1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/io_uring.h>
88 #include <uapi/linux/io_uring.h>
93 #define IORING_MAX_ENTRIES 32768
94 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
97 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
99 #define IORING_FILE_TABLE_SHIFT 9
100 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
101 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
102 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
104 IORING_REGISTER_LAST + IORING_OP_LAST)
107 u32 head ____cacheline_aligned_in_smp;
108 u32 tail ____cacheline_aligned_in_smp;
112 * This data is shared with the application through the mmap at offsets
113 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
115 * The offsets to the member fields are published through struct
116 * io_sqring_offsets when calling io_uring_setup.
120 * Head and tail offsets into the ring; the offsets need to be
121 * masked to get valid indices.
123 * The kernel controls head of the sq ring and the tail of the cq ring,
124 * and the application controls tail of the sq ring and the head of the
127 struct io_uring sq, cq;
129 * Bitmasks to apply to head and tail offsets (constant, equals
132 u32 sq_ring_mask, cq_ring_mask;
133 /* Ring sizes (constant, power of 2) */
134 u32 sq_ring_entries, cq_ring_entries;
136 * Number of invalid entries dropped by the kernel due to
137 * invalid index stored in array
139 * Written by the kernel, shouldn't be modified by the
140 * application (i.e. get number of "new events" by comparing to
143 * After a new SQ head value was read by the application this
144 * counter includes all submissions that were dropped reaching
145 * the new SQ head (and possibly more).
151 * Written by the kernel, shouldn't be modified by the
154 * The application needs a full memory barrier before checking
155 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
161 * Written by the application, shouldn't be modified by the
166 * Number of completion events lost because the queue was full;
167 * this should be avoided by the application by making sure
168 * there are not more requests pending than there is space in
169 * the completion queue.
171 * Written by the kernel, shouldn't be modified by the
172 * application (i.e. get number of "new events" by comparing to
175 * As completion events come in out of order this counter is not
176 * ordered with any other data.
180 * Ring buffer of completion events.
182 * The kernel writes completion events fresh every time they are
183 * produced, so the application is allowed to modify pending
186 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
189 struct io_mapped_ubuf {
192 struct bio_vec *bvec;
193 unsigned int nr_bvecs;
194 unsigned long acct_pages;
197 struct fixed_file_table {
201 struct fixed_file_ref_node {
202 struct percpu_ref refs;
203 struct list_head node;
204 struct list_head file_list;
205 struct fixed_file_data *file_data;
206 struct llist_node llist;
209 struct fixed_file_data {
210 struct fixed_file_table *table;
211 struct io_ring_ctx *ctx;
213 struct percpu_ref *cur_refs;
214 struct percpu_ref refs;
215 struct completion done;
216 struct list_head ref_list;
221 struct list_head list;
227 struct io_restriction {
228 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
229 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
230 u8 sqe_flags_allowed;
231 u8 sqe_flags_required;
239 /* ctx's that are using this sqd */
240 struct list_head ctx_list;
241 struct list_head ctx_new_list;
242 struct mutex ctx_lock;
244 struct task_struct *thread;
245 struct wait_queue_head wait;
250 struct percpu_ref refs;
251 } ____cacheline_aligned_in_smp;
255 unsigned int compat: 1;
256 unsigned int limit_mem: 1;
257 unsigned int cq_overflow_flushed: 1;
258 unsigned int drain_next: 1;
259 unsigned int eventfd_async: 1;
260 unsigned int restricted: 1;
263 * Ring buffer of indices into array of io_uring_sqe, which is
264 * mmapped by the application using the IORING_OFF_SQES offset.
266 * This indirection could e.g. be used to assign fixed
267 * io_uring_sqe entries to operations and only submit them to
268 * the queue when needed.
270 * The kernel modifies neither the indices array nor the entries
274 unsigned cached_sq_head;
277 unsigned sq_thread_idle;
278 unsigned cached_sq_dropped;
279 atomic_t cached_cq_overflow;
280 unsigned long sq_check_overflow;
282 struct list_head defer_list;
283 struct list_head timeout_list;
284 struct list_head cq_overflow_list;
286 wait_queue_head_t inflight_wait;
287 struct io_uring_sqe *sq_sqes;
288 } ____cacheline_aligned_in_smp;
290 struct io_rings *rings;
296 * For SQPOLL usage - we hold a reference to the parent task, so we
297 * have access to the ->files
299 struct task_struct *sqo_task;
301 /* Only used for accounting purposes */
302 struct mm_struct *mm_account;
304 #ifdef CONFIG_BLK_CGROUP
305 struct cgroup_subsys_state *sqo_blkcg_css;
308 struct io_sq_data *sq_data; /* if using sq thread polling */
310 struct wait_queue_head sqo_sq_wait;
311 struct wait_queue_entry sqo_wait_entry;
312 struct list_head sqd_list;
315 * If used, fixed file set. Writers must ensure that ->refs is dead,
316 * readers must ensure that ->refs is alive as long as the file* is
317 * used. Only updated through io_uring_register(2).
319 struct fixed_file_data *file_data;
320 unsigned nr_user_files;
322 /* if used, fixed mapped user buffers */
323 unsigned nr_user_bufs;
324 struct io_mapped_ubuf *user_bufs;
326 struct user_struct *user;
328 const struct cred *creds;
330 struct completion ref_comp;
331 struct completion sq_thread_comp;
333 /* if all else fails... */
334 struct io_kiocb *fallback_req;
336 #if defined(CONFIG_UNIX)
337 struct socket *ring_sock;
340 struct idr io_buffer_idr;
342 struct idr personality_idr;
345 unsigned cached_cq_tail;
348 atomic_t cq_timeouts;
349 unsigned long cq_check_overflow;
350 struct wait_queue_head cq_wait;
351 struct fasync_struct *cq_fasync;
352 struct eventfd_ctx *cq_ev_fd;
353 } ____cacheline_aligned_in_smp;
356 struct mutex uring_lock;
357 wait_queue_head_t wait;
358 } ____cacheline_aligned_in_smp;
361 spinlock_t completion_lock;
364 * ->iopoll_list is protected by the ctx->uring_lock for
365 * io_uring instances that don't use IORING_SETUP_SQPOLL.
366 * For SQPOLL, only the single threaded io_sq_thread() will
367 * manipulate the list, hence no extra locking is needed there.
369 struct list_head iopoll_list;
370 struct hlist_head *cancel_hash;
371 unsigned cancel_hash_bits;
372 bool poll_multi_file;
374 spinlock_t inflight_lock;
375 struct list_head inflight_list;
376 } ____cacheline_aligned_in_smp;
378 struct delayed_work file_put_work;
379 struct llist_head file_put_llist;
381 struct work_struct exit_work;
382 struct io_restriction restrictions;
386 * First field must be the file pointer in all the
387 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
389 struct io_poll_iocb {
392 struct wait_queue_head *head;
398 struct wait_queue_entry wait;
403 struct file *put_file;
407 struct io_timeout_data {
408 struct io_kiocb *req;
409 struct hrtimer timer;
410 struct timespec64 ts;
411 enum hrtimer_mode mode;
416 struct sockaddr __user *addr;
417 int __user *addr_len;
419 unsigned long nofile;
441 struct list_head list;
445 /* NOTE: kiocb has the file as the first member, so don't do it here */
453 struct sockaddr __user *addr;
460 struct user_msghdr __user *umsg;
466 struct io_buffer *kbuf;
472 struct filename *filename;
474 unsigned long nofile;
477 struct io_files_update {
503 struct epoll_event event;
507 struct file *file_out;
508 struct file *file_in;
515 struct io_provide_buf {
529 const char __user *filename;
530 struct statx __user *buffer;
533 struct io_completion {
535 struct list_head list;
539 struct io_async_connect {
540 struct sockaddr_storage address;
543 struct io_async_msghdr {
544 struct iovec fast_iov[UIO_FASTIOV];
546 struct sockaddr __user *uaddr;
548 struct sockaddr_storage addr;
552 struct iovec fast_iov[UIO_FASTIOV];
553 const struct iovec *free_iovec;
554 struct iov_iter iter;
556 struct wait_page_queue wpq;
560 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
561 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
562 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
563 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
564 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
565 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
572 REQ_F_LINK_TIMEOUT_BIT,
574 REQ_F_COMP_LOCKED_BIT,
575 REQ_F_NEED_CLEANUP_BIT,
577 REQ_F_BUFFER_SELECTED_BIT,
578 REQ_F_NO_FILE_TABLE_BIT,
579 REQ_F_WORK_INITIALIZED_BIT,
581 /* not a real bit, just to check we're not overflowing the space */
587 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
588 /* drain existing IO first */
589 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
591 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
592 /* doesn't sever on completion < 0 */
593 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
595 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
596 /* IOSQE_BUFFER_SELECT */
597 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
600 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
601 /* fail rest of links */
602 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
603 /* on inflight list */
604 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
605 /* read/write uses file position */
606 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
607 /* must not punt to workers */
608 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
609 /* has linked timeout */
610 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
612 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
613 /* completion under lock */
614 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
616 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
617 /* already went through poll handler */
618 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
619 /* buffer already selected */
620 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
621 /* doesn't need file table for this request */
622 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
623 /* io_wq_work is initialized */
624 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
628 struct io_poll_iocb poll;
629 struct io_poll_iocb *double_poll;
633 * NOTE! Each of the iocb union members has the file pointer
634 * as the first entry in their struct definition. So you can
635 * access the file pointer through any of the sub-structs,
636 * or directly as just 'ki_filp' in this struct.
642 struct io_poll_iocb poll;
643 struct io_accept accept;
645 struct io_cancel cancel;
646 struct io_timeout timeout;
647 struct io_connect connect;
648 struct io_sr_msg sr_msg;
650 struct io_close close;
651 struct io_files_update files_update;
652 struct io_fadvise fadvise;
653 struct io_madvise madvise;
654 struct io_epoll epoll;
655 struct io_splice splice;
656 struct io_provide_buf pbuf;
657 struct io_statx statx;
658 /* use only after cleaning per-op data, see io_clean_op() */
659 struct io_completion compl;
662 /* opcode allocated if it needs to store data for async defer */
665 /* polled IO has completed */
671 struct io_ring_ctx *ctx;
674 struct task_struct *task;
677 struct list_head link_list;
680 * 1. used with ctx->iopoll_list with reads/writes
681 * 2. to track reqs with ->files (see io_op_def::file_table)
683 struct list_head inflight_entry;
685 struct percpu_ref *fixed_file_refs;
686 struct callback_head task_work;
687 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
688 struct hlist_node hash_node;
689 struct async_poll *apoll;
690 struct io_wq_work work;
693 struct io_defer_entry {
694 struct list_head list;
695 struct io_kiocb *req;
699 #define IO_IOPOLL_BATCH 8
701 struct io_comp_state {
703 struct list_head list;
704 struct io_ring_ctx *ctx;
707 struct io_submit_state {
708 struct blk_plug plug;
711 * io_kiocb alloc cache
713 void *reqs[IO_IOPOLL_BATCH];
714 unsigned int free_reqs;
717 * Batch completion logic
719 struct io_comp_state comp;
722 * File reference cache
726 unsigned int has_refs;
727 unsigned int ios_left;
731 /* needs current->mm setup, does mm access */
732 unsigned needs_mm : 1;
733 /* needs req->file assigned */
734 unsigned needs_file : 1;
735 /* don't fail if file grab fails */
736 unsigned needs_file_no_error : 1;
737 /* hash wq insertion if file is a regular file */
738 unsigned hash_reg_file : 1;
739 /* unbound wq insertion if file is a non-regular file */
740 unsigned unbound_nonreg_file : 1;
741 /* opcode is not supported by this kernel */
742 unsigned not_supported : 1;
743 /* needs file table */
744 unsigned file_table : 1;
746 unsigned needs_fs : 1;
747 /* set if opcode supports polled "wait" */
749 unsigned pollout : 1;
750 /* op supports buffer selection */
751 unsigned buffer_select : 1;
752 /* needs rlimit(RLIMIT_FSIZE) assigned */
753 unsigned needs_fsize : 1;
754 /* must always have async data allocated */
755 unsigned needs_async_data : 1;
756 /* needs blkcg context, issues async io potentially */
757 unsigned needs_blkcg : 1;
758 /* size of async data needed, if any */
759 unsigned short async_size;
762 static const struct io_op_def io_op_defs[] __read_mostly = {
763 [IORING_OP_NOP] = {},
764 [IORING_OP_READV] = {
767 .unbound_nonreg_file = 1,
770 .needs_async_data = 1,
772 .async_size = sizeof(struct io_async_rw),
774 [IORING_OP_WRITEV] = {
778 .unbound_nonreg_file = 1,
781 .needs_async_data = 1,
783 .async_size = sizeof(struct io_async_rw),
785 [IORING_OP_FSYNC] = {
789 [IORING_OP_READ_FIXED] = {
791 .unbound_nonreg_file = 1,
794 .async_size = sizeof(struct io_async_rw),
796 [IORING_OP_WRITE_FIXED] = {
799 .unbound_nonreg_file = 1,
803 .async_size = sizeof(struct io_async_rw),
805 [IORING_OP_POLL_ADD] = {
807 .unbound_nonreg_file = 1,
809 [IORING_OP_POLL_REMOVE] = {},
810 [IORING_OP_SYNC_FILE_RANGE] = {
814 [IORING_OP_SENDMSG] = {
817 .unbound_nonreg_file = 1,
820 .needs_async_data = 1,
822 .async_size = sizeof(struct io_async_msghdr),
824 [IORING_OP_RECVMSG] = {
827 .unbound_nonreg_file = 1,
831 .needs_async_data = 1,
833 .async_size = sizeof(struct io_async_msghdr),
835 [IORING_OP_TIMEOUT] = {
837 .needs_async_data = 1,
838 .async_size = sizeof(struct io_timeout_data),
840 [IORING_OP_TIMEOUT_REMOVE] = {},
841 [IORING_OP_ACCEPT] = {
844 .unbound_nonreg_file = 1,
848 [IORING_OP_ASYNC_CANCEL] = {},
849 [IORING_OP_LINK_TIMEOUT] = {
851 .needs_async_data = 1,
852 .async_size = sizeof(struct io_timeout_data),
854 [IORING_OP_CONNECT] = {
857 .unbound_nonreg_file = 1,
859 .needs_async_data = 1,
860 .async_size = sizeof(struct io_async_connect),
862 [IORING_OP_FALLOCATE] = {
867 [IORING_OP_OPENAT] = {
872 [IORING_OP_CLOSE] = {
874 .needs_file_no_error = 1,
878 [IORING_OP_FILES_UPDATE] = {
882 [IORING_OP_STATX] = {
891 .unbound_nonreg_file = 1,
895 .async_size = sizeof(struct io_async_rw),
897 [IORING_OP_WRITE] = {
900 .unbound_nonreg_file = 1,
904 .async_size = sizeof(struct io_async_rw),
906 [IORING_OP_FADVISE] = {
910 [IORING_OP_MADVISE] = {
917 .unbound_nonreg_file = 1,
924 .unbound_nonreg_file = 1,
929 [IORING_OP_OPENAT2] = {
934 [IORING_OP_EPOLL_CTL] = {
935 .unbound_nonreg_file = 1,
938 [IORING_OP_SPLICE] = {
941 .unbound_nonreg_file = 1,
944 [IORING_OP_PROVIDE_BUFFERS] = {},
945 [IORING_OP_REMOVE_BUFFERS] = {},
949 .unbound_nonreg_file = 1,
953 enum io_mem_account {
958 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
959 struct io_comp_state *cs);
960 static void io_cqring_fill_event(struct io_kiocb *req, long res);
961 static void io_put_req(struct io_kiocb *req);
962 static void io_double_put_req(struct io_kiocb *req);
963 static void __io_double_put_req(struct io_kiocb *req);
964 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
965 static void __io_queue_linked_timeout(struct io_kiocb *req);
966 static void io_queue_linked_timeout(struct io_kiocb *req);
967 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
968 struct io_uring_files_update *ip,
970 static int io_prep_work_files(struct io_kiocb *req);
971 static void __io_clean_op(struct io_kiocb *req);
972 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
973 int fd, struct file **out_file, bool fixed);
974 static void __io_queue_sqe(struct io_kiocb *req,
975 const struct io_uring_sqe *sqe,
976 struct io_comp_state *cs);
977 static void io_file_put_work(struct work_struct *work);
979 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
980 struct iovec **iovec, struct iov_iter *iter,
982 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
983 const struct iovec *fast_iov,
984 struct iov_iter *iter, bool force);
986 static struct kmem_cache *req_cachep;
988 static const struct file_operations io_uring_fops __read_mostly;
990 struct sock *io_uring_get_socket(struct file *file)
992 #if defined(CONFIG_UNIX)
993 if (file->f_op == &io_uring_fops) {
994 struct io_ring_ctx *ctx = file->private_data;
996 return ctx->ring_sock->sk;
1001 EXPORT_SYMBOL(io_uring_get_socket);
1003 static inline void io_clean_op(struct io_kiocb *req)
1005 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
1010 static void io_sq_thread_drop_mm(void)
1012 struct mm_struct *mm = current->mm;
1015 kthread_unuse_mm(mm);
1020 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1023 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
1024 !ctx->sqo_task->mm ||
1025 !mmget_not_zero(ctx->sqo_task->mm)))
1027 kthread_use_mm(ctx->sqo_task->mm);
1033 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1034 struct io_kiocb *req)
1036 if (!io_op_defs[req->opcode].needs_mm)
1038 return __io_sq_thread_acquire_mm(ctx);
1041 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1042 struct cgroup_subsys_state **cur_css)
1045 #ifdef CONFIG_BLK_CGROUP
1046 /* puts the old one when swapping */
1047 if (*cur_css != ctx->sqo_blkcg_css) {
1048 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1049 *cur_css = ctx->sqo_blkcg_css;
1054 static void io_sq_thread_unassociate_blkcg(void)
1056 #ifdef CONFIG_BLK_CGROUP
1057 kthread_associate_blkcg(NULL);
1061 static inline void req_set_fail_links(struct io_kiocb *req)
1063 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1064 req->flags |= REQ_F_FAIL_LINK;
1068 * Note: must call io_req_init_async() for the first time you
1069 * touch any members of io_wq_work.
1071 static inline void io_req_init_async(struct io_kiocb *req)
1073 if (req->flags & REQ_F_WORK_INITIALIZED)
1076 memset(&req->work, 0, sizeof(req->work));
1077 req->flags |= REQ_F_WORK_INITIALIZED;
1080 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1082 return ctx->flags & IORING_SETUP_SQPOLL;
1085 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1087 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1089 complete(&ctx->ref_comp);
1092 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1094 return !req->timeout.off;
1097 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1099 struct io_ring_ctx *ctx;
1102 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1106 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1107 if (!ctx->fallback_req)
1111 * Use 5 bits less than the max cq entries, that should give us around
1112 * 32 entries per hash list if totally full and uniformly spread.
1114 hash_bits = ilog2(p->cq_entries);
1118 ctx->cancel_hash_bits = hash_bits;
1119 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1121 if (!ctx->cancel_hash)
1123 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1125 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1126 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1129 ctx->flags = p->flags;
1130 init_waitqueue_head(&ctx->sqo_sq_wait);
1131 INIT_LIST_HEAD(&ctx->sqd_list);
1132 init_waitqueue_head(&ctx->cq_wait);
1133 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1134 init_completion(&ctx->ref_comp);
1135 init_completion(&ctx->sq_thread_comp);
1136 idr_init(&ctx->io_buffer_idr);
1137 idr_init(&ctx->personality_idr);
1138 mutex_init(&ctx->uring_lock);
1139 init_waitqueue_head(&ctx->wait);
1140 spin_lock_init(&ctx->completion_lock);
1141 INIT_LIST_HEAD(&ctx->iopoll_list);
1142 INIT_LIST_HEAD(&ctx->defer_list);
1143 INIT_LIST_HEAD(&ctx->timeout_list);
1144 init_waitqueue_head(&ctx->inflight_wait);
1145 spin_lock_init(&ctx->inflight_lock);
1146 INIT_LIST_HEAD(&ctx->inflight_list);
1147 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1148 init_llist_head(&ctx->file_put_llist);
1151 if (ctx->fallback_req)
1152 kmem_cache_free(req_cachep, ctx->fallback_req);
1153 kfree(ctx->cancel_hash);
1158 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1160 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1161 struct io_ring_ctx *ctx = req->ctx;
1163 return seq != ctx->cached_cq_tail
1164 + atomic_read(&ctx->cached_cq_overflow);
1170 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1172 struct io_rings *rings = ctx->rings;
1174 /* order cqe stores with ring update */
1175 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1177 if (wq_has_sleeper(&ctx->cq_wait)) {
1178 wake_up_interruptible(&ctx->cq_wait);
1179 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1184 * Returns true if we need to defer file table putting. This can only happen
1185 * from the error path with REQ_F_COMP_LOCKED set.
1187 static bool io_req_clean_work(struct io_kiocb *req)
1189 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1192 req->flags &= ~REQ_F_WORK_INITIALIZED;
1195 mmdrop(req->work.mm);
1196 req->work.mm = NULL;
1198 #ifdef CONFIG_BLK_CGROUP
1199 if (req->work.blkcg_css)
1200 css_put(req->work.blkcg_css);
1202 if (req->work.creds) {
1203 put_cred(req->work.creds);
1204 req->work.creds = NULL;
1207 struct fs_struct *fs = req->work.fs;
1209 if (req->flags & REQ_F_COMP_LOCKED)
1212 spin_lock(&req->work.fs->lock);
1215 spin_unlock(&req->work.fs->lock);
1218 req->work.fs = NULL;
1224 static void io_prep_async_work(struct io_kiocb *req)
1226 const struct io_op_def *def = &io_op_defs[req->opcode];
1228 io_req_init_async(req);
1230 if (req->flags & REQ_F_ISREG) {
1231 if (def->hash_reg_file || (req->ctx->flags & IORING_SETUP_IOPOLL))
1232 io_wq_hash_work(&req->work, file_inode(req->file));
1234 if (def->unbound_nonreg_file)
1235 req->work.flags |= IO_WQ_WORK_UNBOUND;
1237 if (!req->work.mm && def->needs_mm) {
1238 mmgrab(current->mm);
1239 req->work.mm = current->mm;
1241 #ifdef CONFIG_BLK_CGROUP
1242 if (!req->work.blkcg_css && def->needs_blkcg) {
1244 req->work.blkcg_css = blkcg_css();
1246 * This should be rare, either the cgroup is dying or the task
1247 * is moving cgroups. Just punt to root for the handful of ios.
1249 if (!css_tryget_online(req->work.blkcg_css))
1250 req->work.blkcg_css = NULL;
1254 if (!req->work.creds)
1255 req->work.creds = get_current_cred();
1256 if (!req->work.fs && def->needs_fs) {
1257 spin_lock(¤t->fs->lock);
1258 if (!current->fs->in_exec) {
1259 req->work.fs = current->fs;
1260 req->work.fs->users++;
1262 req->work.flags |= IO_WQ_WORK_CANCEL;
1264 spin_unlock(¤t->fs->lock);
1266 if (def->needs_fsize)
1267 req->work.fsize = rlimit(RLIMIT_FSIZE);
1269 req->work.fsize = RLIM_INFINITY;
1272 static void io_prep_async_link(struct io_kiocb *req)
1274 struct io_kiocb *cur;
1276 io_prep_async_work(req);
1277 if (req->flags & REQ_F_LINK_HEAD)
1278 list_for_each_entry(cur, &req->link_list, link_list)
1279 io_prep_async_work(cur);
1282 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1284 struct io_ring_ctx *ctx = req->ctx;
1285 struct io_kiocb *link = io_prep_linked_timeout(req);
1287 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1288 &req->work, req->flags);
1289 io_wq_enqueue(ctx->io_wq, &req->work);
1293 static void io_queue_async_work(struct io_kiocb *req)
1295 struct io_kiocb *link;
1297 /* init ->work of the whole link before punting */
1298 io_prep_async_link(req);
1299 link = __io_queue_async_work(req);
1302 io_queue_linked_timeout(link);
1305 static void io_kill_timeout(struct io_kiocb *req)
1307 struct io_timeout_data *io = req->async_data;
1310 ret = hrtimer_try_to_cancel(&io->timer);
1312 atomic_set(&req->ctx->cq_timeouts,
1313 atomic_read(&req->ctx->cq_timeouts) + 1);
1314 list_del_init(&req->timeout.list);
1315 req->flags |= REQ_F_COMP_LOCKED;
1316 io_cqring_fill_event(req, 0);
1321 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1323 struct io_ring_ctx *ctx = req->ctx;
1325 if (!tsk || req->task == tsk)
1327 if (ctx->flags & IORING_SETUP_SQPOLL) {
1328 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1335 * Returns true if we found and killed one or more timeouts
1337 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1339 struct io_kiocb *req, *tmp;
1342 spin_lock_irq(&ctx->completion_lock);
1343 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1344 if (io_task_match(req, tsk)) {
1345 io_kill_timeout(req);
1349 spin_unlock_irq(&ctx->completion_lock);
1350 return canceled != 0;
1353 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1356 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1357 struct io_defer_entry, list);
1358 struct io_kiocb *link;
1360 if (req_need_defer(de->req, de->seq))
1362 list_del_init(&de->list);
1363 /* punt-init is done before queueing for defer */
1364 link = __io_queue_async_work(de->req);
1366 __io_queue_linked_timeout(link);
1367 /* drop submission reference */
1368 link->flags |= REQ_F_COMP_LOCKED;
1372 } while (!list_empty(&ctx->defer_list));
1375 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1377 while (!list_empty(&ctx->timeout_list)) {
1378 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1379 struct io_kiocb, timeout.list);
1381 if (io_is_timeout_noseq(req))
1383 if (req->timeout.target_seq != ctx->cached_cq_tail
1384 - atomic_read(&ctx->cq_timeouts))
1387 list_del_init(&req->timeout.list);
1388 io_kill_timeout(req);
1392 static void io_commit_cqring(struct io_ring_ctx *ctx)
1394 io_flush_timeouts(ctx);
1395 __io_commit_cqring(ctx);
1397 if (unlikely(!list_empty(&ctx->defer_list)))
1398 __io_queue_deferred(ctx);
1401 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1403 struct io_rings *r = ctx->rings;
1405 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1408 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1410 struct io_rings *rings = ctx->rings;
1413 tail = ctx->cached_cq_tail;
1415 * writes to the cq entry need to come after reading head; the
1416 * control dependency is enough as we're using WRITE_ONCE to
1419 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1422 ctx->cached_cq_tail++;
1423 return &rings->cqes[tail & ctx->cq_mask];
1426 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1430 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1432 if (!ctx->eventfd_async)
1434 return io_wq_current_is_worker();
1437 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1439 if (waitqueue_active(&ctx->wait))
1440 wake_up(&ctx->wait);
1441 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1442 wake_up(&ctx->sq_data->wait);
1443 if (io_should_trigger_evfd(ctx))
1444 eventfd_signal(ctx->cq_ev_fd, 1);
1447 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1449 if (list_empty(&ctx->cq_overflow_list)) {
1450 clear_bit(0, &ctx->sq_check_overflow);
1451 clear_bit(0, &ctx->cq_check_overflow);
1452 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1456 static inline bool io_match_files(struct io_kiocb *req,
1457 struct files_struct *files)
1461 if (req->flags & REQ_F_WORK_INITIALIZED)
1462 return req->work.files == files;
1466 /* Returns true if there are no backlogged entries after the flush */
1467 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1468 struct task_struct *tsk,
1469 struct files_struct *files)
1471 struct io_rings *rings = ctx->rings;
1472 struct io_kiocb *req, *tmp;
1473 struct io_uring_cqe *cqe;
1474 unsigned long flags;
1478 if (list_empty_careful(&ctx->cq_overflow_list))
1480 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1481 rings->cq_ring_entries))
1485 spin_lock_irqsave(&ctx->completion_lock, flags);
1487 /* if force is set, the ring is going away. always drop after that */
1489 ctx->cq_overflow_flushed = 1;
1492 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1493 if (tsk && req->task != tsk)
1495 if (!io_match_files(req, files))
1498 cqe = io_get_cqring(ctx);
1502 list_move(&req->compl.list, &list);
1504 WRITE_ONCE(cqe->user_data, req->user_data);
1505 WRITE_ONCE(cqe->res, req->result);
1506 WRITE_ONCE(cqe->flags, req->compl.cflags);
1508 WRITE_ONCE(ctx->rings->cq_overflow,
1509 atomic_inc_return(&ctx->cached_cq_overflow));
1513 io_commit_cqring(ctx);
1514 io_cqring_mark_overflow(ctx);
1516 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1517 io_cqring_ev_posted(ctx);
1519 while (!list_empty(&list)) {
1520 req = list_first_entry(&list, struct io_kiocb, compl.list);
1521 list_del(&req->compl.list);
1528 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1530 struct io_ring_ctx *ctx = req->ctx;
1531 struct io_uring_cqe *cqe;
1533 trace_io_uring_complete(ctx, req->user_data, res);
1536 * If we can't get a cq entry, userspace overflowed the
1537 * submission (by quite a lot). Increment the overflow count in
1540 cqe = io_get_cqring(ctx);
1542 WRITE_ONCE(cqe->user_data, req->user_data);
1543 WRITE_ONCE(cqe->res, res);
1544 WRITE_ONCE(cqe->flags, cflags);
1545 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1547 * If we're in ring overflow flush mode, or in task cancel mode,
1548 * then we cannot store the request for later flushing, we need
1549 * to drop it on the floor.
1551 WRITE_ONCE(ctx->rings->cq_overflow,
1552 atomic_inc_return(&ctx->cached_cq_overflow));
1554 if (list_empty(&ctx->cq_overflow_list)) {
1555 set_bit(0, &ctx->sq_check_overflow);
1556 set_bit(0, &ctx->cq_check_overflow);
1557 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1561 req->compl.cflags = cflags;
1562 refcount_inc(&req->refs);
1563 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1567 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1569 __io_cqring_fill_event(req, res, 0);
1572 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1574 struct io_ring_ctx *ctx = req->ctx;
1575 unsigned long flags;
1577 spin_lock_irqsave(&ctx->completion_lock, flags);
1578 __io_cqring_fill_event(req, res, cflags);
1579 io_commit_cqring(ctx);
1580 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1582 io_cqring_ev_posted(ctx);
1585 static void io_submit_flush_completions(struct io_comp_state *cs)
1587 struct io_ring_ctx *ctx = cs->ctx;
1589 spin_lock_irq(&ctx->completion_lock);
1590 while (!list_empty(&cs->list)) {
1591 struct io_kiocb *req;
1593 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1594 list_del(&req->compl.list);
1595 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1596 if (!(req->flags & REQ_F_LINK_HEAD)) {
1597 req->flags |= REQ_F_COMP_LOCKED;
1600 spin_unlock_irq(&ctx->completion_lock);
1602 spin_lock_irq(&ctx->completion_lock);
1605 io_commit_cqring(ctx);
1606 spin_unlock_irq(&ctx->completion_lock);
1608 io_cqring_ev_posted(ctx);
1612 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1613 struct io_comp_state *cs)
1616 io_cqring_add_event(req, res, cflags);
1621 req->compl.cflags = cflags;
1622 list_add_tail(&req->compl.list, &cs->list);
1624 io_submit_flush_completions(cs);
1628 static void io_req_complete(struct io_kiocb *req, long res)
1630 __io_req_complete(req, res, 0, NULL);
1633 static inline bool io_is_fallback_req(struct io_kiocb *req)
1635 return req == (struct io_kiocb *)
1636 ((unsigned long) req->ctx->fallback_req & ~1UL);
1639 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1641 struct io_kiocb *req;
1643 req = ctx->fallback_req;
1644 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1650 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1651 struct io_submit_state *state)
1653 if (!state->free_reqs) {
1654 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1658 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1659 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1662 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1663 * retry single alloc to be on the safe side.
1665 if (unlikely(ret <= 0)) {
1666 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1667 if (!state->reqs[0])
1671 state->free_reqs = ret;
1675 return state->reqs[state->free_reqs];
1677 return io_get_fallback_req(ctx);
1680 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1684 percpu_ref_put(req->fixed_file_refs);
1689 static bool io_dismantle_req(struct io_kiocb *req)
1693 if (req->async_data)
1694 kfree(req->async_data);
1696 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1698 return io_req_clean_work(req);
1701 static void __io_free_req_finish(struct io_kiocb *req)
1703 struct io_uring_task *tctx = req->task->io_uring;
1704 struct io_ring_ctx *ctx = req->ctx;
1706 atomic_long_inc(&tctx->req_complete);
1708 wake_up(&tctx->wait);
1709 put_task_struct(req->task);
1711 if (likely(!io_is_fallback_req(req)))
1712 kmem_cache_free(req_cachep, req);
1714 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1715 percpu_ref_put(&ctx->refs);
1718 static void io_req_task_file_table_put(struct callback_head *cb)
1720 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1721 struct fs_struct *fs = req->work.fs;
1723 spin_lock(&req->work.fs->lock);
1726 spin_unlock(&req->work.fs->lock);
1729 req->work.fs = NULL;
1730 __io_free_req_finish(req);
1733 static void __io_free_req(struct io_kiocb *req)
1735 if (!io_dismantle_req(req)) {
1736 __io_free_req_finish(req);
1740 init_task_work(&req->task_work, io_req_task_file_table_put);
1741 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1742 if (unlikely(ret)) {
1743 struct task_struct *tsk;
1745 tsk = io_wq_get_task(req->ctx->io_wq);
1746 task_work_add(tsk, &req->task_work, 0);
1751 static bool io_link_cancel_timeout(struct io_kiocb *req)
1753 struct io_timeout_data *io = req->async_data;
1754 struct io_ring_ctx *ctx = req->ctx;
1757 ret = hrtimer_try_to_cancel(&io->timer);
1759 io_cqring_fill_event(req, -ECANCELED);
1760 io_commit_cqring(ctx);
1761 req->flags &= ~REQ_F_LINK_HEAD;
1769 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1771 struct io_kiocb *link;
1774 if (list_empty(&req->link_list))
1776 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1777 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1780 list_del_init(&link->link_list);
1781 link->flags |= REQ_F_COMP_LOCKED;
1782 wake_ev = io_link_cancel_timeout(link);
1783 req->flags &= ~REQ_F_LINK_TIMEOUT;
1787 static void io_kill_linked_timeout(struct io_kiocb *req)
1789 struct io_ring_ctx *ctx = req->ctx;
1792 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1793 unsigned long flags;
1795 spin_lock_irqsave(&ctx->completion_lock, flags);
1796 wake_ev = __io_kill_linked_timeout(req);
1797 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1799 wake_ev = __io_kill_linked_timeout(req);
1803 io_cqring_ev_posted(ctx);
1806 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1808 struct io_kiocb *nxt;
1811 * The list should never be empty when we are called here. But could
1812 * potentially happen if the chain is messed up, check to be on the
1815 if (unlikely(list_empty(&req->link_list)))
1818 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1819 list_del_init(&req->link_list);
1820 if (!list_empty(&nxt->link_list))
1821 nxt->flags |= REQ_F_LINK_HEAD;
1826 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1828 static void __io_fail_links(struct io_kiocb *req)
1830 struct io_ring_ctx *ctx = req->ctx;
1832 while (!list_empty(&req->link_list)) {
1833 struct io_kiocb *link = list_first_entry(&req->link_list,
1834 struct io_kiocb, link_list);
1836 list_del_init(&link->link_list);
1837 trace_io_uring_fail_link(req, link);
1839 io_cqring_fill_event(link, -ECANCELED);
1840 link->flags |= REQ_F_COMP_LOCKED;
1841 __io_double_put_req(link);
1842 req->flags &= ~REQ_F_LINK_TIMEOUT;
1845 io_commit_cqring(ctx);
1846 io_cqring_ev_posted(ctx);
1849 static void io_fail_links(struct io_kiocb *req)
1851 struct io_ring_ctx *ctx = req->ctx;
1853 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1854 unsigned long flags;
1856 spin_lock_irqsave(&ctx->completion_lock, flags);
1857 __io_fail_links(req);
1858 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1860 __io_fail_links(req);
1863 io_cqring_ev_posted(ctx);
1866 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1868 req->flags &= ~REQ_F_LINK_HEAD;
1869 if (req->flags & REQ_F_LINK_TIMEOUT)
1870 io_kill_linked_timeout(req);
1873 * If LINK is set, we have dependent requests in this chain. If we
1874 * didn't fail this request, queue the first one up, moving any other
1875 * dependencies to the next request. In case of failure, fail the rest
1878 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1879 return io_req_link_next(req);
1884 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1886 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1888 return __io_req_find_next(req);
1891 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb,
1894 struct task_struct *tsk = req->task;
1895 struct io_ring_ctx *ctx = req->ctx;
1898 if (tsk->flags & PF_EXITING)
1902 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1903 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1904 * processing task_work. There's no reliable way to tell if TWA_RESUME
1908 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1909 notify = TWA_SIGNAL;
1911 ret = task_work_add(tsk, cb, notify);
1913 wake_up_process(tsk);
1918 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1920 struct io_ring_ctx *ctx = req->ctx;
1922 spin_lock_irq(&ctx->completion_lock);
1923 io_cqring_fill_event(req, error);
1924 io_commit_cqring(ctx);
1925 spin_unlock_irq(&ctx->completion_lock);
1927 io_cqring_ev_posted(ctx);
1928 req_set_fail_links(req);
1929 io_double_put_req(req);
1932 static void io_req_task_cancel(struct callback_head *cb)
1934 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1935 struct io_ring_ctx *ctx = req->ctx;
1937 __io_req_task_cancel(req, -ECANCELED);
1938 percpu_ref_put(&ctx->refs);
1941 static void __io_req_task_submit(struct io_kiocb *req)
1943 struct io_ring_ctx *ctx = req->ctx;
1945 if (!__io_sq_thread_acquire_mm(ctx)) {
1946 mutex_lock(&ctx->uring_lock);
1947 __io_queue_sqe(req, NULL, NULL);
1948 mutex_unlock(&ctx->uring_lock);
1950 __io_req_task_cancel(req, -EFAULT);
1954 static void io_req_task_submit(struct callback_head *cb)
1956 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1957 struct io_ring_ctx *ctx = req->ctx;
1959 __io_req_task_submit(req);
1960 percpu_ref_put(&ctx->refs);
1963 static void io_req_task_queue(struct io_kiocb *req)
1967 init_task_work(&req->task_work, io_req_task_submit);
1968 percpu_ref_get(&req->ctx->refs);
1970 ret = io_req_task_work_add(req, &req->task_work, true);
1971 if (unlikely(ret)) {
1972 struct task_struct *tsk;
1974 init_task_work(&req->task_work, io_req_task_cancel);
1975 tsk = io_wq_get_task(req->ctx->io_wq);
1976 task_work_add(tsk, &req->task_work, 0);
1977 wake_up_process(tsk);
1981 static void io_queue_next(struct io_kiocb *req)
1983 struct io_kiocb *nxt = io_req_find_next(req);
1986 io_req_task_queue(nxt);
1989 static void io_free_req(struct io_kiocb *req)
1996 void *reqs[IO_IOPOLL_BATCH];
1999 struct task_struct *task;
2003 static inline void io_init_req_batch(struct req_batch *rb)
2010 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2011 struct req_batch *rb)
2013 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2014 percpu_ref_put_many(&ctx->refs, rb->to_free);
2018 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2019 struct req_batch *rb)
2022 __io_req_free_batch_flush(ctx, rb);
2024 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2025 put_task_struct_many(rb->task, rb->task_refs);
2030 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2032 if (unlikely(io_is_fallback_req(req))) {
2036 if (req->flags & REQ_F_LINK_HEAD)
2039 if (req->task != rb->task) {
2041 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2042 put_task_struct_many(rb->task, rb->task_refs);
2044 rb->task = req->task;
2049 WARN_ON_ONCE(io_dismantle_req(req));
2050 rb->reqs[rb->to_free++] = req;
2051 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2052 __io_req_free_batch_flush(req->ctx, rb);
2056 * Drop reference to request, return next in chain (if there is one) if this
2057 * was the last reference to this request.
2059 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2061 struct io_kiocb *nxt = NULL;
2063 if (refcount_dec_and_test(&req->refs)) {
2064 nxt = io_req_find_next(req);
2070 static void io_put_req(struct io_kiocb *req)
2072 if (refcount_dec_and_test(&req->refs))
2076 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2078 struct io_kiocb *nxt;
2081 * A ref is owned by io-wq in which context we're. So, if that's the
2082 * last one, it's safe to steal next work. False negatives are Ok,
2083 * it just will be re-punted async in io_put_work()
2085 if (refcount_read(&req->refs) != 1)
2088 nxt = io_req_find_next(req);
2089 return nxt ? &nxt->work : NULL;
2093 * Must only be used if we don't need to care about links, usually from
2094 * within the completion handling itself.
2096 static void __io_double_put_req(struct io_kiocb *req)
2098 /* drop both submit and complete references */
2099 if (refcount_sub_and_test(2, &req->refs))
2103 static void io_double_put_req(struct io_kiocb *req)
2105 /* drop both submit and complete references */
2106 if (refcount_sub_and_test(2, &req->refs))
2110 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2112 struct io_rings *rings = ctx->rings;
2114 if (test_bit(0, &ctx->cq_check_overflow)) {
2116 * noflush == true is from the waitqueue handler, just ensure
2117 * we wake up the task, and the next invocation will flush the
2118 * entries. We cannot safely to it from here.
2120 if (noflush && !list_empty(&ctx->cq_overflow_list))
2123 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2126 /* See comment at the top of this file */
2128 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2131 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2133 struct io_rings *rings = ctx->rings;
2135 /* make sure SQ entry isn't read before tail */
2136 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2139 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2141 unsigned int cflags;
2143 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2144 cflags |= IORING_CQE_F_BUFFER;
2145 req->flags &= ~REQ_F_BUFFER_SELECTED;
2150 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2152 struct io_buffer *kbuf;
2154 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2155 return io_put_kbuf(req, kbuf);
2158 static inline bool io_run_task_work(void)
2161 * Not safe to run on exiting task, and the task_work handling will
2162 * not add work to such a task.
2164 if (unlikely(current->flags & PF_EXITING))
2166 if (current->task_works) {
2167 __set_current_state(TASK_RUNNING);
2175 static void io_iopoll_queue(struct list_head *again)
2177 struct io_kiocb *req;
2180 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2181 list_del(&req->inflight_entry);
2182 __io_complete_rw(req, -EAGAIN, 0, NULL);
2183 } while (!list_empty(again));
2187 * Find and free completed poll iocbs
2189 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2190 struct list_head *done)
2192 struct req_batch rb;
2193 struct io_kiocb *req;
2196 /* order with ->result store in io_complete_rw_iopoll() */
2199 io_init_req_batch(&rb);
2200 while (!list_empty(done)) {
2203 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2204 if (READ_ONCE(req->result) == -EAGAIN) {
2206 req->iopoll_completed = 0;
2207 list_move_tail(&req->inflight_entry, &again);
2210 list_del(&req->inflight_entry);
2212 if (req->flags & REQ_F_BUFFER_SELECTED)
2213 cflags = io_put_rw_kbuf(req);
2215 __io_cqring_fill_event(req, req->result, cflags);
2218 if (refcount_dec_and_test(&req->refs))
2219 io_req_free_batch(&rb, req);
2222 io_commit_cqring(ctx);
2223 if (ctx->flags & IORING_SETUP_SQPOLL)
2224 io_cqring_ev_posted(ctx);
2225 io_req_free_batch_finish(ctx, &rb);
2227 if (!list_empty(&again))
2228 io_iopoll_queue(&again);
2231 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2234 struct io_kiocb *req, *tmp;
2240 * Only spin for completions if we don't have multiple devices hanging
2241 * off our complete list, and we're under the requested amount.
2243 spin = !ctx->poll_multi_file && *nr_events < min;
2246 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2247 struct kiocb *kiocb = &req->rw.kiocb;
2250 * Move completed and retryable entries to our local lists.
2251 * If we find a request that requires polling, break out
2252 * and complete those lists first, if we have entries there.
2254 if (READ_ONCE(req->iopoll_completed)) {
2255 list_move_tail(&req->inflight_entry, &done);
2258 if (!list_empty(&done))
2261 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2265 /* iopoll may have completed current req */
2266 if (READ_ONCE(req->iopoll_completed))
2267 list_move_tail(&req->inflight_entry, &done);
2274 if (!list_empty(&done))
2275 io_iopoll_complete(ctx, nr_events, &done);
2281 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2282 * non-spinning poll check - we'll still enter the driver poll loop, but only
2283 * as a non-spinning completion check.
2285 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2288 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2291 ret = io_do_iopoll(ctx, nr_events, min);
2294 if (*nr_events >= min)
2302 * We can't just wait for polled events to come to us, we have to actively
2303 * find and complete them.
2305 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2307 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2310 mutex_lock(&ctx->uring_lock);
2311 while (!list_empty(&ctx->iopoll_list)) {
2312 unsigned int nr_events = 0;
2314 io_do_iopoll(ctx, &nr_events, 0);
2316 /* let it sleep and repeat later if can't complete a request */
2320 * Ensure we allow local-to-the-cpu processing to take place,
2321 * in this case we need to ensure that we reap all events.
2322 * Also let task_work, etc. to progress by releasing the mutex
2324 if (need_resched()) {
2325 mutex_unlock(&ctx->uring_lock);
2327 mutex_lock(&ctx->uring_lock);
2330 mutex_unlock(&ctx->uring_lock);
2333 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2335 unsigned int nr_events = 0;
2336 int iters = 0, ret = 0;
2339 * We disallow the app entering submit/complete with polling, but we
2340 * still need to lock the ring to prevent racing with polled issue
2341 * that got punted to a workqueue.
2343 mutex_lock(&ctx->uring_lock);
2346 * Don't enter poll loop if we already have events pending.
2347 * If we do, we can potentially be spinning for commands that
2348 * already triggered a CQE (eg in error).
2350 if (io_cqring_events(ctx, false))
2354 * If a submit got punted to a workqueue, we can have the
2355 * application entering polling for a command before it gets
2356 * issued. That app will hold the uring_lock for the duration
2357 * of the poll right here, so we need to take a breather every
2358 * now and then to ensure that the issue has a chance to add
2359 * the poll to the issued list. Otherwise we can spin here
2360 * forever, while the workqueue is stuck trying to acquire the
2363 if (!(++iters & 7)) {
2364 mutex_unlock(&ctx->uring_lock);
2366 mutex_lock(&ctx->uring_lock);
2369 ret = io_iopoll_getevents(ctx, &nr_events, min);
2373 } while (min && !nr_events && !need_resched());
2375 mutex_unlock(&ctx->uring_lock);
2379 static void kiocb_end_write(struct io_kiocb *req)
2382 * Tell lockdep we inherited freeze protection from submission
2385 if (req->flags & REQ_F_ISREG) {
2386 struct inode *inode = file_inode(req->file);
2388 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2390 file_end_write(req->file);
2393 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2394 struct io_comp_state *cs)
2396 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2399 if (kiocb->ki_flags & IOCB_WRITE)
2400 kiocb_end_write(req);
2402 if (res != req->result)
2403 req_set_fail_links(req);
2404 if (req->flags & REQ_F_BUFFER_SELECTED)
2405 cflags = io_put_rw_kbuf(req);
2406 __io_req_complete(req, res, cflags, cs);
2410 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2412 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2413 ssize_t ret = -ECANCELED;
2414 struct iov_iter iter;
2422 switch (req->opcode) {
2423 case IORING_OP_READV:
2424 case IORING_OP_READ_FIXED:
2425 case IORING_OP_READ:
2428 case IORING_OP_WRITEV:
2429 case IORING_OP_WRITE_FIXED:
2430 case IORING_OP_WRITE:
2434 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2439 if (!req->async_data) {
2440 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2443 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2451 req_set_fail_links(req);
2452 io_req_complete(req, ret);
2457 static bool io_rw_reissue(struct io_kiocb *req, long res)
2460 umode_t mode = file_inode(req->file)->i_mode;
2463 if (!S_ISBLK(mode) && !S_ISREG(mode))
2465 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2468 ret = io_sq_thread_acquire_mm(req->ctx, req);
2470 if (io_resubmit_prep(req, ret)) {
2471 refcount_inc(&req->refs);
2472 io_queue_async_work(req);
2480 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2481 struct io_comp_state *cs)
2483 if (!io_rw_reissue(req, res))
2484 io_complete_rw_common(&req->rw.kiocb, res, cs);
2487 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2489 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2491 __io_complete_rw(req, res, res2, NULL);
2494 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2496 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2498 if (kiocb->ki_flags & IOCB_WRITE)
2499 kiocb_end_write(req);
2501 if (res != -EAGAIN && res != req->result)
2502 req_set_fail_links(req);
2504 WRITE_ONCE(req->result, res);
2505 /* order with io_poll_complete() checking ->result */
2507 WRITE_ONCE(req->iopoll_completed, 1);
2511 * After the iocb has been issued, it's safe to be found on the poll list.
2512 * Adding the kiocb to the list AFTER submission ensures that we don't
2513 * find it from a io_iopoll_getevents() thread before the issuer is done
2514 * accessing the kiocb cookie.
2516 static void io_iopoll_req_issued(struct io_kiocb *req)
2518 struct io_ring_ctx *ctx = req->ctx;
2521 * Track whether we have multiple files in our lists. This will impact
2522 * how we do polling eventually, not spinning if we're on potentially
2523 * different devices.
2525 if (list_empty(&ctx->iopoll_list)) {
2526 ctx->poll_multi_file = false;
2527 } else if (!ctx->poll_multi_file) {
2528 struct io_kiocb *list_req;
2530 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2532 if (list_req->file != req->file)
2533 ctx->poll_multi_file = true;
2537 * For fast devices, IO may have already completed. If it has, add
2538 * it to the front so we find it first.
2540 if (READ_ONCE(req->iopoll_completed))
2541 list_add(&req->inflight_entry, &ctx->iopoll_list);
2543 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2545 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2546 wq_has_sleeper(&ctx->sq_data->wait))
2547 wake_up(&ctx->sq_data->wait);
2550 static void __io_state_file_put(struct io_submit_state *state)
2552 if (state->has_refs)
2553 fput_many(state->file, state->has_refs);
2557 static inline void io_state_file_put(struct io_submit_state *state)
2560 __io_state_file_put(state);
2564 * Get as many references to a file as we have IOs left in this submission,
2565 * assuming most submissions are for one file, or at least that each file
2566 * has more than one submission.
2568 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2574 if (state->fd == fd) {
2579 __io_state_file_put(state);
2581 state->file = fget_many(fd, state->ios_left);
2587 state->has_refs = state->ios_left;
2591 static bool io_bdev_nowait(struct block_device *bdev)
2594 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2601 * If we tracked the file through the SCM inflight mechanism, we could support
2602 * any file. For now, just ensure that anything potentially problematic is done
2605 static bool io_file_supports_async(struct file *file, int rw)
2607 umode_t mode = file_inode(file)->i_mode;
2609 if (S_ISBLK(mode)) {
2610 if (io_bdev_nowait(file->f_inode->i_bdev))
2614 if (S_ISCHR(mode) || S_ISSOCK(mode))
2616 if (S_ISREG(mode)) {
2617 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2618 file->f_op != &io_uring_fops)
2623 /* any ->read/write should understand O_NONBLOCK */
2624 if (file->f_flags & O_NONBLOCK)
2627 if (!(file->f_mode & FMODE_NOWAIT))
2631 return file->f_op->read_iter != NULL;
2633 return file->f_op->write_iter != NULL;
2636 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2638 struct io_ring_ctx *ctx = req->ctx;
2639 struct kiocb *kiocb = &req->rw.kiocb;
2643 if (S_ISREG(file_inode(req->file)->i_mode))
2644 req->flags |= REQ_F_ISREG;
2646 kiocb->ki_pos = READ_ONCE(sqe->off);
2647 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2648 req->flags |= REQ_F_CUR_POS;
2649 kiocb->ki_pos = req->file->f_pos;
2651 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2652 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2653 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2657 ioprio = READ_ONCE(sqe->ioprio);
2659 ret = ioprio_check_cap(ioprio);
2663 kiocb->ki_ioprio = ioprio;
2665 kiocb->ki_ioprio = get_current_ioprio();
2667 /* don't allow async punt if RWF_NOWAIT was requested */
2668 if (kiocb->ki_flags & IOCB_NOWAIT)
2669 req->flags |= REQ_F_NOWAIT;
2671 if (ctx->flags & IORING_SETUP_IOPOLL) {
2672 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2673 !kiocb->ki_filp->f_op->iopoll)
2676 kiocb->ki_flags |= IOCB_HIPRI;
2677 kiocb->ki_complete = io_complete_rw_iopoll;
2678 req->iopoll_completed = 0;
2680 if (kiocb->ki_flags & IOCB_HIPRI)
2682 kiocb->ki_complete = io_complete_rw;
2685 req->rw.addr = READ_ONCE(sqe->addr);
2686 req->rw.len = READ_ONCE(sqe->len);
2687 req->buf_index = READ_ONCE(sqe->buf_index);
2691 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2697 case -ERESTARTNOINTR:
2698 case -ERESTARTNOHAND:
2699 case -ERESTART_RESTARTBLOCK:
2701 * We can't just restart the syscall, since previously
2702 * submitted sqes may already be in progress. Just fail this
2708 kiocb->ki_complete(kiocb, ret, 0);
2712 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2713 struct io_comp_state *cs)
2715 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2716 struct io_async_rw *io = req->async_data;
2718 /* add previously done IO, if any */
2719 if (io && io->bytes_done > 0) {
2721 ret = io->bytes_done;
2723 ret += io->bytes_done;
2726 if (req->flags & REQ_F_CUR_POS)
2727 req->file->f_pos = kiocb->ki_pos;
2728 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2729 __io_complete_rw(req, ret, 0, cs);
2731 io_rw_done(kiocb, ret);
2734 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2735 struct iov_iter *iter)
2737 struct io_ring_ctx *ctx = req->ctx;
2738 size_t len = req->rw.len;
2739 struct io_mapped_ubuf *imu;
2740 u16 index, buf_index = req->buf_index;
2744 if (unlikely(buf_index >= ctx->nr_user_bufs))
2746 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2747 imu = &ctx->user_bufs[index];
2748 buf_addr = req->rw.addr;
2751 if (buf_addr + len < buf_addr)
2753 /* not inside the mapped region */
2754 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2758 * May not be a start of buffer, set size appropriately
2759 * and advance us to the beginning.
2761 offset = buf_addr - imu->ubuf;
2762 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2766 * Don't use iov_iter_advance() here, as it's really slow for
2767 * using the latter parts of a big fixed buffer - it iterates
2768 * over each segment manually. We can cheat a bit here, because
2771 * 1) it's a BVEC iter, we set it up
2772 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2773 * first and last bvec
2775 * So just find our index, and adjust the iterator afterwards.
2776 * If the offset is within the first bvec (or the whole first
2777 * bvec, just use iov_iter_advance(). This makes it easier
2778 * since we can just skip the first segment, which may not
2779 * be PAGE_SIZE aligned.
2781 const struct bio_vec *bvec = imu->bvec;
2783 if (offset <= bvec->bv_len) {
2784 iov_iter_advance(iter, offset);
2786 unsigned long seg_skip;
2788 /* skip first vec */
2789 offset -= bvec->bv_len;
2790 seg_skip = 1 + (offset >> PAGE_SHIFT);
2792 iter->bvec = bvec + seg_skip;
2793 iter->nr_segs -= seg_skip;
2794 iter->count -= bvec->bv_len + offset;
2795 iter->iov_offset = offset & ~PAGE_MASK;
2802 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2805 mutex_unlock(&ctx->uring_lock);
2808 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2811 * "Normal" inline submissions always hold the uring_lock, since we
2812 * grab it from the system call. Same is true for the SQPOLL offload.
2813 * The only exception is when we've detached the request and issue it
2814 * from an async worker thread, grab the lock for that case.
2817 mutex_lock(&ctx->uring_lock);
2820 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2821 int bgid, struct io_buffer *kbuf,
2824 struct io_buffer *head;
2826 if (req->flags & REQ_F_BUFFER_SELECTED)
2829 io_ring_submit_lock(req->ctx, needs_lock);
2831 lockdep_assert_held(&req->ctx->uring_lock);
2833 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2835 if (!list_empty(&head->list)) {
2836 kbuf = list_last_entry(&head->list, struct io_buffer,
2838 list_del(&kbuf->list);
2841 idr_remove(&req->ctx->io_buffer_idr, bgid);
2843 if (*len > kbuf->len)
2846 kbuf = ERR_PTR(-ENOBUFS);
2849 io_ring_submit_unlock(req->ctx, needs_lock);
2854 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2857 struct io_buffer *kbuf;
2860 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2861 bgid = req->buf_index;
2862 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2865 req->rw.addr = (u64) (unsigned long) kbuf;
2866 req->flags |= REQ_F_BUFFER_SELECTED;
2867 return u64_to_user_ptr(kbuf->addr);
2870 #ifdef CONFIG_COMPAT
2871 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2874 struct compat_iovec __user *uiov;
2875 compat_ssize_t clen;
2879 uiov = u64_to_user_ptr(req->rw.addr);
2880 if (!access_ok(uiov, sizeof(*uiov)))
2882 if (__get_user(clen, &uiov->iov_len))
2888 buf = io_rw_buffer_select(req, &len, needs_lock);
2890 return PTR_ERR(buf);
2891 iov[0].iov_base = buf;
2892 iov[0].iov_len = (compat_size_t) len;
2897 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2900 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2904 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2907 len = iov[0].iov_len;
2910 buf = io_rw_buffer_select(req, &len, needs_lock);
2912 return PTR_ERR(buf);
2913 iov[0].iov_base = buf;
2914 iov[0].iov_len = len;
2918 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2921 if (req->flags & REQ_F_BUFFER_SELECTED) {
2922 struct io_buffer *kbuf;
2924 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2925 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2926 iov[0].iov_len = kbuf->len;
2931 else if (req->rw.len > 1)
2934 #ifdef CONFIG_COMPAT
2935 if (req->ctx->compat)
2936 return io_compat_import(req, iov, needs_lock);
2939 return __io_iov_buffer_select(req, iov, needs_lock);
2942 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2943 struct iovec **iovec, struct iov_iter *iter,
2946 void __user *buf = u64_to_user_ptr(req->rw.addr);
2947 size_t sqe_len = req->rw.len;
2951 opcode = req->opcode;
2952 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2954 return io_import_fixed(req, rw, iter);
2957 /* buffer index only valid with fixed read/write, or buffer select */
2958 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2961 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2962 if (req->flags & REQ_F_BUFFER_SELECT) {
2963 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2965 return PTR_ERR(buf);
2966 req->rw.len = sqe_len;
2969 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2971 return ret < 0 ? ret : sqe_len;
2974 if (req->flags & REQ_F_BUFFER_SELECT) {
2975 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2977 ret = (*iovec)->iov_len;
2978 iov_iter_init(iter, rw, *iovec, 1, ret);
2984 #ifdef CONFIG_COMPAT
2985 if (req->ctx->compat)
2986 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2990 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2993 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2994 struct iovec **iovec, struct iov_iter *iter,
2997 struct io_async_rw *iorw = req->async_data;
3000 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3002 return iov_iter_count(&iorw->iter);
3005 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3007 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3011 * For files that don't have ->read_iter() and ->write_iter(), handle them
3012 * by looping over ->read() or ->write() manually.
3014 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
3015 struct iov_iter *iter)
3020 * Don't support polled IO through this interface, and we can't
3021 * support non-blocking either. For the latter, this just causes
3022 * the kiocb to be handled from an async context.
3024 if (kiocb->ki_flags & IOCB_HIPRI)
3026 if (kiocb->ki_flags & IOCB_NOWAIT)
3029 while (iov_iter_count(iter)) {
3033 if (!iov_iter_is_bvec(iter)) {
3034 iovec = iov_iter_iovec(iter);
3036 /* fixed buffers import bvec */
3037 iovec.iov_base = kmap(iter->bvec->bv_page)
3039 iovec.iov_len = min(iter->count,
3040 iter->bvec->bv_len - iter->iov_offset);
3044 nr = file->f_op->read(file, iovec.iov_base,
3045 iovec.iov_len, io_kiocb_ppos(kiocb));
3047 nr = file->f_op->write(file, iovec.iov_base,
3048 iovec.iov_len, io_kiocb_ppos(kiocb));
3051 if (iov_iter_is_bvec(iter))
3052 kunmap(iter->bvec->bv_page);
3060 if (nr != iovec.iov_len)
3062 iov_iter_advance(iter, nr);
3068 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3069 const struct iovec *fast_iov, struct iov_iter *iter)
3071 struct io_async_rw *rw = req->async_data;
3073 memcpy(&rw->iter, iter, sizeof(*iter));
3074 rw->free_iovec = iovec;
3076 /* can only be fixed buffers, no need to do anything */
3077 if (iter->type == ITER_BVEC)
3080 unsigned iov_off = 0;
3082 rw->iter.iov = rw->fast_iov;
3083 if (iter->iov != fast_iov) {
3084 iov_off = iter->iov - fast_iov;
3085 rw->iter.iov += iov_off;
3087 if (rw->fast_iov != fast_iov)
3088 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3089 sizeof(struct iovec) * iter->nr_segs);
3091 req->flags |= REQ_F_NEED_CLEANUP;
3095 static inline int __io_alloc_async_data(struct io_kiocb *req)
3097 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3098 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3099 return req->async_data == NULL;
3102 static int io_alloc_async_data(struct io_kiocb *req)
3104 if (!io_op_defs[req->opcode].needs_async_data)
3107 return __io_alloc_async_data(req);
3110 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3111 const struct iovec *fast_iov,
3112 struct iov_iter *iter, bool force)
3114 if (!force && !io_op_defs[req->opcode].needs_async_data)
3116 if (!req->async_data) {
3117 if (__io_alloc_async_data(req))
3120 io_req_map_rw(req, iovec, fast_iov, iter);
3125 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
3126 bool force_nonblock)
3128 struct io_async_rw *iorw = req->async_data;
3129 struct iovec *iov = iorw->fast_iov;
3132 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, !force_nonblock);
3133 if (unlikely(ret < 0))
3136 iorw->bytes_done = 0;
3137 iorw->free_iovec = iov;
3139 req->flags |= REQ_F_NEED_CLEANUP;
3143 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3144 bool force_nonblock)
3148 ret = io_prep_rw(req, sqe);
3152 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3155 /* either don't need iovec imported or already have it */
3156 if (!req->async_data)
3158 return io_rw_prep_async(req, READ, force_nonblock);
3162 * This is our waitqueue callback handler, registered through lock_page_async()
3163 * when we initially tried to do the IO with the iocb armed our waitqueue.
3164 * This gets called when the page is unlocked, and we generally expect that to
3165 * happen when the page IO is completed and the page is now uptodate. This will
3166 * queue a task_work based retry of the operation, attempting to copy the data
3167 * again. If the latter fails because the page was NOT uptodate, then we will
3168 * do a thread based blocking retry of the operation. That's the unexpected
3171 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3172 int sync, void *arg)
3174 struct wait_page_queue *wpq;
3175 struct io_kiocb *req = wait->private;
3176 struct wait_page_key *key = arg;
3179 wpq = container_of(wait, struct wait_page_queue, wait);
3181 if (!wake_page_match(wpq, key))
3184 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3185 list_del_init(&wait->entry);
3187 init_task_work(&req->task_work, io_req_task_submit);
3188 percpu_ref_get(&req->ctx->refs);
3190 /* submit ref gets dropped, acquire a new one */
3191 refcount_inc(&req->refs);
3192 ret = io_req_task_work_add(req, &req->task_work, true);
3193 if (unlikely(ret)) {
3194 struct task_struct *tsk;
3196 /* queue just for cancelation */
3197 init_task_work(&req->task_work, io_req_task_cancel);
3198 tsk = io_wq_get_task(req->ctx->io_wq);
3199 task_work_add(tsk, &req->task_work, 0);
3200 wake_up_process(tsk);
3206 * This controls whether a given IO request should be armed for async page
3207 * based retry. If we return false here, the request is handed to the async
3208 * worker threads for retry. If we're doing buffered reads on a regular file,
3209 * we prepare a private wait_page_queue entry and retry the operation. This
3210 * will either succeed because the page is now uptodate and unlocked, or it
3211 * will register a callback when the page is unlocked at IO completion. Through
3212 * that callback, io_uring uses task_work to setup a retry of the operation.
3213 * That retry will attempt the buffered read again. The retry will generally
3214 * succeed, or in rare cases where it fails, we then fall back to using the
3215 * async worker threads for a blocking retry.
3217 static bool io_rw_should_retry(struct io_kiocb *req)
3219 struct io_async_rw *rw = req->async_data;
3220 struct wait_page_queue *wait = &rw->wpq;
3221 struct kiocb *kiocb = &req->rw.kiocb;
3223 /* never retry for NOWAIT, we just complete with -EAGAIN */
3224 if (req->flags & REQ_F_NOWAIT)
3227 /* Only for buffered IO */
3228 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3232 * just use poll if we can, and don't attempt if the fs doesn't
3233 * support callback based unlocks
3235 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3238 wait->wait.func = io_async_buf_func;
3239 wait->wait.private = req;
3240 wait->wait.flags = 0;
3241 INIT_LIST_HEAD(&wait->wait.entry);
3242 kiocb->ki_flags |= IOCB_WAITQ;
3243 kiocb->ki_flags &= ~IOCB_NOWAIT;
3244 kiocb->ki_waitq = wait;
3248 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3250 if (req->file->f_op->read_iter)
3251 return call_read_iter(req->file, &req->rw.kiocb, iter);
3252 else if (req->file->f_op->read)
3253 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3258 static int io_read(struct io_kiocb *req, bool force_nonblock,
3259 struct io_comp_state *cs)
3261 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3262 struct kiocb *kiocb = &req->rw.kiocb;
3263 struct iov_iter __iter, *iter = &__iter;
3264 struct io_async_rw *rw = req->async_data;
3265 ssize_t io_size, ret, ret2;
3272 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3275 iov_count = iov_iter_count(iter);
3277 req->result = io_size;
3280 /* Ensure we clear previously set non-block flag */
3281 if (!force_nonblock)
3282 kiocb->ki_flags &= ~IOCB_NOWAIT;
3284 kiocb->ki_flags |= IOCB_NOWAIT;
3287 /* If the file doesn't support async, just async punt */
3288 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3292 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3296 ret = io_iter_do_read(req, iter);
3300 } else if (ret == -EIOCBQUEUED) {
3303 } else if (ret == -EAGAIN) {
3304 /* IOPOLL retry should happen for io-wq threads */
3305 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3307 /* no retry on NONBLOCK marked file */
3308 if (req->file->f_flags & O_NONBLOCK)
3310 /* some cases will consume bytes even on error returns */
3311 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3314 } else if (ret < 0) {
3315 /* make sure -ERESTARTSYS -> -EINTR is done */
3319 /* read it all, or we did blocking attempt. no retry. */
3320 if (!iov_iter_count(iter) || !force_nonblock ||
3321 (req->file->f_flags & O_NONBLOCK))
3326 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3333 rw = req->async_data;
3334 /* it's copied and will be cleaned with ->io */
3336 /* now use our persistent iterator, if we aren't already */
3339 rw->bytes_done += ret;
3340 /* if we can retry, do so with the callbacks armed */
3341 if (!io_rw_should_retry(req)) {
3342 kiocb->ki_flags &= ~IOCB_WAITQ;
3347 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3348 * get -EIOCBQUEUED, then we'll get a notification when the desired
3349 * page gets unlocked. We can also get a partial read here, and if we
3350 * do, then just retry at the new offset.
3352 ret = io_iter_do_read(req, iter);
3353 if (ret == -EIOCBQUEUED) {
3356 } else if (ret > 0 && ret < io_size) {
3357 /* we got some bytes, but not all. retry. */
3361 kiocb_done(kiocb, ret, cs);
3364 /* it's reportedly faster than delegating the null check to kfree() */
3370 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3371 bool force_nonblock)
3375 ret = io_prep_rw(req, sqe);
3379 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3382 /* either don't need iovec imported or already have it */
3383 if (!req->async_data)
3385 return io_rw_prep_async(req, WRITE, force_nonblock);
3388 static int io_write(struct io_kiocb *req, bool force_nonblock,
3389 struct io_comp_state *cs)
3391 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3392 struct kiocb *kiocb = &req->rw.kiocb;
3393 struct iov_iter __iter, *iter = &__iter;
3394 struct io_async_rw *rw = req->async_data;
3396 ssize_t ret, ret2, io_size;
3401 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3404 iov_count = iov_iter_count(iter);
3406 req->result = io_size;
3408 /* Ensure we clear previously set non-block flag */
3409 if (!force_nonblock)
3410 kiocb->ki_flags &= ~IOCB_NOWAIT;
3412 kiocb->ki_flags |= IOCB_NOWAIT;
3414 /* If the file doesn't support async, just async punt */
3415 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3418 /* file path doesn't support NOWAIT for non-direct_IO */
3419 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3420 (req->flags & REQ_F_ISREG))
3423 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3428 * Open-code file_start_write here to grab freeze protection,
3429 * which will be released by another thread in
3430 * io_complete_rw(). Fool lockdep by telling it the lock got
3431 * released so that it doesn't complain about the held lock when
3432 * we return to userspace.
3434 if (req->flags & REQ_F_ISREG) {
3435 __sb_start_write(file_inode(req->file)->i_sb,
3436 SB_FREEZE_WRITE, true);
3437 __sb_writers_release(file_inode(req->file)->i_sb,
3440 kiocb->ki_flags |= IOCB_WRITE;
3442 if (req->file->f_op->write_iter)
3443 ret2 = call_write_iter(req->file, kiocb, iter);
3444 else if (req->file->f_op->write)
3445 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3450 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3451 * retry them without IOCB_NOWAIT.
3453 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3455 /* no retry on NONBLOCK marked file */
3456 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3458 if (!force_nonblock || ret2 != -EAGAIN) {
3459 /* IOPOLL retry should happen for io-wq threads */
3460 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3463 kiocb_done(kiocb, ret2, cs);
3466 /* some cases will consume bytes even on error returns */
3467 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3468 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3473 /* it's reportedly faster than delegating the null check to kfree() */
3479 static int __io_splice_prep(struct io_kiocb *req,
3480 const struct io_uring_sqe *sqe)
3482 struct io_splice* sp = &req->splice;
3483 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3486 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3490 sp->len = READ_ONCE(sqe->len);
3491 sp->flags = READ_ONCE(sqe->splice_flags);
3493 if (unlikely(sp->flags & ~valid_flags))
3496 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3497 (sp->flags & SPLICE_F_FD_IN_FIXED));
3500 req->flags |= REQ_F_NEED_CLEANUP;
3502 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3504 * Splice operation will be punted aync, and here need to
3505 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3507 io_req_init_async(req);
3508 req->work.flags |= IO_WQ_WORK_UNBOUND;
3514 static int io_tee_prep(struct io_kiocb *req,
3515 const struct io_uring_sqe *sqe)
3517 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3519 return __io_splice_prep(req, sqe);
3522 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3524 struct io_splice *sp = &req->splice;
3525 struct file *in = sp->file_in;
3526 struct file *out = sp->file_out;
3527 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3533 ret = do_tee(in, out, sp->len, flags);
3535 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3536 req->flags &= ~REQ_F_NEED_CLEANUP;
3539 req_set_fail_links(req);
3540 io_req_complete(req, ret);
3544 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3546 struct io_splice* sp = &req->splice;
3548 sp->off_in = READ_ONCE(sqe->splice_off_in);
3549 sp->off_out = READ_ONCE(sqe->off);
3550 return __io_splice_prep(req, sqe);
3553 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3555 struct io_splice *sp = &req->splice;
3556 struct file *in = sp->file_in;
3557 struct file *out = sp->file_out;
3558 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3559 loff_t *poff_in, *poff_out;
3565 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3566 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3569 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3571 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3572 req->flags &= ~REQ_F_NEED_CLEANUP;
3575 req_set_fail_links(req);
3576 io_req_complete(req, ret);
3581 * IORING_OP_NOP just posts a completion event, nothing else.
3583 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3585 struct io_ring_ctx *ctx = req->ctx;
3587 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3590 __io_req_complete(req, 0, 0, cs);
3594 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3596 struct io_ring_ctx *ctx = req->ctx;
3601 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3603 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3606 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3607 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3610 req->sync.off = READ_ONCE(sqe->off);
3611 req->sync.len = READ_ONCE(sqe->len);
3615 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3617 loff_t end = req->sync.off + req->sync.len;
3620 /* fsync always requires a blocking context */
3624 ret = vfs_fsync_range(req->file, req->sync.off,
3625 end > 0 ? end : LLONG_MAX,
3626 req->sync.flags & IORING_FSYNC_DATASYNC);
3628 req_set_fail_links(req);
3629 io_req_complete(req, ret);
3633 static int io_fallocate_prep(struct io_kiocb *req,
3634 const struct io_uring_sqe *sqe)
3636 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3638 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3641 req->sync.off = READ_ONCE(sqe->off);
3642 req->sync.len = READ_ONCE(sqe->addr);
3643 req->sync.mode = READ_ONCE(sqe->len);
3647 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3651 /* fallocate always requiring blocking context */
3654 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3657 req_set_fail_links(req);
3658 io_req_complete(req, ret);
3662 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3664 const char __user *fname;
3667 if (unlikely(sqe->ioprio || sqe->buf_index))
3669 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3672 /* open.how should be already initialised */
3673 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3674 req->open.how.flags |= O_LARGEFILE;
3676 req->open.dfd = READ_ONCE(sqe->fd);
3677 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3678 req->open.filename = getname(fname);
3679 if (IS_ERR(req->open.filename)) {
3680 ret = PTR_ERR(req->open.filename);
3681 req->open.filename = NULL;
3684 req->open.nofile = rlimit(RLIMIT_NOFILE);
3685 req->flags |= REQ_F_NEED_CLEANUP;
3689 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3693 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3695 mode = READ_ONCE(sqe->len);
3696 flags = READ_ONCE(sqe->open_flags);
3697 req->open.how = build_open_how(flags, mode);
3698 return __io_openat_prep(req, sqe);
3701 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3703 struct open_how __user *how;
3707 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3709 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3710 len = READ_ONCE(sqe->len);
3711 if (len < OPEN_HOW_SIZE_VER0)
3714 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3719 return __io_openat_prep(req, sqe);
3722 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3724 struct open_flags op;
3731 ret = build_open_flags(&req->open.how, &op);
3735 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3739 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3742 ret = PTR_ERR(file);
3744 fsnotify_open(file);
3745 fd_install(ret, file);
3748 putname(req->open.filename);
3749 req->flags &= ~REQ_F_NEED_CLEANUP;
3751 req_set_fail_links(req);
3752 io_req_complete(req, ret);
3756 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3758 return io_openat2(req, force_nonblock);
3761 static int io_remove_buffers_prep(struct io_kiocb *req,
3762 const struct io_uring_sqe *sqe)
3764 struct io_provide_buf *p = &req->pbuf;
3767 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3770 tmp = READ_ONCE(sqe->fd);
3771 if (!tmp || tmp > USHRT_MAX)
3774 memset(p, 0, sizeof(*p));
3776 p->bgid = READ_ONCE(sqe->buf_group);
3780 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3781 int bgid, unsigned nbufs)
3785 /* shouldn't happen */
3789 /* the head kbuf is the list itself */
3790 while (!list_empty(&buf->list)) {
3791 struct io_buffer *nxt;
3793 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3794 list_del(&nxt->list);
3801 idr_remove(&ctx->io_buffer_idr, bgid);
3806 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3807 struct io_comp_state *cs)
3809 struct io_provide_buf *p = &req->pbuf;
3810 struct io_ring_ctx *ctx = req->ctx;
3811 struct io_buffer *head;
3814 io_ring_submit_lock(ctx, !force_nonblock);
3816 lockdep_assert_held(&ctx->uring_lock);
3819 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3821 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3823 io_ring_submit_lock(ctx, !force_nonblock);
3825 req_set_fail_links(req);
3826 __io_req_complete(req, ret, 0, cs);
3830 static int io_provide_buffers_prep(struct io_kiocb *req,
3831 const struct io_uring_sqe *sqe)
3833 struct io_provide_buf *p = &req->pbuf;
3836 if (sqe->ioprio || sqe->rw_flags)
3839 tmp = READ_ONCE(sqe->fd);
3840 if (!tmp || tmp > USHRT_MAX)
3843 p->addr = READ_ONCE(sqe->addr);
3844 p->len = READ_ONCE(sqe->len);
3846 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3849 p->bgid = READ_ONCE(sqe->buf_group);
3850 tmp = READ_ONCE(sqe->off);
3851 if (tmp > USHRT_MAX)
3857 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3859 struct io_buffer *buf;
3860 u64 addr = pbuf->addr;
3861 int i, bid = pbuf->bid;
3863 for (i = 0; i < pbuf->nbufs; i++) {
3864 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3869 buf->len = pbuf->len;
3874 INIT_LIST_HEAD(&buf->list);
3877 list_add_tail(&buf->list, &(*head)->list);
3881 return i ? i : -ENOMEM;
3884 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3885 struct io_comp_state *cs)
3887 struct io_provide_buf *p = &req->pbuf;
3888 struct io_ring_ctx *ctx = req->ctx;
3889 struct io_buffer *head, *list;
3892 io_ring_submit_lock(ctx, !force_nonblock);
3894 lockdep_assert_held(&ctx->uring_lock);
3896 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3898 ret = io_add_buffers(p, &head);
3903 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3906 __io_remove_buffers(ctx, head, p->bgid, -1U);
3911 io_ring_submit_unlock(ctx, !force_nonblock);
3913 req_set_fail_links(req);
3914 __io_req_complete(req, ret, 0, cs);
3918 static int io_epoll_ctl_prep(struct io_kiocb *req,
3919 const struct io_uring_sqe *sqe)
3921 #if defined(CONFIG_EPOLL)
3922 if (sqe->ioprio || sqe->buf_index)
3924 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3927 req->epoll.epfd = READ_ONCE(sqe->fd);
3928 req->epoll.op = READ_ONCE(sqe->len);
3929 req->epoll.fd = READ_ONCE(sqe->off);
3931 if (ep_op_has_event(req->epoll.op)) {
3932 struct epoll_event __user *ev;
3934 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3935 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3945 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3946 struct io_comp_state *cs)
3948 #if defined(CONFIG_EPOLL)
3949 struct io_epoll *ie = &req->epoll;
3952 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3953 if (force_nonblock && ret == -EAGAIN)
3957 req_set_fail_links(req);
3958 __io_req_complete(req, ret, 0, cs);
3965 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3967 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3968 if (sqe->ioprio || sqe->buf_index || sqe->off)
3970 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3973 req->madvise.addr = READ_ONCE(sqe->addr);
3974 req->madvise.len = READ_ONCE(sqe->len);
3975 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3982 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3984 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3985 struct io_madvise *ma = &req->madvise;
3991 ret = do_madvise(ma->addr, ma->len, ma->advice);
3993 req_set_fail_links(req);
3994 io_req_complete(req, ret);
4001 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4003 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4005 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4008 req->fadvise.offset = READ_ONCE(sqe->off);
4009 req->fadvise.len = READ_ONCE(sqe->len);
4010 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4014 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4016 struct io_fadvise *fa = &req->fadvise;
4019 if (force_nonblock) {
4020 switch (fa->advice) {
4021 case POSIX_FADV_NORMAL:
4022 case POSIX_FADV_RANDOM:
4023 case POSIX_FADV_SEQUENTIAL:
4030 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4032 req_set_fail_links(req);
4033 io_req_complete(req, ret);
4037 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4039 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4041 if (sqe->ioprio || sqe->buf_index)
4043 if (req->flags & REQ_F_FIXED_FILE)
4046 req->statx.dfd = READ_ONCE(sqe->fd);
4047 req->statx.mask = READ_ONCE(sqe->len);
4048 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4049 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4050 req->statx.flags = READ_ONCE(sqe->statx_flags);
4055 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4057 struct io_statx *ctx = &req->statx;
4060 if (force_nonblock) {
4061 /* only need file table for an actual valid fd */
4062 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4063 req->flags |= REQ_F_NO_FILE_TABLE;
4067 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4071 req_set_fail_links(req);
4072 io_req_complete(req, ret);
4076 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4079 * If we queue this for async, it must not be cancellable. That would
4080 * leave the 'file' in an undeterminate state, and here need to modify
4081 * io_wq_work.flags, so initialize io_wq_work firstly.
4083 io_req_init_async(req);
4084 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4086 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4088 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4089 sqe->rw_flags || sqe->buf_index)
4091 if (req->flags & REQ_F_FIXED_FILE)
4094 req->close.fd = READ_ONCE(sqe->fd);
4095 if ((req->file && req->file->f_op == &io_uring_fops))
4098 req->close.put_file = NULL;
4102 static int io_close(struct io_kiocb *req, bool force_nonblock,
4103 struct io_comp_state *cs)
4105 struct io_close *close = &req->close;
4108 /* might be already done during nonblock submission */
4109 if (!close->put_file) {
4110 ret = __close_fd_get_file(close->fd, &close->put_file);
4112 return (ret == -ENOENT) ? -EBADF : ret;
4115 /* if the file has a flush method, be safe and punt to async */
4116 if (close->put_file->f_op->flush && force_nonblock) {
4117 /* was never set, but play safe */
4118 req->flags &= ~REQ_F_NOWAIT;
4119 /* avoid grabbing files - we don't need the files */
4120 req->flags |= REQ_F_NO_FILE_TABLE;
4124 /* No ->flush() or already async, safely close from here */
4125 ret = filp_close(close->put_file, req->work.files);
4127 req_set_fail_links(req);
4128 fput(close->put_file);
4129 close->put_file = NULL;
4130 __io_req_complete(req, ret, 0, cs);
4134 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4136 struct io_ring_ctx *ctx = req->ctx;
4141 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4143 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4146 req->sync.off = READ_ONCE(sqe->off);
4147 req->sync.len = READ_ONCE(sqe->len);
4148 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4152 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4156 /* sync_file_range always requires a blocking context */
4160 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4163 req_set_fail_links(req);
4164 io_req_complete(req, ret);
4168 #if defined(CONFIG_NET)
4169 static int io_setup_async_msg(struct io_kiocb *req,
4170 struct io_async_msghdr *kmsg)
4172 struct io_async_msghdr *async_msg = req->async_data;
4176 if (io_alloc_async_data(req)) {
4177 if (kmsg->iov != kmsg->fast_iov)
4181 async_msg = req->async_data;
4182 req->flags |= REQ_F_NEED_CLEANUP;
4183 memcpy(async_msg, kmsg, sizeof(*kmsg));
4187 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4188 struct io_async_msghdr *iomsg)
4190 iomsg->iov = iomsg->fast_iov;
4191 iomsg->msg.msg_name = &iomsg->addr;
4192 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4193 req->sr_msg.msg_flags, &iomsg->iov);
4196 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4198 struct io_async_msghdr *async_msg = req->async_data;
4199 struct io_sr_msg *sr = &req->sr_msg;
4202 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4205 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4206 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4207 sr->len = READ_ONCE(sqe->len);
4209 #ifdef CONFIG_COMPAT
4210 if (req->ctx->compat)
4211 sr->msg_flags |= MSG_CMSG_COMPAT;
4214 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4216 ret = io_sendmsg_copy_hdr(req, async_msg);
4218 req->flags |= REQ_F_NEED_CLEANUP;
4222 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4223 struct io_comp_state *cs)
4225 struct io_async_msghdr iomsg, *kmsg;
4226 struct socket *sock;
4230 sock = sock_from_file(req->file, &ret);
4231 if (unlikely(!sock))
4234 if (req->async_data) {
4235 kmsg = req->async_data;
4236 kmsg->msg.msg_name = &kmsg->addr;
4237 /* if iov is set, it's allocated already */
4239 kmsg->iov = kmsg->fast_iov;
4240 kmsg->msg.msg_iter.iov = kmsg->iov;
4242 ret = io_sendmsg_copy_hdr(req, &iomsg);
4248 flags = req->sr_msg.msg_flags;
4249 if (flags & MSG_DONTWAIT)
4250 req->flags |= REQ_F_NOWAIT;
4251 else if (force_nonblock)
4252 flags |= MSG_DONTWAIT;
4254 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4255 if (force_nonblock && ret == -EAGAIN)
4256 return io_setup_async_msg(req, kmsg);
4257 if (ret == -ERESTARTSYS)
4260 if (kmsg->iov != kmsg->fast_iov)
4262 req->flags &= ~REQ_F_NEED_CLEANUP;
4264 req_set_fail_links(req);
4265 __io_req_complete(req, ret, 0, cs);
4269 static int io_send(struct io_kiocb *req, bool force_nonblock,
4270 struct io_comp_state *cs)
4272 struct io_sr_msg *sr = &req->sr_msg;
4275 struct socket *sock;
4279 sock = sock_from_file(req->file, &ret);
4280 if (unlikely(!sock))
4283 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4287 msg.msg_name = NULL;
4288 msg.msg_control = NULL;
4289 msg.msg_controllen = 0;
4290 msg.msg_namelen = 0;
4292 flags = req->sr_msg.msg_flags;
4293 if (flags & MSG_DONTWAIT)
4294 req->flags |= REQ_F_NOWAIT;
4295 else if (force_nonblock)
4296 flags |= MSG_DONTWAIT;
4298 msg.msg_flags = flags;
4299 ret = sock_sendmsg(sock, &msg);
4300 if (force_nonblock && ret == -EAGAIN)
4302 if (ret == -ERESTARTSYS)
4306 req_set_fail_links(req);
4307 __io_req_complete(req, ret, 0, cs);
4311 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4312 struct io_async_msghdr *iomsg)
4314 struct io_sr_msg *sr = &req->sr_msg;
4315 struct iovec __user *uiov;
4319 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4320 &iomsg->uaddr, &uiov, &iov_len);
4324 if (req->flags & REQ_F_BUFFER_SELECT) {
4327 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4329 sr->len = iomsg->iov[0].iov_len;
4330 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4334 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4335 &iomsg->iov, &iomsg->msg.msg_iter);
4343 #ifdef CONFIG_COMPAT
4344 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4345 struct io_async_msghdr *iomsg)
4347 struct compat_msghdr __user *msg_compat;
4348 struct io_sr_msg *sr = &req->sr_msg;
4349 struct compat_iovec __user *uiov;
4354 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4355 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4360 uiov = compat_ptr(ptr);
4361 if (req->flags & REQ_F_BUFFER_SELECT) {
4362 compat_ssize_t clen;
4366 if (!access_ok(uiov, sizeof(*uiov)))
4368 if (__get_user(clen, &uiov->iov_len))
4372 sr->len = iomsg->iov[0].iov_len;
4375 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4377 &iomsg->msg.msg_iter);
4386 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4387 struct io_async_msghdr *iomsg)
4389 iomsg->msg.msg_name = &iomsg->addr;
4390 iomsg->iov = iomsg->fast_iov;
4392 #ifdef CONFIG_COMPAT
4393 if (req->ctx->compat)
4394 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4397 return __io_recvmsg_copy_hdr(req, iomsg);
4400 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4403 struct io_sr_msg *sr = &req->sr_msg;
4404 struct io_buffer *kbuf;
4406 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4411 req->flags |= REQ_F_BUFFER_SELECTED;
4415 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4417 return io_put_kbuf(req, req->sr_msg.kbuf);
4420 static int io_recvmsg_prep(struct io_kiocb *req,
4421 const struct io_uring_sqe *sqe)
4423 struct io_async_msghdr *async_msg = req->async_data;
4424 struct io_sr_msg *sr = &req->sr_msg;
4427 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4430 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4431 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4432 sr->len = READ_ONCE(sqe->len);
4433 sr->bgid = READ_ONCE(sqe->buf_group);
4435 #ifdef CONFIG_COMPAT
4436 if (req->ctx->compat)
4437 sr->msg_flags |= MSG_CMSG_COMPAT;
4440 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4442 ret = io_recvmsg_copy_hdr(req, async_msg);
4444 req->flags |= REQ_F_NEED_CLEANUP;
4448 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4449 struct io_comp_state *cs)
4451 struct io_async_msghdr iomsg, *kmsg;
4452 struct socket *sock;
4453 struct io_buffer *kbuf;
4455 int ret, cflags = 0;
4457 sock = sock_from_file(req->file, &ret);
4458 if (unlikely(!sock))
4461 if (req->async_data) {
4462 kmsg = req->async_data;
4463 kmsg->msg.msg_name = &kmsg->addr;
4464 /* if iov is set, it's allocated already */
4466 kmsg->iov = kmsg->fast_iov;
4467 kmsg->msg.msg_iter.iov = kmsg->iov;
4469 ret = io_recvmsg_copy_hdr(req, &iomsg);
4475 if (req->flags & REQ_F_BUFFER_SELECT) {
4476 kbuf = io_recv_buffer_select(req, !force_nonblock);
4478 return PTR_ERR(kbuf);
4479 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4480 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4481 1, req->sr_msg.len);
4484 flags = req->sr_msg.msg_flags;
4485 if (flags & MSG_DONTWAIT)
4486 req->flags |= REQ_F_NOWAIT;
4487 else if (force_nonblock)
4488 flags |= MSG_DONTWAIT;
4490 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4491 kmsg->uaddr, flags);
4492 if (force_nonblock && ret == -EAGAIN)
4493 return io_setup_async_msg(req, kmsg);
4494 if (ret == -ERESTARTSYS)
4497 if (req->flags & REQ_F_BUFFER_SELECTED)
4498 cflags = io_put_recv_kbuf(req);
4499 if (kmsg->iov != kmsg->fast_iov)
4501 req->flags &= ~REQ_F_NEED_CLEANUP;
4503 req_set_fail_links(req);
4504 __io_req_complete(req, ret, cflags, cs);
4508 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4509 struct io_comp_state *cs)
4511 struct io_buffer *kbuf;
4512 struct io_sr_msg *sr = &req->sr_msg;
4514 void __user *buf = sr->buf;
4515 struct socket *sock;
4518 int ret, cflags = 0;
4520 sock = sock_from_file(req->file, &ret);
4521 if (unlikely(!sock))
4524 if (req->flags & REQ_F_BUFFER_SELECT) {
4525 kbuf = io_recv_buffer_select(req, !force_nonblock);
4527 return PTR_ERR(kbuf);
4528 buf = u64_to_user_ptr(kbuf->addr);
4531 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4535 msg.msg_name = NULL;
4536 msg.msg_control = NULL;
4537 msg.msg_controllen = 0;
4538 msg.msg_namelen = 0;
4539 msg.msg_iocb = NULL;
4542 flags = req->sr_msg.msg_flags;
4543 if (flags & MSG_DONTWAIT)
4544 req->flags |= REQ_F_NOWAIT;
4545 else if (force_nonblock)
4546 flags |= MSG_DONTWAIT;
4548 ret = sock_recvmsg(sock, &msg, flags);
4549 if (force_nonblock && ret == -EAGAIN)
4551 if (ret == -ERESTARTSYS)
4554 if (req->flags & REQ_F_BUFFER_SELECTED)
4555 cflags = io_put_recv_kbuf(req);
4557 req_set_fail_links(req);
4558 __io_req_complete(req, ret, cflags, cs);
4562 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4564 struct io_accept *accept = &req->accept;
4566 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4568 if (sqe->ioprio || sqe->len || sqe->buf_index)
4571 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4572 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4573 accept->flags = READ_ONCE(sqe->accept_flags);
4574 accept->nofile = rlimit(RLIMIT_NOFILE);
4578 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4579 struct io_comp_state *cs)
4581 struct io_accept *accept = &req->accept;
4582 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4585 if (req->file->f_flags & O_NONBLOCK)
4586 req->flags |= REQ_F_NOWAIT;
4588 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4589 accept->addr_len, accept->flags,
4591 if (ret == -EAGAIN && force_nonblock)
4594 if (ret == -ERESTARTSYS)
4596 req_set_fail_links(req);
4598 __io_req_complete(req, ret, 0, cs);
4602 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4604 struct io_connect *conn = &req->connect;
4605 struct io_async_connect *io = req->async_data;
4607 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4609 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4612 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4613 conn->addr_len = READ_ONCE(sqe->addr2);
4618 return move_addr_to_kernel(conn->addr, conn->addr_len,
4622 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4623 struct io_comp_state *cs)
4625 struct io_async_connect __io, *io;
4626 unsigned file_flags;
4629 if (req->async_data) {
4630 io = req->async_data;
4632 ret = move_addr_to_kernel(req->connect.addr,
4633 req->connect.addr_len,
4640 file_flags = force_nonblock ? O_NONBLOCK : 0;
4642 ret = __sys_connect_file(req->file, &io->address,
4643 req->connect.addr_len, file_flags);
4644 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4645 if (req->async_data)
4647 if (io_alloc_async_data(req)) {
4651 io = req->async_data;
4652 memcpy(req->async_data, &__io, sizeof(__io));
4655 if (ret == -ERESTARTSYS)
4659 req_set_fail_links(req);
4660 __io_req_complete(req, ret, 0, cs);
4663 #else /* !CONFIG_NET */
4664 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4669 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4670 struct io_comp_state *cs)
4675 static int io_send(struct io_kiocb *req, bool force_nonblock,
4676 struct io_comp_state *cs)
4681 static int io_recvmsg_prep(struct io_kiocb *req,
4682 const struct io_uring_sqe *sqe)
4687 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4688 struct io_comp_state *cs)
4693 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4694 struct io_comp_state *cs)
4699 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4704 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4705 struct io_comp_state *cs)
4710 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4715 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4716 struct io_comp_state *cs)
4720 #endif /* CONFIG_NET */
4722 struct io_poll_table {
4723 struct poll_table_struct pt;
4724 struct io_kiocb *req;
4728 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4729 __poll_t mask, task_work_func_t func)
4734 /* for instances that support it check for an event match first: */
4735 if (mask && !(mask & poll->events))
4738 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4740 list_del_init(&poll->wait.entry);
4743 init_task_work(&req->task_work, func);
4744 percpu_ref_get(&req->ctx->refs);
4747 * If we using the signalfd wait_queue_head for this wakeup, then
4748 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4749 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4750 * either, as the normal wakeup will suffice.
4752 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4755 * If this fails, then the task is exiting. When a task exits, the
4756 * work gets canceled, so just cancel this request as well instead
4757 * of executing it. We can't safely execute it anyway, as we may not
4758 * have the needed state needed for it anyway.
4760 ret = io_req_task_work_add(req, &req->task_work, twa_signal_ok);
4761 if (unlikely(ret)) {
4762 struct task_struct *tsk;
4764 WRITE_ONCE(poll->canceled, true);
4765 tsk = io_wq_get_task(req->ctx->io_wq);
4766 task_work_add(tsk, &req->task_work, 0);
4767 wake_up_process(tsk);
4772 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4773 __acquires(&req->ctx->completion_lock)
4775 struct io_ring_ctx *ctx = req->ctx;
4777 if (!req->result && !READ_ONCE(poll->canceled)) {
4778 struct poll_table_struct pt = { ._key = poll->events };
4780 req->result = vfs_poll(req->file, &pt) & poll->events;
4783 spin_lock_irq(&ctx->completion_lock);
4784 if (!req->result && !READ_ONCE(poll->canceled)) {
4785 add_wait_queue(poll->head, &poll->wait);
4792 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4794 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4795 if (req->opcode == IORING_OP_POLL_ADD)
4796 return req->async_data;
4797 return req->apoll->double_poll;
4800 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4802 if (req->opcode == IORING_OP_POLL_ADD)
4804 return &req->apoll->poll;
4807 static void io_poll_remove_double(struct io_kiocb *req)
4809 struct io_poll_iocb *poll = io_poll_get_double(req);
4811 lockdep_assert_held(&req->ctx->completion_lock);
4813 if (poll && poll->head) {
4814 struct wait_queue_head *head = poll->head;
4816 spin_lock(&head->lock);
4817 list_del_init(&poll->wait.entry);
4818 if (poll->wait.private)
4819 refcount_dec(&req->refs);
4821 spin_unlock(&head->lock);
4825 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4827 struct io_ring_ctx *ctx = req->ctx;
4829 io_poll_remove_double(req);
4830 req->poll.done = true;
4831 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4832 io_commit_cqring(ctx);
4835 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4837 struct io_ring_ctx *ctx = req->ctx;
4839 if (io_poll_rewait(req, &req->poll)) {
4840 spin_unlock_irq(&ctx->completion_lock);
4844 hash_del(&req->hash_node);
4845 io_poll_complete(req, req->result, 0);
4846 req->flags |= REQ_F_COMP_LOCKED;
4847 *nxt = io_put_req_find_next(req);
4848 spin_unlock_irq(&ctx->completion_lock);
4850 io_cqring_ev_posted(ctx);
4853 static void io_poll_task_func(struct callback_head *cb)
4855 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4856 struct io_ring_ctx *ctx = req->ctx;
4857 struct io_kiocb *nxt = NULL;
4859 io_poll_task_handler(req, &nxt);
4861 __io_req_task_submit(nxt);
4862 percpu_ref_put(&ctx->refs);
4865 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4866 int sync, void *key)
4868 struct io_kiocb *req = wait->private;
4869 struct io_poll_iocb *poll = io_poll_get_single(req);
4870 __poll_t mask = key_to_poll(key);
4872 /* for instances that support it check for an event match first: */
4873 if (mask && !(mask & poll->events))
4876 list_del_init(&wait->entry);
4878 if (poll && poll->head) {
4881 spin_lock(&poll->head->lock);
4882 done = list_empty(&poll->wait.entry);
4884 list_del_init(&poll->wait.entry);
4885 /* make sure double remove sees this as being gone */
4886 wait->private = NULL;
4887 spin_unlock(&poll->head->lock);
4889 __io_async_wake(req, poll, mask, io_poll_task_func);
4891 refcount_dec(&req->refs);
4895 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4896 wait_queue_func_t wake_func)
4900 poll->canceled = false;
4901 poll->events = events;
4902 INIT_LIST_HEAD(&poll->wait.entry);
4903 init_waitqueue_func_entry(&poll->wait, wake_func);
4906 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4907 struct wait_queue_head *head,
4908 struct io_poll_iocb **poll_ptr)
4910 struct io_kiocb *req = pt->req;
4913 * If poll->head is already set, it's because the file being polled
4914 * uses multiple waitqueues for poll handling (eg one for read, one
4915 * for write). Setup a separate io_poll_iocb if this happens.
4917 if (unlikely(poll->head)) {
4918 /* already have a 2nd entry, fail a third attempt */
4920 pt->error = -EINVAL;
4923 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4925 pt->error = -ENOMEM;
4928 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4929 refcount_inc(&req->refs);
4930 poll->wait.private = req;
4937 if (poll->events & EPOLLEXCLUSIVE)
4938 add_wait_queue_exclusive(head, &poll->wait);
4940 add_wait_queue(head, &poll->wait);
4943 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4944 struct poll_table_struct *p)
4946 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4947 struct async_poll *apoll = pt->req->apoll;
4949 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4952 static void io_async_task_func(struct callback_head *cb)
4954 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4955 struct async_poll *apoll = req->apoll;
4956 struct io_ring_ctx *ctx = req->ctx;
4958 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4960 if (io_poll_rewait(req, &apoll->poll)) {
4961 spin_unlock_irq(&ctx->completion_lock);
4962 percpu_ref_put(&ctx->refs);
4966 /* If req is still hashed, it cannot have been canceled. Don't check. */
4967 if (hash_hashed(&req->hash_node))
4968 hash_del(&req->hash_node);
4970 io_poll_remove_double(req);
4971 spin_unlock_irq(&ctx->completion_lock);
4973 if (!READ_ONCE(apoll->poll.canceled))
4974 __io_req_task_submit(req);
4976 __io_req_task_cancel(req, -ECANCELED);
4978 percpu_ref_put(&ctx->refs);
4979 kfree(apoll->double_poll);
4983 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4986 struct io_kiocb *req = wait->private;
4987 struct io_poll_iocb *poll = &req->apoll->poll;
4989 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4992 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4995 static void io_poll_req_insert(struct io_kiocb *req)
4997 struct io_ring_ctx *ctx = req->ctx;
4998 struct hlist_head *list;
5000 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5001 hlist_add_head(&req->hash_node, list);
5004 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5005 struct io_poll_iocb *poll,
5006 struct io_poll_table *ipt, __poll_t mask,
5007 wait_queue_func_t wake_func)
5008 __acquires(&ctx->completion_lock)
5010 struct io_ring_ctx *ctx = req->ctx;
5011 bool cancel = false;
5013 io_init_poll_iocb(poll, mask, wake_func);
5014 poll->file = req->file;
5015 poll->wait.private = req;
5017 ipt->pt._key = mask;
5019 ipt->error = -EINVAL;
5021 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5023 spin_lock_irq(&ctx->completion_lock);
5024 if (likely(poll->head)) {
5025 spin_lock(&poll->head->lock);
5026 if (unlikely(list_empty(&poll->wait.entry))) {
5032 if (mask || ipt->error)
5033 list_del_init(&poll->wait.entry);
5035 WRITE_ONCE(poll->canceled, true);
5036 else if (!poll->done) /* actually waiting for an event */
5037 io_poll_req_insert(req);
5038 spin_unlock(&poll->head->lock);
5044 static bool io_arm_poll_handler(struct io_kiocb *req)
5046 const struct io_op_def *def = &io_op_defs[req->opcode];
5047 struct io_ring_ctx *ctx = req->ctx;
5048 struct async_poll *apoll;
5049 struct io_poll_table ipt;
5053 if (!req->file || !file_can_poll(req->file))
5055 if (req->flags & REQ_F_POLLED)
5059 else if (def->pollout)
5063 /* if we can't nonblock try, then no point in arming a poll handler */
5064 if (!io_file_supports_async(req->file, rw))
5067 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5068 if (unlikely(!apoll))
5070 apoll->double_poll = NULL;
5072 req->flags |= REQ_F_POLLED;
5074 INIT_HLIST_NODE(&req->hash_node);
5078 mask |= POLLIN | POLLRDNORM;
5080 mask |= POLLOUT | POLLWRNORM;
5081 mask |= POLLERR | POLLPRI;
5083 ipt.pt._qproc = io_async_queue_proc;
5085 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5087 if (ret || ipt.error) {
5088 io_poll_remove_double(req);
5089 spin_unlock_irq(&ctx->completion_lock);
5090 kfree(apoll->double_poll);
5094 spin_unlock_irq(&ctx->completion_lock);
5095 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5096 apoll->poll.events);
5100 static bool __io_poll_remove_one(struct io_kiocb *req,
5101 struct io_poll_iocb *poll)
5103 bool do_complete = false;
5105 spin_lock(&poll->head->lock);
5106 WRITE_ONCE(poll->canceled, true);
5107 if (!list_empty(&poll->wait.entry)) {
5108 list_del_init(&poll->wait.entry);
5111 spin_unlock(&poll->head->lock);
5112 hash_del(&req->hash_node);
5116 static bool io_poll_remove_one(struct io_kiocb *req)
5120 io_poll_remove_double(req);
5122 if (req->opcode == IORING_OP_POLL_ADD) {
5123 do_complete = __io_poll_remove_one(req, &req->poll);
5125 struct async_poll *apoll = req->apoll;
5127 /* non-poll requests have submit ref still */
5128 do_complete = __io_poll_remove_one(req, &apoll->poll);
5131 kfree(apoll->double_poll);
5137 io_cqring_fill_event(req, -ECANCELED);
5138 io_commit_cqring(req->ctx);
5139 req->flags |= REQ_F_COMP_LOCKED;
5140 req_set_fail_links(req);
5148 * Returns true if we found and killed one or more poll requests
5150 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5152 struct hlist_node *tmp;
5153 struct io_kiocb *req;
5156 spin_lock_irq(&ctx->completion_lock);
5157 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5158 struct hlist_head *list;
5160 list = &ctx->cancel_hash[i];
5161 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5162 if (io_task_match(req, tsk))
5163 posted += io_poll_remove_one(req);
5166 spin_unlock_irq(&ctx->completion_lock);
5169 io_cqring_ev_posted(ctx);
5174 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5176 struct hlist_head *list;
5177 struct io_kiocb *req;
5179 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5180 hlist_for_each_entry(req, list, hash_node) {
5181 if (sqe_addr != req->user_data)
5183 if (io_poll_remove_one(req))
5191 static int io_poll_remove_prep(struct io_kiocb *req,
5192 const struct io_uring_sqe *sqe)
5194 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5196 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5200 req->poll.addr = READ_ONCE(sqe->addr);
5205 * Find a running poll command that matches one specified in sqe->addr,
5206 * and remove it if found.
5208 static int io_poll_remove(struct io_kiocb *req)
5210 struct io_ring_ctx *ctx = req->ctx;
5214 addr = req->poll.addr;
5215 spin_lock_irq(&ctx->completion_lock);
5216 ret = io_poll_cancel(ctx, addr);
5217 spin_unlock_irq(&ctx->completion_lock);
5220 req_set_fail_links(req);
5221 io_req_complete(req, ret);
5225 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5228 struct io_kiocb *req = wait->private;
5229 struct io_poll_iocb *poll = &req->poll;
5231 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5234 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5235 struct poll_table_struct *p)
5237 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5239 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5242 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5244 struct io_poll_iocb *poll = &req->poll;
5247 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5249 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5254 events = READ_ONCE(sqe->poll32_events);
5256 events = swahw32(events);
5258 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5259 (events & EPOLLEXCLUSIVE);
5263 static int io_poll_add(struct io_kiocb *req)
5265 struct io_poll_iocb *poll = &req->poll;
5266 struct io_ring_ctx *ctx = req->ctx;
5267 struct io_poll_table ipt;
5270 INIT_HLIST_NODE(&req->hash_node);
5271 ipt.pt._qproc = io_poll_queue_proc;
5273 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5276 if (mask) { /* no async, we'd stolen it */
5278 io_poll_complete(req, mask, 0);
5280 spin_unlock_irq(&ctx->completion_lock);
5283 io_cqring_ev_posted(ctx);
5289 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5291 struct io_timeout_data *data = container_of(timer,
5292 struct io_timeout_data, timer);
5293 struct io_kiocb *req = data->req;
5294 struct io_ring_ctx *ctx = req->ctx;
5295 unsigned long flags;
5297 spin_lock_irqsave(&ctx->completion_lock, flags);
5298 atomic_set(&req->ctx->cq_timeouts,
5299 atomic_read(&req->ctx->cq_timeouts) + 1);
5302 * We could be racing with timeout deletion. If the list is empty,
5303 * then timeout lookup already found it and will be handling it.
5305 if (!list_empty(&req->timeout.list))
5306 list_del_init(&req->timeout.list);
5308 io_cqring_fill_event(req, -ETIME);
5309 io_commit_cqring(ctx);
5310 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5312 io_cqring_ev_posted(ctx);
5313 req_set_fail_links(req);
5315 return HRTIMER_NORESTART;
5318 static int __io_timeout_cancel(struct io_kiocb *req)
5320 struct io_timeout_data *io = req->async_data;
5323 list_del_init(&req->timeout.list);
5325 ret = hrtimer_try_to_cancel(&io->timer);
5329 req_set_fail_links(req);
5330 req->flags |= REQ_F_COMP_LOCKED;
5331 io_cqring_fill_event(req, -ECANCELED);
5336 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5338 struct io_kiocb *req;
5341 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5342 if (user_data == req->user_data) {
5351 return __io_timeout_cancel(req);
5354 static int io_timeout_remove_prep(struct io_kiocb *req,
5355 const struct io_uring_sqe *sqe)
5357 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5359 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5361 if (sqe->ioprio || sqe->buf_index || sqe->len)
5364 req->timeout.addr = READ_ONCE(sqe->addr);
5365 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5366 if (req->timeout.flags)
5373 * Remove or update an existing timeout command
5375 static int io_timeout_remove(struct io_kiocb *req)
5377 struct io_ring_ctx *ctx = req->ctx;
5380 spin_lock_irq(&ctx->completion_lock);
5381 ret = io_timeout_cancel(ctx, req->timeout.addr);
5383 io_cqring_fill_event(req, ret);
5384 io_commit_cqring(ctx);
5385 spin_unlock_irq(&ctx->completion_lock);
5386 io_cqring_ev_posted(ctx);
5388 req_set_fail_links(req);
5393 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5394 bool is_timeout_link)
5396 struct io_timeout_data *data;
5398 u32 off = READ_ONCE(sqe->off);
5400 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5402 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5404 if (off && is_timeout_link)
5406 flags = READ_ONCE(sqe->timeout_flags);
5407 if (flags & ~IORING_TIMEOUT_ABS)
5410 req->timeout.off = off;
5412 if (!req->async_data && io_alloc_async_data(req))
5415 data = req->async_data;
5418 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5421 if (flags & IORING_TIMEOUT_ABS)
5422 data->mode = HRTIMER_MODE_ABS;
5424 data->mode = HRTIMER_MODE_REL;
5426 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5430 static int io_timeout(struct io_kiocb *req)
5432 struct io_ring_ctx *ctx = req->ctx;
5433 struct io_timeout_data *data = req->async_data;
5434 struct list_head *entry;
5435 u32 tail, off = req->timeout.off;
5437 spin_lock_irq(&ctx->completion_lock);
5440 * sqe->off holds how many events that need to occur for this
5441 * timeout event to be satisfied. If it isn't set, then this is
5442 * a pure timeout request, sequence isn't used.
5444 if (io_is_timeout_noseq(req)) {
5445 entry = ctx->timeout_list.prev;
5449 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5450 req->timeout.target_seq = tail + off;
5453 * Insertion sort, ensuring the first entry in the list is always
5454 * the one we need first.
5456 list_for_each_prev(entry, &ctx->timeout_list) {
5457 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5460 if (io_is_timeout_noseq(nxt))
5462 /* nxt.seq is behind @tail, otherwise would've been completed */
5463 if (off >= nxt->timeout.target_seq - tail)
5467 list_add(&req->timeout.list, entry);
5468 data->timer.function = io_timeout_fn;
5469 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5470 spin_unlock_irq(&ctx->completion_lock);
5474 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5476 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5478 return req->user_data == (unsigned long) data;
5481 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5483 enum io_wq_cancel cancel_ret;
5486 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5487 switch (cancel_ret) {
5488 case IO_WQ_CANCEL_OK:
5491 case IO_WQ_CANCEL_RUNNING:
5494 case IO_WQ_CANCEL_NOTFOUND:
5502 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5503 struct io_kiocb *req, __u64 sqe_addr,
5506 unsigned long flags;
5509 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5510 if (ret != -ENOENT) {
5511 spin_lock_irqsave(&ctx->completion_lock, flags);
5515 spin_lock_irqsave(&ctx->completion_lock, flags);
5516 ret = io_timeout_cancel(ctx, sqe_addr);
5519 ret = io_poll_cancel(ctx, sqe_addr);
5523 io_cqring_fill_event(req, ret);
5524 io_commit_cqring(ctx);
5525 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5526 io_cqring_ev_posted(ctx);
5529 req_set_fail_links(req);
5533 static int io_async_cancel_prep(struct io_kiocb *req,
5534 const struct io_uring_sqe *sqe)
5536 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5538 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5540 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5543 req->cancel.addr = READ_ONCE(sqe->addr);
5547 static int io_async_cancel(struct io_kiocb *req)
5549 struct io_ring_ctx *ctx = req->ctx;
5551 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5555 static int io_files_update_prep(struct io_kiocb *req,
5556 const struct io_uring_sqe *sqe)
5558 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5560 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5562 if (sqe->ioprio || sqe->rw_flags)
5565 req->files_update.offset = READ_ONCE(sqe->off);
5566 req->files_update.nr_args = READ_ONCE(sqe->len);
5567 if (!req->files_update.nr_args)
5569 req->files_update.arg = READ_ONCE(sqe->addr);
5573 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5574 struct io_comp_state *cs)
5576 struct io_ring_ctx *ctx = req->ctx;
5577 struct io_uring_files_update up;
5583 up.offset = req->files_update.offset;
5584 up.fds = req->files_update.arg;
5586 mutex_lock(&ctx->uring_lock);
5587 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5588 mutex_unlock(&ctx->uring_lock);
5591 req_set_fail_links(req);
5592 __io_req_complete(req, ret, 0, cs);
5596 static int io_req_defer_prep(struct io_kiocb *req,
5597 const struct io_uring_sqe *sqe)
5604 if (io_alloc_async_data(req))
5606 ret = io_prep_work_files(req);
5610 io_prep_async_work(req);
5612 switch (req->opcode) {
5615 case IORING_OP_READV:
5616 case IORING_OP_READ_FIXED:
5617 case IORING_OP_READ:
5618 ret = io_read_prep(req, sqe, true);
5620 case IORING_OP_WRITEV:
5621 case IORING_OP_WRITE_FIXED:
5622 case IORING_OP_WRITE:
5623 ret = io_write_prep(req, sqe, true);
5625 case IORING_OP_POLL_ADD:
5626 ret = io_poll_add_prep(req, sqe);
5628 case IORING_OP_POLL_REMOVE:
5629 ret = io_poll_remove_prep(req, sqe);
5631 case IORING_OP_FSYNC:
5632 ret = io_prep_fsync(req, sqe);
5634 case IORING_OP_SYNC_FILE_RANGE:
5635 ret = io_prep_sfr(req, sqe);
5637 case IORING_OP_SENDMSG:
5638 case IORING_OP_SEND:
5639 ret = io_sendmsg_prep(req, sqe);
5641 case IORING_OP_RECVMSG:
5642 case IORING_OP_RECV:
5643 ret = io_recvmsg_prep(req, sqe);
5645 case IORING_OP_CONNECT:
5646 ret = io_connect_prep(req, sqe);
5648 case IORING_OP_TIMEOUT:
5649 ret = io_timeout_prep(req, sqe, false);
5651 case IORING_OP_TIMEOUT_REMOVE:
5652 ret = io_timeout_remove_prep(req, sqe);
5654 case IORING_OP_ASYNC_CANCEL:
5655 ret = io_async_cancel_prep(req, sqe);
5657 case IORING_OP_LINK_TIMEOUT:
5658 ret = io_timeout_prep(req, sqe, true);
5660 case IORING_OP_ACCEPT:
5661 ret = io_accept_prep(req, sqe);
5663 case IORING_OP_FALLOCATE:
5664 ret = io_fallocate_prep(req, sqe);
5666 case IORING_OP_OPENAT:
5667 ret = io_openat_prep(req, sqe);
5669 case IORING_OP_CLOSE:
5670 ret = io_close_prep(req, sqe);
5672 case IORING_OP_FILES_UPDATE:
5673 ret = io_files_update_prep(req, sqe);
5675 case IORING_OP_STATX:
5676 ret = io_statx_prep(req, sqe);
5678 case IORING_OP_FADVISE:
5679 ret = io_fadvise_prep(req, sqe);
5681 case IORING_OP_MADVISE:
5682 ret = io_madvise_prep(req, sqe);
5684 case IORING_OP_OPENAT2:
5685 ret = io_openat2_prep(req, sqe);
5687 case IORING_OP_EPOLL_CTL:
5688 ret = io_epoll_ctl_prep(req, sqe);
5690 case IORING_OP_SPLICE:
5691 ret = io_splice_prep(req, sqe);
5693 case IORING_OP_PROVIDE_BUFFERS:
5694 ret = io_provide_buffers_prep(req, sqe);
5696 case IORING_OP_REMOVE_BUFFERS:
5697 ret = io_remove_buffers_prep(req, sqe);
5700 ret = io_tee_prep(req, sqe);
5703 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5712 static u32 io_get_sequence(struct io_kiocb *req)
5714 struct io_kiocb *pos;
5715 struct io_ring_ctx *ctx = req->ctx;
5716 u32 total_submitted, nr_reqs = 1;
5718 if (req->flags & REQ_F_LINK_HEAD)
5719 list_for_each_entry(pos, &req->link_list, link_list)
5722 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5723 return total_submitted - nr_reqs;
5726 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5728 struct io_ring_ctx *ctx = req->ctx;
5729 struct io_defer_entry *de;
5733 /* Still need defer if there is pending req in defer list. */
5734 if (likely(list_empty_careful(&ctx->defer_list) &&
5735 !(req->flags & REQ_F_IO_DRAIN)))
5738 seq = io_get_sequence(req);
5739 /* Still a chance to pass the sequence check */
5740 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5743 if (!req->async_data) {
5744 ret = io_req_defer_prep(req, sqe);
5748 io_prep_async_link(req);
5749 de = kmalloc(sizeof(*de), GFP_KERNEL);
5753 spin_lock_irq(&ctx->completion_lock);
5754 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5755 spin_unlock_irq(&ctx->completion_lock);
5757 io_queue_async_work(req);
5758 return -EIOCBQUEUED;
5761 trace_io_uring_defer(ctx, req, req->user_data);
5764 list_add_tail(&de->list, &ctx->defer_list);
5765 spin_unlock_irq(&ctx->completion_lock);
5766 return -EIOCBQUEUED;
5769 static void io_req_drop_files(struct io_kiocb *req)
5771 struct io_ring_ctx *ctx = req->ctx;
5772 unsigned long flags;
5774 spin_lock_irqsave(&ctx->inflight_lock, flags);
5775 list_del(&req->inflight_entry);
5776 if (waitqueue_active(&ctx->inflight_wait))
5777 wake_up(&ctx->inflight_wait);
5778 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5779 req->flags &= ~REQ_F_INFLIGHT;
5780 put_files_struct(req->work.files);
5781 put_nsproxy(req->work.nsproxy);
5782 req->work.files = NULL;
5785 static void __io_clean_op(struct io_kiocb *req)
5787 if (req->flags & REQ_F_BUFFER_SELECTED) {
5788 switch (req->opcode) {
5789 case IORING_OP_READV:
5790 case IORING_OP_READ_FIXED:
5791 case IORING_OP_READ:
5792 kfree((void *)(unsigned long)req->rw.addr);
5794 case IORING_OP_RECVMSG:
5795 case IORING_OP_RECV:
5796 kfree(req->sr_msg.kbuf);
5799 req->flags &= ~REQ_F_BUFFER_SELECTED;
5802 if (req->flags & REQ_F_NEED_CLEANUP) {
5803 switch (req->opcode) {
5804 case IORING_OP_READV:
5805 case IORING_OP_READ_FIXED:
5806 case IORING_OP_READ:
5807 case IORING_OP_WRITEV:
5808 case IORING_OP_WRITE_FIXED:
5809 case IORING_OP_WRITE: {
5810 struct io_async_rw *io = req->async_data;
5812 kfree(io->free_iovec);
5815 case IORING_OP_RECVMSG:
5816 case IORING_OP_SENDMSG: {
5817 struct io_async_msghdr *io = req->async_data;
5818 if (io->iov != io->fast_iov)
5822 case IORING_OP_SPLICE:
5824 io_put_file(req, req->splice.file_in,
5825 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5827 case IORING_OP_OPENAT:
5828 case IORING_OP_OPENAT2:
5829 if (req->open.filename)
5830 putname(req->open.filename);
5833 req->flags &= ~REQ_F_NEED_CLEANUP;
5836 if (req->flags & REQ_F_INFLIGHT)
5837 io_req_drop_files(req);
5840 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5841 bool force_nonblock, struct io_comp_state *cs)
5843 struct io_ring_ctx *ctx = req->ctx;
5846 switch (req->opcode) {
5848 ret = io_nop(req, cs);
5850 case IORING_OP_READV:
5851 case IORING_OP_READ_FIXED:
5852 case IORING_OP_READ:
5854 ret = io_read_prep(req, sqe, force_nonblock);
5858 ret = io_read(req, force_nonblock, cs);
5860 case IORING_OP_WRITEV:
5861 case IORING_OP_WRITE_FIXED:
5862 case IORING_OP_WRITE:
5864 ret = io_write_prep(req, sqe, force_nonblock);
5868 ret = io_write(req, force_nonblock, cs);
5870 case IORING_OP_FSYNC:
5872 ret = io_prep_fsync(req, sqe);
5876 ret = io_fsync(req, force_nonblock);
5878 case IORING_OP_POLL_ADD:
5880 ret = io_poll_add_prep(req, sqe);
5884 ret = io_poll_add(req);
5886 case IORING_OP_POLL_REMOVE:
5888 ret = io_poll_remove_prep(req, sqe);
5892 ret = io_poll_remove(req);
5894 case IORING_OP_SYNC_FILE_RANGE:
5896 ret = io_prep_sfr(req, sqe);
5900 ret = io_sync_file_range(req, force_nonblock);
5902 case IORING_OP_SENDMSG:
5903 case IORING_OP_SEND:
5905 ret = io_sendmsg_prep(req, sqe);
5909 if (req->opcode == IORING_OP_SENDMSG)
5910 ret = io_sendmsg(req, force_nonblock, cs);
5912 ret = io_send(req, force_nonblock, cs);
5914 case IORING_OP_RECVMSG:
5915 case IORING_OP_RECV:
5917 ret = io_recvmsg_prep(req, sqe);
5921 if (req->opcode == IORING_OP_RECVMSG)
5922 ret = io_recvmsg(req, force_nonblock, cs);
5924 ret = io_recv(req, force_nonblock, cs);
5926 case IORING_OP_TIMEOUT:
5928 ret = io_timeout_prep(req, sqe, false);
5932 ret = io_timeout(req);
5934 case IORING_OP_TIMEOUT_REMOVE:
5936 ret = io_timeout_remove_prep(req, sqe);
5940 ret = io_timeout_remove(req);
5942 case IORING_OP_ACCEPT:
5944 ret = io_accept_prep(req, sqe);
5948 ret = io_accept(req, force_nonblock, cs);
5950 case IORING_OP_CONNECT:
5952 ret = io_connect_prep(req, sqe);
5956 ret = io_connect(req, force_nonblock, cs);
5958 case IORING_OP_ASYNC_CANCEL:
5960 ret = io_async_cancel_prep(req, sqe);
5964 ret = io_async_cancel(req);
5966 case IORING_OP_FALLOCATE:
5968 ret = io_fallocate_prep(req, sqe);
5972 ret = io_fallocate(req, force_nonblock);
5974 case IORING_OP_OPENAT:
5976 ret = io_openat_prep(req, sqe);
5980 ret = io_openat(req, force_nonblock);
5982 case IORING_OP_CLOSE:
5984 ret = io_close_prep(req, sqe);
5988 ret = io_close(req, force_nonblock, cs);
5990 case IORING_OP_FILES_UPDATE:
5992 ret = io_files_update_prep(req, sqe);
5996 ret = io_files_update(req, force_nonblock, cs);
5998 case IORING_OP_STATX:
6000 ret = io_statx_prep(req, sqe);
6004 ret = io_statx(req, force_nonblock);
6006 case IORING_OP_FADVISE:
6008 ret = io_fadvise_prep(req, sqe);
6012 ret = io_fadvise(req, force_nonblock);
6014 case IORING_OP_MADVISE:
6016 ret = io_madvise_prep(req, sqe);
6020 ret = io_madvise(req, force_nonblock);
6022 case IORING_OP_OPENAT2:
6024 ret = io_openat2_prep(req, sqe);
6028 ret = io_openat2(req, force_nonblock);
6030 case IORING_OP_EPOLL_CTL:
6032 ret = io_epoll_ctl_prep(req, sqe);
6036 ret = io_epoll_ctl(req, force_nonblock, cs);
6038 case IORING_OP_SPLICE:
6040 ret = io_splice_prep(req, sqe);
6044 ret = io_splice(req, force_nonblock);
6046 case IORING_OP_PROVIDE_BUFFERS:
6048 ret = io_provide_buffers_prep(req, sqe);
6052 ret = io_provide_buffers(req, force_nonblock, cs);
6054 case IORING_OP_REMOVE_BUFFERS:
6056 ret = io_remove_buffers_prep(req, sqe);
6060 ret = io_remove_buffers(req, force_nonblock, cs);
6064 ret = io_tee_prep(req, sqe);
6068 ret = io_tee(req, force_nonblock);
6078 /* If the op doesn't have a file, we're not polling for it */
6079 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6080 const bool in_async = io_wq_current_is_worker();
6082 /* workqueue context doesn't hold uring_lock, grab it now */
6084 mutex_lock(&ctx->uring_lock);
6086 io_iopoll_req_issued(req);
6089 mutex_unlock(&ctx->uring_lock);
6095 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6097 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6098 struct io_kiocb *timeout;
6101 timeout = io_prep_linked_timeout(req);
6103 io_queue_linked_timeout(timeout);
6105 /* if NO_CANCEL is set, we must still run the work */
6106 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6107 IO_WQ_WORK_CANCEL) {
6113 ret = io_issue_sqe(req, NULL, false, NULL);
6115 * We can get EAGAIN for polled IO even though we're
6116 * forcing a sync submission from here, since we can't
6117 * wait for request slots on the block side.
6126 req_set_fail_links(req);
6127 io_req_complete(req, ret);
6130 return io_steal_work(req);
6133 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6136 struct fixed_file_table *table;
6138 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6139 return table->files[index & IORING_FILE_TABLE_MASK];
6142 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
6143 int fd, struct file **out_file, bool fixed)
6145 struct io_ring_ctx *ctx = req->ctx;
6149 if (unlikely(!ctx->file_data ||
6150 (unsigned) fd >= ctx->nr_user_files))
6152 fd = array_index_nospec(fd, ctx->nr_user_files);
6153 file = io_file_from_index(ctx, fd);
6155 req->fixed_file_refs = ctx->file_data->cur_refs;
6156 percpu_ref_get(req->fixed_file_refs);
6159 trace_io_uring_file_get(ctx, fd);
6160 file = __io_file_get(state, fd);
6163 if (file || io_op_defs[req->opcode].needs_file_no_error) {
6170 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6175 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6176 if (unlikely(!fixed && io_async_submit(req->ctx)))
6179 return io_file_get(state, req, fd, &req->file, fixed);
6182 static int io_grab_files(struct io_kiocb *req)
6184 struct io_ring_ctx *ctx = req->ctx;
6186 io_req_init_async(req);
6188 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
6191 req->work.files = get_files_struct(current);
6192 get_nsproxy(current->nsproxy);
6193 req->work.nsproxy = current->nsproxy;
6194 req->flags |= REQ_F_INFLIGHT;
6196 spin_lock_irq(&ctx->inflight_lock);
6197 list_add(&req->inflight_entry, &ctx->inflight_list);
6198 spin_unlock_irq(&ctx->inflight_lock);
6202 static inline int io_prep_work_files(struct io_kiocb *req)
6204 if (!io_op_defs[req->opcode].file_table)
6206 return io_grab_files(req);
6209 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6211 struct io_timeout_data *data = container_of(timer,
6212 struct io_timeout_data, timer);
6213 struct io_kiocb *req = data->req;
6214 struct io_ring_ctx *ctx = req->ctx;
6215 struct io_kiocb *prev = NULL;
6216 unsigned long flags;
6218 spin_lock_irqsave(&ctx->completion_lock, flags);
6221 * We don't expect the list to be empty, that will only happen if we
6222 * race with the completion of the linked work.
6224 if (!list_empty(&req->link_list)) {
6225 prev = list_entry(req->link_list.prev, struct io_kiocb,
6227 if (refcount_inc_not_zero(&prev->refs)) {
6228 list_del_init(&req->link_list);
6229 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6234 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6237 req_set_fail_links(prev);
6238 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6241 io_req_complete(req, -ETIME);
6243 return HRTIMER_NORESTART;
6246 static void __io_queue_linked_timeout(struct io_kiocb *req)
6249 * If the list is now empty, then our linked request finished before
6250 * we got a chance to setup the timer
6252 if (!list_empty(&req->link_list)) {
6253 struct io_timeout_data *data = req->async_data;
6255 data->timer.function = io_link_timeout_fn;
6256 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6261 static void io_queue_linked_timeout(struct io_kiocb *req)
6263 struct io_ring_ctx *ctx = req->ctx;
6265 spin_lock_irq(&ctx->completion_lock);
6266 __io_queue_linked_timeout(req);
6267 spin_unlock_irq(&ctx->completion_lock);
6269 /* drop submission reference */
6273 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6275 struct io_kiocb *nxt;
6277 if (!(req->flags & REQ_F_LINK_HEAD))
6279 if (req->flags & REQ_F_LINK_TIMEOUT)
6282 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6284 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6287 req->flags |= REQ_F_LINK_TIMEOUT;
6291 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6292 struct io_comp_state *cs)
6294 struct io_kiocb *linked_timeout;
6295 struct io_kiocb *nxt;
6296 const struct cred *old_creds = NULL;
6300 linked_timeout = io_prep_linked_timeout(req);
6302 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6303 req->work.creds != current_cred()) {
6305 revert_creds(old_creds);
6306 if (old_creds == req->work.creds)
6307 old_creds = NULL; /* restored original creds */
6309 old_creds = override_creds(req->work.creds);
6312 ret = io_issue_sqe(req, sqe, true, cs);
6315 * We async punt it if the file wasn't marked NOWAIT, or if the file
6316 * doesn't support non-blocking read/write attempts
6318 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6319 if (!io_arm_poll_handler(req)) {
6321 ret = io_prep_work_files(req);
6325 * Queued up for async execution, worker will release
6326 * submit reference when the iocb is actually submitted.
6328 io_queue_async_work(req);
6332 io_queue_linked_timeout(linked_timeout);
6336 if (unlikely(ret)) {
6338 /* un-prep timeout, so it'll be killed as any other linked */
6339 req->flags &= ~REQ_F_LINK_TIMEOUT;
6340 req_set_fail_links(req);
6342 io_req_complete(req, ret);
6346 /* drop submission reference */
6347 nxt = io_put_req_find_next(req);
6349 io_queue_linked_timeout(linked_timeout);
6354 if (req->flags & REQ_F_FORCE_ASYNC)
6360 revert_creds(old_creds);
6363 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6364 struct io_comp_state *cs)
6368 ret = io_req_defer(req, sqe);
6370 if (ret != -EIOCBQUEUED) {
6372 req_set_fail_links(req);
6374 io_req_complete(req, ret);
6376 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6377 if (!req->async_data) {
6378 ret = io_req_defer_prep(req, sqe);
6384 * Never try inline submit of IOSQE_ASYNC is set, go straight
6385 * to async execution.
6387 io_req_init_async(req);
6388 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6389 io_queue_async_work(req);
6391 __io_queue_sqe(req, sqe, cs);
6395 static inline void io_queue_link_head(struct io_kiocb *req,
6396 struct io_comp_state *cs)
6398 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6400 io_req_complete(req, -ECANCELED);
6402 io_queue_sqe(req, NULL, cs);
6405 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6406 struct io_kiocb **link, struct io_comp_state *cs)
6408 struct io_ring_ctx *ctx = req->ctx;
6412 * If we already have a head request, queue this one for async
6413 * submittal once the head completes. If we don't have a head but
6414 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6415 * submitted sync once the chain is complete. If none of those
6416 * conditions are true (normal request), then just queue it.
6419 struct io_kiocb *head = *link;
6422 * Taking sequential execution of a link, draining both sides
6423 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6424 * requests in the link. So, it drains the head and the
6425 * next after the link request. The last one is done via
6426 * drain_next flag to persist the effect across calls.
6428 if (req->flags & REQ_F_IO_DRAIN) {
6429 head->flags |= REQ_F_IO_DRAIN;
6430 ctx->drain_next = 1;
6432 ret = io_req_defer_prep(req, sqe);
6433 if (unlikely(ret)) {
6434 /* fail even hard links since we don't submit */
6435 head->flags |= REQ_F_FAIL_LINK;
6438 trace_io_uring_link(ctx, req, head);
6439 list_add_tail(&req->link_list, &head->link_list);
6441 /* last request of a link, enqueue the link */
6442 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6443 io_queue_link_head(head, cs);
6447 if (unlikely(ctx->drain_next)) {
6448 req->flags |= REQ_F_IO_DRAIN;
6449 ctx->drain_next = 0;
6451 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6452 req->flags |= REQ_F_LINK_HEAD;
6453 INIT_LIST_HEAD(&req->link_list);
6455 ret = io_req_defer_prep(req, sqe);
6457 req->flags |= REQ_F_FAIL_LINK;
6460 io_queue_sqe(req, sqe, cs);
6468 * Batched submission is done, ensure local IO is flushed out.
6470 static void io_submit_state_end(struct io_submit_state *state)
6472 if (!list_empty(&state->comp.list))
6473 io_submit_flush_completions(&state->comp);
6474 blk_finish_plug(&state->plug);
6475 io_state_file_put(state);
6476 if (state->free_reqs)
6477 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6481 * Start submission side cache.
6483 static void io_submit_state_start(struct io_submit_state *state,
6484 struct io_ring_ctx *ctx, unsigned int max_ios)
6486 blk_start_plug(&state->plug);
6488 INIT_LIST_HEAD(&state->comp.list);
6489 state->comp.ctx = ctx;
6490 state->free_reqs = 0;
6492 state->ios_left = max_ios;
6495 static void io_commit_sqring(struct io_ring_ctx *ctx)
6497 struct io_rings *rings = ctx->rings;
6500 * Ensure any loads from the SQEs are done at this point,
6501 * since once we write the new head, the application could
6502 * write new data to them.
6504 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6508 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6509 * that is mapped by userspace. This means that care needs to be taken to
6510 * ensure that reads are stable, as we cannot rely on userspace always
6511 * being a good citizen. If members of the sqe are validated and then later
6512 * used, it's important that those reads are done through READ_ONCE() to
6513 * prevent a re-load down the line.
6515 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6517 u32 *sq_array = ctx->sq_array;
6521 * The cached sq head (or cq tail) serves two purposes:
6523 * 1) allows us to batch the cost of updating the user visible
6525 * 2) allows the kernel side to track the head on its own, even
6526 * though the application is the one updating it.
6528 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6529 if (likely(head < ctx->sq_entries))
6530 return &ctx->sq_sqes[head];
6532 /* drop invalid entries */
6533 ctx->cached_sq_dropped++;
6534 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6538 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6540 ctx->cached_sq_head++;
6544 * Check SQE restrictions (opcode and flags).
6546 * Returns 'true' if SQE is allowed, 'false' otherwise.
6548 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6549 struct io_kiocb *req,
6550 unsigned int sqe_flags)
6552 if (!ctx->restricted)
6555 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6558 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6559 ctx->restrictions.sqe_flags_required)
6562 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6563 ctx->restrictions.sqe_flags_required))
6569 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6570 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6571 IOSQE_BUFFER_SELECT)
6573 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6574 const struct io_uring_sqe *sqe,
6575 struct io_submit_state *state)
6577 unsigned int sqe_flags;
6580 req->opcode = READ_ONCE(sqe->opcode);
6581 req->user_data = READ_ONCE(sqe->user_data);
6582 req->async_data = NULL;
6586 /* one is dropped after submission, the other at completion */
6587 refcount_set(&req->refs, 2);
6588 req->task = current;
6589 get_task_struct(req->task);
6590 atomic_long_inc(&req->task->io_uring->req_issue);
6593 if (unlikely(req->opcode >= IORING_OP_LAST))
6596 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6599 sqe_flags = READ_ONCE(sqe->flags);
6600 /* enforce forwards compatibility on users */
6601 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6604 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6607 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6608 !io_op_defs[req->opcode].buffer_select)
6611 id = READ_ONCE(sqe->personality);
6613 io_req_init_async(req);
6614 req->work.creds = idr_find(&ctx->personality_idr, id);
6615 if (unlikely(!req->work.creds))
6617 get_cred(req->work.creds);
6620 /* same numerical values with corresponding REQ_F_*, safe to copy */
6621 req->flags |= sqe_flags;
6623 if (!io_op_defs[req->opcode].needs_file)
6626 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6629 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6631 struct io_submit_state state;
6632 struct io_kiocb *link = NULL;
6633 int i, submitted = 0;
6635 /* if we have a backlog and couldn't flush it all, return BUSY */
6636 if (test_bit(0, &ctx->sq_check_overflow)) {
6637 if (!list_empty(&ctx->cq_overflow_list) &&
6638 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6642 /* make sure SQ entry isn't read before tail */
6643 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6645 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6648 io_submit_state_start(&state, ctx, nr);
6650 for (i = 0; i < nr; i++) {
6651 const struct io_uring_sqe *sqe;
6652 struct io_kiocb *req;
6655 sqe = io_get_sqe(ctx);
6656 if (unlikely(!sqe)) {
6657 io_consume_sqe(ctx);
6660 req = io_alloc_req(ctx, &state);
6661 if (unlikely(!req)) {
6663 submitted = -EAGAIN;
6667 err = io_init_req(ctx, req, sqe, &state);
6668 io_consume_sqe(ctx);
6669 /* will complete beyond this point, count as submitted */
6672 if (unlikely(err)) {
6675 io_req_complete(req, err);
6679 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6680 true, io_async_submit(ctx));
6681 err = io_submit_sqe(req, sqe, &link, &state.comp);
6686 if (unlikely(submitted != nr)) {
6687 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6689 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6692 io_queue_link_head(link, &state.comp);
6693 io_submit_state_end(&state);
6695 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6696 io_commit_sqring(ctx);
6701 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6703 /* Tell userspace we may need a wakeup call */
6704 spin_lock_irq(&ctx->completion_lock);
6705 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6706 spin_unlock_irq(&ctx->completion_lock);
6709 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6711 spin_lock_irq(&ctx->completion_lock);
6712 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6713 spin_unlock_irq(&ctx->completion_lock);
6716 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6717 int sync, void *key)
6719 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6722 ret = autoremove_wake_function(wqe, mode, sync, key);
6724 unsigned long flags;
6726 spin_lock_irqsave(&ctx->completion_lock, flags);
6727 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6728 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6739 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6740 unsigned long start_jiffies, bool cap_entries)
6742 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6743 struct io_sq_data *sqd = ctx->sq_data;
6744 unsigned int to_submit;
6748 if (!list_empty(&ctx->iopoll_list)) {
6749 unsigned nr_events = 0;
6751 mutex_lock(&ctx->uring_lock);
6752 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6753 io_do_iopoll(ctx, &nr_events, 0);
6754 mutex_unlock(&ctx->uring_lock);
6757 to_submit = io_sqring_entries(ctx);
6760 * If submit got -EBUSY, flag us as needing the application
6761 * to enter the kernel to reap and flush events.
6763 if (!to_submit || ret == -EBUSY || need_resched()) {
6765 * Drop cur_mm before scheduling, we can't hold it for
6766 * long periods (or over schedule()). Do this before
6767 * adding ourselves to the waitqueue, as the unuse/drop
6770 io_sq_thread_drop_mm();
6773 * We're polling. If we're within the defined idle
6774 * period, then let us spin without work before going
6775 * to sleep. The exception is if we got EBUSY doing
6776 * more IO, we should wait for the application to
6777 * reap events and wake us up.
6779 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6780 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6781 !percpu_ref_is_dying(&ctx->refs)))
6784 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6785 TASK_INTERRUPTIBLE);
6788 * While doing polled IO, before going to sleep, we need
6789 * to check if there are new reqs added to iopoll_list,
6790 * it is because reqs may have been punted to io worker
6791 * and will be added to iopoll_list later, hence check
6792 * the iopoll_list again.
6794 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6795 !list_empty_careful(&ctx->iopoll_list)) {
6796 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6800 to_submit = io_sqring_entries(ctx);
6801 if (!to_submit || ret == -EBUSY)
6805 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6806 io_ring_clear_wakeup_flag(ctx);
6808 /* if we're handling multiple rings, cap submit size for fairness */
6809 if (cap_entries && to_submit > 8)
6812 mutex_lock(&ctx->uring_lock);
6813 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6814 ret = io_submit_sqes(ctx, to_submit);
6815 mutex_unlock(&ctx->uring_lock);
6817 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6818 wake_up(&ctx->sqo_sq_wait);
6820 return SQT_DID_WORK;
6823 static void io_sqd_init_new(struct io_sq_data *sqd)
6825 struct io_ring_ctx *ctx;
6827 while (!list_empty(&sqd->ctx_new_list)) {
6828 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6829 init_wait(&ctx->sqo_wait_entry);
6830 ctx->sqo_wait_entry.func = io_sq_wake_function;
6831 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6832 complete(&ctx->sq_thread_comp);
6836 static int io_sq_thread(void *data)
6838 struct cgroup_subsys_state *cur_css = NULL;
6839 const struct cred *old_cred = NULL;
6840 struct io_sq_data *sqd = data;
6841 struct io_ring_ctx *ctx;
6842 unsigned long start_jiffies;
6844 start_jiffies = jiffies;
6845 while (!kthread_should_stop()) {
6846 enum sq_ret ret = 0;
6850 * Any changes to the sqd lists are synchronized through the
6851 * kthread parking. This synchronizes the thread vs users,
6852 * the users are synchronized on the sqd->ctx_lock.
6854 if (kthread_should_park())
6857 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6858 io_sqd_init_new(sqd);
6860 cap_entries = !list_is_singular(&sqd->ctx_list);
6862 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6863 if (current->cred != ctx->creds) {
6865 revert_creds(old_cred);
6866 old_cred = override_creds(ctx->creds);
6868 io_sq_thread_associate_blkcg(ctx, &cur_css);
6870 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6872 io_sq_thread_drop_mm();
6875 if (ret & SQT_SPIN) {
6878 } else if (ret == SQT_IDLE) {
6879 if (kthread_should_park())
6881 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6882 io_ring_set_wakeup_flag(ctx);
6884 start_jiffies = jiffies;
6885 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6886 io_ring_clear_wakeup_flag(ctx);
6893 io_sq_thread_unassociate_blkcg();
6895 revert_creds(old_cred);
6902 struct io_wait_queue {
6903 struct wait_queue_entry wq;
6904 struct io_ring_ctx *ctx;
6906 unsigned nr_timeouts;
6909 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6911 struct io_ring_ctx *ctx = iowq->ctx;
6914 * Wake up if we have enough events, or if a timeout occurred since we
6915 * started waiting. For timeouts, we always want to return to userspace,
6916 * regardless of event count.
6918 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6919 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6922 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6923 int wake_flags, void *key)
6925 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6928 /* use noflush == true, as we can't safely rely on locking context */
6929 if (!io_should_wake(iowq, true))
6932 return autoremove_wake_function(curr, mode, wake_flags, key);
6935 static int io_run_task_work_sig(void)
6937 if (io_run_task_work())
6939 if (!signal_pending(current))
6941 if (current->jobctl & JOBCTL_TASK_WORK) {
6942 spin_lock_irq(¤t->sighand->siglock);
6943 current->jobctl &= ~JOBCTL_TASK_WORK;
6944 recalc_sigpending();
6945 spin_unlock_irq(¤t->sighand->siglock);
6952 * Wait until events become available, if we don't already have some. The
6953 * application must reap them itself, as they reside on the shared cq ring.
6955 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6956 const sigset_t __user *sig, size_t sigsz)
6958 struct io_wait_queue iowq = {
6961 .func = io_wake_function,
6962 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6965 .to_wait = min_events,
6967 struct io_rings *rings = ctx->rings;
6971 if (io_cqring_events(ctx, false) >= min_events)
6973 if (!io_run_task_work())
6978 #ifdef CONFIG_COMPAT
6979 if (in_compat_syscall())
6980 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6984 ret = set_user_sigmask(sig, sigsz);
6990 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6991 trace_io_uring_cqring_wait(ctx, min_events);
6993 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6994 TASK_INTERRUPTIBLE);
6995 /* make sure we run task_work before checking for signals */
6996 ret = io_run_task_work_sig();
7001 if (io_should_wake(&iowq, false))
7005 finish_wait(&ctx->wait, &iowq.wq);
7007 restore_saved_sigmask_unless(ret == -EINTR);
7009 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7012 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7014 #if defined(CONFIG_UNIX)
7015 if (ctx->ring_sock) {
7016 struct sock *sock = ctx->ring_sock->sk;
7017 struct sk_buff *skb;
7019 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7025 for (i = 0; i < ctx->nr_user_files; i++) {
7028 file = io_file_from_index(ctx, i);
7035 static void io_file_ref_kill(struct percpu_ref *ref)
7037 struct fixed_file_data *data;
7039 data = container_of(ref, struct fixed_file_data, refs);
7040 complete(&data->done);
7043 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7045 struct fixed_file_data *data = ctx->file_data;
7046 struct fixed_file_ref_node *ref_node = NULL;
7047 unsigned nr_tables, i;
7052 spin_lock(&data->lock);
7053 if (!list_empty(&data->ref_list))
7054 ref_node = list_first_entry(&data->ref_list,
7055 struct fixed_file_ref_node, node);
7056 spin_unlock(&data->lock);
7058 percpu_ref_kill(&ref_node->refs);
7060 percpu_ref_kill(&data->refs);
7062 /* wait for all refs nodes to complete */
7063 flush_delayed_work(&ctx->file_put_work);
7064 wait_for_completion(&data->done);
7066 __io_sqe_files_unregister(ctx);
7067 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7068 for (i = 0; i < nr_tables; i++)
7069 kfree(data->table[i].files);
7071 percpu_ref_exit(&data->refs);
7073 ctx->file_data = NULL;
7074 ctx->nr_user_files = 0;
7078 static void io_put_sq_data(struct io_sq_data *sqd)
7080 if (refcount_dec_and_test(&sqd->refs)) {
7082 * The park is a bit of a work-around, without it we get
7083 * warning spews on shutdown with SQPOLL set and affinity
7084 * set to a single CPU.
7087 kthread_park(sqd->thread);
7088 kthread_stop(sqd->thread);
7095 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7097 struct io_ring_ctx *ctx_attach;
7098 struct io_sq_data *sqd;
7101 f = fdget(p->wq_fd);
7103 return ERR_PTR(-ENXIO);
7104 if (f.file->f_op != &io_uring_fops) {
7106 return ERR_PTR(-EINVAL);
7109 ctx_attach = f.file->private_data;
7110 sqd = ctx_attach->sq_data;
7113 return ERR_PTR(-EINVAL);
7116 refcount_inc(&sqd->refs);
7121 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7123 struct io_sq_data *sqd;
7125 if (p->flags & IORING_SETUP_ATTACH_WQ)
7126 return io_attach_sq_data(p);
7128 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7130 return ERR_PTR(-ENOMEM);
7132 refcount_set(&sqd->refs, 1);
7133 INIT_LIST_HEAD(&sqd->ctx_list);
7134 INIT_LIST_HEAD(&sqd->ctx_new_list);
7135 mutex_init(&sqd->ctx_lock);
7136 mutex_init(&sqd->lock);
7137 init_waitqueue_head(&sqd->wait);
7141 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7142 __releases(&sqd->lock)
7146 kthread_unpark(sqd->thread);
7147 mutex_unlock(&sqd->lock);
7150 static void io_sq_thread_park(struct io_sq_data *sqd)
7151 __acquires(&sqd->lock)
7155 mutex_lock(&sqd->lock);
7156 kthread_park(sqd->thread);
7159 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7161 struct io_sq_data *sqd = ctx->sq_data;
7166 * We may arrive here from the error branch in
7167 * io_sq_offload_create() where the kthread is created
7168 * without being waked up, thus wake it up now to make
7169 * sure the wait will complete.
7171 wake_up_process(sqd->thread);
7172 wait_for_completion(&ctx->sq_thread_comp);
7174 io_sq_thread_park(sqd);
7177 mutex_lock(&sqd->ctx_lock);
7178 list_del(&ctx->sqd_list);
7179 mutex_unlock(&sqd->ctx_lock);
7182 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7183 io_sq_thread_unpark(sqd);
7186 io_put_sq_data(sqd);
7187 ctx->sq_data = NULL;
7191 static void io_finish_async(struct io_ring_ctx *ctx)
7193 io_sq_thread_stop(ctx);
7196 io_wq_destroy(ctx->io_wq);
7201 #if defined(CONFIG_UNIX)
7203 * Ensure the UNIX gc is aware of our file set, so we are certain that
7204 * the io_uring can be safely unregistered on process exit, even if we have
7205 * loops in the file referencing.
7207 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7209 struct sock *sk = ctx->ring_sock->sk;
7210 struct scm_fp_list *fpl;
7211 struct sk_buff *skb;
7214 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7218 skb = alloc_skb(0, GFP_KERNEL);
7227 fpl->user = get_uid(ctx->user);
7228 for (i = 0; i < nr; i++) {
7229 struct file *file = io_file_from_index(ctx, i + offset);
7233 fpl->fp[nr_files] = get_file(file);
7234 unix_inflight(fpl->user, fpl->fp[nr_files]);
7239 fpl->max = SCM_MAX_FD;
7240 fpl->count = nr_files;
7241 UNIXCB(skb).fp = fpl;
7242 skb->destructor = unix_destruct_scm;
7243 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7244 skb_queue_head(&sk->sk_receive_queue, skb);
7246 for (i = 0; i < nr_files; i++)
7257 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7258 * causes regular reference counting to break down. We rely on the UNIX
7259 * garbage collection to take care of this problem for us.
7261 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7263 unsigned left, total;
7267 left = ctx->nr_user_files;
7269 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7271 ret = __io_sqe_files_scm(ctx, this_files, total);
7275 total += this_files;
7281 while (total < ctx->nr_user_files) {
7282 struct file *file = io_file_from_index(ctx, total);
7292 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7298 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
7303 for (i = 0; i < nr_tables; i++) {
7304 struct fixed_file_table *table = &ctx->file_data->table[i];
7305 unsigned this_files;
7307 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7308 table->files = kcalloc(this_files, sizeof(struct file *),
7312 nr_files -= this_files;
7318 for (i = 0; i < nr_tables; i++) {
7319 struct fixed_file_table *table = &ctx->file_data->table[i];
7320 kfree(table->files);
7325 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7327 #if defined(CONFIG_UNIX)
7328 struct sock *sock = ctx->ring_sock->sk;
7329 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7330 struct sk_buff *skb;
7333 __skb_queue_head_init(&list);
7336 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7337 * remove this entry and rearrange the file array.
7339 skb = skb_dequeue(head);
7341 struct scm_fp_list *fp;
7343 fp = UNIXCB(skb).fp;
7344 for (i = 0; i < fp->count; i++) {
7347 if (fp->fp[i] != file)
7350 unix_notinflight(fp->user, fp->fp[i]);
7351 left = fp->count - 1 - i;
7353 memmove(&fp->fp[i], &fp->fp[i + 1],
7354 left * sizeof(struct file *));
7361 __skb_queue_tail(&list, skb);
7371 __skb_queue_tail(&list, skb);
7373 skb = skb_dequeue(head);
7376 if (skb_peek(&list)) {
7377 spin_lock_irq(&head->lock);
7378 while ((skb = __skb_dequeue(&list)) != NULL)
7379 __skb_queue_tail(head, skb);
7380 spin_unlock_irq(&head->lock);
7387 struct io_file_put {
7388 struct list_head list;
7392 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7394 struct fixed_file_data *file_data = ref_node->file_data;
7395 struct io_ring_ctx *ctx = file_data->ctx;
7396 struct io_file_put *pfile, *tmp;
7398 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7399 list_del(&pfile->list);
7400 io_ring_file_put(ctx, pfile->file);
7404 spin_lock(&file_data->lock);
7405 list_del(&ref_node->node);
7406 spin_unlock(&file_data->lock);
7408 percpu_ref_exit(&ref_node->refs);
7410 percpu_ref_put(&file_data->refs);
7413 static void io_file_put_work(struct work_struct *work)
7415 struct io_ring_ctx *ctx;
7416 struct llist_node *node;
7418 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7419 node = llist_del_all(&ctx->file_put_llist);
7422 struct fixed_file_ref_node *ref_node;
7423 struct llist_node *next = node->next;
7425 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7426 __io_file_put_work(ref_node);
7431 static void io_file_data_ref_zero(struct percpu_ref *ref)
7433 struct fixed_file_ref_node *ref_node;
7434 struct io_ring_ctx *ctx;
7438 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7439 ctx = ref_node->file_data->ctx;
7441 if (percpu_ref_is_dying(&ctx->file_data->refs))
7444 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7446 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7448 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7451 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7452 struct io_ring_ctx *ctx)
7454 struct fixed_file_ref_node *ref_node;
7456 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7458 return ERR_PTR(-ENOMEM);
7460 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7463 return ERR_PTR(-ENOMEM);
7465 INIT_LIST_HEAD(&ref_node->node);
7466 INIT_LIST_HEAD(&ref_node->file_list);
7467 ref_node->file_data = ctx->file_data;
7471 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7473 percpu_ref_exit(&ref_node->refs);
7477 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7480 __s32 __user *fds = (__s32 __user *) arg;
7485 struct fixed_file_ref_node *ref_node;
7491 if (nr_args > IORING_MAX_FIXED_FILES)
7494 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7495 if (!ctx->file_data)
7497 ctx->file_data->ctx = ctx;
7498 init_completion(&ctx->file_data->done);
7499 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7500 spin_lock_init(&ctx->file_data->lock);
7502 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7503 ctx->file_data->table = kcalloc(nr_tables,
7504 sizeof(struct fixed_file_table),
7506 if (!ctx->file_data->table) {
7507 kfree(ctx->file_data);
7508 ctx->file_data = NULL;
7512 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7513 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7514 kfree(ctx->file_data->table);
7515 kfree(ctx->file_data);
7516 ctx->file_data = NULL;
7520 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7521 percpu_ref_exit(&ctx->file_data->refs);
7522 kfree(ctx->file_data->table);
7523 kfree(ctx->file_data);
7524 ctx->file_data = NULL;
7528 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7529 struct fixed_file_table *table;
7533 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7535 /* allow sparse sets */
7541 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7542 index = i & IORING_FILE_TABLE_MASK;
7550 * Don't allow io_uring instances to be registered. If UNIX
7551 * isn't enabled, then this causes a reference cycle and this
7552 * instance can never get freed. If UNIX is enabled we'll
7553 * handle it just fine, but there's still no point in allowing
7554 * a ring fd as it doesn't support regular read/write anyway.
7556 if (file->f_op == &io_uring_fops) {
7561 table->files[index] = file;
7565 for (i = 0; i < ctx->nr_user_files; i++) {
7566 file = io_file_from_index(ctx, i);
7570 for (i = 0; i < nr_tables; i++)
7571 kfree(ctx->file_data->table[i].files);
7573 percpu_ref_exit(&ctx->file_data->refs);
7574 kfree(ctx->file_data->table);
7575 kfree(ctx->file_data);
7576 ctx->file_data = NULL;
7577 ctx->nr_user_files = 0;
7581 ret = io_sqe_files_scm(ctx);
7583 io_sqe_files_unregister(ctx);
7587 ref_node = alloc_fixed_file_ref_node(ctx);
7588 if (IS_ERR(ref_node)) {
7589 io_sqe_files_unregister(ctx);
7590 return PTR_ERR(ref_node);
7593 ctx->file_data->cur_refs = &ref_node->refs;
7594 spin_lock(&ctx->file_data->lock);
7595 list_add(&ref_node->node, &ctx->file_data->ref_list);
7596 spin_unlock(&ctx->file_data->lock);
7597 percpu_ref_get(&ctx->file_data->refs);
7601 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7604 #if defined(CONFIG_UNIX)
7605 struct sock *sock = ctx->ring_sock->sk;
7606 struct sk_buff_head *head = &sock->sk_receive_queue;
7607 struct sk_buff *skb;
7610 * See if we can merge this file into an existing skb SCM_RIGHTS
7611 * file set. If there's no room, fall back to allocating a new skb
7612 * and filling it in.
7614 spin_lock_irq(&head->lock);
7615 skb = skb_peek(head);
7617 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7619 if (fpl->count < SCM_MAX_FD) {
7620 __skb_unlink(skb, head);
7621 spin_unlock_irq(&head->lock);
7622 fpl->fp[fpl->count] = get_file(file);
7623 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7625 spin_lock_irq(&head->lock);
7626 __skb_queue_head(head, skb);
7631 spin_unlock_irq(&head->lock);
7638 return __io_sqe_files_scm(ctx, 1, index);
7644 static int io_queue_file_removal(struct fixed_file_data *data,
7647 struct io_file_put *pfile;
7648 struct percpu_ref *refs = data->cur_refs;
7649 struct fixed_file_ref_node *ref_node;
7651 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7655 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7657 list_add(&pfile->list, &ref_node->file_list);
7662 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7663 struct io_uring_files_update *up,
7666 struct fixed_file_data *data = ctx->file_data;
7667 struct fixed_file_ref_node *ref_node;
7672 bool needs_switch = false;
7674 if (check_add_overflow(up->offset, nr_args, &done))
7676 if (done > ctx->nr_user_files)
7679 ref_node = alloc_fixed_file_ref_node(ctx);
7680 if (IS_ERR(ref_node))
7681 return PTR_ERR(ref_node);
7684 fds = u64_to_user_ptr(up->fds);
7686 struct fixed_file_table *table;
7690 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7694 i = array_index_nospec(up->offset, ctx->nr_user_files);
7695 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7696 index = i & IORING_FILE_TABLE_MASK;
7697 if (table->files[index]) {
7698 file = table->files[index];
7699 err = io_queue_file_removal(data, file);
7702 table->files[index] = NULL;
7703 needs_switch = true;
7712 * Don't allow io_uring instances to be registered. If
7713 * UNIX isn't enabled, then this causes a reference
7714 * cycle and this instance can never get freed. If UNIX
7715 * is enabled we'll handle it just fine, but there's
7716 * still no point in allowing a ring fd as it doesn't
7717 * support regular read/write anyway.
7719 if (file->f_op == &io_uring_fops) {
7724 table->files[index] = file;
7725 err = io_sqe_file_register(ctx, file, i);
7727 table->files[index] = NULL;
7738 percpu_ref_kill(data->cur_refs);
7739 spin_lock(&data->lock);
7740 list_add(&ref_node->node, &data->ref_list);
7741 data->cur_refs = &ref_node->refs;
7742 spin_unlock(&data->lock);
7743 percpu_ref_get(&ctx->file_data->refs);
7745 destroy_fixed_file_ref_node(ref_node);
7747 return done ? done : err;
7750 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7753 struct io_uring_files_update up;
7755 if (!ctx->file_data)
7759 if (copy_from_user(&up, arg, sizeof(up)))
7764 return __io_sqe_files_update(ctx, &up, nr_args);
7767 static void io_free_work(struct io_wq_work *work)
7769 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7771 /* Consider that io_steal_work() relies on this ref */
7775 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7776 struct io_uring_params *p)
7778 struct io_wq_data data;
7780 struct io_ring_ctx *ctx_attach;
7781 unsigned int concurrency;
7784 data.user = ctx->user;
7785 data.free_work = io_free_work;
7786 data.do_work = io_wq_submit_work;
7788 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7789 /* Do QD, or 4 * CPUS, whatever is smallest */
7790 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7792 ctx->io_wq = io_wq_create(concurrency, &data);
7793 if (IS_ERR(ctx->io_wq)) {
7794 ret = PTR_ERR(ctx->io_wq);
7800 f = fdget(p->wq_fd);
7804 if (f.file->f_op != &io_uring_fops) {
7809 ctx_attach = f.file->private_data;
7810 /* @io_wq is protected by holding the fd */
7811 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7816 ctx->io_wq = ctx_attach->io_wq;
7822 static int io_uring_alloc_task_context(struct task_struct *task)
7824 struct io_uring_task *tctx;
7826 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7827 if (unlikely(!tctx))
7831 init_waitqueue_head(&tctx->wait);
7834 atomic_long_set(&tctx->req_issue, 0);
7835 atomic_long_set(&tctx->req_complete, 0);
7836 task->io_uring = tctx;
7840 void __io_uring_free(struct task_struct *tsk)
7842 struct io_uring_task *tctx = tsk->io_uring;
7844 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7845 xa_destroy(&tctx->xa);
7847 tsk->io_uring = NULL;
7850 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7851 struct io_uring_params *p)
7855 if (ctx->flags & IORING_SETUP_SQPOLL) {
7856 struct io_sq_data *sqd;
7859 if (!capable(CAP_SYS_ADMIN))
7862 sqd = io_get_sq_data(p);
7869 io_sq_thread_park(sqd);
7870 mutex_lock(&sqd->ctx_lock);
7871 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7872 mutex_unlock(&sqd->ctx_lock);
7873 io_sq_thread_unpark(sqd);
7875 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7876 if (!ctx->sq_thread_idle)
7877 ctx->sq_thread_idle = HZ;
7882 if (p->flags & IORING_SETUP_SQ_AFF) {
7883 int cpu = p->sq_thread_cpu;
7886 if (cpu >= nr_cpu_ids)
7888 if (!cpu_online(cpu))
7891 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7892 cpu, "io_uring-sq");
7894 sqd->thread = kthread_create(io_sq_thread, sqd,
7897 if (IS_ERR(sqd->thread)) {
7898 ret = PTR_ERR(sqd->thread);
7902 ret = io_uring_alloc_task_context(sqd->thread);
7905 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7906 /* Can't have SQ_AFF without SQPOLL */
7912 ret = io_init_wq_offload(ctx, p);
7918 io_finish_async(ctx);
7922 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7924 struct io_sq_data *sqd = ctx->sq_data;
7926 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7927 wake_up_process(sqd->thread);
7930 static inline void __io_unaccount_mem(struct user_struct *user,
7931 unsigned long nr_pages)
7933 atomic_long_sub(nr_pages, &user->locked_vm);
7936 static inline int __io_account_mem(struct user_struct *user,
7937 unsigned long nr_pages)
7939 unsigned long page_limit, cur_pages, new_pages;
7941 /* Don't allow more pages than we can safely lock */
7942 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7945 cur_pages = atomic_long_read(&user->locked_vm);
7946 new_pages = cur_pages + nr_pages;
7947 if (new_pages > page_limit)
7949 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7950 new_pages) != cur_pages);
7955 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7956 enum io_mem_account acct)
7959 __io_unaccount_mem(ctx->user, nr_pages);
7961 if (ctx->mm_account) {
7962 if (acct == ACCT_LOCKED)
7963 ctx->mm_account->locked_vm -= nr_pages;
7964 else if (acct == ACCT_PINNED)
7965 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7969 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7970 enum io_mem_account acct)
7974 if (ctx->limit_mem) {
7975 ret = __io_account_mem(ctx->user, nr_pages);
7980 if (ctx->mm_account) {
7981 if (acct == ACCT_LOCKED)
7982 ctx->mm_account->locked_vm += nr_pages;
7983 else if (acct == ACCT_PINNED)
7984 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7990 static void io_mem_free(void *ptr)
7997 page = virt_to_head_page(ptr);
7998 if (put_page_testzero(page))
7999 free_compound_page(page);
8002 static void *io_mem_alloc(size_t size)
8004 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8007 return (void *) __get_free_pages(gfp_flags, get_order(size));
8010 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8013 struct io_rings *rings;
8014 size_t off, sq_array_size;
8016 off = struct_size(rings, cqes, cq_entries);
8017 if (off == SIZE_MAX)
8021 off = ALIGN(off, SMP_CACHE_BYTES);
8029 sq_array_size = array_size(sizeof(u32), sq_entries);
8030 if (sq_array_size == SIZE_MAX)
8033 if (check_add_overflow(off, sq_array_size, &off))
8039 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8043 pages = (size_t)1 << get_order(
8044 rings_size(sq_entries, cq_entries, NULL));
8045 pages += (size_t)1 << get_order(
8046 array_size(sizeof(struct io_uring_sqe), sq_entries));
8051 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8055 if (!ctx->user_bufs)
8058 for (i = 0; i < ctx->nr_user_bufs; i++) {
8059 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8061 for (j = 0; j < imu->nr_bvecs; j++)
8062 unpin_user_page(imu->bvec[j].bv_page);
8064 if (imu->acct_pages)
8065 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8070 kfree(ctx->user_bufs);
8071 ctx->user_bufs = NULL;
8072 ctx->nr_user_bufs = 0;
8076 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8077 void __user *arg, unsigned index)
8079 struct iovec __user *src;
8081 #ifdef CONFIG_COMPAT
8083 struct compat_iovec __user *ciovs;
8084 struct compat_iovec ciov;
8086 ciovs = (struct compat_iovec __user *) arg;
8087 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8090 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8091 dst->iov_len = ciov.iov_len;
8095 src = (struct iovec __user *) arg;
8096 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8102 * Not super efficient, but this is just a registration time. And we do cache
8103 * the last compound head, so generally we'll only do a full search if we don't
8106 * We check if the given compound head page has already been accounted, to
8107 * avoid double accounting it. This allows us to account the full size of the
8108 * page, not just the constituent pages of a huge page.
8110 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8111 int nr_pages, struct page *hpage)
8115 /* check current page array */
8116 for (i = 0; i < nr_pages; i++) {
8117 if (!PageCompound(pages[i]))
8119 if (compound_head(pages[i]) == hpage)
8123 /* check previously registered pages */
8124 for (i = 0; i < ctx->nr_user_bufs; i++) {
8125 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8127 for (j = 0; j < imu->nr_bvecs; j++) {
8128 if (!PageCompound(imu->bvec[j].bv_page))
8130 if (compound_head(imu->bvec[j].bv_page) == hpage)
8138 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8139 int nr_pages, struct io_mapped_ubuf *imu,
8140 struct page **last_hpage)
8144 for (i = 0; i < nr_pages; i++) {
8145 if (!PageCompound(pages[i])) {
8150 hpage = compound_head(pages[i]);
8151 if (hpage == *last_hpage)
8153 *last_hpage = hpage;
8154 if (headpage_already_acct(ctx, pages, i, hpage))
8156 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8160 if (!imu->acct_pages)
8163 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8165 imu->acct_pages = 0;
8169 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8172 struct vm_area_struct **vmas = NULL;
8173 struct page **pages = NULL;
8174 struct page *last_hpage = NULL;
8175 int i, j, got_pages = 0;
8180 if (!nr_args || nr_args > UIO_MAXIOV)
8183 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8185 if (!ctx->user_bufs)
8188 for (i = 0; i < nr_args; i++) {
8189 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8190 unsigned long off, start, end, ubuf;
8195 ret = io_copy_iov(ctx, &iov, arg, i);
8200 * Don't impose further limits on the size and buffer
8201 * constraints here, we'll -EINVAL later when IO is
8202 * submitted if they are wrong.
8205 if (!iov.iov_base || !iov.iov_len)
8208 /* arbitrary limit, but we need something */
8209 if (iov.iov_len > SZ_1G)
8212 ubuf = (unsigned long) iov.iov_base;
8213 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8214 start = ubuf >> PAGE_SHIFT;
8215 nr_pages = end - start;
8218 if (!pages || nr_pages > got_pages) {
8221 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8223 vmas = kvmalloc_array(nr_pages,
8224 sizeof(struct vm_area_struct *),
8226 if (!pages || !vmas) {
8230 got_pages = nr_pages;
8233 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8240 mmap_read_lock(current->mm);
8241 pret = pin_user_pages(ubuf, nr_pages,
8242 FOLL_WRITE | FOLL_LONGTERM,
8244 if (pret == nr_pages) {
8245 /* don't support file backed memory */
8246 for (j = 0; j < nr_pages; j++) {
8247 struct vm_area_struct *vma = vmas[j];
8250 !is_file_hugepages(vma->vm_file)) {
8256 ret = pret < 0 ? pret : -EFAULT;
8258 mmap_read_unlock(current->mm);
8261 * if we did partial map, or found file backed vmas,
8262 * release any pages we did get
8265 unpin_user_pages(pages, pret);
8270 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8272 unpin_user_pages(pages, pret);
8277 off = ubuf & ~PAGE_MASK;
8279 for (j = 0; j < nr_pages; j++) {
8282 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8283 imu->bvec[j].bv_page = pages[j];
8284 imu->bvec[j].bv_len = vec_len;
8285 imu->bvec[j].bv_offset = off;
8289 /* store original address for later verification */
8291 imu->len = iov.iov_len;
8292 imu->nr_bvecs = nr_pages;
8294 ctx->nr_user_bufs++;
8302 io_sqe_buffer_unregister(ctx);
8306 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8308 __s32 __user *fds = arg;
8314 if (copy_from_user(&fd, fds, sizeof(*fds)))
8317 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8318 if (IS_ERR(ctx->cq_ev_fd)) {
8319 int ret = PTR_ERR(ctx->cq_ev_fd);
8320 ctx->cq_ev_fd = NULL;
8327 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8329 if (ctx->cq_ev_fd) {
8330 eventfd_ctx_put(ctx->cq_ev_fd);
8331 ctx->cq_ev_fd = NULL;
8338 static int __io_destroy_buffers(int id, void *p, void *data)
8340 struct io_ring_ctx *ctx = data;
8341 struct io_buffer *buf = p;
8343 __io_remove_buffers(ctx, buf, id, -1U);
8347 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8349 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8350 idr_destroy(&ctx->io_buffer_idr);
8353 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8355 io_finish_async(ctx);
8356 io_sqe_buffer_unregister(ctx);
8358 if (ctx->sqo_task) {
8359 put_task_struct(ctx->sqo_task);
8360 ctx->sqo_task = NULL;
8361 mmdrop(ctx->mm_account);
8362 ctx->mm_account = NULL;
8365 #ifdef CONFIG_BLK_CGROUP
8366 if (ctx->sqo_blkcg_css)
8367 css_put(ctx->sqo_blkcg_css);
8370 io_sqe_files_unregister(ctx);
8371 io_eventfd_unregister(ctx);
8372 io_destroy_buffers(ctx);
8373 idr_destroy(&ctx->personality_idr);
8375 #if defined(CONFIG_UNIX)
8376 if (ctx->ring_sock) {
8377 ctx->ring_sock->file = NULL; /* so that iput() is called */
8378 sock_release(ctx->ring_sock);
8382 io_mem_free(ctx->rings);
8383 io_mem_free(ctx->sq_sqes);
8385 percpu_ref_exit(&ctx->refs);
8386 free_uid(ctx->user);
8387 put_cred(ctx->creds);
8388 kfree(ctx->cancel_hash);
8389 kmem_cache_free(req_cachep, ctx->fallback_req);
8393 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8395 struct io_ring_ctx *ctx = file->private_data;
8398 poll_wait(file, &ctx->cq_wait, wait);
8400 * synchronizes with barrier from wq_has_sleeper call in
8404 if (!io_sqring_full(ctx))
8405 mask |= EPOLLOUT | EPOLLWRNORM;
8406 if (io_cqring_events(ctx, false))
8407 mask |= EPOLLIN | EPOLLRDNORM;
8412 static int io_uring_fasync(int fd, struct file *file, int on)
8414 struct io_ring_ctx *ctx = file->private_data;
8416 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8419 static int io_remove_personalities(int id, void *p, void *data)
8421 struct io_ring_ctx *ctx = data;
8422 const struct cred *cred;
8424 cred = idr_remove(&ctx->personality_idr, id);
8430 static void io_ring_exit_work(struct work_struct *work)
8432 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8436 * If we're doing polled IO and end up having requests being
8437 * submitted async (out-of-line), then completions can come in while
8438 * we're waiting for refs to drop. We need to reap these manually,
8439 * as nobody else will be looking for them.
8443 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8444 io_iopoll_try_reap_events(ctx);
8445 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8446 io_ring_ctx_free(ctx);
8449 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8451 mutex_lock(&ctx->uring_lock);
8452 percpu_ref_kill(&ctx->refs);
8453 mutex_unlock(&ctx->uring_lock);
8455 io_kill_timeouts(ctx, NULL);
8456 io_poll_remove_all(ctx, NULL);
8459 io_wq_cancel_all(ctx->io_wq);
8461 /* if we failed setting up the ctx, we might not have any rings */
8463 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8464 io_iopoll_try_reap_events(ctx);
8465 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8468 * Do this upfront, so we won't have a grace period where the ring
8469 * is closed but resources aren't reaped yet. This can cause
8470 * spurious failure in setting up a new ring.
8472 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8475 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8477 * Use system_unbound_wq to avoid spawning tons of event kworkers
8478 * if we're exiting a ton of rings at the same time. It just adds
8479 * noise and overhead, there's no discernable change in runtime
8480 * over using system_wq.
8482 queue_work(system_unbound_wq, &ctx->exit_work);
8485 static int io_uring_release(struct inode *inode, struct file *file)
8487 struct io_ring_ctx *ctx = file->private_data;
8489 file->private_data = NULL;
8490 io_ring_ctx_wait_and_kill(ctx);
8494 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8496 struct files_struct *files = data;
8498 return !files || work->files == files;
8502 * Returns true if 'preq' is the link parent of 'req'
8504 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8506 struct io_kiocb *link;
8508 if (!(preq->flags & REQ_F_LINK_HEAD))
8511 list_for_each_entry(link, &preq->link_list, link_list) {
8519 static bool io_match_link_files(struct io_kiocb *req,
8520 struct files_struct *files)
8522 struct io_kiocb *link;
8524 if (io_match_files(req, files))
8526 if (req->flags & REQ_F_LINK_HEAD) {
8527 list_for_each_entry(link, &req->link_list, link_list) {
8528 if (io_match_files(link, files))
8536 * We're looking to cancel 'req' because it's holding on to our files, but
8537 * 'req' could be a link to another request. See if it is, and cancel that
8538 * parent request if so.
8540 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8542 struct hlist_node *tmp;
8543 struct io_kiocb *preq;
8547 spin_lock_irq(&ctx->completion_lock);
8548 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8549 struct hlist_head *list;
8551 list = &ctx->cancel_hash[i];
8552 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8553 found = io_match_link(preq, req);
8555 io_poll_remove_one(preq);
8560 spin_unlock_irq(&ctx->completion_lock);
8564 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8565 struct io_kiocb *req)
8567 struct io_kiocb *preq;
8570 spin_lock_irq(&ctx->completion_lock);
8571 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8572 found = io_match_link(preq, req);
8574 __io_timeout_cancel(preq);
8578 spin_unlock_irq(&ctx->completion_lock);
8582 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8584 return io_match_link(container_of(work, struct io_kiocb, work), data);
8587 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8589 enum io_wq_cancel cret;
8591 /* cancel this particular work, if it's running */
8592 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8593 if (cret != IO_WQ_CANCEL_NOTFOUND)
8596 /* find links that hold this pending, cancel those */
8597 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8598 if (cret != IO_WQ_CANCEL_NOTFOUND)
8601 /* if we have a poll link holding this pending, cancel that */
8602 if (io_poll_remove_link(ctx, req))
8605 /* final option, timeout link is holding this req pending */
8606 io_timeout_remove_link(ctx, req);
8609 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8610 struct files_struct *files)
8612 struct io_defer_entry *de = NULL;
8615 spin_lock_irq(&ctx->completion_lock);
8616 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8617 if (io_match_link_files(de->req, files)) {
8618 list_cut_position(&list, &ctx->defer_list, &de->list);
8622 spin_unlock_irq(&ctx->completion_lock);
8624 while (!list_empty(&list)) {
8625 de = list_first_entry(&list, struct io_defer_entry, list);
8626 list_del_init(&de->list);
8627 req_set_fail_links(de->req);
8628 io_put_req(de->req);
8629 io_req_complete(de->req, -ECANCELED);
8635 * Returns true if we found and killed one or more files pinning requests
8637 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8638 struct files_struct *files)
8640 if (list_empty_careful(&ctx->inflight_list))
8643 io_cancel_defer_files(ctx, files);
8644 /* cancel all at once, should be faster than doing it one by one*/
8645 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8647 while (!list_empty_careful(&ctx->inflight_list)) {
8648 struct io_kiocb *cancel_req = NULL, *req;
8651 spin_lock_irq(&ctx->inflight_lock);
8652 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8653 if (files && req->work.files != files)
8655 /* req is being completed, ignore */
8656 if (!refcount_inc_not_zero(&req->refs))
8662 prepare_to_wait(&ctx->inflight_wait, &wait,
8663 TASK_UNINTERRUPTIBLE);
8664 spin_unlock_irq(&ctx->inflight_lock);
8666 /* We need to keep going until we don't find a matching req */
8669 /* cancel this request, or head link requests */
8670 io_attempt_cancel(ctx, cancel_req);
8671 io_put_req(cancel_req);
8672 /* cancellations _may_ trigger task work */
8675 finish_wait(&ctx->inflight_wait, &wait);
8681 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8683 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8684 struct task_struct *task = data;
8686 return io_task_match(req, task);
8689 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8690 struct task_struct *task,
8691 struct files_struct *files)
8695 ret = io_uring_cancel_files(ctx, files);
8697 enum io_wq_cancel cret;
8699 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8700 if (cret != IO_WQ_CANCEL_NOTFOUND)
8703 /* SQPOLL thread does its own polling */
8704 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8705 while (!list_empty_careful(&ctx->iopoll_list)) {
8706 io_iopoll_try_reap_events(ctx);
8711 ret |= io_poll_remove_all(ctx, task);
8712 ret |= io_kill_timeouts(ctx, task);
8719 * We need to iteratively cancel requests, in case a request has dependent
8720 * hard links. These persist even for failure of cancelations, hence keep
8721 * looping until none are found.
8723 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8724 struct files_struct *files)
8726 struct task_struct *task = current;
8728 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8729 task = ctx->sq_data->thread;
8731 io_cqring_overflow_flush(ctx, true, task, files);
8733 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8740 * Note that this task has used io_uring. We use it for cancelation purposes.
8742 static int io_uring_add_task_file(struct file *file)
8744 if (unlikely(!current->io_uring)) {
8747 ret = io_uring_alloc_task_context(current);
8751 if (current->io_uring->last != file) {
8752 XA_STATE(xas, ¤t->io_uring->xa, (unsigned long) file);
8756 old = xas_load(&xas);
8760 xas_store(&xas, file);
8764 current->io_uring->last = file;
8771 * Remove this io_uring_file -> task mapping.
8773 static void io_uring_del_task_file(struct file *file)
8775 struct io_uring_task *tctx = current->io_uring;
8776 XA_STATE(xas, &tctx->xa, (unsigned long) file);
8778 if (tctx->last == file)
8782 file = xas_store(&xas, NULL);
8789 static void __io_uring_attempt_task_drop(struct file *file)
8791 XA_STATE(xas, ¤t->io_uring->xa, (unsigned long) file);
8795 old = xas_load(&xas);
8799 io_uring_del_task_file(file);
8803 * Drop task note for this file if we're the only ones that hold it after
8806 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8808 if (!current->io_uring)
8811 * fput() is pending, will be 2 if the only other ref is our potential
8812 * task file note. If the task is exiting, drop regardless of count.
8814 if (!exiting && atomic_long_read(&file->f_count) != 2)
8817 __io_uring_attempt_task_drop(file);
8820 void __io_uring_files_cancel(struct files_struct *files)
8822 struct io_uring_task *tctx = current->io_uring;
8823 XA_STATE(xas, &tctx->xa, 0);
8825 /* make sure overflow events are dropped */
8826 tctx->in_idle = true;
8829 struct io_ring_ctx *ctx;
8833 file = xas_next_entry(&xas, ULONG_MAX);
8839 ctx = file->private_data;
8841 io_uring_cancel_task_requests(ctx, files);
8843 io_uring_del_task_file(file);
8847 static inline bool io_uring_task_idle(struct io_uring_task *tctx)
8849 return atomic_long_read(&tctx->req_issue) ==
8850 atomic_long_read(&tctx->req_complete);
8854 * Find any io_uring fd that this task has registered or done IO on, and cancel
8857 void __io_uring_task_cancel(void)
8859 struct io_uring_task *tctx = current->io_uring;
8863 /* make sure overflow events are dropped */
8864 tctx->in_idle = true;
8866 while (!io_uring_task_idle(tctx)) {
8867 /* read completions before cancelations */
8868 completions = atomic_long_read(&tctx->req_complete);
8869 __io_uring_files_cancel(NULL);
8871 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8874 * If we've seen completions, retry. This avoids a race where
8875 * a completion comes in before we did prepare_to_wait().
8877 if (completions != atomic_long_read(&tctx->req_complete))
8879 if (io_uring_task_idle(tctx))
8884 finish_wait(&tctx->wait, &wait);
8885 tctx->in_idle = false;
8888 static int io_uring_flush(struct file *file, void *data)
8890 struct io_ring_ctx *ctx = file->private_data;
8893 * If the task is going away, cancel work it may have pending
8895 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8898 io_uring_cancel_task_requests(ctx, data);
8899 io_uring_attempt_task_drop(file, !data);
8903 static void *io_uring_validate_mmap_request(struct file *file,
8904 loff_t pgoff, size_t sz)
8906 struct io_ring_ctx *ctx = file->private_data;
8907 loff_t offset = pgoff << PAGE_SHIFT;
8912 case IORING_OFF_SQ_RING:
8913 case IORING_OFF_CQ_RING:
8916 case IORING_OFF_SQES:
8920 return ERR_PTR(-EINVAL);
8923 page = virt_to_head_page(ptr);
8924 if (sz > page_size(page))
8925 return ERR_PTR(-EINVAL);
8932 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8934 size_t sz = vma->vm_end - vma->vm_start;
8938 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8940 return PTR_ERR(ptr);
8942 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8943 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8946 #else /* !CONFIG_MMU */
8948 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8950 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8953 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8955 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8958 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8959 unsigned long addr, unsigned long len,
8960 unsigned long pgoff, unsigned long flags)
8964 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8966 return PTR_ERR(ptr);
8968 return (unsigned long) ptr;
8971 #endif /* !CONFIG_MMU */
8973 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8978 if (!io_sqring_full(ctx))
8981 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8983 if (!io_sqring_full(ctx))
8987 } while (!signal_pending(current));
8989 finish_wait(&ctx->sqo_sq_wait, &wait);
8992 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8993 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8996 struct io_ring_ctx *ctx;
9003 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9004 IORING_ENTER_SQ_WAIT))
9012 if (f.file->f_op != &io_uring_fops)
9016 ctx = f.file->private_data;
9017 if (!percpu_ref_tryget(&ctx->refs))
9021 if (ctx->flags & IORING_SETUP_R_DISABLED)
9025 * For SQ polling, the thread will do all submissions and completions.
9026 * Just return the requested submit count, and wake the thread if
9030 if (ctx->flags & IORING_SETUP_SQPOLL) {
9031 if (!list_empty_careful(&ctx->cq_overflow_list))
9032 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9033 if (flags & IORING_ENTER_SQ_WAKEUP)
9034 wake_up(&ctx->sq_data->wait);
9035 if (flags & IORING_ENTER_SQ_WAIT)
9036 io_sqpoll_wait_sq(ctx);
9037 submitted = to_submit;
9038 } else if (to_submit) {
9039 ret = io_uring_add_task_file(f.file);
9042 mutex_lock(&ctx->uring_lock);
9043 submitted = io_submit_sqes(ctx, to_submit);
9044 mutex_unlock(&ctx->uring_lock);
9046 if (submitted != to_submit)
9049 if (flags & IORING_ENTER_GETEVENTS) {
9050 min_complete = min(min_complete, ctx->cq_entries);
9053 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9054 * space applications don't need to do io completion events
9055 * polling again, they can rely on io_sq_thread to do polling
9056 * work, which can reduce cpu usage and uring_lock contention.
9058 if (ctx->flags & IORING_SETUP_IOPOLL &&
9059 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9060 ret = io_iopoll_check(ctx, min_complete);
9062 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
9067 percpu_ref_put(&ctx->refs);
9070 return submitted ? submitted : ret;
9073 #ifdef CONFIG_PROC_FS
9074 static int io_uring_show_cred(int id, void *p, void *data)
9076 const struct cred *cred = p;
9077 struct seq_file *m = data;
9078 struct user_namespace *uns = seq_user_ns(m);
9079 struct group_info *gi;
9084 seq_printf(m, "%5d\n", id);
9085 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9086 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9087 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9088 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9089 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9090 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9091 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9092 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9093 seq_puts(m, "\n\tGroups:\t");
9094 gi = cred->group_info;
9095 for (g = 0; g < gi->ngroups; g++) {
9096 seq_put_decimal_ull(m, g ? " " : "",
9097 from_kgid_munged(uns, gi->gid[g]));
9099 seq_puts(m, "\n\tCapEff:\t");
9100 cap = cred->cap_effective;
9101 CAP_FOR_EACH_U32(__capi)
9102 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9107 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9109 struct io_sq_data *sq = NULL;
9114 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9115 * since fdinfo case grabs it in the opposite direction of normal use
9116 * cases. If we fail to get the lock, we just don't iterate any
9117 * structures that could be going away outside the io_uring mutex.
9119 has_lock = mutex_trylock(&ctx->uring_lock);
9121 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9124 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9125 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9126 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9127 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9128 struct fixed_file_table *table;
9131 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9132 f = table->files[i & IORING_FILE_TABLE_MASK];
9134 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9136 seq_printf(m, "%5u: <none>\n", i);
9138 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9139 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9140 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9142 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9143 (unsigned int) buf->len);
9145 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9146 seq_printf(m, "Personalities:\n");
9147 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9149 seq_printf(m, "PollList:\n");
9150 spin_lock_irq(&ctx->completion_lock);
9151 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9152 struct hlist_head *list = &ctx->cancel_hash[i];
9153 struct io_kiocb *req;
9155 hlist_for_each_entry(req, list, hash_node)
9156 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9157 req->task->task_works != NULL);
9159 spin_unlock_irq(&ctx->completion_lock);
9161 mutex_unlock(&ctx->uring_lock);
9164 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9166 struct io_ring_ctx *ctx = f->private_data;
9168 if (percpu_ref_tryget(&ctx->refs)) {
9169 __io_uring_show_fdinfo(ctx, m);
9170 percpu_ref_put(&ctx->refs);
9175 static const struct file_operations io_uring_fops = {
9176 .release = io_uring_release,
9177 .flush = io_uring_flush,
9178 .mmap = io_uring_mmap,
9180 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9181 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9183 .poll = io_uring_poll,
9184 .fasync = io_uring_fasync,
9185 #ifdef CONFIG_PROC_FS
9186 .show_fdinfo = io_uring_show_fdinfo,
9190 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9191 struct io_uring_params *p)
9193 struct io_rings *rings;
9194 size_t size, sq_array_offset;
9196 /* make sure these are sane, as we already accounted them */
9197 ctx->sq_entries = p->sq_entries;
9198 ctx->cq_entries = p->cq_entries;
9200 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9201 if (size == SIZE_MAX)
9204 rings = io_mem_alloc(size);
9209 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9210 rings->sq_ring_mask = p->sq_entries - 1;
9211 rings->cq_ring_mask = p->cq_entries - 1;
9212 rings->sq_ring_entries = p->sq_entries;
9213 rings->cq_ring_entries = p->cq_entries;
9214 ctx->sq_mask = rings->sq_ring_mask;
9215 ctx->cq_mask = rings->cq_ring_mask;
9217 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9218 if (size == SIZE_MAX) {
9219 io_mem_free(ctx->rings);
9224 ctx->sq_sqes = io_mem_alloc(size);
9225 if (!ctx->sq_sqes) {
9226 io_mem_free(ctx->rings);
9235 * Allocate an anonymous fd, this is what constitutes the application
9236 * visible backing of an io_uring instance. The application mmaps this
9237 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9238 * we have to tie this fd to a socket for file garbage collection purposes.
9240 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9245 #if defined(CONFIG_UNIX)
9246 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9252 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9256 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9257 O_RDWR | O_CLOEXEC);
9261 ret = PTR_ERR(file);
9265 #if defined(CONFIG_UNIX)
9266 ctx->ring_sock->file = file;
9268 if (unlikely(io_uring_add_task_file(file))) {
9269 file = ERR_PTR(-ENOMEM);
9272 fd_install(ret, file);
9275 #if defined(CONFIG_UNIX)
9276 sock_release(ctx->ring_sock);
9277 ctx->ring_sock = NULL;
9282 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9283 struct io_uring_params __user *params)
9285 struct user_struct *user = NULL;
9286 struct io_ring_ctx *ctx;
9292 if (entries > IORING_MAX_ENTRIES) {
9293 if (!(p->flags & IORING_SETUP_CLAMP))
9295 entries = IORING_MAX_ENTRIES;
9299 * Use twice as many entries for the CQ ring. It's possible for the
9300 * application to drive a higher depth than the size of the SQ ring,
9301 * since the sqes are only used at submission time. This allows for
9302 * some flexibility in overcommitting a bit. If the application has
9303 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9304 * of CQ ring entries manually.
9306 p->sq_entries = roundup_pow_of_two(entries);
9307 if (p->flags & IORING_SETUP_CQSIZE) {
9309 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9310 * to a power-of-two, if it isn't already. We do NOT impose
9311 * any cq vs sq ring sizing.
9313 if (p->cq_entries < p->sq_entries)
9315 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9316 if (!(p->flags & IORING_SETUP_CLAMP))
9318 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9320 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9322 p->cq_entries = 2 * p->sq_entries;
9325 user = get_uid(current_user());
9326 limit_mem = !capable(CAP_IPC_LOCK);
9329 ret = __io_account_mem(user,
9330 ring_pages(p->sq_entries, p->cq_entries));
9337 ctx = io_ring_ctx_alloc(p);
9340 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9345 ctx->compat = in_compat_syscall();
9347 ctx->creds = get_current_cred();
9349 ctx->sqo_task = get_task_struct(current);
9352 * This is just grabbed for accounting purposes. When a process exits,
9353 * the mm is exited and dropped before the files, hence we need to hang
9354 * on to this mm purely for the purposes of being able to unaccount
9355 * memory (locked/pinned vm). It's not used for anything else.
9357 mmgrab(current->mm);
9358 ctx->mm_account = current->mm;
9360 #ifdef CONFIG_BLK_CGROUP
9362 * The sq thread will belong to the original cgroup it was inited in.
9363 * If the cgroup goes offline (e.g. disabling the io controller), then
9364 * issued bios will be associated with the closest cgroup later in the
9368 ctx->sqo_blkcg_css = blkcg_css();
9369 ret = css_tryget_online(ctx->sqo_blkcg_css);
9372 /* don't init against a dying cgroup, have the user try again */
9373 ctx->sqo_blkcg_css = NULL;
9380 * Account memory _before_ installing the file descriptor. Once
9381 * the descriptor is installed, it can get closed at any time. Also
9382 * do this before hitting the general error path, as ring freeing
9383 * will un-account as well.
9385 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9387 ctx->limit_mem = limit_mem;
9389 ret = io_allocate_scq_urings(ctx, p);
9393 ret = io_sq_offload_create(ctx, p);
9397 if (!(p->flags & IORING_SETUP_R_DISABLED))
9398 io_sq_offload_start(ctx);
9400 memset(&p->sq_off, 0, sizeof(p->sq_off));
9401 p->sq_off.head = offsetof(struct io_rings, sq.head);
9402 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9403 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9404 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9405 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9406 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9407 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9409 memset(&p->cq_off, 0, sizeof(p->cq_off));
9410 p->cq_off.head = offsetof(struct io_rings, cq.head);
9411 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9412 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9413 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9414 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9415 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9416 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9418 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9419 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9420 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9421 IORING_FEAT_POLL_32BITS;
9423 if (copy_to_user(params, p, sizeof(*p))) {
9429 * Install ring fd as the very last thing, so we don't risk someone
9430 * having closed it before we finish setup
9432 ret = io_uring_get_fd(ctx);
9436 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9439 io_ring_ctx_wait_and_kill(ctx);
9444 * Sets up an aio uring context, and returns the fd. Applications asks for a
9445 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9446 * params structure passed in.
9448 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9450 struct io_uring_params p;
9453 if (copy_from_user(&p, params, sizeof(p)))
9455 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9460 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9461 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9462 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9463 IORING_SETUP_R_DISABLED))
9466 return io_uring_create(entries, &p, params);
9469 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9470 struct io_uring_params __user *, params)
9472 return io_uring_setup(entries, params);
9475 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9477 struct io_uring_probe *p;
9481 size = struct_size(p, ops, nr_args);
9482 if (size == SIZE_MAX)
9484 p = kzalloc(size, GFP_KERNEL);
9489 if (copy_from_user(p, arg, size))
9492 if (memchr_inv(p, 0, size))
9495 p->last_op = IORING_OP_LAST - 1;
9496 if (nr_args > IORING_OP_LAST)
9497 nr_args = IORING_OP_LAST;
9499 for (i = 0; i < nr_args; i++) {
9501 if (!io_op_defs[i].not_supported)
9502 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9507 if (copy_to_user(arg, p, size))
9514 static int io_register_personality(struct io_ring_ctx *ctx)
9516 const struct cred *creds = get_current_cred();
9519 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9520 USHRT_MAX, GFP_KERNEL);
9526 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9528 const struct cred *old_creds;
9530 old_creds = idr_remove(&ctx->personality_idr, id);
9532 put_cred(old_creds);
9539 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9540 unsigned int nr_args)
9542 struct io_uring_restriction *res;
9546 /* Restrictions allowed only if rings started disabled */
9547 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9550 /* We allow only a single restrictions registration */
9551 if (ctx->restrictions.registered)
9554 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9557 size = array_size(nr_args, sizeof(*res));
9558 if (size == SIZE_MAX)
9561 res = memdup_user(arg, size);
9563 return PTR_ERR(res);
9567 for (i = 0; i < nr_args; i++) {
9568 switch (res[i].opcode) {
9569 case IORING_RESTRICTION_REGISTER_OP:
9570 if (res[i].register_op >= IORING_REGISTER_LAST) {
9575 __set_bit(res[i].register_op,
9576 ctx->restrictions.register_op);
9578 case IORING_RESTRICTION_SQE_OP:
9579 if (res[i].sqe_op >= IORING_OP_LAST) {
9584 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9586 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9587 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9589 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9590 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9599 /* Reset all restrictions if an error happened */
9601 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9603 ctx->restrictions.registered = true;
9609 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9611 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9614 if (ctx->restrictions.registered)
9615 ctx->restricted = 1;
9617 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9619 io_sq_offload_start(ctx);
9624 static bool io_register_op_must_quiesce(int op)
9627 case IORING_UNREGISTER_FILES:
9628 case IORING_REGISTER_FILES_UPDATE:
9629 case IORING_REGISTER_PROBE:
9630 case IORING_REGISTER_PERSONALITY:
9631 case IORING_UNREGISTER_PERSONALITY:
9638 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9639 void __user *arg, unsigned nr_args)
9640 __releases(ctx->uring_lock)
9641 __acquires(ctx->uring_lock)
9646 * We're inside the ring mutex, if the ref is already dying, then
9647 * someone else killed the ctx or is already going through
9648 * io_uring_register().
9650 if (percpu_ref_is_dying(&ctx->refs))
9653 if (io_register_op_must_quiesce(opcode)) {
9654 percpu_ref_kill(&ctx->refs);
9657 * Drop uring mutex before waiting for references to exit. If
9658 * another thread is currently inside io_uring_enter() it might
9659 * need to grab the uring_lock to make progress. If we hold it
9660 * here across the drain wait, then we can deadlock. It's safe
9661 * to drop the mutex here, since no new references will come in
9662 * after we've killed the percpu ref.
9664 mutex_unlock(&ctx->uring_lock);
9666 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9669 if (io_run_task_work_sig() > 0)
9673 mutex_lock(&ctx->uring_lock);
9676 percpu_ref_resurrect(&ctx->refs);
9682 if (ctx->restricted) {
9683 if (opcode >= IORING_REGISTER_LAST) {
9688 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9695 case IORING_REGISTER_BUFFERS:
9696 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9698 case IORING_UNREGISTER_BUFFERS:
9702 ret = io_sqe_buffer_unregister(ctx);
9704 case IORING_REGISTER_FILES:
9705 ret = io_sqe_files_register(ctx, arg, nr_args);
9707 case IORING_UNREGISTER_FILES:
9711 ret = io_sqe_files_unregister(ctx);
9713 case IORING_REGISTER_FILES_UPDATE:
9714 ret = io_sqe_files_update(ctx, arg, nr_args);
9716 case IORING_REGISTER_EVENTFD:
9717 case IORING_REGISTER_EVENTFD_ASYNC:
9721 ret = io_eventfd_register(ctx, arg);
9724 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9725 ctx->eventfd_async = 1;
9727 ctx->eventfd_async = 0;
9729 case IORING_UNREGISTER_EVENTFD:
9733 ret = io_eventfd_unregister(ctx);
9735 case IORING_REGISTER_PROBE:
9737 if (!arg || nr_args > 256)
9739 ret = io_probe(ctx, arg, nr_args);
9741 case IORING_REGISTER_PERSONALITY:
9745 ret = io_register_personality(ctx);
9747 case IORING_UNREGISTER_PERSONALITY:
9751 ret = io_unregister_personality(ctx, nr_args);
9753 case IORING_REGISTER_ENABLE_RINGS:
9757 ret = io_register_enable_rings(ctx);
9759 case IORING_REGISTER_RESTRICTIONS:
9760 ret = io_register_restrictions(ctx, arg, nr_args);
9768 if (io_register_op_must_quiesce(opcode)) {
9769 /* bring the ctx back to life */
9770 percpu_ref_reinit(&ctx->refs);
9772 reinit_completion(&ctx->ref_comp);
9777 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9778 void __user *, arg, unsigned int, nr_args)
9780 struct io_ring_ctx *ctx;
9789 if (f.file->f_op != &io_uring_fops)
9792 ctx = f.file->private_data;
9794 mutex_lock(&ctx->uring_lock);
9795 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9796 mutex_unlock(&ctx->uring_lock);
9797 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9798 ctx->cq_ev_fd != NULL, ret);
9804 static int __init io_uring_init(void)
9806 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9807 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9808 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9811 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9812 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9813 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9814 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9815 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9816 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9817 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9818 BUILD_BUG_SQE_ELEM(8, __u64, off);
9819 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9820 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9821 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9822 BUILD_BUG_SQE_ELEM(24, __u32, len);
9823 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9824 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9825 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9826 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9827 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9828 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9829 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9830 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9831 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9832 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9833 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9834 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9835 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9836 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9837 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9838 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9839 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9840 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9841 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9843 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9844 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9845 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9848 __initcall(io_uring_init);