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);
1184 static void io_req_clean_work(struct io_kiocb *req)
1186 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1189 req->flags &= ~REQ_F_WORK_INITIALIZED;
1192 mmdrop(req->work.mm);
1193 req->work.mm = NULL;
1195 #ifdef CONFIG_BLK_CGROUP
1196 if (req->work.blkcg_css)
1197 css_put(req->work.blkcg_css);
1199 if (req->work.creds) {
1200 put_cred(req->work.creds);
1201 req->work.creds = NULL;
1204 struct fs_struct *fs = req->work.fs;
1206 spin_lock(&req->work.fs->lock);
1209 spin_unlock(&req->work.fs->lock);
1212 req->work.fs = NULL;
1216 static void io_prep_async_work(struct io_kiocb *req)
1218 const struct io_op_def *def = &io_op_defs[req->opcode];
1219 struct io_ring_ctx *ctx = req->ctx;
1221 io_req_init_async(req);
1223 if (req->flags & REQ_F_ISREG) {
1224 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1225 io_wq_hash_work(&req->work, file_inode(req->file));
1227 if (def->unbound_nonreg_file)
1228 req->work.flags |= IO_WQ_WORK_UNBOUND;
1230 if (!req->work.files && io_op_defs[req->opcode].file_table &&
1231 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1232 req->work.files = get_files_struct(current);
1233 get_nsproxy(current->nsproxy);
1234 req->work.nsproxy = current->nsproxy;
1235 req->flags |= REQ_F_INFLIGHT;
1237 spin_lock_irq(&ctx->inflight_lock);
1238 list_add(&req->inflight_entry, &ctx->inflight_list);
1239 spin_unlock_irq(&ctx->inflight_lock);
1241 if (!req->work.mm && def->needs_mm) {
1242 mmgrab(current->mm);
1243 req->work.mm = current->mm;
1245 #ifdef CONFIG_BLK_CGROUP
1246 if (!req->work.blkcg_css && def->needs_blkcg) {
1248 req->work.blkcg_css = blkcg_css();
1250 * This should be rare, either the cgroup is dying or the task
1251 * is moving cgroups. Just punt to root for the handful of ios.
1253 if (!css_tryget_online(req->work.blkcg_css))
1254 req->work.blkcg_css = NULL;
1258 if (!req->work.creds)
1259 req->work.creds = get_current_cred();
1260 if (!req->work.fs && def->needs_fs) {
1261 spin_lock(¤t->fs->lock);
1262 if (!current->fs->in_exec) {
1263 req->work.fs = current->fs;
1264 req->work.fs->users++;
1266 req->work.flags |= IO_WQ_WORK_CANCEL;
1268 spin_unlock(¤t->fs->lock);
1270 if (def->needs_fsize)
1271 req->work.fsize = rlimit(RLIMIT_FSIZE);
1273 req->work.fsize = RLIM_INFINITY;
1276 static void io_prep_async_link(struct io_kiocb *req)
1278 struct io_kiocb *cur;
1280 io_prep_async_work(req);
1281 if (req->flags & REQ_F_LINK_HEAD)
1282 list_for_each_entry(cur, &req->link_list, link_list)
1283 io_prep_async_work(cur);
1286 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1288 struct io_ring_ctx *ctx = req->ctx;
1289 struct io_kiocb *link = io_prep_linked_timeout(req);
1291 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1292 &req->work, req->flags);
1293 io_wq_enqueue(ctx->io_wq, &req->work);
1297 static void io_queue_async_work(struct io_kiocb *req)
1299 struct io_kiocb *link;
1301 /* init ->work of the whole link before punting */
1302 io_prep_async_link(req);
1303 link = __io_queue_async_work(req);
1306 io_queue_linked_timeout(link);
1309 static void io_kill_timeout(struct io_kiocb *req)
1311 struct io_timeout_data *io = req->async_data;
1314 ret = hrtimer_try_to_cancel(&io->timer);
1316 atomic_set(&req->ctx->cq_timeouts,
1317 atomic_read(&req->ctx->cq_timeouts) + 1);
1318 list_del_init(&req->timeout.list);
1319 req->flags |= REQ_F_COMP_LOCKED;
1320 io_cqring_fill_event(req, 0);
1325 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1327 struct io_ring_ctx *ctx = req->ctx;
1329 if (!tsk || req->task == tsk)
1331 if (ctx->flags & IORING_SETUP_SQPOLL) {
1332 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1339 * Returns true if we found and killed one or more timeouts
1341 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1343 struct io_kiocb *req, *tmp;
1346 spin_lock_irq(&ctx->completion_lock);
1347 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1348 if (io_task_match(req, tsk)) {
1349 io_kill_timeout(req);
1353 spin_unlock_irq(&ctx->completion_lock);
1354 return canceled != 0;
1357 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1360 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1361 struct io_defer_entry, list);
1362 struct io_kiocb *link;
1364 if (req_need_defer(de->req, de->seq))
1366 list_del_init(&de->list);
1367 /* punt-init is done before queueing for defer */
1368 link = __io_queue_async_work(de->req);
1370 __io_queue_linked_timeout(link);
1371 /* drop submission reference */
1372 link->flags |= REQ_F_COMP_LOCKED;
1376 } while (!list_empty(&ctx->defer_list));
1379 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1381 while (!list_empty(&ctx->timeout_list)) {
1382 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1383 struct io_kiocb, timeout.list);
1385 if (io_is_timeout_noseq(req))
1387 if (req->timeout.target_seq != ctx->cached_cq_tail
1388 - atomic_read(&ctx->cq_timeouts))
1391 list_del_init(&req->timeout.list);
1392 io_kill_timeout(req);
1396 static void io_commit_cqring(struct io_ring_ctx *ctx)
1398 io_flush_timeouts(ctx);
1399 __io_commit_cqring(ctx);
1401 if (unlikely(!list_empty(&ctx->defer_list)))
1402 __io_queue_deferred(ctx);
1405 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1407 struct io_rings *r = ctx->rings;
1409 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1412 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1414 struct io_rings *rings = ctx->rings;
1417 tail = ctx->cached_cq_tail;
1419 * writes to the cq entry need to come after reading head; the
1420 * control dependency is enough as we're using WRITE_ONCE to
1423 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1426 ctx->cached_cq_tail++;
1427 return &rings->cqes[tail & ctx->cq_mask];
1430 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1434 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1436 if (!ctx->eventfd_async)
1438 return io_wq_current_is_worker();
1441 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1443 if (waitqueue_active(&ctx->wait))
1444 wake_up(&ctx->wait);
1445 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1446 wake_up(&ctx->sq_data->wait);
1447 if (io_should_trigger_evfd(ctx))
1448 eventfd_signal(ctx->cq_ev_fd, 1);
1451 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1453 if (list_empty(&ctx->cq_overflow_list)) {
1454 clear_bit(0, &ctx->sq_check_overflow);
1455 clear_bit(0, &ctx->cq_check_overflow);
1456 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1460 static inline bool io_match_files(struct io_kiocb *req,
1461 struct files_struct *files)
1465 if (req->flags & REQ_F_WORK_INITIALIZED)
1466 return req->work.files == files;
1470 /* Returns true if there are no backlogged entries after the flush */
1471 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1472 struct task_struct *tsk,
1473 struct files_struct *files)
1475 struct io_rings *rings = ctx->rings;
1476 struct io_kiocb *req, *tmp;
1477 struct io_uring_cqe *cqe;
1478 unsigned long flags;
1482 if (list_empty_careful(&ctx->cq_overflow_list))
1484 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1485 rings->cq_ring_entries))
1489 spin_lock_irqsave(&ctx->completion_lock, flags);
1491 /* if force is set, the ring is going away. always drop after that */
1493 ctx->cq_overflow_flushed = 1;
1496 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1497 if (tsk && req->task != tsk)
1499 if (!io_match_files(req, files))
1502 cqe = io_get_cqring(ctx);
1506 list_move(&req->compl.list, &list);
1508 WRITE_ONCE(cqe->user_data, req->user_data);
1509 WRITE_ONCE(cqe->res, req->result);
1510 WRITE_ONCE(cqe->flags, req->compl.cflags);
1512 WRITE_ONCE(ctx->rings->cq_overflow,
1513 atomic_inc_return(&ctx->cached_cq_overflow));
1517 io_commit_cqring(ctx);
1518 io_cqring_mark_overflow(ctx);
1520 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1521 io_cqring_ev_posted(ctx);
1523 while (!list_empty(&list)) {
1524 req = list_first_entry(&list, struct io_kiocb, compl.list);
1525 list_del(&req->compl.list);
1532 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1534 struct io_ring_ctx *ctx = req->ctx;
1535 struct io_uring_cqe *cqe;
1537 trace_io_uring_complete(ctx, req->user_data, res);
1540 * If we can't get a cq entry, userspace overflowed the
1541 * submission (by quite a lot). Increment the overflow count in
1544 cqe = io_get_cqring(ctx);
1546 WRITE_ONCE(cqe->user_data, req->user_data);
1547 WRITE_ONCE(cqe->res, res);
1548 WRITE_ONCE(cqe->flags, cflags);
1549 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1551 * If we're in ring overflow flush mode, or in task cancel mode,
1552 * then we cannot store the request for later flushing, we need
1553 * to drop it on the floor.
1555 WRITE_ONCE(ctx->rings->cq_overflow,
1556 atomic_inc_return(&ctx->cached_cq_overflow));
1558 if (list_empty(&ctx->cq_overflow_list)) {
1559 set_bit(0, &ctx->sq_check_overflow);
1560 set_bit(0, &ctx->cq_check_overflow);
1561 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1565 req->compl.cflags = cflags;
1566 refcount_inc(&req->refs);
1567 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1571 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1573 __io_cqring_fill_event(req, res, 0);
1576 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1578 struct io_ring_ctx *ctx = req->ctx;
1579 unsigned long flags;
1581 spin_lock_irqsave(&ctx->completion_lock, flags);
1582 __io_cqring_fill_event(req, res, cflags);
1583 io_commit_cqring(ctx);
1584 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1586 io_cqring_ev_posted(ctx);
1589 static void io_submit_flush_completions(struct io_comp_state *cs)
1591 struct io_ring_ctx *ctx = cs->ctx;
1593 spin_lock_irq(&ctx->completion_lock);
1594 while (!list_empty(&cs->list)) {
1595 struct io_kiocb *req;
1597 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1598 list_del(&req->compl.list);
1599 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1600 if (!(req->flags & REQ_F_LINK_HEAD)) {
1601 req->flags |= REQ_F_COMP_LOCKED;
1604 spin_unlock_irq(&ctx->completion_lock);
1606 spin_lock_irq(&ctx->completion_lock);
1609 io_commit_cqring(ctx);
1610 spin_unlock_irq(&ctx->completion_lock);
1612 io_cqring_ev_posted(ctx);
1616 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1617 struct io_comp_state *cs)
1620 io_cqring_add_event(req, res, cflags);
1625 req->compl.cflags = cflags;
1626 list_add_tail(&req->compl.list, &cs->list);
1628 io_submit_flush_completions(cs);
1632 static void io_req_complete(struct io_kiocb *req, long res)
1634 __io_req_complete(req, res, 0, NULL);
1637 static inline bool io_is_fallback_req(struct io_kiocb *req)
1639 return req == (struct io_kiocb *)
1640 ((unsigned long) req->ctx->fallback_req & ~1UL);
1643 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1645 struct io_kiocb *req;
1647 req = ctx->fallback_req;
1648 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1654 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1655 struct io_submit_state *state)
1657 if (!state->free_reqs) {
1658 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1662 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1663 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1666 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1667 * retry single alloc to be on the safe side.
1669 if (unlikely(ret <= 0)) {
1670 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1671 if (!state->reqs[0])
1675 state->free_reqs = ret;
1679 return state->reqs[state->free_reqs];
1681 return io_get_fallback_req(ctx);
1684 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1688 percpu_ref_put(req->fixed_file_refs);
1693 static void io_dismantle_req(struct io_kiocb *req)
1697 if (req->async_data)
1698 kfree(req->async_data);
1700 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1702 io_req_clean_work(req);
1705 static void __io_free_req_finish(struct io_kiocb *req)
1707 struct io_uring_task *tctx = req->task->io_uring;
1708 struct io_ring_ctx *ctx = req->ctx;
1710 atomic_long_inc(&tctx->req_complete);
1712 wake_up(&tctx->wait);
1713 put_task_struct(req->task);
1715 if (likely(!io_is_fallback_req(req)))
1716 kmem_cache_free(req_cachep, req);
1718 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1719 percpu_ref_put(&ctx->refs);
1722 static void io_req_task_file_table_put(struct callback_head *cb)
1724 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1726 io_dismantle_req(req);
1727 __io_free_req_finish(req);
1730 static void __io_free_req(struct io_kiocb *req)
1732 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1733 io_dismantle_req(req);
1734 __io_free_req_finish(req);
1738 init_task_work(&req->task_work, io_req_task_file_table_put);
1739 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1740 if (unlikely(ret)) {
1741 struct task_struct *tsk;
1743 tsk = io_wq_get_task(req->ctx->io_wq);
1744 task_work_add(tsk, &req->task_work, 0);
1749 static bool io_link_cancel_timeout(struct io_kiocb *req)
1751 struct io_timeout_data *io = req->async_data;
1752 struct io_ring_ctx *ctx = req->ctx;
1755 ret = hrtimer_try_to_cancel(&io->timer);
1757 req->flags |= REQ_F_COMP_LOCKED;
1758 io_cqring_fill_event(req, -ECANCELED);
1759 io_commit_cqring(ctx);
1760 req->flags &= ~REQ_F_LINK_HEAD;
1768 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1770 struct io_kiocb *link;
1773 if (list_empty(&req->link_list))
1775 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1776 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1779 list_del_init(&link->link_list);
1780 wake_ev = io_link_cancel_timeout(link);
1781 req->flags &= ~REQ_F_LINK_TIMEOUT;
1785 static void io_kill_linked_timeout(struct io_kiocb *req)
1787 struct io_ring_ctx *ctx = req->ctx;
1790 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1791 unsigned long flags;
1793 spin_lock_irqsave(&ctx->completion_lock, flags);
1794 wake_ev = __io_kill_linked_timeout(req);
1795 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1797 wake_ev = __io_kill_linked_timeout(req);
1801 io_cqring_ev_posted(ctx);
1804 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1806 struct io_kiocb *nxt;
1809 * The list should never be empty when we are called here. But could
1810 * potentially happen if the chain is messed up, check to be on the
1813 if (unlikely(list_empty(&req->link_list)))
1816 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1817 list_del_init(&req->link_list);
1818 if (!list_empty(&nxt->link_list))
1819 nxt->flags |= REQ_F_LINK_HEAD;
1824 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1826 static void __io_fail_links(struct io_kiocb *req)
1828 struct io_ring_ctx *ctx = req->ctx;
1830 while (!list_empty(&req->link_list)) {
1831 struct io_kiocb *link = list_first_entry(&req->link_list,
1832 struct io_kiocb, link_list);
1834 list_del_init(&link->link_list);
1835 trace_io_uring_fail_link(req, link);
1837 io_cqring_fill_event(link, -ECANCELED);
1838 link->flags |= REQ_F_COMP_LOCKED;
1839 __io_double_put_req(link);
1842 io_commit_cqring(ctx);
1843 io_cqring_ev_posted(ctx);
1846 static void io_fail_links(struct io_kiocb *req)
1848 struct io_ring_ctx *ctx = req->ctx;
1850 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1851 unsigned long flags;
1853 spin_lock_irqsave(&ctx->completion_lock, flags);
1854 __io_fail_links(req);
1855 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1857 __io_fail_links(req);
1860 io_cqring_ev_posted(ctx);
1863 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1865 req->flags &= ~REQ_F_LINK_HEAD;
1866 if (req->flags & REQ_F_LINK_TIMEOUT)
1867 io_kill_linked_timeout(req);
1870 * If LINK is set, we have dependent requests in this chain. If we
1871 * didn't fail this request, queue the first one up, moving any other
1872 * dependencies to the next request. In case of failure, fail the rest
1875 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1876 return io_req_link_next(req);
1881 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1883 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1885 return __io_req_find_next(req);
1888 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1890 struct task_struct *tsk = req->task;
1891 struct io_ring_ctx *ctx = req->ctx;
1894 if (tsk->flags & PF_EXITING)
1898 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1899 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1900 * processing task_work. There's no reliable way to tell if TWA_RESUME
1904 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1905 notify = TWA_SIGNAL;
1907 ret = task_work_add(tsk, &req->task_work, notify);
1909 wake_up_process(tsk);
1914 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1916 struct io_ring_ctx *ctx = req->ctx;
1918 spin_lock_irq(&ctx->completion_lock);
1919 io_cqring_fill_event(req, error);
1920 io_commit_cqring(ctx);
1921 spin_unlock_irq(&ctx->completion_lock);
1923 io_cqring_ev_posted(ctx);
1924 req_set_fail_links(req);
1925 io_double_put_req(req);
1928 static void io_req_task_cancel(struct callback_head *cb)
1930 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1931 struct io_ring_ctx *ctx = req->ctx;
1933 __io_req_task_cancel(req, -ECANCELED);
1934 percpu_ref_put(&ctx->refs);
1937 static void __io_req_task_submit(struct io_kiocb *req)
1939 struct io_ring_ctx *ctx = req->ctx;
1941 if (!__io_sq_thread_acquire_mm(ctx)) {
1942 mutex_lock(&ctx->uring_lock);
1943 __io_queue_sqe(req, NULL);
1944 mutex_unlock(&ctx->uring_lock);
1946 __io_req_task_cancel(req, -EFAULT);
1950 static void io_req_task_submit(struct callback_head *cb)
1952 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1953 struct io_ring_ctx *ctx = req->ctx;
1955 __io_req_task_submit(req);
1956 percpu_ref_put(&ctx->refs);
1959 static void io_req_task_queue(struct io_kiocb *req)
1963 init_task_work(&req->task_work, io_req_task_submit);
1964 percpu_ref_get(&req->ctx->refs);
1966 ret = io_req_task_work_add(req, true);
1967 if (unlikely(ret)) {
1968 struct task_struct *tsk;
1970 init_task_work(&req->task_work, io_req_task_cancel);
1971 tsk = io_wq_get_task(req->ctx->io_wq);
1972 task_work_add(tsk, &req->task_work, 0);
1973 wake_up_process(tsk);
1977 static void io_queue_next(struct io_kiocb *req)
1979 struct io_kiocb *nxt = io_req_find_next(req);
1982 io_req_task_queue(nxt);
1985 static void io_free_req(struct io_kiocb *req)
1992 void *reqs[IO_IOPOLL_BATCH];
1995 struct task_struct *task;
1999 static inline void io_init_req_batch(struct req_batch *rb)
2006 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2007 struct req_batch *rb)
2009 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2010 percpu_ref_put_many(&ctx->refs, rb->to_free);
2014 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2015 struct req_batch *rb)
2018 __io_req_free_batch_flush(ctx, rb);
2020 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2021 put_task_struct_many(rb->task, rb->task_refs);
2026 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2028 if (unlikely(io_is_fallback_req(req))) {
2032 if (req->flags & REQ_F_LINK_HEAD)
2035 if (req->task != rb->task) {
2037 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2038 put_task_struct_many(rb->task, rb->task_refs);
2040 rb->task = req->task;
2045 io_dismantle_req(req);
2046 rb->reqs[rb->to_free++] = req;
2047 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2048 __io_req_free_batch_flush(req->ctx, rb);
2052 * Drop reference to request, return next in chain (if there is one) if this
2053 * was the last reference to this request.
2055 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2057 struct io_kiocb *nxt = NULL;
2059 if (refcount_dec_and_test(&req->refs)) {
2060 nxt = io_req_find_next(req);
2066 static void io_put_req(struct io_kiocb *req)
2068 if (refcount_dec_and_test(&req->refs))
2072 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2074 struct io_kiocb *nxt;
2077 * A ref is owned by io-wq in which context we're. So, if that's the
2078 * last one, it's safe to steal next work. False negatives are Ok,
2079 * it just will be re-punted async in io_put_work()
2081 if (refcount_read(&req->refs) != 1)
2084 nxt = io_req_find_next(req);
2085 return nxt ? &nxt->work : NULL;
2089 * Must only be used if we don't need to care about links, usually from
2090 * within the completion handling itself.
2092 static void __io_double_put_req(struct io_kiocb *req)
2094 /* drop both submit and complete references */
2095 if (refcount_sub_and_test(2, &req->refs))
2099 static void io_double_put_req(struct io_kiocb *req)
2101 /* drop both submit and complete references */
2102 if (refcount_sub_and_test(2, &req->refs))
2106 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2108 struct io_rings *rings = ctx->rings;
2110 if (test_bit(0, &ctx->cq_check_overflow)) {
2112 * noflush == true is from the waitqueue handler, just ensure
2113 * we wake up the task, and the next invocation will flush the
2114 * entries. We cannot safely to it from here.
2116 if (noflush && !list_empty(&ctx->cq_overflow_list))
2119 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2122 /* See comment at the top of this file */
2124 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2127 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2129 struct io_rings *rings = ctx->rings;
2131 /* make sure SQ entry isn't read before tail */
2132 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2135 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2137 unsigned int cflags;
2139 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2140 cflags |= IORING_CQE_F_BUFFER;
2141 req->flags &= ~REQ_F_BUFFER_SELECTED;
2146 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2148 struct io_buffer *kbuf;
2150 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2151 return io_put_kbuf(req, kbuf);
2154 static inline bool io_run_task_work(void)
2157 * Not safe to run on exiting task, and the task_work handling will
2158 * not add work to such a task.
2160 if (unlikely(current->flags & PF_EXITING))
2162 if (current->task_works) {
2163 __set_current_state(TASK_RUNNING);
2171 static void io_iopoll_queue(struct list_head *again)
2173 struct io_kiocb *req;
2176 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2177 list_del(&req->inflight_entry);
2178 __io_complete_rw(req, -EAGAIN, 0, NULL);
2179 } while (!list_empty(again));
2183 * Find and free completed poll iocbs
2185 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2186 struct list_head *done)
2188 struct req_batch rb;
2189 struct io_kiocb *req;
2192 /* order with ->result store in io_complete_rw_iopoll() */
2195 io_init_req_batch(&rb);
2196 while (!list_empty(done)) {
2199 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2200 if (READ_ONCE(req->result) == -EAGAIN) {
2202 req->iopoll_completed = 0;
2203 list_move_tail(&req->inflight_entry, &again);
2206 list_del(&req->inflight_entry);
2208 if (req->flags & REQ_F_BUFFER_SELECTED)
2209 cflags = io_put_rw_kbuf(req);
2211 __io_cqring_fill_event(req, req->result, cflags);
2214 if (refcount_dec_and_test(&req->refs))
2215 io_req_free_batch(&rb, req);
2218 io_commit_cqring(ctx);
2219 if (ctx->flags & IORING_SETUP_SQPOLL)
2220 io_cqring_ev_posted(ctx);
2221 io_req_free_batch_finish(ctx, &rb);
2223 if (!list_empty(&again))
2224 io_iopoll_queue(&again);
2227 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2230 struct io_kiocb *req, *tmp;
2236 * Only spin for completions if we don't have multiple devices hanging
2237 * off our complete list, and we're under the requested amount.
2239 spin = !ctx->poll_multi_file && *nr_events < min;
2242 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2243 struct kiocb *kiocb = &req->rw.kiocb;
2246 * Move completed and retryable entries to our local lists.
2247 * If we find a request that requires polling, break out
2248 * and complete those lists first, if we have entries there.
2250 if (READ_ONCE(req->iopoll_completed)) {
2251 list_move_tail(&req->inflight_entry, &done);
2254 if (!list_empty(&done))
2257 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2261 /* iopoll may have completed current req */
2262 if (READ_ONCE(req->iopoll_completed))
2263 list_move_tail(&req->inflight_entry, &done);
2270 if (!list_empty(&done))
2271 io_iopoll_complete(ctx, nr_events, &done);
2277 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2278 * non-spinning poll check - we'll still enter the driver poll loop, but only
2279 * as a non-spinning completion check.
2281 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2284 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2287 ret = io_do_iopoll(ctx, nr_events, min);
2290 if (*nr_events >= min)
2298 * We can't just wait for polled events to come to us, we have to actively
2299 * find and complete them.
2301 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2303 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2306 mutex_lock(&ctx->uring_lock);
2307 while (!list_empty(&ctx->iopoll_list)) {
2308 unsigned int nr_events = 0;
2310 io_do_iopoll(ctx, &nr_events, 0);
2312 /* let it sleep and repeat later if can't complete a request */
2316 * Ensure we allow local-to-the-cpu processing to take place,
2317 * in this case we need to ensure that we reap all events.
2318 * Also let task_work, etc. to progress by releasing the mutex
2320 if (need_resched()) {
2321 mutex_unlock(&ctx->uring_lock);
2323 mutex_lock(&ctx->uring_lock);
2326 mutex_unlock(&ctx->uring_lock);
2329 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2331 unsigned int nr_events = 0;
2332 int iters = 0, ret = 0;
2335 * We disallow the app entering submit/complete with polling, but we
2336 * still need to lock the ring to prevent racing with polled issue
2337 * that got punted to a workqueue.
2339 mutex_lock(&ctx->uring_lock);
2342 * Don't enter poll loop if we already have events pending.
2343 * If we do, we can potentially be spinning for commands that
2344 * already triggered a CQE (eg in error).
2346 if (io_cqring_events(ctx, false))
2350 * If a submit got punted to a workqueue, we can have the
2351 * application entering polling for a command before it gets
2352 * issued. That app will hold the uring_lock for the duration
2353 * of the poll right here, so we need to take a breather every
2354 * now and then to ensure that the issue has a chance to add
2355 * the poll to the issued list. Otherwise we can spin here
2356 * forever, while the workqueue is stuck trying to acquire the
2359 if (!(++iters & 7)) {
2360 mutex_unlock(&ctx->uring_lock);
2362 mutex_lock(&ctx->uring_lock);
2365 ret = io_iopoll_getevents(ctx, &nr_events, min);
2369 } while (min && !nr_events && !need_resched());
2371 mutex_unlock(&ctx->uring_lock);
2375 static void kiocb_end_write(struct io_kiocb *req)
2378 * Tell lockdep we inherited freeze protection from submission
2381 if (req->flags & REQ_F_ISREG) {
2382 struct inode *inode = file_inode(req->file);
2384 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2386 file_end_write(req->file);
2389 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2390 struct io_comp_state *cs)
2392 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2395 if (kiocb->ki_flags & IOCB_WRITE)
2396 kiocb_end_write(req);
2398 if (res != req->result)
2399 req_set_fail_links(req);
2400 if (req->flags & REQ_F_BUFFER_SELECTED)
2401 cflags = io_put_rw_kbuf(req);
2402 __io_req_complete(req, res, cflags, cs);
2406 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2408 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2409 ssize_t ret = -ECANCELED;
2410 struct iov_iter iter;
2418 switch (req->opcode) {
2419 case IORING_OP_READV:
2420 case IORING_OP_READ_FIXED:
2421 case IORING_OP_READ:
2424 case IORING_OP_WRITEV:
2425 case IORING_OP_WRITE_FIXED:
2426 case IORING_OP_WRITE:
2430 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2435 if (!req->async_data) {
2436 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2439 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2447 req_set_fail_links(req);
2448 io_req_complete(req, ret);
2453 static bool io_rw_reissue(struct io_kiocb *req, long res)
2456 umode_t mode = file_inode(req->file)->i_mode;
2459 if (!S_ISBLK(mode) && !S_ISREG(mode))
2461 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2464 ret = io_sq_thread_acquire_mm(req->ctx, req);
2466 if (io_resubmit_prep(req, ret)) {
2467 refcount_inc(&req->refs);
2468 io_queue_async_work(req);
2476 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2477 struct io_comp_state *cs)
2479 if (!io_rw_reissue(req, res))
2480 io_complete_rw_common(&req->rw.kiocb, res, cs);
2483 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2485 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2487 __io_complete_rw(req, res, res2, NULL);
2490 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2492 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2494 if (kiocb->ki_flags & IOCB_WRITE)
2495 kiocb_end_write(req);
2497 if (res != -EAGAIN && res != req->result)
2498 req_set_fail_links(req);
2500 WRITE_ONCE(req->result, res);
2501 /* order with io_poll_complete() checking ->result */
2503 WRITE_ONCE(req->iopoll_completed, 1);
2507 * After the iocb has been issued, it's safe to be found on the poll list.
2508 * Adding the kiocb to the list AFTER submission ensures that we don't
2509 * find it from a io_iopoll_getevents() thread before the issuer is done
2510 * accessing the kiocb cookie.
2512 static void io_iopoll_req_issued(struct io_kiocb *req)
2514 struct io_ring_ctx *ctx = req->ctx;
2517 * Track whether we have multiple files in our lists. This will impact
2518 * how we do polling eventually, not spinning if we're on potentially
2519 * different devices.
2521 if (list_empty(&ctx->iopoll_list)) {
2522 ctx->poll_multi_file = false;
2523 } else if (!ctx->poll_multi_file) {
2524 struct io_kiocb *list_req;
2526 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2528 if (list_req->file != req->file)
2529 ctx->poll_multi_file = true;
2533 * For fast devices, IO may have already completed. If it has, add
2534 * it to the front so we find it first.
2536 if (READ_ONCE(req->iopoll_completed))
2537 list_add(&req->inflight_entry, &ctx->iopoll_list);
2539 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2541 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2542 wq_has_sleeper(&ctx->sq_data->wait))
2543 wake_up(&ctx->sq_data->wait);
2546 static void __io_state_file_put(struct io_submit_state *state)
2548 if (state->has_refs)
2549 fput_many(state->file, state->has_refs);
2553 static inline void io_state_file_put(struct io_submit_state *state)
2556 __io_state_file_put(state);
2560 * Get as many references to a file as we have IOs left in this submission,
2561 * assuming most submissions are for one file, or at least that each file
2562 * has more than one submission.
2564 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2570 if (state->fd == fd) {
2574 __io_state_file_put(state);
2576 state->file = fget_many(fd, state->ios_left);
2581 state->has_refs = state->ios_left - 1;
2585 static bool io_bdev_nowait(struct block_device *bdev)
2588 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2595 * If we tracked the file through the SCM inflight mechanism, we could support
2596 * any file. For now, just ensure that anything potentially problematic is done
2599 static bool io_file_supports_async(struct file *file, int rw)
2601 umode_t mode = file_inode(file)->i_mode;
2603 if (S_ISBLK(mode)) {
2604 if (io_bdev_nowait(file->f_inode->i_bdev))
2608 if (S_ISCHR(mode) || S_ISSOCK(mode))
2610 if (S_ISREG(mode)) {
2611 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2612 file->f_op != &io_uring_fops)
2617 /* any ->read/write should understand O_NONBLOCK */
2618 if (file->f_flags & O_NONBLOCK)
2621 if (!(file->f_mode & FMODE_NOWAIT))
2625 return file->f_op->read_iter != NULL;
2627 return file->f_op->write_iter != NULL;
2630 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2632 struct io_ring_ctx *ctx = req->ctx;
2633 struct kiocb *kiocb = &req->rw.kiocb;
2637 if (S_ISREG(file_inode(req->file)->i_mode))
2638 req->flags |= REQ_F_ISREG;
2640 kiocb->ki_pos = READ_ONCE(sqe->off);
2641 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2642 req->flags |= REQ_F_CUR_POS;
2643 kiocb->ki_pos = req->file->f_pos;
2645 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2646 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2647 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2651 ioprio = READ_ONCE(sqe->ioprio);
2653 ret = ioprio_check_cap(ioprio);
2657 kiocb->ki_ioprio = ioprio;
2659 kiocb->ki_ioprio = get_current_ioprio();
2661 /* don't allow async punt if RWF_NOWAIT was requested */
2662 if (kiocb->ki_flags & IOCB_NOWAIT)
2663 req->flags |= REQ_F_NOWAIT;
2665 if (ctx->flags & IORING_SETUP_IOPOLL) {
2666 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2667 !kiocb->ki_filp->f_op->iopoll)
2670 kiocb->ki_flags |= IOCB_HIPRI;
2671 kiocb->ki_complete = io_complete_rw_iopoll;
2672 req->iopoll_completed = 0;
2674 if (kiocb->ki_flags & IOCB_HIPRI)
2676 kiocb->ki_complete = io_complete_rw;
2679 req->rw.addr = READ_ONCE(sqe->addr);
2680 req->rw.len = READ_ONCE(sqe->len);
2681 req->buf_index = READ_ONCE(sqe->buf_index);
2685 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2691 case -ERESTARTNOINTR:
2692 case -ERESTARTNOHAND:
2693 case -ERESTART_RESTARTBLOCK:
2695 * We can't just restart the syscall, since previously
2696 * submitted sqes may already be in progress. Just fail this
2702 kiocb->ki_complete(kiocb, ret, 0);
2706 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2707 struct io_comp_state *cs)
2709 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2710 struct io_async_rw *io = req->async_data;
2712 /* add previously done IO, if any */
2713 if (io && io->bytes_done > 0) {
2715 ret = io->bytes_done;
2717 ret += io->bytes_done;
2720 if (req->flags & REQ_F_CUR_POS)
2721 req->file->f_pos = kiocb->ki_pos;
2722 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2723 __io_complete_rw(req, ret, 0, cs);
2725 io_rw_done(kiocb, ret);
2728 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2729 struct iov_iter *iter)
2731 struct io_ring_ctx *ctx = req->ctx;
2732 size_t len = req->rw.len;
2733 struct io_mapped_ubuf *imu;
2734 u16 index, buf_index = req->buf_index;
2738 if (unlikely(buf_index >= ctx->nr_user_bufs))
2740 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2741 imu = &ctx->user_bufs[index];
2742 buf_addr = req->rw.addr;
2745 if (buf_addr + len < buf_addr)
2747 /* not inside the mapped region */
2748 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2752 * May not be a start of buffer, set size appropriately
2753 * and advance us to the beginning.
2755 offset = buf_addr - imu->ubuf;
2756 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2760 * Don't use iov_iter_advance() here, as it's really slow for
2761 * using the latter parts of a big fixed buffer - it iterates
2762 * over each segment manually. We can cheat a bit here, because
2765 * 1) it's a BVEC iter, we set it up
2766 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2767 * first and last bvec
2769 * So just find our index, and adjust the iterator afterwards.
2770 * If the offset is within the first bvec (or the whole first
2771 * bvec, just use iov_iter_advance(). This makes it easier
2772 * since we can just skip the first segment, which may not
2773 * be PAGE_SIZE aligned.
2775 const struct bio_vec *bvec = imu->bvec;
2777 if (offset <= bvec->bv_len) {
2778 iov_iter_advance(iter, offset);
2780 unsigned long seg_skip;
2782 /* skip first vec */
2783 offset -= bvec->bv_len;
2784 seg_skip = 1 + (offset >> PAGE_SHIFT);
2786 iter->bvec = bvec + seg_skip;
2787 iter->nr_segs -= seg_skip;
2788 iter->count -= bvec->bv_len + offset;
2789 iter->iov_offset = offset & ~PAGE_MASK;
2796 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2799 mutex_unlock(&ctx->uring_lock);
2802 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2805 * "Normal" inline submissions always hold the uring_lock, since we
2806 * grab it from the system call. Same is true for the SQPOLL offload.
2807 * The only exception is when we've detached the request and issue it
2808 * from an async worker thread, grab the lock for that case.
2811 mutex_lock(&ctx->uring_lock);
2814 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2815 int bgid, struct io_buffer *kbuf,
2818 struct io_buffer *head;
2820 if (req->flags & REQ_F_BUFFER_SELECTED)
2823 io_ring_submit_lock(req->ctx, needs_lock);
2825 lockdep_assert_held(&req->ctx->uring_lock);
2827 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2829 if (!list_empty(&head->list)) {
2830 kbuf = list_last_entry(&head->list, struct io_buffer,
2832 list_del(&kbuf->list);
2835 idr_remove(&req->ctx->io_buffer_idr, bgid);
2837 if (*len > kbuf->len)
2840 kbuf = ERR_PTR(-ENOBUFS);
2843 io_ring_submit_unlock(req->ctx, needs_lock);
2848 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2851 struct io_buffer *kbuf;
2854 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2855 bgid = req->buf_index;
2856 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2859 req->rw.addr = (u64) (unsigned long) kbuf;
2860 req->flags |= REQ_F_BUFFER_SELECTED;
2861 return u64_to_user_ptr(kbuf->addr);
2864 #ifdef CONFIG_COMPAT
2865 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2868 struct compat_iovec __user *uiov;
2869 compat_ssize_t clen;
2873 uiov = u64_to_user_ptr(req->rw.addr);
2874 if (!access_ok(uiov, sizeof(*uiov)))
2876 if (__get_user(clen, &uiov->iov_len))
2882 buf = io_rw_buffer_select(req, &len, needs_lock);
2884 return PTR_ERR(buf);
2885 iov[0].iov_base = buf;
2886 iov[0].iov_len = (compat_size_t) len;
2891 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2894 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2898 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2901 len = iov[0].iov_len;
2904 buf = io_rw_buffer_select(req, &len, needs_lock);
2906 return PTR_ERR(buf);
2907 iov[0].iov_base = buf;
2908 iov[0].iov_len = len;
2912 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2915 if (req->flags & REQ_F_BUFFER_SELECTED) {
2916 struct io_buffer *kbuf;
2918 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2919 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2920 iov[0].iov_len = kbuf->len;
2925 else if (req->rw.len > 1)
2928 #ifdef CONFIG_COMPAT
2929 if (req->ctx->compat)
2930 return io_compat_import(req, iov, needs_lock);
2933 return __io_iov_buffer_select(req, iov, needs_lock);
2936 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2937 struct iovec **iovec, struct iov_iter *iter,
2940 void __user *buf = u64_to_user_ptr(req->rw.addr);
2941 size_t sqe_len = req->rw.len;
2945 opcode = req->opcode;
2946 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2948 return io_import_fixed(req, rw, iter);
2951 /* buffer index only valid with fixed read/write, or buffer select */
2952 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2955 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2956 if (req->flags & REQ_F_BUFFER_SELECT) {
2957 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2959 return PTR_ERR(buf);
2960 req->rw.len = sqe_len;
2963 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2965 return ret < 0 ? ret : sqe_len;
2968 if (req->flags & REQ_F_BUFFER_SELECT) {
2969 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2971 ret = (*iovec)->iov_len;
2972 iov_iter_init(iter, rw, *iovec, 1, ret);
2978 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2982 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2983 struct iovec **iovec, struct iov_iter *iter,
2986 struct io_async_rw *iorw = req->async_data;
2989 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
2991 return iov_iter_count(&iorw->iter);
2994 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2996 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3000 * For files that don't have ->read_iter() and ->write_iter(), handle them
3001 * by looping over ->read() or ->write() manually.
3003 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
3004 struct iov_iter *iter)
3009 * Don't support polled IO through this interface, and we can't
3010 * support non-blocking either. For the latter, this just causes
3011 * the kiocb to be handled from an async context.
3013 if (kiocb->ki_flags & IOCB_HIPRI)
3015 if (kiocb->ki_flags & IOCB_NOWAIT)
3018 while (iov_iter_count(iter)) {
3022 if (!iov_iter_is_bvec(iter)) {
3023 iovec = iov_iter_iovec(iter);
3025 /* fixed buffers import bvec */
3026 iovec.iov_base = kmap(iter->bvec->bv_page)
3028 iovec.iov_len = min(iter->count,
3029 iter->bvec->bv_len - iter->iov_offset);
3033 nr = file->f_op->read(file, iovec.iov_base,
3034 iovec.iov_len, io_kiocb_ppos(kiocb));
3036 nr = file->f_op->write(file, iovec.iov_base,
3037 iovec.iov_len, io_kiocb_ppos(kiocb));
3040 if (iov_iter_is_bvec(iter))
3041 kunmap(iter->bvec->bv_page);
3049 if (nr != iovec.iov_len)
3051 iov_iter_advance(iter, nr);
3057 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3058 const struct iovec *fast_iov, struct iov_iter *iter)
3060 struct io_async_rw *rw = req->async_data;
3062 memcpy(&rw->iter, iter, sizeof(*iter));
3063 rw->free_iovec = iovec;
3065 /* can only be fixed buffers, no need to do anything */
3066 if (iter->type == ITER_BVEC)
3069 unsigned iov_off = 0;
3071 rw->iter.iov = rw->fast_iov;
3072 if (iter->iov != fast_iov) {
3073 iov_off = iter->iov - fast_iov;
3074 rw->iter.iov += iov_off;
3076 if (rw->fast_iov != fast_iov)
3077 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3078 sizeof(struct iovec) * iter->nr_segs);
3080 req->flags |= REQ_F_NEED_CLEANUP;
3084 static inline int __io_alloc_async_data(struct io_kiocb *req)
3086 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3087 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3088 return req->async_data == NULL;
3091 static int io_alloc_async_data(struct io_kiocb *req)
3093 if (!io_op_defs[req->opcode].needs_async_data)
3096 return __io_alloc_async_data(req);
3099 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3100 const struct iovec *fast_iov,
3101 struct iov_iter *iter, bool force)
3103 if (!force && !io_op_defs[req->opcode].needs_async_data)
3105 if (!req->async_data) {
3106 if (__io_alloc_async_data(req))
3109 io_req_map_rw(req, iovec, fast_iov, iter);
3114 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3116 struct io_async_rw *iorw = req->async_data;
3117 struct iovec *iov = iorw->fast_iov;
3120 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3121 if (unlikely(ret < 0))
3124 iorw->bytes_done = 0;
3125 iorw->free_iovec = iov;
3127 req->flags |= REQ_F_NEED_CLEANUP;
3131 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3135 ret = io_prep_rw(req, sqe);
3139 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3142 /* either don't need iovec imported or already have it */
3143 if (!req->async_data)
3145 return io_rw_prep_async(req, READ);
3149 * This is our waitqueue callback handler, registered through lock_page_async()
3150 * when we initially tried to do the IO with the iocb armed our waitqueue.
3151 * This gets called when the page is unlocked, and we generally expect that to
3152 * happen when the page IO is completed and the page is now uptodate. This will
3153 * queue a task_work based retry of the operation, attempting to copy the data
3154 * again. If the latter fails because the page was NOT uptodate, then we will
3155 * do a thread based blocking retry of the operation. That's the unexpected
3158 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3159 int sync, void *arg)
3161 struct wait_page_queue *wpq;
3162 struct io_kiocb *req = wait->private;
3163 struct wait_page_key *key = arg;
3166 wpq = container_of(wait, struct wait_page_queue, wait);
3168 if (!wake_page_match(wpq, key))
3171 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3172 list_del_init(&wait->entry);
3174 init_task_work(&req->task_work, io_req_task_submit);
3175 percpu_ref_get(&req->ctx->refs);
3177 /* submit ref gets dropped, acquire a new one */
3178 refcount_inc(&req->refs);
3179 ret = io_req_task_work_add(req, true);
3180 if (unlikely(ret)) {
3181 struct task_struct *tsk;
3183 /* queue just for cancelation */
3184 init_task_work(&req->task_work, io_req_task_cancel);
3185 tsk = io_wq_get_task(req->ctx->io_wq);
3186 task_work_add(tsk, &req->task_work, 0);
3187 wake_up_process(tsk);
3193 * This controls whether a given IO request should be armed for async page
3194 * based retry. If we return false here, the request is handed to the async
3195 * worker threads for retry. If we're doing buffered reads on a regular file,
3196 * we prepare a private wait_page_queue entry and retry the operation. This
3197 * will either succeed because the page is now uptodate and unlocked, or it
3198 * will register a callback when the page is unlocked at IO completion. Through
3199 * that callback, io_uring uses task_work to setup a retry of the operation.
3200 * That retry will attempt the buffered read again. The retry will generally
3201 * succeed, or in rare cases where it fails, we then fall back to using the
3202 * async worker threads for a blocking retry.
3204 static bool io_rw_should_retry(struct io_kiocb *req)
3206 struct io_async_rw *rw = req->async_data;
3207 struct wait_page_queue *wait = &rw->wpq;
3208 struct kiocb *kiocb = &req->rw.kiocb;
3210 /* never retry for NOWAIT, we just complete with -EAGAIN */
3211 if (req->flags & REQ_F_NOWAIT)
3214 /* Only for buffered IO */
3215 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3219 * just use poll if we can, and don't attempt if the fs doesn't
3220 * support callback based unlocks
3222 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3225 wait->wait.func = io_async_buf_func;
3226 wait->wait.private = req;
3227 wait->wait.flags = 0;
3228 INIT_LIST_HEAD(&wait->wait.entry);
3229 kiocb->ki_flags |= IOCB_WAITQ;
3230 kiocb->ki_flags &= ~IOCB_NOWAIT;
3231 kiocb->ki_waitq = wait;
3235 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3237 if (req->file->f_op->read_iter)
3238 return call_read_iter(req->file, &req->rw.kiocb, iter);
3239 else if (req->file->f_op->read)
3240 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3245 static int io_read(struct io_kiocb *req, bool force_nonblock,
3246 struct io_comp_state *cs)
3248 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3249 struct kiocb *kiocb = &req->rw.kiocb;
3250 struct iov_iter __iter, *iter = &__iter;
3251 struct io_async_rw *rw = req->async_data;
3252 ssize_t io_size, ret, ret2;
3259 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3262 iov_count = iov_iter_count(iter);
3264 req->result = io_size;
3267 /* Ensure we clear previously set non-block flag */
3268 if (!force_nonblock)
3269 kiocb->ki_flags &= ~IOCB_NOWAIT;
3271 kiocb->ki_flags |= IOCB_NOWAIT;
3274 /* If the file doesn't support async, just async punt */
3275 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3279 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3283 ret = io_iter_do_read(req, iter);
3287 } else if (ret == -EIOCBQUEUED) {
3290 } else if (ret == -EAGAIN) {
3291 /* IOPOLL retry should happen for io-wq threads */
3292 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3294 /* no retry on NONBLOCK marked file */
3295 if (req->file->f_flags & O_NONBLOCK)
3297 /* some cases will consume bytes even on error returns */
3298 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3301 } else if (ret < 0) {
3302 /* make sure -ERESTARTSYS -> -EINTR is done */
3306 /* read it all, or we did blocking attempt. no retry. */
3307 if (!iov_iter_count(iter) || !force_nonblock ||
3308 (req->file->f_flags & O_NONBLOCK))
3313 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3320 rw = req->async_data;
3321 /* it's copied and will be cleaned with ->io */
3323 /* now use our persistent iterator, if we aren't already */
3326 rw->bytes_done += ret;
3327 /* if we can retry, do so with the callbacks armed */
3328 if (!io_rw_should_retry(req)) {
3329 kiocb->ki_flags &= ~IOCB_WAITQ;
3334 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3335 * get -EIOCBQUEUED, then we'll get a notification when the desired
3336 * page gets unlocked. We can also get a partial read here, and if we
3337 * do, then just retry at the new offset.
3339 ret = io_iter_do_read(req, iter);
3340 if (ret == -EIOCBQUEUED) {
3343 } else if (ret > 0 && ret < io_size) {
3344 /* we got some bytes, but not all. retry. */
3348 kiocb_done(kiocb, ret, cs);
3351 /* it's reportedly faster than delegating the null check to kfree() */
3357 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3361 ret = io_prep_rw(req, sqe);
3365 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3368 /* either don't need iovec imported or already have it */
3369 if (!req->async_data)
3371 return io_rw_prep_async(req, WRITE);
3374 static int io_write(struct io_kiocb *req, bool force_nonblock,
3375 struct io_comp_state *cs)
3377 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3378 struct kiocb *kiocb = &req->rw.kiocb;
3379 struct iov_iter __iter, *iter = &__iter;
3380 struct io_async_rw *rw = req->async_data;
3382 ssize_t ret, ret2, io_size;
3387 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3390 iov_count = iov_iter_count(iter);
3392 req->result = io_size;
3394 /* Ensure we clear previously set non-block flag */
3395 if (!force_nonblock)
3396 kiocb->ki_flags &= ~IOCB_NOWAIT;
3398 kiocb->ki_flags |= IOCB_NOWAIT;
3400 /* If the file doesn't support async, just async punt */
3401 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3404 /* file path doesn't support NOWAIT for non-direct_IO */
3405 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3406 (req->flags & REQ_F_ISREG))
3409 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3414 * Open-code file_start_write here to grab freeze protection,
3415 * which will be released by another thread in
3416 * io_complete_rw(). Fool lockdep by telling it the lock got
3417 * released so that it doesn't complain about the held lock when
3418 * we return to userspace.
3420 if (req->flags & REQ_F_ISREG) {
3421 __sb_start_write(file_inode(req->file)->i_sb,
3422 SB_FREEZE_WRITE, true);
3423 __sb_writers_release(file_inode(req->file)->i_sb,
3426 kiocb->ki_flags |= IOCB_WRITE;
3428 if (req->file->f_op->write_iter)
3429 ret2 = call_write_iter(req->file, kiocb, iter);
3430 else if (req->file->f_op->write)
3431 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3436 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3437 * retry them without IOCB_NOWAIT.
3439 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3441 /* no retry on NONBLOCK marked file */
3442 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3444 if (!force_nonblock || ret2 != -EAGAIN) {
3445 /* IOPOLL retry should happen for io-wq threads */
3446 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3449 kiocb_done(kiocb, ret2, cs);
3452 /* some cases will consume bytes even on error returns */
3453 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3454 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3459 /* it's reportedly faster than delegating the null check to kfree() */
3465 static int __io_splice_prep(struct io_kiocb *req,
3466 const struct io_uring_sqe *sqe)
3468 struct io_splice* sp = &req->splice;
3469 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3471 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3475 sp->len = READ_ONCE(sqe->len);
3476 sp->flags = READ_ONCE(sqe->splice_flags);
3478 if (unlikely(sp->flags & ~valid_flags))
3481 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3482 (sp->flags & SPLICE_F_FD_IN_FIXED));
3485 req->flags |= REQ_F_NEED_CLEANUP;
3487 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3489 * Splice operation will be punted aync, and here need to
3490 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3492 io_req_init_async(req);
3493 req->work.flags |= IO_WQ_WORK_UNBOUND;
3499 static int io_tee_prep(struct io_kiocb *req,
3500 const struct io_uring_sqe *sqe)
3502 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3504 return __io_splice_prep(req, sqe);
3507 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3509 struct io_splice *sp = &req->splice;
3510 struct file *in = sp->file_in;
3511 struct file *out = sp->file_out;
3512 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3518 ret = do_tee(in, out, sp->len, flags);
3520 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3521 req->flags &= ~REQ_F_NEED_CLEANUP;
3524 req_set_fail_links(req);
3525 io_req_complete(req, ret);
3529 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3531 struct io_splice* sp = &req->splice;
3533 sp->off_in = READ_ONCE(sqe->splice_off_in);
3534 sp->off_out = READ_ONCE(sqe->off);
3535 return __io_splice_prep(req, sqe);
3538 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3540 struct io_splice *sp = &req->splice;
3541 struct file *in = sp->file_in;
3542 struct file *out = sp->file_out;
3543 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3544 loff_t *poff_in, *poff_out;
3550 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3551 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3554 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3556 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3557 req->flags &= ~REQ_F_NEED_CLEANUP;
3560 req_set_fail_links(req);
3561 io_req_complete(req, ret);
3566 * IORING_OP_NOP just posts a completion event, nothing else.
3568 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3570 struct io_ring_ctx *ctx = req->ctx;
3572 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3575 __io_req_complete(req, 0, 0, cs);
3579 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3581 struct io_ring_ctx *ctx = req->ctx;
3586 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3588 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3591 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3592 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3595 req->sync.off = READ_ONCE(sqe->off);
3596 req->sync.len = READ_ONCE(sqe->len);
3600 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3602 loff_t end = req->sync.off + req->sync.len;
3605 /* fsync always requires a blocking context */
3609 ret = vfs_fsync_range(req->file, req->sync.off,
3610 end > 0 ? end : LLONG_MAX,
3611 req->sync.flags & IORING_FSYNC_DATASYNC);
3613 req_set_fail_links(req);
3614 io_req_complete(req, ret);
3618 static int io_fallocate_prep(struct io_kiocb *req,
3619 const struct io_uring_sqe *sqe)
3621 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3623 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3626 req->sync.off = READ_ONCE(sqe->off);
3627 req->sync.len = READ_ONCE(sqe->addr);
3628 req->sync.mode = READ_ONCE(sqe->len);
3632 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3636 /* fallocate always requiring blocking context */
3639 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3642 req_set_fail_links(req);
3643 io_req_complete(req, ret);
3647 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3649 const char __user *fname;
3652 if (unlikely(sqe->ioprio || sqe->buf_index))
3654 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3657 /* open.how should be already initialised */
3658 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3659 req->open.how.flags |= O_LARGEFILE;
3661 req->open.dfd = READ_ONCE(sqe->fd);
3662 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3663 req->open.filename = getname(fname);
3664 if (IS_ERR(req->open.filename)) {
3665 ret = PTR_ERR(req->open.filename);
3666 req->open.filename = NULL;
3669 req->open.nofile = rlimit(RLIMIT_NOFILE);
3670 req->flags |= REQ_F_NEED_CLEANUP;
3674 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3678 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3680 mode = READ_ONCE(sqe->len);
3681 flags = READ_ONCE(sqe->open_flags);
3682 req->open.how = build_open_how(flags, mode);
3683 return __io_openat_prep(req, sqe);
3686 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3688 struct open_how __user *how;
3692 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3694 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3695 len = READ_ONCE(sqe->len);
3696 if (len < OPEN_HOW_SIZE_VER0)
3699 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3704 return __io_openat_prep(req, sqe);
3707 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3709 struct open_flags op;
3716 ret = build_open_flags(&req->open.how, &op);
3720 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3724 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3727 ret = PTR_ERR(file);
3729 fsnotify_open(file);
3730 fd_install(ret, file);
3733 putname(req->open.filename);
3734 req->flags &= ~REQ_F_NEED_CLEANUP;
3736 req_set_fail_links(req);
3737 io_req_complete(req, ret);
3741 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3743 return io_openat2(req, force_nonblock);
3746 static int io_remove_buffers_prep(struct io_kiocb *req,
3747 const struct io_uring_sqe *sqe)
3749 struct io_provide_buf *p = &req->pbuf;
3752 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3755 tmp = READ_ONCE(sqe->fd);
3756 if (!tmp || tmp > USHRT_MAX)
3759 memset(p, 0, sizeof(*p));
3761 p->bgid = READ_ONCE(sqe->buf_group);
3765 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3766 int bgid, unsigned nbufs)
3770 /* shouldn't happen */
3774 /* the head kbuf is the list itself */
3775 while (!list_empty(&buf->list)) {
3776 struct io_buffer *nxt;
3778 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3779 list_del(&nxt->list);
3786 idr_remove(&ctx->io_buffer_idr, bgid);
3791 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3792 struct io_comp_state *cs)
3794 struct io_provide_buf *p = &req->pbuf;
3795 struct io_ring_ctx *ctx = req->ctx;
3796 struct io_buffer *head;
3799 io_ring_submit_lock(ctx, !force_nonblock);
3801 lockdep_assert_held(&ctx->uring_lock);
3804 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3806 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3808 io_ring_submit_lock(ctx, !force_nonblock);
3810 req_set_fail_links(req);
3811 __io_req_complete(req, ret, 0, cs);
3815 static int io_provide_buffers_prep(struct io_kiocb *req,
3816 const struct io_uring_sqe *sqe)
3818 struct io_provide_buf *p = &req->pbuf;
3821 if (sqe->ioprio || sqe->rw_flags)
3824 tmp = READ_ONCE(sqe->fd);
3825 if (!tmp || tmp > USHRT_MAX)
3828 p->addr = READ_ONCE(sqe->addr);
3829 p->len = READ_ONCE(sqe->len);
3831 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3834 p->bgid = READ_ONCE(sqe->buf_group);
3835 tmp = READ_ONCE(sqe->off);
3836 if (tmp > USHRT_MAX)
3842 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3844 struct io_buffer *buf;
3845 u64 addr = pbuf->addr;
3846 int i, bid = pbuf->bid;
3848 for (i = 0; i < pbuf->nbufs; i++) {
3849 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3854 buf->len = pbuf->len;
3859 INIT_LIST_HEAD(&buf->list);
3862 list_add_tail(&buf->list, &(*head)->list);
3866 return i ? i : -ENOMEM;
3869 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3870 struct io_comp_state *cs)
3872 struct io_provide_buf *p = &req->pbuf;
3873 struct io_ring_ctx *ctx = req->ctx;
3874 struct io_buffer *head, *list;
3877 io_ring_submit_lock(ctx, !force_nonblock);
3879 lockdep_assert_held(&ctx->uring_lock);
3881 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3883 ret = io_add_buffers(p, &head);
3888 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3891 __io_remove_buffers(ctx, head, p->bgid, -1U);
3896 io_ring_submit_unlock(ctx, !force_nonblock);
3898 req_set_fail_links(req);
3899 __io_req_complete(req, ret, 0, cs);
3903 static int io_epoll_ctl_prep(struct io_kiocb *req,
3904 const struct io_uring_sqe *sqe)
3906 #if defined(CONFIG_EPOLL)
3907 if (sqe->ioprio || sqe->buf_index)
3909 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3912 req->epoll.epfd = READ_ONCE(sqe->fd);
3913 req->epoll.op = READ_ONCE(sqe->len);
3914 req->epoll.fd = READ_ONCE(sqe->off);
3916 if (ep_op_has_event(req->epoll.op)) {
3917 struct epoll_event __user *ev;
3919 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3920 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3930 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3931 struct io_comp_state *cs)
3933 #if defined(CONFIG_EPOLL)
3934 struct io_epoll *ie = &req->epoll;
3937 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3938 if (force_nonblock && ret == -EAGAIN)
3942 req_set_fail_links(req);
3943 __io_req_complete(req, ret, 0, cs);
3950 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3952 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3953 if (sqe->ioprio || sqe->buf_index || sqe->off)
3955 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3958 req->madvise.addr = READ_ONCE(sqe->addr);
3959 req->madvise.len = READ_ONCE(sqe->len);
3960 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3967 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3969 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3970 struct io_madvise *ma = &req->madvise;
3976 ret = do_madvise(ma->addr, ma->len, ma->advice);
3978 req_set_fail_links(req);
3979 io_req_complete(req, ret);
3986 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3988 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3990 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3993 req->fadvise.offset = READ_ONCE(sqe->off);
3994 req->fadvise.len = READ_ONCE(sqe->len);
3995 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3999 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4001 struct io_fadvise *fa = &req->fadvise;
4004 if (force_nonblock) {
4005 switch (fa->advice) {
4006 case POSIX_FADV_NORMAL:
4007 case POSIX_FADV_RANDOM:
4008 case POSIX_FADV_SEQUENTIAL:
4015 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4017 req_set_fail_links(req);
4018 io_req_complete(req, ret);
4022 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4024 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4026 if (sqe->ioprio || sqe->buf_index)
4028 if (req->flags & REQ_F_FIXED_FILE)
4031 req->statx.dfd = READ_ONCE(sqe->fd);
4032 req->statx.mask = READ_ONCE(sqe->len);
4033 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4034 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4035 req->statx.flags = READ_ONCE(sqe->statx_flags);
4040 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4042 struct io_statx *ctx = &req->statx;
4045 if (force_nonblock) {
4046 /* only need file table for an actual valid fd */
4047 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4048 req->flags |= REQ_F_NO_FILE_TABLE;
4052 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4056 req_set_fail_links(req);
4057 io_req_complete(req, ret);
4061 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4064 * If we queue this for async, it must not be cancellable. That would
4065 * leave the 'file' in an undeterminate state, and here need to modify
4066 * io_wq_work.flags, so initialize io_wq_work firstly.
4068 io_req_init_async(req);
4069 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4071 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4073 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4074 sqe->rw_flags || sqe->buf_index)
4076 if (req->flags & REQ_F_FIXED_FILE)
4079 req->close.fd = READ_ONCE(sqe->fd);
4080 if ((req->file && req->file->f_op == &io_uring_fops))
4083 req->close.put_file = NULL;
4087 static int io_close(struct io_kiocb *req, bool force_nonblock,
4088 struct io_comp_state *cs)
4090 struct io_close *close = &req->close;
4093 /* might be already done during nonblock submission */
4094 if (!close->put_file) {
4095 ret = __close_fd_get_file(close->fd, &close->put_file);
4097 return (ret == -ENOENT) ? -EBADF : ret;
4100 /* if the file has a flush method, be safe and punt to async */
4101 if (close->put_file->f_op->flush && force_nonblock) {
4102 /* was never set, but play safe */
4103 req->flags &= ~REQ_F_NOWAIT;
4104 /* avoid grabbing files - we don't need the files */
4105 req->flags |= REQ_F_NO_FILE_TABLE;
4109 /* No ->flush() or already async, safely close from here */
4110 ret = filp_close(close->put_file, req->work.files);
4112 req_set_fail_links(req);
4113 fput(close->put_file);
4114 close->put_file = NULL;
4115 __io_req_complete(req, ret, 0, cs);
4119 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4121 struct io_ring_ctx *ctx = req->ctx;
4126 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4128 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4131 req->sync.off = READ_ONCE(sqe->off);
4132 req->sync.len = READ_ONCE(sqe->len);
4133 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4137 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4141 /* sync_file_range always requires a blocking context */
4145 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4148 req_set_fail_links(req);
4149 io_req_complete(req, ret);
4153 #if defined(CONFIG_NET)
4154 static int io_setup_async_msg(struct io_kiocb *req,
4155 struct io_async_msghdr *kmsg)
4157 struct io_async_msghdr *async_msg = req->async_data;
4161 if (io_alloc_async_data(req)) {
4162 if (kmsg->iov != kmsg->fast_iov)
4166 async_msg = req->async_data;
4167 req->flags |= REQ_F_NEED_CLEANUP;
4168 memcpy(async_msg, kmsg, sizeof(*kmsg));
4172 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4173 struct io_async_msghdr *iomsg)
4175 iomsg->iov = iomsg->fast_iov;
4176 iomsg->msg.msg_name = &iomsg->addr;
4177 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4178 req->sr_msg.msg_flags, &iomsg->iov);
4181 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4183 struct io_async_msghdr *async_msg = req->async_data;
4184 struct io_sr_msg *sr = &req->sr_msg;
4187 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4190 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4191 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4192 sr->len = READ_ONCE(sqe->len);
4194 #ifdef CONFIG_COMPAT
4195 if (req->ctx->compat)
4196 sr->msg_flags |= MSG_CMSG_COMPAT;
4199 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4201 ret = io_sendmsg_copy_hdr(req, async_msg);
4203 req->flags |= REQ_F_NEED_CLEANUP;
4207 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4208 struct io_comp_state *cs)
4210 struct io_async_msghdr iomsg, *kmsg;
4211 struct socket *sock;
4215 sock = sock_from_file(req->file, &ret);
4216 if (unlikely(!sock))
4219 if (req->async_data) {
4220 kmsg = req->async_data;
4221 kmsg->msg.msg_name = &kmsg->addr;
4222 /* if iov is set, it's allocated already */
4224 kmsg->iov = kmsg->fast_iov;
4225 kmsg->msg.msg_iter.iov = kmsg->iov;
4227 ret = io_sendmsg_copy_hdr(req, &iomsg);
4233 flags = req->sr_msg.msg_flags;
4234 if (flags & MSG_DONTWAIT)
4235 req->flags |= REQ_F_NOWAIT;
4236 else if (force_nonblock)
4237 flags |= MSG_DONTWAIT;
4239 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4240 if (force_nonblock && ret == -EAGAIN)
4241 return io_setup_async_msg(req, kmsg);
4242 if (ret == -ERESTARTSYS)
4245 if (kmsg->iov != kmsg->fast_iov)
4247 req->flags &= ~REQ_F_NEED_CLEANUP;
4249 req_set_fail_links(req);
4250 __io_req_complete(req, ret, 0, cs);
4254 static int io_send(struct io_kiocb *req, bool force_nonblock,
4255 struct io_comp_state *cs)
4257 struct io_sr_msg *sr = &req->sr_msg;
4260 struct socket *sock;
4264 sock = sock_from_file(req->file, &ret);
4265 if (unlikely(!sock))
4268 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4272 msg.msg_name = NULL;
4273 msg.msg_control = NULL;
4274 msg.msg_controllen = 0;
4275 msg.msg_namelen = 0;
4277 flags = req->sr_msg.msg_flags;
4278 if (flags & MSG_DONTWAIT)
4279 req->flags |= REQ_F_NOWAIT;
4280 else if (force_nonblock)
4281 flags |= MSG_DONTWAIT;
4283 msg.msg_flags = flags;
4284 ret = sock_sendmsg(sock, &msg);
4285 if (force_nonblock && ret == -EAGAIN)
4287 if (ret == -ERESTARTSYS)
4291 req_set_fail_links(req);
4292 __io_req_complete(req, ret, 0, cs);
4296 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4297 struct io_async_msghdr *iomsg)
4299 struct io_sr_msg *sr = &req->sr_msg;
4300 struct iovec __user *uiov;
4304 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4305 &iomsg->uaddr, &uiov, &iov_len);
4309 if (req->flags & REQ_F_BUFFER_SELECT) {
4312 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4314 sr->len = iomsg->iov[0].iov_len;
4315 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4319 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4320 &iomsg->iov, &iomsg->msg.msg_iter,
4329 #ifdef CONFIG_COMPAT
4330 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4331 struct io_async_msghdr *iomsg)
4333 struct compat_msghdr __user *msg_compat;
4334 struct io_sr_msg *sr = &req->sr_msg;
4335 struct compat_iovec __user *uiov;
4340 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4341 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4346 uiov = compat_ptr(ptr);
4347 if (req->flags & REQ_F_BUFFER_SELECT) {
4348 compat_ssize_t clen;
4352 if (!access_ok(uiov, sizeof(*uiov)))
4354 if (__get_user(clen, &uiov->iov_len))
4358 sr->len = iomsg->iov[0].iov_len;
4361 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4362 UIO_FASTIOV, &iomsg->iov,
4363 &iomsg->msg.msg_iter, true);
4372 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4373 struct io_async_msghdr *iomsg)
4375 iomsg->msg.msg_name = &iomsg->addr;
4376 iomsg->iov = iomsg->fast_iov;
4378 #ifdef CONFIG_COMPAT
4379 if (req->ctx->compat)
4380 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4383 return __io_recvmsg_copy_hdr(req, iomsg);
4386 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4389 struct io_sr_msg *sr = &req->sr_msg;
4390 struct io_buffer *kbuf;
4392 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4397 req->flags |= REQ_F_BUFFER_SELECTED;
4401 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4403 return io_put_kbuf(req, req->sr_msg.kbuf);
4406 static int io_recvmsg_prep(struct io_kiocb *req,
4407 const struct io_uring_sqe *sqe)
4409 struct io_async_msghdr *async_msg = req->async_data;
4410 struct io_sr_msg *sr = &req->sr_msg;
4413 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4416 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4417 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4418 sr->len = READ_ONCE(sqe->len);
4419 sr->bgid = READ_ONCE(sqe->buf_group);
4421 #ifdef CONFIG_COMPAT
4422 if (req->ctx->compat)
4423 sr->msg_flags |= MSG_CMSG_COMPAT;
4426 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4428 ret = io_recvmsg_copy_hdr(req, async_msg);
4430 req->flags |= REQ_F_NEED_CLEANUP;
4434 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4435 struct io_comp_state *cs)
4437 struct io_async_msghdr iomsg, *kmsg;
4438 struct socket *sock;
4439 struct io_buffer *kbuf;
4441 int ret, cflags = 0;
4443 sock = sock_from_file(req->file, &ret);
4444 if (unlikely(!sock))
4447 if (req->async_data) {
4448 kmsg = req->async_data;
4449 kmsg->msg.msg_name = &kmsg->addr;
4450 /* if iov is set, it's allocated already */
4452 kmsg->iov = kmsg->fast_iov;
4453 kmsg->msg.msg_iter.iov = kmsg->iov;
4455 ret = io_recvmsg_copy_hdr(req, &iomsg);
4461 if (req->flags & REQ_F_BUFFER_SELECT) {
4462 kbuf = io_recv_buffer_select(req, !force_nonblock);
4464 return PTR_ERR(kbuf);
4465 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4466 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4467 1, req->sr_msg.len);
4470 flags = req->sr_msg.msg_flags;
4471 if (flags & MSG_DONTWAIT)
4472 req->flags |= REQ_F_NOWAIT;
4473 else if (force_nonblock)
4474 flags |= MSG_DONTWAIT;
4476 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4477 kmsg->uaddr, flags);
4478 if (force_nonblock && ret == -EAGAIN)
4479 return io_setup_async_msg(req, kmsg);
4480 if (ret == -ERESTARTSYS)
4483 if (req->flags & REQ_F_BUFFER_SELECTED)
4484 cflags = io_put_recv_kbuf(req);
4485 if (kmsg->iov != kmsg->fast_iov)
4487 req->flags &= ~REQ_F_NEED_CLEANUP;
4489 req_set_fail_links(req);
4490 __io_req_complete(req, ret, cflags, cs);
4494 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4495 struct io_comp_state *cs)
4497 struct io_buffer *kbuf;
4498 struct io_sr_msg *sr = &req->sr_msg;
4500 void __user *buf = sr->buf;
4501 struct socket *sock;
4504 int ret, cflags = 0;
4506 sock = sock_from_file(req->file, &ret);
4507 if (unlikely(!sock))
4510 if (req->flags & REQ_F_BUFFER_SELECT) {
4511 kbuf = io_recv_buffer_select(req, !force_nonblock);
4513 return PTR_ERR(kbuf);
4514 buf = u64_to_user_ptr(kbuf->addr);
4517 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4521 msg.msg_name = NULL;
4522 msg.msg_control = NULL;
4523 msg.msg_controllen = 0;
4524 msg.msg_namelen = 0;
4525 msg.msg_iocb = NULL;
4528 flags = req->sr_msg.msg_flags;
4529 if (flags & MSG_DONTWAIT)
4530 req->flags |= REQ_F_NOWAIT;
4531 else if (force_nonblock)
4532 flags |= MSG_DONTWAIT;
4534 ret = sock_recvmsg(sock, &msg, flags);
4535 if (force_nonblock && ret == -EAGAIN)
4537 if (ret == -ERESTARTSYS)
4540 if (req->flags & REQ_F_BUFFER_SELECTED)
4541 cflags = io_put_recv_kbuf(req);
4543 req_set_fail_links(req);
4544 __io_req_complete(req, ret, cflags, cs);
4548 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4550 struct io_accept *accept = &req->accept;
4552 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4554 if (sqe->ioprio || sqe->len || sqe->buf_index)
4557 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4558 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4559 accept->flags = READ_ONCE(sqe->accept_flags);
4560 accept->nofile = rlimit(RLIMIT_NOFILE);
4564 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4565 struct io_comp_state *cs)
4567 struct io_accept *accept = &req->accept;
4568 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4571 if (req->file->f_flags & O_NONBLOCK)
4572 req->flags |= REQ_F_NOWAIT;
4574 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4575 accept->addr_len, accept->flags,
4577 if (ret == -EAGAIN && force_nonblock)
4580 if (ret == -ERESTARTSYS)
4582 req_set_fail_links(req);
4584 __io_req_complete(req, ret, 0, cs);
4588 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4590 struct io_connect *conn = &req->connect;
4591 struct io_async_connect *io = req->async_data;
4593 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4595 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4598 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4599 conn->addr_len = READ_ONCE(sqe->addr2);
4604 return move_addr_to_kernel(conn->addr, conn->addr_len,
4608 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4609 struct io_comp_state *cs)
4611 struct io_async_connect __io, *io;
4612 unsigned file_flags;
4615 if (req->async_data) {
4616 io = req->async_data;
4618 ret = move_addr_to_kernel(req->connect.addr,
4619 req->connect.addr_len,
4626 file_flags = force_nonblock ? O_NONBLOCK : 0;
4628 ret = __sys_connect_file(req->file, &io->address,
4629 req->connect.addr_len, file_flags);
4630 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4631 if (req->async_data)
4633 if (io_alloc_async_data(req)) {
4637 io = req->async_data;
4638 memcpy(req->async_data, &__io, sizeof(__io));
4641 if (ret == -ERESTARTSYS)
4645 req_set_fail_links(req);
4646 __io_req_complete(req, ret, 0, cs);
4649 #else /* !CONFIG_NET */
4650 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4655 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4656 struct io_comp_state *cs)
4661 static int io_send(struct io_kiocb *req, bool force_nonblock,
4662 struct io_comp_state *cs)
4667 static int io_recvmsg_prep(struct io_kiocb *req,
4668 const struct io_uring_sqe *sqe)
4673 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4674 struct io_comp_state *cs)
4679 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4680 struct io_comp_state *cs)
4685 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4690 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4691 struct io_comp_state *cs)
4696 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4701 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4702 struct io_comp_state *cs)
4706 #endif /* CONFIG_NET */
4708 struct io_poll_table {
4709 struct poll_table_struct pt;
4710 struct io_kiocb *req;
4714 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4715 __poll_t mask, task_work_func_t func)
4720 /* for instances that support it check for an event match first: */
4721 if (mask && !(mask & poll->events))
4724 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4726 list_del_init(&poll->wait.entry);
4729 init_task_work(&req->task_work, func);
4730 percpu_ref_get(&req->ctx->refs);
4733 * If we using the signalfd wait_queue_head for this wakeup, then
4734 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4735 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4736 * either, as the normal wakeup will suffice.
4738 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4741 * If this fails, then the task is exiting. When a task exits, the
4742 * work gets canceled, so just cancel this request as well instead
4743 * of executing it. We can't safely execute it anyway, as we may not
4744 * have the needed state needed for it anyway.
4746 ret = io_req_task_work_add(req, twa_signal_ok);
4747 if (unlikely(ret)) {
4748 struct task_struct *tsk;
4750 WRITE_ONCE(poll->canceled, true);
4751 tsk = io_wq_get_task(req->ctx->io_wq);
4752 task_work_add(tsk, &req->task_work, 0);
4753 wake_up_process(tsk);
4758 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4759 __acquires(&req->ctx->completion_lock)
4761 struct io_ring_ctx *ctx = req->ctx;
4763 if (!req->result && !READ_ONCE(poll->canceled)) {
4764 struct poll_table_struct pt = { ._key = poll->events };
4766 req->result = vfs_poll(req->file, &pt) & poll->events;
4769 spin_lock_irq(&ctx->completion_lock);
4770 if (!req->result && !READ_ONCE(poll->canceled)) {
4771 add_wait_queue(poll->head, &poll->wait);
4778 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4780 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4781 if (req->opcode == IORING_OP_POLL_ADD)
4782 return req->async_data;
4783 return req->apoll->double_poll;
4786 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4788 if (req->opcode == IORING_OP_POLL_ADD)
4790 return &req->apoll->poll;
4793 static void io_poll_remove_double(struct io_kiocb *req)
4795 struct io_poll_iocb *poll = io_poll_get_double(req);
4797 lockdep_assert_held(&req->ctx->completion_lock);
4799 if (poll && poll->head) {
4800 struct wait_queue_head *head = poll->head;
4802 spin_lock(&head->lock);
4803 list_del_init(&poll->wait.entry);
4804 if (poll->wait.private)
4805 refcount_dec(&req->refs);
4807 spin_unlock(&head->lock);
4811 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4813 struct io_ring_ctx *ctx = req->ctx;
4815 io_poll_remove_double(req);
4816 req->poll.done = true;
4817 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4818 io_commit_cqring(ctx);
4821 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4823 struct io_ring_ctx *ctx = req->ctx;
4825 if (io_poll_rewait(req, &req->poll)) {
4826 spin_unlock_irq(&ctx->completion_lock);
4830 hash_del(&req->hash_node);
4831 io_poll_complete(req, req->result, 0);
4832 spin_unlock_irq(&ctx->completion_lock);
4834 *nxt = io_put_req_find_next(req);
4835 io_cqring_ev_posted(ctx);
4838 static void io_poll_task_func(struct callback_head *cb)
4840 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4841 struct io_ring_ctx *ctx = req->ctx;
4842 struct io_kiocb *nxt = NULL;
4844 io_poll_task_handler(req, &nxt);
4846 __io_req_task_submit(nxt);
4847 percpu_ref_put(&ctx->refs);
4850 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4851 int sync, void *key)
4853 struct io_kiocb *req = wait->private;
4854 struct io_poll_iocb *poll = io_poll_get_single(req);
4855 __poll_t mask = key_to_poll(key);
4857 /* for instances that support it check for an event match first: */
4858 if (mask && !(mask & poll->events))
4861 list_del_init(&wait->entry);
4863 if (poll && poll->head) {
4866 spin_lock(&poll->head->lock);
4867 done = list_empty(&poll->wait.entry);
4869 list_del_init(&poll->wait.entry);
4870 /* make sure double remove sees this as being gone */
4871 wait->private = NULL;
4872 spin_unlock(&poll->head->lock);
4874 __io_async_wake(req, poll, mask, io_poll_task_func);
4876 refcount_dec(&req->refs);
4880 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4881 wait_queue_func_t wake_func)
4885 poll->canceled = false;
4886 poll->events = events;
4887 INIT_LIST_HEAD(&poll->wait.entry);
4888 init_waitqueue_func_entry(&poll->wait, wake_func);
4891 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4892 struct wait_queue_head *head,
4893 struct io_poll_iocb **poll_ptr)
4895 struct io_kiocb *req = pt->req;
4898 * If poll->head is already set, it's because the file being polled
4899 * uses multiple waitqueues for poll handling (eg one for read, one
4900 * for write). Setup a separate io_poll_iocb if this happens.
4902 if (unlikely(poll->head)) {
4903 /* already have a 2nd entry, fail a third attempt */
4905 pt->error = -EINVAL;
4908 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4910 pt->error = -ENOMEM;
4913 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4914 refcount_inc(&req->refs);
4915 poll->wait.private = req;
4922 if (poll->events & EPOLLEXCLUSIVE)
4923 add_wait_queue_exclusive(head, &poll->wait);
4925 add_wait_queue(head, &poll->wait);
4928 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4929 struct poll_table_struct *p)
4931 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4932 struct async_poll *apoll = pt->req->apoll;
4934 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4937 static void io_async_task_func(struct callback_head *cb)
4939 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4940 struct async_poll *apoll = req->apoll;
4941 struct io_ring_ctx *ctx = req->ctx;
4943 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4945 if (io_poll_rewait(req, &apoll->poll)) {
4946 spin_unlock_irq(&ctx->completion_lock);
4947 percpu_ref_put(&ctx->refs);
4951 /* If req is still hashed, it cannot have been canceled. Don't check. */
4952 if (hash_hashed(&req->hash_node))
4953 hash_del(&req->hash_node);
4955 io_poll_remove_double(req);
4956 spin_unlock_irq(&ctx->completion_lock);
4958 if (!READ_ONCE(apoll->poll.canceled))
4959 __io_req_task_submit(req);
4961 __io_req_task_cancel(req, -ECANCELED);
4963 percpu_ref_put(&ctx->refs);
4964 kfree(apoll->double_poll);
4968 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4971 struct io_kiocb *req = wait->private;
4972 struct io_poll_iocb *poll = &req->apoll->poll;
4974 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4977 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4980 static void io_poll_req_insert(struct io_kiocb *req)
4982 struct io_ring_ctx *ctx = req->ctx;
4983 struct hlist_head *list;
4985 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4986 hlist_add_head(&req->hash_node, list);
4989 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4990 struct io_poll_iocb *poll,
4991 struct io_poll_table *ipt, __poll_t mask,
4992 wait_queue_func_t wake_func)
4993 __acquires(&ctx->completion_lock)
4995 struct io_ring_ctx *ctx = req->ctx;
4996 bool cancel = false;
4998 io_init_poll_iocb(poll, mask, wake_func);
4999 poll->file = req->file;
5000 poll->wait.private = req;
5002 ipt->pt._key = mask;
5004 ipt->error = -EINVAL;
5006 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5008 spin_lock_irq(&ctx->completion_lock);
5009 if (likely(poll->head)) {
5010 spin_lock(&poll->head->lock);
5011 if (unlikely(list_empty(&poll->wait.entry))) {
5017 if (mask || ipt->error)
5018 list_del_init(&poll->wait.entry);
5020 WRITE_ONCE(poll->canceled, true);
5021 else if (!poll->done) /* actually waiting for an event */
5022 io_poll_req_insert(req);
5023 spin_unlock(&poll->head->lock);
5029 static bool io_arm_poll_handler(struct io_kiocb *req)
5031 const struct io_op_def *def = &io_op_defs[req->opcode];
5032 struct io_ring_ctx *ctx = req->ctx;
5033 struct async_poll *apoll;
5034 struct io_poll_table ipt;
5038 if (!req->file || !file_can_poll(req->file))
5040 if (req->flags & REQ_F_POLLED)
5044 else if (def->pollout)
5048 /* if we can't nonblock try, then no point in arming a poll handler */
5049 if (!io_file_supports_async(req->file, rw))
5052 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5053 if (unlikely(!apoll))
5055 apoll->double_poll = NULL;
5057 req->flags |= REQ_F_POLLED;
5059 INIT_HLIST_NODE(&req->hash_node);
5063 mask |= POLLIN | POLLRDNORM;
5065 mask |= POLLOUT | POLLWRNORM;
5067 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5068 if ((req->opcode == IORING_OP_RECVMSG) &&
5069 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5072 mask |= POLLERR | POLLPRI;
5074 ipt.pt._qproc = io_async_queue_proc;
5076 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5078 if (ret || ipt.error) {
5079 io_poll_remove_double(req);
5080 spin_unlock_irq(&ctx->completion_lock);
5081 kfree(apoll->double_poll);
5085 spin_unlock_irq(&ctx->completion_lock);
5086 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5087 apoll->poll.events);
5091 static bool __io_poll_remove_one(struct io_kiocb *req,
5092 struct io_poll_iocb *poll)
5094 bool do_complete = false;
5096 spin_lock(&poll->head->lock);
5097 WRITE_ONCE(poll->canceled, true);
5098 if (!list_empty(&poll->wait.entry)) {
5099 list_del_init(&poll->wait.entry);
5102 spin_unlock(&poll->head->lock);
5103 hash_del(&req->hash_node);
5107 static bool io_poll_remove_one(struct io_kiocb *req)
5111 io_poll_remove_double(req);
5113 if (req->opcode == IORING_OP_POLL_ADD) {
5114 do_complete = __io_poll_remove_one(req, &req->poll);
5116 struct async_poll *apoll = req->apoll;
5118 /* non-poll requests have submit ref still */
5119 do_complete = __io_poll_remove_one(req, &apoll->poll);
5122 kfree(apoll->double_poll);
5128 io_cqring_fill_event(req, -ECANCELED);
5129 io_commit_cqring(req->ctx);
5130 req->flags |= REQ_F_COMP_LOCKED;
5131 req_set_fail_links(req);
5139 * Returns true if we found and killed one or more poll requests
5141 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5143 struct hlist_node *tmp;
5144 struct io_kiocb *req;
5147 spin_lock_irq(&ctx->completion_lock);
5148 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5149 struct hlist_head *list;
5151 list = &ctx->cancel_hash[i];
5152 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5153 if (io_task_match(req, tsk))
5154 posted += io_poll_remove_one(req);
5157 spin_unlock_irq(&ctx->completion_lock);
5160 io_cqring_ev_posted(ctx);
5165 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5167 struct hlist_head *list;
5168 struct io_kiocb *req;
5170 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5171 hlist_for_each_entry(req, list, hash_node) {
5172 if (sqe_addr != req->user_data)
5174 if (io_poll_remove_one(req))
5182 static int io_poll_remove_prep(struct io_kiocb *req,
5183 const struct io_uring_sqe *sqe)
5185 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5187 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5191 req->poll.addr = READ_ONCE(sqe->addr);
5196 * Find a running poll command that matches one specified in sqe->addr,
5197 * and remove it if found.
5199 static int io_poll_remove(struct io_kiocb *req)
5201 struct io_ring_ctx *ctx = req->ctx;
5205 addr = req->poll.addr;
5206 spin_lock_irq(&ctx->completion_lock);
5207 ret = io_poll_cancel(ctx, addr);
5208 spin_unlock_irq(&ctx->completion_lock);
5211 req_set_fail_links(req);
5212 io_req_complete(req, ret);
5216 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5219 struct io_kiocb *req = wait->private;
5220 struct io_poll_iocb *poll = &req->poll;
5222 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5225 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5226 struct poll_table_struct *p)
5228 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5230 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5233 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5235 struct io_poll_iocb *poll = &req->poll;
5238 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5240 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5245 events = READ_ONCE(sqe->poll32_events);
5247 events = swahw32(events);
5249 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5250 (events & EPOLLEXCLUSIVE);
5254 static int io_poll_add(struct io_kiocb *req)
5256 struct io_poll_iocb *poll = &req->poll;
5257 struct io_ring_ctx *ctx = req->ctx;
5258 struct io_poll_table ipt;
5261 INIT_HLIST_NODE(&req->hash_node);
5262 ipt.pt._qproc = io_poll_queue_proc;
5264 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5267 if (mask) { /* no async, we'd stolen it */
5269 io_poll_complete(req, mask, 0);
5271 spin_unlock_irq(&ctx->completion_lock);
5274 io_cqring_ev_posted(ctx);
5280 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5282 struct io_timeout_data *data = container_of(timer,
5283 struct io_timeout_data, timer);
5284 struct io_kiocb *req = data->req;
5285 struct io_ring_ctx *ctx = req->ctx;
5286 unsigned long flags;
5288 spin_lock_irqsave(&ctx->completion_lock, flags);
5289 list_del_init(&req->timeout.list);
5290 atomic_set(&req->ctx->cq_timeouts,
5291 atomic_read(&req->ctx->cq_timeouts) + 1);
5293 io_cqring_fill_event(req, -ETIME);
5294 io_commit_cqring(ctx);
5295 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5297 io_cqring_ev_posted(ctx);
5298 req_set_fail_links(req);
5300 return HRTIMER_NORESTART;
5303 static int __io_timeout_cancel(struct io_kiocb *req)
5305 struct io_timeout_data *io = req->async_data;
5308 ret = hrtimer_try_to_cancel(&io->timer);
5311 list_del_init(&req->timeout.list);
5313 req_set_fail_links(req);
5314 req->flags |= REQ_F_COMP_LOCKED;
5315 io_cqring_fill_event(req, -ECANCELED);
5320 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5322 struct io_kiocb *req;
5325 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5326 if (user_data == req->user_data) {
5335 return __io_timeout_cancel(req);
5338 static int io_timeout_remove_prep(struct io_kiocb *req,
5339 const struct io_uring_sqe *sqe)
5341 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5343 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5345 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5348 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5353 * Remove or update an existing timeout command
5355 static int io_timeout_remove(struct io_kiocb *req)
5357 struct io_ring_ctx *ctx = req->ctx;
5360 spin_lock_irq(&ctx->completion_lock);
5361 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5363 io_cqring_fill_event(req, ret);
5364 io_commit_cqring(ctx);
5365 spin_unlock_irq(&ctx->completion_lock);
5366 io_cqring_ev_posted(ctx);
5368 req_set_fail_links(req);
5373 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5374 bool is_timeout_link)
5376 struct io_timeout_data *data;
5378 u32 off = READ_ONCE(sqe->off);
5380 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5382 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5384 if (off && is_timeout_link)
5386 flags = READ_ONCE(sqe->timeout_flags);
5387 if (flags & ~IORING_TIMEOUT_ABS)
5390 req->timeout.off = off;
5392 if (!req->async_data && io_alloc_async_data(req))
5395 data = req->async_data;
5398 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5401 if (flags & IORING_TIMEOUT_ABS)
5402 data->mode = HRTIMER_MODE_ABS;
5404 data->mode = HRTIMER_MODE_REL;
5406 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5410 static int io_timeout(struct io_kiocb *req)
5412 struct io_ring_ctx *ctx = req->ctx;
5413 struct io_timeout_data *data = req->async_data;
5414 struct list_head *entry;
5415 u32 tail, off = req->timeout.off;
5417 spin_lock_irq(&ctx->completion_lock);
5420 * sqe->off holds how many events that need to occur for this
5421 * timeout event to be satisfied. If it isn't set, then this is
5422 * a pure timeout request, sequence isn't used.
5424 if (io_is_timeout_noseq(req)) {
5425 entry = ctx->timeout_list.prev;
5429 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5430 req->timeout.target_seq = tail + off;
5433 * Insertion sort, ensuring the first entry in the list is always
5434 * the one we need first.
5436 list_for_each_prev(entry, &ctx->timeout_list) {
5437 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5440 if (io_is_timeout_noseq(nxt))
5442 /* nxt.seq is behind @tail, otherwise would've been completed */
5443 if (off >= nxt->timeout.target_seq - tail)
5447 list_add(&req->timeout.list, entry);
5448 data->timer.function = io_timeout_fn;
5449 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5450 spin_unlock_irq(&ctx->completion_lock);
5454 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5456 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5458 return req->user_data == (unsigned long) data;
5461 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5463 enum io_wq_cancel cancel_ret;
5466 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5467 switch (cancel_ret) {
5468 case IO_WQ_CANCEL_OK:
5471 case IO_WQ_CANCEL_RUNNING:
5474 case IO_WQ_CANCEL_NOTFOUND:
5482 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5483 struct io_kiocb *req, __u64 sqe_addr,
5486 unsigned long flags;
5489 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5490 if (ret != -ENOENT) {
5491 spin_lock_irqsave(&ctx->completion_lock, flags);
5495 spin_lock_irqsave(&ctx->completion_lock, flags);
5496 ret = io_timeout_cancel(ctx, sqe_addr);
5499 ret = io_poll_cancel(ctx, sqe_addr);
5503 io_cqring_fill_event(req, ret);
5504 io_commit_cqring(ctx);
5505 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5506 io_cqring_ev_posted(ctx);
5509 req_set_fail_links(req);
5513 static int io_async_cancel_prep(struct io_kiocb *req,
5514 const struct io_uring_sqe *sqe)
5516 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5518 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5520 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5523 req->cancel.addr = READ_ONCE(sqe->addr);
5527 static int io_async_cancel(struct io_kiocb *req)
5529 struct io_ring_ctx *ctx = req->ctx;
5531 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5535 static int io_files_update_prep(struct io_kiocb *req,
5536 const struct io_uring_sqe *sqe)
5538 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5540 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5542 if (sqe->ioprio || sqe->rw_flags)
5545 req->files_update.offset = READ_ONCE(sqe->off);
5546 req->files_update.nr_args = READ_ONCE(sqe->len);
5547 if (!req->files_update.nr_args)
5549 req->files_update.arg = READ_ONCE(sqe->addr);
5553 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5554 struct io_comp_state *cs)
5556 struct io_ring_ctx *ctx = req->ctx;
5557 struct io_uring_files_update up;
5563 up.offset = req->files_update.offset;
5564 up.fds = req->files_update.arg;
5566 mutex_lock(&ctx->uring_lock);
5567 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5568 mutex_unlock(&ctx->uring_lock);
5571 req_set_fail_links(req);
5572 __io_req_complete(req, ret, 0, cs);
5576 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5578 switch (req->opcode) {
5581 case IORING_OP_READV:
5582 case IORING_OP_READ_FIXED:
5583 case IORING_OP_READ:
5584 return io_read_prep(req, sqe);
5585 case IORING_OP_WRITEV:
5586 case IORING_OP_WRITE_FIXED:
5587 case IORING_OP_WRITE:
5588 return io_write_prep(req, sqe);
5589 case IORING_OP_POLL_ADD:
5590 return io_poll_add_prep(req, sqe);
5591 case IORING_OP_POLL_REMOVE:
5592 return io_poll_remove_prep(req, sqe);
5593 case IORING_OP_FSYNC:
5594 return io_prep_fsync(req, sqe);
5595 case IORING_OP_SYNC_FILE_RANGE:
5596 return io_prep_sfr(req, sqe);
5597 case IORING_OP_SENDMSG:
5598 case IORING_OP_SEND:
5599 return io_sendmsg_prep(req, sqe);
5600 case IORING_OP_RECVMSG:
5601 case IORING_OP_RECV:
5602 return io_recvmsg_prep(req, sqe);
5603 case IORING_OP_CONNECT:
5604 return io_connect_prep(req, sqe);
5605 case IORING_OP_TIMEOUT:
5606 return io_timeout_prep(req, sqe, false);
5607 case IORING_OP_TIMEOUT_REMOVE:
5608 return io_timeout_remove_prep(req, sqe);
5609 case IORING_OP_ASYNC_CANCEL:
5610 return io_async_cancel_prep(req, sqe);
5611 case IORING_OP_LINK_TIMEOUT:
5612 return io_timeout_prep(req, sqe, true);
5613 case IORING_OP_ACCEPT:
5614 return io_accept_prep(req, sqe);
5615 case IORING_OP_FALLOCATE:
5616 return io_fallocate_prep(req, sqe);
5617 case IORING_OP_OPENAT:
5618 return io_openat_prep(req, sqe);
5619 case IORING_OP_CLOSE:
5620 return io_close_prep(req, sqe);
5621 case IORING_OP_FILES_UPDATE:
5622 return io_files_update_prep(req, sqe);
5623 case IORING_OP_STATX:
5624 return io_statx_prep(req, sqe);
5625 case IORING_OP_FADVISE:
5626 return io_fadvise_prep(req, sqe);
5627 case IORING_OP_MADVISE:
5628 return io_madvise_prep(req, sqe);
5629 case IORING_OP_OPENAT2:
5630 return io_openat2_prep(req, sqe);
5631 case IORING_OP_EPOLL_CTL:
5632 return io_epoll_ctl_prep(req, sqe);
5633 case IORING_OP_SPLICE:
5634 return io_splice_prep(req, sqe);
5635 case IORING_OP_PROVIDE_BUFFERS:
5636 return io_provide_buffers_prep(req, sqe);
5637 case IORING_OP_REMOVE_BUFFERS:
5638 return io_remove_buffers_prep(req, sqe);
5640 return io_tee_prep(req, sqe);
5643 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5648 static int io_req_defer_prep(struct io_kiocb *req,
5649 const struct io_uring_sqe *sqe)
5653 if (io_alloc_async_data(req))
5655 return io_req_prep(req, sqe);
5658 static u32 io_get_sequence(struct io_kiocb *req)
5660 struct io_kiocb *pos;
5661 struct io_ring_ctx *ctx = req->ctx;
5662 u32 total_submitted, nr_reqs = 1;
5664 if (req->flags & REQ_F_LINK_HEAD)
5665 list_for_each_entry(pos, &req->link_list, link_list)
5668 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5669 return total_submitted - nr_reqs;
5672 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5674 struct io_ring_ctx *ctx = req->ctx;
5675 struct io_defer_entry *de;
5679 /* Still need defer if there is pending req in defer list. */
5680 if (likely(list_empty_careful(&ctx->defer_list) &&
5681 !(req->flags & REQ_F_IO_DRAIN)))
5684 seq = io_get_sequence(req);
5685 /* Still a chance to pass the sequence check */
5686 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5689 if (!req->async_data) {
5690 ret = io_req_defer_prep(req, sqe);
5694 io_prep_async_link(req);
5695 de = kmalloc(sizeof(*de), GFP_KERNEL);
5699 spin_lock_irq(&ctx->completion_lock);
5700 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5701 spin_unlock_irq(&ctx->completion_lock);
5703 io_queue_async_work(req);
5704 return -EIOCBQUEUED;
5707 trace_io_uring_defer(ctx, req, req->user_data);
5710 list_add_tail(&de->list, &ctx->defer_list);
5711 spin_unlock_irq(&ctx->completion_lock);
5712 return -EIOCBQUEUED;
5715 static void io_req_drop_files(struct io_kiocb *req)
5717 struct io_ring_ctx *ctx = req->ctx;
5718 unsigned long flags;
5720 spin_lock_irqsave(&ctx->inflight_lock, flags);
5721 list_del(&req->inflight_entry);
5722 if (waitqueue_active(&ctx->inflight_wait))
5723 wake_up(&ctx->inflight_wait);
5724 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5725 req->flags &= ~REQ_F_INFLIGHT;
5726 put_files_struct(req->work.files);
5727 put_nsproxy(req->work.nsproxy);
5728 req->work.files = NULL;
5731 static void __io_clean_op(struct io_kiocb *req)
5733 if (req->flags & REQ_F_BUFFER_SELECTED) {
5734 switch (req->opcode) {
5735 case IORING_OP_READV:
5736 case IORING_OP_READ_FIXED:
5737 case IORING_OP_READ:
5738 kfree((void *)(unsigned long)req->rw.addr);
5740 case IORING_OP_RECVMSG:
5741 case IORING_OP_RECV:
5742 kfree(req->sr_msg.kbuf);
5745 req->flags &= ~REQ_F_BUFFER_SELECTED;
5748 if (req->flags & REQ_F_NEED_CLEANUP) {
5749 switch (req->opcode) {
5750 case IORING_OP_READV:
5751 case IORING_OP_READ_FIXED:
5752 case IORING_OP_READ:
5753 case IORING_OP_WRITEV:
5754 case IORING_OP_WRITE_FIXED:
5755 case IORING_OP_WRITE: {
5756 struct io_async_rw *io = req->async_data;
5758 kfree(io->free_iovec);
5761 case IORING_OP_RECVMSG:
5762 case IORING_OP_SENDMSG: {
5763 struct io_async_msghdr *io = req->async_data;
5764 if (io->iov != io->fast_iov)
5768 case IORING_OP_SPLICE:
5770 io_put_file(req, req->splice.file_in,
5771 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5773 case IORING_OP_OPENAT:
5774 case IORING_OP_OPENAT2:
5775 if (req->open.filename)
5776 putname(req->open.filename);
5779 req->flags &= ~REQ_F_NEED_CLEANUP;
5782 if (req->flags & REQ_F_INFLIGHT)
5783 io_req_drop_files(req);
5786 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5787 struct io_comp_state *cs)
5789 struct io_ring_ctx *ctx = req->ctx;
5792 switch (req->opcode) {
5794 ret = io_nop(req, cs);
5796 case IORING_OP_READV:
5797 case IORING_OP_READ_FIXED:
5798 case IORING_OP_READ:
5799 ret = io_read(req, force_nonblock, cs);
5801 case IORING_OP_WRITEV:
5802 case IORING_OP_WRITE_FIXED:
5803 case IORING_OP_WRITE:
5804 ret = io_write(req, force_nonblock, cs);
5806 case IORING_OP_FSYNC:
5807 ret = io_fsync(req, force_nonblock);
5809 case IORING_OP_POLL_ADD:
5810 ret = io_poll_add(req);
5812 case IORING_OP_POLL_REMOVE:
5813 ret = io_poll_remove(req);
5815 case IORING_OP_SYNC_FILE_RANGE:
5816 ret = io_sync_file_range(req, force_nonblock);
5818 case IORING_OP_SENDMSG:
5819 ret = io_sendmsg(req, force_nonblock, cs);
5821 case IORING_OP_SEND:
5822 ret = io_send(req, force_nonblock, cs);
5824 case IORING_OP_RECVMSG:
5825 ret = io_recvmsg(req, force_nonblock, cs);
5827 case IORING_OP_RECV:
5828 ret = io_recv(req, force_nonblock, cs);
5830 case IORING_OP_TIMEOUT:
5831 ret = io_timeout(req);
5833 case IORING_OP_TIMEOUT_REMOVE:
5834 ret = io_timeout_remove(req);
5836 case IORING_OP_ACCEPT:
5837 ret = io_accept(req, force_nonblock, cs);
5839 case IORING_OP_CONNECT:
5840 ret = io_connect(req, force_nonblock, cs);
5842 case IORING_OP_ASYNC_CANCEL:
5843 ret = io_async_cancel(req);
5845 case IORING_OP_FALLOCATE:
5846 ret = io_fallocate(req, force_nonblock);
5848 case IORING_OP_OPENAT:
5849 ret = io_openat(req, force_nonblock);
5851 case IORING_OP_CLOSE:
5852 ret = io_close(req, force_nonblock, cs);
5854 case IORING_OP_FILES_UPDATE:
5855 ret = io_files_update(req, force_nonblock, cs);
5857 case IORING_OP_STATX:
5858 ret = io_statx(req, force_nonblock);
5860 case IORING_OP_FADVISE:
5861 ret = io_fadvise(req, force_nonblock);
5863 case IORING_OP_MADVISE:
5864 ret = io_madvise(req, force_nonblock);
5866 case IORING_OP_OPENAT2:
5867 ret = io_openat2(req, force_nonblock);
5869 case IORING_OP_EPOLL_CTL:
5870 ret = io_epoll_ctl(req, force_nonblock, cs);
5872 case IORING_OP_SPLICE:
5873 ret = io_splice(req, force_nonblock);
5875 case IORING_OP_PROVIDE_BUFFERS:
5876 ret = io_provide_buffers(req, force_nonblock, cs);
5878 case IORING_OP_REMOVE_BUFFERS:
5879 ret = io_remove_buffers(req, force_nonblock, cs);
5882 ret = io_tee(req, force_nonblock);
5892 /* If the op doesn't have a file, we're not polling for it */
5893 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5894 const bool in_async = io_wq_current_is_worker();
5896 /* workqueue context doesn't hold uring_lock, grab it now */
5898 mutex_lock(&ctx->uring_lock);
5900 io_iopoll_req_issued(req);
5903 mutex_unlock(&ctx->uring_lock);
5909 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5911 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5912 struct io_kiocb *timeout;
5915 timeout = io_prep_linked_timeout(req);
5917 io_queue_linked_timeout(timeout);
5919 /* if NO_CANCEL is set, we must still run the work */
5920 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5921 IO_WQ_WORK_CANCEL) {
5927 ret = io_issue_sqe(req, false, NULL);
5929 * We can get EAGAIN for polled IO even though we're
5930 * forcing a sync submission from here, since we can't
5931 * wait for request slots on the block side.
5940 req_set_fail_links(req);
5941 io_req_complete(req, ret);
5944 return io_steal_work(req);
5947 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5950 struct fixed_file_table *table;
5952 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5953 return table->files[index & IORING_FILE_TABLE_MASK];
5956 static struct file *io_file_get(struct io_submit_state *state,
5957 struct io_kiocb *req, int fd, bool fixed)
5959 struct io_ring_ctx *ctx = req->ctx;
5963 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
5965 fd = array_index_nospec(fd, ctx->nr_user_files);
5966 file = io_file_from_index(ctx, fd);
5968 req->fixed_file_refs = &ctx->file_data->node->refs;
5969 percpu_ref_get(req->fixed_file_refs);
5972 trace_io_uring_file_get(ctx, fd);
5973 file = __io_file_get(state, fd);
5979 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5984 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5985 if (unlikely(!fixed && io_async_submit(req->ctx)))
5988 req->file = io_file_get(state, req, fd, fixed);
5989 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
5994 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5996 struct io_timeout_data *data = container_of(timer,
5997 struct io_timeout_data, timer);
5998 struct io_kiocb *req = data->req;
5999 struct io_ring_ctx *ctx = req->ctx;
6000 struct io_kiocb *prev = NULL;
6001 unsigned long flags;
6003 spin_lock_irqsave(&ctx->completion_lock, flags);
6006 * We don't expect the list to be empty, that will only happen if we
6007 * race with the completion of the linked work.
6009 if (!list_empty(&req->link_list)) {
6010 prev = list_entry(req->link_list.prev, struct io_kiocb,
6012 if (refcount_inc_not_zero(&prev->refs)) {
6013 list_del_init(&req->link_list);
6014 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6019 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6022 req_set_fail_links(prev);
6023 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6026 io_req_complete(req, -ETIME);
6028 return HRTIMER_NORESTART;
6031 static void __io_queue_linked_timeout(struct io_kiocb *req)
6034 * If the list is now empty, then our linked request finished before
6035 * we got a chance to setup the timer
6037 if (!list_empty(&req->link_list)) {
6038 struct io_timeout_data *data = req->async_data;
6040 data->timer.function = io_link_timeout_fn;
6041 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6046 static void io_queue_linked_timeout(struct io_kiocb *req)
6048 struct io_ring_ctx *ctx = req->ctx;
6050 spin_lock_irq(&ctx->completion_lock);
6051 __io_queue_linked_timeout(req);
6052 spin_unlock_irq(&ctx->completion_lock);
6054 /* drop submission reference */
6058 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6060 struct io_kiocb *nxt;
6062 if (!(req->flags & REQ_F_LINK_HEAD))
6064 if (req->flags & REQ_F_LINK_TIMEOUT)
6067 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6069 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6072 req->flags |= REQ_F_LINK_TIMEOUT;
6076 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6078 struct io_kiocb *linked_timeout;
6079 struct io_kiocb *nxt;
6080 const struct cred *old_creds = NULL;
6084 linked_timeout = io_prep_linked_timeout(req);
6086 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6087 req->work.creds != current_cred()) {
6089 revert_creds(old_creds);
6090 if (old_creds == req->work.creds)
6091 old_creds = NULL; /* restored original creds */
6093 old_creds = override_creds(req->work.creds);
6096 ret = io_issue_sqe(req, true, cs);
6099 * We async punt it if the file wasn't marked NOWAIT, or if the file
6100 * doesn't support non-blocking read/write attempts
6102 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6103 if (!io_arm_poll_handler(req)) {
6106 * Queued up for async execution, worker will release
6107 * submit reference when the iocb is actually submitted.
6109 io_queue_async_work(req);
6113 io_queue_linked_timeout(linked_timeout);
6117 if (unlikely(ret)) {
6118 /* un-prep timeout, so it'll be killed as any other linked */
6119 req->flags &= ~REQ_F_LINK_TIMEOUT;
6120 req_set_fail_links(req);
6122 io_req_complete(req, ret);
6126 /* drop submission reference */
6127 nxt = io_put_req_find_next(req);
6129 io_queue_linked_timeout(linked_timeout);
6134 if (req->flags & REQ_F_FORCE_ASYNC)
6140 revert_creds(old_creds);
6143 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6144 struct io_comp_state *cs)
6148 ret = io_req_defer(req, sqe);
6150 if (ret != -EIOCBQUEUED) {
6152 req_set_fail_links(req);
6154 io_req_complete(req, ret);
6156 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6157 if (!req->async_data) {
6158 ret = io_req_defer_prep(req, sqe);
6164 * Never try inline submit of IOSQE_ASYNC is set, go straight
6165 * to async execution.
6167 io_req_init_async(req);
6168 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6169 io_queue_async_work(req);
6172 ret = io_req_prep(req, sqe);
6176 __io_queue_sqe(req, cs);
6180 static inline void io_queue_link_head(struct io_kiocb *req,
6181 struct io_comp_state *cs)
6183 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6185 io_req_complete(req, -ECANCELED);
6187 io_queue_sqe(req, NULL, cs);
6190 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6191 struct io_kiocb **link, struct io_comp_state *cs)
6193 struct io_ring_ctx *ctx = req->ctx;
6197 * If we already have a head request, queue this one for async
6198 * submittal once the head completes. If we don't have a head but
6199 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6200 * submitted sync once the chain is complete. If none of those
6201 * conditions are true (normal request), then just queue it.
6204 struct io_kiocb *head = *link;
6207 * Taking sequential execution of a link, draining both sides
6208 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6209 * requests in the link. So, it drains the head and the
6210 * next after the link request. The last one is done via
6211 * drain_next flag to persist the effect across calls.
6213 if (req->flags & REQ_F_IO_DRAIN) {
6214 head->flags |= REQ_F_IO_DRAIN;
6215 ctx->drain_next = 1;
6217 ret = io_req_defer_prep(req, sqe);
6218 if (unlikely(ret)) {
6219 /* fail even hard links since we don't submit */
6220 head->flags |= REQ_F_FAIL_LINK;
6223 trace_io_uring_link(ctx, req, head);
6224 list_add_tail(&req->link_list, &head->link_list);
6226 /* last request of a link, enqueue the link */
6227 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6228 io_queue_link_head(head, cs);
6232 if (unlikely(ctx->drain_next)) {
6233 req->flags |= REQ_F_IO_DRAIN;
6234 ctx->drain_next = 0;
6236 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6237 req->flags |= REQ_F_LINK_HEAD;
6238 INIT_LIST_HEAD(&req->link_list);
6240 ret = io_req_defer_prep(req, sqe);
6242 req->flags |= REQ_F_FAIL_LINK;
6245 io_queue_sqe(req, sqe, cs);
6253 * Batched submission is done, ensure local IO is flushed out.
6255 static void io_submit_state_end(struct io_submit_state *state)
6257 if (!list_empty(&state->comp.list))
6258 io_submit_flush_completions(&state->comp);
6259 blk_finish_plug(&state->plug);
6260 io_state_file_put(state);
6261 if (state->free_reqs)
6262 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6266 * Start submission side cache.
6268 static void io_submit_state_start(struct io_submit_state *state,
6269 struct io_ring_ctx *ctx, unsigned int max_ios)
6271 blk_start_plug(&state->plug);
6273 INIT_LIST_HEAD(&state->comp.list);
6274 state->comp.ctx = ctx;
6275 state->free_reqs = 0;
6277 state->ios_left = max_ios;
6280 static void io_commit_sqring(struct io_ring_ctx *ctx)
6282 struct io_rings *rings = ctx->rings;
6285 * Ensure any loads from the SQEs are done at this point,
6286 * since once we write the new head, the application could
6287 * write new data to them.
6289 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6293 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6294 * that is mapped by userspace. This means that care needs to be taken to
6295 * ensure that reads are stable, as we cannot rely on userspace always
6296 * being a good citizen. If members of the sqe are validated and then later
6297 * used, it's important that those reads are done through READ_ONCE() to
6298 * prevent a re-load down the line.
6300 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6302 u32 *sq_array = ctx->sq_array;
6306 * The cached sq head (or cq tail) serves two purposes:
6308 * 1) allows us to batch the cost of updating the user visible
6310 * 2) allows the kernel side to track the head on its own, even
6311 * though the application is the one updating it.
6313 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6314 if (likely(head < ctx->sq_entries))
6315 return &ctx->sq_sqes[head];
6317 /* drop invalid entries */
6318 ctx->cached_sq_dropped++;
6319 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6323 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6325 ctx->cached_sq_head++;
6329 * Check SQE restrictions (opcode and flags).
6331 * Returns 'true' if SQE is allowed, 'false' otherwise.
6333 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6334 struct io_kiocb *req,
6335 unsigned int sqe_flags)
6337 if (!ctx->restricted)
6340 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6343 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6344 ctx->restrictions.sqe_flags_required)
6347 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6348 ctx->restrictions.sqe_flags_required))
6354 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6355 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6356 IOSQE_BUFFER_SELECT)
6358 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6359 const struct io_uring_sqe *sqe,
6360 struct io_submit_state *state)
6362 unsigned int sqe_flags;
6365 req->opcode = READ_ONCE(sqe->opcode);
6366 req->user_data = READ_ONCE(sqe->user_data);
6367 req->async_data = NULL;
6371 /* one is dropped after submission, the other at completion */
6372 refcount_set(&req->refs, 2);
6373 req->task = current;
6376 if (unlikely(req->opcode >= IORING_OP_LAST))
6379 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6382 sqe_flags = READ_ONCE(sqe->flags);
6383 /* enforce forwards compatibility on users */
6384 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6387 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6390 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6391 !io_op_defs[req->opcode].buffer_select)
6394 id = READ_ONCE(sqe->personality);
6396 io_req_init_async(req);
6397 req->work.creds = idr_find(&ctx->personality_idr, id);
6398 if (unlikely(!req->work.creds))
6400 get_cred(req->work.creds);
6403 /* same numerical values with corresponding REQ_F_*, safe to copy */
6404 req->flags |= sqe_flags;
6406 if (!io_op_defs[req->opcode].needs_file)
6409 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6414 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6416 struct io_submit_state state;
6417 struct io_kiocb *link = NULL;
6418 int i, submitted = 0;
6420 /* if we have a backlog and couldn't flush it all, return BUSY */
6421 if (test_bit(0, &ctx->sq_check_overflow)) {
6422 if (!list_empty(&ctx->cq_overflow_list) &&
6423 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6427 /* make sure SQ entry isn't read before tail */
6428 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6430 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6433 atomic_long_add(nr, ¤t->io_uring->req_issue);
6434 refcount_add(nr, ¤t->usage);
6436 io_submit_state_start(&state, ctx, nr);
6438 for (i = 0; i < nr; i++) {
6439 const struct io_uring_sqe *sqe;
6440 struct io_kiocb *req;
6443 sqe = io_get_sqe(ctx);
6444 if (unlikely(!sqe)) {
6445 io_consume_sqe(ctx);
6448 req = io_alloc_req(ctx, &state);
6449 if (unlikely(!req)) {
6451 submitted = -EAGAIN;
6454 io_consume_sqe(ctx);
6455 /* will complete beyond this point, count as submitted */
6458 err = io_init_req(ctx, req, sqe, &state);
6459 if (unlikely(err)) {
6462 io_req_complete(req, err);
6466 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6467 true, io_async_submit(ctx));
6468 err = io_submit_sqe(req, sqe, &link, &state.comp);
6473 if (unlikely(submitted != nr)) {
6474 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6476 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6477 atomic_long_sub(nr - ref_used, ¤t->io_uring->req_issue);
6478 put_task_struct_many(current, nr - ref_used);
6481 io_queue_link_head(link, &state.comp);
6482 io_submit_state_end(&state);
6484 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6485 io_commit_sqring(ctx);
6490 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6492 /* Tell userspace we may need a wakeup call */
6493 spin_lock_irq(&ctx->completion_lock);
6494 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6495 spin_unlock_irq(&ctx->completion_lock);
6498 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6500 spin_lock_irq(&ctx->completion_lock);
6501 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6502 spin_unlock_irq(&ctx->completion_lock);
6505 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6506 int sync, void *key)
6508 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6511 ret = autoremove_wake_function(wqe, mode, sync, key);
6513 unsigned long flags;
6515 spin_lock_irqsave(&ctx->completion_lock, flags);
6516 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6517 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6528 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6529 unsigned long start_jiffies, bool cap_entries)
6531 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6532 struct io_sq_data *sqd = ctx->sq_data;
6533 unsigned int to_submit;
6537 if (!list_empty(&ctx->iopoll_list)) {
6538 unsigned nr_events = 0;
6540 mutex_lock(&ctx->uring_lock);
6541 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6542 io_do_iopoll(ctx, &nr_events, 0);
6543 mutex_unlock(&ctx->uring_lock);
6546 to_submit = io_sqring_entries(ctx);
6549 * If submit got -EBUSY, flag us as needing the application
6550 * to enter the kernel to reap and flush events.
6552 if (!to_submit || ret == -EBUSY || need_resched()) {
6554 * Drop cur_mm before scheduling, we can't hold it for
6555 * long periods (or over schedule()). Do this before
6556 * adding ourselves to the waitqueue, as the unuse/drop
6559 io_sq_thread_drop_mm();
6562 * We're polling. If we're within the defined idle
6563 * period, then let us spin without work before going
6564 * to sleep. The exception is if we got EBUSY doing
6565 * more IO, we should wait for the application to
6566 * reap events and wake us up.
6568 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6569 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6570 !percpu_ref_is_dying(&ctx->refs)))
6573 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6574 TASK_INTERRUPTIBLE);
6577 * While doing polled IO, before going to sleep, we need
6578 * to check if there are new reqs added to iopoll_list,
6579 * it is because reqs may have been punted to io worker
6580 * and will be added to iopoll_list later, hence check
6581 * the iopoll_list again.
6583 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6584 !list_empty_careful(&ctx->iopoll_list)) {
6585 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6589 to_submit = io_sqring_entries(ctx);
6590 if (!to_submit || ret == -EBUSY)
6594 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6595 io_ring_clear_wakeup_flag(ctx);
6597 /* if we're handling multiple rings, cap submit size for fairness */
6598 if (cap_entries && to_submit > 8)
6601 mutex_lock(&ctx->uring_lock);
6602 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6603 ret = io_submit_sqes(ctx, to_submit);
6604 mutex_unlock(&ctx->uring_lock);
6606 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6607 wake_up(&ctx->sqo_sq_wait);
6609 return SQT_DID_WORK;
6612 static void io_sqd_init_new(struct io_sq_data *sqd)
6614 struct io_ring_ctx *ctx;
6616 while (!list_empty(&sqd->ctx_new_list)) {
6617 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6618 init_wait(&ctx->sqo_wait_entry);
6619 ctx->sqo_wait_entry.func = io_sq_wake_function;
6620 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6621 complete(&ctx->sq_thread_comp);
6625 static int io_sq_thread(void *data)
6627 struct cgroup_subsys_state *cur_css = NULL;
6628 const struct cred *old_cred = NULL;
6629 struct io_sq_data *sqd = data;
6630 struct io_ring_ctx *ctx;
6631 unsigned long start_jiffies;
6633 start_jiffies = jiffies;
6634 while (!kthread_should_stop()) {
6635 enum sq_ret ret = 0;
6639 * Any changes to the sqd lists are synchronized through the
6640 * kthread parking. This synchronizes the thread vs users,
6641 * the users are synchronized on the sqd->ctx_lock.
6643 if (kthread_should_park())
6646 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6647 io_sqd_init_new(sqd);
6649 cap_entries = !list_is_singular(&sqd->ctx_list);
6651 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6652 if (current->cred != ctx->creds) {
6654 revert_creds(old_cred);
6655 old_cred = override_creds(ctx->creds);
6657 io_sq_thread_associate_blkcg(ctx, &cur_css);
6659 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6661 io_sq_thread_drop_mm();
6664 if (ret & SQT_SPIN) {
6667 } else if (ret == SQT_IDLE) {
6668 if (kthread_should_park())
6670 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6671 io_ring_set_wakeup_flag(ctx);
6673 start_jiffies = jiffies;
6674 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6675 io_ring_clear_wakeup_flag(ctx);
6682 io_sq_thread_unassociate_blkcg();
6684 revert_creds(old_cred);
6691 struct io_wait_queue {
6692 struct wait_queue_entry wq;
6693 struct io_ring_ctx *ctx;
6695 unsigned nr_timeouts;
6698 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6700 struct io_ring_ctx *ctx = iowq->ctx;
6703 * Wake up if we have enough events, or if a timeout occurred since we
6704 * started waiting. For timeouts, we always want to return to userspace,
6705 * regardless of event count.
6707 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6708 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6711 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6712 int wake_flags, void *key)
6714 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6717 /* use noflush == true, as we can't safely rely on locking context */
6718 if (!io_should_wake(iowq, true))
6721 return autoremove_wake_function(curr, mode, wake_flags, key);
6724 static int io_run_task_work_sig(void)
6726 if (io_run_task_work())
6728 if (!signal_pending(current))
6730 if (current->jobctl & JOBCTL_TASK_WORK) {
6731 spin_lock_irq(¤t->sighand->siglock);
6732 current->jobctl &= ~JOBCTL_TASK_WORK;
6733 recalc_sigpending();
6734 spin_unlock_irq(¤t->sighand->siglock);
6741 * Wait until events become available, if we don't already have some. The
6742 * application must reap them itself, as they reside on the shared cq ring.
6744 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6745 const sigset_t __user *sig, size_t sigsz)
6747 struct io_wait_queue iowq = {
6750 .func = io_wake_function,
6751 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6754 .to_wait = min_events,
6756 struct io_rings *rings = ctx->rings;
6760 if (io_cqring_events(ctx, false) >= min_events)
6762 if (!io_run_task_work())
6767 #ifdef CONFIG_COMPAT
6768 if (in_compat_syscall())
6769 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6773 ret = set_user_sigmask(sig, sigsz);
6779 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6780 trace_io_uring_cqring_wait(ctx, min_events);
6782 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6783 TASK_INTERRUPTIBLE);
6784 /* make sure we run task_work before checking for signals */
6785 ret = io_run_task_work_sig();
6790 if (io_should_wake(&iowq, false))
6794 finish_wait(&ctx->wait, &iowq.wq);
6796 restore_saved_sigmask_unless(ret == -EINTR);
6798 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6801 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6803 #if defined(CONFIG_UNIX)
6804 if (ctx->ring_sock) {
6805 struct sock *sock = ctx->ring_sock->sk;
6806 struct sk_buff *skb;
6808 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6814 for (i = 0; i < ctx->nr_user_files; i++) {
6817 file = io_file_from_index(ctx, i);
6824 static void io_file_ref_kill(struct percpu_ref *ref)
6826 struct fixed_file_data *data;
6828 data = container_of(ref, struct fixed_file_data, refs);
6829 complete(&data->done);
6832 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6834 struct fixed_file_data *data = ctx->file_data;
6835 struct fixed_file_ref_node *ref_node = NULL;
6836 unsigned nr_tables, i;
6841 spin_lock(&data->lock);
6842 if (!list_empty(&data->ref_list))
6843 ref_node = list_first_entry(&data->ref_list,
6844 struct fixed_file_ref_node, node);
6845 spin_unlock(&data->lock);
6847 percpu_ref_kill(&ref_node->refs);
6849 percpu_ref_kill(&data->refs);
6851 /* wait for all refs nodes to complete */
6852 flush_delayed_work(&ctx->file_put_work);
6853 wait_for_completion(&data->done);
6855 __io_sqe_files_unregister(ctx);
6856 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6857 for (i = 0; i < nr_tables; i++)
6858 kfree(data->table[i].files);
6860 percpu_ref_exit(&data->refs);
6862 ctx->file_data = NULL;
6863 ctx->nr_user_files = 0;
6867 static void io_put_sq_data(struct io_sq_data *sqd)
6869 if (refcount_dec_and_test(&sqd->refs)) {
6871 * The park is a bit of a work-around, without it we get
6872 * warning spews on shutdown with SQPOLL set and affinity
6873 * set to a single CPU.
6876 kthread_park(sqd->thread);
6877 kthread_stop(sqd->thread);
6884 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
6886 struct io_ring_ctx *ctx_attach;
6887 struct io_sq_data *sqd;
6890 f = fdget(p->wq_fd);
6892 return ERR_PTR(-ENXIO);
6893 if (f.file->f_op != &io_uring_fops) {
6895 return ERR_PTR(-EINVAL);
6898 ctx_attach = f.file->private_data;
6899 sqd = ctx_attach->sq_data;
6902 return ERR_PTR(-EINVAL);
6905 refcount_inc(&sqd->refs);
6910 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
6912 struct io_sq_data *sqd;
6914 if (p->flags & IORING_SETUP_ATTACH_WQ)
6915 return io_attach_sq_data(p);
6917 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
6919 return ERR_PTR(-ENOMEM);
6921 refcount_set(&sqd->refs, 1);
6922 INIT_LIST_HEAD(&sqd->ctx_list);
6923 INIT_LIST_HEAD(&sqd->ctx_new_list);
6924 mutex_init(&sqd->ctx_lock);
6925 mutex_init(&sqd->lock);
6926 init_waitqueue_head(&sqd->wait);
6930 static void io_sq_thread_unpark(struct io_sq_data *sqd)
6931 __releases(&sqd->lock)
6935 kthread_unpark(sqd->thread);
6936 mutex_unlock(&sqd->lock);
6939 static void io_sq_thread_park(struct io_sq_data *sqd)
6940 __acquires(&sqd->lock)
6944 mutex_lock(&sqd->lock);
6945 kthread_park(sqd->thread);
6948 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6950 struct io_sq_data *sqd = ctx->sq_data;
6955 * We may arrive here from the error branch in
6956 * io_sq_offload_create() where the kthread is created
6957 * without being waked up, thus wake it up now to make
6958 * sure the wait will complete.
6960 wake_up_process(sqd->thread);
6961 wait_for_completion(&ctx->sq_thread_comp);
6963 io_sq_thread_park(sqd);
6966 mutex_lock(&sqd->ctx_lock);
6967 list_del(&ctx->sqd_list);
6968 mutex_unlock(&sqd->ctx_lock);
6971 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6972 io_sq_thread_unpark(sqd);
6975 io_put_sq_data(sqd);
6976 ctx->sq_data = NULL;
6980 static void io_finish_async(struct io_ring_ctx *ctx)
6982 io_sq_thread_stop(ctx);
6985 io_wq_destroy(ctx->io_wq);
6990 #if defined(CONFIG_UNIX)
6992 * Ensure the UNIX gc is aware of our file set, so we are certain that
6993 * the io_uring can be safely unregistered on process exit, even if we have
6994 * loops in the file referencing.
6996 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6998 struct sock *sk = ctx->ring_sock->sk;
6999 struct scm_fp_list *fpl;
7000 struct sk_buff *skb;
7003 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7007 skb = alloc_skb(0, GFP_KERNEL);
7016 fpl->user = get_uid(ctx->user);
7017 for (i = 0; i < nr; i++) {
7018 struct file *file = io_file_from_index(ctx, i + offset);
7022 fpl->fp[nr_files] = get_file(file);
7023 unix_inflight(fpl->user, fpl->fp[nr_files]);
7028 fpl->max = SCM_MAX_FD;
7029 fpl->count = nr_files;
7030 UNIXCB(skb).fp = fpl;
7031 skb->destructor = unix_destruct_scm;
7032 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7033 skb_queue_head(&sk->sk_receive_queue, skb);
7035 for (i = 0; i < nr_files; i++)
7046 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7047 * causes regular reference counting to break down. We rely on the UNIX
7048 * garbage collection to take care of this problem for us.
7050 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7052 unsigned left, total;
7056 left = ctx->nr_user_files;
7058 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7060 ret = __io_sqe_files_scm(ctx, this_files, total);
7064 total += this_files;
7070 while (total < ctx->nr_user_files) {
7071 struct file *file = io_file_from_index(ctx, total);
7081 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7087 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7088 unsigned nr_tables, unsigned nr_files)
7092 for (i = 0; i < nr_tables; i++) {
7093 struct fixed_file_table *table = &file_data->table[i];
7094 unsigned this_files;
7096 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7097 table->files = kcalloc(this_files, sizeof(struct file *),
7101 nr_files -= this_files;
7107 for (i = 0; i < nr_tables; i++) {
7108 struct fixed_file_table *table = &file_data->table[i];
7109 kfree(table->files);
7114 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7116 #if defined(CONFIG_UNIX)
7117 struct sock *sock = ctx->ring_sock->sk;
7118 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7119 struct sk_buff *skb;
7122 __skb_queue_head_init(&list);
7125 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7126 * remove this entry and rearrange the file array.
7128 skb = skb_dequeue(head);
7130 struct scm_fp_list *fp;
7132 fp = UNIXCB(skb).fp;
7133 for (i = 0; i < fp->count; i++) {
7136 if (fp->fp[i] != file)
7139 unix_notinflight(fp->user, fp->fp[i]);
7140 left = fp->count - 1 - i;
7142 memmove(&fp->fp[i], &fp->fp[i + 1],
7143 left * sizeof(struct file *));
7150 __skb_queue_tail(&list, skb);
7160 __skb_queue_tail(&list, skb);
7162 skb = skb_dequeue(head);
7165 if (skb_peek(&list)) {
7166 spin_lock_irq(&head->lock);
7167 while ((skb = __skb_dequeue(&list)) != NULL)
7168 __skb_queue_tail(head, skb);
7169 spin_unlock_irq(&head->lock);
7176 struct io_file_put {
7177 struct list_head list;
7181 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7183 struct fixed_file_data *file_data = ref_node->file_data;
7184 struct io_ring_ctx *ctx = file_data->ctx;
7185 struct io_file_put *pfile, *tmp;
7187 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7188 list_del(&pfile->list);
7189 io_ring_file_put(ctx, pfile->file);
7193 spin_lock(&file_data->lock);
7194 list_del(&ref_node->node);
7195 spin_unlock(&file_data->lock);
7197 percpu_ref_exit(&ref_node->refs);
7199 percpu_ref_put(&file_data->refs);
7202 static void io_file_put_work(struct work_struct *work)
7204 struct io_ring_ctx *ctx;
7205 struct llist_node *node;
7207 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7208 node = llist_del_all(&ctx->file_put_llist);
7211 struct fixed_file_ref_node *ref_node;
7212 struct llist_node *next = node->next;
7214 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7215 __io_file_put_work(ref_node);
7220 static void io_file_data_ref_zero(struct percpu_ref *ref)
7222 struct fixed_file_ref_node *ref_node;
7223 struct io_ring_ctx *ctx;
7227 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7228 ctx = ref_node->file_data->ctx;
7230 if (percpu_ref_is_dying(&ctx->file_data->refs))
7233 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7235 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7237 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7240 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7241 struct io_ring_ctx *ctx)
7243 struct fixed_file_ref_node *ref_node;
7245 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7247 return ERR_PTR(-ENOMEM);
7249 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7252 return ERR_PTR(-ENOMEM);
7254 INIT_LIST_HEAD(&ref_node->node);
7255 INIT_LIST_HEAD(&ref_node->file_list);
7256 ref_node->file_data = ctx->file_data;
7260 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7262 percpu_ref_exit(&ref_node->refs);
7266 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7269 __s32 __user *fds = (__s32 __user *) arg;
7270 unsigned nr_tables, i;
7272 int fd, ret = -ENOMEM;
7273 struct fixed_file_ref_node *ref_node;
7274 struct fixed_file_data *file_data;
7280 if (nr_args > IORING_MAX_FIXED_FILES)
7283 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7286 file_data->ctx = ctx;
7287 init_completion(&file_data->done);
7288 INIT_LIST_HEAD(&file_data->ref_list);
7289 spin_lock_init(&file_data->lock);
7291 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7292 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7294 if (!file_data->table)
7297 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7298 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7301 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7304 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7305 struct fixed_file_table *table;
7308 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7312 /* allow sparse sets */
7322 * Don't allow io_uring instances to be registered. If UNIX
7323 * isn't enabled, then this causes a reference cycle and this
7324 * instance can never get freed. If UNIX is enabled we'll
7325 * handle it just fine, but there's still no point in allowing
7326 * a ring fd as it doesn't support regular read/write anyway.
7328 if (file->f_op == &io_uring_fops) {
7332 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7333 index = i & IORING_FILE_TABLE_MASK;
7334 table->files[index] = file;
7337 ctx->file_data = file_data;
7338 ret = io_sqe_files_scm(ctx);
7340 io_sqe_files_unregister(ctx);
7344 ref_node = alloc_fixed_file_ref_node(ctx);
7345 if (IS_ERR(ref_node)) {
7346 io_sqe_files_unregister(ctx);
7347 return PTR_ERR(ref_node);
7350 file_data->node = ref_node;
7351 spin_lock(&file_data->lock);
7352 list_add(&ref_node->node, &file_data->ref_list);
7353 spin_unlock(&file_data->lock);
7354 percpu_ref_get(&file_data->refs);
7357 for (i = 0; i < ctx->nr_user_files; i++) {
7358 file = io_file_from_index(ctx, i);
7362 for (i = 0; i < nr_tables; i++)
7363 kfree(file_data->table[i].files);
7364 ctx->nr_user_files = 0;
7366 percpu_ref_exit(&file_data->refs);
7368 kfree(file_data->table);
7373 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7376 #if defined(CONFIG_UNIX)
7377 struct sock *sock = ctx->ring_sock->sk;
7378 struct sk_buff_head *head = &sock->sk_receive_queue;
7379 struct sk_buff *skb;
7382 * See if we can merge this file into an existing skb SCM_RIGHTS
7383 * file set. If there's no room, fall back to allocating a new skb
7384 * and filling it in.
7386 spin_lock_irq(&head->lock);
7387 skb = skb_peek(head);
7389 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7391 if (fpl->count < SCM_MAX_FD) {
7392 __skb_unlink(skb, head);
7393 spin_unlock_irq(&head->lock);
7394 fpl->fp[fpl->count] = get_file(file);
7395 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7397 spin_lock_irq(&head->lock);
7398 __skb_queue_head(head, skb);
7403 spin_unlock_irq(&head->lock);
7410 return __io_sqe_files_scm(ctx, 1, index);
7416 static int io_queue_file_removal(struct fixed_file_data *data,
7419 struct io_file_put *pfile;
7420 struct fixed_file_ref_node *ref_node = data->node;
7422 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7427 list_add(&pfile->list, &ref_node->file_list);
7432 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7433 struct io_uring_files_update *up,
7436 struct fixed_file_data *data = ctx->file_data;
7437 struct fixed_file_ref_node *ref_node;
7442 bool needs_switch = false;
7444 if (check_add_overflow(up->offset, nr_args, &done))
7446 if (done > ctx->nr_user_files)
7449 ref_node = alloc_fixed_file_ref_node(ctx);
7450 if (IS_ERR(ref_node))
7451 return PTR_ERR(ref_node);
7454 fds = u64_to_user_ptr(up->fds);
7456 struct fixed_file_table *table;
7460 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7464 i = array_index_nospec(up->offset, ctx->nr_user_files);
7465 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7466 index = i & IORING_FILE_TABLE_MASK;
7467 if (table->files[index]) {
7468 file = table->files[index];
7469 err = io_queue_file_removal(data, file);
7472 table->files[index] = NULL;
7473 needs_switch = true;
7482 * Don't allow io_uring instances to be registered. If
7483 * UNIX isn't enabled, then this causes a reference
7484 * cycle and this instance can never get freed. If UNIX
7485 * is enabled we'll handle it just fine, but there's
7486 * still no point in allowing a ring fd as it doesn't
7487 * support regular read/write anyway.
7489 if (file->f_op == &io_uring_fops) {
7494 table->files[index] = file;
7495 err = io_sqe_file_register(ctx, file, i);
7497 table->files[index] = NULL;
7508 percpu_ref_kill(&data->node->refs);
7509 spin_lock(&data->lock);
7510 list_add(&ref_node->node, &data->ref_list);
7511 data->node = ref_node;
7512 spin_unlock(&data->lock);
7513 percpu_ref_get(&ctx->file_data->refs);
7515 destroy_fixed_file_ref_node(ref_node);
7517 return done ? done : err;
7520 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7523 struct io_uring_files_update up;
7525 if (!ctx->file_data)
7529 if (copy_from_user(&up, arg, sizeof(up)))
7534 return __io_sqe_files_update(ctx, &up, nr_args);
7537 static void io_free_work(struct io_wq_work *work)
7539 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7541 /* Consider that io_steal_work() relies on this ref */
7545 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7546 struct io_uring_params *p)
7548 struct io_wq_data data;
7550 struct io_ring_ctx *ctx_attach;
7551 unsigned int concurrency;
7554 data.user = ctx->user;
7555 data.free_work = io_free_work;
7556 data.do_work = io_wq_submit_work;
7558 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7559 /* Do QD, or 4 * CPUS, whatever is smallest */
7560 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7562 ctx->io_wq = io_wq_create(concurrency, &data);
7563 if (IS_ERR(ctx->io_wq)) {
7564 ret = PTR_ERR(ctx->io_wq);
7570 f = fdget(p->wq_fd);
7574 if (f.file->f_op != &io_uring_fops) {
7579 ctx_attach = f.file->private_data;
7580 /* @io_wq is protected by holding the fd */
7581 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7586 ctx->io_wq = ctx_attach->io_wq;
7592 static int io_uring_alloc_task_context(struct task_struct *task)
7594 struct io_uring_task *tctx;
7596 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7597 if (unlikely(!tctx))
7601 init_waitqueue_head(&tctx->wait);
7604 atomic_long_set(&tctx->req_issue, 0);
7605 atomic_long_set(&tctx->req_complete, 0);
7606 task->io_uring = tctx;
7610 void __io_uring_free(struct task_struct *tsk)
7612 struct io_uring_task *tctx = tsk->io_uring;
7614 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7616 tsk->io_uring = NULL;
7619 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7620 struct io_uring_params *p)
7624 if (ctx->flags & IORING_SETUP_SQPOLL) {
7625 struct io_sq_data *sqd;
7628 if (!capable(CAP_SYS_ADMIN))
7631 sqd = io_get_sq_data(p);
7638 io_sq_thread_park(sqd);
7639 mutex_lock(&sqd->ctx_lock);
7640 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7641 mutex_unlock(&sqd->ctx_lock);
7642 io_sq_thread_unpark(sqd);
7644 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7645 if (!ctx->sq_thread_idle)
7646 ctx->sq_thread_idle = HZ;
7651 if (p->flags & IORING_SETUP_SQ_AFF) {
7652 int cpu = p->sq_thread_cpu;
7655 if (cpu >= nr_cpu_ids)
7657 if (!cpu_online(cpu))
7660 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7661 cpu, "io_uring-sq");
7663 sqd->thread = kthread_create(io_sq_thread, sqd,
7666 if (IS_ERR(sqd->thread)) {
7667 ret = PTR_ERR(sqd->thread);
7671 ret = io_uring_alloc_task_context(sqd->thread);
7674 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7675 /* Can't have SQ_AFF without SQPOLL */
7681 ret = io_init_wq_offload(ctx, p);
7687 io_finish_async(ctx);
7691 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7693 struct io_sq_data *sqd = ctx->sq_data;
7695 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7696 wake_up_process(sqd->thread);
7699 static inline void __io_unaccount_mem(struct user_struct *user,
7700 unsigned long nr_pages)
7702 atomic_long_sub(nr_pages, &user->locked_vm);
7705 static inline int __io_account_mem(struct user_struct *user,
7706 unsigned long nr_pages)
7708 unsigned long page_limit, cur_pages, new_pages;
7710 /* Don't allow more pages than we can safely lock */
7711 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7714 cur_pages = atomic_long_read(&user->locked_vm);
7715 new_pages = cur_pages + nr_pages;
7716 if (new_pages > page_limit)
7718 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7719 new_pages) != cur_pages);
7724 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7725 enum io_mem_account acct)
7728 __io_unaccount_mem(ctx->user, nr_pages);
7730 if (ctx->mm_account) {
7731 if (acct == ACCT_LOCKED)
7732 ctx->mm_account->locked_vm -= nr_pages;
7733 else if (acct == ACCT_PINNED)
7734 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7738 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7739 enum io_mem_account acct)
7743 if (ctx->limit_mem) {
7744 ret = __io_account_mem(ctx->user, nr_pages);
7749 if (ctx->mm_account) {
7750 if (acct == ACCT_LOCKED)
7751 ctx->mm_account->locked_vm += nr_pages;
7752 else if (acct == ACCT_PINNED)
7753 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7759 static void io_mem_free(void *ptr)
7766 page = virt_to_head_page(ptr);
7767 if (put_page_testzero(page))
7768 free_compound_page(page);
7771 static void *io_mem_alloc(size_t size)
7773 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7776 return (void *) __get_free_pages(gfp_flags, get_order(size));
7779 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7782 struct io_rings *rings;
7783 size_t off, sq_array_size;
7785 off = struct_size(rings, cqes, cq_entries);
7786 if (off == SIZE_MAX)
7790 off = ALIGN(off, SMP_CACHE_BYTES);
7798 sq_array_size = array_size(sizeof(u32), sq_entries);
7799 if (sq_array_size == SIZE_MAX)
7802 if (check_add_overflow(off, sq_array_size, &off))
7808 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7812 pages = (size_t)1 << get_order(
7813 rings_size(sq_entries, cq_entries, NULL));
7814 pages += (size_t)1 << get_order(
7815 array_size(sizeof(struct io_uring_sqe), sq_entries));
7820 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7824 if (!ctx->user_bufs)
7827 for (i = 0; i < ctx->nr_user_bufs; i++) {
7828 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7830 for (j = 0; j < imu->nr_bvecs; j++)
7831 unpin_user_page(imu->bvec[j].bv_page);
7833 if (imu->acct_pages)
7834 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7839 kfree(ctx->user_bufs);
7840 ctx->user_bufs = NULL;
7841 ctx->nr_user_bufs = 0;
7845 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7846 void __user *arg, unsigned index)
7848 struct iovec __user *src;
7850 #ifdef CONFIG_COMPAT
7852 struct compat_iovec __user *ciovs;
7853 struct compat_iovec ciov;
7855 ciovs = (struct compat_iovec __user *) arg;
7856 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7859 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7860 dst->iov_len = ciov.iov_len;
7864 src = (struct iovec __user *) arg;
7865 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7871 * Not super efficient, but this is just a registration time. And we do cache
7872 * the last compound head, so generally we'll only do a full search if we don't
7875 * We check if the given compound head page has already been accounted, to
7876 * avoid double accounting it. This allows us to account the full size of the
7877 * page, not just the constituent pages of a huge page.
7879 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
7880 int nr_pages, struct page *hpage)
7884 /* check current page array */
7885 for (i = 0; i < nr_pages; i++) {
7886 if (!PageCompound(pages[i]))
7888 if (compound_head(pages[i]) == hpage)
7892 /* check previously registered pages */
7893 for (i = 0; i < ctx->nr_user_bufs; i++) {
7894 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7896 for (j = 0; j < imu->nr_bvecs; j++) {
7897 if (!PageCompound(imu->bvec[j].bv_page))
7899 if (compound_head(imu->bvec[j].bv_page) == hpage)
7907 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
7908 int nr_pages, struct io_mapped_ubuf *imu,
7909 struct page **last_hpage)
7913 for (i = 0; i < nr_pages; i++) {
7914 if (!PageCompound(pages[i])) {
7919 hpage = compound_head(pages[i]);
7920 if (hpage == *last_hpage)
7922 *last_hpage = hpage;
7923 if (headpage_already_acct(ctx, pages, i, hpage))
7925 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
7929 if (!imu->acct_pages)
7932 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
7934 imu->acct_pages = 0;
7938 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7941 struct vm_area_struct **vmas = NULL;
7942 struct page **pages = NULL;
7943 struct page *last_hpage = NULL;
7944 int i, j, got_pages = 0;
7949 if (!nr_args || nr_args > UIO_MAXIOV)
7952 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7954 if (!ctx->user_bufs)
7957 for (i = 0; i < nr_args; i++) {
7958 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7959 unsigned long off, start, end, ubuf;
7964 ret = io_copy_iov(ctx, &iov, arg, i);
7969 * Don't impose further limits on the size and buffer
7970 * constraints here, we'll -EINVAL later when IO is
7971 * submitted if they are wrong.
7974 if (!iov.iov_base || !iov.iov_len)
7977 /* arbitrary limit, but we need something */
7978 if (iov.iov_len > SZ_1G)
7981 ubuf = (unsigned long) iov.iov_base;
7982 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7983 start = ubuf >> PAGE_SHIFT;
7984 nr_pages = end - start;
7987 if (!pages || nr_pages > got_pages) {
7990 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7992 vmas = kvmalloc_array(nr_pages,
7993 sizeof(struct vm_area_struct *),
7995 if (!pages || !vmas) {
7999 got_pages = nr_pages;
8002 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8009 mmap_read_lock(current->mm);
8010 pret = pin_user_pages(ubuf, nr_pages,
8011 FOLL_WRITE | FOLL_LONGTERM,
8013 if (pret == nr_pages) {
8014 /* don't support file backed memory */
8015 for (j = 0; j < nr_pages; j++) {
8016 struct vm_area_struct *vma = vmas[j];
8019 !is_file_hugepages(vma->vm_file)) {
8025 ret = pret < 0 ? pret : -EFAULT;
8027 mmap_read_unlock(current->mm);
8030 * if we did partial map, or found file backed vmas,
8031 * release any pages we did get
8034 unpin_user_pages(pages, pret);
8039 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8041 unpin_user_pages(pages, pret);
8046 off = ubuf & ~PAGE_MASK;
8048 for (j = 0; j < nr_pages; j++) {
8051 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8052 imu->bvec[j].bv_page = pages[j];
8053 imu->bvec[j].bv_len = vec_len;
8054 imu->bvec[j].bv_offset = off;
8058 /* store original address for later verification */
8060 imu->len = iov.iov_len;
8061 imu->nr_bvecs = nr_pages;
8063 ctx->nr_user_bufs++;
8071 io_sqe_buffer_unregister(ctx);
8075 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8077 __s32 __user *fds = arg;
8083 if (copy_from_user(&fd, fds, sizeof(*fds)))
8086 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8087 if (IS_ERR(ctx->cq_ev_fd)) {
8088 int ret = PTR_ERR(ctx->cq_ev_fd);
8089 ctx->cq_ev_fd = NULL;
8096 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8098 if (ctx->cq_ev_fd) {
8099 eventfd_ctx_put(ctx->cq_ev_fd);
8100 ctx->cq_ev_fd = NULL;
8107 static int __io_destroy_buffers(int id, void *p, void *data)
8109 struct io_ring_ctx *ctx = data;
8110 struct io_buffer *buf = p;
8112 __io_remove_buffers(ctx, buf, id, -1U);
8116 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8118 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8119 idr_destroy(&ctx->io_buffer_idr);
8122 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8124 io_finish_async(ctx);
8125 io_sqe_buffer_unregister(ctx);
8127 if (ctx->sqo_task) {
8128 put_task_struct(ctx->sqo_task);
8129 ctx->sqo_task = NULL;
8130 mmdrop(ctx->mm_account);
8131 ctx->mm_account = NULL;
8134 #ifdef CONFIG_BLK_CGROUP
8135 if (ctx->sqo_blkcg_css)
8136 css_put(ctx->sqo_blkcg_css);
8139 io_sqe_files_unregister(ctx);
8140 io_eventfd_unregister(ctx);
8141 io_destroy_buffers(ctx);
8142 idr_destroy(&ctx->personality_idr);
8144 #if defined(CONFIG_UNIX)
8145 if (ctx->ring_sock) {
8146 ctx->ring_sock->file = NULL; /* so that iput() is called */
8147 sock_release(ctx->ring_sock);
8151 io_mem_free(ctx->rings);
8152 io_mem_free(ctx->sq_sqes);
8154 percpu_ref_exit(&ctx->refs);
8155 free_uid(ctx->user);
8156 put_cred(ctx->creds);
8157 kfree(ctx->cancel_hash);
8158 kmem_cache_free(req_cachep, ctx->fallback_req);
8162 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8164 struct io_ring_ctx *ctx = file->private_data;
8167 poll_wait(file, &ctx->cq_wait, wait);
8169 * synchronizes with barrier from wq_has_sleeper call in
8173 if (!io_sqring_full(ctx))
8174 mask |= EPOLLOUT | EPOLLWRNORM;
8175 if (io_cqring_events(ctx, false))
8176 mask |= EPOLLIN | EPOLLRDNORM;
8181 static int io_uring_fasync(int fd, struct file *file, int on)
8183 struct io_ring_ctx *ctx = file->private_data;
8185 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8188 static int io_remove_personalities(int id, void *p, void *data)
8190 struct io_ring_ctx *ctx = data;
8191 const struct cred *cred;
8193 cred = idr_remove(&ctx->personality_idr, id);
8199 static void io_ring_exit_work(struct work_struct *work)
8201 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8205 * If we're doing polled IO and end up having requests being
8206 * submitted async (out-of-line), then completions can come in while
8207 * we're waiting for refs to drop. We need to reap these manually,
8208 * as nobody else will be looking for them.
8212 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8213 io_iopoll_try_reap_events(ctx);
8214 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8215 io_ring_ctx_free(ctx);
8218 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8220 mutex_lock(&ctx->uring_lock);
8221 percpu_ref_kill(&ctx->refs);
8222 mutex_unlock(&ctx->uring_lock);
8224 io_kill_timeouts(ctx, NULL);
8225 io_poll_remove_all(ctx, NULL);
8228 io_wq_cancel_all(ctx->io_wq);
8230 /* if we failed setting up the ctx, we might not have any rings */
8232 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8233 io_iopoll_try_reap_events(ctx);
8234 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8237 * Do this upfront, so we won't have a grace period where the ring
8238 * is closed but resources aren't reaped yet. This can cause
8239 * spurious failure in setting up a new ring.
8241 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8244 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8246 * Use system_unbound_wq to avoid spawning tons of event kworkers
8247 * if we're exiting a ton of rings at the same time. It just adds
8248 * noise and overhead, there's no discernable change in runtime
8249 * over using system_wq.
8251 queue_work(system_unbound_wq, &ctx->exit_work);
8254 static int io_uring_release(struct inode *inode, struct file *file)
8256 struct io_ring_ctx *ctx = file->private_data;
8258 file->private_data = NULL;
8259 io_ring_ctx_wait_and_kill(ctx);
8263 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8265 struct files_struct *files = data;
8267 return !files || work->files == files;
8271 * Returns true if 'preq' is the link parent of 'req'
8273 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8275 struct io_kiocb *link;
8277 if (!(preq->flags & REQ_F_LINK_HEAD))
8280 list_for_each_entry(link, &preq->link_list, link_list) {
8288 static bool io_match_link_files(struct io_kiocb *req,
8289 struct files_struct *files)
8291 struct io_kiocb *link;
8293 if (io_match_files(req, files))
8295 if (req->flags & REQ_F_LINK_HEAD) {
8296 list_for_each_entry(link, &req->link_list, link_list) {
8297 if (io_match_files(link, files))
8305 * We're looking to cancel 'req' because it's holding on to our files, but
8306 * 'req' could be a link to another request. See if it is, and cancel that
8307 * parent request if so.
8309 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8311 struct hlist_node *tmp;
8312 struct io_kiocb *preq;
8316 spin_lock_irq(&ctx->completion_lock);
8317 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8318 struct hlist_head *list;
8320 list = &ctx->cancel_hash[i];
8321 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8322 found = io_match_link(preq, req);
8324 io_poll_remove_one(preq);
8329 spin_unlock_irq(&ctx->completion_lock);
8333 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8334 struct io_kiocb *req)
8336 struct io_kiocb *preq;
8339 spin_lock_irq(&ctx->completion_lock);
8340 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8341 found = io_match_link(preq, req);
8343 __io_timeout_cancel(preq);
8347 spin_unlock_irq(&ctx->completion_lock);
8351 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8353 return io_match_link(container_of(work, struct io_kiocb, work), data);
8356 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8358 enum io_wq_cancel cret;
8360 /* cancel this particular work, if it's running */
8361 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8362 if (cret != IO_WQ_CANCEL_NOTFOUND)
8365 /* find links that hold this pending, cancel those */
8366 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8367 if (cret != IO_WQ_CANCEL_NOTFOUND)
8370 /* if we have a poll link holding this pending, cancel that */
8371 if (io_poll_remove_link(ctx, req))
8374 /* final option, timeout link is holding this req pending */
8375 io_timeout_remove_link(ctx, req);
8378 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8379 struct files_struct *files)
8381 struct io_defer_entry *de = NULL;
8384 spin_lock_irq(&ctx->completion_lock);
8385 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8386 if (io_match_link_files(de->req, files)) {
8387 list_cut_position(&list, &ctx->defer_list, &de->list);
8391 spin_unlock_irq(&ctx->completion_lock);
8393 while (!list_empty(&list)) {
8394 de = list_first_entry(&list, struct io_defer_entry, list);
8395 list_del_init(&de->list);
8396 req_set_fail_links(de->req);
8397 io_put_req(de->req);
8398 io_req_complete(de->req, -ECANCELED);
8404 * Returns true if we found and killed one or more files pinning requests
8406 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8407 struct files_struct *files)
8409 if (list_empty_careful(&ctx->inflight_list))
8412 io_cancel_defer_files(ctx, files);
8413 /* cancel all at once, should be faster than doing it one by one*/
8414 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8416 while (!list_empty_careful(&ctx->inflight_list)) {
8417 struct io_kiocb *cancel_req = NULL, *req;
8420 spin_lock_irq(&ctx->inflight_lock);
8421 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8422 if (files && req->work.files != files)
8424 /* req is being completed, ignore */
8425 if (!refcount_inc_not_zero(&req->refs))
8431 prepare_to_wait(&ctx->inflight_wait, &wait,
8432 TASK_UNINTERRUPTIBLE);
8433 spin_unlock_irq(&ctx->inflight_lock);
8435 /* We need to keep going until we don't find a matching req */
8438 /* cancel this request, or head link requests */
8439 io_attempt_cancel(ctx, cancel_req);
8440 io_put_req(cancel_req);
8441 /* cancellations _may_ trigger task work */
8444 finish_wait(&ctx->inflight_wait, &wait);
8450 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8452 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8453 struct task_struct *task = data;
8455 return io_task_match(req, task);
8458 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8459 struct task_struct *task,
8460 struct files_struct *files)
8464 ret = io_uring_cancel_files(ctx, files);
8466 enum io_wq_cancel cret;
8468 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8469 if (cret != IO_WQ_CANCEL_NOTFOUND)
8472 /* SQPOLL thread does its own polling */
8473 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8474 while (!list_empty_careful(&ctx->iopoll_list)) {
8475 io_iopoll_try_reap_events(ctx);
8480 ret |= io_poll_remove_all(ctx, task);
8481 ret |= io_kill_timeouts(ctx, task);
8488 * We need to iteratively cancel requests, in case a request has dependent
8489 * hard links. These persist even for failure of cancelations, hence keep
8490 * looping until none are found.
8492 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8493 struct files_struct *files)
8495 struct task_struct *task = current;
8497 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8498 task = ctx->sq_data->thread;
8500 io_cqring_overflow_flush(ctx, true, task, files);
8502 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8509 * Note that this task has used io_uring. We use it for cancelation purposes.
8511 static int io_uring_add_task_file(struct file *file)
8513 struct io_uring_task *tctx = current->io_uring;
8515 if (unlikely(!tctx)) {
8518 ret = io_uring_alloc_task_context(current);
8521 tctx = current->io_uring;
8523 if (tctx->last != file) {
8524 void *old = xa_load(&tctx->xa, (unsigned long)file);
8528 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8537 * Remove this io_uring_file -> task mapping.
8539 static void io_uring_del_task_file(struct file *file)
8541 struct io_uring_task *tctx = current->io_uring;
8543 if (tctx->last == file)
8545 file = xa_erase(&tctx->xa, (unsigned long)file);
8550 static void __io_uring_attempt_task_drop(struct file *file)
8552 struct file *old = xa_load(¤t->io_uring->xa, (unsigned long)file);
8555 io_uring_del_task_file(file);
8559 * Drop task note for this file if we're the only ones that hold it after
8562 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8564 if (!current->io_uring)
8567 * fput() is pending, will be 2 if the only other ref is our potential
8568 * task file note. If the task is exiting, drop regardless of count.
8570 if (!exiting && atomic_long_read(&file->f_count) != 2)
8573 __io_uring_attempt_task_drop(file);
8576 void __io_uring_files_cancel(struct files_struct *files)
8578 struct io_uring_task *tctx = current->io_uring;
8580 unsigned long index;
8582 /* make sure overflow events are dropped */
8583 tctx->in_idle = true;
8585 xa_for_each(&tctx->xa, index, file) {
8586 struct io_ring_ctx *ctx = file->private_data;
8588 io_uring_cancel_task_requests(ctx, files);
8590 io_uring_del_task_file(file);
8594 static inline bool io_uring_task_idle(struct io_uring_task *tctx)
8596 return atomic_long_read(&tctx->req_issue) ==
8597 atomic_long_read(&tctx->req_complete);
8601 * Find any io_uring fd that this task has registered or done IO on, and cancel
8604 void __io_uring_task_cancel(void)
8606 struct io_uring_task *tctx = current->io_uring;
8610 /* make sure overflow events are dropped */
8611 tctx->in_idle = true;
8613 while (!io_uring_task_idle(tctx)) {
8614 /* read completions before cancelations */
8615 completions = atomic_long_read(&tctx->req_complete);
8616 __io_uring_files_cancel(NULL);
8618 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8621 * If we've seen completions, retry. This avoids a race where
8622 * a completion comes in before we did prepare_to_wait().
8624 if (completions != atomic_long_read(&tctx->req_complete))
8626 if (io_uring_task_idle(tctx))
8631 finish_wait(&tctx->wait, &wait);
8632 tctx->in_idle = false;
8635 static int io_uring_flush(struct file *file, void *data)
8637 struct io_ring_ctx *ctx = file->private_data;
8640 * If the task is going away, cancel work it may have pending
8642 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8645 io_uring_cancel_task_requests(ctx, data);
8646 io_uring_attempt_task_drop(file, !data);
8650 static void *io_uring_validate_mmap_request(struct file *file,
8651 loff_t pgoff, size_t sz)
8653 struct io_ring_ctx *ctx = file->private_data;
8654 loff_t offset = pgoff << PAGE_SHIFT;
8659 case IORING_OFF_SQ_RING:
8660 case IORING_OFF_CQ_RING:
8663 case IORING_OFF_SQES:
8667 return ERR_PTR(-EINVAL);
8670 page = virt_to_head_page(ptr);
8671 if (sz > page_size(page))
8672 return ERR_PTR(-EINVAL);
8679 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8681 size_t sz = vma->vm_end - vma->vm_start;
8685 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8687 return PTR_ERR(ptr);
8689 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8690 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8693 #else /* !CONFIG_MMU */
8695 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8697 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8700 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8702 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8705 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8706 unsigned long addr, unsigned long len,
8707 unsigned long pgoff, unsigned long flags)
8711 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8713 return PTR_ERR(ptr);
8715 return (unsigned long) ptr;
8718 #endif /* !CONFIG_MMU */
8720 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8725 if (!io_sqring_full(ctx))
8728 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8730 if (!io_sqring_full(ctx))
8734 } while (!signal_pending(current));
8736 finish_wait(&ctx->sqo_sq_wait, &wait);
8739 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8740 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8743 struct io_ring_ctx *ctx;
8750 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8751 IORING_ENTER_SQ_WAIT))
8759 if (f.file->f_op != &io_uring_fops)
8763 ctx = f.file->private_data;
8764 if (!percpu_ref_tryget(&ctx->refs))
8768 if (ctx->flags & IORING_SETUP_R_DISABLED)
8772 * For SQ polling, the thread will do all submissions and completions.
8773 * Just return the requested submit count, and wake the thread if
8777 if (ctx->flags & IORING_SETUP_SQPOLL) {
8778 if (!list_empty_careful(&ctx->cq_overflow_list))
8779 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8780 if (flags & IORING_ENTER_SQ_WAKEUP)
8781 wake_up(&ctx->sq_data->wait);
8782 if (flags & IORING_ENTER_SQ_WAIT)
8783 io_sqpoll_wait_sq(ctx);
8784 submitted = to_submit;
8785 } else if (to_submit) {
8786 ret = io_uring_add_task_file(f.file);
8789 mutex_lock(&ctx->uring_lock);
8790 submitted = io_submit_sqes(ctx, to_submit);
8791 mutex_unlock(&ctx->uring_lock);
8793 if (submitted != to_submit)
8796 if (flags & IORING_ENTER_GETEVENTS) {
8797 min_complete = min(min_complete, ctx->cq_entries);
8800 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8801 * space applications don't need to do io completion events
8802 * polling again, they can rely on io_sq_thread to do polling
8803 * work, which can reduce cpu usage and uring_lock contention.
8805 if (ctx->flags & IORING_SETUP_IOPOLL &&
8806 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8807 ret = io_iopoll_check(ctx, min_complete);
8809 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8814 percpu_ref_put(&ctx->refs);
8817 return submitted ? submitted : ret;
8820 #ifdef CONFIG_PROC_FS
8821 static int io_uring_show_cred(int id, void *p, void *data)
8823 const struct cred *cred = p;
8824 struct seq_file *m = data;
8825 struct user_namespace *uns = seq_user_ns(m);
8826 struct group_info *gi;
8831 seq_printf(m, "%5d\n", id);
8832 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8833 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8834 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8835 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8836 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8837 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8838 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8839 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8840 seq_puts(m, "\n\tGroups:\t");
8841 gi = cred->group_info;
8842 for (g = 0; g < gi->ngroups; g++) {
8843 seq_put_decimal_ull(m, g ? " " : "",
8844 from_kgid_munged(uns, gi->gid[g]));
8846 seq_puts(m, "\n\tCapEff:\t");
8847 cap = cred->cap_effective;
8848 CAP_FOR_EACH_U32(__capi)
8849 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8854 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8856 struct io_sq_data *sq = NULL;
8861 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8862 * since fdinfo case grabs it in the opposite direction of normal use
8863 * cases. If we fail to get the lock, we just don't iterate any
8864 * structures that could be going away outside the io_uring mutex.
8866 has_lock = mutex_trylock(&ctx->uring_lock);
8868 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
8871 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
8872 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
8873 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8874 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8875 struct fixed_file_table *table;
8878 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8879 f = table->files[i & IORING_FILE_TABLE_MASK];
8881 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8883 seq_printf(m, "%5u: <none>\n", i);
8885 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8886 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8887 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8889 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8890 (unsigned int) buf->len);
8892 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8893 seq_printf(m, "Personalities:\n");
8894 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8896 seq_printf(m, "PollList:\n");
8897 spin_lock_irq(&ctx->completion_lock);
8898 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8899 struct hlist_head *list = &ctx->cancel_hash[i];
8900 struct io_kiocb *req;
8902 hlist_for_each_entry(req, list, hash_node)
8903 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8904 req->task->task_works != NULL);
8906 spin_unlock_irq(&ctx->completion_lock);
8908 mutex_unlock(&ctx->uring_lock);
8911 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8913 struct io_ring_ctx *ctx = f->private_data;
8915 if (percpu_ref_tryget(&ctx->refs)) {
8916 __io_uring_show_fdinfo(ctx, m);
8917 percpu_ref_put(&ctx->refs);
8922 static const struct file_operations io_uring_fops = {
8923 .release = io_uring_release,
8924 .flush = io_uring_flush,
8925 .mmap = io_uring_mmap,
8927 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8928 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8930 .poll = io_uring_poll,
8931 .fasync = io_uring_fasync,
8932 #ifdef CONFIG_PROC_FS
8933 .show_fdinfo = io_uring_show_fdinfo,
8937 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8938 struct io_uring_params *p)
8940 struct io_rings *rings;
8941 size_t size, sq_array_offset;
8943 /* make sure these are sane, as we already accounted them */
8944 ctx->sq_entries = p->sq_entries;
8945 ctx->cq_entries = p->cq_entries;
8947 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8948 if (size == SIZE_MAX)
8951 rings = io_mem_alloc(size);
8956 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8957 rings->sq_ring_mask = p->sq_entries - 1;
8958 rings->cq_ring_mask = p->cq_entries - 1;
8959 rings->sq_ring_entries = p->sq_entries;
8960 rings->cq_ring_entries = p->cq_entries;
8961 ctx->sq_mask = rings->sq_ring_mask;
8962 ctx->cq_mask = rings->cq_ring_mask;
8964 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8965 if (size == SIZE_MAX) {
8966 io_mem_free(ctx->rings);
8971 ctx->sq_sqes = io_mem_alloc(size);
8972 if (!ctx->sq_sqes) {
8973 io_mem_free(ctx->rings);
8982 * Allocate an anonymous fd, this is what constitutes the application
8983 * visible backing of an io_uring instance. The application mmaps this
8984 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8985 * we have to tie this fd to a socket for file garbage collection purposes.
8987 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8992 #if defined(CONFIG_UNIX)
8993 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8999 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9003 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9004 O_RDWR | O_CLOEXEC);
9008 ret = PTR_ERR(file);
9012 #if defined(CONFIG_UNIX)
9013 ctx->ring_sock->file = file;
9015 if (unlikely(io_uring_add_task_file(file))) {
9016 file = ERR_PTR(-ENOMEM);
9019 fd_install(ret, file);
9022 #if defined(CONFIG_UNIX)
9023 sock_release(ctx->ring_sock);
9024 ctx->ring_sock = NULL;
9029 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9030 struct io_uring_params __user *params)
9032 struct user_struct *user = NULL;
9033 struct io_ring_ctx *ctx;
9039 if (entries > IORING_MAX_ENTRIES) {
9040 if (!(p->flags & IORING_SETUP_CLAMP))
9042 entries = IORING_MAX_ENTRIES;
9046 * Use twice as many entries for the CQ ring. It's possible for the
9047 * application to drive a higher depth than the size of the SQ ring,
9048 * since the sqes are only used at submission time. This allows for
9049 * some flexibility in overcommitting a bit. If the application has
9050 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9051 * of CQ ring entries manually.
9053 p->sq_entries = roundup_pow_of_two(entries);
9054 if (p->flags & IORING_SETUP_CQSIZE) {
9056 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9057 * to a power-of-two, if it isn't already. We do NOT impose
9058 * any cq vs sq ring sizing.
9060 if (p->cq_entries < p->sq_entries)
9062 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9063 if (!(p->flags & IORING_SETUP_CLAMP))
9065 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9067 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9069 p->cq_entries = 2 * p->sq_entries;
9072 user = get_uid(current_user());
9073 limit_mem = !capable(CAP_IPC_LOCK);
9076 ret = __io_account_mem(user,
9077 ring_pages(p->sq_entries, p->cq_entries));
9084 ctx = io_ring_ctx_alloc(p);
9087 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9092 ctx->compat = in_compat_syscall();
9094 ctx->creds = get_current_cred();
9096 ctx->sqo_task = get_task_struct(current);
9099 * This is just grabbed for accounting purposes. When a process exits,
9100 * the mm is exited and dropped before the files, hence we need to hang
9101 * on to this mm purely for the purposes of being able to unaccount
9102 * memory (locked/pinned vm). It's not used for anything else.
9104 mmgrab(current->mm);
9105 ctx->mm_account = current->mm;
9107 #ifdef CONFIG_BLK_CGROUP
9109 * The sq thread will belong to the original cgroup it was inited in.
9110 * If the cgroup goes offline (e.g. disabling the io controller), then
9111 * issued bios will be associated with the closest cgroup later in the
9115 ctx->sqo_blkcg_css = blkcg_css();
9116 ret = css_tryget_online(ctx->sqo_blkcg_css);
9119 /* don't init against a dying cgroup, have the user try again */
9120 ctx->sqo_blkcg_css = NULL;
9127 * Account memory _before_ installing the file descriptor. Once
9128 * the descriptor is installed, it can get closed at any time. Also
9129 * do this before hitting the general error path, as ring freeing
9130 * will un-account as well.
9132 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9134 ctx->limit_mem = limit_mem;
9136 ret = io_allocate_scq_urings(ctx, p);
9140 ret = io_sq_offload_create(ctx, p);
9144 if (!(p->flags & IORING_SETUP_R_DISABLED))
9145 io_sq_offload_start(ctx);
9147 memset(&p->sq_off, 0, sizeof(p->sq_off));
9148 p->sq_off.head = offsetof(struct io_rings, sq.head);
9149 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9150 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9151 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9152 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9153 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9154 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9156 memset(&p->cq_off, 0, sizeof(p->cq_off));
9157 p->cq_off.head = offsetof(struct io_rings, cq.head);
9158 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9159 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9160 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9161 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9162 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9163 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9165 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9166 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9167 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9168 IORING_FEAT_POLL_32BITS;
9170 if (copy_to_user(params, p, sizeof(*p))) {
9176 * Install ring fd as the very last thing, so we don't risk someone
9177 * having closed it before we finish setup
9179 ret = io_uring_get_fd(ctx);
9183 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9186 io_ring_ctx_wait_and_kill(ctx);
9191 * Sets up an aio uring context, and returns the fd. Applications asks for a
9192 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9193 * params structure passed in.
9195 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9197 struct io_uring_params p;
9200 if (copy_from_user(&p, params, sizeof(p)))
9202 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9207 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9208 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9209 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9210 IORING_SETUP_R_DISABLED))
9213 return io_uring_create(entries, &p, params);
9216 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9217 struct io_uring_params __user *, params)
9219 return io_uring_setup(entries, params);
9222 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9224 struct io_uring_probe *p;
9228 size = struct_size(p, ops, nr_args);
9229 if (size == SIZE_MAX)
9231 p = kzalloc(size, GFP_KERNEL);
9236 if (copy_from_user(p, arg, size))
9239 if (memchr_inv(p, 0, size))
9242 p->last_op = IORING_OP_LAST - 1;
9243 if (nr_args > IORING_OP_LAST)
9244 nr_args = IORING_OP_LAST;
9246 for (i = 0; i < nr_args; i++) {
9248 if (!io_op_defs[i].not_supported)
9249 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9254 if (copy_to_user(arg, p, size))
9261 static int io_register_personality(struct io_ring_ctx *ctx)
9263 const struct cred *creds = get_current_cred();
9266 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9267 USHRT_MAX, GFP_KERNEL);
9273 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9275 const struct cred *old_creds;
9277 old_creds = idr_remove(&ctx->personality_idr, id);
9279 put_cred(old_creds);
9286 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9287 unsigned int nr_args)
9289 struct io_uring_restriction *res;
9293 /* Restrictions allowed only if rings started disabled */
9294 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9297 /* We allow only a single restrictions registration */
9298 if (ctx->restrictions.registered)
9301 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9304 size = array_size(nr_args, sizeof(*res));
9305 if (size == SIZE_MAX)
9308 res = memdup_user(arg, size);
9310 return PTR_ERR(res);
9314 for (i = 0; i < nr_args; i++) {
9315 switch (res[i].opcode) {
9316 case IORING_RESTRICTION_REGISTER_OP:
9317 if (res[i].register_op >= IORING_REGISTER_LAST) {
9322 __set_bit(res[i].register_op,
9323 ctx->restrictions.register_op);
9325 case IORING_RESTRICTION_SQE_OP:
9326 if (res[i].sqe_op >= IORING_OP_LAST) {
9331 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9333 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9334 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9336 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9337 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9346 /* Reset all restrictions if an error happened */
9348 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9350 ctx->restrictions.registered = true;
9356 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9358 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9361 if (ctx->restrictions.registered)
9362 ctx->restricted = 1;
9364 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9366 io_sq_offload_start(ctx);
9371 static bool io_register_op_must_quiesce(int op)
9374 case IORING_UNREGISTER_FILES:
9375 case IORING_REGISTER_FILES_UPDATE:
9376 case IORING_REGISTER_PROBE:
9377 case IORING_REGISTER_PERSONALITY:
9378 case IORING_UNREGISTER_PERSONALITY:
9385 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9386 void __user *arg, unsigned nr_args)
9387 __releases(ctx->uring_lock)
9388 __acquires(ctx->uring_lock)
9393 * We're inside the ring mutex, if the ref is already dying, then
9394 * someone else killed the ctx or is already going through
9395 * io_uring_register().
9397 if (percpu_ref_is_dying(&ctx->refs))
9400 if (io_register_op_must_quiesce(opcode)) {
9401 percpu_ref_kill(&ctx->refs);
9404 * Drop uring mutex before waiting for references to exit. If
9405 * another thread is currently inside io_uring_enter() it might
9406 * need to grab the uring_lock to make progress. If we hold it
9407 * here across the drain wait, then we can deadlock. It's safe
9408 * to drop the mutex here, since no new references will come in
9409 * after we've killed the percpu ref.
9411 mutex_unlock(&ctx->uring_lock);
9413 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9416 ret = io_run_task_work_sig();
9421 mutex_lock(&ctx->uring_lock);
9424 percpu_ref_resurrect(&ctx->refs);
9429 if (ctx->restricted) {
9430 if (opcode >= IORING_REGISTER_LAST) {
9435 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9442 case IORING_REGISTER_BUFFERS:
9443 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9445 case IORING_UNREGISTER_BUFFERS:
9449 ret = io_sqe_buffer_unregister(ctx);
9451 case IORING_REGISTER_FILES:
9452 ret = io_sqe_files_register(ctx, arg, nr_args);
9454 case IORING_UNREGISTER_FILES:
9458 ret = io_sqe_files_unregister(ctx);
9460 case IORING_REGISTER_FILES_UPDATE:
9461 ret = io_sqe_files_update(ctx, arg, nr_args);
9463 case IORING_REGISTER_EVENTFD:
9464 case IORING_REGISTER_EVENTFD_ASYNC:
9468 ret = io_eventfd_register(ctx, arg);
9471 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9472 ctx->eventfd_async = 1;
9474 ctx->eventfd_async = 0;
9476 case IORING_UNREGISTER_EVENTFD:
9480 ret = io_eventfd_unregister(ctx);
9482 case IORING_REGISTER_PROBE:
9484 if (!arg || nr_args > 256)
9486 ret = io_probe(ctx, arg, nr_args);
9488 case IORING_REGISTER_PERSONALITY:
9492 ret = io_register_personality(ctx);
9494 case IORING_UNREGISTER_PERSONALITY:
9498 ret = io_unregister_personality(ctx, nr_args);
9500 case IORING_REGISTER_ENABLE_RINGS:
9504 ret = io_register_enable_rings(ctx);
9506 case IORING_REGISTER_RESTRICTIONS:
9507 ret = io_register_restrictions(ctx, arg, nr_args);
9515 if (io_register_op_must_quiesce(opcode)) {
9516 /* bring the ctx back to life */
9517 percpu_ref_reinit(&ctx->refs);
9519 reinit_completion(&ctx->ref_comp);
9524 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9525 void __user *, arg, unsigned int, nr_args)
9527 struct io_ring_ctx *ctx;
9536 if (f.file->f_op != &io_uring_fops)
9539 ctx = f.file->private_data;
9541 mutex_lock(&ctx->uring_lock);
9542 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9543 mutex_unlock(&ctx->uring_lock);
9544 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9545 ctx->cq_ev_fd != NULL, ret);
9551 static int __init io_uring_init(void)
9553 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9554 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9555 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9558 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9559 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9560 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9561 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9562 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9563 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9564 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9565 BUILD_BUG_SQE_ELEM(8, __u64, off);
9566 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9567 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9568 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9569 BUILD_BUG_SQE_ELEM(24, __u32, len);
9570 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9571 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9572 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9573 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9574 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9575 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9576 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9577 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9578 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9579 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9580 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9581 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9582 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9583 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9584 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9585 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9586 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9587 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9588 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9590 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9591 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9592 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9595 __initcall(io_uring_init);