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 io_cqring_fill_event(req, -ECANCELED);
1773 io_commit_cqring(ctx);
1774 req->flags &= ~REQ_F_LINK_HEAD;
1782 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1784 struct io_kiocb *link;
1787 if (list_empty(&req->link_list))
1789 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1790 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1793 list_del_init(&link->link_list);
1794 link->flags |= REQ_F_COMP_LOCKED;
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);
1855 req->flags &= ~REQ_F_LINK_TIMEOUT;
1858 io_commit_cqring(ctx);
1859 io_cqring_ev_posted(ctx);
1862 static void io_fail_links(struct io_kiocb *req)
1864 struct io_ring_ctx *ctx = req->ctx;
1866 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1867 unsigned long flags;
1869 spin_lock_irqsave(&ctx->completion_lock, flags);
1870 __io_fail_links(req);
1871 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1873 __io_fail_links(req);
1876 io_cqring_ev_posted(ctx);
1879 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1881 req->flags &= ~REQ_F_LINK_HEAD;
1882 if (req->flags & REQ_F_LINK_TIMEOUT)
1883 io_kill_linked_timeout(req);
1886 * If LINK is set, we have dependent requests in this chain. If we
1887 * didn't fail this request, queue the first one up, moving any other
1888 * dependencies to the next request. In case of failure, fail the rest
1891 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1892 return io_req_link_next(req);
1897 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1899 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1901 return __io_req_find_next(req);
1904 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1906 struct task_struct *tsk = req->task;
1907 struct io_ring_ctx *ctx = req->ctx;
1910 if (tsk->flags & PF_EXITING)
1914 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1915 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1916 * processing task_work. There's no reliable way to tell if TWA_RESUME
1920 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1921 notify = TWA_SIGNAL;
1923 ret = task_work_add(tsk, &req->task_work, notify);
1925 wake_up_process(tsk);
1930 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1932 struct io_ring_ctx *ctx = req->ctx;
1934 spin_lock_irq(&ctx->completion_lock);
1935 io_cqring_fill_event(req, error);
1936 io_commit_cqring(ctx);
1937 spin_unlock_irq(&ctx->completion_lock);
1939 io_cqring_ev_posted(ctx);
1940 req_set_fail_links(req);
1941 io_double_put_req(req);
1944 static void io_req_task_cancel(struct callback_head *cb)
1946 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1947 struct io_ring_ctx *ctx = req->ctx;
1949 __io_req_task_cancel(req, -ECANCELED);
1950 percpu_ref_put(&ctx->refs);
1953 static void __io_req_task_submit(struct io_kiocb *req)
1955 struct io_ring_ctx *ctx = req->ctx;
1957 if (!__io_sq_thread_acquire_mm(ctx)) {
1958 mutex_lock(&ctx->uring_lock);
1959 __io_queue_sqe(req, NULL);
1960 mutex_unlock(&ctx->uring_lock);
1962 __io_req_task_cancel(req, -EFAULT);
1966 static void io_req_task_submit(struct callback_head *cb)
1968 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1969 struct io_ring_ctx *ctx = req->ctx;
1971 __io_req_task_submit(req);
1972 percpu_ref_put(&ctx->refs);
1975 static void io_req_task_queue(struct io_kiocb *req)
1979 init_task_work(&req->task_work, io_req_task_submit);
1980 percpu_ref_get(&req->ctx->refs);
1982 ret = io_req_task_work_add(req, true);
1983 if (unlikely(ret)) {
1984 struct task_struct *tsk;
1986 init_task_work(&req->task_work, io_req_task_cancel);
1987 tsk = io_wq_get_task(req->ctx->io_wq);
1988 task_work_add(tsk, &req->task_work, 0);
1989 wake_up_process(tsk);
1993 static void io_queue_next(struct io_kiocb *req)
1995 struct io_kiocb *nxt = io_req_find_next(req);
1998 io_req_task_queue(nxt);
2001 static void io_free_req(struct io_kiocb *req)
2008 void *reqs[IO_IOPOLL_BATCH];
2011 struct task_struct *task;
2015 static inline void io_init_req_batch(struct req_batch *rb)
2022 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2023 struct req_batch *rb)
2025 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2026 percpu_ref_put_many(&ctx->refs, rb->to_free);
2030 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2031 struct req_batch *rb)
2034 __io_req_free_batch_flush(ctx, rb);
2036 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2037 put_task_struct_many(rb->task, rb->task_refs);
2042 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2044 if (unlikely(io_is_fallback_req(req))) {
2048 if (req->flags & REQ_F_LINK_HEAD)
2051 if (req->task != rb->task) {
2053 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2054 put_task_struct_many(rb->task, rb->task_refs);
2056 rb->task = req->task;
2061 WARN_ON_ONCE(io_dismantle_req(req));
2062 rb->reqs[rb->to_free++] = req;
2063 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2064 __io_req_free_batch_flush(req->ctx, rb);
2068 * Drop reference to request, return next in chain (if there is one) if this
2069 * was the last reference to this request.
2071 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2073 struct io_kiocb *nxt = NULL;
2075 if (refcount_dec_and_test(&req->refs)) {
2076 nxt = io_req_find_next(req);
2082 static void io_put_req(struct io_kiocb *req)
2084 if (refcount_dec_and_test(&req->refs))
2088 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2090 struct io_kiocb *nxt;
2093 * A ref is owned by io-wq in which context we're. So, if that's the
2094 * last one, it's safe to steal next work. False negatives are Ok,
2095 * it just will be re-punted async in io_put_work()
2097 if (refcount_read(&req->refs) != 1)
2100 nxt = io_req_find_next(req);
2101 return nxt ? &nxt->work : NULL;
2105 * Must only be used if we don't need to care about links, usually from
2106 * within the completion handling itself.
2108 static void __io_double_put_req(struct io_kiocb *req)
2110 /* drop both submit and complete references */
2111 if (refcount_sub_and_test(2, &req->refs))
2115 static void io_double_put_req(struct io_kiocb *req)
2117 /* drop both submit and complete references */
2118 if (refcount_sub_and_test(2, &req->refs))
2122 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2124 struct io_rings *rings = ctx->rings;
2126 if (test_bit(0, &ctx->cq_check_overflow)) {
2128 * noflush == true is from the waitqueue handler, just ensure
2129 * we wake up the task, and the next invocation will flush the
2130 * entries. We cannot safely to it from here.
2132 if (noflush && !list_empty(&ctx->cq_overflow_list))
2135 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2138 /* See comment at the top of this file */
2140 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2143 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2145 struct io_rings *rings = ctx->rings;
2147 /* make sure SQ entry isn't read before tail */
2148 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2151 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2153 unsigned int cflags;
2155 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2156 cflags |= IORING_CQE_F_BUFFER;
2157 req->flags &= ~REQ_F_BUFFER_SELECTED;
2162 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2164 struct io_buffer *kbuf;
2166 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2167 return io_put_kbuf(req, kbuf);
2170 static inline bool io_run_task_work(void)
2173 * Not safe to run on exiting task, and the task_work handling will
2174 * not add work to such a task.
2176 if (unlikely(current->flags & PF_EXITING))
2178 if (current->task_works) {
2179 __set_current_state(TASK_RUNNING);
2187 static void io_iopoll_queue(struct list_head *again)
2189 struct io_kiocb *req;
2192 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2193 list_del(&req->inflight_entry);
2194 __io_complete_rw(req, -EAGAIN, 0, NULL);
2195 } while (!list_empty(again));
2199 * Find and free completed poll iocbs
2201 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2202 struct list_head *done)
2204 struct req_batch rb;
2205 struct io_kiocb *req;
2208 /* order with ->result store in io_complete_rw_iopoll() */
2211 io_init_req_batch(&rb);
2212 while (!list_empty(done)) {
2215 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2216 if (READ_ONCE(req->result) == -EAGAIN) {
2218 req->iopoll_completed = 0;
2219 list_move_tail(&req->inflight_entry, &again);
2222 list_del(&req->inflight_entry);
2224 if (req->flags & REQ_F_BUFFER_SELECTED)
2225 cflags = io_put_rw_kbuf(req);
2227 __io_cqring_fill_event(req, req->result, cflags);
2230 if (refcount_dec_and_test(&req->refs))
2231 io_req_free_batch(&rb, req);
2234 io_commit_cqring(ctx);
2235 if (ctx->flags & IORING_SETUP_SQPOLL)
2236 io_cqring_ev_posted(ctx);
2237 io_req_free_batch_finish(ctx, &rb);
2239 if (!list_empty(&again))
2240 io_iopoll_queue(&again);
2243 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2246 struct io_kiocb *req, *tmp;
2252 * Only spin for completions if we don't have multiple devices hanging
2253 * off our complete list, and we're under the requested amount.
2255 spin = !ctx->poll_multi_file && *nr_events < min;
2258 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2259 struct kiocb *kiocb = &req->rw.kiocb;
2262 * Move completed and retryable entries to our local lists.
2263 * If we find a request that requires polling, break out
2264 * and complete those lists first, if we have entries there.
2266 if (READ_ONCE(req->iopoll_completed)) {
2267 list_move_tail(&req->inflight_entry, &done);
2270 if (!list_empty(&done))
2273 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2277 /* iopoll may have completed current req */
2278 if (READ_ONCE(req->iopoll_completed))
2279 list_move_tail(&req->inflight_entry, &done);
2286 if (!list_empty(&done))
2287 io_iopoll_complete(ctx, nr_events, &done);
2293 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2294 * non-spinning poll check - we'll still enter the driver poll loop, but only
2295 * as a non-spinning completion check.
2297 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2300 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2303 ret = io_do_iopoll(ctx, nr_events, min);
2306 if (*nr_events >= min)
2314 * We can't just wait for polled events to come to us, we have to actively
2315 * find and complete them.
2317 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2319 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2322 mutex_lock(&ctx->uring_lock);
2323 while (!list_empty(&ctx->iopoll_list)) {
2324 unsigned int nr_events = 0;
2326 io_do_iopoll(ctx, &nr_events, 0);
2328 /* let it sleep and repeat later if can't complete a request */
2332 * Ensure we allow local-to-the-cpu processing to take place,
2333 * in this case we need to ensure that we reap all events.
2334 * Also let task_work, etc. to progress by releasing the mutex
2336 if (need_resched()) {
2337 mutex_unlock(&ctx->uring_lock);
2339 mutex_lock(&ctx->uring_lock);
2342 mutex_unlock(&ctx->uring_lock);
2345 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2347 unsigned int nr_events = 0;
2348 int iters = 0, ret = 0;
2351 * We disallow the app entering submit/complete with polling, but we
2352 * still need to lock the ring to prevent racing with polled issue
2353 * that got punted to a workqueue.
2355 mutex_lock(&ctx->uring_lock);
2358 * Don't enter poll loop if we already have events pending.
2359 * If we do, we can potentially be spinning for commands that
2360 * already triggered a CQE (eg in error).
2362 if (io_cqring_events(ctx, false))
2366 * If a submit got punted to a workqueue, we can have the
2367 * application entering polling for a command before it gets
2368 * issued. That app will hold the uring_lock for the duration
2369 * of the poll right here, so we need to take a breather every
2370 * now and then to ensure that the issue has a chance to add
2371 * the poll to the issued list. Otherwise we can spin here
2372 * forever, while the workqueue is stuck trying to acquire the
2375 if (!(++iters & 7)) {
2376 mutex_unlock(&ctx->uring_lock);
2378 mutex_lock(&ctx->uring_lock);
2381 ret = io_iopoll_getevents(ctx, &nr_events, min);
2385 } while (min && !nr_events && !need_resched());
2387 mutex_unlock(&ctx->uring_lock);
2391 static void kiocb_end_write(struct io_kiocb *req)
2394 * Tell lockdep we inherited freeze protection from submission
2397 if (req->flags & REQ_F_ISREG) {
2398 struct inode *inode = file_inode(req->file);
2400 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2402 file_end_write(req->file);
2405 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2406 struct io_comp_state *cs)
2408 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2411 if (kiocb->ki_flags & IOCB_WRITE)
2412 kiocb_end_write(req);
2414 if (res != req->result)
2415 req_set_fail_links(req);
2416 if (req->flags & REQ_F_BUFFER_SELECTED)
2417 cflags = io_put_rw_kbuf(req);
2418 __io_req_complete(req, res, cflags, cs);
2422 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2424 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2425 ssize_t ret = -ECANCELED;
2426 struct iov_iter iter;
2434 switch (req->opcode) {
2435 case IORING_OP_READV:
2436 case IORING_OP_READ_FIXED:
2437 case IORING_OP_READ:
2440 case IORING_OP_WRITEV:
2441 case IORING_OP_WRITE_FIXED:
2442 case IORING_OP_WRITE:
2446 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2451 if (!req->async_data) {
2452 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2455 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2463 req_set_fail_links(req);
2464 io_req_complete(req, ret);
2469 static bool io_rw_reissue(struct io_kiocb *req, long res)
2472 umode_t mode = file_inode(req->file)->i_mode;
2475 if (!S_ISBLK(mode) && !S_ISREG(mode))
2477 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2480 ret = io_sq_thread_acquire_mm(req->ctx, req);
2482 if (io_resubmit_prep(req, ret)) {
2483 refcount_inc(&req->refs);
2484 io_queue_async_work(req);
2492 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2493 struct io_comp_state *cs)
2495 if (!io_rw_reissue(req, res))
2496 io_complete_rw_common(&req->rw.kiocb, res, cs);
2499 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2501 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2503 __io_complete_rw(req, res, res2, NULL);
2506 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2508 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2510 if (kiocb->ki_flags & IOCB_WRITE)
2511 kiocb_end_write(req);
2513 if (res != -EAGAIN && res != req->result)
2514 req_set_fail_links(req);
2516 WRITE_ONCE(req->result, res);
2517 /* order with io_poll_complete() checking ->result */
2519 WRITE_ONCE(req->iopoll_completed, 1);
2523 * After the iocb has been issued, it's safe to be found on the poll list.
2524 * Adding the kiocb to the list AFTER submission ensures that we don't
2525 * find it from a io_iopoll_getevents() thread before the issuer is done
2526 * accessing the kiocb cookie.
2528 static void io_iopoll_req_issued(struct io_kiocb *req)
2530 struct io_ring_ctx *ctx = req->ctx;
2533 * Track whether we have multiple files in our lists. This will impact
2534 * how we do polling eventually, not spinning if we're on potentially
2535 * different devices.
2537 if (list_empty(&ctx->iopoll_list)) {
2538 ctx->poll_multi_file = false;
2539 } else if (!ctx->poll_multi_file) {
2540 struct io_kiocb *list_req;
2542 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2544 if (list_req->file != req->file)
2545 ctx->poll_multi_file = true;
2549 * For fast devices, IO may have already completed. If it has, add
2550 * it to the front so we find it first.
2552 if (READ_ONCE(req->iopoll_completed))
2553 list_add(&req->inflight_entry, &ctx->iopoll_list);
2555 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2557 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2558 wq_has_sleeper(&ctx->sq_data->wait))
2559 wake_up(&ctx->sq_data->wait);
2562 static void __io_state_file_put(struct io_submit_state *state)
2564 if (state->has_refs)
2565 fput_many(state->file, state->has_refs);
2569 static inline void io_state_file_put(struct io_submit_state *state)
2572 __io_state_file_put(state);
2576 * Get as many references to a file as we have IOs left in this submission,
2577 * assuming most submissions are for one file, or at least that each file
2578 * has more than one submission.
2580 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2586 if (state->fd == fd) {
2590 __io_state_file_put(state);
2592 state->file = fget_many(fd, state->ios_left);
2597 state->has_refs = state->ios_left - 1;
2601 static bool io_bdev_nowait(struct block_device *bdev)
2604 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2611 * If we tracked the file through the SCM inflight mechanism, we could support
2612 * any file. For now, just ensure that anything potentially problematic is done
2615 static bool io_file_supports_async(struct file *file, int rw)
2617 umode_t mode = file_inode(file)->i_mode;
2619 if (S_ISBLK(mode)) {
2620 if (io_bdev_nowait(file->f_inode->i_bdev))
2624 if (S_ISCHR(mode) || S_ISSOCK(mode))
2626 if (S_ISREG(mode)) {
2627 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2628 file->f_op != &io_uring_fops)
2633 /* any ->read/write should understand O_NONBLOCK */
2634 if (file->f_flags & O_NONBLOCK)
2637 if (!(file->f_mode & FMODE_NOWAIT))
2641 return file->f_op->read_iter != NULL;
2643 return file->f_op->write_iter != NULL;
2646 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2648 struct io_ring_ctx *ctx = req->ctx;
2649 struct kiocb *kiocb = &req->rw.kiocb;
2653 if (S_ISREG(file_inode(req->file)->i_mode))
2654 req->flags |= REQ_F_ISREG;
2656 kiocb->ki_pos = READ_ONCE(sqe->off);
2657 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2658 req->flags |= REQ_F_CUR_POS;
2659 kiocb->ki_pos = req->file->f_pos;
2661 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2662 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2663 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2667 ioprio = READ_ONCE(sqe->ioprio);
2669 ret = ioprio_check_cap(ioprio);
2673 kiocb->ki_ioprio = ioprio;
2675 kiocb->ki_ioprio = get_current_ioprio();
2677 /* don't allow async punt if RWF_NOWAIT was requested */
2678 if (kiocb->ki_flags & IOCB_NOWAIT)
2679 req->flags |= REQ_F_NOWAIT;
2681 if (ctx->flags & IORING_SETUP_IOPOLL) {
2682 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2683 !kiocb->ki_filp->f_op->iopoll)
2686 kiocb->ki_flags |= IOCB_HIPRI;
2687 kiocb->ki_complete = io_complete_rw_iopoll;
2688 req->iopoll_completed = 0;
2690 if (kiocb->ki_flags & IOCB_HIPRI)
2692 kiocb->ki_complete = io_complete_rw;
2695 req->rw.addr = READ_ONCE(sqe->addr);
2696 req->rw.len = READ_ONCE(sqe->len);
2697 req->buf_index = READ_ONCE(sqe->buf_index);
2701 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2707 case -ERESTARTNOINTR:
2708 case -ERESTARTNOHAND:
2709 case -ERESTART_RESTARTBLOCK:
2711 * We can't just restart the syscall, since previously
2712 * submitted sqes may already be in progress. Just fail this
2718 kiocb->ki_complete(kiocb, ret, 0);
2722 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2723 struct io_comp_state *cs)
2725 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2726 struct io_async_rw *io = req->async_data;
2728 /* add previously done IO, if any */
2729 if (io && io->bytes_done > 0) {
2731 ret = io->bytes_done;
2733 ret += io->bytes_done;
2736 if (req->flags & REQ_F_CUR_POS)
2737 req->file->f_pos = kiocb->ki_pos;
2738 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2739 __io_complete_rw(req, ret, 0, cs);
2741 io_rw_done(kiocb, ret);
2744 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2745 struct iov_iter *iter)
2747 struct io_ring_ctx *ctx = req->ctx;
2748 size_t len = req->rw.len;
2749 struct io_mapped_ubuf *imu;
2750 u16 index, buf_index = req->buf_index;
2754 if (unlikely(buf_index >= ctx->nr_user_bufs))
2756 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2757 imu = &ctx->user_bufs[index];
2758 buf_addr = req->rw.addr;
2761 if (buf_addr + len < buf_addr)
2763 /* not inside the mapped region */
2764 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2768 * May not be a start of buffer, set size appropriately
2769 * and advance us to the beginning.
2771 offset = buf_addr - imu->ubuf;
2772 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2776 * Don't use iov_iter_advance() here, as it's really slow for
2777 * using the latter parts of a big fixed buffer - it iterates
2778 * over each segment manually. We can cheat a bit here, because
2781 * 1) it's a BVEC iter, we set it up
2782 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2783 * first and last bvec
2785 * So just find our index, and adjust the iterator afterwards.
2786 * If the offset is within the first bvec (or the whole first
2787 * bvec, just use iov_iter_advance(). This makes it easier
2788 * since we can just skip the first segment, which may not
2789 * be PAGE_SIZE aligned.
2791 const struct bio_vec *bvec = imu->bvec;
2793 if (offset <= bvec->bv_len) {
2794 iov_iter_advance(iter, offset);
2796 unsigned long seg_skip;
2798 /* skip first vec */
2799 offset -= bvec->bv_len;
2800 seg_skip = 1 + (offset >> PAGE_SHIFT);
2802 iter->bvec = bvec + seg_skip;
2803 iter->nr_segs -= seg_skip;
2804 iter->count -= bvec->bv_len + offset;
2805 iter->iov_offset = offset & ~PAGE_MASK;
2812 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2815 mutex_unlock(&ctx->uring_lock);
2818 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2821 * "Normal" inline submissions always hold the uring_lock, since we
2822 * grab it from the system call. Same is true for the SQPOLL offload.
2823 * The only exception is when we've detached the request and issue it
2824 * from an async worker thread, grab the lock for that case.
2827 mutex_lock(&ctx->uring_lock);
2830 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2831 int bgid, struct io_buffer *kbuf,
2834 struct io_buffer *head;
2836 if (req->flags & REQ_F_BUFFER_SELECTED)
2839 io_ring_submit_lock(req->ctx, needs_lock);
2841 lockdep_assert_held(&req->ctx->uring_lock);
2843 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2845 if (!list_empty(&head->list)) {
2846 kbuf = list_last_entry(&head->list, struct io_buffer,
2848 list_del(&kbuf->list);
2851 idr_remove(&req->ctx->io_buffer_idr, bgid);
2853 if (*len > kbuf->len)
2856 kbuf = ERR_PTR(-ENOBUFS);
2859 io_ring_submit_unlock(req->ctx, needs_lock);
2864 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2867 struct io_buffer *kbuf;
2870 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2871 bgid = req->buf_index;
2872 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2875 req->rw.addr = (u64) (unsigned long) kbuf;
2876 req->flags |= REQ_F_BUFFER_SELECTED;
2877 return u64_to_user_ptr(kbuf->addr);
2880 #ifdef CONFIG_COMPAT
2881 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2884 struct compat_iovec __user *uiov;
2885 compat_ssize_t clen;
2889 uiov = u64_to_user_ptr(req->rw.addr);
2890 if (!access_ok(uiov, sizeof(*uiov)))
2892 if (__get_user(clen, &uiov->iov_len))
2898 buf = io_rw_buffer_select(req, &len, needs_lock);
2900 return PTR_ERR(buf);
2901 iov[0].iov_base = buf;
2902 iov[0].iov_len = (compat_size_t) len;
2907 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2910 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2914 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2917 len = iov[0].iov_len;
2920 buf = io_rw_buffer_select(req, &len, needs_lock);
2922 return PTR_ERR(buf);
2923 iov[0].iov_base = buf;
2924 iov[0].iov_len = len;
2928 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2931 if (req->flags & REQ_F_BUFFER_SELECTED) {
2932 struct io_buffer *kbuf;
2934 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2935 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2936 iov[0].iov_len = kbuf->len;
2941 else if (req->rw.len > 1)
2944 #ifdef CONFIG_COMPAT
2945 if (req->ctx->compat)
2946 return io_compat_import(req, iov, needs_lock);
2949 return __io_iov_buffer_select(req, iov, needs_lock);
2952 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2953 struct iovec **iovec, struct iov_iter *iter,
2956 void __user *buf = u64_to_user_ptr(req->rw.addr);
2957 size_t sqe_len = req->rw.len;
2961 opcode = req->opcode;
2962 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2964 return io_import_fixed(req, rw, iter);
2967 /* buffer index only valid with fixed read/write, or buffer select */
2968 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2971 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2972 if (req->flags & REQ_F_BUFFER_SELECT) {
2973 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2975 return PTR_ERR(buf);
2976 req->rw.len = sqe_len;
2979 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2981 return ret < 0 ? ret : sqe_len;
2984 if (req->flags & REQ_F_BUFFER_SELECT) {
2985 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2987 ret = (*iovec)->iov_len;
2988 iov_iter_init(iter, rw, *iovec, 1, ret);
2994 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2998 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2999 struct iovec **iovec, struct iov_iter *iter,
3002 struct io_async_rw *iorw = req->async_data;
3005 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3007 return iov_iter_count(&iorw->iter);
3010 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3012 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3016 * For files that don't have ->read_iter() and ->write_iter(), handle them
3017 * by looping over ->read() or ->write() manually.
3019 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
3020 struct iov_iter *iter)
3025 * Don't support polled IO through this interface, and we can't
3026 * support non-blocking either. For the latter, this just causes
3027 * the kiocb to be handled from an async context.
3029 if (kiocb->ki_flags & IOCB_HIPRI)
3031 if (kiocb->ki_flags & IOCB_NOWAIT)
3034 while (iov_iter_count(iter)) {
3038 if (!iov_iter_is_bvec(iter)) {
3039 iovec = iov_iter_iovec(iter);
3041 /* fixed buffers import bvec */
3042 iovec.iov_base = kmap(iter->bvec->bv_page)
3044 iovec.iov_len = min(iter->count,
3045 iter->bvec->bv_len - iter->iov_offset);
3049 nr = file->f_op->read(file, iovec.iov_base,
3050 iovec.iov_len, io_kiocb_ppos(kiocb));
3052 nr = file->f_op->write(file, iovec.iov_base,
3053 iovec.iov_len, io_kiocb_ppos(kiocb));
3056 if (iov_iter_is_bvec(iter))
3057 kunmap(iter->bvec->bv_page);
3065 if (nr != iovec.iov_len)
3067 iov_iter_advance(iter, nr);
3073 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3074 const struct iovec *fast_iov, struct iov_iter *iter)
3076 struct io_async_rw *rw = req->async_data;
3078 memcpy(&rw->iter, iter, sizeof(*iter));
3079 rw->free_iovec = iovec;
3081 /* can only be fixed buffers, no need to do anything */
3082 if (iter->type == ITER_BVEC)
3085 unsigned iov_off = 0;
3087 rw->iter.iov = rw->fast_iov;
3088 if (iter->iov != fast_iov) {
3089 iov_off = iter->iov - fast_iov;
3090 rw->iter.iov += iov_off;
3092 if (rw->fast_iov != fast_iov)
3093 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3094 sizeof(struct iovec) * iter->nr_segs);
3096 req->flags |= REQ_F_NEED_CLEANUP;
3100 static inline int __io_alloc_async_data(struct io_kiocb *req)
3102 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3103 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3104 return req->async_data == NULL;
3107 static int io_alloc_async_data(struct io_kiocb *req)
3109 if (!io_op_defs[req->opcode].needs_async_data)
3112 return __io_alloc_async_data(req);
3115 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3116 const struct iovec *fast_iov,
3117 struct iov_iter *iter, bool force)
3119 if (!force && !io_op_defs[req->opcode].needs_async_data)
3121 if (!req->async_data) {
3122 if (__io_alloc_async_data(req))
3125 io_req_map_rw(req, iovec, fast_iov, iter);
3130 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3132 struct io_async_rw *iorw = req->async_data;
3133 struct iovec *iov = iorw->fast_iov;
3136 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3137 if (unlikely(ret < 0))
3140 iorw->bytes_done = 0;
3141 iorw->free_iovec = iov;
3143 req->flags |= REQ_F_NEED_CLEANUP;
3147 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3151 ret = io_prep_rw(req, sqe);
3155 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3158 /* either don't need iovec imported or already have it */
3159 if (!req->async_data)
3161 return io_rw_prep_async(req, READ);
3165 * This is our waitqueue callback handler, registered through lock_page_async()
3166 * when we initially tried to do the IO with the iocb armed our waitqueue.
3167 * This gets called when the page is unlocked, and we generally expect that to
3168 * happen when the page IO is completed and the page is now uptodate. This will
3169 * queue a task_work based retry of the operation, attempting to copy the data
3170 * again. If the latter fails because the page was NOT uptodate, then we will
3171 * do a thread based blocking retry of the operation. That's the unexpected
3174 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3175 int sync, void *arg)
3177 struct wait_page_queue *wpq;
3178 struct io_kiocb *req = wait->private;
3179 struct wait_page_key *key = arg;
3182 wpq = container_of(wait, struct wait_page_queue, wait);
3184 if (!wake_page_match(wpq, key))
3187 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3188 list_del_init(&wait->entry);
3190 init_task_work(&req->task_work, io_req_task_submit);
3191 percpu_ref_get(&req->ctx->refs);
3193 /* submit ref gets dropped, acquire a new one */
3194 refcount_inc(&req->refs);
3195 ret = io_req_task_work_add(req, true);
3196 if (unlikely(ret)) {
3197 struct task_struct *tsk;
3199 /* queue just for cancelation */
3200 init_task_work(&req->task_work, io_req_task_cancel);
3201 tsk = io_wq_get_task(req->ctx->io_wq);
3202 task_work_add(tsk, &req->task_work, 0);
3203 wake_up_process(tsk);
3209 * This controls whether a given IO request should be armed for async page
3210 * based retry. If we return false here, the request is handed to the async
3211 * worker threads for retry. If we're doing buffered reads on a regular file,
3212 * we prepare a private wait_page_queue entry and retry the operation. This
3213 * will either succeed because the page is now uptodate and unlocked, or it
3214 * will register a callback when the page is unlocked at IO completion. Through
3215 * that callback, io_uring uses task_work to setup a retry of the operation.
3216 * That retry will attempt the buffered read again. The retry will generally
3217 * succeed, or in rare cases where it fails, we then fall back to using the
3218 * async worker threads for a blocking retry.
3220 static bool io_rw_should_retry(struct io_kiocb *req)
3222 struct io_async_rw *rw = req->async_data;
3223 struct wait_page_queue *wait = &rw->wpq;
3224 struct kiocb *kiocb = &req->rw.kiocb;
3226 /* never retry for NOWAIT, we just complete with -EAGAIN */
3227 if (req->flags & REQ_F_NOWAIT)
3230 /* Only for buffered IO */
3231 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3235 * just use poll if we can, and don't attempt if the fs doesn't
3236 * support callback based unlocks
3238 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3241 wait->wait.func = io_async_buf_func;
3242 wait->wait.private = req;
3243 wait->wait.flags = 0;
3244 INIT_LIST_HEAD(&wait->wait.entry);
3245 kiocb->ki_flags |= IOCB_WAITQ;
3246 kiocb->ki_flags &= ~IOCB_NOWAIT;
3247 kiocb->ki_waitq = wait;
3251 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3253 if (req->file->f_op->read_iter)
3254 return call_read_iter(req->file, &req->rw.kiocb, iter);
3255 else if (req->file->f_op->read)
3256 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3261 static int io_read(struct io_kiocb *req, bool force_nonblock,
3262 struct io_comp_state *cs)
3264 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3265 struct kiocb *kiocb = &req->rw.kiocb;
3266 struct iov_iter __iter, *iter = &__iter;
3267 struct io_async_rw *rw = req->async_data;
3268 ssize_t io_size, ret, ret2;
3275 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3278 iov_count = iov_iter_count(iter);
3280 req->result = io_size;
3283 /* Ensure we clear previously set non-block flag */
3284 if (!force_nonblock)
3285 kiocb->ki_flags &= ~IOCB_NOWAIT;
3287 kiocb->ki_flags |= IOCB_NOWAIT;
3290 /* If the file doesn't support async, just async punt */
3291 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3295 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3299 ret = io_iter_do_read(req, iter);
3303 } else if (ret == -EIOCBQUEUED) {
3306 } else if (ret == -EAGAIN) {
3307 /* IOPOLL retry should happen for io-wq threads */
3308 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3310 /* no retry on NONBLOCK marked file */
3311 if (req->file->f_flags & O_NONBLOCK)
3313 /* some cases will consume bytes even on error returns */
3314 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3317 } else if (ret < 0) {
3318 /* make sure -ERESTARTSYS -> -EINTR is done */
3322 /* read it all, or we did blocking attempt. no retry. */
3323 if (!iov_iter_count(iter) || !force_nonblock ||
3324 (req->file->f_flags & O_NONBLOCK))
3329 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3336 rw = req->async_data;
3337 /* it's copied and will be cleaned with ->io */
3339 /* now use our persistent iterator, if we aren't already */
3342 rw->bytes_done += ret;
3343 /* if we can retry, do so with the callbacks armed */
3344 if (!io_rw_should_retry(req)) {
3345 kiocb->ki_flags &= ~IOCB_WAITQ;
3350 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3351 * get -EIOCBQUEUED, then we'll get a notification when the desired
3352 * page gets unlocked. We can also get a partial read here, and if we
3353 * do, then just retry at the new offset.
3355 ret = io_iter_do_read(req, iter);
3356 if (ret == -EIOCBQUEUED) {
3359 } else if (ret > 0 && ret < io_size) {
3360 /* we got some bytes, but not all. retry. */
3364 kiocb_done(kiocb, ret, cs);
3367 /* it's reportedly faster than delegating the null check to kfree() */
3373 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3377 ret = io_prep_rw(req, sqe);
3381 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3384 /* either don't need iovec imported or already have it */
3385 if (!req->async_data)
3387 return io_rw_prep_async(req, WRITE);
3390 static int io_write(struct io_kiocb *req, bool force_nonblock,
3391 struct io_comp_state *cs)
3393 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3394 struct kiocb *kiocb = &req->rw.kiocb;
3395 struct iov_iter __iter, *iter = &__iter;
3396 struct io_async_rw *rw = req->async_data;
3398 ssize_t ret, ret2, io_size;
3403 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3406 iov_count = iov_iter_count(iter);
3408 req->result = io_size;
3410 /* Ensure we clear previously set non-block flag */
3411 if (!force_nonblock)
3412 kiocb->ki_flags &= ~IOCB_NOWAIT;
3414 kiocb->ki_flags |= IOCB_NOWAIT;
3416 /* If the file doesn't support async, just async punt */
3417 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3420 /* file path doesn't support NOWAIT for non-direct_IO */
3421 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3422 (req->flags & REQ_F_ISREG))
3425 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3430 * Open-code file_start_write here to grab freeze protection,
3431 * which will be released by another thread in
3432 * io_complete_rw(). Fool lockdep by telling it the lock got
3433 * released so that it doesn't complain about the held lock when
3434 * we return to userspace.
3436 if (req->flags & REQ_F_ISREG) {
3437 __sb_start_write(file_inode(req->file)->i_sb,
3438 SB_FREEZE_WRITE, true);
3439 __sb_writers_release(file_inode(req->file)->i_sb,
3442 kiocb->ki_flags |= IOCB_WRITE;
3444 if (req->file->f_op->write_iter)
3445 ret2 = call_write_iter(req->file, kiocb, iter);
3446 else if (req->file->f_op->write)
3447 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3452 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3453 * retry them without IOCB_NOWAIT.
3455 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3457 /* no retry on NONBLOCK marked file */
3458 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3460 if (!force_nonblock || ret2 != -EAGAIN) {
3461 /* IOPOLL retry should happen for io-wq threads */
3462 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3465 kiocb_done(kiocb, ret2, cs);
3468 /* some cases will consume bytes even on error returns */
3469 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3470 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3475 /* it's reportedly faster than delegating the null check to kfree() */
3481 static int __io_splice_prep(struct io_kiocb *req,
3482 const struct io_uring_sqe *sqe)
3484 struct io_splice* sp = &req->splice;
3485 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3487 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3491 sp->len = READ_ONCE(sqe->len);
3492 sp->flags = READ_ONCE(sqe->splice_flags);
3494 if (unlikely(sp->flags & ~valid_flags))
3497 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3498 (sp->flags & SPLICE_F_FD_IN_FIXED));
3501 req->flags |= REQ_F_NEED_CLEANUP;
3503 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3505 * Splice operation will be punted aync, and here need to
3506 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3508 io_req_init_async(req);
3509 req->work.flags |= IO_WQ_WORK_UNBOUND;
3515 static int io_tee_prep(struct io_kiocb *req,
3516 const struct io_uring_sqe *sqe)
3518 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3520 return __io_splice_prep(req, sqe);
3523 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3525 struct io_splice *sp = &req->splice;
3526 struct file *in = sp->file_in;
3527 struct file *out = sp->file_out;
3528 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3534 ret = do_tee(in, out, sp->len, flags);
3536 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3537 req->flags &= ~REQ_F_NEED_CLEANUP;
3540 req_set_fail_links(req);
3541 io_req_complete(req, ret);
3545 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3547 struct io_splice* sp = &req->splice;
3549 sp->off_in = READ_ONCE(sqe->splice_off_in);
3550 sp->off_out = READ_ONCE(sqe->off);
3551 return __io_splice_prep(req, sqe);
3554 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3556 struct io_splice *sp = &req->splice;
3557 struct file *in = sp->file_in;
3558 struct file *out = sp->file_out;
3559 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3560 loff_t *poff_in, *poff_out;
3566 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3567 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3570 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3572 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3573 req->flags &= ~REQ_F_NEED_CLEANUP;
3576 req_set_fail_links(req);
3577 io_req_complete(req, ret);
3582 * IORING_OP_NOP just posts a completion event, nothing else.
3584 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3586 struct io_ring_ctx *ctx = req->ctx;
3588 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3591 __io_req_complete(req, 0, 0, cs);
3595 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3597 struct io_ring_ctx *ctx = req->ctx;
3602 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3604 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3607 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3608 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3611 req->sync.off = READ_ONCE(sqe->off);
3612 req->sync.len = READ_ONCE(sqe->len);
3616 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3618 loff_t end = req->sync.off + req->sync.len;
3621 /* fsync always requires a blocking context */
3625 ret = vfs_fsync_range(req->file, req->sync.off,
3626 end > 0 ? end : LLONG_MAX,
3627 req->sync.flags & IORING_FSYNC_DATASYNC);
3629 req_set_fail_links(req);
3630 io_req_complete(req, ret);
3634 static int io_fallocate_prep(struct io_kiocb *req,
3635 const struct io_uring_sqe *sqe)
3637 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3639 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3642 req->sync.off = READ_ONCE(sqe->off);
3643 req->sync.len = READ_ONCE(sqe->addr);
3644 req->sync.mode = READ_ONCE(sqe->len);
3648 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3652 /* fallocate always requiring blocking context */
3655 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3658 req_set_fail_links(req);
3659 io_req_complete(req, ret);
3663 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3665 const char __user *fname;
3668 if (unlikely(sqe->ioprio || sqe->buf_index))
3670 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3673 /* open.how should be already initialised */
3674 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3675 req->open.how.flags |= O_LARGEFILE;
3677 req->open.dfd = READ_ONCE(sqe->fd);
3678 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3679 req->open.filename = getname(fname);
3680 if (IS_ERR(req->open.filename)) {
3681 ret = PTR_ERR(req->open.filename);
3682 req->open.filename = NULL;
3685 req->open.nofile = rlimit(RLIMIT_NOFILE);
3686 req->flags |= REQ_F_NEED_CLEANUP;
3690 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3694 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3696 mode = READ_ONCE(sqe->len);
3697 flags = READ_ONCE(sqe->open_flags);
3698 req->open.how = build_open_how(flags, mode);
3699 return __io_openat_prep(req, sqe);
3702 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3704 struct open_how __user *how;
3708 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3710 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3711 len = READ_ONCE(sqe->len);
3712 if (len < OPEN_HOW_SIZE_VER0)
3715 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3720 return __io_openat_prep(req, sqe);
3723 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3725 struct open_flags op;
3732 ret = build_open_flags(&req->open.how, &op);
3736 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3740 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3743 ret = PTR_ERR(file);
3745 fsnotify_open(file);
3746 fd_install(ret, file);
3749 putname(req->open.filename);
3750 req->flags &= ~REQ_F_NEED_CLEANUP;
3752 req_set_fail_links(req);
3753 io_req_complete(req, ret);
3757 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3759 return io_openat2(req, force_nonblock);
3762 static int io_remove_buffers_prep(struct io_kiocb *req,
3763 const struct io_uring_sqe *sqe)
3765 struct io_provide_buf *p = &req->pbuf;
3768 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3771 tmp = READ_ONCE(sqe->fd);
3772 if (!tmp || tmp > USHRT_MAX)
3775 memset(p, 0, sizeof(*p));
3777 p->bgid = READ_ONCE(sqe->buf_group);
3781 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3782 int bgid, unsigned nbufs)
3786 /* shouldn't happen */
3790 /* the head kbuf is the list itself */
3791 while (!list_empty(&buf->list)) {
3792 struct io_buffer *nxt;
3794 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3795 list_del(&nxt->list);
3802 idr_remove(&ctx->io_buffer_idr, bgid);
3807 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3808 struct io_comp_state *cs)
3810 struct io_provide_buf *p = &req->pbuf;
3811 struct io_ring_ctx *ctx = req->ctx;
3812 struct io_buffer *head;
3815 io_ring_submit_lock(ctx, !force_nonblock);
3817 lockdep_assert_held(&ctx->uring_lock);
3820 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3822 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3824 io_ring_submit_lock(ctx, !force_nonblock);
3826 req_set_fail_links(req);
3827 __io_req_complete(req, ret, 0, cs);
3831 static int io_provide_buffers_prep(struct io_kiocb *req,
3832 const struct io_uring_sqe *sqe)
3834 struct io_provide_buf *p = &req->pbuf;
3837 if (sqe->ioprio || sqe->rw_flags)
3840 tmp = READ_ONCE(sqe->fd);
3841 if (!tmp || tmp > USHRT_MAX)
3844 p->addr = READ_ONCE(sqe->addr);
3845 p->len = READ_ONCE(sqe->len);
3847 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3850 p->bgid = READ_ONCE(sqe->buf_group);
3851 tmp = READ_ONCE(sqe->off);
3852 if (tmp > USHRT_MAX)
3858 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3860 struct io_buffer *buf;
3861 u64 addr = pbuf->addr;
3862 int i, bid = pbuf->bid;
3864 for (i = 0; i < pbuf->nbufs; i++) {
3865 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3870 buf->len = pbuf->len;
3875 INIT_LIST_HEAD(&buf->list);
3878 list_add_tail(&buf->list, &(*head)->list);
3882 return i ? i : -ENOMEM;
3885 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3886 struct io_comp_state *cs)
3888 struct io_provide_buf *p = &req->pbuf;
3889 struct io_ring_ctx *ctx = req->ctx;
3890 struct io_buffer *head, *list;
3893 io_ring_submit_lock(ctx, !force_nonblock);
3895 lockdep_assert_held(&ctx->uring_lock);
3897 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3899 ret = io_add_buffers(p, &head);
3904 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3907 __io_remove_buffers(ctx, head, p->bgid, -1U);
3912 io_ring_submit_unlock(ctx, !force_nonblock);
3914 req_set_fail_links(req);
3915 __io_req_complete(req, ret, 0, cs);
3919 static int io_epoll_ctl_prep(struct io_kiocb *req,
3920 const struct io_uring_sqe *sqe)
3922 #if defined(CONFIG_EPOLL)
3923 if (sqe->ioprio || sqe->buf_index)
3925 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3928 req->epoll.epfd = READ_ONCE(sqe->fd);
3929 req->epoll.op = READ_ONCE(sqe->len);
3930 req->epoll.fd = READ_ONCE(sqe->off);
3932 if (ep_op_has_event(req->epoll.op)) {
3933 struct epoll_event __user *ev;
3935 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3936 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3946 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3947 struct io_comp_state *cs)
3949 #if defined(CONFIG_EPOLL)
3950 struct io_epoll *ie = &req->epoll;
3953 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3954 if (force_nonblock && ret == -EAGAIN)
3958 req_set_fail_links(req);
3959 __io_req_complete(req, ret, 0, cs);
3966 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3968 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3969 if (sqe->ioprio || sqe->buf_index || sqe->off)
3971 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3974 req->madvise.addr = READ_ONCE(sqe->addr);
3975 req->madvise.len = READ_ONCE(sqe->len);
3976 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3983 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3985 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3986 struct io_madvise *ma = &req->madvise;
3992 ret = do_madvise(ma->addr, ma->len, ma->advice);
3994 req_set_fail_links(req);
3995 io_req_complete(req, ret);
4002 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4004 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4006 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4009 req->fadvise.offset = READ_ONCE(sqe->off);
4010 req->fadvise.len = READ_ONCE(sqe->len);
4011 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4015 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4017 struct io_fadvise *fa = &req->fadvise;
4020 if (force_nonblock) {
4021 switch (fa->advice) {
4022 case POSIX_FADV_NORMAL:
4023 case POSIX_FADV_RANDOM:
4024 case POSIX_FADV_SEQUENTIAL:
4031 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4033 req_set_fail_links(req);
4034 io_req_complete(req, ret);
4038 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4040 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4042 if (sqe->ioprio || sqe->buf_index)
4044 if (req->flags & REQ_F_FIXED_FILE)
4047 req->statx.dfd = READ_ONCE(sqe->fd);
4048 req->statx.mask = READ_ONCE(sqe->len);
4049 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4050 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4051 req->statx.flags = READ_ONCE(sqe->statx_flags);
4056 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4058 struct io_statx *ctx = &req->statx;
4061 if (force_nonblock) {
4062 /* only need file table for an actual valid fd */
4063 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4064 req->flags |= REQ_F_NO_FILE_TABLE;
4068 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4072 req_set_fail_links(req);
4073 io_req_complete(req, ret);
4077 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4080 * If we queue this for async, it must not be cancellable. That would
4081 * leave the 'file' in an undeterminate state, and here need to modify
4082 * io_wq_work.flags, so initialize io_wq_work firstly.
4084 io_req_init_async(req);
4085 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4087 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4089 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4090 sqe->rw_flags || sqe->buf_index)
4092 if (req->flags & REQ_F_FIXED_FILE)
4095 req->close.fd = READ_ONCE(sqe->fd);
4096 if ((req->file && req->file->f_op == &io_uring_fops))
4099 req->close.put_file = NULL;
4103 static int io_close(struct io_kiocb *req, bool force_nonblock,
4104 struct io_comp_state *cs)
4106 struct io_close *close = &req->close;
4109 /* might be already done during nonblock submission */
4110 if (!close->put_file) {
4111 ret = __close_fd_get_file(close->fd, &close->put_file);
4113 return (ret == -ENOENT) ? -EBADF : ret;
4116 /* if the file has a flush method, be safe and punt to async */
4117 if (close->put_file->f_op->flush && force_nonblock) {
4118 /* was never set, but play safe */
4119 req->flags &= ~REQ_F_NOWAIT;
4120 /* avoid grabbing files - we don't need the files */
4121 req->flags |= REQ_F_NO_FILE_TABLE;
4125 /* No ->flush() or already async, safely close from here */
4126 ret = filp_close(close->put_file, req->work.files);
4128 req_set_fail_links(req);
4129 fput(close->put_file);
4130 close->put_file = NULL;
4131 __io_req_complete(req, ret, 0, cs);
4135 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4137 struct io_ring_ctx *ctx = req->ctx;
4142 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4144 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4147 req->sync.off = READ_ONCE(sqe->off);
4148 req->sync.len = READ_ONCE(sqe->len);
4149 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4153 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4157 /* sync_file_range always requires a blocking context */
4161 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4164 req_set_fail_links(req);
4165 io_req_complete(req, ret);
4169 #if defined(CONFIG_NET)
4170 static int io_setup_async_msg(struct io_kiocb *req,
4171 struct io_async_msghdr *kmsg)
4173 struct io_async_msghdr *async_msg = req->async_data;
4177 if (io_alloc_async_data(req)) {
4178 if (kmsg->iov != kmsg->fast_iov)
4182 async_msg = req->async_data;
4183 req->flags |= REQ_F_NEED_CLEANUP;
4184 memcpy(async_msg, kmsg, sizeof(*kmsg));
4188 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4189 struct io_async_msghdr *iomsg)
4191 iomsg->iov = iomsg->fast_iov;
4192 iomsg->msg.msg_name = &iomsg->addr;
4193 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4194 req->sr_msg.msg_flags, &iomsg->iov);
4197 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4199 struct io_async_msghdr *async_msg = req->async_data;
4200 struct io_sr_msg *sr = &req->sr_msg;
4203 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4206 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4207 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4208 sr->len = READ_ONCE(sqe->len);
4210 #ifdef CONFIG_COMPAT
4211 if (req->ctx->compat)
4212 sr->msg_flags |= MSG_CMSG_COMPAT;
4215 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4217 ret = io_sendmsg_copy_hdr(req, async_msg);
4219 req->flags |= REQ_F_NEED_CLEANUP;
4223 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4224 struct io_comp_state *cs)
4226 struct io_async_msghdr iomsg, *kmsg;
4227 struct socket *sock;
4231 sock = sock_from_file(req->file, &ret);
4232 if (unlikely(!sock))
4235 if (req->async_data) {
4236 kmsg = req->async_data;
4237 kmsg->msg.msg_name = &kmsg->addr;
4238 /* if iov is set, it's allocated already */
4240 kmsg->iov = kmsg->fast_iov;
4241 kmsg->msg.msg_iter.iov = kmsg->iov;
4243 ret = io_sendmsg_copy_hdr(req, &iomsg);
4249 flags = req->sr_msg.msg_flags;
4250 if (flags & MSG_DONTWAIT)
4251 req->flags |= REQ_F_NOWAIT;
4252 else if (force_nonblock)
4253 flags |= MSG_DONTWAIT;
4255 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4256 if (force_nonblock && ret == -EAGAIN)
4257 return io_setup_async_msg(req, kmsg);
4258 if (ret == -ERESTARTSYS)
4261 if (kmsg->iov != kmsg->fast_iov)
4263 req->flags &= ~REQ_F_NEED_CLEANUP;
4265 req_set_fail_links(req);
4266 __io_req_complete(req, ret, 0, cs);
4270 static int io_send(struct io_kiocb *req, bool force_nonblock,
4271 struct io_comp_state *cs)
4273 struct io_sr_msg *sr = &req->sr_msg;
4276 struct socket *sock;
4280 sock = sock_from_file(req->file, &ret);
4281 if (unlikely(!sock))
4284 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4288 msg.msg_name = NULL;
4289 msg.msg_control = NULL;
4290 msg.msg_controllen = 0;
4291 msg.msg_namelen = 0;
4293 flags = req->sr_msg.msg_flags;
4294 if (flags & MSG_DONTWAIT)
4295 req->flags |= REQ_F_NOWAIT;
4296 else if (force_nonblock)
4297 flags |= MSG_DONTWAIT;
4299 msg.msg_flags = flags;
4300 ret = sock_sendmsg(sock, &msg);
4301 if (force_nonblock && ret == -EAGAIN)
4303 if (ret == -ERESTARTSYS)
4307 req_set_fail_links(req);
4308 __io_req_complete(req, ret, 0, cs);
4312 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4313 struct io_async_msghdr *iomsg)
4315 struct io_sr_msg *sr = &req->sr_msg;
4316 struct iovec __user *uiov;
4320 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4321 &iomsg->uaddr, &uiov, &iov_len);
4325 if (req->flags & REQ_F_BUFFER_SELECT) {
4328 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4330 sr->len = iomsg->iov[0].iov_len;
4331 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4335 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4336 &iomsg->iov, &iomsg->msg.msg_iter,
4345 #ifdef CONFIG_COMPAT
4346 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4347 struct io_async_msghdr *iomsg)
4349 struct compat_msghdr __user *msg_compat;
4350 struct io_sr_msg *sr = &req->sr_msg;
4351 struct compat_iovec __user *uiov;
4356 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4357 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4362 uiov = compat_ptr(ptr);
4363 if (req->flags & REQ_F_BUFFER_SELECT) {
4364 compat_ssize_t clen;
4368 if (!access_ok(uiov, sizeof(*uiov)))
4370 if (__get_user(clen, &uiov->iov_len))
4374 sr->len = iomsg->iov[0].iov_len;
4377 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4378 UIO_FASTIOV, &iomsg->iov,
4379 &iomsg->msg.msg_iter, true);
4388 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4389 struct io_async_msghdr *iomsg)
4391 iomsg->msg.msg_name = &iomsg->addr;
4392 iomsg->iov = iomsg->fast_iov;
4394 #ifdef CONFIG_COMPAT
4395 if (req->ctx->compat)
4396 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4399 return __io_recvmsg_copy_hdr(req, iomsg);
4402 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4405 struct io_sr_msg *sr = &req->sr_msg;
4406 struct io_buffer *kbuf;
4408 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4413 req->flags |= REQ_F_BUFFER_SELECTED;
4417 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4419 return io_put_kbuf(req, req->sr_msg.kbuf);
4422 static int io_recvmsg_prep(struct io_kiocb *req,
4423 const struct io_uring_sqe *sqe)
4425 struct io_async_msghdr *async_msg = req->async_data;
4426 struct io_sr_msg *sr = &req->sr_msg;
4429 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4432 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4433 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4434 sr->len = READ_ONCE(sqe->len);
4435 sr->bgid = READ_ONCE(sqe->buf_group);
4437 #ifdef CONFIG_COMPAT
4438 if (req->ctx->compat)
4439 sr->msg_flags |= MSG_CMSG_COMPAT;
4442 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4444 ret = io_recvmsg_copy_hdr(req, async_msg);
4446 req->flags |= REQ_F_NEED_CLEANUP;
4450 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4451 struct io_comp_state *cs)
4453 struct io_async_msghdr iomsg, *kmsg;
4454 struct socket *sock;
4455 struct io_buffer *kbuf;
4457 int ret, cflags = 0;
4459 sock = sock_from_file(req->file, &ret);
4460 if (unlikely(!sock))
4463 if (req->async_data) {
4464 kmsg = req->async_data;
4465 kmsg->msg.msg_name = &kmsg->addr;
4466 /* if iov is set, it's allocated already */
4468 kmsg->iov = kmsg->fast_iov;
4469 kmsg->msg.msg_iter.iov = kmsg->iov;
4471 ret = io_recvmsg_copy_hdr(req, &iomsg);
4477 if (req->flags & REQ_F_BUFFER_SELECT) {
4478 kbuf = io_recv_buffer_select(req, !force_nonblock);
4480 return PTR_ERR(kbuf);
4481 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4482 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4483 1, req->sr_msg.len);
4486 flags = req->sr_msg.msg_flags;
4487 if (flags & MSG_DONTWAIT)
4488 req->flags |= REQ_F_NOWAIT;
4489 else if (force_nonblock)
4490 flags |= MSG_DONTWAIT;
4492 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4493 kmsg->uaddr, flags);
4494 if (force_nonblock && ret == -EAGAIN)
4495 return io_setup_async_msg(req, kmsg);
4496 if (ret == -ERESTARTSYS)
4499 if (req->flags & REQ_F_BUFFER_SELECTED)
4500 cflags = io_put_recv_kbuf(req);
4501 if (kmsg->iov != kmsg->fast_iov)
4503 req->flags &= ~REQ_F_NEED_CLEANUP;
4505 req_set_fail_links(req);
4506 __io_req_complete(req, ret, cflags, cs);
4510 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4511 struct io_comp_state *cs)
4513 struct io_buffer *kbuf;
4514 struct io_sr_msg *sr = &req->sr_msg;
4516 void __user *buf = sr->buf;
4517 struct socket *sock;
4520 int ret, cflags = 0;
4522 sock = sock_from_file(req->file, &ret);
4523 if (unlikely(!sock))
4526 if (req->flags & REQ_F_BUFFER_SELECT) {
4527 kbuf = io_recv_buffer_select(req, !force_nonblock);
4529 return PTR_ERR(kbuf);
4530 buf = u64_to_user_ptr(kbuf->addr);
4533 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4537 msg.msg_name = NULL;
4538 msg.msg_control = NULL;
4539 msg.msg_controllen = 0;
4540 msg.msg_namelen = 0;
4541 msg.msg_iocb = NULL;
4544 flags = req->sr_msg.msg_flags;
4545 if (flags & MSG_DONTWAIT)
4546 req->flags |= REQ_F_NOWAIT;
4547 else if (force_nonblock)
4548 flags |= MSG_DONTWAIT;
4550 ret = sock_recvmsg(sock, &msg, flags);
4551 if (force_nonblock && ret == -EAGAIN)
4553 if (ret == -ERESTARTSYS)
4556 if (req->flags & REQ_F_BUFFER_SELECTED)
4557 cflags = io_put_recv_kbuf(req);
4559 req_set_fail_links(req);
4560 __io_req_complete(req, ret, cflags, cs);
4564 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4566 struct io_accept *accept = &req->accept;
4568 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4570 if (sqe->ioprio || sqe->len || sqe->buf_index)
4573 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4574 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4575 accept->flags = READ_ONCE(sqe->accept_flags);
4576 accept->nofile = rlimit(RLIMIT_NOFILE);
4580 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4581 struct io_comp_state *cs)
4583 struct io_accept *accept = &req->accept;
4584 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4587 if (req->file->f_flags & O_NONBLOCK)
4588 req->flags |= REQ_F_NOWAIT;
4590 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4591 accept->addr_len, accept->flags,
4593 if (ret == -EAGAIN && force_nonblock)
4596 if (ret == -ERESTARTSYS)
4598 req_set_fail_links(req);
4600 __io_req_complete(req, ret, 0, cs);
4604 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4606 struct io_connect *conn = &req->connect;
4607 struct io_async_connect *io = req->async_data;
4609 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4611 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4614 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4615 conn->addr_len = READ_ONCE(sqe->addr2);
4620 return move_addr_to_kernel(conn->addr, conn->addr_len,
4624 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4625 struct io_comp_state *cs)
4627 struct io_async_connect __io, *io;
4628 unsigned file_flags;
4631 if (req->async_data) {
4632 io = req->async_data;
4634 ret = move_addr_to_kernel(req->connect.addr,
4635 req->connect.addr_len,
4642 file_flags = force_nonblock ? O_NONBLOCK : 0;
4644 ret = __sys_connect_file(req->file, &io->address,
4645 req->connect.addr_len, file_flags);
4646 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4647 if (req->async_data)
4649 if (io_alloc_async_data(req)) {
4653 io = req->async_data;
4654 memcpy(req->async_data, &__io, sizeof(__io));
4657 if (ret == -ERESTARTSYS)
4661 req_set_fail_links(req);
4662 __io_req_complete(req, ret, 0, cs);
4665 #else /* !CONFIG_NET */
4666 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4671 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4672 struct io_comp_state *cs)
4677 static int io_send(struct io_kiocb *req, bool force_nonblock,
4678 struct io_comp_state *cs)
4683 static int io_recvmsg_prep(struct io_kiocb *req,
4684 const struct io_uring_sqe *sqe)
4689 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4690 struct io_comp_state *cs)
4695 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4696 struct io_comp_state *cs)
4701 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4706 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4707 struct io_comp_state *cs)
4712 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4717 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4718 struct io_comp_state *cs)
4722 #endif /* CONFIG_NET */
4724 struct io_poll_table {
4725 struct poll_table_struct pt;
4726 struct io_kiocb *req;
4730 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4731 __poll_t mask, task_work_func_t func)
4736 /* for instances that support it check for an event match first: */
4737 if (mask && !(mask & poll->events))
4740 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4742 list_del_init(&poll->wait.entry);
4745 init_task_work(&req->task_work, func);
4746 percpu_ref_get(&req->ctx->refs);
4749 * If we using the signalfd wait_queue_head for this wakeup, then
4750 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4751 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4752 * either, as the normal wakeup will suffice.
4754 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4757 * If this fails, then the task is exiting. When a task exits, the
4758 * work gets canceled, so just cancel this request as well instead
4759 * of executing it. We can't safely execute it anyway, as we may not
4760 * have the needed state needed for it anyway.
4762 ret = io_req_task_work_add(req, twa_signal_ok);
4763 if (unlikely(ret)) {
4764 struct task_struct *tsk;
4766 WRITE_ONCE(poll->canceled, true);
4767 tsk = io_wq_get_task(req->ctx->io_wq);
4768 task_work_add(tsk, &req->task_work, 0);
4769 wake_up_process(tsk);
4774 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4775 __acquires(&req->ctx->completion_lock)
4777 struct io_ring_ctx *ctx = req->ctx;
4779 if (!req->result && !READ_ONCE(poll->canceled)) {
4780 struct poll_table_struct pt = { ._key = poll->events };
4782 req->result = vfs_poll(req->file, &pt) & poll->events;
4785 spin_lock_irq(&ctx->completion_lock);
4786 if (!req->result && !READ_ONCE(poll->canceled)) {
4787 add_wait_queue(poll->head, &poll->wait);
4794 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4796 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4797 if (req->opcode == IORING_OP_POLL_ADD)
4798 return req->async_data;
4799 return req->apoll->double_poll;
4802 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4804 if (req->opcode == IORING_OP_POLL_ADD)
4806 return &req->apoll->poll;
4809 static void io_poll_remove_double(struct io_kiocb *req)
4811 struct io_poll_iocb *poll = io_poll_get_double(req);
4813 lockdep_assert_held(&req->ctx->completion_lock);
4815 if (poll && poll->head) {
4816 struct wait_queue_head *head = poll->head;
4818 spin_lock(&head->lock);
4819 list_del_init(&poll->wait.entry);
4820 if (poll->wait.private)
4821 refcount_dec(&req->refs);
4823 spin_unlock(&head->lock);
4827 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4829 struct io_ring_ctx *ctx = req->ctx;
4831 io_poll_remove_double(req);
4832 req->poll.done = true;
4833 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4834 io_commit_cqring(ctx);
4837 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4839 struct io_ring_ctx *ctx = req->ctx;
4841 if (io_poll_rewait(req, &req->poll)) {
4842 spin_unlock_irq(&ctx->completion_lock);
4846 hash_del(&req->hash_node);
4847 io_poll_complete(req, req->result, 0);
4848 req->flags |= REQ_F_COMP_LOCKED;
4849 *nxt = io_put_req_find_next(req);
4850 spin_unlock_irq(&ctx->completion_lock);
4852 io_cqring_ev_posted(ctx);
4855 static void io_poll_task_func(struct callback_head *cb)
4857 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4858 struct io_ring_ctx *ctx = req->ctx;
4859 struct io_kiocb *nxt = NULL;
4861 io_poll_task_handler(req, &nxt);
4863 __io_req_task_submit(nxt);
4864 percpu_ref_put(&ctx->refs);
4867 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4868 int sync, void *key)
4870 struct io_kiocb *req = wait->private;
4871 struct io_poll_iocb *poll = io_poll_get_single(req);
4872 __poll_t mask = key_to_poll(key);
4874 /* for instances that support it check for an event match first: */
4875 if (mask && !(mask & poll->events))
4878 list_del_init(&wait->entry);
4880 if (poll && poll->head) {
4883 spin_lock(&poll->head->lock);
4884 done = list_empty(&poll->wait.entry);
4886 list_del_init(&poll->wait.entry);
4887 /* make sure double remove sees this as being gone */
4888 wait->private = NULL;
4889 spin_unlock(&poll->head->lock);
4891 __io_async_wake(req, poll, mask, io_poll_task_func);
4893 refcount_dec(&req->refs);
4897 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4898 wait_queue_func_t wake_func)
4902 poll->canceled = false;
4903 poll->events = events;
4904 INIT_LIST_HEAD(&poll->wait.entry);
4905 init_waitqueue_func_entry(&poll->wait, wake_func);
4908 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4909 struct wait_queue_head *head,
4910 struct io_poll_iocb **poll_ptr)
4912 struct io_kiocb *req = pt->req;
4915 * If poll->head is already set, it's because the file being polled
4916 * uses multiple waitqueues for poll handling (eg one for read, one
4917 * for write). Setup a separate io_poll_iocb if this happens.
4919 if (unlikely(poll->head)) {
4920 /* already have a 2nd entry, fail a third attempt */
4922 pt->error = -EINVAL;
4925 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4927 pt->error = -ENOMEM;
4930 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4931 refcount_inc(&req->refs);
4932 poll->wait.private = req;
4939 if (poll->events & EPOLLEXCLUSIVE)
4940 add_wait_queue_exclusive(head, &poll->wait);
4942 add_wait_queue(head, &poll->wait);
4945 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4946 struct poll_table_struct *p)
4948 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4949 struct async_poll *apoll = pt->req->apoll;
4951 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4954 static void io_async_task_func(struct callback_head *cb)
4956 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4957 struct async_poll *apoll = req->apoll;
4958 struct io_ring_ctx *ctx = req->ctx;
4960 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4962 if (io_poll_rewait(req, &apoll->poll)) {
4963 spin_unlock_irq(&ctx->completion_lock);
4964 percpu_ref_put(&ctx->refs);
4968 /* If req is still hashed, it cannot have been canceled. Don't check. */
4969 if (hash_hashed(&req->hash_node))
4970 hash_del(&req->hash_node);
4972 io_poll_remove_double(req);
4973 spin_unlock_irq(&ctx->completion_lock);
4975 if (!READ_ONCE(apoll->poll.canceled))
4976 __io_req_task_submit(req);
4978 __io_req_task_cancel(req, -ECANCELED);
4980 percpu_ref_put(&ctx->refs);
4981 kfree(apoll->double_poll);
4985 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4988 struct io_kiocb *req = wait->private;
4989 struct io_poll_iocb *poll = &req->apoll->poll;
4991 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4994 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4997 static void io_poll_req_insert(struct io_kiocb *req)
4999 struct io_ring_ctx *ctx = req->ctx;
5000 struct hlist_head *list;
5002 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5003 hlist_add_head(&req->hash_node, list);
5006 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5007 struct io_poll_iocb *poll,
5008 struct io_poll_table *ipt, __poll_t mask,
5009 wait_queue_func_t wake_func)
5010 __acquires(&ctx->completion_lock)
5012 struct io_ring_ctx *ctx = req->ctx;
5013 bool cancel = false;
5015 io_init_poll_iocb(poll, mask, wake_func);
5016 poll->file = req->file;
5017 poll->wait.private = req;
5019 ipt->pt._key = mask;
5021 ipt->error = -EINVAL;
5023 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5025 spin_lock_irq(&ctx->completion_lock);
5026 if (likely(poll->head)) {
5027 spin_lock(&poll->head->lock);
5028 if (unlikely(list_empty(&poll->wait.entry))) {
5034 if (mask || ipt->error)
5035 list_del_init(&poll->wait.entry);
5037 WRITE_ONCE(poll->canceled, true);
5038 else if (!poll->done) /* actually waiting for an event */
5039 io_poll_req_insert(req);
5040 spin_unlock(&poll->head->lock);
5046 static bool io_arm_poll_handler(struct io_kiocb *req)
5048 const struct io_op_def *def = &io_op_defs[req->opcode];
5049 struct io_ring_ctx *ctx = req->ctx;
5050 struct async_poll *apoll;
5051 struct io_poll_table ipt;
5055 if (!req->file || !file_can_poll(req->file))
5057 if (req->flags & REQ_F_POLLED)
5061 else if (def->pollout)
5065 /* if we can't nonblock try, then no point in arming a poll handler */
5066 if (!io_file_supports_async(req->file, rw))
5069 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5070 if (unlikely(!apoll))
5072 apoll->double_poll = NULL;
5074 req->flags |= REQ_F_POLLED;
5076 INIT_HLIST_NODE(&req->hash_node);
5080 mask |= POLLIN | POLLRDNORM;
5082 mask |= POLLOUT | POLLWRNORM;
5083 mask |= POLLERR | POLLPRI;
5085 ipt.pt._qproc = io_async_queue_proc;
5087 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5089 if (ret || ipt.error) {
5090 io_poll_remove_double(req);
5091 spin_unlock_irq(&ctx->completion_lock);
5092 kfree(apoll->double_poll);
5096 spin_unlock_irq(&ctx->completion_lock);
5097 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5098 apoll->poll.events);
5102 static bool __io_poll_remove_one(struct io_kiocb *req,
5103 struct io_poll_iocb *poll)
5105 bool do_complete = false;
5107 spin_lock(&poll->head->lock);
5108 WRITE_ONCE(poll->canceled, true);
5109 if (!list_empty(&poll->wait.entry)) {
5110 list_del_init(&poll->wait.entry);
5113 spin_unlock(&poll->head->lock);
5114 hash_del(&req->hash_node);
5118 static bool io_poll_remove_one(struct io_kiocb *req)
5122 io_poll_remove_double(req);
5124 if (req->opcode == IORING_OP_POLL_ADD) {
5125 do_complete = __io_poll_remove_one(req, &req->poll);
5127 struct async_poll *apoll = req->apoll;
5129 /* non-poll requests have submit ref still */
5130 do_complete = __io_poll_remove_one(req, &apoll->poll);
5133 kfree(apoll->double_poll);
5139 io_cqring_fill_event(req, -ECANCELED);
5140 io_commit_cqring(req->ctx);
5141 req->flags |= REQ_F_COMP_LOCKED;
5142 req_set_fail_links(req);
5150 * Returns true if we found and killed one or more poll requests
5152 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5154 struct hlist_node *tmp;
5155 struct io_kiocb *req;
5158 spin_lock_irq(&ctx->completion_lock);
5159 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5160 struct hlist_head *list;
5162 list = &ctx->cancel_hash[i];
5163 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5164 if (io_task_match(req, tsk))
5165 posted += io_poll_remove_one(req);
5168 spin_unlock_irq(&ctx->completion_lock);
5171 io_cqring_ev_posted(ctx);
5176 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5178 struct hlist_head *list;
5179 struct io_kiocb *req;
5181 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5182 hlist_for_each_entry(req, list, hash_node) {
5183 if (sqe_addr != req->user_data)
5185 if (io_poll_remove_one(req))
5193 static int io_poll_remove_prep(struct io_kiocb *req,
5194 const struct io_uring_sqe *sqe)
5196 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5198 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5202 req->poll.addr = READ_ONCE(sqe->addr);
5207 * Find a running poll command that matches one specified in sqe->addr,
5208 * and remove it if found.
5210 static int io_poll_remove(struct io_kiocb *req)
5212 struct io_ring_ctx *ctx = req->ctx;
5216 addr = req->poll.addr;
5217 spin_lock_irq(&ctx->completion_lock);
5218 ret = io_poll_cancel(ctx, addr);
5219 spin_unlock_irq(&ctx->completion_lock);
5222 req_set_fail_links(req);
5223 io_req_complete(req, ret);
5227 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5230 struct io_kiocb *req = wait->private;
5231 struct io_poll_iocb *poll = &req->poll;
5233 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5236 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5237 struct poll_table_struct *p)
5239 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5241 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5244 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5246 struct io_poll_iocb *poll = &req->poll;
5249 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5251 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5256 events = READ_ONCE(sqe->poll32_events);
5258 events = swahw32(events);
5260 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5261 (events & EPOLLEXCLUSIVE);
5265 static int io_poll_add(struct io_kiocb *req)
5267 struct io_poll_iocb *poll = &req->poll;
5268 struct io_ring_ctx *ctx = req->ctx;
5269 struct io_poll_table ipt;
5272 INIT_HLIST_NODE(&req->hash_node);
5273 ipt.pt._qproc = io_poll_queue_proc;
5275 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5278 if (mask) { /* no async, we'd stolen it */
5280 io_poll_complete(req, mask, 0);
5282 spin_unlock_irq(&ctx->completion_lock);
5285 io_cqring_ev_posted(ctx);
5291 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5293 struct io_timeout_data *data = container_of(timer,
5294 struct io_timeout_data, timer);
5295 struct io_kiocb *req = data->req;
5296 struct io_ring_ctx *ctx = req->ctx;
5297 unsigned long flags;
5299 spin_lock_irqsave(&ctx->completion_lock, flags);
5300 list_del_init(&req->timeout.list);
5301 atomic_set(&req->ctx->cq_timeouts,
5302 atomic_read(&req->ctx->cq_timeouts) + 1);
5304 io_cqring_fill_event(req, -ETIME);
5305 io_commit_cqring(ctx);
5306 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5308 io_cqring_ev_posted(ctx);
5309 req_set_fail_links(req);
5311 return HRTIMER_NORESTART;
5314 static int __io_timeout_cancel(struct io_kiocb *req)
5316 struct io_timeout_data *io = req->async_data;
5319 ret = hrtimer_try_to_cancel(&io->timer);
5322 list_del_init(&req->timeout.list);
5324 req_set_fail_links(req);
5325 req->flags |= REQ_F_COMP_LOCKED;
5326 io_cqring_fill_event(req, -ECANCELED);
5331 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5333 struct io_kiocb *req;
5336 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5337 if (user_data == req->user_data) {
5346 return __io_timeout_cancel(req);
5349 static int io_timeout_remove_prep(struct io_kiocb *req,
5350 const struct io_uring_sqe *sqe)
5352 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5354 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5356 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5359 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5364 * Remove or update an existing timeout command
5366 static int io_timeout_remove(struct io_kiocb *req)
5368 struct io_ring_ctx *ctx = req->ctx;
5371 spin_lock_irq(&ctx->completion_lock);
5372 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5374 io_cqring_fill_event(req, ret);
5375 io_commit_cqring(ctx);
5376 spin_unlock_irq(&ctx->completion_lock);
5377 io_cqring_ev_posted(ctx);
5379 req_set_fail_links(req);
5384 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5385 bool is_timeout_link)
5387 struct io_timeout_data *data;
5389 u32 off = READ_ONCE(sqe->off);
5391 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5393 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5395 if (off && is_timeout_link)
5397 flags = READ_ONCE(sqe->timeout_flags);
5398 if (flags & ~IORING_TIMEOUT_ABS)
5401 req->timeout.off = off;
5403 if (!req->async_data && io_alloc_async_data(req))
5406 data = req->async_data;
5409 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5412 if (flags & IORING_TIMEOUT_ABS)
5413 data->mode = HRTIMER_MODE_ABS;
5415 data->mode = HRTIMER_MODE_REL;
5417 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5421 static int io_timeout(struct io_kiocb *req)
5423 struct io_ring_ctx *ctx = req->ctx;
5424 struct io_timeout_data *data = req->async_data;
5425 struct list_head *entry;
5426 u32 tail, off = req->timeout.off;
5428 spin_lock_irq(&ctx->completion_lock);
5431 * sqe->off holds how many events that need to occur for this
5432 * timeout event to be satisfied. If it isn't set, then this is
5433 * a pure timeout request, sequence isn't used.
5435 if (io_is_timeout_noseq(req)) {
5436 entry = ctx->timeout_list.prev;
5440 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5441 req->timeout.target_seq = tail + off;
5444 * Insertion sort, ensuring the first entry in the list is always
5445 * the one we need first.
5447 list_for_each_prev(entry, &ctx->timeout_list) {
5448 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5451 if (io_is_timeout_noseq(nxt))
5453 /* nxt.seq is behind @tail, otherwise would've been completed */
5454 if (off >= nxt->timeout.target_seq - tail)
5458 list_add(&req->timeout.list, entry);
5459 data->timer.function = io_timeout_fn;
5460 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5461 spin_unlock_irq(&ctx->completion_lock);
5465 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5467 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5469 return req->user_data == (unsigned long) data;
5472 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5474 enum io_wq_cancel cancel_ret;
5477 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5478 switch (cancel_ret) {
5479 case IO_WQ_CANCEL_OK:
5482 case IO_WQ_CANCEL_RUNNING:
5485 case IO_WQ_CANCEL_NOTFOUND:
5493 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5494 struct io_kiocb *req, __u64 sqe_addr,
5497 unsigned long flags;
5500 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5501 if (ret != -ENOENT) {
5502 spin_lock_irqsave(&ctx->completion_lock, flags);
5506 spin_lock_irqsave(&ctx->completion_lock, flags);
5507 ret = io_timeout_cancel(ctx, sqe_addr);
5510 ret = io_poll_cancel(ctx, sqe_addr);
5514 io_cqring_fill_event(req, ret);
5515 io_commit_cqring(ctx);
5516 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5517 io_cqring_ev_posted(ctx);
5520 req_set_fail_links(req);
5524 static int io_async_cancel_prep(struct io_kiocb *req,
5525 const struct io_uring_sqe *sqe)
5527 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5529 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5531 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5534 req->cancel.addr = READ_ONCE(sqe->addr);
5538 static int io_async_cancel(struct io_kiocb *req)
5540 struct io_ring_ctx *ctx = req->ctx;
5542 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5546 static int io_files_update_prep(struct io_kiocb *req,
5547 const struct io_uring_sqe *sqe)
5549 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5551 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5553 if (sqe->ioprio || sqe->rw_flags)
5556 req->files_update.offset = READ_ONCE(sqe->off);
5557 req->files_update.nr_args = READ_ONCE(sqe->len);
5558 if (!req->files_update.nr_args)
5560 req->files_update.arg = READ_ONCE(sqe->addr);
5564 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5565 struct io_comp_state *cs)
5567 struct io_ring_ctx *ctx = req->ctx;
5568 struct io_uring_files_update up;
5574 up.offset = req->files_update.offset;
5575 up.fds = req->files_update.arg;
5577 mutex_lock(&ctx->uring_lock);
5578 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5579 mutex_unlock(&ctx->uring_lock);
5582 req_set_fail_links(req);
5583 __io_req_complete(req, ret, 0, cs);
5587 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5589 switch (req->opcode) {
5592 case IORING_OP_READV:
5593 case IORING_OP_READ_FIXED:
5594 case IORING_OP_READ:
5595 return io_read_prep(req, sqe);
5596 case IORING_OP_WRITEV:
5597 case IORING_OP_WRITE_FIXED:
5598 case IORING_OP_WRITE:
5599 return io_write_prep(req, sqe);
5600 case IORING_OP_POLL_ADD:
5601 return io_poll_add_prep(req, sqe);
5602 case IORING_OP_POLL_REMOVE:
5603 return io_poll_remove_prep(req, sqe);
5604 case IORING_OP_FSYNC:
5605 return io_prep_fsync(req, sqe);
5606 case IORING_OP_SYNC_FILE_RANGE:
5607 return io_prep_sfr(req, sqe);
5608 case IORING_OP_SENDMSG:
5609 case IORING_OP_SEND:
5610 return io_sendmsg_prep(req, sqe);
5611 case IORING_OP_RECVMSG:
5612 case IORING_OP_RECV:
5613 return io_recvmsg_prep(req, sqe);
5614 case IORING_OP_CONNECT:
5615 return io_connect_prep(req, sqe);
5616 case IORING_OP_TIMEOUT:
5617 return io_timeout_prep(req, sqe, false);
5618 case IORING_OP_TIMEOUT_REMOVE:
5619 return io_timeout_remove_prep(req, sqe);
5620 case IORING_OP_ASYNC_CANCEL:
5621 return io_async_cancel_prep(req, sqe);
5622 case IORING_OP_LINK_TIMEOUT:
5623 return io_timeout_prep(req, sqe, true);
5624 case IORING_OP_ACCEPT:
5625 return io_accept_prep(req, sqe);
5626 case IORING_OP_FALLOCATE:
5627 return io_fallocate_prep(req, sqe);
5628 case IORING_OP_OPENAT:
5629 return io_openat_prep(req, sqe);
5630 case IORING_OP_CLOSE:
5631 return io_close_prep(req, sqe);
5632 case IORING_OP_FILES_UPDATE:
5633 return io_files_update_prep(req, sqe);
5634 case IORING_OP_STATX:
5635 return io_statx_prep(req, sqe);
5636 case IORING_OP_FADVISE:
5637 return io_fadvise_prep(req, sqe);
5638 case IORING_OP_MADVISE:
5639 return io_madvise_prep(req, sqe);
5640 case IORING_OP_OPENAT2:
5641 return io_openat2_prep(req, sqe);
5642 case IORING_OP_EPOLL_CTL:
5643 return io_epoll_ctl_prep(req, sqe);
5644 case IORING_OP_SPLICE:
5645 return io_splice_prep(req, sqe);
5646 case IORING_OP_PROVIDE_BUFFERS:
5647 return io_provide_buffers_prep(req, sqe);
5648 case IORING_OP_REMOVE_BUFFERS:
5649 return io_remove_buffers_prep(req, sqe);
5651 return io_tee_prep(req, sqe);
5654 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5659 static int io_req_defer_prep(struct io_kiocb *req,
5660 const struct io_uring_sqe *sqe)
5664 if (io_alloc_async_data(req))
5666 return io_req_prep(req, sqe);
5669 static u32 io_get_sequence(struct io_kiocb *req)
5671 struct io_kiocb *pos;
5672 struct io_ring_ctx *ctx = req->ctx;
5673 u32 total_submitted, nr_reqs = 1;
5675 if (req->flags & REQ_F_LINK_HEAD)
5676 list_for_each_entry(pos, &req->link_list, link_list)
5679 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5680 return total_submitted - nr_reqs;
5683 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5685 struct io_ring_ctx *ctx = req->ctx;
5686 struct io_defer_entry *de;
5690 /* Still need defer if there is pending req in defer list. */
5691 if (likely(list_empty_careful(&ctx->defer_list) &&
5692 !(req->flags & REQ_F_IO_DRAIN)))
5695 seq = io_get_sequence(req);
5696 /* Still a chance to pass the sequence check */
5697 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5700 if (!req->async_data) {
5701 ret = io_req_defer_prep(req, sqe);
5705 io_prep_async_link(req);
5706 de = kmalloc(sizeof(*de), GFP_KERNEL);
5710 spin_lock_irq(&ctx->completion_lock);
5711 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5712 spin_unlock_irq(&ctx->completion_lock);
5714 io_queue_async_work(req);
5715 return -EIOCBQUEUED;
5718 trace_io_uring_defer(ctx, req, req->user_data);
5721 list_add_tail(&de->list, &ctx->defer_list);
5722 spin_unlock_irq(&ctx->completion_lock);
5723 return -EIOCBQUEUED;
5726 static void io_req_drop_files(struct io_kiocb *req)
5728 struct io_ring_ctx *ctx = req->ctx;
5729 unsigned long flags;
5731 spin_lock_irqsave(&ctx->inflight_lock, flags);
5732 list_del(&req->inflight_entry);
5733 if (waitqueue_active(&ctx->inflight_wait))
5734 wake_up(&ctx->inflight_wait);
5735 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5736 req->flags &= ~REQ_F_INFLIGHT;
5737 put_files_struct(req->work.files);
5738 put_nsproxy(req->work.nsproxy);
5739 req->work.files = NULL;
5742 static void __io_clean_op(struct io_kiocb *req)
5744 if (req->flags & REQ_F_BUFFER_SELECTED) {
5745 switch (req->opcode) {
5746 case IORING_OP_READV:
5747 case IORING_OP_READ_FIXED:
5748 case IORING_OP_READ:
5749 kfree((void *)(unsigned long)req->rw.addr);
5751 case IORING_OP_RECVMSG:
5752 case IORING_OP_RECV:
5753 kfree(req->sr_msg.kbuf);
5756 req->flags &= ~REQ_F_BUFFER_SELECTED;
5759 if (req->flags & REQ_F_NEED_CLEANUP) {
5760 switch (req->opcode) {
5761 case IORING_OP_READV:
5762 case IORING_OP_READ_FIXED:
5763 case IORING_OP_READ:
5764 case IORING_OP_WRITEV:
5765 case IORING_OP_WRITE_FIXED:
5766 case IORING_OP_WRITE: {
5767 struct io_async_rw *io = req->async_data;
5769 kfree(io->free_iovec);
5772 case IORING_OP_RECVMSG:
5773 case IORING_OP_SENDMSG: {
5774 struct io_async_msghdr *io = req->async_data;
5775 if (io->iov != io->fast_iov)
5779 case IORING_OP_SPLICE:
5781 io_put_file(req, req->splice.file_in,
5782 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5784 case IORING_OP_OPENAT:
5785 case IORING_OP_OPENAT2:
5786 if (req->open.filename)
5787 putname(req->open.filename);
5790 req->flags &= ~REQ_F_NEED_CLEANUP;
5793 if (req->flags & REQ_F_INFLIGHT)
5794 io_req_drop_files(req);
5797 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5798 struct io_comp_state *cs)
5800 struct io_ring_ctx *ctx = req->ctx;
5803 switch (req->opcode) {
5805 ret = io_nop(req, cs);
5807 case IORING_OP_READV:
5808 case IORING_OP_READ_FIXED:
5809 case IORING_OP_READ:
5810 ret = io_read(req, force_nonblock, cs);
5812 case IORING_OP_WRITEV:
5813 case IORING_OP_WRITE_FIXED:
5814 case IORING_OP_WRITE:
5815 ret = io_write(req, force_nonblock, cs);
5817 case IORING_OP_FSYNC:
5818 ret = io_fsync(req, force_nonblock);
5820 case IORING_OP_POLL_ADD:
5821 ret = io_poll_add(req);
5823 case IORING_OP_POLL_REMOVE:
5824 ret = io_poll_remove(req);
5826 case IORING_OP_SYNC_FILE_RANGE:
5827 ret = io_sync_file_range(req, force_nonblock);
5829 case IORING_OP_SENDMSG:
5830 ret = io_sendmsg(req, force_nonblock, cs);
5832 case IORING_OP_SEND:
5833 ret = io_send(req, force_nonblock, cs);
5835 case IORING_OP_RECVMSG:
5836 ret = io_recvmsg(req, force_nonblock, cs);
5838 case IORING_OP_RECV:
5839 ret = io_recv(req, force_nonblock, cs);
5841 case IORING_OP_TIMEOUT:
5842 ret = io_timeout(req);
5844 case IORING_OP_TIMEOUT_REMOVE:
5845 ret = io_timeout_remove(req);
5847 case IORING_OP_ACCEPT:
5848 ret = io_accept(req, force_nonblock, cs);
5850 case IORING_OP_CONNECT:
5851 ret = io_connect(req, force_nonblock, cs);
5853 case IORING_OP_ASYNC_CANCEL:
5854 ret = io_async_cancel(req);
5856 case IORING_OP_FALLOCATE:
5857 ret = io_fallocate(req, force_nonblock);
5859 case IORING_OP_OPENAT:
5860 ret = io_openat(req, force_nonblock);
5862 case IORING_OP_CLOSE:
5863 ret = io_close(req, force_nonblock, cs);
5865 case IORING_OP_FILES_UPDATE:
5866 ret = io_files_update(req, force_nonblock, cs);
5868 case IORING_OP_STATX:
5869 ret = io_statx(req, force_nonblock);
5871 case IORING_OP_FADVISE:
5872 ret = io_fadvise(req, force_nonblock);
5874 case IORING_OP_MADVISE:
5875 ret = io_madvise(req, force_nonblock);
5877 case IORING_OP_OPENAT2:
5878 ret = io_openat2(req, force_nonblock);
5880 case IORING_OP_EPOLL_CTL:
5881 ret = io_epoll_ctl(req, force_nonblock, cs);
5883 case IORING_OP_SPLICE:
5884 ret = io_splice(req, force_nonblock);
5886 case IORING_OP_PROVIDE_BUFFERS:
5887 ret = io_provide_buffers(req, force_nonblock, cs);
5889 case IORING_OP_REMOVE_BUFFERS:
5890 ret = io_remove_buffers(req, force_nonblock, cs);
5893 ret = io_tee(req, force_nonblock);
5903 /* If the op doesn't have a file, we're not polling for it */
5904 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5905 const bool in_async = io_wq_current_is_worker();
5907 /* workqueue context doesn't hold uring_lock, grab it now */
5909 mutex_lock(&ctx->uring_lock);
5911 io_iopoll_req_issued(req);
5914 mutex_unlock(&ctx->uring_lock);
5920 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5922 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5923 struct io_kiocb *timeout;
5926 timeout = io_prep_linked_timeout(req);
5928 io_queue_linked_timeout(timeout);
5930 /* if NO_CANCEL is set, we must still run the work */
5931 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5932 IO_WQ_WORK_CANCEL) {
5938 ret = io_issue_sqe(req, false, NULL);
5940 * We can get EAGAIN for polled IO even though we're
5941 * forcing a sync submission from here, since we can't
5942 * wait for request slots on the block side.
5951 req_set_fail_links(req);
5952 io_req_complete(req, ret);
5955 return io_steal_work(req);
5958 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5961 struct fixed_file_table *table;
5963 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5964 return table->files[index & IORING_FILE_TABLE_MASK];
5967 static struct file *io_file_get(struct io_submit_state *state,
5968 struct io_kiocb *req, int fd, bool fixed)
5970 struct io_ring_ctx *ctx = req->ctx;
5974 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
5976 fd = array_index_nospec(fd, ctx->nr_user_files);
5977 file = io_file_from_index(ctx, fd);
5979 req->fixed_file_refs = &ctx->file_data->node->refs;
5980 percpu_ref_get(req->fixed_file_refs);
5983 trace_io_uring_file_get(ctx, fd);
5984 file = __io_file_get(state, fd);
5990 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5995 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5996 if (unlikely(!fixed && io_async_submit(req->ctx)))
5999 req->file = io_file_get(state, req, fd, fixed);
6000 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6005 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6007 struct io_timeout_data *data = container_of(timer,
6008 struct io_timeout_data, timer);
6009 struct io_kiocb *req = data->req;
6010 struct io_ring_ctx *ctx = req->ctx;
6011 struct io_kiocb *prev = NULL;
6012 unsigned long flags;
6014 spin_lock_irqsave(&ctx->completion_lock, flags);
6017 * We don't expect the list to be empty, that will only happen if we
6018 * race with the completion of the linked work.
6020 if (!list_empty(&req->link_list)) {
6021 prev = list_entry(req->link_list.prev, struct io_kiocb,
6023 if (refcount_inc_not_zero(&prev->refs)) {
6024 list_del_init(&req->link_list);
6025 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6030 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6033 req_set_fail_links(prev);
6034 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6037 io_req_complete(req, -ETIME);
6039 return HRTIMER_NORESTART;
6042 static void __io_queue_linked_timeout(struct io_kiocb *req)
6045 * If the list is now empty, then our linked request finished before
6046 * we got a chance to setup the timer
6048 if (!list_empty(&req->link_list)) {
6049 struct io_timeout_data *data = req->async_data;
6051 data->timer.function = io_link_timeout_fn;
6052 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6057 static void io_queue_linked_timeout(struct io_kiocb *req)
6059 struct io_ring_ctx *ctx = req->ctx;
6061 spin_lock_irq(&ctx->completion_lock);
6062 __io_queue_linked_timeout(req);
6063 spin_unlock_irq(&ctx->completion_lock);
6065 /* drop submission reference */
6069 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6071 struct io_kiocb *nxt;
6073 if (!(req->flags & REQ_F_LINK_HEAD))
6075 if (req->flags & REQ_F_LINK_TIMEOUT)
6078 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6080 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6083 req->flags |= REQ_F_LINK_TIMEOUT;
6087 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6089 struct io_kiocb *linked_timeout;
6090 struct io_kiocb *nxt;
6091 const struct cred *old_creds = NULL;
6095 linked_timeout = io_prep_linked_timeout(req);
6097 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6098 req->work.creds != current_cred()) {
6100 revert_creds(old_creds);
6101 if (old_creds == req->work.creds)
6102 old_creds = NULL; /* restored original creds */
6104 old_creds = override_creds(req->work.creds);
6107 ret = io_issue_sqe(req, true, cs);
6110 * We async punt it if the file wasn't marked NOWAIT, or if the file
6111 * doesn't support non-blocking read/write attempts
6113 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6114 if (!io_arm_poll_handler(req)) {
6117 * Queued up for async execution, worker will release
6118 * submit reference when the iocb is actually submitted.
6120 io_queue_async_work(req);
6124 io_queue_linked_timeout(linked_timeout);
6128 if (unlikely(ret)) {
6129 /* un-prep timeout, so it'll be killed as any other linked */
6130 req->flags &= ~REQ_F_LINK_TIMEOUT;
6131 req_set_fail_links(req);
6133 io_req_complete(req, ret);
6137 /* drop submission reference */
6138 nxt = io_put_req_find_next(req);
6140 io_queue_linked_timeout(linked_timeout);
6145 if (req->flags & REQ_F_FORCE_ASYNC)
6151 revert_creds(old_creds);
6154 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6155 struct io_comp_state *cs)
6159 ret = io_req_defer(req, sqe);
6161 if (ret != -EIOCBQUEUED) {
6163 req_set_fail_links(req);
6165 io_req_complete(req, ret);
6167 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6168 if (!req->async_data) {
6169 ret = io_req_defer_prep(req, sqe);
6175 * Never try inline submit of IOSQE_ASYNC is set, go straight
6176 * to async execution.
6178 io_req_init_async(req);
6179 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6180 io_queue_async_work(req);
6183 ret = io_req_prep(req, sqe);
6187 __io_queue_sqe(req, cs);
6191 static inline void io_queue_link_head(struct io_kiocb *req,
6192 struct io_comp_state *cs)
6194 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6196 io_req_complete(req, -ECANCELED);
6198 io_queue_sqe(req, NULL, cs);
6201 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6202 struct io_kiocb **link, struct io_comp_state *cs)
6204 struct io_ring_ctx *ctx = req->ctx;
6208 * If we already have a head request, queue this one for async
6209 * submittal once the head completes. If we don't have a head but
6210 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6211 * submitted sync once the chain is complete. If none of those
6212 * conditions are true (normal request), then just queue it.
6215 struct io_kiocb *head = *link;
6218 * Taking sequential execution of a link, draining both sides
6219 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6220 * requests in the link. So, it drains the head and the
6221 * next after the link request. The last one is done via
6222 * drain_next flag to persist the effect across calls.
6224 if (req->flags & REQ_F_IO_DRAIN) {
6225 head->flags |= REQ_F_IO_DRAIN;
6226 ctx->drain_next = 1;
6228 ret = io_req_defer_prep(req, sqe);
6229 if (unlikely(ret)) {
6230 /* fail even hard links since we don't submit */
6231 head->flags |= REQ_F_FAIL_LINK;
6234 trace_io_uring_link(ctx, req, head);
6235 list_add_tail(&req->link_list, &head->link_list);
6237 /* last request of a link, enqueue the link */
6238 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6239 io_queue_link_head(head, cs);
6243 if (unlikely(ctx->drain_next)) {
6244 req->flags |= REQ_F_IO_DRAIN;
6245 ctx->drain_next = 0;
6247 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6248 req->flags |= REQ_F_LINK_HEAD;
6249 INIT_LIST_HEAD(&req->link_list);
6251 ret = io_req_defer_prep(req, sqe);
6253 req->flags |= REQ_F_FAIL_LINK;
6256 io_queue_sqe(req, sqe, cs);
6264 * Batched submission is done, ensure local IO is flushed out.
6266 static void io_submit_state_end(struct io_submit_state *state)
6268 if (!list_empty(&state->comp.list))
6269 io_submit_flush_completions(&state->comp);
6270 blk_finish_plug(&state->plug);
6271 io_state_file_put(state);
6272 if (state->free_reqs)
6273 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6277 * Start submission side cache.
6279 static void io_submit_state_start(struct io_submit_state *state,
6280 struct io_ring_ctx *ctx, unsigned int max_ios)
6282 blk_start_plug(&state->plug);
6284 INIT_LIST_HEAD(&state->comp.list);
6285 state->comp.ctx = ctx;
6286 state->free_reqs = 0;
6288 state->ios_left = max_ios;
6291 static void io_commit_sqring(struct io_ring_ctx *ctx)
6293 struct io_rings *rings = ctx->rings;
6296 * Ensure any loads from the SQEs are done at this point,
6297 * since once we write the new head, the application could
6298 * write new data to them.
6300 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6304 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6305 * that is mapped by userspace. This means that care needs to be taken to
6306 * ensure that reads are stable, as we cannot rely on userspace always
6307 * being a good citizen. If members of the sqe are validated and then later
6308 * used, it's important that those reads are done through READ_ONCE() to
6309 * prevent a re-load down the line.
6311 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6313 u32 *sq_array = ctx->sq_array;
6317 * The cached sq head (or cq tail) serves two purposes:
6319 * 1) allows us to batch the cost of updating the user visible
6321 * 2) allows the kernel side to track the head on its own, even
6322 * though the application is the one updating it.
6324 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6325 if (likely(head < ctx->sq_entries))
6326 return &ctx->sq_sqes[head];
6328 /* drop invalid entries */
6329 ctx->cached_sq_dropped++;
6330 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6334 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6336 ctx->cached_sq_head++;
6340 * Check SQE restrictions (opcode and flags).
6342 * Returns 'true' if SQE is allowed, 'false' otherwise.
6344 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6345 struct io_kiocb *req,
6346 unsigned int sqe_flags)
6348 if (!ctx->restricted)
6351 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6354 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6355 ctx->restrictions.sqe_flags_required)
6358 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6359 ctx->restrictions.sqe_flags_required))
6365 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6366 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6367 IOSQE_BUFFER_SELECT)
6369 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6370 const struct io_uring_sqe *sqe,
6371 struct io_submit_state *state)
6373 unsigned int sqe_flags;
6376 req->opcode = READ_ONCE(sqe->opcode);
6377 req->user_data = READ_ONCE(sqe->user_data);
6378 req->async_data = NULL;
6382 /* one is dropped after submission, the other at completion */
6383 refcount_set(&req->refs, 2);
6384 req->task = current;
6387 if (unlikely(req->opcode >= IORING_OP_LAST))
6390 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6393 sqe_flags = READ_ONCE(sqe->flags);
6394 /* enforce forwards compatibility on users */
6395 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6398 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6401 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6402 !io_op_defs[req->opcode].buffer_select)
6405 id = READ_ONCE(sqe->personality);
6407 io_req_init_async(req);
6408 req->work.creds = idr_find(&ctx->personality_idr, id);
6409 if (unlikely(!req->work.creds))
6411 get_cred(req->work.creds);
6414 /* same numerical values with corresponding REQ_F_*, safe to copy */
6415 req->flags |= sqe_flags;
6417 if (!io_op_defs[req->opcode].needs_file)
6420 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6425 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6427 struct io_submit_state state;
6428 struct io_kiocb *link = NULL;
6429 int i, submitted = 0;
6431 /* if we have a backlog and couldn't flush it all, return BUSY */
6432 if (test_bit(0, &ctx->sq_check_overflow)) {
6433 if (!list_empty(&ctx->cq_overflow_list) &&
6434 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6438 /* make sure SQ entry isn't read before tail */
6439 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6441 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6444 atomic_long_add(nr, ¤t->io_uring->req_issue);
6445 refcount_add(nr, ¤t->usage);
6447 io_submit_state_start(&state, ctx, nr);
6449 for (i = 0; i < nr; i++) {
6450 const struct io_uring_sqe *sqe;
6451 struct io_kiocb *req;
6454 sqe = io_get_sqe(ctx);
6455 if (unlikely(!sqe)) {
6456 io_consume_sqe(ctx);
6459 req = io_alloc_req(ctx, &state);
6460 if (unlikely(!req)) {
6462 submitted = -EAGAIN;
6465 io_consume_sqe(ctx);
6466 /* will complete beyond this point, count as submitted */
6469 err = io_init_req(ctx, req, sqe, &state);
6470 if (unlikely(err)) {
6473 io_req_complete(req, err);
6477 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6478 true, io_async_submit(ctx));
6479 err = io_submit_sqe(req, sqe, &link, &state.comp);
6484 if (unlikely(submitted != nr)) {
6485 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6487 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6488 atomic_long_sub(nr - ref_used, ¤t->io_uring->req_issue);
6489 put_task_struct_many(current, nr - ref_used);
6492 io_queue_link_head(link, &state.comp);
6493 io_submit_state_end(&state);
6495 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6496 io_commit_sqring(ctx);
6501 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6503 /* Tell userspace we may need a wakeup call */
6504 spin_lock_irq(&ctx->completion_lock);
6505 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6506 spin_unlock_irq(&ctx->completion_lock);
6509 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6511 spin_lock_irq(&ctx->completion_lock);
6512 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6513 spin_unlock_irq(&ctx->completion_lock);
6516 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6517 int sync, void *key)
6519 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6522 ret = autoremove_wake_function(wqe, mode, sync, key);
6524 unsigned long flags;
6526 spin_lock_irqsave(&ctx->completion_lock, flags);
6527 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6528 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6539 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6540 unsigned long start_jiffies, bool cap_entries)
6542 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6543 struct io_sq_data *sqd = ctx->sq_data;
6544 unsigned int to_submit;
6548 if (!list_empty(&ctx->iopoll_list)) {
6549 unsigned nr_events = 0;
6551 mutex_lock(&ctx->uring_lock);
6552 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6553 io_do_iopoll(ctx, &nr_events, 0);
6554 mutex_unlock(&ctx->uring_lock);
6557 to_submit = io_sqring_entries(ctx);
6560 * If submit got -EBUSY, flag us as needing the application
6561 * to enter the kernel to reap and flush events.
6563 if (!to_submit || ret == -EBUSY || need_resched()) {
6565 * Drop cur_mm before scheduling, we can't hold it for
6566 * long periods (or over schedule()). Do this before
6567 * adding ourselves to the waitqueue, as the unuse/drop
6570 io_sq_thread_drop_mm();
6573 * We're polling. If we're within the defined idle
6574 * period, then let us spin without work before going
6575 * to sleep. The exception is if we got EBUSY doing
6576 * more IO, we should wait for the application to
6577 * reap events and wake us up.
6579 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6580 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6581 !percpu_ref_is_dying(&ctx->refs)))
6584 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6585 TASK_INTERRUPTIBLE);
6588 * While doing polled IO, before going to sleep, we need
6589 * to check if there are new reqs added to iopoll_list,
6590 * it is because reqs may have been punted to io worker
6591 * and will be added to iopoll_list later, hence check
6592 * the iopoll_list again.
6594 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6595 !list_empty_careful(&ctx->iopoll_list)) {
6596 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6600 to_submit = io_sqring_entries(ctx);
6601 if (!to_submit || ret == -EBUSY)
6605 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6606 io_ring_clear_wakeup_flag(ctx);
6608 /* if we're handling multiple rings, cap submit size for fairness */
6609 if (cap_entries && to_submit > 8)
6612 mutex_lock(&ctx->uring_lock);
6613 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6614 ret = io_submit_sqes(ctx, to_submit);
6615 mutex_unlock(&ctx->uring_lock);
6617 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6618 wake_up(&ctx->sqo_sq_wait);
6620 return SQT_DID_WORK;
6623 static void io_sqd_init_new(struct io_sq_data *sqd)
6625 struct io_ring_ctx *ctx;
6627 while (!list_empty(&sqd->ctx_new_list)) {
6628 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6629 init_wait(&ctx->sqo_wait_entry);
6630 ctx->sqo_wait_entry.func = io_sq_wake_function;
6631 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6632 complete(&ctx->sq_thread_comp);
6636 static int io_sq_thread(void *data)
6638 struct cgroup_subsys_state *cur_css = NULL;
6639 const struct cred *old_cred = NULL;
6640 struct io_sq_data *sqd = data;
6641 struct io_ring_ctx *ctx;
6642 unsigned long start_jiffies;
6644 start_jiffies = jiffies;
6645 while (!kthread_should_stop()) {
6646 enum sq_ret ret = 0;
6650 * Any changes to the sqd lists are synchronized through the
6651 * kthread parking. This synchronizes the thread vs users,
6652 * the users are synchronized on the sqd->ctx_lock.
6654 if (kthread_should_park())
6657 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6658 io_sqd_init_new(sqd);
6660 cap_entries = !list_is_singular(&sqd->ctx_list);
6662 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6663 if (current->cred != ctx->creds) {
6665 revert_creds(old_cred);
6666 old_cred = override_creds(ctx->creds);
6668 io_sq_thread_associate_blkcg(ctx, &cur_css);
6670 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6672 io_sq_thread_drop_mm();
6675 if (ret & SQT_SPIN) {
6678 } else if (ret == SQT_IDLE) {
6679 if (kthread_should_park())
6681 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6682 io_ring_set_wakeup_flag(ctx);
6684 start_jiffies = jiffies;
6685 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6686 io_ring_clear_wakeup_flag(ctx);
6693 io_sq_thread_unassociate_blkcg();
6695 revert_creds(old_cred);
6702 struct io_wait_queue {
6703 struct wait_queue_entry wq;
6704 struct io_ring_ctx *ctx;
6706 unsigned nr_timeouts;
6709 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6711 struct io_ring_ctx *ctx = iowq->ctx;
6714 * Wake up if we have enough events, or if a timeout occurred since we
6715 * started waiting. For timeouts, we always want to return to userspace,
6716 * regardless of event count.
6718 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6719 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6722 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6723 int wake_flags, void *key)
6725 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6728 /* use noflush == true, as we can't safely rely on locking context */
6729 if (!io_should_wake(iowq, true))
6732 return autoremove_wake_function(curr, mode, wake_flags, key);
6735 static int io_run_task_work_sig(void)
6737 if (io_run_task_work())
6739 if (!signal_pending(current))
6741 if (current->jobctl & JOBCTL_TASK_WORK) {
6742 spin_lock_irq(¤t->sighand->siglock);
6743 current->jobctl &= ~JOBCTL_TASK_WORK;
6744 recalc_sigpending();
6745 spin_unlock_irq(¤t->sighand->siglock);
6752 * Wait until events become available, if we don't already have some. The
6753 * application must reap them itself, as they reside on the shared cq ring.
6755 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6756 const sigset_t __user *sig, size_t sigsz)
6758 struct io_wait_queue iowq = {
6761 .func = io_wake_function,
6762 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6765 .to_wait = min_events,
6767 struct io_rings *rings = ctx->rings;
6771 if (io_cqring_events(ctx, false) >= min_events)
6773 if (!io_run_task_work())
6778 #ifdef CONFIG_COMPAT
6779 if (in_compat_syscall())
6780 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6784 ret = set_user_sigmask(sig, sigsz);
6790 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6791 trace_io_uring_cqring_wait(ctx, min_events);
6793 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6794 TASK_INTERRUPTIBLE);
6795 /* make sure we run task_work before checking for signals */
6796 ret = io_run_task_work_sig();
6801 if (io_should_wake(&iowq, false))
6805 finish_wait(&ctx->wait, &iowq.wq);
6807 restore_saved_sigmask_unless(ret == -EINTR);
6809 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6812 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6814 #if defined(CONFIG_UNIX)
6815 if (ctx->ring_sock) {
6816 struct sock *sock = ctx->ring_sock->sk;
6817 struct sk_buff *skb;
6819 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6825 for (i = 0; i < ctx->nr_user_files; i++) {
6828 file = io_file_from_index(ctx, i);
6835 static void io_file_ref_kill(struct percpu_ref *ref)
6837 struct fixed_file_data *data;
6839 data = container_of(ref, struct fixed_file_data, refs);
6840 complete(&data->done);
6843 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6845 struct fixed_file_data *data = ctx->file_data;
6846 struct fixed_file_ref_node *ref_node = NULL;
6847 unsigned nr_tables, i;
6852 spin_lock(&data->lock);
6853 if (!list_empty(&data->ref_list))
6854 ref_node = list_first_entry(&data->ref_list,
6855 struct fixed_file_ref_node, node);
6856 spin_unlock(&data->lock);
6858 percpu_ref_kill(&ref_node->refs);
6860 percpu_ref_kill(&data->refs);
6862 /* wait for all refs nodes to complete */
6863 flush_delayed_work(&ctx->file_put_work);
6864 wait_for_completion(&data->done);
6866 __io_sqe_files_unregister(ctx);
6867 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6868 for (i = 0; i < nr_tables; i++)
6869 kfree(data->table[i].files);
6871 percpu_ref_exit(&data->refs);
6873 ctx->file_data = NULL;
6874 ctx->nr_user_files = 0;
6878 static void io_put_sq_data(struct io_sq_data *sqd)
6880 if (refcount_dec_and_test(&sqd->refs)) {
6882 * The park is a bit of a work-around, without it we get
6883 * warning spews on shutdown with SQPOLL set and affinity
6884 * set to a single CPU.
6887 kthread_park(sqd->thread);
6888 kthread_stop(sqd->thread);
6895 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
6897 struct io_ring_ctx *ctx_attach;
6898 struct io_sq_data *sqd;
6901 f = fdget(p->wq_fd);
6903 return ERR_PTR(-ENXIO);
6904 if (f.file->f_op != &io_uring_fops) {
6906 return ERR_PTR(-EINVAL);
6909 ctx_attach = f.file->private_data;
6910 sqd = ctx_attach->sq_data;
6913 return ERR_PTR(-EINVAL);
6916 refcount_inc(&sqd->refs);
6921 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
6923 struct io_sq_data *sqd;
6925 if (p->flags & IORING_SETUP_ATTACH_WQ)
6926 return io_attach_sq_data(p);
6928 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
6930 return ERR_PTR(-ENOMEM);
6932 refcount_set(&sqd->refs, 1);
6933 INIT_LIST_HEAD(&sqd->ctx_list);
6934 INIT_LIST_HEAD(&sqd->ctx_new_list);
6935 mutex_init(&sqd->ctx_lock);
6936 mutex_init(&sqd->lock);
6937 init_waitqueue_head(&sqd->wait);
6941 static void io_sq_thread_unpark(struct io_sq_data *sqd)
6942 __releases(&sqd->lock)
6946 kthread_unpark(sqd->thread);
6947 mutex_unlock(&sqd->lock);
6950 static void io_sq_thread_park(struct io_sq_data *sqd)
6951 __acquires(&sqd->lock)
6955 mutex_lock(&sqd->lock);
6956 kthread_park(sqd->thread);
6959 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6961 struct io_sq_data *sqd = ctx->sq_data;
6966 * We may arrive here from the error branch in
6967 * io_sq_offload_create() where the kthread is created
6968 * without being waked up, thus wake it up now to make
6969 * sure the wait will complete.
6971 wake_up_process(sqd->thread);
6972 wait_for_completion(&ctx->sq_thread_comp);
6974 io_sq_thread_park(sqd);
6977 mutex_lock(&sqd->ctx_lock);
6978 list_del(&ctx->sqd_list);
6979 mutex_unlock(&sqd->ctx_lock);
6982 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6983 io_sq_thread_unpark(sqd);
6986 io_put_sq_data(sqd);
6987 ctx->sq_data = NULL;
6991 static void io_finish_async(struct io_ring_ctx *ctx)
6993 io_sq_thread_stop(ctx);
6996 io_wq_destroy(ctx->io_wq);
7001 #if defined(CONFIG_UNIX)
7003 * Ensure the UNIX gc is aware of our file set, so we are certain that
7004 * the io_uring can be safely unregistered on process exit, even if we have
7005 * loops in the file referencing.
7007 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7009 struct sock *sk = ctx->ring_sock->sk;
7010 struct scm_fp_list *fpl;
7011 struct sk_buff *skb;
7014 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7018 skb = alloc_skb(0, GFP_KERNEL);
7027 fpl->user = get_uid(ctx->user);
7028 for (i = 0; i < nr; i++) {
7029 struct file *file = io_file_from_index(ctx, i + offset);
7033 fpl->fp[nr_files] = get_file(file);
7034 unix_inflight(fpl->user, fpl->fp[nr_files]);
7039 fpl->max = SCM_MAX_FD;
7040 fpl->count = nr_files;
7041 UNIXCB(skb).fp = fpl;
7042 skb->destructor = unix_destruct_scm;
7043 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7044 skb_queue_head(&sk->sk_receive_queue, skb);
7046 for (i = 0; i < nr_files; i++)
7057 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7058 * causes regular reference counting to break down. We rely on the UNIX
7059 * garbage collection to take care of this problem for us.
7061 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7063 unsigned left, total;
7067 left = ctx->nr_user_files;
7069 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7071 ret = __io_sqe_files_scm(ctx, this_files, total);
7075 total += this_files;
7081 while (total < ctx->nr_user_files) {
7082 struct file *file = io_file_from_index(ctx, total);
7092 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7098 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7099 unsigned nr_tables, unsigned nr_files)
7103 for (i = 0; i < nr_tables; i++) {
7104 struct fixed_file_table *table = &file_data->table[i];
7105 unsigned this_files;
7107 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7108 table->files = kcalloc(this_files, sizeof(struct file *),
7112 nr_files -= this_files;
7118 for (i = 0; i < nr_tables; i++) {
7119 struct fixed_file_table *table = &file_data->table[i];
7120 kfree(table->files);
7125 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7127 #if defined(CONFIG_UNIX)
7128 struct sock *sock = ctx->ring_sock->sk;
7129 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7130 struct sk_buff *skb;
7133 __skb_queue_head_init(&list);
7136 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7137 * remove this entry and rearrange the file array.
7139 skb = skb_dequeue(head);
7141 struct scm_fp_list *fp;
7143 fp = UNIXCB(skb).fp;
7144 for (i = 0; i < fp->count; i++) {
7147 if (fp->fp[i] != file)
7150 unix_notinflight(fp->user, fp->fp[i]);
7151 left = fp->count - 1 - i;
7153 memmove(&fp->fp[i], &fp->fp[i + 1],
7154 left * sizeof(struct file *));
7161 __skb_queue_tail(&list, skb);
7171 __skb_queue_tail(&list, skb);
7173 skb = skb_dequeue(head);
7176 if (skb_peek(&list)) {
7177 spin_lock_irq(&head->lock);
7178 while ((skb = __skb_dequeue(&list)) != NULL)
7179 __skb_queue_tail(head, skb);
7180 spin_unlock_irq(&head->lock);
7187 struct io_file_put {
7188 struct list_head list;
7192 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7194 struct fixed_file_data *file_data = ref_node->file_data;
7195 struct io_ring_ctx *ctx = file_data->ctx;
7196 struct io_file_put *pfile, *tmp;
7198 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7199 list_del(&pfile->list);
7200 io_ring_file_put(ctx, pfile->file);
7204 spin_lock(&file_data->lock);
7205 list_del(&ref_node->node);
7206 spin_unlock(&file_data->lock);
7208 percpu_ref_exit(&ref_node->refs);
7210 percpu_ref_put(&file_data->refs);
7213 static void io_file_put_work(struct work_struct *work)
7215 struct io_ring_ctx *ctx;
7216 struct llist_node *node;
7218 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7219 node = llist_del_all(&ctx->file_put_llist);
7222 struct fixed_file_ref_node *ref_node;
7223 struct llist_node *next = node->next;
7225 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7226 __io_file_put_work(ref_node);
7231 static void io_file_data_ref_zero(struct percpu_ref *ref)
7233 struct fixed_file_ref_node *ref_node;
7234 struct io_ring_ctx *ctx;
7238 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7239 ctx = ref_node->file_data->ctx;
7241 if (percpu_ref_is_dying(&ctx->file_data->refs))
7244 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7246 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7248 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7251 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7252 struct io_ring_ctx *ctx)
7254 struct fixed_file_ref_node *ref_node;
7256 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7258 return ERR_PTR(-ENOMEM);
7260 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7263 return ERR_PTR(-ENOMEM);
7265 INIT_LIST_HEAD(&ref_node->node);
7266 INIT_LIST_HEAD(&ref_node->file_list);
7267 ref_node->file_data = ctx->file_data;
7271 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7273 percpu_ref_exit(&ref_node->refs);
7277 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7280 __s32 __user *fds = (__s32 __user *) arg;
7281 unsigned nr_tables, i;
7283 int fd, ret = -ENOMEM;
7284 struct fixed_file_ref_node *ref_node;
7285 struct fixed_file_data *file_data;
7291 if (nr_args > IORING_MAX_FIXED_FILES)
7294 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7297 file_data->ctx = ctx;
7298 init_completion(&file_data->done);
7299 INIT_LIST_HEAD(&file_data->ref_list);
7300 spin_lock_init(&file_data->lock);
7302 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7303 file_data->table = kcalloc(nr_tables, sizeof(file_data->table),
7305 if (!file_data->table)
7308 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7309 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7312 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7315 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7316 struct fixed_file_table *table;
7319 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7323 /* allow sparse sets */
7333 * Don't allow io_uring instances to be registered. If UNIX
7334 * isn't enabled, then this causes a reference cycle and this
7335 * instance can never get freed. If UNIX is enabled we'll
7336 * handle it just fine, but there's still no point in allowing
7337 * a ring fd as it doesn't support regular read/write anyway.
7339 if (file->f_op == &io_uring_fops) {
7343 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7344 index = i & IORING_FILE_TABLE_MASK;
7345 table->files[index] = file;
7348 ctx->file_data = file_data;
7349 ret = io_sqe_files_scm(ctx);
7351 io_sqe_files_unregister(ctx);
7355 ref_node = alloc_fixed_file_ref_node(ctx);
7356 if (IS_ERR(ref_node)) {
7357 io_sqe_files_unregister(ctx);
7358 return PTR_ERR(ref_node);
7361 file_data->node = ref_node;
7362 spin_lock(&file_data->lock);
7363 list_add(&ref_node->node, &file_data->ref_list);
7364 spin_unlock(&file_data->lock);
7365 percpu_ref_get(&file_data->refs);
7368 for (i = 0; i < ctx->nr_user_files; i++) {
7369 file = io_file_from_index(ctx, i);
7373 for (i = 0; i < nr_tables; i++)
7374 kfree(file_data->table[i].files);
7375 ctx->nr_user_files = 0;
7377 percpu_ref_exit(&file_data->refs);
7379 kfree(file_data->table);
7384 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7387 #if defined(CONFIG_UNIX)
7388 struct sock *sock = ctx->ring_sock->sk;
7389 struct sk_buff_head *head = &sock->sk_receive_queue;
7390 struct sk_buff *skb;
7393 * See if we can merge this file into an existing skb SCM_RIGHTS
7394 * file set. If there's no room, fall back to allocating a new skb
7395 * and filling it in.
7397 spin_lock_irq(&head->lock);
7398 skb = skb_peek(head);
7400 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7402 if (fpl->count < SCM_MAX_FD) {
7403 __skb_unlink(skb, head);
7404 spin_unlock_irq(&head->lock);
7405 fpl->fp[fpl->count] = get_file(file);
7406 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7408 spin_lock_irq(&head->lock);
7409 __skb_queue_head(head, skb);
7414 spin_unlock_irq(&head->lock);
7421 return __io_sqe_files_scm(ctx, 1, index);
7427 static int io_queue_file_removal(struct fixed_file_data *data,
7430 struct io_file_put *pfile;
7431 struct fixed_file_ref_node *ref_node = data->node;
7433 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7438 list_add(&pfile->list, &ref_node->file_list);
7443 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7444 struct io_uring_files_update *up,
7447 struct fixed_file_data *data = ctx->file_data;
7448 struct fixed_file_ref_node *ref_node;
7453 bool needs_switch = false;
7455 if (check_add_overflow(up->offset, nr_args, &done))
7457 if (done > ctx->nr_user_files)
7460 ref_node = alloc_fixed_file_ref_node(ctx);
7461 if (IS_ERR(ref_node))
7462 return PTR_ERR(ref_node);
7465 fds = u64_to_user_ptr(up->fds);
7467 struct fixed_file_table *table;
7471 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7475 i = array_index_nospec(up->offset, ctx->nr_user_files);
7476 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7477 index = i & IORING_FILE_TABLE_MASK;
7478 if (table->files[index]) {
7479 file = table->files[index];
7480 err = io_queue_file_removal(data, file);
7483 table->files[index] = NULL;
7484 needs_switch = true;
7493 * Don't allow io_uring instances to be registered. If
7494 * UNIX isn't enabled, then this causes a reference
7495 * cycle and this instance can never get freed. If UNIX
7496 * is enabled we'll handle it just fine, but there's
7497 * still no point in allowing a ring fd as it doesn't
7498 * support regular read/write anyway.
7500 if (file->f_op == &io_uring_fops) {
7505 table->files[index] = file;
7506 err = io_sqe_file_register(ctx, file, i);
7508 table->files[index] = NULL;
7519 percpu_ref_kill(&data->node->refs);
7520 spin_lock(&data->lock);
7521 list_add(&ref_node->node, &data->ref_list);
7522 data->node = ref_node;
7523 spin_unlock(&data->lock);
7524 percpu_ref_get(&ctx->file_data->refs);
7526 destroy_fixed_file_ref_node(ref_node);
7528 return done ? done : err;
7531 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7534 struct io_uring_files_update up;
7536 if (!ctx->file_data)
7540 if (copy_from_user(&up, arg, sizeof(up)))
7545 return __io_sqe_files_update(ctx, &up, nr_args);
7548 static void io_free_work(struct io_wq_work *work)
7550 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7552 /* Consider that io_steal_work() relies on this ref */
7556 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7557 struct io_uring_params *p)
7559 struct io_wq_data data;
7561 struct io_ring_ctx *ctx_attach;
7562 unsigned int concurrency;
7565 data.user = ctx->user;
7566 data.free_work = io_free_work;
7567 data.do_work = io_wq_submit_work;
7569 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7570 /* Do QD, or 4 * CPUS, whatever is smallest */
7571 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7573 ctx->io_wq = io_wq_create(concurrency, &data);
7574 if (IS_ERR(ctx->io_wq)) {
7575 ret = PTR_ERR(ctx->io_wq);
7581 f = fdget(p->wq_fd);
7585 if (f.file->f_op != &io_uring_fops) {
7590 ctx_attach = f.file->private_data;
7591 /* @io_wq is protected by holding the fd */
7592 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7597 ctx->io_wq = ctx_attach->io_wq;
7603 static int io_uring_alloc_task_context(struct task_struct *task)
7605 struct io_uring_task *tctx;
7607 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7608 if (unlikely(!tctx))
7612 init_waitqueue_head(&tctx->wait);
7615 atomic_long_set(&tctx->req_issue, 0);
7616 atomic_long_set(&tctx->req_complete, 0);
7617 task->io_uring = tctx;
7621 void __io_uring_free(struct task_struct *tsk)
7623 struct io_uring_task *tctx = tsk->io_uring;
7625 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7627 tsk->io_uring = NULL;
7630 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7631 struct io_uring_params *p)
7635 if (ctx->flags & IORING_SETUP_SQPOLL) {
7636 struct io_sq_data *sqd;
7639 if (!capable(CAP_SYS_ADMIN))
7642 sqd = io_get_sq_data(p);
7649 io_sq_thread_park(sqd);
7650 mutex_lock(&sqd->ctx_lock);
7651 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7652 mutex_unlock(&sqd->ctx_lock);
7653 io_sq_thread_unpark(sqd);
7655 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7656 if (!ctx->sq_thread_idle)
7657 ctx->sq_thread_idle = HZ;
7662 if (p->flags & IORING_SETUP_SQ_AFF) {
7663 int cpu = p->sq_thread_cpu;
7666 if (cpu >= nr_cpu_ids)
7668 if (!cpu_online(cpu))
7671 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7672 cpu, "io_uring-sq");
7674 sqd->thread = kthread_create(io_sq_thread, sqd,
7677 if (IS_ERR(sqd->thread)) {
7678 ret = PTR_ERR(sqd->thread);
7682 ret = io_uring_alloc_task_context(sqd->thread);
7685 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7686 /* Can't have SQ_AFF without SQPOLL */
7692 ret = io_init_wq_offload(ctx, p);
7698 io_finish_async(ctx);
7702 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7704 struct io_sq_data *sqd = ctx->sq_data;
7706 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7707 wake_up_process(sqd->thread);
7710 static inline void __io_unaccount_mem(struct user_struct *user,
7711 unsigned long nr_pages)
7713 atomic_long_sub(nr_pages, &user->locked_vm);
7716 static inline int __io_account_mem(struct user_struct *user,
7717 unsigned long nr_pages)
7719 unsigned long page_limit, cur_pages, new_pages;
7721 /* Don't allow more pages than we can safely lock */
7722 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7725 cur_pages = atomic_long_read(&user->locked_vm);
7726 new_pages = cur_pages + nr_pages;
7727 if (new_pages > page_limit)
7729 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7730 new_pages) != cur_pages);
7735 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7736 enum io_mem_account acct)
7739 __io_unaccount_mem(ctx->user, nr_pages);
7741 if (ctx->mm_account) {
7742 if (acct == ACCT_LOCKED)
7743 ctx->mm_account->locked_vm -= nr_pages;
7744 else if (acct == ACCT_PINNED)
7745 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7749 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7750 enum io_mem_account acct)
7754 if (ctx->limit_mem) {
7755 ret = __io_account_mem(ctx->user, nr_pages);
7760 if (ctx->mm_account) {
7761 if (acct == ACCT_LOCKED)
7762 ctx->mm_account->locked_vm += nr_pages;
7763 else if (acct == ACCT_PINNED)
7764 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7770 static void io_mem_free(void *ptr)
7777 page = virt_to_head_page(ptr);
7778 if (put_page_testzero(page))
7779 free_compound_page(page);
7782 static void *io_mem_alloc(size_t size)
7784 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7787 return (void *) __get_free_pages(gfp_flags, get_order(size));
7790 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7793 struct io_rings *rings;
7794 size_t off, sq_array_size;
7796 off = struct_size(rings, cqes, cq_entries);
7797 if (off == SIZE_MAX)
7801 off = ALIGN(off, SMP_CACHE_BYTES);
7809 sq_array_size = array_size(sizeof(u32), sq_entries);
7810 if (sq_array_size == SIZE_MAX)
7813 if (check_add_overflow(off, sq_array_size, &off))
7819 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7823 pages = (size_t)1 << get_order(
7824 rings_size(sq_entries, cq_entries, NULL));
7825 pages += (size_t)1 << get_order(
7826 array_size(sizeof(struct io_uring_sqe), sq_entries));
7831 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7835 if (!ctx->user_bufs)
7838 for (i = 0; i < ctx->nr_user_bufs; i++) {
7839 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7841 for (j = 0; j < imu->nr_bvecs; j++)
7842 unpin_user_page(imu->bvec[j].bv_page);
7844 if (imu->acct_pages)
7845 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7850 kfree(ctx->user_bufs);
7851 ctx->user_bufs = NULL;
7852 ctx->nr_user_bufs = 0;
7856 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7857 void __user *arg, unsigned index)
7859 struct iovec __user *src;
7861 #ifdef CONFIG_COMPAT
7863 struct compat_iovec __user *ciovs;
7864 struct compat_iovec ciov;
7866 ciovs = (struct compat_iovec __user *) arg;
7867 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7870 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7871 dst->iov_len = ciov.iov_len;
7875 src = (struct iovec __user *) arg;
7876 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7882 * Not super efficient, but this is just a registration time. And we do cache
7883 * the last compound head, so generally we'll only do a full search if we don't
7886 * We check if the given compound head page has already been accounted, to
7887 * avoid double accounting it. This allows us to account the full size of the
7888 * page, not just the constituent pages of a huge page.
7890 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
7891 int nr_pages, struct page *hpage)
7895 /* check current page array */
7896 for (i = 0; i < nr_pages; i++) {
7897 if (!PageCompound(pages[i]))
7899 if (compound_head(pages[i]) == hpage)
7903 /* check previously registered pages */
7904 for (i = 0; i < ctx->nr_user_bufs; i++) {
7905 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7907 for (j = 0; j < imu->nr_bvecs; j++) {
7908 if (!PageCompound(imu->bvec[j].bv_page))
7910 if (compound_head(imu->bvec[j].bv_page) == hpage)
7918 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
7919 int nr_pages, struct io_mapped_ubuf *imu,
7920 struct page **last_hpage)
7924 for (i = 0; i < nr_pages; i++) {
7925 if (!PageCompound(pages[i])) {
7930 hpage = compound_head(pages[i]);
7931 if (hpage == *last_hpage)
7933 *last_hpage = hpage;
7934 if (headpage_already_acct(ctx, pages, i, hpage))
7936 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
7940 if (!imu->acct_pages)
7943 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
7945 imu->acct_pages = 0;
7949 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7952 struct vm_area_struct **vmas = NULL;
7953 struct page **pages = NULL;
7954 struct page *last_hpage = NULL;
7955 int i, j, got_pages = 0;
7960 if (!nr_args || nr_args > UIO_MAXIOV)
7963 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7965 if (!ctx->user_bufs)
7968 for (i = 0; i < nr_args; i++) {
7969 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7970 unsigned long off, start, end, ubuf;
7975 ret = io_copy_iov(ctx, &iov, arg, i);
7980 * Don't impose further limits on the size and buffer
7981 * constraints here, we'll -EINVAL later when IO is
7982 * submitted if they are wrong.
7985 if (!iov.iov_base || !iov.iov_len)
7988 /* arbitrary limit, but we need something */
7989 if (iov.iov_len > SZ_1G)
7992 ubuf = (unsigned long) iov.iov_base;
7993 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7994 start = ubuf >> PAGE_SHIFT;
7995 nr_pages = end - start;
7998 if (!pages || nr_pages > got_pages) {
8001 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8003 vmas = kvmalloc_array(nr_pages,
8004 sizeof(struct vm_area_struct *),
8006 if (!pages || !vmas) {
8010 got_pages = nr_pages;
8013 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8020 mmap_read_lock(current->mm);
8021 pret = pin_user_pages(ubuf, nr_pages,
8022 FOLL_WRITE | FOLL_LONGTERM,
8024 if (pret == nr_pages) {
8025 /* don't support file backed memory */
8026 for (j = 0; j < nr_pages; j++) {
8027 struct vm_area_struct *vma = vmas[j];
8030 !is_file_hugepages(vma->vm_file)) {
8036 ret = pret < 0 ? pret : -EFAULT;
8038 mmap_read_unlock(current->mm);
8041 * if we did partial map, or found file backed vmas,
8042 * release any pages we did get
8045 unpin_user_pages(pages, pret);
8050 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8052 unpin_user_pages(pages, pret);
8057 off = ubuf & ~PAGE_MASK;
8059 for (j = 0; j < nr_pages; j++) {
8062 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8063 imu->bvec[j].bv_page = pages[j];
8064 imu->bvec[j].bv_len = vec_len;
8065 imu->bvec[j].bv_offset = off;
8069 /* store original address for later verification */
8071 imu->len = iov.iov_len;
8072 imu->nr_bvecs = nr_pages;
8074 ctx->nr_user_bufs++;
8082 io_sqe_buffer_unregister(ctx);
8086 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8088 __s32 __user *fds = arg;
8094 if (copy_from_user(&fd, fds, sizeof(*fds)))
8097 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8098 if (IS_ERR(ctx->cq_ev_fd)) {
8099 int ret = PTR_ERR(ctx->cq_ev_fd);
8100 ctx->cq_ev_fd = NULL;
8107 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8109 if (ctx->cq_ev_fd) {
8110 eventfd_ctx_put(ctx->cq_ev_fd);
8111 ctx->cq_ev_fd = NULL;
8118 static int __io_destroy_buffers(int id, void *p, void *data)
8120 struct io_ring_ctx *ctx = data;
8121 struct io_buffer *buf = p;
8123 __io_remove_buffers(ctx, buf, id, -1U);
8127 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8129 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8130 idr_destroy(&ctx->io_buffer_idr);
8133 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8135 io_finish_async(ctx);
8136 io_sqe_buffer_unregister(ctx);
8138 if (ctx->sqo_task) {
8139 put_task_struct(ctx->sqo_task);
8140 ctx->sqo_task = NULL;
8141 mmdrop(ctx->mm_account);
8142 ctx->mm_account = NULL;
8145 #ifdef CONFIG_BLK_CGROUP
8146 if (ctx->sqo_blkcg_css)
8147 css_put(ctx->sqo_blkcg_css);
8150 io_sqe_files_unregister(ctx);
8151 io_eventfd_unregister(ctx);
8152 io_destroy_buffers(ctx);
8153 idr_destroy(&ctx->personality_idr);
8155 #if defined(CONFIG_UNIX)
8156 if (ctx->ring_sock) {
8157 ctx->ring_sock->file = NULL; /* so that iput() is called */
8158 sock_release(ctx->ring_sock);
8162 io_mem_free(ctx->rings);
8163 io_mem_free(ctx->sq_sqes);
8165 percpu_ref_exit(&ctx->refs);
8166 free_uid(ctx->user);
8167 put_cred(ctx->creds);
8168 kfree(ctx->cancel_hash);
8169 kmem_cache_free(req_cachep, ctx->fallback_req);
8173 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8175 struct io_ring_ctx *ctx = file->private_data;
8178 poll_wait(file, &ctx->cq_wait, wait);
8180 * synchronizes with barrier from wq_has_sleeper call in
8184 if (!io_sqring_full(ctx))
8185 mask |= EPOLLOUT | EPOLLWRNORM;
8186 if (io_cqring_events(ctx, false))
8187 mask |= EPOLLIN | EPOLLRDNORM;
8192 static int io_uring_fasync(int fd, struct file *file, int on)
8194 struct io_ring_ctx *ctx = file->private_data;
8196 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8199 static int io_remove_personalities(int id, void *p, void *data)
8201 struct io_ring_ctx *ctx = data;
8202 const struct cred *cred;
8204 cred = idr_remove(&ctx->personality_idr, id);
8210 static void io_ring_exit_work(struct work_struct *work)
8212 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8216 * If we're doing polled IO and end up having requests being
8217 * submitted async (out-of-line), then completions can come in while
8218 * we're waiting for refs to drop. We need to reap these manually,
8219 * as nobody else will be looking for them.
8223 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8224 io_iopoll_try_reap_events(ctx);
8225 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8226 io_ring_ctx_free(ctx);
8229 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8231 mutex_lock(&ctx->uring_lock);
8232 percpu_ref_kill(&ctx->refs);
8233 mutex_unlock(&ctx->uring_lock);
8235 io_kill_timeouts(ctx, NULL);
8236 io_poll_remove_all(ctx, NULL);
8239 io_wq_cancel_all(ctx->io_wq);
8241 /* if we failed setting up the ctx, we might not have any rings */
8243 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8244 io_iopoll_try_reap_events(ctx);
8245 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8248 * Do this upfront, so we won't have a grace period where the ring
8249 * is closed but resources aren't reaped yet. This can cause
8250 * spurious failure in setting up a new ring.
8252 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8255 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8257 * Use system_unbound_wq to avoid spawning tons of event kworkers
8258 * if we're exiting a ton of rings at the same time. It just adds
8259 * noise and overhead, there's no discernable change in runtime
8260 * over using system_wq.
8262 queue_work(system_unbound_wq, &ctx->exit_work);
8265 static int io_uring_release(struct inode *inode, struct file *file)
8267 struct io_ring_ctx *ctx = file->private_data;
8269 file->private_data = NULL;
8270 io_ring_ctx_wait_and_kill(ctx);
8274 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8276 struct files_struct *files = data;
8278 return !files || work->files == files;
8282 * Returns true if 'preq' is the link parent of 'req'
8284 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8286 struct io_kiocb *link;
8288 if (!(preq->flags & REQ_F_LINK_HEAD))
8291 list_for_each_entry(link, &preq->link_list, link_list) {
8299 static bool io_match_link_files(struct io_kiocb *req,
8300 struct files_struct *files)
8302 struct io_kiocb *link;
8304 if (io_match_files(req, files))
8306 if (req->flags & REQ_F_LINK_HEAD) {
8307 list_for_each_entry(link, &req->link_list, link_list) {
8308 if (io_match_files(link, files))
8316 * We're looking to cancel 'req' because it's holding on to our files, but
8317 * 'req' could be a link to another request. See if it is, and cancel that
8318 * parent request if so.
8320 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8322 struct hlist_node *tmp;
8323 struct io_kiocb *preq;
8327 spin_lock_irq(&ctx->completion_lock);
8328 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8329 struct hlist_head *list;
8331 list = &ctx->cancel_hash[i];
8332 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8333 found = io_match_link(preq, req);
8335 io_poll_remove_one(preq);
8340 spin_unlock_irq(&ctx->completion_lock);
8344 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8345 struct io_kiocb *req)
8347 struct io_kiocb *preq;
8350 spin_lock_irq(&ctx->completion_lock);
8351 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8352 found = io_match_link(preq, req);
8354 __io_timeout_cancel(preq);
8358 spin_unlock_irq(&ctx->completion_lock);
8362 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8364 return io_match_link(container_of(work, struct io_kiocb, work), data);
8367 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8369 enum io_wq_cancel cret;
8371 /* cancel this particular work, if it's running */
8372 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8373 if (cret != IO_WQ_CANCEL_NOTFOUND)
8376 /* find links that hold this pending, cancel those */
8377 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8378 if (cret != IO_WQ_CANCEL_NOTFOUND)
8381 /* if we have a poll link holding this pending, cancel that */
8382 if (io_poll_remove_link(ctx, req))
8385 /* final option, timeout link is holding this req pending */
8386 io_timeout_remove_link(ctx, req);
8389 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8390 struct files_struct *files)
8392 struct io_defer_entry *de = NULL;
8395 spin_lock_irq(&ctx->completion_lock);
8396 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8397 if (io_match_link_files(de->req, files)) {
8398 list_cut_position(&list, &ctx->defer_list, &de->list);
8402 spin_unlock_irq(&ctx->completion_lock);
8404 while (!list_empty(&list)) {
8405 de = list_first_entry(&list, struct io_defer_entry, list);
8406 list_del_init(&de->list);
8407 req_set_fail_links(de->req);
8408 io_put_req(de->req);
8409 io_req_complete(de->req, -ECANCELED);
8415 * Returns true if we found and killed one or more files pinning requests
8417 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8418 struct files_struct *files)
8420 if (list_empty_careful(&ctx->inflight_list))
8423 io_cancel_defer_files(ctx, files);
8424 /* cancel all at once, should be faster than doing it one by one*/
8425 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8427 while (!list_empty_careful(&ctx->inflight_list)) {
8428 struct io_kiocb *cancel_req = NULL, *req;
8431 spin_lock_irq(&ctx->inflight_lock);
8432 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8433 if (files && req->work.files != files)
8435 /* req is being completed, ignore */
8436 if (!refcount_inc_not_zero(&req->refs))
8442 prepare_to_wait(&ctx->inflight_wait, &wait,
8443 TASK_UNINTERRUPTIBLE);
8444 spin_unlock_irq(&ctx->inflight_lock);
8446 /* We need to keep going until we don't find a matching req */
8449 /* cancel this request, or head link requests */
8450 io_attempt_cancel(ctx, cancel_req);
8451 io_put_req(cancel_req);
8452 /* cancellations _may_ trigger task work */
8455 finish_wait(&ctx->inflight_wait, &wait);
8461 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8463 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8464 struct task_struct *task = data;
8466 return io_task_match(req, task);
8469 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8470 struct task_struct *task,
8471 struct files_struct *files)
8475 ret = io_uring_cancel_files(ctx, files);
8477 enum io_wq_cancel cret;
8479 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8480 if (cret != IO_WQ_CANCEL_NOTFOUND)
8483 /* SQPOLL thread does its own polling */
8484 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8485 while (!list_empty_careful(&ctx->iopoll_list)) {
8486 io_iopoll_try_reap_events(ctx);
8491 ret |= io_poll_remove_all(ctx, task);
8492 ret |= io_kill_timeouts(ctx, task);
8499 * We need to iteratively cancel requests, in case a request has dependent
8500 * hard links. These persist even for failure of cancelations, hence keep
8501 * looping until none are found.
8503 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8504 struct files_struct *files)
8506 struct task_struct *task = current;
8508 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8509 task = ctx->sq_data->thread;
8511 io_cqring_overflow_flush(ctx, true, task, files);
8513 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8520 * Note that this task has used io_uring. We use it for cancelation purposes.
8522 static int io_uring_add_task_file(struct file *file)
8524 struct io_uring_task *tctx = current->io_uring;
8526 if (unlikely(!tctx)) {
8529 ret = io_uring_alloc_task_context(current);
8532 tctx = current->io_uring;
8534 if (tctx->last != file) {
8535 void *old = xa_load(&tctx->xa, (unsigned long)file);
8539 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8548 * Remove this io_uring_file -> task mapping.
8550 static void io_uring_del_task_file(struct file *file)
8552 struct io_uring_task *tctx = current->io_uring;
8554 if (tctx->last == file)
8556 file = xa_erase(&tctx->xa, (unsigned long)file);
8561 static void __io_uring_attempt_task_drop(struct file *file)
8563 struct file *old = xa_load(¤t->io_uring->xa, (unsigned long)file);
8566 io_uring_del_task_file(file);
8570 * Drop task note for this file if we're the only ones that hold it after
8573 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8575 if (!current->io_uring)
8578 * fput() is pending, will be 2 if the only other ref is our potential
8579 * task file note. If the task is exiting, drop regardless of count.
8581 if (!exiting && atomic_long_read(&file->f_count) != 2)
8584 __io_uring_attempt_task_drop(file);
8587 void __io_uring_files_cancel(struct files_struct *files)
8589 struct io_uring_task *tctx = current->io_uring;
8591 unsigned long index;
8593 /* make sure overflow events are dropped */
8594 tctx->in_idle = true;
8596 xa_for_each(&tctx->xa, index, file) {
8597 struct io_ring_ctx *ctx = file->private_data;
8599 io_uring_cancel_task_requests(ctx, files);
8601 io_uring_del_task_file(file);
8605 static inline bool io_uring_task_idle(struct io_uring_task *tctx)
8607 return atomic_long_read(&tctx->req_issue) ==
8608 atomic_long_read(&tctx->req_complete);
8612 * Find any io_uring fd that this task has registered or done IO on, and cancel
8615 void __io_uring_task_cancel(void)
8617 struct io_uring_task *tctx = current->io_uring;
8621 /* make sure overflow events are dropped */
8622 tctx->in_idle = true;
8624 while (!io_uring_task_idle(tctx)) {
8625 /* read completions before cancelations */
8626 completions = atomic_long_read(&tctx->req_complete);
8627 __io_uring_files_cancel(NULL);
8629 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8632 * If we've seen completions, retry. This avoids a race where
8633 * a completion comes in before we did prepare_to_wait().
8635 if (completions != atomic_long_read(&tctx->req_complete))
8637 if (io_uring_task_idle(tctx))
8642 finish_wait(&tctx->wait, &wait);
8643 tctx->in_idle = false;
8646 static int io_uring_flush(struct file *file, void *data)
8648 struct io_ring_ctx *ctx = file->private_data;
8651 * If the task is going away, cancel work it may have pending
8653 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8656 io_uring_cancel_task_requests(ctx, data);
8657 io_uring_attempt_task_drop(file, !data);
8661 static void *io_uring_validate_mmap_request(struct file *file,
8662 loff_t pgoff, size_t sz)
8664 struct io_ring_ctx *ctx = file->private_data;
8665 loff_t offset = pgoff << PAGE_SHIFT;
8670 case IORING_OFF_SQ_RING:
8671 case IORING_OFF_CQ_RING:
8674 case IORING_OFF_SQES:
8678 return ERR_PTR(-EINVAL);
8681 page = virt_to_head_page(ptr);
8682 if (sz > page_size(page))
8683 return ERR_PTR(-EINVAL);
8690 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8692 size_t sz = vma->vm_end - vma->vm_start;
8696 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8698 return PTR_ERR(ptr);
8700 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8701 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8704 #else /* !CONFIG_MMU */
8706 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8708 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8711 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8713 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8716 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8717 unsigned long addr, unsigned long len,
8718 unsigned long pgoff, unsigned long flags)
8722 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8724 return PTR_ERR(ptr);
8726 return (unsigned long) ptr;
8729 #endif /* !CONFIG_MMU */
8731 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8736 if (!io_sqring_full(ctx))
8739 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8741 if (!io_sqring_full(ctx))
8745 } while (!signal_pending(current));
8747 finish_wait(&ctx->sqo_sq_wait, &wait);
8750 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8751 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8754 struct io_ring_ctx *ctx;
8761 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8762 IORING_ENTER_SQ_WAIT))
8770 if (f.file->f_op != &io_uring_fops)
8774 ctx = f.file->private_data;
8775 if (!percpu_ref_tryget(&ctx->refs))
8779 if (ctx->flags & IORING_SETUP_R_DISABLED)
8783 * For SQ polling, the thread will do all submissions and completions.
8784 * Just return the requested submit count, and wake the thread if
8788 if (ctx->flags & IORING_SETUP_SQPOLL) {
8789 if (!list_empty_careful(&ctx->cq_overflow_list))
8790 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8791 if (flags & IORING_ENTER_SQ_WAKEUP)
8792 wake_up(&ctx->sq_data->wait);
8793 if (flags & IORING_ENTER_SQ_WAIT)
8794 io_sqpoll_wait_sq(ctx);
8795 submitted = to_submit;
8796 } else if (to_submit) {
8797 ret = io_uring_add_task_file(f.file);
8800 mutex_lock(&ctx->uring_lock);
8801 submitted = io_submit_sqes(ctx, to_submit);
8802 mutex_unlock(&ctx->uring_lock);
8804 if (submitted != to_submit)
8807 if (flags & IORING_ENTER_GETEVENTS) {
8808 min_complete = min(min_complete, ctx->cq_entries);
8811 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8812 * space applications don't need to do io completion events
8813 * polling again, they can rely on io_sq_thread to do polling
8814 * work, which can reduce cpu usage and uring_lock contention.
8816 if (ctx->flags & IORING_SETUP_IOPOLL &&
8817 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8818 ret = io_iopoll_check(ctx, min_complete);
8820 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8825 percpu_ref_put(&ctx->refs);
8828 return submitted ? submitted : ret;
8831 #ifdef CONFIG_PROC_FS
8832 static int io_uring_show_cred(int id, void *p, void *data)
8834 const struct cred *cred = p;
8835 struct seq_file *m = data;
8836 struct user_namespace *uns = seq_user_ns(m);
8837 struct group_info *gi;
8842 seq_printf(m, "%5d\n", id);
8843 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8844 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8845 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8846 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8847 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8848 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8849 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8850 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8851 seq_puts(m, "\n\tGroups:\t");
8852 gi = cred->group_info;
8853 for (g = 0; g < gi->ngroups; g++) {
8854 seq_put_decimal_ull(m, g ? " " : "",
8855 from_kgid_munged(uns, gi->gid[g]));
8857 seq_puts(m, "\n\tCapEff:\t");
8858 cap = cred->cap_effective;
8859 CAP_FOR_EACH_U32(__capi)
8860 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8865 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8867 struct io_sq_data *sq = NULL;
8872 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8873 * since fdinfo case grabs it in the opposite direction of normal use
8874 * cases. If we fail to get the lock, we just don't iterate any
8875 * structures that could be going away outside the io_uring mutex.
8877 has_lock = mutex_trylock(&ctx->uring_lock);
8879 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
8882 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
8883 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
8884 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8885 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8886 struct fixed_file_table *table;
8889 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8890 f = table->files[i & IORING_FILE_TABLE_MASK];
8892 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8894 seq_printf(m, "%5u: <none>\n", i);
8896 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8897 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8898 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8900 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8901 (unsigned int) buf->len);
8903 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8904 seq_printf(m, "Personalities:\n");
8905 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8907 seq_printf(m, "PollList:\n");
8908 spin_lock_irq(&ctx->completion_lock);
8909 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8910 struct hlist_head *list = &ctx->cancel_hash[i];
8911 struct io_kiocb *req;
8913 hlist_for_each_entry(req, list, hash_node)
8914 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8915 req->task->task_works != NULL);
8917 spin_unlock_irq(&ctx->completion_lock);
8919 mutex_unlock(&ctx->uring_lock);
8922 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8924 struct io_ring_ctx *ctx = f->private_data;
8926 if (percpu_ref_tryget(&ctx->refs)) {
8927 __io_uring_show_fdinfo(ctx, m);
8928 percpu_ref_put(&ctx->refs);
8933 static const struct file_operations io_uring_fops = {
8934 .release = io_uring_release,
8935 .flush = io_uring_flush,
8936 .mmap = io_uring_mmap,
8938 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8939 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8941 .poll = io_uring_poll,
8942 .fasync = io_uring_fasync,
8943 #ifdef CONFIG_PROC_FS
8944 .show_fdinfo = io_uring_show_fdinfo,
8948 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8949 struct io_uring_params *p)
8951 struct io_rings *rings;
8952 size_t size, sq_array_offset;
8954 /* make sure these are sane, as we already accounted them */
8955 ctx->sq_entries = p->sq_entries;
8956 ctx->cq_entries = p->cq_entries;
8958 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8959 if (size == SIZE_MAX)
8962 rings = io_mem_alloc(size);
8967 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8968 rings->sq_ring_mask = p->sq_entries - 1;
8969 rings->cq_ring_mask = p->cq_entries - 1;
8970 rings->sq_ring_entries = p->sq_entries;
8971 rings->cq_ring_entries = p->cq_entries;
8972 ctx->sq_mask = rings->sq_ring_mask;
8973 ctx->cq_mask = rings->cq_ring_mask;
8975 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8976 if (size == SIZE_MAX) {
8977 io_mem_free(ctx->rings);
8982 ctx->sq_sqes = io_mem_alloc(size);
8983 if (!ctx->sq_sqes) {
8984 io_mem_free(ctx->rings);
8993 * Allocate an anonymous fd, this is what constitutes the application
8994 * visible backing of an io_uring instance. The application mmaps this
8995 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8996 * we have to tie this fd to a socket for file garbage collection purposes.
8998 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9003 #if defined(CONFIG_UNIX)
9004 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9010 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9014 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9015 O_RDWR | O_CLOEXEC);
9019 ret = PTR_ERR(file);
9023 #if defined(CONFIG_UNIX)
9024 ctx->ring_sock->file = file;
9026 if (unlikely(io_uring_add_task_file(file))) {
9027 file = ERR_PTR(-ENOMEM);
9030 fd_install(ret, file);
9033 #if defined(CONFIG_UNIX)
9034 sock_release(ctx->ring_sock);
9035 ctx->ring_sock = NULL;
9040 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9041 struct io_uring_params __user *params)
9043 struct user_struct *user = NULL;
9044 struct io_ring_ctx *ctx;
9050 if (entries > IORING_MAX_ENTRIES) {
9051 if (!(p->flags & IORING_SETUP_CLAMP))
9053 entries = IORING_MAX_ENTRIES;
9057 * Use twice as many entries for the CQ ring. It's possible for the
9058 * application to drive a higher depth than the size of the SQ ring,
9059 * since the sqes are only used at submission time. This allows for
9060 * some flexibility in overcommitting a bit. If the application has
9061 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9062 * of CQ ring entries manually.
9064 p->sq_entries = roundup_pow_of_two(entries);
9065 if (p->flags & IORING_SETUP_CQSIZE) {
9067 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9068 * to a power-of-two, if it isn't already. We do NOT impose
9069 * any cq vs sq ring sizing.
9071 if (p->cq_entries < p->sq_entries)
9073 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9074 if (!(p->flags & IORING_SETUP_CLAMP))
9076 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9078 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9080 p->cq_entries = 2 * p->sq_entries;
9083 user = get_uid(current_user());
9084 limit_mem = !capable(CAP_IPC_LOCK);
9087 ret = __io_account_mem(user,
9088 ring_pages(p->sq_entries, p->cq_entries));
9095 ctx = io_ring_ctx_alloc(p);
9098 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9103 ctx->compat = in_compat_syscall();
9105 ctx->creds = get_current_cred();
9107 ctx->sqo_task = get_task_struct(current);
9110 * This is just grabbed for accounting purposes. When a process exits,
9111 * the mm is exited and dropped before the files, hence we need to hang
9112 * on to this mm purely for the purposes of being able to unaccount
9113 * memory (locked/pinned vm). It's not used for anything else.
9115 mmgrab(current->mm);
9116 ctx->mm_account = current->mm;
9118 #ifdef CONFIG_BLK_CGROUP
9120 * The sq thread will belong to the original cgroup it was inited in.
9121 * If the cgroup goes offline (e.g. disabling the io controller), then
9122 * issued bios will be associated with the closest cgroup later in the
9126 ctx->sqo_blkcg_css = blkcg_css();
9127 ret = css_tryget_online(ctx->sqo_blkcg_css);
9130 /* don't init against a dying cgroup, have the user try again */
9131 ctx->sqo_blkcg_css = NULL;
9138 * Account memory _before_ installing the file descriptor. Once
9139 * the descriptor is installed, it can get closed at any time. Also
9140 * do this before hitting the general error path, as ring freeing
9141 * will un-account as well.
9143 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9145 ctx->limit_mem = limit_mem;
9147 ret = io_allocate_scq_urings(ctx, p);
9151 ret = io_sq_offload_create(ctx, p);
9155 if (!(p->flags & IORING_SETUP_R_DISABLED))
9156 io_sq_offload_start(ctx);
9158 memset(&p->sq_off, 0, sizeof(p->sq_off));
9159 p->sq_off.head = offsetof(struct io_rings, sq.head);
9160 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9161 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9162 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9163 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9164 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9165 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9167 memset(&p->cq_off, 0, sizeof(p->cq_off));
9168 p->cq_off.head = offsetof(struct io_rings, cq.head);
9169 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9170 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9171 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9172 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9173 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9174 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9176 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9177 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9178 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9179 IORING_FEAT_POLL_32BITS;
9181 if (copy_to_user(params, p, sizeof(*p))) {
9187 * Install ring fd as the very last thing, so we don't risk someone
9188 * having closed it before we finish setup
9190 ret = io_uring_get_fd(ctx);
9194 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9197 io_ring_ctx_wait_and_kill(ctx);
9202 * Sets up an aio uring context, and returns the fd. Applications asks for a
9203 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9204 * params structure passed in.
9206 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9208 struct io_uring_params p;
9211 if (copy_from_user(&p, params, sizeof(p)))
9213 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9218 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9219 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9220 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9221 IORING_SETUP_R_DISABLED))
9224 return io_uring_create(entries, &p, params);
9227 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9228 struct io_uring_params __user *, params)
9230 return io_uring_setup(entries, params);
9233 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9235 struct io_uring_probe *p;
9239 size = struct_size(p, ops, nr_args);
9240 if (size == SIZE_MAX)
9242 p = kzalloc(size, GFP_KERNEL);
9247 if (copy_from_user(p, arg, size))
9250 if (memchr_inv(p, 0, size))
9253 p->last_op = IORING_OP_LAST - 1;
9254 if (nr_args > IORING_OP_LAST)
9255 nr_args = IORING_OP_LAST;
9257 for (i = 0; i < nr_args; i++) {
9259 if (!io_op_defs[i].not_supported)
9260 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9265 if (copy_to_user(arg, p, size))
9272 static int io_register_personality(struct io_ring_ctx *ctx)
9274 const struct cred *creds = get_current_cred();
9277 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9278 USHRT_MAX, GFP_KERNEL);
9284 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9286 const struct cred *old_creds;
9288 old_creds = idr_remove(&ctx->personality_idr, id);
9290 put_cred(old_creds);
9297 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9298 unsigned int nr_args)
9300 struct io_uring_restriction *res;
9304 /* Restrictions allowed only if rings started disabled */
9305 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9308 /* We allow only a single restrictions registration */
9309 if (ctx->restrictions.registered)
9312 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9315 size = array_size(nr_args, sizeof(*res));
9316 if (size == SIZE_MAX)
9319 res = memdup_user(arg, size);
9321 return PTR_ERR(res);
9325 for (i = 0; i < nr_args; i++) {
9326 switch (res[i].opcode) {
9327 case IORING_RESTRICTION_REGISTER_OP:
9328 if (res[i].register_op >= IORING_REGISTER_LAST) {
9333 __set_bit(res[i].register_op,
9334 ctx->restrictions.register_op);
9336 case IORING_RESTRICTION_SQE_OP:
9337 if (res[i].sqe_op >= IORING_OP_LAST) {
9342 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9344 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9345 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9347 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9348 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9357 /* Reset all restrictions if an error happened */
9359 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9361 ctx->restrictions.registered = true;
9367 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9369 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9372 if (ctx->restrictions.registered)
9373 ctx->restricted = 1;
9375 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9377 io_sq_offload_start(ctx);
9382 static bool io_register_op_must_quiesce(int op)
9385 case IORING_UNREGISTER_FILES:
9386 case IORING_REGISTER_FILES_UPDATE:
9387 case IORING_REGISTER_PROBE:
9388 case IORING_REGISTER_PERSONALITY:
9389 case IORING_UNREGISTER_PERSONALITY:
9396 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9397 void __user *arg, unsigned nr_args)
9398 __releases(ctx->uring_lock)
9399 __acquires(ctx->uring_lock)
9404 * We're inside the ring mutex, if the ref is already dying, then
9405 * someone else killed the ctx or is already going through
9406 * io_uring_register().
9408 if (percpu_ref_is_dying(&ctx->refs))
9411 if (io_register_op_must_quiesce(opcode)) {
9412 percpu_ref_kill(&ctx->refs);
9415 * Drop uring mutex before waiting for references to exit. If
9416 * another thread is currently inside io_uring_enter() it might
9417 * need to grab the uring_lock to make progress. If we hold it
9418 * here across the drain wait, then we can deadlock. It's safe
9419 * to drop the mutex here, since no new references will come in
9420 * after we've killed the percpu ref.
9422 mutex_unlock(&ctx->uring_lock);
9424 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9427 ret = io_run_task_work_sig();
9432 mutex_lock(&ctx->uring_lock);
9435 percpu_ref_resurrect(&ctx->refs);
9440 if (ctx->restricted) {
9441 if (opcode >= IORING_REGISTER_LAST) {
9446 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9453 case IORING_REGISTER_BUFFERS:
9454 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9456 case IORING_UNREGISTER_BUFFERS:
9460 ret = io_sqe_buffer_unregister(ctx);
9462 case IORING_REGISTER_FILES:
9463 ret = io_sqe_files_register(ctx, arg, nr_args);
9465 case IORING_UNREGISTER_FILES:
9469 ret = io_sqe_files_unregister(ctx);
9471 case IORING_REGISTER_FILES_UPDATE:
9472 ret = io_sqe_files_update(ctx, arg, nr_args);
9474 case IORING_REGISTER_EVENTFD:
9475 case IORING_REGISTER_EVENTFD_ASYNC:
9479 ret = io_eventfd_register(ctx, arg);
9482 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9483 ctx->eventfd_async = 1;
9485 ctx->eventfd_async = 0;
9487 case IORING_UNREGISTER_EVENTFD:
9491 ret = io_eventfd_unregister(ctx);
9493 case IORING_REGISTER_PROBE:
9495 if (!arg || nr_args > 256)
9497 ret = io_probe(ctx, arg, nr_args);
9499 case IORING_REGISTER_PERSONALITY:
9503 ret = io_register_personality(ctx);
9505 case IORING_UNREGISTER_PERSONALITY:
9509 ret = io_unregister_personality(ctx, nr_args);
9511 case IORING_REGISTER_ENABLE_RINGS:
9515 ret = io_register_enable_rings(ctx);
9517 case IORING_REGISTER_RESTRICTIONS:
9518 ret = io_register_restrictions(ctx, arg, nr_args);
9526 if (io_register_op_must_quiesce(opcode)) {
9527 /* bring the ctx back to life */
9528 percpu_ref_reinit(&ctx->refs);
9530 reinit_completion(&ctx->ref_comp);
9535 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9536 void __user *, arg, unsigned int, nr_args)
9538 struct io_ring_ctx *ctx;
9547 if (f.file->f_op != &io_uring_fops)
9550 ctx = f.file->private_data;
9552 mutex_lock(&ctx->uring_lock);
9553 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9554 mutex_unlock(&ctx->uring_lock);
9555 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9556 ctx->cq_ev_fd != NULL, ret);
9562 static int __init io_uring_init(void)
9564 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9565 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9566 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9569 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9570 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9571 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9572 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9573 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9574 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9575 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9576 BUILD_BUG_SQE_ELEM(8, __u64, off);
9577 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9578 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9579 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9580 BUILD_BUG_SQE_ELEM(24, __u32, len);
9581 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9582 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9583 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9584 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9585 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9586 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9587 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9588 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9589 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9590 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9591 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9592 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9593 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9594 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9595 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9596 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9597 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9598 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9599 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9601 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9602 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9603 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9606 __initcall(io_uring_init);