1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
84 #include <linux/audit.h>
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/io_uring.h>
89 #include <uapi/linux/io_uring.h>
94 #define IORING_MAX_ENTRIES 32768
95 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
98 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
100 #define IORING_FILE_TABLE_SHIFT 9
101 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
102 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
103 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 struct io_mapped_ubuf {
193 struct bio_vec *bvec;
194 unsigned int nr_bvecs;
195 unsigned long acct_pages;
198 struct fixed_file_table {
202 struct fixed_file_ref_node {
203 struct percpu_ref refs;
204 struct list_head node;
205 struct list_head file_list;
206 struct fixed_file_data *file_data;
207 struct llist_node llist;
211 struct fixed_file_data {
212 struct fixed_file_table *table;
213 struct io_ring_ctx *ctx;
215 struct fixed_file_ref_node *node;
216 struct percpu_ref refs;
217 struct completion done;
218 struct list_head ref_list;
223 struct list_head list;
229 struct io_restriction {
230 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
231 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
232 u8 sqe_flags_allowed;
233 u8 sqe_flags_required;
241 /* ctx's that are using this sqd */
242 struct list_head ctx_list;
243 struct list_head ctx_new_list;
244 struct mutex ctx_lock;
246 struct task_struct *thread;
247 struct wait_queue_head wait;
252 struct percpu_ref refs;
253 } ____cacheline_aligned_in_smp;
257 unsigned int compat: 1;
258 unsigned int limit_mem: 1;
259 unsigned int cq_overflow_flushed: 1;
260 unsigned int drain_next: 1;
261 unsigned int eventfd_async: 1;
262 unsigned int restricted: 1;
263 unsigned int sqo_dead: 1;
266 * Ring buffer of indices into array of io_uring_sqe, which is
267 * mmapped by the application using the IORING_OFF_SQES offset.
269 * This indirection could e.g. be used to assign fixed
270 * io_uring_sqe entries to operations and only submit them to
271 * the queue when needed.
273 * The kernel modifies neither the indices array nor the entries
277 unsigned cached_sq_head;
280 unsigned sq_thread_idle;
281 unsigned cached_sq_dropped;
282 unsigned cached_cq_overflow;
283 unsigned long sq_check_overflow;
285 struct list_head defer_list;
286 struct list_head timeout_list;
287 struct list_head cq_overflow_list;
289 struct io_uring_sqe *sq_sqes;
290 } ____cacheline_aligned_in_smp;
292 struct io_rings *rings;
298 * For SQPOLL usage - we hold a reference to the parent task, so we
299 * have access to the ->files
301 struct task_struct *sqo_task;
303 /* Only used for accounting purposes */
304 struct mm_struct *mm_account;
306 #ifdef CONFIG_BLK_CGROUP
307 struct cgroup_subsys_state *sqo_blkcg_css;
310 struct io_sq_data *sq_data; /* if using sq thread polling */
312 struct wait_queue_head sqo_sq_wait;
313 struct wait_queue_entry sqo_wait_entry;
314 struct list_head sqd_list;
317 * If used, fixed file set. Writers must ensure that ->refs is dead,
318 * readers must ensure that ->refs is alive as long as the file* is
319 * used. Only updated through io_uring_register(2).
321 struct fixed_file_data *file_data;
322 unsigned nr_user_files;
324 /* if used, fixed mapped user buffers */
325 unsigned nr_user_bufs;
326 struct io_mapped_ubuf *user_bufs;
328 struct user_struct *user;
330 const struct cred *creds;
334 unsigned int sessionid;
337 struct completion ref_comp;
338 struct completion sq_thread_comp;
340 /* if all else fails... */
341 struct io_kiocb *fallback_req;
343 #if defined(CONFIG_UNIX)
344 struct socket *ring_sock;
347 struct xarray io_buffers;
349 struct xarray personalities;
353 unsigned cached_cq_tail;
356 atomic_t cq_timeouts;
357 unsigned cq_last_tm_flush;
358 unsigned long cq_check_overflow;
359 struct wait_queue_head cq_wait;
360 struct fasync_struct *cq_fasync;
361 struct eventfd_ctx *cq_ev_fd;
362 } ____cacheline_aligned_in_smp;
365 struct mutex uring_lock;
366 wait_queue_head_t wait;
367 } ____cacheline_aligned_in_smp;
370 spinlock_t completion_lock;
373 * ->iopoll_list is protected by the ctx->uring_lock for
374 * io_uring instances that don't use IORING_SETUP_SQPOLL.
375 * For SQPOLL, only the single threaded io_sq_thread() will
376 * manipulate the list, hence no extra locking is needed there.
378 struct list_head iopoll_list;
379 struct hlist_head *cancel_hash;
380 unsigned cancel_hash_bits;
381 bool poll_multi_file;
383 spinlock_t inflight_lock;
384 struct list_head inflight_list;
385 } ____cacheline_aligned_in_smp;
387 struct delayed_work file_put_work;
388 struct llist_head file_put_llist;
390 struct work_struct exit_work;
391 struct io_restriction restrictions;
395 * First field must be the file pointer in all the
396 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
398 struct io_poll_iocb {
401 struct wait_queue_head *head;
407 struct wait_queue_entry wait;
412 struct file *put_file;
416 struct io_timeout_data {
417 struct io_kiocb *req;
418 struct hrtimer timer;
419 struct timespec64 ts;
420 enum hrtimer_mode mode;
425 struct sockaddr __user *addr;
426 int __user *addr_len;
428 unsigned long nofile;
448 struct list_head list;
451 struct io_timeout_rem {
457 /* NOTE: kiocb has the file as the first member, so don't do it here */
465 struct sockaddr __user *addr;
472 struct user_msghdr __user *umsg;
478 struct io_buffer *kbuf;
484 bool ignore_nonblock;
485 struct filename *filename;
487 unsigned long nofile;
490 struct io_files_update {
516 struct epoll_event event;
520 struct file *file_out;
521 struct file *file_in;
528 struct io_provide_buf {
542 const char __user *filename;
543 struct statx __user *buffer;
546 struct io_completion {
548 struct list_head list;
552 struct io_async_connect {
553 struct sockaddr_storage address;
556 struct io_async_msghdr {
557 struct iovec fast_iov[UIO_FASTIOV];
559 struct sockaddr __user *uaddr;
561 struct sockaddr_storage addr;
565 struct iovec fast_iov[UIO_FASTIOV];
566 const struct iovec *free_iovec;
567 struct iov_iter iter;
569 struct wait_page_queue wpq;
573 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
574 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
575 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
576 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
577 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
578 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
585 REQ_F_LINK_TIMEOUT_BIT,
587 REQ_F_NEED_CLEANUP_BIT,
589 REQ_F_BUFFER_SELECTED_BIT,
590 REQ_F_NO_FILE_TABLE_BIT,
591 REQ_F_WORK_INITIALIZED_BIT,
592 REQ_F_LTIMEOUT_ACTIVE_BIT,
594 /* not a real bit, just to check we're not overflowing the space */
600 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
601 /* drain existing IO first */
602 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
604 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
605 /* doesn't sever on completion < 0 */
606 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
608 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
609 /* IOSQE_BUFFER_SELECT */
610 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
613 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
614 /* fail rest of links */
615 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
616 /* on inflight list */
617 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
618 /* read/write uses file position */
619 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
620 /* must not punt to workers */
621 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
622 /* has or had linked timeout */
623 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
625 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
627 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
628 /* already went through poll handler */
629 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
630 /* buffer already selected */
631 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
632 /* doesn't need file table for this request */
633 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
634 /* io_wq_work is initialized */
635 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
636 /* linked timeout is active, i.e. prepared by link's head */
637 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
641 struct io_poll_iocb poll;
642 struct io_poll_iocb *double_poll;
646 * NOTE! Each of the iocb union members has the file pointer
647 * as the first entry in their struct definition. So you can
648 * access the file pointer through any of the sub-structs,
649 * or directly as just 'ki_filp' in this struct.
655 struct io_poll_iocb poll;
656 struct io_accept accept;
658 struct io_cancel cancel;
659 struct io_timeout timeout;
660 struct io_timeout_rem timeout_rem;
661 struct io_connect connect;
662 struct io_sr_msg sr_msg;
664 struct io_close close;
665 struct io_files_update files_update;
666 struct io_fadvise fadvise;
667 struct io_madvise madvise;
668 struct io_epoll epoll;
669 struct io_splice splice;
670 struct io_provide_buf pbuf;
671 struct io_statx statx;
672 /* use only after cleaning per-op data, see io_clean_op() */
673 struct io_completion compl;
676 /* opcode allocated if it needs to store data for async defer */
679 /* polled IO has completed */
685 struct io_ring_ctx *ctx;
688 struct task_struct *task;
691 struct list_head link_list;
694 * 1. used with ctx->iopoll_list with reads/writes
695 * 2. to track reqs with ->files (see io_op_def::file_table)
697 struct list_head inflight_entry;
699 struct percpu_ref *fixed_file_refs;
700 struct callback_head task_work;
701 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
702 struct hlist_node hash_node;
703 struct async_poll *apoll;
704 struct io_wq_work work;
707 struct io_defer_entry {
708 struct list_head list;
709 struct io_kiocb *req;
713 #define IO_IOPOLL_BATCH 8
715 struct io_comp_state {
717 struct list_head list;
718 struct io_ring_ctx *ctx;
721 struct io_submit_state {
722 struct blk_plug plug;
725 * io_kiocb alloc cache
727 void *reqs[IO_IOPOLL_BATCH];
728 unsigned int free_reqs;
731 * Batch completion logic
733 struct io_comp_state comp;
736 * File reference cache
740 unsigned int has_refs;
741 unsigned int ios_left;
745 /* needs req->file assigned */
746 unsigned needs_file : 1;
747 /* don't fail if file grab fails */
748 unsigned needs_file_no_error : 1;
749 /* hash wq insertion if file is a regular file */
750 unsigned hash_reg_file : 1;
751 /* unbound wq insertion if file is a non-regular file */
752 unsigned unbound_nonreg_file : 1;
753 /* opcode is not supported by this kernel */
754 unsigned not_supported : 1;
755 /* set if opcode supports polled "wait" */
757 unsigned pollout : 1;
758 /* op supports buffer selection */
759 unsigned buffer_select : 1;
760 /* must always have async data allocated */
761 unsigned needs_async_data : 1;
762 /* size of async data needed, if any */
763 unsigned short async_size;
767 static const struct io_op_def io_op_defs[] = {
768 [IORING_OP_NOP] = {},
769 [IORING_OP_READV] = {
771 .unbound_nonreg_file = 1,
774 .needs_async_data = 1,
775 .async_size = sizeof(struct io_async_rw),
776 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
778 [IORING_OP_WRITEV] = {
781 .unbound_nonreg_file = 1,
783 .needs_async_data = 1,
784 .async_size = sizeof(struct io_async_rw),
785 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
788 [IORING_OP_FSYNC] = {
790 .work_flags = IO_WQ_WORK_BLKCG,
792 [IORING_OP_READ_FIXED] = {
794 .unbound_nonreg_file = 1,
796 .async_size = sizeof(struct io_async_rw),
797 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
799 [IORING_OP_WRITE_FIXED] = {
802 .unbound_nonreg_file = 1,
804 .async_size = sizeof(struct io_async_rw),
805 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
808 [IORING_OP_POLL_ADD] = {
810 .unbound_nonreg_file = 1,
812 [IORING_OP_POLL_REMOVE] = {},
813 [IORING_OP_SYNC_FILE_RANGE] = {
815 .work_flags = IO_WQ_WORK_BLKCG,
817 [IORING_OP_SENDMSG] = {
819 .unbound_nonreg_file = 1,
821 .needs_async_data = 1,
822 .async_size = sizeof(struct io_async_msghdr),
823 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
826 [IORING_OP_RECVMSG] = {
828 .unbound_nonreg_file = 1,
831 .needs_async_data = 1,
832 .async_size = sizeof(struct io_async_msghdr),
833 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
836 [IORING_OP_TIMEOUT] = {
837 .needs_async_data = 1,
838 .async_size = sizeof(struct io_timeout_data),
839 .work_flags = IO_WQ_WORK_MM,
841 [IORING_OP_TIMEOUT_REMOVE] = {},
842 [IORING_OP_ACCEPT] = {
844 .unbound_nonreg_file = 1,
846 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
848 [IORING_OP_ASYNC_CANCEL] = {},
849 [IORING_OP_LINK_TIMEOUT] = {
850 .needs_async_data = 1,
851 .async_size = sizeof(struct io_timeout_data),
852 .work_flags = IO_WQ_WORK_MM,
854 [IORING_OP_CONNECT] = {
856 .unbound_nonreg_file = 1,
858 .needs_async_data = 1,
859 .async_size = sizeof(struct io_async_connect),
860 .work_flags = IO_WQ_WORK_MM,
862 [IORING_OP_FALLOCATE] = {
864 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
866 [IORING_OP_OPENAT] = {
867 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
870 [IORING_OP_CLOSE] = {
872 .needs_file_no_error = 1,
873 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
875 [IORING_OP_FILES_UPDATE] = {
876 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
878 [IORING_OP_STATX] = {
879 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
880 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
884 .unbound_nonreg_file = 1,
887 .async_size = sizeof(struct io_async_rw),
888 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
890 [IORING_OP_WRITE] = {
893 .unbound_nonreg_file = 1,
895 .async_size = sizeof(struct io_async_rw),
896 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
899 [IORING_OP_FADVISE] = {
901 .work_flags = IO_WQ_WORK_BLKCG,
903 [IORING_OP_MADVISE] = {
904 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
908 .unbound_nonreg_file = 1,
910 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
914 .unbound_nonreg_file = 1,
917 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
919 [IORING_OP_OPENAT2] = {
920 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
923 [IORING_OP_EPOLL_CTL] = {
924 .unbound_nonreg_file = 1,
925 .work_flags = IO_WQ_WORK_FILES,
927 [IORING_OP_SPLICE] = {
930 .unbound_nonreg_file = 1,
931 .work_flags = IO_WQ_WORK_BLKCG,
933 [IORING_OP_PROVIDE_BUFFERS] = {},
934 [IORING_OP_REMOVE_BUFFERS] = {},
938 .unbound_nonreg_file = 1,
942 enum io_mem_account {
947 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node);
948 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
949 struct io_ring_ctx *ctx);
951 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
952 struct io_comp_state *cs);
953 static void io_cqring_fill_event(struct io_kiocb *req, long res);
954 static void io_put_req(struct io_kiocb *req);
955 static void io_put_req_deferred(struct io_kiocb *req, int nr);
956 static void io_double_put_req(struct io_kiocb *req);
957 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
958 static void __io_queue_linked_timeout(struct io_kiocb *req);
959 static void io_queue_linked_timeout(struct io_kiocb *req);
960 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
961 struct io_uring_files_update *ip,
963 static void __io_clean_op(struct io_kiocb *req);
964 static struct file *io_file_get(struct io_submit_state *state,
965 struct io_kiocb *req, int fd, bool fixed);
966 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
967 static void io_file_put_work(struct work_struct *work);
969 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
970 struct iovec **iovec, struct iov_iter *iter,
972 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
973 const struct iovec *fast_iov,
974 struct iov_iter *iter, bool force);
975 static void io_req_drop_files(struct io_kiocb *req);
976 static void io_req_task_queue(struct io_kiocb *req);
978 static struct kmem_cache *req_cachep;
980 static const struct file_operations io_uring_fops;
982 struct sock *io_uring_get_socket(struct file *file)
984 #if defined(CONFIG_UNIX)
985 if (file->f_op == &io_uring_fops) {
986 struct io_ring_ctx *ctx = file->private_data;
988 return ctx->ring_sock->sk;
993 EXPORT_SYMBOL(io_uring_get_socket);
995 static inline void io_clean_op(struct io_kiocb *req)
997 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1001 static inline bool __io_match_files(struct io_kiocb *req,
1002 struct files_struct *files)
1004 if (req->file && req->file->f_op == &io_uring_fops)
1007 return ((req->flags & REQ_F_WORK_INITIALIZED) &&
1008 (req->work.flags & IO_WQ_WORK_FILES)) &&
1009 req->work.identity->files == files;
1012 static bool io_match_task(struct io_kiocb *head,
1013 struct task_struct *task,
1014 struct files_struct *files)
1016 struct io_kiocb *link;
1018 if (task && head->task != task) {
1019 /* in terms of cancelation, always match if req task is dead */
1020 if (head->task->flags & PF_EXITING)
1026 if (__io_match_files(head, files))
1028 if (head->flags & REQ_F_LINK_HEAD) {
1029 list_for_each_entry(link, &head->link_list, link_list) {
1030 if (__io_match_files(link, files))
1038 static void io_sq_thread_drop_mm(void)
1040 struct mm_struct *mm = current->mm;
1043 kthread_unuse_mm(mm);
1049 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1051 struct mm_struct *mm;
1053 if (current->flags & PF_EXITING)
1058 /* Should never happen */
1059 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL)))
1062 task_lock(ctx->sqo_task);
1063 mm = ctx->sqo_task->mm;
1064 if (unlikely(!mm || !mmget_not_zero(mm)))
1066 task_unlock(ctx->sqo_task);
1076 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1077 struct io_kiocb *req)
1079 if (!(io_op_defs[req->opcode].work_flags & IO_WQ_WORK_MM))
1081 return __io_sq_thread_acquire_mm(ctx);
1084 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1085 struct cgroup_subsys_state **cur_css)
1088 #ifdef CONFIG_BLK_CGROUP
1089 /* puts the old one when swapping */
1090 if (*cur_css != ctx->sqo_blkcg_css) {
1091 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1092 *cur_css = ctx->sqo_blkcg_css;
1097 static void io_sq_thread_unassociate_blkcg(void)
1099 #ifdef CONFIG_BLK_CGROUP
1100 kthread_associate_blkcg(NULL);
1104 static inline void req_set_fail_links(struct io_kiocb *req)
1106 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1107 req->flags |= REQ_F_FAIL_LINK;
1111 * None of these are dereferenced, they are simply used to check if any of
1112 * them have changed. If we're under current and check they are still the
1113 * same, we're fine to grab references to them for actual out-of-line use.
1115 static void io_init_identity(struct io_identity *id)
1117 id->files = current->files;
1118 id->mm = current->mm;
1119 #ifdef CONFIG_BLK_CGROUP
1121 id->blkcg_css = blkcg_css();
1124 id->creds = current_cred();
1125 id->nsproxy = current->nsproxy;
1126 id->fs = current->fs;
1127 id->fsize = rlimit(RLIMIT_FSIZE);
1129 id->loginuid = current->loginuid;
1130 id->sessionid = current->sessionid;
1132 refcount_set(&id->count, 1);
1135 static inline void __io_req_init_async(struct io_kiocb *req)
1137 memset(&req->work, 0, sizeof(req->work));
1138 req->flags |= REQ_F_WORK_INITIALIZED;
1142 * Note: must call io_req_init_async() for the first time you
1143 * touch any members of io_wq_work.
1145 static inline void io_req_init_async(struct io_kiocb *req)
1147 struct io_uring_task *tctx = current->io_uring;
1149 if (req->flags & REQ_F_WORK_INITIALIZED)
1152 __io_req_init_async(req);
1154 /* Grab a ref if this isn't our static identity */
1155 req->work.identity = tctx->identity;
1156 if (tctx->identity != &tctx->__identity)
1157 refcount_inc(&req->work.identity->count);
1160 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1162 return ctx->flags & IORING_SETUP_SQPOLL;
1165 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1167 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1169 complete(&ctx->ref_comp);
1172 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1174 return !req->timeout.off;
1177 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1179 struct io_ring_ctx *ctx;
1182 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1186 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1187 if (!ctx->fallback_req)
1191 * Use 5 bits less than the max cq entries, that should give us around
1192 * 32 entries per hash list if totally full and uniformly spread.
1194 hash_bits = ilog2(p->cq_entries);
1198 ctx->cancel_hash_bits = hash_bits;
1199 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1201 if (!ctx->cancel_hash)
1203 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1205 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1206 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1209 ctx->flags = p->flags;
1210 init_waitqueue_head(&ctx->sqo_sq_wait);
1211 INIT_LIST_HEAD(&ctx->sqd_list);
1212 init_waitqueue_head(&ctx->cq_wait);
1213 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1214 init_completion(&ctx->ref_comp);
1215 init_completion(&ctx->sq_thread_comp);
1216 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1217 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1218 mutex_init(&ctx->uring_lock);
1219 init_waitqueue_head(&ctx->wait);
1220 spin_lock_init(&ctx->completion_lock);
1221 INIT_LIST_HEAD(&ctx->iopoll_list);
1222 INIT_LIST_HEAD(&ctx->defer_list);
1223 INIT_LIST_HEAD(&ctx->timeout_list);
1224 spin_lock_init(&ctx->inflight_lock);
1225 INIT_LIST_HEAD(&ctx->inflight_list);
1226 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1227 init_llist_head(&ctx->file_put_llist);
1230 if (ctx->fallback_req)
1231 kmem_cache_free(req_cachep, ctx->fallback_req);
1232 kfree(ctx->cancel_hash);
1237 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1239 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1240 struct io_ring_ctx *ctx = req->ctx;
1242 return seq != ctx->cached_cq_tail
1243 + READ_ONCE(ctx->cached_cq_overflow);
1249 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1251 struct io_rings *rings = ctx->rings;
1253 /* order cqe stores with ring update */
1254 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1257 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1259 if (req->work.identity == &tctx->__identity)
1261 if (refcount_dec_and_test(&req->work.identity->count))
1262 kfree(req->work.identity);
1265 static void io_req_clean_work(struct io_kiocb *req)
1267 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1270 req->flags &= ~REQ_F_WORK_INITIALIZED;
1272 if (req->work.flags & IO_WQ_WORK_MM) {
1273 mmdrop(req->work.identity->mm);
1274 req->work.flags &= ~IO_WQ_WORK_MM;
1276 #ifdef CONFIG_BLK_CGROUP
1277 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1278 css_put(req->work.identity->blkcg_css);
1279 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1282 if (req->work.flags & IO_WQ_WORK_CREDS) {
1283 put_cred(req->work.identity->creds);
1284 req->work.flags &= ~IO_WQ_WORK_CREDS;
1286 if (req->work.flags & IO_WQ_WORK_FS) {
1287 struct fs_struct *fs = req->work.identity->fs;
1289 spin_lock(&req->work.identity->fs->lock);
1292 spin_unlock(&req->work.identity->fs->lock);
1295 req->work.flags &= ~IO_WQ_WORK_FS;
1297 if (req->flags & REQ_F_INFLIGHT)
1298 io_req_drop_files(req);
1300 io_put_identity(req->task->io_uring, req);
1304 * Create a private copy of io_identity, since some fields don't match
1305 * the current context.
1307 static bool io_identity_cow(struct io_kiocb *req)
1309 struct io_uring_task *tctx = current->io_uring;
1310 const struct cred *creds = NULL;
1311 struct io_identity *id;
1313 if (req->work.flags & IO_WQ_WORK_CREDS)
1314 creds = req->work.identity->creds;
1316 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1317 if (unlikely(!id)) {
1318 req->work.flags |= IO_WQ_WORK_CANCEL;
1323 * We can safely just re-init the creds we copied Either the field
1324 * matches the current one, or we haven't grabbed it yet. The only
1325 * exception is ->creds, through registered personalities, so handle
1326 * that one separately.
1328 io_init_identity(id);
1332 /* add one for this request */
1333 refcount_inc(&id->count);
1335 /* drop tctx and req identity references, if needed */
1336 if (tctx->identity != &tctx->__identity &&
1337 refcount_dec_and_test(&tctx->identity->count))
1338 kfree(tctx->identity);
1339 if (req->work.identity != &tctx->__identity &&
1340 refcount_dec_and_test(&req->work.identity->count))
1341 kfree(req->work.identity);
1343 req->work.identity = id;
1344 tctx->identity = id;
1348 static bool io_grab_identity(struct io_kiocb *req)
1350 const struct io_op_def *def = &io_op_defs[req->opcode];
1351 struct io_identity *id = req->work.identity;
1352 struct io_ring_ctx *ctx = req->ctx;
1354 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1355 if (id->fsize != rlimit(RLIMIT_FSIZE))
1357 req->work.flags |= IO_WQ_WORK_FSIZE;
1359 #ifdef CONFIG_BLK_CGROUP
1360 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1361 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1363 if (id->blkcg_css != blkcg_css()) {
1368 * This should be rare, either the cgroup is dying or the task
1369 * is moving cgroups. Just punt to root for the handful of ios.
1371 if (css_tryget_online(id->blkcg_css))
1372 req->work.flags |= IO_WQ_WORK_BLKCG;
1376 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1377 if (id->creds != current_cred())
1379 get_cred(id->creds);
1380 req->work.flags |= IO_WQ_WORK_CREDS;
1383 if (!uid_eq(current->loginuid, id->loginuid) ||
1384 current->sessionid != id->sessionid)
1387 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1388 (def->work_flags & IO_WQ_WORK_FS)) {
1389 if (current->fs != id->fs)
1391 spin_lock(&id->fs->lock);
1392 if (!id->fs->in_exec) {
1394 req->work.flags |= IO_WQ_WORK_FS;
1396 req->work.flags |= IO_WQ_WORK_CANCEL;
1398 spin_unlock(¤t->fs->lock);
1400 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1401 (def->work_flags & IO_WQ_WORK_FILES) &&
1402 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1403 if (id->files != current->files ||
1404 id->nsproxy != current->nsproxy)
1406 atomic_inc(&id->files->count);
1407 get_nsproxy(id->nsproxy);
1409 if (!(req->flags & REQ_F_INFLIGHT)) {
1410 req->flags |= REQ_F_INFLIGHT;
1412 spin_lock_irq(&ctx->inflight_lock);
1413 list_add(&req->inflight_entry, &ctx->inflight_list);
1414 spin_unlock_irq(&ctx->inflight_lock);
1416 req->work.flags |= IO_WQ_WORK_FILES;
1418 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1419 (def->work_flags & IO_WQ_WORK_MM)) {
1420 if (id->mm != current->mm)
1423 req->work.flags |= IO_WQ_WORK_MM;
1429 static void io_prep_async_work(struct io_kiocb *req)
1431 const struct io_op_def *def = &io_op_defs[req->opcode];
1432 struct io_ring_ctx *ctx = req->ctx;
1433 struct io_identity *id;
1435 io_req_init_async(req);
1436 id = req->work.identity;
1438 if (req->flags & REQ_F_FORCE_ASYNC)
1439 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1441 if (req->flags & REQ_F_ISREG) {
1442 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1443 io_wq_hash_work(&req->work, file_inode(req->file));
1444 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1445 if (def->unbound_nonreg_file)
1446 req->work.flags |= IO_WQ_WORK_UNBOUND;
1449 /* if we fail grabbing identity, we must COW, regrab, and retry */
1450 if (io_grab_identity(req))
1453 if (!io_identity_cow(req))
1456 /* can't fail at this point */
1457 if (!io_grab_identity(req))
1461 static void io_prep_async_link(struct io_kiocb *req)
1463 struct io_kiocb *cur;
1465 io_prep_async_work(req);
1466 if (req->flags & REQ_F_LINK_HEAD)
1467 list_for_each_entry(cur, &req->link_list, link_list)
1468 io_prep_async_work(cur);
1471 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1473 struct io_ring_ctx *ctx = req->ctx;
1474 struct io_kiocb *link = io_prep_linked_timeout(req);
1476 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1477 &req->work, req->flags);
1478 io_wq_enqueue(ctx->io_wq, &req->work);
1482 static void io_queue_async_work(struct io_kiocb *req)
1484 struct io_kiocb *link;
1486 /* init ->work of the whole link before punting */
1487 io_prep_async_link(req);
1488 link = __io_queue_async_work(req);
1491 io_queue_linked_timeout(link);
1494 static void io_kill_timeout(struct io_kiocb *req, int status)
1496 struct io_timeout_data *io = req->async_data;
1499 ret = hrtimer_try_to_cancel(&io->timer);
1502 req_set_fail_links(req);
1503 atomic_set(&req->ctx->cq_timeouts,
1504 atomic_read(&req->ctx->cq_timeouts) + 1);
1505 list_del_init(&req->timeout.list);
1506 io_cqring_fill_event(req, status);
1507 io_put_req_deferred(req, 1);
1512 * Returns true if we found and killed one or more timeouts
1514 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1515 struct files_struct *files)
1517 struct io_kiocb *req, *tmp;
1520 spin_lock_irq(&ctx->completion_lock);
1521 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1522 if (io_match_task(req, tsk, files)) {
1523 io_kill_timeout(req, -ECANCELED);
1527 spin_unlock_irq(&ctx->completion_lock);
1528 return canceled != 0;
1531 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1534 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1535 struct io_defer_entry, list);
1537 if (req_need_defer(de->req, de->seq))
1539 list_del_init(&de->list);
1540 io_req_task_queue(de->req);
1542 } while (!list_empty(&ctx->defer_list));
1545 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1549 if (list_empty(&ctx->timeout_list))
1552 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1555 u32 events_needed, events_got;
1556 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1557 struct io_kiocb, timeout.list);
1559 if (io_is_timeout_noseq(req))
1563 * Since seq can easily wrap around over time, subtract
1564 * the last seq at which timeouts were flushed before comparing.
1565 * Assuming not more than 2^31-1 events have happened since,
1566 * these subtractions won't have wrapped, so we can check if
1567 * target is in [last_seq, current_seq] by comparing the two.
1569 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1570 events_got = seq - ctx->cq_last_tm_flush;
1571 if (events_got < events_needed)
1574 list_del_init(&req->timeout.list);
1575 io_kill_timeout(req, 0);
1576 } while (!list_empty(&ctx->timeout_list));
1578 ctx->cq_last_tm_flush = seq;
1581 static void io_commit_cqring(struct io_ring_ctx *ctx)
1583 io_flush_timeouts(ctx);
1584 __io_commit_cqring(ctx);
1586 if (unlikely(!list_empty(&ctx->defer_list)))
1587 __io_queue_deferred(ctx);
1590 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1592 struct io_rings *r = ctx->rings;
1594 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1597 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1599 struct io_rings *rings = ctx->rings;
1602 tail = ctx->cached_cq_tail;
1604 * writes to the cq entry need to come after reading head; the
1605 * control dependency is enough as we're using WRITE_ONCE to
1608 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1611 ctx->cached_cq_tail++;
1612 return &rings->cqes[tail & ctx->cq_mask];
1615 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1619 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1621 if (!ctx->eventfd_async)
1623 return io_wq_current_is_worker();
1626 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1628 if (wq_has_sleeper(&ctx->cq_wait)) {
1629 wake_up_interruptible(&ctx->cq_wait);
1630 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1632 if (waitqueue_active(&ctx->wait))
1633 wake_up(&ctx->wait);
1634 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1635 wake_up(&ctx->sq_data->wait);
1636 if (io_should_trigger_evfd(ctx))
1637 eventfd_signal(ctx->cq_ev_fd, 1);
1640 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1642 if (list_empty(&ctx->cq_overflow_list)) {
1643 clear_bit(0, &ctx->sq_check_overflow);
1644 clear_bit(0, &ctx->cq_check_overflow);
1645 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1649 /* Returns true if there are no backlogged entries after the flush */
1650 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1651 struct task_struct *tsk,
1652 struct files_struct *files)
1654 struct io_rings *rings = ctx->rings;
1655 struct io_kiocb *req, *tmp;
1656 struct io_uring_cqe *cqe;
1657 unsigned long flags;
1661 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1662 rings->cq_ring_entries))
1666 spin_lock_irqsave(&ctx->completion_lock, flags);
1669 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1670 if (!io_match_task(req, tsk, files))
1673 cqe = io_get_cqring(ctx);
1677 list_move(&req->compl.list, &list);
1679 WRITE_ONCE(cqe->user_data, req->user_data);
1680 WRITE_ONCE(cqe->res, req->result);
1681 WRITE_ONCE(cqe->flags, req->compl.cflags);
1683 ctx->cached_cq_overflow++;
1684 WRITE_ONCE(ctx->rings->cq_overflow,
1685 ctx->cached_cq_overflow);
1689 io_commit_cqring(ctx);
1690 io_cqring_mark_overflow(ctx);
1692 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1693 io_cqring_ev_posted(ctx);
1695 while (!list_empty(&list)) {
1696 req = list_first_entry(&list, struct io_kiocb, compl.list);
1697 list_del(&req->compl.list);
1704 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1705 struct task_struct *tsk,
1706 struct files_struct *files)
1708 if (test_bit(0, &ctx->cq_check_overflow)) {
1709 /* iopoll syncs against uring_lock, not completion_lock */
1710 if (ctx->flags & IORING_SETUP_IOPOLL)
1711 mutex_lock(&ctx->uring_lock);
1712 __io_cqring_overflow_flush(ctx, force, tsk, files);
1713 if (ctx->flags & IORING_SETUP_IOPOLL)
1714 mutex_unlock(&ctx->uring_lock);
1718 static void __io_cqring_fill_event(struct io_kiocb *req, long res,
1719 unsigned int cflags)
1721 struct io_ring_ctx *ctx = req->ctx;
1722 struct io_uring_cqe *cqe;
1724 trace_io_uring_complete(ctx, req->user_data, res);
1727 * If we can't get a cq entry, userspace overflowed the
1728 * submission (by quite a lot). Increment the overflow count in
1731 cqe = io_get_cqring(ctx);
1733 WRITE_ONCE(cqe->user_data, req->user_data);
1734 WRITE_ONCE(cqe->res, res);
1735 WRITE_ONCE(cqe->flags, cflags);
1736 } else if (ctx->cq_overflow_flushed ||
1737 atomic_read(&req->task->io_uring->in_idle)) {
1739 * If we're in ring overflow flush mode, or in task cancel mode,
1740 * then we cannot store the request for later flushing, we need
1741 * to drop it on the floor.
1743 ctx->cached_cq_overflow++;
1744 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1746 if (list_empty(&ctx->cq_overflow_list)) {
1747 set_bit(0, &ctx->sq_check_overflow);
1748 set_bit(0, &ctx->cq_check_overflow);
1749 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1753 req->compl.cflags = cflags;
1754 refcount_inc(&req->refs);
1755 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1759 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1761 __io_cqring_fill_event(req, res, 0);
1764 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1766 struct io_ring_ctx *ctx = req->ctx;
1767 unsigned long flags;
1769 spin_lock_irqsave(&ctx->completion_lock, flags);
1770 __io_cqring_fill_event(req, res, cflags);
1771 io_commit_cqring(ctx);
1772 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1774 io_cqring_ev_posted(ctx);
1777 static void io_submit_flush_completions(struct io_comp_state *cs)
1779 struct io_ring_ctx *ctx = cs->ctx;
1781 spin_lock_irq(&ctx->completion_lock);
1782 while (!list_empty(&cs->list)) {
1783 struct io_kiocb *req;
1785 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1786 list_del(&req->compl.list);
1787 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1790 * io_free_req() doesn't care about completion_lock unless one
1791 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1792 * because of a potential deadlock with req->work.fs->lock
1794 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1795 |REQ_F_WORK_INITIALIZED)) {
1796 spin_unlock_irq(&ctx->completion_lock);
1798 spin_lock_irq(&ctx->completion_lock);
1803 io_commit_cqring(ctx);
1804 spin_unlock_irq(&ctx->completion_lock);
1806 io_cqring_ev_posted(ctx);
1810 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1811 struct io_comp_state *cs)
1814 io_cqring_add_event(req, res, cflags);
1819 req->compl.cflags = cflags;
1820 list_add_tail(&req->compl.list, &cs->list);
1822 io_submit_flush_completions(cs);
1826 static void io_req_complete(struct io_kiocb *req, long res)
1828 __io_req_complete(req, res, 0, NULL);
1831 static inline bool io_is_fallback_req(struct io_kiocb *req)
1833 return req == (struct io_kiocb *)
1834 ((unsigned long) req->ctx->fallback_req & ~1UL);
1837 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1839 struct io_kiocb *req;
1841 req = ctx->fallback_req;
1842 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1848 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1849 struct io_submit_state *state)
1851 if (!state->free_reqs) {
1852 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1856 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1857 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1860 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1861 * retry single alloc to be on the safe side.
1863 if (unlikely(ret <= 0)) {
1864 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1865 if (!state->reqs[0])
1869 state->free_reqs = ret;
1873 return state->reqs[state->free_reqs];
1875 return io_get_fallback_req(ctx);
1878 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1882 percpu_ref_put(req->fixed_file_refs);
1887 static void io_dismantle_req(struct io_kiocb *req)
1891 if (req->async_data)
1892 kfree(req->async_data);
1894 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1896 io_req_clean_work(req);
1899 static void __io_free_req(struct io_kiocb *req)
1901 struct io_uring_task *tctx = req->task->io_uring;
1902 struct io_ring_ctx *ctx = req->ctx;
1904 io_dismantle_req(req);
1906 percpu_counter_dec(&tctx->inflight);
1907 if (atomic_read(&tctx->in_idle))
1908 wake_up(&tctx->wait);
1909 put_task_struct(req->task);
1911 if (likely(!io_is_fallback_req(req)))
1912 kmem_cache_free(req_cachep, req);
1914 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1915 percpu_ref_put(&ctx->refs);
1918 static void io_kill_linked_timeout(struct io_kiocb *req)
1920 struct io_ring_ctx *ctx = req->ctx;
1921 struct io_kiocb *link;
1922 bool cancelled = false;
1923 unsigned long flags;
1925 spin_lock_irqsave(&ctx->completion_lock, flags);
1926 link = list_first_entry_or_null(&req->link_list, struct io_kiocb,
1929 * Can happen if a linked timeout fired and link had been like
1930 * req -> link t-out -> link t-out [-> ...]
1932 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1933 struct io_timeout_data *io = link->async_data;
1936 list_del_init(&link->link_list);
1937 ret = hrtimer_try_to_cancel(&io->timer);
1939 io_cqring_fill_event(link, -ECANCELED);
1940 io_commit_cqring(ctx);
1944 req->flags &= ~REQ_F_LINK_TIMEOUT;
1945 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1948 io_cqring_ev_posted(ctx);
1953 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1955 struct io_kiocb *nxt;
1958 * The list should never be empty when we are called here. But could
1959 * potentially happen if the chain is messed up, check to be on the
1962 if (unlikely(list_empty(&req->link_list)))
1965 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1966 list_del_init(&req->link_list);
1967 if (!list_empty(&nxt->link_list))
1968 nxt->flags |= REQ_F_LINK_HEAD;
1973 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1975 static void io_fail_links(struct io_kiocb *req)
1977 struct io_ring_ctx *ctx = req->ctx;
1978 unsigned long flags;
1980 spin_lock_irqsave(&ctx->completion_lock, flags);
1981 while (!list_empty(&req->link_list)) {
1982 struct io_kiocb *link = list_first_entry(&req->link_list,
1983 struct io_kiocb, link_list);
1985 list_del_init(&link->link_list);
1986 trace_io_uring_fail_link(req, link);
1988 io_cqring_fill_event(link, -ECANCELED);
1991 * It's ok to free under spinlock as they're not linked anymore,
1992 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1995 if (link->flags & REQ_F_WORK_INITIALIZED)
1996 io_put_req_deferred(link, 2);
1998 io_double_put_req(link);
2001 io_commit_cqring(ctx);
2002 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2004 io_cqring_ev_posted(ctx);
2007 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2009 req->flags &= ~REQ_F_LINK_HEAD;
2010 if (req->flags & REQ_F_LINK_TIMEOUT)
2011 io_kill_linked_timeout(req);
2014 * If LINK is set, we have dependent requests in this chain. If we
2015 * didn't fail this request, queue the first one up, moving any other
2016 * dependencies to the next request. In case of failure, fail the rest
2019 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
2020 return io_req_link_next(req);
2025 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2027 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
2029 return __io_req_find_next(req);
2032 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
2034 struct task_struct *tsk = req->task;
2035 struct io_ring_ctx *ctx = req->ctx;
2036 enum task_work_notify_mode notify;
2039 if (tsk->flags & PF_EXITING)
2043 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2044 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2045 * processing task_work. There's no reliable way to tell if TWA_RESUME
2049 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
2050 notify = TWA_SIGNAL;
2052 ret = task_work_add(tsk, &req->task_work, notify);
2054 wake_up_process(tsk);
2059 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2061 struct io_ring_ctx *ctx = req->ctx;
2063 spin_lock_irq(&ctx->completion_lock);
2064 io_cqring_fill_event(req, error);
2065 io_commit_cqring(ctx);
2066 spin_unlock_irq(&ctx->completion_lock);
2068 io_cqring_ev_posted(ctx);
2069 req_set_fail_links(req);
2070 io_double_put_req(req);
2073 static void io_req_task_cancel(struct callback_head *cb)
2075 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2076 struct io_ring_ctx *ctx = req->ctx;
2078 mutex_lock(&ctx->uring_lock);
2079 __io_req_task_cancel(req, -ECANCELED);
2080 mutex_unlock(&ctx->uring_lock);
2081 percpu_ref_put(&ctx->refs);
2084 static void __io_req_task_submit(struct io_kiocb *req)
2086 struct io_ring_ctx *ctx = req->ctx;
2088 mutex_lock(&ctx->uring_lock);
2089 if (!ctx->sqo_dead && !__io_sq_thread_acquire_mm(ctx))
2090 __io_queue_sqe(req, NULL);
2092 __io_req_task_cancel(req, -EFAULT);
2093 mutex_unlock(&ctx->uring_lock);
2095 if (ctx->flags & IORING_SETUP_SQPOLL)
2096 io_sq_thread_drop_mm();
2099 static void io_req_task_submit(struct callback_head *cb)
2101 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2102 struct io_ring_ctx *ctx = req->ctx;
2104 __io_req_task_submit(req);
2105 percpu_ref_put(&ctx->refs);
2108 static void io_req_task_queue(struct io_kiocb *req)
2112 init_task_work(&req->task_work, io_req_task_submit);
2113 percpu_ref_get(&req->ctx->refs);
2115 ret = io_req_task_work_add(req, true);
2116 if (unlikely(ret)) {
2117 struct task_struct *tsk;
2119 init_task_work(&req->task_work, io_req_task_cancel);
2120 tsk = io_wq_get_task(req->ctx->io_wq);
2121 task_work_add(tsk, &req->task_work, TWA_NONE);
2122 wake_up_process(tsk);
2126 static void io_queue_next(struct io_kiocb *req)
2128 struct io_kiocb *nxt = io_req_find_next(req);
2131 io_req_task_queue(nxt);
2134 static void io_free_req(struct io_kiocb *req)
2141 void *reqs[IO_IOPOLL_BATCH];
2144 struct task_struct *task;
2148 static inline void io_init_req_batch(struct req_batch *rb)
2155 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2156 struct req_batch *rb)
2158 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2159 percpu_ref_put_many(&ctx->refs, rb->to_free);
2163 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2164 struct req_batch *rb)
2167 __io_req_free_batch_flush(ctx, rb);
2169 struct io_uring_task *tctx = rb->task->io_uring;
2171 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2172 if (atomic_read(&tctx->in_idle))
2173 wake_up(&tctx->wait);
2174 put_task_struct_many(rb->task, rb->task_refs);
2179 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2181 if (unlikely(io_is_fallback_req(req))) {
2185 if (req->flags & REQ_F_LINK_HEAD)
2188 if (req->task != rb->task) {
2190 struct io_uring_task *tctx = rb->task->io_uring;
2192 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2193 if (atomic_read(&tctx->in_idle))
2194 wake_up(&tctx->wait);
2195 put_task_struct_many(rb->task, rb->task_refs);
2197 rb->task = req->task;
2202 io_dismantle_req(req);
2203 rb->reqs[rb->to_free++] = req;
2204 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2205 __io_req_free_batch_flush(req->ctx, rb);
2209 * Drop reference to request, return next in chain (if there is one) if this
2210 * was the last reference to this request.
2212 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2214 struct io_kiocb *nxt = NULL;
2216 if (refcount_dec_and_test(&req->refs)) {
2217 nxt = io_req_find_next(req);
2223 static void io_put_req(struct io_kiocb *req)
2225 if (refcount_dec_and_test(&req->refs))
2229 static void io_put_req_deferred_cb(struct callback_head *cb)
2231 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2236 static void io_free_req_deferred(struct io_kiocb *req)
2240 init_task_work(&req->task_work, io_put_req_deferred_cb);
2241 ret = io_req_task_work_add(req, true);
2242 if (unlikely(ret)) {
2243 struct task_struct *tsk;
2245 tsk = io_wq_get_task(req->ctx->io_wq);
2246 task_work_add(tsk, &req->task_work, TWA_NONE);
2247 wake_up_process(tsk);
2251 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2253 if (refcount_sub_and_test(refs, &req->refs))
2254 io_free_req_deferred(req);
2257 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2259 struct io_kiocb *nxt;
2262 * A ref is owned by io-wq in which context we're. So, if that's the
2263 * last one, it's safe to steal next work. False negatives are Ok,
2264 * it just will be re-punted async in io_put_work()
2266 if (refcount_read(&req->refs) != 1)
2269 nxt = io_req_find_next(req);
2270 return nxt ? &nxt->work : NULL;
2273 static void io_double_put_req(struct io_kiocb *req)
2275 /* drop both submit and complete references */
2276 if (refcount_sub_and_test(2, &req->refs))
2280 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2282 struct io_rings *rings = ctx->rings;
2284 /* See comment at the top of this file */
2286 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2289 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2291 struct io_rings *rings = ctx->rings;
2293 /* make sure SQ entry isn't read before tail */
2294 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2297 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2299 unsigned int cflags;
2301 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2302 cflags |= IORING_CQE_F_BUFFER;
2303 req->flags &= ~REQ_F_BUFFER_SELECTED;
2308 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2310 struct io_buffer *kbuf;
2312 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2313 return io_put_kbuf(req, kbuf);
2316 static inline bool io_run_task_work(void)
2319 * Not safe to run on exiting task, and the task_work handling will
2320 * not add work to such a task.
2322 if (unlikely(current->flags & PF_EXITING))
2324 if (current->task_works) {
2325 __set_current_state(TASK_RUNNING);
2333 static void io_iopoll_queue(struct list_head *again)
2335 struct io_kiocb *req;
2338 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2339 list_del(&req->inflight_entry);
2340 __io_complete_rw(req, -EAGAIN, 0, NULL);
2341 } while (!list_empty(again));
2345 * Find and free completed poll iocbs
2347 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2348 struct list_head *done)
2350 struct req_batch rb;
2351 struct io_kiocb *req;
2354 /* order with ->result store in io_complete_rw_iopoll() */
2357 io_init_req_batch(&rb);
2358 while (!list_empty(done)) {
2361 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2362 if (READ_ONCE(req->result) == -EAGAIN) {
2364 req->iopoll_completed = 0;
2365 list_move_tail(&req->inflight_entry, &again);
2368 list_del(&req->inflight_entry);
2370 if (req->flags & REQ_F_BUFFER_SELECTED)
2371 cflags = io_put_rw_kbuf(req);
2373 __io_cqring_fill_event(req, req->result, cflags);
2376 if (refcount_dec_and_test(&req->refs))
2377 io_req_free_batch(&rb, req);
2380 io_commit_cqring(ctx);
2381 if (ctx->flags & IORING_SETUP_SQPOLL)
2382 io_cqring_ev_posted(ctx);
2383 io_req_free_batch_finish(ctx, &rb);
2385 if (!list_empty(&again))
2386 io_iopoll_queue(&again);
2389 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2392 struct io_kiocb *req, *tmp;
2398 * Only spin for completions if we don't have multiple devices hanging
2399 * off our complete list, and we're under the requested amount.
2401 spin = !ctx->poll_multi_file && *nr_events < min;
2404 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2405 struct kiocb *kiocb = &req->rw.kiocb;
2408 * Move completed and retryable entries to our local lists.
2409 * If we find a request that requires polling, break out
2410 * and complete those lists first, if we have entries there.
2412 if (READ_ONCE(req->iopoll_completed)) {
2413 list_move_tail(&req->inflight_entry, &done);
2416 if (!list_empty(&done))
2419 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2423 /* iopoll may have completed current req */
2424 if (READ_ONCE(req->iopoll_completed))
2425 list_move_tail(&req->inflight_entry, &done);
2432 if (!list_empty(&done))
2433 io_iopoll_complete(ctx, nr_events, &done);
2439 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2440 * non-spinning poll check - we'll still enter the driver poll loop, but only
2441 * as a non-spinning completion check.
2443 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2446 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2449 ret = io_do_iopoll(ctx, nr_events, min);
2452 if (*nr_events >= min)
2460 * We can't just wait for polled events to come to us, we have to actively
2461 * find and complete them.
2463 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2465 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2468 mutex_lock(&ctx->uring_lock);
2469 while (!list_empty(&ctx->iopoll_list)) {
2470 unsigned int nr_events = 0;
2472 io_do_iopoll(ctx, &nr_events, 0);
2474 /* let it sleep and repeat later if can't complete a request */
2478 * Ensure we allow local-to-the-cpu processing to take place,
2479 * in this case we need to ensure that we reap all events.
2480 * Also let task_work, etc. to progress by releasing the mutex
2482 if (need_resched()) {
2483 mutex_unlock(&ctx->uring_lock);
2485 mutex_lock(&ctx->uring_lock);
2488 mutex_unlock(&ctx->uring_lock);
2491 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2493 unsigned int nr_events = 0;
2494 int iters = 0, ret = 0;
2497 * We disallow the app entering submit/complete with polling, but we
2498 * still need to lock the ring to prevent racing with polled issue
2499 * that got punted to a workqueue.
2501 mutex_lock(&ctx->uring_lock);
2504 * Don't enter poll loop if we already have events pending.
2505 * If we do, we can potentially be spinning for commands that
2506 * already triggered a CQE (eg in error).
2508 if (test_bit(0, &ctx->cq_check_overflow))
2509 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2510 if (io_cqring_events(ctx))
2514 * If a submit got punted to a workqueue, we can have the
2515 * application entering polling for a command before it gets
2516 * issued. That app will hold the uring_lock for the duration
2517 * of the poll right here, so we need to take a breather every
2518 * now and then to ensure that the issue has a chance to add
2519 * the poll to the issued list. Otherwise we can spin here
2520 * forever, while the workqueue is stuck trying to acquire the
2523 if (!(++iters & 7)) {
2524 mutex_unlock(&ctx->uring_lock);
2526 mutex_lock(&ctx->uring_lock);
2529 ret = io_iopoll_getevents(ctx, &nr_events, min);
2533 } while (min && !nr_events && !need_resched());
2535 mutex_unlock(&ctx->uring_lock);
2539 static void kiocb_end_write(struct io_kiocb *req)
2542 * Tell lockdep we inherited freeze protection from submission
2545 if (req->flags & REQ_F_ISREG) {
2546 struct inode *inode = file_inode(req->file);
2548 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2550 file_end_write(req->file);
2553 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2554 struct io_comp_state *cs)
2556 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2559 if (kiocb->ki_flags & IOCB_WRITE)
2560 kiocb_end_write(req);
2562 if (res != req->result)
2563 req_set_fail_links(req);
2564 if (req->flags & REQ_F_BUFFER_SELECTED)
2565 cflags = io_put_rw_kbuf(req);
2566 __io_req_complete(req, res, cflags, cs);
2570 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2572 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2573 ssize_t ret = -ECANCELED;
2574 struct iov_iter iter;
2582 switch (req->opcode) {
2583 case IORING_OP_READV:
2584 case IORING_OP_READ_FIXED:
2585 case IORING_OP_READ:
2588 case IORING_OP_WRITEV:
2589 case IORING_OP_WRITE_FIXED:
2590 case IORING_OP_WRITE:
2594 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2599 if (!req->async_data) {
2600 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2603 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2611 req_set_fail_links(req);
2616 static bool io_rw_reissue(struct io_kiocb *req, long res)
2619 umode_t mode = file_inode(req->file)->i_mode;
2622 if (!S_ISBLK(mode) && !S_ISREG(mode))
2624 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2627 * If ref is dying, we might be running poll reap from the exit work.
2628 * Don't attempt to reissue from that path, just let it fail with
2631 if (percpu_ref_is_dying(&req->ctx->refs))
2634 ret = io_sq_thread_acquire_mm(req->ctx, req);
2636 if (io_resubmit_prep(req, ret)) {
2637 refcount_inc(&req->refs);
2638 io_queue_async_work(req);
2646 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2647 struct io_comp_state *cs)
2649 if (!io_rw_reissue(req, res))
2650 io_complete_rw_common(&req->rw.kiocb, res, cs);
2653 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2655 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2657 __io_complete_rw(req, res, res2, NULL);
2660 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2662 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2664 if (kiocb->ki_flags & IOCB_WRITE)
2665 kiocb_end_write(req);
2667 if (res != -EAGAIN && res != req->result)
2668 req_set_fail_links(req);
2670 WRITE_ONCE(req->result, res);
2671 /* order with io_poll_complete() checking ->result */
2673 WRITE_ONCE(req->iopoll_completed, 1);
2677 * After the iocb has been issued, it's safe to be found on the poll list.
2678 * Adding the kiocb to the list AFTER submission ensures that we don't
2679 * find it from a io_iopoll_getevents() thread before the issuer is done
2680 * accessing the kiocb cookie.
2682 static void io_iopoll_req_issued(struct io_kiocb *req)
2684 struct io_ring_ctx *ctx = req->ctx;
2687 * Track whether we have multiple files in our lists. This will impact
2688 * how we do polling eventually, not spinning if we're on potentially
2689 * different devices.
2691 if (list_empty(&ctx->iopoll_list)) {
2692 ctx->poll_multi_file = false;
2693 } else if (!ctx->poll_multi_file) {
2694 struct io_kiocb *list_req;
2696 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2698 if (list_req->file != req->file)
2699 ctx->poll_multi_file = true;
2703 * For fast devices, IO may have already completed. If it has, add
2704 * it to the front so we find it first.
2706 if (READ_ONCE(req->iopoll_completed))
2707 list_add(&req->inflight_entry, &ctx->iopoll_list);
2709 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2711 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2712 wq_has_sleeper(&ctx->sq_data->wait))
2713 wake_up(&ctx->sq_data->wait);
2716 static void __io_state_file_put(struct io_submit_state *state)
2718 if (state->has_refs)
2719 fput_many(state->file, state->has_refs);
2723 static inline void io_state_file_put(struct io_submit_state *state)
2726 __io_state_file_put(state);
2730 * Get as many references to a file as we have IOs left in this submission,
2731 * assuming most submissions are for one file, or at least that each file
2732 * has more than one submission.
2734 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2740 if (state->fd == fd) {
2744 __io_state_file_put(state);
2746 state->file = fget_many(fd, state->ios_left);
2751 state->has_refs = state->ios_left - 1;
2755 static bool io_bdev_nowait(struct block_device *bdev)
2758 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2765 * If we tracked the file through the SCM inflight mechanism, we could support
2766 * any file. For now, just ensure that anything potentially problematic is done
2769 static bool io_file_supports_async(struct file *file, int rw)
2771 umode_t mode = file_inode(file)->i_mode;
2773 if (S_ISBLK(mode)) {
2774 if (io_bdev_nowait(file->f_inode->i_bdev))
2780 if (S_ISREG(mode)) {
2781 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2782 file->f_op != &io_uring_fops)
2787 /* any ->read/write should understand O_NONBLOCK */
2788 if (file->f_flags & O_NONBLOCK)
2791 if (!(file->f_mode & FMODE_NOWAIT))
2795 return file->f_op->read_iter != NULL;
2797 return file->f_op->write_iter != NULL;
2800 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2802 struct io_ring_ctx *ctx = req->ctx;
2803 struct kiocb *kiocb = &req->rw.kiocb;
2807 if (S_ISREG(file_inode(req->file)->i_mode))
2808 req->flags |= REQ_F_ISREG;
2810 kiocb->ki_pos = READ_ONCE(sqe->off);
2811 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2812 req->flags |= REQ_F_CUR_POS;
2813 kiocb->ki_pos = req->file->f_pos;
2815 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2816 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2817 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2821 ioprio = READ_ONCE(sqe->ioprio);
2823 ret = ioprio_check_cap(ioprio);
2827 kiocb->ki_ioprio = ioprio;
2829 kiocb->ki_ioprio = get_current_ioprio();
2831 /* don't allow async punt if RWF_NOWAIT was requested */
2832 if (kiocb->ki_flags & IOCB_NOWAIT)
2833 req->flags |= REQ_F_NOWAIT;
2835 if (ctx->flags & IORING_SETUP_IOPOLL) {
2836 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2837 !kiocb->ki_filp->f_op->iopoll)
2840 kiocb->ki_flags |= IOCB_HIPRI;
2841 kiocb->ki_complete = io_complete_rw_iopoll;
2842 req->iopoll_completed = 0;
2844 if (kiocb->ki_flags & IOCB_HIPRI)
2846 kiocb->ki_complete = io_complete_rw;
2849 req->rw.addr = READ_ONCE(sqe->addr);
2850 req->rw.len = READ_ONCE(sqe->len);
2851 req->buf_index = READ_ONCE(sqe->buf_index);
2855 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2861 case -ERESTARTNOINTR:
2862 case -ERESTARTNOHAND:
2863 case -ERESTART_RESTARTBLOCK:
2865 * We can't just restart the syscall, since previously
2866 * submitted sqes may already be in progress. Just fail this
2872 kiocb->ki_complete(kiocb, ret, 0);
2876 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2877 struct io_comp_state *cs)
2879 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2880 struct io_async_rw *io = req->async_data;
2882 /* add previously done IO, if any */
2883 if (io && io->bytes_done > 0) {
2885 ret = io->bytes_done;
2887 ret += io->bytes_done;
2890 if (req->flags & REQ_F_CUR_POS)
2891 req->file->f_pos = kiocb->ki_pos;
2892 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2893 __io_complete_rw(req, ret, 0, cs);
2895 io_rw_done(kiocb, ret);
2898 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2899 struct iov_iter *iter)
2901 struct io_ring_ctx *ctx = req->ctx;
2902 size_t len = req->rw.len;
2903 struct io_mapped_ubuf *imu;
2904 u16 index, buf_index = req->buf_index;
2908 if (unlikely(buf_index >= ctx->nr_user_bufs))
2910 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2911 imu = &ctx->user_bufs[index];
2912 buf_addr = req->rw.addr;
2915 if (buf_addr + len < buf_addr)
2917 /* not inside the mapped region */
2918 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2922 * May not be a start of buffer, set size appropriately
2923 * and advance us to the beginning.
2925 offset = buf_addr - imu->ubuf;
2926 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2930 * Don't use iov_iter_advance() here, as it's really slow for
2931 * using the latter parts of a big fixed buffer - it iterates
2932 * over each segment manually. We can cheat a bit here, because
2935 * 1) it's a BVEC iter, we set it up
2936 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2937 * first and last bvec
2939 * So just find our index, and adjust the iterator afterwards.
2940 * If the offset is within the first bvec (or the whole first
2941 * bvec, just use iov_iter_advance(). This makes it easier
2942 * since we can just skip the first segment, which may not
2943 * be PAGE_SIZE aligned.
2945 const struct bio_vec *bvec = imu->bvec;
2947 if (offset <= bvec->bv_len) {
2948 iov_iter_advance(iter, offset);
2950 unsigned long seg_skip;
2952 /* skip first vec */
2953 offset -= bvec->bv_len;
2954 seg_skip = 1 + (offset >> PAGE_SHIFT);
2956 iter->bvec = bvec + seg_skip;
2957 iter->nr_segs -= seg_skip;
2958 iter->count -= bvec->bv_len + offset;
2959 iter->iov_offset = offset & ~PAGE_MASK;
2966 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2969 mutex_unlock(&ctx->uring_lock);
2972 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2975 * "Normal" inline submissions always hold the uring_lock, since we
2976 * grab it from the system call. Same is true for the SQPOLL offload.
2977 * The only exception is when we've detached the request and issue it
2978 * from an async worker thread, grab the lock for that case.
2981 mutex_lock(&ctx->uring_lock);
2984 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2985 int bgid, struct io_buffer *kbuf,
2988 struct io_buffer *head;
2990 if (req->flags & REQ_F_BUFFER_SELECTED)
2993 io_ring_submit_lock(req->ctx, needs_lock);
2995 lockdep_assert_held(&req->ctx->uring_lock);
2997 head = xa_load(&req->ctx->io_buffers, bgid);
2999 if (!list_empty(&head->list)) {
3000 kbuf = list_last_entry(&head->list, struct io_buffer,
3002 list_del(&kbuf->list);
3005 xa_erase(&req->ctx->io_buffers, bgid);
3007 if (*len > kbuf->len)
3010 kbuf = ERR_PTR(-ENOBUFS);
3013 io_ring_submit_unlock(req->ctx, needs_lock);
3018 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3021 struct io_buffer *kbuf;
3024 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3025 bgid = req->buf_index;
3026 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3029 req->rw.addr = (u64) (unsigned long) kbuf;
3030 req->flags |= REQ_F_BUFFER_SELECTED;
3031 return u64_to_user_ptr(kbuf->addr);
3034 #ifdef CONFIG_COMPAT
3035 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3038 struct compat_iovec __user *uiov;
3039 compat_ssize_t clen;
3043 uiov = u64_to_user_ptr(req->rw.addr);
3044 if (!access_ok(uiov, sizeof(*uiov)))
3046 if (__get_user(clen, &uiov->iov_len))
3052 buf = io_rw_buffer_select(req, &len, needs_lock);
3054 return PTR_ERR(buf);
3055 iov[0].iov_base = buf;
3056 iov[0].iov_len = (compat_size_t) len;
3061 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3064 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3068 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3071 len = iov[0].iov_len;
3074 buf = io_rw_buffer_select(req, &len, needs_lock);
3076 return PTR_ERR(buf);
3077 iov[0].iov_base = buf;
3078 iov[0].iov_len = len;
3082 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3085 if (req->flags & REQ_F_BUFFER_SELECTED) {
3086 struct io_buffer *kbuf;
3088 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3089 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3090 iov[0].iov_len = kbuf->len;
3093 if (req->rw.len != 1)
3096 #ifdef CONFIG_COMPAT
3097 if (req->ctx->compat)
3098 return io_compat_import(req, iov, needs_lock);
3101 return __io_iov_buffer_select(req, iov, needs_lock);
3104 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3105 struct iovec **iovec, struct iov_iter *iter,
3108 void __user *buf = u64_to_user_ptr(req->rw.addr);
3109 size_t sqe_len = req->rw.len;
3113 opcode = req->opcode;
3114 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3116 return io_import_fixed(req, rw, iter);
3119 /* buffer index only valid with fixed read/write, or buffer select */
3120 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3123 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3124 if (req->flags & REQ_F_BUFFER_SELECT) {
3125 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3127 return PTR_ERR(buf);
3128 req->rw.len = sqe_len;
3131 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3136 if (req->flags & REQ_F_BUFFER_SELECT) {
3137 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3139 ret = (*iovec)->iov_len;
3140 iov_iter_init(iter, rw, *iovec, 1, ret);
3146 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3150 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3151 struct iovec **iovec, struct iov_iter *iter,
3154 struct io_async_rw *iorw = req->async_data;
3157 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3162 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3164 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3168 * For files that don't have ->read_iter() and ->write_iter(), handle them
3169 * by looping over ->read() or ->write() manually.
3171 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3173 struct kiocb *kiocb = &req->rw.kiocb;
3174 struct file *file = req->file;
3178 * Don't support polled IO through this interface, and we can't
3179 * support non-blocking either. For the latter, this just causes
3180 * the kiocb to be handled from an async context.
3182 if (kiocb->ki_flags & IOCB_HIPRI)
3184 if (kiocb->ki_flags & IOCB_NOWAIT)
3187 while (iov_iter_count(iter)) {
3191 if (!iov_iter_is_bvec(iter)) {
3192 iovec = iov_iter_iovec(iter);
3194 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3195 iovec.iov_len = req->rw.len;
3199 nr = file->f_op->read(file, iovec.iov_base,
3200 iovec.iov_len, io_kiocb_ppos(kiocb));
3202 nr = file->f_op->write(file, iovec.iov_base,
3203 iovec.iov_len, io_kiocb_ppos(kiocb));
3211 if (!iov_iter_is_bvec(iter)) {
3212 iov_iter_advance(iter, nr);
3218 if (nr != iovec.iov_len)
3225 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3226 const struct iovec *fast_iov, struct iov_iter *iter)
3228 struct io_async_rw *rw = req->async_data;
3230 memcpy(&rw->iter, iter, sizeof(*iter));
3231 rw->free_iovec = iovec;
3233 /* can only be fixed buffers, no need to do anything */
3234 if (iov_iter_is_bvec(iter))
3237 unsigned iov_off = 0;
3239 rw->iter.iov = rw->fast_iov;
3240 if (iter->iov != fast_iov) {
3241 iov_off = iter->iov - fast_iov;
3242 rw->iter.iov += iov_off;
3244 if (rw->fast_iov != fast_iov)
3245 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3246 sizeof(struct iovec) * iter->nr_segs);
3248 req->flags |= REQ_F_NEED_CLEANUP;
3252 static inline int __io_alloc_async_data(struct io_kiocb *req)
3254 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3255 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3256 return req->async_data == NULL;
3259 static int io_alloc_async_data(struct io_kiocb *req)
3261 if (!io_op_defs[req->opcode].needs_async_data)
3264 return __io_alloc_async_data(req);
3267 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3268 const struct iovec *fast_iov,
3269 struct iov_iter *iter, bool force)
3271 if (!force && !io_op_defs[req->opcode].needs_async_data)
3273 if (!req->async_data) {
3274 if (__io_alloc_async_data(req))
3277 io_req_map_rw(req, iovec, fast_iov, iter);
3282 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3284 struct io_async_rw *iorw = req->async_data;
3285 struct iovec *iov = iorw->fast_iov;
3288 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3289 if (unlikely(ret < 0))
3292 iorw->bytes_done = 0;
3293 iorw->free_iovec = iov;
3295 req->flags |= REQ_F_NEED_CLEANUP;
3299 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3303 ret = io_prep_rw(req, sqe);
3307 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3310 /* either don't need iovec imported or already have it */
3311 if (!req->async_data)
3313 return io_rw_prep_async(req, READ);
3317 * This is our waitqueue callback handler, registered through lock_page_async()
3318 * when we initially tried to do the IO with the iocb armed our waitqueue.
3319 * This gets called when the page is unlocked, and we generally expect that to
3320 * happen when the page IO is completed and the page is now uptodate. This will
3321 * queue a task_work based retry of the operation, attempting to copy the data
3322 * again. If the latter fails because the page was NOT uptodate, then we will
3323 * do a thread based blocking retry of the operation. That's the unexpected
3326 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3327 int sync, void *arg)
3329 struct wait_page_queue *wpq;
3330 struct io_kiocb *req = wait->private;
3331 struct wait_page_key *key = arg;
3334 wpq = container_of(wait, struct wait_page_queue, wait);
3336 if (!wake_page_match(wpq, key))
3339 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3340 list_del_init(&wait->entry);
3342 init_task_work(&req->task_work, io_req_task_submit);
3343 percpu_ref_get(&req->ctx->refs);
3345 /* submit ref gets dropped, acquire a new one */
3346 refcount_inc(&req->refs);
3347 ret = io_req_task_work_add(req, true);
3348 if (unlikely(ret)) {
3349 struct task_struct *tsk;
3351 /* queue just for cancelation */
3352 init_task_work(&req->task_work, io_req_task_cancel);
3353 tsk = io_wq_get_task(req->ctx->io_wq);
3354 task_work_add(tsk, &req->task_work, TWA_NONE);
3355 wake_up_process(tsk);
3361 * This controls whether a given IO request should be armed for async page
3362 * based retry. If we return false here, the request is handed to the async
3363 * worker threads for retry. If we're doing buffered reads on a regular file,
3364 * we prepare a private wait_page_queue entry and retry the operation. This
3365 * will either succeed because the page is now uptodate and unlocked, or it
3366 * will register a callback when the page is unlocked at IO completion. Through
3367 * that callback, io_uring uses task_work to setup a retry of the operation.
3368 * That retry will attempt the buffered read again. The retry will generally
3369 * succeed, or in rare cases where it fails, we then fall back to using the
3370 * async worker threads for a blocking retry.
3372 static bool io_rw_should_retry(struct io_kiocb *req)
3374 struct io_async_rw *rw = req->async_data;
3375 struct wait_page_queue *wait = &rw->wpq;
3376 struct kiocb *kiocb = &req->rw.kiocb;
3378 /* never retry for NOWAIT, we just complete with -EAGAIN */
3379 if (req->flags & REQ_F_NOWAIT)
3382 /* Only for buffered IO */
3383 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3387 * just use poll if we can, and don't attempt if the fs doesn't
3388 * support callback based unlocks
3390 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3393 wait->wait.func = io_async_buf_func;
3394 wait->wait.private = req;
3395 wait->wait.flags = 0;
3396 INIT_LIST_HEAD(&wait->wait.entry);
3397 kiocb->ki_flags |= IOCB_WAITQ;
3398 kiocb->ki_flags &= ~IOCB_NOWAIT;
3399 kiocb->ki_waitq = wait;
3403 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3405 if (req->file->f_op->read_iter)
3406 return call_read_iter(req->file, &req->rw.kiocb, iter);
3407 else if (req->file->f_op->read)
3408 return loop_rw_iter(READ, req, iter);
3413 static int io_read(struct io_kiocb *req, bool force_nonblock,
3414 struct io_comp_state *cs)
3416 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3417 struct kiocb *kiocb = &req->rw.kiocb;
3418 struct iov_iter __iter, *iter = &__iter;
3419 struct io_async_rw *rw = req->async_data;
3420 ssize_t io_size, ret, ret2;
3426 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3429 io_size = iov_iter_count(iter);
3430 req->result = io_size;
3433 /* Ensure we clear previously set non-block flag */
3434 if (!force_nonblock)
3435 kiocb->ki_flags &= ~IOCB_NOWAIT;
3437 kiocb->ki_flags |= IOCB_NOWAIT;
3440 /* If the file doesn't support async, just async punt */
3441 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3445 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3449 ret = io_iter_do_read(req, iter);
3453 } else if (ret == -EIOCBQUEUED) {
3456 } else if (ret == -EAGAIN) {
3457 /* IOPOLL retry should happen for io-wq threads */
3458 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3460 /* no retry on NONBLOCK marked file */
3461 if (req->file->f_flags & O_NONBLOCK)
3463 /* some cases will consume bytes even on error returns */
3464 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3467 } else if (ret < 0) {
3468 /* make sure -ERESTARTSYS -> -EINTR is done */
3472 /* read it all, or we did blocking attempt. no retry. */
3473 if (!iov_iter_count(iter) || !force_nonblock ||
3474 (req->file->f_flags & O_NONBLOCK) || !(req->flags & REQ_F_ISREG))
3479 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3486 rw = req->async_data;
3487 /* it's copied and will be cleaned with ->io */
3489 /* now use our persistent iterator, if we aren't already */
3492 rw->bytes_done += ret;
3493 /* if we can retry, do so with the callbacks armed */
3494 if (!io_rw_should_retry(req)) {
3495 kiocb->ki_flags &= ~IOCB_WAITQ;
3500 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3501 * get -EIOCBQUEUED, then we'll get a notification when the desired
3502 * page gets unlocked. We can also get a partial read here, and if we
3503 * do, then just retry at the new offset.
3505 ret = io_iter_do_read(req, iter);
3506 if (ret == -EIOCBQUEUED) {
3509 } else if (ret > 0 && ret < io_size) {
3510 /* we got some bytes, but not all. retry. */
3511 kiocb->ki_flags &= ~IOCB_WAITQ;
3515 kiocb_done(kiocb, ret, cs);
3518 /* it's reportedly faster than delegating the null check to kfree() */
3524 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3528 ret = io_prep_rw(req, sqe);
3532 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3535 /* either don't need iovec imported or already have it */
3536 if (!req->async_data)
3538 return io_rw_prep_async(req, WRITE);
3541 static int io_write(struct io_kiocb *req, bool force_nonblock,
3542 struct io_comp_state *cs)
3544 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3545 struct kiocb *kiocb = &req->rw.kiocb;
3546 struct iov_iter __iter, *iter = &__iter;
3547 struct io_async_rw *rw = req->async_data;
3548 ssize_t ret, ret2, io_size;
3553 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3556 io_size = iov_iter_count(iter);
3557 req->result = io_size;
3559 /* Ensure we clear previously set non-block flag */
3560 if (!force_nonblock)
3561 kiocb->ki_flags &= ~IOCB_NOWAIT;
3563 kiocb->ki_flags |= IOCB_NOWAIT;
3565 /* If the file doesn't support async, just async punt */
3566 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3569 /* file path doesn't support NOWAIT for non-direct_IO */
3570 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3571 (req->flags & REQ_F_ISREG))
3574 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3579 * Open-code file_start_write here to grab freeze protection,
3580 * which will be released by another thread in
3581 * io_complete_rw(). Fool lockdep by telling it the lock got
3582 * released so that it doesn't complain about the held lock when
3583 * we return to userspace.
3585 if (req->flags & REQ_F_ISREG) {
3586 sb_start_write(file_inode(req->file)->i_sb);
3587 __sb_writers_release(file_inode(req->file)->i_sb,
3590 kiocb->ki_flags |= IOCB_WRITE;
3592 if (req->file->f_op->write_iter)
3593 ret2 = call_write_iter(req->file, kiocb, iter);
3594 else if (req->file->f_op->write)
3595 ret2 = loop_rw_iter(WRITE, req, iter);
3600 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3601 * retry them without IOCB_NOWAIT.
3603 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3605 /* no retry on NONBLOCK marked file */
3606 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3608 if (!force_nonblock || ret2 != -EAGAIN) {
3609 /* IOPOLL retry should happen for io-wq threads */
3610 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3613 kiocb_done(kiocb, ret2, cs);
3616 /* some cases will consume bytes even on error returns */
3617 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3618 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3623 /* it's reportedly faster than delegating the null check to kfree() */
3629 static int __io_splice_prep(struct io_kiocb *req,
3630 const struct io_uring_sqe *sqe)
3632 struct io_splice* sp = &req->splice;
3633 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3635 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3639 sp->len = READ_ONCE(sqe->len);
3640 sp->flags = READ_ONCE(sqe->splice_flags);
3642 if (unlikely(sp->flags & ~valid_flags))
3645 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3646 (sp->flags & SPLICE_F_FD_IN_FIXED));
3649 req->flags |= REQ_F_NEED_CLEANUP;
3651 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3653 * Splice operation will be punted aync, and here need to
3654 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3656 io_req_init_async(req);
3657 req->work.flags |= IO_WQ_WORK_UNBOUND;
3663 static int io_tee_prep(struct io_kiocb *req,
3664 const struct io_uring_sqe *sqe)
3666 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3668 return __io_splice_prep(req, sqe);
3671 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3673 struct io_splice *sp = &req->splice;
3674 struct file *in = sp->file_in;
3675 struct file *out = sp->file_out;
3676 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3682 ret = do_tee(in, out, sp->len, flags);
3684 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3685 req->flags &= ~REQ_F_NEED_CLEANUP;
3688 req_set_fail_links(req);
3689 io_req_complete(req, ret);
3693 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3695 struct io_splice* sp = &req->splice;
3697 sp->off_in = READ_ONCE(sqe->splice_off_in);
3698 sp->off_out = READ_ONCE(sqe->off);
3699 return __io_splice_prep(req, sqe);
3702 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3704 struct io_splice *sp = &req->splice;
3705 struct file *in = sp->file_in;
3706 struct file *out = sp->file_out;
3707 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3708 loff_t *poff_in, *poff_out;
3714 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3715 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3718 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3720 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3721 req->flags &= ~REQ_F_NEED_CLEANUP;
3724 req_set_fail_links(req);
3725 io_req_complete(req, ret);
3730 * IORING_OP_NOP just posts a completion event, nothing else.
3732 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3734 struct io_ring_ctx *ctx = req->ctx;
3736 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3739 __io_req_complete(req, 0, 0, cs);
3743 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3745 struct io_ring_ctx *ctx = req->ctx;
3750 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3752 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
3756 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3757 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3760 req->sync.off = READ_ONCE(sqe->off);
3761 req->sync.len = READ_ONCE(sqe->len);
3765 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3767 loff_t end = req->sync.off + req->sync.len;
3770 /* fsync always requires a blocking context */
3774 ret = vfs_fsync_range(req->file, req->sync.off,
3775 end > 0 ? end : LLONG_MAX,
3776 req->sync.flags & IORING_FSYNC_DATASYNC);
3778 req_set_fail_links(req);
3779 io_req_complete(req, ret);
3783 static int io_fallocate_prep(struct io_kiocb *req,
3784 const struct io_uring_sqe *sqe)
3786 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
3789 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3792 req->sync.off = READ_ONCE(sqe->off);
3793 req->sync.len = READ_ONCE(sqe->addr);
3794 req->sync.mode = READ_ONCE(sqe->len);
3798 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3802 /* fallocate always requiring blocking context */
3805 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3808 req_set_fail_links(req);
3809 io_req_complete(req, ret);
3813 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3815 const char __user *fname;
3818 if (unlikely(sqe->ioprio || sqe->buf_index || sqe->splice_fd_in))
3820 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3823 /* open.how should be already initialised */
3824 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3825 req->open.how.flags |= O_LARGEFILE;
3827 req->open.dfd = READ_ONCE(sqe->fd);
3828 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3829 req->open.filename = getname(fname);
3830 if (IS_ERR(req->open.filename)) {
3831 ret = PTR_ERR(req->open.filename);
3832 req->open.filename = NULL;
3835 req->open.nofile = rlimit(RLIMIT_NOFILE);
3836 req->open.ignore_nonblock = false;
3837 req->flags |= REQ_F_NEED_CLEANUP;
3841 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3845 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3847 mode = READ_ONCE(sqe->len);
3848 flags = READ_ONCE(sqe->open_flags);
3849 req->open.how = build_open_how(flags, mode);
3850 return __io_openat_prep(req, sqe);
3853 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3855 struct open_how __user *how;
3859 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3861 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3862 len = READ_ONCE(sqe->len);
3863 if (len < OPEN_HOW_SIZE_VER0)
3866 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3871 return __io_openat_prep(req, sqe);
3874 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3876 struct open_flags op;
3880 if (force_nonblock && !req->open.ignore_nonblock)
3883 ret = build_open_flags(&req->open.how, &op);
3887 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3891 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3894 ret = PTR_ERR(file);
3896 * A work-around to ensure that /proc/self works that way
3897 * that it should - if we get -EOPNOTSUPP back, then assume
3898 * that proc_self_get_link() failed us because we're in async
3899 * context. We should be safe to retry this from the task
3900 * itself with force_nonblock == false set, as it should not
3901 * block on lookup. Would be nice to know this upfront and
3902 * avoid the async dance, but doesn't seem feasible.
3904 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
3905 req->open.ignore_nonblock = true;
3906 refcount_inc(&req->refs);
3907 io_req_task_queue(req);
3911 fsnotify_open(file);
3912 fd_install(ret, file);
3915 putname(req->open.filename);
3916 req->flags &= ~REQ_F_NEED_CLEANUP;
3918 req_set_fail_links(req);
3919 io_req_complete(req, ret);
3923 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3925 return io_openat2(req, force_nonblock);
3928 static int io_remove_buffers_prep(struct io_kiocb *req,
3929 const struct io_uring_sqe *sqe)
3931 struct io_provide_buf *p = &req->pbuf;
3934 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
3938 tmp = READ_ONCE(sqe->fd);
3939 if (!tmp || tmp > USHRT_MAX)
3942 memset(p, 0, sizeof(*p));
3944 p->bgid = READ_ONCE(sqe->buf_group);
3948 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3949 int bgid, unsigned nbufs)
3953 /* shouldn't happen */
3957 /* the head kbuf is the list itself */
3958 while (!list_empty(&buf->list)) {
3959 struct io_buffer *nxt;
3961 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3962 list_del(&nxt->list);
3969 xa_erase(&ctx->io_buffers, bgid);
3974 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3975 struct io_comp_state *cs)
3977 struct io_provide_buf *p = &req->pbuf;
3978 struct io_ring_ctx *ctx = req->ctx;
3979 struct io_buffer *head;
3982 io_ring_submit_lock(ctx, !force_nonblock);
3984 lockdep_assert_held(&ctx->uring_lock);
3987 head = xa_load(&ctx->io_buffers, p->bgid);
3989 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3991 req_set_fail_links(req);
3993 /* need to hold the lock to complete IOPOLL requests */
3994 if (ctx->flags & IORING_SETUP_IOPOLL) {
3995 __io_req_complete(req, ret, 0, cs);
3996 io_ring_submit_unlock(ctx, !force_nonblock);
3998 io_ring_submit_unlock(ctx, !force_nonblock);
3999 __io_req_complete(req, ret, 0, cs);
4004 static int io_provide_buffers_prep(struct io_kiocb *req,
4005 const struct io_uring_sqe *sqe)
4007 unsigned long size, tmp_check;
4008 struct io_provide_buf *p = &req->pbuf;
4011 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4014 tmp = READ_ONCE(sqe->fd);
4015 if (!tmp || tmp > USHRT_MAX)
4018 p->addr = READ_ONCE(sqe->addr);
4019 p->len = READ_ONCE(sqe->len);
4021 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4024 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4027 size = (unsigned long)p->len * p->nbufs;
4028 if (!access_ok(u64_to_user_ptr(p->addr), size))
4031 p->bgid = READ_ONCE(sqe->buf_group);
4032 tmp = READ_ONCE(sqe->off);
4033 if (tmp > USHRT_MAX)
4039 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4041 struct io_buffer *buf;
4042 u64 addr = pbuf->addr;
4043 int i, bid = pbuf->bid;
4045 for (i = 0; i < pbuf->nbufs; i++) {
4046 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4051 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4056 INIT_LIST_HEAD(&buf->list);
4059 list_add_tail(&buf->list, &(*head)->list);
4063 return i ? i : -ENOMEM;
4066 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4067 struct io_comp_state *cs)
4069 struct io_provide_buf *p = &req->pbuf;
4070 struct io_ring_ctx *ctx = req->ctx;
4071 struct io_buffer *head, *list;
4074 io_ring_submit_lock(ctx, !force_nonblock);
4076 lockdep_assert_held(&ctx->uring_lock);
4078 list = head = xa_load(&ctx->io_buffers, p->bgid);
4080 ret = io_add_buffers(p, &head);
4081 if (ret >= 0 && !list) {
4082 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4084 __io_remove_buffers(ctx, head, p->bgid, -1U);
4087 req_set_fail_links(req);
4089 /* need to hold the lock to complete IOPOLL requests */
4090 if (ctx->flags & IORING_SETUP_IOPOLL) {
4091 __io_req_complete(req, ret, 0, cs);
4092 io_ring_submit_unlock(ctx, !force_nonblock);
4094 io_ring_submit_unlock(ctx, !force_nonblock);
4095 __io_req_complete(req, ret, 0, cs);
4100 static int io_epoll_ctl_prep(struct io_kiocb *req,
4101 const struct io_uring_sqe *sqe)
4103 #if defined(CONFIG_EPOLL)
4104 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4106 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4109 req->epoll.epfd = READ_ONCE(sqe->fd);
4110 req->epoll.op = READ_ONCE(sqe->len);
4111 req->epoll.fd = READ_ONCE(sqe->off);
4113 if (ep_op_has_event(req->epoll.op)) {
4114 struct epoll_event __user *ev;
4116 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4117 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4127 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4128 struct io_comp_state *cs)
4130 #if defined(CONFIG_EPOLL)
4131 struct io_epoll *ie = &req->epoll;
4134 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4135 if (force_nonblock && ret == -EAGAIN)
4139 req_set_fail_links(req);
4140 __io_req_complete(req, ret, 0, cs);
4147 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4149 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4150 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4152 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4155 req->madvise.addr = READ_ONCE(sqe->addr);
4156 req->madvise.len = READ_ONCE(sqe->len);
4157 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4164 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4166 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4167 struct io_madvise *ma = &req->madvise;
4173 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4175 req_set_fail_links(req);
4176 io_req_complete(req, ret);
4183 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4185 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4187 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4190 req->fadvise.offset = READ_ONCE(sqe->off);
4191 req->fadvise.len = READ_ONCE(sqe->len);
4192 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4196 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4198 struct io_fadvise *fa = &req->fadvise;
4201 if (force_nonblock) {
4202 switch (fa->advice) {
4203 case POSIX_FADV_NORMAL:
4204 case POSIX_FADV_RANDOM:
4205 case POSIX_FADV_SEQUENTIAL:
4212 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4214 req_set_fail_links(req);
4215 io_req_complete(req, ret);
4219 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4221 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4223 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4225 if (req->flags & REQ_F_FIXED_FILE)
4228 req->statx.dfd = READ_ONCE(sqe->fd);
4229 req->statx.mask = READ_ONCE(sqe->len);
4230 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4231 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4232 req->statx.flags = READ_ONCE(sqe->statx_flags);
4237 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4239 struct io_statx *ctx = &req->statx;
4242 if (force_nonblock) {
4243 /* only need file table for an actual valid fd */
4244 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4245 req->flags |= REQ_F_NO_FILE_TABLE;
4249 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4253 req_set_fail_links(req);
4254 io_req_complete(req, ret);
4258 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4261 * If we queue this for async, it must not be cancellable. That would
4262 * leave the 'file' in an undeterminate state, and here need to modify
4263 * io_wq_work.flags, so initialize io_wq_work firstly.
4265 io_req_init_async(req);
4267 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4269 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4270 sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
4272 if (req->flags & REQ_F_FIXED_FILE)
4275 req->close.fd = READ_ONCE(sqe->fd);
4276 if ((req->file && req->file->f_op == &io_uring_fops))
4279 req->close.put_file = NULL;
4283 static int io_close(struct io_kiocb *req, bool force_nonblock,
4284 struct io_comp_state *cs)
4286 struct io_close *close = &req->close;
4289 /* might be already done during nonblock submission */
4290 if (!close->put_file) {
4291 ret = __close_fd_get_file(close->fd, &close->put_file);
4293 return (ret == -ENOENT) ? -EBADF : ret;
4296 /* if the file has a flush method, be safe and punt to async */
4297 if (close->put_file->f_op->flush && force_nonblock) {
4298 /* not safe to cancel at this point */
4299 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4300 /* was never set, but play safe */
4301 req->flags &= ~REQ_F_NOWAIT;
4302 /* avoid grabbing files - we don't need the files */
4303 req->flags |= REQ_F_NO_FILE_TABLE;
4307 /* No ->flush() or already async, safely close from here */
4308 ret = filp_close(close->put_file, req->work.identity->files);
4310 req_set_fail_links(req);
4311 fput(close->put_file);
4312 close->put_file = NULL;
4313 __io_req_complete(req, ret, 0, cs);
4317 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4319 struct io_ring_ctx *ctx = req->ctx;
4324 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4326 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4330 req->sync.off = READ_ONCE(sqe->off);
4331 req->sync.len = READ_ONCE(sqe->len);
4332 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4336 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4340 /* sync_file_range always requires a blocking context */
4344 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4347 req_set_fail_links(req);
4348 io_req_complete(req, ret);
4352 #if defined(CONFIG_NET)
4353 static int io_setup_async_msg(struct io_kiocb *req,
4354 struct io_async_msghdr *kmsg)
4356 struct io_async_msghdr *async_msg = req->async_data;
4360 if (io_alloc_async_data(req)) {
4361 if (kmsg->iov != kmsg->fast_iov)
4365 async_msg = req->async_data;
4366 req->flags |= REQ_F_NEED_CLEANUP;
4367 memcpy(async_msg, kmsg, sizeof(*kmsg));
4371 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4372 struct io_async_msghdr *iomsg)
4374 iomsg->iov = iomsg->fast_iov;
4375 iomsg->msg.msg_name = &iomsg->addr;
4376 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4377 req->sr_msg.msg_flags, &iomsg->iov);
4380 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4382 struct io_async_msghdr *async_msg = req->async_data;
4383 struct io_sr_msg *sr = &req->sr_msg;
4386 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4389 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4390 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4391 sr->len = READ_ONCE(sqe->len);
4393 #ifdef CONFIG_COMPAT
4394 if (req->ctx->compat)
4395 sr->msg_flags |= MSG_CMSG_COMPAT;
4398 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4400 ret = io_sendmsg_copy_hdr(req, async_msg);
4402 req->flags |= REQ_F_NEED_CLEANUP;
4406 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4407 struct io_comp_state *cs)
4409 struct io_async_msghdr iomsg, *kmsg;
4410 struct socket *sock;
4415 sock = sock_from_file(req->file, &ret);
4416 if (unlikely(!sock))
4419 if (req->async_data) {
4420 kmsg = req->async_data;
4421 kmsg->msg.msg_name = &kmsg->addr;
4422 /* if iov is set, it's allocated already */
4424 kmsg->iov = kmsg->fast_iov;
4425 kmsg->msg.msg_iter.iov = kmsg->iov;
4427 ret = io_sendmsg_copy_hdr(req, &iomsg);
4433 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4434 if (flags & MSG_DONTWAIT)
4435 req->flags |= REQ_F_NOWAIT;
4436 else if (force_nonblock)
4437 flags |= MSG_DONTWAIT;
4439 if (flags & MSG_WAITALL)
4440 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4442 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4443 if (force_nonblock && ret == -EAGAIN)
4444 return io_setup_async_msg(req, kmsg);
4445 if (ret == -ERESTARTSYS)
4448 if (kmsg->iov != kmsg->fast_iov)
4450 req->flags &= ~REQ_F_NEED_CLEANUP;
4452 req_set_fail_links(req);
4453 __io_req_complete(req, ret, 0, cs);
4457 static int io_send(struct io_kiocb *req, bool force_nonblock,
4458 struct io_comp_state *cs)
4460 struct io_sr_msg *sr = &req->sr_msg;
4463 struct socket *sock;
4468 sock = sock_from_file(req->file, &ret);
4469 if (unlikely(!sock))
4472 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4476 msg.msg_name = NULL;
4477 msg.msg_control = NULL;
4478 msg.msg_controllen = 0;
4479 msg.msg_namelen = 0;
4481 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4482 if (flags & MSG_DONTWAIT)
4483 req->flags |= REQ_F_NOWAIT;
4484 else if (force_nonblock)
4485 flags |= MSG_DONTWAIT;
4487 if (flags & MSG_WAITALL)
4488 min_ret = iov_iter_count(&msg.msg_iter);
4490 msg.msg_flags = flags;
4491 ret = sock_sendmsg(sock, &msg);
4492 if (force_nonblock && ret == -EAGAIN)
4494 if (ret == -ERESTARTSYS)
4498 req_set_fail_links(req);
4499 __io_req_complete(req, ret, 0, cs);
4503 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4504 struct io_async_msghdr *iomsg)
4506 struct io_sr_msg *sr = &req->sr_msg;
4507 struct iovec __user *uiov;
4511 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4512 &iomsg->uaddr, &uiov, &iov_len);
4516 if (req->flags & REQ_F_BUFFER_SELECT) {
4519 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4521 sr->len = iomsg->iov[0].iov_len;
4522 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4526 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4527 &iomsg->iov, &iomsg->msg.msg_iter,
4536 #ifdef CONFIG_COMPAT
4537 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4538 struct io_async_msghdr *iomsg)
4540 struct compat_msghdr __user *msg_compat;
4541 struct io_sr_msg *sr = &req->sr_msg;
4542 struct compat_iovec __user *uiov;
4547 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4548 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4553 uiov = compat_ptr(ptr);
4554 if (req->flags & REQ_F_BUFFER_SELECT) {
4555 compat_ssize_t clen;
4559 if (!access_ok(uiov, sizeof(*uiov)))
4561 if (__get_user(clen, &uiov->iov_len))
4566 iomsg->iov[0].iov_len = clen;
4569 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4570 UIO_FASTIOV, &iomsg->iov,
4571 &iomsg->msg.msg_iter, true);
4580 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4581 struct io_async_msghdr *iomsg)
4583 iomsg->msg.msg_name = &iomsg->addr;
4584 iomsg->iov = iomsg->fast_iov;
4586 #ifdef CONFIG_COMPAT
4587 if (req->ctx->compat)
4588 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4591 return __io_recvmsg_copy_hdr(req, iomsg);
4594 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4597 struct io_sr_msg *sr = &req->sr_msg;
4598 struct io_buffer *kbuf;
4600 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4605 req->flags |= REQ_F_BUFFER_SELECTED;
4609 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4611 return io_put_kbuf(req, req->sr_msg.kbuf);
4614 static int io_recvmsg_prep(struct io_kiocb *req,
4615 const struct io_uring_sqe *sqe)
4617 struct io_async_msghdr *async_msg = req->async_data;
4618 struct io_sr_msg *sr = &req->sr_msg;
4621 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4624 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4625 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4626 sr->len = READ_ONCE(sqe->len);
4627 sr->bgid = READ_ONCE(sqe->buf_group);
4629 #ifdef CONFIG_COMPAT
4630 if (req->ctx->compat)
4631 sr->msg_flags |= MSG_CMSG_COMPAT;
4634 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4636 ret = io_recvmsg_copy_hdr(req, async_msg);
4638 req->flags |= REQ_F_NEED_CLEANUP;
4642 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4643 struct io_comp_state *cs)
4645 struct io_async_msghdr iomsg, *kmsg;
4646 struct socket *sock;
4647 struct io_buffer *kbuf;
4650 int ret, cflags = 0;
4652 sock = sock_from_file(req->file, &ret);
4653 if (unlikely(!sock))
4656 if (req->async_data) {
4657 kmsg = req->async_data;
4658 kmsg->msg.msg_name = &kmsg->addr;
4659 /* if iov is set, it's allocated already */
4661 kmsg->iov = kmsg->fast_iov;
4662 kmsg->msg.msg_iter.iov = kmsg->iov;
4664 ret = io_recvmsg_copy_hdr(req, &iomsg);
4670 if (req->flags & REQ_F_BUFFER_SELECT) {
4671 kbuf = io_recv_buffer_select(req, !force_nonblock);
4673 return PTR_ERR(kbuf);
4674 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4675 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4676 1, req->sr_msg.len);
4679 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4680 if (flags & MSG_DONTWAIT)
4681 req->flags |= REQ_F_NOWAIT;
4682 else if (force_nonblock)
4683 flags |= MSG_DONTWAIT;
4685 if (flags & MSG_WAITALL)
4686 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4688 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4689 kmsg->uaddr, flags);
4690 if (force_nonblock && ret == -EAGAIN)
4691 return io_setup_async_msg(req, kmsg);
4692 if (ret == -ERESTARTSYS)
4695 if (req->flags & REQ_F_BUFFER_SELECTED)
4696 cflags = io_put_recv_kbuf(req);
4697 if (kmsg->iov != kmsg->fast_iov)
4699 req->flags &= ~REQ_F_NEED_CLEANUP;
4700 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4701 req_set_fail_links(req);
4702 __io_req_complete(req, ret, cflags, cs);
4706 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4707 struct io_comp_state *cs)
4709 struct io_buffer *kbuf;
4710 struct io_sr_msg *sr = &req->sr_msg;
4712 void __user *buf = sr->buf;
4713 struct socket *sock;
4717 int ret, cflags = 0;
4719 sock = sock_from_file(req->file, &ret);
4720 if (unlikely(!sock))
4723 if (req->flags & REQ_F_BUFFER_SELECT) {
4724 kbuf = io_recv_buffer_select(req, !force_nonblock);
4726 return PTR_ERR(kbuf);
4727 buf = u64_to_user_ptr(kbuf->addr);
4730 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4734 msg.msg_name = NULL;
4735 msg.msg_control = NULL;
4736 msg.msg_controllen = 0;
4737 msg.msg_namelen = 0;
4738 msg.msg_iocb = NULL;
4741 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4742 if (flags & MSG_DONTWAIT)
4743 req->flags |= REQ_F_NOWAIT;
4744 else if (force_nonblock)
4745 flags |= MSG_DONTWAIT;
4747 if (flags & MSG_WAITALL)
4748 min_ret = iov_iter_count(&msg.msg_iter);
4750 ret = sock_recvmsg(sock, &msg, flags);
4751 if (force_nonblock && ret == -EAGAIN)
4753 if (ret == -ERESTARTSYS)
4756 if (req->flags & REQ_F_BUFFER_SELECTED)
4757 cflags = io_put_recv_kbuf(req);
4758 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4759 req_set_fail_links(req);
4760 __io_req_complete(req, ret, cflags, cs);
4764 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4766 struct io_accept *accept = &req->accept;
4768 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4770 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->splice_fd_in)
4773 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4774 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4775 accept->flags = READ_ONCE(sqe->accept_flags);
4776 accept->nofile = rlimit(RLIMIT_NOFILE);
4780 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4781 struct io_comp_state *cs)
4783 struct io_accept *accept = &req->accept;
4784 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4787 if (req->file->f_flags & O_NONBLOCK)
4788 req->flags |= REQ_F_NOWAIT;
4790 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4791 accept->addr_len, accept->flags,
4793 if (ret == -EAGAIN && force_nonblock)
4796 if (ret == -ERESTARTSYS)
4798 req_set_fail_links(req);
4800 __io_req_complete(req, ret, 0, cs);
4804 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4806 struct io_connect *conn = &req->connect;
4807 struct io_async_connect *io = req->async_data;
4809 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4811 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
4815 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4816 conn->addr_len = READ_ONCE(sqe->addr2);
4821 return move_addr_to_kernel(conn->addr, conn->addr_len,
4825 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4826 struct io_comp_state *cs)
4828 struct io_async_connect __io, *io;
4829 unsigned file_flags;
4832 if (req->async_data) {
4833 io = req->async_data;
4835 ret = move_addr_to_kernel(req->connect.addr,
4836 req->connect.addr_len,
4843 file_flags = force_nonblock ? O_NONBLOCK : 0;
4845 ret = __sys_connect_file(req->file, &io->address,
4846 req->connect.addr_len, file_flags);
4847 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4848 if (req->async_data)
4850 if (io_alloc_async_data(req)) {
4854 io = req->async_data;
4855 memcpy(req->async_data, &__io, sizeof(__io));
4858 if (ret == -ERESTARTSYS)
4862 req_set_fail_links(req);
4863 __io_req_complete(req, ret, 0, cs);
4866 #else /* !CONFIG_NET */
4867 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4872 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4873 struct io_comp_state *cs)
4878 static int io_send(struct io_kiocb *req, bool force_nonblock,
4879 struct io_comp_state *cs)
4884 static int io_recvmsg_prep(struct io_kiocb *req,
4885 const struct io_uring_sqe *sqe)
4890 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4891 struct io_comp_state *cs)
4896 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4897 struct io_comp_state *cs)
4902 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4907 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4908 struct io_comp_state *cs)
4913 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4918 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4919 struct io_comp_state *cs)
4923 #endif /* CONFIG_NET */
4925 struct io_poll_table {
4926 struct poll_table_struct pt;
4927 struct io_kiocb *req;
4932 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4933 __poll_t mask, task_work_func_t func)
4938 /* for instances that support it check for an event match first: */
4939 if (mask && !(mask & poll->events))
4942 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4944 list_del_init(&poll->wait.entry);
4947 init_task_work(&req->task_work, func);
4948 percpu_ref_get(&req->ctx->refs);
4951 * If we using the signalfd wait_queue_head for this wakeup, then
4952 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4953 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4954 * either, as the normal wakeup will suffice.
4956 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4959 * If this fails, then the task is exiting. When a task exits, the
4960 * work gets canceled, so just cancel this request as well instead
4961 * of executing it. We can't safely execute it anyway, as we may not
4962 * have the needed state needed for it anyway.
4964 ret = io_req_task_work_add(req, twa_signal_ok);
4965 if (unlikely(ret)) {
4966 struct task_struct *tsk;
4968 WRITE_ONCE(poll->canceled, true);
4969 tsk = io_wq_get_task(req->ctx->io_wq);
4970 task_work_add(tsk, &req->task_work, TWA_NONE);
4971 wake_up_process(tsk);
4976 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4977 __acquires(&req->ctx->completion_lock)
4979 struct io_ring_ctx *ctx = req->ctx;
4981 if (!req->result && !READ_ONCE(poll->canceled)) {
4982 struct poll_table_struct pt = { ._key = poll->events };
4984 req->result = vfs_poll(req->file, &pt) & poll->events;
4987 spin_lock_irq(&ctx->completion_lock);
4988 if (!req->result && !READ_ONCE(poll->canceled)) {
4989 add_wait_queue(poll->head, &poll->wait);
4996 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4998 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4999 if (req->opcode == IORING_OP_POLL_ADD)
5000 return req->async_data;
5001 return req->apoll->double_poll;
5004 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5006 if (req->opcode == IORING_OP_POLL_ADD)
5008 return &req->apoll->poll;
5011 static void io_poll_remove_double(struct io_kiocb *req)
5013 struct io_poll_iocb *poll = io_poll_get_double(req);
5015 lockdep_assert_held(&req->ctx->completion_lock);
5017 if (poll && poll->head) {
5018 struct wait_queue_head *head = poll->head;
5020 spin_lock(&head->lock);
5021 list_del_init(&poll->wait.entry);
5022 if (poll->wait.private)
5023 refcount_dec(&req->refs);
5025 spin_unlock(&head->lock);
5029 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5031 struct io_ring_ctx *ctx = req->ctx;
5033 io_poll_remove_double(req);
5034 req->poll.done = true;
5035 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5036 io_commit_cqring(ctx);
5039 static void io_poll_task_func(struct callback_head *cb)
5041 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5042 struct io_ring_ctx *ctx = req->ctx;
5043 struct io_kiocb *nxt;
5045 if (io_poll_rewait(req, &req->poll)) {
5046 spin_unlock_irq(&ctx->completion_lock);
5048 hash_del(&req->hash_node);
5049 io_poll_complete(req, req->result, 0);
5050 spin_unlock_irq(&ctx->completion_lock);
5052 nxt = io_put_req_find_next(req);
5053 io_cqring_ev_posted(ctx);
5055 __io_req_task_submit(nxt);
5058 percpu_ref_put(&ctx->refs);
5061 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5062 int sync, void *key)
5064 struct io_kiocb *req = wait->private;
5065 struct io_poll_iocb *poll = io_poll_get_single(req);
5066 __poll_t mask = key_to_poll(key);
5068 /* for instances that support it check for an event match first: */
5069 if (mask && !(mask & poll->events))
5072 list_del_init(&wait->entry);
5074 if (poll && poll->head) {
5077 spin_lock(&poll->head->lock);
5078 done = list_empty(&poll->wait.entry);
5080 list_del_init(&poll->wait.entry);
5081 /* make sure double remove sees this as being gone */
5082 wait->private = NULL;
5083 spin_unlock(&poll->head->lock);
5085 /* use wait func handler, so it matches the rq type */
5086 poll->wait.func(&poll->wait, mode, sync, key);
5089 refcount_dec(&req->refs);
5093 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5094 wait_queue_func_t wake_func)
5098 poll->canceled = false;
5099 poll->events = events;
5100 INIT_LIST_HEAD(&poll->wait.entry);
5101 init_waitqueue_func_entry(&poll->wait, wake_func);
5104 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5105 struct wait_queue_head *head,
5106 struct io_poll_iocb **poll_ptr)
5108 struct io_kiocb *req = pt->req;
5111 * The file being polled uses multiple waitqueues for poll handling
5112 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5115 if (unlikely(pt->nr_entries)) {
5116 struct io_poll_iocb *poll_one = poll;
5118 /* already have a 2nd entry, fail a third attempt */
5120 pt->error = -EINVAL;
5123 /* double add on the same waitqueue head, ignore */
5124 if (poll->head == head)
5126 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5128 pt->error = -ENOMEM;
5131 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5132 refcount_inc(&req->refs);
5133 poll->wait.private = req;
5140 if (poll->events & EPOLLEXCLUSIVE)
5141 add_wait_queue_exclusive(head, &poll->wait);
5143 add_wait_queue(head, &poll->wait);
5146 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5147 struct poll_table_struct *p)
5149 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5150 struct async_poll *apoll = pt->req->apoll;
5152 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5155 static void io_async_task_func(struct callback_head *cb)
5157 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5158 struct async_poll *apoll = req->apoll;
5159 struct io_ring_ctx *ctx = req->ctx;
5161 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5163 if (io_poll_rewait(req, &apoll->poll)) {
5164 spin_unlock_irq(&ctx->completion_lock);
5165 percpu_ref_put(&ctx->refs);
5169 /* If req is still hashed, it cannot have been canceled. Don't check. */
5170 if (hash_hashed(&req->hash_node))
5171 hash_del(&req->hash_node);
5173 io_poll_remove_double(req);
5174 spin_unlock_irq(&ctx->completion_lock);
5176 if (!READ_ONCE(apoll->poll.canceled))
5177 __io_req_task_submit(req);
5179 __io_req_task_cancel(req, -ECANCELED);
5181 percpu_ref_put(&ctx->refs);
5182 kfree(apoll->double_poll);
5186 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5189 struct io_kiocb *req = wait->private;
5190 struct io_poll_iocb *poll = &req->apoll->poll;
5192 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5195 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5198 static void io_poll_req_insert(struct io_kiocb *req)
5200 struct io_ring_ctx *ctx = req->ctx;
5201 struct hlist_head *list;
5203 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5204 hlist_add_head(&req->hash_node, list);
5207 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5208 struct io_poll_iocb *poll,
5209 struct io_poll_table *ipt, __poll_t mask,
5210 wait_queue_func_t wake_func)
5211 __acquires(&ctx->completion_lock)
5213 struct io_ring_ctx *ctx = req->ctx;
5214 bool cancel = false;
5216 INIT_HLIST_NODE(&req->hash_node);
5217 io_init_poll_iocb(poll, mask, wake_func);
5218 poll->file = req->file;
5219 poll->wait.private = req;
5221 ipt->pt._key = mask;
5224 ipt->nr_entries = 0;
5226 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5227 if (unlikely(!ipt->nr_entries) && !ipt->error)
5228 ipt->error = -EINVAL;
5230 spin_lock_irq(&ctx->completion_lock);
5232 io_poll_remove_double(req);
5233 if (likely(poll->head)) {
5234 spin_lock(&poll->head->lock);
5235 if (unlikely(list_empty(&poll->wait.entry))) {
5241 if (mask || ipt->error)
5242 list_del_init(&poll->wait.entry);
5244 WRITE_ONCE(poll->canceled, true);
5245 else if (!poll->done) /* actually waiting for an event */
5246 io_poll_req_insert(req);
5247 spin_unlock(&poll->head->lock);
5253 static bool io_arm_poll_handler(struct io_kiocb *req)
5255 const struct io_op_def *def = &io_op_defs[req->opcode];
5256 struct io_ring_ctx *ctx = req->ctx;
5257 struct async_poll *apoll;
5258 struct io_poll_table ipt;
5262 if (!req->file || !file_can_poll(req->file))
5264 if (req->flags & REQ_F_POLLED)
5268 else if (def->pollout)
5272 /* if we can't nonblock try, then no point in arming a poll handler */
5273 if (!io_file_supports_async(req->file, rw))
5276 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5277 if (unlikely(!apoll))
5279 apoll->double_poll = NULL;
5281 req->flags |= REQ_F_POLLED;
5286 mask |= POLLIN | POLLRDNORM;
5288 mask |= POLLOUT | POLLWRNORM;
5290 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5291 if ((req->opcode == IORING_OP_RECVMSG) &&
5292 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5295 mask |= POLLERR | POLLPRI;
5297 ipt.pt._qproc = io_async_queue_proc;
5299 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5301 if (ret || ipt.error) {
5302 io_poll_remove_double(req);
5303 spin_unlock_irq(&ctx->completion_lock);
5304 kfree(apoll->double_poll);
5308 spin_unlock_irq(&ctx->completion_lock);
5309 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5310 apoll->poll.events);
5314 static bool __io_poll_remove_one(struct io_kiocb *req,
5315 struct io_poll_iocb *poll)
5317 bool do_complete = false;
5319 spin_lock(&poll->head->lock);
5320 WRITE_ONCE(poll->canceled, true);
5321 if (!list_empty(&poll->wait.entry)) {
5322 list_del_init(&poll->wait.entry);
5325 spin_unlock(&poll->head->lock);
5326 hash_del(&req->hash_node);
5330 static bool io_poll_remove_one(struct io_kiocb *req)
5334 io_poll_remove_double(req);
5336 if (req->opcode == IORING_OP_POLL_ADD) {
5337 do_complete = __io_poll_remove_one(req, &req->poll);
5339 struct async_poll *apoll = req->apoll;
5341 /* non-poll requests have submit ref still */
5342 do_complete = __io_poll_remove_one(req, &apoll->poll);
5345 kfree(apoll->double_poll);
5351 io_cqring_fill_event(req, -ECANCELED);
5352 io_commit_cqring(req->ctx);
5353 req_set_fail_links(req);
5354 io_put_req_deferred(req, 1);
5361 * Returns true if we found and killed one or more poll requests
5363 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5364 struct files_struct *files)
5366 struct hlist_node *tmp;
5367 struct io_kiocb *req;
5370 spin_lock_irq(&ctx->completion_lock);
5371 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5372 struct hlist_head *list;
5374 list = &ctx->cancel_hash[i];
5375 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5376 if (io_match_task(req, tsk, files))
5377 posted += io_poll_remove_one(req);
5380 spin_unlock_irq(&ctx->completion_lock);
5383 io_cqring_ev_posted(ctx);
5388 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5390 struct hlist_head *list;
5391 struct io_kiocb *req;
5393 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5394 hlist_for_each_entry(req, list, hash_node) {
5395 if (sqe_addr != req->user_data)
5397 if (io_poll_remove_one(req))
5405 static int io_poll_remove_prep(struct io_kiocb *req,
5406 const struct io_uring_sqe *sqe)
5408 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5410 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5414 req->poll.addr = READ_ONCE(sqe->addr);
5419 * Find a running poll command that matches one specified in sqe->addr,
5420 * and remove it if found.
5422 static int io_poll_remove(struct io_kiocb *req)
5424 struct io_ring_ctx *ctx = req->ctx;
5428 addr = req->poll.addr;
5429 spin_lock_irq(&ctx->completion_lock);
5430 ret = io_poll_cancel(ctx, addr);
5431 spin_unlock_irq(&ctx->completion_lock);
5434 req_set_fail_links(req);
5435 io_req_complete(req, ret);
5439 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5442 struct io_kiocb *req = wait->private;
5443 struct io_poll_iocb *poll = &req->poll;
5445 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5448 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5449 struct poll_table_struct *p)
5451 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5453 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5456 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5458 struct io_poll_iocb *poll = &req->poll;
5461 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5463 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5466 events = READ_ONCE(sqe->poll32_events);
5468 events = swahw32(events);
5470 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5471 (events & EPOLLEXCLUSIVE);
5475 static int io_poll_add(struct io_kiocb *req)
5477 struct io_poll_iocb *poll = &req->poll;
5478 struct io_ring_ctx *ctx = req->ctx;
5479 struct io_poll_table ipt;
5482 ipt.pt._qproc = io_poll_queue_proc;
5484 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5487 if (mask) { /* no async, we'd stolen it */
5489 io_poll_complete(req, mask, 0);
5491 spin_unlock_irq(&ctx->completion_lock);
5494 io_cqring_ev_posted(ctx);
5500 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5502 struct io_timeout_data *data = container_of(timer,
5503 struct io_timeout_data, timer);
5504 struct io_kiocb *req = data->req;
5505 struct io_ring_ctx *ctx = req->ctx;
5506 unsigned long flags;
5508 spin_lock_irqsave(&ctx->completion_lock, flags);
5509 list_del_init(&req->timeout.list);
5510 atomic_set(&req->ctx->cq_timeouts,
5511 atomic_read(&req->ctx->cq_timeouts) + 1);
5513 io_cqring_fill_event(req, -ETIME);
5514 io_commit_cqring(ctx);
5515 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5517 io_cqring_ev_posted(ctx);
5518 req_set_fail_links(req);
5520 return HRTIMER_NORESTART;
5523 static int __io_timeout_cancel(struct io_kiocb *req)
5525 struct io_timeout_data *io = req->async_data;
5528 ret = hrtimer_try_to_cancel(&io->timer);
5531 list_del_init(&req->timeout.list);
5533 req_set_fail_links(req);
5534 io_cqring_fill_event(req, -ECANCELED);
5535 io_put_req_deferred(req, 1);
5539 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5541 struct io_kiocb *req;
5544 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5545 if (user_data == req->user_data) {
5554 return __io_timeout_cancel(req);
5557 static int io_timeout_remove_prep(struct io_kiocb *req,
5558 const struct io_uring_sqe *sqe)
5560 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5562 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5564 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags ||
5568 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5573 * Remove or update an existing timeout command
5575 static int io_timeout_remove(struct io_kiocb *req)
5577 struct io_ring_ctx *ctx = req->ctx;
5580 spin_lock_irq(&ctx->completion_lock);
5581 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5583 io_cqring_fill_event(req, ret);
5584 io_commit_cqring(ctx);
5585 spin_unlock_irq(&ctx->completion_lock);
5586 io_cqring_ev_posted(ctx);
5588 req_set_fail_links(req);
5593 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5594 bool is_timeout_link)
5596 struct io_timeout_data *data;
5598 u32 off = READ_ONCE(sqe->off);
5600 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5602 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
5605 if (off && is_timeout_link)
5607 flags = READ_ONCE(sqe->timeout_flags);
5608 if (flags & ~IORING_TIMEOUT_ABS)
5611 req->timeout.off = off;
5613 if (!req->async_data && io_alloc_async_data(req))
5616 data = req->async_data;
5619 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5622 if (flags & IORING_TIMEOUT_ABS)
5623 data->mode = HRTIMER_MODE_ABS;
5625 data->mode = HRTIMER_MODE_REL;
5627 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5631 static int io_timeout(struct io_kiocb *req)
5633 struct io_ring_ctx *ctx = req->ctx;
5634 struct io_timeout_data *data = req->async_data;
5635 struct list_head *entry;
5636 u32 tail, off = req->timeout.off;
5638 spin_lock_irq(&ctx->completion_lock);
5641 * sqe->off holds how many events that need to occur for this
5642 * timeout event to be satisfied. If it isn't set, then this is
5643 * a pure timeout request, sequence isn't used.
5645 if (io_is_timeout_noseq(req)) {
5646 entry = ctx->timeout_list.prev;
5650 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5651 req->timeout.target_seq = tail + off;
5653 /* Update the last seq here in case io_flush_timeouts() hasn't.
5654 * This is safe because ->completion_lock is held, and submissions
5655 * and completions are never mixed in the same ->completion_lock section.
5657 ctx->cq_last_tm_flush = tail;
5660 * Insertion sort, ensuring the first entry in the list is always
5661 * the one we need first.
5663 list_for_each_prev(entry, &ctx->timeout_list) {
5664 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5667 if (io_is_timeout_noseq(nxt))
5669 /* nxt.seq is behind @tail, otherwise would've been completed */
5670 if (off >= nxt->timeout.target_seq - tail)
5674 list_add(&req->timeout.list, entry);
5675 data->timer.function = io_timeout_fn;
5676 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5677 spin_unlock_irq(&ctx->completion_lock);
5681 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5683 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5685 return req->user_data == (unsigned long) data;
5688 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5690 enum io_wq_cancel cancel_ret;
5693 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5694 switch (cancel_ret) {
5695 case IO_WQ_CANCEL_OK:
5698 case IO_WQ_CANCEL_RUNNING:
5701 case IO_WQ_CANCEL_NOTFOUND:
5709 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5710 struct io_kiocb *req, __u64 sqe_addr,
5713 unsigned long flags;
5716 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5717 if (ret != -ENOENT) {
5718 spin_lock_irqsave(&ctx->completion_lock, flags);
5722 spin_lock_irqsave(&ctx->completion_lock, flags);
5723 ret = io_timeout_cancel(ctx, sqe_addr);
5726 ret = io_poll_cancel(ctx, sqe_addr);
5730 io_cqring_fill_event(req, ret);
5731 io_commit_cqring(ctx);
5732 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5733 io_cqring_ev_posted(ctx);
5736 req_set_fail_links(req);
5740 static int io_async_cancel_prep(struct io_kiocb *req,
5741 const struct io_uring_sqe *sqe)
5743 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5745 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5747 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
5751 req->cancel.addr = READ_ONCE(sqe->addr);
5755 static int io_async_cancel(struct io_kiocb *req)
5757 struct io_ring_ctx *ctx = req->ctx;
5759 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5763 static int io_files_update_prep(struct io_kiocb *req,
5764 const struct io_uring_sqe *sqe)
5766 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5768 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5770 if (sqe->ioprio || sqe->rw_flags)
5773 req->files_update.offset = READ_ONCE(sqe->off);
5774 req->files_update.nr_args = READ_ONCE(sqe->len);
5775 if (!req->files_update.nr_args)
5777 req->files_update.arg = READ_ONCE(sqe->addr);
5781 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5782 struct io_comp_state *cs)
5784 struct io_ring_ctx *ctx = req->ctx;
5785 struct io_uring_files_update up;
5791 up.offset = req->files_update.offset;
5792 up.fds = req->files_update.arg;
5794 mutex_lock(&ctx->uring_lock);
5795 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5796 mutex_unlock(&ctx->uring_lock);
5799 req_set_fail_links(req);
5800 __io_req_complete(req, ret, 0, cs);
5804 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5806 switch (req->opcode) {
5809 case IORING_OP_READV:
5810 case IORING_OP_READ_FIXED:
5811 case IORING_OP_READ:
5812 return io_read_prep(req, sqe);
5813 case IORING_OP_WRITEV:
5814 case IORING_OP_WRITE_FIXED:
5815 case IORING_OP_WRITE:
5816 return io_write_prep(req, sqe);
5817 case IORING_OP_POLL_ADD:
5818 return io_poll_add_prep(req, sqe);
5819 case IORING_OP_POLL_REMOVE:
5820 return io_poll_remove_prep(req, sqe);
5821 case IORING_OP_FSYNC:
5822 return io_prep_fsync(req, sqe);
5823 case IORING_OP_SYNC_FILE_RANGE:
5824 return io_prep_sfr(req, sqe);
5825 case IORING_OP_SENDMSG:
5826 case IORING_OP_SEND:
5827 return io_sendmsg_prep(req, sqe);
5828 case IORING_OP_RECVMSG:
5829 case IORING_OP_RECV:
5830 return io_recvmsg_prep(req, sqe);
5831 case IORING_OP_CONNECT:
5832 return io_connect_prep(req, sqe);
5833 case IORING_OP_TIMEOUT:
5834 return io_timeout_prep(req, sqe, false);
5835 case IORING_OP_TIMEOUT_REMOVE:
5836 return io_timeout_remove_prep(req, sqe);
5837 case IORING_OP_ASYNC_CANCEL:
5838 return io_async_cancel_prep(req, sqe);
5839 case IORING_OP_LINK_TIMEOUT:
5840 return io_timeout_prep(req, sqe, true);
5841 case IORING_OP_ACCEPT:
5842 return io_accept_prep(req, sqe);
5843 case IORING_OP_FALLOCATE:
5844 return io_fallocate_prep(req, sqe);
5845 case IORING_OP_OPENAT:
5846 return io_openat_prep(req, sqe);
5847 case IORING_OP_CLOSE:
5848 return io_close_prep(req, sqe);
5849 case IORING_OP_FILES_UPDATE:
5850 return io_files_update_prep(req, sqe);
5851 case IORING_OP_STATX:
5852 return io_statx_prep(req, sqe);
5853 case IORING_OP_FADVISE:
5854 return io_fadvise_prep(req, sqe);
5855 case IORING_OP_MADVISE:
5856 return io_madvise_prep(req, sqe);
5857 case IORING_OP_OPENAT2:
5858 return io_openat2_prep(req, sqe);
5859 case IORING_OP_EPOLL_CTL:
5860 return io_epoll_ctl_prep(req, sqe);
5861 case IORING_OP_SPLICE:
5862 return io_splice_prep(req, sqe);
5863 case IORING_OP_PROVIDE_BUFFERS:
5864 return io_provide_buffers_prep(req, sqe);
5865 case IORING_OP_REMOVE_BUFFERS:
5866 return io_remove_buffers_prep(req, sqe);
5868 return io_tee_prep(req, sqe);
5871 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5876 static int io_req_defer_prep(struct io_kiocb *req,
5877 const struct io_uring_sqe *sqe)
5881 if (io_alloc_async_data(req))
5883 return io_req_prep(req, sqe);
5886 static u32 io_get_sequence(struct io_kiocb *req)
5888 struct io_kiocb *pos;
5889 struct io_ring_ctx *ctx = req->ctx;
5890 u32 total_submitted, nr_reqs = 1;
5892 if (req->flags & REQ_F_LINK_HEAD)
5893 list_for_each_entry(pos, &req->link_list, link_list)
5896 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5897 return total_submitted - nr_reqs;
5900 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5902 struct io_ring_ctx *ctx = req->ctx;
5903 struct io_defer_entry *de;
5907 /* Still need defer if there is pending req in defer list. */
5908 if (likely(list_empty_careful(&ctx->defer_list) &&
5909 !(req->flags & REQ_F_IO_DRAIN)))
5912 seq = io_get_sequence(req);
5913 /* Still a chance to pass the sequence check */
5914 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5917 if (!req->async_data) {
5918 ret = io_req_defer_prep(req, sqe);
5922 io_prep_async_link(req);
5923 de = kmalloc(sizeof(*de), GFP_KERNEL);
5927 spin_lock_irq(&ctx->completion_lock);
5928 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5929 spin_unlock_irq(&ctx->completion_lock);
5931 io_queue_async_work(req);
5932 return -EIOCBQUEUED;
5935 trace_io_uring_defer(ctx, req, req->user_data);
5938 list_add_tail(&de->list, &ctx->defer_list);
5939 spin_unlock_irq(&ctx->completion_lock);
5940 return -EIOCBQUEUED;
5943 static void io_req_drop_files(struct io_kiocb *req)
5945 struct io_ring_ctx *ctx = req->ctx;
5946 struct io_uring_task *tctx = req->task->io_uring;
5947 unsigned long flags;
5949 if (req->work.flags & IO_WQ_WORK_FILES) {
5950 put_files_struct(req->work.identity->files);
5951 put_nsproxy(req->work.identity->nsproxy);
5953 spin_lock_irqsave(&ctx->inflight_lock, flags);
5954 list_del(&req->inflight_entry);
5955 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5956 req->flags &= ~REQ_F_INFLIGHT;
5957 req->work.flags &= ~IO_WQ_WORK_FILES;
5958 if (atomic_read(&tctx->in_idle))
5959 wake_up(&tctx->wait);
5962 static void __io_clean_op(struct io_kiocb *req)
5964 if (req->flags & REQ_F_BUFFER_SELECTED) {
5965 switch (req->opcode) {
5966 case IORING_OP_READV:
5967 case IORING_OP_READ_FIXED:
5968 case IORING_OP_READ:
5969 kfree((void *)(unsigned long)req->rw.addr);
5971 case IORING_OP_RECVMSG:
5972 case IORING_OP_RECV:
5973 kfree(req->sr_msg.kbuf);
5976 req->flags &= ~REQ_F_BUFFER_SELECTED;
5979 if (req->flags & REQ_F_NEED_CLEANUP) {
5980 switch (req->opcode) {
5981 case IORING_OP_READV:
5982 case IORING_OP_READ_FIXED:
5983 case IORING_OP_READ:
5984 case IORING_OP_WRITEV:
5985 case IORING_OP_WRITE_FIXED:
5986 case IORING_OP_WRITE: {
5987 struct io_async_rw *io = req->async_data;
5989 kfree(io->free_iovec);
5992 case IORING_OP_RECVMSG:
5993 case IORING_OP_SENDMSG: {
5994 struct io_async_msghdr *io = req->async_data;
5995 if (io->iov != io->fast_iov)
5999 case IORING_OP_SPLICE:
6001 io_put_file(req, req->splice.file_in,
6002 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6004 case IORING_OP_OPENAT:
6005 case IORING_OP_OPENAT2:
6006 if (req->open.filename)
6007 putname(req->open.filename);
6010 req->flags &= ~REQ_F_NEED_CLEANUP;
6014 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
6015 struct io_comp_state *cs)
6017 struct io_ring_ctx *ctx = req->ctx;
6020 switch (req->opcode) {
6022 ret = io_nop(req, cs);
6024 case IORING_OP_READV:
6025 case IORING_OP_READ_FIXED:
6026 case IORING_OP_READ:
6027 ret = io_read(req, force_nonblock, cs);
6029 case IORING_OP_WRITEV:
6030 case IORING_OP_WRITE_FIXED:
6031 case IORING_OP_WRITE:
6032 ret = io_write(req, force_nonblock, cs);
6034 case IORING_OP_FSYNC:
6035 ret = io_fsync(req, force_nonblock);
6037 case IORING_OP_POLL_ADD:
6038 ret = io_poll_add(req);
6040 case IORING_OP_POLL_REMOVE:
6041 ret = io_poll_remove(req);
6043 case IORING_OP_SYNC_FILE_RANGE:
6044 ret = io_sync_file_range(req, force_nonblock);
6046 case IORING_OP_SENDMSG:
6047 ret = io_sendmsg(req, force_nonblock, cs);
6049 case IORING_OP_SEND:
6050 ret = io_send(req, force_nonblock, cs);
6052 case IORING_OP_RECVMSG:
6053 ret = io_recvmsg(req, force_nonblock, cs);
6055 case IORING_OP_RECV:
6056 ret = io_recv(req, force_nonblock, cs);
6058 case IORING_OP_TIMEOUT:
6059 ret = io_timeout(req);
6061 case IORING_OP_TIMEOUT_REMOVE:
6062 ret = io_timeout_remove(req);
6064 case IORING_OP_ACCEPT:
6065 ret = io_accept(req, force_nonblock, cs);
6067 case IORING_OP_CONNECT:
6068 ret = io_connect(req, force_nonblock, cs);
6070 case IORING_OP_ASYNC_CANCEL:
6071 ret = io_async_cancel(req);
6073 case IORING_OP_FALLOCATE:
6074 ret = io_fallocate(req, force_nonblock);
6076 case IORING_OP_OPENAT:
6077 ret = io_openat(req, force_nonblock);
6079 case IORING_OP_CLOSE:
6080 ret = io_close(req, force_nonblock, cs);
6082 case IORING_OP_FILES_UPDATE:
6083 ret = io_files_update(req, force_nonblock, cs);
6085 case IORING_OP_STATX:
6086 ret = io_statx(req, force_nonblock);
6088 case IORING_OP_FADVISE:
6089 ret = io_fadvise(req, force_nonblock);
6091 case IORING_OP_MADVISE:
6092 ret = io_madvise(req, force_nonblock);
6094 case IORING_OP_OPENAT2:
6095 ret = io_openat2(req, force_nonblock);
6097 case IORING_OP_EPOLL_CTL:
6098 ret = io_epoll_ctl(req, force_nonblock, cs);
6100 case IORING_OP_SPLICE:
6101 ret = io_splice(req, force_nonblock);
6103 case IORING_OP_PROVIDE_BUFFERS:
6104 ret = io_provide_buffers(req, force_nonblock, cs);
6106 case IORING_OP_REMOVE_BUFFERS:
6107 ret = io_remove_buffers(req, force_nonblock, cs);
6110 ret = io_tee(req, force_nonblock);
6120 /* If the op doesn't have a file, we're not polling for it */
6121 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6122 const bool in_async = io_wq_current_is_worker();
6124 /* workqueue context doesn't hold uring_lock, grab it now */
6126 mutex_lock(&ctx->uring_lock);
6128 io_iopoll_req_issued(req);
6131 mutex_unlock(&ctx->uring_lock);
6137 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6139 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6140 struct io_kiocb *timeout;
6143 timeout = io_prep_linked_timeout(req);
6145 io_queue_linked_timeout(timeout);
6147 /* if NO_CANCEL is set, we must still run the work */
6148 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6149 IO_WQ_WORK_CANCEL) {
6155 ret = io_issue_sqe(req, false, NULL);
6157 * We can get EAGAIN for polled IO even though we're
6158 * forcing a sync submission from here, since we can't
6159 * wait for request slots on the block side.
6168 struct io_ring_ctx *lock_ctx = NULL;
6170 if (req->ctx->flags & IORING_SETUP_IOPOLL)
6171 lock_ctx = req->ctx;
6174 * io_iopoll_complete() does not hold completion_lock to
6175 * complete polled io, so here for polled io, we can not call
6176 * io_req_complete() directly, otherwise there maybe concurrent
6177 * access to cqring, defer_list, etc, which is not safe. Given
6178 * that io_iopoll_complete() is always called under uring_lock,
6179 * so here for polled io, we also get uring_lock to complete
6183 mutex_lock(&lock_ctx->uring_lock);
6185 req_set_fail_links(req);
6186 io_req_complete(req, ret);
6189 mutex_unlock(&lock_ctx->uring_lock);
6192 return io_steal_work(req);
6195 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6198 struct fixed_file_table *table;
6200 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6201 return table->files[index & IORING_FILE_TABLE_MASK];
6204 static struct file *io_file_get(struct io_submit_state *state,
6205 struct io_kiocb *req, int fd, bool fixed)
6207 struct io_ring_ctx *ctx = req->ctx;
6211 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6213 fd = array_index_nospec(fd, ctx->nr_user_files);
6214 file = io_file_from_index(ctx, fd);
6216 req->fixed_file_refs = &ctx->file_data->node->refs;
6217 percpu_ref_get(req->fixed_file_refs);
6220 trace_io_uring_file_get(ctx, fd);
6221 file = __io_file_get(state, fd);
6224 if (file && file->f_op == &io_uring_fops &&
6225 !(req->flags & REQ_F_INFLIGHT)) {
6226 io_req_init_async(req);
6227 req->flags |= REQ_F_INFLIGHT;
6229 spin_lock_irq(&ctx->inflight_lock);
6230 list_add(&req->inflight_entry, &ctx->inflight_list);
6231 spin_unlock_irq(&ctx->inflight_lock);
6237 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6242 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6243 if (unlikely(!fixed && io_async_submit(req->ctx)))
6246 req->file = io_file_get(state, req, fd, fixed);
6247 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6252 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6254 struct io_timeout_data *data = container_of(timer,
6255 struct io_timeout_data, timer);
6256 struct io_kiocb *req = data->req;
6257 struct io_ring_ctx *ctx = req->ctx;
6258 struct io_kiocb *prev = NULL;
6259 unsigned long flags;
6261 spin_lock_irqsave(&ctx->completion_lock, flags);
6264 * We don't expect the list to be empty, that will only happen if we
6265 * race with the completion of the linked work.
6267 if (!list_empty(&req->link_list)) {
6268 prev = list_entry(req->link_list.prev, struct io_kiocb,
6270 if (refcount_inc_not_zero(&prev->refs))
6271 list_del_init(&req->link_list);
6276 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6279 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6280 io_put_req_deferred(prev, 1);
6282 io_cqring_add_event(req, -ETIME, 0);
6283 io_put_req_deferred(req, 1);
6285 return HRTIMER_NORESTART;
6288 static void __io_queue_linked_timeout(struct io_kiocb *req)
6291 * If the list is now empty, then our linked request finished before
6292 * we got a chance to setup the timer
6294 if (!list_empty(&req->link_list)) {
6295 struct io_timeout_data *data = req->async_data;
6297 data->timer.function = io_link_timeout_fn;
6298 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6303 static void io_queue_linked_timeout(struct io_kiocb *req)
6305 struct io_ring_ctx *ctx = req->ctx;
6307 spin_lock_irq(&ctx->completion_lock);
6308 __io_queue_linked_timeout(req);
6309 spin_unlock_irq(&ctx->completion_lock);
6311 /* drop submission reference */
6315 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6317 struct io_kiocb *nxt;
6319 if (!(req->flags & REQ_F_LINK_HEAD))
6321 if (req->flags & REQ_F_LINK_TIMEOUT)
6324 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6326 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6329 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6330 req->flags |= REQ_F_LINK_TIMEOUT;
6334 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6336 struct io_kiocb *linked_timeout;
6337 const struct cred *old_creds = NULL;
6341 linked_timeout = io_prep_linked_timeout(req);
6343 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6344 (req->work.flags & IO_WQ_WORK_CREDS) &&
6345 req->work.identity->creds != current_cred()) {
6347 revert_creds(old_creds);
6348 if (old_creds == req->work.identity->creds)
6349 old_creds = NULL; /* restored original creds */
6351 old_creds = override_creds(req->work.identity->creds);
6354 ret = io_issue_sqe(req, true, cs);
6357 * We async punt it if the file wasn't marked NOWAIT, or if the file
6358 * doesn't support non-blocking read/write attempts
6360 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6361 if (!io_arm_poll_handler(req)) {
6363 * Queued up for async execution, worker will release
6364 * submit reference when the iocb is actually submitted.
6366 io_queue_async_work(req);
6370 io_queue_linked_timeout(linked_timeout);
6371 } else if (likely(!ret)) {
6372 /* drop submission reference */
6373 req = io_put_req_find_next(req);
6375 io_queue_linked_timeout(linked_timeout);
6378 if (!(req->flags & REQ_F_FORCE_ASYNC))
6380 io_queue_async_work(req);
6383 /* un-prep timeout, so it'll be killed as any other linked */
6384 req->flags &= ~REQ_F_LINK_TIMEOUT;
6385 req_set_fail_links(req);
6387 io_req_complete(req, ret);
6391 revert_creds(old_creds);
6394 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6395 struct io_comp_state *cs)
6399 ret = io_req_defer(req, sqe);
6401 if (ret != -EIOCBQUEUED) {
6403 req_set_fail_links(req);
6405 io_req_complete(req, ret);
6407 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6408 if (!req->async_data) {
6409 ret = io_req_defer_prep(req, sqe);
6413 io_queue_async_work(req);
6416 ret = io_req_prep(req, sqe);
6420 __io_queue_sqe(req, cs);
6424 static inline void io_queue_link_head(struct io_kiocb *req,
6425 struct io_comp_state *cs)
6427 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6429 io_req_complete(req, -ECANCELED);
6431 io_queue_sqe(req, NULL, cs);
6434 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6435 struct io_kiocb **link, struct io_comp_state *cs)
6437 struct io_ring_ctx *ctx = req->ctx;
6441 * If we already have a head request, queue this one for async
6442 * submittal once the head completes. If we don't have a head but
6443 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6444 * submitted sync once the chain is complete. If none of those
6445 * conditions are true (normal request), then just queue it.
6448 struct io_kiocb *head = *link;
6451 * Taking sequential execution of a link, draining both sides
6452 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6453 * requests in the link. So, it drains the head and the
6454 * next after the link request. The last one is done via
6455 * drain_next flag to persist the effect across calls.
6457 if (req->flags & REQ_F_IO_DRAIN) {
6458 head->flags |= REQ_F_IO_DRAIN;
6459 ctx->drain_next = 1;
6461 ret = io_req_defer_prep(req, sqe);
6462 if (unlikely(ret)) {
6463 /* fail even hard links since we don't submit */
6464 head->flags |= REQ_F_FAIL_LINK;
6467 trace_io_uring_link(ctx, req, head);
6468 list_add_tail(&req->link_list, &head->link_list);
6470 /* last request of a link, enqueue the link */
6471 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6472 io_queue_link_head(head, cs);
6476 if (unlikely(ctx->drain_next)) {
6477 req->flags |= REQ_F_IO_DRAIN;
6478 ctx->drain_next = 0;
6480 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6481 req->flags |= REQ_F_LINK_HEAD;
6482 INIT_LIST_HEAD(&req->link_list);
6484 ret = io_req_defer_prep(req, sqe);
6486 req->flags |= REQ_F_FAIL_LINK;
6489 io_queue_sqe(req, sqe, cs);
6497 * Batched submission is done, ensure local IO is flushed out.
6499 static void io_submit_state_end(struct io_submit_state *state)
6501 if (!list_empty(&state->comp.list))
6502 io_submit_flush_completions(&state->comp);
6503 blk_finish_plug(&state->plug);
6504 io_state_file_put(state);
6505 if (state->free_reqs)
6506 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6510 * Start submission side cache.
6512 static void io_submit_state_start(struct io_submit_state *state,
6513 struct io_ring_ctx *ctx, unsigned int max_ios)
6515 blk_start_plug(&state->plug);
6517 INIT_LIST_HEAD(&state->comp.list);
6518 state->comp.ctx = ctx;
6519 state->free_reqs = 0;
6521 state->ios_left = max_ios;
6524 static void io_commit_sqring(struct io_ring_ctx *ctx)
6526 struct io_rings *rings = ctx->rings;
6529 * Ensure any loads from the SQEs are done at this point,
6530 * since once we write the new head, the application could
6531 * write new data to them.
6533 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6537 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6538 * that is mapped by userspace. This means that care needs to be taken to
6539 * ensure that reads are stable, as we cannot rely on userspace always
6540 * being a good citizen. If members of the sqe are validated and then later
6541 * used, it's important that those reads are done through READ_ONCE() to
6542 * prevent a re-load down the line.
6544 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6546 u32 *sq_array = ctx->sq_array;
6550 * The cached sq head (or cq tail) serves two purposes:
6552 * 1) allows us to batch the cost of updating the user visible
6554 * 2) allows the kernel side to track the head on its own, even
6555 * though the application is the one updating it.
6557 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6558 if (likely(head < ctx->sq_entries))
6559 return &ctx->sq_sqes[head];
6561 /* drop invalid entries */
6562 ctx->cached_sq_dropped++;
6563 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6567 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6569 ctx->cached_sq_head++;
6573 * Check SQE restrictions (opcode and flags).
6575 * Returns 'true' if SQE is allowed, 'false' otherwise.
6577 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6578 struct io_kiocb *req,
6579 unsigned int sqe_flags)
6581 if (!ctx->restricted)
6584 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6587 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6588 ctx->restrictions.sqe_flags_required)
6591 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6592 ctx->restrictions.sqe_flags_required))
6598 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6599 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6600 IOSQE_BUFFER_SELECT)
6602 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6603 const struct io_uring_sqe *sqe,
6604 struct io_submit_state *state)
6606 unsigned int sqe_flags;
6609 req->opcode = READ_ONCE(sqe->opcode);
6610 req->user_data = READ_ONCE(sqe->user_data);
6611 req->async_data = NULL;
6615 /* one is dropped after submission, the other at completion */
6616 refcount_set(&req->refs, 2);
6617 req->task = current;
6620 if (unlikely(req->opcode >= IORING_OP_LAST))
6623 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6626 sqe_flags = READ_ONCE(sqe->flags);
6627 /* enforce forwards compatibility on users */
6628 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6631 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6634 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6635 !io_op_defs[req->opcode].buffer_select)
6638 id = READ_ONCE(sqe->personality);
6640 struct io_identity *iod;
6642 iod = xa_load(&ctx->personalities, id);
6645 refcount_inc(&iod->count);
6647 __io_req_init_async(req);
6648 get_cred(iod->creds);
6649 req->work.identity = iod;
6650 req->work.flags |= IO_WQ_WORK_CREDS;
6653 /* same numerical values with corresponding REQ_F_*, safe to copy */
6654 req->flags |= sqe_flags;
6656 if (!io_op_defs[req->opcode].needs_file)
6659 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6664 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6666 struct io_submit_state state;
6667 struct io_kiocb *link = NULL;
6668 int i, submitted = 0;
6670 /* if we have a backlog and couldn't flush it all, return BUSY */
6671 if (test_bit(0, &ctx->sq_check_overflow)) {
6672 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6676 /* make sure SQ entry isn't read before tail */
6677 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6679 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6682 percpu_counter_add(¤t->io_uring->inflight, nr);
6683 refcount_add(nr, ¤t->usage);
6685 io_submit_state_start(&state, ctx, nr);
6687 for (i = 0; i < nr; i++) {
6688 const struct io_uring_sqe *sqe;
6689 struct io_kiocb *req;
6692 sqe = io_get_sqe(ctx);
6693 if (unlikely(!sqe)) {
6694 io_consume_sqe(ctx);
6697 req = io_alloc_req(ctx, &state);
6698 if (unlikely(!req)) {
6700 submitted = -EAGAIN;
6703 io_consume_sqe(ctx);
6704 /* will complete beyond this point, count as submitted */
6707 err = io_init_req(ctx, req, sqe, &state);
6708 if (unlikely(err)) {
6711 io_req_complete(req, err);
6715 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6716 true, io_async_submit(ctx));
6717 err = io_submit_sqe(req, sqe, &link, &state.comp);
6722 if (unlikely(submitted != nr)) {
6723 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6724 struct io_uring_task *tctx = current->io_uring;
6725 int unused = nr - ref_used;
6727 percpu_ref_put_many(&ctx->refs, unused);
6728 percpu_counter_sub(&tctx->inflight, unused);
6729 put_task_struct_many(current, unused);
6732 io_queue_link_head(link, &state.comp);
6733 io_submit_state_end(&state);
6735 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6736 io_commit_sqring(ctx);
6741 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6743 /* Tell userspace we may need a wakeup call */
6744 spin_lock_irq(&ctx->completion_lock);
6745 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6746 spin_unlock_irq(&ctx->completion_lock);
6749 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6751 spin_lock_irq(&ctx->completion_lock);
6752 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6753 spin_unlock_irq(&ctx->completion_lock);
6756 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6757 int sync, void *key)
6759 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6762 ret = autoremove_wake_function(wqe, mode, sync, key);
6764 unsigned long flags;
6766 spin_lock_irqsave(&ctx->completion_lock, flags);
6767 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6768 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6779 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6780 unsigned long start_jiffies, bool cap_entries)
6782 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6783 struct io_sq_data *sqd = ctx->sq_data;
6784 unsigned int to_submit;
6788 if (!list_empty(&ctx->iopoll_list)) {
6789 unsigned nr_events = 0;
6791 mutex_lock(&ctx->uring_lock);
6792 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6793 io_do_iopoll(ctx, &nr_events, 0);
6794 mutex_unlock(&ctx->uring_lock);
6797 to_submit = io_sqring_entries(ctx);
6800 * If submit got -EBUSY, flag us as needing the application
6801 * to enter the kernel to reap and flush events.
6803 if (!to_submit || ret == -EBUSY || need_resched()) {
6805 * Drop cur_mm before scheduling, we can't hold it for
6806 * long periods (or over schedule()). Do this before
6807 * adding ourselves to the waitqueue, as the unuse/drop
6810 io_sq_thread_drop_mm();
6813 * We're polling. If we're within the defined idle
6814 * period, then let us spin without work before going
6815 * to sleep. The exception is if we got EBUSY doing
6816 * more IO, we should wait for the application to
6817 * reap events and wake us up.
6819 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6820 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6821 !percpu_ref_is_dying(&ctx->refs)))
6824 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6825 TASK_INTERRUPTIBLE);
6828 * While doing polled IO, before going to sleep, we need
6829 * to check if there are new reqs added to iopoll_list,
6830 * it is because reqs may have been punted to io worker
6831 * and will be added to iopoll_list later, hence check
6832 * the iopoll_list again.
6834 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6835 !list_empty_careful(&ctx->iopoll_list)) {
6836 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6840 to_submit = io_sqring_entries(ctx);
6841 if (!to_submit || ret == -EBUSY)
6845 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6846 io_ring_clear_wakeup_flag(ctx);
6848 /* if we're handling multiple rings, cap submit size for fairness */
6849 if (cap_entries && to_submit > 8)
6852 mutex_lock(&ctx->uring_lock);
6853 if (likely(!percpu_ref_is_dying(&ctx->refs) && !ctx->sqo_dead))
6854 ret = io_submit_sqes(ctx, to_submit);
6855 mutex_unlock(&ctx->uring_lock);
6857 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6858 wake_up(&ctx->sqo_sq_wait);
6860 return SQT_DID_WORK;
6863 static void io_sqd_init_new(struct io_sq_data *sqd)
6865 struct io_ring_ctx *ctx;
6867 while (!list_empty(&sqd->ctx_new_list)) {
6868 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6869 init_wait(&ctx->sqo_wait_entry);
6870 ctx->sqo_wait_entry.func = io_sq_wake_function;
6871 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6872 complete(&ctx->sq_thread_comp);
6876 static int io_sq_thread(void *data)
6878 struct cgroup_subsys_state *cur_css = NULL;
6879 const struct cred *old_cred = NULL;
6880 struct io_sq_data *sqd = data;
6881 struct io_ring_ctx *ctx;
6882 unsigned long start_jiffies;
6884 start_jiffies = jiffies;
6885 while (!kthread_should_stop()) {
6886 enum sq_ret ret = 0;
6890 * Any changes to the sqd lists are synchronized through the
6891 * kthread parking. This synchronizes the thread vs users,
6892 * the users are synchronized on the sqd->ctx_lock.
6894 if (kthread_should_park()) {
6897 * When sq thread is unparked, in case the previous park operation
6898 * comes from io_put_sq_data(), which means that sq thread is going
6899 * to be stopped, so here needs to have a check.
6901 if (kthread_should_stop())
6905 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6906 io_sqd_init_new(sqd);
6908 cap_entries = !list_is_singular(&sqd->ctx_list);
6910 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6911 if (current->cred != ctx->creds) {
6913 revert_creds(old_cred);
6914 old_cred = override_creds(ctx->creds);
6916 io_sq_thread_associate_blkcg(ctx, &cur_css);
6918 current->loginuid = ctx->loginuid;
6919 current->sessionid = ctx->sessionid;
6922 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6924 io_sq_thread_drop_mm();
6927 if (ret & SQT_SPIN) {
6929 io_sq_thread_drop_mm();
6931 } else if (ret == SQT_IDLE) {
6932 if (kthread_should_park())
6934 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6935 io_ring_set_wakeup_flag(ctx);
6937 start_jiffies = jiffies;
6938 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6939 io_ring_clear_wakeup_flag(ctx);
6944 io_sq_thread_drop_mm();
6947 io_sq_thread_unassociate_blkcg();
6949 revert_creds(old_cred);
6956 struct io_wait_queue {
6957 struct wait_queue_entry wq;
6958 struct io_ring_ctx *ctx;
6960 unsigned nr_timeouts;
6963 static inline bool io_should_wake(struct io_wait_queue *iowq)
6965 struct io_ring_ctx *ctx = iowq->ctx;
6968 * Wake up if we have enough events, or if a timeout occurred since we
6969 * started waiting. For timeouts, we always want to return to userspace,
6970 * regardless of event count.
6972 return io_cqring_events(ctx) >= iowq->to_wait ||
6973 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6976 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6977 int wake_flags, void *key)
6979 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6983 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6984 * the task, and the next invocation will do it.
6986 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6987 return autoremove_wake_function(curr, mode, wake_flags, key);
6991 static int io_run_task_work_sig(void)
6993 if (io_run_task_work())
6995 if (!signal_pending(current))
6997 if (current->jobctl & JOBCTL_TASK_WORK) {
6998 spin_lock_irq(¤t->sighand->siglock);
6999 current->jobctl &= ~JOBCTL_TASK_WORK;
7000 recalc_sigpending();
7001 spin_unlock_irq(¤t->sighand->siglock);
7008 * Wait until events become available, if we don't already have some. The
7009 * application must reap them itself, as they reside on the shared cq ring.
7011 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7012 const sigset_t __user *sig, size_t sigsz)
7014 struct io_wait_queue iowq = {
7017 .func = io_wake_function,
7018 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7021 .to_wait = min_events,
7023 struct io_rings *rings = ctx->rings;
7027 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7028 if (io_cqring_events(ctx) >= min_events)
7030 if (!io_run_task_work())
7035 #ifdef CONFIG_COMPAT
7036 if (in_compat_syscall())
7037 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7041 ret = set_user_sigmask(sig, sigsz);
7047 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7048 trace_io_uring_cqring_wait(ctx, min_events);
7050 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7051 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7052 TASK_INTERRUPTIBLE);
7053 /* make sure we run task_work before checking for signals */
7054 ret = io_run_task_work_sig();
7056 finish_wait(&ctx->wait, &iowq.wq);
7061 if (io_should_wake(&iowq))
7063 if (test_bit(0, &ctx->cq_check_overflow)) {
7064 finish_wait(&ctx->wait, &iowq.wq);
7069 finish_wait(&ctx->wait, &iowq.wq);
7071 restore_saved_sigmask_unless(ret == -EINTR);
7073 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7076 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7078 #if defined(CONFIG_UNIX)
7079 if (ctx->ring_sock) {
7080 struct sock *sock = ctx->ring_sock->sk;
7081 struct sk_buff *skb;
7083 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7089 for (i = 0; i < ctx->nr_user_files; i++) {
7092 file = io_file_from_index(ctx, i);
7099 static void io_file_ref_kill(struct percpu_ref *ref)
7101 struct fixed_file_data *data;
7103 data = container_of(ref, struct fixed_file_data, refs);
7104 complete(&data->done);
7107 static void io_sqe_files_set_node(struct fixed_file_data *file_data,
7108 struct fixed_file_ref_node *ref_node)
7110 spin_lock_bh(&file_data->lock);
7111 file_data->node = ref_node;
7112 list_add_tail(&ref_node->node, &file_data->ref_list);
7113 spin_unlock_bh(&file_data->lock);
7114 percpu_ref_get(&file_data->refs);
7117 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7119 struct fixed_file_data *data = ctx->file_data;
7120 struct fixed_file_ref_node *backup_node, *ref_node = NULL;
7121 unsigned nr_tables, i;
7126 backup_node = alloc_fixed_file_ref_node(ctx);
7130 spin_lock_bh(&data->lock);
7131 ref_node = data->node;
7132 spin_unlock_bh(&data->lock);
7134 percpu_ref_kill(&ref_node->refs);
7136 percpu_ref_kill(&data->refs);
7138 /* wait for all refs nodes to complete */
7139 flush_delayed_work(&ctx->file_put_work);
7141 ret = wait_for_completion_interruptible(&data->done);
7144 ret = io_run_task_work_sig();
7146 percpu_ref_resurrect(&data->refs);
7147 reinit_completion(&data->done);
7148 io_sqe_files_set_node(data, backup_node);
7153 __io_sqe_files_unregister(ctx);
7154 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7155 for (i = 0; i < nr_tables; i++)
7156 kfree(data->table[i].files);
7158 percpu_ref_exit(&data->refs);
7160 ctx->file_data = NULL;
7161 ctx->nr_user_files = 0;
7162 destroy_fixed_file_ref_node(backup_node);
7166 static void io_put_sq_data(struct io_sq_data *sqd)
7168 if (refcount_dec_and_test(&sqd->refs)) {
7170 * The park is a bit of a work-around, without it we get
7171 * warning spews on shutdown with SQPOLL set and affinity
7172 * set to a single CPU.
7175 kthread_park(sqd->thread);
7176 kthread_stop(sqd->thread);
7183 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7185 struct io_ring_ctx *ctx_attach;
7186 struct io_sq_data *sqd;
7189 f = fdget(p->wq_fd);
7191 return ERR_PTR(-ENXIO);
7192 if (f.file->f_op != &io_uring_fops) {
7194 return ERR_PTR(-EINVAL);
7197 ctx_attach = f.file->private_data;
7198 sqd = ctx_attach->sq_data;
7201 return ERR_PTR(-EINVAL);
7204 refcount_inc(&sqd->refs);
7209 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7211 struct io_sq_data *sqd;
7213 if (p->flags & IORING_SETUP_ATTACH_WQ)
7214 return io_attach_sq_data(p);
7216 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7218 return ERR_PTR(-ENOMEM);
7220 refcount_set(&sqd->refs, 1);
7221 INIT_LIST_HEAD(&sqd->ctx_list);
7222 INIT_LIST_HEAD(&sqd->ctx_new_list);
7223 mutex_init(&sqd->ctx_lock);
7224 mutex_init(&sqd->lock);
7225 init_waitqueue_head(&sqd->wait);
7229 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7230 __releases(&sqd->lock)
7234 kthread_unpark(sqd->thread);
7235 mutex_unlock(&sqd->lock);
7238 static void io_sq_thread_park(struct io_sq_data *sqd)
7239 __acquires(&sqd->lock)
7243 mutex_lock(&sqd->lock);
7244 kthread_park(sqd->thread);
7247 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7249 struct io_sq_data *sqd = ctx->sq_data;
7254 * We may arrive here from the error branch in
7255 * io_sq_offload_create() where the kthread is created
7256 * without being waked up, thus wake it up now to make
7257 * sure the wait will complete.
7259 wake_up_process(sqd->thread);
7260 wait_for_completion(&ctx->sq_thread_comp);
7262 io_sq_thread_park(sqd);
7265 mutex_lock(&sqd->ctx_lock);
7266 list_del(&ctx->sqd_list);
7267 mutex_unlock(&sqd->ctx_lock);
7270 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7271 io_sq_thread_unpark(sqd);
7274 io_put_sq_data(sqd);
7275 ctx->sq_data = NULL;
7279 static void io_finish_async(struct io_ring_ctx *ctx)
7281 io_sq_thread_stop(ctx);
7284 io_wq_destroy(ctx->io_wq);
7289 #if defined(CONFIG_UNIX)
7291 * Ensure the UNIX gc is aware of our file set, so we are certain that
7292 * the io_uring can be safely unregistered on process exit, even if we have
7293 * loops in the file referencing.
7295 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7297 struct sock *sk = ctx->ring_sock->sk;
7298 struct scm_fp_list *fpl;
7299 struct sk_buff *skb;
7302 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7306 skb = alloc_skb(0, GFP_KERNEL);
7315 fpl->user = get_uid(ctx->user);
7316 for (i = 0; i < nr; i++) {
7317 struct file *file = io_file_from_index(ctx, i + offset);
7321 fpl->fp[nr_files] = get_file(file);
7322 unix_inflight(fpl->user, fpl->fp[nr_files]);
7327 fpl->max = SCM_MAX_FD;
7328 fpl->count = nr_files;
7329 UNIXCB(skb).fp = fpl;
7330 skb->destructor = unix_destruct_scm;
7331 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7332 skb_queue_head(&sk->sk_receive_queue, skb);
7334 for (i = 0; i < nr_files; i++)
7345 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7346 * causes regular reference counting to break down. We rely on the UNIX
7347 * garbage collection to take care of this problem for us.
7349 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7351 unsigned left, total;
7355 left = ctx->nr_user_files;
7357 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7359 ret = __io_sqe_files_scm(ctx, this_files, total);
7363 total += this_files;
7369 while (total < ctx->nr_user_files) {
7370 struct file *file = io_file_from_index(ctx, total);
7380 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7386 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7387 unsigned nr_tables, unsigned nr_files)
7391 for (i = 0; i < nr_tables; i++) {
7392 struct fixed_file_table *table = &file_data->table[i];
7393 unsigned this_files;
7395 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7396 table->files = kcalloc(this_files, sizeof(struct file *),
7397 GFP_KERNEL_ACCOUNT);
7400 nr_files -= this_files;
7406 for (i = 0; i < nr_tables; i++) {
7407 struct fixed_file_table *table = &file_data->table[i];
7408 kfree(table->files);
7413 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7415 #if defined(CONFIG_UNIX)
7416 struct sock *sock = ctx->ring_sock->sk;
7417 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7418 struct sk_buff *skb;
7421 __skb_queue_head_init(&list);
7424 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7425 * remove this entry and rearrange the file array.
7427 skb = skb_dequeue(head);
7429 struct scm_fp_list *fp;
7431 fp = UNIXCB(skb).fp;
7432 for (i = 0; i < fp->count; i++) {
7435 if (fp->fp[i] != file)
7438 unix_notinflight(fp->user, fp->fp[i]);
7439 left = fp->count - 1 - i;
7441 memmove(&fp->fp[i], &fp->fp[i + 1],
7442 left * sizeof(struct file *));
7449 __skb_queue_tail(&list, skb);
7459 __skb_queue_tail(&list, skb);
7461 skb = skb_dequeue(head);
7464 if (skb_peek(&list)) {
7465 spin_lock_irq(&head->lock);
7466 while ((skb = __skb_dequeue(&list)) != NULL)
7467 __skb_queue_tail(head, skb);
7468 spin_unlock_irq(&head->lock);
7475 struct io_file_put {
7476 struct list_head list;
7480 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7482 struct fixed_file_data *file_data = ref_node->file_data;
7483 struct io_ring_ctx *ctx = file_data->ctx;
7484 struct io_file_put *pfile, *tmp;
7486 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7487 list_del(&pfile->list);
7488 io_ring_file_put(ctx, pfile->file);
7492 percpu_ref_exit(&ref_node->refs);
7494 percpu_ref_put(&file_data->refs);
7497 static void io_file_put_work(struct work_struct *work)
7499 struct io_ring_ctx *ctx;
7500 struct llist_node *node;
7502 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7503 node = llist_del_all(&ctx->file_put_llist);
7506 struct fixed_file_ref_node *ref_node;
7507 struct llist_node *next = node->next;
7509 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7510 __io_file_put_work(ref_node);
7515 static void io_file_data_ref_zero(struct percpu_ref *ref)
7517 struct fixed_file_ref_node *ref_node;
7518 struct fixed_file_data *data;
7519 struct io_ring_ctx *ctx;
7520 bool first_add = false;
7523 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7524 data = ref_node->file_data;
7527 spin_lock_bh(&data->lock);
7528 ref_node->done = true;
7530 while (!list_empty(&data->ref_list)) {
7531 ref_node = list_first_entry(&data->ref_list,
7532 struct fixed_file_ref_node, node);
7533 /* recycle ref nodes in order */
7534 if (!ref_node->done)
7536 list_del(&ref_node->node);
7537 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7539 spin_unlock_bh(&data->lock);
7541 if (percpu_ref_is_dying(&data->refs))
7545 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7547 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7550 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7551 struct io_ring_ctx *ctx)
7553 struct fixed_file_ref_node *ref_node;
7555 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7559 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7564 INIT_LIST_HEAD(&ref_node->node);
7565 INIT_LIST_HEAD(&ref_node->file_list);
7566 ref_node->file_data = ctx->file_data;
7567 ref_node->done = false;
7571 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7573 percpu_ref_exit(&ref_node->refs);
7577 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7580 __s32 __user *fds = (__s32 __user *) arg;
7581 unsigned nr_tables, i;
7583 int fd, ret = -ENOMEM;
7584 struct fixed_file_ref_node *ref_node;
7585 struct fixed_file_data *file_data;
7591 if (nr_args > IORING_MAX_FIXED_FILES)
7593 if (nr_args > rlimit(RLIMIT_NOFILE))
7596 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL_ACCOUNT);
7599 file_data->ctx = ctx;
7600 init_completion(&file_data->done);
7601 INIT_LIST_HEAD(&file_data->ref_list);
7602 spin_lock_init(&file_data->lock);
7604 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7605 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7606 GFP_KERNEL_ACCOUNT);
7607 if (!file_data->table)
7610 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7611 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7614 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7616 ctx->file_data = file_data;
7618 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7619 struct fixed_file_table *table;
7622 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7626 /* allow sparse sets */
7636 * Don't allow io_uring instances to be registered. If UNIX
7637 * isn't enabled, then this causes a reference cycle and this
7638 * instance can never get freed. If UNIX is enabled we'll
7639 * handle it just fine, but there's still no point in allowing
7640 * a ring fd as it doesn't support regular read/write anyway.
7642 if (file->f_op == &io_uring_fops) {
7646 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7647 index = i & IORING_FILE_TABLE_MASK;
7648 table->files[index] = file;
7651 ret = io_sqe_files_scm(ctx);
7653 io_sqe_files_unregister(ctx);
7657 ref_node = alloc_fixed_file_ref_node(ctx);
7659 io_sqe_files_unregister(ctx);
7663 io_sqe_files_set_node(file_data, ref_node);
7666 for (i = 0; i < ctx->nr_user_files; i++) {
7667 file = io_file_from_index(ctx, i);
7671 for (i = 0; i < nr_tables; i++)
7672 kfree(file_data->table[i].files);
7673 ctx->nr_user_files = 0;
7675 percpu_ref_exit(&file_data->refs);
7677 kfree(file_data->table);
7679 ctx->file_data = NULL;
7683 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7686 #if defined(CONFIG_UNIX)
7687 struct sock *sock = ctx->ring_sock->sk;
7688 struct sk_buff_head *head = &sock->sk_receive_queue;
7689 struct sk_buff *skb;
7692 * See if we can merge this file into an existing skb SCM_RIGHTS
7693 * file set. If there's no room, fall back to allocating a new skb
7694 * and filling it in.
7696 spin_lock_irq(&head->lock);
7697 skb = skb_peek(head);
7699 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7701 if (fpl->count < SCM_MAX_FD) {
7702 __skb_unlink(skb, head);
7703 spin_unlock_irq(&head->lock);
7704 fpl->fp[fpl->count] = get_file(file);
7705 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7707 spin_lock_irq(&head->lock);
7708 __skb_queue_head(head, skb);
7713 spin_unlock_irq(&head->lock);
7720 return __io_sqe_files_scm(ctx, 1, index);
7726 static int io_queue_file_removal(struct fixed_file_data *data,
7729 struct io_file_put *pfile;
7730 struct fixed_file_ref_node *ref_node = data->node;
7732 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7737 list_add(&pfile->list, &ref_node->file_list);
7742 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7743 struct io_uring_files_update *up,
7746 struct fixed_file_data *data = ctx->file_data;
7747 struct fixed_file_ref_node *ref_node;
7752 bool needs_switch = false;
7754 if (check_add_overflow(up->offset, nr_args, &done))
7756 if (done > ctx->nr_user_files)
7759 ref_node = alloc_fixed_file_ref_node(ctx);
7764 fds = u64_to_user_ptr(up->fds);
7766 struct fixed_file_table *table;
7770 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7774 i = array_index_nospec(up->offset, ctx->nr_user_files);
7775 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7776 index = i & IORING_FILE_TABLE_MASK;
7777 if (table->files[index]) {
7778 file = table->files[index];
7779 err = io_queue_file_removal(data, file);
7782 table->files[index] = NULL;
7783 needs_switch = true;
7792 * Don't allow io_uring instances to be registered. If
7793 * UNIX isn't enabled, then this causes a reference
7794 * cycle and this instance can never get freed. If UNIX
7795 * is enabled we'll handle it just fine, but there's
7796 * still no point in allowing a ring fd as it doesn't
7797 * support regular read/write anyway.
7799 if (file->f_op == &io_uring_fops) {
7804 table->files[index] = file;
7805 err = io_sqe_file_register(ctx, file, i);
7807 table->files[index] = NULL;
7818 percpu_ref_kill(&data->node->refs);
7819 io_sqe_files_set_node(data, ref_node);
7821 destroy_fixed_file_ref_node(ref_node);
7823 return done ? done : err;
7826 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7829 struct io_uring_files_update up;
7831 if (!ctx->file_data)
7835 if (copy_from_user(&up, arg, sizeof(up)))
7840 return __io_sqe_files_update(ctx, &up, nr_args);
7843 static void io_free_work(struct io_wq_work *work)
7845 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7847 /* Consider that io_steal_work() relies on this ref */
7851 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7852 struct io_uring_params *p)
7854 struct io_wq_data data;
7856 struct io_ring_ctx *ctx_attach;
7857 unsigned int concurrency;
7860 data.user = ctx->user;
7861 data.free_work = io_free_work;
7862 data.do_work = io_wq_submit_work;
7864 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7865 /* Do QD, or 4 * CPUS, whatever is smallest */
7866 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7868 ctx->io_wq = io_wq_create(concurrency, &data);
7869 if (IS_ERR(ctx->io_wq)) {
7870 ret = PTR_ERR(ctx->io_wq);
7876 f = fdget(p->wq_fd);
7880 if (f.file->f_op != &io_uring_fops) {
7885 ctx_attach = f.file->private_data;
7886 /* @io_wq is protected by holding the fd */
7887 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7892 ctx->io_wq = ctx_attach->io_wq;
7898 static int io_uring_alloc_task_context(struct task_struct *task)
7900 struct io_uring_task *tctx;
7903 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7904 if (unlikely(!tctx))
7907 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7908 if (unlikely(ret)) {
7914 init_waitqueue_head(&tctx->wait);
7916 atomic_set(&tctx->in_idle, 0);
7917 tctx->sqpoll = false;
7918 io_init_identity(&tctx->__identity);
7919 tctx->identity = &tctx->__identity;
7920 task->io_uring = tctx;
7924 void __io_uring_free(struct task_struct *tsk)
7926 struct io_uring_task *tctx = tsk->io_uring;
7928 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7929 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7930 if (tctx->identity != &tctx->__identity)
7931 kfree(tctx->identity);
7932 percpu_counter_destroy(&tctx->inflight);
7934 tsk->io_uring = NULL;
7937 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7938 struct io_uring_params *p)
7942 if (ctx->flags & IORING_SETUP_SQPOLL) {
7943 struct io_sq_data *sqd;
7946 if (!capable(CAP_SYS_ADMIN))
7949 sqd = io_get_sq_data(p);
7956 io_sq_thread_park(sqd);
7957 mutex_lock(&sqd->ctx_lock);
7958 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7959 mutex_unlock(&sqd->ctx_lock);
7960 io_sq_thread_unpark(sqd);
7962 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7963 if (!ctx->sq_thread_idle)
7964 ctx->sq_thread_idle = HZ;
7969 if (p->flags & IORING_SETUP_SQ_AFF) {
7970 int cpu = p->sq_thread_cpu;
7973 if (cpu >= nr_cpu_ids)
7975 if (!cpu_online(cpu))
7978 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7979 cpu, "io_uring-sq");
7981 sqd->thread = kthread_create(io_sq_thread, sqd,
7984 if (IS_ERR(sqd->thread)) {
7985 ret = PTR_ERR(sqd->thread);
7989 ret = io_uring_alloc_task_context(sqd->thread);
7992 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7993 /* Can't have SQ_AFF without SQPOLL */
7999 ret = io_init_wq_offload(ctx, p);
8005 io_finish_async(ctx);
8009 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8011 struct io_sq_data *sqd = ctx->sq_data;
8013 ctx->flags &= ~IORING_SETUP_R_DISABLED;
8014 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd && sqd->thread)
8015 wake_up_process(sqd->thread);
8018 static inline void __io_unaccount_mem(struct user_struct *user,
8019 unsigned long nr_pages)
8021 atomic_long_sub(nr_pages, &user->locked_vm);
8024 static inline int __io_account_mem(struct user_struct *user,
8025 unsigned long nr_pages)
8027 unsigned long page_limit, cur_pages, new_pages;
8029 /* Don't allow more pages than we can safely lock */
8030 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8033 cur_pages = atomic_long_read(&user->locked_vm);
8034 new_pages = cur_pages + nr_pages;
8035 if (new_pages > page_limit)
8037 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8038 new_pages) != cur_pages);
8043 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8044 enum io_mem_account acct)
8047 __io_unaccount_mem(ctx->user, nr_pages);
8049 if (ctx->mm_account) {
8050 if (acct == ACCT_LOCKED)
8051 ctx->mm_account->locked_vm -= nr_pages;
8052 else if (acct == ACCT_PINNED)
8053 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8057 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8058 enum io_mem_account acct)
8062 if (ctx->limit_mem) {
8063 ret = __io_account_mem(ctx->user, nr_pages);
8068 if (ctx->mm_account) {
8069 if (acct == ACCT_LOCKED)
8070 ctx->mm_account->locked_vm += nr_pages;
8071 else if (acct == ACCT_PINNED)
8072 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8078 static void io_mem_free(void *ptr)
8085 page = virt_to_head_page(ptr);
8086 if (put_page_testzero(page))
8087 free_compound_page(page);
8090 static void *io_mem_alloc(size_t size)
8092 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8095 return (void *) __get_free_pages(gfp_flags, get_order(size));
8098 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8101 struct io_rings *rings;
8102 size_t off, sq_array_size;
8104 off = struct_size(rings, cqes, cq_entries);
8105 if (off == SIZE_MAX)
8109 off = ALIGN(off, SMP_CACHE_BYTES);
8117 sq_array_size = array_size(sizeof(u32), sq_entries);
8118 if (sq_array_size == SIZE_MAX)
8121 if (check_add_overflow(off, sq_array_size, &off))
8127 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8131 pages = (size_t)1 << get_order(
8132 rings_size(sq_entries, cq_entries, NULL));
8133 pages += (size_t)1 << get_order(
8134 array_size(sizeof(struct io_uring_sqe), sq_entries));
8139 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8143 if (!ctx->user_bufs)
8146 for (i = 0; i < ctx->nr_user_bufs; i++) {
8147 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8149 for (j = 0; j < imu->nr_bvecs; j++)
8150 unpin_user_page(imu->bvec[j].bv_page);
8152 if (imu->acct_pages)
8153 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8158 kfree(ctx->user_bufs);
8159 ctx->user_bufs = NULL;
8160 ctx->nr_user_bufs = 0;
8164 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8165 void __user *arg, unsigned index)
8167 struct iovec __user *src;
8169 #ifdef CONFIG_COMPAT
8171 struct compat_iovec __user *ciovs;
8172 struct compat_iovec ciov;
8174 ciovs = (struct compat_iovec __user *) arg;
8175 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8178 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8179 dst->iov_len = ciov.iov_len;
8183 src = (struct iovec __user *) arg;
8184 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8190 * Not super efficient, but this is just a registration time. And we do cache
8191 * the last compound head, so generally we'll only do a full search if we don't
8194 * We check if the given compound head page has already been accounted, to
8195 * avoid double accounting it. This allows us to account the full size of the
8196 * page, not just the constituent pages of a huge page.
8198 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8199 int nr_pages, struct page *hpage)
8203 /* check current page array */
8204 for (i = 0; i < nr_pages; i++) {
8205 if (!PageCompound(pages[i]))
8207 if (compound_head(pages[i]) == hpage)
8211 /* check previously registered pages */
8212 for (i = 0; i < ctx->nr_user_bufs; i++) {
8213 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8215 for (j = 0; j < imu->nr_bvecs; j++) {
8216 if (!PageCompound(imu->bvec[j].bv_page))
8218 if (compound_head(imu->bvec[j].bv_page) == hpage)
8226 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8227 int nr_pages, struct io_mapped_ubuf *imu,
8228 struct page **last_hpage)
8232 for (i = 0; i < nr_pages; i++) {
8233 if (!PageCompound(pages[i])) {
8238 hpage = compound_head(pages[i]);
8239 if (hpage == *last_hpage)
8241 *last_hpage = hpage;
8242 if (headpage_already_acct(ctx, pages, i, hpage))
8244 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8248 if (!imu->acct_pages)
8251 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8253 imu->acct_pages = 0;
8257 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8260 struct vm_area_struct **vmas = NULL;
8261 struct page **pages = NULL;
8262 struct page *last_hpage = NULL;
8263 int i, j, got_pages = 0;
8268 if (!nr_args || nr_args > UIO_MAXIOV)
8271 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8273 if (!ctx->user_bufs)
8276 for (i = 0; i < nr_args; i++) {
8277 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8278 unsigned long off, start, end, ubuf;
8283 ret = io_copy_iov(ctx, &iov, arg, i);
8288 * Don't impose further limits on the size and buffer
8289 * constraints here, we'll -EINVAL later when IO is
8290 * submitted if they are wrong.
8293 if (!iov.iov_base || !iov.iov_len)
8296 /* arbitrary limit, but we need something */
8297 if (iov.iov_len > SZ_1G)
8300 ubuf = (unsigned long) iov.iov_base;
8301 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8302 start = ubuf >> PAGE_SHIFT;
8303 nr_pages = end - start;
8306 if (!pages || nr_pages > got_pages) {
8309 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8311 vmas = kvmalloc_array(nr_pages,
8312 sizeof(struct vm_area_struct *),
8314 if (!pages || !vmas) {
8318 got_pages = nr_pages;
8321 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8328 mmap_read_lock(current->mm);
8329 pret = pin_user_pages(ubuf, nr_pages,
8330 FOLL_WRITE | FOLL_LONGTERM,
8332 if (pret == nr_pages) {
8333 /* don't support file backed memory */
8334 for (j = 0; j < nr_pages; j++) {
8335 struct vm_area_struct *vma = vmas[j];
8338 !is_file_hugepages(vma->vm_file)) {
8344 ret = pret < 0 ? pret : -EFAULT;
8346 mmap_read_unlock(current->mm);
8349 * if we did partial map, or found file backed vmas,
8350 * release any pages we did get
8353 unpin_user_pages(pages, pret);
8358 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8360 unpin_user_pages(pages, pret);
8365 off = ubuf & ~PAGE_MASK;
8367 for (j = 0; j < nr_pages; j++) {
8370 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8371 imu->bvec[j].bv_page = pages[j];
8372 imu->bvec[j].bv_len = vec_len;
8373 imu->bvec[j].bv_offset = off;
8377 /* store original address for later verification */
8379 imu->len = iov.iov_len;
8380 imu->nr_bvecs = nr_pages;
8382 ctx->nr_user_bufs++;
8390 io_sqe_buffer_unregister(ctx);
8394 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8396 __s32 __user *fds = arg;
8402 if (copy_from_user(&fd, fds, sizeof(*fds)))
8405 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8406 if (IS_ERR(ctx->cq_ev_fd)) {
8407 int ret = PTR_ERR(ctx->cq_ev_fd);
8408 ctx->cq_ev_fd = NULL;
8415 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8417 if (ctx->cq_ev_fd) {
8418 eventfd_ctx_put(ctx->cq_ev_fd);
8419 ctx->cq_ev_fd = NULL;
8426 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8428 struct io_buffer *buf;
8429 unsigned long index;
8431 xa_for_each(&ctx->io_buffers, index, buf)
8432 __io_remove_buffers(ctx, buf, index, -1U);
8435 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8437 io_finish_async(ctx);
8438 io_sqe_buffer_unregister(ctx);
8440 if (ctx->sqo_task) {
8441 put_task_struct(ctx->sqo_task);
8442 ctx->sqo_task = NULL;
8443 mmdrop(ctx->mm_account);
8444 ctx->mm_account = NULL;
8447 #ifdef CONFIG_BLK_CGROUP
8448 if (ctx->sqo_blkcg_css)
8449 css_put(ctx->sqo_blkcg_css);
8452 io_sqe_files_unregister(ctx);
8453 io_eventfd_unregister(ctx);
8454 io_destroy_buffers(ctx);
8456 #if defined(CONFIG_UNIX)
8457 if (ctx->ring_sock) {
8458 ctx->ring_sock->file = NULL; /* so that iput() is called */
8459 sock_release(ctx->ring_sock);
8463 io_mem_free(ctx->rings);
8464 io_mem_free(ctx->sq_sqes);
8466 percpu_ref_exit(&ctx->refs);
8467 free_uid(ctx->user);
8468 put_cred(ctx->creds);
8469 kfree(ctx->cancel_hash);
8470 kmem_cache_free(req_cachep, ctx->fallback_req);
8474 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8476 struct io_ring_ctx *ctx = file->private_data;
8479 poll_wait(file, &ctx->cq_wait, wait);
8481 * synchronizes with barrier from wq_has_sleeper call in
8485 if (!io_sqring_full(ctx))
8486 mask |= EPOLLOUT | EPOLLWRNORM;
8489 * Don't flush cqring overflow list here, just do a simple check.
8490 * Otherwise there could possible be ABBA deadlock:
8493 * lock(&ctx->uring_lock);
8495 * lock(&ctx->uring_lock);
8498 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8499 * pushs them to do the flush.
8501 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8502 mask |= EPOLLIN | EPOLLRDNORM;
8507 static int io_uring_fasync(int fd, struct file *file, int on)
8509 struct io_ring_ctx *ctx = file->private_data;
8511 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8514 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8516 struct io_identity *iod;
8518 iod = xa_erase(&ctx->personalities, id);
8520 put_cred(iod->creds);
8521 if (refcount_dec_and_test(&iod->count))
8529 static void io_ring_exit_work(struct work_struct *work)
8531 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8535 * If we're doing polled IO and end up having requests being
8536 * submitted async (out-of-line), then completions can come in while
8537 * we're waiting for refs to drop. We need to reap these manually,
8538 * as nobody else will be looking for them.
8541 io_iopoll_try_reap_events(ctx);
8542 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8543 io_ring_ctx_free(ctx);
8546 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8548 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8550 return req->ctx == data;
8553 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8555 unsigned long index;
8556 struct io_identify *iod;
8558 mutex_lock(&ctx->uring_lock);
8559 percpu_ref_kill(&ctx->refs);
8560 /* if force is set, the ring is going away. always drop after that */
8562 if (WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) && !ctx->sqo_dead))
8565 ctx->cq_overflow_flushed = 1;
8567 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8568 mutex_unlock(&ctx->uring_lock);
8570 io_kill_timeouts(ctx, NULL, NULL);
8571 io_poll_remove_all(ctx, NULL, NULL);
8574 io_wq_cancel_cb(ctx->io_wq, io_cancel_ctx_cb, ctx, true);
8576 /* if we failed setting up the ctx, we might not have any rings */
8577 io_iopoll_try_reap_events(ctx);
8578 xa_for_each(&ctx->personalities, index, iod)
8579 io_unregister_personality(ctx, index);
8582 * Do this upfront, so we won't have a grace period where the ring
8583 * is closed but resources aren't reaped yet. This can cause
8584 * spurious failure in setting up a new ring.
8586 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8589 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8591 * Use system_unbound_wq to avoid spawning tons of event kworkers
8592 * if we're exiting a ton of rings at the same time. It just adds
8593 * noise and overhead, there's no discernable change in runtime
8594 * over using system_wq.
8596 queue_work(system_unbound_wq, &ctx->exit_work);
8599 static int io_uring_release(struct inode *inode, struct file *file)
8601 struct io_ring_ctx *ctx = file->private_data;
8603 file->private_data = NULL;
8604 io_ring_ctx_wait_and_kill(ctx);
8608 struct io_task_cancel {
8609 struct task_struct *task;
8610 struct files_struct *files;
8613 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8615 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8616 struct io_task_cancel *cancel = data;
8619 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8620 unsigned long flags;
8621 struct io_ring_ctx *ctx = req->ctx;
8623 /* protect against races with linked timeouts */
8624 spin_lock_irqsave(&ctx->completion_lock, flags);
8625 ret = io_match_task(req, cancel->task, cancel->files);
8626 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8628 ret = io_match_task(req, cancel->task, cancel->files);
8633 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8634 struct task_struct *task,
8635 struct files_struct *files)
8637 struct io_defer_entry *de = NULL;
8640 spin_lock_irq(&ctx->completion_lock);
8641 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8642 if (io_match_task(de->req, task, files)) {
8643 list_cut_position(&list, &ctx->defer_list, &de->list);
8647 spin_unlock_irq(&ctx->completion_lock);
8649 while (!list_empty(&list)) {
8650 de = list_first_entry(&list, struct io_defer_entry, list);
8651 list_del_init(&de->list);
8652 req_set_fail_links(de->req);
8653 io_put_req(de->req);
8654 io_req_complete(de->req, -ECANCELED);
8659 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8660 struct task_struct *task,
8661 struct files_struct *files)
8663 struct io_kiocb *req;
8666 spin_lock_irq(&ctx->inflight_lock);
8667 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8668 cnt += io_match_task(req, task, files);
8669 spin_unlock_irq(&ctx->inflight_lock);
8673 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8674 struct task_struct *task,
8675 struct files_struct *files)
8677 while (!list_empty_careful(&ctx->inflight_list)) {
8678 struct io_task_cancel cancel = { .task = task, .files = files };
8682 inflight = io_uring_count_inflight(ctx, task, files);
8686 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8687 io_poll_remove_all(ctx, task, files);
8688 io_kill_timeouts(ctx, task, files);
8689 /* cancellations _may_ trigger task work */
8692 prepare_to_wait(&task->io_uring->wait, &wait,
8693 TASK_UNINTERRUPTIBLE);
8694 if (inflight == io_uring_count_inflight(ctx, task, files))
8696 finish_wait(&task->io_uring->wait, &wait);
8700 static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8701 struct task_struct *task)
8704 struct io_task_cancel cancel = { .task = task, .files = NULL, };
8705 enum io_wq_cancel cret;
8708 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8709 if (cret != IO_WQ_CANCEL_NOTFOUND)
8712 /* SQPOLL thread does its own polling */
8713 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8714 while (!list_empty_careful(&ctx->iopoll_list)) {
8715 io_iopoll_try_reap_events(ctx);
8720 ret |= io_poll_remove_all(ctx, task, NULL);
8721 ret |= io_kill_timeouts(ctx, task, NULL);
8729 static void io_disable_sqo_submit(struct io_ring_ctx *ctx)
8731 mutex_lock(&ctx->uring_lock);
8733 if (ctx->flags & IORING_SETUP_R_DISABLED)
8734 io_sq_offload_start(ctx);
8735 mutex_unlock(&ctx->uring_lock);
8737 /* make sure callers enter the ring to get error */
8739 io_ring_set_wakeup_flag(ctx);
8743 * We need to iteratively cancel requests, in case a request has dependent
8744 * hard links. These persist even for failure of cancelations, hence keep
8745 * looping until none are found.
8747 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8748 struct files_struct *files)
8750 struct task_struct *task = current;
8752 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8753 io_disable_sqo_submit(ctx);
8754 task = ctx->sq_data->thread;
8755 atomic_inc(&task->io_uring->in_idle);
8756 io_sq_thread_park(ctx->sq_data);
8759 io_cancel_defer_files(ctx, task, files);
8760 io_cqring_overflow_flush(ctx, true, task, files);
8763 __io_uring_cancel_task_requests(ctx, task);
8765 io_uring_cancel_files(ctx, task, files);
8767 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8768 atomic_dec(&task->io_uring->in_idle);
8769 io_sq_thread_unpark(ctx->sq_data);
8774 * Note that this task has used io_uring. We use it for cancelation purposes.
8776 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8778 struct io_uring_task *tctx = current->io_uring;
8781 if (unlikely(!tctx)) {
8782 ret = io_uring_alloc_task_context(current);
8785 tctx = current->io_uring;
8787 if (tctx->last != file) {
8788 void *old = xa_load(&tctx->xa, (unsigned long)file);
8792 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
8803 * This is race safe in that the task itself is doing this, hence it
8804 * cannot be going through the exit/cancel paths at the same time.
8805 * This cannot be modified while exit/cancel is running.
8807 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8808 tctx->sqpoll = true;
8814 * Remove this io_uring_file -> task mapping.
8816 static void io_uring_del_task_file(struct file *file)
8818 struct io_uring_task *tctx = current->io_uring;
8820 if (tctx->last == file)
8822 file = xa_erase(&tctx->xa, (unsigned long)file);
8827 static void io_uring_remove_task_files(struct io_uring_task *tctx)
8830 unsigned long index;
8832 xa_for_each(&tctx->xa, index, file)
8833 io_uring_del_task_file(file);
8836 void __io_uring_files_cancel(struct files_struct *files)
8838 struct io_uring_task *tctx = current->io_uring;
8840 unsigned long index;
8842 /* make sure overflow events are dropped */
8843 atomic_inc(&tctx->in_idle);
8844 xa_for_each(&tctx->xa, index, file)
8845 io_uring_cancel_task_requests(file->private_data, files);
8846 atomic_dec(&tctx->in_idle);
8849 io_uring_remove_task_files(tctx);
8852 static s64 tctx_inflight(struct io_uring_task *tctx)
8854 unsigned long index;
8858 inflight = percpu_counter_sum(&tctx->inflight);
8863 * If we have SQPOLL rings, then we need to iterate and find them, and
8864 * add the pending count for those.
8866 xa_for_each(&tctx->xa, index, file) {
8867 struct io_ring_ctx *ctx = file->private_data;
8869 if (ctx->flags & IORING_SETUP_SQPOLL) {
8870 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
8872 inflight += percpu_counter_sum(&__tctx->inflight);
8880 * Find any io_uring fd that this task has registered or done IO on, and cancel
8883 void __io_uring_task_cancel(void)
8885 struct io_uring_task *tctx = current->io_uring;
8889 /* make sure overflow events are dropped */
8890 atomic_inc(&tctx->in_idle);
8892 /* trigger io_disable_sqo_submit() */
8894 __io_uring_files_cancel(NULL);
8897 /* read completions before cancelations */
8898 inflight = tctx_inflight(tctx);
8901 __io_uring_files_cancel(NULL);
8903 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8906 * If we've seen completions, retry without waiting. This
8907 * avoids a race where a completion comes in before we did
8908 * prepare_to_wait().
8910 if (inflight == tctx_inflight(tctx))
8912 finish_wait(&tctx->wait, &wait);
8915 atomic_dec(&tctx->in_idle);
8917 io_uring_remove_task_files(tctx);
8920 static int io_uring_flush(struct file *file, void *data)
8922 struct io_uring_task *tctx = current->io_uring;
8923 struct io_ring_ctx *ctx = file->private_data;
8925 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8926 io_uring_cancel_task_requests(ctx, NULL);
8931 /* we should have cancelled and erased it before PF_EXITING */
8932 WARN_ON_ONCE((current->flags & PF_EXITING) &&
8933 xa_load(&tctx->xa, (unsigned long)file));
8936 * fput() is pending, will be 2 if the only other ref is our potential
8937 * task file note. If the task is exiting, drop regardless of count.
8939 if (atomic_long_read(&file->f_count) != 2)
8942 if (ctx->flags & IORING_SETUP_SQPOLL) {
8943 /* there is only one file note, which is owned by sqo_task */
8944 WARN_ON_ONCE(ctx->sqo_task != current &&
8945 xa_load(&tctx->xa, (unsigned long)file));
8946 /* sqo_dead check is for when this happens after cancellation */
8947 WARN_ON_ONCE(ctx->sqo_task == current && !ctx->sqo_dead &&
8948 !xa_load(&tctx->xa, (unsigned long)file));
8950 io_disable_sqo_submit(ctx);
8953 if (!(ctx->flags & IORING_SETUP_SQPOLL) || ctx->sqo_task == current)
8954 io_uring_del_task_file(file);
8958 static void *io_uring_validate_mmap_request(struct file *file,
8959 loff_t pgoff, size_t sz)
8961 struct io_ring_ctx *ctx = file->private_data;
8962 loff_t offset = pgoff << PAGE_SHIFT;
8967 case IORING_OFF_SQ_RING:
8968 case IORING_OFF_CQ_RING:
8971 case IORING_OFF_SQES:
8975 return ERR_PTR(-EINVAL);
8978 page = virt_to_head_page(ptr);
8979 if (sz > page_size(page))
8980 return ERR_PTR(-EINVAL);
8987 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8989 size_t sz = vma->vm_end - vma->vm_start;
8993 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8995 return PTR_ERR(ptr);
8997 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8998 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9001 #else /* !CONFIG_MMU */
9003 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9005 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9008 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9010 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9013 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9014 unsigned long addr, unsigned long len,
9015 unsigned long pgoff, unsigned long flags)
9019 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9021 return PTR_ERR(ptr);
9023 return (unsigned long) ptr;
9026 #endif /* !CONFIG_MMU */
9028 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9034 if (!io_sqring_full(ctx))
9037 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9039 if (unlikely(ctx->sqo_dead)) {
9044 if (!io_sqring_full(ctx))
9048 } while (!signal_pending(current));
9050 finish_wait(&ctx->sqo_sq_wait, &wait);
9055 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9056 u32, min_complete, u32, flags, const sigset_t __user *, sig,
9059 struct io_ring_ctx *ctx;
9066 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9067 IORING_ENTER_SQ_WAIT))
9075 if (f.file->f_op != &io_uring_fops)
9079 ctx = f.file->private_data;
9080 if (!percpu_ref_tryget(&ctx->refs))
9084 if (ctx->flags & IORING_SETUP_R_DISABLED)
9088 * For SQ polling, the thread will do all submissions and completions.
9089 * Just return the requested submit count, and wake the thread if
9093 if (ctx->flags & IORING_SETUP_SQPOLL) {
9094 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9096 if (unlikely(ctx->sqo_dead)) {
9100 if (flags & IORING_ENTER_SQ_WAKEUP)
9101 wake_up(&ctx->sq_data->wait);
9102 if (flags & IORING_ENTER_SQ_WAIT) {
9103 ret = io_sqpoll_wait_sq(ctx);
9107 submitted = to_submit;
9108 } else if (to_submit) {
9109 ret = io_uring_add_task_file(ctx, f.file);
9112 mutex_lock(&ctx->uring_lock);
9113 submitted = io_submit_sqes(ctx, to_submit);
9114 mutex_unlock(&ctx->uring_lock);
9116 if (submitted != to_submit)
9119 if (flags & IORING_ENTER_GETEVENTS) {
9120 min_complete = min(min_complete, ctx->cq_entries);
9123 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9124 * space applications don't need to do io completion events
9125 * polling again, they can rely on io_sq_thread to do polling
9126 * work, which can reduce cpu usage and uring_lock contention.
9128 if (ctx->flags & IORING_SETUP_IOPOLL &&
9129 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9130 ret = io_iopoll_check(ctx, min_complete);
9132 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
9137 percpu_ref_put(&ctx->refs);
9140 return submitted ? submitted : ret;
9143 #ifdef CONFIG_PROC_FS
9144 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9145 const struct io_identity *iod)
9147 const struct cred *cred = iod->creds;
9148 struct user_namespace *uns = seq_user_ns(m);
9149 struct group_info *gi;
9154 seq_printf(m, "%5d\n", id);
9155 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9156 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9157 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9158 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9159 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9160 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9161 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9162 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9163 seq_puts(m, "\n\tGroups:\t");
9164 gi = cred->group_info;
9165 for (g = 0; g < gi->ngroups; g++) {
9166 seq_put_decimal_ull(m, g ? " " : "",
9167 from_kgid_munged(uns, gi->gid[g]));
9169 seq_puts(m, "\n\tCapEff:\t");
9170 cap = cred->cap_effective;
9171 CAP_FOR_EACH_U32(__capi)
9172 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9177 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9179 struct io_sq_data *sq = NULL;
9184 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9185 * since fdinfo case grabs it in the opposite direction of normal use
9186 * cases. If we fail to get the lock, we just don't iterate any
9187 * structures that could be going away outside the io_uring mutex.
9189 has_lock = mutex_trylock(&ctx->uring_lock);
9191 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9194 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9195 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9196 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9197 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9198 struct fixed_file_table *table;
9201 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9202 f = table->files[i & IORING_FILE_TABLE_MASK];
9204 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9206 seq_printf(m, "%5u: <none>\n", i);
9208 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9209 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9210 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9212 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9213 (unsigned int) buf->len);
9215 if (has_lock && !xa_empty(&ctx->personalities)) {
9216 unsigned long index;
9217 const struct io_identity *iod;
9219 seq_printf(m, "Personalities:\n");
9220 xa_for_each(&ctx->personalities, index, iod)
9221 io_uring_show_cred(m, index, iod);
9223 seq_printf(m, "PollList:\n");
9224 spin_lock_irq(&ctx->completion_lock);
9225 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9226 struct hlist_head *list = &ctx->cancel_hash[i];
9227 struct io_kiocb *req;
9229 hlist_for_each_entry(req, list, hash_node)
9230 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9231 req->task->task_works != NULL);
9233 spin_unlock_irq(&ctx->completion_lock);
9235 mutex_unlock(&ctx->uring_lock);
9238 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9240 struct io_ring_ctx *ctx = f->private_data;
9242 if (percpu_ref_tryget(&ctx->refs)) {
9243 __io_uring_show_fdinfo(ctx, m);
9244 percpu_ref_put(&ctx->refs);
9249 static const struct file_operations io_uring_fops = {
9250 .release = io_uring_release,
9251 .flush = io_uring_flush,
9252 .mmap = io_uring_mmap,
9254 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9255 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9257 .poll = io_uring_poll,
9258 .fasync = io_uring_fasync,
9259 #ifdef CONFIG_PROC_FS
9260 .show_fdinfo = io_uring_show_fdinfo,
9264 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9265 struct io_uring_params *p)
9267 struct io_rings *rings;
9268 size_t size, sq_array_offset;
9270 /* make sure these are sane, as we already accounted them */
9271 ctx->sq_entries = p->sq_entries;
9272 ctx->cq_entries = p->cq_entries;
9274 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9275 if (size == SIZE_MAX)
9278 rings = io_mem_alloc(size);
9283 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9284 rings->sq_ring_mask = p->sq_entries - 1;
9285 rings->cq_ring_mask = p->cq_entries - 1;
9286 rings->sq_ring_entries = p->sq_entries;
9287 rings->cq_ring_entries = p->cq_entries;
9288 ctx->sq_mask = rings->sq_ring_mask;
9289 ctx->cq_mask = rings->cq_ring_mask;
9291 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9292 if (size == SIZE_MAX) {
9293 io_mem_free(ctx->rings);
9298 ctx->sq_sqes = io_mem_alloc(size);
9299 if (!ctx->sq_sqes) {
9300 io_mem_free(ctx->rings);
9308 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9312 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9316 ret = io_uring_add_task_file(ctx, file);
9321 fd_install(fd, file);
9326 * Allocate an anonymous fd, this is what constitutes the application
9327 * visible backing of an io_uring instance. The application mmaps this
9328 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9329 * we have to tie this fd to a socket for file garbage collection purposes.
9331 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9334 #if defined(CONFIG_UNIX)
9337 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9340 return ERR_PTR(ret);
9343 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9344 O_RDWR | O_CLOEXEC);
9345 #if defined(CONFIG_UNIX)
9347 sock_release(ctx->ring_sock);
9348 ctx->ring_sock = NULL;
9350 ctx->ring_sock->file = file;
9356 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9357 struct io_uring_params __user *params)
9359 struct user_struct *user = NULL;
9360 struct io_ring_ctx *ctx;
9367 if (entries > IORING_MAX_ENTRIES) {
9368 if (!(p->flags & IORING_SETUP_CLAMP))
9370 entries = IORING_MAX_ENTRIES;
9374 * Use twice as many entries for the CQ ring. It's possible for the
9375 * application to drive a higher depth than the size of the SQ ring,
9376 * since the sqes are only used at submission time. This allows for
9377 * some flexibility in overcommitting a bit. If the application has
9378 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9379 * of CQ ring entries manually.
9381 p->sq_entries = roundup_pow_of_two(entries);
9382 if (p->flags & IORING_SETUP_CQSIZE) {
9384 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9385 * to a power-of-two, if it isn't already. We do NOT impose
9386 * any cq vs sq ring sizing.
9390 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9391 if (!(p->flags & IORING_SETUP_CLAMP))
9393 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9395 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9396 if (p->cq_entries < p->sq_entries)
9399 p->cq_entries = 2 * p->sq_entries;
9402 user = get_uid(current_user());
9403 limit_mem = !capable(CAP_IPC_LOCK);
9406 ret = __io_account_mem(user,
9407 ring_pages(p->sq_entries, p->cq_entries));
9414 ctx = io_ring_ctx_alloc(p);
9417 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9422 ctx->compat = in_compat_syscall();
9424 ctx->creds = get_current_cred();
9426 ctx->loginuid = current->loginuid;
9427 ctx->sessionid = current->sessionid;
9429 ctx->sqo_task = get_task_struct(current);
9432 * This is just grabbed for accounting purposes. When a process exits,
9433 * the mm is exited and dropped before the files, hence we need to hang
9434 * on to this mm purely for the purposes of being able to unaccount
9435 * memory (locked/pinned vm). It's not used for anything else.
9437 mmgrab(current->mm);
9438 ctx->mm_account = current->mm;
9440 #ifdef CONFIG_BLK_CGROUP
9442 * The sq thread will belong to the original cgroup it was inited in.
9443 * If the cgroup goes offline (e.g. disabling the io controller), then
9444 * issued bios will be associated with the closest cgroup later in the
9448 ctx->sqo_blkcg_css = blkcg_css();
9449 ret = css_tryget_online(ctx->sqo_blkcg_css);
9452 /* don't init against a dying cgroup, have the user try again */
9453 ctx->sqo_blkcg_css = NULL;
9460 * Account memory _before_ installing the file descriptor. Once
9461 * the descriptor is installed, it can get closed at any time. Also
9462 * do this before hitting the general error path, as ring freeing
9463 * will un-account as well.
9465 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9467 ctx->limit_mem = limit_mem;
9469 ret = io_allocate_scq_urings(ctx, p);
9473 ret = io_sq_offload_create(ctx, p);
9477 if (!(p->flags & IORING_SETUP_R_DISABLED))
9478 io_sq_offload_start(ctx);
9480 memset(&p->sq_off, 0, sizeof(p->sq_off));
9481 p->sq_off.head = offsetof(struct io_rings, sq.head);
9482 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9483 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9484 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9485 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9486 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9487 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9489 memset(&p->cq_off, 0, sizeof(p->cq_off));
9490 p->cq_off.head = offsetof(struct io_rings, cq.head);
9491 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9492 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9493 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9494 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9495 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9496 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9498 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9499 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9500 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9501 IORING_FEAT_POLL_32BITS;
9503 if (copy_to_user(params, p, sizeof(*p))) {
9508 file = io_uring_get_file(ctx);
9510 ret = PTR_ERR(file);
9515 * Install ring fd as the very last thing, so we don't risk someone
9516 * having closed it before we finish setup
9518 ret = io_uring_install_fd(ctx, file);
9520 io_disable_sqo_submit(ctx);
9521 /* fput will clean it up */
9526 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9529 io_disable_sqo_submit(ctx);
9530 io_ring_ctx_wait_and_kill(ctx);
9535 * Sets up an aio uring context, and returns the fd. Applications asks for a
9536 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9537 * params structure passed in.
9539 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9541 struct io_uring_params p;
9544 if (copy_from_user(&p, params, sizeof(p)))
9546 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9551 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9552 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9553 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9554 IORING_SETUP_R_DISABLED))
9557 return io_uring_create(entries, &p, params);
9560 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9561 struct io_uring_params __user *, params)
9563 return io_uring_setup(entries, params);
9566 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9568 struct io_uring_probe *p;
9572 size = struct_size(p, ops, nr_args);
9573 if (size == SIZE_MAX)
9575 p = kzalloc(size, GFP_KERNEL);
9580 if (copy_from_user(p, arg, size))
9583 if (memchr_inv(p, 0, size))
9586 p->last_op = IORING_OP_LAST - 1;
9587 if (nr_args > IORING_OP_LAST)
9588 nr_args = IORING_OP_LAST;
9590 for (i = 0; i < nr_args; i++) {
9592 if (!io_op_defs[i].not_supported)
9593 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9598 if (copy_to_user(arg, p, size))
9605 static int io_register_personality(struct io_ring_ctx *ctx)
9607 struct io_identity *iod;
9611 iod = kmalloc(sizeof(*iod), GFP_KERNEL);
9615 io_init_identity(iod);
9616 iod->creds = get_current_cred();
9618 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)iod,
9619 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9621 put_cred(iod->creds);
9628 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9629 unsigned int nr_args)
9631 struct io_uring_restriction *res;
9635 /* Restrictions allowed only if rings started disabled */
9636 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9639 /* We allow only a single restrictions registration */
9640 if (ctx->restrictions.registered)
9643 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9646 size = array_size(nr_args, sizeof(*res));
9647 if (size == SIZE_MAX)
9650 res = memdup_user(arg, size);
9652 return PTR_ERR(res);
9656 for (i = 0; i < nr_args; i++) {
9657 switch (res[i].opcode) {
9658 case IORING_RESTRICTION_REGISTER_OP:
9659 if (res[i].register_op >= IORING_REGISTER_LAST) {
9664 __set_bit(res[i].register_op,
9665 ctx->restrictions.register_op);
9667 case IORING_RESTRICTION_SQE_OP:
9668 if (res[i].sqe_op >= IORING_OP_LAST) {
9673 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9675 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9676 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9678 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9679 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9688 /* Reset all restrictions if an error happened */
9690 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9692 ctx->restrictions.registered = true;
9698 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9700 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9703 if (ctx->restrictions.registered)
9704 ctx->restricted = 1;
9706 io_sq_offload_start(ctx);
9710 static bool io_register_op_must_quiesce(int op)
9713 case IORING_UNREGISTER_FILES:
9714 case IORING_REGISTER_FILES_UPDATE:
9715 case IORING_REGISTER_PROBE:
9716 case IORING_REGISTER_PERSONALITY:
9717 case IORING_UNREGISTER_PERSONALITY:
9724 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9725 void __user *arg, unsigned nr_args)
9726 __releases(ctx->uring_lock)
9727 __acquires(ctx->uring_lock)
9732 * We're inside the ring mutex, if the ref is already dying, then
9733 * someone else killed the ctx or is already going through
9734 * io_uring_register().
9736 if (percpu_ref_is_dying(&ctx->refs))
9739 if (io_register_op_must_quiesce(opcode)) {
9740 percpu_ref_kill(&ctx->refs);
9743 * Drop uring mutex before waiting for references to exit. If
9744 * another thread is currently inside io_uring_enter() it might
9745 * need to grab the uring_lock to make progress. If we hold it
9746 * here across the drain wait, then we can deadlock. It's safe
9747 * to drop the mutex here, since no new references will come in
9748 * after we've killed the percpu ref.
9750 mutex_unlock(&ctx->uring_lock);
9752 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9755 ret = io_run_task_work_sig();
9760 mutex_lock(&ctx->uring_lock);
9763 percpu_ref_resurrect(&ctx->refs);
9768 if (ctx->restricted) {
9769 if (opcode >= IORING_REGISTER_LAST) {
9774 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9781 case IORING_REGISTER_BUFFERS:
9782 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9784 case IORING_UNREGISTER_BUFFERS:
9788 ret = io_sqe_buffer_unregister(ctx);
9790 case IORING_REGISTER_FILES:
9791 ret = io_sqe_files_register(ctx, arg, nr_args);
9793 case IORING_UNREGISTER_FILES:
9797 ret = io_sqe_files_unregister(ctx);
9799 case IORING_REGISTER_FILES_UPDATE:
9800 ret = io_sqe_files_update(ctx, arg, nr_args);
9802 case IORING_REGISTER_EVENTFD:
9803 case IORING_REGISTER_EVENTFD_ASYNC:
9807 ret = io_eventfd_register(ctx, arg);
9810 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9811 ctx->eventfd_async = 1;
9813 ctx->eventfd_async = 0;
9815 case IORING_UNREGISTER_EVENTFD:
9819 ret = io_eventfd_unregister(ctx);
9821 case IORING_REGISTER_PROBE:
9823 if (!arg || nr_args > 256)
9825 ret = io_probe(ctx, arg, nr_args);
9827 case IORING_REGISTER_PERSONALITY:
9831 ret = io_register_personality(ctx);
9833 case IORING_UNREGISTER_PERSONALITY:
9837 ret = io_unregister_personality(ctx, nr_args);
9839 case IORING_REGISTER_ENABLE_RINGS:
9843 ret = io_register_enable_rings(ctx);
9845 case IORING_REGISTER_RESTRICTIONS:
9846 ret = io_register_restrictions(ctx, arg, nr_args);
9854 if (io_register_op_must_quiesce(opcode)) {
9855 /* bring the ctx back to life */
9856 percpu_ref_reinit(&ctx->refs);
9858 reinit_completion(&ctx->ref_comp);
9863 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9864 void __user *, arg, unsigned int, nr_args)
9866 struct io_ring_ctx *ctx;
9875 if (f.file->f_op != &io_uring_fops)
9878 ctx = f.file->private_data;
9880 mutex_lock(&ctx->uring_lock);
9881 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9882 mutex_unlock(&ctx->uring_lock);
9883 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9884 ctx->cq_ev_fd != NULL, ret);
9890 static int __init io_uring_init(void)
9892 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9893 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9894 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9897 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9898 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9899 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9900 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9901 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9902 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9903 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9904 BUILD_BUG_SQE_ELEM(8, __u64, off);
9905 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9906 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9907 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9908 BUILD_BUG_SQE_ELEM(24, __u32, len);
9909 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9910 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9911 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9912 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9913 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9914 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9915 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9916 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9917 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9918 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9919 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9920 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9921 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9922 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9923 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9924 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9925 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9926 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9927 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9929 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9930 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9931 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9934 __initcall(io_uring_init);