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 fixed_file_ref_node *node;
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;
439 struct list_head list;
442 struct io_timeout_rem {
448 /* NOTE: kiocb has the file as the first member, so don't do it here */
456 struct sockaddr __user *addr;
463 struct user_msghdr __user *umsg;
469 struct io_buffer *kbuf;
475 struct filename *filename;
477 unsigned long nofile;
480 struct io_files_update {
506 struct epoll_event event;
510 struct file *file_out;
511 struct file *file_in;
518 struct io_provide_buf {
532 const char __user *filename;
533 struct statx __user *buffer;
536 struct io_completion {
538 struct list_head list;
542 struct io_async_connect {
543 struct sockaddr_storage address;
546 struct io_async_msghdr {
547 struct iovec fast_iov[UIO_FASTIOV];
549 struct sockaddr __user *uaddr;
551 struct sockaddr_storage addr;
555 struct iovec fast_iov[UIO_FASTIOV];
556 const struct iovec *free_iovec;
557 struct iov_iter iter;
559 struct wait_page_queue wpq;
563 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
564 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
565 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
566 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
567 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
568 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
575 REQ_F_LINK_TIMEOUT_BIT,
577 REQ_F_COMP_LOCKED_BIT,
578 REQ_F_NEED_CLEANUP_BIT,
580 REQ_F_BUFFER_SELECTED_BIT,
581 REQ_F_NO_FILE_TABLE_BIT,
582 REQ_F_WORK_INITIALIZED_BIT,
584 /* not a real bit, just to check we're not overflowing the space */
590 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
591 /* drain existing IO first */
592 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
594 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
595 /* doesn't sever on completion < 0 */
596 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
598 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
599 /* IOSQE_BUFFER_SELECT */
600 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
603 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
604 /* fail rest of links */
605 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
606 /* on inflight list */
607 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
608 /* read/write uses file position */
609 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
610 /* must not punt to workers */
611 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
612 /* has linked timeout */
613 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
615 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
616 /* completion under lock */
617 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
619 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
620 /* already went through poll handler */
621 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
622 /* buffer already selected */
623 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
624 /* doesn't need file table for this request */
625 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
626 /* io_wq_work is initialized */
627 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
631 struct io_poll_iocb poll;
632 struct io_poll_iocb *double_poll;
636 * NOTE! Each of the iocb union members has the file pointer
637 * as the first entry in their struct definition. So you can
638 * access the file pointer through any of the sub-structs,
639 * or directly as just 'ki_filp' in this struct.
645 struct io_poll_iocb poll;
646 struct io_accept accept;
648 struct io_cancel cancel;
649 struct io_timeout timeout;
650 struct io_timeout_rem timeout_rem;
651 struct io_connect connect;
652 struct io_sr_msg sr_msg;
654 struct io_close close;
655 struct io_files_update files_update;
656 struct io_fadvise fadvise;
657 struct io_madvise madvise;
658 struct io_epoll epoll;
659 struct io_splice splice;
660 struct io_provide_buf pbuf;
661 struct io_statx statx;
662 /* use only after cleaning per-op data, see io_clean_op() */
663 struct io_completion compl;
666 /* opcode allocated if it needs to store data for async defer */
669 /* polled IO has completed */
675 struct io_ring_ctx *ctx;
678 struct task_struct *task;
681 struct list_head link_list;
684 * 1. used with ctx->iopoll_list with reads/writes
685 * 2. to track reqs with ->files (see io_op_def::file_table)
687 struct list_head inflight_entry;
689 struct percpu_ref *fixed_file_refs;
690 struct callback_head task_work;
691 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
692 struct hlist_node hash_node;
693 struct async_poll *apoll;
694 struct io_wq_work work;
697 struct io_defer_entry {
698 struct list_head list;
699 struct io_kiocb *req;
703 #define IO_IOPOLL_BATCH 8
705 struct io_comp_state {
707 struct list_head list;
708 struct io_ring_ctx *ctx;
711 struct io_submit_state {
712 struct blk_plug plug;
715 * io_kiocb alloc cache
717 void *reqs[IO_IOPOLL_BATCH];
718 unsigned int free_reqs;
721 * Batch completion logic
723 struct io_comp_state comp;
726 * File reference cache
730 unsigned int has_refs;
731 unsigned int ios_left;
735 /* needs current->mm setup, does mm access */
736 unsigned needs_mm : 1;
737 /* needs req->file assigned */
738 unsigned needs_file : 1;
739 /* don't fail if file grab fails */
740 unsigned needs_file_no_error : 1;
741 /* hash wq insertion if file is a regular file */
742 unsigned hash_reg_file : 1;
743 /* unbound wq insertion if file is a non-regular file */
744 unsigned unbound_nonreg_file : 1;
745 /* opcode is not supported by this kernel */
746 unsigned not_supported : 1;
747 /* needs file table */
748 unsigned file_table : 1;
750 unsigned needs_fs : 1;
751 /* set if opcode supports polled "wait" */
753 unsigned pollout : 1;
754 /* op supports buffer selection */
755 unsigned buffer_select : 1;
756 /* needs rlimit(RLIMIT_FSIZE) assigned */
757 unsigned needs_fsize : 1;
758 /* must always have async data allocated */
759 unsigned needs_async_data : 1;
760 /* needs blkcg context, issues async io potentially */
761 unsigned needs_blkcg : 1;
762 /* size of async data needed, if any */
763 unsigned short async_size;
766 static const struct io_op_def io_op_defs[] __read_mostly = {
767 [IORING_OP_NOP] = {},
768 [IORING_OP_READV] = {
771 .unbound_nonreg_file = 1,
774 .needs_async_data = 1,
776 .async_size = sizeof(struct io_async_rw),
778 [IORING_OP_WRITEV] = {
782 .unbound_nonreg_file = 1,
785 .needs_async_data = 1,
787 .async_size = sizeof(struct io_async_rw),
789 [IORING_OP_FSYNC] = {
793 [IORING_OP_READ_FIXED] = {
795 .unbound_nonreg_file = 1,
798 .async_size = sizeof(struct io_async_rw),
800 [IORING_OP_WRITE_FIXED] = {
803 .unbound_nonreg_file = 1,
807 .async_size = sizeof(struct io_async_rw),
809 [IORING_OP_POLL_ADD] = {
811 .unbound_nonreg_file = 1,
813 [IORING_OP_POLL_REMOVE] = {},
814 [IORING_OP_SYNC_FILE_RANGE] = {
818 [IORING_OP_SENDMSG] = {
821 .unbound_nonreg_file = 1,
824 .needs_async_data = 1,
826 .async_size = sizeof(struct io_async_msghdr),
828 [IORING_OP_RECVMSG] = {
831 .unbound_nonreg_file = 1,
835 .needs_async_data = 1,
837 .async_size = sizeof(struct io_async_msghdr),
839 [IORING_OP_TIMEOUT] = {
841 .needs_async_data = 1,
842 .async_size = sizeof(struct io_timeout_data),
844 [IORING_OP_TIMEOUT_REMOVE] = {},
845 [IORING_OP_ACCEPT] = {
848 .unbound_nonreg_file = 1,
852 [IORING_OP_ASYNC_CANCEL] = {},
853 [IORING_OP_LINK_TIMEOUT] = {
855 .needs_async_data = 1,
856 .async_size = sizeof(struct io_timeout_data),
858 [IORING_OP_CONNECT] = {
861 .unbound_nonreg_file = 1,
863 .needs_async_data = 1,
864 .async_size = sizeof(struct io_async_connect),
866 [IORING_OP_FALLOCATE] = {
871 [IORING_OP_OPENAT] = {
876 [IORING_OP_CLOSE] = {
878 .needs_file_no_error = 1,
882 [IORING_OP_FILES_UPDATE] = {
886 [IORING_OP_STATX] = {
895 .unbound_nonreg_file = 1,
899 .async_size = sizeof(struct io_async_rw),
901 [IORING_OP_WRITE] = {
904 .unbound_nonreg_file = 1,
908 .async_size = sizeof(struct io_async_rw),
910 [IORING_OP_FADVISE] = {
914 [IORING_OP_MADVISE] = {
921 .unbound_nonreg_file = 1,
928 .unbound_nonreg_file = 1,
933 [IORING_OP_OPENAT2] = {
938 [IORING_OP_EPOLL_CTL] = {
939 .unbound_nonreg_file = 1,
942 [IORING_OP_SPLICE] = {
945 .unbound_nonreg_file = 1,
948 [IORING_OP_PROVIDE_BUFFERS] = {},
949 [IORING_OP_REMOVE_BUFFERS] = {},
953 .unbound_nonreg_file = 1,
957 enum io_mem_account {
962 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
963 struct io_comp_state *cs);
964 static void io_cqring_fill_event(struct io_kiocb *req, long res);
965 static void io_put_req(struct io_kiocb *req);
966 static void io_double_put_req(struct io_kiocb *req);
967 static void __io_double_put_req(struct io_kiocb *req);
968 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
969 static void __io_queue_linked_timeout(struct io_kiocb *req);
970 static void io_queue_linked_timeout(struct io_kiocb *req);
971 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
972 struct io_uring_files_update *ip,
974 static void __io_clean_op(struct io_kiocb *req);
975 static struct file *io_file_get(struct io_submit_state *state,
976 struct io_kiocb *req, int fd, bool fixed);
977 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
978 static void io_file_put_work(struct work_struct *work);
980 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
981 struct iovec **iovec, struct iov_iter *iter,
983 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
984 const struct iovec *fast_iov,
985 struct iov_iter *iter, bool force);
987 static struct kmem_cache *req_cachep;
989 static const struct file_operations io_uring_fops __read_mostly;
991 struct sock *io_uring_get_socket(struct file *file)
993 #if defined(CONFIG_UNIX)
994 if (file->f_op == &io_uring_fops) {
995 struct io_ring_ctx *ctx = file->private_data;
997 return ctx->ring_sock->sk;
1002 EXPORT_SYMBOL(io_uring_get_socket);
1004 static inline void io_clean_op(struct io_kiocb *req)
1006 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
1011 static void io_sq_thread_drop_mm(void)
1013 struct mm_struct *mm = current->mm;
1016 kthread_unuse_mm(mm);
1021 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1024 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
1025 !ctx->sqo_task->mm ||
1026 !mmget_not_zero(ctx->sqo_task->mm)))
1028 kthread_use_mm(ctx->sqo_task->mm);
1034 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1035 struct io_kiocb *req)
1037 if (!io_op_defs[req->opcode].needs_mm)
1039 return __io_sq_thread_acquire_mm(ctx);
1042 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1043 struct cgroup_subsys_state **cur_css)
1046 #ifdef CONFIG_BLK_CGROUP
1047 /* puts the old one when swapping */
1048 if (*cur_css != ctx->sqo_blkcg_css) {
1049 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1050 *cur_css = ctx->sqo_blkcg_css;
1055 static void io_sq_thread_unassociate_blkcg(void)
1057 #ifdef CONFIG_BLK_CGROUP
1058 kthread_associate_blkcg(NULL);
1062 static inline void req_set_fail_links(struct io_kiocb *req)
1064 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1065 req->flags |= REQ_F_FAIL_LINK;
1069 * Note: must call io_req_init_async() for the first time you
1070 * touch any members of io_wq_work.
1072 static inline void io_req_init_async(struct io_kiocb *req)
1074 if (req->flags & REQ_F_WORK_INITIALIZED)
1077 memset(&req->work, 0, sizeof(req->work));
1078 req->flags |= REQ_F_WORK_INITIALIZED;
1081 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1083 return ctx->flags & IORING_SETUP_SQPOLL;
1086 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1088 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1090 complete(&ctx->ref_comp);
1093 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1095 return !req->timeout.off;
1098 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1100 struct io_ring_ctx *ctx;
1103 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1107 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1108 if (!ctx->fallback_req)
1112 * Use 5 bits less than the max cq entries, that should give us around
1113 * 32 entries per hash list if totally full and uniformly spread.
1115 hash_bits = ilog2(p->cq_entries);
1119 ctx->cancel_hash_bits = hash_bits;
1120 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1122 if (!ctx->cancel_hash)
1124 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1126 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1127 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1130 ctx->flags = p->flags;
1131 init_waitqueue_head(&ctx->sqo_sq_wait);
1132 INIT_LIST_HEAD(&ctx->sqd_list);
1133 init_waitqueue_head(&ctx->cq_wait);
1134 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1135 init_completion(&ctx->ref_comp);
1136 init_completion(&ctx->sq_thread_comp);
1137 idr_init(&ctx->io_buffer_idr);
1138 idr_init(&ctx->personality_idr);
1139 mutex_init(&ctx->uring_lock);
1140 init_waitqueue_head(&ctx->wait);
1141 spin_lock_init(&ctx->completion_lock);
1142 INIT_LIST_HEAD(&ctx->iopoll_list);
1143 INIT_LIST_HEAD(&ctx->defer_list);
1144 INIT_LIST_HEAD(&ctx->timeout_list);
1145 init_waitqueue_head(&ctx->inflight_wait);
1146 spin_lock_init(&ctx->inflight_lock);
1147 INIT_LIST_HEAD(&ctx->inflight_list);
1148 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1149 init_llist_head(&ctx->file_put_llist);
1152 if (ctx->fallback_req)
1153 kmem_cache_free(req_cachep, ctx->fallback_req);
1154 kfree(ctx->cancel_hash);
1159 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1161 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1162 struct io_ring_ctx *ctx = req->ctx;
1164 return seq != ctx->cached_cq_tail
1165 + atomic_read(&ctx->cached_cq_overflow);
1171 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1173 struct io_rings *rings = ctx->rings;
1175 /* order cqe stores with ring update */
1176 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1178 if (wq_has_sleeper(&ctx->cq_wait)) {
1179 wake_up_interruptible(&ctx->cq_wait);
1180 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1185 * Returns true if we need to defer file table putting. This can only happen
1186 * from the error path with REQ_F_COMP_LOCKED set.
1188 static bool io_req_clean_work(struct io_kiocb *req)
1190 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1193 req->flags &= ~REQ_F_WORK_INITIALIZED;
1196 mmdrop(req->work.mm);
1197 req->work.mm = NULL;
1199 #ifdef CONFIG_BLK_CGROUP
1200 if (req->work.blkcg_css)
1201 css_put(req->work.blkcg_css);
1203 if (req->work.creds) {
1204 put_cred(req->work.creds);
1205 req->work.creds = NULL;
1208 struct fs_struct *fs = req->work.fs;
1210 if (req->flags & REQ_F_COMP_LOCKED)
1213 spin_lock(&req->work.fs->lock);
1216 spin_unlock(&req->work.fs->lock);
1219 req->work.fs = NULL;
1225 static void io_prep_async_work(struct io_kiocb *req)
1227 const struct io_op_def *def = &io_op_defs[req->opcode];
1228 struct io_ring_ctx *ctx = req->ctx;
1230 io_req_init_async(req);
1232 if (req->flags & REQ_F_ISREG) {
1233 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1234 io_wq_hash_work(&req->work, file_inode(req->file));
1236 if (def->unbound_nonreg_file)
1237 req->work.flags |= IO_WQ_WORK_UNBOUND;
1239 if (!req->work.files && io_op_defs[req->opcode].file_table &&
1240 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1241 req->work.files = get_files_struct(current);
1242 get_nsproxy(current->nsproxy);
1243 req->work.nsproxy = current->nsproxy;
1244 req->flags |= REQ_F_INFLIGHT;
1246 spin_lock_irq(&ctx->inflight_lock);
1247 list_add(&req->inflight_entry, &ctx->inflight_list);
1248 spin_unlock_irq(&ctx->inflight_lock);
1250 if (!req->work.mm && def->needs_mm) {
1251 mmgrab(current->mm);
1252 req->work.mm = current->mm;
1254 #ifdef CONFIG_BLK_CGROUP
1255 if (!req->work.blkcg_css && def->needs_blkcg) {
1257 req->work.blkcg_css = blkcg_css();
1259 * This should be rare, either the cgroup is dying or the task
1260 * is moving cgroups. Just punt to root for the handful of ios.
1262 if (!css_tryget_online(req->work.blkcg_css))
1263 req->work.blkcg_css = NULL;
1267 if (!req->work.creds)
1268 req->work.creds = get_current_cred();
1269 if (!req->work.fs && def->needs_fs) {
1270 spin_lock(¤t->fs->lock);
1271 if (!current->fs->in_exec) {
1272 req->work.fs = current->fs;
1273 req->work.fs->users++;
1275 req->work.flags |= IO_WQ_WORK_CANCEL;
1277 spin_unlock(¤t->fs->lock);
1279 if (def->needs_fsize)
1280 req->work.fsize = rlimit(RLIMIT_FSIZE);
1282 req->work.fsize = RLIM_INFINITY;
1285 static void io_prep_async_link(struct io_kiocb *req)
1287 struct io_kiocb *cur;
1289 io_prep_async_work(req);
1290 if (req->flags & REQ_F_LINK_HEAD)
1291 list_for_each_entry(cur, &req->link_list, link_list)
1292 io_prep_async_work(cur);
1295 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1297 struct io_ring_ctx *ctx = req->ctx;
1298 struct io_kiocb *link = io_prep_linked_timeout(req);
1300 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1301 &req->work, req->flags);
1302 io_wq_enqueue(ctx->io_wq, &req->work);
1306 static void io_queue_async_work(struct io_kiocb *req)
1308 struct io_kiocb *link;
1310 /* init ->work of the whole link before punting */
1311 io_prep_async_link(req);
1312 link = __io_queue_async_work(req);
1315 io_queue_linked_timeout(link);
1318 static void io_kill_timeout(struct io_kiocb *req)
1320 struct io_timeout_data *io = req->async_data;
1323 ret = hrtimer_try_to_cancel(&io->timer);
1325 atomic_set(&req->ctx->cq_timeouts,
1326 atomic_read(&req->ctx->cq_timeouts) + 1);
1327 list_del_init(&req->timeout.list);
1328 req->flags |= REQ_F_COMP_LOCKED;
1329 io_cqring_fill_event(req, 0);
1334 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1336 struct io_ring_ctx *ctx = req->ctx;
1338 if (!tsk || req->task == tsk)
1340 if (ctx->flags & IORING_SETUP_SQPOLL) {
1341 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1348 * Returns true if we found and killed one or more timeouts
1350 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1352 struct io_kiocb *req, *tmp;
1355 spin_lock_irq(&ctx->completion_lock);
1356 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1357 if (io_task_match(req, tsk)) {
1358 io_kill_timeout(req);
1362 spin_unlock_irq(&ctx->completion_lock);
1363 return canceled != 0;
1366 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1369 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1370 struct io_defer_entry, list);
1371 struct io_kiocb *link;
1373 if (req_need_defer(de->req, de->seq))
1375 list_del_init(&de->list);
1376 /* punt-init is done before queueing for defer */
1377 link = __io_queue_async_work(de->req);
1379 __io_queue_linked_timeout(link);
1380 /* drop submission reference */
1381 link->flags |= REQ_F_COMP_LOCKED;
1385 } while (!list_empty(&ctx->defer_list));
1388 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1390 while (!list_empty(&ctx->timeout_list)) {
1391 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1392 struct io_kiocb, timeout.list);
1394 if (io_is_timeout_noseq(req))
1396 if (req->timeout.target_seq != ctx->cached_cq_tail
1397 - atomic_read(&ctx->cq_timeouts))
1400 list_del_init(&req->timeout.list);
1401 io_kill_timeout(req);
1405 static void io_commit_cqring(struct io_ring_ctx *ctx)
1407 io_flush_timeouts(ctx);
1408 __io_commit_cqring(ctx);
1410 if (unlikely(!list_empty(&ctx->defer_list)))
1411 __io_queue_deferred(ctx);
1414 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1416 struct io_rings *r = ctx->rings;
1418 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1421 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1423 struct io_rings *rings = ctx->rings;
1426 tail = ctx->cached_cq_tail;
1428 * writes to the cq entry need to come after reading head; the
1429 * control dependency is enough as we're using WRITE_ONCE to
1432 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1435 ctx->cached_cq_tail++;
1436 return &rings->cqes[tail & ctx->cq_mask];
1439 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1443 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1445 if (!ctx->eventfd_async)
1447 return io_wq_current_is_worker();
1450 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1452 if (waitqueue_active(&ctx->wait))
1453 wake_up(&ctx->wait);
1454 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1455 wake_up(&ctx->sq_data->wait);
1456 if (io_should_trigger_evfd(ctx))
1457 eventfd_signal(ctx->cq_ev_fd, 1);
1460 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1462 if (list_empty(&ctx->cq_overflow_list)) {
1463 clear_bit(0, &ctx->sq_check_overflow);
1464 clear_bit(0, &ctx->cq_check_overflow);
1465 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1469 static inline bool io_match_files(struct io_kiocb *req,
1470 struct files_struct *files)
1474 if (req->flags & REQ_F_WORK_INITIALIZED)
1475 return req->work.files == files;
1479 /* Returns true if there are no backlogged entries after the flush */
1480 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1481 struct task_struct *tsk,
1482 struct files_struct *files)
1484 struct io_rings *rings = ctx->rings;
1485 struct io_kiocb *req, *tmp;
1486 struct io_uring_cqe *cqe;
1487 unsigned long flags;
1491 if (list_empty_careful(&ctx->cq_overflow_list))
1493 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1494 rings->cq_ring_entries))
1498 spin_lock_irqsave(&ctx->completion_lock, flags);
1500 /* if force is set, the ring is going away. always drop after that */
1502 ctx->cq_overflow_flushed = 1;
1505 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1506 if (tsk && req->task != tsk)
1508 if (!io_match_files(req, files))
1511 cqe = io_get_cqring(ctx);
1515 list_move(&req->compl.list, &list);
1517 WRITE_ONCE(cqe->user_data, req->user_data);
1518 WRITE_ONCE(cqe->res, req->result);
1519 WRITE_ONCE(cqe->flags, req->compl.cflags);
1521 WRITE_ONCE(ctx->rings->cq_overflow,
1522 atomic_inc_return(&ctx->cached_cq_overflow));
1526 io_commit_cqring(ctx);
1527 io_cqring_mark_overflow(ctx);
1529 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1530 io_cqring_ev_posted(ctx);
1532 while (!list_empty(&list)) {
1533 req = list_first_entry(&list, struct io_kiocb, compl.list);
1534 list_del(&req->compl.list);
1541 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1543 struct io_ring_ctx *ctx = req->ctx;
1544 struct io_uring_cqe *cqe;
1546 trace_io_uring_complete(ctx, req->user_data, res);
1549 * If we can't get a cq entry, userspace overflowed the
1550 * submission (by quite a lot). Increment the overflow count in
1553 cqe = io_get_cqring(ctx);
1555 WRITE_ONCE(cqe->user_data, req->user_data);
1556 WRITE_ONCE(cqe->res, res);
1557 WRITE_ONCE(cqe->flags, cflags);
1558 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1560 * If we're in ring overflow flush mode, or in task cancel mode,
1561 * then we cannot store the request for later flushing, we need
1562 * to drop it on the floor.
1564 WRITE_ONCE(ctx->rings->cq_overflow,
1565 atomic_inc_return(&ctx->cached_cq_overflow));
1567 if (list_empty(&ctx->cq_overflow_list)) {
1568 set_bit(0, &ctx->sq_check_overflow);
1569 set_bit(0, &ctx->cq_check_overflow);
1570 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1574 req->compl.cflags = cflags;
1575 refcount_inc(&req->refs);
1576 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1580 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1582 __io_cqring_fill_event(req, res, 0);
1585 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1587 struct io_ring_ctx *ctx = req->ctx;
1588 unsigned long flags;
1590 spin_lock_irqsave(&ctx->completion_lock, flags);
1591 __io_cqring_fill_event(req, res, cflags);
1592 io_commit_cqring(ctx);
1593 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1595 io_cqring_ev_posted(ctx);
1598 static void io_submit_flush_completions(struct io_comp_state *cs)
1600 struct io_ring_ctx *ctx = cs->ctx;
1602 spin_lock_irq(&ctx->completion_lock);
1603 while (!list_empty(&cs->list)) {
1604 struct io_kiocb *req;
1606 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1607 list_del(&req->compl.list);
1608 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1609 if (!(req->flags & REQ_F_LINK_HEAD)) {
1610 req->flags |= REQ_F_COMP_LOCKED;
1613 spin_unlock_irq(&ctx->completion_lock);
1615 spin_lock_irq(&ctx->completion_lock);
1618 io_commit_cqring(ctx);
1619 spin_unlock_irq(&ctx->completion_lock);
1621 io_cqring_ev_posted(ctx);
1625 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1626 struct io_comp_state *cs)
1629 io_cqring_add_event(req, res, cflags);
1634 req->compl.cflags = cflags;
1635 list_add_tail(&req->compl.list, &cs->list);
1637 io_submit_flush_completions(cs);
1641 static void io_req_complete(struct io_kiocb *req, long res)
1643 __io_req_complete(req, res, 0, NULL);
1646 static inline bool io_is_fallback_req(struct io_kiocb *req)
1648 return req == (struct io_kiocb *)
1649 ((unsigned long) req->ctx->fallback_req & ~1UL);
1652 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1654 struct io_kiocb *req;
1656 req = ctx->fallback_req;
1657 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1663 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1664 struct io_submit_state *state)
1666 if (!state->free_reqs) {
1667 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1671 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1672 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1675 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1676 * retry single alloc to be on the safe side.
1678 if (unlikely(ret <= 0)) {
1679 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1680 if (!state->reqs[0])
1684 state->free_reqs = ret;
1688 return state->reqs[state->free_reqs];
1690 return io_get_fallback_req(ctx);
1693 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1697 percpu_ref_put(req->fixed_file_refs);
1702 static bool io_dismantle_req(struct io_kiocb *req)
1706 if (req->async_data)
1707 kfree(req->async_data);
1709 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1711 return io_req_clean_work(req);
1714 static void __io_free_req_finish(struct io_kiocb *req)
1716 struct io_uring_task *tctx = req->task->io_uring;
1717 struct io_ring_ctx *ctx = req->ctx;
1719 atomic_long_inc(&tctx->req_complete);
1721 wake_up(&tctx->wait);
1722 put_task_struct(req->task);
1724 if (likely(!io_is_fallback_req(req)))
1725 kmem_cache_free(req_cachep, req);
1727 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1728 percpu_ref_put(&ctx->refs);
1731 static void io_req_task_file_table_put(struct callback_head *cb)
1733 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1734 struct fs_struct *fs = req->work.fs;
1736 spin_lock(&req->work.fs->lock);
1739 spin_unlock(&req->work.fs->lock);
1742 req->work.fs = NULL;
1743 __io_free_req_finish(req);
1746 static void __io_free_req(struct io_kiocb *req)
1748 if (!io_dismantle_req(req)) {
1749 __io_free_req_finish(req);
1753 init_task_work(&req->task_work, io_req_task_file_table_put);
1754 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1755 if (unlikely(ret)) {
1756 struct task_struct *tsk;
1758 tsk = io_wq_get_task(req->ctx->io_wq);
1759 task_work_add(tsk, &req->task_work, 0);
1764 static bool io_link_cancel_timeout(struct io_kiocb *req)
1766 struct io_timeout_data *io = req->async_data;
1767 struct io_ring_ctx *ctx = req->ctx;
1770 ret = hrtimer_try_to_cancel(&io->timer);
1772 req->flags |= REQ_F_COMP_LOCKED;
1773 io_cqring_fill_event(req, -ECANCELED);
1774 io_commit_cqring(ctx);
1775 req->flags &= ~REQ_F_LINK_HEAD;
1783 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1785 struct io_kiocb *link;
1788 if (list_empty(&req->link_list))
1790 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1791 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1794 list_del_init(&link->link_list);
1795 wake_ev = io_link_cancel_timeout(link);
1796 req->flags &= ~REQ_F_LINK_TIMEOUT;
1800 static void io_kill_linked_timeout(struct io_kiocb *req)
1802 struct io_ring_ctx *ctx = req->ctx;
1805 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1806 unsigned long flags;
1808 spin_lock_irqsave(&ctx->completion_lock, flags);
1809 wake_ev = __io_kill_linked_timeout(req);
1810 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1812 wake_ev = __io_kill_linked_timeout(req);
1816 io_cqring_ev_posted(ctx);
1819 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1821 struct io_kiocb *nxt;
1824 * The list should never be empty when we are called here. But could
1825 * potentially happen if the chain is messed up, check to be on the
1828 if (unlikely(list_empty(&req->link_list)))
1831 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1832 list_del_init(&req->link_list);
1833 if (!list_empty(&nxt->link_list))
1834 nxt->flags |= REQ_F_LINK_HEAD;
1839 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1841 static void __io_fail_links(struct io_kiocb *req)
1843 struct io_ring_ctx *ctx = req->ctx;
1845 while (!list_empty(&req->link_list)) {
1846 struct io_kiocb *link = list_first_entry(&req->link_list,
1847 struct io_kiocb, link_list);
1849 list_del_init(&link->link_list);
1850 trace_io_uring_fail_link(req, link);
1852 io_cqring_fill_event(link, -ECANCELED);
1853 link->flags |= REQ_F_COMP_LOCKED;
1854 __io_double_put_req(link);
1857 io_commit_cqring(ctx);
1858 io_cqring_ev_posted(ctx);
1861 static void io_fail_links(struct io_kiocb *req)
1863 struct io_ring_ctx *ctx = req->ctx;
1865 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1866 unsigned long flags;
1868 spin_lock_irqsave(&ctx->completion_lock, flags);
1869 __io_fail_links(req);
1870 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1872 __io_fail_links(req);
1875 io_cqring_ev_posted(ctx);
1878 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1880 req->flags &= ~REQ_F_LINK_HEAD;
1881 if (req->flags & REQ_F_LINK_TIMEOUT)
1882 io_kill_linked_timeout(req);
1885 * If LINK is set, we have dependent requests in this chain. If we
1886 * didn't fail this request, queue the first one up, moving any other
1887 * dependencies to the next request. In case of failure, fail the rest
1890 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1891 return io_req_link_next(req);
1896 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1898 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1900 return __io_req_find_next(req);
1903 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1905 struct task_struct *tsk = req->task;
1906 struct io_ring_ctx *ctx = req->ctx;
1909 if (tsk->flags & PF_EXITING)
1913 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1914 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1915 * processing task_work. There's no reliable way to tell if TWA_RESUME
1919 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1920 notify = TWA_SIGNAL;
1922 ret = task_work_add(tsk, &req->task_work, notify);
1924 wake_up_process(tsk);
1929 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1931 struct io_ring_ctx *ctx = req->ctx;
1933 spin_lock_irq(&ctx->completion_lock);
1934 io_cqring_fill_event(req, error);
1935 io_commit_cqring(ctx);
1936 spin_unlock_irq(&ctx->completion_lock);
1938 io_cqring_ev_posted(ctx);
1939 req_set_fail_links(req);
1940 io_double_put_req(req);
1943 static void io_req_task_cancel(struct callback_head *cb)
1945 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1946 struct io_ring_ctx *ctx = req->ctx;
1948 __io_req_task_cancel(req, -ECANCELED);
1949 percpu_ref_put(&ctx->refs);
1952 static void __io_req_task_submit(struct io_kiocb *req)
1954 struct io_ring_ctx *ctx = req->ctx;
1956 if (!__io_sq_thread_acquire_mm(ctx)) {
1957 mutex_lock(&ctx->uring_lock);
1958 __io_queue_sqe(req, NULL);
1959 mutex_unlock(&ctx->uring_lock);
1961 __io_req_task_cancel(req, -EFAULT);
1965 static void io_req_task_submit(struct callback_head *cb)
1967 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1968 struct io_ring_ctx *ctx = req->ctx;
1970 __io_req_task_submit(req);
1971 percpu_ref_put(&ctx->refs);
1974 static void io_req_task_queue(struct io_kiocb *req)
1978 init_task_work(&req->task_work, io_req_task_submit);
1979 percpu_ref_get(&req->ctx->refs);
1981 ret = io_req_task_work_add(req, true);
1982 if (unlikely(ret)) {
1983 struct task_struct *tsk;
1985 init_task_work(&req->task_work, io_req_task_cancel);
1986 tsk = io_wq_get_task(req->ctx->io_wq);
1987 task_work_add(tsk, &req->task_work, 0);
1988 wake_up_process(tsk);
1992 static void io_queue_next(struct io_kiocb *req)
1994 struct io_kiocb *nxt = io_req_find_next(req);
1997 io_req_task_queue(nxt);
2000 static void io_free_req(struct io_kiocb *req)
2007 void *reqs[IO_IOPOLL_BATCH];
2010 struct task_struct *task;
2014 static inline void io_init_req_batch(struct req_batch *rb)
2021 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2022 struct req_batch *rb)
2024 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2025 percpu_ref_put_many(&ctx->refs, rb->to_free);
2029 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2030 struct req_batch *rb)
2033 __io_req_free_batch_flush(ctx, rb);
2035 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2036 put_task_struct_many(rb->task, rb->task_refs);
2041 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2043 if (unlikely(io_is_fallback_req(req))) {
2047 if (req->flags & REQ_F_LINK_HEAD)
2050 if (req->task != rb->task) {
2052 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2053 put_task_struct_many(rb->task, rb->task_refs);
2055 rb->task = req->task;
2060 WARN_ON_ONCE(io_dismantle_req(req));
2061 rb->reqs[rb->to_free++] = req;
2062 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2063 __io_req_free_batch_flush(req->ctx, rb);
2067 * Drop reference to request, return next in chain (if there is one) if this
2068 * was the last reference to this request.
2070 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2072 struct io_kiocb *nxt = NULL;
2074 if (refcount_dec_and_test(&req->refs)) {
2075 nxt = io_req_find_next(req);
2081 static void io_put_req(struct io_kiocb *req)
2083 if (refcount_dec_and_test(&req->refs))
2087 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2089 struct io_kiocb *nxt;
2092 * A ref is owned by io-wq in which context we're. So, if that's the
2093 * last one, it's safe to steal next work. False negatives are Ok,
2094 * it just will be re-punted async in io_put_work()
2096 if (refcount_read(&req->refs) != 1)
2099 nxt = io_req_find_next(req);
2100 return nxt ? &nxt->work : NULL;
2104 * Must only be used if we don't need to care about links, usually from
2105 * within the completion handling itself.
2107 static void __io_double_put_req(struct io_kiocb *req)
2109 /* drop both submit and complete references */
2110 if (refcount_sub_and_test(2, &req->refs))
2114 static void io_double_put_req(struct io_kiocb *req)
2116 /* drop both submit and complete references */
2117 if (refcount_sub_and_test(2, &req->refs))
2121 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2123 struct io_rings *rings = ctx->rings;
2125 if (test_bit(0, &ctx->cq_check_overflow)) {
2127 * noflush == true is from the waitqueue handler, just ensure
2128 * we wake up the task, and the next invocation will flush the
2129 * entries. We cannot safely to it from here.
2131 if (noflush && !list_empty(&ctx->cq_overflow_list))
2134 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2137 /* See comment at the top of this file */
2139 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2142 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2144 struct io_rings *rings = ctx->rings;
2146 /* make sure SQ entry isn't read before tail */
2147 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2150 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2152 unsigned int cflags;
2154 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2155 cflags |= IORING_CQE_F_BUFFER;
2156 req->flags &= ~REQ_F_BUFFER_SELECTED;
2161 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2163 struct io_buffer *kbuf;
2165 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2166 return io_put_kbuf(req, kbuf);
2169 static inline bool io_run_task_work(void)
2172 * Not safe to run on exiting task, and the task_work handling will
2173 * not add work to such a task.
2175 if (unlikely(current->flags & PF_EXITING))
2177 if (current->task_works) {
2178 __set_current_state(TASK_RUNNING);
2186 static void io_iopoll_queue(struct list_head *again)
2188 struct io_kiocb *req;
2191 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2192 list_del(&req->inflight_entry);
2193 __io_complete_rw(req, -EAGAIN, 0, NULL);
2194 } while (!list_empty(again));
2198 * Find and free completed poll iocbs
2200 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2201 struct list_head *done)
2203 struct req_batch rb;
2204 struct io_kiocb *req;
2207 /* order with ->result store in io_complete_rw_iopoll() */
2210 io_init_req_batch(&rb);
2211 while (!list_empty(done)) {
2214 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2215 if (READ_ONCE(req->result) == -EAGAIN) {
2217 req->iopoll_completed = 0;
2218 list_move_tail(&req->inflight_entry, &again);
2221 list_del(&req->inflight_entry);
2223 if (req->flags & REQ_F_BUFFER_SELECTED)
2224 cflags = io_put_rw_kbuf(req);
2226 __io_cqring_fill_event(req, req->result, cflags);
2229 if (refcount_dec_and_test(&req->refs))
2230 io_req_free_batch(&rb, req);
2233 io_commit_cqring(ctx);
2234 if (ctx->flags & IORING_SETUP_SQPOLL)
2235 io_cqring_ev_posted(ctx);
2236 io_req_free_batch_finish(ctx, &rb);
2238 if (!list_empty(&again))
2239 io_iopoll_queue(&again);
2242 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2245 struct io_kiocb *req, *tmp;
2251 * Only spin for completions if we don't have multiple devices hanging
2252 * off our complete list, and we're under the requested amount.
2254 spin = !ctx->poll_multi_file && *nr_events < min;
2257 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2258 struct kiocb *kiocb = &req->rw.kiocb;
2261 * Move completed and retryable entries to our local lists.
2262 * If we find a request that requires polling, break out
2263 * and complete those lists first, if we have entries there.
2265 if (READ_ONCE(req->iopoll_completed)) {
2266 list_move_tail(&req->inflight_entry, &done);
2269 if (!list_empty(&done))
2272 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2276 /* iopoll may have completed current req */
2277 if (READ_ONCE(req->iopoll_completed))
2278 list_move_tail(&req->inflight_entry, &done);
2285 if (!list_empty(&done))
2286 io_iopoll_complete(ctx, nr_events, &done);
2292 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2293 * non-spinning poll check - we'll still enter the driver poll loop, but only
2294 * as a non-spinning completion check.
2296 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2299 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2302 ret = io_do_iopoll(ctx, nr_events, min);
2305 if (*nr_events >= min)
2313 * We can't just wait for polled events to come to us, we have to actively
2314 * find and complete them.
2316 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2318 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2321 mutex_lock(&ctx->uring_lock);
2322 while (!list_empty(&ctx->iopoll_list)) {
2323 unsigned int nr_events = 0;
2325 io_do_iopoll(ctx, &nr_events, 0);
2327 /* let it sleep and repeat later if can't complete a request */
2331 * Ensure we allow local-to-the-cpu processing to take place,
2332 * in this case we need to ensure that we reap all events.
2333 * Also let task_work, etc. to progress by releasing the mutex
2335 if (need_resched()) {
2336 mutex_unlock(&ctx->uring_lock);
2338 mutex_lock(&ctx->uring_lock);
2341 mutex_unlock(&ctx->uring_lock);
2344 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2346 unsigned int nr_events = 0;
2347 int iters = 0, ret = 0;
2350 * We disallow the app entering submit/complete with polling, but we
2351 * still need to lock the ring to prevent racing with polled issue
2352 * that got punted to a workqueue.
2354 mutex_lock(&ctx->uring_lock);
2357 * Don't enter poll loop if we already have events pending.
2358 * If we do, we can potentially be spinning for commands that
2359 * already triggered a CQE (eg in error).
2361 if (io_cqring_events(ctx, false))
2365 * If a submit got punted to a workqueue, we can have the
2366 * application entering polling for a command before it gets
2367 * issued. That app will hold the uring_lock for the duration
2368 * of the poll right here, so we need to take a breather every
2369 * now and then to ensure that the issue has a chance to add
2370 * the poll to the issued list. Otherwise we can spin here
2371 * forever, while the workqueue is stuck trying to acquire the
2374 if (!(++iters & 7)) {
2375 mutex_unlock(&ctx->uring_lock);
2377 mutex_lock(&ctx->uring_lock);
2380 ret = io_iopoll_getevents(ctx, &nr_events, min);
2384 } while (min && !nr_events && !need_resched());
2386 mutex_unlock(&ctx->uring_lock);
2390 static void kiocb_end_write(struct io_kiocb *req)
2393 * Tell lockdep we inherited freeze protection from submission
2396 if (req->flags & REQ_F_ISREG) {
2397 struct inode *inode = file_inode(req->file);
2399 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2401 file_end_write(req->file);
2404 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2405 struct io_comp_state *cs)
2407 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2410 if (kiocb->ki_flags & IOCB_WRITE)
2411 kiocb_end_write(req);
2413 if (res != req->result)
2414 req_set_fail_links(req);
2415 if (req->flags & REQ_F_BUFFER_SELECTED)
2416 cflags = io_put_rw_kbuf(req);
2417 __io_req_complete(req, res, cflags, cs);
2421 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2423 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2424 ssize_t ret = -ECANCELED;
2425 struct iov_iter iter;
2433 switch (req->opcode) {
2434 case IORING_OP_READV:
2435 case IORING_OP_READ_FIXED:
2436 case IORING_OP_READ:
2439 case IORING_OP_WRITEV:
2440 case IORING_OP_WRITE_FIXED:
2441 case IORING_OP_WRITE:
2445 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2450 if (!req->async_data) {
2451 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2454 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2462 req_set_fail_links(req);
2463 io_req_complete(req, ret);
2468 static bool io_rw_reissue(struct io_kiocb *req, long res)
2471 umode_t mode = file_inode(req->file)->i_mode;
2474 if (!S_ISBLK(mode) && !S_ISREG(mode))
2476 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2479 ret = io_sq_thread_acquire_mm(req->ctx, req);
2481 if (io_resubmit_prep(req, ret)) {
2482 refcount_inc(&req->refs);
2483 io_queue_async_work(req);
2491 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2492 struct io_comp_state *cs)
2494 if (!io_rw_reissue(req, res))
2495 io_complete_rw_common(&req->rw.kiocb, res, cs);
2498 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2500 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2502 __io_complete_rw(req, res, res2, NULL);
2505 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2507 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2509 if (kiocb->ki_flags & IOCB_WRITE)
2510 kiocb_end_write(req);
2512 if (res != -EAGAIN && res != req->result)
2513 req_set_fail_links(req);
2515 WRITE_ONCE(req->result, res);
2516 /* order with io_poll_complete() checking ->result */
2518 WRITE_ONCE(req->iopoll_completed, 1);
2522 * After the iocb has been issued, it's safe to be found on the poll list.
2523 * Adding the kiocb to the list AFTER submission ensures that we don't
2524 * find it from a io_iopoll_getevents() thread before the issuer is done
2525 * accessing the kiocb cookie.
2527 static void io_iopoll_req_issued(struct io_kiocb *req)
2529 struct io_ring_ctx *ctx = req->ctx;
2532 * Track whether we have multiple files in our lists. This will impact
2533 * how we do polling eventually, not spinning if we're on potentially
2534 * different devices.
2536 if (list_empty(&ctx->iopoll_list)) {
2537 ctx->poll_multi_file = false;
2538 } else if (!ctx->poll_multi_file) {
2539 struct io_kiocb *list_req;
2541 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2543 if (list_req->file != req->file)
2544 ctx->poll_multi_file = true;
2548 * For fast devices, IO may have already completed. If it has, add
2549 * it to the front so we find it first.
2551 if (READ_ONCE(req->iopoll_completed))
2552 list_add(&req->inflight_entry, &ctx->iopoll_list);
2554 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2556 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2557 wq_has_sleeper(&ctx->sq_data->wait))
2558 wake_up(&ctx->sq_data->wait);
2561 static void __io_state_file_put(struct io_submit_state *state)
2563 if (state->has_refs)
2564 fput_many(state->file, state->has_refs);
2568 static inline void io_state_file_put(struct io_submit_state *state)
2571 __io_state_file_put(state);
2575 * Get as many references to a file as we have IOs left in this submission,
2576 * assuming most submissions are for one file, or at least that each file
2577 * has more than one submission.
2579 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2585 if (state->fd == fd) {
2589 __io_state_file_put(state);
2591 state->file = fget_many(fd, state->ios_left);
2596 state->has_refs = state->ios_left - 1;
2600 static bool io_bdev_nowait(struct block_device *bdev)
2603 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2610 * If we tracked the file through the SCM inflight mechanism, we could support
2611 * any file. For now, just ensure that anything potentially problematic is done
2614 static bool io_file_supports_async(struct file *file, int rw)
2616 umode_t mode = file_inode(file)->i_mode;
2618 if (S_ISBLK(mode)) {
2619 if (io_bdev_nowait(file->f_inode->i_bdev))
2623 if (S_ISCHR(mode) || S_ISSOCK(mode))
2625 if (S_ISREG(mode)) {
2626 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2627 file->f_op != &io_uring_fops)
2632 /* any ->read/write should understand O_NONBLOCK */
2633 if (file->f_flags & O_NONBLOCK)
2636 if (!(file->f_mode & FMODE_NOWAIT))
2640 return file->f_op->read_iter != NULL;
2642 return file->f_op->write_iter != NULL;
2645 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2647 struct io_ring_ctx *ctx = req->ctx;
2648 struct kiocb *kiocb = &req->rw.kiocb;
2652 if (S_ISREG(file_inode(req->file)->i_mode))
2653 req->flags |= REQ_F_ISREG;
2655 kiocb->ki_pos = READ_ONCE(sqe->off);
2656 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2657 req->flags |= REQ_F_CUR_POS;
2658 kiocb->ki_pos = req->file->f_pos;
2660 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2661 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2662 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2666 ioprio = READ_ONCE(sqe->ioprio);
2668 ret = ioprio_check_cap(ioprio);
2672 kiocb->ki_ioprio = ioprio;
2674 kiocb->ki_ioprio = get_current_ioprio();
2676 /* don't allow async punt if RWF_NOWAIT was requested */
2677 if (kiocb->ki_flags & IOCB_NOWAIT)
2678 req->flags |= REQ_F_NOWAIT;
2680 if (ctx->flags & IORING_SETUP_IOPOLL) {
2681 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2682 !kiocb->ki_filp->f_op->iopoll)
2685 kiocb->ki_flags |= IOCB_HIPRI;
2686 kiocb->ki_complete = io_complete_rw_iopoll;
2687 req->iopoll_completed = 0;
2689 if (kiocb->ki_flags & IOCB_HIPRI)
2691 kiocb->ki_complete = io_complete_rw;
2694 req->rw.addr = READ_ONCE(sqe->addr);
2695 req->rw.len = READ_ONCE(sqe->len);
2696 req->buf_index = READ_ONCE(sqe->buf_index);
2700 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2706 case -ERESTARTNOINTR:
2707 case -ERESTARTNOHAND:
2708 case -ERESTART_RESTARTBLOCK:
2710 * We can't just restart the syscall, since previously
2711 * submitted sqes may already be in progress. Just fail this
2717 kiocb->ki_complete(kiocb, ret, 0);
2721 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2722 struct io_comp_state *cs)
2724 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2725 struct io_async_rw *io = req->async_data;
2727 /* add previously done IO, if any */
2728 if (io && io->bytes_done > 0) {
2730 ret = io->bytes_done;
2732 ret += io->bytes_done;
2735 if (req->flags & REQ_F_CUR_POS)
2736 req->file->f_pos = kiocb->ki_pos;
2737 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2738 __io_complete_rw(req, ret, 0, cs);
2740 io_rw_done(kiocb, ret);
2743 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2744 struct iov_iter *iter)
2746 struct io_ring_ctx *ctx = req->ctx;
2747 size_t len = req->rw.len;
2748 struct io_mapped_ubuf *imu;
2749 u16 index, buf_index = req->buf_index;
2753 if (unlikely(buf_index >= ctx->nr_user_bufs))
2755 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2756 imu = &ctx->user_bufs[index];
2757 buf_addr = req->rw.addr;
2760 if (buf_addr + len < buf_addr)
2762 /* not inside the mapped region */
2763 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2767 * May not be a start of buffer, set size appropriately
2768 * and advance us to the beginning.
2770 offset = buf_addr - imu->ubuf;
2771 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2775 * Don't use iov_iter_advance() here, as it's really slow for
2776 * using the latter parts of a big fixed buffer - it iterates
2777 * over each segment manually. We can cheat a bit here, because
2780 * 1) it's a BVEC iter, we set it up
2781 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2782 * first and last bvec
2784 * So just find our index, and adjust the iterator afterwards.
2785 * If the offset is within the first bvec (or the whole first
2786 * bvec, just use iov_iter_advance(). This makes it easier
2787 * since we can just skip the first segment, which may not
2788 * be PAGE_SIZE aligned.
2790 const struct bio_vec *bvec = imu->bvec;
2792 if (offset <= bvec->bv_len) {
2793 iov_iter_advance(iter, offset);
2795 unsigned long seg_skip;
2797 /* skip first vec */
2798 offset -= bvec->bv_len;
2799 seg_skip = 1 + (offset >> PAGE_SHIFT);
2801 iter->bvec = bvec + seg_skip;
2802 iter->nr_segs -= seg_skip;
2803 iter->count -= bvec->bv_len + offset;
2804 iter->iov_offset = offset & ~PAGE_MASK;
2811 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2814 mutex_unlock(&ctx->uring_lock);
2817 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2820 * "Normal" inline submissions always hold the uring_lock, since we
2821 * grab it from the system call. Same is true for the SQPOLL offload.
2822 * The only exception is when we've detached the request and issue it
2823 * from an async worker thread, grab the lock for that case.
2826 mutex_lock(&ctx->uring_lock);
2829 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2830 int bgid, struct io_buffer *kbuf,
2833 struct io_buffer *head;
2835 if (req->flags & REQ_F_BUFFER_SELECTED)
2838 io_ring_submit_lock(req->ctx, needs_lock);
2840 lockdep_assert_held(&req->ctx->uring_lock);
2842 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2844 if (!list_empty(&head->list)) {
2845 kbuf = list_last_entry(&head->list, struct io_buffer,
2847 list_del(&kbuf->list);
2850 idr_remove(&req->ctx->io_buffer_idr, bgid);
2852 if (*len > kbuf->len)
2855 kbuf = ERR_PTR(-ENOBUFS);
2858 io_ring_submit_unlock(req->ctx, needs_lock);
2863 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2866 struct io_buffer *kbuf;
2869 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2870 bgid = req->buf_index;
2871 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2874 req->rw.addr = (u64) (unsigned long) kbuf;
2875 req->flags |= REQ_F_BUFFER_SELECTED;
2876 return u64_to_user_ptr(kbuf->addr);
2879 #ifdef CONFIG_COMPAT
2880 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2883 struct compat_iovec __user *uiov;
2884 compat_ssize_t clen;
2888 uiov = u64_to_user_ptr(req->rw.addr);
2889 if (!access_ok(uiov, sizeof(*uiov)))
2891 if (__get_user(clen, &uiov->iov_len))
2897 buf = io_rw_buffer_select(req, &len, needs_lock);
2899 return PTR_ERR(buf);
2900 iov[0].iov_base = buf;
2901 iov[0].iov_len = (compat_size_t) len;
2906 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2909 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2913 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2916 len = iov[0].iov_len;
2919 buf = io_rw_buffer_select(req, &len, needs_lock);
2921 return PTR_ERR(buf);
2922 iov[0].iov_base = buf;
2923 iov[0].iov_len = len;
2927 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2930 if (req->flags & REQ_F_BUFFER_SELECTED) {
2931 struct io_buffer *kbuf;
2933 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2934 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2935 iov[0].iov_len = kbuf->len;
2940 else if (req->rw.len > 1)
2943 #ifdef CONFIG_COMPAT
2944 if (req->ctx->compat)
2945 return io_compat_import(req, iov, needs_lock);
2948 return __io_iov_buffer_select(req, iov, needs_lock);
2951 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2952 struct iovec **iovec, struct iov_iter *iter,
2955 void __user *buf = u64_to_user_ptr(req->rw.addr);
2956 size_t sqe_len = req->rw.len;
2960 opcode = req->opcode;
2961 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2963 return io_import_fixed(req, rw, iter);
2966 /* buffer index only valid with fixed read/write, or buffer select */
2967 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2970 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2971 if (req->flags & REQ_F_BUFFER_SELECT) {
2972 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2974 return PTR_ERR(buf);
2975 req->rw.len = sqe_len;
2978 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2980 return ret < 0 ? ret : sqe_len;
2983 if (req->flags & REQ_F_BUFFER_SELECT) {
2984 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2986 ret = (*iovec)->iov_len;
2987 iov_iter_init(iter, rw, *iovec, 1, ret);
2993 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2997 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2998 struct iovec **iovec, struct iov_iter *iter,
3001 struct io_async_rw *iorw = req->async_data;
3004 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3006 return iov_iter_count(&iorw->iter);
3009 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3011 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3015 * For files that don't have ->read_iter() and ->write_iter(), handle them
3016 * by looping over ->read() or ->write() manually.
3018 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
3019 struct iov_iter *iter)
3024 * Don't support polled IO through this interface, and we can't
3025 * support non-blocking either. For the latter, this just causes
3026 * the kiocb to be handled from an async context.
3028 if (kiocb->ki_flags & IOCB_HIPRI)
3030 if (kiocb->ki_flags & IOCB_NOWAIT)
3033 while (iov_iter_count(iter)) {
3037 if (!iov_iter_is_bvec(iter)) {
3038 iovec = iov_iter_iovec(iter);
3040 /* fixed buffers import bvec */
3041 iovec.iov_base = kmap(iter->bvec->bv_page)
3043 iovec.iov_len = min(iter->count,
3044 iter->bvec->bv_len - iter->iov_offset);
3048 nr = file->f_op->read(file, iovec.iov_base,
3049 iovec.iov_len, io_kiocb_ppos(kiocb));
3051 nr = file->f_op->write(file, iovec.iov_base,
3052 iovec.iov_len, io_kiocb_ppos(kiocb));
3055 if (iov_iter_is_bvec(iter))
3056 kunmap(iter->bvec->bv_page);
3064 if (nr != iovec.iov_len)
3066 iov_iter_advance(iter, nr);
3072 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3073 const struct iovec *fast_iov, struct iov_iter *iter)
3075 struct io_async_rw *rw = req->async_data;
3077 memcpy(&rw->iter, iter, sizeof(*iter));
3078 rw->free_iovec = iovec;
3080 /* can only be fixed buffers, no need to do anything */
3081 if (iter->type == ITER_BVEC)
3084 unsigned iov_off = 0;
3086 rw->iter.iov = rw->fast_iov;
3087 if (iter->iov != fast_iov) {
3088 iov_off = iter->iov - fast_iov;
3089 rw->iter.iov += iov_off;
3091 if (rw->fast_iov != fast_iov)
3092 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3093 sizeof(struct iovec) * iter->nr_segs);
3095 req->flags |= REQ_F_NEED_CLEANUP;
3099 static inline int __io_alloc_async_data(struct io_kiocb *req)
3101 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3102 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3103 return req->async_data == NULL;
3106 static int io_alloc_async_data(struct io_kiocb *req)
3108 if (!io_op_defs[req->opcode].needs_async_data)
3111 return __io_alloc_async_data(req);
3114 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3115 const struct iovec *fast_iov,
3116 struct iov_iter *iter, bool force)
3118 if (!force && !io_op_defs[req->opcode].needs_async_data)
3120 if (!req->async_data) {
3121 if (__io_alloc_async_data(req))
3124 io_req_map_rw(req, iovec, fast_iov, iter);
3129 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3131 struct io_async_rw *iorw = req->async_data;
3132 struct iovec *iov = iorw->fast_iov;
3135 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3136 if (unlikely(ret < 0))
3139 iorw->bytes_done = 0;
3140 iorw->free_iovec = iov;
3142 req->flags |= REQ_F_NEED_CLEANUP;
3146 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3150 ret = io_prep_rw(req, sqe);
3154 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3157 /* either don't need iovec imported or already have it */
3158 if (!req->async_data)
3160 return io_rw_prep_async(req, READ);
3164 * This is our waitqueue callback handler, registered through lock_page_async()
3165 * when we initially tried to do the IO with the iocb armed our waitqueue.
3166 * This gets called when the page is unlocked, and we generally expect that to
3167 * happen when the page IO is completed and the page is now uptodate. This will
3168 * queue a task_work based retry of the operation, attempting to copy the data
3169 * again. If the latter fails because the page was NOT uptodate, then we will
3170 * do a thread based blocking retry of the operation. That's the unexpected
3173 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3174 int sync, void *arg)
3176 struct wait_page_queue *wpq;
3177 struct io_kiocb *req = wait->private;
3178 struct wait_page_key *key = arg;
3181 wpq = container_of(wait, struct wait_page_queue, wait);
3183 if (!wake_page_match(wpq, key))
3186 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3187 list_del_init(&wait->entry);
3189 init_task_work(&req->task_work, io_req_task_submit);
3190 percpu_ref_get(&req->ctx->refs);
3192 /* submit ref gets dropped, acquire a new one */
3193 refcount_inc(&req->refs);
3194 ret = io_req_task_work_add(req, true);
3195 if (unlikely(ret)) {
3196 struct task_struct *tsk;
3198 /* queue just for cancelation */
3199 init_task_work(&req->task_work, io_req_task_cancel);
3200 tsk = io_wq_get_task(req->ctx->io_wq);
3201 task_work_add(tsk, &req->task_work, 0);
3202 wake_up_process(tsk);
3208 * This controls whether a given IO request should be armed for async page
3209 * based retry. If we return false here, the request is handed to the async
3210 * worker threads for retry. If we're doing buffered reads on a regular file,
3211 * we prepare a private wait_page_queue entry and retry the operation. This
3212 * will either succeed because the page is now uptodate and unlocked, or it
3213 * will register a callback when the page is unlocked at IO completion. Through
3214 * that callback, io_uring uses task_work to setup a retry of the operation.
3215 * That retry will attempt the buffered read again. The retry will generally
3216 * succeed, or in rare cases where it fails, we then fall back to using the
3217 * async worker threads for a blocking retry.
3219 static bool io_rw_should_retry(struct io_kiocb *req)
3221 struct io_async_rw *rw = req->async_data;
3222 struct wait_page_queue *wait = &rw->wpq;
3223 struct kiocb *kiocb = &req->rw.kiocb;
3225 /* never retry for NOWAIT, we just complete with -EAGAIN */
3226 if (req->flags & REQ_F_NOWAIT)
3229 /* Only for buffered IO */
3230 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3234 * just use poll if we can, and don't attempt if the fs doesn't
3235 * support callback based unlocks
3237 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3240 wait->wait.func = io_async_buf_func;
3241 wait->wait.private = req;
3242 wait->wait.flags = 0;
3243 INIT_LIST_HEAD(&wait->wait.entry);
3244 kiocb->ki_flags |= IOCB_WAITQ;
3245 kiocb->ki_flags &= ~IOCB_NOWAIT;
3246 kiocb->ki_waitq = wait;
3250 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3252 if (req->file->f_op->read_iter)
3253 return call_read_iter(req->file, &req->rw.kiocb, iter);
3254 else if (req->file->f_op->read)
3255 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3260 static int io_read(struct io_kiocb *req, bool force_nonblock,
3261 struct io_comp_state *cs)
3263 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3264 struct kiocb *kiocb = &req->rw.kiocb;
3265 struct iov_iter __iter, *iter = &__iter;
3266 struct io_async_rw *rw = req->async_data;
3267 ssize_t io_size, ret, ret2;
3274 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3277 iov_count = iov_iter_count(iter);
3279 req->result = io_size;
3282 /* Ensure we clear previously set non-block flag */
3283 if (!force_nonblock)
3284 kiocb->ki_flags &= ~IOCB_NOWAIT;
3286 kiocb->ki_flags |= IOCB_NOWAIT;
3289 /* If the file doesn't support async, just async punt */
3290 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3294 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3298 ret = io_iter_do_read(req, iter);
3302 } else if (ret == -EIOCBQUEUED) {
3305 } else if (ret == -EAGAIN) {
3306 /* IOPOLL retry should happen for io-wq threads */
3307 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3309 /* no retry on NONBLOCK marked file */
3310 if (req->file->f_flags & O_NONBLOCK)
3312 /* some cases will consume bytes even on error returns */
3313 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3316 } else if (ret < 0) {
3317 /* make sure -ERESTARTSYS -> -EINTR is done */
3321 /* read it all, or we did blocking attempt. no retry. */
3322 if (!iov_iter_count(iter) || !force_nonblock ||
3323 (req->file->f_flags & O_NONBLOCK))
3328 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3335 rw = req->async_data;
3336 /* it's copied and will be cleaned with ->io */
3338 /* now use our persistent iterator, if we aren't already */
3341 rw->bytes_done += ret;
3342 /* if we can retry, do so with the callbacks armed */
3343 if (!io_rw_should_retry(req)) {
3344 kiocb->ki_flags &= ~IOCB_WAITQ;
3349 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3350 * get -EIOCBQUEUED, then we'll get a notification when the desired
3351 * page gets unlocked. We can also get a partial read here, and if we
3352 * do, then just retry at the new offset.
3354 ret = io_iter_do_read(req, iter);
3355 if (ret == -EIOCBQUEUED) {
3358 } else if (ret > 0 && ret < io_size) {
3359 /* we got some bytes, but not all. retry. */
3363 kiocb_done(kiocb, ret, cs);
3366 /* it's reportedly faster than delegating the null check to kfree() */
3372 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3376 ret = io_prep_rw(req, sqe);
3380 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3383 /* either don't need iovec imported or already have it */
3384 if (!req->async_data)
3386 return io_rw_prep_async(req, WRITE);
3389 static int io_write(struct io_kiocb *req, bool force_nonblock,
3390 struct io_comp_state *cs)
3392 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3393 struct kiocb *kiocb = &req->rw.kiocb;
3394 struct iov_iter __iter, *iter = &__iter;
3395 struct io_async_rw *rw = req->async_data;
3397 ssize_t ret, ret2, io_size;
3402 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3405 iov_count = iov_iter_count(iter);
3407 req->result = io_size;
3409 /* Ensure we clear previously set non-block flag */
3410 if (!force_nonblock)
3411 kiocb->ki_flags &= ~IOCB_NOWAIT;
3413 kiocb->ki_flags |= IOCB_NOWAIT;
3415 /* If the file doesn't support async, just async punt */
3416 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3419 /* file path doesn't support NOWAIT for non-direct_IO */
3420 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3421 (req->flags & REQ_F_ISREG))
3424 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3429 * Open-code file_start_write here to grab freeze protection,
3430 * which will be released by another thread in
3431 * io_complete_rw(). Fool lockdep by telling it the lock got
3432 * released so that it doesn't complain about the held lock when
3433 * we return to userspace.
3435 if (req->flags & REQ_F_ISREG) {
3436 __sb_start_write(file_inode(req->file)->i_sb,
3437 SB_FREEZE_WRITE, true);
3438 __sb_writers_release(file_inode(req->file)->i_sb,
3441 kiocb->ki_flags |= IOCB_WRITE;
3443 if (req->file->f_op->write_iter)
3444 ret2 = call_write_iter(req->file, kiocb, iter);
3445 else if (req->file->f_op->write)
3446 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3451 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3452 * retry them without IOCB_NOWAIT.
3454 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3456 /* no retry on NONBLOCK marked file */
3457 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3459 if (!force_nonblock || ret2 != -EAGAIN) {
3460 /* IOPOLL retry should happen for io-wq threads */
3461 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3464 kiocb_done(kiocb, ret2, cs);
3467 /* some cases will consume bytes even on error returns */
3468 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3469 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3474 /* it's reportedly faster than delegating the null check to kfree() */
3480 static int __io_splice_prep(struct io_kiocb *req,
3481 const struct io_uring_sqe *sqe)
3483 struct io_splice* sp = &req->splice;
3484 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3486 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3490 sp->len = READ_ONCE(sqe->len);
3491 sp->flags = READ_ONCE(sqe->splice_flags);
3493 if (unlikely(sp->flags & ~valid_flags))
3496 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3497 (sp->flags & SPLICE_F_FD_IN_FIXED));
3500 req->flags |= REQ_F_NEED_CLEANUP;
3502 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3504 * Splice operation will be punted aync, and here need to
3505 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3507 io_req_init_async(req);
3508 req->work.flags |= IO_WQ_WORK_UNBOUND;
3514 static int io_tee_prep(struct io_kiocb *req,
3515 const struct io_uring_sqe *sqe)
3517 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3519 return __io_splice_prep(req, sqe);
3522 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3524 struct io_splice *sp = &req->splice;
3525 struct file *in = sp->file_in;
3526 struct file *out = sp->file_out;
3527 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3533 ret = do_tee(in, out, sp->len, flags);
3535 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3536 req->flags &= ~REQ_F_NEED_CLEANUP;
3539 req_set_fail_links(req);
3540 io_req_complete(req, ret);
3544 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3546 struct io_splice* sp = &req->splice;
3548 sp->off_in = READ_ONCE(sqe->splice_off_in);
3549 sp->off_out = READ_ONCE(sqe->off);
3550 return __io_splice_prep(req, sqe);
3553 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3555 struct io_splice *sp = &req->splice;
3556 struct file *in = sp->file_in;
3557 struct file *out = sp->file_out;
3558 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3559 loff_t *poff_in, *poff_out;
3565 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3566 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3569 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3571 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3572 req->flags &= ~REQ_F_NEED_CLEANUP;
3575 req_set_fail_links(req);
3576 io_req_complete(req, ret);
3581 * IORING_OP_NOP just posts a completion event, nothing else.
3583 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3585 struct io_ring_ctx *ctx = req->ctx;
3587 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3590 __io_req_complete(req, 0, 0, cs);
3594 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3596 struct io_ring_ctx *ctx = req->ctx;
3601 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3603 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3606 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3607 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3610 req->sync.off = READ_ONCE(sqe->off);
3611 req->sync.len = READ_ONCE(sqe->len);
3615 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3617 loff_t end = req->sync.off + req->sync.len;
3620 /* fsync always requires a blocking context */
3624 ret = vfs_fsync_range(req->file, req->sync.off,
3625 end > 0 ? end : LLONG_MAX,
3626 req->sync.flags & IORING_FSYNC_DATASYNC);
3628 req_set_fail_links(req);
3629 io_req_complete(req, ret);
3633 static int io_fallocate_prep(struct io_kiocb *req,
3634 const struct io_uring_sqe *sqe)
3636 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3638 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3641 req->sync.off = READ_ONCE(sqe->off);
3642 req->sync.len = READ_ONCE(sqe->addr);
3643 req->sync.mode = READ_ONCE(sqe->len);
3647 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3651 /* fallocate always requiring blocking context */
3654 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3657 req_set_fail_links(req);
3658 io_req_complete(req, ret);
3662 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3664 const char __user *fname;
3667 if (unlikely(sqe->ioprio || sqe->buf_index))
3669 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3672 /* open.how should be already initialised */
3673 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3674 req->open.how.flags |= O_LARGEFILE;
3676 req->open.dfd = READ_ONCE(sqe->fd);
3677 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3678 req->open.filename = getname(fname);
3679 if (IS_ERR(req->open.filename)) {
3680 ret = PTR_ERR(req->open.filename);
3681 req->open.filename = NULL;
3684 req->open.nofile = rlimit(RLIMIT_NOFILE);
3685 req->flags |= REQ_F_NEED_CLEANUP;
3689 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3693 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3695 mode = READ_ONCE(sqe->len);
3696 flags = READ_ONCE(sqe->open_flags);
3697 req->open.how = build_open_how(flags, mode);
3698 return __io_openat_prep(req, sqe);
3701 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3703 struct open_how __user *how;
3707 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3709 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3710 len = READ_ONCE(sqe->len);
3711 if (len < OPEN_HOW_SIZE_VER0)
3714 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3719 return __io_openat_prep(req, sqe);
3722 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3724 struct open_flags op;
3731 ret = build_open_flags(&req->open.how, &op);
3735 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3739 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3742 ret = PTR_ERR(file);
3744 fsnotify_open(file);
3745 fd_install(ret, file);
3748 putname(req->open.filename);
3749 req->flags &= ~REQ_F_NEED_CLEANUP;
3751 req_set_fail_links(req);
3752 io_req_complete(req, ret);
3756 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3758 return io_openat2(req, force_nonblock);
3761 static int io_remove_buffers_prep(struct io_kiocb *req,
3762 const struct io_uring_sqe *sqe)
3764 struct io_provide_buf *p = &req->pbuf;
3767 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3770 tmp = READ_ONCE(sqe->fd);
3771 if (!tmp || tmp > USHRT_MAX)
3774 memset(p, 0, sizeof(*p));
3776 p->bgid = READ_ONCE(sqe->buf_group);
3780 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3781 int bgid, unsigned nbufs)
3785 /* shouldn't happen */
3789 /* the head kbuf is the list itself */
3790 while (!list_empty(&buf->list)) {
3791 struct io_buffer *nxt;
3793 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3794 list_del(&nxt->list);
3801 idr_remove(&ctx->io_buffer_idr, bgid);
3806 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3807 struct io_comp_state *cs)
3809 struct io_provide_buf *p = &req->pbuf;
3810 struct io_ring_ctx *ctx = req->ctx;
3811 struct io_buffer *head;
3814 io_ring_submit_lock(ctx, !force_nonblock);
3816 lockdep_assert_held(&ctx->uring_lock);
3819 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3821 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3823 io_ring_submit_lock(ctx, !force_nonblock);
3825 req_set_fail_links(req);
3826 __io_req_complete(req, ret, 0, cs);
3830 static int io_provide_buffers_prep(struct io_kiocb *req,
3831 const struct io_uring_sqe *sqe)
3833 struct io_provide_buf *p = &req->pbuf;
3836 if (sqe->ioprio || sqe->rw_flags)
3839 tmp = READ_ONCE(sqe->fd);
3840 if (!tmp || tmp > USHRT_MAX)
3843 p->addr = READ_ONCE(sqe->addr);
3844 p->len = READ_ONCE(sqe->len);
3846 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3849 p->bgid = READ_ONCE(sqe->buf_group);
3850 tmp = READ_ONCE(sqe->off);
3851 if (tmp > USHRT_MAX)
3857 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3859 struct io_buffer *buf;
3860 u64 addr = pbuf->addr;
3861 int i, bid = pbuf->bid;
3863 for (i = 0; i < pbuf->nbufs; i++) {
3864 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3869 buf->len = pbuf->len;
3874 INIT_LIST_HEAD(&buf->list);
3877 list_add_tail(&buf->list, &(*head)->list);
3881 return i ? i : -ENOMEM;
3884 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3885 struct io_comp_state *cs)
3887 struct io_provide_buf *p = &req->pbuf;
3888 struct io_ring_ctx *ctx = req->ctx;
3889 struct io_buffer *head, *list;
3892 io_ring_submit_lock(ctx, !force_nonblock);
3894 lockdep_assert_held(&ctx->uring_lock);
3896 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3898 ret = io_add_buffers(p, &head);
3903 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3906 __io_remove_buffers(ctx, head, p->bgid, -1U);
3911 io_ring_submit_unlock(ctx, !force_nonblock);
3913 req_set_fail_links(req);
3914 __io_req_complete(req, ret, 0, cs);
3918 static int io_epoll_ctl_prep(struct io_kiocb *req,
3919 const struct io_uring_sqe *sqe)
3921 #if defined(CONFIG_EPOLL)
3922 if (sqe->ioprio || sqe->buf_index)
3924 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3927 req->epoll.epfd = READ_ONCE(sqe->fd);
3928 req->epoll.op = READ_ONCE(sqe->len);
3929 req->epoll.fd = READ_ONCE(sqe->off);
3931 if (ep_op_has_event(req->epoll.op)) {
3932 struct epoll_event __user *ev;
3934 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3935 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3945 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3946 struct io_comp_state *cs)
3948 #if defined(CONFIG_EPOLL)
3949 struct io_epoll *ie = &req->epoll;
3952 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3953 if (force_nonblock && ret == -EAGAIN)
3957 req_set_fail_links(req);
3958 __io_req_complete(req, ret, 0, cs);
3965 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3967 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3968 if (sqe->ioprio || sqe->buf_index || sqe->off)
3970 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3973 req->madvise.addr = READ_ONCE(sqe->addr);
3974 req->madvise.len = READ_ONCE(sqe->len);
3975 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3982 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3984 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3985 struct io_madvise *ma = &req->madvise;
3991 ret = do_madvise(ma->addr, ma->len, ma->advice);
3993 req_set_fail_links(req);
3994 io_req_complete(req, ret);
4001 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4003 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4005 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4008 req->fadvise.offset = READ_ONCE(sqe->off);
4009 req->fadvise.len = READ_ONCE(sqe->len);
4010 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4014 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4016 struct io_fadvise *fa = &req->fadvise;
4019 if (force_nonblock) {
4020 switch (fa->advice) {
4021 case POSIX_FADV_NORMAL:
4022 case POSIX_FADV_RANDOM:
4023 case POSIX_FADV_SEQUENTIAL:
4030 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4032 req_set_fail_links(req);
4033 io_req_complete(req, ret);
4037 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4039 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4041 if (sqe->ioprio || sqe->buf_index)
4043 if (req->flags & REQ_F_FIXED_FILE)
4046 req->statx.dfd = READ_ONCE(sqe->fd);
4047 req->statx.mask = READ_ONCE(sqe->len);
4048 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4049 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4050 req->statx.flags = READ_ONCE(sqe->statx_flags);
4055 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4057 struct io_statx *ctx = &req->statx;
4060 if (force_nonblock) {
4061 /* only need file table for an actual valid fd */
4062 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4063 req->flags |= REQ_F_NO_FILE_TABLE;
4067 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4071 req_set_fail_links(req);
4072 io_req_complete(req, ret);
4076 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4079 * If we queue this for async, it must not be cancellable. That would
4080 * leave the 'file' in an undeterminate state, and here need to modify
4081 * io_wq_work.flags, so initialize io_wq_work firstly.
4083 io_req_init_async(req);
4084 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4086 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4088 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4089 sqe->rw_flags || sqe->buf_index)
4091 if (req->flags & REQ_F_FIXED_FILE)
4094 req->close.fd = READ_ONCE(sqe->fd);
4095 if ((req->file && req->file->f_op == &io_uring_fops))
4098 req->close.put_file = NULL;
4102 static int io_close(struct io_kiocb *req, bool force_nonblock,
4103 struct io_comp_state *cs)
4105 struct io_close *close = &req->close;
4108 /* might be already done during nonblock submission */
4109 if (!close->put_file) {
4110 ret = __close_fd_get_file(close->fd, &close->put_file);
4112 return (ret == -ENOENT) ? -EBADF : ret;
4115 /* if the file has a flush method, be safe and punt to async */
4116 if (close->put_file->f_op->flush && force_nonblock) {
4117 /* was never set, but play safe */
4118 req->flags &= ~REQ_F_NOWAIT;
4119 /* avoid grabbing files - we don't need the files */
4120 req->flags |= REQ_F_NO_FILE_TABLE;
4124 /* No ->flush() or already async, safely close from here */
4125 ret = filp_close(close->put_file, req->work.files);
4127 req_set_fail_links(req);
4128 fput(close->put_file);
4129 close->put_file = NULL;
4130 __io_req_complete(req, ret, 0, cs);
4134 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4136 struct io_ring_ctx *ctx = req->ctx;
4141 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4143 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4146 req->sync.off = READ_ONCE(sqe->off);
4147 req->sync.len = READ_ONCE(sqe->len);
4148 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4152 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4156 /* sync_file_range always requires a blocking context */
4160 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4163 req_set_fail_links(req);
4164 io_req_complete(req, ret);
4168 #if defined(CONFIG_NET)
4169 static int io_setup_async_msg(struct io_kiocb *req,
4170 struct io_async_msghdr *kmsg)
4172 struct io_async_msghdr *async_msg = req->async_data;
4176 if (io_alloc_async_data(req)) {
4177 if (kmsg->iov != kmsg->fast_iov)
4181 async_msg = req->async_data;
4182 req->flags |= REQ_F_NEED_CLEANUP;
4183 memcpy(async_msg, kmsg, sizeof(*kmsg));
4187 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4188 struct io_async_msghdr *iomsg)
4190 iomsg->iov = iomsg->fast_iov;
4191 iomsg->msg.msg_name = &iomsg->addr;
4192 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4193 req->sr_msg.msg_flags, &iomsg->iov);
4196 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4198 struct io_async_msghdr *async_msg = req->async_data;
4199 struct io_sr_msg *sr = &req->sr_msg;
4202 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4205 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4206 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4207 sr->len = READ_ONCE(sqe->len);
4209 #ifdef CONFIG_COMPAT
4210 if (req->ctx->compat)
4211 sr->msg_flags |= MSG_CMSG_COMPAT;
4214 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4216 ret = io_sendmsg_copy_hdr(req, async_msg);
4218 req->flags |= REQ_F_NEED_CLEANUP;
4222 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4223 struct io_comp_state *cs)
4225 struct io_async_msghdr iomsg, *kmsg;
4226 struct socket *sock;
4230 sock = sock_from_file(req->file, &ret);
4231 if (unlikely(!sock))
4234 if (req->async_data) {
4235 kmsg = req->async_data;
4236 kmsg->msg.msg_name = &kmsg->addr;
4237 /* if iov is set, it's allocated already */
4239 kmsg->iov = kmsg->fast_iov;
4240 kmsg->msg.msg_iter.iov = kmsg->iov;
4242 ret = io_sendmsg_copy_hdr(req, &iomsg);
4248 flags = req->sr_msg.msg_flags;
4249 if (flags & MSG_DONTWAIT)
4250 req->flags |= REQ_F_NOWAIT;
4251 else if (force_nonblock)
4252 flags |= MSG_DONTWAIT;
4254 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4255 if (force_nonblock && ret == -EAGAIN)
4256 return io_setup_async_msg(req, kmsg);
4257 if (ret == -ERESTARTSYS)
4260 if (kmsg->iov != kmsg->fast_iov)
4262 req->flags &= ~REQ_F_NEED_CLEANUP;
4264 req_set_fail_links(req);
4265 __io_req_complete(req, ret, 0, cs);
4269 static int io_send(struct io_kiocb *req, bool force_nonblock,
4270 struct io_comp_state *cs)
4272 struct io_sr_msg *sr = &req->sr_msg;
4275 struct socket *sock;
4279 sock = sock_from_file(req->file, &ret);
4280 if (unlikely(!sock))
4283 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4287 msg.msg_name = NULL;
4288 msg.msg_control = NULL;
4289 msg.msg_controllen = 0;
4290 msg.msg_namelen = 0;
4292 flags = req->sr_msg.msg_flags;
4293 if (flags & MSG_DONTWAIT)
4294 req->flags |= REQ_F_NOWAIT;
4295 else if (force_nonblock)
4296 flags |= MSG_DONTWAIT;
4298 msg.msg_flags = flags;
4299 ret = sock_sendmsg(sock, &msg);
4300 if (force_nonblock && ret == -EAGAIN)
4302 if (ret == -ERESTARTSYS)
4306 req_set_fail_links(req);
4307 __io_req_complete(req, ret, 0, cs);
4311 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4312 struct io_async_msghdr *iomsg)
4314 struct io_sr_msg *sr = &req->sr_msg;
4315 struct iovec __user *uiov;
4319 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4320 &iomsg->uaddr, &uiov, &iov_len);
4324 if (req->flags & REQ_F_BUFFER_SELECT) {
4327 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4329 sr->len = iomsg->iov[0].iov_len;
4330 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4334 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4335 &iomsg->iov, &iomsg->msg.msg_iter,
4344 #ifdef CONFIG_COMPAT
4345 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4346 struct io_async_msghdr *iomsg)
4348 struct compat_msghdr __user *msg_compat;
4349 struct io_sr_msg *sr = &req->sr_msg;
4350 struct compat_iovec __user *uiov;
4355 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4356 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4361 uiov = compat_ptr(ptr);
4362 if (req->flags & REQ_F_BUFFER_SELECT) {
4363 compat_ssize_t clen;
4367 if (!access_ok(uiov, sizeof(*uiov)))
4369 if (__get_user(clen, &uiov->iov_len))
4373 sr->len = iomsg->iov[0].iov_len;
4376 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4377 UIO_FASTIOV, &iomsg->iov,
4378 &iomsg->msg.msg_iter, true);
4387 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4388 struct io_async_msghdr *iomsg)
4390 iomsg->msg.msg_name = &iomsg->addr;
4391 iomsg->iov = iomsg->fast_iov;
4393 #ifdef CONFIG_COMPAT
4394 if (req->ctx->compat)
4395 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4398 return __io_recvmsg_copy_hdr(req, iomsg);
4401 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4404 struct io_sr_msg *sr = &req->sr_msg;
4405 struct io_buffer *kbuf;
4407 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4412 req->flags |= REQ_F_BUFFER_SELECTED;
4416 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4418 return io_put_kbuf(req, req->sr_msg.kbuf);
4421 static int io_recvmsg_prep(struct io_kiocb *req,
4422 const struct io_uring_sqe *sqe)
4424 struct io_async_msghdr *async_msg = req->async_data;
4425 struct io_sr_msg *sr = &req->sr_msg;
4428 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4431 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4432 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4433 sr->len = READ_ONCE(sqe->len);
4434 sr->bgid = READ_ONCE(sqe->buf_group);
4436 #ifdef CONFIG_COMPAT
4437 if (req->ctx->compat)
4438 sr->msg_flags |= MSG_CMSG_COMPAT;
4441 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4443 ret = io_recvmsg_copy_hdr(req, async_msg);
4445 req->flags |= REQ_F_NEED_CLEANUP;
4449 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4450 struct io_comp_state *cs)
4452 struct io_async_msghdr iomsg, *kmsg;
4453 struct socket *sock;
4454 struct io_buffer *kbuf;
4456 int ret, cflags = 0;
4458 sock = sock_from_file(req->file, &ret);
4459 if (unlikely(!sock))
4462 if (req->async_data) {
4463 kmsg = req->async_data;
4464 kmsg->msg.msg_name = &kmsg->addr;
4465 /* if iov is set, it's allocated already */
4467 kmsg->iov = kmsg->fast_iov;
4468 kmsg->msg.msg_iter.iov = kmsg->iov;
4470 ret = io_recvmsg_copy_hdr(req, &iomsg);
4476 if (req->flags & REQ_F_BUFFER_SELECT) {
4477 kbuf = io_recv_buffer_select(req, !force_nonblock);
4479 return PTR_ERR(kbuf);
4480 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4481 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4482 1, req->sr_msg.len);
4485 flags = req->sr_msg.msg_flags;
4486 if (flags & MSG_DONTWAIT)
4487 req->flags |= REQ_F_NOWAIT;
4488 else if (force_nonblock)
4489 flags |= MSG_DONTWAIT;
4491 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4492 kmsg->uaddr, flags);
4493 if (force_nonblock && ret == -EAGAIN)
4494 return io_setup_async_msg(req, kmsg);
4495 if (ret == -ERESTARTSYS)
4498 if (req->flags & REQ_F_BUFFER_SELECTED)
4499 cflags = io_put_recv_kbuf(req);
4500 if (kmsg->iov != kmsg->fast_iov)
4502 req->flags &= ~REQ_F_NEED_CLEANUP;
4504 req_set_fail_links(req);
4505 __io_req_complete(req, ret, cflags, cs);
4509 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4510 struct io_comp_state *cs)
4512 struct io_buffer *kbuf;
4513 struct io_sr_msg *sr = &req->sr_msg;
4515 void __user *buf = sr->buf;
4516 struct socket *sock;
4519 int ret, cflags = 0;
4521 sock = sock_from_file(req->file, &ret);
4522 if (unlikely(!sock))
4525 if (req->flags & REQ_F_BUFFER_SELECT) {
4526 kbuf = io_recv_buffer_select(req, !force_nonblock);
4528 return PTR_ERR(kbuf);
4529 buf = u64_to_user_ptr(kbuf->addr);
4532 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4536 msg.msg_name = NULL;
4537 msg.msg_control = NULL;
4538 msg.msg_controllen = 0;
4539 msg.msg_namelen = 0;
4540 msg.msg_iocb = NULL;
4543 flags = req->sr_msg.msg_flags;
4544 if (flags & MSG_DONTWAIT)
4545 req->flags |= REQ_F_NOWAIT;
4546 else if (force_nonblock)
4547 flags |= MSG_DONTWAIT;
4549 ret = sock_recvmsg(sock, &msg, flags);
4550 if (force_nonblock && ret == -EAGAIN)
4552 if (ret == -ERESTARTSYS)
4555 if (req->flags & REQ_F_BUFFER_SELECTED)
4556 cflags = io_put_recv_kbuf(req);
4558 req_set_fail_links(req);
4559 __io_req_complete(req, ret, cflags, cs);
4563 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4565 struct io_accept *accept = &req->accept;
4567 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4569 if (sqe->ioprio || sqe->len || sqe->buf_index)
4572 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4573 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4574 accept->flags = READ_ONCE(sqe->accept_flags);
4575 accept->nofile = rlimit(RLIMIT_NOFILE);
4579 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4580 struct io_comp_state *cs)
4582 struct io_accept *accept = &req->accept;
4583 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4586 if (req->file->f_flags & O_NONBLOCK)
4587 req->flags |= REQ_F_NOWAIT;
4589 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4590 accept->addr_len, accept->flags,
4592 if (ret == -EAGAIN && force_nonblock)
4595 if (ret == -ERESTARTSYS)
4597 req_set_fail_links(req);
4599 __io_req_complete(req, ret, 0, cs);
4603 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4605 struct io_connect *conn = &req->connect;
4606 struct io_async_connect *io = req->async_data;
4608 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4610 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4613 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4614 conn->addr_len = READ_ONCE(sqe->addr2);
4619 return move_addr_to_kernel(conn->addr, conn->addr_len,
4623 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4624 struct io_comp_state *cs)
4626 struct io_async_connect __io, *io;
4627 unsigned file_flags;
4630 if (req->async_data) {
4631 io = req->async_data;
4633 ret = move_addr_to_kernel(req->connect.addr,
4634 req->connect.addr_len,
4641 file_flags = force_nonblock ? O_NONBLOCK : 0;
4643 ret = __sys_connect_file(req->file, &io->address,
4644 req->connect.addr_len, file_flags);
4645 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4646 if (req->async_data)
4648 if (io_alloc_async_data(req)) {
4652 io = req->async_data;
4653 memcpy(req->async_data, &__io, sizeof(__io));
4656 if (ret == -ERESTARTSYS)
4660 req_set_fail_links(req);
4661 __io_req_complete(req, ret, 0, cs);
4664 #else /* !CONFIG_NET */
4665 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4670 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4671 struct io_comp_state *cs)
4676 static int io_send(struct io_kiocb *req, bool force_nonblock,
4677 struct io_comp_state *cs)
4682 static int io_recvmsg_prep(struct io_kiocb *req,
4683 const struct io_uring_sqe *sqe)
4688 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4689 struct io_comp_state *cs)
4694 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4695 struct io_comp_state *cs)
4700 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4705 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4706 struct io_comp_state *cs)
4711 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4716 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4717 struct io_comp_state *cs)
4721 #endif /* CONFIG_NET */
4723 struct io_poll_table {
4724 struct poll_table_struct pt;
4725 struct io_kiocb *req;
4729 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4730 __poll_t mask, task_work_func_t func)
4735 /* for instances that support it check for an event match first: */
4736 if (mask && !(mask & poll->events))
4739 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4741 list_del_init(&poll->wait.entry);
4744 init_task_work(&req->task_work, func);
4745 percpu_ref_get(&req->ctx->refs);
4748 * If we using the signalfd wait_queue_head for this wakeup, then
4749 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4750 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4751 * either, as the normal wakeup will suffice.
4753 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4756 * If this fails, then the task is exiting. When a task exits, the
4757 * work gets canceled, so just cancel this request as well instead
4758 * of executing it. We can't safely execute it anyway, as we may not
4759 * have the needed state needed for it anyway.
4761 ret = io_req_task_work_add(req, twa_signal_ok);
4762 if (unlikely(ret)) {
4763 struct task_struct *tsk;
4765 WRITE_ONCE(poll->canceled, true);
4766 tsk = io_wq_get_task(req->ctx->io_wq);
4767 task_work_add(tsk, &req->task_work, 0);
4768 wake_up_process(tsk);
4773 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4774 __acquires(&req->ctx->completion_lock)
4776 struct io_ring_ctx *ctx = req->ctx;
4778 if (!req->result && !READ_ONCE(poll->canceled)) {
4779 struct poll_table_struct pt = { ._key = poll->events };
4781 req->result = vfs_poll(req->file, &pt) & poll->events;
4784 spin_lock_irq(&ctx->completion_lock);
4785 if (!req->result && !READ_ONCE(poll->canceled)) {
4786 add_wait_queue(poll->head, &poll->wait);
4793 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4795 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4796 if (req->opcode == IORING_OP_POLL_ADD)
4797 return req->async_data;
4798 return req->apoll->double_poll;
4801 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4803 if (req->opcode == IORING_OP_POLL_ADD)
4805 return &req->apoll->poll;
4808 static void io_poll_remove_double(struct io_kiocb *req)
4810 struct io_poll_iocb *poll = io_poll_get_double(req);
4812 lockdep_assert_held(&req->ctx->completion_lock);
4814 if (poll && poll->head) {
4815 struct wait_queue_head *head = poll->head;
4817 spin_lock(&head->lock);
4818 list_del_init(&poll->wait.entry);
4819 if (poll->wait.private)
4820 refcount_dec(&req->refs);
4822 spin_unlock(&head->lock);
4826 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4828 struct io_ring_ctx *ctx = req->ctx;
4830 io_poll_remove_double(req);
4831 req->poll.done = true;
4832 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4833 io_commit_cqring(ctx);
4836 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4838 struct io_ring_ctx *ctx = req->ctx;
4840 if (io_poll_rewait(req, &req->poll)) {
4841 spin_unlock_irq(&ctx->completion_lock);
4845 hash_del(&req->hash_node);
4846 io_poll_complete(req, req->result, 0);
4847 req->flags |= REQ_F_COMP_LOCKED;
4848 *nxt = io_put_req_find_next(req);
4849 spin_unlock_irq(&ctx->completion_lock);
4851 io_cqring_ev_posted(ctx);
4854 static void io_poll_task_func(struct callback_head *cb)
4856 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4857 struct io_ring_ctx *ctx = req->ctx;
4858 struct io_kiocb *nxt = NULL;
4860 io_poll_task_handler(req, &nxt);
4862 __io_req_task_submit(nxt);
4863 percpu_ref_put(&ctx->refs);
4866 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4867 int sync, void *key)
4869 struct io_kiocb *req = wait->private;
4870 struct io_poll_iocb *poll = io_poll_get_single(req);
4871 __poll_t mask = key_to_poll(key);
4873 /* for instances that support it check for an event match first: */
4874 if (mask && !(mask & poll->events))
4877 list_del_init(&wait->entry);
4879 if (poll && poll->head) {
4882 spin_lock(&poll->head->lock);
4883 done = list_empty(&poll->wait.entry);
4885 list_del_init(&poll->wait.entry);
4886 /* make sure double remove sees this as being gone */
4887 wait->private = NULL;
4888 spin_unlock(&poll->head->lock);
4890 __io_async_wake(req, poll, mask, io_poll_task_func);
4892 refcount_dec(&req->refs);
4896 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4897 wait_queue_func_t wake_func)
4901 poll->canceled = false;
4902 poll->events = events;
4903 INIT_LIST_HEAD(&poll->wait.entry);
4904 init_waitqueue_func_entry(&poll->wait, wake_func);
4907 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4908 struct wait_queue_head *head,
4909 struct io_poll_iocb **poll_ptr)
4911 struct io_kiocb *req = pt->req;
4914 * If poll->head is already set, it's because the file being polled
4915 * uses multiple waitqueues for poll handling (eg one for read, one
4916 * for write). Setup a separate io_poll_iocb if this happens.
4918 if (unlikely(poll->head)) {
4919 /* already have a 2nd entry, fail a third attempt */
4921 pt->error = -EINVAL;
4924 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4926 pt->error = -ENOMEM;
4929 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4930 refcount_inc(&req->refs);
4931 poll->wait.private = req;
4938 if (poll->events & EPOLLEXCLUSIVE)
4939 add_wait_queue_exclusive(head, &poll->wait);
4941 add_wait_queue(head, &poll->wait);
4944 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4945 struct poll_table_struct *p)
4947 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4948 struct async_poll *apoll = pt->req->apoll;
4950 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4953 static void io_async_task_func(struct callback_head *cb)
4955 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4956 struct async_poll *apoll = req->apoll;
4957 struct io_ring_ctx *ctx = req->ctx;
4959 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4961 if (io_poll_rewait(req, &apoll->poll)) {
4962 spin_unlock_irq(&ctx->completion_lock);
4963 percpu_ref_put(&ctx->refs);
4967 /* If req is still hashed, it cannot have been canceled. Don't check. */
4968 if (hash_hashed(&req->hash_node))
4969 hash_del(&req->hash_node);
4971 io_poll_remove_double(req);
4972 spin_unlock_irq(&ctx->completion_lock);
4974 if (!READ_ONCE(apoll->poll.canceled))
4975 __io_req_task_submit(req);
4977 __io_req_task_cancel(req, -ECANCELED);
4979 percpu_ref_put(&ctx->refs);
4980 kfree(apoll->double_poll);
4984 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4987 struct io_kiocb *req = wait->private;
4988 struct io_poll_iocb *poll = &req->apoll->poll;
4990 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4993 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4996 static void io_poll_req_insert(struct io_kiocb *req)
4998 struct io_ring_ctx *ctx = req->ctx;
4999 struct hlist_head *list;
5001 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5002 hlist_add_head(&req->hash_node, list);
5005 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5006 struct io_poll_iocb *poll,
5007 struct io_poll_table *ipt, __poll_t mask,
5008 wait_queue_func_t wake_func)
5009 __acquires(&ctx->completion_lock)
5011 struct io_ring_ctx *ctx = req->ctx;
5012 bool cancel = false;
5014 io_init_poll_iocb(poll, mask, wake_func);
5015 poll->file = req->file;
5016 poll->wait.private = req;
5018 ipt->pt._key = mask;
5020 ipt->error = -EINVAL;
5022 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5024 spin_lock_irq(&ctx->completion_lock);
5025 if (likely(poll->head)) {
5026 spin_lock(&poll->head->lock);
5027 if (unlikely(list_empty(&poll->wait.entry))) {
5033 if (mask || ipt->error)
5034 list_del_init(&poll->wait.entry);
5036 WRITE_ONCE(poll->canceled, true);
5037 else if (!poll->done) /* actually waiting for an event */
5038 io_poll_req_insert(req);
5039 spin_unlock(&poll->head->lock);
5045 static bool io_arm_poll_handler(struct io_kiocb *req)
5047 const struct io_op_def *def = &io_op_defs[req->opcode];
5048 struct io_ring_ctx *ctx = req->ctx;
5049 struct async_poll *apoll;
5050 struct io_poll_table ipt;
5054 if (!req->file || !file_can_poll(req->file))
5056 if (req->flags & REQ_F_POLLED)
5060 else if (def->pollout)
5064 /* if we can't nonblock try, then no point in arming a poll handler */
5065 if (!io_file_supports_async(req->file, rw))
5068 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5069 if (unlikely(!apoll))
5071 apoll->double_poll = NULL;
5073 req->flags |= REQ_F_POLLED;
5075 INIT_HLIST_NODE(&req->hash_node);
5079 mask |= POLLIN | POLLRDNORM;
5081 mask |= POLLOUT | POLLWRNORM;
5083 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5084 if ((req->opcode == IORING_OP_RECVMSG) &&
5085 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5088 mask |= POLLERR | POLLPRI;
5090 ipt.pt._qproc = io_async_queue_proc;
5092 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5094 if (ret || ipt.error) {
5095 io_poll_remove_double(req);
5096 spin_unlock_irq(&ctx->completion_lock);
5097 kfree(apoll->double_poll);
5101 spin_unlock_irq(&ctx->completion_lock);
5102 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5103 apoll->poll.events);
5107 static bool __io_poll_remove_one(struct io_kiocb *req,
5108 struct io_poll_iocb *poll)
5110 bool do_complete = false;
5112 spin_lock(&poll->head->lock);
5113 WRITE_ONCE(poll->canceled, true);
5114 if (!list_empty(&poll->wait.entry)) {
5115 list_del_init(&poll->wait.entry);
5118 spin_unlock(&poll->head->lock);
5119 hash_del(&req->hash_node);
5123 static bool io_poll_remove_one(struct io_kiocb *req)
5127 io_poll_remove_double(req);
5129 if (req->opcode == IORING_OP_POLL_ADD) {
5130 do_complete = __io_poll_remove_one(req, &req->poll);
5132 struct async_poll *apoll = req->apoll;
5134 /* non-poll requests have submit ref still */
5135 do_complete = __io_poll_remove_one(req, &apoll->poll);
5138 kfree(apoll->double_poll);
5144 io_cqring_fill_event(req, -ECANCELED);
5145 io_commit_cqring(req->ctx);
5146 req->flags |= REQ_F_COMP_LOCKED;
5147 req_set_fail_links(req);
5155 * Returns true if we found and killed one or more poll requests
5157 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5159 struct hlist_node *tmp;
5160 struct io_kiocb *req;
5163 spin_lock_irq(&ctx->completion_lock);
5164 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5165 struct hlist_head *list;
5167 list = &ctx->cancel_hash[i];
5168 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5169 if (io_task_match(req, tsk))
5170 posted += io_poll_remove_one(req);
5173 spin_unlock_irq(&ctx->completion_lock);
5176 io_cqring_ev_posted(ctx);
5181 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5183 struct hlist_head *list;
5184 struct io_kiocb *req;
5186 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5187 hlist_for_each_entry(req, list, hash_node) {
5188 if (sqe_addr != req->user_data)
5190 if (io_poll_remove_one(req))
5198 static int io_poll_remove_prep(struct io_kiocb *req,
5199 const struct io_uring_sqe *sqe)
5201 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5203 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5207 req->poll.addr = READ_ONCE(sqe->addr);
5212 * Find a running poll command that matches one specified in sqe->addr,
5213 * and remove it if found.
5215 static int io_poll_remove(struct io_kiocb *req)
5217 struct io_ring_ctx *ctx = req->ctx;
5221 addr = req->poll.addr;
5222 spin_lock_irq(&ctx->completion_lock);
5223 ret = io_poll_cancel(ctx, addr);
5224 spin_unlock_irq(&ctx->completion_lock);
5227 req_set_fail_links(req);
5228 io_req_complete(req, ret);
5232 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5235 struct io_kiocb *req = wait->private;
5236 struct io_poll_iocb *poll = &req->poll;
5238 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5241 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5242 struct poll_table_struct *p)
5244 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5246 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5249 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5251 struct io_poll_iocb *poll = &req->poll;
5254 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5256 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5261 events = READ_ONCE(sqe->poll32_events);
5263 events = swahw32(events);
5265 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5266 (events & EPOLLEXCLUSIVE);
5270 static int io_poll_add(struct io_kiocb *req)
5272 struct io_poll_iocb *poll = &req->poll;
5273 struct io_ring_ctx *ctx = req->ctx;
5274 struct io_poll_table ipt;
5277 INIT_HLIST_NODE(&req->hash_node);
5278 ipt.pt._qproc = io_poll_queue_proc;
5280 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5283 if (mask) { /* no async, we'd stolen it */
5285 io_poll_complete(req, mask, 0);
5287 spin_unlock_irq(&ctx->completion_lock);
5290 io_cqring_ev_posted(ctx);
5296 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5298 struct io_timeout_data *data = container_of(timer,
5299 struct io_timeout_data, timer);
5300 struct io_kiocb *req = data->req;
5301 struct io_ring_ctx *ctx = req->ctx;
5302 unsigned long flags;
5304 spin_lock_irqsave(&ctx->completion_lock, flags);
5305 list_del_init(&req->timeout.list);
5306 atomic_set(&req->ctx->cq_timeouts,
5307 atomic_read(&req->ctx->cq_timeouts) + 1);
5309 io_cqring_fill_event(req, -ETIME);
5310 io_commit_cqring(ctx);
5311 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5313 io_cqring_ev_posted(ctx);
5314 req_set_fail_links(req);
5316 return HRTIMER_NORESTART;
5319 static int __io_timeout_cancel(struct io_kiocb *req)
5321 struct io_timeout_data *io = req->async_data;
5324 ret = hrtimer_try_to_cancel(&io->timer);
5327 list_del_init(&req->timeout.list);
5329 req_set_fail_links(req);
5330 req->flags |= REQ_F_COMP_LOCKED;
5331 io_cqring_fill_event(req, -ECANCELED);
5336 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5338 struct io_kiocb *req;
5341 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5342 if (user_data == req->user_data) {
5351 return __io_timeout_cancel(req);
5354 static int io_timeout_remove_prep(struct io_kiocb *req,
5355 const struct io_uring_sqe *sqe)
5357 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5359 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5361 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5364 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5369 * Remove or update an existing timeout command
5371 static int io_timeout_remove(struct io_kiocb *req)
5373 struct io_ring_ctx *ctx = req->ctx;
5376 spin_lock_irq(&ctx->completion_lock);
5377 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5379 io_cqring_fill_event(req, ret);
5380 io_commit_cqring(ctx);
5381 spin_unlock_irq(&ctx->completion_lock);
5382 io_cqring_ev_posted(ctx);
5384 req_set_fail_links(req);
5389 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5390 bool is_timeout_link)
5392 struct io_timeout_data *data;
5394 u32 off = READ_ONCE(sqe->off);
5396 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5398 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5400 if (off && is_timeout_link)
5402 flags = READ_ONCE(sqe->timeout_flags);
5403 if (flags & ~IORING_TIMEOUT_ABS)
5406 req->timeout.off = off;
5408 if (!req->async_data && io_alloc_async_data(req))
5411 data = req->async_data;
5414 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5417 if (flags & IORING_TIMEOUT_ABS)
5418 data->mode = HRTIMER_MODE_ABS;
5420 data->mode = HRTIMER_MODE_REL;
5422 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5426 static int io_timeout(struct io_kiocb *req)
5428 struct io_ring_ctx *ctx = req->ctx;
5429 struct io_timeout_data *data = req->async_data;
5430 struct list_head *entry;
5431 u32 tail, off = req->timeout.off;
5433 spin_lock_irq(&ctx->completion_lock);
5436 * sqe->off holds how many events that need to occur for this
5437 * timeout event to be satisfied. If it isn't set, then this is
5438 * a pure timeout request, sequence isn't used.
5440 if (io_is_timeout_noseq(req)) {
5441 entry = ctx->timeout_list.prev;
5445 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5446 req->timeout.target_seq = tail + off;
5449 * Insertion sort, ensuring the first entry in the list is always
5450 * the one we need first.
5452 list_for_each_prev(entry, &ctx->timeout_list) {
5453 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5456 if (io_is_timeout_noseq(nxt))
5458 /* nxt.seq is behind @tail, otherwise would've been completed */
5459 if (off >= nxt->timeout.target_seq - tail)
5463 list_add(&req->timeout.list, entry);
5464 data->timer.function = io_timeout_fn;
5465 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5466 spin_unlock_irq(&ctx->completion_lock);
5470 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5472 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5474 return req->user_data == (unsigned long) data;
5477 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5479 enum io_wq_cancel cancel_ret;
5482 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5483 switch (cancel_ret) {
5484 case IO_WQ_CANCEL_OK:
5487 case IO_WQ_CANCEL_RUNNING:
5490 case IO_WQ_CANCEL_NOTFOUND:
5498 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5499 struct io_kiocb *req, __u64 sqe_addr,
5502 unsigned long flags;
5505 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5506 if (ret != -ENOENT) {
5507 spin_lock_irqsave(&ctx->completion_lock, flags);
5511 spin_lock_irqsave(&ctx->completion_lock, flags);
5512 ret = io_timeout_cancel(ctx, sqe_addr);
5515 ret = io_poll_cancel(ctx, sqe_addr);
5519 io_cqring_fill_event(req, ret);
5520 io_commit_cqring(ctx);
5521 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5522 io_cqring_ev_posted(ctx);
5525 req_set_fail_links(req);
5529 static int io_async_cancel_prep(struct io_kiocb *req,
5530 const struct io_uring_sqe *sqe)
5532 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5534 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5536 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5539 req->cancel.addr = READ_ONCE(sqe->addr);
5543 static int io_async_cancel(struct io_kiocb *req)
5545 struct io_ring_ctx *ctx = req->ctx;
5547 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5551 static int io_files_update_prep(struct io_kiocb *req,
5552 const struct io_uring_sqe *sqe)
5554 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5556 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5558 if (sqe->ioprio || sqe->rw_flags)
5561 req->files_update.offset = READ_ONCE(sqe->off);
5562 req->files_update.nr_args = READ_ONCE(sqe->len);
5563 if (!req->files_update.nr_args)
5565 req->files_update.arg = READ_ONCE(sqe->addr);
5569 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5570 struct io_comp_state *cs)
5572 struct io_ring_ctx *ctx = req->ctx;
5573 struct io_uring_files_update up;
5579 up.offset = req->files_update.offset;
5580 up.fds = req->files_update.arg;
5582 mutex_lock(&ctx->uring_lock);
5583 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5584 mutex_unlock(&ctx->uring_lock);
5587 req_set_fail_links(req);
5588 __io_req_complete(req, ret, 0, cs);
5592 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5594 switch (req->opcode) {
5597 case IORING_OP_READV:
5598 case IORING_OP_READ_FIXED:
5599 case IORING_OP_READ:
5600 return io_read_prep(req, sqe);
5601 case IORING_OP_WRITEV:
5602 case IORING_OP_WRITE_FIXED:
5603 case IORING_OP_WRITE:
5604 return io_write_prep(req, sqe);
5605 case IORING_OP_POLL_ADD:
5606 return io_poll_add_prep(req, sqe);
5607 case IORING_OP_POLL_REMOVE:
5608 return io_poll_remove_prep(req, sqe);
5609 case IORING_OP_FSYNC:
5610 return io_prep_fsync(req, sqe);
5611 case IORING_OP_SYNC_FILE_RANGE:
5612 return io_prep_sfr(req, sqe);
5613 case IORING_OP_SENDMSG:
5614 case IORING_OP_SEND:
5615 return io_sendmsg_prep(req, sqe);
5616 case IORING_OP_RECVMSG:
5617 case IORING_OP_RECV:
5618 return io_recvmsg_prep(req, sqe);
5619 case IORING_OP_CONNECT:
5620 return io_connect_prep(req, sqe);
5621 case IORING_OP_TIMEOUT:
5622 return io_timeout_prep(req, sqe, false);
5623 case IORING_OP_TIMEOUT_REMOVE:
5624 return io_timeout_remove_prep(req, sqe);
5625 case IORING_OP_ASYNC_CANCEL:
5626 return io_async_cancel_prep(req, sqe);
5627 case IORING_OP_LINK_TIMEOUT:
5628 return io_timeout_prep(req, sqe, true);
5629 case IORING_OP_ACCEPT:
5630 return io_accept_prep(req, sqe);
5631 case IORING_OP_FALLOCATE:
5632 return io_fallocate_prep(req, sqe);
5633 case IORING_OP_OPENAT:
5634 return io_openat_prep(req, sqe);
5635 case IORING_OP_CLOSE:
5636 return io_close_prep(req, sqe);
5637 case IORING_OP_FILES_UPDATE:
5638 return io_files_update_prep(req, sqe);
5639 case IORING_OP_STATX:
5640 return io_statx_prep(req, sqe);
5641 case IORING_OP_FADVISE:
5642 return io_fadvise_prep(req, sqe);
5643 case IORING_OP_MADVISE:
5644 return io_madvise_prep(req, sqe);
5645 case IORING_OP_OPENAT2:
5646 return io_openat2_prep(req, sqe);
5647 case IORING_OP_EPOLL_CTL:
5648 return io_epoll_ctl_prep(req, sqe);
5649 case IORING_OP_SPLICE:
5650 return io_splice_prep(req, sqe);
5651 case IORING_OP_PROVIDE_BUFFERS:
5652 return io_provide_buffers_prep(req, sqe);
5653 case IORING_OP_REMOVE_BUFFERS:
5654 return io_remove_buffers_prep(req, sqe);
5656 return io_tee_prep(req, sqe);
5659 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5664 static int io_req_defer_prep(struct io_kiocb *req,
5665 const struct io_uring_sqe *sqe)
5669 if (io_alloc_async_data(req))
5671 return io_req_prep(req, sqe);
5674 static u32 io_get_sequence(struct io_kiocb *req)
5676 struct io_kiocb *pos;
5677 struct io_ring_ctx *ctx = req->ctx;
5678 u32 total_submitted, nr_reqs = 1;
5680 if (req->flags & REQ_F_LINK_HEAD)
5681 list_for_each_entry(pos, &req->link_list, link_list)
5684 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5685 return total_submitted - nr_reqs;
5688 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5690 struct io_ring_ctx *ctx = req->ctx;
5691 struct io_defer_entry *de;
5695 /* Still need defer if there is pending req in defer list. */
5696 if (likely(list_empty_careful(&ctx->defer_list) &&
5697 !(req->flags & REQ_F_IO_DRAIN)))
5700 seq = io_get_sequence(req);
5701 /* Still a chance to pass the sequence check */
5702 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5705 if (!req->async_data) {
5706 ret = io_req_defer_prep(req, sqe);
5710 io_prep_async_link(req);
5711 de = kmalloc(sizeof(*de), GFP_KERNEL);
5715 spin_lock_irq(&ctx->completion_lock);
5716 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5717 spin_unlock_irq(&ctx->completion_lock);
5719 io_queue_async_work(req);
5720 return -EIOCBQUEUED;
5723 trace_io_uring_defer(ctx, req, req->user_data);
5726 list_add_tail(&de->list, &ctx->defer_list);
5727 spin_unlock_irq(&ctx->completion_lock);
5728 return -EIOCBQUEUED;
5731 static void io_req_drop_files(struct io_kiocb *req)
5733 struct io_ring_ctx *ctx = req->ctx;
5734 unsigned long flags;
5736 spin_lock_irqsave(&ctx->inflight_lock, flags);
5737 list_del(&req->inflight_entry);
5738 if (waitqueue_active(&ctx->inflight_wait))
5739 wake_up(&ctx->inflight_wait);
5740 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5741 req->flags &= ~REQ_F_INFLIGHT;
5742 put_files_struct(req->work.files);
5743 put_nsproxy(req->work.nsproxy);
5744 req->work.files = NULL;
5747 static void __io_clean_op(struct io_kiocb *req)
5749 if (req->flags & REQ_F_BUFFER_SELECTED) {
5750 switch (req->opcode) {
5751 case IORING_OP_READV:
5752 case IORING_OP_READ_FIXED:
5753 case IORING_OP_READ:
5754 kfree((void *)(unsigned long)req->rw.addr);
5756 case IORING_OP_RECVMSG:
5757 case IORING_OP_RECV:
5758 kfree(req->sr_msg.kbuf);
5761 req->flags &= ~REQ_F_BUFFER_SELECTED;
5764 if (req->flags & REQ_F_NEED_CLEANUP) {
5765 switch (req->opcode) {
5766 case IORING_OP_READV:
5767 case IORING_OP_READ_FIXED:
5768 case IORING_OP_READ:
5769 case IORING_OP_WRITEV:
5770 case IORING_OP_WRITE_FIXED:
5771 case IORING_OP_WRITE: {
5772 struct io_async_rw *io = req->async_data;
5774 kfree(io->free_iovec);
5777 case IORING_OP_RECVMSG:
5778 case IORING_OP_SENDMSG: {
5779 struct io_async_msghdr *io = req->async_data;
5780 if (io->iov != io->fast_iov)
5784 case IORING_OP_SPLICE:
5786 io_put_file(req, req->splice.file_in,
5787 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5789 case IORING_OP_OPENAT:
5790 case IORING_OP_OPENAT2:
5791 if (req->open.filename)
5792 putname(req->open.filename);
5795 req->flags &= ~REQ_F_NEED_CLEANUP;
5798 if (req->flags & REQ_F_INFLIGHT)
5799 io_req_drop_files(req);
5802 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5803 struct io_comp_state *cs)
5805 struct io_ring_ctx *ctx = req->ctx;
5808 switch (req->opcode) {
5810 ret = io_nop(req, cs);
5812 case IORING_OP_READV:
5813 case IORING_OP_READ_FIXED:
5814 case IORING_OP_READ:
5815 ret = io_read(req, force_nonblock, cs);
5817 case IORING_OP_WRITEV:
5818 case IORING_OP_WRITE_FIXED:
5819 case IORING_OP_WRITE:
5820 ret = io_write(req, force_nonblock, cs);
5822 case IORING_OP_FSYNC:
5823 ret = io_fsync(req, force_nonblock);
5825 case IORING_OP_POLL_ADD:
5826 ret = io_poll_add(req);
5828 case IORING_OP_POLL_REMOVE:
5829 ret = io_poll_remove(req);
5831 case IORING_OP_SYNC_FILE_RANGE:
5832 ret = io_sync_file_range(req, force_nonblock);
5834 case IORING_OP_SENDMSG:
5835 ret = io_sendmsg(req, force_nonblock, cs);
5837 case IORING_OP_SEND:
5838 ret = io_send(req, force_nonblock, cs);
5840 case IORING_OP_RECVMSG:
5841 ret = io_recvmsg(req, force_nonblock, cs);
5843 case IORING_OP_RECV:
5844 ret = io_recv(req, force_nonblock, cs);
5846 case IORING_OP_TIMEOUT:
5847 ret = io_timeout(req);
5849 case IORING_OP_TIMEOUT_REMOVE:
5850 ret = io_timeout_remove(req);
5852 case IORING_OP_ACCEPT:
5853 ret = io_accept(req, force_nonblock, cs);
5855 case IORING_OP_CONNECT:
5856 ret = io_connect(req, force_nonblock, cs);
5858 case IORING_OP_ASYNC_CANCEL:
5859 ret = io_async_cancel(req);
5861 case IORING_OP_FALLOCATE:
5862 ret = io_fallocate(req, force_nonblock);
5864 case IORING_OP_OPENAT:
5865 ret = io_openat(req, force_nonblock);
5867 case IORING_OP_CLOSE:
5868 ret = io_close(req, force_nonblock, cs);
5870 case IORING_OP_FILES_UPDATE:
5871 ret = io_files_update(req, force_nonblock, cs);
5873 case IORING_OP_STATX:
5874 ret = io_statx(req, force_nonblock);
5876 case IORING_OP_FADVISE:
5877 ret = io_fadvise(req, force_nonblock);
5879 case IORING_OP_MADVISE:
5880 ret = io_madvise(req, force_nonblock);
5882 case IORING_OP_OPENAT2:
5883 ret = io_openat2(req, force_nonblock);
5885 case IORING_OP_EPOLL_CTL:
5886 ret = io_epoll_ctl(req, force_nonblock, cs);
5888 case IORING_OP_SPLICE:
5889 ret = io_splice(req, force_nonblock);
5891 case IORING_OP_PROVIDE_BUFFERS:
5892 ret = io_provide_buffers(req, force_nonblock, cs);
5894 case IORING_OP_REMOVE_BUFFERS:
5895 ret = io_remove_buffers(req, force_nonblock, cs);
5898 ret = io_tee(req, force_nonblock);
5908 /* If the op doesn't have a file, we're not polling for it */
5909 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5910 const bool in_async = io_wq_current_is_worker();
5912 /* workqueue context doesn't hold uring_lock, grab it now */
5914 mutex_lock(&ctx->uring_lock);
5916 io_iopoll_req_issued(req);
5919 mutex_unlock(&ctx->uring_lock);
5925 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5927 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5928 struct io_kiocb *timeout;
5931 timeout = io_prep_linked_timeout(req);
5933 io_queue_linked_timeout(timeout);
5935 /* if NO_CANCEL is set, we must still run the work */
5936 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5937 IO_WQ_WORK_CANCEL) {
5943 ret = io_issue_sqe(req, false, NULL);
5945 * We can get EAGAIN for polled IO even though we're
5946 * forcing a sync submission from here, since we can't
5947 * wait for request slots on the block side.
5956 req_set_fail_links(req);
5957 io_req_complete(req, ret);
5960 return io_steal_work(req);
5963 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5966 struct fixed_file_table *table;
5968 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5969 return table->files[index & IORING_FILE_TABLE_MASK];
5972 static struct file *io_file_get(struct io_submit_state *state,
5973 struct io_kiocb *req, int fd, bool fixed)
5975 struct io_ring_ctx *ctx = req->ctx;
5979 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
5981 fd = array_index_nospec(fd, ctx->nr_user_files);
5982 file = io_file_from_index(ctx, fd);
5984 req->fixed_file_refs = &ctx->file_data->node->refs;
5985 percpu_ref_get(req->fixed_file_refs);
5988 trace_io_uring_file_get(ctx, fd);
5989 file = __io_file_get(state, fd);
5995 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6000 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6001 if (unlikely(!fixed && io_async_submit(req->ctx)))
6004 req->file = io_file_get(state, req, fd, fixed);
6005 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6010 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6012 struct io_timeout_data *data = container_of(timer,
6013 struct io_timeout_data, timer);
6014 struct io_kiocb *req = data->req;
6015 struct io_ring_ctx *ctx = req->ctx;
6016 struct io_kiocb *prev = NULL;
6017 unsigned long flags;
6019 spin_lock_irqsave(&ctx->completion_lock, flags);
6022 * We don't expect the list to be empty, that will only happen if we
6023 * race with the completion of the linked work.
6025 if (!list_empty(&req->link_list)) {
6026 prev = list_entry(req->link_list.prev, struct io_kiocb,
6028 if (refcount_inc_not_zero(&prev->refs)) {
6029 list_del_init(&req->link_list);
6030 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6035 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6038 req_set_fail_links(prev);
6039 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6042 io_req_complete(req, -ETIME);
6044 return HRTIMER_NORESTART;
6047 static void __io_queue_linked_timeout(struct io_kiocb *req)
6050 * If the list is now empty, then our linked request finished before
6051 * we got a chance to setup the timer
6053 if (!list_empty(&req->link_list)) {
6054 struct io_timeout_data *data = req->async_data;
6056 data->timer.function = io_link_timeout_fn;
6057 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6062 static void io_queue_linked_timeout(struct io_kiocb *req)
6064 struct io_ring_ctx *ctx = req->ctx;
6066 spin_lock_irq(&ctx->completion_lock);
6067 __io_queue_linked_timeout(req);
6068 spin_unlock_irq(&ctx->completion_lock);
6070 /* drop submission reference */
6074 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6076 struct io_kiocb *nxt;
6078 if (!(req->flags & REQ_F_LINK_HEAD))
6080 if (req->flags & REQ_F_LINK_TIMEOUT)
6083 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6085 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6088 req->flags |= REQ_F_LINK_TIMEOUT;
6092 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6094 struct io_kiocb *linked_timeout;
6095 struct io_kiocb *nxt;
6096 const struct cred *old_creds = NULL;
6100 linked_timeout = io_prep_linked_timeout(req);
6102 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6103 req->work.creds != current_cred()) {
6105 revert_creds(old_creds);
6106 if (old_creds == req->work.creds)
6107 old_creds = NULL; /* restored original creds */
6109 old_creds = override_creds(req->work.creds);
6112 ret = io_issue_sqe(req, true, cs);
6115 * We async punt it if the file wasn't marked NOWAIT, or if the file
6116 * doesn't support non-blocking read/write attempts
6118 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6119 if (!io_arm_poll_handler(req)) {
6122 * Queued up for async execution, worker will release
6123 * submit reference when the iocb is actually submitted.
6125 io_queue_async_work(req);
6129 io_queue_linked_timeout(linked_timeout);
6133 if (unlikely(ret)) {
6134 /* un-prep timeout, so it'll be killed as any other linked */
6135 req->flags &= ~REQ_F_LINK_TIMEOUT;
6136 req_set_fail_links(req);
6138 io_req_complete(req, ret);
6142 /* drop submission reference */
6143 nxt = io_put_req_find_next(req);
6145 io_queue_linked_timeout(linked_timeout);
6150 if (req->flags & REQ_F_FORCE_ASYNC)
6156 revert_creds(old_creds);
6159 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6160 struct io_comp_state *cs)
6164 ret = io_req_defer(req, sqe);
6166 if (ret != -EIOCBQUEUED) {
6168 req_set_fail_links(req);
6170 io_req_complete(req, ret);
6172 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6173 if (!req->async_data) {
6174 ret = io_req_defer_prep(req, sqe);
6180 * Never try inline submit of IOSQE_ASYNC is set, go straight
6181 * to async execution.
6183 io_req_init_async(req);
6184 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6185 io_queue_async_work(req);
6188 ret = io_req_prep(req, sqe);
6192 __io_queue_sqe(req, cs);
6196 static inline void io_queue_link_head(struct io_kiocb *req,
6197 struct io_comp_state *cs)
6199 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6201 io_req_complete(req, -ECANCELED);
6203 io_queue_sqe(req, NULL, cs);
6206 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6207 struct io_kiocb **link, struct io_comp_state *cs)
6209 struct io_ring_ctx *ctx = req->ctx;
6213 * If we already have a head request, queue this one for async
6214 * submittal once the head completes. If we don't have a head but
6215 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6216 * submitted sync once the chain is complete. If none of those
6217 * conditions are true (normal request), then just queue it.
6220 struct io_kiocb *head = *link;
6223 * Taking sequential execution of a link, draining both sides
6224 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6225 * requests in the link. So, it drains the head and the
6226 * next after the link request. The last one is done via
6227 * drain_next flag to persist the effect across calls.
6229 if (req->flags & REQ_F_IO_DRAIN) {
6230 head->flags |= REQ_F_IO_DRAIN;
6231 ctx->drain_next = 1;
6233 ret = io_req_defer_prep(req, sqe);
6234 if (unlikely(ret)) {
6235 /* fail even hard links since we don't submit */
6236 head->flags |= REQ_F_FAIL_LINK;
6239 trace_io_uring_link(ctx, req, head);
6240 list_add_tail(&req->link_list, &head->link_list);
6242 /* last request of a link, enqueue the link */
6243 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6244 io_queue_link_head(head, cs);
6248 if (unlikely(ctx->drain_next)) {
6249 req->flags |= REQ_F_IO_DRAIN;
6250 ctx->drain_next = 0;
6252 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6253 req->flags |= REQ_F_LINK_HEAD;
6254 INIT_LIST_HEAD(&req->link_list);
6256 ret = io_req_defer_prep(req, sqe);
6258 req->flags |= REQ_F_FAIL_LINK;
6261 io_queue_sqe(req, sqe, cs);
6269 * Batched submission is done, ensure local IO is flushed out.
6271 static void io_submit_state_end(struct io_submit_state *state)
6273 if (!list_empty(&state->comp.list))
6274 io_submit_flush_completions(&state->comp);
6275 blk_finish_plug(&state->plug);
6276 io_state_file_put(state);
6277 if (state->free_reqs)
6278 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6282 * Start submission side cache.
6284 static void io_submit_state_start(struct io_submit_state *state,
6285 struct io_ring_ctx *ctx, unsigned int max_ios)
6287 blk_start_plug(&state->plug);
6289 INIT_LIST_HEAD(&state->comp.list);
6290 state->comp.ctx = ctx;
6291 state->free_reqs = 0;
6293 state->ios_left = max_ios;
6296 static void io_commit_sqring(struct io_ring_ctx *ctx)
6298 struct io_rings *rings = ctx->rings;
6301 * Ensure any loads from the SQEs are done at this point,
6302 * since once we write the new head, the application could
6303 * write new data to them.
6305 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6309 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6310 * that is mapped by userspace. This means that care needs to be taken to
6311 * ensure that reads are stable, as we cannot rely on userspace always
6312 * being a good citizen. If members of the sqe are validated and then later
6313 * used, it's important that those reads are done through READ_ONCE() to
6314 * prevent a re-load down the line.
6316 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6318 u32 *sq_array = ctx->sq_array;
6322 * The cached sq head (or cq tail) serves two purposes:
6324 * 1) allows us to batch the cost of updating the user visible
6326 * 2) allows the kernel side to track the head on its own, even
6327 * though the application is the one updating it.
6329 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6330 if (likely(head < ctx->sq_entries))
6331 return &ctx->sq_sqes[head];
6333 /* drop invalid entries */
6334 ctx->cached_sq_dropped++;
6335 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6339 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6341 ctx->cached_sq_head++;
6345 * Check SQE restrictions (opcode and flags).
6347 * Returns 'true' if SQE is allowed, 'false' otherwise.
6349 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6350 struct io_kiocb *req,
6351 unsigned int sqe_flags)
6353 if (!ctx->restricted)
6356 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6359 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6360 ctx->restrictions.sqe_flags_required)
6363 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6364 ctx->restrictions.sqe_flags_required))
6370 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6371 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6372 IOSQE_BUFFER_SELECT)
6374 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6375 const struct io_uring_sqe *sqe,
6376 struct io_submit_state *state)
6378 unsigned int sqe_flags;
6381 req->opcode = READ_ONCE(sqe->opcode);
6382 req->user_data = READ_ONCE(sqe->user_data);
6383 req->async_data = NULL;
6387 /* one is dropped after submission, the other at completion */
6388 refcount_set(&req->refs, 2);
6389 req->task = current;
6392 if (unlikely(req->opcode >= IORING_OP_LAST))
6395 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6398 sqe_flags = READ_ONCE(sqe->flags);
6399 /* enforce forwards compatibility on users */
6400 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6403 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6406 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6407 !io_op_defs[req->opcode].buffer_select)
6410 id = READ_ONCE(sqe->personality);
6412 io_req_init_async(req);
6413 req->work.creds = idr_find(&ctx->personality_idr, id);
6414 if (unlikely(!req->work.creds))
6416 get_cred(req->work.creds);
6419 /* same numerical values with corresponding REQ_F_*, safe to copy */
6420 req->flags |= sqe_flags;
6422 if (!io_op_defs[req->opcode].needs_file)
6425 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6430 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6432 struct io_submit_state state;
6433 struct io_kiocb *link = NULL;
6434 int i, submitted = 0;
6436 /* if we have a backlog and couldn't flush it all, return BUSY */
6437 if (test_bit(0, &ctx->sq_check_overflow)) {
6438 if (!list_empty(&ctx->cq_overflow_list) &&
6439 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6443 /* make sure SQ entry isn't read before tail */
6444 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6446 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6449 atomic_long_add(nr, ¤t->io_uring->req_issue);
6450 refcount_add(nr, ¤t->usage);
6452 io_submit_state_start(&state, ctx, nr);
6454 for (i = 0; i < nr; i++) {
6455 const struct io_uring_sqe *sqe;
6456 struct io_kiocb *req;
6459 sqe = io_get_sqe(ctx);
6460 if (unlikely(!sqe)) {
6461 io_consume_sqe(ctx);
6464 req = io_alloc_req(ctx, &state);
6465 if (unlikely(!req)) {
6467 submitted = -EAGAIN;
6470 io_consume_sqe(ctx);
6471 /* will complete beyond this point, count as submitted */
6474 err = io_init_req(ctx, req, sqe, &state);
6475 if (unlikely(err)) {
6478 io_req_complete(req, err);
6482 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6483 true, io_async_submit(ctx));
6484 err = io_submit_sqe(req, sqe, &link, &state.comp);
6489 if (unlikely(submitted != nr)) {
6490 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6492 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6493 atomic_long_sub(nr - ref_used, ¤t->io_uring->req_issue);
6494 put_task_struct_many(current, nr - ref_used);
6497 io_queue_link_head(link, &state.comp);
6498 io_submit_state_end(&state);
6500 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6501 io_commit_sqring(ctx);
6506 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6508 /* Tell userspace we may need a wakeup call */
6509 spin_lock_irq(&ctx->completion_lock);
6510 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6511 spin_unlock_irq(&ctx->completion_lock);
6514 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6516 spin_lock_irq(&ctx->completion_lock);
6517 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6518 spin_unlock_irq(&ctx->completion_lock);
6521 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6522 int sync, void *key)
6524 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6527 ret = autoremove_wake_function(wqe, mode, sync, key);
6529 unsigned long flags;
6531 spin_lock_irqsave(&ctx->completion_lock, flags);
6532 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6533 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6544 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6545 unsigned long start_jiffies, bool cap_entries)
6547 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6548 struct io_sq_data *sqd = ctx->sq_data;
6549 unsigned int to_submit;
6553 if (!list_empty(&ctx->iopoll_list)) {
6554 unsigned nr_events = 0;
6556 mutex_lock(&ctx->uring_lock);
6557 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6558 io_do_iopoll(ctx, &nr_events, 0);
6559 mutex_unlock(&ctx->uring_lock);
6562 to_submit = io_sqring_entries(ctx);
6565 * If submit got -EBUSY, flag us as needing the application
6566 * to enter the kernel to reap and flush events.
6568 if (!to_submit || ret == -EBUSY || need_resched()) {
6570 * Drop cur_mm before scheduling, we can't hold it for
6571 * long periods (or over schedule()). Do this before
6572 * adding ourselves to the waitqueue, as the unuse/drop
6575 io_sq_thread_drop_mm();
6578 * We're polling. If we're within the defined idle
6579 * period, then let us spin without work before going
6580 * to sleep. The exception is if we got EBUSY doing
6581 * more IO, we should wait for the application to
6582 * reap events and wake us up.
6584 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6585 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6586 !percpu_ref_is_dying(&ctx->refs)))
6589 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6590 TASK_INTERRUPTIBLE);
6593 * While doing polled IO, before going to sleep, we need
6594 * to check if there are new reqs added to iopoll_list,
6595 * it is because reqs may have been punted to io worker
6596 * and will be added to iopoll_list later, hence check
6597 * the iopoll_list again.
6599 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6600 !list_empty_careful(&ctx->iopoll_list)) {
6601 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6605 to_submit = io_sqring_entries(ctx);
6606 if (!to_submit || ret == -EBUSY)
6610 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6611 io_ring_clear_wakeup_flag(ctx);
6613 /* if we're handling multiple rings, cap submit size for fairness */
6614 if (cap_entries && to_submit > 8)
6617 mutex_lock(&ctx->uring_lock);
6618 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6619 ret = io_submit_sqes(ctx, to_submit);
6620 mutex_unlock(&ctx->uring_lock);
6622 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6623 wake_up(&ctx->sqo_sq_wait);
6625 return SQT_DID_WORK;
6628 static void io_sqd_init_new(struct io_sq_data *sqd)
6630 struct io_ring_ctx *ctx;
6632 while (!list_empty(&sqd->ctx_new_list)) {
6633 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6634 init_wait(&ctx->sqo_wait_entry);
6635 ctx->sqo_wait_entry.func = io_sq_wake_function;
6636 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6637 complete(&ctx->sq_thread_comp);
6641 static int io_sq_thread(void *data)
6643 struct cgroup_subsys_state *cur_css = NULL;
6644 const struct cred *old_cred = NULL;
6645 struct io_sq_data *sqd = data;
6646 struct io_ring_ctx *ctx;
6647 unsigned long start_jiffies;
6649 start_jiffies = jiffies;
6650 while (!kthread_should_stop()) {
6651 enum sq_ret ret = 0;
6655 * Any changes to the sqd lists are synchronized through the
6656 * kthread parking. This synchronizes the thread vs users,
6657 * the users are synchronized on the sqd->ctx_lock.
6659 if (kthread_should_park())
6662 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6663 io_sqd_init_new(sqd);
6665 cap_entries = !list_is_singular(&sqd->ctx_list);
6667 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6668 if (current->cred != ctx->creds) {
6670 revert_creds(old_cred);
6671 old_cred = override_creds(ctx->creds);
6673 io_sq_thread_associate_blkcg(ctx, &cur_css);
6675 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6677 io_sq_thread_drop_mm();
6680 if (ret & SQT_SPIN) {
6683 } else if (ret == SQT_IDLE) {
6684 if (kthread_should_park())
6686 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6687 io_ring_set_wakeup_flag(ctx);
6689 start_jiffies = jiffies;
6690 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6691 io_ring_clear_wakeup_flag(ctx);
6698 io_sq_thread_unassociate_blkcg();
6700 revert_creds(old_cred);
6707 struct io_wait_queue {
6708 struct wait_queue_entry wq;
6709 struct io_ring_ctx *ctx;
6711 unsigned nr_timeouts;
6714 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6716 struct io_ring_ctx *ctx = iowq->ctx;
6719 * Wake up if we have enough events, or if a timeout occurred since we
6720 * started waiting. For timeouts, we always want to return to userspace,
6721 * regardless of event count.
6723 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6724 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6727 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6728 int wake_flags, void *key)
6730 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6733 /* use noflush == true, as we can't safely rely on locking context */
6734 if (!io_should_wake(iowq, true))
6737 return autoremove_wake_function(curr, mode, wake_flags, key);
6740 static int io_run_task_work_sig(void)
6742 if (io_run_task_work())
6744 if (!signal_pending(current))
6746 if (current->jobctl & JOBCTL_TASK_WORK) {
6747 spin_lock_irq(¤t->sighand->siglock);
6748 current->jobctl &= ~JOBCTL_TASK_WORK;
6749 recalc_sigpending();
6750 spin_unlock_irq(¤t->sighand->siglock);
6757 * Wait until events become available, if we don't already have some. The
6758 * application must reap them itself, as they reside on the shared cq ring.
6760 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6761 const sigset_t __user *sig, size_t sigsz)
6763 struct io_wait_queue iowq = {
6766 .func = io_wake_function,
6767 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6770 .to_wait = min_events,
6772 struct io_rings *rings = ctx->rings;
6776 if (io_cqring_events(ctx, false) >= min_events)
6778 if (!io_run_task_work())
6783 #ifdef CONFIG_COMPAT
6784 if (in_compat_syscall())
6785 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6789 ret = set_user_sigmask(sig, sigsz);
6795 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6796 trace_io_uring_cqring_wait(ctx, min_events);
6798 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6799 TASK_INTERRUPTIBLE);
6800 /* make sure we run task_work before checking for signals */
6801 ret = io_run_task_work_sig();
6806 if (io_should_wake(&iowq, false))
6810 finish_wait(&ctx->wait, &iowq.wq);
6812 restore_saved_sigmask_unless(ret == -EINTR);
6814 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6817 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6819 #if defined(CONFIG_UNIX)
6820 if (ctx->ring_sock) {
6821 struct sock *sock = ctx->ring_sock->sk;
6822 struct sk_buff *skb;
6824 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6830 for (i = 0; i < ctx->nr_user_files; i++) {
6833 file = io_file_from_index(ctx, i);
6840 static void io_file_ref_kill(struct percpu_ref *ref)
6842 struct fixed_file_data *data;
6844 data = container_of(ref, struct fixed_file_data, refs);
6845 complete(&data->done);
6848 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6850 struct fixed_file_data *data = ctx->file_data;
6851 struct fixed_file_ref_node *ref_node = NULL;
6852 unsigned nr_tables, i;
6857 spin_lock(&data->lock);
6858 if (!list_empty(&data->ref_list))
6859 ref_node = list_first_entry(&data->ref_list,
6860 struct fixed_file_ref_node, node);
6861 spin_unlock(&data->lock);
6863 percpu_ref_kill(&ref_node->refs);
6865 percpu_ref_kill(&data->refs);
6867 /* wait for all refs nodes to complete */
6868 flush_delayed_work(&ctx->file_put_work);
6869 wait_for_completion(&data->done);
6871 __io_sqe_files_unregister(ctx);
6872 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6873 for (i = 0; i < nr_tables; i++)
6874 kfree(data->table[i].files);
6876 percpu_ref_exit(&data->refs);
6878 ctx->file_data = NULL;
6879 ctx->nr_user_files = 0;
6883 static void io_put_sq_data(struct io_sq_data *sqd)
6885 if (refcount_dec_and_test(&sqd->refs)) {
6887 * The park is a bit of a work-around, without it we get
6888 * warning spews on shutdown with SQPOLL set and affinity
6889 * set to a single CPU.
6892 kthread_park(sqd->thread);
6893 kthread_stop(sqd->thread);
6900 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
6902 struct io_ring_ctx *ctx_attach;
6903 struct io_sq_data *sqd;
6906 f = fdget(p->wq_fd);
6908 return ERR_PTR(-ENXIO);
6909 if (f.file->f_op != &io_uring_fops) {
6911 return ERR_PTR(-EINVAL);
6914 ctx_attach = f.file->private_data;
6915 sqd = ctx_attach->sq_data;
6918 return ERR_PTR(-EINVAL);
6921 refcount_inc(&sqd->refs);
6926 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
6928 struct io_sq_data *sqd;
6930 if (p->flags & IORING_SETUP_ATTACH_WQ)
6931 return io_attach_sq_data(p);
6933 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
6935 return ERR_PTR(-ENOMEM);
6937 refcount_set(&sqd->refs, 1);
6938 INIT_LIST_HEAD(&sqd->ctx_list);
6939 INIT_LIST_HEAD(&sqd->ctx_new_list);
6940 mutex_init(&sqd->ctx_lock);
6941 mutex_init(&sqd->lock);
6942 init_waitqueue_head(&sqd->wait);
6946 static void io_sq_thread_unpark(struct io_sq_data *sqd)
6947 __releases(&sqd->lock)
6951 kthread_unpark(sqd->thread);
6952 mutex_unlock(&sqd->lock);
6955 static void io_sq_thread_park(struct io_sq_data *sqd)
6956 __acquires(&sqd->lock)
6960 mutex_lock(&sqd->lock);
6961 kthread_park(sqd->thread);
6964 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6966 struct io_sq_data *sqd = ctx->sq_data;
6971 * We may arrive here from the error branch in
6972 * io_sq_offload_create() where the kthread is created
6973 * without being waked up, thus wake it up now to make
6974 * sure the wait will complete.
6976 wake_up_process(sqd->thread);
6977 wait_for_completion(&ctx->sq_thread_comp);
6979 io_sq_thread_park(sqd);
6982 mutex_lock(&sqd->ctx_lock);
6983 list_del(&ctx->sqd_list);
6984 mutex_unlock(&sqd->ctx_lock);
6987 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6988 io_sq_thread_unpark(sqd);
6991 io_put_sq_data(sqd);
6992 ctx->sq_data = NULL;
6996 static void io_finish_async(struct io_ring_ctx *ctx)
6998 io_sq_thread_stop(ctx);
7001 io_wq_destroy(ctx->io_wq);
7006 #if defined(CONFIG_UNIX)
7008 * Ensure the UNIX gc is aware of our file set, so we are certain that
7009 * the io_uring can be safely unregistered on process exit, even if we have
7010 * loops in the file referencing.
7012 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7014 struct sock *sk = ctx->ring_sock->sk;
7015 struct scm_fp_list *fpl;
7016 struct sk_buff *skb;
7019 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7023 skb = alloc_skb(0, GFP_KERNEL);
7032 fpl->user = get_uid(ctx->user);
7033 for (i = 0; i < nr; i++) {
7034 struct file *file = io_file_from_index(ctx, i + offset);
7038 fpl->fp[nr_files] = get_file(file);
7039 unix_inflight(fpl->user, fpl->fp[nr_files]);
7044 fpl->max = SCM_MAX_FD;
7045 fpl->count = nr_files;
7046 UNIXCB(skb).fp = fpl;
7047 skb->destructor = unix_destruct_scm;
7048 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7049 skb_queue_head(&sk->sk_receive_queue, skb);
7051 for (i = 0; i < nr_files; i++)
7062 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7063 * causes regular reference counting to break down. We rely on the UNIX
7064 * garbage collection to take care of this problem for us.
7066 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7068 unsigned left, total;
7072 left = ctx->nr_user_files;
7074 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7076 ret = __io_sqe_files_scm(ctx, this_files, total);
7080 total += this_files;
7086 while (total < ctx->nr_user_files) {
7087 struct file *file = io_file_from_index(ctx, total);
7097 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7103 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7104 unsigned nr_tables, unsigned nr_files)
7108 for (i = 0; i < nr_tables; i++) {
7109 struct fixed_file_table *table = &file_data->table[i];
7110 unsigned this_files;
7112 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7113 table->files = kcalloc(this_files, sizeof(struct file *),
7117 nr_files -= this_files;
7123 for (i = 0; i < nr_tables; i++) {
7124 struct fixed_file_table *table = &file_data->table[i];
7125 kfree(table->files);
7130 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7132 #if defined(CONFIG_UNIX)
7133 struct sock *sock = ctx->ring_sock->sk;
7134 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7135 struct sk_buff *skb;
7138 __skb_queue_head_init(&list);
7141 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7142 * remove this entry and rearrange the file array.
7144 skb = skb_dequeue(head);
7146 struct scm_fp_list *fp;
7148 fp = UNIXCB(skb).fp;
7149 for (i = 0; i < fp->count; i++) {
7152 if (fp->fp[i] != file)
7155 unix_notinflight(fp->user, fp->fp[i]);
7156 left = fp->count - 1 - i;
7158 memmove(&fp->fp[i], &fp->fp[i + 1],
7159 left * sizeof(struct file *));
7166 __skb_queue_tail(&list, skb);
7176 __skb_queue_tail(&list, skb);
7178 skb = skb_dequeue(head);
7181 if (skb_peek(&list)) {
7182 spin_lock_irq(&head->lock);
7183 while ((skb = __skb_dequeue(&list)) != NULL)
7184 __skb_queue_tail(head, skb);
7185 spin_unlock_irq(&head->lock);
7192 struct io_file_put {
7193 struct list_head list;
7197 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7199 struct fixed_file_data *file_data = ref_node->file_data;
7200 struct io_ring_ctx *ctx = file_data->ctx;
7201 struct io_file_put *pfile, *tmp;
7203 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7204 list_del(&pfile->list);
7205 io_ring_file_put(ctx, pfile->file);
7209 spin_lock(&file_data->lock);
7210 list_del(&ref_node->node);
7211 spin_unlock(&file_data->lock);
7213 percpu_ref_exit(&ref_node->refs);
7215 percpu_ref_put(&file_data->refs);
7218 static void io_file_put_work(struct work_struct *work)
7220 struct io_ring_ctx *ctx;
7221 struct llist_node *node;
7223 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7224 node = llist_del_all(&ctx->file_put_llist);
7227 struct fixed_file_ref_node *ref_node;
7228 struct llist_node *next = node->next;
7230 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7231 __io_file_put_work(ref_node);
7236 static void io_file_data_ref_zero(struct percpu_ref *ref)
7238 struct fixed_file_ref_node *ref_node;
7239 struct io_ring_ctx *ctx;
7243 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7244 ctx = ref_node->file_data->ctx;
7246 if (percpu_ref_is_dying(&ctx->file_data->refs))
7249 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7251 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7253 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7256 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7257 struct io_ring_ctx *ctx)
7259 struct fixed_file_ref_node *ref_node;
7261 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7263 return ERR_PTR(-ENOMEM);
7265 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7268 return ERR_PTR(-ENOMEM);
7270 INIT_LIST_HEAD(&ref_node->node);
7271 INIT_LIST_HEAD(&ref_node->file_list);
7272 ref_node->file_data = ctx->file_data;
7276 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7278 percpu_ref_exit(&ref_node->refs);
7282 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7285 __s32 __user *fds = (__s32 __user *) arg;
7286 unsigned nr_tables, i;
7288 int fd, ret = -ENOMEM;
7289 struct fixed_file_ref_node *ref_node;
7290 struct fixed_file_data *file_data;
7296 if (nr_args > IORING_MAX_FIXED_FILES)
7299 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7302 file_data->ctx = ctx;
7303 init_completion(&file_data->done);
7304 INIT_LIST_HEAD(&file_data->ref_list);
7305 spin_lock_init(&file_data->lock);
7307 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7308 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7310 if (!file_data->table)
7313 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7314 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7317 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7320 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7321 struct fixed_file_table *table;
7324 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7328 /* allow sparse sets */
7338 * Don't allow io_uring instances to be registered. If UNIX
7339 * isn't enabled, then this causes a reference cycle and this
7340 * instance can never get freed. If UNIX is enabled we'll
7341 * handle it just fine, but there's still no point in allowing
7342 * a ring fd as it doesn't support regular read/write anyway.
7344 if (file->f_op == &io_uring_fops) {
7348 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7349 index = i & IORING_FILE_TABLE_MASK;
7350 table->files[index] = file;
7353 ctx->file_data = file_data;
7354 ret = io_sqe_files_scm(ctx);
7356 io_sqe_files_unregister(ctx);
7360 ref_node = alloc_fixed_file_ref_node(ctx);
7361 if (IS_ERR(ref_node)) {
7362 io_sqe_files_unregister(ctx);
7363 return PTR_ERR(ref_node);
7366 file_data->node = ref_node;
7367 spin_lock(&file_data->lock);
7368 list_add(&ref_node->node, &file_data->ref_list);
7369 spin_unlock(&file_data->lock);
7370 percpu_ref_get(&file_data->refs);
7373 for (i = 0; i < ctx->nr_user_files; i++) {
7374 file = io_file_from_index(ctx, i);
7378 for (i = 0; i < nr_tables; i++)
7379 kfree(file_data->table[i].files);
7380 ctx->nr_user_files = 0;
7382 percpu_ref_exit(&file_data->refs);
7384 kfree(file_data->table);
7389 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7392 #if defined(CONFIG_UNIX)
7393 struct sock *sock = ctx->ring_sock->sk;
7394 struct sk_buff_head *head = &sock->sk_receive_queue;
7395 struct sk_buff *skb;
7398 * See if we can merge this file into an existing skb SCM_RIGHTS
7399 * file set. If there's no room, fall back to allocating a new skb
7400 * and filling it in.
7402 spin_lock_irq(&head->lock);
7403 skb = skb_peek(head);
7405 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7407 if (fpl->count < SCM_MAX_FD) {
7408 __skb_unlink(skb, head);
7409 spin_unlock_irq(&head->lock);
7410 fpl->fp[fpl->count] = get_file(file);
7411 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7413 spin_lock_irq(&head->lock);
7414 __skb_queue_head(head, skb);
7419 spin_unlock_irq(&head->lock);
7426 return __io_sqe_files_scm(ctx, 1, index);
7432 static int io_queue_file_removal(struct fixed_file_data *data,
7435 struct io_file_put *pfile;
7436 struct fixed_file_ref_node *ref_node = data->node;
7438 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7443 list_add(&pfile->list, &ref_node->file_list);
7448 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7449 struct io_uring_files_update *up,
7452 struct fixed_file_data *data = ctx->file_data;
7453 struct fixed_file_ref_node *ref_node;
7458 bool needs_switch = false;
7460 if (check_add_overflow(up->offset, nr_args, &done))
7462 if (done > ctx->nr_user_files)
7465 ref_node = alloc_fixed_file_ref_node(ctx);
7466 if (IS_ERR(ref_node))
7467 return PTR_ERR(ref_node);
7470 fds = u64_to_user_ptr(up->fds);
7472 struct fixed_file_table *table;
7476 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7480 i = array_index_nospec(up->offset, ctx->nr_user_files);
7481 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7482 index = i & IORING_FILE_TABLE_MASK;
7483 if (table->files[index]) {
7484 file = table->files[index];
7485 err = io_queue_file_removal(data, file);
7488 table->files[index] = NULL;
7489 needs_switch = true;
7498 * Don't allow io_uring instances to be registered. If
7499 * UNIX isn't enabled, then this causes a reference
7500 * cycle and this instance can never get freed. If UNIX
7501 * is enabled we'll handle it just fine, but there's
7502 * still no point in allowing a ring fd as it doesn't
7503 * support regular read/write anyway.
7505 if (file->f_op == &io_uring_fops) {
7510 table->files[index] = file;
7511 err = io_sqe_file_register(ctx, file, i);
7513 table->files[index] = NULL;
7524 percpu_ref_kill(&data->node->refs);
7525 spin_lock(&data->lock);
7526 list_add(&ref_node->node, &data->ref_list);
7527 data->node = ref_node;
7528 spin_unlock(&data->lock);
7529 percpu_ref_get(&ctx->file_data->refs);
7531 destroy_fixed_file_ref_node(ref_node);
7533 return done ? done : err;
7536 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7539 struct io_uring_files_update up;
7541 if (!ctx->file_data)
7545 if (copy_from_user(&up, arg, sizeof(up)))
7550 return __io_sqe_files_update(ctx, &up, nr_args);
7553 static void io_free_work(struct io_wq_work *work)
7555 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7557 /* Consider that io_steal_work() relies on this ref */
7561 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7562 struct io_uring_params *p)
7564 struct io_wq_data data;
7566 struct io_ring_ctx *ctx_attach;
7567 unsigned int concurrency;
7570 data.user = ctx->user;
7571 data.free_work = io_free_work;
7572 data.do_work = io_wq_submit_work;
7574 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7575 /* Do QD, or 4 * CPUS, whatever is smallest */
7576 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7578 ctx->io_wq = io_wq_create(concurrency, &data);
7579 if (IS_ERR(ctx->io_wq)) {
7580 ret = PTR_ERR(ctx->io_wq);
7586 f = fdget(p->wq_fd);
7590 if (f.file->f_op != &io_uring_fops) {
7595 ctx_attach = f.file->private_data;
7596 /* @io_wq is protected by holding the fd */
7597 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7602 ctx->io_wq = ctx_attach->io_wq;
7608 static int io_uring_alloc_task_context(struct task_struct *task)
7610 struct io_uring_task *tctx;
7612 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7613 if (unlikely(!tctx))
7617 init_waitqueue_head(&tctx->wait);
7620 atomic_long_set(&tctx->req_issue, 0);
7621 atomic_long_set(&tctx->req_complete, 0);
7622 task->io_uring = tctx;
7626 void __io_uring_free(struct task_struct *tsk)
7628 struct io_uring_task *tctx = tsk->io_uring;
7630 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7632 tsk->io_uring = NULL;
7635 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7636 struct io_uring_params *p)
7640 if (ctx->flags & IORING_SETUP_SQPOLL) {
7641 struct io_sq_data *sqd;
7644 if (!capable(CAP_SYS_ADMIN))
7647 sqd = io_get_sq_data(p);
7654 io_sq_thread_park(sqd);
7655 mutex_lock(&sqd->ctx_lock);
7656 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7657 mutex_unlock(&sqd->ctx_lock);
7658 io_sq_thread_unpark(sqd);
7660 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7661 if (!ctx->sq_thread_idle)
7662 ctx->sq_thread_idle = HZ;
7667 if (p->flags & IORING_SETUP_SQ_AFF) {
7668 int cpu = p->sq_thread_cpu;
7671 if (cpu >= nr_cpu_ids)
7673 if (!cpu_online(cpu))
7676 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7677 cpu, "io_uring-sq");
7679 sqd->thread = kthread_create(io_sq_thread, sqd,
7682 if (IS_ERR(sqd->thread)) {
7683 ret = PTR_ERR(sqd->thread);
7687 ret = io_uring_alloc_task_context(sqd->thread);
7690 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7691 /* Can't have SQ_AFF without SQPOLL */
7697 ret = io_init_wq_offload(ctx, p);
7703 io_finish_async(ctx);
7707 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7709 struct io_sq_data *sqd = ctx->sq_data;
7711 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7712 wake_up_process(sqd->thread);
7715 static inline void __io_unaccount_mem(struct user_struct *user,
7716 unsigned long nr_pages)
7718 atomic_long_sub(nr_pages, &user->locked_vm);
7721 static inline int __io_account_mem(struct user_struct *user,
7722 unsigned long nr_pages)
7724 unsigned long page_limit, cur_pages, new_pages;
7726 /* Don't allow more pages than we can safely lock */
7727 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7730 cur_pages = atomic_long_read(&user->locked_vm);
7731 new_pages = cur_pages + nr_pages;
7732 if (new_pages > page_limit)
7734 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7735 new_pages) != cur_pages);
7740 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7741 enum io_mem_account acct)
7744 __io_unaccount_mem(ctx->user, nr_pages);
7746 if (ctx->mm_account) {
7747 if (acct == ACCT_LOCKED)
7748 ctx->mm_account->locked_vm -= nr_pages;
7749 else if (acct == ACCT_PINNED)
7750 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7754 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7755 enum io_mem_account acct)
7759 if (ctx->limit_mem) {
7760 ret = __io_account_mem(ctx->user, nr_pages);
7765 if (ctx->mm_account) {
7766 if (acct == ACCT_LOCKED)
7767 ctx->mm_account->locked_vm += nr_pages;
7768 else if (acct == ACCT_PINNED)
7769 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7775 static void io_mem_free(void *ptr)
7782 page = virt_to_head_page(ptr);
7783 if (put_page_testzero(page))
7784 free_compound_page(page);
7787 static void *io_mem_alloc(size_t size)
7789 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7792 return (void *) __get_free_pages(gfp_flags, get_order(size));
7795 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7798 struct io_rings *rings;
7799 size_t off, sq_array_size;
7801 off = struct_size(rings, cqes, cq_entries);
7802 if (off == SIZE_MAX)
7806 off = ALIGN(off, SMP_CACHE_BYTES);
7814 sq_array_size = array_size(sizeof(u32), sq_entries);
7815 if (sq_array_size == SIZE_MAX)
7818 if (check_add_overflow(off, sq_array_size, &off))
7824 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7828 pages = (size_t)1 << get_order(
7829 rings_size(sq_entries, cq_entries, NULL));
7830 pages += (size_t)1 << get_order(
7831 array_size(sizeof(struct io_uring_sqe), sq_entries));
7836 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7840 if (!ctx->user_bufs)
7843 for (i = 0; i < ctx->nr_user_bufs; i++) {
7844 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7846 for (j = 0; j < imu->nr_bvecs; j++)
7847 unpin_user_page(imu->bvec[j].bv_page);
7849 if (imu->acct_pages)
7850 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7855 kfree(ctx->user_bufs);
7856 ctx->user_bufs = NULL;
7857 ctx->nr_user_bufs = 0;
7861 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7862 void __user *arg, unsigned index)
7864 struct iovec __user *src;
7866 #ifdef CONFIG_COMPAT
7868 struct compat_iovec __user *ciovs;
7869 struct compat_iovec ciov;
7871 ciovs = (struct compat_iovec __user *) arg;
7872 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7875 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7876 dst->iov_len = ciov.iov_len;
7880 src = (struct iovec __user *) arg;
7881 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7887 * Not super efficient, but this is just a registration time. And we do cache
7888 * the last compound head, so generally we'll only do a full search if we don't
7891 * We check if the given compound head page has already been accounted, to
7892 * avoid double accounting it. This allows us to account the full size of the
7893 * page, not just the constituent pages of a huge page.
7895 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
7896 int nr_pages, struct page *hpage)
7900 /* check current page array */
7901 for (i = 0; i < nr_pages; i++) {
7902 if (!PageCompound(pages[i]))
7904 if (compound_head(pages[i]) == hpage)
7908 /* check previously registered pages */
7909 for (i = 0; i < ctx->nr_user_bufs; i++) {
7910 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7912 for (j = 0; j < imu->nr_bvecs; j++) {
7913 if (!PageCompound(imu->bvec[j].bv_page))
7915 if (compound_head(imu->bvec[j].bv_page) == hpage)
7923 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
7924 int nr_pages, struct io_mapped_ubuf *imu,
7925 struct page **last_hpage)
7929 for (i = 0; i < nr_pages; i++) {
7930 if (!PageCompound(pages[i])) {
7935 hpage = compound_head(pages[i]);
7936 if (hpage == *last_hpage)
7938 *last_hpage = hpage;
7939 if (headpage_already_acct(ctx, pages, i, hpage))
7941 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
7945 if (!imu->acct_pages)
7948 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
7950 imu->acct_pages = 0;
7954 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7957 struct vm_area_struct **vmas = NULL;
7958 struct page **pages = NULL;
7959 struct page *last_hpage = NULL;
7960 int i, j, got_pages = 0;
7965 if (!nr_args || nr_args > UIO_MAXIOV)
7968 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7970 if (!ctx->user_bufs)
7973 for (i = 0; i < nr_args; i++) {
7974 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7975 unsigned long off, start, end, ubuf;
7980 ret = io_copy_iov(ctx, &iov, arg, i);
7985 * Don't impose further limits on the size and buffer
7986 * constraints here, we'll -EINVAL later when IO is
7987 * submitted if they are wrong.
7990 if (!iov.iov_base || !iov.iov_len)
7993 /* arbitrary limit, but we need something */
7994 if (iov.iov_len > SZ_1G)
7997 ubuf = (unsigned long) iov.iov_base;
7998 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7999 start = ubuf >> PAGE_SHIFT;
8000 nr_pages = end - start;
8003 if (!pages || nr_pages > got_pages) {
8006 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8008 vmas = kvmalloc_array(nr_pages,
8009 sizeof(struct vm_area_struct *),
8011 if (!pages || !vmas) {
8015 got_pages = nr_pages;
8018 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8025 mmap_read_lock(current->mm);
8026 pret = pin_user_pages(ubuf, nr_pages,
8027 FOLL_WRITE | FOLL_LONGTERM,
8029 if (pret == nr_pages) {
8030 /* don't support file backed memory */
8031 for (j = 0; j < nr_pages; j++) {
8032 struct vm_area_struct *vma = vmas[j];
8035 !is_file_hugepages(vma->vm_file)) {
8041 ret = pret < 0 ? pret : -EFAULT;
8043 mmap_read_unlock(current->mm);
8046 * if we did partial map, or found file backed vmas,
8047 * release any pages we did get
8050 unpin_user_pages(pages, pret);
8055 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8057 unpin_user_pages(pages, pret);
8062 off = ubuf & ~PAGE_MASK;
8064 for (j = 0; j < nr_pages; j++) {
8067 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8068 imu->bvec[j].bv_page = pages[j];
8069 imu->bvec[j].bv_len = vec_len;
8070 imu->bvec[j].bv_offset = off;
8074 /* store original address for later verification */
8076 imu->len = iov.iov_len;
8077 imu->nr_bvecs = nr_pages;
8079 ctx->nr_user_bufs++;
8087 io_sqe_buffer_unregister(ctx);
8091 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8093 __s32 __user *fds = arg;
8099 if (copy_from_user(&fd, fds, sizeof(*fds)))
8102 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8103 if (IS_ERR(ctx->cq_ev_fd)) {
8104 int ret = PTR_ERR(ctx->cq_ev_fd);
8105 ctx->cq_ev_fd = NULL;
8112 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8114 if (ctx->cq_ev_fd) {
8115 eventfd_ctx_put(ctx->cq_ev_fd);
8116 ctx->cq_ev_fd = NULL;
8123 static int __io_destroy_buffers(int id, void *p, void *data)
8125 struct io_ring_ctx *ctx = data;
8126 struct io_buffer *buf = p;
8128 __io_remove_buffers(ctx, buf, id, -1U);
8132 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8134 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8135 idr_destroy(&ctx->io_buffer_idr);
8138 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8140 io_finish_async(ctx);
8141 io_sqe_buffer_unregister(ctx);
8143 if (ctx->sqo_task) {
8144 put_task_struct(ctx->sqo_task);
8145 ctx->sqo_task = NULL;
8146 mmdrop(ctx->mm_account);
8147 ctx->mm_account = NULL;
8150 #ifdef CONFIG_BLK_CGROUP
8151 if (ctx->sqo_blkcg_css)
8152 css_put(ctx->sqo_blkcg_css);
8155 io_sqe_files_unregister(ctx);
8156 io_eventfd_unregister(ctx);
8157 io_destroy_buffers(ctx);
8158 idr_destroy(&ctx->personality_idr);
8160 #if defined(CONFIG_UNIX)
8161 if (ctx->ring_sock) {
8162 ctx->ring_sock->file = NULL; /* so that iput() is called */
8163 sock_release(ctx->ring_sock);
8167 io_mem_free(ctx->rings);
8168 io_mem_free(ctx->sq_sqes);
8170 percpu_ref_exit(&ctx->refs);
8171 free_uid(ctx->user);
8172 put_cred(ctx->creds);
8173 kfree(ctx->cancel_hash);
8174 kmem_cache_free(req_cachep, ctx->fallback_req);
8178 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8180 struct io_ring_ctx *ctx = file->private_data;
8183 poll_wait(file, &ctx->cq_wait, wait);
8185 * synchronizes with barrier from wq_has_sleeper call in
8189 if (!io_sqring_full(ctx))
8190 mask |= EPOLLOUT | EPOLLWRNORM;
8191 if (io_cqring_events(ctx, false))
8192 mask |= EPOLLIN | EPOLLRDNORM;
8197 static int io_uring_fasync(int fd, struct file *file, int on)
8199 struct io_ring_ctx *ctx = file->private_data;
8201 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8204 static int io_remove_personalities(int id, void *p, void *data)
8206 struct io_ring_ctx *ctx = data;
8207 const struct cred *cred;
8209 cred = idr_remove(&ctx->personality_idr, id);
8215 static void io_ring_exit_work(struct work_struct *work)
8217 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8221 * If we're doing polled IO and end up having requests being
8222 * submitted async (out-of-line), then completions can come in while
8223 * we're waiting for refs to drop. We need to reap these manually,
8224 * as nobody else will be looking for them.
8228 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8229 io_iopoll_try_reap_events(ctx);
8230 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8231 io_ring_ctx_free(ctx);
8234 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8236 mutex_lock(&ctx->uring_lock);
8237 percpu_ref_kill(&ctx->refs);
8238 mutex_unlock(&ctx->uring_lock);
8240 io_kill_timeouts(ctx, NULL);
8241 io_poll_remove_all(ctx, NULL);
8244 io_wq_cancel_all(ctx->io_wq);
8246 /* if we failed setting up the ctx, we might not have any rings */
8248 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8249 io_iopoll_try_reap_events(ctx);
8250 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8253 * Do this upfront, so we won't have a grace period where the ring
8254 * is closed but resources aren't reaped yet. This can cause
8255 * spurious failure in setting up a new ring.
8257 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8260 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8262 * Use system_unbound_wq to avoid spawning tons of event kworkers
8263 * if we're exiting a ton of rings at the same time. It just adds
8264 * noise and overhead, there's no discernable change in runtime
8265 * over using system_wq.
8267 queue_work(system_unbound_wq, &ctx->exit_work);
8270 static int io_uring_release(struct inode *inode, struct file *file)
8272 struct io_ring_ctx *ctx = file->private_data;
8274 file->private_data = NULL;
8275 io_ring_ctx_wait_and_kill(ctx);
8279 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8281 struct files_struct *files = data;
8283 return !files || work->files == files;
8287 * Returns true if 'preq' is the link parent of 'req'
8289 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8291 struct io_kiocb *link;
8293 if (!(preq->flags & REQ_F_LINK_HEAD))
8296 list_for_each_entry(link, &preq->link_list, link_list) {
8304 static bool io_match_link_files(struct io_kiocb *req,
8305 struct files_struct *files)
8307 struct io_kiocb *link;
8309 if (io_match_files(req, files))
8311 if (req->flags & REQ_F_LINK_HEAD) {
8312 list_for_each_entry(link, &req->link_list, link_list) {
8313 if (io_match_files(link, files))
8321 * We're looking to cancel 'req' because it's holding on to our files, but
8322 * 'req' could be a link to another request. See if it is, and cancel that
8323 * parent request if so.
8325 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8327 struct hlist_node *tmp;
8328 struct io_kiocb *preq;
8332 spin_lock_irq(&ctx->completion_lock);
8333 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8334 struct hlist_head *list;
8336 list = &ctx->cancel_hash[i];
8337 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8338 found = io_match_link(preq, req);
8340 io_poll_remove_one(preq);
8345 spin_unlock_irq(&ctx->completion_lock);
8349 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8350 struct io_kiocb *req)
8352 struct io_kiocb *preq;
8355 spin_lock_irq(&ctx->completion_lock);
8356 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8357 found = io_match_link(preq, req);
8359 __io_timeout_cancel(preq);
8363 spin_unlock_irq(&ctx->completion_lock);
8367 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8369 return io_match_link(container_of(work, struct io_kiocb, work), data);
8372 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8374 enum io_wq_cancel cret;
8376 /* cancel this particular work, if it's running */
8377 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8378 if (cret != IO_WQ_CANCEL_NOTFOUND)
8381 /* find links that hold this pending, cancel those */
8382 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8383 if (cret != IO_WQ_CANCEL_NOTFOUND)
8386 /* if we have a poll link holding this pending, cancel that */
8387 if (io_poll_remove_link(ctx, req))
8390 /* final option, timeout link is holding this req pending */
8391 io_timeout_remove_link(ctx, req);
8394 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8395 struct files_struct *files)
8397 struct io_defer_entry *de = NULL;
8400 spin_lock_irq(&ctx->completion_lock);
8401 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8402 if (io_match_link_files(de->req, files)) {
8403 list_cut_position(&list, &ctx->defer_list, &de->list);
8407 spin_unlock_irq(&ctx->completion_lock);
8409 while (!list_empty(&list)) {
8410 de = list_first_entry(&list, struct io_defer_entry, list);
8411 list_del_init(&de->list);
8412 req_set_fail_links(de->req);
8413 io_put_req(de->req);
8414 io_req_complete(de->req, -ECANCELED);
8420 * Returns true if we found and killed one or more files pinning requests
8422 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8423 struct files_struct *files)
8425 if (list_empty_careful(&ctx->inflight_list))
8428 io_cancel_defer_files(ctx, files);
8429 /* cancel all at once, should be faster than doing it one by one*/
8430 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8432 while (!list_empty_careful(&ctx->inflight_list)) {
8433 struct io_kiocb *cancel_req = NULL, *req;
8436 spin_lock_irq(&ctx->inflight_lock);
8437 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8438 if (files && req->work.files != files)
8440 /* req is being completed, ignore */
8441 if (!refcount_inc_not_zero(&req->refs))
8447 prepare_to_wait(&ctx->inflight_wait, &wait,
8448 TASK_UNINTERRUPTIBLE);
8449 spin_unlock_irq(&ctx->inflight_lock);
8451 /* We need to keep going until we don't find a matching req */
8454 /* cancel this request, or head link requests */
8455 io_attempt_cancel(ctx, cancel_req);
8456 io_put_req(cancel_req);
8457 /* cancellations _may_ trigger task work */
8460 finish_wait(&ctx->inflight_wait, &wait);
8466 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8468 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8469 struct task_struct *task = data;
8471 return io_task_match(req, task);
8474 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8475 struct task_struct *task,
8476 struct files_struct *files)
8480 ret = io_uring_cancel_files(ctx, files);
8482 enum io_wq_cancel cret;
8484 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8485 if (cret != IO_WQ_CANCEL_NOTFOUND)
8488 /* SQPOLL thread does its own polling */
8489 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8490 while (!list_empty_careful(&ctx->iopoll_list)) {
8491 io_iopoll_try_reap_events(ctx);
8496 ret |= io_poll_remove_all(ctx, task);
8497 ret |= io_kill_timeouts(ctx, task);
8504 * We need to iteratively cancel requests, in case a request has dependent
8505 * hard links. These persist even for failure of cancelations, hence keep
8506 * looping until none are found.
8508 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8509 struct files_struct *files)
8511 struct task_struct *task = current;
8513 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8514 task = ctx->sq_data->thread;
8516 io_cqring_overflow_flush(ctx, true, task, files);
8518 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8525 * Note that this task has used io_uring. We use it for cancelation purposes.
8527 static int io_uring_add_task_file(struct file *file)
8529 struct io_uring_task *tctx = current->io_uring;
8531 if (unlikely(!tctx)) {
8534 ret = io_uring_alloc_task_context(current);
8537 tctx = current->io_uring;
8539 if (tctx->last != file) {
8540 void *old = xa_load(&tctx->xa, (unsigned long)file);
8544 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8553 * Remove this io_uring_file -> task mapping.
8555 static void io_uring_del_task_file(struct file *file)
8557 struct io_uring_task *tctx = current->io_uring;
8559 if (tctx->last == file)
8561 file = xa_erase(&tctx->xa, (unsigned long)file);
8566 static void __io_uring_attempt_task_drop(struct file *file)
8568 struct file *old = xa_load(¤t->io_uring->xa, (unsigned long)file);
8571 io_uring_del_task_file(file);
8575 * Drop task note for this file if we're the only ones that hold it after
8578 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8580 if (!current->io_uring)
8583 * fput() is pending, will be 2 if the only other ref is our potential
8584 * task file note. If the task is exiting, drop regardless of count.
8586 if (!exiting && atomic_long_read(&file->f_count) != 2)
8589 __io_uring_attempt_task_drop(file);
8592 void __io_uring_files_cancel(struct files_struct *files)
8594 struct io_uring_task *tctx = current->io_uring;
8596 unsigned long index;
8598 /* make sure overflow events are dropped */
8599 tctx->in_idle = true;
8601 xa_for_each(&tctx->xa, index, file) {
8602 struct io_ring_ctx *ctx = file->private_data;
8604 io_uring_cancel_task_requests(ctx, files);
8606 io_uring_del_task_file(file);
8610 static inline bool io_uring_task_idle(struct io_uring_task *tctx)
8612 return atomic_long_read(&tctx->req_issue) ==
8613 atomic_long_read(&tctx->req_complete);
8617 * Find any io_uring fd that this task has registered or done IO on, and cancel
8620 void __io_uring_task_cancel(void)
8622 struct io_uring_task *tctx = current->io_uring;
8626 /* make sure overflow events are dropped */
8627 tctx->in_idle = true;
8629 while (!io_uring_task_idle(tctx)) {
8630 /* read completions before cancelations */
8631 completions = atomic_long_read(&tctx->req_complete);
8632 __io_uring_files_cancel(NULL);
8634 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8637 * If we've seen completions, retry. This avoids a race where
8638 * a completion comes in before we did prepare_to_wait().
8640 if (completions != atomic_long_read(&tctx->req_complete))
8642 if (io_uring_task_idle(tctx))
8647 finish_wait(&tctx->wait, &wait);
8648 tctx->in_idle = false;
8651 static int io_uring_flush(struct file *file, void *data)
8653 struct io_ring_ctx *ctx = file->private_data;
8656 * If the task is going away, cancel work it may have pending
8658 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8661 io_uring_cancel_task_requests(ctx, data);
8662 io_uring_attempt_task_drop(file, !data);
8666 static void *io_uring_validate_mmap_request(struct file *file,
8667 loff_t pgoff, size_t sz)
8669 struct io_ring_ctx *ctx = file->private_data;
8670 loff_t offset = pgoff << PAGE_SHIFT;
8675 case IORING_OFF_SQ_RING:
8676 case IORING_OFF_CQ_RING:
8679 case IORING_OFF_SQES:
8683 return ERR_PTR(-EINVAL);
8686 page = virt_to_head_page(ptr);
8687 if (sz > page_size(page))
8688 return ERR_PTR(-EINVAL);
8695 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8697 size_t sz = vma->vm_end - vma->vm_start;
8701 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8703 return PTR_ERR(ptr);
8705 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8706 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8709 #else /* !CONFIG_MMU */
8711 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8713 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8716 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8718 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8721 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8722 unsigned long addr, unsigned long len,
8723 unsigned long pgoff, unsigned long flags)
8727 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8729 return PTR_ERR(ptr);
8731 return (unsigned long) ptr;
8734 #endif /* !CONFIG_MMU */
8736 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8741 if (!io_sqring_full(ctx))
8744 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8746 if (!io_sqring_full(ctx))
8750 } while (!signal_pending(current));
8752 finish_wait(&ctx->sqo_sq_wait, &wait);
8755 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8756 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8759 struct io_ring_ctx *ctx;
8766 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8767 IORING_ENTER_SQ_WAIT))
8775 if (f.file->f_op != &io_uring_fops)
8779 ctx = f.file->private_data;
8780 if (!percpu_ref_tryget(&ctx->refs))
8784 if (ctx->flags & IORING_SETUP_R_DISABLED)
8788 * For SQ polling, the thread will do all submissions and completions.
8789 * Just return the requested submit count, and wake the thread if
8793 if (ctx->flags & IORING_SETUP_SQPOLL) {
8794 if (!list_empty_careful(&ctx->cq_overflow_list))
8795 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8796 if (flags & IORING_ENTER_SQ_WAKEUP)
8797 wake_up(&ctx->sq_data->wait);
8798 if (flags & IORING_ENTER_SQ_WAIT)
8799 io_sqpoll_wait_sq(ctx);
8800 submitted = to_submit;
8801 } else if (to_submit) {
8802 ret = io_uring_add_task_file(f.file);
8805 mutex_lock(&ctx->uring_lock);
8806 submitted = io_submit_sqes(ctx, to_submit);
8807 mutex_unlock(&ctx->uring_lock);
8809 if (submitted != to_submit)
8812 if (flags & IORING_ENTER_GETEVENTS) {
8813 min_complete = min(min_complete, ctx->cq_entries);
8816 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8817 * space applications don't need to do io completion events
8818 * polling again, they can rely on io_sq_thread to do polling
8819 * work, which can reduce cpu usage and uring_lock contention.
8821 if (ctx->flags & IORING_SETUP_IOPOLL &&
8822 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8823 ret = io_iopoll_check(ctx, min_complete);
8825 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8830 percpu_ref_put(&ctx->refs);
8833 return submitted ? submitted : ret;
8836 #ifdef CONFIG_PROC_FS
8837 static int io_uring_show_cred(int id, void *p, void *data)
8839 const struct cred *cred = p;
8840 struct seq_file *m = data;
8841 struct user_namespace *uns = seq_user_ns(m);
8842 struct group_info *gi;
8847 seq_printf(m, "%5d\n", id);
8848 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8849 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8850 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8851 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8852 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8853 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8854 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8855 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8856 seq_puts(m, "\n\tGroups:\t");
8857 gi = cred->group_info;
8858 for (g = 0; g < gi->ngroups; g++) {
8859 seq_put_decimal_ull(m, g ? " " : "",
8860 from_kgid_munged(uns, gi->gid[g]));
8862 seq_puts(m, "\n\tCapEff:\t");
8863 cap = cred->cap_effective;
8864 CAP_FOR_EACH_U32(__capi)
8865 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8870 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8872 struct io_sq_data *sq = NULL;
8877 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8878 * since fdinfo case grabs it in the opposite direction of normal use
8879 * cases. If we fail to get the lock, we just don't iterate any
8880 * structures that could be going away outside the io_uring mutex.
8882 has_lock = mutex_trylock(&ctx->uring_lock);
8884 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
8887 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
8888 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
8889 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8890 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8891 struct fixed_file_table *table;
8894 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8895 f = table->files[i & IORING_FILE_TABLE_MASK];
8897 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8899 seq_printf(m, "%5u: <none>\n", i);
8901 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8902 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8903 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8905 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8906 (unsigned int) buf->len);
8908 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8909 seq_printf(m, "Personalities:\n");
8910 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8912 seq_printf(m, "PollList:\n");
8913 spin_lock_irq(&ctx->completion_lock);
8914 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8915 struct hlist_head *list = &ctx->cancel_hash[i];
8916 struct io_kiocb *req;
8918 hlist_for_each_entry(req, list, hash_node)
8919 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8920 req->task->task_works != NULL);
8922 spin_unlock_irq(&ctx->completion_lock);
8924 mutex_unlock(&ctx->uring_lock);
8927 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8929 struct io_ring_ctx *ctx = f->private_data;
8931 if (percpu_ref_tryget(&ctx->refs)) {
8932 __io_uring_show_fdinfo(ctx, m);
8933 percpu_ref_put(&ctx->refs);
8938 static const struct file_operations io_uring_fops = {
8939 .release = io_uring_release,
8940 .flush = io_uring_flush,
8941 .mmap = io_uring_mmap,
8943 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8944 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8946 .poll = io_uring_poll,
8947 .fasync = io_uring_fasync,
8948 #ifdef CONFIG_PROC_FS
8949 .show_fdinfo = io_uring_show_fdinfo,
8953 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8954 struct io_uring_params *p)
8956 struct io_rings *rings;
8957 size_t size, sq_array_offset;
8959 /* make sure these are sane, as we already accounted them */
8960 ctx->sq_entries = p->sq_entries;
8961 ctx->cq_entries = p->cq_entries;
8963 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8964 if (size == SIZE_MAX)
8967 rings = io_mem_alloc(size);
8972 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8973 rings->sq_ring_mask = p->sq_entries - 1;
8974 rings->cq_ring_mask = p->cq_entries - 1;
8975 rings->sq_ring_entries = p->sq_entries;
8976 rings->cq_ring_entries = p->cq_entries;
8977 ctx->sq_mask = rings->sq_ring_mask;
8978 ctx->cq_mask = rings->cq_ring_mask;
8980 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8981 if (size == SIZE_MAX) {
8982 io_mem_free(ctx->rings);
8987 ctx->sq_sqes = io_mem_alloc(size);
8988 if (!ctx->sq_sqes) {
8989 io_mem_free(ctx->rings);
8998 * Allocate an anonymous fd, this is what constitutes the application
8999 * visible backing of an io_uring instance. The application mmaps this
9000 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9001 * we have to tie this fd to a socket for file garbage collection purposes.
9003 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9008 #if defined(CONFIG_UNIX)
9009 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9015 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9019 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9020 O_RDWR | O_CLOEXEC);
9024 ret = PTR_ERR(file);
9028 #if defined(CONFIG_UNIX)
9029 ctx->ring_sock->file = file;
9031 if (unlikely(io_uring_add_task_file(file))) {
9032 file = ERR_PTR(-ENOMEM);
9035 fd_install(ret, file);
9038 #if defined(CONFIG_UNIX)
9039 sock_release(ctx->ring_sock);
9040 ctx->ring_sock = NULL;
9045 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9046 struct io_uring_params __user *params)
9048 struct user_struct *user = NULL;
9049 struct io_ring_ctx *ctx;
9055 if (entries > IORING_MAX_ENTRIES) {
9056 if (!(p->flags & IORING_SETUP_CLAMP))
9058 entries = IORING_MAX_ENTRIES;
9062 * Use twice as many entries for the CQ ring. It's possible for the
9063 * application to drive a higher depth than the size of the SQ ring,
9064 * since the sqes are only used at submission time. This allows for
9065 * some flexibility in overcommitting a bit. If the application has
9066 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9067 * of CQ ring entries manually.
9069 p->sq_entries = roundup_pow_of_two(entries);
9070 if (p->flags & IORING_SETUP_CQSIZE) {
9072 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9073 * to a power-of-two, if it isn't already. We do NOT impose
9074 * any cq vs sq ring sizing.
9076 if (p->cq_entries < p->sq_entries)
9078 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9079 if (!(p->flags & IORING_SETUP_CLAMP))
9081 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9083 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9085 p->cq_entries = 2 * p->sq_entries;
9088 user = get_uid(current_user());
9089 limit_mem = !capable(CAP_IPC_LOCK);
9092 ret = __io_account_mem(user,
9093 ring_pages(p->sq_entries, p->cq_entries));
9100 ctx = io_ring_ctx_alloc(p);
9103 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9108 ctx->compat = in_compat_syscall();
9110 ctx->creds = get_current_cred();
9112 ctx->sqo_task = get_task_struct(current);
9115 * This is just grabbed for accounting purposes. When a process exits,
9116 * the mm is exited and dropped before the files, hence we need to hang
9117 * on to this mm purely for the purposes of being able to unaccount
9118 * memory (locked/pinned vm). It's not used for anything else.
9120 mmgrab(current->mm);
9121 ctx->mm_account = current->mm;
9123 #ifdef CONFIG_BLK_CGROUP
9125 * The sq thread will belong to the original cgroup it was inited in.
9126 * If the cgroup goes offline (e.g. disabling the io controller), then
9127 * issued bios will be associated with the closest cgroup later in the
9131 ctx->sqo_blkcg_css = blkcg_css();
9132 ret = css_tryget_online(ctx->sqo_blkcg_css);
9135 /* don't init against a dying cgroup, have the user try again */
9136 ctx->sqo_blkcg_css = NULL;
9143 * Account memory _before_ installing the file descriptor. Once
9144 * the descriptor is installed, it can get closed at any time. Also
9145 * do this before hitting the general error path, as ring freeing
9146 * will un-account as well.
9148 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9150 ctx->limit_mem = limit_mem;
9152 ret = io_allocate_scq_urings(ctx, p);
9156 ret = io_sq_offload_create(ctx, p);
9160 if (!(p->flags & IORING_SETUP_R_DISABLED))
9161 io_sq_offload_start(ctx);
9163 memset(&p->sq_off, 0, sizeof(p->sq_off));
9164 p->sq_off.head = offsetof(struct io_rings, sq.head);
9165 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9166 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9167 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9168 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9169 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9170 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9172 memset(&p->cq_off, 0, sizeof(p->cq_off));
9173 p->cq_off.head = offsetof(struct io_rings, cq.head);
9174 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9175 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9176 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9177 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9178 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9179 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9181 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9182 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9183 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9184 IORING_FEAT_POLL_32BITS;
9186 if (copy_to_user(params, p, sizeof(*p))) {
9192 * Install ring fd as the very last thing, so we don't risk someone
9193 * having closed it before we finish setup
9195 ret = io_uring_get_fd(ctx);
9199 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9202 io_ring_ctx_wait_and_kill(ctx);
9207 * Sets up an aio uring context, and returns the fd. Applications asks for a
9208 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9209 * params structure passed in.
9211 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9213 struct io_uring_params p;
9216 if (copy_from_user(&p, params, sizeof(p)))
9218 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9223 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9224 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9225 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9226 IORING_SETUP_R_DISABLED))
9229 return io_uring_create(entries, &p, params);
9232 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9233 struct io_uring_params __user *, params)
9235 return io_uring_setup(entries, params);
9238 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9240 struct io_uring_probe *p;
9244 size = struct_size(p, ops, nr_args);
9245 if (size == SIZE_MAX)
9247 p = kzalloc(size, GFP_KERNEL);
9252 if (copy_from_user(p, arg, size))
9255 if (memchr_inv(p, 0, size))
9258 p->last_op = IORING_OP_LAST - 1;
9259 if (nr_args > IORING_OP_LAST)
9260 nr_args = IORING_OP_LAST;
9262 for (i = 0; i < nr_args; i++) {
9264 if (!io_op_defs[i].not_supported)
9265 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9270 if (copy_to_user(arg, p, size))
9277 static int io_register_personality(struct io_ring_ctx *ctx)
9279 const struct cred *creds = get_current_cred();
9282 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9283 USHRT_MAX, GFP_KERNEL);
9289 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9291 const struct cred *old_creds;
9293 old_creds = idr_remove(&ctx->personality_idr, id);
9295 put_cred(old_creds);
9302 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9303 unsigned int nr_args)
9305 struct io_uring_restriction *res;
9309 /* Restrictions allowed only if rings started disabled */
9310 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9313 /* We allow only a single restrictions registration */
9314 if (ctx->restrictions.registered)
9317 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9320 size = array_size(nr_args, sizeof(*res));
9321 if (size == SIZE_MAX)
9324 res = memdup_user(arg, size);
9326 return PTR_ERR(res);
9330 for (i = 0; i < nr_args; i++) {
9331 switch (res[i].opcode) {
9332 case IORING_RESTRICTION_REGISTER_OP:
9333 if (res[i].register_op >= IORING_REGISTER_LAST) {
9338 __set_bit(res[i].register_op,
9339 ctx->restrictions.register_op);
9341 case IORING_RESTRICTION_SQE_OP:
9342 if (res[i].sqe_op >= IORING_OP_LAST) {
9347 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9349 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9350 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9352 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9353 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9362 /* Reset all restrictions if an error happened */
9364 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9366 ctx->restrictions.registered = true;
9372 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9374 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9377 if (ctx->restrictions.registered)
9378 ctx->restricted = 1;
9380 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9382 io_sq_offload_start(ctx);
9387 static bool io_register_op_must_quiesce(int op)
9390 case IORING_UNREGISTER_FILES:
9391 case IORING_REGISTER_FILES_UPDATE:
9392 case IORING_REGISTER_PROBE:
9393 case IORING_REGISTER_PERSONALITY:
9394 case IORING_UNREGISTER_PERSONALITY:
9401 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9402 void __user *arg, unsigned nr_args)
9403 __releases(ctx->uring_lock)
9404 __acquires(ctx->uring_lock)
9409 * We're inside the ring mutex, if the ref is already dying, then
9410 * someone else killed the ctx or is already going through
9411 * io_uring_register().
9413 if (percpu_ref_is_dying(&ctx->refs))
9416 if (io_register_op_must_quiesce(opcode)) {
9417 percpu_ref_kill(&ctx->refs);
9420 * Drop uring mutex before waiting for references to exit. If
9421 * another thread is currently inside io_uring_enter() it might
9422 * need to grab the uring_lock to make progress. If we hold it
9423 * here across the drain wait, then we can deadlock. It's safe
9424 * to drop the mutex here, since no new references will come in
9425 * after we've killed the percpu ref.
9427 mutex_unlock(&ctx->uring_lock);
9429 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9432 ret = io_run_task_work_sig();
9437 mutex_lock(&ctx->uring_lock);
9440 percpu_ref_resurrect(&ctx->refs);
9445 if (ctx->restricted) {
9446 if (opcode >= IORING_REGISTER_LAST) {
9451 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9458 case IORING_REGISTER_BUFFERS:
9459 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9461 case IORING_UNREGISTER_BUFFERS:
9465 ret = io_sqe_buffer_unregister(ctx);
9467 case IORING_REGISTER_FILES:
9468 ret = io_sqe_files_register(ctx, arg, nr_args);
9470 case IORING_UNREGISTER_FILES:
9474 ret = io_sqe_files_unregister(ctx);
9476 case IORING_REGISTER_FILES_UPDATE:
9477 ret = io_sqe_files_update(ctx, arg, nr_args);
9479 case IORING_REGISTER_EVENTFD:
9480 case IORING_REGISTER_EVENTFD_ASYNC:
9484 ret = io_eventfd_register(ctx, arg);
9487 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9488 ctx->eventfd_async = 1;
9490 ctx->eventfd_async = 0;
9492 case IORING_UNREGISTER_EVENTFD:
9496 ret = io_eventfd_unregister(ctx);
9498 case IORING_REGISTER_PROBE:
9500 if (!arg || nr_args > 256)
9502 ret = io_probe(ctx, arg, nr_args);
9504 case IORING_REGISTER_PERSONALITY:
9508 ret = io_register_personality(ctx);
9510 case IORING_UNREGISTER_PERSONALITY:
9514 ret = io_unregister_personality(ctx, nr_args);
9516 case IORING_REGISTER_ENABLE_RINGS:
9520 ret = io_register_enable_rings(ctx);
9522 case IORING_REGISTER_RESTRICTIONS:
9523 ret = io_register_restrictions(ctx, arg, nr_args);
9531 if (io_register_op_must_quiesce(opcode)) {
9532 /* bring the ctx back to life */
9533 percpu_ref_reinit(&ctx->refs);
9535 reinit_completion(&ctx->ref_comp);
9540 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9541 void __user *, arg, unsigned int, nr_args)
9543 struct io_ring_ctx *ctx;
9552 if (f.file->f_op != &io_uring_fops)
9555 ctx = f.file->private_data;
9557 mutex_lock(&ctx->uring_lock);
9558 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9559 mutex_unlock(&ctx->uring_lock);
9560 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9561 ctx->cq_ev_fd != NULL, ret);
9567 static int __init io_uring_init(void)
9569 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9570 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9571 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9574 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9575 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9576 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9577 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9578 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9579 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9580 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9581 BUILD_BUG_SQE_ELEM(8, __u64, off);
9582 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9583 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9584 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9585 BUILD_BUG_SQE_ELEM(24, __u32, len);
9586 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9587 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9588 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9589 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9590 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9591 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9592 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9593 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9594 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9595 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9596 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9597 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9598 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9599 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9600 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9601 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9602 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9603 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9604 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9606 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9607 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9608 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9611 __initcall(io_uring_init);