqe_flags)
6464 if (!ctx->restricted)
6467 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6470 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6471 ctx->restrictions.sqe_flags_required)
6474 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6475 ctx->restrictions.sqe_flags_required))
6481 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6482 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6483 IOSQE_BUFFER_SELECT)
6485 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6486 const struct io_uring_sqe *sqe,
6487 struct io_submit_state *state)
6489 unsigned int sqe_flags;
6492 req->opcode = READ_ONCE(sqe->opcode);
6493 req->user_data = READ_ONCE(sqe->user_data);
6498 /* one is dropped after submission, the other at completion */
6499 refcount_set(&req->refs, 2);
6500 req->task = current;
6501 get_task_struct(req->task);
6502 atomic_long_inc(&req->task->io_uring->req_issue);
6505 if (unlikely(req->opcode >= IORING_OP_LAST))
6508 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6511 sqe_flags = READ_ONCE(sqe->flags);
6512 /* enforce forwards compatibility on users */
6513 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6516 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6519 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6520 !io_op_defs[req->opcode].buffer_select)
6523 id = READ_ONCE(sqe->personality);
6525 io_req_init_async(req);
6526 req->work.creds = idr_find(&ctx->personality_idr, id);
6527 if (unlikely(!req->work.creds))
6529 get_cred(req->work.creds);
6532 /* same numerical values with corresponding REQ_F_*, safe to copy */
6533 req->flags |= sqe_flags;
6535 if (!io_op_defs[req->opcode].needs_file)
6538 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6541 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6543 struct io_submit_state state;
6544 struct io_kiocb *link = NULL;
6545 int i, submitted = 0;
6547 /* if we have a backlog and couldn't flush it all, return BUSY */
6548 if (test_bit(0, &ctx->sq_check_overflow)) {
6549 if (!list_empty(&ctx->cq_overflow_list) &&
6550 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6554 /* make sure SQ entry isn't read before tail */
6555 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6557 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6560 io_submit_state_start(&state, ctx, nr);
6562 for (i = 0; i < nr; i++) {
6563 const struct io_uring_sqe *sqe;
6564 struct io_kiocb *req;
6567 sqe = io_get_sqe(ctx);
6568 if (unlikely(!sqe)) {
6569 io_consume_sqe(ctx);
6572 req = io_alloc_req(ctx, &state);
6573 if (unlikely(!req)) {
6575 submitted = -EAGAIN;
6579 err = io_init_req(ctx, req, sqe, &state);
6580 io_consume_sqe(ctx);
6581 /* will complete beyond this point, count as submitted */
6584 if (unlikely(err)) {
6587 io_req_complete(req, err);
6591 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6592 true, io_async_submit(ctx));
6593 err = io_submit_sqe(req, sqe, &link, &state.comp);
6598 if (unlikely(submitted != nr)) {
6599 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6601 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6604 io_queue_link_head(link, &state.comp);
6605 io_submit_state_end(&state);
6607 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6608 io_commit_sqring(ctx);
6613 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6615 /* Tell userspace we may need a wakeup call */
6616 spin_lock_irq(&ctx->completion_lock);
6617 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6618 spin_unlock_irq(&ctx->completion_lock);
6621 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6623 spin_lock_irq(&ctx->completion_lock);
6624 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6625 spin_unlock_irq(&ctx->completion_lock);
6628 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6629 int sync, void *key)
6631 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6634 ret = autoremove_wake_function(wqe, mode, sync, key);
6636 unsigned long flags;
6638 spin_lock_irqsave(&ctx->completion_lock, flags);
6639 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6640 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6651 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6652 unsigned long start_jiffies)
6654 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6655 unsigned int to_submit;
6659 if (!list_empty(&ctx->iopoll_list)) {
6660 unsigned nr_events = 0;
6662 mutex_lock(&ctx->uring_lock);
6663 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6664 io_do_iopoll(ctx, &nr_events, 0);
6665 mutex_unlock(&ctx->uring_lock);
6668 to_submit = io_sqring_entries(ctx);
6671 * If submit got -EBUSY, flag us as needing the application
6672 * to enter the kernel to reap and flush events.
6674 if (!to_submit || ret == -EBUSY || need_resched()) {
6676 * Drop cur_mm before scheduling, we can't hold it for
6677 * long periods (or over schedule()). Do this before
6678 * adding ourselves to the waitqueue, as the unuse/drop
6681 io_sq_thread_drop_mm();
6684 * We're polling. If we're within the defined idle
6685 * period, then let us spin without work before going
6686 * to sleep. The exception is if we got EBUSY doing
6687 * more IO, we should wait for the application to
6688 * reap events and wake us up.
6690 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6691 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6692 !percpu_ref_is_dying(&ctx->refs)))
6695 prepare_to_wait(ctx->sqo_wait, &ctx->sqo_wait_entry,
6696 TASK_INTERRUPTIBLE);
6699 * While doing polled IO, before going to sleep, we need
6700 * to check if there are new reqs added to iopoll_list,
6701 * it is because reqs may have been punted to io worker
6702 * and will be added to iopoll_list later, hence check
6703 * the iopoll_list again.
6705 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6706 !list_empty_careful(&ctx->iopoll_list)) {
6707 finish_wait(ctx->sqo_wait, &ctx->sqo_wait_entry);
6711 io_ring_set_wakeup_flag(ctx);
6713 to_submit = io_sqring_entries(ctx);
6714 if (!to_submit || ret == -EBUSY)
6718 finish_wait(ctx->sqo_wait, &ctx->sqo_wait_entry);
6719 io_ring_clear_wakeup_flag(ctx);
6721 mutex_lock(&ctx->uring_lock);
6722 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6723 ret = io_submit_sqes(ctx, to_submit);
6724 mutex_unlock(&ctx->uring_lock);
6725 return SQT_DID_WORK;
6728 static int io_sq_thread(void *data)
6730 struct io_ring_ctx *ctx = data;
6731 const struct cred *old_cred;
6732 unsigned long start_jiffies;
6734 init_wait(&ctx->sqo_wait_entry);
6735 ctx->sqo_wait_entry.func = io_sq_wake_function;
6737 complete(&ctx->sq_thread_comp);
6739 old_cred = override_creds(ctx->creds);
6741 start_jiffies = jiffies;
6742 while (!kthread_should_park()) {
6745 ret = __io_sq_thread(ctx, start_jiffies);
6749 start_jiffies = jiffies;
6762 io_sq_thread_drop_mm();
6763 revert_creds(old_cred);
6770 struct io_wait_queue {
6771 struct wait_queue_entry wq;
6772 struct io_ring_ctx *ctx;
6774 unsigned nr_timeouts;
6777 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6779 struct io_ring_ctx *ctx = iowq->ctx;
6782 * Wake up if we have enough events, or if a timeout occurred since we
6783 * started waiting. For timeouts, we always want to return to userspace,
6784 * regardless of event count.
6786 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6787 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6790 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6791 int wake_flags, void *key)
6793 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6796 /* use noflush == true, as we can't safely rely on locking context */
6797 if (!io_should_wake(iowq, true))
6800 return autoremove_wake_function(curr, mode, wake_flags, key);
6804 * Wait until events become available, if we don't already have some. The
6805 * application must reap them itself, as they reside on the shared cq ring.
6807 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6808 const sigset_t __user *sig, size_t sigsz)
6810 struct io_wait_queue iowq = {
6813 .func = io_wake_function,
6814 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6817 .to_wait = min_events,
6819 struct io_rings *rings = ctx->rings;
6823 if (io_cqring_events(ctx, false) >= min_events)
6825 if (!io_run_task_work())
6830 #ifdef CONFIG_COMPAT
6831 if (in_compat_syscall())
6832 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6836 ret = set_user_sigmask(sig, sigsz);
6842 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6843 trace_io_uring_cqring_wait(ctx, min_events);
6845 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6846 TASK_INTERRUPTIBLE);
6847 /* make sure we run task_work before checking for signals */
6848 if (io_run_task_work())
6850 if (signal_pending(current)) {
6851 if (current->jobctl & JOBCTL_TASK_WORK) {
6852 spin_lock_irq(¤t->sighand->siglock);
6853 current->jobctl &= ~JOBCTL_TASK_WORK;
6854 recalc_sigpending();
6855 spin_unlock_irq(¤t->sighand->siglock);
6861 if (io_should_wake(&iowq, false))
6865 finish_wait(&ctx->wait, &iowq.wq);
6867 restore_saved_sigmask_unless(ret == -EINTR);
6869 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6872 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6874 #if defined(CONFIG_UNIX)
6875 if (ctx->ring_sock) {
6876 struct sock *sock = ctx->ring_sock->sk;
6877 struct sk_buff *skb;
6879 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6885 for (i = 0; i < ctx->nr_user_files; i++) {
6888 file = io_file_from_index(ctx, i);
6895 static void io_file_ref_kill(struct percpu_ref *ref)
6897 struct fixed_file_data *data;
6899 data = container_of(ref, struct fixed_file_data, refs);
6900 complete(&data->done);
6903 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6905 struct fixed_file_data *data = ctx->file_data;
6906 struct fixed_file_ref_node *ref_node = NULL;
6907 unsigned nr_tables, i;
6912 spin_lock(&data->lock);
6913 if (!list_empty(&data->ref_list))
6914 ref_node = list_first_entry(&data->ref_list,
6915 struct fixed_file_ref_node, node);
6916 spin_unlock(&data->lock);
6918 percpu_ref_kill(&ref_node->refs);
6920 percpu_ref_kill(&data->refs);
6922 /* wait for all refs nodes to complete */
6923 flush_delayed_work(&ctx->file_put_work);
6924 wait_for_completion(&data->done);
6926 __io_sqe_files_unregister(ctx);
6927 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6928 for (i = 0; i < nr_tables; i++)
6929 kfree(data->table[i].files);
6931 percpu_ref_exit(&data->refs);
6933 ctx->file_data = NULL;
6934 ctx->nr_user_files = 0;
6938 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6940 if (ctx->sqo_thread) {
6942 * We may arrive here from the error branch in
6943 * io_sq_offload_create() where the kthread is created
6944 * without being waked up, thus wake it up now to make
6945 * sure the wait will complete.
6947 wake_up_process(ctx->sqo_thread);
6949 wait_for_completion(&ctx->sq_thread_comp);
6951 * The park is a bit of a work-around, without it we get
6952 * warning spews on shutdown with SQPOLL set and affinity
6953 * set to a single CPU.
6955 kthread_park(ctx->sqo_thread);
6956 kthread_stop(ctx->sqo_thread);
6957 ctx->sqo_thread = NULL;
6961 static void io_finish_async(struct io_ring_ctx *ctx)
6963 io_sq_thread_stop(ctx);
6966 io_wq_destroy(ctx->io_wq);
6971 #if defined(CONFIG_UNIX)
6973 * Ensure the UNIX gc is aware of our file set, so we are certain that
6974 * the io_uring can be safely unregistered on process exit, even if we have
6975 * loops in the file referencing.
6977 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6979 struct sock *sk = ctx->ring_sock->sk;
6980 struct scm_fp_list *fpl;
6981 struct sk_buff *skb;
6984 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6988 skb = alloc_skb(0, GFP_KERNEL);
6997 fpl->user = get_uid(ctx->user);
6998 for (i = 0; i < nr; i++) {
6999 struct file *file = io_file_from_index(ctx, i + offset);
7003 fpl->fp[nr_files] = get_file(file);
7004 unix_inflight(fpl->user, fpl->fp[nr_files]);
7009 fpl->max = SCM_MAX_FD;
7010 fpl->count = nr_files;
7011 UNIXCB(skb).fp = fpl;
7012 skb->destructor = unix_destruct_scm;
7013 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7014 skb_queue_head(&sk->sk_receive_queue, skb);
7016 for (i = 0; i < nr_files; i++)
7027 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7028 * causes regular reference counting to break down. We rely on the UNIX
7029 * garbage collection to take care of this problem for us.
7031 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7033 unsigned left, total;
7037 left = ctx->nr_user_files;
7039 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7041 ret = __io_sqe_files_scm(ctx, this_files, total);
7045 total += this_files;
7051 while (total < ctx->nr_user_files) {
7052 struct file *file = io_file_from_index(ctx, total);
7062 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7068 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
7073 for (i = 0; i < nr_tables; i++) {
7074 struct fixed_file_table *table = &ctx->file_data->table[i];
7075 unsigned this_files;
7077 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7078 table->files = kcalloc(this_files, sizeof(struct file *),
7082 nr_files -= this_files;
7088 for (i = 0; i < nr_tables; i++) {
7089 struct fixed_file_table *table = &ctx->file_data->table[i];
7090 kfree(table->files);
7095 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7097 #if defined(CONFIG_UNIX)
7098 struct sock *sock = ctx->ring_sock->sk;
7099 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7100 struct sk_buff *skb;
7103 __skb_queue_head_init(&list);
7106 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7107 * remove this entry and rearrange the file array.
7109 skb = skb_dequeue(head);
7111 struct scm_fp_list *fp;
7113 fp = UNIXCB(skb).fp;
7114 for (i = 0; i < fp->count; i++) {
7117 if (fp->fp[i] != file)
7120 unix_notinflight(fp->user, fp->fp[i]);
7121 left = fp->count - 1 - i;
7123 memmove(&fp->fp[i], &fp->fp[i + 1],
7124 left * sizeof(struct file *));
7131 __skb_queue_tail(&list, skb);
7141 __skb_queue_tail(&list, skb);
7143 skb = skb_dequeue(head);
7146 if (skb_peek(&list)) {
7147 spin_lock_irq(&head->lock);
7148 while ((skb = __skb_dequeue(&list)) != NULL)
7149 __skb_queue_tail(head, skb);
7150 spin_unlock_irq(&head->lock);
7157 struct io_file_put {
7158 struct list_head list;
7162 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7164 struct fixed_file_data *file_data = ref_node->file_data;
7165 struct io_ring_ctx *ctx = file_data->ctx;
7166 struct io_file_put *pfile, *tmp;
7168 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7169 list_del(&pfile->list);
7170 io_ring_file_put(ctx, pfile->file);
7174 spin_lock(&file_data->lock);
7175 list_del(&ref_node->node);
7176 spin_unlock(&file_data->lock);
7178 percpu_ref_exit(&ref_node->refs);
7180 percpu_ref_put(&file_data->refs);
7183 static void io_file_put_work(struct work_struct *work)
7185 struct io_ring_ctx *ctx;
7186 struct llist_node *node;
7188 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7189 node = llist_del_all(&ctx->file_put_llist);
7192 struct fixed_file_ref_node *ref_node;
7193 struct llist_node *next = node->next;
7195 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7196 __io_file_put_work(ref_node);
7201 static void io_file_data_ref_zero(struct percpu_ref *ref)
7203 struct fixed_file_ref_node *ref_node;
7204 struct io_ring_ctx *ctx;
7208 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7209 ctx = ref_node->file_data->ctx;
7211 if (percpu_ref_is_dying(&ctx->file_data->refs))
7214 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7216 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7218 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7221 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7222 struct io_ring_ctx *ctx)
7224 struct fixed_file_ref_node *ref_node;
7226 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7228 return ERR_PTR(-ENOMEM);
7230 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7233 return ERR_PTR(-ENOMEM);
7235 INIT_LIST_HEAD(&ref_node->node);
7236 INIT_LIST_HEAD(&ref_node->file_list);
7237 ref_node->file_data = ctx->file_data;
7241 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7243 percpu_ref_exit(&ref_node->refs);
7247 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7250 __s32 __user *fds = (__s32 __user *) arg;
7255 struct fixed_file_ref_node *ref_node;
7261 if (nr_args > IORING_MAX_FIXED_FILES)
7264 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7265 if (!ctx->file_data)
7267 ctx->file_data->ctx = ctx;
7268 init_completion(&ctx->file_data->done);
7269 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7270 spin_lock_init(&ctx->file_data->lock);
7272 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7273 ctx->file_data->table = kcalloc(nr_tables,
7274 sizeof(struct fixed_file_table),
7276 if (!ctx->file_data->table) {
7277 kfree(ctx->file_data);
7278 ctx->file_data = NULL;
7282 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7283 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7284 kfree(ctx->file_data->table);
7285 kfree(ctx->file_data);
7286 ctx->file_data = NULL;
7290 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7291 percpu_ref_exit(&ctx->file_data->refs);
7292 kfree(ctx->file_data->table);
7293 kfree(ctx->file_data);
7294 ctx->file_data = NULL;
7298 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7299 struct fixed_file_table *table;
7303 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7305 /* allow sparse sets */
7311 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7312 index = i & IORING_FILE_TABLE_MASK;
7320 * Don't allow io_uring instances to be registered. If UNIX
7321 * isn't enabled, then this causes a reference cycle and this
7322 * instance can never get freed. If UNIX is enabled we'll
7323 * handle it just fine, but there's still no point in allowing
7324 * a ring fd as it doesn't support regular read/write anyway.
7326 if (file->f_op == &io_uring_fops) {
7331 table->files[index] = file;
7335 for (i = 0; i < ctx->nr_user_files; i++) {
7336 file = io_file_from_index(ctx, i);
7340 for (i = 0; i < nr_tables; i++)
7341 kfree(ctx->file_data->table[i].files);
7343 percpu_ref_exit(&ctx->file_data->refs);
7344 kfree(ctx->file_data->table);
7345 kfree(ctx->file_data);
7346 ctx->file_data = NULL;
7347 ctx->nr_user_files = 0;
7351 ret = io_sqe_files_scm(ctx);
7353 io_sqe_files_unregister(ctx);
7357 ref_node = alloc_fixed_file_ref_node(ctx);
7358 if (IS_ERR(ref_node)) {
7359 io_sqe_files_unregister(ctx);
7360 return PTR_ERR(ref_node);
7363 ctx->file_data->cur_refs = &ref_node->refs;
7364 spin_lock(&ctx->file_data->lock);
7365 list_add(&ref_node->node, &ctx->file_data->ref_list);
7366 spin_unlock(&ctx->file_data->lock);
7367 percpu_ref_get(&ctx->file_data->refs);
7371 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7374 #if defined(CONFIG_UNIX)
7375 struct sock *sock = ctx->ring_sock->sk;
7376 struct sk_buff_head *head = &sock->sk_receive_queue;
7377 struct sk_buff *skb;
7380 * See if we can merge this file into an existing skb SCM_RIGHTS
7381 * file set. If there's no room, fall back to allocating a new skb
7382 * and filling it in.
7384 spin_lock_irq(&head->lock);
7385 skb = skb_peek(head);
7387 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7389 if (fpl->count < SCM_MAX_FD) {
7390 __skb_unlink(skb, head);
7391 spin_unlock_irq(&head->lock);
7392 fpl->fp[fpl->count] = get_file(file);
7393 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7395 spin_lock_irq(&head->lock);
7396 __skb_queue_head(head, skb);
7401 spin_unlock_irq(&head->lock);
7408 return __io_sqe_files_scm(ctx, 1, index);
7414 static int io_queue_file_removal(struct fixed_file_data *data,
7417 struct io_file_put *pfile;
7418 struct percpu_ref *refs = data->cur_refs;
7419 struct fixed_file_ref_node *ref_node;
7421 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7425 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7427 list_add(&pfile->list, &ref_node->file_list);
7432 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7433 struct io_uring_files_update *up,
7436 struct fixed_file_data *data = ctx->file_data;
7437 struct fixed_file_ref_node *ref_node;
7442 bool needs_switch = false;
7444 if (check_add_overflow(up->offset, nr_args, &done))
7446 if (done > ctx->nr_user_files)
7449 ref_node = alloc_fixed_file_ref_node(ctx);
7450 if (IS_ERR(ref_node))
7451 return PTR_ERR(ref_node);
7454 fds = u64_to_user_ptr(up->fds);
7456 struct fixed_file_table *table;
7460 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7464 i = array_index_nospec(up->offset, ctx->nr_user_files);
7465 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7466 index = i & IORING_FILE_TABLE_MASK;
7467 if (table->files[index]) {
7468 file = table->files[index];
7469 err = io_queue_file_removal(data, file);
7472 table->files[index] = NULL;
7473 needs_switch = true;
7482 * Don't allow io_uring instances to be registered. If
7483 * UNIX isn't enabled, then this causes a reference
7484 * cycle and this instance can never get freed. If UNIX
7485 * is enabled we'll handle it just fine, but there's
7486 * still no point in allowing a ring fd as it doesn't
7487 * support regular read/write anyway.
7489 if (file->f_op == &io_uring_fops) {
7494 table->files[index] = file;
7495 err = io_sqe_file_register(ctx, file, i);
7497 table->files[index] = NULL;
7508 percpu_ref_kill(data->cur_refs);
7509 spin_lock(&data->lock);
7510 list_add(&ref_node->node, &data->ref_list);
7511 data->cur_refs = &ref_node->refs;
7512 spin_unlock(&data->lock);
7513 percpu_ref_get(&ctx->file_data->refs);
7515 destroy_fixed_file_ref_node(ref_node);
7517 return done ? done : err;
7520 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7523 struct io_uring_files_update up;
7525 if (!ctx->file_data)
7529 if (copy_from_user(&up, arg, sizeof(up)))
7534 return __io_sqe_files_update(ctx, &up, nr_args);
7537 static void io_free_work(struct io_wq_work *work)
7539 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7541 /* Consider that io_steal_work() relies on this ref */
7545 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7546 struct io_uring_params *p)
7548 struct io_wq_data data;
7550 struct io_ring_ctx *ctx_attach;
7551 unsigned int concurrency;
7554 data.user = ctx->user;
7555 data.free_work = io_free_work;
7556 data.do_work = io_wq_submit_work;
7558 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7559 /* Do QD, or 4 * CPUS, whatever is smallest */
7560 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7562 ctx->io_wq = io_wq_create(concurrency, &data);
7563 if (IS_ERR(ctx->io_wq)) {
7564 ret = PTR_ERR(ctx->io_wq);
7570 f = fdget(p->wq_fd);
7574 if (f.file->f_op != &io_uring_fops) {
7579 ctx_attach = f.file->private_data;
7580 /* @io_wq is protected by holding the fd */
7581 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7586 ctx->io_wq = ctx_attach->io_wq;
7592 static int io_uring_alloc_task_context(struct task_struct *task)
7594 struct io_uring_task *tctx;
7596 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7597 if (unlikely(!tctx))
7601 init_waitqueue_head(&tctx->wait);
7604 atomic_long_set(&tctx->req_issue, 0);
7605 atomic_long_set(&tctx->req_complete, 0);
7606 task->io_uring = tctx;
7610 void __io_uring_free(struct task_struct *tsk)
7612 struct io_uring_task *tctx = tsk->io_uring;
7614 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7615 xa_destroy(&tctx->xa);
7617 tsk->io_uring = NULL;
7620 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7621 struct io_uring_params *p)
7625 if (ctx->flags & IORING_SETUP_SQPOLL) {
7627 if (!capable(CAP_SYS_ADMIN))
7630 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7631 if (!ctx->sq_thread_idle)
7632 ctx->sq_thread_idle = HZ;
7634 if (p->flags & IORING_SETUP_SQ_AFF) {
7635 int cpu = p->sq_thread_cpu;
7638 if (cpu >= nr_cpu_ids)
7640 if (!cpu_online(cpu))
7643 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7647 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7650 if (IS_ERR(ctx->sqo_thread)) {
7651 ret = PTR_ERR(ctx->sqo_thread);
7652 ctx->sqo_thread = NULL;
7655 ret = io_uring_alloc_task_context(ctx->sqo_thread);
7658 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7659 /* Can't have SQ_AFF without SQPOLL */
7664 ret = io_init_wq_offload(ctx, p);
7670 io_finish_async(ctx);
7674 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7676 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sqo_thread)
7677 wake_up_process(ctx->sqo_thread);
7680 static inline void __io_unaccount_mem(struct user_struct *user,
7681 unsigned long nr_pages)
7683 atomic_long_sub(nr_pages, &user->locked_vm);
7686 static inline int __io_account_mem(struct user_struct *user,
7687 unsigned long nr_pages)
7689 unsigned long page_limit, cur_pages, new_pages;
7691 /* Don't allow more pages than we can safely lock */
7692 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7695 cur_pages = atomic_long_read(&user->locked_vm);
7696 new_pages = cur_pages + nr_pages;
7697 if (new_pages > page_limit)
7699 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7700 new_pages) != cur_pages);
7705 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7706 enum io_mem_account acct)
7709 __io_unaccount_mem(ctx->user, nr_pages);
7711 if (ctx->mm_account) {
7712 if (acct == ACCT_LOCKED)
7713 ctx->mm_account->locked_vm -= nr_pages;
7714 else if (acct == ACCT_PINNED)
7715 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7719 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7720 enum io_mem_account acct)
7724 if (ctx->limit_mem) {
7725 ret = __io_account_mem(ctx->user, nr_pages);
7730 if (ctx->mm_account) {
7731 if (acct == ACCT_LOCKED)
7732 ctx->mm_account->locked_vm += nr_pages;
7733 else if (acct == ACCT_PINNED)
7734 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7740 static void io_mem_free(void *ptr)
7747 page = virt_to_head_page(ptr);
7748 if (put_page_testzero(page))
7749 free_compound_page(page);
7752 static void *io_mem_alloc(size_t size)
7754 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7757 return (void *) __get_free_pages(gfp_flags, get_order(size));
7760 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7763 struct io_rings *rings;
7764 size_t off, sq_array_size;
7766 off = struct_size(rings, cqes, cq_entries);
7767 if (off == SIZE_MAX)
7771 off = ALIGN(off, SMP_CACHE_BYTES);
7779 sq_array_size = array_size(sizeof(u32), sq_entries);
7780 if (sq_array_size == SIZE_MAX)
7783 if (check_add_overflow(off, sq_array_size, &off))
7789 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7793 pages = (size_t)1 << get_order(
7794 rings_size(sq_entries, cq_entries, NULL));
7795 pages += (size_t)1 << get_order(
7796 array_size(sizeof(struct io_uring_sqe), sq_entries));
7801 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7805 if (!ctx->user_bufs)
7808 for (i = 0; i < ctx->nr_user_bufs; i++) {
7809 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7811 for (j = 0; j < imu->nr_bvecs; j++)
7812 unpin_user_page(imu->bvec[j].bv_page);
7814 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7819 kfree(ctx->user_bufs);
7820 ctx->user_bufs = NULL;
7821 ctx->nr_user_bufs = 0;
7825 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7826 void __user *arg, unsigned index)
7828 struct iovec __user *src;
7830 #ifdef CONFIG_COMPAT
7832 struct compat_iovec __user *ciovs;
7833 struct compat_iovec ciov;
7835 ciovs = (struct compat_iovec __user *) arg;
7836 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7839 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7840 dst->iov_len = ciov.iov_len;
7844 src = (struct iovec __user *) arg;
7845 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7850 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7853 struct vm_area_struct **vmas = NULL;
7854 struct page **pages = NULL;
7855 int i, j, got_pages = 0;
7860 if (!nr_args || nr_args > UIO_MAXIOV)
7863 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7865 if (!ctx->user_bufs)
7868 for (i = 0; i < nr_args; i++) {
7869 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7870 unsigned long off, start, end, ubuf;
7875 ret = io_copy_iov(ctx, &iov, arg, i);
7880 * Don't impose further limits on the size and buffer
7881 * constraints here, we'll -EINVAL later when IO is
7882 * submitted if they are wrong.
7885 if (!iov.iov_base || !iov.iov_len)
7888 /* arbitrary limit, but we need something */
7889 if (iov.iov_len > SZ_1G)
7892 ubuf = (unsigned long) iov.iov_base;
7893 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7894 start = ubuf >> PAGE_SHIFT;
7895 nr_pages = end - start;
7897 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7902 if (!pages || nr_pages > got_pages) {
7905 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7907 vmas = kvmalloc_array(nr_pages,
7908 sizeof(struct vm_area_struct *),
7910 if (!pages || !vmas) {
7912 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7915 got_pages = nr_pages;
7918 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7922 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7927 mmap_read_lock(current->mm);
7928 pret = pin_user_pages(ubuf, nr_pages,
7929 FOLL_WRITE | FOLL_LONGTERM,
7931 if (pret == nr_pages) {
7932 /* don't support file backed memory */
7933 for (j = 0; j < nr_pages; j++) {
7934 struct vm_area_struct *vma = vmas[j];
7937 !is_file_hugepages(vma->vm_file)) {
7943 ret = pret < 0 ? pret : -EFAULT;
7945 mmap_read_unlock(current->mm);
7948 * if we did partial map, or found file backed vmas,
7949 * release any pages we did get
7952 unpin_user_pages(pages, pret);
7953 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7958 off = ubuf & ~PAGE_MASK;
7960 for (j = 0; j < nr_pages; j++) {
7963 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7964 imu->bvec[j].bv_page = pages[j];
7965 imu->bvec[j].bv_len = vec_len;
7966 imu->bvec[j].bv_offset = off;
7970 /* store original address for later verification */
7972 imu->len = iov.iov_len;
7973 imu->nr_bvecs = nr_pages;
7975 ctx->nr_user_bufs++;
7983 io_sqe_buffer_unregister(ctx);
7987 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7989 __s32 __user *fds = arg;
7995 if (copy_from_user(&fd, fds, sizeof(*fds)))
7998 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7999 if (IS_ERR(ctx->cq_ev_fd)) {
8000 int ret = PTR_ERR(ctx->cq_ev_fd);
8001 ctx->cq_ev_fd = NULL;
8008 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8010 if (ctx->cq_ev_fd) {
8011 eventfd_ctx_put(ctx->cq_ev_fd);
8012 ctx->cq_ev_fd = NULL;
8019 static int __io_destroy_buffers(int id, void *p, void *data)
8021 struct io_ring_ctx *ctx = data;
8022 struct io_buffer *buf = p;
8024 __io_remove_buffers(ctx, buf, id, -1U);
8028 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8030 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8031 idr_destroy(&ctx->io_buffer_idr);
8034 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8036 io_finish_async(ctx);
8037 io_sqe_buffer_unregister(ctx);
8039 if (ctx->sqo_task) {
8040 put_task_struct(ctx->sqo_task);
8041 ctx->sqo_task = NULL;
8042 mmdrop(ctx->mm_account);
8043 ctx->mm_account = NULL;
8046 io_sqe_files_unregister(ctx);
8047 io_eventfd_unregister(ctx);
8048 io_destroy_buffers(ctx);
8049 idr_destroy(&ctx->personality_idr);
8051 #if defined(CONFIG_UNIX)
8052 if (ctx->ring_sock) {
8053 ctx->ring_sock->file = NULL; /* so that iput() is called */
8054 sock_release(ctx->ring_sock);
8058 io_mem_free(ctx->rings);
8059 io_mem_free(ctx->sq_sqes);
8061 percpu_ref_exit(&ctx->refs);
8062 free_uid(ctx->user);
8063 put_cred(ctx->creds);
8064 kfree(ctx->cancel_hash);
8065 kmem_cache_free(req_cachep, ctx->fallback_req);
8069 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8071 struct io_ring_ctx *ctx = file->private_data;
8074 poll_wait(file, &ctx->cq_wait, wait);
8076 * synchronizes with barrier from wq_has_sleeper call in
8080 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
8081 ctx->rings->sq_ring_entries)
8082 mask |= EPOLLOUT | EPOLLWRNORM;
8083 if (io_cqring_events(ctx, false))
8084 mask |= EPOLLIN | EPOLLRDNORM;
8089 static int io_uring_fasync(int fd, struct file *file, int on)
8091 struct io_ring_ctx *ctx = file->private_data;
8093 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8096 static int io_remove_personalities(int id, void *p, void *data)
8098 struct io_ring_ctx *ctx = data;
8099 const struct cred *cred;
8101 cred = idr_remove(&ctx->personality_idr, id);
8107 static void io_ring_exit_work(struct work_struct *work)
8109 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8113 * If we're doing polled IO and end up having requests being
8114 * submitted async (out-of-line), then completions can come in while
8115 * we're waiting for refs to drop. We need to reap these manually,
8116 * as nobody else will be looking for them.
8120 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8121 io_iopoll_try_reap_events(ctx);
8122 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8123 io_ring_ctx_free(ctx);
8126 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8128 mutex_lock(&ctx->uring_lock);
8129 percpu_ref_kill(&ctx->refs);
8130 mutex_unlock(&ctx->uring_lock);
8132 io_kill_timeouts(ctx, NULL);
8133 io_poll_remove_all(ctx, NULL);
8136 io_wq_cancel_all(ctx->io_wq);
8138 /* if we failed setting up the ctx, we might not have any rings */
8140 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8141 io_iopoll_try_reap_events(ctx);
8142 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8145 * Do this upfront, so we won't have a grace period where the ring
8146 * is closed but resources aren't reaped yet. This can cause
8147 * spurious failure in setting up a new ring.
8149 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8152 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8154 * Use system_unbound_wq to avoid spawning tons of event kworkers
8155 * if we're exiting a ton of rings at the same time. It just adds
8156 * noise and overhead, there's no discernable change in runtime
8157 * over using system_wq.
8159 queue_work(system_unbound_wq, &ctx->exit_work);
8162 static int io_uring_release(struct inode *inode, struct file *file)
8164 struct io_ring_ctx *ctx = file->private_data;
8166 file->private_data = NULL;
8167 io_ring_ctx_wait_and_kill(ctx);
8171 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8173 struct files_struct *files = data;
8175 return !files || work->files == files;
8179 * Returns true if 'preq' is the link parent of 'req'
8181 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8183 struct io_kiocb *link;
8185 if (!(preq->flags & REQ_F_LINK_HEAD))
8188 list_for_each_entry(link, &preq->link_list, link_list) {
8196 static bool io_match_link_files(struct io_kiocb *req,
8197 struct files_struct *files)
8199 struct io_kiocb *link;
8201 if (io_match_files(req, files))
8203 if (req->flags & REQ_F_LINK_HEAD) {
8204 list_for_each_entry(link, &req->link_list, link_list) {
8205 if (io_match_files(link, files))
8213 * We're looking to cancel 'req' because it's holding on to our files, but
8214 * 'req' could be a link to another request. See if it is, and cancel that
8215 * parent request if so.
8217 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8219 struct hlist_node *tmp;
8220 struct io_kiocb *preq;
8224 spin_lock_irq(&ctx->completion_lock);
8225 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8226 struct hlist_head *list;
8228 list = &ctx->cancel_hash[i];
8229 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8230 found = io_match_link(preq, req);
8232 io_poll_remove_one(preq);
8237 spin_unlock_irq(&ctx->completion_lock);
8241 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8242 struct io_kiocb *req)
8244 struct io_kiocb *preq;
8247 spin_lock_irq(&ctx->completion_lock);
8248 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8249 found = io_match_link(preq, req);
8251 __io_timeout_cancel(preq);
8255 spin_unlock_irq(&ctx->completion_lock);
8259 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8261 return io_match_link(container_of(work, struct io_kiocb, work), data);
8264 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8266 enum io_wq_cancel cret;
8268 /* cancel this particular work, if it's running */
8269 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8270 if (cret != IO_WQ_CANCEL_NOTFOUND)
8273 /* find links that hold this pending, cancel those */
8274 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8275 if (cret != IO_WQ_CANCEL_NOTFOUND)
8278 /* if we have a poll link holding this pending, cancel that */
8279 if (io_poll_remove_link(ctx, req))
8282 /* final option, timeout link is holding this req pending */
8283 io_timeout_remove_link(ctx, req);
8286 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8287 struct files_struct *files)
8289 struct io_defer_entry *de = NULL;
8292 spin_lock_irq(&ctx->completion_lock);
8293 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8294 if (io_match_link_files(de->req, files)) {
8295 list_cut_position(&list, &ctx->defer_list, &de->list);
8299 spin_unlock_irq(&ctx->completion_lock);
8301 while (!list_empty(&list)) {
8302 de = list_first_entry(&list, struct io_defer_entry, list);
8303 list_del_init(&de->list);
8304 req_set_fail_links(de->req);
8305 io_put_req(de->req);
8306 io_req_complete(de->req, -ECANCELED);
8312 * Returns true if we found and killed one or more files pinning requests
8314 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8315 struct files_struct *files)
8317 if (list_empty_careful(&ctx->inflight_list))
8320 io_cancel_defer_files(ctx, files);
8321 /* cancel all at once, should be faster than doing it one by one*/
8322 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8324 while (!list_empty_careful(&ctx->inflight_list)) {
8325 struct io_kiocb *cancel_req = NULL, *req;
8328 spin_lock_irq(&ctx->inflight_lock);
8329 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8330 if (files && req->work.files != files)
8332 /* req is being completed, ignore */
8333 if (!refcount_inc_not_zero(&req->refs))
8339 prepare_to_wait(&ctx->inflight_wait, &wait,
8340 TASK_UNINTERRUPTIBLE);
8341 spin_unlock_irq(&ctx->inflight_lock);
8343 /* We need to keep going until we don't find a matching req */
8346 /* cancel this request, or head link requests */
8347 io_attempt_cancel(ctx, cancel_req);
8348 io_put_req(cancel_req);
8349 /* cancellations _may_ trigger task work */
8352 finish_wait(&ctx->inflight_wait, &wait);
8358 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8360 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8361 struct task_struct *task = data;
8363 return io_task_match(req, task);
8366 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8367 struct task_struct *task,
8368 struct files_struct *files)
8372 ret = io_uring_cancel_files(ctx, files);
8374 enum io_wq_cancel cret;
8376 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8377 if (cret != IO_WQ_CANCEL_NOTFOUND)
8380 /* SQPOLL thread does its own polling */
8381 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8382 while (!list_empty_careful(&ctx->iopoll_list)) {
8383 io_iopoll_try_reap_events(ctx);
8388 ret |= io_poll_remove_all(ctx, task);
8389 ret |= io_kill_timeouts(ctx, task);
8396 * We need to iteratively cancel requests, in case a request has dependent
8397 * hard links. These persist even for failure of cancelations, hence keep
8398 * looping until none are found.
8400 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8401 struct files_struct *files)
8403 struct task_struct *task = current;
8405 if (ctx->flags & IORING_SETUP_SQPOLL)
8406 task = ctx->sqo_thread;
8408 io_cqring_overflow_flush(ctx, true, task, files);
8410 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8417 * Note that this task has used io_uring. We use it for cancelation purposes.
8419 static int io_uring_add_task_file(struct file *file)
8421 if (unlikely(!current->io_uring)) {
8424 ret = io_uring_alloc_task_context(current);
8428 if (current->io_uring->last != file) {
8429 XA_STATE(xas, ¤t->io_uring->xa, (unsigned long) file);
8433 old = xas_load(&xas);
8437 xas_store(&xas, file);
8441 current->io_uring->last = file;
8448 * Remove this io_uring_file -> task mapping.
8450 static void io_uring_del_task_file(struct file *file)
8452 struct io_uring_task *tctx = current->io_uring;
8453 XA_STATE(xas, &tctx->xa, (unsigned long) file);
8455 if (tctx->last == file)
8459 file = xas_store(&xas, NULL);
8466 static void __io_uring_attempt_task_drop(struct file *file)
8468 XA_STATE(xas, ¤t->io_uring->xa, (unsigned long) file);
8472 old = xas_load(&xas);
8476 io_uring_del_task_file(file);
8480 * Drop task note for this file if we're the only ones that hold it after
8483 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8485 if (!current->io_uring)
8488 * fput() is pending, will be 2 if the only other ref is our potential
8489 * task file note. If the task is exiting, drop regardless of count.
8491 if (!exiting && atomic_long_read(&file->f_count) != 2)
8494 __io_uring_attempt_task_drop(file);
8497 void __io_uring_files_cancel(struct files_struct *files)
8499 struct io_uring_task *tctx = current->io_uring;
8500 XA_STATE(xas, &tctx->xa, 0);
8502 /* make sure overflow events are dropped */
8503 tctx->in_idle = true;
8506 struct io_ring_ctx *ctx;
8510 file = xas_next_entry(&xas, ULONG_MAX);
8516 ctx = file->private_data;
8518 io_uring_cancel_task_requests(ctx, files);
8520 io_uring_del_task_file(file);
8524 static inline bool io_uring_task_idle(struct io_uring_task *tctx)
8526 return atomic_long_read(&tctx->req_issue) ==
8527 atomic_long_read(&tctx->req_complete);
8531 * Find any io_uring fd that this task has registered or done IO on, and cancel
8534 void __io_uring_task_cancel(void)
8536 struct io_uring_task *tctx = current->io_uring;
8540 /* make sure overflow events are dropped */
8541 tctx->in_idle = true;
8543 while (!io_uring_task_idle(tctx)) {
8544 /* read completions before cancelations */
8545 completions = atomic_long_read(&tctx->req_complete);
8546 __io_uring_files_cancel(NULL);
8548 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8551 * If we've seen completions, retry. This avoids a race where
8552 * a completion comes in before we did prepare_to_wait().
8554 if (completions != atomic_long_read(&tctx->req_complete))
8556 if (io_uring_task_idle(tctx))
8561 finish_wait(&tctx->wait, &wait);
8562 tctx->in_idle = false;
8565 static int io_uring_flush(struct file *file, void *data)
8567 struct io_ring_ctx *ctx = file->private_data;
8570 * If the task is going away, cancel work it may have pending
8572 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8575 io_uring_cancel_task_requests(ctx, data);
8576 io_uring_attempt_task_drop(file, !data);
8580 static void *io_uring_validate_mmap_request(struct file *file,
8581 loff_t pgoff, size_t sz)
8583 struct io_ring_ctx *ctx = file->private_data;
8584 loff_t offset = pgoff << PAGE_SHIFT;
8589 case IORING_OFF_SQ_RING:
8590 case IORING_OFF_CQ_RING:
8593 case IORING_OFF_SQES:
8597 return ERR_PTR(-EINVAL);
8600 page = virt_to_head_page(ptr);
8601 if (sz > page_size(page))
8602 return ERR_PTR(-EINVAL);
8609 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8611 size_t sz = vma->vm_end - vma->vm_start;
8615 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8617 return PTR_ERR(ptr);
8619 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8620 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8623 #else /* !CONFIG_MMU */
8625 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8627 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8630 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8632 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8635 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8636 unsigned long addr, unsigned long len,
8637 unsigned long pgoff, unsigned long flags)
8641 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8643 return PTR_ERR(ptr);
8645 return (unsigned long) ptr;
8648 #endif /* !CONFIG_MMU */
8650 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8651 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8654 struct io_ring_ctx *ctx;
8661 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8669 if (f.file->f_op != &io_uring_fops)
8673 ctx = f.file->private_data;
8674 if (!percpu_ref_tryget(&ctx->refs))
8678 if (ctx->flags & IORING_SETUP_R_DISABLED)
8682 * For SQ polling, the thread will do all submissions and completions.
8683 * Just return the requested submit count, and wake the thread if
8687 if (ctx->flags & IORING_SETUP_SQPOLL) {
8688 if (!list_empty_careful(&ctx->cq_overflow_list))
8689 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8690 if (flags & IORING_ENTER_SQ_WAKEUP)
8691 wake_up(ctx->sqo_wait);
8692 submitted = to_submit;
8693 } else if (to_submit) {
8694 ret = io_uring_add_task_file(f.file);
8697 mutex_lock(&ctx->uring_lock);
8698 submitted = io_submit_sqes(ctx, to_submit);
8699 mutex_unlock(&ctx->uring_lock);
8701 if (submitted != to_submit)
8704 if (flags & IORING_ENTER_GETEVENTS) {
8705 min_complete = min(min_complete, ctx->cq_entries);
8708 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8709 * space applications don't need to do io completion events
8710 * polling again, they can rely on io_sq_thread to do polling
8711 * work, which can reduce cpu usage and uring_lock contention.
8713 if (ctx->flags & IORING_SETUP_IOPOLL &&
8714 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8715 ret = io_iopoll_check(ctx, min_complete);
8717 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8722 percpu_ref_put(&ctx->refs);
8725 return submitted ? submitted : ret;
8728 #ifdef CONFIG_PROC_FS
8729 static int io_uring_show_cred(int id, void *p, void *data)
8731 const struct cred *cred = p;
8732 struct seq_file *m = data;
8733 struct user_namespace *uns = seq_user_ns(m);
8734 struct group_info *gi;
8739 seq_printf(m, "%5d\n", id);
8740 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8741 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8742 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8743 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8744 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8745 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8746 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8747 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8748 seq_puts(m, "\n\tGroups:\t");
8749 gi = cred->group_info;
8750 for (g = 0; g < gi->ngroups; g++) {
8751 seq_put_decimal_ull(m, g ? " " : "",
8752 from_kgid_munged(uns, gi->gid[g]));
8754 seq_puts(m, "\n\tCapEff:\t");
8755 cap = cred->cap_effective;
8756 CAP_FOR_EACH_U32(__capi)
8757 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8762 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8768 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8769 * since fdinfo case grabs it in the opposite direction of normal use
8770 * cases. If we fail to get the lock, we just don't iterate any
8771 * structures that could be going away outside the io_uring mutex.
8773 has_lock = mutex_trylock(&ctx->uring_lock);
8775 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8776 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8777 struct fixed_file_table *table;
8780 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8781 f = table->files[i & IORING_FILE_TABLE_MASK];
8783 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8785 seq_printf(m, "%5u: <none>\n", i);
8787 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8788 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8789 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8791 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8792 (unsigned int) buf->len);
8794 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8795 seq_printf(m, "Personalities:\n");
8796 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8798 seq_printf(m, "PollList:\n");
8799 spin_lock_irq(&ctx->completion_lock);
8800 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8801 struct hlist_head *list = &ctx->cancel_hash[i];
8802 struct io_kiocb *req;
8804 hlist_for_each_entry(req, list, hash_node)
8805 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8806 req->task->task_works != NULL);
8808 spin_unlock_irq(&ctx->completion_lock);
8810 mutex_unlock(&ctx->uring_lock);
8813 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8815 struct io_ring_ctx *ctx = f->private_data;
8817 if (percpu_ref_tryget(&ctx->refs)) {
8818 __io_uring_show_fdinfo(ctx, m);
8819 percpu_ref_put(&ctx->refs);
8824 static const struct file_operations io_uring_fops = {
8825 .release = io_uring_release,
8826 .flush = io_uring_flush,
8827 .mmap = io_uring_mmap,
8829 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8830 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8832 .poll = io_uring_poll,
8833 .fasync = io_uring_fasync,
8834 #ifdef CONFIG_PROC_FS
8835 .show_fdinfo = io_uring_show_fdinfo,
8839 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8840 struct io_uring_params *p)
8842 struct io_rings *rings;
8843 size_t size, sq_array_offset;
8845 /* make sure these are sane, as we already accounted them */
8846 ctx->sq_entries = p->sq_entries;
8847 ctx->cq_entries = p->cq_entries;
8849 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8850 if (size == SIZE_MAX)
8853 rings = io_mem_alloc(size);
8858 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8859 rings->sq_ring_mask = p->sq_entries - 1;
8860 rings->cq_ring_mask = p->cq_entries - 1;
8861 rings->sq_ring_entries = p->sq_entries;
8862 rings->cq_ring_entries = p->cq_entries;
8863 ctx->sq_mask = rings->sq_ring_mask;
8864 ctx->cq_mask = rings->cq_ring_mask;
8866 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8867 if (size == SIZE_MAX) {
8868 io_mem_free(ctx->rings);
8873 ctx->sq_sqes = io_mem_alloc(size);
8874 if (!ctx->sq_sqes) {
8875 io_mem_free(ctx->rings);
8884 * Allocate an anonymous fd, this is what constitutes the application
8885 * visible backing of an io_uring instance. The application mmaps this
8886 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8887 * we have to tie this fd to a socket for file garbage collection purposes.
8889 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8894 #if defined(CONFIG_UNIX)
8895 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8901 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8905 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8906 O_RDWR | O_CLOEXEC);
8910 ret = PTR_ERR(file);
8914 #if defined(CONFIG_UNIX)
8915 ctx->ring_sock->file = file;
8917 if (unlikely(io_uring_add_task_file(file))) {
8918 file = ERR_PTR(-ENOMEM);
8921 fd_install(ret, file);
8924 #if defined(CONFIG_UNIX)
8925 sock_release(ctx->ring_sock);
8926 ctx->ring_sock = NULL;
8931 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8932 struct io_uring_params __user *params)
8934 struct user_struct *user = NULL;
8935 struct io_ring_ctx *ctx;
8941 if (entries > IORING_MAX_ENTRIES) {
8942 if (!(p->flags & IORING_SETUP_CLAMP))
8944 entries = IORING_MAX_ENTRIES;
8948 * Use twice as many entries for the CQ ring. It's possible for the
8949 * application to drive a higher depth than the size of the SQ ring,
8950 * since the sqes are only used at submission time. This allows for
8951 * some flexibility in overcommitting a bit. If the application has
8952 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8953 * of CQ ring entries manually.
8955 p->sq_entries = roundup_pow_of_two(entries);
8956 if (p->flags & IORING_SETUP_CQSIZE) {
8958 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8959 * to a power-of-two, if it isn't already. We do NOT impose
8960 * any cq vs sq ring sizing.
8962 if (p->cq_entries < p->sq_entries)
8964 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8965 if (!(p->flags & IORING_SETUP_CLAMP))
8967 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8969 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8971 p->cq_entries = 2 * p->sq_entries;
8974 user = get_uid(current_user());
8975 limit_mem = !capable(CAP_IPC_LOCK);
8978 ret = __io_account_mem(user,
8979 ring_pages(p->sq_entries, p->cq_entries));
8986 ctx = io_ring_ctx_alloc(p);
8989 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8994 ctx->compat = in_compat_syscall();
8996 ctx->creds = get_current_cred();
8998 ctx->sqo_task = get_task_struct(current);
9001 * This is just grabbed for accounting purposes. When a process exits,
9002 * the mm is exited and dropped before the files, hence we need to hang
9003 * on to this mm purely for the purposes of being able to unaccount
9004 * memory (locked/pinned vm). It's not used for anything else.
9006 mmgrab(current->mm);
9007 ctx->mm_account = current->mm;
9010 * Account memory _before_ installing the file descriptor. Once
9011 * the descriptor is installed, it can get closed at any time. Also
9012 * do this before hitting the general error path, as ring freeing
9013 * will un-account as well.
9015 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9017 ctx->limit_mem = limit_mem;
9019 ret = io_allocate_scq_urings(ctx, p);
9023 ret = io_sq_offload_create(ctx, p);
9027 if (!(p->flags & IORING_SETUP_R_DISABLED))
9028 io_sq_offload_start(ctx);
9030 memset(&p->sq_off, 0, sizeof(p->sq_off));
9031 p->sq_off.head = offsetof(struct io_rings, sq.head);
9032 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9033 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9034 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9035 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9036 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9037 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9039 memset(&p->cq_off, 0, sizeof(p->cq_off));
9040 p->cq_off.head = offsetof(struct io_rings, cq.head);
9041 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9042 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9043 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9044 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9045 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9046 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9048 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9049 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9050 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9051 IORING_FEAT_POLL_32BITS;
9053 if (copy_to_user(params, p, sizeof(*p))) {
9059 * Install ring fd as the very last thing, so we don't risk someone
9060 * having closed it before we finish setup
9062 ret = io_uring_get_fd(ctx);
9066 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9069 io_ring_ctx_wait_and_kill(ctx);
9074 * Sets up an aio uring context, and returns the fd. Applications asks for a
9075 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9076 * params structure passed in.
9078 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9080 struct io_uring_params p;
9083 if (copy_from_user(&p, params, sizeof(p)))
9085 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9090 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9091 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9092 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9093 IORING_SETUP_R_DISABLED))
9096 return io_uring_create(entries, &p, params);
9099 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9100 struct io_uring_params __user *, params)
9102 return io_uring_setup(entries, params);
9105 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9107 struct io_uring_probe *p;
9111 size = struct_size(p, ops, nr_args);
9112 if (size == SIZE_MAX)
9114 p = kzalloc(size, GFP_KERNEL);
9119 if (copy_from_user(p, arg, size))
9122 if (memchr_inv(p, 0, size))
9125 p->last_op = IORING_OP_LAST - 1;
9126 if (nr_args > IORING_OP_LAST)
9127 nr_args = IORING_OP_LAST;
9129 for (i = 0; i < nr_args; i++) {
9131 if (!io_op_defs[i].not_supported)
9132 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9137 if (copy_to_user(arg, p, size))
9144 static int io_register_personality(struct io_ring_ctx *ctx)
9146 const struct cred *creds = get_current_cred();
9149 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9150 USHRT_MAX, GFP_KERNEL);
9156 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9158 const struct cred *old_creds;
9160 old_creds = idr_remove(&ctx->personality_idr, id);
9162 put_cred(old_creds);
9169 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9170 unsigned int nr_args)
9172 struct io_uring_restriction *res;
9176 /* Restrictions allowed only if rings started disabled */
9177 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9180 /* We allow only a single restrictions registration */
9181 if (ctx->restrictions.registered)
9184 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9187 size = array_size(nr_args, sizeof(*res));
9188 if (size == SIZE_MAX)
9191 res = memdup_user(arg, size);
9193 return PTR_ERR(res);
9197 for (i = 0; i < nr_args; i++) {
9198 switch (res[i].opcode) {
9199 case IORING_RESTRICTION_REGISTER_OP:
9200 if (res[i].register_op >= IORING_REGISTER_LAST) {
9205 __set_bit(res[i].register_op,
9206 ctx->restrictions.register_op);
9208 case IORING_RESTRICTION_SQE_OP:
9209 if (res[i].sqe_op >= IORING_OP_LAST) {
9214 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9216 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9217 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9219 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9220 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9229 /* Reset all restrictions if an error happened */
9231 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9233 ctx->restrictions.registered = true;
9239 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9241 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9244 if (ctx->restrictions.registered)
9245 ctx->restricted = 1;
9247 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9249 io_sq_offload_start(ctx);
9254 static bool io_register_op_must_quiesce(int op)
9257 case IORING_UNREGISTER_FILES:
9258 case IORING_REGISTER_FILES_UPDATE:
9259 case IORING_REGISTER_PROBE:
9260 case IORING_REGISTER_PERSONALITY:
9261 case IORING_UNREGISTER_PERSONALITY:
9268 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9269 void __user *arg, unsigned nr_args)
9270 __releases(ctx->uring_lock)
9271 __acquires(ctx->uring_lock)
9276 * We're inside the ring mutex, if the ref is already dying, then
9277 * someone else killed the ctx or is already going through
9278 * io_uring_register().
9280 if (percpu_ref_is_dying(&ctx->refs))
9283 if (io_register_op_must_quiesce(opcode)) {
9284 percpu_ref_kill(&ctx->refs);
9287 * Drop uring mutex before waiting for references to exit. If
9288 * another thread is currently inside io_uring_enter() it might
9289 * need to grab the uring_lock to make progress. If we hold it
9290 * here across the drain wait, then we can deadlock. It's safe
9291 * to drop the mutex here, since no new references will come in
9292 * after we've killed the percpu ref.
9294 mutex_unlock(&ctx->uring_lock);
9295 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9296 mutex_lock(&ctx->uring_lock);
9298 percpu_ref_resurrect(&ctx->refs);
9304 if (ctx->restricted) {
9305 if (opcode >= IORING_REGISTER_LAST) {
9310 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9317 case IORING_REGISTER_BUFFERS:
9318 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9320 case IORING_UNREGISTER_BUFFERS:
9324 ret = io_sqe_buffer_unregister(ctx);
9326 case IORING_REGISTER_FILES:
9327 ret = io_sqe_files_register(ctx, arg, nr_args);
9329 case IORING_UNREGISTER_FILES:
9333 ret = io_sqe_files_unregister(ctx);
9335 case IORING_REGISTER_FILES_UPDATE:
9336 ret = io_sqe_files_update(ctx, arg, nr_args);
9338 case IORING_REGISTER_EVENTFD:
9339 case IORING_REGISTER_EVENTFD_ASYNC:
9343 ret = io_eventfd_register(ctx, arg);
9346 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9347 ctx->eventfd_async = 1;
9349 ctx->eventfd_async = 0;
9351 case IORING_UNREGISTER_EVENTFD:
9355 ret = io_eventfd_unregister(ctx);
9357 case IORING_REGISTER_PROBE:
9359 if (!arg || nr_args > 256)
9361 ret = io_probe(ctx, arg, nr_args);
9363 case IORING_REGISTER_PERSONALITY:
9367 ret = io_register_personality(ctx);
9369 case IORING_UNREGISTER_PERSONALITY:
9373 ret = io_unregister_personality(ctx, nr_args);
9375 case IORING_REGISTER_ENABLE_RINGS:
9379 ret = io_register_enable_rings(ctx);
9381 case IORING_REGISTER_RESTRICTIONS:
9382 ret = io_register_restrictions(ctx, arg, nr_args);
9390 if (io_register_op_must_quiesce(opcode)) {
9391 /* bring the ctx back to life */
9392 percpu_ref_reinit(&ctx->refs);
9394 reinit_completion(&ctx->ref_comp);
9399 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9400 void __user *, arg, unsigned int, nr_args)
9402 struct io_ring_ctx *ctx;
9411 if (f.file->f_op != &io_uring_fops)
9414 ctx = f.file->private_data;
9416 mutex_lock(&ctx->uring_lock);
9417 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9418 mutex_unlock(&ctx->uring_lock);
9419 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9420 ctx->cq_ev_fd != NULL, ret);
9426 static int __init io_uring_init(void)
9428 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9429 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9430 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9433 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9434 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9435 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9436 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9437 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9438 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9439 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9440 BUILD_BUG_SQE_ELEM(8, __u64, off);
9441 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9442 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9443 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9444 BUILD_BUG_SQE_ELEM(24, __u32, len);
9445 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9446 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9447 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9448 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9449 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9450 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9451 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9452 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9453 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9454 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9455 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9456 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9457 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9458 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9459 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9460 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9461 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9462 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9463 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9465 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9466 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9467 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9470 __initcall(io_uring_init);