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>
84 #define CREATE_TRACE_POINTS
85 #include <trace/events/io_uring.h>
87 #include <uapi/linux/io_uring.h>
92 #define IORING_MAX_ENTRIES 32768
93 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
96 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
98 #define IORING_FILE_TABLE_SHIFT 9
99 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
100 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
101 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
102 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
103 IORING_REGISTER_LAST + IORING_OP_LAST)
106 u32 head ____cacheline_aligned_in_smp;
107 u32 tail ____cacheline_aligned_in_smp;
111 * This data is shared with the application through the mmap at offsets
112 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
114 * The offsets to the member fields are published through struct
115 * io_sqring_offsets when calling io_uring_setup.
119 * Head and tail offsets into the ring; the offsets need to be
120 * masked to get valid indices.
122 * The kernel controls head of the sq ring and the tail of the cq ring,
123 * and the application controls tail of the sq ring and the head of the
126 struct io_uring sq, cq;
128 * Bitmasks to apply to head and tail offsets (constant, equals
131 u32 sq_ring_mask, cq_ring_mask;
132 /* Ring sizes (constant, power of 2) */
133 u32 sq_ring_entries, cq_ring_entries;
135 * Number of invalid entries dropped by the kernel due to
136 * invalid index stored in array
138 * Written by the kernel, shouldn't be modified by the
139 * application (i.e. get number of "new events" by comparing to
142 * After a new SQ head value was read by the application this
143 * counter includes all submissions that were dropped reaching
144 * the new SQ head (and possibly more).
150 * Written by the kernel, shouldn't be modified by the
153 * The application needs a full memory barrier before checking
154 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
160 * Written by the application, shouldn't be modified by the
165 * Number of completion events lost because the queue was full;
166 * this should be avoided by the application by making sure
167 * there are not more requests pending than there is space in
168 * the completion queue.
170 * Written by the kernel, shouldn't be modified by the
171 * application (i.e. get number of "new events" by comparing to
174 * As completion events come in out of order this counter is not
175 * ordered with any other data.
179 * Ring buffer of completion events.
181 * The kernel writes completion events fresh every time they are
182 * produced, so the application is allowed to modify pending
185 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
188 struct io_mapped_ubuf {
191 struct bio_vec *bvec;
192 unsigned int nr_bvecs;
195 struct fixed_file_table {
199 struct fixed_file_ref_node {
200 struct percpu_ref refs;
201 struct list_head node;
202 struct list_head file_list;
203 struct fixed_file_data *file_data;
204 struct llist_node llist;
207 struct fixed_file_data {
208 struct fixed_file_table *table;
209 struct io_ring_ctx *ctx;
211 struct percpu_ref *cur_refs;
212 struct percpu_ref refs;
213 struct completion done;
214 struct list_head ref_list;
219 struct list_head list;
225 struct io_restriction {
226 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
227 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
228 u8 sqe_flags_allowed;
229 u8 sqe_flags_required;
235 struct percpu_ref refs;
236 } ____cacheline_aligned_in_smp;
240 unsigned int compat: 1;
241 unsigned int limit_mem: 1;
242 unsigned int cq_overflow_flushed: 1;
243 unsigned int drain_next: 1;
244 unsigned int eventfd_async: 1;
245 unsigned int restricted: 1;
248 * Ring buffer of indices into array of io_uring_sqe, which is
249 * mmapped by the application using the IORING_OFF_SQES offset.
251 * This indirection could e.g. be used to assign fixed
252 * io_uring_sqe entries to operations and only submit them to
253 * the queue when needed.
255 * The kernel modifies neither the indices array nor the entries
259 unsigned cached_sq_head;
262 unsigned sq_thread_idle;
263 unsigned cached_sq_dropped;
264 atomic_t cached_cq_overflow;
265 unsigned long sq_check_overflow;
267 struct list_head defer_list;
268 struct list_head timeout_list;
269 struct list_head cq_overflow_list;
271 wait_queue_head_t inflight_wait;
272 struct io_uring_sqe *sq_sqes;
273 } ____cacheline_aligned_in_smp;
275 struct io_rings *rings;
279 struct task_struct *sqo_thread; /* if using sq thread polling */
282 * For SQPOLL usage - we hold a reference to the parent task, so we
283 * have access to the ->files
285 struct task_struct *sqo_task;
287 /* Only used for accounting purposes */
288 struct mm_struct *mm_account;
290 wait_queue_head_t sqo_wait;
293 * If used, fixed file set. Writers must ensure that ->refs is dead,
294 * readers must ensure that ->refs is alive as long as the file* is
295 * used. Only updated through io_uring_register(2).
297 struct fixed_file_data *file_data;
298 unsigned nr_user_files;
300 /* if used, fixed mapped user buffers */
301 unsigned nr_user_bufs;
302 struct io_mapped_ubuf *user_bufs;
304 struct user_struct *user;
306 const struct cred *creds;
308 struct completion ref_comp;
309 struct completion sq_thread_comp;
311 /* if all else fails... */
312 struct io_kiocb *fallback_req;
314 #if defined(CONFIG_UNIX)
315 struct socket *ring_sock;
318 struct idr io_buffer_idr;
320 struct idr personality_idr;
323 unsigned cached_cq_tail;
326 atomic_t cq_timeouts;
327 unsigned long cq_check_overflow;
328 struct wait_queue_head cq_wait;
329 struct fasync_struct *cq_fasync;
330 struct eventfd_ctx *cq_ev_fd;
331 } ____cacheline_aligned_in_smp;
334 struct mutex uring_lock;
335 wait_queue_head_t wait;
336 } ____cacheline_aligned_in_smp;
339 spinlock_t completion_lock;
342 * ->iopoll_list is protected by the ctx->uring_lock for
343 * io_uring instances that don't use IORING_SETUP_SQPOLL.
344 * For SQPOLL, only the single threaded io_sq_thread() will
345 * manipulate the list, hence no extra locking is needed there.
347 struct list_head iopoll_list;
348 struct hlist_head *cancel_hash;
349 unsigned cancel_hash_bits;
350 bool poll_multi_file;
352 spinlock_t inflight_lock;
353 struct list_head inflight_list;
354 } ____cacheline_aligned_in_smp;
356 struct delayed_work file_put_work;
357 struct llist_head file_put_llist;
359 struct work_struct exit_work;
360 struct io_restriction restrictions;
364 * First field must be the file pointer in all the
365 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
367 struct io_poll_iocb {
370 struct wait_queue_head *head;
376 struct wait_queue_entry wait;
381 struct file *put_file;
385 struct io_timeout_data {
386 struct io_kiocb *req;
387 struct hrtimer timer;
388 struct timespec64 ts;
389 enum hrtimer_mode mode;
394 struct sockaddr __user *addr;
395 int __user *addr_len;
397 unsigned long nofile;
419 struct list_head list;
423 /* NOTE: kiocb has the file as the first member, so don't do it here */
431 struct sockaddr __user *addr;
438 struct user_msghdr __user *umsg;
444 struct io_buffer *kbuf;
450 struct filename *filename;
452 unsigned long nofile;
455 struct io_files_update {
481 struct epoll_event event;
485 struct file *file_out;
486 struct file *file_in;
493 struct io_provide_buf {
507 const char __user *filename;
508 struct statx __user *buffer;
511 struct io_completion {
513 struct list_head list;
517 struct io_async_connect {
518 struct sockaddr_storage address;
521 struct io_async_msghdr {
522 struct iovec fast_iov[UIO_FASTIOV];
524 struct sockaddr __user *uaddr;
526 struct sockaddr_storage addr;
530 struct iovec fast_iov[UIO_FASTIOV];
531 const struct iovec *free_iovec;
532 struct iov_iter iter;
534 struct wait_page_queue wpq;
537 struct io_async_ctx {
539 struct io_async_rw rw;
540 struct io_async_msghdr msg;
541 struct io_async_connect connect;
542 struct io_timeout_data timeout;
547 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
548 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
549 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
550 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
551 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
552 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
559 REQ_F_LINK_TIMEOUT_BIT,
561 REQ_F_COMP_LOCKED_BIT,
562 REQ_F_NEED_CLEANUP_BIT,
564 REQ_F_BUFFER_SELECTED_BIT,
565 REQ_F_NO_FILE_TABLE_BIT,
566 REQ_F_WORK_INITIALIZED_BIT,
568 /* not a real bit, just to check we're not overflowing the space */
574 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
575 /* drain existing IO first */
576 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
578 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
579 /* doesn't sever on completion < 0 */
580 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
582 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
583 /* IOSQE_BUFFER_SELECT */
584 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
587 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
588 /* fail rest of links */
589 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
590 /* on inflight list */
591 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
592 /* read/write uses file position */
593 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
594 /* must not punt to workers */
595 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
596 /* has linked timeout */
597 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
599 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
600 /* completion under lock */
601 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
603 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
604 /* already went through poll handler */
605 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
606 /* buffer already selected */
607 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
608 /* doesn't need file table for this request */
609 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
610 /* io_wq_work is initialized */
611 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
615 struct io_poll_iocb poll;
616 struct io_poll_iocb *double_poll;
620 * NOTE! Each of the iocb union members has the file pointer
621 * as the first entry in their struct definition. So you can
622 * access the file pointer through any of the sub-structs,
623 * or directly as just 'ki_filp' in this struct.
629 struct io_poll_iocb poll;
630 struct io_accept accept;
632 struct io_cancel cancel;
633 struct io_timeout timeout;
634 struct io_connect connect;
635 struct io_sr_msg sr_msg;
637 struct io_close close;
638 struct io_files_update files_update;
639 struct io_fadvise fadvise;
640 struct io_madvise madvise;
641 struct io_epoll epoll;
642 struct io_splice splice;
643 struct io_provide_buf pbuf;
644 struct io_statx statx;
645 /* use only after cleaning per-op data, see io_clean_op() */
646 struct io_completion compl;
649 struct io_async_ctx *io;
651 /* polled IO has completed */
657 struct io_ring_ctx *ctx;
660 struct task_struct *task;
663 struct list_head link_list;
666 * 1. used with ctx->iopoll_list with reads/writes
667 * 2. to track reqs with ->files (see io_op_def::file_table)
669 struct list_head inflight_entry;
671 struct percpu_ref *fixed_file_refs;
672 struct callback_head task_work;
673 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
674 struct hlist_node hash_node;
675 struct async_poll *apoll;
676 struct io_wq_work work;
679 struct io_defer_entry {
680 struct list_head list;
681 struct io_kiocb *req;
685 #define IO_IOPOLL_BATCH 8
687 struct io_comp_state {
689 struct list_head list;
690 struct io_ring_ctx *ctx;
693 struct io_submit_state {
694 struct blk_plug plug;
697 * io_kiocb alloc cache
699 void *reqs[IO_IOPOLL_BATCH];
700 unsigned int free_reqs;
703 * Batch completion logic
705 struct io_comp_state comp;
708 * File reference cache
712 unsigned int has_refs;
713 unsigned int ios_left;
717 /* needs req->io allocated for deferral/async */
718 unsigned async_ctx : 1;
719 /* needs current->mm setup, does mm access */
720 unsigned needs_mm : 1;
721 /* needs req->file assigned */
722 unsigned needs_file : 1;
723 /* don't fail if file grab fails */
724 unsigned needs_file_no_error : 1;
725 /* hash wq insertion if file is a regular file */
726 unsigned hash_reg_file : 1;
727 /* unbound wq insertion if file is a non-regular file */
728 unsigned unbound_nonreg_file : 1;
729 /* opcode is not supported by this kernel */
730 unsigned not_supported : 1;
731 /* needs file table */
732 unsigned file_table : 1;
734 unsigned needs_fs : 1;
735 /* set if opcode supports polled "wait" */
737 unsigned pollout : 1;
738 /* op supports buffer selection */
739 unsigned buffer_select : 1;
740 unsigned needs_fsize : 1;
743 static const struct io_op_def io_op_defs[] = {
744 [IORING_OP_NOP] = {},
745 [IORING_OP_READV] = {
749 .unbound_nonreg_file = 1,
753 [IORING_OP_WRITEV] = {
758 .unbound_nonreg_file = 1,
762 [IORING_OP_FSYNC] = {
765 [IORING_OP_READ_FIXED] = {
767 .unbound_nonreg_file = 1,
770 [IORING_OP_WRITE_FIXED] = {
773 .unbound_nonreg_file = 1,
777 [IORING_OP_POLL_ADD] = {
779 .unbound_nonreg_file = 1,
781 [IORING_OP_POLL_REMOVE] = {},
782 [IORING_OP_SYNC_FILE_RANGE] = {
785 [IORING_OP_SENDMSG] = {
789 .unbound_nonreg_file = 1,
793 [IORING_OP_RECVMSG] = {
797 .unbound_nonreg_file = 1,
802 [IORING_OP_TIMEOUT] = {
806 [IORING_OP_TIMEOUT_REMOVE] = {},
807 [IORING_OP_ACCEPT] = {
810 .unbound_nonreg_file = 1,
814 [IORING_OP_ASYNC_CANCEL] = {},
815 [IORING_OP_LINK_TIMEOUT] = {
819 [IORING_OP_CONNECT] = {
823 .unbound_nonreg_file = 1,
826 [IORING_OP_FALLOCATE] = {
830 [IORING_OP_OPENAT] = {
834 [IORING_OP_CLOSE] = {
836 .needs_file_no_error = 1,
839 [IORING_OP_FILES_UPDATE] = {
843 [IORING_OP_STATX] = {
851 .unbound_nonreg_file = 1,
855 [IORING_OP_WRITE] = {
858 .unbound_nonreg_file = 1,
862 [IORING_OP_FADVISE] = {
865 [IORING_OP_MADVISE] = {
871 .unbound_nonreg_file = 1,
877 .unbound_nonreg_file = 1,
881 [IORING_OP_OPENAT2] = {
885 [IORING_OP_EPOLL_CTL] = {
886 .unbound_nonreg_file = 1,
889 [IORING_OP_SPLICE] = {
892 .unbound_nonreg_file = 1,
894 [IORING_OP_PROVIDE_BUFFERS] = {},
895 [IORING_OP_REMOVE_BUFFERS] = {},
899 .unbound_nonreg_file = 1,
903 enum io_mem_account {
908 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
909 struct io_comp_state *cs);
910 static void io_cqring_fill_event(struct io_kiocb *req, long res);
911 static void io_put_req(struct io_kiocb *req);
912 static void io_double_put_req(struct io_kiocb *req);
913 static void __io_double_put_req(struct io_kiocb *req);
914 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
915 static void __io_queue_linked_timeout(struct io_kiocb *req);
916 static void io_queue_linked_timeout(struct io_kiocb *req);
917 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
918 struct io_uring_files_update *ip,
920 static int io_prep_work_files(struct io_kiocb *req);
921 static void __io_clean_op(struct io_kiocb *req);
922 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
923 int fd, struct file **out_file, bool fixed);
924 static void __io_queue_sqe(struct io_kiocb *req,
925 const struct io_uring_sqe *sqe,
926 struct io_comp_state *cs);
927 static void io_file_put_work(struct work_struct *work);
929 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
930 struct iovec **iovec, struct iov_iter *iter,
932 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
933 const struct iovec *fast_iov,
934 struct iov_iter *iter, bool force);
936 static struct kmem_cache *req_cachep;
938 static const struct file_operations io_uring_fops;
940 struct sock *io_uring_get_socket(struct file *file)
942 #if defined(CONFIG_UNIX)
943 if (file->f_op == &io_uring_fops) {
944 struct io_ring_ctx *ctx = file->private_data;
946 return ctx->ring_sock->sk;
951 EXPORT_SYMBOL(io_uring_get_socket);
953 static inline void io_clean_op(struct io_kiocb *req)
955 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
960 static void io_sq_thread_drop_mm(void)
962 struct mm_struct *mm = current->mm;
965 kthread_unuse_mm(mm);
970 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
973 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
974 !ctx->sqo_task->mm ||
975 !mmget_not_zero(ctx->sqo_task->mm)))
977 kthread_use_mm(ctx->sqo_task->mm);
983 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
984 struct io_kiocb *req)
986 if (!io_op_defs[req->opcode].needs_mm)
988 return __io_sq_thread_acquire_mm(ctx);
991 static inline void req_set_fail_links(struct io_kiocb *req)
993 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
994 req->flags |= REQ_F_FAIL_LINK;
998 * Note: must call io_req_init_async() for the first time you
999 * touch any members of io_wq_work.
1001 static inline void io_req_init_async(struct io_kiocb *req)
1003 if (req->flags & REQ_F_WORK_INITIALIZED)
1006 memset(&req->work, 0, sizeof(req->work));
1007 req->flags |= REQ_F_WORK_INITIALIZED;
1010 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1012 return ctx->flags & IORING_SETUP_SQPOLL;
1015 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1017 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1019 complete(&ctx->ref_comp);
1022 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1024 return !req->timeout.off;
1027 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1029 struct io_ring_ctx *ctx;
1032 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1036 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1037 if (!ctx->fallback_req)
1041 * Use 5 bits less than the max cq entries, that should give us around
1042 * 32 entries per hash list if totally full and uniformly spread.
1044 hash_bits = ilog2(p->cq_entries);
1048 ctx->cancel_hash_bits = hash_bits;
1049 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1051 if (!ctx->cancel_hash)
1053 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1055 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1056 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1059 ctx->flags = p->flags;
1060 init_waitqueue_head(&ctx->sqo_wait);
1061 init_waitqueue_head(&ctx->cq_wait);
1062 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1063 init_completion(&ctx->ref_comp);
1064 init_completion(&ctx->sq_thread_comp);
1065 idr_init(&ctx->io_buffer_idr);
1066 idr_init(&ctx->personality_idr);
1067 mutex_init(&ctx->uring_lock);
1068 init_waitqueue_head(&ctx->wait);
1069 spin_lock_init(&ctx->completion_lock);
1070 INIT_LIST_HEAD(&ctx->iopoll_list);
1071 INIT_LIST_HEAD(&ctx->defer_list);
1072 INIT_LIST_HEAD(&ctx->timeout_list);
1073 init_waitqueue_head(&ctx->inflight_wait);
1074 spin_lock_init(&ctx->inflight_lock);
1075 INIT_LIST_HEAD(&ctx->inflight_list);
1076 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1077 init_llist_head(&ctx->file_put_llist);
1080 if (ctx->fallback_req)
1081 kmem_cache_free(req_cachep, ctx->fallback_req);
1082 kfree(ctx->cancel_hash);
1087 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1089 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1090 struct io_ring_ctx *ctx = req->ctx;
1092 return seq != ctx->cached_cq_tail
1093 + atomic_read(&ctx->cached_cq_overflow);
1099 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1101 struct io_rings *rings = ctx->rings;
1103 /* order cqe stores with ring update */
1104 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1106 if (wq_has_sleeper(&ctx->cq_wait)) {
1107 wake_up_interruptible(&ctx->cq_wait);
1108 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1113 * Returns true if we need to defer file table putting. This can only happen
1114 * from the error path with REQ_F_COMP_LOCKED set.
1116 static bool io_req_clean_work(struct io_kiocb *req)
1118 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1121 req->flags &= ~REQ_F_WORK_INITIALIZED;
1124 mmdrop(req->work.mm);
1125 req->work.mm = NULL;
1127 if (req->work.creds) {
1128 put_cred(req->work.creds);
1129 req->work.creds = NULL;
1132 struct fs_struct *fs = req->work.fs;
1134 if (req->flags & REQ_F_COMP_LOCKED)
1137 spin_lock(&req->work.fs->lock);
1140 spin_unlock(&req->work.fs->lock);
1143 req->work.fs = NULL;
1149 static void io_prep_async_work(struct io_kiocb *req)
1151 const struct io_op_def *def = &io_op_defs[req->opcode];
1153 io_req_init_async(req);
1155 if (req->flags & REQ_F_ISREG) {
1156 if (def->hash_reg_file || (req->ctx->flags & IORING_SETUP_IOPOLL))
1157 io_wq_hash_work(&req->work, file_inode(req->file));
1159 if (def->unbound_nonreg_file)
1160 req->work.flags |= IO_WQ_WORK_UNBOUND;
1162 if (!req->work.mm && def->needs_mm) {
1163 mmgrab(current->mm);
1164 req->work.mm = current->mm;
1166 if (!req->work.creds)
1167 req->work.creds = get_current_cred();
1168 if (!req->work.fs && def->needs_fs) {
1169 spin_lock(¤t->fs->lock);
1170 if (!current->fs->in_exec) {
1171 req->work.fs = current->fs;
1172 req->work.fs->users++;
1174 req->work.flags |= IO_WQ_WORK_CANCEL;
1176 spin_unlock(¤t->fs->lock);
1178 if (def->needs_fsize)
1179 req->work.fsize = rlimit(RLIMIT_FSIZE);
1181 req->work.fsize = RLIM_INFINITY;
1184 static void io_prep_async_link(struct io_kiocb *req)
1186 struct io_kiocb *cur;
1188 io_prep_async_work(req);
1189 if (req->flags & REQ_F_LINK_HEAD)
1190 list_for_each_entry(cur, &req->link_list, link_list)
1191 io_prep_async_work(cur);
1194 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1196 struct io_ring_ctx *ctx = req->ctx;
1197 struct io_kiocb *link = io_prep_linked_timeout(req);
1199 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1200 &req->work, req->flags);
1201 io_wq_enqueue(ctx->io_wq, &req->work);
1205 static void io_queue_async_work(struct io_kiocb *req)
1207 struct io_kiocb *link;
1209 /* init ->work of the whole link before punting */
1210 io_prep_async_link(req);
1211 link = __io_queue_async_work(req);
1214 io_queue_linked_timeout(link);
1217 static void io_kill_timeout(struct io_kiocb *req)
1221 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1223 atomic_set(&req->ctx->cq_timeouts,
1224 atomic_read(&req->ctx->cq_timeouts) + 1);
1225 list_del_init(&req->timeout.list);
1226 req->flags |= REQ_F_COMP_LOCKED;
1227 io_cqring_fill_event(req, 0);
1232 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1234 struct io_ring_ctx *ctx = req->ctx;
1236 if (!tsk || req->task == tsk)
1238 if ((ctx->flags & IORING_SETUP_SQPOLL) && req->task == ctx->sqo_thread)
1244 * Returns true if we found and killed one or more timeouts
1246 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1248 struct io_kiocb *req, *tmp;
1251 spin_lock_irq(&ctx->completion_lock);
1252 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1253 if (io_task_match(req, tsk)) {
1254 io_kill_timeout(req);
1258 spin_unlock_irq(&ctx->completion_lock);
1259 return canceled != 0;
1262 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1265 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1266 struct io_defer_entry, list);
1267 struct io_kiocb *link;
1269 if (req_need_defer(de->req, de->seq))
1271 list_del_init(&de->list);
1272 /* punt-init is done before queueing for defer */
1273 link = __io_queue_async_work(de->req);
1275 __io_queue_linked_timeout(link);
1276 /* drop submission reference */
1277 link->flags |= REQ_F_COMP_LOCKED;
1281 } while (!list_empty(&ctx->defer_list));
1284 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1286 while (!list_empty(&ctx->timeout_list)) {
1287 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1288 struct io_kiocb, timeout.list);
1290 if (io_is_timeout_noseq(req))
1292 if (req->timeout.target_seq != ctx->cached_cq_tail
1293 - atomic_read(&ctx->cq_timeouts))
1296 list_del_init(&req->timeout.list);
1297 io_kill_timeout(req);
1301 static void io_commit_cqring(struct io_ring_ctx *ctx)
1303 io_flush_timeouts(ctx);
1304 __io_commit_cqring(ctx);
1306 if (unlikely(!list_empty(&ctx->defer_list)))
1307 __io_queue_deferred(ctx);
1310 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1312 struct io_rings *rings = ctx->rings;
1315 tail = ctx->cached_cq_tail;
1317 * writes to the cq entry need to come after reading head; the
1318 * control dependency is enough as we're using WRITE_ONCE to
1321 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1324 ctx->cached_cq_tail++;
1325 return &rings->cqes[tail & ctx->cq_mask];
1328 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1332 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1334 if (!ctx->eventfd_async)
1336 return io_wq_current_is_worker();
1339 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1341 if (waitqueue_active(&ctx->wait))
1342 wake_up(&ctx->wait);
1343 if (waitqueue_active(&ctx->sqo_wait))
1344 wake_up(&ctx->sqo_wait);
1345 if (io_should_trigger_evfd(ctx))
1346 eventfd_signal(ctx->cq_ev_fd, 1);
1349 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1351 if (list_empty(&ctx->cq_overflow_list)) {
1352 clear_bit(0, &ctx->sq_check_overflow);
1353 clear_bit(0, &ctx->cq_check_overflow);
1354 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1358 static inline bool io_match_files(struct io_kiocb *req,
1359 struct files_struct *files)
1363 if (req->flags & REQ_F_WORK_INITIALIZED)
1364 return req->work.files == files;
1368 /* Returns true if there are no backlogged entries after the flush */
1369 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1370 struct task_struct *tsk,
1371 struct files_struct *files)
1373 struct io_rings *rings = ctx->rings;
1374 struct io_kiocb *req, *tmp;
1375 struct io_uring_cqe *cqe;
1376 unsigned long flags;
1380 if (list_empty_careful(&ctx->cq_overflow_list))
1382 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1383 rings->cq_ring_entries))
1387 spin_lock_irqsave(&ctx->completion_lock, flags);
1389 /* if force is set, the ring is going away. always drop after that */
1391 ctx->cq_overflow_flushed = 1;
1394 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1395 if (tsk && req->task != tsk)
1397 if (!io_match_files(req, files))
1400 cqe = io_get_cqring(ctx);
1404 list_move(&req->compl.list, &list);
1406 WRITE_ONCE(cqe->user_data, req->user_data);
1407 WRITE_ONCE(cqe->res, req->result);
1408 WRITE_ONCE(cqe->flags, req->compl.cflags);
1410 WRITE_ONCE(ctx->rings->cq_overflow,
1411 atomic_inc_return(&ctx->cached_cq_overflow));
1415 io_commit_cqring(ctx);
1416 io_cqring_mark_overflow(ctx);
1418 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1419 io_cqring_ev_posted(ctx);
1421 while (!list_empty(&list)) {
1422 req = list_first_entry(&list, struct io_kiocb, compl.list);
1423 list_del(&req->compl.list);
1430 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1432 struct io_ring_ctx *ctx = req->ctx;
1433 struct io_uring_cqe *cqe;
1435 trace_io_uring_complete(ctx, req->user_data, res);
1438 * If we can't get a cq entry, userspace overflowed the
1439 * submission (by quite a lot). Increment the overflow count in
1442 cqe = io_get_cqring(ctx);
1444 WRITE_ONCE(cqe->user_data, req->user_data);
1445 WRITE_ONCE(cqe->res, res);
1446 WRITE_ONCE(cqe->flags, cflags);
1447 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1449 * If we're in ring overflow flush mode, or in task cancel mode,
1450 * then we cannot store the request for later flushing, we need
1451 * to drop it on the floor.
1453 WRITE_ONCE(ctx->rings->cq_overflow,
1454 atomic_inc_return(&ctx->cached_cq_overflow));
1456 if (list_empty(&ctx->cq_overflow_list)) {
1457 set_bit(0, &ctx->sq_check_overflow);
1458 set_bit(0, &ctx->cq_check_overflow);
1459 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1463 req->compl.cflags = cflags;
1464 refcount_inc(&req->refs);
1465 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1469 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1471 __io_cqring_fill_event(req, res, 0);
1474 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1476 struct io_ring_ctx *ctx = req->ctx;
1477 unsigned long flags;
1479 spin_lock_irqsave(&ctx->completion_lock, flags);
1480 __io_cqring_fill_event(req, res, cflags);
1481 io_commit_cqring(ctx);
1482 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1484 io_cqring_ev_posted(ctx);
1487 static void io_submit_flush_completions(struct io_comp_state *cs)
1489 struct io_ring_ctx *ctx = cs->ctx;
1491 spin_lock_irq(&ctx->completion_lock);
1492 while (!list_empty(&cs->list)) {
1493 struct io_kiocb *req;
1495 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1496 list_del(&req->compl.list);
1497 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1498 if (!(req->flags & REQ_F_LINK_HEAD)) {
1499 req->flags |= REQ_F_COMP_LOCKED;
1502 spin_unlock_irq(&ctx->completion_lock);
1504 spin_lock_irq(&ctx->completion_lock);
1507 io_commit_cqring(ctx);
1508 spin_unlock_irq(&ctx->completion_lock);
1510 io_cqring_ev_posted(ctx);
1514 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1515 struct io_comp_state *cs)
1518 io_cqring_add_event(req, res, cflags);
1523 req->compl.cflags = cflags;
1524 list_add_tail(&req->compl.list, &cs->list);
1526 io_submit_flush_completions(cs);
1530 static void io_req_complete(struct io_kiocb *req, long res)
1532 __io_req_complete(req, res, 0, NULL);
1535 static inline bool io_is_fallback_req(struct io_kiocb *req)
1537 return req == (struct io_kiocb *)
1538 ((unsigned long) req->ctx->fallback_req & ~1UL);
1541 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1543 struct io_kiocb *req;
1545 req = ctx->fallback_req;
1546 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1552 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1553 struct io_submit_state *state)
1555 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1556 struct io_kiocb *req;
1558 if (!state->free_reqs) {
1562 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1563 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1566 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1567 * retry single alloc to be on the safe side.
1569 if (unlikely(ret <= 0)) {
1570 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1571 if (!state->reqs[0])
1575 state->free_reqs = ret - 1;
1576 req = state->reqs[ret - 1];
1579 req = state->reqs[state->free_reqs];
1584 return io_get_fallback_req(ctx);
1587 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1591 percpu_ref_put(req->fixed_file_refs);
1596 static bool io_dismantle_req(struct io_kiocb *req)
1603 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1605 return io_req_clean_work(req);
1608 static void __io_free_req_finish(struct io_kiocb *req)
1610 struct io_uring_task *tctx = req->task->io_uring;
1611 struct io_ring_ctx *ctx = req->ctx;
1613 atomic_long_inc(&tctx->req_complete);
1615 wake_up(&tctx->wait);
1616 put_task_struct(req->task);
1618 if (likely(!io_is_fallback_req(req)))
1619 kmem_cache_free(req_cachep, req);
1621 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1622 percpu_ref_put(&ctx->refs);
1625 static void io_req_task_file_table_put(struct callback_head *cb)
1627 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1628 struct fs_struct *fs = req->work.fs;
1630 spin_lock(&req->work.fs->lock);
1633 spin_unlock(&req->work.fs->lock);
1636 req->work.fs = NULL;
1637 __io_free_req_finish(req);
1640 static void __io_free_req(struct io_kiocb *req)
1642 if (!io_dismantle_req(req)) {
1643 __io_free_req_finish(req);
1647 init_task_work(&req->task_work, io_req_task_file_table_put);
1648 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1649 if (unlikely(ret)) {
1650 struct task_struct *tsk;
1652 tsk = io_wq_get_task(req->ctx->io_wq);
1653 task_work_add(tsk, &req->task_work, 0);
1658 static bool io_link_cancel_timeout(struct io_kiocb *req)
1660 struct io_ring_ctx *ctx = req->ctx;
1663 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1665 io_cqring_fill_event(req, -ECANCELED);
1666 io_commit_cqring(ctx);
1667 req->flags &= ~REQ_F_LINK_HEAD;
1675 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1677 struct io_kiocb *link;
1680 if (list_empty(&req->link_list))
1682 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1683 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1686 list_del_init(&link->link_list);
1687 link->flags |= REQ_F_COMP_LOCKED;
1688 wake_ev = io_link_cancel_timeout(link);
1689 req->flags &= ~REQ_F_LINK_TIMEOUT;
1693 static void io_kill_linked_timeout(struct io_kiocb *req)
1695 struct io_ring_ctx *ctx = req->ctx;
1698 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1699 unsigned long flags;
1701 spin_lock_irqsave(&ctx->completion_lock, flags);
1702 wake_ev = __io_kill_linked_timeout(req);
1703 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1705 wake_ev = __io_kill_linked_timeout(req);
1709 io_cqring_ev_posted(ctx);
1712 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1714 struct io_kiocb *nxt;
1717 * The list should never be empty when we are called here. But could
1718 * potentially happen if the chain is messed up, check to be on the
1721 if (unlikely(list_empty(&req->link_list)))
1724 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1725 list_del_init(&req->link_list);
1726 if (!list_empty(&nxt->link_list))
1727 nxt->flags |= REQ_F_LINK_HEAD;
1732 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1734 static void __io_fail_links(struct io_kiocb *req)
1736 struct io_ring_ctx *ctx = req->ctx;
1738 while (!list_empty(&req->link_list)) {
1739 struct io_kiocb *link = list_first_entry(&req->link_list,
1740 struct io_kiocb, link_list);
1742 list_del_init(&link->link_list);
1743 trace_io_uring_fail_link(req, link);
1745 io_cqring_fill_event(link, -ECANCELED);
1746 link->flags |= REQ_F_COMP_LOCKED;
1747 __io_double_put_req(link);
1748 req->flags &= ~REQ_F_LINK_TIMEOUT;
1751 io_commit_cqring(ctx);
1752 io_cqring_ev_posted(ctx);
1755 static void io_fail_links(struct io_kiocb *req)
1757 struct io_ring_ctx *ctx = req->ctx;
1759 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1760 unsigned long flags;
1762 spin_lock_irqsave(&ctx->completion_lock, flags);
1763 __io_fail_links(req);
1764 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1766 __io_fail_links(req);
1769 io_cqring_ev_posted(ctx);
1772 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1774 req->flags &= ~REQ_F_LINK_HEAD;
1775 if (req->flags & REQ_F_LINK_TIMEOUT)
1776 io_kill_linked_timeout(req);
1779 * If LINK is set, we have dependent requests in this chain. If we
1780 * didn't fail this request, queue the first one up, moving any other
1781 * dependencies to the next request. In case of failure, fail the rest
1784 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1785 return io_req_link_next(req);
1790 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1792 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1794 return __io_req_find_next(req);
1797 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb,
1800 struct task_struct *tsk = req->task;
1801 struct io_ring_ctx *ctx = req->ctx;
1804 if (tsk->flags & PF_EXITING)
1808 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1809 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1810 * processing task_work. There's no reliable way to tell if TWA_RESUME
1814 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1815 notify = TWA_SIGNAL;
1817 ret = task_work_add(tsk, cb, notify);
1819 wake_up_process(tsk);
1824 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1826 struct io_ring_ctx *ctx = req->ctx;
1828 spin_lock_irq(&ctx->completion_lock);
1829 io_cqring_fill_event(req, error);
1830 io_commit_cqring(ctx);
1831 spin_unlock_irq(&ctx->completion_lock);
1833 io_cqring_ev_posted(ctx);
1834 req_set_fail_links(req);
1835 io_double_put_req(req);
1838 static void io_req_task_cancel(struct callback_head *cb)
1840 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1841 struct io_ring_ctx *ctx = req->ctx;
1843 __io_req_task_cancel(req, -ECANCELED);
1844 percpu_ref_put(&ctx->refs);
1847 static void __io_req_task_submit(struct io_kiocb *req)
1849 struct io_ring_ctx *ctx = req->ctx;
1851 if (!__io_sq_thread_acquire_mm(ctx)) {
1852 mutex_lock(&ctx->uring_lock);
1853 __io_queue_sqe(req, NULL, NULL);
1854 mutex_unlock(&ctx->uring_lock);
1856 __io_req_task_cancel(req, -EFAULT);
1860 static void io_req_task_submit(struct callback_head *cb)
1862 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1863 struct io_ring_ctx *ctx = req->ctx;
1865 __io_req_task_submit(req);
1866 percpu_ref_put(&ctx->refs);
1869 static void io_req_task_queue(struct io_kiocb *req)
1873 init_task_work(&req->task_work, io_req_task_submit);
1874 percpu_ref_get(&req->ctx->refs);
1876 ret = io_req_task_work_add(req, &req->task_work, true);
1877 if (unlikely(ret)) {
1878 struct task_struct *tsk;
1880 init_task_work(&req->task_work, io_req_task_cancel);
1881 tsk = io_wq_get_task(req->ctx->io_wq);
1882 task_work_add(tsk, &req->task_work, 0);
1883 wake_up_process(tsk);
1887 static void io_queue_next(struct io_kiocb *req)
1889 struct io_kiocb *nxt = io_req_find_next(req);
1892 io_req_task_queue(nxt);
1895 static void io_free_req(struct io_kiocb *req)
1902 void *reqs[IO_IOPOLL_BATCH];
1905 struct task_struct *task;
1909 static inline void io_init_req_batch(struct req_batch *rb)
1916 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1917 struct req_batch *rb)
1919 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1920 percpu_ref_put_many(&ctx->refs, rb->to_free);
1924 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1925 struct req_batch *rb)
1928 __io_req_free_batch_flush(ctx, rb);
1930 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
1931 put_task_struct_many(rb->task, rb->task_refs);
1936 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1938 if (unlikely(io_is_fallback_req(req))) {
1942 if (req->flags & REQ_F_LINK_HEAD)
1945 if (req->task != rb->task) {
1947 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
1948 put_task_struct_many(rb->task, rb->task_refs);
1950 rb->task = req->task;
1955 WARN_ON_ONCE(io_dismantle_req(req));
1956 rb->reqs[rb->to_free++] = req;
1957 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1958 __io_req_free_batch_flush(req->ctx, rb);
1962 * Drop reference to request, return next in chain (if there is one) if this
1963 * was the last reference to this request.
1965 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1967 struct io_kiocb *nxt = NULL;
1969 if (refcount_dec_and_test(&req->refs)) {
1970 nxt = io_req_find_next(req);
1976 static void io_put_req(struct io_kiocb *req)
1978 if (refcount_dec_and_test(&req->refs))
1982 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1984 struct io_kiocb *nxt;
1987 * A ref is owned by io-wq in which context we're. So, if that's the
1988 * last one, it's safe to steal next work. False negatives are Ok,
1989 * it just will be re-punted async in io_put_work()
1991 if (refcount_read(&req->refs) != 1)
1994 nxt = io_req_find_next(req);
1995 return nxt ? &nxt->work : NULL;
1999 * Must only be used if we don't need to care about links, usually from
2000 * within the completion handling itself.
2002 static void __io_double_put_req(struct io_kiocb *req)
2004 /* drop both submit and complete references */
2005 if (refcount_sub_and_test(2, &req->refs))
2009 static void io_double_put_req(struct io_kiocb *req)
2011 /* drop both submit and complete references */
2012 if (refcount_sub_and_test(2, &req->refs))
2016 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2018 struct io_rings *rings = ctx->rings;
2020 if (test_bit(0, &ctx->cq_check_overflow)) {
2022 * noflush == true is from the waitqueue handler, just ensure
2023 * we wake up the task, and the next invocation will flush the
2024 * entries. We cannot safely to it from here.
2026 if (noflush && !list_empty(&ctx->cq_overflow_list))
2029 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2032 /* See comment at the top of this file */
2034 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2037 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2039 struct io_rings *rings = ctx->rings;
2041 /* make sure SQ entry isn't read before tail */
2042 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2045 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2047 unsigned int cflags;
2049 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2050 cflags |= IORING_CQE_F_BUFFER;
2051 req->flags &= ~REQ_F_BUFFER_SELECTED;
2056 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2058 struct io_buffer *kbuf;
2060 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2061 return io_put_kbuf(req, kbuf);
2064 static inline bool io_run_task_work(void)
2067 * Not safe to run on exiting task, and the task_work handling will
2068 * not add work to such a task.
2070 if (unlikely(current->flags & PF_EXITING))
2072 if (current->task_works) {
2073 __set_current_state(TASK_RUNNING);
2081 static void io_iopoll_queue(struct list_head *again)
2083 struct io_kiocb *req;
2086 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2087 list_del(&req->inflight_entry);
2088 __io_complete_rw(req, -EAGAIN, 0, NULL);
2089 } while (!list_empty(again));
2093 * Find and free completed poll iocbs
2095 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2096 struct list_head *done)
2098 struct req_batch rb;
2099 struct io_kiocb *req;
2102 /* order with ->result store in io_complete_rw_iopoll() */
2105 io_init_req_batch(&rb);
2106 while (!list_empty(done)) {
2109 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2110 if (READ_ONCE(req->result) == -EAGAIN) {
2112 req->iopoll_completed = 0;
2113 list_move_tail(&req->inflight_entry, &again);
2116 list_del(&req->inflight_entry);
2118 if (req->flags & REQ_F_BUFFER_SELECTED)
2119 cflags = io_put_rw_kbuf(req);
2121 __io_cqring_fill_event(req, req->result, cflags);
2124 if (refcount_dec_and_test(&req->refs))
2125 io_req_free_batch(&rb, req);
2128 io_commit_cqring(ctx);
2129 if (ctx->flags & IORING_SETUP_SQPOLL)
2130 io_cqring_ev_posted(ctx);
2131 io_req_free_batch_finish(ctx, &rb);
2133 if (!list_empty(&again))
2134 io_iopoll_queue(&again);
2137 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2140 struct io_kiocb *req, *tmp;
2146 * Only spin for completions if we don't have multiple devices hanging
2147 * off our complete list, and we're under the requested amount.
2149 spin = !ctx->poll_multi_file && *nr_events < min;
2152 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2153 struct kiocb *kiocb = &req->rw.kiocb;
2156 * Move completed and retryable entries to our local lists.
2157 * If we find a request that requires polling, break out
2158 * and complete those lists first, if we have entries there.
2160 if (READ_ONCE(req->iopoll_completed)) {
2161 list_move_tail(&req->inflight_entry, &done);
2164 if (!list_empty(&done))
2167 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2171 /* iopoll may have completed current req */
2172 if (READ_ONCE(req->iopoll_completed))
2173 list_move_tail(&req->inflight_entry, &done);
2180 if (!list_empty(&done))
2181 io_iopoll_complete(ctx, nr_events, &done);
2187 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2188 * non-spinning poll check - we'll still enter the driver poll loop, but only
2189 * as a non-spinning completion check.
2191 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2194 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2197 ret = io_do_iopoll(ctx, nr_events, min);
2200 if (*nr_events >= min)
2208 * We can't just wait for polled events to come to us, we have to actively
2209 * find and complete them.
2211 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2213 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2216 mutex_lock(&ctx->uring_lock);
2217 while (!list_empty(&ctx->iopoll_list)) {
2218 unsigned int nr_events = 0;
2220 io_do_iopoll(ctx, &nr_events, 0);
2222 /* let it sleep and repeat later if can't complete a request */
2226 * Ensure we allow local-to-the-cpu processing to take place,
2227 * in this case we need to ensure that we reap all events.
2228 * Also let task_work, etc. to progress by releasing the mutex
2230 if (need_resched()) {
2231 mutex_unlock(&ctx->uring_lock);
2233 mutex_lock(&ctx->uring_lock);
2236 mutex_unlock(&ctx->uring_lock);
2239 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2241 unsigned int nr_events = 0;
2242 int iters = 0, ret = 0;
2245 * We disallow the app entering submit/complete with polling, but we
2246 * still need to lock the ring to prevent racing with polled issue
2247 * that got punted to a workqueue.
2249 mutex_lock(&ctx->uring_lock);
2252 * Don't enter poll loop if we already have events pending.
2253 * If we do, we can potentially be spinning for commands that
2254 * already triggered a CQE (eg in error).
2256 if (io_cqring_events(ctx, false))
2260 * If a submit got punted to a workqueue, we can have the
2261 * application entering polling for a command before it gets
2262 * issued. That app will hold the uring_lock for the duration
2263 * of the poll right here, so we need to take a breather every
2264 * now and then to ensure that the issue has a chance to add
2265 * the poll to the issued list. Otherwise we can spin here
2266 * forever, while the workqueue is stuck trying to acquire the
2269 if (!(++iters & 7)) {
2270 mutex_unlock(&ctx->uring_lock);
2272 mutex_lock(&ctx->uring_lock);
2275 ret = io_iopoll_getevents(ctx, &nr_events, min);
2279 } while (min && !nr_events && !need_resched());
2281 mutex_unlock(&ctx->uring_lock);
2285 static void kiocb_end_write(struct io_kiocb *req)
2288 * Tell lockdep we inherited freeze protection from submission
2291 if (req->flags & REQ_F_ISREG) {
2292 struct inode *inode = file_inode(req->file);
2294 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2296 file_end_write(req->file);
2299 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2300 struct io_comp_state *cs)
2302 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2305 if (kiocb->ki_flags & IOCB_WRITE)
2306 kiocb_end_write(req);
2308 if (res != req->result)
2309 req_set_fail_links(req);
2310 if (req->flags & REQ_F_BUFFER_SELECTED)
2311 cflags = io_put_rw_kbuf(req);
2312 __io_req_complete(req, res, cflags, cs);
2316 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2318 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2319 ssize_t ret = -ECANCELED;
2320 struct iov_iter iter;
2328 switch (req->opcode) {
2329 case IORING_OP_READV:
2330 case IORING_OP_READ_FIXED:
2331 case IORING_OP_READ:
2334 case IORING_OP_WRITEV:
2335 case IORING_OP_WRITE_FIXED:
2336 case IORING_OP_WRITE:
2340 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2346 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2349 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2357 req_set_fail_links(req);
2358 io_req_complete(req, ret);
2363 static bool io_rw_reissue(struct io_kiocb *req, long res)
2366 umode_t mode = file_inode(req->file)->i_mode;
2369 if (!S_ISBLK(mode) && !S_ISREG(mode))
2371 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2374 ret = io_sq_thread_acquire_mm(req->ctx, req);
2376 if (io_resubmit_prep(req, ret)) {
2377 refcount_inc(&req->refs);
2378 io_queue_async_work(req);
2386 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2387 struct io_comp_state *cs)
2389 if (!io_rw_reissue(req, res))
2390 io_complete_rw_common(&req->rw.kiocb, res, cs);
2393 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2395 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2397 __io_complete_rw(req, res, res2, NULL);
2400 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2402 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2404 if (kiocb->ki_flags & IOCB_WRITE)
2405 kiocb_end_write(req);
2407 if (res != -EAGAIN && res != req->result)
2408 req_set_fail_links(req);
2410 WRITE_ONCE(req->result, res);
2411 /* order with io_poll_complete() checking ->result */
2413 WRITE_ONCE(req->iopoll_completed, 1);
2417 * After the iocb has been issued, it's safe to be found on the poll list.
2418 * Adding the kiocb to the list AFTER submission ensures that we don't
2419 * find it from a io_iopoll_getevents() thread before the issuer is done
2420 * accessing the kiocb cookie.
2422 static void io_iopoll_req_issued(struct io_kiocb *req)
2424 struct io_ring_ctx *ctx = req->ctx;
2427 * Track whether we have multiple files in our lists. This will impact
2428 * how we do polling eventually, not spinning if we're on potentially
2429 * different devices.
2431 if (list_empty(&ctx->iopoll_list)) {
2432 ctx->poll_multi_file = false;
2433 } else if (!ctx->poll_multi_file) {
2434 struct io_kiocb *list_req;
2436 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2438 if (list_req->file != req->file)
2439 ctx->poll_multi_file = true;
2443 * For fast devices, IO may have already completed. If it has, add
2444 * it to the front so we find it first.
2446 if (READ_ONCE(req->iopoll_completed))
2447 list_add(&req->inflight_entry, &ctx->iopoll_list);
2449 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2451 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2452 wq_has_sleeper(&ctx->sqo_wait))
2453 wake_up(&ctx->sqo_wait);
2456 static void __io_state_file_put(struct io_submit_state *state)
2458 if (state->has_refs)
2459 fput_many(state->file, state->has_refs);
2463 static inline void io_state_file_put(struct io_submit_state *state)
2466 __io_state_file_put(state);
2470 * Get as many references to a file as we have IOs left in this submission,
2471 * assuming most submissions are for one file, or at least that each file
2472 * has more than one submission.
2474 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2480 if (state->fd == fd) {
2485 __io_state_file_put(state);
2487 state->file = fget_many(fd, state->ios_left);
2493 state->has_refs = state->ios_left;
2497 static bool io_bdev_nowait(struct block_device *bdev)
2500 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2507 * If we tracked the file through the SCM inflight mechanism, we could support
2508 * any file. For now, just ensure that anything potentially problematic is done
2511 static bool io_file_supports_async(struct file *file, int rw)
2513 umode_t mode = file_inode(file)->i_mode;
2515 if (S_ISBLK(mode)) {
2516 if (io_bdev_nowait(file->f_inode->i_bdev))
2520 if (S_ISCHR(mode) || S_ISSOCK(mode))
2522 if (S_ISREG(mode)) {
2523 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2524 file->f_op != &io_uring_fops)
2529 /* any ->read/write should understand O_NONBLOCK */
2530 if (file->f_flags & O_NONBLOCK)
2533 if (!(file->f_mode & FMODE_NOWAIT))
2537 return file->f_op->read_iter != NULL;
2539 return file->f_op->write_iter != NULL;
2542 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2543 bool force_nonblock)
2545 struct io_ring_ctx *ctx = req->ctx;
2546 struct kiocb *kiocb = &req->rw.kiocb;
2550 if (S_ISREG(file_inode(req->file)->i_mode))
2551 req->flags |= REQ_F_ISREG;
2553 kiocb->ki_pos = READ_ONCE(sqe->off);
2554 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2555 req->flags |= REQ_F_CUR_POS;
2556 kiocb->ki_pos = req->file->f_pos;
2558 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2559 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2560 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2564 ioprio = READ_ONCE(sqe->ioprio);
2566 ret = ioprio_check_cap(ioprio);
2570 kiocb->ki_ioprio = ioprio;
2572 kiocb->ki_ioprio = get_current_ioprio();
2574 /* don't allow async punt if RWF_NOWAIT was requested */
2575 if (kiocb->ki_flags & IOCB_NOWAIT)
2576 req->flags |= REQ_F_NOWAIT;
2579 kiocb->ki_flags |= IOCB_NOWAIT;
2581 if (ctx->flags & IORING_SETUP_IOPOLL) {
2582 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2583 !kiocb->ki_filp->f_op->iopoll)
2586 kiocb->ki_flags |= IOCB_HIPRI;
2587 kiocb->ki_complete = io_complete_rw_iopoll;
2588 req->iopoll_completed = 0;
2590 if (kiocb->ki_flags & IOCB_HIPRI)
2592 kiocb->ki_complete = io_complete_rw;
2595 req->rw.addr = READ_ONCE(sqe->addr);
2596 req->rw.len = READ_ONCE(sqe->len);
2597 req->buf_index = READ_ONCE(sqe->buf_index);
2601 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2607 case -ERESTARTNOINTR:
2608 case -ERESTARTNOHAND:
2609 case -ERESTART_RESTARTBLOCK:
2611 * We can't just restart the syscall, since previously
2612 * submitted sqes may already be in progress. Just fail this
2618 kiocb->ki_complete(kiocb, ret, 0);
2622 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2623 struct io_comp_state *cs)
2625 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2627 /* add previously done IO, if any */
2628 if (req->io && req->io->rw.bytes_done > 0) {
2630 ret = req->io->rw.bytes_done;
2632 ret += req->io->rw.bytes_done;
2635 if (req->flags & REQ_F_CUR_POS)
2636 req->file->f_pos = kiocb->ki_pos;
2637 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2638 __io_complete_rw(req, ret, 0, cs);
2640 io_rw_done(kiocb, ret);
2643 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2644 struct iov_iter *iter)
2646 struct io_ring_ctx *ctx = req->ctx;
2647 size_t len = req->rw.len;
2648 struct io_mapped_ubuf *imu;
2649 u16 index, buf_index;
2653 /* attempt to use fixed buffers without having provided iovecs */
2654 if (unlikely(!ctx->user_bufs))
2657 buf_index = req->buf_index;
2658 if (unlikely(buf_index >= ctx->nr_user_bufs))
2661 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2662 imu = &ctx->user_bufs[index];
2663 buf_addr = req->rw.addr;
2666 if (buf_addr + len < buf_addr)
2668 /* not inside the mapped region */
2669 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2673 * May not be a start of buffer, set size appropriately
2674 * and advance us to the beginning.
2676 offset = buf_addr - imu->ubuf;
2677 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2681 * Don't use iov_iter_advance() here, as it's really slow for
2682 * using the latter parts of a big fixed buffer - it iterates
2683 * over each segment manually. We can cheat a bit here, because
2686 * 1) it's a BVEC iter, we set it up
2687 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2688 * first and last bvec
2690 * So just find our index, and adjust the iterator afterwards.
2691 * If the offset is within the first bvec (or the whole first
2692 * bvec, just use iov_iter_advance(). This makes it easier
2693 * since we can just skip the first segment, which may not
2694 * be PAGE_SIZE aligned.
2696 const struct bio_vec *bvec = imu->bvec;
2698 if (offset <= bvec->bv_len) {
2699 iov_iter_advance(iter, offset);
2701 unsigned long seg_skip;
2703 /* skip first vec */
2704 offset -= bvec->bv_len;
2705 seg_skip = 1 + (offset >> PAGE_SHIFT);
2707 iter->bvec = bvec + seg_skip;
2708 iter->nr_segs -= seg_skip;
2709 iter->count -= bvec->bv_len + offset;
2710 iter->iov_offset = offset & ~PAGE_MASK;
2717 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2720 mutex_unlock(&ctx->uring_lock);
2723 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2726 * "Normal" inline submissions always hold the uring_lock, since we
2727 * grab it from the system call. Same is true for the SQPOLL offload.
2728 * The only exception is when we've detached the request and issue it
2729 * from an async worker thread, grab the lock for that case.
2732 mutex_lock(&ctx->uring_lock);
2735 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2736 int bgid, struct io_buffer *kbuf,
2739 struct io_buffer *head;
2741 if (req->flags & REQ_F_BUFFER_SELECTED)
2744 io_ring_submit_lock(req->ctx, needs_lock);
2746 lockdep_assert_held(&req->ctx->uring_lock);
2748 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2750 if (!list_empty(&head->list)) {
2751 kbuf = list_last_entry(&head->list, struct io_buffer,
2753 list_del(&kbuf->list);
2756 idr_remove(&req->ctx->io_buffer_idr, bgid);
2758 if (*len > kbuf->len)
2761 kbuf = ERR_PTR(-ENOBUFS);
2764 io_ring_submit_unlock(req->ctx, needs_lock);
2769 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2772 struct io_buffer *kbuf;
2775 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2776 bgid = req->buf_index;
2777 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2780 req->rw.addr = (u64) (unsigned long) kbuf;
2781 req->flags |= REQ_F_BUFFER_SELECTED;
2782 return u64_to_user_ptr(kbuf->addr);
2785 #ifdef CONFIG_COMPAT
2786 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2789 struct compat_iovec __user *uiov;
2790 compat_ssize_t clen;
2794 uiov = u64_to_user_ptr(req->rw.addr);
2795 if (!access_ok(uiov, sizeof(*uiov)))
2797 if (__get_user(clen, &uiov->iov_len))
2803 buf = io_rw_buffer_select(req, &len, needs_lock);
2805 return PTR_ERR(buf);
2806 iov[0].iov_base = buf;
2807 iov[0].iov_len = (compat_size_t) len;
2812 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2815 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2819 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2822 len = iov[0].iov_len;
2825 buf = io_rw_buffer_select(req, &len, needs_lock);
2827 return PTR_ERR(buf);
2828 iov[0].iov_base = buf;
2829 iov[0].iov_len = len;
2833 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2836 if (req->flags & REQ_F_BUFFER_SELECTED) {
2837 struct io_buffer *kbuf;
2839 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2840 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2841 iov[0].iov_len = kbuf->len;
2846 else if (req->rw.len > 1)
2849 #ifdef CONFIG_COMPAT
2850 if (req->ctx->compat)
2851 return io_compat_import(req, iov, needs_lock);
2854 return __io_iov_buffer_select(req, iov, needs_lock);
2857 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2858 struct iovec **iovec, struct iov_iter *iter,
2861 void __user *buf = u64_to_user_ptr(req->rw.addr);
2862 size_t sqe_len = req->rw.len;
2866 opcode = req->opcode;
2867 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2869 return io_import_fixed(req, rw, iter);
2872 /* buffer index only valid with fixed read/write, or buffer select */
2873 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2876 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2877 if (req->flags & REQ_F_BUFFER_SELECT) {
2878 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2880 return PTR_ERR(buf);
2881 req->rw.len = sqe_len;
2884 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2886 return ret < 0 ? ret : sqe_len;
2889 if (req->flags & REQ_F_BUFFER_SELECT) {
2890 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2892 ret = (*iovec)->iov_len;
2893 iov_iter_init(iter, rw, *iovec, 1, ret);
2899 #ifdef CONFIG_COMPAT
2900 if (req->ctx->compat)
2901 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2905 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2908 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2909 struct iovec **iovec, struct iov_iter *iter,
2913 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
2915 return iov_iter_count(&req->io->rw.iter);
2918 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2920 return kiocb->ki_filp->f_mode & FMODE_STREAM ? NULL : &kiocb->ki_pos;
2924 * For files that don't have ->read_iter() and ->write_iter(), handle them
2925 * by looping over ->read() or ->write() manually.
2927 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2928 struct iov_iter *iter)
2933 * Don't support polled IO through this interface, and we can't
2934 * support non-blocking either. For the latter, this just causes
2935 * the kiocb to be handled from an async context.
2937 if (kiocb->ki_flags & IOCB_HIPRI)
2939 if (kiocb->ki_flags & IOCB_NOWAIT)
2942 while (iov_iter_count(iter)) {
2946 if (!iov_iter_is_bvec(iter)) {
2947 iovec = iov_iter_iovec(iter);
2949 /* fixed buffers import bvec */
2950 iovec.iov_base = kmap(iter->bvec->bv_page)
2952 iovec.iov_len = min(iter->count,
2953 iter->bvec->bv_len - iter->iov_offset);
2957 nr = file->f_op->read(file, iovec.iov_base,
2958 iovec.iov_len, io_kiocb_ppos(kiocb));
2960 nr = file->f_op->write(file, iovec.iov_base,
2961 iovec.iov_len, io_kiocb_ppos(kiocb));
2964 if (iov_iter_is_bvec(iter))
2965 kunmap(iter->bvec->bv_page);
2973 if (nr != iovec.iov_len)
2975 iov_iter_advance(iter, nr);
2981 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
2982 const struct iovec *fast_iov, struct iov_iter *iter)
2984 struct io_async_rw *rw = &req->io->rw;
2986 memcpy(&rw->iter, iter, sizeof(*iter));
2987 rw->free_iovec = NULL;
2989 /* can only be fixed buffers, no need to do anything */
2990 if (iter->type == ITER_BVEC)
2993 unsigned iov_off = 0;
2995 rw->iter.iov = rw->fast_iov;
2996 if (iter->iov != fast_iov) {
2997 iov_off = iter->iov - fast_iov;
2998 rw->iter.iov += iov_off;
3000 if (rw->fast_iov != fast_iov)
3001 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3002 sizeof(struct iovec) * iter->nr_segs);
3004 rw->free_iovec = iovec;
3005 req->flags |= REQ_F_NEED_CLEANUP;
3009 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
3011 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
3012 return req->io == NULL;
3015 static int io_alloc_async_ctx(struct io_kiocb *req)
3017 if (!io_op_defs[req->opcode].async_ctx)
3020 return __io_alloc_async_ctx(req);
3023 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3024 const struct iovec *fast_iov,
3025 struct iov_iter *iter, bool force)
3027 if (!force && !io_op_defs[req->opcode].async_ctx)
3030 if (__io_alloc_async_ctx(req))
3033 io_req_map_rw(req, iovec, fast_iov, iter);
3038 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
3039 bool force_nonblock)
3041 struct io_async_rw *iorw = &req->io->rw;
3045 iorw->iter.iov = iov = iorw->fast_iov;
3046 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, !force_nonblock);
3047 if (unlikely(ret < 0))
3050 iorw->iter.iov = iov;
3051 io_req_map_rw(req, iorw->iter.iov, iorw->fast_iov, &iorw->iter);
3055 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3056 bool force_nonblock)
3060 ret = io_prep_rw(req, sqe, force_nonblock);
3064 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3067 /* either don't need iovec imported or already have it */
3068 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3070 return io_rw_prep_async(req, READ, force_nonblock);
3074 * This is our waitqueue callback handler, registered through lock_page_async()
3075 * when we initially tried to do the IO with the iocb armed our waitqueue.
3076 * This gets called when the page is unlocked, and we generally expect that to
3077 * happen when the page IO is completed and the page is now uptodate. This will
3078 * queue a task_work based retry of the operation, attempting to copy the data
3079 * again. If the latter fails because the page was NOT uptodate, then we will
3080 * do a thread based blocking retry of the operation. That's the unexpected
3083 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3084 int sync, void *arg)
3086 struct wait_page_queue *wpq;
3087 struct io_kiocb *req = wait->private;
3088 struct wait_page_key *key = arg;
3091 wpq = container_of(wait, struct wait_page_queue, wait);
3093 if (!wake_page_match(wpq, key))
3096 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3097 list_del_init(&wait->entry);
3099 init_task_work(&req->task_work, io_req_task_submit);
3100 percpu_ref_get(&req->ctx->refs);
3102 /* submit ref gets dropped, acquire a new one */
3103 refcount_inc(&req->refs);
3104 ret = io_req_task_work_add(req, &req->task_work, true);
3105 if (unlikely(ret)) {
3106 struct task_struct *tsk;
3108 /* queue just for cancelation */
3109 init_task_work(&req->task_work, io_req_task_cancel);
3110 tsk = io_wq_get_task(req->ctx->io_wq);
3111 task_work_add(tsk, &req->task_work, 0);
3112 wake_up_process(tsk);
3118 * This controls whether a given IO request should be armed for async page
3119 * based retry. If we return false here, the request is handed to the async
3120 * worker threads for retry. If we're doing buffered reads on a regular file,
3121 * we prepare a private wait_page_queue entry and retry the operation. This
3122 * will either succeed because the page is now uptodate and unlocked, or it
3123 * will register a callback when the page is unlocked at IO completion. Through
3124 * that callback, io_uring uses task_work to setup a retry of the operation.
3125 * That retry will attempt the buffered read again. The retry will generally
3126 * succeed, or in rare cases where it fails, we then fall back to using the
3127 * async worker threads for a blocking retry.
3129 static bool io_rw_should_retry(struct io_kiocb *req)
3131 struct wait_page_queue *wait = &req->io->rw.wpq;
3132 struct kiocb *kiocb = &req->rw.kiocb;
3134 /* never retry for NOWAIT, we just complete with -EAGAIN */
3135 if (req->flags & REQ_F_NOWAIT)
3138 /* Only for buffered IO */
3139 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3143 * just use poll if we can, and don't attempt if the fs doesn't
3144 * support callback based unlocks
3146 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3149 wait->wait.func = io_async_buf_func;
3150 wait->wait.private = req;
3151 wait->wait.flags = 0;
3152 INIT_LIST_HEAD(&wait->wait.entry);
3153 kiocb->ki_flags |= IOCB_WAITQ;
3154 kiocb->ki_flags &= ~IOCB_NOWAIT;
3155 kiocb->ki_waitq = wait;
3159 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3161 if (req->file->f_op->read_iter)
3162 return call_read_iter(req->file, &req->rw.kiocb, iter);
3163 else if (req->file->f_op->read)
3164 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3169 static int io_read(struct io_kiocb *req, bool force_nonblock,
3170 struct io_comp_state *cs)
3172 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3173 struct kiocb *kiocb = &req->rw.kiocb;
3174 struct iov_iter __iter, *iter = &__iter;
3175 ssize_t io_size, ret, ret2;
3180 iter = &req->io->rw.iter;
3182 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3185 iov_count = iov_iter_count(iter);
3187 req->result = io_size;
3190 /* Ensure we clear previously set non-block flag */
3191 if (!force_nonblock)
3192 kiocb->ki_flags &= ~IOCB_NOWAIT;
3194 /* If the file doesn't support async, just async punt */
3195 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3199 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3203 ret = io_iter_do_read(req, iter);
3207 } else if (ret == -EIOCBQUEUED) {
3210 } else if (ret == -EAGAIN) {
3211 /* IOPOLL retry should happen for io-wq threads */
3212 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3214 /* no retry on NONBLOCK marked file */
3215 if (req->file->f_flags & O_NONBLOCK)
3217 /* some cases will consume bytes even on error returns */
3218 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3221 } else if (ret < 0) {
3222 /* make sure -ERESTARTSYS -> -EINTR is done */
3226 /* read it all, or we did blocking attempt. no retry. */
3227 if (!iov_iter_count(iter) || !force_nonblock ||
3228 (req->file->f_flags & O_NONBLOCK))
3233 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3240 /* it's copied and will be cleaned with ->io */
3242 /* now use our persistent iterator, if we aren't already */
3243 iter = &req->io->rw.iter;
3245 req->io->rw.bytes_done += ret;
3246 /* if we can retry, do so with the callbacks armed */
3247 if (!io_rw_should_retry(req)) {
3248 kiocb->ki_flags &= ~IOCB_WAITQ;
3253 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3254 * get -EIOCBQUEUED, then we'll get a notification when the desired
3255 * page gets unlocked. We can also get a partial read here, and if we
3256 * do, then just retry at the new offset.
3258 ret = io_iter_do_read(req, iter);
3259 if (ret == -EIOCBQUEUED) {
3262 } else if (ret > 0 && ret < io_size) {
3263 /* we got some bytes, but not all. retry. */
3267 kiocb_done(kiocb, ret, cs);
3270 /* it's reportedly faster than delegating the null check to kfree() */
3276 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3277 bool force_nonblock)
3281 ret = io_prep_rw(req, sqe, force_nonblock);
3285 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3288 /* either don't need iovec imported or already have it */
3289 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3291 return io_rw_prep_async(req, WRITE, force_nonblock);
3294 static int io_write(struct io_kiocb *req, bool force_nonblock,
3295 struct io_comp_state *cs)
3297 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3298 struct kiocb *kiocb = &req->rw.kiocb;
3299 struct iov_iter __iter, *iter = &__iter;
3301 ssize_t ret, ret2, io_size;
3304 iter = &req->io->rw.iter;
3306 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3309 iov_count = iov_iter_count(iter);
3311 req->result = io_size;
3313 /* Ensure we clear previously set non-block flag */
3314 if (!force_nonblock)
3315 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3317 /* If the file doesn't support async, just async punt */
3318 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3321 /* file path doesn't support NOWAIT for non-direct_IO */
3322 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3323 (req->flags & REQ_F_ISREG))
3326 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3331 * Open-code file_start_write here to grab freeze protection,
3332 * which will be released by another thread in
3333 * io_complete_rw(). Fool lockdep by telling it the lock got
3334 * released so that it doesn't complain about the held lock when
3335 * we return to userspace.
3337 if (req->flags & REQ_F_ISREG) {
3338 __sb_start_write(file_inode(req->file)->i_sb,
3339 SB_FREEZE_WRITE, true);
3340 __sb_writers_release(file_inode(req->file)->i_sb,
3343 kiocb->ki_flags |= IOCB_WRITE;
3345 if (req->file->f_op->write_iter)
3346 ret2 = call_write_iter(req->file, kiocb, iter);
3347 else if (req->file->f_op->write)
3348 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3353 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3354 * retry them without IOCB_NOWAIT.
3356 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3358 /* no retry on NONBLOCK marked file */
3359 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3361 if (!force_nonblock || ret2 != -EAGAIN) {
3362 /* IOPOLL retry should happen for io-wq threads */
3363 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3366 kiocb_done(kiocb, ret2, cs);
3369 /* some cases will consume bytes even on error returns */
3370 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3371 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3376 /* it's reportedly faster than delegating the null check to kfree() */
3382 static int __io_splice_prep(struct io_kiocb *req,
3383 const struct io_uring_sqe *sqe)
3385 struct io_splice* sp = &req->splice;
3386 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3389 if (req->flags & REQ_F_NEED_CLEANUP)
3391 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3395 sp->len = READ_ONCE(sqe->len);
3396 sp->flags = READ_ONCE(sqe->splice_flags);
3398 if (unlikely(sp->flags & ~valid_flags))
3401 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3402 (sp->flags & SPLICE_F_FD_IN_FIXED));
3405 req->flags |= REQ_F_NEED_CLEANUP;
3407 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3409 * Splice operation will be punted aync, and here need to
3410 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3412 io_req_init_async(req);
3413 req->work.flags |= IO_WQ_WORK_UNBOUND;
3419 static int io_tee_prep(struct io_kiocb *req,
3420 const struct io_uring_sqe *sqe)
3422 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3424 return __io_splice_prep(req, sqe);
3427 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3429 struct io_splice *sp = &req->splice;
3430 struct file *in = sp->file_in;
3431 struct file *out = sp->file_out;
3432 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3438 ret = do_tee(in, out, sp->len, flags);
3440 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3441 req->flags &= ~REQ_F_NEED_CLEANUP;
3444 req_set_fail_links(req);
3445 io_req_complete(req, ret);
3449 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3451 struct io_splice* sp = &req->splice;
3453 sp->off_in = READ_ONCE(sqe->splice_off_in);
3454 sp->off_out = READ_ONCE(sqe->off);
3455 return __io_splice_prep(req, sqe);
3458 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3460 struct io_splice *sp = &req->splice;
3461 struct file *in = sp->file_in;
3462 struct file *out = sp->file_out;
3463 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3464 loff_t *poff_in, *poff_out;
3470 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3471 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3474 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3476 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3477 req->flags &= ~REQ_F_NEED_CLEANUP;
3480 req_set_fail_links(req);
3481 io_req_complete(req, ret);
3486 * IORING_OP_NOP just posts a completion event, nothing else.
3488 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3490 struct io_ring_ctx *ctx = req->ctx;
3492 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3495 __io_req_complete(req, 0, 0, cs);
3499 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3501 struct io_ring_ctx *ctx = req->ctx;
3506 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3508 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3511 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3512 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3515 req->sync.off = READ_ONCE(sqe->off);
3516 req->sync.len = READ_ONCE(sqe->len);
3520 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3522 loff_t end = req->sync.off + req->sync.len;
3525 /* fsync always requires a blocking context */
3529 ret = vfs_fsync_range(req->file, req->sync.off,
3530 end > 0 ? end : LLONG_MAX,
3531 req->sync.flags & IORING_FSYNC_DATASYNC);
3533 req_set_fail_links(req);
3534 io_req_complete(req, ret);
3538 static int io_fallocate_prep(struct io_kiocb *req,
3539 const struct io_uring_sqe *sqe)
3541 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3543 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3546 req->sync.off = READ_ONCE(sqe->off);
3547 req->sync.len = READ_ONCE(sqe->addr);
3548 req->sync.mode = READ_ONCE(sqe->len);
3552 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3556 /* fallocate always requiring blocking context */
3559 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3562 req_set_fail_links(req);
3563 io_req_complete(req, ret);
3567 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3569 const char __user *fname;
3572 if (unlikely(sqe->ioprio || sqe->buf_index))
3574 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3577 /* open.how should be already initialised */
3578 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3579 req->open.how.flags |= O_LARGEFILE;
3581 req->open.dfd = READ_ONCE(sqe->fd);
3582 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3583 req->open.filename = getname(fname);
3584 if (IS_ERR(req->open.filename)) {
3585 ret = PTR_ERR(req->open.filename);
3586 req->open.filename = NULL;
3589 req->open.nofile = rlimit(RLIMIT_NOFILE);
3590 req->flags |= REQ_F_NEED_CLEANUP;
3594 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3598 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3600 if (req->flags & REQ_F_NEED_CLEANUP)
3602 mode = READ_ONCE(sqe->len);
3603 flags = READ_ONCE(sqe->open_flags);
3604 req->open.how = build_open_how(flags, mode);
3605 return __io_openat_prep(req, sqe);
3608 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3610 struct open_how __user *how;
3614 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3616 if (req->flags & REQ_F_NEED_CLEANUP)
3618 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3619 len = READ_ONCE(sqe->len);
3620 if (len < OPEN_HOW_SIZE_VER0)
3623 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3628 return __io_openat_prep(req, sqe);
3631 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3633 struct open_flags op;
3640 ret = build_open_flags(&req->open.how, &op);
3644 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3648 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3651 ret = PTR_ERR(file);
3653 fsnotify_open(file);
3654 fd_install(ret, file);
3657 putname(req->open.filename);
3658 req->flags &= ~REQ_F_NEED_CLEANUP;
3660 req_set_fail_links(req);
3661 io_req_complete(req, ret);
3665 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3667 return io_openat2(req, force_nonblock);
3670 static int io_remove_buffers_prep(struct io_kiocb *req,
3671 const struct io_uring_sqe *sqe)
3673 struct io_provide_buf *p = &req->pbuf;
3676 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3679 tmp = READ_ONCE(sqe->fd);
3680 if (!tmp || tmp > USHRT_MAX)
3683 memset(p, 0, sizeof(*p));
3685 p->bgid = READ_ONCE(sqe->buf_group);
3689 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3690 int bgid, unsigned nbufs)
3694 /* shouldn't happen */
3698 /* the head kbuf is the list itself */
3699 while (!list_empty(&buf->list)) {
3700 struct io_buffer *nxt;
3702 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3703 list_del(&nxt->list);
3710 idr_remove(&ctx->io_buffer_idr, bgid);
3715 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3716 struct io_comp_state *cs)
3718 struct io_provide_buf *p = &req->pbuf;
3719 struct io_ring_ctx *ctx = req->ctx;
3720 struct io_buffer *head;
3723 io_ring_submit_lock(ctx, !force_nonblock);
3725 lockdep_assert_held(&ctx->uring_lock);
3728 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3730 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3732 io_ring_submit_lock(ctx, !force_nonblock);
3734 req_set_fail_links(req);
3735 __io_req_complete(req, ret, 0, cs);
3739 static int io_provide_buffers_prep(struct io_kiocb *req,
3740 const struct io_uring_sqe *sqe)
3742 struct io_provide_buf *p = &req->pbuf;
3745 if (sqe->ioprio || sqe->rw_flags)
3748 tmp = READ_ONCE(sqe->fd);
3749 if (!tmp || tmp > USHRT_MAX)
3752 p->addr = READ_ONCE(sqe->addr);
3753 p->len = READ_ONCE(sqe->len);
3755 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3758 p->bgid = READ_ONCE(sqe->buf_group);
3759 tmp = READ_ONCE(sqe->off);
3760 if (tmp > USHRT_MAX)
3766 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3768 struct io_buffer *buf;
3769 u64 addr = pbuf->addr;
3770 int i, bid = pbuf->bid;
3772 for (i = 0; i < pbuf->nbufs; i++) {
3773 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3778 buf->len = pbuf->len;
3783 INIT_LIST_HEAD(&buf->list);
3786 list_add_tail(&buf->list, &(*head)->list);
3790 return i ? i : -ENOMEM;
3793 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3794 struct io_comp_state *cs)
3796 struct io_provide_buf *p = &req->pbuf;
3797 struct io_ring_ctx *ctx = req->ctx;
3798 struct io_buffer *head, *list;
3801 io_ring_submit_lock(ctx, !force_nonblock);
3803 lockdep_assert_held(&ctx->uring_lock);
3805 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3807 ret = io_add_buffers(p, &head);
3812 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3815 __io_remove_buffers(ctx, head, p->bgid, -1U);
3820 io_ring_submit_unlock(ctx, !force_nonblock);
3822 req_set_fail_links(req);
3823 __io_req_complete(req, ret, 0, cs);
3827 static int io_epoll_ctl_prep(struct io_kiocb *req,
3828 const struct io_uring_sqe *sqe)
3830 #if defined(CONFIG_EPOLL)
3831 if (sqe->ioprio || sqe->buf_index)
3833 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3836 req->epoll.epfd = READ_ONCE(sqe->fd);
3837 req->epoll.op = READ_ONCE(sqe->len);
3838 req->epoll.fd = READ_ONCE(sqe->off);
3840 if (ep_op_has_event(req->epoll.op)) {
3841 struct epoll_event __user *ev;
3843 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3844 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3854 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3855 struct io_comp_state *cs)
3857 #if defined(CONFIG_EPOLL)
3858 struct io_epoll *ie = &req->epoll;
3861 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3862 if (force_nonblock && ret == -EAGAIN)
3866 req_set_fail_links(req);
3867 __io_req_complete(req, ret, 0, cs);
3874 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3876 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3877 if (sqe->ioprio || sqe->buf_index || sqe->off)
3879 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3882 req->madvise.addr = READ_ONCE(sqe->addr);
3883 req->madvise.len = READ_ONCE(sqe->len);
3884 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3891 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3893 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3894 struct io_madvise *ma = &req->madvise;
3900 ret = do_madvise(ma->addr, ma->len, ma->advice);
3902 req_set_fail_links(req);
3903 io_req_complete(req, ret);
3910 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3912 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3914 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3917 req->fadvise.offset = READ_ONCE(sqe->off);
3918 req->fadvise.len = READ_ONCE(sqe->len);
3919 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3923 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3925 struct io_fadvise *fa = &req->fadvise;
3928 if (force_nonblock) {
3929 switch (fa->advice) {
3930 case POSIX_FADV_NORMAL:
3931 case POSIX_FADV_RANDOM:
3932 case POSIX_FADV_SEQUENTIAL:
3939 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3941 req_set_fail_links(req);
3942 io_req_complete(req, ret);
3946 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3948 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3950 if (sqe->ioprio || sqe->buf_index)
3952 if (req->flags & REQ_F_FIXED_FILE)
3955 req->statx.dfd = READ_ONCE(sqe->fd);
3956 req->statx.mask = READ_ONCE(sqe->len);
3957 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3958 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3959 req->statx.flags = READ_ONCE(sqe->statx_flags);
3964 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3966 struct io_statx *ctx = &req->statx;
3969 if (force_nonblock) {
3970 /* only need file table for an actual valid fd */
3971 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3972 req->flags |= REQ_F_NO_FILE_TABLE;
3976 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3980 req_set_fail_links(req);
3981 io_req_complete(req, ret);
3985 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3988 * If we queue this for async, it must not be cancellable. That would
3989 * leave the 'file' in an undeterminate state, and here need to modify
3990 * io_wq_work.flags, so initialize io_wq_work firstly.
3992 io_req_init_async(req);
3993 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3995 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3997 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3998 sqe->rw_flags || sqe->buf_index)
4000 if (req->flags & REQ_F_FIXED_FILE)
4003 req->close.fd = READ_ONCE(sqe->fd);
4004 if ((req->file && req->file->f_op == &io_uring_fops))
4007 req->close.put_file = NULL;
4011 static int io_close(struct io_kiocb *req, bool force_nonblock,
4012 struct io_comp_state *cs)
4014 struct io_close *close = &req->close;
4017 /* might be already done during nonblock submission */
4018 if (!close->put_file) {
4019 ret = __close_fd_get_file(close->fd, &close->put_file);
4021 return (ret == -ENOENT) ? -EBADF : ret;
4024 /* if the file has a flush method, be safe and punt to async */
4025 if (close->put_file->f_op->flush && force_nonblock) {
4026 /* was never set, but play safe */
4027 req->flags &= ~REQ_F_NOWAIT;
4028 /* avoid grabbing files - we don't need the files */
4029 req->flags |= REQ_F_NO_FILE_TABLE;
4033 /* No ->flush() or already async, safely close from here */
4034 ret = filp_close(close->put_file, req->work.files);
4036 req_set_fail_links(req);
4037 fput(close->put_file);
4038 close->put_file = NULL;
4039 __io_req_complete(req, ret, 0, cs);
4043 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4045 struct io_ring_ctx *ctx = req->ctx;
4050 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4052 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4055 req->sync.off = READ_ONCE(sqe->off);
4056 req->sync.len = READ_ONCE(sqe->len);
4057 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4061 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4065 /* sync_file_range always requires a blocking context */
4069 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4072 req_set_fail_links(req);
4073 io_req_complete(req, ret);
4077 #if defined(CONFIG_NET)
4078 static int io_setup_async_msg(struct io_kiocb *req,
4079 struct io_async_msghdr *kmsg)
4083 if (io_alloc_async_ctx(req)) {
4084 if (kmsg->iov != kmsg->fast_iov)
4088 req->flags |= REQ_F_NEED_CLEANUP;
4089 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
4093 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4094 struct io_async_msghdr *iomsg)
4096 iomsg->iov = iomsg->fast_iov;
4097 iomsg->msg.msg_name = &iomsg->addr;
4098 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4099 req->sr_msg.msg_flags, &iomsg->iov);
4102 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4104 struct io_sr_msg *sr = &req->sr_msg;
4105 struct io_async_ctx *io = req->io;
4108 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4111 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4112 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4113 sr->len = READ_ONCE(sqe->len);
4115 #ifdef CONFIG_COMPAT
4116 if (req->ctx->compat)
4117 sr->msg_flags |= MSG_CMSG_COMPAT;
4120 if (!io || req->opcode == IORING_OP_SEND)
4122 /* iovec is already imported */
4123 if (req->flags & REQ_F_NEED_CLEANUP)
4126 ret = io_sendmsg_copy_hdr(req, &io->msg);
4128 req->flags |= REQ_F_NEED_CLEANUP;
4132 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4133 struct io_comp_state *cs)
4135 struct io_async_msghdr iomsg, *kmsg;
4136 struct socket *sock;
4140 sock = sock_from_file(req->file, &ret);
4141 if (unlikely(!sock))
4145 kmsg = &req->io->msg;
4146 kmsg->msg.msg_name = &req->io->msg.addr;
4147 /* if iov is set, it's allocated already */
4149 kmsg->iov = kmsg->fast_iov;
4150 kmsg->msg.msg_iter.iov = kmsg->iov;
4152 ret = io_sendmsg_copy_hdr(req, &iomsg);
4158 flags = req->sr_msg.msg_flags;
4159 if (flags & MSG_DONTWAIT)
4160 req->flags |= REQ_F_NOWAIT;
4161 else if (force_nonblock)
4162 flags |= MSG_DONTWAIT;
4164 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4165 if (force_nonblock && ret == -EAGAIN)
4166 return io_setup_async_msg(req, kmsg);
4167 if (ret == -ERESTARTSYS)
4170 if (kmsg->iov != kmsg->fast_iov)
4172 req->flags &= ~REQ_F_NEED_CLEANUP;
4174 req_set_fail_links(req);
4175 __io_req_complete(req, ret, 0, cs);
4179 static int io_send(struct io_kiocb *req, bool force_nonblock,
4180 struct io_comp_state *cs)
4182 struct io_sr_msg *sr = &req->sr_msg;
4185 struct socket *sock;
4189 sock = sock_from_file(req->file, &ret);
4190 if (unlikely(!sock))
4193 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4197 msg.msg_name = NULL;
4198 msg.msg_control = NULL;
4199 msg.msg_controllen = 0;
4200 msg.msg_namelen = 0;
4202 flags = req->sr_msg.msg_flags;
4203 if (flags & MSG_DONTWAIT)
4204 req->flags |= REQ_F_NOWAIT;
4205 else if (force_nonblock)
4206 flags |= MSG_DONTWAIT;
4208 msg.msg_flags = flags;
4209 ret = sock_sendmsg(sock, &msg);
4210 if (force_nonblock && ret == -EAGAIN)
4212 if (ret == -ERESTARTSYS)
4216 req_set_fail_links(req);
4217 __io_req_complete(req, ret, 0, cs);
4221 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4222 struct io_async_msghdr *iomsg)
4224 struct io_sr_msg *sr = &req->sr_msg;
4225 struct iovec __user *uiov;
4229 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4230 &iomsg->uaddr, &uiov, &iov_len);
4234 if (req->flags & REQ_F_BUFFER_SELECT) {
4237 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4239 sr->len = iomsg->iov[0].iov_len;
4240 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4244 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4245 &iomsg->iov, &iomsg->msg.msg_iter);
4253 #ifdef CONFIG_COMPAT
4254 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4255 struct io_async_msghdr *iomsg)
4257 struct compat_msghdr __user *msg_compat;
4258 struct io_sr_msg *sr = &req->sr_msg;
4259 struct compat_iovec __user *uiov;
4264 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4265 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4270 uiov = compat_ptr(ptr);
4271 if (req->flags & REQ_F_BUFFER_SELECT) {
4272 compat_ssize_t clen;
4276 if (!access_ok(uiov, sizeof(*uiov)))
4278 if (__get_user(clen, &uiov->iov_len))
4282 sr->len = iomsg->iov[0].iov_len;
4285 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4287 &iomsg->msg.msg_iter);
4296 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4297 struct io_async_msghdr *iomsg)
4299 iomsg->msg.msg_name = &iomsg->addr;
4300 iomsg->iov = iomsg->fast_iov;
4302 #ifdef CONFIG_COMPAT
4303 if (req->ctx->compat)
4304 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4307 return __io_recvmsg_copy_hdr(req, iomsg);
4310 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4313 struct io_sr_msg *sr = &req->sr_msg;
4314 struct io_buffer *kbuf;
4316 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4321 req->flags |= REQ_F_BUFFER_SELECTED;
4325 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4327 return io_put_kbuf(req, req->sr_msg.kbuf);
4330 static int io_recvmsg_prep(struct io_kiocb *req,
4331 const struct io_uring_sqe *sqe)
4333 struct io_sr_msg *sr = &req->sr_msg;
4334 struct io_async_ctx *io = req->io;
4337 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4340 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4341 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4342 sr->len = READ_ONCE(sqe->len);
4343 sr->bgid = READ_ONCE(sqe->buf_group);
4345 #ifdef CONFIG_COMPAT
4346 if (req->ctx->compat)
4347 sr->msg_flags |= MSG_CMSG_COMPAT;
4350 if (!io || req->opcode == IORING_OP_RECV)
4352 /* iovec is already imported */
4353 if (req->flags & REQ_F_NEED_CLEANUP)
4356 ret = io_recvmsg_copy_hdr(req, &io->msg);
4358 req->flags |= REQ_F_NEED_CLEANUP;
4362 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4363 struct io_comp_state *cs)
4365 struct io_async_msghdr iomsg, *kmsg;
4366 struct socket *sock;
4367 struct io_buffer *kbuf;
4369 int ret, cflags = 0;
4371 sock = sock_from_file(req->file, &ret);
4372 if (unlikely(!sock))
4376 kmsg = &req->io->msg;
4377 kmsg->msg.msg_name = &req->io->msg.addr;
4378 /* if iov is set, it's allocated already */
4380 kmsg->iov = kmsg->fast_iov;
4381 kmsg->msg.msg_iter.iov = kmsg->iov;
4383 ret = io_recvmsg_copy_hdr(req, &iomsg);
4389 if (req->flags & REQ_F_BUFFER_SELECT) {
4390 kbuf = io_recv_buffer_select(req, !force_nonblock);
4392 return PTR_ERR(kbuf);
4393 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4394 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4395 1, req->sr_msg.len);
4398 flags = req->sr_msg.msg_flags;
4399 if (flags & MSG_DONTWAIT)
4400 req->flags |= REQ_F_NOWAIT;
4401 else if (force_nonblock)
4402 flags |= MSG_DONTWAIT;
4404 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4405 kmsg->uaddr, flags);
4406 if (force_nonblock && ret == -EAGAIN)
4407 return io_setup_async_msg(req, kmsg);
4408 if (ret == -ERESTARTSYS)
4411 if (req->flags & REQ_F_BUFFER_SELECTED)
4412 cflags = io_put_recv_kbuf(req);
4413 if (kmsg->iov != kmsg->fast_iov)
4415 req->flags &= ~REQ_F_NEED_CLEANUP;
4417 req_set_fail_links(req);
4418 __io_req_complete(req, ret, cflags, cs);
4422 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4423 struct io_comp_state *cs)
4425 struct io_buffer *kbuf;
4426 struct io_sr_msg *sr = &req->sr_msg;
4428 void __user *buf = sr->buf;
4429 struct socket *sock;
4432 int ret, cflags = 0;
4434 sock = sock_from_file(req->file, &ret);
4435 if (unlikely(!sock))
4438 if (req->flags & REQ_F_BUFFER_SELECT) {
4439 kbuf = io_recv_buffer_select(req, !force_nonblock);
4441 return PTR_ERR(kbuf);
4442 buf = u64_to_user_ptr(kbuf->addr);
4445 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4449 msg.msg_name = NULL;
4450 msg.msg_control = NULL;
4451 msg.msg_controllen = 0;
4452 msg.msg_namelen = 0;
4453 msg.msg_iocb = NULL;
4456 flags = req->sr_msg.msg_flags;
4457 if (flags & MSG_DONTWAIT)
4458 req->flags |= REQ_F_NOWAIT;
4459 else if (force_nonblock)
4460 flags |= MSG_DONTWAIT;
4462 ret = sock_recvmsg(sock, &msg, flags);
4463 if (force_nonblock && ret == -EAGAIN)
4465 if (ret == -ERESTARTSYS)
4468 if (req->flags & REQ_F_BUFFER_SELECTED)
4469 cflags = io_put_recv_kbuf(req);
4471 req_set_fail_links(req);
4472 __io_req_complete(req, ret, cflags, cs);
4476 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4478 struct io_accept *accept = &req->accept;
4480 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4482 if (sqe->ioprio || sqe->len || sqe->buf_index)
4485 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4486 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4487 accept->flags = READ_ONCE(sqe->accept_flags);
4488 accept->nofile = rlimit(RLIMIT_NOFILE);
4492 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4493 struct io_comp_state *cs)
4495 struct io_accept *accept = &req->accept;
4496 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4499 if (req->file->f_flags & O_NONBLOCK)
4500 req->flags |= REQ_F_NOWAIT;
4502 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4503 accept->addr_len, accept->flags,
4505 if (ret == -EAGAIN && force_nonblock)
4508 if (ret == -ERESTARTSYS)
4510 req_set_fail_links(req);
4512 __io_req_complete(req, ret, 0, cs);
4516 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4518 struct io_connect *conn = &req->connect;
4519 struct io_async_ctx *io = req->io;
4521 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4523 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4526 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4527 conn->addr_len = READ_ONCE(sqe->addr2);
4532 return move_addr_to_kernel(conn->addr, conn->addr_len,
4533 &io->connect.address);
4536 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4537 struct io_comp_state *cs)
4539 struct io_async_ctx __io, *io;
4540 unsigned file_flags;
4546 ret = move_addr_to_kernel(req->connect.addr,
4547 req->connect.addr_len,
4548 &__io.connect.address);
4554 file_flags = force_nonblock ? O_NONBLOCK : 0;
4556 ret = __sys_connect_file(req->file, &io->connect.address,
4557 req->connect.addr_len, file_flags);
4558 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4561 if (io_alloc_async_ctx(req)) {
4565 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4568 if (ret == -ERESTARTSYS)
4572 req_set_fail_links(req);
4573 __io_req_complete(req, ret, 0, cs);
4576 #else /* !CONFIG_NET */
4577 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4582 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4583 struct io_comp_state *cs)
4588 static int io_send(struct io_kiocb *req, bool force_nonblock,
4589 struct io_comp_state *cs)
4594 static int io_recvmsg_prep(struct io_kiocb *req,
4595 const struct io_uring_sqe *sqe)
4600 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4601 struct io_comp_state *cs)
4606 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4607 struct io_comp_state *cs)
4612 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4617 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4618 struct io_comp_state *cs)
4623 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4628 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4629 struct io_comp_state *cs)
4633 #endif /* CONFIG_NET */
4635 struct io_poll_table {
4636 struct poll_table_struct pt;
4637 struct io_kiocb *req;
4641 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4642 __poll_t mask, task_work_func_t func)
4647 /* for instances that support it check for an event match first: */
4648 if (mask && !(mask & poll->events))
4651 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4653 list_del_init(&poll->wait.entry);
4656 init_task_work(&req->task_work, func);
4657 percpu_ref_get(&req->ctx->refs);
4660 * If we using the signalfd wait_queue_head for this wakeup, then
4661 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4662 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4663 * either, as the normal wakeup will suffice.
4665 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4668 * If this fails, then the task is exiting. When a task exits, the
4669 * work gets canceled, so just cancel this request as well instead
4670 * of executing it. We can't safely execute it anyway, as we may not
4671 * have the needed state needed for it anyway.
4673 ret = io_req_task_work_add(req, &req->task_work, twa_signal_ok);
4674 if (unlikely(ret)) {
4675 struct task_struct *tsk;
4677 WRITE_ONCE(poll->canceled, true);
4678 tsk = io_wq_get_task(req->ctx->io_wq);
4679 task_work_add(tsk, &req->task_work, 0);
4680 wake_up_process(tsk);
4685 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4686 __acquires(&req->ctx->completion_lock)
4688 struct io_ring_ctx *ctx = req->ctx;
4690 if (!req->result && !READ_ONCE(poll->canceled)) {
4691 struct poll_table_struct pt = { ._key = poll->events };
4693 req->result = vfs_poll(req->file, &pt) & poll->events;
4696 spin_lock_irq(&ctx->completion_lock);
4697 if (!req->result && !READ_ONCE(poll->canceled)) {
4698 add_wait_queue(poll->head, &poll->wait);
4705 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4707 /* pure poll stashes this in ->io, poll driven retry elsewhere */
4708 if (req->opcode == IORING_OP_POLL_ADD)
4709 return (struct io_poll_iocb *) req->io;
4710 return req->apoll->double_poll;
4713 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4715 if (req->opcode == IORING_OP_POLL_ADD)
4717 return &req->apoll->poll;
4720 static void io_poll_remove_double(struct io_kiocb *req)
4722 struct io_poll_iocb *poll = io_poll_get_double(req);
4724 lockdep_assert_held(&req->ctx->completion_lock);
4726 if (poll && poll->head) {
4727 struct wait_queue_head *head = poll->head;
4729 spin_lock(&head->lock);
4730 list_del_init(&poll->wait.entry);
4731 if (poll->wait.private)
4732 refcount_dec(&req->refs);
4734 spin_unlock(&head->lock);
4738 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4740 struct io_ring_ctx *ctx = req->ctx;
4742 io_poll_remove_double(req);
4743 req->poll.done = true;
4744 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4745 io_commit_cqring(ctx);
4748 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4750 struct io_ring_ctx *ctx = req->ctx;
4752 if (io_poll_rewait(req, &req->poll)) {
4753 spin_unlock_irq(&ctx->completion_lock);
4757 hash_del(&req->hash_node);
4758 io_poll_complete(req, req->result, 0);
4759 req->flags |= REQ_F_COMP_LOCKED;
4760 *nxt = io_put_req_find_next(req);
4761 spin_unlock_irq(&ctx->completion_lock);
4763 io_cqring_ev_posted(ctx);
4766 static void io_poll_task_func(struct callback_head *cb)
4768 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4769 struct io_ring_ctx *ctx = req->ctx;
4770 struct io_kiocb *nxt = NULL;
4772 io_poll_task_handler(req, &nxt);
4774 __io_req_task_submit(nxt);
4775 percpu_ref_put(&ctx->refs);
4778 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4779 int sync, void *key)
4781 struct io_kiocb *req = wait->private;
4782 struct io_poll_iocb *poll = io_poll_get_single(req);
4783 __poll_t mask = key_to_poll(key);
4785 /* for instances that support it check for an event match first: */
4786 if (mask && !(mask & poll->events))
4789 list_del_init(&wait->entry);
4791 if (poll && poll->head) {
4794 spin_lock(&poll->head->lock);
4795 done = list_empty(&poll->wait.entry);
4797 list_del_init(&poll->wait.entry);
4798 /* make sure double remove sees this as being gone */
4799 wait->private = NULL;
4800 spin_unlock(&poll->head->lock);
4802 __io_async_wake(req, poll, mask, io_poll_task_func);
4804 refcount_dec(&req->refs);
4808 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4809 wait_queue_func_t wake_func)
4813 poll->canceled = false;
4814 poll->events = events;
4815 INIT_LIST_HEAD(&poll->wait.entry);
4816 init_waitqueue_func_entry(&poll->wait, wake_func);
4819 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4820 struct wait_queue_head *head,
4821 struct io_poll_iocb **poll_ptr)
4823 struct io_kiocb *req = pt->req;
4826 * If poll->head is already set, it's because the file being polled
4827 * uses multiple waitqueues for poll handling (eg one for read, one
4828 * for write). Setup a separate io_poll_iocb if this happens.
4830 if (unlikely(poll->head)) {
4831 /* already have a 2nd entry, fail a third attempt */
4833 pt->error = -EINVAL;
4836 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4838 pt->error = -ENOMEM;
4841 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4842 refcount_inc(&req->refs);
4843 poll->wait.private = req;
4850 if (poll->events & EPOLLEXCLUSIVE)
4851 add_wait_queue_exclusive(head, &poll->wait);
4853 add_wait_queue(head, &poll->wait);
4856 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4857 struct poll_table_struct *p)
4859 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4860 struct async_poll *apoll = pt->req->apoll;
4862 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4865 static void io_async_task_func(struct callback_head *cb)
4867 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4868 struct async_poll *apoll = req->apoll;
4869 struct io_ring_ctx *ctx = req->ctx;
4871 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4873 if (io_poll_rewait(req, &apoll->poll)) {
4874 spin_unlock_irq(&ctx->completion_lock);
4875 percpu_ref_put(&ctx->refs);
4879 /* If req is still hashed, it cannot have been canceled. Don't check. */
4880 if (hash_hashed(&req->hash_node))
4881 hash_del(&req->hash_node);
4883 io_poll_remove_double(req);
4884 spin_unlock_irq(&ctx->completion_lock);
4886 if (!READ_ONCE(apoll->poll.canceled))
4887 __io_req_task_submit(req);
4889 __io_req_task_cancel(req, -ECANCELED);
4891 percpu_ref_put(&ctx->refs);
4892 kfree(apoll->double_poll);
4896 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4899 struct io_kiocb *req = wait->private;
4900 struct io_poll_iocb *poll = &req->apoll->poll;
4902 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4905 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4908 static void io_poll_req_insert(struct io_kiocb *req)
4910 struct io_ring_ctx *ctx = req->ctx;
4911 struct hlist_head *list;
4913 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4914 hlist_add_head(&req->hash_node, list);
4917 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4918 struct io_poll_iocb *poll,
4919 struct io_poll_table *ipt, __poll_t mask,
4920 wait_queue_func_t wake_func)
4921 __acquires(&ctx->completion_lock)
4923 struct io_ring_ctx *ctx = req->ctx;
4924 bool cancel = false;
4926 io_init_poll_iocb(poll, mask, wake_func);
4927 poll->file = req->file;
4928 poll->wait.private = req;
4930 ipt->pt._key = mask;
4932 ipt->error = -EINVAL;
4934 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4936 spin_lock_irq(&ctx->completion_lock);
4937 if (likely(poll->head)) {
4938 spin_lock(&poll->head->lock);
4939 if (unlikely(list_empty(&poll->wait.entry))) {
4945 if (mask || ipt->error)
4946 list_del_init(&poll->wait.entry);
4948 WRITE_ONCE(poll->canceled, true);
4949 else if (!poll->done) /* actually waiting for an event */
4950 io_poll_req_insert(req);
4951 spin_unlock(&poll->head->lock);
4957 static bool io_arm_poll_handler(struct io_kiocb *req)
4959 const struct io_op_def *def = &io_op_defs[req->opcode];
4960 struct io_ring_ctx *ctx = req->ctx;
4961 struct async_poll *apoll;
4962 struct io_poll_table ipt;
4966 if (!req->file || !file_can_poll(req->file))
4968 if (req->flags & REQ_F_POLLED)
4972 else if (def->pollout)
4976 /* if we can't nonblock try, then no point in arming a poll handler */
4977 if (!io_file_supports_async(req->file, rw))
4980 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4981 if (unlikely(!apoll))
4983 apoll->double_poll = NULL;
4985 req->flags |= REQ_F_POLLED;
4987 INIT_HLIST_NODE(&req->hash_node);
4991 mask |= POLLIN | POLLRDNORM;
4993 mask |= POLLOUT | POLLWRNORM;
4994 mask |= POLLERR | POLLPRI;
4996 ipt.pt._qproc = io_async_queue_proc;
4998 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5000 if (ret || ipt.error) {
5001 io_poll_remove_double(req);
5002 spin_unlock_irq(&ctx->completion_lock);
5003 kfree(apoll->double_poll);
5007 spin_unlock_irq(&ctx->completion_lock);
5008 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5009 apoll->poll.events);
5013 static bool __io_poll_remove_one(struct io_kiocb *req,
5014 struct io_poll_iocb *poll)
5016 bool do_complete = false;
5018 spin_lock(&poll->head->lock);
5019 WRITE_ONCE(poll->canceled, true);
5020 if (!list_empty(&poll->wait.entry)) {
5021 list_del_init(&poll->wait.entry);
5024 spin_unlock(&poll->head->lock);
5025 hash_del(&req->hash_node);
5029 static bool io_poll_remove_one(struct io_kiocb *req)
5033 io_poll_remove_double(req);
5035 if (req->opcode == IORING_OP_POLL_ADD) {
5036 do_complete = __io_poll_remove_one(req, &req->poll);
5038 struct async_poll *apoll = req->apoll;
5040 /* non-poll requests have submit ref still */
5041 do_complete = __io_poll_remove_one(req, &apoll->poll);
5044 kfree(apoll->double_poll);
5050 io_cqring_fill_event(req, -ECANCELED);
5051 io_commit_cqring(req->ctx);
5052 req->flags |= REQ_F_COMP_LOCKED;
5053 req_set_fail_links(req);
5061 * Returns true if we found and killed one or more poll requests
5063 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5065 struct hlist_node *tmp;
5066 struct io_kiocb *req;
5069 spin_lock_irq(&ctx->completion_lock);
5070 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5071 struct hlist_head *list;
5073 list = &ctx->cancel_hash[i];
5074 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5075 if (io_task_match(req, tsk))
5076 posted += io_poll_remove_one(req);
5079 spin_unlock_irq(&ctx->completion_lock);
5082 io_cqring_ev_posted(ctx);
5087 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5089 struct hlist_head *list;
5090 struct io_kiocb *req;
5092 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5093 hlist_for_each_entry(req, list, hash_node) {
5094 if (sqe_addr != req->user_data)
5096 if (io_poll_remove_one(req))
5104 static int io_poll_remove_prep(struct io_kiocb *req,
5105 const struct io_uring_sqe *sqe)
5107 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5109 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5113 req->poll.addr = READ_ONCE(sqe->addr);
5118 * Find a running poll command that matches one specified in sqe->addr,
5119 * and remove it if found.
5121 static int io_poll_remove(struct io_kiocb *req)
5123 struct io_ring_ctx *ctx = req->ctx;
5127 addr = req->poll.addr;
5128 spin_lock_irq(&ctx->completion_lock);
5129 ret = io_poll_cancel(ctx, addr);
5130 spin_unlock_irq(&ctx->completion_lock);
5133 req_set_fail_links(req);
5134 io_req_complete(req, ret);
5138 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5141 struct io_kiocb *req = wait->private;
5142 struct io_poll_iocb *poll = &req->poll;
5144 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5147 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5148 struct poll_table_struct *p)
5150 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5152 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
5155 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5157 struct io_poll_iocb *poll = &req->poll;
5160 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5162 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5167 events = READ_ONCE(sqe->poll32_events);
5169 events = swahw32(events);
5171 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5172 (events & EPOLLEXCLUSIVE);
5176 static int io_poll_add(struct io_kiocb *req)
5178 struct io_poll_iocb *poll = &req->poll;
5179 struct io_ring_ctx *ctx = req->ctx;
5180 struct io_poll_table ipt;
5183 INIT_HLIST_NODE(&req->hash_node);
5184 ipt.pt._qproc = io_poll_queue_proc;
5186 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5189 if (mask) { /* no async, we'd stolen it */
5191 io_poll_complete(req, mask, 0);
5193 spin_unlock_irq(&ctx->completion_lock);
5196 io_cqring_ev_posted(ctx);
5202 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5204 struct io_timeout_data *data = container_of(timer,
5205 struct io_timeout_data, timer);
5206 struct io_kiocb *req = data->req;
5207 struct io_ring_ctx *ctx = req->ctx;
5208 unsigned long flags;
5210 spin_lock_irqsave(&ctx->completion_lock, flags);
5211 atomic_set(&req->ctx->cq_timeouts,
5212 atomic_read(&req->ctx->cq_timeouts) + 1);
5215 * We could be racing with timeout deletion. If the list is empty,
5216 * then timeout lookup already found it and will be handling it.
5218 if (!list_empty(&req->timeout.list))
5219 list_del_init(&req->timeout.list);
5221 io_cqring_fill_event(req, -ETIME);
5222 io_commit_cqring(ctx);
5223 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5225 io_cqring_ev_posted(ctx);
5226 req_set_fail_links(req);
5228 return HRTIMER_NORESTART;
5231 static int __io_timeout_cancel(struct io_kiocb *req)
5235 list_del_init(&req->timeout.list);
5237 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5241 req_set_fail_links(req);
5242 req->flags |= REQ_F_COMP_LOCKED;
5243 io_cqring_fill_event(req, -ECANCELED);
5248 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5250 struct io_kiocb *req;
5253 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5254 if (user_data == req->user_data) {
5263 return __io_timeout_cancel(req);
5266 static int io_timeout_remove_prep(struct io_kiocb *req,
5267 const struct io_uring_sqe *sqe)
5269 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5271 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5273 if (sqe->ioprio || sqe->buf_index || sqe->len)
5276 req->timeout.addr = READ_ONCE(sqe->addr);
5277 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5278 if (req->timeout.flags)
5285 * Remove or update an existing timeout command
5287 static int io_timeout_remove(struct io_kiocb *req)
5289 struct io_ring_ctx *ctx = req->ctx;
5292 spin_lock_irq(&ctx->completion_lock);
5293 ret = io_timeout_cancel(ctx, req->timeout.addr);
5295 io_cqring_fill_event(req, ret);
5296 io_commit_cqring(ctx);
5297 spin_unlock_irq(&ctx->completion_lock);
5298 io_cqring_ev_posted(ctx);
5300 req_set_fail_links(req);
5305 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5306 bool is_timeout_link)
5308 struct io_timeout_data *data;
5310 u32 off = READ_ONCE(sqe->off);
5312 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5314 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5316 if (off && is_timeout_link)
5318 flags = READ_ONCE(sqe->timeout_flags);
5319 if (flags & ~IORING_TIMEOUT_ABS)
5322 req->timeout.off = off;
5324 if (!req->io && io_alloc_async_ctx(req))
5327 data = &req->io->timeout;
5330 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5333 if (flags & IORING_TIMEOUT_ABS)
5334 data->mode = HRTIMER_MODE_ABS;
5336 data->mode = HRTIMER_MODE_REL;
5338 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5342 static int io_timeout(struct io_kiocb *req)
5344 struct io_ring_ctx *ctx = req->ctx;
5345 struct io_timeout_data *data = &req->io->timeout;
5346 struct list_head *entry;
5347 u32 tail, off = req->timeout.off;
5349 spin_lock_irq(&ctx->completion_lock);
5352 * sqe->off holds how many events that need to occur for this
5353 * timeout event to be satisfied. If it isn't set, then this is
5354 * a pure timeout request, sequence isn't used.
5356 if (io_is_timeout_noseq(req)) {
5357 entry = ctx->timeout_list.prev;
5361 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5362 req->timeout.target_seq = tail + off;
5365 * Insertion sort, ensuring the first entry in the list is always
5366 * the one we need first.
5368 list_for_each_prev(entry, &ctx->timeout_list) {
5369 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5372 if (io_is_timeout_noseq(nxt))
5374 /* nxt.seq is behind @tail, otherwise would've been completed */
5375 if (off >= nxt->timeout.target_seq - tail)
5379 list_add(&req->timeout.list, entry);
5380 data->timer.function = io_timeout_fn;
5381 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5382 spin_unlock_irq(&ctx->completion_lock);
5386 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5388 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5390 return req->user_data == (unsigned long) data;
5393 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5395 enum io_wq_cancel cancel_ret;
5398 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5399 switch (cancel_ret) {
5400 case IO_WQ_CANCEL_OK:
5403 case IO_WQ_CANCEL_RUNNING:
5406 case IO_WQ_CANCEL_NOTFOUND:
5414 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5415 struct io_kiocb *req, __u64 sqe_addr,
5418 unsigned long flags;
5421 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5422 if (ret != -ENOENT) {
5423 spin_lock_irqsave(&ctx->completion_lock, flags);
5427 spin_lock_irqsave(&ctx->completion_lock, flags);
5428 ret = io_timeout_cancel(ctx, sqe_addr);
5431 ret = io_poll_cancel(ctx, sqe_addr);
5435 io_cqring_fill_event(req, ret);
5436 io_commit_cqring(ctx);
5437 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5438 io_cqring_ev_posted(ctx);
5441 req_set_fail_links(req);
5445 static int io_async_cancel_prep(struct io_kiocb *req,
5446 const struct io_uring_sqe *sqe)
5448 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5450 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5452 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5455 req->cancel.addr = READ_ONCE(sqe->addr);
5459 static int io_async_cancel(struct io_kiocb *req)
5461 struct io_ring_ctx *ctx = req->ctx;
5463 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5467 static int io_files_update_prep(struct io_kiocb *req,
5468 const struct io_uring_sqe *sqe)
5470 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5472 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5474 if (sqe->ioprio || sqe->rw_flags)
5477 req->files_update.offset = READ_ONCE(sqe->off);
5478 req->files_update.nr_args = READ_ONCE(sqe->len);
5479 if (!req->files_update.nr_args)
5481 req->files_update.arg = READ_ONCE(sqe->addr);
5485 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5486 struct io_comp_state *cs)
5488 struct io_ring_ctx *ctx = req->ctx;
5489 struct io_uring_files_update up;
5495 up.offset = req->files_update.offset;
5496 up.fds = req->files_update.arg;
5498 mutex_lock(&ctx->uring_lock);
5499 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5500 mutex_unlock(&ctx->uring_lock);
5503 req_set_fail_links(req);
5504 __io_req_complete(req, ret, 0, cs);
5508 static int io_req_defer_prep(struct io_kiocb *req,
5509 const struct io_uring_sqe *sqe)
5516 if (io_alloc_async_ctx(req))
5518 ret = io_prep_work_files(req);
5522 io_prep_async_work(req);
5524 switch (req->opcode) {
5527 case IORING_OP_READV:
5528 case IORING_OP_READ_FIXED:
5529 case IORING_OP_READ:
5530 ret = io_read_prep(req, sqe, true);
5532 case IORING_OP_WRITEV:
5533 case IORING_OP_WRITE_FIXED:
5534 case IORING_OP_WRITE:
5535 ret = io_write_prep(req, sqe, true);
5537 case IORING_OP_POLL_ADD:
5538 ret = io_poll_add_prep(req, sqe);
5540 case IORING_OP_POLL_REMOVE:
5541 ret = io_poll_remove_prep(req, sqe);
5543 case IORING_OP_FSYNC:
5544 ret = io_prep_fsync(req, sqe);
5546 case IORING_OP_SYNC_FILE_RANGE:
5547 ret = io_prep_sfr(req, sqe);
5549 case IORING_OP_SENDMSG:
5550 case IORING_OP_SEND:
5551 ret = io_sendmsg_prep(req, sqe);
5553 case IORING_OP_RECVMSG:
5554 case IORING_OP_RECV:
5555 ret = io_recvmsg_prep(req, sqe);
5557 case IORING_OP_CONNECT:
5558 ret = io_connect_prep(req, sqe);
5560 case IORING_OP_TIMEOUT:
5561 ret = io_timeout_prep(req, sqe, false);
5563 case IORING_OP_TIMEOUT_REMOVE:
5564 ret = io_timeout_remove_prep(req, sqe);
5566 case IORING_OP_ASYNC_CANCEL:
5567 ret = io_async_cancel_prep(req, sqe);
5569 case IORING_OP_LINK_TIMEOUT:
5570 ret = io_timeout_prep(req, sqe, true);
5572 case IORING_OP_ACCEPT:
5573 ret = io_accept_prep(req, sqe);
5575 case IORING_OP_FALLOCATE:
5576 ret = io_fallocate_prep(req, sqe);
5578 case IORING_OP_OPENAT:
5579 ret = io_openat_prep(req, sqe);
5581 case IORING_OP_CLOSE:
5582 ret = io_close_prep(req, sqe);
5584 case IORING_OP_FILES_UPDATE:
5585 ret = io_files_update_prep(req, sqe);
5587 case IORING_OP_STATX:
5588 ret = io_statx_prep(req, sqe);
5590 case IORING_OP_FADVISE:
5591 ret = io_fadvise_prep(req, sqe);
5593 case IORING_OP_MADVISE:
5594 ret = io_madvise_prep(req, sqe);
5596 case IORING_OP_OPENAT2:
5597 ret = io_openat2_prep(req, sqe);
5599 case IORING_OP_EPOLL_CTL:
5600 ret = io_epoll_ctl_prep(req, sqe);
5602 case IORING_OP_SPLICE:
5603 ret = io_splice_prep(req, sqe);
5605 case IORING_OP_PROVIDE_BUFFERS:
5606 ret = io_provide_buffers_prep(req, sqe);
5608 case IORING_OP_REMOVE_BUFFERS:
5609 ret = io_remove_buffers_prep(req, sqe);
5612 ret = io_tee_prep(req, sqe);
5615 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5624 static u32 io_get_sequence(struct io_kiocb *req)
5626 struct io_kiocb *pos;
5627 struct io_ring_ctx *ctx = req->ctx;
5628 u32 total_submitted, nr_reqs = 1;
5630 if (req->flags & REQ_F_LINK_HEAD)
5631 list_for_each_entry(pos, &req->link_list, link_list)
5634 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5635 return total_submitted - nr_reqs;
5638 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5640 struct io_ring_ctx *ctx = req->ctx;
5641 struct io_defer_entry *de;
5645 /* Still need defer if there is pending req in defer list. */
5646 if (likely(list_empty_careful(&ctx->defer_list) &&
5647 !(req->flags & REQ_F_IO_DRAIN)))
5650 seq = io_get_sequence(req);
5651 /* Still a chance to pass the sequence check */
5652 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5656 ret = io_req_defer_prep(req, sqe);
5660 io_prep_async_link(req);
5661 de = kmalloc(sizeof(*de), GFP_KERNEL);
5665 spin_lock_irq(&ctx->completion_lock);
5666 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5667 spin_unlock_irq(&ctx->completion_lock);
5669 io_queue_async_work(req);
5670 return -EIOCBQUEUED;
5673 trace_io_uring_defer(ctx, req, req->user_data);
5676 list_add_tail(&de->list, &ctx->defer_list);
5677 spin_unlock_irq(&ctx->completion_lock);
5678 return -EIOCBQUEUED;
5681 static void io_req_drop_files(struct io_kiocb *req)
5683 struct io_ring_ctx *ctx = req->ctx;
5684 unsigned long flags;
5686 spin_lock_irqsave(&ctx->inflight_lock, flags);
5687 list_del(&req->inflight_entry);
5688 if (waitqueue_active(&ctx->inflight_wait))
5689 wake_up(&ctx->inflight_wait);
5690 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5691 req->flags &= ~REQ_F_INFLIGHT;
5692 put_files_struct(req->work.files);
5693 put_nsproxy(req->work.nsproxy);
5694 req->work.files = NULL;
5697 static void __io_clean_op(struct io_kiocb *req)
5699 struct io_async_ctx *io = req->io;
5701 if (req->flags & REQ_F_BUFFER_SELECTED) {
5702 switch (req->opcode) {
5703 case IORING_OP_READV:
5704 case IORING_OP_READ_FIXED:
5705 case IORING_OP_READ:
5706 kfree((void *)(unsigned long)req->rw.addr);
5708 case IORING_OP_RECVMSG:
5709 case IORING_OP_RECV:
5710 kfree(req->sr_msg.kbuf);
5713 req->flags &= ~REQ_F_BUFFER_SELECTED;
5716 if (req->flags & REQ_F_NEED_CLEANUP) {
5717 switch (req->opcode) {
5718 case IORING_OP_READV:
5719 case IORING_OP_READ_FIXED:
5720 case IORING_OP_READ:
5721 case IORING_OP_WRITEV:
5722 case IORING_OP_WRITE_FIXED:
5723 case IORING_OP_WRITE:
5724 if (io->rw.free_iovec)
5725 kfree(io->rw.free_iovec);
5727 case IORING_OP_RECVMSG:
5728 case IORING_OP_SENDMSG:
5729 if (io->msg.iov != io->msg.fast_iov)
5732 case IORING_OP_SPLICE:
5734 io_put_file(req, req->splice.file_in,
5735 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5737 case IORING_OP_OPENAT:
5738 case IORING_OP_OPENAT2:
5739 if (req->open.filename)
5740 putname(req->open.filename);
5743 req->flags &= ~REQ_F_NEED_CLEANUP;
5746 if (req->flags & REQ_F_INFLIGHT)
5747 io_req_drop_files(req);
5750 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5751 bool force_nonblock, struct io_comp_state *cs)
5753 struct io_ring_ctx *ctx = req->ctx;
5756 switch (req->opcode) {
5758 ret = io_nop(req, cs);
5760 case IORING_OP_READV:
5761 case IORING_OP_READ_FIXED:
5762 case IORING_OP_READ:
5764 ret = io_read_prep(req, sqe, force_nonblock);
5768 ret = io_read(req, force_nonblock, cs);
5770 case IORING_OP_WRITEV:
5771 case IORING_OP_WRITE_FIXED:
5772 case IORING_OP_WRITE:
5774 ret = io_write_prep(req, sqe, force_nonblock);
5778 ret = io_write(req, force_nonblock, cs);
5780 case IORING_OP_FSYNC:
5782 ret = io_prep_fsync(req, sqe);
5786 ret = io_fsync(req, force_nonblock);
5788 case IORING_OP_POLL_ADD:
5790 ret = io_poll_add_prep(req, sqe);
5794 ret = io_poll_add(req);
5796 case IORING_OP_POLL_REMOVE:
5798 ret = io_poll_remove_prep(req, sqe);
5802 ret = io_poll_remove(req);
5804 case IORING_OP_SYNC_FILE_RANGE:
5806 ret = io_prep_sfr(req, sqe);
5810 ret = io_sync_file_range(req, force_nonblock);
5812 case IORING_OP_SENDMSG:
5813 case IORING_OP_SEND:
5815 ret = io_sendmsg_prep(req, sqe);
5819 if (req->opcode == IORING_OP_SENDMSG)
5820 ret = io_sendmsg(req, force_nonblock, cs);
5822 ret = io_send(req, force_nonblock, cs);
5824 case IORING_OP_RECVMSG:
5825 case IORING_OP_RECV:
5827 ret = io_recvmsg_prep(req, sqe);
5831 if (req->opcode == IORING_OP_RECVMSG)
5832 ret = io_recvmsg(req, force_nonblock, cs);
5834 ret = io_recv(req, force_nonblock, cs);
5836 case IORING_OP_TIMEOUT:
5838 ret = io_timeout_prep(req, sqe, false);
5842 ret = io_timeout(req);
5844 case IORING_OP_TIMEOUT_REMOVE:
5846 ret = io_timeout_remove_prep(req, sqe);
5850 ret = io_timeout_remove(req);
5852 case IORING_OP_ACCEPT:
5854 ret = io_accept_prep(req, sqe);
5858 ret = io_accept(req, force_nonblock, cs);
5860 case IORING_OP_CONNECT:
5862 ret = io_connect_prep(req, sqe);
5866 ret = io_connect(req, force_nonblock, cs);
5868 case IORING_OP_ASYNC_CANCEL:
5870 ret = io_async_cancel_prep(req, sqe);
5874 ret = io_async_cancel(req);
5876 case IORING_OP_FALLOCATE:
5878 ret = io_fallocate_prep(req, sqe);
5882 ret = io_fallocate(req, force_nonblock);
5884 case IORING_OP_OPENAT:
5886 ret = io_openat_prep(req, sqe);
5890 ret = io_openat(req, force_nonblock);
5892 case IORING_OP_CLOSE:
5894 ret = io_close_prep(req, sqe);
5898 ret = io_close(req, force_nonblock, cs);
5900 case IORING_OP_FILES_UPDATE:
5902 ret = io_files_update_prep(req, sqe);
5906 ret = io_files_update(req, force_nonblock, cs);
5908 case IORING_OP_STATX:
5910 ret = io_statx_prep(req, sqe);
5914 ret = io_statx(req, force_nonblock);
5916 case IORING_OP_FADVISE:
5918 ret = io_fadvise_prep(req, sqe);
5922 ret = io_fadvise(req, force_nonblock);
5924 case IORING_OP_MADVISE:
5926 ret = io_madvise_prep(req, sqe);
5930 ret = io_madvise(req, force_nonblock);
5932 case IORING_OP_OPENAT2:
5934 ret = io_openat2_prep(req, sqe);
5938 ret = io_openat2(req, force_nonblock);
5940 case IORING_OP_EPOLL_CTL:
5942 ret = io_epoll_ctl_prep(req, sqe);
5946 ret = io_epoll_ctl(req, force_nonblock, cs);
5948 case IORING_OP_SPLICE:
5950 ret = io_splice_prep(req, sqe);
5954 ret = io_splice(req, force_nonblock);
5956 case IORING_OP_PROVIDE_BUFFERS:
5958 ret = io_provide_buffers_prep(req, sqe);
5962 ret = io_provide_buffers(req, force_nonblock, cs);
5964 case IORING_OP_REMOVE_BUFFERS:
5966 ret = io_remove_buffers_prep(req, sqe);
5970 ret = io_remove_buffers(req, force_nonblock, cs);
5974 ret = io_tee_prep(req, sqe);
5978 ret = io_tee(req, force_nonblock);
5988 /* If the op doesn't have a file, we're not polling for it */
5989 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5990 const bool in_async = io_wq_current_is_worker();
5992 /* workqueue context doesn't hold uring_lock, grab it now */
5994 mutex_lock(&ctx->uring_lock);
5996 io_iopoll_req_issued(req);
5999 mutex_unlock(&ctx->uring_lock);
6005 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6007 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6008 struct io_kiocb *timeout;
6011 timeout = io_prep_linked_timeout(req);
6013 io_queue_linked_timeout(timeout);
6015 /* if NO_CANCEL is set, we must still run the work */
6016 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6017 IO_WQ_WORK_CANCEL) {
6023 ret = io_issue_sqe(req, NULL, false, NULL);
6025 * We can get EAGAIN for polled IO even though we're
6026 * forcing a sync submission from here, since we can't
6027 * wait for request slots on the block side.
6036 req_set_fail_links(req);
6037 io_req_complete(req, ret);
6040 return io_steal_work(req);
6043 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6046 struct fixed_file_table *table;
6048 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6049 return table->files[index & IORING_FILE_TABLE_MASK];
6052 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
6053 int fd, struct file **out_file, bool fixed)
6055 struct io_ring_ctx *ctx = req->ctx;
6059 if (unlikely(!ctx->file_data ||
6060 (unsigned) fd >= ctx->nr_user_files))
6062 fd = array_index_nospec(fd, ctx->nr_user_files);
6063 file = io_file_from_index(ctx, fd);
6065 req->fixed_file_refs = ctx->file_data->cur_refs;
6066 percpu_ref_get(req->fixed_file_refs);
6069 trace_io_uring_file_get(ctx, fd);
6070 file = __io_file_get(state, fd);
6073 if (file || io_op_defs[req->opcode].needs_file_no_error) {
6080 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6085 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6086 if (unlikely(!fixed && io_async_submit(req->ctx)))
6089 return io_file_get(state, req, fd, &req->file, fixed);
6092 static int io_grab_files(struct io_kiocb *req)
6094 struct io_ring_ctx *ctx = req->ctx;
6096 io_req_init_async(req);
6098 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
6101 req->work.files = get_files_struct(current);
6102 get_nsproxy(current->nsproxy);
6103 req->work.nsproxy = current->nsproxy;
6104 req->flags |= REQ_F_INFLIGHT;
6106 spin_lock_irq(&ctx->inflight_lock);
6107 list_add(&req->inflight_entry, &ctx->inflight_list);
6108 spin_unlock_irq(&ctx->inflight_lock);
6112 static inline int io_prep_work_files(struct io_kiocb *req)
6114 if (!io_op_defs[req->opcode].file_table)
6116 return io_grab_files(req);
6119 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6121 struct io_timeout_data *data = container_of(timer,
6122 struct io_timeout_data, timer);
6123 struct io_kiocb *req = data->req;
6124 struct io_ring_ctx *ctx = req->ctx;
6125 struct io_kiocb *prev = NULL;
6126 unsigned long flags;
6128 spin_lock_irqsave(&ctx->completion_lock, flags);
6131 * We don't expect the list to be empty, that will only happen if we
6132 * race with the completion of the linked work.
6134 if (!list_empty(&req->link_list)) {
6135 prev = list_entry(req->link_list.prev, struct io_kiocb,
6137 if (refcount_inc_not_zero(&prev->refs)) {
6138 list_del_init(&req->link_list);
6139 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6144 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6147 req_set_fail_links(prev);
6148 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6151 io_req_complete(req, -ETIME);
6153 return HRTIMER_NORESTART;
6156 static void __io_queue_linked_timeout(struct io_kiocb *req)
6159 * If the list is now empty, then our linked request finished before
6160 * we got a chance to setup the timer
6162 if (!list_empty(&req->link_list)) {
6163 struct io_timeout_data *data = &req->io->timeout;
6165 data->timer.function = io_link_timeout_fn;
6166 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6171 static void io_queue_linked_timeout(struct io_kiocb *req)
6173 struct io_ring_ctx *ctx = req->ctx;
6175 spin_lock_irq(&ctx->completion_lock);
6176 __io_queue_linked_timeout(req);
6177 spin_unlock_irq(&ctx->completion_lock);
6179 /* drop submission reference */
6183 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6185 struct io_kiocb *nxt;
6187 if (!(req->flags & REQ_F_LINK_HEAD))
6189 if (req->flags & REQ_F_LINK_TIMEOUT)
6192 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6194 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6197 req->flags |= REQ_F_LINK_TIMEOUT;
6201 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6202 struct io_comp_state *cs)
6204 struct io_kiocb *linked_timeout;
6205 struct io_kiocb *nxt;
6206 const struct cred *old_creds = NULL;
6210 linked_timeout = io_prep_linked_timeout(req);
6212 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6213 req->work.creds != current_cred()) {
6215 revert_creds(old_creds);
6216 if (old_creds == req->work.creds)
6217 old_creds = NULL; /* restored original creds */
6219 old_creds = override_creds(req->work.creds);
6222 ret = io_issue_sqe(req, sqe, true, cs);
6225 * We async punt it if the file wasn't marked NOWAIT, or if the file
6226 * doesn't support non-blocking read/write attempts
6228 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6229 if (!io_arm_poll_handler(req)) {
6231 ret = io_prep_work_files(req);
6235 * Queued up for async execution, worker will release
6236 * submit reference when the iocb is actually submitted.
6238 io_queue_async_work(req);
6242 io_queue_linked_timeout(linked_timeout);
6246 if (unlikely(ret)) {
6248 /* un-prep timeout, so it'll be killed as any other linked */
6249 req->flags &= ~REQ_F_LINK_TIMEOUT;
6250 req_set_fail_links(req);
6252 io_req_complete(req, ret);
6256 /* drop submission reference */
6257 nxt = io_put_req_find_next(req);
6259 io_queue_linked_timeout(linked_timeout);
6264 if (req->flags & REQ_F_FORCE_ASYNC)
6270 revert_creds(old_creds);
6273 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6274 struct io_comp_state *cs)
6278 ret = io_req_defer(req, sqe);
6280 if (ret != -EIOCBQUEUED) {
6282 req_set_fail_links(req);
6284 io_req_complete(req, ret);
6286 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6288 ret = io_req_defer_prep(req, sqe);
6294 * Never try inline submit of IOSQE_ASYNC is set, go straight
6295 * to async execution.
6297 io_req_init_async(req);
6298 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6299 io_queue_async_work(req);
6301 __io_queue_sqe(req, sqe, cs);
6305 static inline void io_queue_link_head(struct io_kiocb *req,
6306 struct io_comp_state *cs)
6308 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6310 io_req_complete(req, -ECANCELED);
6312 io_queue_sqe(req, NULL, cs);
6315 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6316 struct io_kiocb **link, struct io_comp_state *cs)
6318 struct io_ring_ctx *ctx = req->ctx;
6322 * If we already have a head request, queue this one for async
6323 * submittal once the head completes. If we don't have a head but
6324 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6325 * submitted sync once the chain is complete. If none of those
6326 * conditions are true (normal request), then just queue it.
6329 struct io_kiocb *head = *link;
6332 * Taking sequential execution of a link, draining both sides
6333 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6334 * requests in the link. So, it drains the head and the
6335 * next after the link request. The last one is done via
6336 * drain_next flag to persist the effect across calls.
6338 if (req->flags & REQ_F_IO_DRAIN) {
6339 head->flags |= REQ_F_IO_DRAIN;
6340 ctx->drain_next = 1;
6342 ret = io_req_defer_prep(req, sqe);
6343 if (unlikely(ret)) {
6344 /* fail even hard links since we don't submit */
6345 head->flags |= REQ_F_FAIL_LINK;
6348 trace_io_uring_link(ctx, req, head);
6349 list_add_tail(&req->link_list, &head->link_list);
6351 /* last request of a link, enqueue the link */
6352 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6353 io_queue_link_head(head, cs);
6357 if (unlikely(ctx->drain_next)) {
6358 req->flags |= REQ_F_IO_DRAIN;
6359 ctx->drain_next = 0;
6361 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6362 req->flags |= REQ_F_LINK_HEAD;
6363 INIT_LIST_HEAD(&req->link_list);
6365 ret = io_req_defer_prep(req, sqe);
6367 req->flags |= REQ_F_FAIL_LINK;
6370 io_queue_sqe(req, sqe, cs);
6378 * Batched submission is done, ensure local IO is flushed out.
6380 static void io_submit_state_end(struct io_submit_state *state)
6382 if (!list_empty(&state->comp.list))
6383 io_submit_flush_completions(&state->comp);
6384 blk_finish_plug(&state->plug);
6385 io_state_file_put(state);
6386 if (state->free_reqs)
6387 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6391 * Start submission side cache.
6393 static void io_submit_state_start(struct io_submit_state *state,
6394 struct io_ring_ctx *ctx, unsigned int max_ios)
6396 blk_start_plug(&state->plug);
6398 INIT_LIST_HEAD(&state->comp.list);
6399 state->comp.ctx = ctx;
6400 state->free_reqs = 0;
6402 state->ios_left = max_ios;
6405 static void io_commit_sqring(struct io_ring_ctx *ctx)
6407 struct io_rings *rings = ctx->rings;
6410 * Ensure any loads from the SQEs are done at this point,
6411 * since once we write the new head, the application could
6412 * write new data to them.
6414 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6418 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6419 * that is mapped by userspace. This means that care needs to be taken to
6420 * ensure that reads are stable, as we cannot rely on userspace always
6421 * being a good citizen. If members of the sqe are validated and then later
6422 * used, it's important that those reads are done through READ_ONCE() to
6423 * prevent a re-load down the line.
6425 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6427 u32 *sq_array = ctx->sq_array;
6431 * The cached sq head (or cq tail) serves two purposes:
6433 * 1) allows us to batch the cost of updating the user visible
6435 * 2) allows the kernel side to track the head on its own, even
6436 * though the application is the one updating it.
6438 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6439 if (likely(head < ctx->sq_entries))
6440 return &ctx->sq_sqes[head];
6442 /* drop invalid entries */
6443 ctx->cached_sq_dropped++;
6444 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6448 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6450 ctx->cached_sq_head++;
6454 * Check SQE restrictions (opcode and flags).
6456 * Returns 'true' if SQE is allowed, 'false' otherwise.
6458 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6459 struct io_kiocb *req,
6460 unsigned int sqe_flags)
6462 if (!ctx->restricted)
6465 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6468 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6469 ctx->restrictions.sqe_flags_required)
6472 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6473 ctx->restrictions.sqe_flags_required))
6479 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6480 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6481 IOSQE_BUFFER_SELECT)
6483 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6484 const struct io_uring_sqe *sqe,
6485 struct io_submit_state *state)
6487 unsigned int sqe_flags;
6490 req->opcode = READ_ONCE(sqe->opcode);
6491 req->user_data = READ_ONCE(sqe->user_data);
6496 /* one is dropped after submission, the other at completion */
6497 refcount_set(&req->refs, 2);
6498 req->task = current;
6499 get_task_struct(req->task);
6500 atomic_long_inc(&req->task->io_uring->req_issue);
6503 if (unlikely(req->opcode >= IORING_OP_LAST))
6506 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6509 sqe_flags = READ_ONCE(sqe->flags);
6510 /* enforce forwards compatibility on users */
6511 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6514 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6517 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6518 !io_op_defs[req->opcode].buffer_select)
6521 id = READ_ONCE(sqe->personality);
6523 io_req_init_async(req);
6524 req->work.creds = idr_find(&ctx->personality_idr, id);
6525 if (unlikely(!req->work.creds))
6527 get_cred(req->work.creds);
6530 /* same numerical values with corresponding REQ_F_*, safe to copy */
6531 req->flags |= sqe_flags;
6533 if (!io_op_defs[req->opcode].needs_file)
6536 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6539 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6541 struct io_submit_state state;
6542 struct io_kiocb *link = NULL;
6543 int i, submitted = 0;
6545 /* if we have a backlog and couldn't flush it all, return BUSY */
6546 if (test_bit(0, &ctx->sq_check_overflow)) {
6547 if (!list_empty(&ctx->cq_overflow_list) &&
6548 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6552 /* make sure SQ entry isn't read before tail */
6553 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6555 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6558 io_submit_state_start(&state, ctx, nr);
6560 for (i = 0; i < nr; i++) {
6561 const struct io_uring_sqe *sqe;
6562 struct io_kiocb *req;
6565 sqe = io_get_sqe(ctx);
6566 if (unlikely(!sqe)) {
6567 io_consume_sqe(ctx);
6570 req = io_alloc_req(ctx, &state);
6571 if (unlikely(!req)) {
6573 submitted = -EAGAIN;
6577 err = io_init_req(ctx, req, sqe, &state);
6578 io_consume_sqe(ctx);
6579 /* will complete beyond this point, count as submitted */
6582 if (unlikely(err)) {
6585 io_req_complete(req, err);
6589 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6590 true, io_async_submit(ctx));
6591 err = io_submit_sqe(req, sqe, &link, &state.comp);
6596 if (unlikely(submitted != nr)) {
6597 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6599 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6602 io_queue_link_head(link, &state.comp);
6603 io_submit_state_end(&state);
6605 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6606 io_commit_sqring(ctx);
6611 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6613 /* Tell userspace we may need a wakeup call */
6614 spin_lock_irq(&ctx->completion_lock);
6615 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6616 spin_unlock_irq(&ctx->completion_lock);
6619 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6621 spin_lock_irq(&ctx->completion_lock);
6622 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6623 spin_unlock_irq(&ctx->completion_lock);
6626 static int io_sq_thread(void *data)
6628 struct io_ring_ctx *ctx = data;
6629 const struct cred *old_cred;
6631 unsigned long timeout;
6634 complete(&ctx->sq_thread_comp);
6636 old_cred = override_creds(ctx->creds);
6638 timeout = jiffies + ctx->sq_thread_idle;
6639 while (!kthread_should_park()) {
6640 unsigned int to_submit;
6642 if (!list_empty(&ctx->iopoll_list)) {
6643 unsigned nr_events = 0;
6645 mutex_lock(&ctx->uring_lock);
6646 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6647 io_do_iopoll(ctx, &nr_events, 0);
6649 timeout = jiffies + ctx->sq_thread_idle;
6650 mutex_unlock(&ctx->uring_lock);
6653 to_submit = io_sqring_entries(ctx);
6656 * If submit got -EBUSY, flag us as needing the application
6657 * to enter the kernel to reap and flush events.
6659 if (!to_submit || ret == -EBUSY || need_resched()) {
6661 * Drop cur_mm before scheduling, we can't hold it for
6662 * long periods (or over schedule()). Do this before
6663 * adding ourselves to the waitqueue, as the unuse/drop
6666 io_sq_thread_drop_mm();
6669 * We're polling. If we're within the defined idle
6670 * period, then let us spin without work before going
6671 * to sleep. The exception is if we got EBUSY doing
6672 * more IO, we should wait for the application to
6673 * reap events and wake us up.
6675 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6676 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6677 !percpu_ref_is_dying(&ctx->refs))) {
6683 prepare_to_wait(&ctx->sqo_wait, &wait,
6684 TASK_INTERRUPTIBLE);
6687 * While doing polled IO, before going to sleep, we need
6688 * to check if there are new reqs added to iopoll_list,
6689 * it is because reqs may have been punted to io worker
6690 * and will be added to iopoll_list later, hence check
6691 * the iopoll_list again.
6693 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6694 !list_empty_careful(&ctx->iopoll_list)) {
6695 finish_wait(&ctx->sqo_wait, &wait);
6699 io_ring_set_wakeup_flag(ctx);
6701 to_submit = io_sqring_entries(ctx);
6702 if (!to_submit || ret == -EBUSY) {
6703 if (kthread_should_park()) {
6704 finish_wait(&ctx->sqo_wait, &wait);
6707 if (io_run_task_work()) {
6708 finish_wait(&ctx->sqo_wait, &wait);
6709 io_ring_clear_wakeup_flag(ctx);
6713 finish_wait(&ctx->sqo_wait, &wait);
6715 io_ring_clear_wakeup_flag(ctx);
6719 finish_wait(&ctx->sqo_wait, &wait);
6721 io_ring_clear_wakeup_flag(ctx);
6724 mutex_lock(&ctx->uring_lock);
6725 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6726 ret = io_submit_sqes(ctx, to_submit);
6727 mutex_unlock(&ctx->uring_lock);
6728 timeout = jiffies + ctx->sq_thread_idle;
6733 io_sq_thread_drop_mm();
6734 revert_creds(old_cred);
6741 struct io_wait_queue {
6742 struct wait_queue_entry wq;
6743 struct io_ring_ctx *ctx;
6745 unsigned nr_timeouts;
6748 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6750 struct io_ring_ctx *ctx = iowq->ctx;
6753 * Wake up if we have enough events, or if a timeout occurred since we
6754 * started waiting. For timeouts, we always want to return to userspace,
6755 * regardless of event count.
6757 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6758 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6761 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6762 int wake_flags, void *key)
6764 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6767 /* use noflush == true, as we can't safely rely on locking context */
6768 if (!io_should_wake(iowq, true))
6771 return autoremove_wake_function(curr, mode, wake_flags, key);
6775 * Wait until events become available, if we don't already have some. The
6776 * application must reap them itself, as they reside on the shared cq ring.
6778 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6779 const sigset_t __user *sig, size_t sigsz)
6781 struct io_wait_queue iowq = {
6784 .func = io_wake_function,
6785 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6788 .to_wait = min_events,
6790 struct io_rings *rings = ctx->rings;
6794 if (io_cqring_events(ctx, false) >= min_events)
6796 if (!io_run_task_work())
6801 #ifdef CONFIG_COMPAT
6802 if (in_compat_syscall())
6803 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6807 ret = set_user_sigmask(sig, sigsz);
6813 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6814 trace_io_uring_cqring_wait(ctx, min_events);
6816 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6817 TASK_INTERRUPTIBLE);
6818 /* make sure we run task_work before checking for signals */
6819 if (io_run_task_work())
6821 if (signal_pending(current)) {
6822 if (current->jobctl & JOBCTL_TASK_WORK) {
6823 spin_lock_irq(¤t->sighand->siglock);
6824 current->jobctl &= ~JOBCTL_TASK_WORK;
6825 recalc_sigpending();
6826 spin_unlock_irq(¤t->sighand->siglock);
6832 if (io_should_wake(&iowq, false))
6836 finish_wait(&ctx->wait, &iowq.wq);
6838 restore_saved_sigmask_unless(ret == -EINTR);
6840 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6843 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6845 #if defined(CONFIG_UNIX)
6846 if (ctx->ring_sock) {
6847 struct sock *sock = ctx->ring_sock->sk;
6848 struct sk_buff *skb;
6850 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6856 for (i = 0; i < ctx->nr_user_files; i++) {
6859 file = io_file_from_index(ctx, i);
6866 static void io_file_ref_kill(struct percpu_ref *ref)
6868 struct fixed_file_data *data;
6870 data = container_of(ref, struct fixed_file_data, refs);
6871 complete(&data->done);
6874 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6876 struct fixed_file_data *data = ctx->file_data;
6877 struct fixed_file_ref_node *ref_node = NULL;
6878 unsigned nr_tables, i;
6883 spin_lock(&data->lock);
6884 if (!list_empty(&data->ref_list))
6885 ref_node = list_first_entry(&data->ref_list,
6886 struct fixed_file_ref_node, node);
6887 spin_unlock(&data->lock);
6889 percpu_ref_kill(&ref_node->refs);
6891 percpu_ref_kill(&data->refs);
6893 /* wait for all refs nodes to complete */
6894 flush_delayed_work(&ctx->file_put_work);
6895 wait_for_completion(&data->done);
6897 __io_sqe_files_unregister(ctx);
6898 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6899 for (i = 0; i < nr_tables; i++)
6900 kfree(data->table[i].files);
6902 percpu_ref_exit(&data->refs);
6904 ctx->file_data = NULL;
6905 ctx->nr_user_files = 0;
6909 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6911 if (ctx->sqo_thread) {
6913 * We may arrive here from the error branch in
6914 * io_sq_offload_create() where the kthread is created
6915 * without being waked up, thus wake it up now to make
6916 * sure the wait will complete.
6918 wake_up_process(ctx->sqo_thread);
6920 wait_for_completion(&ctx->sq_thread_comp);
6922 * The park is a bit of a work-around, without it we get
6923 * warning spews on shutdown with SQPOLL set and affinity
6924 * set to a single CPU.
6926 kthread_park(ctx->sqo_thread);
6927 kthread_stop(ctx->sqo_thread);
6928 ctx->sqo_thread = NULL;
6932 static void io_finish_async(struct io_ring_ctx *ctx)
6934 io_sq_thread_stop(ctx);
6937 io_wq_destroy(ctx->io_wq);
6942 #if defined(CONFIG_UNIX)
6944 * Ensure the UNIX gc is aware of our file set, so we are certain that
6945 * the io_uring can be safely unregistered on process exit, even if we have
6946 * loops in the file referencing.
6948 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6950 struct sock *sk = ctx->ring_sock->sk;
6951 struct scm_fp_list *fpl;
6952 struct sk_buff *skb;
6955 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6959 skb = alloc_skb(0, GFP_KERNEL);
6968 fpl->user = get_uid(ctx->user);
6969 for (i = 0; i < nr; i++) {
6970 struct file *file = io_file_from_index(ctx, i + offset);
6974 fpl->fp[nr_files] = get_file(file);
6975 unix_inflight(fpl->user, fpl->fp[nr_files]);
6980 fpl->max = SCM_MAX_FD;
6981 fpl->count = nr_files;
6982 UNIXCB(skb).fp = fpl;
6983 skb->destructor = unix_destruct_scm;
6984 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6985 skb_queue_head(&sk->sk_receive_queue, skb);
6987 for (i = 0; i < nr_files; i++)
6998 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6999 * causes regular reference counting to break down. We rely on the UNIX
7000 * garbage collection to take care of this problem for us.
7002 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7004 unsigned left, total;
7008 left = ctx->nr_user_files;
7010 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7012 ret = __io_sqe_files_scm(ctx, this_files, total);
7016 total += this_files;
7022 while (total < ctx->nr_user_files) {
7023 struct file *file = io_file_from_index(ctx, total);
7033 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7039 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
7044 for (i = 0; i < nr_tables; i++) {
7045 struct fixed_file_table *table = &ctx->file_data->table[i];
7046 unsigned this_files;
7048 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7049 table->files = kcalloc(this_files, sizeof(struct file *),
7053 nr_files -= this_files;
7059 for (i = 0; i < nr_tables; i++) {
7060 struct fixed_file_table *table = &ctx->file_data->table[i];
7061 kfree(table->files);
7066 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7068 #if defined(CONFIG_UNIX)
7069 struct sock *sock = ctx->ring_sock->sk;
7070 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7071 struct sk_buff *skb;
7074 __skb_queue_head_init(&list);
7077 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7078 * remove this entry and rearrange the file array.
7080 skb = skb_dequeue(head);
7082 struct scm_fp_list *fp;
7084 fp = UNIXCB(skb).fp;
7085 for (i = 0; i < fp->count; i++) {
7088 if (fp->fp[i] != file)
7091 unix_notinflight(fp->user, fp->fp[i]);
7092 left = fp->count - 1 - i;
7094 memmove(&fp->fp[i], &fp->fp[i + 1],
7095 left * sizeof(struct file *));
7102 __skb_queue_tail(&list, skb);
7112 __skb_queue_tail(&list, skb);
7114 skb = skb_dequeue(head);
7117 if (skb_peek(&list)) {
7118 spin_lock_irq(&head->lock);
7119 while ((skb = __skb_dequeue(&list)) != NULL)
7120 __skb_queue_tail(head, skb);
7121 spin_unlock_irq(&head->lock);
7128 struct io_file_put {
7129 struct list_head list;
7133 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7135 struct fixed_file_data *file_data = ref_node->file_data;
7136 struct io_ring_ctx *ctx = file_data->ctx;
7137 struct io_file_put *pfile, *tmp;
7139 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7140 list_del(&pfile->list);
7141 io_ring_file_put(ctx, pfile->file);
7145 spin_lock(&file_data->lock);
7146 list_del(&ref_node->node);
7147 spin_unlock(&file_data->lock);
7149 percpu_ref_exit(&ref_node->refs);
7151 percpu_ref_put(&file_data->refs);
7154 static void io_file_put_work(struct work_struct *work)
7156 struct io_ring_ctx *ctx;
7157 struct llist_node *node;
7159 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7160 node = llist_del_all(&ctx->file_put_llist);
7163 struct fixed_file_ref_node *ref_node;
7164 struct llist_node *next = node->next;
7166 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7167 __io_file_put_work(ref_node);
7172 static void io_file_data_ref_zero(struct percpu_ref *ref)
7174 struct fixed_file_ref_node *ref_node;
7175 struct io_ring_ctx *ctx;
7179 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7180 ctx = ref_node->file_data->ctx;
7182 if (percpu_ref_is_dying(&ctx->file_data->refs))
7185 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7187 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7189 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7192 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7193 struct io_ring_ctx *ctx)
7195 struct fixed_file_ref_node *ref_node;
7197 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7199 return ERR_PTR(-ENOMEM);
7201 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7204 return ERR_PTR(-ENOMEM);
7206 INIT_LIST_HEAD(&ref_node->node);
7207 INIT_LIST_HEAD(&ref_node->file_list);
7208 ref_node->file_data = ctx->file_data;
7212 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7214 percpu_ref_exit(&ref_node->refs);
7218 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7221 __s32 __user *fds = (__s32 __user *) arg;
7226 struct fixed_file_ref_node *ref_node;
7232 if (nr_args > IORING_MAX_FIXED_FILES)
7235 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7236 if (!ctx->file_data)
7238 ctx->file_data->ctx = ctx;
7239 init_completion(&ctx->file_data->done);
7240 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7241 spin_lock_init(&ctx->file_data->lock);
7243 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7244 ctx->file_data->table = kcalloc(nr_tables,
7245 sizeof(struct fixed_file_table),
7247 if (!ctx->file_data->table) {
7248 kfree(ctx->file_data);
7249 ctx->file_data = NULL;
7253 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7254 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7255 kfree(ctx->file_data->table);
7256 kfree(ctx->file_data);
7257 ctx->file_data = NULL;
7261 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7262 percpu_ref_exit(&ctx->file_data->refs);
7263 kfree(ctx->file_data->table);
7264 kfree(ctx->file_data);
7265 ctx->file_data = NULL;
7269 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7270 struct fixed_file_table *table;
7274 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7276 /* allow sparse sets */
7282 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7283 index = i & IORING_FILE_TABLE_MASK;
7291 * Don't allow io_uring instances to be registered. If UNIX
7292 * isn't enabled, then this causes a reference cycle and this
7293 * instance can never get freed. If UNIX is enabled we'll
7294 * handle it just fine, but there's still no point in allowing
7295 * a ring fd as it doesn't support regular read/write anyway.
7297 if (file->f_op == &io_uring_fops) {
7302 table->files[index] = file;
7306 for (i = 0; i < ctx->nr_user_files; i++) {
7307 file = io_file_from_index(ctx, i);
7311 for (i = 0; i < nr_tables; i++)
7312 kfree(ctx->file_data->table[i].files);
7314 percpu_ref_exit(&ctx->file_data->refs);
7315 kfree(ctx->file_data->table);
7316 kfree(ctx->file_data);
7317 ctx->file_data = NULL;
7318 ctx->nr_user_files = 0;
7322 ret = io_sqe_files_scm(ctx);
7324 io_sqe_files_unregister(ctx);
7328 ref_node = alloc_fixed_file_ref_node(ctx);
7329 if (IS_ERR(ref_node)) {
7330 io_sqe_files_unregister(ctx);
7331 return PTR_ERR(ref_node);
7334 ctx->file_data->cur_refs = &ref_node->refs;
7335 spin_lock(&ctx->file_data->lock);
7336 list_add(&ref_node->node, &ctx->file_data->ref_list);
7337 spin_unlock(&ctx->file_data->lock);
7338 percpu_ref_get(&ctx->file_data->refs);
7342 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7345 #if defined(CONFIG_UNIX)
7346 struct sock *sock = ctx->ring_sock->sk;
7347 struct sk_buff_head *head = &sock->sk_receive_queue;
7348 struct sk_buff *skb;
7351 * See if we can merge this file into an existing skb SCM_RIGHTS
7352 * file set. If there's no room, fall back to allocating a new skb
7353 * and filling it in.
7355 spin_lock_irq(&head->lock);
7356 skb = skb_peek(head);
7358 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7360 if (fpl->count < SCM_MAX_FD) {
7361 __skb_unlink(skb, head);
7362 spin_unlock_irq(&head->lock);
7363 fpl->fp[fpl->count] = get_file(file);
7364 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7366 spin_lock_irq(&head->lock);
7367 __skb_queue_head(head, skb);
7372 spin_unlock_irq(&head->lock);
7379 return __io_sqe_files_scm(ctx, 1, index);
7385 static int io_queue_file_removal(struct fixed_file_data *data,
7388 struct io_file_put *pfile;
7389 struct percpu_ref *refs = data->cur_refs;
7390 struct fixed_file_ref_node *ref_node;
7392 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7396 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7398 list_add(&pfile->list, &ref_node->file_list);
7403 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7404 struct io_uring_files_update *up,
7407 struct fixed_file_data *data = ctx->file_data;
7408 struct fixed_file_ref_node *ref_node;
7413 bool needs_switch = false;
7415 if (check_add_overflow(up->offset, nr_args, &done))
7417 if (done > ctx->nr_user_files)
7420 ref_node = alloc_fixed_file_ref_node(ctx);
7421 if (IS_ERR(ref_node))
7422 return PTR_ERR(ref_node);
7425 fds = u64_to_user_ptr(up->fds);
7427 struct fixed_file_table *table;
7431 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7435 i = array_index_nospec(up->offset, ctx->nr_user_files);
7436 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7437 index = i & IORING_FILE_TABLE_MASK;
7438 if (table->files[index]) {
7439 file = table->files[index];
7440 err = io_queue_file_removal(data, file);
7443 table->files[index] = NULL;
7444 needs_switch = true;
7453 * Don't allow io_uring instances to be registered. If
7454 * UNIX isn't enabled, then this causes a reference
7455 * cycle and this instance can never get freed. If UNIX
7456 * is enabled we'll handle it just fine, but there's
7457 * still no point in allowing a ring fd as it doesn't
7458 * support regular read/write anyway.
7460 if (file->f_op == &io_uring_fops) {
7465 table->files[index] = file;
7466 err = io_sqe_file_register(ctx, file, i);
7468 table->files[index] = NULL;
7479 percpu_ref_kill(data->cur_refs);
7480 spin_lock(&data->lock);
7481 list_add(&ref_node->node, &data->ref_list);
7482 data->cur_refs = &ref_node->refs;
7483 spin_unlock(&data->lock);
7484 percpu_ref_get(&ctx->file_data->refs);
7486 destroy_fixed_file_ref_node(ref_node);
7488 return done ? done : err;
7491 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7494 struct io_uring_files_update up;
7496 if (!ctx->file_data)
7500 if (copy_from_user(&up, arg, sizeof(up)))
7505 return __io_sqe_files_update(ctx, &up, nr_args);
7508 static void io_free_work(struct io_wq_work *work)
7510 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7512 /* Consider that io_steal_work() relies on this ref */
7516 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7517 struct io_uring_params *p)
7519 struct io_wq_data data;
7521 struct io_ring_ctx *ctx_attach;
7522 unsigned int concurrency;
7525 data.user = ctx->user;
7526 data.free_work = io_free_work;
7527 data.do_work = io_wq_submit_work;
7529 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7530 /* Do QD, or 4 * CPUS, whatever is smallest */
7531 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7533 ctx->io_wq = io_wq_create(concurrency, &data);
7534 if (IS_ERR(ctx->io_wq)) {
7535 ret = PTR_ERR(ctx->io_wq);
7541 f = fdget(p->wq_fd);
7545 if (f.file->f_op != &io_uring_fops) {
7550 ctx_attach = f.file->private_data;
7551 /* @io_wq is protected by holding the fd */
7552 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7557 ctx->io_wq = ctx_attach->io_wq;
7563 static int io_uring_alloc_task_context(struct task_struct *task)
7565 struct io_uring_task *tctx;
7567 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7568 if (unlikely(!tctx))
7572 init_waitqueue_head(&tctx->wait);
7575 atomic_long_set(&tctx->req_issue, 0);
7576 atomic_long_set(&tctx->req_complete, 0);
7577 task->io_uring = tctx;
7581 void __io_uring_free(struct task_struct *tsk)
7583 struct io_uring_task *tctx = tsk->io_uring;
7585 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7586 xa_destroy(&tctx->xa);
7588 tsk->io_uring = NULL;
7591 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7592 struct io_uring_params *p)
7596 if (ctx->flags & IORING_SETUP_SQPOLL) {
7598 if (!capable(CAP_SYS_ADMIN))
7601 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7602 if (!ctx->sq_thread_idle)
7603 ctx->sq_thread_idle = HZ;
7605 if (p->flags & IORING_SETUP_SQ_AFF) {
7606 int cpu = p->sq_thread_cpu;
7609 if (cpu >= nr_cpu_ids)
7611 if (!cpu_online(cpu))
7614 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7618 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7621 if (IS_ERR(ctx->sqo_thread)) {
7622 ret = PTR_ERR(ctx->sqo_thread);
7623 ctx->sqo_thread = NULL;
7626 ret = io_uring_alloc_task_context(ctx->sqo_thread);
7629 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7630 /* Can't have SQ_AFF without SQPOLL */
7635 ret = io_init_wq_offload(ctx, p);
7641 io_finish_async(ctx);
7645 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7647 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sqo_thread)
7648 wake_up_process(ctx->sqo_thread);
7651 static inline void __io_unaccount_mem(struct user_struct *user,
7652 unsigned long nr_pages)
7654 atomic_long_sub(nr_pages, &user->locked_vm);
7657 static inline int __io_account_mem(struct user_struct *user,
7658 unsigned long nr_pages)
7660 unsigned long page_limit, cur_pages, new_pages;
7662 /* Don't allow more pages than we can safely lock */
7663 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7666 cur_pages = atomic_long_read(&user->locked_vm);
7667 new_pages = cur_pages + nr_pages;
7668 if (new_pages > page_limit)
7670 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7671 new_pages) != cur_pages);
7676 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7677 enum io_mem_account acct)
7680 __io_unaccount_mem(ctx->user, nr_pages);
7682 if (ctx->mm_account) {
7683 if (acct == ACCT_LOCKED)
7684 ctx->mm_account->locked_vm -= nr_pages;
7685 else if (acct == ACCT_PINNED)
7686 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7690 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7691 enum io_mem_account acct)
7695 if (ctx->limit_mem) {
7696 ret = __io_account_mem(ctx->user, nr_pages);
7701 if (ctx->mm_account) {
7702 if (acct == ACCT_LOCKED)
7703 ctx->mm_account->locked_vm += nr_pages;
7704 else if (acct == ACCT_PINNED)
7705 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7711 static void io_mem_free(void *ptr)
7718 page = virt_to_head_page(ptr);
7719 if (put_page_testzero(page))
7720 free_compound_page(page);
7723 static void *io_mem_alloc(size_t size)
7725 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7728 return (void *) __get_free_pages(gfp_flags, get_order(size));
7731 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7734 struct io_rings *rings;
7735 size_t off, sq_array_size;
7737 off = struct_size(rings, cqes, cq_entries);
7738 if (off == SIZE_MAX)
7742 off = ALIGN(off, SMP_CACHE_BYTES);
7750 sq_array_size = array_size(sizeof(u32), sq_entries);
7751 if (sq_array_size == SIZE_MAX)
7754 if (check_add_overflow(off, sq_array_size, &off))
7760 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7764 pages = (size_t)1 << get_order(
7765 rings_size(sq_entries, cq_entries, NULL));
7766 pages += (size_t)1 << get_order(
7767 array_size(sizeof(struct io_uring_sqe), sq_entries));
7772 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7776 if (!ctx->user_bufs)
7779 for (i = 0; i < ctx->nr_user_bufs; i++) {
7780 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7782 for (j = 0; j < imu->nr_bvecs; j++)
7783 unpin_user_page(imu->bvec[j].bv_page);
7785 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7790 kfree(ctx->user_bufs);
7791 ctx->user_bufs = NULL;
7792 ctx->nr_user_bufs = 0;
7796 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7797 void __user *arg, unsigned index)
7799 struct iovec __user *src;
7801 #ifdef CONFIG_COMPAT
7803 struct compat_iovec __user *ciovs;
7804 struct compat_iovec ciov;
7806 ciovs = (struct compat_iovec __user *) arg;
7807 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7810 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7811 dst->iov_len = ciov.iov_len;
7815 src = (struct iovec __user *) arg;
7816 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7821 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7824 struct vm_area_struct **vmas = NULL;
7825 struct page **pages = NULL;
7826 int i, j, got_pages = 0;
7831 if (!nr_args || nr_args > UIO_MAXIOV)
7834 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7836 if (!ctx->user_bufs)
7839 for (i = 0; i < nr_args; i++) {
7840 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7841 unsigned long off, start, end, ubuf;
7846 ret = io_copy_iov(ctx, &iov, arg, i);
7851 * Don't impose further limits on the size and buffer
7852 * constraints here, we'll -EINVAL later when IO is
7853 * submitted if they are wrong.
7856 if (!iov.iov_base || !iov.iov_len)
7859 /* arbitrary limit, but we need something */
7860 if (iov.iov_len > SZ_1G)
7863 ubuf = (unsigned long) iov.iov_base;
7864 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7865 start = ubuf >> PAGE_SHIFT;
7866 nr_pages = end - start;
7868 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7873 if (!pages || nr_pages > got_pages) {
7876 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7878 vmas = kvmalloc_array(nr_pages,
7879 sizeof(struct vm_area_struct *),
7881 if (!pages || !vmas) {
7883 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7886 got_pages = nr_pages;
7889 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7893 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7898 mmap_read_lock(current->mm);
7899 pret = pin_user_pages(ubuf, nr_pages,
7900 FOLL_WRITE | FOLL_LONGTERM,
7902 if (pret == nr_pages) {
7903 /* don't support file backed memory */
7904 for (j = 0; j < nr_pages; j++) {
7905 struct vm_area_struct *vma = vmas[j];
7908 !is_file_hugepages(vma->vm_file)) {
7914 ret = pret < 0 ? pret : -EFAULT;
7916 mmap_read_unlock(current->mm);
7919 * if we did partial map, or found file backed vmas,
7920 * release any pages we did get
7923 unpin_user_pages(pages, pret);
7924 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7929 off = ubuf & ~PAGE_MASK;
7931 for (j = 0; j < nr_pages; j++) {
7934 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7935 imu->bvec[j].bv_page = pages[j];
7936 imu->bvec[j].bv_len = vec_len;
7937 imu->bvec[j].bv_offset = off;
7941 /* store original address for later verification */
7943 imu->len = iov.iov_len;
7944 imu->nr_bvecs = nr_pages;
7946 ctx->nr_user_bufs++;
7954 io_sqe_buffer_unregister(ctx);
7958 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7960 __s32 __user *fds = arg;
7966 if (copy_from_user(&fd, fds, sizeof(*fds)))
7969 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7970 if (IS_ERR(ctx->cq_ev_fd)) {
7971 int ret = PTR_ERR(ctx->cq_ev_fd);
7972 ctx->cq_ev_fd = NULL;
7979 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7981 if (ctx->cq_ev_fd) {
7982 eventfd_ctx_put(ctx->cq_ev_fd);
7983 ctx->cq_ev_fd = NULL;
7990 static int __io_destroy_buffers(int id, void *p, void *data)
7992 struct io_ring_ctx *ctx = data;
7993 struct io_buffer *buf = p;
7995 __io_remove_buffers(ctx, buf, id, -1U);
7999 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8001 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8002 idr_destroy(&ctx->io_buffer_idr);
8005 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8007 io_finish_async(ctx);
8008 io_sqe_buffer_unregister(ctx);
8010 if (ctx->sqo_task) {
8011 put_task_struct(ctx->sqo_task);
8012 ctx->sqo_task = NULL;
8013 mmdrop(ctx->mm_account);
8014 ctx->mm_account = NULL;
8017 io_sqe_files_unregister(ctx);
8018 io_eventfd_unregister(ctx);
8019 io_destroy_buffers(ctx);
8020 idr_destroy(&ctx->personality_idr);
8022 #if defined(CONFIG_UNIX)
8023 if (ctx->ring_sock) {
8024 ctx->ring_sock->file = NULL; /* so that iput() is called */
8025 sock_release(ctx->ring_sock);
8029 io_mem_free(ctx->rings);
8030 io_mem_free(ctx->sq_sqes);
8032 percpu_ref_exit(&ctx->refs);
8033 free_uid(ctx->user);
8034 put_cred(ctx->creds);
8035 kfree(ctx->cancel_hash);
8036 kmem_cache_free(req_cachep, ctx->fallback_req);
8040 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8042 struct io_ring_ctx *ctx = file->private_data;
8045 poll_wait(file, &ctx->cq_wait, wait);
8047 * synchronizes with barrier from wq_has_sleeper call in
8051 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
8052 ctx->rings->sq_ring_entries)
8053 mask |= EPOLLOUT | EPOLLWRNORM;
8054 if (io_cqring_events(ctx, false))
8055 mask |= EPOLLIN | EPOLLRDNORM;
8060 static int io_uring_fasync(int fd, struct file *file, int on)
8062 struct io_ring_ctx *ctx = file->private_data;
8064 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8067 static int io_remove_personalities(int id, void *p, void *data)
8069 struct io_ring_ctx *ctx = data;
8070 const struct cred *cred;
8072 cred = idr_remove(&ctx->personality_idr, id);
8078 static void io_ring_exit_work(struct work_struct *work)
8080 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8084 * If we're doing polled IO and end up having requests being
8085 * submitted async (out-of-line), then completions can come in while
8086 * we're waiting for refs to drop. We need to reap these manually,
8087 * as nobody else will be looking for them.
8091 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8092 io_iopoll_try_reap_events(ctx);
8093 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8094 io_ring_ctx_free(ctx);
8097 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8099 mutex_lock(&ctx->uring_lock);
8100 percpu_ref_kill(&ctx->refs);
8101 mutex_unlock(&ctx->uring_lock);
8103 io_kill_timeouts(ctx, NULL);
8104 io_poll_remove_all(ctx, NULL);
8107 io_wq_cancel_all(ctx->io_wq);
8109 /* if we failed setting up the ctx, we might not have any rings */
8111 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8112 io_iopoll_try_reap_events(ctx);
8113 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8116 * Do this upfront, so we won't have a grace period where the ring
8117 * is closed but resources aren't reaped yet. This can cause
8118 * spurious failure in setting up a new ring.
8120 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8123 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8125 * Use system_unbound_wq to avoid spawning tons of event kworkers
8126 * if we're exiting a ton of rings at the same time. It just adds
8127 * noise and overhead, there's no discernable change in runtime
8128 * over using system_wq.
8130 queue_work(system_unbound_wq, &ctx->exit_work);
8133 static int io_uring_release(struct inode *inode, struct file *file)
8135 struct io_ring_ctx *ctx = file->private_data;
8137 file->private_data = NULL;
8138 io_ring_ctx_wait_and_kill(ctx);
8142 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8144 struct files_struct *files = data;
8146 return !files || work->files == files;
8150 * Returns true if 'preq' is the link parent of 'req'
8152 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8154 struct io_kiocb *link;
8156 if (!(preq->flags & REQ_F_LINK_HEAD))
8159 list_for_each_entry(link, &preq->link_list, link_list) {
8167 static bool io_match_link_files(struct io_kiocb *req,
8168 struct files_struct *files)
8170 struct io_kiocb *link;
8172 if (io_match_files(req, files))
8174 if (req->flags & REQ_F_LINK_HEAD) {
8175 list_for_each_entry(link, &req->link_list, link_list) {
8176 if (io_match_files(link, files))
8184 * We're looking to cancel 'req' because it's holding on to our files, but
8185 * 'req' could be a link to another request. See if it is, and cancel that
8186 * parent request if so.
8188 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8190 struct hlist_node *tmp;
8191 struct io_kiocb *preq;
8195 spin_lock_irq(&ctx->completion_lock);
8196 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8197 struct hlist_head *list;
8199 list = &ctx->cancel_hash[i];
8200 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8201 found = io_match_link(preq, req);
8203 io_poll_remove_one(preq);
8208 spin_unlock_irq(&ctx->completion_lock);
8212 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8213 struct io_kiocb *req)
8215 struct io_kiocb *preq;
8218 spin_lock_irq(&ctx->completion_lock);
8219 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8220 found = io_match_link(preq, req);
8222 __io_timeout_cancel(preq);
8226 spin_unlock_irq(&ctx->completion_lock);
8230 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8232 return io_match_link(container_of(work, struct io_kiocb, work), data);
8235 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8237 enum io_wq_cancel cret;
8239 /* cancel this particular work, if it's running */
8240 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8241 if (cret != IO_WQ_CANCEL_NOTFOUND)
8244 /* find links that hold this pending, cancel those */
8245 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8246 if (cret != IO_WQ_CANCEL_NOTFOUND)
8249 /* if we have a poll link holding this pending, cancel that */
8250 if (io_poll_remove_link(ctx, req))
8253 /* final option, timeout link is holding this req pending */
8254 io_timeout_remove_link(ctx, req);
8257 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8258 struct files_struct *files)
8260 struct io_defer_entry *de = NULL;
8263 spin_lock_irq(&ctx->completion_lock);
8264 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8265 if (io_match_link_files(de->req, files)) {
8266 list_cut_position(&list, &ctx->defer_list, &de->list);
8270 spin_unlock_irq(&ctx->completion_lock);
8272 while (!list_empty(&list)) {
8273 de = list_first_entry(&list, struct io_defer_entry, list);
8274 list_del_init(&de->list);
8275 req_set_fail_links(de->req);
8276 io_put_req(de->req);
8277 io_req_complete(de->req, -ECANCELED);
8283 * Returns true if we found and killed one or more files pinning requests
8285 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8286 struct files_struct *files)
8288 if (list_empty_careful(&ctx->inflight_list))
8291 io_cancel_defer_files(ctx, files);
8292 /* cancel all at once, should be faster than doing it one by one*/
8293 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8295 while (!list_empty_careful(&ctx->inflight_list)) {
8296 struct io_kiocb *cancel_req = NULL, *req;
8299 spin_lock_irq(&ctx->inflight_lock);
8300 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8301 if (files && req->work.files != files)
8303 /* req is being completed, ignore */
8304 if (!refcount_inc_not_zero(&req->refs))
8310 prepare_to_wait(&ctx->inflight_wait, &wait,
8311 TASK_UNINTERRUPTIBLE);
8312 spin_unlock_irq(&ctx->inflight_lock);
8314 /* We need to keep going until we don't find a matching req */
8317 /* cancel this request, or head link requests */
8318 io_attempt_cancel(ctx, cancel_req);
8319 io_put_req(cancel_req);
8320 /* cancellations _may_ trigger task work */
8323 finish_wait(&ctx->inflight_wait, &wait);
8329 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8331 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8332 struct task_struct *task = data;
8334 return io_task_match(req, task);
8337 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8338 struct task_struct *task,
8339 struct files_struct *files)
8343 ret = io_uring_cancel_files(ctx, files);
8345 enum io_wq_cancel cret;
8347 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8348 if (cret != IO_WQ_CANCEL_NOTFOUND)
8351 /* SQPOLL thread does its own polling */
8352 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8353 while (!list_empty_careful(&ctx->iopoll_list)) {
8354 io_iopoll_try_reap_events(ctx);
8359 ret |= io_poll_remove_all(ctx, task);
8360 ret |= io_kill_timeouts(ctx, task);
8367 * We need to iteratively cancel requests, in case a request has dependent
8368 * hard links. These persist even for failure of cancelations, hence keep
8369 * looping until none are found.
8371 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8372 struct files_struct *files)
8374 struct task_struct *task = current;
8376 if (ctx->flags & IORING_SETUP_SQPOLL)
8377 task = ctx->sqo_thread;
8379 io_cqring_overflow_flush(ctx, true, task, files);
8381 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8388 * Note that this task has used io_uring. We use it for cancelation purposes.
8390 static int io_uring_add_task_file(struct file *file)
8392 if (unlikely(!current->io_uring)) {
8395 ret = io_uring_alloc_task_context(current);
8399 if (current->io_uring->last != file) {
8400 XA_STATE(xas, ¤t->io_uring->xa, (unsigned long) file);
8404 old = xas_load(&xas);
8408 xas_store(&xas, file);
8412 current->io_uring->last = file;
8419 * Remove this io_uring_file -> task mapping.
8421 static void io_uring_del_task_file(struct file *file)
8423 struct io_uring_task *tctx = current->io_uring;
8424 XA_STATE(xas, &tctx->xa, (unsigned long) file);
8426 if (tctx->last == file)
8430 file = xas_store(&xas, NULL);
8437 static void __io_uring_attempt_task_drop(struct file *file)
8439 XA_STATE(xas, ¤t->io_uring->xa, (unsigned long) file);
8443 old = xas_load(&xas);
8447 io_uring_del_task_file(file);
8451 * Drop task note for this file if we're the only ones that hold it after
8454 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8456 if (!current->io_uring)
8459 * fput() is pending, will be 2 if the only other ref is our potential
8460 * task file note. If the task is exiting, drop regardless of count.
8462 if (!exiting && atomic_long_read(&file->f_count) != 2)
8465 __io_uring_attempt_task_drop(file);
8468 void __io_uring_files_cancel(struct files_struct *files)
8470 struct io_uring_task *tctx = current->io_uring;
8471 XA_STATE(xas, &tctx->xa, 0);
8473 /* make sure overflow events are dropped */
8474 tctx->in_idle = true;
8477 struct io_ring_ctx *ctx;
8481 file = xas_next_entry(&xas, ULONG_MAX);
8487 ctx = file->private_data;
8489 io_uring_cancel_task_requests(ctx, files);
8491 io_uring_del_task_file(file);
8495 static inline bool io_uring_task_idle(struct io_uring_task *tctx)
8497 return atomic_long_read(&tctx->req_issue) ==
8498 atomic_long_read(&tctx->req_complete);
8502 * Find any io_uring fd that this task has registered or done IO on, and cancel
8505 void __io_uring_task_cancel(void)
8507 struct io_uring_task *tctx = current->io_uring;
8511 /* make sure overflow events are dropped */
8512 tctx->in_idle = true;
8514 while (!io_uring_task_idle(tctx)) {
8515 /* read completions before cancelations */
8516 completions = atomic_long_read(&tctx->req_complete);
8517 __io_uring_files_cancel(NULL);
8519 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8522 * If we've seen completions, retry. This avoids a race where
8523 * a completion comes in before we did prepare_to_wait().
8525 if (completions != atomic_long_read(&tctx->req_complete))
8527 if (io_uring_task_idle(tctx))
8532 finish_wait(&tctx->wait, &wait);
8533 tctx->in_idle = false;
8536 static int io_uring_flush(struct file *file, void *data)
8538 struct io_ring_ctx *ctx = file->private_data;
8541 * If the task is going away, cancel work it may have pending
8543 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8546 io_uring_cancel_task_requests(ctx, data);
8547 io_uring_attempt_task_drop(file, !data);
8551 static void *io_uring_validate_mmap_request(struct file *file,
8552 loff_t pgoff, size_t sz)
8554 struct io_ring_ctx *ctx = file->private_data;
8555 loff_t offset = pgoff << PAGE_SHIFT;
8560 case IORING_OFF_SQ_RING:
8561 case IORING_OFF_CQ_RING:
8564 case IORING_OFF_SQES:
8568 return ERR_PTR(-EINVAL);
8571 page = virt_to_head_page(ptr);
8572 if (sz > page_size(page))
8573 return ERR_PTR(-EINVAL);
8580 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8582 size_t sz = vma->vm_end - vma->vm_start;
8586 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8588 return PTR_ERR(ptr);
8590 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8591 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8594 #else /* !CONFIG_MMU */
8596 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8598 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8601 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8603 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8606 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8607 unsigned long addr, unsigned long len,
8608 unsigned long pgoff, unsigned long flags)
8612 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8614 return PTR_ERR(ptr);
8616 return (unsigned long) ptr;
8619 #endif /* !CONFIG_MMU */
8621 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8622 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8625 struct io_ring_ctx *ctx;
8632 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8640 if (f.file->f_op != &io_uring_fops)
8644 ctx = f.file->private_data;
8645 if (!percpu_ref_tryget(&ctx->refs))
8649 if (ctx->flags & IORING_SETUP_R_DISABLED)
8653 * For SQ polling, the thread will do all submissions and completions.
8654 * Just return the requested submit count, and wake the thread if
8658 if (ctx->flags & IORING_SETUP_SQPOLL) {
8659 if (!list_empty_careful(&ctx->cq_overflow_list))
8660 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8661 if (flags & IORING_ENTER_SQ_WAKEUP)
8662 wake_up(&ctx->sqo_wait);
8663 submitted = to_submit;
8664 } else if (to_submit) {
8665 ret = io_uring_add_task_file(f.file);
8668 mutex_lock(&ctx->uring_lock);
8669 submitted = io_submit_sqes(ctx, to_submit);
8670 mutex_unlock(&ctx->uring_lock);
8672 if (submitted != to_submit)
8675 if (flags & IORING_ENTER_GETEVENTS) {
8676 min_complete = min(min_complete, ctx->cq_entries);
8679 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8680 * space applications don't need to do io completion events
8681 * polling again, they can rely on io_sq_thread to do polling
8682 * work, which can reduce cpu usage and uring_lock contention.
8684 if (ctx->flags & IORING_SETUP_IOPOLL &&
8685 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8686 ret = io_iopoll_check(ctx, min_complete);
8688 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8693 percpu_ref_put(&ctx->refs);
8696 return submitted ? submitted : ret;
8699 #ifdef CONFIG_PROC_FS
8700 static int io_uring_show_cred(int id, void *p, void *data)
8702 const struct cred *cred = p;
8703 struct seq_file *m = data;
8704 struct user_namespace *uns = seq_user_ns(m);
8705 struct group_info *gi;
8710 seq_printf(m, "%5d\n", id);
8711 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8712 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8713 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8714 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8715 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8716 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8717 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8718 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8719 seq_puts(m, "\n\tGroups:\t");
8720 gi = cred->group_info;
8721 for (g = 0; g < gi->ngroups; g++) {
8722 seq_put_decimal_ull(m, g ? " " : "",
8723 from_kgid_munged(uns, gi->gid[g]));
8725 seq_puts(m, "\n\tCapEff:\t");
8726 cap = cred->cap_effective;
8727 CAP_FOR_EACH_U32(__capi)
8728 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8733 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8739 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8740 * since fdinfo case grabs it in the opposite direction of normal use
8741 * cases. If we fail to get the lock, we just don't iterate any
8742 * structures that could be going away outside the io_uring mutex.
8744 has_lock = mutex_trylock(&ctx->uring_lock);
8746 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8747 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8748 struct fixed_file_table *table;
8751 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8752 f = table->files[i & IORING_FILE_TABLE_MASK];
8754 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8756 seq_printf(m, "%5u: <none>\n", i);
8758 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8759 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8760 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8762 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8763 (unsigned int) buf->len);
8765 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8766 seq_printf(m, "Personalities:\n");
8767 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8769 seq_printf(m, "PollList:\n");
8770 spin_lock_irq(&ctx->completion_lock);
8771 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8772 struct hlist_head *list = &ctx->cancel_hash[i];
8773 struct io_kiocb *req;
8775 hlist_for_each_entry(req, list, hash_node)
8776 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8777 req->task->task_works != NULL);
8779 spin_unlock_irq(&ctx->completion_lock);
8781 mutex_unlock(&ctx->uring_lock);
8784 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8786 struct io_ring_ctx *ctx = f->private_data;
8788 if (percpu_ref_tryget(&ctx->refs)) {
8789 __io_uring_show_fdinfo(ctx, m);
8790 percpu_ref_put(&ctx->refs);
8795 static const struct file_operations io_uring_fops = {
8796 .release = io_uring_release,
8797 .flush = io_uring_flush,
8798 .mmap = io_uring_mmap,
8800 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8801 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8803 .poll = io_uring_poll,
8804 .fasync = io_uring_fasync,
8805 #ifdef CONFIG_PROC_FS
8806 .show_fdinfo = io_uring_show_fdinfo,
8810 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8811 struct io_uring_params *p)
8813 struct io_rings *rings;
8814 size_t size, sq_array_offset;
8816 /* make sure these are sane, as we already accounted them */
8817 ctx->sq_entries = p->sq_entries;
8818 ctx->cq_entries = p->cq_entries;
8820 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8821 if (size == SIZE_MAX)
8824 rings = io_mem_alloc(size);
8829 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8830 rings->sq_ring_mask = p->sq_entries - 1;
8831 rings->cq_ring_mask = p->cq_entries - 1;
8832 rings->sq_ring_entries = p->sq_entries;
8833 rings->cq_ring_entries = p->cq_entries;
8834 ctx->sq_mask = rings->sq_ring_mask;
8835 ctx->cq_mask = rings->cq_ring_mask;
8837 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8838 if (size == SIZE_MAX) {
8839 io_mem_free(ctx->rings);
8844 ctx->sq_sqes = io_mem_alloc(size);
8845 if (!ctx->sq_sqes) {
8846 io_mem_free(ctx->rings);
8855 * Allocate an anonymous fd, this is what constitutes the application
8856 * visible backing of an io_uring instance. The application mmaps this
8857 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8858 * we have to tie this fd to a socket for file garbage collection purposes.
8860 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8865 #if defined(CONFIG_UNIX)
8866 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8872 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8876 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8877 O_RDWR | O_CLOEXEC);
8881 ret = PTR_ERR(file);
8885 #if defined(CONFIG_UNIX)
8886 ctx->ring_sock->file = file;
8888 if (unlikely(io_uring_add_task_file(file))) {
8889 file = ERR_PTR(-ENOMEM);
8892 fd_install(ret, file);
8895 #if defined(CONFIG_UNIX)
8896 sock_release(ctx->ring_sock);
8897 ctx->ring_sock = NULL;
8902 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8903 struct io_uring_params __user *params)
8905 struct user_struct *user = NULL;
8906 struct io_ring_ctx *ctx;
8912 if (entries > IORING_MAX_ENTRIES) {
8913 if (!(p->flags & IORING_SETUP_CLAMP))
8915 entries = IORING_MAX_ENTRIES;
8919 * Use twice as many entries for the CQ ring. It's possible for the
8920 * application to drive a higher depth than the size of the SQ ring,
8921 * since the sqes are only used at submission time. This allows for
8922 * some flexibility in overcommitting a bit. If the application has
8923 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8924 * of CQ ring entries manually.
8926 p->sq_entries = roundup_pow_of_two(entries);
8927 if (p->flags & IORING_SETUP_CQSIZE) {
8929 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8930 * to a power-of-two, if it isn't already. We do NOT impose
8931 * any cq vs sq ring sizing.
8933 if (p->cq_entries < p->sq_entries)
8935 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8936 if (!(p->flags & IORING_SETUP_CLAMP))
8938 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8940 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8942 p->cq_entries = 2 * p->sq_entries;
8945 user = get_uid(current_user());
8946 limit_mem = !capable(CAP_IPC_LOCK);
8949 ret = __io_account_mem(user,
8950 ring_pages(p->sq_entries, p->cq_entries));
8957 ctx = io_ring_ctx_alloc(p);
8960 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8965 ctx->compat = in_compat_syscall();
8967 ctx->creds = get_current_cred();
8969 ctx->sqo_task = get_task_struct(current);
8972 * This is just grabbed for accounting purposes. When a process exits,
8973 * the mm is exited and dropped before the files, hence we need to hang
8974 * on to this mm purely for the purposes of being able to unaccount
8975 * memory (locked/pinned vm). It's not used for anything else.
8977 mmgrab(current->mm);
8978 ctx->mm_account = current->mm;
8981 * Account memory _before_ installing the file descriptor. Once
8982 * the descriptor is installed, it can get closed at any time. Also
8983 * do this before hitting the general error path, as ring freeing
8984 * will un-account as well.
8986 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8988 ctx->limit_mem = limit_mem;
8990 ret = io_allocate_scq_urings(ctx, p);
8994 ret = io_sq_offload_create(ctx, p);
8998 if (!(p->flags & IORING_SETUP_R_DISABLED))
8999 io_sq_offload_start(ctx);
9001 memset(&p->sq_off, 0, sizeof(p->sq_off));
9002 p->sq_off.head = offsetof(struct io_rings, sq.head);
9003 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9004 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9005 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9006 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9007 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9008 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9010 memset(&p->cq_off, 0, sizeof(p->cq_off));
9011 p->cq_off.head = offsetof(struct io_rings, cq.head);
9012 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9013 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9014 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9015 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9016 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9017 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9019 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9020 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9021 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9022 IORING_FEAT_POLL_32BITS;
9024 if (copy_to_user(params, p, sizeof(*p))) {
9030 * Install ring fd as the very last thing, so we don't risk someone
9031 * having closed it before we finish setup
9033 ret = io_uring_get_fd(ctx);
9037 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9040 io_ring_ctx_wait_and_kill(ctx);
9045 * Sets up an aio uring context, and returns the fd. Applications asks for a
9046 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9047 * params structure passed in.
9049 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9051 struct io_uring_params p;
9054 if (copy_from_user(&p, params, sizeof(p)))
9056 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9061 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9062 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9063 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9064 IORING_SETUP_R_DISABLED))
9067 return io_uring_create(entries, &p, params);
9070 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9071 struct io_uring_params __user *, params)
9073 return io_uring_setup(entries, params);
9076 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9078 struct io_uring_probe *p;
9082 size = struct_size(p, ops, nr_args);
9083 if (size == SIZE_MAX)
9085 p = kzalloc(size, GFP_KERNEL);
9090 if (copy_from_user(p, arg, size))
9093 if (memchr_inv(p, 0, size))
9096 p->last_op = IORING_OP_LAST - 1;
9097 if (nr_args > IORING_OP_LAST)
9098 nr_args = IORING_OP_LAST;
9100 for (i = 0; i < nr_args; i++) {
9102 if (!io_op_defs[i].not_supported)
9103 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9108 if (copy_to_user(arg, p, size))
9115 static int io_register_personality(struct io_ring_ctx *ctx)
9117 const struct cred *creds = get_current_cred();
9120 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9121 USHRT_MAX, GFP_KERNEL);
9127 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9129 const struct cred *old_creds;
9131 old_creds = idr_remove(&ctx->personality_idr, id);
9133 put_cred(old_creds);
9140 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9141 unsigned int nr_args)
9143 struct io_uring_restriction *res;
9147 /* Restrictions allowed only if rings started disabled */
9148 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9151 /* We allow only a single restrictions registration */
9152 if (ctx->restrictions.registered)
9155 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9158 size = array_size(nr_args, sizeof(*res));
9159 if (size == SIZE_MAX)
9162 res = memdup_user(arg, size);
9164 return PTR_ERR(res);
9168 for (i = 0; i < nr_args; i++) {
9169 switch (res[i].opcode) {
9170 case IORING_RESTRICTION_REGISTER_OP:
9171 if (res[i].register_op >= IORING_REGISTER_LAST) {
9176 __set_bit(res[i].register_op,
9177 ctx->restrictions.register_op);
9179 case IORING_RESTRICTION_SQE_OP:
9180 if (res[i].sqe_op >= IORING_OP_LAST) {
9185 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9187 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9188 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9190 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9191 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9200 /* Reset all restrictions if an error happened */
9202 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9204 ctx->restrictions.registered = true;
9210 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9212 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9215 if (ctx->restrictions.registered)
9216 ctx->restricted = 1;
9218 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9220 io_sq_offload_start(ctx);
9225 static bool io_register_op_must_quiesce(int op)
9228 case IORING_UNREGISTER_FILES:
9229 case IORING_REGISTER_FILES_UPDATE:
9230 case IORING_REGISTER_PROBE:
9231 case IORING_REGISTER_PERSONALITY:
9232 case IORING_UNREGISTER_PERSONALITY:
9239 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9240 void __user *arg, unsigned nr_args)
9241 __releases(ctx->uring_lock)
9242 __acquires(ctx->uring_lock)
9247 * We're inside the ring mutex, if the ref is already dying, then
9248 * someone else killed the ctx or is already going through
9249 * io_uring_register().
9251 if (percpu_ref_is_dying(&ctx->refs))
9254 if (io_register_op_must_quiesce(opcode)) {
9255 percpu_ref_kill(&ctx->refs);
9258 * Drop uring mutex before waiting for references to exit. If
9259 * another thread is currently inside io_uring_enter() it might
9260 * need to grab the uring_lock to make progress. If we hold it
9261 * here across the drain wait, then we can deadlock. It's safe
9262 * to drop the mutex here, since no new references will come in
9263 * after we've killed the percpu ref.
9265 mutex_unlock(&ctx->uring_lock);
9266 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9267 mutex_lock(&ctx->uring_lock);
9269 percpu_ref_resurrect(&ctx->refs);
9275 if (ctx->restricted) {
9276 if (opcode >= IORING_REGISTER_LAST) {
9281 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9288 case IORING_REGISTER_BUFFERS:
9289 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9291 case IORING_UNREGISTER_BUFFERS:
9295 ret = io_sqe_buffer_unregister(ctx);
9297 case IORING_REGISTER_FILES:
9298 ret = io_sqe_files_register(ctx, arg, nr_args);
9300 case IORING_UNREGISTER_FILES:
9304 ret = io_sqe_files_unregister(ctx);
9306 case IORING_REGISTER_FILES_UPDATE:
9307 ret = io_sqe_files_update(ctx, arg, nr_args);
9309 case IORING_REGISTER_EVENTFD:
9310 case IORING_REGISTER_EVENTFD_ASYNC:
9314 ret = io_eventfd_register(ctx, arg);
9317 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9318 ctx->eventfd_async = 1;
9320 ctx->eventfd_async = 0;
9322 case IORING_UNREGISTER_EVENTFD:
9326 ret = io_eventfd_unregister(ctx);
9328 case IORING_REGISTER_PROBE:
9330 if (!arg || nr_args > 256)
9332 ret = io_probe(ctx, arg, nr_args);
9334 case IORING_REGISTER_PERSONALITY:
9338 ret = io_register_personality(ctx);
9340 case IORING_UNREGISTER_PERSONALITY:
9344 ret = io_unregister_personality(ctx, nr_args);
9346 case IORING_REGISTER_ENABLE_RINGS:
9350 ret = io_register_enable_rings(ctx);
9352 case IORING_REGISTER_RESTRICTIONS:
9353 ret = io_register_restrictions(ctx, arg, nr_args);
9361 if (io_register_op_must_quiesce(opcode)) {
9362 /* bring the ctx back to life */
9363 percpu_ref_reinit(&ctx->refs);
9365 reinit_completion(&ctx->ref_comp);
9370 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9371 void __user *, arg, unsigned int, nr_args)
9373 struct io_ring_ctx *ctx;
9382 if (f.file->f_op != &io_uring_fops)
9385 ctx = f.file->private_data;
9387 mutex_lock(&ctx->uring_lock);
9388 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9389 mutex_unlock(&ctx->uring_lock);
9390 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9391 ctx->cq_ev_fd != NULL, ret);
9397 static int __init io_uring_init(void)
9399 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9400 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9401 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9404 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9405 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9406 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9407 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9408 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9409 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9410 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9411 BUILD_BUG_SQE_ELEM(8, __u64, off);
9412 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9413 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9414 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9415 BUILD_BUG_SQE_ELEM(24, __u32, len);
9416 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9417 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9418 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9419 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9420 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9421 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9422 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9423 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9424 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9425 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9426 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9427 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9428 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9429 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9430 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9431 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9432 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9433 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9434 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9436 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9437 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9438 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9441 __initcall(io_uring_init);