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>
83 #include <linux/blk-cgroup.h>
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/io_uring.h>
88 #include <uapi/linux/io_uring.h>
93 #define IORING_MAX_ENTRIES 32768
94 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
97 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
99 #define IORING_FILE_TABLE_SHIFT 9
100 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
101 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
102 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
104 IORING_REGISTER_LAST + IORING_OP_LAST)
107 u32 head ____cacheline_aligned_in_smp;
108 u32 tail ____cacheline_aligned_in_smp;
112 * This data is shared with the application through the mmap at offsets
113 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
115 * The offsets to the member fields are published through struct
116 * io_sqring_offsets when calling io_uring_setup.
120 * Head and tail offsets into the ring; the offsets need to be
121 * masked to get valid indices.
123 * The kernel controls head of the sq ring and the tail of the cq ring,
124 * and the application controls tail of the sq ring and the head of the
127 struct io_uring sq, cq;
129 * Bitmasks to apply to head and tail offsets (constant, equals
132 u32 sq_ring_mask, cq_ring_mask;
133 /* Ring sizes (constant, power of 2) */
134 u32 sq_ring_entries, cq_ring_entries;
136 * Number of invalid entries dropped by the kernel due to
137 * invalid index stored in array
139 * Written by the kernel, shouldn't be modified by the
140 * application (i.e. get number of "new events" by comparing to
143 * After a new SQ head value was read by the application this
144 * counter includes all submissions that were dropped reaching
145 * the new SQ head (and possibly more).
151 * Written by the kernel, shouldn't be modified by the
154 * The application needs a full memory barrier before checking
155 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
161 * Written by the application, shouldn't be modified by the
166 * Number of completion events lost because the queue was full;
167 * this should be avoided by the application by making sure
168 * there are not more requests pending than there is space in
169 * the completion queue.
171 * Written by the kernel, shouldn't be modified by the
172 * application (i.e. get number of "new events" by comparing to
175 * As completion events come in out of order this counter is not
176 * ordered with any other data.
180 * Ring buffer of completion events.
182 * The kernel writes completion events fresh every time they are
183 * produced, so the application is allowed to modify pending
186 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
189 struct io_mapped_ubuf {
192 struct bio_vec *bvec;
193 unsigned int nr_bvecs;
194 unsigned long acct_pages;
197 struct fixed_file_table {
201 struct fixed_file_ref_node {
202 struct percpu_ref refs;
203 struct list_head node;
204 struct list_head file_list;
205 struct fixed_file_data *file_data;
206 struct llist_node llist;
209 struct fixed_file_data {
210 struct fixed_file_table *table;
211 struct io_ring_ctx *ctx;
213 struct percpu_ref *cur_refs;
214 struct percpu_ref refs;
215 struct completion done;
216 struct list_head ref_list;
221 struct list_head list;
227 struct io_restriction {
228 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
229 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
230 u8 sqe_flags_allowed;
231 u8 sqe_flags_required;
239 /* ctx's that are using this sqd */
240 struct list_head ctx_list;
241 struct list_head ctx_new_list;
242 struct mutex ctx_lock;
244 struct task_struct *thread;
245 struct wait_queue_head wait;
250 struct percpu_ref refs;
251 } ____cacheline_aligned_in_smp;
255 unsigned int compat: 1;
256 unsigned int limit_mem: 1;
257 unsigned int cq_overflow_flushed: 1;
258 unsigned int drain_next: 1;
259 unsigned int eventfd_async: 1;
260 unsigned int restricted: 1;
263 * Ring buffer of indices into array of io_uring_sqe, which is
264 * mmapped by the application using the IORING_OFF_SQES offset.
266 * This indirection could e.g. be used to assign fixed
267 * io_uring_sqe entries to operations and only submit them to
268 * the queue when needed.
270 * The kernel modifies neither the indices array nor the entries
274 unsigned cached_sq_head;
277 unsigned sq_thread_idle;
278 unsigned cached_sq_dropped;
279 atomic_t cached_cq_overflow;
280 unsigned long sq_check_overflow;
282 struct list_head defer_list;
283 struct list_head timeout_list;
284 struct list_head cq_overflow_list;
286 wait_queue_head_t inflight_wait;
287 struct io_uring_sqe *sq_sqes;
288 } ____cacheline_aligned_in_smp;
290 struct io_rings *rings;
296 * For SQPOLL usage - we hold a reference to the parent task, so we
297 * have access to the ->files
299 struct task_struct *sqo_task;
301 /* Only used for accounting purposes */
302 struct mm_struct *mm_account;
304 #ifdef CONFIG_BLK_CGROUP
305 struct cgroup_subsys_state *sqo_blkcg_css;
308 struct io_sq_data *sq_data; /* if using sq thread polling */
310 struct wait_queue_head sqo_sq_wait;
311 struct wait_queue_entry sqo_wait_entry;
312 struct list_head sqd_list;
315 * If used, fixed file set. Writers must ensure that ->refs is dead,
316 * readers must ensure that ->refs is alive as long as the file* is
317 * used. Only updated through io_uring_register(2).
319 struct fixed_file_data *file_data;
320 unsigned nr_user_files;
322 /* if used, fixed mapped user buffers */
323 unsigned nr_user_bufs;
324 struct io_mapped_ubuf *user_bufs;
326 struct user_struct *user;
328 const struct cred *creds;
330 struct completion ref_comp;
331 struct completion sq_thread_comp;
333 /* if all else fails... */
334 struct io_kiocb *fallback_req;
336 #if defined(CONFIG_UNIX)
337 struct socket *ring_sock;
340 struct idr io_buffer_idr;
342 struct idr personality_idr;
345 unsigned cached_cq_tail;
348 atomic_t cq_timeouts;
349 unsigned long cq_check_overflow;
350 struct wait_queue_head cq_wait;
351 struct fasync_struct *cq_fasync;
352 struct eventfd_ctx *cq_ev_fd;
353 } ____cacheline_aligned_in_smp;
356 struct mutex uring_lock;
357 wait_queue_head_t wait;
358 } ____cacheline_aligned_in_smp;
361 spinlock_t completion_lock;
364 * ->iopoll_list is protected by the ctx->uring_lock for
365 * io_uring instances that don't use IORING_SETUP_SQPOLL.
366 * For SQPOLL, only the single threaded io_sq_thread() will
367 * manipulate the list, hence no extra locking is needed there.
369 struct list_head iopoll_list;
370 struct hlist_head *cancel_hash;
371 unsigned cancel_hash_bits;
372 bool poll_multi_file;
374 spinlock_t inflight_lock;
375 struct list_head inflight_list;
376 } ____cacheline_aligned_in_smp;
378 struct delayed_work file_put_work;
379 struct llist_head file_put_llist;
381 struct work_struct exit_work;
382 struct io_restriction restrictions;
386 * First field must be the file pointer in all the
387 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
389 struct io_poll_iocb {
392 struct wait_queue_head *head;
398 struct wait_queue_entry wait;
403 struct file *put_file;
407 struct io_timeout_data {
408 struct io_kiocb *req;
409 struct hrtimer timer;
410 struct timespec64 ts;
411 enum hrtimer_mode mode;
416 struct sockaddr __user *addr;
417 int __user *addr_len;
419 unsigned long nofile;
441 struct list_head list;
445 /* NOTE: kiocb has the file as the first member, so don't do it here */
453 struct sockaddr __user *addr;
460 struct user_msghdr __user *umsg;
466 struct io_buffer *kbuf;
472 struct filename *filename;
474 unsigned long nofile;
477 struct io_files_update {
503 struct epoll_event event;
507 struct file *file_out;
508 struct file *file_in;
515 struct io_provide_buf {
529 const char __user *filename;
530 struct statx __user *buffer;
533 struct io_completion {
535 struct list_head list;
539 struct io_async_connect {
540 struct sockaddr_storage address;
543 struct io_async_msghdr {
544 struct iovec fast_iov[UIO_FASTIOV];
546 struct sockaddr __user *uaddr;
548 struct sockaddr_storage addr;
552 struct iovec fast_iov[UIO_FASTIOV];
553 const struct iovec *free_iovec;
554 struct iov_iter iter;
556 struct wait_page_queue wpq;
560 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
561 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
562 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
563 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
564 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
565 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
572 REQ_F_LINK_TIMEOUT_BIT,
574 REQ_F_COMP_LOCKED_BIT,
575 REQ_F_NEED_CLEANUP_BIT,
577 REQ_F_BUFFER_SELECTED_BIT,
578 REQ_F_NO_FILE_TABLE_BIT,
579 REQ_F_WORK_INITIALIZED_BIT,
581 /* not a real bit, just to check we're not overflowing the space */
587 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
588 /* drain existing IO first */
589 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
591 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
592 /* doesn't sever on completion < 0 */
593 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
595 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
596 /* IOSQE_BUFFER_SELECT */
597 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
600 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
601 /* fail rest of links */
602 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
603 /* on inflight list */
604 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
605 /* read/write uses file position */
606 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
607 /* must not punt to workers */
608 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
609 /* has linked timeout */
610 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
612 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
613 /* completion under lock */
614 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
616 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
617 /* already went through poll handler */
618 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
619 /* buffer already selected */
620 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
621 /* doesn't need file table for this request */
622 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
623 /* io_wq_work is initialized */
624 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
628 struct io_poll_iocb poll;
629 struct io_poll_iocb *double_poll;
633 * NOTE! Each of the iocb union members has the file pointer
634 * as the first entry in their struct definition. So you can
635 * access the file pointer through any of the sub-structs,
636 * or directly as just 'ki_filp' in this struct.
642 struct io_poll_iocb poll;
643 struct io_accept accept;
645 struct io_cancel cancel;
646 struct io_timeout timeout;
647 struct io_connect connect;
648 struct io_sr_msg sr_msg;
650 struct io_close close;
651 struct io_files_update files_update;
652 struct io_fadvise fadvise;
653 struct io_madvise madvise;
654 struct io_epoll epoll;
655 struct io_splice splice;
656 struct io_provide_buf pbuf;
657 struct io_statx statx;
658 /* use only after cleaning per-op data, see io_clean_op() */
659 struct io_completion compl;
662 /* opcode allocated if it needs to store data for async defer */
665 /* polled IO has completed */
671 struct io_ring_ctx *ctx;
674 struct task_struct *task;
677 struct list_head link_list;
680 * 1. used with ctx->iopoll_list with reads/writes
681 * 2. to track reqs with ->files (see io_op_def::file_table)
683 struct list_head inflight_entry;
685 struct percpu_ref *fixed_file_refs;
686 struct callback_head task_work;
687 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
688 struct hlist_node hash_node;
689 struct async_poll *apoll;
690 struct io_wq_work work;
693 struct io_defer_entry {
694 struct list_head list;
695 struct io_kiocb *req;
699 #define IO_IOPOLL_BATCH 8
701 struct io_comp_state {
703 struct list_head list;
704 struct io_ring_ctx *ctx;
707 struct io_submit_state {
708 struct blk_plug plug;
711 * io_kiocb alloc cache
713 void *reqs[IO_IOPOLL_BATCH];
714 unsigned int free_reqs;
717 * Batch completion logic
719 struct io_comp_state comp;
722 * File reference cache
726 unsigned int has_refs;
727 unsigned int ios_left;
731 /* needs current->mm setup, does mm access */
732 unsigned needs_mm : 1;
733 /* needs req->file assigned */
734 unsigned needs_file : 1;
735 /* don't fail if file grab fails */
736 unsigned needs_file_no_error : 1;
737 /* hash wq insertion if file is a regular file */
738 unsigned hash_reg_file : 1;
739 /* unbound wq insertion if file is a non-regular file */
740 unsigned unbound_nonreg_file : 1;
741 /* opcode is not supported by this kernel */
742 unsigned not_supported : 1;
743 /* needs file table */
744 unsigned file_table : 1;
746 unsigned needs_fs : 1;
747 /* set if opcode supports polled "wait" */
749 unsigned pollout : 1;
750 /* op supports buffer selection */
751 unsigned buffer_select : 1;
752 /* needs rlimit(RLIMIT_FSIZE) assigned */
753 unsigned needs_fsize : 1;
754 /* must always have async data allocated */
755 unsigned needs_async_data : 1;
756 /* needs blkcg context, issues async io potentially */
757 unsigned needs_blkcg : 1;
758 /* size of async data needed, if any */
759 unsigned short async_size;
762 static const struct io_op_def io_op_defs[] __read_mostly = {
763 [IORING_OP_NOP] = {},
764 [IORING_OP_READV] = {
767 .unbound_nonreg_file = 1,
770 .needs_async_data = 1,
772 .async_size = sizeof(struct io_async_rw),
774 [IORING_OP_WRITEV] = {
778 .unbound_nonreg_file = 1,
781 .needs_async_data = 1,
783 .async_size = sizeof(struct io_async_rw),
785 [IORING_OP_FSYNC] = {
789 [IORING_OP_READ_FIXED] = {
791 .unbound_nonreg_file = 1,
794 .async_size = sizeof(struct io_async_rw),
796 [IORING_OP_WRITE_FIXED] = {
799 .unbound_nonreg_file = 1,
803 .async_size = sizeof(struct io_async_rw),
805 [IORING_OP_POLL_ADD] = {
807 .unbound_nonreg_file = 1,
809 [IORING_OP_POLL_REMOVE] = {},
810 [IORING_OP_SYNC_FILE_RANGE] = {
814 [IORING_OP_SENDMSG] = {
817 .unbound_nonreg_file = 1,
820 .needs_async_data = 1,
822 .async_size = sizeof(struct io_async_msghdr),
824 [IORING_OP_RECVMSG] = {
827 .unbound_nonreg_file = 1,
831 .needs_async_data = 1,
833 .async_size = sizeof(struct io_async_msghdr),
835 [IORING_OP_TIMEOUT] = {
837 .needs_async_data = 1,
838 .async_size = sizeof(struct io_timeout_data),
840 [IORING_OP_TIMEOUT_REMOVE] = {},
841 [IORING_OP_ACCEPT] = {
844 .unbound_nonreg_file = 1,
848 [IORING_OP_ASYNC_CANCEL] = {},
849 [IORING_OP_LINK_TIMEOUT] = {
851 .needs_async_data = 1,
852 .async_size = sizeof(struct io_timeout_data),
854 [IORING_OP_CONNECT] = {
857 .unbound_nonreg_file = 1,
859 .needs_async_data = 1,
860 .async_size = sizeof(struct io_async_connect),
862 [IORING_OP_FALLOCATE] = {
867 [IORING_OP_OPENAT] = {
872 [IORING_OP_CLOSE] = {
874 .needs_file_no_error = 1,
878 [IORING_OP_FILES_UPDATE] = {
882 [IORING_OP_STATX] = {
891 .unbound_nonreg_file = 1,
895 .async_size = sizeof(struct io_async_rw),
897 [IORING_OP_WRITE] = {
900 .unbound_nonreg_file = 1,
904 .async_size = sizeof(struct io_async_rw),
906 [IORING_OP_FADVISE] = {
910 [IORING_OP_MADVISE] = {
917 .unbound_nonreg_file = 1,
924 .unbound_nonreg_file = 1,
929 [IORING_OP_OPENAT2] = {
934 [IORING_OP_EPOLL_CTL] = {
935 .unbound_nonreg_file = 1,
938 [IORING_OP_SPLICE] = {
941 .unbound_nonreg_file = 1,
944 [IORING_OP_PROVIDE_BUFFERS] = {},
945 [IORING_OP_REMOVE_BUFFERS] = {},
949 .unbound_nonreg_file = 1,
953 enum io_mem_account {
958 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
959 struct io_comp_state *cs);
960 static void io_cqring_fill_event(struct io_kiocb *req, long res);
961 static void io_put_req(struct io_kiocb *req);
962 static void io_double_put_req(struct io_kiocb *req);
963 static void __io_double_put_req(struct io_kiocb *req);
964 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
965 static void __io_queue_linked_timeout(struct io_kiocb *req);
966 static void io_queue_linked_timeout(struct io_kiocb *req);
967 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
968 struct io_uring_files_update *ip,
970 static void __io_clean_op(struct io_kiocb *req);
971 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
972 int fd, struct file **out_file, bool fixed);
973 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
974 static void io_file_put_work(struct work_struct *work);
976 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
977 struct iovec **iovec, struct iov_iter *iter,
979 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
980 const struct iovec *fast_iov,
981 struct iov_iter *iter, bool force);
983 static struct kmem_cache *req_cachep;
985 static const struct file_operations io_uring_fops __read_mostly;
987 struct sock *io_uring_get_socket(struct file *file)
989 #if defined(CONFIG_UNIX)
990 if (file->f_op == &io_uring_fops) {
991 struct io_ring_ctx *ctx = file->private_data;
993 return ctx->ring_sock->sk;
998 EXPORT_SYMBOL(io_uring_get_socket);
1000 static inline void io_clean_op(struct io_kiocb *req)
1002 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
1007 static void io_sq_thread_drop_mm(void)
1009 struct mm_struct *mm = current->mm;
1012 kthread_unuse_mm(mm);
1017 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1020 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
1021 !ctx->sqo_task->mm ||
1022 !mmget_not_zero(ctx->sqo_task->mm)))
1024 kthread_use_mm(ctx->sqo_task->mm);
1030 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1031 struct io_kiocb *req)
1033 if (!io_op_defs[req->opcode].needs_mm)
1035 return __io_sq_thread_acquire_mm(ctx);
1038 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1039 struct cgroup_subsys_state **cur_css)
1042 #ifdef CONFIG_BLK_CGROUP
1043 /* puts the old one when swapping */
1044 if (*cur_css != ctx->sqo_blkcg_css) {
1045 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1046 *cur_css = ctx->sqo_blkcg_css;
1051 static void io_sq_thread_unassociate_blkcg(void)
1053 #ifdef CONFIG_BLK_CGROUP
1054 kthread_associate_blkcg(NULL);
1058 static inline void req_set_fail_links(struct io_kiocb *req)
1060 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1061 req->flags |= REQ_F_FAIL_LINK;
1065 * Note: must call io_req_init_async() for the first time you
1066 * touch any members of io_wq_work.
1068 static inline void io_req_init_async(struct io_kiocb *req)
1070 if (req->flags & REQ_F_WORK_INITIALIZED)
1073 memset(&req->work, 0, sizeof(req->work));
1074 req->flags |= REQ_F_WORK_INITIALIZED;
1077 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1079 return ctx->flags & IORING_SETUP_SQPOLL;
1082 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1084 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1086 complete(&ctx->ref_comp);
1089 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1091 return !req->timeout.off;
1094 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1096 struct io_ring_ctx *ctx;
1099 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1103 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1104 if (!ctx->fallback_req)
1108 * Use 5 bits less than the max cq entries, that should give us around
1109 * 32 entries per hash list if totally full and uniformly spread.
1111 hash_bits = ilog2(p->cq_entries);
1115 ctx->cancel_hash_bits = hash_bits;
1116 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1118 if (!ctx->cancel_hash)
1120 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1122 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1123 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1126 ctx->flags = p->flags;
1127 init_waitqueue_head(&ctx->sqo_sq_wait);
1128 INIT_LIST_HEAD(&ctx->sqd_list);
1129 init_waitqueue_head(&ctx->cq_wait);
1130 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1131 init_completion(&ctx->ref_comp);
1132 init_completion(&ctx->sq_thread_comp);
1133 idr_init(&ctx->io_buffer_idr);
1134 idr_init(&ctx->personality_idr);
1135 mutex_init(&ctx->uring_lock);
1136 init_waitqueue_head(&ctx->wait);
1137 spin_lock_init(&ctx->completion_lock);
1138 INIT_LIST_HEAD(&ctx->iopoll_list);
1139 INIT_LIST_HEAD(&ctx->defer_list);
1140 INIT_LIST_HEAD(&ctx->timeout_list);
1141 init_waitqueue_head(&ctx->inflight_wait);
1142 spin_lock_init(&ctx->inflight_lock);
1143 INIT_LIST_HEAD(&ctx->inflight_list);
1144 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1145 init_llist_head(&ctx->file_put_llist);
1148 if (ctx->fallback_req)
1149 kmem_cache_free(req_cachep, ctx->fallback_req);
1150 kfree(ctx->cancel_hash);
1155 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1157 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1158 struct io_ring_ctx *ctx = req->ctx;
1160 return seq != ctx->cached_cq_tail
1161 + atomic_read(&ctx->cached_cq_overflow);
1167 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1169 struct io_rings *rings = ctx->rings;
1171 /* order cqe stores with ring update */
1172 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1174 if (wq_has_sleeper(&ctx->cq_wait)) {
1175 wake_up_interruptible(&ctx->cq_wait);
1176 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1181 * Returns true if we need to defer file table putting. This can only happen
1182 * from the error path with REQ_F_COMP_LOCKED set.
1184 static bool io_req_clean_work(struct io_kiocb *req)
1186 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1189 req->flags &= ~REQ_F_WORK_INITIALIZED;
1192 mmdrop(req->work.mm);
1193 req->work.mm = NULL;
1195 #ifdef CONFIG_BLK_CGROUP
1196 if (req->work.blkcg_css)
1197 css_put(req->work.blkcg_css);
1199 if (req->work.creds) {
1200 put_cred(req->work.creds);
1201 req->work.creds = NULL;
1204 struct fs_struct *fs = req->work.fs;
1206 if (req->flags & REQ_F_COMP_LOCKED)
1209 spin_lock(&req->work.fs->lock);
1212 spin_unlock(&req->work.fs->lock);
1215 req->work.fs = NULL;
1221 static void io_prep_async_work(struct io_kiocb *req)
1223 const struct io_op_def *def = &io_op_defs[req->opcode];
1224 struct io_ring_ctx *ctx = req->ctx;
1226 io_req_init_async(req);
1228 if (req->flags & REQ_F_ISREG) {
1229 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1230 io_wq_hash_work(&req->work, file_inode(req->file));
1232 if (def->unbound_nonreg_file)
1233 req->work.flags |= IO_WQ_WORK_UNBOUND;
1235 if (!req->work.files && io_op_defs[req->opcode].file_table &&
1236 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1237 req->work.files = get_files_struct(current);
1238 get_nsproxy(current->nsproxy);
1239 req->work.nsproxy = current->nsproxy;
1240 req->flags |= REQ_F_INFLIGHT;
1242 spin_lock_irq(&ctx->inflight_lock);
1243 list_add(&req->inflight_entry, &ctx->inflight_list);
1244 spin_unlock_irq(&ctx->inflight_lock);
1246 if (!req->work.mm && def->needs_mm) {
1247 mmgrab(current->mm);
1248 req->work.mm = current->mm;
1250 #ifdef CONFIG_BLK_CGROUP
1251 if (!req->work.blkcg_css && def->needs_blkcg) {
1253 req->work.blkcg_css = blkcg_css();
1255 * This should be rare, either the cgroup is dying or the task
1256 * is moving cgroups. Just punt to root for the handful of ios.
1258 if (!css_tryget_online(req->work.blkcg_css))
1259 req->work.blkcg_css = NULL;
1263 if (!req->work.creds)
1264 req->work.creds = get_current_cred();
1265 if (!req->work.fs && def->needs_fs) {
1266 spin_lock(¤t->fs->lock);
1267 if (!current->fs->in_exec) {
1268 req->work.fs = current->fs;
1269 req->work.fs->users++;
1271 req->work.flags |= IO_WQ_WORK_CANCEL;
1273 spin_unlock(¤t->fs->lock);
1275 if (def->needs_fsize)
1276 req->work.fsize = rlimit(RLIMIT_FSIZE);
1278 req->work.fsize = RLIM_INFINITY;
1281 static void io_prep_async_link(struct io_kiocb *req)
1283 struct io_kiocb *cur;
1285 io_prep_async_work(req);
1286 if (req->flags & REQ_F_LINK_HEAD)
1287 list_for_each_entry(cur, &req->link_list, link_list)
1288 io_prep_async_work(cur);
1291 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1293 struct io_ring_ctx *ctx = req->ctx;
1294 struct io_kiocb *link = io_prep_linked_timeout(req);
1296 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1297 &req->work, req->flags);
1298 io_wq_enqueue(ctx->io_wq, &req->work);
1302 static void io_queue_async_work(struct io_kiocb *req)
1304 struct io_kiocb *link;
1306 /* init ->work of the whole link before punting */
1307 io_prep_async_link(req);
1308 link = __io_queue_async_work(req);
1311 io_queue_linked_timeout(link);
1314 static void io_kill_timeout(struct io_kiocb *req)
1316 struct io_timeout_data *io = req->async_data;
1319 ret = hrtimer_try_to_cancel(&io->timer);
1321 atomic_set(&req->ctx->cq_timeouts,
1322 atomic_read(&req->ctx->cq_timeouts) + 1);
1323 list_del_init(&req->timeout.list);
1324 req->flags |= REQ_F_COMP_LOCKED;
1325 io_cqring_fill_event(req, 0);
1330 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1332 struct io_ring_ctx *ctx = req->ctx;
1334 if (!tsk || req->task == tsk)
1336 if (ctx->flags & IORING_SETUP_SQPOLL) {
1337 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1344 * Returns true if we found and killed one or more timeouts
1346 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1348 struct io_kiocb *req, *tmp;
1351 spin_lock_irq(&ctx->completion_lock);
1352 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1353 if (io_task_match(req, tsk)) {
1354 io_kill_timeout(req);
1358 spin_unlock_irq(&ctx->completion_lock);
1359 return canceled != 0;
1362 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1365 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1366 struct io_defer_entry, list);
1367 struct io_kiocb *link;
1369 if (req_need_defer(de->req, de->seq))
1371 list_del_init(&de->list);
1372 /* punt-init is done before queueing for defer */
1373 link = __io_queue_async_work(de->req);
1375 __io_queue_linked_timeout(link);
1376 /* drop submission reference */
1377 link->flags |= REQ_F_COMP_LOCKED;
1381 } while (!list_empty(&ctx->defer_list));
1384 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1386 while (!list_empty(&ctx->timeout_list)) {
1387 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1388 struct io_kiocb, timeout.list);
1390 if (io_is_timeout_noseq(req))
1392 if (req->timeout.target_seq != ctx->cached_cq_tail
1393 - atomic_read(&ctx->cq_timeouts))
1396 list_del_init(&req->timeout.list);
1397 io_kill_timeout(req);
1401 static void io_commit_cqring(struct io_ring_ctx *ctx)
1403 io_flush_timeouts(ctx);
1404 __io_commit_cqring(ctx);
1406 if (unlikely(!list_empty(&ctx->defer_list)))
1407 __io_queue_deferred(ctx);
1410 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1412 struct io_rings *r = ctx->rings;
1414 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1417 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1419 struct io_rings *rings = ctx->rings;
1422 tail = ctx->cached_cq_tail;
1424 * writes to the cq entry need to come after reading head; the
1425 * control dependency is enough as we're using WRITE_ONCE to
1428 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1431 ctx->cached_cq_tail++;
1432 return &rings->cqes[tail & ctx->cq_mask];
1435 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1439 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1441 if (!ctx->eventfd_async)
1443 return io_wq_current_is_worker();
1446 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1448 if (waitqueue_active(&ctx->wait))
1449 wake_up(&ctx->wait);
1450 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1451 wake_up(&ctx->sq_data->wait);
1452 if (io_should_trigger_evfd(ctx))
1453 eventfd_signal(ctx->cq_ev_fd, 1);
1456 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1458 if (list_empty(&ctx->cq_overflow_list)) {
1459 clear_bit(0, &ctx->sq_check_overflow);
1460 clear_bit(0, &ctx->cq_check_overflow);
1461 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1465 static inline bool io_match_files(struct io_kiocb *req,
1466 struct files_struct *files)
1470 if (req->flags & REQ_F_WORK_INITIALIZED)
1471 return req->work.files == files;
1475 /* Returns true if there are no backlogged entries after the flush */
1476 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1477 struct task_struct *tsk,
1478 struct files_struct *files)
1480 struct io_rings *rings = ctx->rings;
1481 struct io_kiocb *req, *tmp;
1482 struct io_uring_cqe *cqe;
1483 unsigned long flags;
1487 if (list_empty_careful(&ctx->cq_overflow_list))
1489 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1490 rings->cq_ring_entries))
1494 spin_lock_irqsave(&ctx->completion_lock, flags);
1496 /* if force is set, the ring is going away. always drop after that */
1498 ctx->cq_overflow_flushed = 1;
1501 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1502 if (tsk && req->task != tsk)
1504 if (!io_match_files(req, files))
1507 cqe = io_get_cqring(ctx);
1511 list_move(&req->compl.list, &list);
1513 WRITE_ONCE(cqe->user_data, req->user_data);
1514 WRITE_ONCE(cqe->res, req->result);
1515 WRITE_ONCE(cqe->flags, req->compl.cflags);
1517 WRITE_ONCE(ctx->rings->cq_overflow,
1518 atomic_inc_return(&ctx->cached_cq_overflow));
1522 io_commit_cqring(ctx);
1523 io_cqring_mark_overflow(ctx);
1525 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1526 io_cqring_ev_posted(ctx);
1528 while (!list_empty(&list)) {
1529 req = list_first_entry(&list, struct io_kiocb, compl.list);
1530 list_del(&req->compl.list);
1537 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1539 struct io_ring_ctx *ctx = req->ctx;
1540 struct io_uring_cqe *cqe;
1542 trace_io_uring_complete(ctx, req->user_data, res);
1545 * If we can't get a cq entry, userspace overflowed the
1546 * submission (by quite a lot). Increment the overflow count in
1549 cqe = io_get_cqring(ctx);
1551 WRITE_ONCE(cqe->user_data, req->user_data);
1552 WRITE_ONCE(cqe->res, res);
1553 WRITE_ONCE(cqe->flags, cflags);
1554 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1556 * If we're in ring overflow flush mode, or in task cancel mode,
1557 * then we cannot store the request for later flushing, we need
1558 * to drop it on the floor.
1560 WRITE_ONCE(ctx->rings->cq_overflow,
1561 atomic_inc_return(&ctx->cached_cq_overflow));
1563 if (list_empty(&ctx->cq_overflow_list)) {
1564 set_bit(0, &ctx->sq_check_overflow);
1565 set_bit(0, &ctx->cq_check_overflow);
1566 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1570 req->compl.cflags = cflags;
1571 refcount_inc(&req->refs);
1572 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1576 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1578 __io_cqring_fill_event(req, res, 0);
1581 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1583 struct io_ring_ctx *ctx = req->ctx;
1584 unsigned long flags;
1586 spin_lock_irqsave(&ctx->completion_lock, flags);
1587 __io_cqring_fill_event(req, res, cflags);
1588 io_commit_cqring(ctx);
1589 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1591 io_cqring_ev_posted(ctx);
1594 static void io_submit_flush_completions(struct io_comp_state *cs)
1596 struct io_ring_ctx *ctx = cs->ctx;
1598 spin_lock_irq(&ctx->completion_lock);
1599 while (!list_empty(&cs->list)) {
1600 struct io_kiocb *req;
1602 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1603 list_del(&req->compl.list);
1604 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1605 if (!(req->flags & REQ_F_LINK_HEAD)) {
1606 req->flags |= REQ_F_COMP_LOCKED;
1609 spin_unlock_irq(&ctx->completion_lock);
1611 spin_lock_irq(&ctx->completion_lock);
1614 io_commit_cqring(ctx);
1615 spin_unlock_irq(&ctx->completion_lock);
1617 io_cqring_ev_posted(ctx);
1621 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1622 struct io_comp_state *cs)
1625 io_cqring_add_event(req, res, cflags);
1630 req->compl.cflags = cflags;
1631 list_add_tail(&req->compl.list, &cs->list);
1633 io_submit_flush_completions(cs);
1637 static void io_req_complete(struct io_kiocb *req, long res)
1639 __io_req_complete(req, res, 0, NULL);
1642 static inline bool io_is_fallback_req(struct io_kiocb *req)
1644 return req == (struct io_kiocb *)
1645 ((unsigned long) req->ctx->fallback_req & ~1UL);
1648 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1650 struct io_kiocb *req;
1652 req = ctx->fallback_req;
1653 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1659 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1660 struct io_submit_state *state)
1662 if (!state->free_reqs) {
1663 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1667 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1668 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1671 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1672 * retry single alloc to be on the safe side.
1674 if (unlikely(ret <= 0)) {
1675 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1676 if (!state->reqs[0])
1680 state->free_reqs = ret;
1684 return state->reqs[state->free_reqs];
1686 return io_get_fallback_req(ctx);
1689 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1693 percpu_ref_put(req->fixed_file_refs);
1698 static bool io_dismantle_req(struct io_kiocb *req)
1702 if (req->async_data)
1703 kfree(req->async_data);
1705 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1707 return io_req_clean_work(req);
1710 static void __io_free_req_finish(struct io_kiocb *req)
1712 struct io_uring_task *tctx = req->task->io_uring;
1713 struct io_ring_ctx *ctx = req->ctx;
1715 atomic_long_inc(&tctx->req_complete);
1717 wake_up(&tctx->wait);
1718 put_task_struct(req->task);
1720 if (likely(!io_is_fallback_req(req)))
1721 kmem_cache_free(req_cachep, req);
1723 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1724 percpu_ref_put(&ctx->refs);
1727 static void io_req_task_file_table_put(struct callback_head *cb)
1729 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1730 struct fs_struct *fs = req->work.fs;
1732 spin_lock(&req->work.fs->lock);
1735 spin_unlock(&req->work.fs->lock);
1738 req->work.fs = NULL;
1739 __io_free_req_finish(req);
1742 static void __io_free_req(struct io_kiocb *req)
1744 if (!io_dismantle_req(req)) {
1745 __io_free_req_finish(req);
1749 init_task_work(&req->task_work, io_req_task_file_table_put);
1750 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1751 if (unlikely(ret)) {
1752 struct task_struct *tsk;
1754 tsk = io_wq_get_task(req->ctx->io_wq);
1755 task_work_add(tsk, &req->task_work, 0);
1760 static bool io_link_cancel_timeout(struct io_kiocb *req)
1762 struct io_timeout_data *io = req->async_data;
1763 struct io_ring_ctx *ctx = req->ctx;
1766 ret = hrtimer_try_to_cancel(&io->timer);
1768 io_cqring_fill_event(req, -ECANCELED);
1769 io_commit_cqring(ctx);
1770 req->flags &= ~REQ_F_LINK_HEAD;
1778 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1780 struct io_kiocb *link;
1783 if (list_empty(&req->link_list))
1785 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1786 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1789 list_del_init(&link->link_list);
1790 link->flags |= REQ_F_COMP_LOCKED;
1791 wake_ev = io_link_cancel_timeout(link);
1792 req->flags &= ~REQ_F_LINK_TIMEOUT;
1796 static void io_kill_linked_timeout(struct io_kiocb *req)
1798 struct io_ring_ctx *ctx = req->ctx;
1801 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1802 unsigned long flags;
1804 spin_lock_irqsave(&ctx->completion_lock, flags);
1805 wake_ev = __io_kill_linked_timeout(req);
1806 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1808 wake_ev = __io_kill_linked_timeout(req);
1812 io_cqring_ev_posted(ctx);
1815 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1817 struct io_kiocb *nxt;
1820 * The list should never be empty when we are called here. But could
1821 * potentially happen if the chain is messed up, check to be on the
1824 if (unlikely(list_empty(&req->link_list)))
1827 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1828 list_del_init(&req->link_list);
1829 if (!list_empty(&nxt->link_list))
1830 nxt->flags |= REQ_F_LINK_HEAD;
1835 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1837 static void __io_fail_links(struct io_kiocb *req)
1839 struct io_ring_ctx *ctx = req->ctx;
1841 while (!list_empty(&req->link_list)) {
1842 struct io_kiocb *link = list_first_entry(&req->link_list,
1843 struct io_kiocb, link_list);
1845 list_del_init(&link->link_list);
1846 trace_io_uring_fail_link(req, link);
1848 io_cqring_fill_event(link, -ECANCELED);
1849 link->flags |= REQ_F_COMP_LOCKED;
1850 __io_double_put_req(link);
1851 req->flags &= ~REQ_F_LINK_TIMEOUT;
1854 io_commit_cqring(ctx);
1855 io_cqring_ev_posted(ctx);
1858 static void io_fail_links(struct io_kiocb *req)
1860 struct io_ring_ctx *ctx = req->ctx;
1862 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1863 unsigned long flags;
1865 spin_lock_irqsave(&ctx->completion_lock, flags);
1866 __io_fail_links(req);
1867 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1869 __io_fail_links(req);
1872 io_cqring_ev_posted(ctx);
1875 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1877 req->flags &= ~REQ_F_LINK_HEAD;
1878 if (req->flags & REQ_F_LINK_TIMEOUT)
1879 io_kill_linked_timeout(req);
1882 * If LINK is set, we have dependent requests in this chain. If we
1883 * didn't fail this request, queue the first one up, moving any other
1884 * dependencies to the next request. In case of failure, fail the rest
1887 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1888 return io_req_link_next(req);
1893 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1895 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1897 return __io_req_find_next(req);
1900 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1902 struct task_struct *tsk = req->task;
1903 struct io_ring_ctx *ctx = req->ctx;
1906 if (tsk->flags & PF_EXITING)
1910 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1911 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1912 * processing task_work. There's no reliable way to tell if TWA_RESUME
1916 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1917 notify = TWA_SIGNAL;
1919 ret = task_work_add(tsk, &req->task_work, notify);
1921 wake_up_process(tsk);
1926 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1928 struct io_ring_ctx *ctx = req->ctx;
1930 spin_lock_irq(&ctx->completion_lock);
1931 io_cqring_fill_event(req, error);
1932 io_commit_cqring(ctx);
1933 spin_unlock_irq(&ctx->completion_lock);
1935 io_cqring_ev_posted(ctx);
1936 req_set_fail_links(req);
1937 io_double_put_req(req);
1940 static void io_req_task_cancel(struct callback_head *cb)
1942 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1943 struct io_ring_ctx *ctx = req->ctx;
1945 __io_req_task_cancel(req, -ECANCELED);
1946 percpu_ref_put(&ctx->refs);
1949 static void __io_req_task_submit(struct io_kiocb *req)
1951 struct io_ring_ctx *ctx = req->ctx;
1953 if (!__io_sq_thread_acquire_mm(ctx)) {
1954 mutex_lock(&ctx->uring_lock);
1955 __io_queue_sqe(req, NULL);
1956 mutex_unlock(&ctx->uring_lock);
1958 __io_req_task_cancel(req, -EFAULT);
1962 static void io_req_task_submit(struct callback_head *cb)
1964 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1965 struct io_ring_ctx *ctx = req->ctx;
1967 __io_req_task_submit(req);
1968 percpu_ref_put(&ctx->refs);
1971 static void io_req_task_queue(struct io_kiocb *req)
1975 init_task_work(&req->task_work, io_req_task_submit);
1976 percpu_ref_get(&req->ctx->refs);
1978 ret = io_req_task_work_add(req, true);
1979 if (unlikely(ret)) {
1980 struct task_struct *tsk;
1982 init_task_work(&req->task_work, io_req_task_cancel);
1983 tsk = io_wq_get_task(req->ctx->io_wq);
1984 task_work_add(tsk, &req->task_work, 0);
1985 wake_up_process(tsk);
1989 static void io_queue_next(struct io_kiocb *req)
1991 struct io_kiocb *nxt = io_req_find_next(req);
1994 io_req_task_queue(nxt);
1997 static void io_free_req(struct io_kiocb *req)
2004 void *reqs[IO_IOPOLL_BATCH];
2007 struct task_struct *task;
2011 static inline void io_init_req_batch(struct req_batch *rb)
2018 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2019 struct req_batch *rb)
2021 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2022 percpu_ref_put_many(&ctx->refs, rb->to_free);
2026 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2027 struct req_batch *rb)
2030 __io_req_free_batch_flush(ctx, rb);
2032 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2033 put_task_struct_many(rb->task, rb->task_refs);
2038 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2040 if (unlikely(io_is_fallback_req(req))) {
2044 if (req->flags & REQ_F_LINK_HEAD)
2047 if (req->task != rb->task) {
2049 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2050 put_task_struct_many(rb->task, rb->task_refs);
2052 rb->task = req->task;
2057 WARN_ON_ONCE(io_dismantle_req(req));
2058 rb->reqs[rb->to_free++] = req;
2059 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2060 __io_req_free_batch_flush(req->ctx, rb);
2064 * Drop reference to request, return next in chain (if there is one) if this
2065 * was the last reference to this request.
2067 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2069 struct io_kiocb *nxt = NULL;
2071 if (refcount_dec_and_test(&req->refs)) {
2072 nxt = io_req_find_next(req);
2078 static void io_put_req(struct io_kiocb *req)
2080 if (refcount_dec_and_test(&req->refs))
2084 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2086 struct io_kiocb *nxt;
2089 * A ref is owned by io-wq in which context we're. So, if that's the
2090 * last one, it's safe to steal next work. False negatives are Ok,
2091 * it just will be re-punted async in io_put_work()
2093 if (refcount_read(&req->refs) != 1)
2096 nxt = io_req_find_next(req);
2097 return nxt ? &nxt->work : NULL;
2101 * Must only be used if we don't need to care about links, usually from
2102 * within the completion handling itself.
2104 static void __io_double_put_req(struct io_kiocb *req)
2106 /* drop both submit and complete references */
2107 if (refcount_sub_and_test(2, &req->refs))
2111 static void io_double_put_req(struct io_kiocb *req)
2113 /* drop both submit and complete references */
2114 if (refcount_sub_and_test(2, &req->refs))
2118 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2120 struct io_rings *rings = ctx->rings;
2122 if (test_bit(0, &ctx->cq_check_overflow)) {
2124 * noflush == true is from the waitqueue handler, just ensure
2125 * we wake up the task, and the next invocation will flush the
2126 * entries. We cannot safely to it from here.
2128 if (noflush && !list_empty(&ctx->cq_overflow_list))
2131 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2134 /* See comment at the top of this file */
2136 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2139 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2141 struct io_rings *rings = ctx->rings;
2143 /* make sure SQ entry isn't read before tail */
2144 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2147 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2149 unsigned int cflags;
2151 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2152 cflags |= IORING_CQE_F_BUFFER;
2153 req->flags &= ~REQ_F_BUFFER_SELECTED;
2158 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2160 struct io_buffer *kbuf;
2162 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2163 return io_put_kbuf(req, kbuf);
2166 static inline bool io_run_task_work(void)
2169 * Not safe to run on exiting task, and the task_work handling will
2170 * not add work to such a task.
2172 if (unlikely(current->flags & PF_EXITING))
2174 if (current->task_works) {
2175 __set_current_state(TASK_RUNNING);
2183 static void io_iopoll_queue(struct list_head *again)
2185 struct io_kiocb *req;
2188 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2189 list_del(&req->inflight_entry);
2190 __io_complete_rw(req, -EAGAIN, 0, NULL);
2191 } while (!list_empty(again));
2195 * Find and free completed poll iocbs
2197 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2198 struct list_head *done)
2200 struct req_batch rb;
2201 struct io_kiocb *req;
2204 /* order with ->result store in io_complete_rw_iopoll() */
2207 io_init_req_batch(&rb);
2208 while (!list_empty(done)) {
2211 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2212 if (READ_ONCE(req->result) == -EAGAIN) {
2214 req->iopoll_completed = 0;
2215 list_move_tail(&req->inflight_entry, &again);
2218 list_del(&req->inflight_entry);
2220 if (req->flags & REQ_F_BUFFER_SELECTED)
2221 cflags = io_put_rw_kbuf(req);
2223 __io_cqring_fill_event(req, req->result, cflags);
2226 if (refcount_dec_and_test(&req->refs))
2227 io_req_free_batch(&rb, req);
2230 io_commit_cqring(ctx);
2231 if (ctx->flags & IORING_SETUP_SQPOLL)
2232 io_cqring_ev_posted(ctx);
2233 io_req_free_batch_finish(ctx, &rb);
2235 if (!list_empty(&again))
2236 io_iopoll_queue(&again);
2239 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2242 struct io_kiocb *req, *tmp;
2248 * Only spin for completions if we don't have multiple devices hanging
2249 * off our complete list, and we're under the requested amount.
2251 spin = !ctx->poll_multi_file && *nr_events < min;
2254 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2255 struct kiocb *kiocb = &req->rw.kiocb;
2258 * Move completed and retryable entries to our local lists.
2259 * If we find a request that requires polling, break out
2260 * and complete those lists first, if we have entries there.
2262 if (READ_ONCE(req->iopoll_completed)) {
2263 list_move_tail(&req->inflight_entry, &done);
2266 if (!list_empty(&done))
2269 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2273 /* iopoll may have completed current req */
2274 if (READ_ONCE(req->iopoll_completed))
2275 list_move_tail(&req->inflight_entry, &done);
2282 if (!list_empty(&done))
2283 io_iopoll_complete(ctx, nr_events, &done);
2289 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2290 * non-spinning poll check - we'll still enter the driver poll loop, but only
2291 * as a non-spinning completion check.
2293 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2296 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2299 ret = io_do_iopoll(ctx, nr_events, min);
2302 if (*nr_events >= min)
2310 * We can't just wait for polled events to come to us, we have to actively
2311 * find and complete them.
2313 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2315 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2318 mutex_lock(&ctx->uring_lock);
2319 while (!list_empty(&ctx->iopoll_list)) {
2320 unsigned int nr_events = 0;
2322 io_do_iopoll(ctx, &nr_events, 0);
2324 /* let it sleep and repeat later if can't complete a request */
2328 * Ensure we allow local-to-the-cpu processing to take place,
2329 * in this case we need to ensure that we reap all events.
2330 * Also let task_work, etc. to progress by releasing the mutex
2332 if (need_resched()) {
2333 mutex_unlock(&ctx->uring_lock);
2335 mutex_lock(&ctx->uring_lock);
2338 mutex_unlock(&ctx->uring_lock);
2341 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2343 unsigned int nr_events = 0;
2344 int iters = 0, ret = 0;
2347 * We disallow the app entering submit/complete with polling, but we
2348 * still need to lock the ring to prevent racing with polled issue
2349 * that got punted to a workqueue.
2351 mutex_lock(&ctx->uring_lock);
2354 * Don't enter poll loop if we already have events pending.
2355 * If we do, we can potentially be spinning for commands that
2356 * already triggered a CQE (eg in error).
2358 if (io_cqring_events(ctx, false))
2362 * If a submit got punted to a workqueue, we can have the
2363 * application entering polling for a command before it gets
2364 * issued. That app will hold the uring_lock for the duration
2365 * of the poll right here, so we need to take a breather every
2366 * now and then to ensure that the issue has a chance to add
2367 * the poll to the issued list. Otherwise we can spin here
2368 * forever, while the workqueue is stuck trying to acquire the
2371 if (!(++iters & 7)) {
2372 mutex_unlock(&ctx->uring_lock);
2374 mutex_lock(&ctx->uring_lock);
2377 ret = io_iopoll_getevents(ctx, &nr_events, min);
2381 } while (min && !nr_events && !need_resched());
2383 mutex_unlock(&ctx->uring_lock);
2387 static void kiocb_end_write(struct io_kiocb *req)
2390 * Tell lockdep we inherited freeze protection from submission
2393 if (req->flags & REQ_F_ISREG) {
2394 struct inode *inode = file_inode(req->file);
2396 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2398 file_end_write(req->file);
2401 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2402 struct io_comp_state *cs)
2404 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 != req->result)
2411 req_set_fail_links(req);
2412 if (req->flags & REQ_F_BUFFER_SELECTED)
2413 cflags = io_put_rw_kbuf(req);
2414 __io_req_complete(req, res, cflags, cs);
2418 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2420 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2421 ssize_t ret = -ECANCELED;
2422 struct iov_iter iter;
2430 switch (req->opcode) {
2431 case IORING_OP_READV:
2432 case IORING_OP_READ_FIXED:
2433 case IORING_OP_READ:
2436 case IORING_OP_WRITEV:
2437 case IORING_OP_WRITE_FIXED:
2438 case IORING_OP_WRITE:
2442 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2447 if (!req->async_data) {
2448 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2451 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2459 req_set_fail_links(req);
2460 io_req_complete(req, ret);
2465 static bool io_rw_reissue(struct io_kiocb *req, long res)
2468 umode_t mode = file_inode(req->file)->i_mode;
2471 if (!S_ISBLK(mode) && !S_ISREG(mode))
2473 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2476 ret = io_sq_thread_acquire_mm(req->ctx, req);
2478 if (io_resubmit_prep(req, ret)) {
2479 refcount_inc(&req->refs);
2480 io_queue_async_work(req);
2488 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2489 struct io_comp_state *cs)
2491 if (!io_rw_reissue(req, res))
2492 io_complete_rw_common(&req->rw.kiocb, res, cs);
2495 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2497 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2499 __io_complete_rw(req, res, res2, NULL);
2502 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2504 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2506 if (kiocb->ki_flags & IOCB_WRITE)
2507 kiocb_end_write(req);
2509 if (res != -EAGAIN && res != req->result)
2510 req_set_fail_links(req);
2512 WRITE_ONCE(req->result, res);
2513 /* order with io_poll_complete() checking ->result */
2515 WRITE_ONCE(req->iopoll_completed, 1);
2519 * After the iocb has been issued, it's safe to be found on the poll list.
2520 * Adding the kiocb to the list AFTER submission ensures that we don't
2521 * find it from a io_iopoll_getevents() thread before the issuer is done
2522 * accessing the kiocb cookie.
2524 static void io_iopoll_req_issued(struct io_kiocb *req)
2526 struct io_ring_ctx *ctx = req->ctx;
2529 * Track whether we have multiple files in our lists. This will impact
2530 * how we do polling eventually, not spinning if we're on potentially
2531 * different devices.
2533 if (list_empty(&ctx->iopoll_list)) {
2534 ctx->poll_multi_file = false;
2535 } else if (!ctx->poll_multi_file) {
2536 struct io_kiocb *list_req;
2538 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2540 if (list_req->file != req->file)
2541 ctx->poll_multi_file = true;
2545 * For fast devices, IO may have already completed. If it has, add
2546 * it to the front so we find it first.
2548 if (READ_ONCE(req->iopoll_completed))
2549 list_add(&req->inflight_entry, &ctx->iopoll_list);
2551 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2553 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2554 wq_has_sleeper(&ctx->sq_data->wait))
2555 wake_up(&ctx->sq_data->wait);
2558 static void __io_state_file_put(struct io_submit_state *state)
2560 if (state->has_refs)
2561 fput_many(state->file, state->has_refs);
2565 static inline void io_state_file_put(struct io_submit_state *state)
2568 __io_state_file_put(state);
2572 * Get as many references to a file as we have IOs left in this submission,
2573 * assuming most submissions are for one file, or at least that each file
2574 * has more than one submission.
2576 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2582 if (state->fd == fd) {
2587 __io_state_file_put(state);
2589 state->file = fget_many(fd, state->ios_left);
2595 state->has_refs = state->ios_left;
2599 static bool io_bdev_nowait(struct block_device *bdev)
2602 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2609 * If we tracked the file through the SCM inflight mechanism, we could support
2610 * any file. For now, just ensure that anything potentially problematic is done
2613 static bool io_file_supports_async(struct file *file, int rw)
2615 umode_t mode = file_inode(file)->i_mode;
2617 if (S_ISBLK(mode)) {
2618 if (io_bdev_nowait(file->f_inode->i_bdev))
2622 if (S_ISCHR(mode) || S_ISSOCK(mode))
2624 if (S_ISREG(mode)) {
2625 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2626 file->f_op != &io_uring_fops)
2631 /* any ->read/write should understand O_NONBLOCK */
2632 if (file->f_flags & O_NONBLOCK)
2635 if (!(file->f_mode & FMODE_NOWAIT))
2639 return file->f_op->read_iter != NULL;
2641 return file->f_op->write_iter != NULL;
2644 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2646 struct io_ring_ctx *ctx = req->ctx;
2647 struct kiocb *kiocb = &req->rw.kiocb;
2651 if (S_ISREG(file_inode(req->file)->i_mode))
2652 req->flags |= REQ_F_ISREG;
2654 kiocb->ki_pos = READ_ONCE(sqe->off);
2655 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2656 req->flags |= REQ_F_CUR_POS;
2657 kiocb->ki_pos = req->file->f_pos;
2659 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2660 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2661 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2665 ioprio = READ_ONCE(sqe->ioprio);
2667 ret = ioprio_check_cap(ioprio);
2671 kiocb->ki_ioprio = ioprio;
2673 kiocb->ki_ioprio = get_current_ioprio();
2675 /* don't allow async punt if RWF_NOWAIT was requested */
2676 if (kiocb->ki_flags & IOCB_NOWAIT)
2677 req->flags |= REQ_F_NOWAIT;
2679 if (ctx->flags & IORING_SETUP_IOPOLL) {
2680 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2681 !kiocb->ki_filp->f_op->iopoll)
2684 kiocb->ki_flags |= IOCB_HIPRI;
2685 kiocb->ki_complete = io_complete_rw_iopoll;
2686 req->iopoll_completed = 0;
2688 if (kiocb->ki_flags & IOCB_HIPRI)
2690 kiocb->ki_complete = io_complete_rw;
2693 req->rw.addr = READ_ONCE(sqe->addr);
2694 req->rw.len = READ_ONCE(sqe->len);
2695 req->buf_index = READ_ONCE(sqe->buf_index);
2699 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2705 case -ERESTARTNOINTR:
2706 case -ERESTARTNOHAND:
2707 case -ERESTART_RESTARTBLOCK:
2709 * We can't just restart the syscall, since previously
2710 * submitted sqes may already be in progress. Just fail this
2716 kiocb->ki_complete(kiocb, ret, 0);
2720 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2721 struct io_comp_state *cs)
2723 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2724 struct io_async_rw *io = req->async_data;
2726 /* add previously done IO, if any */
2727 if (io && io->bytes_done > 0) {
2729 ret = io->bytes_done;
2731 ret += io->bytes_done;
2734 if (req->flags & REQ_F_CUR_POS)
2735 req->file->f_pos = kiocb->ki_pos;
2736 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2737 __io_complete_rw(req, ret, 0, cs);
2739 io_rw_done(kiocb, ret);
2742 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2743 struct iov_iter *iter)
2745 struct io_ring_ctx *ctx = req->ctx;
2746 size_t len = req->rw.len;
2747 struct io_mapped_ubuf *imu;
2748 u16 index, buf_index = req->buf_index;
2752 if (unlikely(buf_index >= ctx->nr_user_bufs))
2754 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2755 imu = &ctx->user_bufs[index];
2756 buf_addr = req->rw.addr;
2759 if (buf_addr + len < buf_addr)
2761 /* not inside the mapped region */
2762 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2766 * May not be a start of buffer, set size appropriately
2767 * and advance us to the beginning.
2769 offset = buf_addr - imu->ubuf;
2770 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2774 * Don't use iov_iter_advance() here, as it's really slow for
2775 * using the latter parts of a big fixed buffer - it iterates
2776 * over each segment manually. We can cheat a bit here, because
2779 * 1) it's a BVEC iter, we set it up
2780 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2781 * first and last bvec
2783 * So just find our index, and adjust the iterator afterwards.
2784 * If the offset is within the first bvec (or the whole first
2785 * bvec, just use iov_iter_advance(). This makes it easier
2786 * since we can just skip the first segment, which may not
2787 * be PAGE_SIZE aligned.
2789 const struct bio_vec *bvec = imu->bvec;
2791 if (offset <= bvec->bv_len) {
2792 iov_iter_advance(iter, offset);
2794 unsigned long seg_skip;
2796 /* skip first vec */
2797 offset -= bvec->bv_len;
2798 seg_skip = 1 + (offset >> PAGE_SHIFT);
2800 iter->bvec = bvec + seg_skip;
2801 iter->nr_segs -= seg_skip;
2802 iter->count -= bvec->bv_len + offset;
2803 iter->iov_offset = offset & ~PAGE_MASK;
2810 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2813 mutex_unlock(&ctx->uring_lock);
2816 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2819 * "Normal" inline submissions always hold the uring_lock, since we
2820 * grab it from the system call. Same is true for the SQPOLL offload.
2821 * The only exception is when we've detached the request and issue it
2822 * from an async worker thread, grab the lock for that case.
2825 mutex_lock(&ctx->uring_lock);
2828 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2829 int bgid, struct io_buffer *kbuf,
2832 struct io_buffer *head;
2834 if (req->flags & REQ_F_BUFFER_SELECTED)
2837 io_ring_submit_lock(req->ctx, needs_lock);
2839 lockdep_assert_held(&req->ctx->uring_lock);
2841 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2843 if (!list_empty(&head->list)) {
2844 kbuf = list_last_entry(&head->list, struct io_buffer,
2846 list_del(&kbuf->list);
2849 idr_remove(&req->ctx->io_buffer_idr, bgid);
2851 if (*len > kbuf->len)
2854 kbuf = ERR_PTR(-ENOBUFS);
2857 io_ring_submit_unlock(req->ctx, needs_lock);
2862 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2865 struct io_buffer *kbuf;
2868 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2869 bgid = req->buf_index;
2870 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2873 req->rw.addr = (u64) (unsigned long) kbuf;
2874 req->flags |= REQ_F_BUFFER_SELECTED;
2875 return u64_to_user_ptr(kbuf->addr);
2878 #ifdef CONFIG_COMPAT
2879 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2882 struct compat_iovec __user *uiov;
2883 compat_ssize_t clen;
2887 uiov = u64_to_user_ptr(req->rw.addr);
2888 if (!access_ok(uiov, sizeof(*uiov)))
2890 if (__get_user(clen, &uiov->iov_len))
2896 buf = io_rw_buffer_select(req, &len, needs_lock);
2898 return PTR_ERR(buf);
2899 iov[0].iov_base = buf;
2900 iov[0].iov_len = (compat_size_t) len;
2905 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2908 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2912 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2915 len = iov[0].iov_len;
2918 buf = io_rw_buffer_select(req, &len, needs_lock);
2920 return PTR_ERR(buf);
2921 iov[0].iov_base = buf;
2922 iov[0].iov_len = len;
2926 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2929 if (req->flags & REQ_F_BUFFER_SELECTED) {
2930 struct io_buffer *kbuf;
2932 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2933 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2934 iov[0].iov_len = kbuf->len;
2939 else if (req->rw.len > 1)
2942 #ifdef CONFIG_COMPAT
2943 if (req->ctx->compat)
2944 return io_compat_import(req, iov, needs_lock);
2947 return __io_iov_buffer_select(req, iov, needs_lock);
2950 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2951 struct iovec **iovec, struct iov_iter *iter,
2954 void __user *buf = u64_to_user_ptr(req->rw.addr);
2955 size_t sqe_len = req->rw.len;
2959 opcode = req->opcode;
2960 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2962 return io_import_fixed(req, rw, iter);
2965 /* buffer index only valid with fixed read/write, or buffer select */
2966 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2969 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2970 if (req->flags & REQ_F_BUFFER_SELECT) {
2971 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2973 return PTR_ERR(buf);
2974 req->rw.len = sqe_len;
2977 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2979 return ret < 0 ? ret : sqe_len;
2982 if (req->flags & REQ_F_BUFFER_SELECT) {
2983 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2985 ret = (*iovec)->iov_len;
2986 iov_iter_init(iter, rw, *iovec, 1, ret);
2992 #ifdef CONFIG_COMPAT
2993 if (req->ctx->compat)
2994 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2998 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
3001 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3002 struct iovec **iovec, struct iov_iter *iter,
3005 struct io_async_rw *iorw = req->async_data;
3008 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3010 return iov_iter_count(&iorw->iter);
3013 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3015 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3019 * For files that don't have ->read_iter() and ->write_iter(), handle them
3020 * by looping over ->read() or ->write() manually.
3022 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
3023 struct iov_iter *iter)
3028 * Don't support polled IO through this interface, and we can't
3029 * support non-blocking either. For the latter, this just causes
3030 * the kiocb to be handled from an async context.
3032 if (kiocb->ki_flags & IOCB_HIPRI)
3034 if (kiocb->ki_flags & IOCB_NOWAIT)
3037 while (iov_iter_count(iter)) {
3041 if (!iov_iter_is_bvec(iter)) {
3042 iovec = iov_iter_iovec(iter);
3044 /* fixed buffers import bvec */
3045 iovec.iov_base = kmap(iter->bvec->bv_page)
3047 iovec.iov_len = min(iter->count,
3048 iter->bvec->bv_len - iter->iov_offset);
3052 nr = file->f_op->read(file, iovec.iov_base,
3053 iovec.iov_len, io_kiocb_ppos(kiocb));
3055 nr = file->f_op->write(file, iovec.iov_base,
3056 iovec.iov_len, io_kiocb_ppos(kiocb));
3059 if (iov_iter_is_bvec(iter))
3060 kunmap(iter->bvec->bv_page);
3068 if (nr != iovec.iov_len)
3070 iov_iter_advance(iter, nr);
3076 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3077 const struct iovec *fast_iov, struct iov_iter *iter)
3079 struct io_async_rw *rw = req->async_data;
3081 memcpy(&rw->iter, iter, sizeof(*iter));
3082 rw->free_iovec = iovec;
3084 /* can only be fixed buffers, no need to do anything */
3085 if (iter->type == ITER_BVEC)
3088 unsigned iov_off = 0;
3090 rw->iter.iov = rw->fast_iov;
3091 if (iter->iov != fast_iov) {
3092 iov_off = iter->iov - fast_iov;
3093 rw->iter.iov += iov_off;
3095 if (rw->fast_iov != fast_iov)
3096 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3097 sizeof(struct iovec) * iter->nr_segs);
3099 req->flags |= REQ_F_NEED_CLEANUP;
3103 static inline int __io_alloc_async_data(struct io_kiocb *req)
3105 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3106 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3107 return req->async_data == NULL;
3110 static int io_alloc_async_data(struct io_kiocb *req)
3112 if (!io_op_defs[req->opcode].needs_async_data)
3115 return __io_alloc_async_data(req);
3118 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3119 const struct iovec *fast_iov,
3120 struct iov_iter *iter, bool force)
3122 if (!force && !io_op_defs[req->opcode].needs_async_data)
3124 if (!req->async_data) {
3125 if (__io_alloc_async_data(req))
3128 io_req_map_rw(req, iovec, fast_iov, iter);
3133 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3135 struct io_async_rw *iorw = req->async_data;
3136 struct iovec *iov = iorw->fast_iov;
3139 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3140 if (unlikely(ret < 0))
3143 iorw->bytes_done = 0;
3144 iorw->free_iovec = iov;
3146 req->flags |= REQ_F_NEED_CLEANUP;
3150 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3154 ret = io_prep_rw(req, sqe);
3158 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3161 /* either don't need iovec imported or already have it */
3162 if (!req->async_data)
3164 return io_rw_prep_async(req, READ);
3168 * This is our waitqueue callback handler, registered through lock_page_async()
3169 * when we initially tried to do the IO with the iocb armed our waitqueue.
3170 * This gets called when the page is unlocked, and we generally expect that to
3171 * happen when the page IO is completed and the page is now uptodate. This will
3172 * queue a task_work based retry of the operation, attempting to copy the data
3173 * again. If the latter fails because the page was NOT uptodate, then we will
3174 * do a thread based blocking retry of the operation. That's the unexpected
3177 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3178 int sync, void *arg)
3180 struct wait_page_queue *wpq;
3181 struct io_kiocb *req = wait->private;
3182 struct wait_page_key *key = arg;
3185 wpq = container_of(wait, struct wait_page_queue, wait);
3187 if (!wake_page_match(wpq, key))
3190 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3191 list_del_init(&wait->entry);
3193 init_task_work(&req->task_work, io_req_task_submit);
3194 percpu_ref_get(&req->ctx->refs);
3196 /* submit ref gets dropped, acquire a new one */
3197 refcount_inc(&req->refs);
3198 ret = io_req_task_work_add(req, true);
3199 if (unlikely(ret)) {
3200 struct task_struct *tsk;
3202 /* queue just for cancelation */
3203 init_task_work(&req->task_work, io_req_task_cancel);
3204 tsk = io_wq_get_task(req->ctx->io_wq);
3205 task_work_add(tsk, &req->task_work, 0);
3206 wake_up_process(tsk);
3212 * This controls whether a given IO request should be armed for async page
3213 * based retry. If we return false here, the request is handed to the async
3214 * worker threads for retry. If we're doing buffered reads on a regular file,
3215 * we prepare a private wait_page_queue entry and retry the operation. This
3216 * will either succeed because the page is now uptodate and unlocked, or it
3217 * will register a callback when the page is unlocked at IO completion. Through
3218 * that callback, io_uring uses task_work to setup a retry of the operation.
3219 * That retry will attempt the buffered read again. The retry will generally
3220 * succeed, or in rare cases where it fails, we then fall back to using the
3221 * async worker threads for a blocking retry.
3223 static bool io_rw_should_retry(struct io_kiocb *req)
3225 struct io_async_rw *rw = req->async_data;
3226 struct wait_page_queue *wait = &rw->wpq;
3227 struct kiocb *kiocb = &req->rw.kiocb;
3229 /* never retry for NOWAIT, we just complete with -EAGAIN */
3230 if (req->flags & REQ_F_NOWAIT)
3233 /* Only for buffered IO */
3234 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3238 * just use poll if we can, and don't attempt if the fs doesn't
3239 * support callback based unlocks
3241 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3244 wait->wait.func = io_async_buf_func;
3245 wait->wait.private = req;
3246 wait->wait.flags = 0;
3247 INIT_LIST_HEAD(&wait->wait.entry);
3248 kiocb->ki_flags |= IOCB_WAITQ;
3249 kiocb->ki_flags &= ~IOCB_NOWAIT;
3250 kiocb->ki_waitq = wait;
3254 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3256 if (req->file->f_op->read_iter)
3257 return call_read_iter(req->file, &req->rw.kiocb, iter);
3258 else if (req->file->f_op->read)
3259 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3264 static int io_read(struct io_kiocb *req, bool force_nonblock,
3265 struct io_comp_state *cs)
3267 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3268 struct kiocb *kiocb = &req->rw.kiocb;
3269 struct iov_iter __iter, *iter = &__iter;
3270 struct io_async_rw *rw = req->async_data;
3271 ssize_t io_size, ret, ret2;
3278 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3281 iov_count = iov_iter_count(iter);
3283 req->result = io_size;
3286 /* Ensure we clear previously set non-block flag */
3287 if (!force_nonblock)
3288 kiocb->ki_flags &= ~IOCB_NOWAIT;
3290 kiocb->ki_flags |= IOCB_NOWAIT;
3293 /* If the file doesn't support async, just async punt */
3294 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3298 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3302 ret = io_iter_do_read(req, iter);
3306 } else if (ret == -EIOCBQUEUED) {
3309 } else if (ret == -EAGAIN) {
3310 /* IOPOLL retry should happen for io-wq threads */
3311 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3313 /* no retry on NONBLOCK marked file */
3314 if (req->file->f_flags & O_NONBLOCK)
3316 /* some cases will consume bytes even on error returns */
3317 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3320 } else if (ret < 0) {
3321 /* make sure -ERESTARTSYS -> -EINTR is done */
3325 /* read it all, or we did blocking attempt. no retry. */
3326 if (!iov_iter_count(iter) || !force_nonblock ||
3327 (req->file->f_flags & O_NONBLOCK))
3332 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3339 rw = req->async_data;
3340 /* it's copied and will be cleaned with ->io */
3342 /* now use our persistent iterator, if we aren't already */
3345 rw->bytes_done += ret;
3346 /* if we can retry, do so with the callbacks armed */
3347 if (!io_rw_should_retry(req)) {
3348 kiocb->ki_flags &= ~IOCB_WAITQ;
3353 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3354 * get -EIOCBQUEUED, then we'll get a notification when the desired
3355 * page gets unlocked. We can also get a partial read here, and if we
3356 * do, then just retry at the new offset.
3358 ret = io_iter_do_read(req, iter);
3359 if (ret == -EIOCBQUEUED) {
3362 } else if (ret > 0 && ret < io_size) {
3363 /* we got some bytes, but not all. retry. */
3367 kiocb_done(kiocb, ret, cs);
3370 /* it's reportedly faster than delegating the null check to kfree() */
3376 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3380 ret = io_prep_rw(req, sqe);
3384 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3387 /* either don't need iovec imported or already have it */
3388 if (!req->async_data)
3390 return io_rw_prep_async(req, WRITE);
3393 static int io_write(struct io_kiocb *req, bool force_nonblock,
3394 struct io_comp_state *cs)
3396 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3397 struct kiocb *kiocb = &req->rw.kiocb;
3398 struct iov_iter __iter, *iter = &__iter;
3399 struct io_async_rw *rw = req->async_data;
3401 ssize_t ret, ret2, io_size;
3406 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3409 iov_count = iov_iter_count(iter);
3411 req->result = io_size;
3413 /* Ensure we clear previously set non-block flag */
3414 if (!force_nonblock)
3415 kiocb->ki_flags &= ~IOCB_NOWAIT;
3417 kiocb->ki_flags |= IOCB_NOWAIT;
3419 /* If the file doesn't support async, just async punt */
3420 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3423 /* file path doesn't support NOWAIT for non-direct_IO */
3424 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3425 (req->flags & REQ_F_ISREG))
3428 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3433 * Open-code file_start_write here to grab freeze protection,
3434 * which will be released by another thread in
3435 * io_complete_rw(). Fool lockdep by telling it the lock got
3436 * released so that it doesn't complain about the held lock when
3437 * we return to userspace.
3439 if (req->flags & REQ_F_ISREG) {
3440 __sb_start_write(file_inode(req->file)->i_sb,
3441 SB_FREEZE_WRITE, true);
3442 __sb_writers_release(file_inode(req->file)->i_sb,
3445 kiocb->ki_flags |= IOCB_WRITE;
3447 if (req->file->f_op->write_iter)
3448 ret2 = call_write_iter(req->file, kiocb, iter);
3449 else if (req->file->f_op->write)
3450 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3455 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3456 * retry them without IOCB_NOWAIT.
3458 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3460 /* no retry on NONBLOCK marked file */
3461 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3463 if (!force_nonblock || ret2 != -EAGAIN) {
3464 /* IOPOLL retry should happen for io-wq threads */
3465 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3468 kiocb_done(kiocb, ret2, cs);
3471 /* some cases will consume bytes even on error returns */
3472 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3473 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3478 /* it's reportedly faster than delegating the null check to kfree() */
3484 static int __io_splice_prep(struct io_kiocb *req,
3485 const struct io_uring_sqe *sqe)
3487 struct io_splice* sp = &req->splice;
3488 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3491 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3495 sp->len = READ_ONCE(sqe->len);
3496 sp->flags = READ_ONCE(sqe->splice_flags);
3498 if (unlikely(sp->flags & ~valid_flags))
3501 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3502 (sp->flags & SPLICE_F_FD_IN_FIXED));
3505 req->flags |= REQ_F_NEED_CLEANUP;
3507 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3509 * Splice operation will be punted aync, and here need to
3510 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3512 io_req_init_async(req);
3513 req->work.flags |= IO_WQ_WORK_UNBOUND;
3519 static int io_tee_prep(struct io_kiocb *req,
3520 const struct io_uring_sqe *sqe)
3522 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3524 return __io_splice_prep(req, sqe);
3527 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3529 struct io_splice *sp = &req->splice;
3530 struct file *in = sp->file_in;
3531 struct file *out = sp->file_out;
3532 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3538 ret = do_tee(in, out, sp->len, flags);
3540 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3541 req->flags &= ~REQ_F_NEED_CLEANUP;
3544 req_set_fail_links(req);
3545 io_req_complete(req, ret);
3549 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3551 struct io_splice* sp = &req->splice;
3553 sp->off_in = READ_ONCE(sqe->splice_off_in);
3554 sp->off_out = READ_ONCE(sqe->off);
3555 return __io_splice_prep(req, sqe);
3558 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3560 struct io_splice *sp = &req->splice;
3561 struct file *in = sp->file_in;
3562 struct file *out = sp->file_out;
3563 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3564 loff_t *poff_in, *poff_out;
3570 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3571 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3574 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3576 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3577 req->flags &= ~REQ_F_NEED_CLEANUP;
3580 req_set_fail_links(req);
3581 io_req_complete(req, ret);
3586 * IORING_OP_NOP just posts a completion event, nothing else.
3588 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3590 struct io_ring_ctx *ctx = req->ctx;
3592 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3595 __io_req_complete(req, 0, 0, cs);
3599 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3601 struct io_ring_ctx *ctx = req->ctx;
3606 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3608 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3611 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3612 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3615 req->sync.off = READ_ONCE(sqe->off);
3616 req->sync.len = READ_ONCE(sqe->len);
3620 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3622 loff_t end = req->sync.off + req->sync.len;
3625 /* fsync always requires a blocking context */
3629 ret = vfs_fsync_range(req->file, req->sync.off,
3630 end > 0 ? end : LLONG_MAX,
3631 req->sync.flags & IORING_FSYNC_DATASYNC);
3633 req_set_fail_links(req);
3634 io_req_complete(req, ret);
3638 static int io_fallocate_prep(struct io_kiocb *req,
3639 const struct io_uring_sqe *sqe)
3641 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3643 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3646 req->sync.off = READ_ONCE(sqe->off);
3647 req->sync.len = READ_ONCE(sqe->addr);
3648 req->sync.mode = READ_ONCE(sqe->len);
3652 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3656 /* fallocate always requiring blocking context */
3659 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3662 req_set_fail_links(req);
3663 io_req_complete(req, ret);
3667 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3669 const char __user *fname;
3672 if (unlikely(sqe->ioprio || sqe->buf_index))
3674 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3677 /* open.how should be already initialised */
3678 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3679 req->open.how.flags |= O_LARGEFILE;
3681 req->open.dfd = READ_ONCE(sqe->fd);
3682 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3683 req->open.filename = getname(fname);
3684 if (IS_ERR(req->open.filename)) {
3685 ret = PTR_ERR(req->open.filename);
3686 req->open.filename = NULL;
3689 req->open.nofile = rlimit(RLIMIT_NOFILE);
3690 req->flags |= REQ_F_NEED_CLEANUP;
3694 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3698 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3700 mode = READ_ONCE(sqe->len);
3701 flags = READ_ONCE(sqe->open_flags);
3702 req->open.how = build_open_how(flags, mode);
3703 return __io_openat_prep(req, sqe);
3706 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3708 struct open_how __user *how;
3712 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3714 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3715 len = READ_ONCE(sqe->len);
3716 if (len < OPEN_HOW_SIZE_VER0)
3719 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3724 return __io_openat_prep(req, sqe);
3727 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3729 struct open_flags op;
3736 ret = build_open_flags(&req->open.how, &op);
3740 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3744 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3747 ret = PTR_ERR(file);
3749 fsnotify_open(file);
3750 fd_install(ret, file);
3753 putname(req->open.filename);
3754 req->flags &= ~REQ_F_NEED_CLEANUP;
3756 req_set_fail_links(req);
3757 io_req_complete(req, ret);
3761 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3763 return io_openat2(req, force_nonblock);
3766 static int io_remove_buffers_prep(struct io_kiocb *req,
3767 const struct io_uring_sqe *sqe)
3769 struct io_provide_buf *p = &req->pbuf;
3772 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3775 tmp = READ_ONCE(sqe->fd);
3776 if (!tmp || tmp > USHRT_MAX)
3779 memset(p, 0, sizeof(*p));
3781 p->bgid = READ_ONCE(sqe->buf_group);
3785 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3786 int bgid, unsigned nbufs)
3790 /* shouldn't happen */
3794 /* the head kbuf is the list itself */
3795 while (!list_empty(&buf->list)) {
3796 struct io_buffer *nxt;
3798 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3799 list_del(&nxt->list);
3806 idr_remove(&ctx->io_buffer_idr, bgid);
3811 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3812 struct io_comp_state *cs)
3814 struct io_provide_buf *p = &req->pbuf;
3815 struct io_ring_ctx *ctx = req->ctx;
3816 struct io_buffer *head;
3819 io_ring_submit_lock(ctx, !force_nonblock);
3821 lockdep_assert_held(&ctx->uring_lock);
3824 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3826 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3828 io_ring_submit_lock(ctx, !force_nonblock);
3830 req_set_fail_links(req);
3831 __io_req_complete(req, ret, 0, cs);
3835 static int io_provide_buffers_prep(struct io_kiocb *req,
3836 const struct io_uring_sqe *sqe)
3838 struct io_provide_buf *p = &req->pbuf;
3841 if (sqe->ioprio || sqe->rw_flags)
3844 tmp = READ_ONCE(sqe->fd);
3845 if (!tmp || tmp > USHRT_MAX)
3848 p->addr = READ_ONCE(sqe->addr);
3849 p->len = READ_ONCE(sqe->len);
3851 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3854 p->bgid = READ_ONCE(sqe->buf_group);
3855 tmp = READ_ONCE(sqe->off);
3856 if (tmp > USHRT_MAX)
3862 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3864 struct io_buffer *buf;
3865 u64 addr = pbuf->addr;
3866 int i, bid = pbuf->bid;
3868 for (i = 0; i < pbuf->nbufs; i++) {
3869 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3874 buf->len = pbuf->len;
3879 INIT_LIST_HEAD(&buf->list);
3882 list_add_tail(&buf->list, &(*head)->list);
3886 return i ? i : -ENOMEM;
3889 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3890 struct io_comp_state *cs)
3892 struct io_provide_buf *p = &req->pbuf;
3893 struct io_ring_ctx *ctx = req->ctx;
3894 struct io_buffer *head, *list;
3897 io_ring_submit_lock(ctx, !force_nonblock);
3899 lockdep_assert_held(&ctx->uring_lock);
3901 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3903 ret = io_add_buffers(p, &head);
3908 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3911 __io_remove_buffers(ctx, head, p->bgid, -1U);
3916 io_ring_submit_unlock(ctx, !force_nonblock);
3918 req_set_fail_links(req);
3919 __io_req_complete(req, ret, 0, cs);
3923 static int io_epoll_ctl_prep(struct io_kiocb *req,
3924 const struct io_uring_sqe *sqe)
3926 #if defined(CONFIG_EPOLL)
3927 if (sqe->ioprio || sqe->buf_index)
3929 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3932 req->epoll.epfd = READ_ONCE(sqe->fd);
3933 req->epoll.op = READ_ONCE(sqe->len);
3934 req->epoll.fd = READ_ONCE(sqe->off);
3936 if (ep_op_has_event(req->epoll.op)) {
3937 struct epoll_event __user *ev;
3939 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3940 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3950 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3951 struct io_comp_state *cs)
3953 #if defined(CONFIG_EPOLL)
3954 struct io_epoll *ie = &req->epoll;
3957 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3958 if (force_nonblock && ret == -EAGAIN)
3962 req_set_fail_links(req);
3963 __io_req_complete(req, ret, 0, cs);
3970 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3972 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3973 if (sqe->ioprio || sqe->buf_index || sqe->off)
3975 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3978 req->madvise.addr = READ_ONCE(sqe->addr);
3979 req->madvise.len = READ_ONCE(sqe->len);
3980 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3987 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3989 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3990 struct io_madvise *ma = &req->madvise;
3996 ret = do_madvise(ma->addr, ma->len, ma->advice);
3998 req_set_fail_links(req);
3999 io_req_complete(req, ret);
4006 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4008 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4010 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4013 req->fadvise.offset = READ_ONCE(sqe->off);
4014 req->fadvise.len = READ_ONCE(sqe->len);
4015 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4019 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4021 struct io_fadvise *fa = &req->fadvise;
4024 if (force_nonblock) {
4025 switch (fa->advice) {
4026 case POSIX_FADV_NORMAL:
4027 case POSIX_FADV_RANDOM:
4028 case POSIX_FADV_SEQUENTIAL:
4035 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4037 req_set_fail_links(req);
4038 io_req_complete(req, ret);
4042 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4044 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4046 if (sqe->ioprio || sqe->buf_index)
4048 if (req->flags & REQ_F_FIXED_FILE)
4051 req->statx.dfd = READ_ONCE(sqe->fd);
4052 req->statx.mask = READ_ONCE(sqe->len);
4053 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4054 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4055 req->statx.flags = READ_ONCE(sqe->statx_flags);
4060 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4062 struct io_statx *ctx = &req->statx;
4065 if (force_nonblock) {
4066 /* only need file table for an actual valid fd */
4067 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4068 req->flags |= REQ_F_NO_FILE_TABLE;
4072 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4076 req_set_fail_links(req);
4077 io_req_complete(req, ret);
4081 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4084 * If we queue this for async, it must not be cancellable. That would
4085 * leave the 'file' in an undeterminate state, and here need to modify
4086 * io_wq_work.flags, so initialize io_wq_work firstly.
4088 io_req_init_async(req);
4089 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4091 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4093 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4094 sqe->rw_flags || sqe->buf_index)
4096 if (req->flags & REQ_F_FIXED_FILE)
4099 req->close.fd = READ_ONCE(sqe->fd);
4100 if ((req->file && req->file->f_op == &io_uring_fops))
4103 req->close.put_file = NULL;
4107 static int io_close(struct io_kiocb *req, bool force_nonblock,
4108 struct io_comp_state *cs)
4110 struct io_close *close = &req->close;
4113 /* might be already done during nonblock submission */
4114 if (!close->put_file) {
4115 ret = __close_fd_get_file(close->fd, &close->put_file);
4117 return (ret == -ENOENT) ? -EBADF : ret;
4120 /* if the file has a flush method, be safe and punt to async */
4121 if (close->put_file->f_op->flush && force_nonblock) {
4122 /* was never set, but play safe */
4123 req->flags &= ~REQ_F_NOWAIT;
4124 /* avoid grabbing files - we don't need the files */
4125 req->flags |= REQ_F_NO_FILE_TABLE;
4129 /* No ->flush() or already async, safely close from here */
4130 ret = filp_close(close->put_file, req->work.files);
4132 req_set_fail_links(req);
4133 fput(close->put_file);
4134 close->put_file = NULL;
4135 __io_req_complete(req, ret, 0, cs);
4139 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4141 struct io_ring_ctx *ctx = req->ctx;
4146 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4148 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4151 req->sync.off = READ_ONCE(sqe->off);
4152 req->sync.len = READ_ONCE(sqe->len);
4153 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4157 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4161 /* sync_file_range always requires a blocking context */
4165 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4168 req_set_fail_links(req);
4169 io_req_complete(req, ret);
4173 #if defined(CONFIG_NET)
4174 static int io_setup_async_msg(struct io_kiocb *req,
4175 struct io_async_msghdr *kmsg)
4177 struct io_async_msghdr *async_msg = req->async_data;
4181 if (io_alloc_async_data(req)) {
4182 if (kmsg->iov != kmsg->fast_iov)
4186 async_msg = req->async_data;
4187 req->flags |= REQ_F_NEED_CLEANUP;
4188 memcpy(async_msg, kmsg, sizeof(*kmsg));
4192 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4193 struct io_async_msghdr *iomsg)
4195 iomsg->iov = iomsg->fast_iov;
4196 iomsg->msg.msg_name = &iomsg->addr;
4197 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4198 req->sr_msg.msg_flags, &iomsg->iov);
4201 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4203 struct io_async_msghdr *async_msg = req->async_data;
4204 struct io_sr_msg *sr = &req->sr_msg;
4207 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4210 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4211 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4212 sr->len = READ_ONCE(sqe->len);
4214 #ifdef CONFIG_COMPAT
4215 if (req->ctx->compat)
4216 sr->msg_flags |= MSG_CMSG_COMPAT;
4219 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4221 ret = io_sendmsg_copy_hdr(req, async_msg);
4223 req->flags |= REQ_F_NEED_CLEANUP;
4227 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4228 struct io_comp_state *cs)
4230 struct io_async_msghdr iomsg, *kmsg;
4231 struct socket *sock;
4235 sock = sock_from_file(req->file, &ret);
4236 if (unlikely(!sock))
4239 if (req->async_data) {
4240 kmsg = req->async_data;
4241 kmsg->msg.msg_name = &kmsg->addr;
4242 /* if iov is set, it's allocated already */
4244 kmsg->iov = kmsg->fast_iov;
4245 kmsg->msg.msg_iter.iov = kmsg->iov;
4247 ret = io_sendmsg_copy_hdr(req, &iomsg);
4253 flags = req->sr_msg.msg_flags;
4254 if (flags & MSG_DONTWAIT)
4255 req->flags |= REQ_F_NOWAIT;
4256 else if (force_nonblock)
4257 flags |= MSG_DONTWAIT;
4259 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4260 if (force_nonblock && ret == -EAGAIN)
4261 return io_setup_async_msg(req, kmsg);
4262 if (ret == -ERESTARTSYS)
4265 if (kmsg->iov != kmsg->fast_iov)
4267 req->flags &= ~REQ_F_NEED_CLEANUP;
4269 req_set_fail_links(req);
4270 __io_req_complete(req, ret, 0, cs);
4274 static int io_send(struct io_kiocb *req, bool force_nonblock,
4275 struct io_comp_state *cs)
4277 struct io_sr_msg *sr = &req->sr_msg;
4280 struct socket *sock;
4284 sock = sock_from_file(req->file, &ret);
4285 if (unlikely(!sock))
4288 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4292 msg.msg_name = NULL;
4293 msg.msg_control = NULL;
4294 msg.msg_controllen = 0;
4295 msg.msg_namelen = 0;
4297 flags = req->sr_msg.msg_flags;
4298 if (flags & MSG_DONTWAIT)
4299 req->flags |= REQ_F_NOWAIT;
4300 else if (force_nonblock)
4301 flags |= MSG_DONTWAIT;
4303 msg.msg_flags = flags;
4304 ret = sock_sendmsg(sock, &msg);
4305 if (force_nonblock && ret == -EAGAIN)
4307 if (ret == -ERESTARTSYS)
4311 req_set_fail_links(req);
4312 __io_req_complete(req, ret, 0, cs);
4316 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4317 struct io_async_msghdr *iomsg)
4319 struct io_sr_msg *sr = &req->sr_msg;
4320 struct iovec __user *uiov;
4324 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4325 &iomsg->uaddr, &uiov, &iov_len);
4329 if (req->flags & REQ_F_BUFFER_SELECT) {
4332 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4334 sr->len = iomsg->iov[0].iov_len;
4335 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4339 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4340 &iomsg->iov, &iomsg->msg.msg_iter);
4348 #ifdef CONFIG_COMPAT
4349 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4350 struct io_async_msghdr *iomsg)
4352 struct compat_msghdr __user *msg_compat;
4353 struct io_sr_msg *sr = &req->sr_msg;
4354 struct compat_iovec __user *uiov;
4359 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4360 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4365 uiov = compat_ptr(ptr);
4366 if (req->flags & REQ_F_BUFFER_SELECT) {
4367 compat_ssize_t clen;
4371 if (!access_ok(uiov, sizeof(*uiov)))
4373 if (__get_user(clen, &uiov->iov_len))
4377 sr->len = iomsg->iov[0].iov_len;
4380 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4382 &iomsg->msg.msg_iter);
4391 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4392 struct io_async_msghdr *iomsg)
4394 iomsg->msg.msg_name = &iomsg->addr;
4395 iomsg->iov = iomsg->fast_iov;
4397 #ifdef CONFIG_COMPAT
4398 if (req->ctx->compat)
4399 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4402 return __io_recvmsg_copy_hdr(req, iomsg);
4405 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4408 struct io_sr_msg *sr = &req->sr_msg;
4409 struct io_buffer *kbuf;
4411 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4416 req->flags |= REQ_F_BUFFER_SELECTED;
4420 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4422 return io_put_kbuf(req, req->sr_msg.kbuf);
4425 static int io_recvmsg_prep(struct io_kiocb *req,
4426 const struct io_uring_sqe *sqe)
4428 struct io_async_msghdr *async_msg = req->async_data;
4429 struct io_sr_msg *sr = &req->sr_msg;
4432 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4435 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4436 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4437 sr->len = READ_ONCE(sqe->len);
4438 sr->bgid = READ_ONCE(sqe->buf_group);
4440 #ifdef CONFIG_COMPAT
4441 if (req->ctx->compat)
4442 sr->msg_flags |= MSG_CMSG_COMPAT;
4445 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4447 ret = io_recvmsg_copy_hdr(req, async_msg);
4449 req->flags |= REQ_F_NEED_CLEANUP;
4453 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4454 struct io_comp_state *cs)
4456 struct io_async_msghdr iomsg, *kmsg;
4457 struct socket *sock;
4458 struct io_buffer *kbuf;
4460 int ret, cflags = 0;
4462 sock = sock_from_file(req->file, &ret);
4463 if (unlikely(!sock))
4466 if (req->async_data) {
4467 kmsg = req->async_data;
4468 kmsg->msg.msg_name = &kmsg->addr;
4469 /* if iov is set, it's allocated already */
4471 kmsg->iov = kmsg->fast_iov;
4472 kmsg->msg.msg_iter.iov = kmsg->iov;
4474 ret = io_recvmsg_copy_hdr(req, &iomsg);
4480 if (req->flags & REQ_F_BUFFER_SELECT) {
4481 kbuf = io_recv_buffer_select(req, !force_nonblock);
4483 return PTR_ERR(kbuf);
4484 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4485 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4486 1, req->sr_msg.len);
4489 flags = req->sr_msg.msg_flags;
4490 if (flags & MSG_DONTWAIT)
4491 req->flags |= REQ_F_NOWAIT;
4492 else if (force_nonblock)
4493 flags |= MSG_DONTWAIT;
4495 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4496 kmsg->uaddr, flags);
4497 if (force_nonblock && ret == -EAGAIN)
4498 return io_setup_async_msg(req, kmsg);
4499 if (ret == -ERESTARTSYS)
4502 if (req->flags & REQ_F_BUFFER_SELECTED)
4503 cflags = io_put_recv_kbuf(req);
4504 if (kmsg->iov != kmsg->fast_iov)
4506 req->flags &= ~REQ_F_NEED_CLEANUP;
4508 req_set_fail_links(req);
4509 __io_req_complete(req, ret, cflags, cs);
4513 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4514 struct io_comp_state *cs)
4516 struct io_buffer *kbuf;
4517 struct io_sr_msg *sr = &req->sr_msg;
4519 void __user *buf = sr->buf;
4520 struct socket *sock;
4523 int ret, cflags = 0;
4525 sock = sock_from_file(req->file, &ret);
4526 if (unlikely(!sock))
4529 if (req->flags & REQ_F_BUFFER_SELECT) {
4530 kbuf = io_recv_buffer_select(req, !force_nonblock);
4532 return PTR_ERR(kbuf);
4533 buf = u64_to_user_ptr(kbuf->addr);
4536 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4540 msg.msg_name = NULL;
4541 msg.msg_control = NULL;
4542 msg.msg_controllen = 0;
4543 msg.msg_namelen = 0;
4544 msg.msg_iocb = NULL;
4547 flags = req->sr_msg.msg_flags;
4548 if (flags & MSG_DONTWAIT)
4549 req->flags |= REQ_F_NOWAIT;
4550 else if (force_nonblock)
4551 flags |= MSG_DONTWAIT;
4553 ret = sock_recvmsg(sock, &msg, flags);
4554 if (force_nonblock && ret == -EAGAIN)
4556 if (ret == -ERESTARTSYS)
4559 if (req->flags & REQ_F_BUFFER_SELECTED)
4560 cflags = io_put_recv_kbuf(req);
4562 req_set_fail_links(req);
4563 __io_req_complete(req, ret, cflags, cs);
4567 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4569 struct io_accept *accept = &req->accept;
4571 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4573 if (sqe->ioprio || sqe->len || sqe->buf_index)
4576 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4577 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4578 accept->flags = READ_ONCE(sqe->accept_flags);
4579 accept->nofile = rlimit(RLIMIT_NOFILE);
4583 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4584 struct io_comp_state *cs)
4586 struct io_accept *accept = &req->accept;
4587 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4590 if (req->file->f_flags & O_NONBLOCK)
4591 req->flags |= REQ_F_NOWAIT;
4593 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4594 accept->addr_len, accept->flags,
4596 if (ret == -EAGAIN && force_nonblock)
4599 if (ret == -ERESTARTSYS)
4601 req_set_fail_links(req);
4603 __io_req_complete(req, ret, 0, cs);
4607 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4609 struct io_connect *conn = &req->connect;
4610 struct io_async_connect *io = req->async_data;
4612 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4614 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4617 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4618 conn->addr_len = READ_ONCE(sqe->addr2);
4623 return move_addr_to_kernel(conn->addr, conn->addr_len,
4627 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4628 struct io_comp_state *cs)
4630 struct io_async_connect __io, *io;
4631 unsigned file_flags;
4634 if (req->async_data) {
4635 io = req->async_data;
4637 ret = move_addr_to_kernel(req->connect.addr,
4638 req->connect.addr_len,
4645 file_flags = force_nonblock ? O_NONBLOCK : 0;
4647 ret = __sys_connect_file(req->file, &io->address,
4648 req->connect.addr_len, file_flags);
4649 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4650 if (req->async_data)
4652 if (io_alloc_async_data(req)) {
4656 io = req->async_data;
4657 memcpy(req->async_data, &__io, sizeof(__io));
4660 if (ret == -ERESTARTSYS)
4664 req_set_fail_links(req);
4665 __io_req_complete(req, ret, 0, cs);
4668 #else /* !CONFIG_NET */
4669 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4674 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4675 struct io_comp_state *cs)
4680 static int io_send(struct io_kiocb *req, bool force_nonblock,
4681 struct io_comp_state *cs)
4686 static int io_recvmsg_prep(struct io_kiocb *req,
4687 const struct io_uring_sqe *sqe)
4692 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4693 struct io_comp_state *cs)
4698 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4699 struct io_comp_state *cs)
4704 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4709 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4710 struct io_comp_state *cs)
4715 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4720 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4721 struct io_comp_state *cs)
4725 #endif /* CONFIG_NET */
4727 struct io_poll_table {
4728 struct poll_table_struct pt;
4729 struct io_kiocb *req;
4733 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4734 __poll_t mask, task_work_func_t func)
4739 /* for instances that support it check for an event match first: */
4740 if (mask && !(mask & poll->events))
4743 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4745 list_del_init(&poll->wait.entry);
4748 init_task_work(&req->task_work, func);
4749 percpu_ref_get(&req->ctx->refs);
4752 * If we using the signalfd wait_queue_head for this wakeup, then
4753 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4754 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4755 * either, as the normal wakeup will suffice.
4757 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4760 * If this fails, then the task is exiting. When a task exits, the
4761 * work gets canceled, so just cancel this request as well instead
4762 * of executing it. We can't safely execute it anyway, as we may not
4763 * have the needed state needed for it anyway.
4765 ret = io_req_task_work_add(req, twa_signal_ok);
4766 if (unlikely(ret)) {
4767 struct task_struct *tsk;
4769 WRITE_ONCE(poll->canceled, true);
4770 tsk = io_wq_get_task(req->ctx->io_wq);
4771 task_work_add(tsk, &req->task_work, 0);
4772 wake_up_process(tsk);
4777 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4778 __acquires(&req->ctx->completion_lock)
4780 struct io_ring_ctx *ctx = req->ctx;
4782 if (!req->result && !READ_ONCE(poll->canceled)) {
4783 struct poll_table_struct pt = { ._key = poll->events };
4785 req->result = vfs_poll(req->file, &pt) & poll->events;
4788 spin_lock_irq(&ctx->completion_lock);
4789 if (!req->result && !READ_ONCE(poll->canceled)) {
4790 add_wait_queue(poll->head, &poll->wait);
4797 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4799 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4800 if (req->opcode == IORING_OP_POLL_ADD)
4801 return req->async_data;
4802 return req->apoll->double_poll;
4805 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4807 if (req->opcode == IORING_OP_POLL_ADD)
4809 return &req->apoll->poll;
4812 static void io_poll_remove_double(struct io_kiocb *req)
4814 struct io_poll_iocb *poll = io_poll_get_double(req);
4816 lockdep_assert_held(&req->ctx->completion_lock);
4818 if (poll && poll->head) {
4819 struct wait_queue_head *head = poll->head;
4821 spin_lock(&head->lock);
4822 list_del_init(&poll->wait.entry);
4823 if (poll->wait.private)
4824 refcount_dec(&req->refs);
4826 spin_unlock(&head->lock);
4830 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4832 struct io_ring_ctx *ctx = req->ctx;
4834 io_poll_remove_double(req);
4835 req->poll.done = true;
4836 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4837 io_commit_cqring(ctx);
4840 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4842 struct io_ring_ctx *ctx = req->ctx;
4844 if (io_poll_rewait(req, &req->poll)) {
4845 spin_unlock_irq(&ctx->completion_lock);
4849 hash_del(&req->hash_node);
4850 io_poll_complete(req, req->result, 0);
4851 req->flags |= REQ_F_COMP_LOCKED;
4852 *nxt = io_put_req_find_next(req);
4853 spin_unlock_irq(&ctx->completion_lock);
4855 io_cqring_ev_posted(ctx);
4858 static void io_poll_task_func(struct callback_head *cb)
4860 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4861 struct io_ring_ctx *ctx = req->ctx;
4862 struct io_kiocb *nxt = NULL;
4864 io_poll_task_handler(req, &nxt);
4866 __io_req_task_submit(nxt);
4867 percpu_ref_put(&ctx->refs);
4870 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4871 int sync, void *key)
4873 struct io_kiocb *req = wait->private;
4874 struct io_poll_iocb *poll = io_poll_get_single(req);
4875 __poll_t mask = key_to_poll(key);
4877 /* for instances that support it check for an event match first: */
4878 if (mask && !(mask & poll->events))
4881 list_del_init(&wait->entry);
4883 if (poll && poll->head) {
4886 spin_lock(&poll->head->lock);
4887 done = list_empty(&poll->wait.entry);
4889 list_del_init(&poll->wait.entry);
4890 /* make sure double remove sees this as being gone */
4891 wait->private = NULL;
4892 spin_unlock(&poll->head->lock);
4894 __io_async_wake(req, poll, mask, io_poll_task_func);
4896 refcount_dec(&req->refs);
4900 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4901 wait_queue_func_t wake_func)
4905 poll->canceled = false;
4906 poll->events = events;
4907 INIT_LIST_HEAD(&poll->wait.entry);
4908 init_waitqueue_func_entry(&poll->wait, wake_func);
4911 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4912 struct wait_queue_head *head,
4913 struct io_poll_iocb **poll_ptr)
4915 struct io_kiocb *req = pt->req;
4918 * If poll->head is already set, it's because the file being polled
4919 * uses multiple waitqueues for poll handling (eg one for read, one
4920 * for write). Setup a separate io_poll_iocb if this happens.
4922 if (unlikely(poll->head)) {
4923 /* already have a 2nd entry, fail a third attempt */
4925 pt->error = -EINVAL;
4928 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4930 pt->error = -ENOMEM;
4933 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4934 refcount_inc(&req->refs);
4935 poll->wait.private = req;
4942 if (poll->events & EPOLLEXCLUSIVE)
4943 add_wait_queue_exclusive(head, &poll->wait);
4945 add_wait_queue(head, &poll->wait);
4948 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4949 struct poll_table_struct *p)
4951 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4952 struct async_poll *apoll = pt->req->apoll;
4954 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4957 static void io_async_task_func(struct callback_head *cb)
4959 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4960 struct async_poll *apoll = req->apoll;
4961 struct io_ring_ctx *ctx = req->ctx;
4963 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4965 if (io_poll_rewait(req, &apoll->poll)) {
4966 spin_unlock_irq(&ctx->completion_lock);
4967 percpu_ref_put(&ctx->refs);
4971 /* If req is still hashed, it cannot have been canceled. Don't check. */
4972 if (hash_hashed(&req->hash_node))
4973 hash_del(&req->hash_node);
4975 io_poll_remove_double(req);
4976 spin_unlock_irq(&ctx->completion_lock);
4978 if (!READ_ONCE(apoll->poll.canceled))
4979 __io_req_task_submit(req);
4981 __io_req_task_cancel(req, -ECANCELED);
4983 percpu_ref_put(&ctx->refs);
4984 kfree(apoll->double_poll);
4988 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4991 struct io_kiocb *req = wait->private;
4992 struct io_poll_iocb *poll = &req->apoll->poll;
4994 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4997 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5000 static void io_poll_req_insert(struct io_kiocb *req)
5002 struct io_ring_ctx *ctx = req->ctx;
5003 struct hlist_head *list;
5005 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5006 hlist_add_head(&req->hash_node, list);
5009 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5010 struct io_poll_iocb *poll,
5011 struct io_poll_table *ipt, __poll_t mask,
5012 wait_queue_func_t wake_func)
5013 __acquires(&ctx->completion_lock)
5015 struct io_ring_ctx *ctx = req->ctx;
5016 bool cancel = false;
5018 io_init_poll_iocb(poll, mask, wake_func);
5019 poll->file = req->file;
5020 poll->wait.private = req;
5022 ipt->pt._key = mask;
5024 ipt->error = -EINVAL;
5026 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5028 spin_lock_irq(&ctx->completion_lock);
5029 if (likely(poll->head)) {
5030 spin_lock(&poll->head->lock);
5031 if (unlikely(list_empty(&poll->wait.entry))) {
5037 if (mask || ipt->error)
5038 list_del_init(&poll->wait.entry);
5040 WRITE_ONCE(poll->canceled, true);
5041 else if (!poll->done) /* actually waiting for an event */
5042 io_poll_req_insert(req);
5043 spin_unlock(&poll->head->lock);
5049 static bool io_arm_poll_handler(struct io_kiocb *req)
5051 const struct io_op_def *def = &io_op_defs[req->opcode];
5052 struct io_ring_ctx *ctx = req->ctx;
5053 struct async_poll *apoll;
5054 struct io_poll_table ipt;
5058 if (!req->file || !file_can_poll(req->file))
5060 if (req->flags & REQ_F_POLLED)
5064 else if (def->pollout)
5068 /* if we can't nonblock try, then no point in arming a poll handler */
5069 if (!io_file_supports_async(req->file, rw))
5072 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5073 if (unlikely(!apoll))
5075 apoll->double_poll = NULL;
5077 req->flags |= REQ_F_POLLED;
5079 INIT_HLIST_NODE(&req->hash_node);
5083 mask |= POLLIN | POLLRDNORM;
5085 mask |= POLLOUT | POLLWRNORM;
5086 mask |= POLLERR | POLLPRI;
5088 ipt.pt._qproc = io_async_queue_proc;
5090 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5092 if (ret || ipt.error) {
5093 io_poll_remove_double(req);
5094 spin_unlock_irq(&ctx->completion_lock);
5095 kfree(apoll->double_poll);
5099 spin_unlock_irq(&ctx->completion_lock);
5100 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5101 apoll->poll.events);
5105 static bool __io_poll_remove_one(struct io_kiocb *req,
5106 struct io_poll_iocb *poll)
5108 bool do_complete = false;
5110 spin_lock(&poll->head->lock);
5111 WRITE_ONCE(poll->canceled, true);
5112 if (!list_empty(&poll->wait.entry)) {
5113 list_del_init(&poll->wait.entry);
5116 spin_unlock(&poll->head->lock);
5117 hash_del(&req->hash_node);
5121 static bool io_poll_remove_one(struct io_kiocb *req)
5125 io_poll_remove_double(req);
5127 if (req->opcode == IORING_OP_POLL_ADD) {
5128 do_complete = __io_poll_remove_one(req, &req->poll);
5130 struct async_poll *apoll = req->apoll;
5132 /* non-poll requests have submit ref still */
5133 do_complete = __io_poll_remove_one(req, &apoll->poll);
5136 kfree(apoll->double_poll);
5142 io_cqring_fill_event(req, -ECANCELED);
5143 io_commit_cqring(req->ctx);
5144 req->flags |= REQ_F_COMP_LOCKED;
5145 req_set_fail_links(req);
5153 * Returns true if we found and killed one or more poll requests
5155 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5157 struct hlist_node *tmp;
5158 struct io_kiocb *req;
5161 spin_lock_irq(&ctx->completion_lock);
5162 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5163 struct hlist_head *list;
5165 list = &ctx->cancel_hash[i];
5166 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5167 if (io_task_match(req, tsk))
5168 posted += io_poll_remove_one(req);
5171 spin_unlock_irq(&ctx->completion_lock);
5174 io_cqring_ev_posted(ctx);
5179 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5181 struct hlist_head *list;
5182 struct io_kiocb *req;
5184 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5185 hlist_for_each_entry(req, list, hash_node) {
5186 if (sqe_addr != req->user_data)
5188 if (io_poll_remove_one(req))
5196 static int io_poll_remove_prep(struct io_kiocb *req,
5197 const struct io_uring_sqe *sqe)
5199 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5201 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5205 req->poll.addr = READ_ONCE(sqe->addr);
5210 * Find a running poll command that matches one specified in sqe->addr,
5211 * and remove it if found.
5213 static int io_poll_remove(struct io_kiocb *req)
5215 struct io_ring_ctx *ctx = req->ctx;
5219 addr = req->poll.addr;
5220 spin_lock_irq(&ctx->completion_lock);
5221 ret = io_poll_cancel(ctx, addr);
5222 spin_unlock_irq(&ctx->completion_lock);
5225 req_set_fail_links(req);
5226 io_req_complete(req, ret);
5230 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5233 struct io_kiocb *req = wait->private;
5234 struct io_poll_iocb *poll = &req->poll;
5236 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5239 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5240 struct poll_table_struct *p)
5242 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5244 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5247 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5249 struct io_poll_iocb *poll = &req->poll;
5252 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5254 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5259 events = READ_ONCE(sqe->poll32_events);
5261 events = swahw32(events);
5263 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5264 (events & EPOLLEXCLUSIVE);
5268 static int io_poll_add(struct io_kiocb *req)
5270 struct io_poll_iocb *poll = &req->poll;
5271 struct io_ring_ctx *ctx = req->ctx;
5272 struct io_poll_table ipt;
5275 INIT_HLIST_NODE(&req->hash_node);
5276 ipt.pt._qproc = io_poll_queue_proc;
5278 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5281 if (mask) { /* no async, we'd stolen it */
5283 io_poll_complete(req, mask, 0);
5285 spin_unlock_irq(&ctx->completion_lock);
5288 io_cqring_ev_posted(ctx);
5294 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5296 struct io_timeout_data *data = container_of(timer,
5297 struct io_timeout_data, timer);
5298 struct io_kiocb *req = data->req;
5299 struct io_ring_ctx *ctx = req->ctx;
5300 unsigned long flags;
5302 spin_lock_irqsave(&ctx->completion_lock, flags);
5303 atomic_set(&req->ctx->cq_timeouts,
5304 atomic_read(&req->ctx->cq_timeouts) + 1);
5307 * We could be racing with timeout deletion. If the list is empty,
5308 * then timeout lookup already found it and will be handling it.
5310 if (!list_empty(&req->timeout.list))
5311 list_del_init(&req->timeout.list);
5313 io_cqring_fill_event(req, -ETIME);
5314 io_commit_cqring(ctx);
5315 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5317 io_cqring_ev_posted(ctx);
5318 req_set_fail_links(req);
5320 return HRTIMER_NORESTART;
5323 static int __io_timeout_cancel(struct io_kiocb *req)
5325 struct io_timeout_data *io = req->async_data;
5328 list_del_init(&req->timeout.list);
5330 ret = hrtimer_try_to_cancel(&io->timer);
5334 req_set_fail_links(req);
5335 req->flags |= REQ_F_COMP_LOCKED;
5336 io_cqring_fill_event(req, -ECANCELED);
5341 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5343 struct io_kiocb *req;
5346 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5347 if (user_data == req->user_data) {
5356 return __io_timeout_cancel(req);
5359 static int io_timeout_remove_prep(struct io_kiocb *req,
5360 const struct io_uring_sqe *sqe)
5362 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5364 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5366 if (sqe->ioprio || sqe->buf_index || sqe->len)
5369 req->timeout.addr = READ_ONCE(sqe->addr);
5370 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5371 if (req->timeout.flags)
5378 * Remove or update an existing timeout command
5380 static int io_timeout_remove(struct io_kiocb *req)
5382 struct io_ring_ctx *ctx = req->ctx;
5385 spin_lock_irq(&ctx->completion_lock);
5386 ret = io_timeout_cancel(ctx, req->timeout.addr);
5388 io_cqring_fill_event(req, ret);
5389 io_commit_cqring(ctx);
5390 spin_unlock_irq(&ctx->completion_lock);
5391 io_cqring_ev_posted(ctx);
5393 req_set_fail_links(req);
5398 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5399 bool is_timeout_link)
5401 struct io_timeout_data *data;
5403 u32 off = READ_ONCE(sqe->off);
5405 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5407 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5409 if (off && is_timeout_link)
5411 flags = READ_ONCE(sqe->timeout_flags);
5412 if (flags & ~IORING_TIMEOUT_ABS)
5415 req->timeout.off = off;
5417 if (!req->async_data && io_alloc_async_data(req))
5420 data = req->async_data;
5423 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5426 if (flags & IORING_TIMEOUT_ABS)
5427 data->mode = HRTIMER_MODE_ABS;
5429 data->mode = HRTIMER_MODE_REL;
5431 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5435 static int io_timeout(struct io_kiocb *req)
5437 struct io_ring_ctx *ctx = req->ctx;
5438 struct io_timeout_data *data = req->async_data;
5439 struct list_head *entry;
5440 u32 tail, off = req->timeout.off;
5442 spin_lock_irq(&ctx->completion_lock);
5445 * sqe->off holds how many events that need to occur for this
5446 * timeout event to be satisfied. If it isn't set, then this is
5447 * a pure timeout request, sequence isn't used.
5449 if (io_is_timeout_noseq(req)) {
5450 entry = ctx->timeout_list.prev;
5454 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5455 req->timeout.target_seq = tail + off;
5458 * Insertion sort, ensuring the first entry in the list is always
5459 * the one we need first.
5461 list_for_each_prev(entry, &ctx->timeout_list) {
5462 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5465 if (io_is_timeout_noseq(nxt))
5467 /* nxt.seq is behind @tail, otherwise would've been completed */
5468 if (off >= nxt->timeout.target_seq - tail)
5472 list_add(&req->timeout.list, entry);
5473 data->timer.function = io_timeout_fn;
5474 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5475 spin_unlock_irq(&ctx->completion_lock);
5479 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5481 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5483 return req->user_data == (unsigned long) data;
5486 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5488 enum io_wq_cancel cancel_ret;
5491 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5492 switch (cancel_ret) {
5493 case IO_WQ_CANCEL_OK:
5496 case IO_WQ_CANCEL_RUNNING:
5499 case IO_WQ_CANCEL_NOTFOUND:
5507 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5508 struct io_kiocb *req, __u64 sqe_addr,
5511 unsigned long flags;
5514 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5515 if (ret != -ENOENT) {
5516 spin_lock_irqsave(&ctx->completion_lock, flags);
5520 spin_lock_irqsave(&ctx->completion_lock, flags);
5521 ret = io_timeout_cancel(ctx, sqe_addr);
5524 ret = io_poll_cancel(ctx, sqe_addr);
5528 io_cqring_fill_event(req, ret);
5529 io_commit_cqring(ctx);
5530 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5531 io_cqring_ev_posted(ctx);
5534 req_set_fail_links(req);
5538 static int io_async_cancel_prep(struct io_kiocb *req,
5539 const struct io_uring_sqe *sqe)
5541 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5543 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5545 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5548 req->cancel.addr = READ_ONCE(sqe->addr);
5552 static int io_async_cancel(struct io_kiocb *req)
5554 struct io_ring_ctx *ctx = req->ctx;
5556 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5560 static int io_files_update_prep(struct io_kiocb *req,
5561 const struct io_uring_sqe *sqe)
5563 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5565 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5567 if (sqe->ioprio || sqe->rw_flags)
5570 req->files_update.offset = READ_ONCE(sqe->off);
5571 req->files_update.nr_args = READ_ONCE(sqe->len);
5572 if (!req->files_update.nr_args)
5574 req->files_update.arg = READ_ONCE(sqe->addr);
5578 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5579 struct io_comp_state *cs)
5581 struct io_ring_ctx *ctx = req->ctx;
5582 struct io_uring_files_update up;
5588 up.offset = req->files_update.offset;
5589 up.fds = req->files_update.arg;
5591 mutex_lock(&ctx->uring_lock);
5592 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5593 mutex_unlock(&ctx->uring_lock);
5596 req_set_fail_links(req);
5597 __io_req_complete(req, ret, 0, cs);
5601 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5603 switch (req->opcode) {
5606 case IORING_OP_READV:
5607 case IORING_OP_READ_FIXED:
5608 case IORING_OP_READ:
5609 return io_read_prep(req, sqe);
5610 case IORING_OP_WRITEV:
5611 case IORING_OP_WRITE_FIXED:
5612 case IORING_OP_WRITE:
5613 return io_write_prep(req, sqe);
5614 case IORING_OP_POLL_ADD:
5615 return io_poll_add_prep(req, sqe);
5616 case IORING_OP_POLL_REMOVE:
5617 return io_poll_remove_prep(req, sqe);
5618 case IORING_OP_FSYNC:
5619 return io_prep_fsync(req, sqe);
5620 case IORING_OP_SYNC_FILE_RANGE:
5621 return io_prep_sfr(req, sqe);
5622 case IORING_OP_SENDMSG:
5623 case IORING_OP_SEND:
5624 return io_sendmsg_prep(req, sqe);
5625 case IORING_OP_RECVMSG:
5626 case IORING_OP_RECV:
5627 return io_recvmsg_prep(req, sqe);
5628 case IORING_OP_CONNECT:
5629 return io_connect_prep(req, sqe);
5630 case IORING_OP_TIMEOUT:
5631 return io_timeout_prep(req, sqe, false);
5632 case IORING_OP_TIMEOUT_REMOVE:
5633 return io_timeout_remove_prep(req, sqe);
5634 case IORING_OP_ASYNC_CANCEL:
5635 return io_async_cancel_prep(req, sqe);
5636 case IORING_OP_LINK_TIMEOUT:
5637 return io_timeout_prep(req, sqe, true);
5638 case IORING_OP_ACCEPT:
5639 return io_accept_prep(req, sqe);
5640 case IORING_OP_FALLOCATE:
5641 return io_fallocate_prep(req, sqe);
5642 case IORING_OP_OPENAT:
5643 return io_openat_prep(req, sqe);
5644 case IORING_OP_CLOSE:
5645 return io_close_prep(req, sqe);
5646 case IORING_OP_FILES_UPDATE:
5647 return io_files_update_prep(req, sqe);
5648 case IORING_OP_STATX:
5649 return io_statx_prep(req, sqe);
5650 case IORING_OP_FADVISE:
5651 return io_fadvise_prep(req, sqe);
5652 case IORING_OP_MADVISE:
5653 return io_madvise_prep(req, sqe);
5654 case IORING_OP_OPENAT2:
5655 return io_openat2_prep(req, sqe);
5656 case IORING_OP_EPOLL_CTL:
5657 return io_epoll_ctl_prep(req, sqe);
5658 case IORING_OP_SPLICE:
5659 return io_splice_prep(req, sqe);
5660 case IORING_OP_PROVIDE_BUFFERS:
5661 return io_provide_buffers_prep(req, sqe);
5662 case IORING_OP_REMOVE_BUFFERS:
5663 return io_remove_buffers_prep(req, sqe);
5665 return io_tee_prep(req, sqe);
5668 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5673 static int io_req_defer_prep(struct io_kiocb *req,
5674 const struct io_uring_sqe *sqe)
5678 if (io_alloc_async_data(req))
5680 return io_req_prep(req, sqe);
5683 static u32 io_get_sequence(struct io_kiocb *req)
5685 struct io_kiocb *pos;
5686 struct io_ring_ctx *ctx = req->ctx;
5687 u32 total_submitted, nr_reqs = 1;
5689 if (req->flags & REQ_F_LINK_HEAD)
5690 list_for_each_entry(pos, &req->link_list, link_list)
5693 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5694 return total_submitted - nr_reqs;
5697 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5699 struct io_ring_ctx *ctx = req->ctx;
5700 struct io_defer_entry *de;
5704 /* Still need defer if there is pending req in defer list. */
5705 if (likely(list_empty_careful(&ctx->defer_list) &&
5706 !(req->flags & REQ_F_IO_DRAIN)))
5709 seq = io_get_sequence(req);
5710 /* Still a chance to pass the sequence check */
5711 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5714 if (!req->async_data) {
5715 ret = io_req_defer_prep(req, sqe);
5719 io_prep_async_link(req);
5720 de = kmalloc(sizeof(*de), GFP_KERNEL);
5724 spin_lock_irq(&ctx->completion_lock);
5725 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5726 spin_unlock_irq(&ctx->completion_lock);
5728 io_queue_async_work(req);
5729 return -EIOCBQUEUED;
5732 trace_io_uring_defer(ctx, req, req->user_data);
5735 list_add_tail(&de->list, &ctx->defer_list);
5736 spin_unlock_irq(&ctx->completion_lock);
5737 return -EIOCBQUEUED;
5740 static void io_req_drop_files(struct io_kiocb *req)
5742 struct io_ring_ctx *ctx = req->ctx;
5743 unsigned long flags;
5745 spin_lock_irqsave(&ctx->inflight_lock, flags);
5746 list_del(&req->inflight_entry);
5747 if (waitqueue_active(&ctx->inflight_wait))
5748 wake_up(&ctx->inflight_wait);
5749 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5750 req->flags &= ~REQ_F_INFLIGHT;
5751 put_files_struct(req->work.files);
5752 put_nsproxy(req->work.nsproxy);
5753 req->work.files = NULL;
5756 static void __io_clean_op(struct io_kiocb *req)
5758 if (req->flags & REQ_F_BUFFER_SELECTED) {
5759 switch (req->opcode) {
5760 case IORING_OP_READV:
5761 case IORING_OP_READ_FIXED:
5762 case IORING_OP_READ:
5763 kfree((void *)(unsigned long)req->rw.addr);
5765 case IORING_OP_RECVMSG:
5766 case IORING_OP_RECV:
5767 kfree(req->sr_msg.kbuf);
5770 req->flags &= ~REQ_F_BUFFER_SELECTED;
5773 if (req->flags & REQ_F_NEED_CLEANUP) {
5774 switch (req->opcode) {
5775 case IORING_OP_READV:
5776 case IORING_OP_READ_FIXED:
5777 case IORING_OP_READ:
5778 case IORING_OP_WRITEV:
5779 case IORING_OP_WRITE_FIXED:
5780 case IORING_OP_WRITE: {
5781 struct io_async_rw *io = req->async_data;
5783 kfree(io->free_iovec);
5786 case IORING_OP_RECVMSG:
5787 case IORING_OP_SENDMSG: {
5788 struct io_async_msghdr *io = req->async_data;
5789 if (io->iov != io->fast_iov)
5793 case IORING_OP_SPLICE:
5795 io_put_file(req, req->splice.file_in,
5796 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5798 case IORING_OP_OPENAT:
5799 case IORING_OP_OPENAT2:
5800 if (req->open.filename)
5801 putname(req->open.filename);
5804 req->flags &= ~REQ_F_NEED_CLEANUP;
5807 if (req->flags & REQ_F_INFLIGHT)
5808 io_req_drop_files(req);
5811 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5812 struct io_comp_state *cs)
5814 struct io_ring_ctx *ctx = req->ctx;
5817 switch (req->opcode) {
5819 ret = io_nop(req, cs);
5821 case IORING_OP_READV:
5822 case IORING_OP_READ_FIXED:
5823 case IORING_OP_READ:
5824 ret = io_read(req, force_nonblock, cs);
5826 case IORING_OP_WRITEV:
5827 case IORING_OP_WRITE_FIXED:
5828 case IORING_OP_WRITE:
5829 ret = io_write(req, force_nonblock, cs);
5831 case IORING_OP_FSYNC:
5832 ret = io_fsync(req, force_nonblock);
5834 case IORING_OP_POLL_ADD:
5835 ret = io_poll_add(req);
5837 case IORING_OP_POLL_REMOVE:
5838 ret = io_poll_remove(req);
5840 case IORING_OP_SYNC_FILE_RANGE:
5841 ret = io_sync_file_range(req, force_nonblock);
5843 case IORING_OP_SENDMSG:
5844 case IORING_OP_SEND:
5845 if (req->opcode == IORING_OP_SENDMSG)
5846 ret = io_sendmsg(req, force_nonblock, cs);
5848 ret = io_send(req, force_nonblock, cs);
5850 case IORING_OP_RECVMSG:
5851 case IORING_OP_RECV:
5852 if (req->opcode == IORING_OP_RECVMSG)
5853 ret = io_recvmsg(req, force_nonblock, cs);
5855 ret = io_recv(req, force_nonblock, cs);
5857 case IORING_OP_TIMEOUT:
5858 ret = io_timeout(req);
5860 case IORING_OP_TIMEOUT_REMOVE:
5861 ret = io_timeout_remove(req);
5863 case IORING_OP_ACCEPT:
5864 ret = io_accept(req, force_nonblock, cs);
5866 case IORING_OP_CONNECT:
5867 ret = io_connect(req, force_nonblock, cs);
5869 case IORING_OP_ASYNC_CANCEL:
5870 ret = io_async_cancel(req);
5872 case IORING_OP_FALLOCATE:
5873 ret = io_fallocate(req, force_nonblock);
5875 case IORING_OP_OPENAT:
5876 ret = io_openat(req, force_nonblock);
5878 case IORING_OP_CLOSE:
5879 ret = io_close(req, force_nonblock, cs);
5881 case IORING_OP_FILES_UPDATE:
5882 ret = io_files_update(req, force_nonblock, cs);
5884 case IORING_OP_STATX:
5885 ret = io_statx(req, force_nonblock);
5887 case IORING_OP_FADVISE:
5888 ret = io_fadvise(req, force_nonblock);
5890 case IORING_OP_MADVISE:
5891 ret = io_madvise(req, force_nonblock);
5893 case IORING_OP_OPENAT2:
5894 ret = io_openat2(req, force_nonblock);
5896 case IORING_OP_EPOLL_CTL:
5897 ret = io_epoll_ctl(req, force_nonblock, cs);
5899 case IORING_OP_SPLICE:
5900 ret = io_splice(req, force_nonblock);
5902 case IORING_OP_PROVIDE_BUFFERS:
5903 ret = io_provide_buffers(req, force_nonblock, cs);
5905 case IORING_OP_REMOVE_BUFFERS:
5906 ret = io_remove_buffers(req, force_nonblock, cs);
5909 ret = io_tee(req, force_nonblock);
5919 /* If the op doesn't have a file, we're not polling for it */
5920 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5921 const bool in_async = io_wq_current_is_worker();
5923 /* workqueue context doesn't hold uring_lock, grab it now */
5925 mutex_lock(&ctx->uring_lock);
5927 io_iopoll_req_issued(req);
5930 mutex_unlock(&ctx->uring_lock);
5936 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5938 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5939 struct io_kiocb *timeout;
5942 timeout = io_prep_linked_timeout(req);
5944 io_queue_linked_timeout(timeout);
5946 /* if NO_CANCEL is set, we must still run the work */
5947 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5948 IO_WQ_WORK_CANCEL) {
5954 ret = io_issue_sqe(req, false, NULL);
5956 * We can get EAGAIN for polled IO even though we're
5957 * forcing a sync submission from here, since we can't
5958 * wait for request slots on the block side.
5967 req_set_fail_links(req);
5968 io_req_complete(req, ret);
5971 return io_steal_work(req);
5974 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5977 struct fixed_file_table *table;
5979 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5980 return table->files[index & IORING_FILE_TABLE_MASK];
5983 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5984 int fd, struct file **out_file, bool fixed)
5986 struct io_ring_ctx *ctx = req->ctx;
5990 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
5992 fd = array_index_nospec(fd, ctx->nr_user_files);
5993 file = io_file_from_index(ctx, fd);
5995 req->fixed_file_refs = ctx->file_data->cur_refs;
5996 percpu_ref_get(req->fixed_file_refs);
5999 trace_io_uring_file_get(ctx, fd);
6000 file = __io_file_get(state, fd);
6003 if (file || io_op_defs[req->opcode].needs_file_no_error) {
6010 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6015 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6016 if (unlikely(!fixed && io_async_submit(req->ctx)))
6019 return io_file_get(state, req, fd, &req->file, fixed);
6022 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6024 struct io_timeout_data *data = container_of(timer,
6025 struct io_timeout_data, timer);
6026 struct io_kiocb *req = data->req;
6027 struct io_ring_ctx *ctx = req->ctx;
6028 struct io_kiocb *prev = NULL;
6029 unsigned long flags;
6031 spin_lock_irqsave(&ctx->completion_lock, flags);
6034 * We don't expect the list to be empty, that will only happen if we
6035 * race with the completion of the linked work.
6037 if (!list_empty(&req->link_list)) {
6038 prev = list_entry(req->link_list.prev, struct io_kiocb,
6040 if (refcount_inc_not_zero(&prev->refs)) {
6041 list_del_init(&req->link_list);
6042 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6047 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6050 req_set_fail_links(prev);
6051 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6054 io_req_complete(req, -ETIME);
6056 return HRTIMER_NORESTART;
6059 static void __io_queue_linked_timeout(struct io_kiocb *req)
6062 * If the list is now empty, then our linked request finished before
6063 * we got a chance to setup the timer
6065 if (!list_empty(&req->link_list)) {
6066 struct io_timeout_data *data = req->async_data;
6068 data->timer.function = io_link_timeout_fn;
6069 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6074 static void io_queue_linked_timeout(struct io_kiocb *req)
6076 struct io_ring_ctx *ctx = req->ctx;
6078 spin_lock_irq(&ctx->completion_lock);
6079 __io_queue_linked_timeout(req);
6080 spin_unlock_irq(&ctx->completion_lock);
6082 /* drop submission reference */
6086 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6088 struct io_kiocb *nxt;
6090 if (!(req->flags & REQ_F_LINK_HEAD))
6092 if (req->flags & REQ_F_LINK_TIMEOUT)
6095 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6097 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6100 req->flags |= REQ_F_LINK_TIMEOUT;
6104 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6106 struct io_kiocb *linked_timeout;
6107 struct io_kiocb *nxt;
6108 const struct cred *old_creds = NULL;
6112 linked_timeout = io_prep_linked_timeout(req);
6114 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6115 req->work.creds != current_cred()) {
6117 revert_creds(old_creds);
6118 if (old_creds == req->work.creds)
6119 old_creds = NULL; /* restored original creds */
6121 old_creds = override_creds(req->work.creds);
6124 ret = io_issue_sqe(req, true, cs);
6127 * We async punt it if the file wasn't marked NOWAIT, or if the file
6128 * doesn't support non-blocking read/write attempts
6130 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6131 if (!io_arm_poll_handler(req)) {
6134 * Queued up for async execution, worker will release
6135 * submit reference when the iocb is actually submitted.
6137 io_queue_async_work(req);
6141 io_queue_linked_timeout(linked_timeout);
6145 if (unlikely(ret)) {
6146 /* un-prep timeout, so it'll be killed as any other linked */
6147 req->flags &= ~REQ_F_LINK_TIMEOUT;
6148 req_set_fail_links(req);
6150 io_req_complete(req, ret);
6154 /* drop submission reference */
6155 nxt = io_put_req_find_next(req);
6157 io_queue_linked_timeout(linked_timeout);
6162 if (req->flags & REQ_F_FORCE_ASYNC)
6168 revert_creds(old_creds);
6171 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6172 struct io_comp_state *cs)
6176 ret = io_req_defer(req, sqe);
6178 if (ret != -EIOCBQUEUED) {
6180 req_set_fail_links(req);
6182 io_req_complete(req, ret);
6184 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6185 if (!req->async_data) {
6186 ret = io_req_defer_prep(req, sqe);
6192 * Never try inline submit of IOSQE_ASYNC is set, go straight
6193 * to async execution.
6195 io_req_init_async(req);
6196 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6197 io_queue_async_work(req);
6200 ret = io_req_prep(req, sqe);
6204 __io_queue_sqe(req, cs);
6208 static inline void io_queue_link_head(struct io_kiocb *req,
6209 struct io_comp_state *cs)
6211 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6213 io_req_complete(req, -ECANCELED);
6215 io_queue_sqe(req, NULL, cs);
6218 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6219 struct io_kiocb **link, struct io_comp_state *cs)
6221 struct io_ring_ctx *ctx = req->ctx;
6225 * If we already have a head request, queue this one for async
6226 * submittal once the head completes. If we don't have a head but
6227 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6228 * submitted sync once the chain is complete. If none of those
6229 * conditions are true (normal request), then just queue it.
6232 struct io_kiocb *head = *link;
6235 * Taking sequential execution of a link, draining both sides
6236 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6237 * requests in the link. So, it drains the head and the
6238 * next after the link request. The last one is done via
6239 * drain_next flag to persist the effect across calls.
6241 if (req->flags & REQ_F_IO_DRAIN) {
6242 head->flags |= REQ_F_IO_DRAIN;
6243 ctx->drain_next = 1;
6245 ret = io_req_defer_prep(req, sqe);
6246 if (unlikely(ret)) {
6247 /* fail even hard links since we don't submit */
6248 head->flags |= REQ_F_FAIL_LINK;
6251 trace_io_uring_link(ctx, req, head);
6252 list_add_tail(&req->link_list, &head->link_list);
6254 /* last request of a link, enqueue the link */
6255 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6256 io_queue_link_head(head, cs);
6260 if (unlikely(ctx->drain_next)) {
6261 req->flags |= REQ_F_IO_DRAIN;
6262 ctx->drain_next = 0;
6264 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6265 req->flags |= REQ_F_LINK_HEAD;
6266 INIT_LIST_HEAD(&req->link_list);
6268 ret = io_req_defer_prep(req, sqe);
6270 req->flags |= REQ_F_FAIL_LINK;
6273 io_queue_sqe(req, sqe, cs);
6281 * Batched submission is done, ensure local IO is flushed out.
6283 static void io_submit_state_end(struct io_submit_state *state)
6285 if (!list_empty(&state->comp.list))
6286 io_submit_flush_completions(&state->comp);
6287 blk_finish_plug(&state->plug);
6288 io_state_file_put(state);
6289 if (state->free_reqs)
6290 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6294 * Start submission side cache.
6296 static void io_submit_state_start(struct io_submit_state *state,
6297 struct io_ring_ctx *ctx, unsigned int max_ios)
6299 blk_start_plug(&state->plug);
6301 INIT_LIST_HEAD(&state->comp.list);
6302 state->comp.ctx = ctx;
6303 state->free_reqs = 0;
6305 state->ios_left = max_ios;
6308 static void io_commit_sqring(struct io_ring_ctx *ctx)
6310 struct io_rings *rings = ctx->rings;
6313 * Ensure any loads from the SQEs are done at this point,
6314 * since once we write the new head, the application could
6315 * write new data to them.
6317 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6321 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6322 * that is mapped by userspace. This means that care needs to be taken to
6323 * ensure that reads are stable, as we cannot rely on userspace always
6324 * being a good citizen. If members of the sqe are validated and then later
6325 * used, it's important that those reads are done through READ_ONCE() to
6326 * prevent a re-load down the line.
6328 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6330 u32 *sq_array = ctx->sq_array;
6334 * The cached sq head (or cq tail) serves two purposes:
6336 * 1) allows us to batch the cost of updating the user visible
6338 * 2) allows the kernel side to track the head on its own, even
6339 * though the application is the one updating it.
6341 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6342 if (likely(head < ctx->sq_entries))
6343 return &ctx->sq_sqes[head];
6345 /* drop invalid entries */
6346 ctx->cached_sq_dropped++;
6347 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6351 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6353 ctx->cached_sq_head++;
6357 * Check SQE restrictions (opcode and flags).
6359 * Returns 'true' if SQE is allowed, 'false' otherwise.
6361 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6362 struct io_kiocb *req,
6363 unsigned int sqe_flags)
6365 if (!ctx->restricted)
6368 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6371 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6372 ctx->restrictions.sqe_flags_required)
6375 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6376 ctx->restrictions.sqe_flags_required))
6382 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6383 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6384 IOSQE_BUFFER_SELECT)
6386 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6387 const struct io_uring_sqe *sqe,
6388 struct io_submit_state *state)
6390 unsigned int sqe_flags;
6393 req->opcode = READ_ONCE(sqe->opcode);
6394 req->user_data = READ_ONCE(sqe->user_data);
6395 req->async_data = NULL;
6399 /* one is dropped after submission, the other at completion */
6400 refcount_set(&req->refs, 2);
6401 req->task = current;
6404 if (unlikely(req->opcode >= IORING_OP_LAST))
6407 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6410 sqe_flags = READ_ONCE(sqe->flags);
6411 /* enforce forwards compatibility on users */
6412 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6415 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6418 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6419 !io_op_defs[req->opcode].buffer_select)
6422 id = READ_ONCE(sqe->personality);
6424 io_req_init_async(req);
6425 req->work.creds = idr_find(&ctx->personality_idr, id);
6426 if (unlikely(!req->work.creds))
6428 get_cred(req->work.creds);
6431 /* same numerical values with corresponding REQ_F_*, safe to copy */
6432 req->flags |= sqe_flags;
6434 if (!io_op_defs[req->opcode].needs_file)
6437 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6440 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6442 struct io_submit_state state;
6443 struct io_kiocb *link = NULL;
6444 int i, submitted = 0;
6446 /* if we have a backlog and couldn't flush it all, return BUSY */
6447 if (test_bit(0, &ctx->sq_check_overflow)) {
6448 if (!list_empty(&ctx->cq_overflow_list) &&
6449 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6453 /* make sure SQ entry isn't read before tail */
6454 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6456 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6459 atomic_long_add(nr, ¤t->io_uring->req_issue);
6460 refcount_add(nr, ¤t->usage);
6462 io_submit_state_start(&state, ctx, nr);
6464 for (i = 0; i < nr; i++) {
6465 const struct io_uring_sqe *sqe;
6466 struct io_kiocb *req;
6469 sqe = io_get_sqe(ctx);
6470 if (unlikely(!sqe)) {
6471 io_consume_sqe(ctx);
6474 req = io_alloc_req(ctx, &state);
6475 if (unlikely(!req)) {
6477 submitted = -EAGAIN;
6481 err = io_init_req(ctx, req, sqe, &state);
6482 io_consume_sqe(ctx);
6483 /* will complete beyond this point, count as submitted */
6486 if (unlikely(err)) {
6489 io_req_complete(req, err);
6493 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6494 true, io_async_submit(ctx));
6495 err = io_submit_sqe(req, sqe, &link, &state.comp);
6500 if (unlikely(submitted != nr)) {
6501 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6503 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6504 atomic_long_sub(nr - ref_used, ¤t->io_uring->req_issue);
6505 put_task_struct_many(current, nr - ref_used);
6508 io_queue_link_head(link, &state.comp);
6509 io_submit_state_end(&state);
6511 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6512 io_commit_sqring(ctx);
6517 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6519 /* Tell userspace we may need a wakeup call */
6520 spin_lock_irq(&ctx->completion_lock);
6521 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6522 spin_unlock_irq(&ctx->completion_lock);
6525 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6527 spin_lock_irq(&ctx->completion_lock);
6528 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6529 spin_unlock_irq(&ctx->completion_lock);
6532 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6533 int sync, void *key)
6535 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6538 ret = autoremove_wake_function(wqe, mode, sync, key);
6540 unsigned long flags;
6542 spin_lock_irqsave(&ctx->completion_lock, flags);
6543 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6544 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6555 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6556 unsigned long start_jiffies, bool cap_entries)
6558 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6559 struct io_sq_data *sqd = ctx->sq_data;
6560 unsigned int to_submit;
6564 if (!list_empty(&ctx->iopoll_list)) {
6565 unsigned nr_events = 0;
6567 mutex_lock(&ctx->uring_lock);
6568 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6569 io_do_iopoll(ctx, &nr_events, 0);
6570 mutex_unlock(&ctx->uring_lock);
6573 to_submit = io_sqring_entries(ctx);
6576 * If submit got -EBUSY, flag us as needing the application
6577 * to enter the kernel to reap and flush events.
6579 if (!to_submit || ret == -EBUSY || need_resched()) {
6581 * Drop cur_mm before scheduling, we can't hold it for
6582 * long periods (or over schedule()). Do this before
6583 * adding ourselves to the waitqueue, as the unuse/drop
6586 io_sq_thread_drop_mm();
6589 * We're polling. If we're within the defined idle
6590 * period, then let us spin without work before going
6591 * to sleep. The exception is if we got EBUSY doing
6592 * more IO, we should wait for the application to
6593 * reap events and wake us up.
6595 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6596 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6597 !percpu_ref_is_dying(&ctx->refs)))
6600 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6601 TASK_INTERRUPTIBLE);
6604 * While doing polled IO, before going to sleep, we need
6605 * to check if there are new reqs added to iopoll_list,
6606 * it is because reqs may have been punted to io worker
6607 * and will be added to iopoll_list later, hence check
6608 * the iopoll_list again.
6610 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6611 !list_empty_careful(&ctx->iopoll_list)) {
6612 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6616 to_submit = io_sqring_entries(ctx);
6617 if (!to_submit || ret == -EBUSY)
6621 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6622 io_ring_clear_wakeup_flag(ctx);
6624 /* if we're handling multiple rings, cap submit size for fairness */
6625 if (cap_entries && to_submit > 8)
6628 mutex_lock(&ctx->uring_lock);
6629 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6630 ret = io_submit_sqes(ctx, to_submit);
6631 mutex_unlock(&ctx->uring_lock);
6633 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6634 wake_up(&ctx->sqo_sq_wait);
6636 return SQT_DID_WORK;
6639 static void io_sqd_init_new(struct io_sq_data *sqd)
6641 struct io_ring_ctx *ctx;
6643 while (!list_empty(&sqd->ctx_new_list)) {
6644 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6645 init_wait(&ctx->sqo_wait_entry);
6646 ctx->sqo_wait_entry.func = io_sq_wake_function;
6647 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6648 complete(&ctx->sq_thread_comp);
6652 static int io_sq_thread(void *data)
6654 struct cgroup_subsys_state *cur_css = NULL;
6655 const struct cred *old_cred = NULL;
6656 struct io_sq_data *sqd = data;
6657 struct io_ring_ctx *ctx;
6658 unsigned long start_jiffies;
6660 start_jiffies = jiffies;
6661 while (!kthread_should_stop()) {
6662 enum sq_ret ret = 0;
6666 * Any changes to the sqd lists are synchronized through the
6667 * kthread parking. This synchronizes the thread vs users,
6668 * the users are synchronized on the sqd->ctx_lock.
6670 if (kthread_should_park())
6673 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6674 io_sqd_init_new(sqd);
6676 cap_entries = !list_is_singular(&sqd->ctx_list);
6678 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6679 if (current->cred != ctx->creds) {
6681 revert_creds(old_cred);
6682 old_cred = override_creds(ctx->creds);
6684 io_sq_thread_associate_blkcg(ctx, &cur_css);
6686 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6688 io_sq_thread_drop_mm();
6691 if (ret & SQT_SPIN) {
6694 } else if (ret == SQT_IDLE) {
6695 if (kthread_should_park())
6697 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6698 io_ring_set_wakeup_flag(ctx);
6700 start_jiffies = jiffies;
6701 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6702 io_ring_clear_wakeup_flag(ctx);
6709 io_sq_thread_unassociate_blkcg();
6711 revert_creds(old_cred);
6718 struct io_wait_queue {
6719 struct wait_queue_entry wq;
6720 struct io_ring_ctx *ctx;
6722 unsigned nr_timeouts;
6725 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6727 struct io_ring_ctx *ctx = iowq->ctx;
6730 * Wake up if we have enough events, or if a timeout occurred since we
6731 * started waiting. For timeouts, we always want to return to userspace,
6732 * regardless of event count.
6734 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6735 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6738 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6739 int wake_flags, void *key)
6741 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6744 /* use noflush == true, as we can't safely rely on locking context */
6745 if (!io_should_wake(iowq, true))
6748 return autoremove_wake_function(curr, mode, wake_flags, key);
6751 static int io_run_task_work_sig(void)
6753 if (io_run_task_work())
6755 if (!signal_pending(current))
6757 if (current->jobctl & JOBCTL_TASK_WORK) {
6758 spin_lock_irq(¤t->sighand->siglock);
6759 current->jobctl &= ~JOBCTL_TASK_WORK;
6760 recalc_sigpending();
6761 spin_unlock_irq(¤t->sighand->siglock);
6768 * Wait until events become available, if we don't already have some. The
6769 * application must reap them itself, as they reside on the shared cq ring.
6771 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6772 const sigset_t __user *sig, size_t sigsz)
6774 struct io_wait_queue iowq = {
6777 .func = io_wake_function,
6778 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6781 .to_wait = min_events,
6783 struct io_rings *rings = ctx->rings;
6787 if (io_cqring_events(ctx, false) >= min_events)
6789 if (!io_run_task_work())
6794 #ifdef CONFIG_COMPAT
6795 if (in_compat_syscall())
6796 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6800 ret = set_user_sigmask(sig, sigsz);
6806 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6807 trace_io_uring_cqring_wait(ctx, min_events);
6809 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6810 TASK_INTERRUPTIBLE);
6811 /* make sure we run task_work before checking for signals */
6812 ret = io_run_task_work_sig();
6817 if (io_should_wake(&iowq, false))
6821 finish_wait(&ctx->wait, &iowq.wq);
6823 restore_saved_sigmask_unless(ret == -EINTR);
6825 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6828 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6830 #if defined(CONFIG_UNIX)
6831 if (ctx->ring_sock) {
6832 struct sock *sock = ctx->ring_sock->sk;
6833 struct sk_buff *skb;
6835 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6841 for (i = 0; i < ctx->nr_user_files; i++) {
6844 file = io_file_from_index(ctx, i);
6851 static void io_file_ref_kill(struct percpu_ref *ref)
6853 struct fixed_file_data *data;
6855 data = container_of(ref, struct fixed_file_data, refs);
6856 complete(&data->done);
6859 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6861 struct fixed_file_data *data = ctx->file_data;
6862 struct fixed_file_ref_node *ref_node = NULL;
6863 unsigned nr_tables, i;
6868 spin_lock(&data->lock);
6869 if (!list_empty(&data->ref_list))
6870 ref_node = list_first_entry(&data->ref_list,
6871 struct fixed_file_ref_node, node);
6872 spin_unlock(&data->lock);
6874 percpu_ref_kill(&ref_node->refs);
6876 percpu_ref_kill(&data->refs);
6878 /* wait for all refs nodes to complete */
6879 flush_delayed_work(&ctx->file_put_work);
6880 wait_for_completion(&data->done);
6882 __io_sqe_files_unregister(ctx);
6883 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6884 for (i = 0; i < nr_tables; i++)
6885 kfree(data->table[i].files);
6887 percpu_ref_exit(&data->refs);
6889 ctx->file_data = NULL;
6890 ctx->nr_user_files = 0;
6894 static void io_put_sq_data(struct io_sq_data *sqd)
6896 if (refcount_dec_and_test(&sqd->refs)) {
6898 * The park is a bit of a work-around, without it we get
6899 * warning spews on shutdown with SQPOLL set and affinity
6900 * set to a single CPU.
6903 kthread_park(sqd->thread);
6904 kthread_stop(sqd->thread);
6911 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
6913 struct io_ring_ctx *ctx_attach;
6914 struct io_sq_data *sqd;
6917 f = fdget(p->wq_fd);
6919 return ERR_PTR(-ENXIO);
6920 if (f.file->f_op != &io_uring_fops) {
6922 return ERR_PTR(-EINVAL);
6925 ctx_attach = f.file->private_data;
6926 sqd = ctx_attach->sq_data;
6929 return ERR_PTR(-EINVAL);
6932 refcount_inc(&sqd->refs);
6937 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
6939 struct io_sq_data *sqd;
6941 if (p->flags & IORING_SETUP_ATTACH_WQ)
6942 return io_attach_sq_data(p);
6944 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
6946 return ERR_PTR(-ENOMEM);
6948 refcount_set(&sqd->refs, 1);
6949 INIT_LIST_HEAD(&sqd->ctx_list);
6950 INIT_LIST_HEAD(&sqd->ctx_new_list);
6951 mutex_init(&sqd->ctx_lock);
6952 mutex_init(&sqd->lock);
6953 init_waitqueue_head(&sqd->wait);
6957 static void io_sq_thread_unpark(struct io_sq_data *sqd)
6958 __releases(&sqd->lock)
6962 kthread_unpark(sqd->thread);
6963 mutex_unlock(&sqd->lock);
6966 static void io_sq_thread_park(struct io_sq_data *sqd)
6967 __acquires(&sqd->lock)
6971 mutex_lock(&sqd->lock);
6972 kthread_park(sqd->thread);
6975 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6977 struct io_sq_data *sqd = ctx->sq_data;
6982 * We may arrive here from the error branch in
6983 * io_sq_offload_create() where the kthread is created
6984 * without being waked up, thus wake it up now to make
6985 * sure the wait will complete.
6987 wake_up_process(sqd->thread);
6988 wait_for_completion(&ctx->sq_thread_comp);
6990 io_sq_thread_park(sqd);
6993 mutex_lock(&sqd->ctx_lock);
6994 list_del(&ctx->sqd_list);
6995 mutex_unlock(&sqd->ctx_lock);
6998 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6999 io_sq_thread_unpark(sqd);
7002 io_put_sq_data(sqd);
7003 ctx->sq_data = NULL;
7007 static void io_finish_async(struct io_ring_ctx *ctx)
7009 io_sq_thread_stop(ctx);
7012 io_wq_destroy(ctx->io_wq);
7017 #if defined(CONFIG_UNIX)
7019 * Ensure the UNIX gc is aware of our file set, so we are certain that
7020 * the io_uring can be safely unregistered on process exit, even if we have
7021 * loops in the file referencing.
7023 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7025 struct sock *sk = ctx->ring_sock->sk;
7026 struct scm_fp_list *fpl;
7027 struct sk_buff *skb;
7030 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7034 skb = alloc_skb(0, GFP_KERNEL);
7043 fpl->user = get_uid(ctx->user);
7044 for (i = 0; i < nr; i++) {
7045 struct file *file = io_file_from_index(ctx, i + offset);
7049 fpl->fp[nr_files] = get_file(file);
7050 unix_inflight(fpl->user, fpl->fp[nr_files]);
7055 fpl->max = SCM_MAX_FD;
7056 fpl->count = nr_files;
7057 UNIXCB(skb).fp = fpl;
7058 skb->destructor = unix_destruct_scm;
7059 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7060 skb_queue_head(&sk->sk_receive_queue, skb);
7062 for (i = 0; i < nr_files; i++)
7073 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7074 * causes regular reference counting to break down. We rely on the UNIX
7075 * garbage collection to take care of this problem for us.
7077 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7079 unsigned left, total;
7083 left = ctx->nr_user_files;
7085 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7087 ret = __io_sqe_files_scm(ctx, this_files, total);
7091 total += this_files;
7097 while (total < ctx->nr_user_files) {
7098 struct file *file = io_file_from_index(ctx, total);
7108 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7114 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
7119 for (i = 0; i < nr_tables; i++) {
7120 struct fixed_file_table *table = &ctx->file_data->table[i];
7121 unsigned this_files;
7123 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7124 table->files = kcalloc(this_files, sizeof(struct file *),
7128 nr_files -= this_files;
7134 for (i = 0; i < nr_tables; i++) {
7135 struct fixed_file_table *table = &ctx->file_data->table[i];
7136 kfree(table->files);
7141 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7143 #if defined(CONFIG_UNIX)
7144 struct sock *sock = ctx->ring_sock->sk;
7145 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7146 struct sk_buff *skb;
7149 __skb_queue_head_init(&list);
7152 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7153 * remove this entry and rearrange the file array.
7155 skb = skb_dequeue(head);
7157 struct scm_fp_list *fp;
7159 fp = UNIXCB(skb).fp;
7160 for (i = 0; i < fp->count; i++) {
7163 if (fp->fp[i] != file)
7166 unix_notinflight(fp->user, fp->fp[i]);
7167 left = fp->count - 1 - i;
7169 memmove(&fp->fp[i], &fp->fp[i + 1],
7170 left * sizeof(struct file *));
7177 __skb_queue_tail(&list, skb);
7187 __skb_queue_tail(&list, skb);
7189 skb = skb_dequeue(head);
7192 if (skb_peek(&list)) {
7193 spin_lock_irq(&head->lock);
7194 while ((skb = __skb_dequeue(&list)) != NULL)
7195 __skb_queue_tail(head, skb);
7196 spin_unlock_irq(&head->lock);
7203 struct io_file_put {
7204 struct list_head list;
7208 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7210 struct fixed_file_data *file_data = ref_node->file_data;
7211 struct io_ring_ctx *ctx = file_data->ctx;
7212 struct io_file_put *pfile, *tmp;
7214 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7215 list_del(&pfile->list);
7216 io_ring_file_put(ctx, pfile->file);
7220 spin_lock(&file_data->lock);
7221 list_del(&ref_node->node);
7222 spin_unlock(&file_data->lock);
7224 percpu_ref_exit(&ref_node->refs);
7226 percpu_ref_put(&file_data->refs);
7229 static void io_file_put_work(struct work_struct *work)
7231 struct io_ring_ctx *ctx;
7232 struct llist_node *node;
7234 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7235 node = llist_del_all(&ctx->file_put_llist);
7238 struct fixed_file_ref_node *ref_node;
7239 struct llist_node *next = node->next;
7241 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7242 __io_file_put_work(ref_node);
7247 static void io_file_data_ref_zero(struct percpu_ref *ref)
7249 struct fixed_file_ref_node *ref_node;
7250 struct io_ring_ctx *ctx;
7254 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7255 ctx = ref_node->file_data->ctx;
7257 if (percpu_ref_is_dying(&ctx->file_data->refs))
7260 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7262 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7264 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7267 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7268 struct io_ring_ctx *ctx)
7270 struct fixed_file_ref_node *ref_node;
7272 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7274 return ERR_PTR(-ENOMEM);
7276 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7279 return ERR_PTR(-ENOMEM);
7281 INIT_LIST_HEAD(&ref_node->node);
7282 INIT_LIST_HEAD(&ref_node->file_list);
7283 ref_node->file_data = ctx->file_data;
7287 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7289 percpu_ref_exit(&ref_node->refs);
7293 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7296 __s32 __user *fds = (__s32 __user *) arg;
7301 struct fixed_file_ref_node *ref_node;
7307 if (nr_args > IORING_MAX_FIXED_FILES)
7310 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7311 if (!ctx->file_data)
7313 ctx->file_data->ctx = ctx;
7314 init_completion(&ctx->file_data->done);
7315 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7316 spin_lock_init(&ctx->file_data->lock);
7318 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7319 ctx->file_data->table = kcalloc(nr_tables,
7320 sizeof(struct fixed_file_table),
7322 if (!ctx->file_data->table) {
7323 kfree(ctx->file_data);
7324 ctx->file_data = NULL;
7328 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7329 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7330 kfree(ctx->file_data->table);
7331 kfree(ctx->file_data);
7332 ctx->file_data = NULL;
7336 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7337 percpu_ref_exit(&ctx->file_data->refs);
7338 kfree(ctx->file_data->table);
7339 kfree(ctx->file_data);
7340 ctx->file_data = NULL;
7344 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7345 struct fixed_file_table *table;
7349 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7351 /* allow sparse sets */
7357 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7358 index = i & IORING_FILE_TABLE_MASK;
7366 * Don't allow io_uring instances to be registered. If UNIX
7367 * isn't enabled, then this causes a reference cycle and this
7368 * instance can never get freed. If UNIX is enabled we'll
7369 * handle it just fine, but there's still no point in allowing
7370 * a ring fd as it doesn't support regular read/write anyway.
7372 if (file->f_op == &io_uring_fops) {
7377 table->files[index] = file;
7381 for (i = 0; i < ctx->nr_user_files; i++) {
7382 file = io_file_from_index(ctx, i);
7386 for (i = 0; i < nr_tables; i++)
7387 kfree(ctx->file_data->table[i].files);
7389 percpu_ref_exit(&ctx->file_data->refs);
7390 kfree(ctx->file_data->table);
7391 kfree(ctx->file_data);
7392 ctx->file_data = NULL;
7393 ctx->nr_user_files = 0;
7397 ret = io_sqe_files_scm(ctx);
7399 io_sqe_files_unregister(ctx);
7403 ref_node = alloc_fixed_file_ref_node(ctx);
7404 if (IS_ERR(ref_node)) {
7405 io_sqe_files_unregister(ctx);
7406 return PTR_ERR(ref_node);
7409 ctx->file_data->cur_refs = &ref_node->refs;
7410 spin_lock(&ctx->file_data->lock);
7411 list_add(&ref_node->node, &ctx->file_data->ref_list);
7412 spin_unlock(&ctx->file_data->lock);
7413 percpu_ref_get(&ctx->file_data->refs);
7417 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7420 #if defined(CONFIG_UNIX)
7421 struct sock *sock = ctx->ring_sock->sk;
7422 struct sk_buff_head *head = &sock->sk_receive_queue;
7423 struct sk_buff *skb;
7426 * See if we can merge this file into an existing skb SCM_RIGHTS
7427 * file set. If there's no room, fall back to allocating a new skb
7428 * and filling it in.
7430 spin_lock_irq(&head->lock);
7431 skb = skb_peek(head);
7433 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7435 if (fpl->count < SCM_MAX_FD) {
7436 __skb_unlink(skb, head);
7437 spin_unlock_irq(&head->lock);
7438 fpl->fp[fpl->count] = get_file(file);
7439 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7441 spin_lock_irq(&head->lock);
7442 __skb_queue_head(head, skb);
7447 spin_unlock_irq(&head->lock);
7454 return __io_sqe_files_scm(ctx, 1, index);
7460 static int io_queue_file_removal(struct fixed_file_data *data,
7463 struct io_file_put *pfile;
7464 struct percpu_ref *refs = data->cur_refs;
7465 struct fixed_file_ref_node *ref_node;
7467 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7471 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7473 list_add(&pfile->list, &ref_node->file_list);
7478 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7479 struct io_uring_files_update *up,
7482 struct fixed_file_data *data = ctx->file_data;
7483 struct fixed_file_ref_node *ref_node;
7488 bool needs_switch = false;
7490 if (check_add_overflow(up->offset, nr_args, &done))
7492 if (done > ctx->nr_user_files)
7495 ref_node = alloc_fixed_file_ref_node(ctx);
7496 if (IS_ERR(ref_node))
7497 return PTR_ERR(ref_node);
7500 fds = u64_to_user_ptr(up->fds);
7502 struct fixed_file_table *table;
7506 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7510 i = array_index_nospec(up->offset, ctx->nr_user_files);
7511 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7512 index = i & IORING_FILE_TABLE_MASK;
7513 if (table->files[index]) {
7514 file = table->files[index];
7515 err = io_queue_file_removal(data, file);
7518 table->files[index] = NULL;
7519 needs_switch = true;
7528 * Don't allow io_uring instances to be registered. If
7529 * UNIX isn't enabled, then this causes a reference
7530 * cycle and this instance can never get freed. If UNIX
7531 * is enabled we'll handle it just fine, but there's
7532 * still no point in allowing a ring fd as it doesn't
7533 * support regular read/write anyway.
7535 if (file->f_op == &io_uring_fops) {
7540 table->files[index] = file;
7541 err = io_sqe_file_register(ctx, file, i);
7543 table->files[index] = NULL;
7554 percpu_ref_kill(data->cur_refs);
7555 spin_lock(&data->lock);
7556 list_add(&ref_node->node, &data->ref_list);
7557 data->cur_refs = &ref_node->refs;
7558 spin_unlock(&data->lock);
7559 percpu_ref_get(&ctx->file_data->refs);
7561 destroy_fixed_file_ref_node(ref_node);
7563 return done ? done : err;
7566 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7569 struct io_uring_files_update up;
7571 if (!ctx->file_data)
7575 if (copy_from_user(&up, arg, sizeof(up)))
7580 return __io_sqe_files_update(ctx, &up, nr_args);
7583 static void io_free_work(struct io_wq_work *work)
7585 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7587 /* Consider that io_steal_work() relies on this ref */
7591 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7592 struct io_uring_params *p)
7594 struct io_wq_data data;
7596 struct io_ring_ctx *ctx_attach;
7597 unsigned int concurrency;
7600 data.user = ctx->user;
7601 data.free_work = io_free_work;
7602 data.do_work = io_wq_submit_work;
7604 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7605 /* Do QD, or 4 * CPUS, whatever is smallest */
7606 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7608 ctx->io_wq = io_wq_create(concurrency, &data);
7609 if (IS_ERR(ctx->io_wq)) {
7610 ret = PTR_ERR(ctx->io_wq);
7616 f = fdget(p->wq_fd);
7620 if (f.file->f_op != &io_uring_fops) {
7625 ctx_attach = f.file->private_data;
7626 /* @io_wq is protected by holding the fd */
7627 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7632 ctx->io_wq = ctx_attach->io_wq;
7638 static int io_uring_alloc_task_context(struct task_struct *task)
7640 struct io_uring_task *tctx;
7642 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7643 if (unlikely(!tctx))
7647 init_waitqueue_head(&tctx->wait);
7650 atomic_long_set(&tctx->req_issue, 0);
7651 atomic_long_set(&tctx->req_complete, 0);
7652 task->io_uring = tctx;
7656 void __io_uring_free(struct task_struct *tsk)
7658 struct io_uring_task *tctx = tsk->io_uring;
7660 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7662 tsk->io_uring = NULL;
7665 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7666 struct io_uring_params *p)
7670 if (ctx->flags & IORING_SETUP_SQPOLL) {
7671 struct io_sq_data *sqd;
7674 if (!capable(CAP_SYS_ADMIN))
7677 sqd = io_get_sq_data(p);
7684 io_sq_thread_park(sqd);
7685 mutex_lock(&sqd->ctx_lock);
7686 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7687 mutex_unlock(&sqd->ctx_lock);
7688 io_sq_thread_unpark(sqd);
7690 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7691 if (!ctx->sq_thread_idle)
7692 ctx->sq_thread_idle = HZ;
7697 if (p->flags & IORING_SETUP_SQ_AFF) {
7698 int cpu = p->sq_thread_cpu;
7701 if (cpu >= nr_cpu_ids)
7703 if (!cpu_online(cpu))
7706 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7707 cpu, "io_uring-sq");
7709 sqd->thread = kthread_create(io_sq_thread, sqd,
7712 if (IS_ERR(sqd->thread)) {
7713 ret = PTR_ERR(sqd->thread);
7717 ret = io_uring_alloc_task_context(sqd->thread);
7720 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7721 /* Can't have SQ_AFF without SQPOLL */
7727 ret = io_init_wq_offload(ctx, p);
7733 io_finish_async(ctx);
7737 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7739 struct io_sq_data *sqd = ctx->sq_data;
7741 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7742 wake_up_process(sqd->thread);
7745 static inline void __io_unaccount_mem(struct user_struct *user,
7746 unsigned long nr_pages)
7748 atomic_long_sub(nr_pages, &user->locked_vm);
7751 static inline int __io_account_mem(struct user_struct *user,
7752 unsigned long nr_pages)
7754 unsigned long page_limit, cur_pages, new_pages;
7756 /* Don't allow more pages than we can safely lock */
7757 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7760 cur_pages = atomic_long_read(&user->locked_vm);
7761 new_pages = cur_pages + nr_pages;
7762 if (new_pages > page_limit)
7764 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7765 new_pages) != cur_pages);
7770 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7771 enum io_mem_account acct)
7774 __io_unaccount_mem(ctx->user, nr_pages);
7776 if (ctx->mm_account) {
7777 if (acct == ACCT_LOCKED)
7778 ctx->mm_account->locked_vm -= nr_pages;
7779 else if (acct == ACCT_PINNED)
7780 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7784 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7785 enum io_mem_account acct)
7789 if (ctx->limit_mem) {
7790 ret = __io_account_mem(ctx->user, nr_pages);
7795 if (ctx->mm_account) {
7796 if (acct == ACCT_LOCKED)
7797 ctx->mm_account->locked_vm += nr_pages;
7798 else if (acct == ACCT_PINNED)
7799 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7805 static void io_mem_free(void *ptr)
7812 page = virt_to_head_page(ptr);
7813 if (put_page_testzero(page))
7814 free_compound_page(page);
7817 static void *io_mem_alloc(size_t size)
7819 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7822 return (void *) __get_free_pages(gfp_flags, get_order(size));
7825 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7828 struct io_rings *rings;
7829 size_t off, sq_array_size;
7831 off = struct_size(rings, cqes, cq_entries);
7832 if (off == SIZE_MAX)
7836 off = ALIGN(off, SMP_CACHE_BYTES);
7844 sq_array_size = array_size(sizeof(u32), sq_entries);
7845 if (sq_array_size == SIZE_MAX)
7848 if (check_add_overflow(off, sq_array_size, &off))
7854 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7858 pages = (size_t)1 << get_order(
7859 rings_size(sq_entries, cq_entries, NULL));
7860 pages += (size_t)1 << get_order(
7861 array_size(sizeof(struct io_uring_sqe), sq_entries));
7866 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7870 if (!ctx->user_bufs)
7873 for (i = 0; i < ctx->nr_user_bufs; i++) {
7874 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7876 for (j = 0; j < imu->nr_bvecs; j++)
7877 unpin_user_page(imu->bvec[j].bv_page);
7879 if (imu->acct_pages)
7880 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7885 kfree(ctx->user_bufs);
7886 ctx->user_bufs = NULL;
7887 ctx->nr_user_bufs = 0;
7891 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7892 void __user *arg, unsigned index)
7894 struct iovec __user *src;
7896 #ifdef CONFIG_COMPAT
7898 struct compat_iovec __user *ciovs;
7899 struct compat_iovec ciov;
7901 ciovs = (struct compat_iovec __user *) arg;
7902 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7905 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7906 dst->iov_len = ciov.iov_len;
7910 src = (struct iovec __user *) arg;
7911 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7917 * Not super efficient, but this is just a registration time. And we do cache
7918 * the last compound head, so generally we'll only do a full search if we don't
7921 * We check if the given compound head page has already been accounted, to
7922 * avoid double accounting it. This allows us to account the full size of the
7923 * page, not just the constituent pages of a huge page.
7925 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
7926 int nr_pages, struct page *hpage)
7930 /* check current page array */
7931 for (i = 0; i < nr_pages; i++) {
7932 if (!PageCompound(pages[i]))
7934 if (compound_head(pages[i]) == hpage)
7938 /* check previously registered pages */
7939 for (i = 0; i < ctx->nr_user_bufs; i++) {
7940 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7942 for (j = 0; j < imu->nr_bvecs; j++) {
7943 if (!PageCompound(imu->bvec[j].bv_page))
7945 if (compound_head(imu->bvec[j].bv_page) == hpage)
7953 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
7954 int nr_pages, struct io_mapped_ubuf *imu,
7955 struct page **last_hpage)
7959 for (i = 0; i < nr_pages; i++) {
7960 if (!PageCompound(pages[i])) {
7965 hpage = compound_head(pages[i]);
7966 if (hpage == *last_hpage)
7968 *last_hpage = hpage;
7969 if (headpage_already_acct(ctx, pages, i, hpage))
7971 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
7975 if (!imu->acct_pages)
7978 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
7980 imu->acct_pages = 0;
7984 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7987 struct vm_area_struct **vmas = NULL;
7988 struct page **pages = NULL;
7989 struct page *last_hpage = NULL;
7990 int i, j, got_pages = 0;
7995 if (!nr_args || nr_args > UIO_MAXIOV)
7998 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8000 if (!ctx->user_bufs)
8003 for (i = 0; i < nr_args; i++) {
8004 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8005 unsigned long off, start, end, ubuf;
8010 ret = io_copy_iov(ctx, &iov, arg, i);
8015 * Don't impose further limits on the size and buffer
8016 * constraints here, we'll -EINVAL later when IO is
8017 * submitted if they are wrong.
8020 if (!iov.iov_base || !iov.iov_len)
8023 /* arbitrary limit, but we need something */
8024 if (iov.iov_len > SZ_1G)
8027 ubuf = (unsigned long) iov.iov_base;
8028 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8029 start = ubuf >> PAGE_SHIFT;
8030 nr_pages = end - start;
8033 if (!pages || nr_pages > got_pages) {
8036 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8038 vmas = kvmalloc_array(nr_pages,
8039 sizeof(struct vm_area_struct *),
8041 if (!pages || !vmas) {
8045 got_pages = nr_pages;
8048 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8055 mmap_read_lock(current->mm);
8056 pret = pin_user_pages(ubuf, nr_pages,
8057 FOLL_WRITE | FOLL_LONGTERM,
8059 if (pret == nr_pages) {
8060 /* don't support file backed memory */
8061 for (j = 0; j < nr_pages; j++) {
8062 struct vm_area_struct *vma = vmas[j];
8065 !is_file_hugepages(vma->vm_file)) {
8071 ret = pret < 0 ? pret : -EFAULT;
8073 mmap_read_unlock(current->mm);
8076 * if we did partial map, or found file backed vmas,
8077 * release any pages we did get
8080 unpin_user_pages(pages, pret);
8085 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8087 unpin_user_pages(pages, pret);
8092 off = ubuf & ~PAGE_MASK;
8094 for (j = 0; j < nr_pages; j++) {
8097 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8098 imu->bvec[j].bv_page = pages[j];
8099 imu->bvec[j].bv_len = vec_len;
8100 imu->bvec[j].bv_offset = off;
8104 /* store original address for later verification */
8106 imu->len = iov.iov_len;
8107 imu->nr_bvecs = nr_pages;
8109 ctx->nr_user_bufs++;
8117 io_sqe_buffer_unregister(ctx);
8121 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8123 __s32 __user *fds = arg;
8129 if (copy_from_user(&fd, fds, sizeof(*fds)))
8132 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8133 if (IS_ERR(ctx->cq_ev_fd)) {
8134 int ret = PTR_ERR(ctx->cq_ev_fd);
8135 ctx->cq_ev_fd = NULL;
8142 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8144 if (ctx->cq_ev_fd) {
8145 eventfd_ctx_put(ctx->cq_ev_fd);
8146 ctx->cq_ev_fd = NULL;
8153 static int __io_destroy_buffers(int id, void *p, void *data)
8155 struct io_ring_ctx *ctx = data;
8156 struct io_buffer *buf = p;
8158 __io_remove_buffers(ctx, buf, id, -1U);
8162 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8164 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8165 idr_destroy(&ctx->io_buffer_idr);
8168 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8170 io_finish_async(ctx);
8171 io_sqe_buffer_unregister(ctx);
8173 if (ctx->sqo_task) {
8174 put_task_struct(ctx->sqo_task);
8175 ctx->sqo_task = NULL;
8176 mmdrop(ctx->mm_account);
8177 ctx->mm_account = NULL;
8180 #ifdef CONFIG_BLK_CGROUP
8181 if (ctx->sqo_blkcg_css)
8182 css_put(ctx->sqo_blkcg_css);
8185 io_sqe_files_unregister(ctx);
8186 io_eventfd_unregister(ctx);
8187 io_destroy_buffers(ctx);
8188 idr_destroy(&ctx->personality_idr);
8190 #if defined(CONFIG_UNIX)
8191 if (ctx->ring_sock) {
8192 ctx->ring_sock->file = NULL; /* so that iput() is called */
8193 sock_release(ctx->ring_sock);
8197 io_mem_free(ctx->rings);
8198 io_mem_free(ctx->sq_sqes);
8200 percpu_ref_exit(&ctx->refs);
8201 free_uid(ctx->user);
8202 put_cred(ctx->creds);
8203 kfree(ctx->cancel_hash);
8204 kmem_cache_free(req_cachep, ctx->fallback_req);
8208 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8210 struct io_ring_ctx *ctx = file->private_data;
8213 poll_wait(file, &ctx->cq_wait, wait);
8215 * synchronizes with barrier from wq_has_sleeper call in
8219 if (!io_sqring_full(ctx))
8220 mask |= EPOLLOUT | EPOLLWRNORM;
8221 if (io_cqring_events(ctx, false))
8222 mask |= EPOLLIN | EPOLLRDNORM;
8227 static int io_uring_fasync(int fd, struct file *file, int on)
8229 struct io_ring_ctx *ctx = file->private_data;
8231 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8234 static int io_remove_personalities(int id, void *p, void *data)
8236 struct io_ring_ctx *ctx = data;
8237 const struct cred *cred;
8239 cred = idr_remove(&ctx->personality_idr, id);
8245 static void io_ring_exit_work(struct work_struct *work)
8247 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8251 * If we're doing polled IO and end up having requests being
8252 * submitted async (out-of-line), then completions can come in while
8253 * we're waiting for refs to drop. We need to reap these manually,
8254 * as nobody else will be looking for them.
8258 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8259 io_iopoll_try_reap_events(ctx);
8260 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8261 io_ring_ctx_free(ctx);
8264 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8266 mutex_lock(&ctx->uring_lock);
8267 percpu_ref_kill(&ctx->refs);
8268 mutex_unlock(&ctx->uring_lock);
8270 io_kill_timeouts(ctx, NULL);
8271 io_poll_remove_all(ctx, NULL);
8274 io_wq_cancel_all(ctx->io_wq);
8276 /* if we failed setting up the ctx, we might not have any rings */
8278 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8279 io_iopoll_try_reap_events(ctx);
8280 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8283 * Do this upfront, so we won't have a grace period where the ring
8284 * is closed but resources aren't reaped yet. This can cause
8285 * spurious failure in setting up a new ring.
8287 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8290 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8292 * Use system_unbound_wq to avoid spawning tons of event kworkers
8293 * if we're exiting a ton of rings at the same time. It just adds
8294 * noise and overhead, there's no discernable change in runtime
8295 * over using system_wq.
8297 queue_work(system_unbound_wq, &ctx->exit_work);
8300 static int io_uring_release(struct inode *inode, struct file *file)
8302 struct io_ring_ctx *ctx = file->private_data;
8304 file->private_data = NULL;
8305 io_ring_ctx_wait_and_kill(ctx);
8309 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8311 struct files_struct *files = data;
8313 return !files || work->files == files;
8317 * Returns true if 'preq' is the link parent of 'req'
8319 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8321 struct io_kiocb *link;
8323 if (!(preq->flags & REQ_F_LINK_HEAD))
8326 list_for_each_entry(link, &preq->link_list, link_list) {
8334 static bool io_match_link_files(struct io_kiocb *req,
8335 struct files_struct *files)
8337 struct io_kiocb *link;
8339 if (io_match_files(req, files))
8341 if (req->flags & REQ_F_LINK_HEAD) {
8342 list_for_each_entry(link, &req->link_list, link_list) {
8343 if (io_match_files(link, files))
8351 * We're looking to cancel 'req' because it's holding on to our files, but
8352 * 'req' could be a link to another request. See if it is, and cancel that
8353 * parent request if so.
8355 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8357 struct hlist_node *tmp;
8358 struct io_kiocb *preq;
8362 spin_lock_irq(&ctx->completion_lock);
8363 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8364 struct hlist_head *list;
8366 list = &ctx->cancel_hash[i];
8367 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8368 found = io_match_link(preq, req);
8370 io_poll_remove_one(preq);
8375 spin_unlock_irq(&ctx->completion_lock);
8379 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8380 struct io_kiocb *req)
8382 struct io_kiocb *preq;
8385 spin_lock_irq(&ctx->completion_lock);
8386 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8387 found = io_match_link(preq, req);
8389 __io_timeout_cancel(preq);
8393 spin_unlock_irq(&ctx->completion_lock);
8397 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8399 return io_match_link(container_of(work, struct io_kiocb, work), data);
8402 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8404 enum io_wq_cancel cret;
8406 /* cancel this particular work, if it's running */
8407 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8408 if (cret != IO_WQ_CANCEL_NOTFOUND)
8411 /* find links that hold this pending, cancel those */
8412 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8413 if (cret != IO_WQ_CANCEL_NOTFOUND)
8416 /* if we have a poll link holding this pending, cancel that */
8417 if (io_poll_remove_link(ctx, req))
8420 /* final option, timeout link is holding this req pending */
8421 io_timeout_remove_link(ctx, req);
8424 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8425 struct files_struct *files)
8427 struct io_defer_entry *de = NULL;
8430 spin_lock_irq(&ctx->completion_lock);
8431 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8432 if (io_match_link_files(de->req, files)) {
8433 list_cut_position(&list, &ctx->defer_list, &de->list);
8437 spin_unlock_irq(&ctx->completion_lock);
8439 while (!list_empty(&list)) {
8440 de = list_first_entry(&list, struct io_defer_entry, list);
8441 list_del_init(&de->list);
8442 req_set_fail_links(de->req);
8443 io_put_req(de->req);
8444 io_req_complete(de->req, -ECANCELED);
8450 * Returns true if we found and killed one or more files pinning requests
8452 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8453 struct files_struct *files)
8455 if (list_empty_careful(&ctx->inflight_list))
8458 io_cancel_defer_files(ctx, files);
8459 /* cancel all at once, should be faster than doing it one by one*/
8460 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8462 while (!list_empty_careful(&ctx->inflight_list)) {
8463 struct io_kiocb *cancel_req = NULL, *req;
8466 spin_lock_irq(&ctx->inflight_lock);
8467 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8468 if (files && req->work.files != files)
8470 /* req is being completed, ignore */
8471 if (!refcount_inc_not_zero(&req->refs))
8477 prepare_to_wait(&ctx->inflight_wait, &wait,
8478 TASK_UNINTERRUPTIBLE);
8479 spin_unlock_irq(&ctx->inflight_lock);
8481 /* We need to keep going until we don't find a matching req */
8484 /* cancel this request, or head link requests */
8485 io_attempt_cancel(ctx, cancel_req);
8486 io_put_req(cancel_req);
8487 /* cancellations _may_ trigger task work */
8490 finish_wait(&ctx->inflight_wait, &wait);
8496 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8498 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8499 struct task_struct *task = data;
8501 return io_task_match(req, task);
8504 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8505 struct task_struct *task,
8506 struct files_struct *files)
8510 ret = io_uring_cancel_files(ctx, files);
8512 enum io_wq_cancel cret;
8514 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8515 if (cret != IO_WQ_CANCEL_NOTFOUND)
8518 /* SQPOLL thread does its own polling */
8519 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8520 while (!list_empty_careful(&ctx->iopoll_list)) {
8521 io_iopoll_try_reap_events(ctx);
8526 ret |= io_poll_remove_all(ctx, task);
8527 ret |= io_kill_timeouts(ctx, task);
8534 * We need to iteratively cancel requests, in case a request has dependent
8535 * hard links. These persist even for failure of cancelations, hence keep
8536 * looping until none are found.
8538 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8539 struct files_struct *files)
8541 struct task_struct *task = current;
8543 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8544 task = ctx->sq_data->thread;
8546 io_cqring_overflow_flush(ctx, true, task, files);
8548 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8555 * Note that this task has used io_uring. We use it for cancelation purposes.
8557 static int io_uring_add_task_file(struct file *file)
8559 struct io_uring_task *tctx = current->io_uring;
8561 if (unlikely(!tctx)) {
8564 ret = io_uring_alloc_task_context(current);
8567 tctx = current->io_uring;
8569 if (tctx->last != file) {
8570 void *old = xa_load(&tctx->xa, (unsigned long)file);
8574 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8583 * Remove this io_uring_file -> task mapping.
8585 static void io_uring_del_task_file(struct file *file)
8587 struct io_uring_task *tctx = current->io_uring;
8589 if (tctx->last == file)
8591 file = xa_erase(&tctx->xa, (unsigned long)file);
8596 static void __io_uring_attempt_task_drop(struct file *file)
8598 struct file *old = xa_load(¤t->io_uring->xa, (unsigned long)file);
8601 io_uring_del_task_file(file);
8605 * Drop task note for this file if we're the only ones that hold it after
8608 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8610 if (!current->io_uring)
8613 * fput() is pending, will be 2 if the only other ref is our potential
8614 * task file note. If the task is exiting, drop regardless of count.
8616 if (!exiting && atomic_long_read(&file->f_count) != 2)
8619 __io_uring_attempt_task_drop(file);
8622 void __io_uring_files_cancel(struct files_struct *files)
8624 struct io_uring_task *tctx = current->io_uring;
8626 unsigned long index;
8628 /* make sure overflow events are dropped */
8629 tctx->in_idle = true;
8631 xa_for_each(&tctx->xa, index, file) {
8632 struct io_ring_ctx *ctx = file->private_data;
8634 io_uring_cancel_task_requests(ctx, files);
8636 io_uring_del_task_file(file);
8640 static inline bool io_uring_task_idle(struct io_uring_task *tctx)
8642 return atomic_long_read(&tctx->req_issue) ==
8643 atomic_long_read(&tctx->req_complete);
8647 * Find any io_uring fd that this task has registered or done IO on, and cancel
8650 void __io_uring_task_cancel(void)
8652 struct io_uring_task *tctx = current->io_uring;
8656 /* make sure overflow events are dropped */
8657 tctx->in_idle = true;
8659 while (!io_uring_task_idle(tctx)) {
8660 /* read completions before cancelations */
8661 completions = atomic_long_read(&tctx->req_complete);
8662 __io_uring_files_cancel(NULL);
8664 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8667 * If we've seen completions, retry. This avoids a race where
8668 * a completion comes in before we did prepare_to_wait().
8670 if (completions != atomic_long_read(&tctx->req_complete))
8672 if (io_uring_task_idle(tctx))
8677 finish_wait(&tctx->wait, &wait);
8678 tctx->in_idle = false;
8681 static int io_uring_flush(struct file *file, void *data)
8683 struct io_ring_ctx *ctx = file->private_data;
8686 * If the task is going away, cancel work it may have pending
8688 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8691 io_uring_cancel_task_requests(ctx, data);
8692 io_uring_attempt_task_drop(file, !data);
8696 static void *io_uring_validate_mmap_request(struct file *file,
8697 loff_t pgoff, size_t sz)
8699 struct io_ring_ctx *ctx = file->private_data;
8700 loff_t offset = pgoff << PAGE_SHIFT;
8705 case IORING_OFF_SQ_RING:
8706 case IORING_OFF_CQ_RING:
8709 case IORING_OFF_SQES:
8713 return ERR_PTR(-EINVAL);
8716 page = virt_to_head_page(ptr);
8717 if (sz > page_size(page))
8718 return ERR_PTR(-EINVAL);
8725 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8727 size_t sz = vma->vm_end - vma->vm_start;
8731 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8733 return PTR_ERR(ptr);
8735 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8736 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8739 #else /* !CONFIG_MMU */
8741 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8743 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8746 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8748 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8751 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8752 unsigned long addr, unsigned long len,
8753 unsigned long pgoff, unsigned long flags)
8757 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8759 return PTR_ERR(ptr);
8761 return (unsigned long) ptr;
8764 #endif /* !CONFIG_MMU */
8766 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8771 if (!io_sqring_full(ctx))
8774 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8776 if (!io_sqring_full(ctx))
8780 } while (!signal_pending(current));
8782 finish_wait(&ctx->sqo_sq_wait, &wait);
8785 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8786 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8789 struct io_ring_ctx *ctx;
8796 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8797 IORING_ENTER_SQ_WAIT))
8805 if (f.file->f_op != &io_uring_fops)
8809 ctx = f.file->private_data;
8810 if (!percpu_ref_tryget(&ctx->refs))
8814 if (ctx->flags & IORING_SETUP_R_DISABLED)
8818 * For SQ polling, the thread will do all submissions and completions.
8819 * Just return the requested submit count, and wake the thread if
8823 if (ctx->flags & IORING_SETUP_SQPOLL) {
8824 if (!list_empty_careful(&ctx->cq_overflow_list))
8825 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8826 if (flags & IORING_ENTER_SQ_WAKEUP)
8827 wake_up(&ctx->sq_data->wait);
8828 if (flags & IORING_ENTER_SQ_WAIT)
8829 io_sqpoll_wait_sq(ctx);
8830 submitted = to_submit;
8831 } else if (to_submit) {
8832 ret = io_uring_add_task_file(f.file);
8835 mutex_lock(&ctx->uring_lock);
8836 submitted = io_submit_sqes(ctx, to_submit);
8837 mutex_unlock(&ctx->uring_lock);
8839 if (submitted != to_submit)
8842 if (flags & IORING_ENTER_GETEVENTS) {
8843 min_complete = min(min_complete, ctx->cq_entries);
8846 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8847 * space applications don't need to do io completion events
8848 * polling again, they can rely on io_sq_thread to do polling
8849 * work, which can reduce cpu usage and uring_lock contention.
8851 if (ctx->flags & IORING_SETUP_IOPOLL &&
8852 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8853 ret = io_iopoll_check(ctx, min_complete);
8855 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8860 percpu_ref_put(&ctx->refs);
8863 return submitted ? submitted : ret;
8866 #ifdef CONFIG_PROC_FS
8867 static int io_uring_show_cred(int id, void *p, void *data)
8869 const struct cred *cred = p;
8870 struct seq_file *m = data;
8871 struct user_namespace *uns = seq_user_ns(m);
8872 struct group_info *gi;
8877 seq_printf(m, "%5d\n", id);
8878 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8879 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8880 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8881 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8882 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8883 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8884 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8885 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8886 seq_puts(m, "\n\tGroups:\t");
8887 gi = cred->group_info;
8888 for (g = 0; g < gi->ngroups; g++) {
8889 seq_put_decimal_ull(m, g ? " " : "",
8890 from_kgid_munged(uns, gi->gid[g]));
8892 seq_puts(m, "\n\tCapEff:\t");
8893 cap = cred->cap_effective;
8894 CAP_FOR_EACH_U32(__capi)
8895 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8900 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8902 struct io_sq_data *sq = NULL;
8907 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8908 * since fdinfo case grabs it in the opposite direction of normal use
8909 * cases. If we fail to get the lock, we just don't iterate any
8910 * structures that could be going away outside the io_uring mutex.
8912 has_lock = mutex_trylock(&ctx->uring_lock);
8914 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
8917 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
8918 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
8919 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8920 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8921 struct fixed_file_table *table;
8924 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8925 f = table->files[i & IORING_FILE_TABLE_MASK];
8927 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8929 seq_printf(m, "%5u: <none>\n", i);
8931 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8932 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8933 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8935 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8936 (unsigned int) buf->len);
8938 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8939 seq_printf(m, "Personalities:\n");
8940 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8942 seq_printf(m, "PollList:\n");
8943 spin_lock_irq(&ctx->completion_lock);
8944 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8945 struct hlist_head *list = &ctx->cancel_hash[i];
8946 struct io_kiocb *req;
8948 hlist_for_each_entry(req, list, hash_node)
8949 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8950 req->task->task_works != NULL);
8952 spin_unlock_irq(&ctx->completion_lock);
8954 mutex_unlock(&ctx->uring_lock);
8957 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8959 struct io_ring_ctx *ctx = f->private_data;
8961 if (percpu_ref_tryget(&ctx->refs)) {
8962 __io_uring_show_fdinfo(ctx, m);
8963 percpu_ref_put(&ctx->refs);
8968 static const struct file_operations io_uring_fops = {
8969 .release = io_uring_release,
8970 .flush = io_uring_flush,
8971 .mmap = io_uring_mmap,
8973 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8974 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8976 .poll = io_uring_poll,
8977 .fasync = io_uring_fasync,
8978 #ifdef CONFIG_PROC_FS
8979 .show_fdinfo = io_uring_show_fdinfo,
8983 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8984 struct io_uring_params *p)
8986 struct io_rings *rings;
8987 size_t size, sq_array_offset;
8989 /* make sure these are sane, as we already accounted them */
8990 ctx->sq_entries = p->sq_entries;
8991 ctx->cq_entries = p->cq_entries;
8993 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8994 if (size == SIZE_MAX)
8997 rings = io_mem_alloc(size);
9002 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9003 rings->sq_ring_mask = p->sq_entries - 1;
9004 rings->cq_ring_mask = p->cq_entries - 1;
9005 rings->sq_ring_entries = p->sq_entries;
9006 rings->cq_ring_entries = p->cq_entries;
9007 ctx->sq_mask = rings->sq_ring_mask;
9008 ctx->cq_mask = rings->cq_ring_mask;
9010 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9011 if (size == SIZE_MAX) {
9012 io_mem_free(ctx->rings);
9017 ctx->sq_sqes = io_mem_alloc(size);
9018 if (!ctx->sq_sqes) {
9019 io_mem_free(ctx->rings);
9028 * Allocate an anonymous fd, this is what constitutes the application
9029 * visible backing of an io_uring instance. The application mmaps this
9030 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9031 * we have to tie this fd to a socket for file garbage collection purposes.
9033 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9038 #if defined(CONFIG_UNIX)
9039 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9045 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9049 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9050 O_RDWR | O_CLOEXEC);
9054 ret = PTR_ERR(file);
9058 #if defined(CONFIG_UNIX)
9059 ctx->ring_sock->file = file;
9061 if (unlikely(io_uring_add_task_file(file))) {
9062 file = ERR_PTR(-ENOMEM);
9065 fd_install(ret, file);
9068 #if defined(CONFIG_UNIX)
9069 sock_release(ctx->ring_sock);
9070 ctx->ring_sock = NULL;
9075 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9076 struct io_uring_params __user *params)
9078 struct user_struct *user = NULL;
9079 struct io_ring_ctx *ctx;
9085 if (entries > IORING_MAX_ENTRIES) {
9086 if (!(p->flags & IORING_SETUP_CLAMP))
9088 entries = IORING_MAX_ENTRIES;
9092 * Use twice as many entries for the CQ ring. It's possible for the
9093 * application to drive a higher depth than the size of the SQ ring,
9094 * since the sqes are only used at submission time. This allows for
9095 * some flexibility in overcommitting a bit. If the application has
9096 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9097 * of CQ ring entries manually.
9099 p->sq_entries = roundup_pow_of_two(entries);
9100 if (p->flags & IORING_SETUP_CQSIZE) {
9102 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9103 * to a power-of-two, if it isn't already. We do NOT impose
9104 * any cq vs sq ring sizing.
9106 if (p->cq_entries < p->sq_entries)
9108 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9109 if (!(p->flags & IORING_SETUP_CLAMP))
9111 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9113 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9115 p->cq_entries = 2 * p->sq_entries;
9118 user = get_uid(current_user());
9119 limit_mem = !capable(CAP_IPC_LOCK);
9122 ret = __io_account_mem(user,
9123 ring_pages(p->sq_entries, p->cq_entries));
9130 ctx = io_ring_ctx_alloc(p);
9133 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9138 ctx->compat = in_compat_syscall();
9140 ctx->creds = get_current_cred();
9142 ctx->sqo_task = get_task_struct(current);
9145 * This is just grabbed for accounting purposes. When a process exits,
9146 * the mm is exited and dropped before the files, hence we need to hang
9147 * on to this mm purely for the purposes of being able to unaccount
9148 * memory (locked/pinned vm). It's not used for anything else.
9150 mmgrab(current->mm);
9151 ctx->mm_account = current->mm;
9153 #ifdef CONFIG_BLK_CGROUP
9155 * The sq thread will belong to the original cgroup it was inited in.
9156 * If the cgroup goes offline (e.g. disabling the io controller), then
9157 * issued bios will be associated with the closest cgroup later in the
9161 ctx->sqo_blkcg_css = blkcg_css();
9162 ret = css_tryget_online(ctx->sqo_blkcg_css);
9165 /* don't init against a dying cgroup, have the user try again */
9166 ctx->sqo_blkcg_css = NULL;
9173 * Account memory _before_ installing the file descriptor. Once
9174 * the descriptor is installed, it can get closed at any time. Also
9175 * do this before hitting the general error path, as ring freeing
9176 * will un-account as well.
9178 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9180 ctx->limit_mem = limit_mem;
9182 ret = io_allocate_scq_urings(ctx, p);
9186 ret = io_sq_offload_create(ctx, p);
9190 if (!(p->flags & IORING_SETUP_R_DISABLED))
9191 io_sq_offload_start(ctx);
9193 memset(&p->sq_off, 0, sizeof(p->sq_off));
9194 p->sq_off.head = offsetof(struct io_rings, sq.head);
9195 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9196 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9197 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9198 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9199 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9200 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9202 memset(&p->cq_off, 0, sizeof(p->cq_off));
9203 p->cq_off.head = offsetof(struct io_rings, cq.head);
9204 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9205 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9206 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9207 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9208 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9209 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9211 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9212 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9213 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9214 IORING_FEAT_POLL_32BITS;
9216 if (copy_to_user(params, p, sizeof(*p))) {
9222 * Install ring fd as the very last thing, so we don't risk someone
9223 * having closed it before we finish setup
9225 ret = io_uring_get_fd(ctx);
9229 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9232 io_ring_ctx_wait_and_kill(ctx);
9237 * Sets up an aio uring context, and returns the fd. Applications asks for a
9238 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9239 * params structure passed in.
9241 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9243 struct io_uring_params p;
9246 if (copy_from_user(&p, params, sizeof(p)))
9248 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9253 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9254 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9255 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9256 IORING_SETUP_R_DISABLED))
9259 return io_uring_create(entries, &p, params);
9262 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9263 struct io_uring_params __user *, params)
9265 return io_uring_setup(entries, params);
9268 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9270 struct io_uring_probe *p;
9274 size = struct_size(p, ops, nr_args);
9275 if (size == SIZE_MAX)
9277 p = kzalloc(size, GFP_KERNEL);
9282 if (copy_from_user(p, arg, size))
9285 if (memchr_inv(p, 0, size))
9288 p->last_op = IORING_OP_LAST - 1;
9289 if (nr_args > IORING_OP_LAST)
9290 nr_args = IORING_OP_LAST;
9292 for (i = 0; i < nr_args; i++) {
9294 if (!io_op_defs[i].not_supported)
9295 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9300 if (copy_to_user(arg, p, size))
9307 static int io_register_personality(struct io_ring_ctx *ctx)
9309 const struct cred *creds = get_current_cred();
9312 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9313 USHRT_MAX, GFP_KERNEL);
9319 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9321 const struct cred *old_creds;
9323 old_creds = idr_remove(&ctx->personality_idr, id);
9325 put_cred(old_creds);
9332 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9333 unsigned int nr_args)
9335 struct io_uring_restriction *res;
9339 /* Restrictions allowed only if rings started disabled */
9340 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9343 /* We allow only a single restrictions registration */
9344 if (ctx->restrictions.registered)
9347 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9350 size = array_size(nr_args, sizeof(*res));
9351 if (size == SIZE_MAX)
9354 res = memdup_user(arg, size);
9356 return PTR_ERR(res);
9360 for (i = 0; i < nr_args; i++) {
9361 switch (res[i].opcode) {
9362 case IORING_RESTRICTION_REGISTER_OP:
9363 if (res[i].register_op >= IORING_REGISTER_LAST) {
9368 __set_bit(res[i].register_op,
9369 ctx->restrictions.register_op);
9371 case IORING_RESTRICTION_SQE_OP:
9372 if (res[i].sqe_op >= IORING_OP_LAST) {
9377 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9379 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9380 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9382 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9383 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9392 /* Reset all restrictions if an error happened */
9394 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9396 ctx->restrictions.registered = true;
9402 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9404 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9407 if (ctx->restrictions.registered)
9408 ctx->restricted = 1;
9410 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9412 io_sq_offload_start(ctx);
9417 static bool io_register_op_must_quiesce(int op)
9420 case IORING_UNREGISTER_FILES:
9421 case IORING_REGISTER_FILES_UPDATE:
9422 case IORING_REGISTER_PROBE:
9423 case IORING_REGISTER_PERSONALITY:
9424 case IORING_UNREGISTER_PERSONALITY:
9431 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9432 void __user *arg, unsigned nr_args)
9433 __releases(ctx->uring_lock)
9434 __acquires(ctx->uring_lock)
9439 * We're inside the ring mutex, if the ref is already dying, then
9440 * someone else killed the ctx or is already going through
9441 * io_uring_register().
9443 if (percpu_ref_is_dying(&ctx->refs))
9446 if (io_register_op_must_quiesce(opcode)) {
9447 percpu_ref_kill(&ctx->refs);
9450 * Drop uring mutex before waiting for references to exit. If
9451 * another thread is currently inside io_uring_enter() it might
9452 * need to grab the uring_lock to make progress. If we hold it
9453 * here across the drain wait, then we can deadlock. It's safe
9454 * to drop the mutex here, since no new references will come in
9455 * after we've killed the percpu ref.
9457 mutex_unlock(&ctx->uring_lock);
9459 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9462 ret = io_run_task_work_sig();
9467 mutex_lock(&ctx->uring_lock);
9470 percpu_ref_resurrect(&ctx->refs);
9475 if (ctx->restricted) {
9476 if (opcode >= IORING_REGISTER_LAST) {
9481 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9488 case IORING_REGISTER_BUFFERS:
9489 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9491 case IORING_UNREGISTER_BUFFERS:
9495 ret = io_sqe_buffer_unregister(ctx);
9497 case IORING_REGISTER_FILES:
9498 ret = io_sqe_files_register(ctx, arg, nr_args);
9500 case IORING_UNREGISTER_FILES:
9504 ret = io_sqe_files_unregister(ctx);
9506 case IORING_REGISTER_FILES_UPDATE:
9507 ret = io_sqe_files_update(ctx, arg, nr_args);
9509 case IORING_REGISTER_EVENTFD:
9510 case IORING_REGISTER_EVENTFD_ASYNC:
9514 ret = io_eventfd_register(ctx, arg);
9517 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9518 ctx->eventfd_async = 1;
9520 ctx->eventfd_async = 0;
9522 case IORING_UNREGISTER_EVENTFD:
9526 ret = io_eventfd_unregister(ctx);
9528 case IORING_REGISTER_PROBE:
9530 if (!arg || nr_args > 256)
9532 ret = io_probe(ctx, arg, nr_args);
9534 case IORING_REGISTER_PERSONALITY:
9538 ret = io_register_personality(ctx);
9540 case IORING_UNREGISTER_PERSONALITY:
9544 ret = io_unregister_personality(ctx, nr_args);
9546 case IORING_REGISTER_ENABLE_RINGS:
9550 ret = io_register_enable_rings(ctx);
9552 case IORING_REGISTER_RESTRICTIONS:
9553 ret = io_register_restrictions(ctx, arg, nr_args);
9561 if (io_register_op_must_quiesce(opcode)) {
9562 /* bring the ctx back to life */
9563 percpu_ref_reinit(&ctx->refs);
9565 reinit_completion(&ctx->ref_comp);
9570 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9571 void __user *, arg, unsigned int, nr_args)
9573 struct io_ring_ctx *ctx;
9582 if (f.file->f_op != &io_uring_fops)
9585 ctx = f.file->private_data;
9587 mutex_lock(&ctx->uring_lock);
9588 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9589 mutex_unlock(&ctx->uring_lock);
9590 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9591 ctx->cq_ev_fd != NULL, ret);
9597 static int __init io_uring_init(void)
9599 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9600 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9601 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9604 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9605 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9606 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9607 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9608 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9609 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9610 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9611 BUILD_BUG_SQE_ELEM(8, __u64, off);
9612 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9613 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9614 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9615 BUILD_BUG_SQE_ELEM(24, __u32, len);
9616 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9617 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9618 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9619 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9620 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9621 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9622 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9623 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9624 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9625 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9626 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9627 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9628 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9629 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9630 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9631 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9632 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9633 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9634 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9636 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9637 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9638 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9641 __initcall(io_uring_init);