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 list_head iopoll_entry;
701 struct percpu_ref *fixed_file_refs;
702 struct callback_head task_work;
703 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
704 struct hlist_node hash_node;
705 struct async_poll *apoll;
706 struct io_wq_work work;
709 struct io_defer_entry {
710 struct list_head list;
711 struct io_kiocb *req;
715 #define IO_IOPOLL_BATCH 8
717 struct io_comp_state {
719 struct list_head list;
720 struct io_ring_ctx *ctx;
723 struct io_submit_state {
724 struct blk_plug plug;
727 * io_kiocb alloc cache
729 void *reqs[IO_IOPOLL_BATCH];
730 unsigned int free_reqs;
733 * Batch completion logic
735 struct io_comp_state comp;
738 * File reference cache
742 unsigned int has_refs;
743 unsigned int ios_left;
747 /* needs req->file assigned */
748 unsigned needs_file : 1;
749 /* don't fail if file grab fails */
750 unsigned needs_file_no_error : 1;
751 /* hash wq insertion if file is a regular file */
752 unsigned hash_reg_file : 1;
753 /* unbound wq insertion if file is a non-regular file */
754 unsigned unbound_nonreg_file : 1;
755 /* opcode is not supported by this kernel */
756 unsigned not_supported : 1;
757 /* set if opcode supports polled "wait" */
759 unsigned pollout : 1;
760 /* op supports buffer selection */
761 unsigned buffer_select : 1;
762 /* must always have async data allocated */
763 unsigned needs_async_data : 1;
764 /* size of async data needed, if any */
765 unsigned short async_size;
769 static const struct io_op_def io_op_defs[] = {
770 [IORING_OP_NOP] = {},
771 [IORING_OP_READV] = {
773 .unbound_nonreg_file = 1,
776 .needs_async_data = 1,
777 .async_size = sizeof(struct io_async_rw),
778 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
781 [IORING_OP_WRITEV] = {
784 .unbound_nonreg_file = 1,
786 .needs_async_data = 1,
787 .async_size = sizeof(struct io_async_rw),
788 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
789 IO_WQ_WORK_FSIZE | IO_WQ_WORK_FILES,
791 [IORING_OP_FSYNC] = {
793 .work_flags = IO_WQ_WORK_BLKCG,
795 [IORING_OP_READ_FIXED] = {
797 .unbound_nonreg_file = 1,
799 .async_size = sizeof(struct io_async_rw),
800 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM |
803 [IORING_OP_WRITE_FIXED] = {
806 .unbound_nonreg_file = 1,
808 .async_size = sizeof(struct io_async_rw),
809 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
810 IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
812 [IORING_OP_POLL_ADD] = {
814 .unbound_nonreg_file = 1,
816 [IORING_OP_POLL_REMOVE] = {},
817 [IORING_OP_SYNC_FILE_RANGE] = {
819 .work_flags = IO_WQ_WORK_BLKCG,
821 [IORING_OP_SENDMSG] = {
823 .unbound_nonreg_file = 1,
825 .needs_async_data = 1,
826 .async_size = sizeof(struct io_async_msghdr),
827 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
830 [IORING_OP_RECVMSG] = {
832 .unbound_nonreg_file = 1,
835 .needs_async_data = 1,
836 .async_size = sizeof(struct io_async_msghdr),
837 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
840 [IORING_OP_TIMEOUT] = {
841 .needs_async_data = 1,
842 .async_size = sizeof(struct io_timeout_data),
843 .work_flags = IO_WQ_WORK_MM,
845 [IORING_OP_TIMEOUT_REMOVE] = {},
846 [IORING_OP_ACCEPT] = {
848 .unbound_nonreg_file = 1,
850 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
852 [IORING_OP_ASYNC_CANCEL] = {},
853 [IORING_OP_LINK_TIMEOUT] = {
854 .needs_async_data = 1,
855 .async_size = sizeof(struct io_timeout_data),
856 .work_flags = IO_WQ_WORK_MM,
858 [IORING_OP_CONNECT] = {
860 .unbound_nonreg_file = 1,
862 .needs_async_data = 1,
863 .async_size = sizeof(struct io_async_connect),
864 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FS,
866 [IORING_OP_FALLOCATE] = {
868 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
870 [IORING_OP_OPENAT] = {
871 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
874 [IORING_OP_CLOSE] = {
876 .needs_file_no_error = 1,
877 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
879 [IORING_OP_FILES_UPDATE] = {
880 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
882 [IORING_OP_STATX] = {
883 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
884 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
888 .unbound_nonreg_file = 1,
891 .async_size = sizeof(struct io_async_rw),
892 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
895 [IORING_OP_WRITE] = {
898 .unbound_nonreg_file = 1,
900 .async_size = sizeof(struct io_async_rw),
901 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
902 IO_WQ_WORK_FSIZE | IO_WQ_WORK_FILES,
904 [IORING_OP_FADVISE] = {
906 .work_flags = IO_WQ_WORK_BLKCG,
908 [IORING_OP_MADVISE] = {
909 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
913 .unbound_nonreg_file = 1,
915 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
920 .unbound_nonreg_file = 1,
923 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
926 [IORING_OP_OPENAT2] = {
927 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
930 [IORING_OP_EPOLL_CTL] = {
931 .unbound_nonreg_file = 1,
932 .work_flags = IO_WQ_WORK_FILES,
934 [IORING_OP_SPLICE] = {
937 .unbound_nonreg_file = 1,
938 .work_flags = IO_WQ_WORK_BLKCG,
940 [IORING_OP_PROVIDE_BUFFERS] = {},
941 [IORING_OP_REMOVE_BUFFERS] = {},
945 .unbound_nonreg_file = 1,
949 enum io_mem_account {
954 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node);
955 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
956 struct io_ring_ctx *ctx);
958 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
959 struct io_comp_state *cs);
960 static void io_cqring_fill_event(struct io_kiocb *req, long res);
961 static void io_put_req(struct io_kiocb *req);
962 static void io_put_req_deferred(struct io_kiocb *req, int nr);
963 static void io_double_put_req(struct io_kiocb *req);
964 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
965 static void __io_queue_linked_timeout(struct io_kiocb *req);
966 static void io_queue_linked_timeout(struct io_kiocb *req);
967 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
968 struct io_uring_files_update *ip,
970 static void __io_clean_op(struct io_kiocb *req);
971 static struct file *io_file_get(struct io_submit_state *state,
972 struct io_kiocb *req, int fd, bool fixed);
973 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
974 static void io_file_put_work(struct work_struct *work);
976 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
977 struct iovec **iovec, struct iov_iter *iter,
979 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
980 const struct iovec *fast_iov,
981 struct iov_iter *iter, bool force);
982 static void io_req_drop_files(struct io_kiocb *req);
983 static void io_req_task_queue(struct io_kiocb *req);
985 static struct kmem_cache *req_cachep;
987 static const struct file_operations io_uring_fops;
989 struct sock *io_uring_get_socket(struct file *file)
991 #if defined(CONFIG_UNIX)
992 if (file->f_op == &io_uring_fops) {
993 struct io_ring_ctx *ctx = file->private_data;
995 return ctx->ring_sock->sk;
1000 EXPORT_SYMBOL(io_uring_get_socket);
1002 static inline void io_clean_op(struct io_kiocb *req)
1004 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1008 static inline bool __io_match_files(struct io_kiocb *req,
1009 struct files_struct *files)
1011 if (req->file && req->file->f_op == &io_uring_fops)
1014 return ((req->flags & REQ_F_WORK_INITIALIZED) &&
1015 (req->work.flags & IO_WQ_WORK_FILES)) &&
1016 req->work.identity->files == files;
1019 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1021 bool got = percpu_ref_tryget(ref);
1023 /* already at zero, wait for ->release() */
1025 wait_for_completion(compl);
1026 percpu_ref_resurrect(ref);
1028 percpu_ref_put(ref);
1031 static bool io_match_task(struct io_kiocb *head,
1032 struct task_struct *task,
1033 struct files_struct *files)
1035 struct io_kiocb *link;
1037 if (task && head->task != task) {
1038 /* in terms of cancelation, always match if req task is dead */
1039 if (head->task->flags & PF_EXITING)
1045 if (__io_match_files(head, files))
1047 if (head->flags & REQ_F_LINK_HEAD) {
1048 list_for_each_entry(link, &head->link_list, link_list) {
1049 if (__io_match_files(link, files))
1057 static void io_sq_thread_drop_mm(void)
1059 struct mm_struct *mm = current->mm;
1062 kthread_unuse_mm(mm);
1068 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1070 struct mm_struct *mm;
1072 if (current->flags & PF_EXITING)
1077 /* Should never happen */
1078 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL)))
1081 task_lock(ctx->sqo_task);
1082 mm = ctx->sqo_task->mm;
1083 if (unlikely(!mm || !mmget_not_zero(mm)))
1085 task_unlock(ctx->sqo_task);
1095 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1096 struct io_kiocb *req)
1098 if (!(io_op_defs[req->opcode].work_flags & IO_WQ_WORK_MM))
1100 return __io_sq_thread_acquire_mm(ctx);
1103 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1104 struct cgroup_subsys_state **cur_css)
1107 #ifdef CONFIG_BLK_CGROUP
1108 /* puts the old one when swapping */
1109 if (*cur_css != ctx->sqo_blkcg_css) {
1110 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1111 *cur_css = ctx->sqo_blkcg_css;
1116 static void io_sq_thread_unassociate_blkcg(void)
1118 #ifdef CONFIG_BLK_CGROUP
1119 kthread_associate_blkcg(NULL);
1123 static inline void req_set_fail_links(struct io_kiocb *req)
1125 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1126 req->flags |= REQ_F_FAIL_LINK;
1130 * None of these are dereferenced, they are simply used to check if any of
1131 * them have changed. If we're under current and check they are still the
1132 * same, we're fine to grab references to them for actual out-of-line use.
1134 static void io_init_identity(struct io_identity *id)
1136 id->files = current->files;
1137 id->mm = current->mm;
1138 #ifdef CONFIG_BLK_CGROUP
1140 id->blkcg_css = blkcg_css();
1143 id->creds = current_cred();
1144 id->nsproxy = current->nsproxy;
1145 id->fs = current->fs;
1146 id->fsize = rlimit(RLIMIT_FSIZE);
1148 id->loginuid = current->loginuid;
1149 id->sessionid = current->sessionid;
1151 refcount_set(&id->count, 1);
1154 static inline void __io_req_init_async(struct io_kiocb *req)
1156 memset(&req->work, 0, sizeof(req->work));
1157 req->flags |= REQ_F_WORK_INITIALIZED;
1161 * Note: must call io_req_init_async() for the first time you
1162 * touch any members of io_wq_work.
1164 static inline void io_req_init_async(struct io_kiocb *req)
1166 struct io_uring_task *tctx = req->task->io_uring;
1168 if (req->flags & REQ_F_WORK_INITIALIZED)
1171 __io_req_init_async(req);
1173 /* Grab a ref if this isn't our static identity */
1174 req->work.identity = tctx->identity;
1175 if (tctx->identity != &tctx->__identity)
1176 refcount_inc(&req->work.identity->count);
1179 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1181 return ctx->flags & IORING_SETUP_SQPOLL;
1184 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1186 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1188 complete(&ctx->ref_comp);
1191 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1193 return !req->timeout.off;
1196 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1198 struct io_ring_ctx *ctx;
1201 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1205 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1206 if (!ctx->fallback_req)
1210 * Use 5 bits less than the max cq entries, that should give us around
1211 * 32 entries per hash list if totally full and uniformly spread.
1213 hash_bits = ilog2(p->cq_entries);
1217 ctx->cancel_hash_bits = hash_bits;
1218 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1220 if (!ctx->cancel_hash)
1222 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1224 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1225 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1228 ctx->flags = p->flags;
1229 init_waitqueue_head(&ctx->sqo_sq_wait);
1230 INIT_LIST_HEAD(&ctx->sqd_list);
1231 init_waitqueue_head(&ctx->cq_wait);
1232 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1233 init_completion(&ctx->ref_comp);
1234 init_completion(&ctx->sq_thread_comp);
1235 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1236 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1237 mutex_init(&ctx->uring_lock);
1238 init_waitqueue_head(&ctx->wait);
1239 spin_lock_init(&ctx->completion_lock);
1240 INIT_LIST_HEAD(&ctx->iopoll_list);
1241 INIT_LIST_HEAD(&ctx->defer_list);
1242 INIT_LIST_HEAD(&ctx->timeout_list);
1243 spin_lock_init(&ctx->inflight_lock);
1244 INIT_LIST_HEAD(&ctx->inflight_list);
1245 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1246 init_llist_head(&ctx->file_put_llist);
1249 if (ctx->fallback_req)
1250 kmem_cache_free(req_cachep, ctx->fallback_req);
1251 kfree(ctx->cancel_hash);
1256 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1258 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1259 struct io_ring_ctx *ctx = req->ctx;
1261 return seq != ctx->cached_cq_tail
1262 + READ_ONCE(ctx->cached_cq_overflow);
1268 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1270 struct io_rings *rings = ctx->rings;
1272 /* order cqe stores with ring update */
1273 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1276 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1278 if (req->work.identity == &tctx->__identity)
1280 if (refcount_dec_and_test(&req->work.identity->count))
1281 kfree(req->work.identity);
1284 static void io_req_clean_work(struct io_kiocb *req)
1286 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1289 req->flags &= ~REQ_F_WORK_INITIALIZED;
1291 if (req->work.flags & IO_WQ_WORK_MM) {
1292 mmdrop(req->work.identity->mm);
1293 req->work.flags &= ~IO_WQ_WORK_MM;
1295 #ifdef CONFIG_BLK_CGROUP
1296 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1297 css_put(req->work.identity->blkcg_css);
1298 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1301 if (req->work.flags & IO_WQ_WORK_CREDS) {
1302 put_cred(req->work.identity->creds);
1303 req->work.flags &= ~IO_WQ_WORK_CREDS;
1305 if (req->work.flags & IO_WQ_WORK_FS) {
1306 struct fs_struct *fs = req->work.identity->fs;
1308 spin_lock(&req->work.identity->fs->lock);
1311 spin_unlock(&req->work.identity->fs->lock);
1314 req->work.flags &= ~IO_WQ_WORK_FS;
1316 if (req->flags & REQ_F_INFLIGHT)
1317 io_req_drop_files(req);
1319 io_put_identity(req->task->io_uring, req);
1323 * Create a private copy of io_identity, since some fields don't match
1324 * the current context.
1326 static bool io_identity_cow(struct io_kiocb *req)
1328 struct io_uring_task *tctx = req->task->io_uring;
1329 const struct cred *creds = NULL;
1330 struct io_identity *id;
1332 if (req->work.flags & IO_WQ_WORK_CREDS)
1333 creds = req->work.identity->creds;
1335 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1336 if (unlikely(!id)) {
1337 req->work.flags |= IO_WQ_WORK_CANCEL;
1342 * We can safely just re-init the creds we copied Either the field
1343 * matches the current one, or we haven't grabbed it yet. The only
1344 * exception is ->creds, through registered personalities, so handle
1345 * that one separately.
1347 io_init_identity(id);
1351 /* add one for this request */
1352 refcount_inc(&id->count);
1354 /* drop tctx and req identity references, if needed */
1355 if (tctx->identity != &tctx->__identity &&
1356 refcount_dec_and_test(&tctx->identity->count))
1357 kfree(tctx->identity);
1358 if (req->work.identity != &tctx->__identity &&
1359 refcount_dec_and_test(&req->work.identity->count))
1360 kfree(req->work.identity);
1362 req->work.identity = id;
1363 tctx->identity = id;
1367 static bool io_grab_identity(struct io_kiocb *req)
1369 const struct io_op_def *def = &io_op_defs[req->opcode];
1370 struct io_identity *id = req->work.identity;
1371 struct io_ring_ctx *ctx = req->ctx;
1373 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1374 if (id->fsize != rlimit(RLIMIT_FSIZE))
1376 req->work.flags |= IO_WQ_WORK_FSIZE;
1378 #ifdef CONFIG_BLK_CGROUP
1379 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1380 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1382 if (id->blkcg_css != blkcg_css()) {
1387 * This should be rare, either the cgroup is dying or the task
1388 * is moving cgroups. Just punt to root for the handful of ios.
1390 if (css_tryget_online(id->blkcg_css))
1391 req->work.flags |= IO_WQ_WORK_BLKCG;
1395 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1396 if (id->creds != current_cred())
1398 get_cred(id->creds);
1399 req->work.flags |= IO_WQ_WORK_CREDS;
1402 if (!uid_eq(current->loginuid, id->loginuid) ||
1403 current->sessionid != id->sessionid)
1406 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1407 (def->work_flags & IO_WQ_WORK_FS)) {
1408 if (current->fs != id->fs)
1410 spin_lock(&id->fs->lock);
1411 if (!id->fs->in_exec) {
1413 req->work.flags |= IO_WQ_WORK_FS;
1415 req->work.flags |= IO_WQ_WORK_CANCEL;
1417 spin_unlock(¤t->fs->lock);
1419 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1420 (def->work_flags & IO_WQ_WORK_FILES) &&
1421 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1422 if (id->files != current->files ||
1423 id->nsproxy != current->nsproxy)
1425 atomic_inc(&id->files->count);
1426 get_nsproxy(id->nsproxy);
1428 if (!(req->flags & REQ_F_INFLIGHT)) {
1429 req->flags |= REQ_F_INFLIGHT;
1431 spin_lock_irq(&ctx->inflight_lock);
1432 list_add(&req->inflight_entry, &ctx->inflight_list);
1433 spin_unlock_irq(&ctx->inflight_lock);
1435 req->work.flags |= IO_WQ_WORK_FILES;
1437 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1438 (def->work_flags & IO_WQ_WORK_MM)) {
1439 if (id->mm != current->mm)
1442 req->work.flags |= IO_WQ_WORK_MM;
1448 static void io_prep_async_work(struct io_kiocb *req)
1450 const struct io_op_def *def = &io_op_defs[req->opcode];
1451 struct io_ring_ctx *ctx = req->ctx;
1452 struct io_identity *id;
1454 io_req_init_async(req);
1455 id = req->work.identity;
1457 if (req->flags & REQ_F_FORCE_ASYNC)
1458 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1460 if (req->flags & REQ_F_ISREG) {
1461 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1462 io_wq_hash_work(&req->work, file_inode(req->file));
1463 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1464 if (def->unbound_nonreg_file)
1465 req->work.flags |= IO_WQ_WORK_UNBOUND;
1468 /* if we fail grabbing identity, we must COW, regrab, and retry */
1469 if (io_grab_identity(req))
1472 if (!io_identity_cow(req))
1475 /* can't fail at this point */
1476 if (!io_grab_identity(req))
1480 static void io_prep_async_link(struct io_kiocb *req)
1482 struct io_kiocb *cur;
1484 io_prep_async_work(req);
1485 if (req->flags & REQ_F_LINK_HEAD)
1486 list_for_each_entry(cur, &req->link_list, link_list)
1487 io_prep_async_work(cur);
1490 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1492 struct io_ring_ctx *ctx = req->ctx;
1493 struct io_kiocb *link = io_prep_linked_timeout(req);
1495 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1496 &req->work, req->flags);
1497 io_wq_enqueue(ctx->io_wq, &req->work);
1501 static void io_queue_async_work(struct io_kiocb *req)
1503 struct io_kiocb *link;
1505 /* init ->work of the whole link before punting */
1506 io_prep_async_link(req);
1507 link = __io_queue_async_work(req);
1510 io_queue_linked_timeout(link);
1513 static void io_kill_timeout(struct io_kiocb *req, int status)
1515 struct io_timeout_data *io = req->async_data;
1518 ret = hrtimer_try_to_cancel(&io->timer);
1521 req_set_fail_links(req);
1522 atomic_set(&req->ctx->cq_timeouts,
1523 atomic_read(&req->ctx->cq_timeouts) + 1);
1524 list_del_init(&req->timeout.list);
1525 io_cqring_fill_event(req, status);
1526 io_put_req_deferred(req, 1);
1531 * Returns true if we found and killed one or more timeouts
1533 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1534 struct files_struct *files)
1536 struct io_kiocb *req, *tmp;
1539 spin_lock_irq(&ctx->completion_lock);
1540 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1541 if (io_match_task(req, tsk, files)) {
1542 io_kill_timeout(req, -ECANCELED);
1546 spin_unlock_irq(&ctx->completion_lock);
1547 return canceled != 0;
1550 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1553 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1554 struct io_defer_entry, list);
1556 if (req_need_defer(de->req, de->seq))
1558 list_del_init(&de->list);
1559 io_req_task_queue(de->req);
1561 } while (!list_empty(&ctx->defer_list));
1564 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1566 struct io_kiocb *req, *tmp;
1569 if (list_empty(&ctx->timeout_list))
1572 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1574 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1575 u32 events_needed, events_got;
1577 if (io_is_timeout_noseq(req))
1581 * Since seq can easily wrap around over time, subtract
1582 * the last seq at which timeouts were flushed before comparing.
1583 * Assuming not more than 2^31-1 events have happened since,
1584 * these subtractions won't have wrapped, so we can check if
1585 * target is in [last_seq, current_seq] by comparing the two.
1587 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1588 events_got = seq - ctx->cq_last_tm_flush;
1589 if (events_got < events_needed)
1592 io_kill_timeout(req, 0);
1595 ctx->cq_last_tm_flush = seq;
1598 static void io_commit_cqring(struct io_ring_ctx *ctx)
1600 io_flush_timeouts(ctx);
1601 __io_commit_cqring(ctx);
1603 if (unlikely(!list_empty(&ctx->defer_list)))
1604 __io_queue_deferred(ctx);
1607 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1609 struct io_rings *r = ctx->rings;
1611 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1614 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1616 struct io_rings *rings = ctx->rings;
1619 tail = ctx->cached_cq_tail;
1621 * writes to the cq entry need to come after reading head; the
1622 * control dependency is enough as we're using WRITE_ONCE to
1625 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1628 ctx->cached_cq_tail++;
1629 return &rings->cqes[tail & ctx->cq_mask];
1632 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1636 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1638 if (!ctx->eventfd_async)
1640 return io_wq_current_is_worker();
1643 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1645 if (wq_has_sleeper(&ctx->cq_wait)) {
1646 wake_up_interruptible(&ctx->cq_wait);
1647 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1649 if (waitqueue_active(&ctx->wait))
1650 wake_up(&ctx->wait);
1651 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1652 wake_up(&ctx->sq_data->wait);
1653 if (io_should_trigger_evfd(ctx))
1654 eventfd_signal(ctx->cq_ev_fd, 1);
1657 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1659 if (list_empty(&ctx->cq_overflow_list)) {
1660 clear_bit(0, &ctx->sq_check_overflow);
1661 clear_bit(0, &ctx->cq_check_overflow);
1662 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1666 /* Returns true if there are no backlogged entries after the flush */
1667 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1668 struct task_struct *tsk,
1669 struct files_struct *files)
1671 struct io_rings *rings = ctx->rings;
1672 struct io_kiocb *req, *tmp;
1673 struct io_uring_cqe *cqe;
1674 unsigned long flags;
1678 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1679 rings->cq_ring_entries))
1683 spin_lock_irqsave(&ctx->completion_lock, flags);
1686 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1687 if (!io_match_task(req, tsk, files))
1690 cqe = io_get_cqring(ctx);
1694 list_move(&req->compl.list, &list);
1696 WRITE_ONCE(cqe->user_data, req->user_data);
1697 WRITE_ONCE(cqe->res, req->result);
1698 WRITE_ONCE(cqe->flags, req->compl.cflags);
1700 ctx->cached_cq_overflow++;
1701 WRITE_ONCE(ctx->rings->cq_overflow,
1702 ctx->cached_cq_overflow);
1706 io_commit_cqring(ctx);
1707 io_cqring_mark_overflow(ctx);
1709 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1710 io_cqring_ev_posted(ctx);
1712 while (!list_empty(&list)) {
1713 req = list_first_entry(&list, struct io_kiocb, compl.list);
1714 list_del(&req->compl.list);
1721 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1722 struct task_struct *tsk,
1723 struct files_struct *files)
1725 if (test_bit(0, &ctx->cq_check_overflow)) {
1726 /* iopoll syncs against uring_lock, not completion_lock */
1727 if (ctx->flags & IORING_SETUP_IOPOLL)
1728 mutex_lock(&ctx->uring_lock);
1729 __io_cqring_overflow_flush(ctx, force, tsk, files);
1730 if (ctx->flags & IORING_SETUP_IOPOLL)
1731 mutex_unlock(&ctx->uring_lock);
1735 static void __io_cqring_fill_event(struct io_kiocb *req, long res,
1736 unsigned int cflags)
1738 struct io_ring_ctx *ctx = req->ctx;
1739 struct io_uring_cqe *cqe;
1741 trace_io_uring_complete(ctx, req->user_data, res);
1744 * If we can't get a cq entry, userspace overflowed the
1745 * submission (by quite a lot). Increment the overflow count in
1748 cqe = io_get_cqring(ctx);
1750 WRITE_ONCE(cqe->user_data, req->user_data);
1751 WRITE_ONCE(cqe->res, res);
1752 WRITE_ONCE(cqe->flags, cflags);
1753 } else if (ctx->cq_overflow_flushed ||
1754 atomic_read(&req->task->io_uring->in_idle)) {
1756 * If we're in ring overflow flush mode, or in task cancel mode,
1757 * then we cannot store the request for later flushing, we need
1758 * to drop it on the floor.
1760 ctx->cached_cq_overflow++;
1761 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1763 if (list_empty(&ctx->cq_overflow_list)) {
1764 set_bit(0, &ctx->sq_check_overflow);
1765 set_bit(0, &ctx->cq_check_overflow);
1766 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1770 req->compl.cflags = cflags;
1771 refcount_inc(&req->refs);
1772 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1776 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1778 __io_cqring_fill_event(req, res, 0);
1781 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1783 struct io_ring_ctx *ctx = req->ctx;
1784 unsigned long flags;
1786 spin_lock_irqsave(&ctx->completion_lock, flags);
1787 __io_cqring_fill_event(req, res, cflags);
1788 io_commit_cqring(ctx);
1789 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1791 io_cqring_ev_posted(ctx);
1794 static void io_submit_flush_completions(struct io_comp_state *cs)
1796 struct io_ring_ctx *ctx = cs->ctx;
1798 spin_lock_irq(&ctx->completion_lock);
1799 while (!list_empty(&cs->list)) {
1800 struct io_kiocb *req;
1802 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1803 list_del(&req->compl.list);
1804 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1807 * io_free_req() doesn't care about completion_lock unless one
1808 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1809 * because of a potential deadlock with req->work.fs->lock
1811 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1812 |REQ_F_WORK_INITIALIZED)) {
1813 spin_unlock_irq(&ctx->completion_lock);
1815 spin_lock_irq(&ctx->completion_lock);
1820 io_commit_cqring(ctx);
1821 spin_unlock_irq(&ctx->completion_lock);
1823 io_cqring_ev_posted(ctx);
1827 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1828 struct io_comp_state *cs)
1831 io_cqring_add_event(req, res, cflags);
1836 req->compl.cflags = cflags;
1837 list_add_tail(&req->compl.list, &cs->list);
1839 io_submit_flush_completions(cs);
1843 static void io_req_complete(struct io_kiocb *req, long res)
1845 __io_req_complete(req, res, 0, NULL);
1848 static inline bool io_is_fallback_req(struct io_kiocb *req)
1850 return req == (struct io_kiocb *)
1851 ((unsigned long) req->ctx->fallback_req & ~1UL);
1854 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1856 struct io_kiocb *req;
1858 req = ctx->fallback_req;
1859 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1865 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1866 struct io_submit_state *state)
1868 if (!state->free_reqs) {
1869 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1873 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1874 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1877 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1878 * retry single alloc to be on the safe side.
1880 if (unlikely(ret <= 0)) {
1881 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1882 if (!state->reqs[0])
1886 state->free_reqs = ret;
1890 return state->reqs[state->free_reqs];
1892 return io_get_fallback_req(ctx);
1895 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1899 percpu_ref_put(req->fixed_file_refs);
1904 static void io_dismantle_req(struct io_kiocb *req)
1908 if (req->async_data)
1909 kfree(req->async_data);
1911 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1913 io_req_clean_work(req);
1916 static void __io_free_req(struct io_kiocb *req)
1918 struct io_uring_task *tctx = req->task->io_uring;
1919 struct io_ring_ctx *ctx = req->ctx;
1921 io_dismantle_req(req);
1923 percpu_counter_dec(&tctx->inflight);
1924 if (atomic_read(&tctx->in_idle))
1925 wake_up(&tctx->wait);
1926 put_task_struct(req->task);
1928 if (likely(!io_is_fallback_req(req)))
1929 kmem_cache_free(req_cachep, req);
1931 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1932 percpu_ref_put(&ctx->refs);
1935 static void io_kill_linked_timeout(struct io_kiocb *req)
1937 struct io_ring_ctx *ctx = req->ctx;
1938 struct io_kiocb *link;
1939 bool cancelled = false;
1940 unsigned long flags;
1942 spin_lock_irqsave(&ctx->completion_lock, flags);
1943 link = list_first_entry_or_null(&req->link_list, struct io_kiocb,
1946 * Can happen if a linked timeout fired and link had been like
1947 * req -> link t-out -> link t-out [-> ...]
1949 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1950 struct io_timeout_data *io = link->async_data;
1953 list_del_init(&link->link_list);
1954 ret = hrtimer_try_to_cancel(&io->timer);
1956 io_cqring_fill_event(link, -ECANCELED);
1957 io_commit_cqring(ctx);
1961 req->flags &= ~REQ_F_LINK_TIMEOUT;
1962 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1965 io_cqring_ev_posted(ctx);
1970 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1972 struct io_kiocb *nxt;
1975 * The list should never be empty when we are called here. But could
1976 * potentially happen if the chain is messed up, check to be on the
1979 if (unlikely(list_empty(&req->link_list)))
1982 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1983 list_del_init(&req->link_list);
1984 if (!list_empty(&nxt->link_list))
1985 nxt->flags |= REQ_F_LINK_HEAD;
1990 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1992 static void io_fail_links(struct io_kiocb *req)
1994 struct io_ring_ctx *ctx = req->ctx;
1995 unsigned long flags;
1997 spin_lock_irqsave(&ctx->completion_lock, flags);
1998 while (!list_empty(&req->link_list)) {
1999 struct io_kiocb *link = list_first_entry(&req->link_list,
2000 struct io_kiocb, link_list);
2002 list_del_init(&link->link_list);
2003 trace_io_uring_fail_link(req, link);
2005 io_cqring_fill_event(link, -ECANCELED);
2008 * It's ok to free under spinlock as they're not linked anymore,
2009 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2012 if (link->flags & REQ_F_WORK_INITIALIZED)
2013 io_put_req_deferred(link, 2);
2015 io_double_put_req(link);
2018 io_commit_cqring(ctx);
2019 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2021 io_cqring_ev_posted(ctx);
2024 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2026 req->flags &= ~REQ_F_LINK_HEAD;
2027 if (req->flags & REQ_F_LINK_TIMEOUT)
2028 io_kill_linked_timeout(req);
2031 * If LINK is set, we have dependent requests in this chain. If we
2032 * didn't fail this request, queue the first one up, moving any other
2033 * dependencies to the next request. In case of failure, fail the rest
2036 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
2037 return io_req_link_next(req);
2042 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2044 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
2046 return __io_req_find_next(req);
2049 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
2051 struct task_struct *tsk = req->task;
2052 struct io_ring_ctx *ctx = req->ctx;
2053 enum task_work_notify_mode notify;
2056 if (tsk->flags & PF_EXITING)
2060 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2061 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2062 * processing task_work. There's no reliable way to tell if TWA_RESUME
2066 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
2067 notify = TWA_SIGNAL;
2069 ret = task_work_add(tsk, &req->task_work, notify);
2071 wake_up_process(tsk);
2076 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2078 struct io_ring_ctx *ctx = req->ctx;
2080 spin_lock_irq(&ctx->completion_lock);
2081 io_cqring_fill_event(req, error);
2082 io_commit_cqring(ctx);
2083 spin_unlock_irq(&ctx->completion_lock);
2085 io_cqring_ev_posted(ctx);
2086 req_set_fail_links(req);
2087 io_double_put_req(req);
2090 static void io_req_task_cancel(struct callback_head *cb)
2092 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2093 struct io_ring_ctx *ctx = req->ctx;
2095 mutex_lock(&ctx->uring_lock);
2096 __io_req_task_cancel(req, -ECANCELED);
2097 mutex_unlock(&ctx->uring_lock);
2098 percpu_ref_put(&ctx->refs);
2101 static void __io_req_task_submit(struct io_kiocb *req)
2103 struct io_ring_ctx *ctx = req->ctx;
2105 mutex_lock(&ctx->uring_lock);
2106 if (!ctx->sqo_dead && !__io_sq_thread_acquire_mm(ctx))
2107 __io_queue_sqe(req, NULL);
2109 __io_req_task_cancel(req, -EFAULT);
2110 mutex_unlock(&ctx->uring_lock);
2112 if (ctx->flags & IORING_SETUP_SQPOLL)
2113 io_sq_thread_drop_mm();
2116 static void io_req_task_submit(struct callback_head *cb)
2118 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2119 struct io_ring_ctx *ctx = req->ctx;
2121 __io_req_task_submit(req);
2122 percpu_ref_put(&ctx->refs);
2125 static void io_req_task_queue(struct io_kiocb *req)
2129 init_task_work(&req->task_work, io_req_task_submit);
2130 percpu_ref_get(&req->ctx->refs);
2132 ret = io_req_task_work_add(req, true);
2133 if (unlikely(ret)) {
2134 struct task_struct *tsk;
2136 init_task_work(&req->task_work, io_req_task_cancel);
2137 tsk = io_wq_get_task(req->ctx->io_wq);
2138 task_work_add(tsk, &req->task_work, TWA_NONE);
2139 wake_up_process(tsk);
2143 static void io_queue_next(struct io_kiocb *req)
2145 struct io_kiocb *nxt = io_req_find_next(req);
2148 io_req_task_queue(nxt);
2151 static void io_free_req(struct io_kiocb *req)
2158 void *reqs[IO_IOPOLL_BATCH];
2161 struct task_struct *task;
2165 static inline void io_init_req_batch(struct req_batch *rb)
2172 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2173 struct req_batch *rb)
2175 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2176 percpu_ref_put_many(&ctx->refs, rb->to_free);
2180 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2181 struct req_batch *rb)
2184 __io_req_free_batch_flush(ctx, rb);
2186 struct io_uring_task *tctx = rb->task->io_uring;
2188 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2189 if (atomic_read(&tctx->in_idle))
2190 wake_up(&tctx->wait);
2191 put_task_struct_many(rb->task, rb->task_refs);
2196 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2198 if (unlikely(io_is_fallback_req(req))) {
2202 if (req->flags & REQ_F_LINK_HEAD)
2205 if (req->task != rb->task) {
2207 struct io_uring_task *tctx = rb->task->io_uring;
2209 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2210 if (atomic_read(&tctx->in_idle))
2211 wake_up(&tctx->wait);
2212 put_task_struct_many(rb->task, rb->task_refs);
2214 rb->task = req->task;
2219 io_dismantle_req(req);
2220 rb->reqs[rb->to_free++] = req;
2221 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2222 __io_req_free_batch_flush(req->ctx, rb);
2226 * Drop reference to request, return next in chain (if there is one) if this
2227 * was the last reference to this request.
2229 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2231 struct io_kiocb *nxt = NULL;
2233 if (refcount_dec_and_test(&req->refs)) {
2234 nxt = io_req_find_next(req);
2240 static void io_put_req(struct io_kiocb *req)
2242 if (refcount_dec_and_test(&req->refs))
2246 static void io_put_req_deferred_cb(struct callback_head *cb)
2248 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2253 static void io_free_req_deferred(struct io_kiocb *req)
2257 init_task_work(&req->task_work, io_put_req_deferred_cb);
2258 ret = io_req_task_work_add(req, true);
2259 if (unlikely(ret)) {
2260 struct task_struct *tsk;
2262 tsk = io_wq_get_task(req->ctx->io_wq);
2263 task_work_add(tsk, &req->task_work, TWA_NONE);
2264 wake_up_process(tsk);
2268 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2270 if (refcount_sub_and_test(refs, &req->refs))
2271 io_free_req_deferred(req);
2274 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2276 struct io_kiocb *nxt;
2279 * A ref is owned by io-wq in which context we're. So, if that's the
2280 * last one, it's safe to steal next work. False negatives are Ok,
2281 * it just will be re-punted async in io_put_work()
2283 if (refcount_read(&req->refs) != 1)
2286 nxt = io_req_find_next(req);
2287 return nxt ? &nxt->work : NULL;
2290 static void io_double_put_req(struct io_kiocb *req)
2292 /* drop both submit and complete references */
2293 if (refcount_sub_and_test(2, &req->refs))
2297 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2299 struct io_rings *rings = ctx->rings;
2301 /* See comment at the top of this file */
2303 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2306 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2308 struct io_rings *rings = ctx->rings;
2310 /* make sure SQ entry isn't read before tail */
2311 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2314 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2316 unsigned int cflags;
2318 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2319 cflags |= IORING_CQE_F_BUFFER;
2320 req->flags &= ~REQ_F_BUFFER_SELECTED;
2325 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2327 struct io_buffer *kbuf;
2329 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2330 return io_put_kbuf(req, kbuf);
2333 static inline bool io_run_task_work(void)
2336 * Not safe to run on exiting task, and the task_work handling will
2337 * not add work to such a task.
2339 if (unlikely(current->flags & PF_EXITING))
2341 if (current->task_works) {
2342 __set_current_state(TASK_RUNNING);
2350 static void io_iopoll_queue(struct list_head *again)
2352 struct io_kiocb *req;
2355 req = list_first_entry(again, struct io_kiocb, iopoll_entry);
2356 list_del(&req->iopoll_entry);
2357 __io_complete_rw(req, -EAGAIN, 0, NULL);
2358 } while (!list_empty(again));
2362 * Find and free completed poll iocbs
2364 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2365 struct list_head *done)
2367 struct req_batch rb;
2368 struct io_kiocb *req;
2371 /* order with ->result store in io_complete_rw_iopoll() */
2374 io_init_req_batch(&rb);
2375 while (!list_empty(done)) {
2378 req = list_first_entry(done, struct io_kiocb, iopoll_entry);
2379 if (READ_ONCE(req->result) == -EAGAIN) {
2381 req->iopoll_completed = 0;
2382 list_move_tail(&req->iopoll_entry, &again);
2385 list_del(&req->iopoll_entry);
2387 if (req->flags & REQ_F_BUFFER_SELECTED)
2388 cflags = io_put_rw_kbuf(req);
2390 __io_cqring_fill_event(req, req->result, cflags);
2393 if (refcount_dec_and_test(&req->refs))
2394 io_req_free_batch(&rb, req);
2397 io_commit_cqring(ctx);
2398 if (ctx->flags & IORING_SETUP_SQPOLL)
2399 io_cqring_ev_posted(ctx);
2400 io_req_free_batch_finish(ctx, &rb);
2402 if (!list_empty(&again))
2403 io_iopoll_queue(&again);
2406 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2409 struct io_kiocb *req, *tmp;
2415 * Only spin for completions if we don't have multiple devices hanging
2416 * off our complete list, and we're under the requested amount.
2418 spin = !ctx->poll_multi_file && *nr_events < min;
2421 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, iopoll_entry) {
2422 struct kiocb *kiocb = &req->rw.kiocb;
2425 * Move completed and retryable entries to our local lists.
2426 * If we find a request that requires polling, break out
2427 * and complete those lists first, if we have entries there.
2429 if (READ_ONCE(req->iopoll_completed)) {
2430 list_move_tail(&req->iopoll_entry, &done);
2433 if (!list_empty(&done))
2436 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2440 /* iopoll may have completed current req */
2441 if (READ_ONCE(req->iopoll_completed))
2442 list_move_tail(&req->iopoll_entry, &done);
2449 if (!list_empty(&done))
2450 io_iopoll_complete(ctx, nr_events, &done);
2456 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2457 * non-spinning poll check - we'll still enter the driver poll loop, but only
2458 * as a non-spinning completion check.
2460 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2463 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2466 ret = io_do_iopoll(ctx, nr_events, min);
2469 if (*nr_events >= min)
2477 * We can't just wait for polled events to come to us, we have to actively
2478 * find and complete them.
2480 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2482 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2485 mutex_lock(&ctx->uring_lock);
2486 while (!list_empty(&ctx->iopoll_list)) {
2487 unsigned int nr_events = 0;
2489 io_do_iopoll(ctx, &nr_events, 0);
2491 /* let it sleep and repeat later if can't complete a request */
2495 * Ensure we allow local-to-the-cpu processing to take place,
2496 * in this case we need to ensure that we reap all events.
2497 * Also let task_work, etc. to progress by releasing the mutex
2499 if (need_resched()) {
2500 mutex_unlock(&ctx->uring_lock);
2502 mutex_lock(&ctx->uring_lock);
2505 mutex_unlock(&ctx->uring_lock);
2508 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2510 unsigned int nr_events = 0;
2511 int iters = 0, ret = 0;
2514 * We disallow the app entering submit/complete with polling, but we
2515 * still need to lock the ring to prevent racing with polled issue
2516 * that got punted to a workqueue.
2518 mutex_lock(&ctx->uring_lock);
2521 * Don't enter poll loop if we already have events pending.
2522 * If we do, we can potentially be spinning for commands that
2523 * already triggered a CQE (eg in error).
2525 if (test_bit(0, &ctx->cq_check_overflow))
2526 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2527 if (io_cqring_events(ctx))
2531 * If a submit got punted to a workqueue, we can have the
2532 * application entering polling for a command before it gets
2533 * issued. That app will hold the uring_lock for the duration
2534 * of the poll right here, so we need to take a breather every
2535 * now and then to ensure that the issue has a chance to add
2536 * the poll to the issued list. Otherwise we can spin here
2537 * forever, while the workqueue is stuck trying to acquire the
2540 if (!(++iters & 7)) {
2541 mutex_unlock(&ctx->uring_lock);
2543 mutex_lock(&ctx->uring_lock);
2546 ret = io_iopoll_getevents(ctx, &nr_events, min);
2550 } while (min && !nr_events && !need_resched());
2552 mutex_unlock(&ctx->uring_lock);
2556 static void kiocb_end_write(struct io_kiocb *req)
2559 * Tell lockdep we inherited freeze protection from submission
2562 if (req->flags & REQ_F_ISREG) {
2563 struct inode *inode = file_inode(req->file);
2565 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2567 file_end_write(req->file);
2570 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2571 struct io_comp_state *cs)
2573 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2576 if (kiocb->ki_flags & IOCB_WRITE)
2577 kiocb_end_write(req);
2579 if (res != req->result)
2580 req_set_fail_links(req);
2581 if (req->flags & REQ_F_BUFFER_SELECTED)
2582 cflags = io_put_rw_kbuf(req);
2583 __io_req_complete(req, res, cflags, cs);
2587 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2589 req_set_fail_links(req);
2594 static bool io_rw_reissue(struct io_kiocb *req, long res)
2597 umode_t mode = file_inode(req->file)->i_mode;
2600 if (!S_ISBLK(mode) && !S_ISREG(mode))
2602 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2605 * If ref is dying, we might be running poll reap from the exit work.
2606 * Don't attempt to reissue from that path, just let it fail with
2609 if (percpu_ref_is_dying(&req->ctx->refs))
2612 ret = io_sq_thread_acquire_mm(req->ctx, req);
2614 if (io_resubmit_prep(req, ret)) {
2615 refcount_inc(&req->refs);
2616 io_queue_async_work(req);
2624 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2625 struct io_comp_state *cs)
2627 if (!io_rw_reissue(req, res))
2628 io_complete_rw_common(&req->rw.kiocb, res, cs);
2631 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2633 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2635 __io_complete_rw(req, res, res2, NULL);
2638 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2640 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2642 if (kiocb->ki_flags & IOCB_WRITE)
2643 kiocb_end_write(req);
2645 if (res != -EAGAIN && res != req->result)
2646 req_set_fail_links(req);
2648 WRITE_ONCE(req->result, res);
2649 /* order with io_poll_complete() checking ->result */
2651 WRITE_ONCE(req->iopoll_completed, 1);
2655 * After the iocb has been issued, it's safe to be found on the poll list.
2656 * Adding the kiocb to the list AFTER submission ensures that we don't
2657 * find it from a io_iopoll_getevents() thread before the issuer is done
2658 * accessing the kiocb cookie.
2660 static void io_iopoll_req_issued(struct io_kiocb *req)
2662 struct io_ring_ctx *ctx = req->ctx;
2665 * Track whether we have multiple files in our lists. This will impact
2666 * how we do polling eventually, not spinning if we're on potentially
2667 * different devices.
2669 if (list_empty(&ctx->iopoll_list)) {
2670 ctx->poll_multi_file = false;
2671 } else if (!ctx->poll_multi_file) {
2672 struct io_kiocb *list_req;
2674 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2676 if (list_req->file != req->file)
2677 ctx->poll_multi_file = true;
2681 * For fast devices, IO may have already completed. If it has, add
2682 * it to the front so we find it first.
2684 if (READ_ONCE(req->iopoll_completed))
2685 list_add(&req->iopoll_entry, &ctx->iopoll_list);
2687 list_add_tail(&req->iopoll_entry, &ctx->iopoll_list);
2689 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2690 wq_has_sleeper(&ctx->sq_data->wait))
2691 wake_up(&ctx->sq_data->wait);
2694 static void __io_state_file_put(struct io_submit_state *state)
2696 if (state->has_refs)
2697 fput_many(state->file, state->has_refs);
2701 static inline void io_state_file_put(struct io_submit_state *state)
2704 __io_state_file_put(state);
2708 * Get as many references to a file as we have IOs left in this submission,
2709 * assuming most submissions are for one file, or at least that each file
2710 * has more than one submission.
2712 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2718 if (state->fd == fd) {
2722 __io_state_file_put(state);
2724 state->file = fget_many(fd, state->ios_left);
2729 state->has_refs = state->ios_left - 1;
2733 static bool io_bdev_nowait(struct block_device *bdev)
2736 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2743 * If we tracked the file through the SCM inflight mechanism, we could support
2744 * any file. For now, just ensure that anything potentially problematic is done
2747 static bool io_file_supports_async(struct file *file, int rw)
2749 umode_t mode = file_inode(file)->i_mode;
2751 if (S_ISBLK(mode)) {
2752 if (io_bdev_nowait(file->f_inode->i_bdev))
2758 if (S_ISREG(mode)) {
2759 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2760 file->f_op != &io_uring_fops)
2765 /* any ->read/write should understand O_NONBLOCK */
2766 if (file->f_flags & O_NONBLOCK)
2769 if (!(file->f_mode & FMODE_NOWAIT))
2773 return file->f_op->read_iter != NULL;
2775 return file->f_op->write_iter != NULL;
2778 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2780 struct io_ring_ctx *ctx = req->ctx;
2781 struct kiocb *kiocb = &req->rw.kiocb;
2785 if (S_ISREG(file_inode(req->file)->i_mode))
2786 req->flags |= REQ_F_ISREG;
2788 kiocb->ki_pos = READ_ONCE(sqe->off);
2789 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2790 req->flags |= REQ_F_CUR_POS;
2791 kiocb->ki_pos = req->file->f_pos;
2793 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2794 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2795 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2799 ioprio = READ_ONCE(sqe->ioprio);
2801 ret = ioprio_check_cap(ioprio);
2805 kiocb->ki_ioprio = ioprio;
2807 kiocb->ki_ioprio = get_current_ioprio();
2809 /* don't allow async punt if RWF_NOWAIT was requested */
2810 if (kiocb->ki_flags & IOCB_NOWAIT)
2811 req->flags |= REQ_F_NOWAIT;
2813 if (ctx->flags & IORING_SETUP_IOPOLL) {
2814 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2815 !kiocb->ki_filp->f_op->iopoll)
2818 kiocb->ki_flags |= IOCB_HIPRI;
2819 kiocb->ki_complete = io_complete_rw_iopoll;
2820 req->iopoll_completed = 0;
2822 if (kiocb->ki_flags & IOCB_HIPRI)
2824 kiocb->ki_complete = io_complete_rw;
2827 req->rw.addr = READ_ONCE(sqe->addr);
2828 req->rw.len = READ_ONCE(sqe->len);
2829 req->buf_index = READ_ONCE(sqe->buf_index);
2833 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2839 case -ERESTARTNOINTR:
2840 case -ERESTARTNOHAND:
2841 case -ERESTART_RESTARTBLOCK:
2843 * We can't just restart the syscall, since previously
2844 * submitted sqes may already be in progress. Just fail this
2850 kiocb->ki_complete(kiocb, ret, 0);
2854 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2855 struct io_comp_state *cs)
2857 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2858 struct io_async_rw *io = req->async_data;
2860 /* add previously done IO, if any */
2861 if (io && io->bytes_done > 0) {
2863 ret = io->bytes_done;
2865 ret += io->bytes_done;
2868 if (req->flags & REQ_F_CUR_POS)
2869 req->file->f_pos = kiocb->ki_pos;
2870 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2871 __io_complete_rw(req, ret, 0, cs);
2873 io_rw_done(kiocb, ret);
2876 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2877 struct iov_iter *iter)
2879 struct io_ring_ctx *ctx = req->ctx;
2880 size_t len = req->rw.len;
2881 struct io_mapped_ubuf *imu;
2882 u16 index, buf_index = req->buf_index;
2886 if (unlikely(buf_index >= ctx->nr_user_bufs))
2888 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2889 imu = &ctx->user_bufs[index];
2890 buf_addr = req->rw.addr;
2893 if (buf_addr + len < buf_addr)
2895 /* not inside the mapped region */
2896 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2900 * May not be a start of buffer, set size appropriately
2901 * and advance us to the beginning.
2903 offset = buf_addr - imu->ubuf;
2904 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2908 * Don't use iov_iter_advance() here, as it's really slow for
2909 * using the latter parts of a big fixed buffer - it iterates
2910 * over each segment manually. We can cheat a bit here, because
2913 * 1) it's a BVEC iter, we set it up
2914 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2915 * first and last bvec
2917 * So just find our index, and adjust the iterator afterwards.
2918 * If the offset is within the first bvec (or the whole first
2919 * bvec, just use iov_iter_advance(). This makes it easier
2920 * since we can just skip the first segment, which may not
2921 * be PAGE_SIZE aligned.
2923 const struct bio_vec *bvec = imu->bvec;
2925 if (offset <= bvec->bv_len) {
2926 iov_iter_advance(iter, offset);
2928 unsigned long seg_skip;
2930 /* skip first vec */
2931 offset -= bvec->bv_len;
2932 seg_skip = 1 + (offset >> PAGE_SHIFT);
2934 iter->bvec = bvec + seg_skip;
2935 iter->nr_segs -= seg_skip;
2936 iter->count -= bvec->bv_len + offset;
2937 iter->iov_offset = offset & ~PAGE_MASK;
2944 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2947 mutex_unlock(&ctx->uring_lock);
2950 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2953 * "Normal" inline submissions always hold the uring_lock, since we
2954 * grab it from the system call. Same is true for the SQPOLL offload.
2955 * The only exception is when we've detached the request and issue it
2956 * from an async worker thread, grab the lock for that case.
2959 mutex_lock(&ctx->uring_lock);
2962 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2963 int bgid, struct io_buffer *kbuf,
2966 struct io_buffer *head;
2968 if (req->flags & REQ_F_BUFFER_SELECTED)
2971 io_ring_submit_lock(req->ctx, needs_lock);
2973 lockdep_assert_held(&req->ctx->uring_lock);
2975 head = xa_load(&req->ctx->io_buffers, bgid);
2977 if (!list_empty(&head->list)) {
2978 kbuf = list_last_entry(&head->list, struct io_buffer,
2980 list_del(&kbuf->list);
2983 xa_erase(&req->ctx->io_buffers, bgid);
2985 if (*len > kbuf->len)
2988 kbuf = ERR_PTR(-ENOBUFS);
2991 io_ring_submit_unlock(req->ctx, needs_lock);
2996 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2999 struct io_buffer *kbuf;
3002 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3003 bgid = req->buf_index;
3004 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3007 req->rw.addr = (u64) (unsigned long) kbuf;
3008 req->flags |= REQ_F_BUFFER_SELECTED;
3009 return u64_to_user_ptr(kbuf->addr);
3012 #ifdef CONFIG_COMPAT
3013 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3016 struct compat_iovec __user *uiov;
3017 compat_ssize_t clen;
3021 uiov = u64_to_user_ptr(req->rw.addr);
3022 if (!access_ok(uiov, sizeof(*uiov)))
3024 if (__get_user(clen, &uiov->iov_len))
3030 buf = io_rw_buffer_select(req, &len, needs_lock);
3032 return PTR_ERR(buf);
3033 iov[0].iov_base = buf;
3034 iov[0].iov_len = (compat_size_t) len;
3039 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3042 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3046 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3049 len = iov[0].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 = len;
3060 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3063 if (req->flags & REQ_F_BUFFER_SELECTED) {
3064 struct io_buffer *kbuf;
3066 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3067 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3068 iov[0].iov_len = kbuf->len;
3071 if (req->rw.len != 1)
3074 #ifdef CONFIG_COMPAT
3075 if (req->ctx->compat)
3076 return io_compat_import(req, iov, needs_lock);
3079 return __io_iov_buffer_select(req, iov, needs_lock);
3082 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3083 struct iovec **iovec, struct iov_iter *iter,
3086 void __user *buf = u64_to_user_ptr(req->rw.addr);
3087 size_t sqe_len = req->rw.len;
3091 opcode = req->opcode;
3092 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3094 return io_import_fixed(req, rw, iter);
3097 /* buffer index only valid with fixed read/write, or buffer select */
3098 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3101 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3102 if (req->flags & REQ_F_BUFFER_SELECT) {
3103 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3105 return PTR_ERR(buf);
3106 req->rw.len = sqe_len;
3109 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3114 if (req->flags & REQ_F_BUFFER_SELECT) {
3115 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3117 ret = (*iovec)->iov_len;
3118 iov_iter_init(iter, rw, *iovec, 1, ret);
3124 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3128 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3129 struct iovec **iovec, struct iov_iter *iter,
3132 struct io_async_rw *iorw = req->async_data;
3135 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3140 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3142 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3146 * For files that don't have ->read_iter() and ->write_iter(), handle them
3147 * by looping over ->read() or ->write() manually.
3149 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3151 struct kiocb *kiocb = &req->rw.kiocb;
3152 struct file *file = req->file;
3156 * Don't support polled IO through this interface, and we can't
3157 * support non-blocking either. For the latter, this just causes
3158 * the kiocb to be handled from an async context.
3160 if (kiocb->ki_flags & IOCB_HIPRI)
3162 if (kiocb->ki_flags & IOCB_NOWAIT)
3165 while (iov_iter_count(iter)) {
3169 if (!iov_iter_is_bvec(iter)) {
3170 iovec = iov_iter_iovec(iter);
3172 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3173 iovec.iov_len = req->rw.len;
3177 nr = file->f_op->read(file, iovec.iov_base,
3178 iovec.iov_len, io_kiocb_ppos(kiocb));
3180 nr = file->f_op->write(file, iovec.iov_base,
3181 iovec.iov_len, io_kiocb_ppos(kiocb));
3190 if (!iov_iter_is_bvec(iter)) {
3191 iov_iter_advance(iter, nr);
3198 if (nr != iovec.iov_len)
3205 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3206 const struct iovec *fast_iov, struct iov_iter *iter)
3208 struct io_async_rw *rw = req->async_data;
3210 memcpy(&rw->iter, iter, sizeof(*iter));
3211 rw->free_iovec = iovec;
3213 /* can only be fixed buffers, no need to do anything */
3214 if (iov_iter_is_bvec(iter))
3217 unsigned iov_off = 0;
3219 rw->iter.iov = rw->fast_iov;
3220 if (iter->iov != fast_iov) {
3221 iov_off = iter->iov - fast_iov;
3222 rw->iter.iov += iov_off;
3224 if (rw->fast_iov != fast_iov)
3225 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3226 sizeof(struct iovec) * iter->nr_segs);
3228 req->flags |= REQ_F_NEED_CLEANUP;
3232 static inline int __io_alloc_async_data(struct io_kiocb *req)
3234 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3235 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3236 return req->async_data == NULL;
3239 static int io_alloc_async_data(struct io_kiocb *req)
3241 if (!io_op_defs[req->opcode].needs_async_data)
3244 return __io_alloc_async_data(req);
3247 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3248 const struct iovec *fast_iov,
3249 struct iov_iter *iter, bool force)
3251 if (!force && !io_op_defs[req->opcode].needs_async_data)
3253 if (!req->async_data) {
3254 if (__io_alloc_async_data(req))
3257 io_req_map_rw(req, iovec, fast_iov, iter);
3262 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3264 struct io_async_rw *iorw = req->async_data;
3265 struct iovec *iov = iorw->fast_iov;
3268 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3269 if (unlikely(ret < 0))
3272 iorw->bytes_done = 0;
3273 iorw->free_iovec = iov;
3275 req->flags |= REQ_F_NEED_CLEANUP;
3279 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3283 ret = io_prep_rw(req, sqe);
3287 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3290 /* either don't need iovec imported or already have it */
3291 if (!req->async_data)
3293 return io_rw_prep_async(req, READ);
3297 * This is our waitqueue callback handler, registered through lock_page_async()
3298 * when we initially tried to do the IO with the iocb armed our waitqueue.
3299 * This gets called when the page is unlocked, and we generally expect that to
3300 * happen when the page IO is completed and the page is now uptodate. This will
3301 * queue a task_work based retry of the operation, attempting to copy the data
3302 * again. If the latter fails because the page was NOT uptodate, then we will
3303 * do a thread based blocking retry of the operation. That's the unexpected
3306 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3307 int sync, void *arg)
3309 struct wait_page_queue *wpq;
3310 struct io_kiocb *req = wait->private;
3311 struct wait_page_key *key = arg;
3314 wpq = container_of(wait, struct wait_page_queue, wait);
3316 if (!wake_page_match(wpq, key))
3319 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3320 list_del_init(&wait->entry);
3322 init_task_work(&req->task_work, io_req_task_submit);
3323 percpu_ref_get(&req->ctx->refs);
3325 /* submit ref gets dropped, acquire a new one */
3326 refcount_inc(&req->refs);
3327 ret = io_req_task_work_add(req, true);
3328 if (unlikely(ret)) {
3329 struct task_struct *tsk;
3331 /* queue just for cancelation */
3332 init_task_work(&req->task_work, io_req_task_cancel);
3333 tsk = io_wq_get_task(req->ctx->io_wq);
3334 task_work_add(tsk, &req->task_work, TWA_NONE);
3335 wake_up_process(tsk);
3341 * This controls whether a given IO request should be armed for async page
3342 * based retry. If we return false here, the request is handed to the async
3343 * worker threads for retry. If we're doing buffered reads on a regular file,
3344 * we prepare a private wait_page_queue entry and retry the operation. This
3345 * will either succeed because the page is now uptodate and unlocked, or it
3346 * will register a callback when the page is unlocked at IO completion. Through
3347 * that callback, io_uring uses task_work to setup a retry of the operation.
3348 * That retry will attempt the buffered read again. The retry will generally
3349 * succeed, or in rare cases where it fails, we then fall back to using the
3350 * async worker threads for a blocking retry.
3352 static bool io_rw_should_retry(struct io_kiocb *req)
3354 struct io_async_rw *rw = req->async_data;
3355 struct wait_page_queue *wait = &rw->wpq;
3356 struct kiocb *kiocb = &req->rw.kiocb;
3358 /* never retry for NOWAIT, we just complete with -EAGAIN */
3359 if (req->flags & REQ_F_NOWAIT)
3362 /* Only for buffered IO */
3363 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3367 * just use poll if we can, and don't attempt if the fs doesn't
3368 * support callback based unlocks
3370 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3373 wait->wait.func = io_async_buf_func;
3374 wait->wait.private = req;
3375 wait->wait.flags = 0;
3376 INIT_LIST_HEAD(&wait->wait.entry);
3377 kiocb->ki_flags |= IOCB_WAITQ;
3378 kiocb->ki_flags &= ~IOCB_NOWAIT;
3379 kiocb->ki_waitq = wait;
3383 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3385 if (req->file->f_op->read_iter)
3386 return call_read_iter(req->file, &req->rw.kiocb, iter);
3387 else if (req->file->f_op->read)
3388 return loop_rw_iter(READ, req, iter);
3393 static int io_read(struct io_kiocb *req, bool force_nonblock,
3394 struct io_comp_state *cs)
3396 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3397 struct kiocb *kiocb = &req->rw.kiocb;
3398 struct iov_iter __iter, *iter = &__iter;
3399 struct iov_iter iter_cp;
3400 struct io_async_rw *rw = req->async_data;
3401 ssize_t io_size, ret, ret2;
3407 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3411 io_size = iov_iter_count(iter);
3412 req->result = io_size;
3415 /* Ensure we clear previously set non-block flag */
3416 if (!force_nonblock)
3417 kiocb->ki_flags &= ~IOCB_NOWAIT;
3419 kiocb->ki_flags |= IOCB_NOWAIT;
3422 /* If the file doesn't support async, just async punt */
3423 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3427 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3431 ret = io_iter_do_read(req, iter);
3435 } else if (ret == -EIOCBQUEUED) {
3438 } else if (ret == -EAGAIN) {
3439 /* IOPOLL retry should happen for io-wq threads */
3440 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3442 /* no retry on NONBLOCK marked file */
3443 if (req->file->f_flags & O_NONBLOCK)
3445 /* some cases will consume bytes even on error returns */
3449 } else if (ret < 0) {
3450 /* make sure -ERESTARTSYS -> -EINTR is done */
3454 /* read it all, or we did blocking attempt. no retry. */
3455 if (!iov_iter_count(iter) || !force_nonblock ||
3456 (req->file->f_flags & O_NONBLOCK) || !(req->flags & REQ_F_ISREG))
3461 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3468 rw = req->async_data;
3469 /* it's copied and will be cleaned with ->io */
3471 /* now use our persistent iterator, if we aren't already */
3474 rw->bytes_done += ret;
3475 /* if we can retry, do so with the callbacks armed */
3476 if (!io_rw_should_retry(req)) {
3477 kiocb->ki_flags &= ~IOCB_WAITQ;
3482 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3483 * get -EIOCBQUEUED, then we'll get a notification when the desired
3484 * page gets unlocked. We can also get a partial read here, and if we
3485 * do, then just retry at the new offset.
3487 ret = io_iter_do_read(req, iter);
3488 if (ret == -EIOCBQUEUED) {
3491 } else if (ret > 0 && ret < io_size) {
3492 /* we got some bytes, but not all. retry. */
3493 kiocb->ki_flags &= ~IOCB_WAITQ;
3497 kiocb_done(kiocb, ret, cs);
3500 /* it's reportedly faster than delegating the null check to kfree() */
3506 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3510 ret = io_prep_rw(req, sqe);
3514 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3517 /* either don't need iovec imported or already have it */
3518 if (!req->async_data)
3520 return io_rw_prep_async(req, WRITE);
3523 static int io_write(struct io_kiocb *req, bool force_nonblock,
3524 struct io_comp_state *cs)
3526 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3527 struct kiocb *kiocb = &req->rw.kiocb;
3528 struct iov_iter __iter, *iter = &__iter;
3529 struct iov_iter iter_cp;
3530 struct io_async_rw *rw = req->async_data;
3531 ssize_t ret, ret2, io_size;
3536 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3540 io_size = iov_iter_count(iter);
3541 req->result = io_size;
3543 /* Ensure we clear previously set non-block flag */
3544 if (!force_nonblock)
3545 kiocb->ki_flags &= ~IOCB_NOWAIT;
3547 kiocb->ki_flags |= IOCB_NOWAIT;
3549 /* If the file doesn't support async, just async punt */
3550 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3553 /* file path doesn't support NOWAIT for non-direct_IO */
3554 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3555 (req->flags & REQ_F_ISREG))
3558 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3563 * Open-code file_start_write here to grab freeze protection,
3564 * which will be released by another thread in
3565 * io_complete_rw(). Fool lockdep by telling it the lock got
3566 * released so that it doesn't complain about the held lock when
3567 * we return to userspace.
3569 if (req->flags & REQ_F_ISREG) {
3570 sb_start_write(file_inode(req->file)->i_sb);
3571 __sb_writers_release(file_inode(req->file)->i_sb,
3574 kiocb->ki_flags |= IOCB_WRITE;
3576 if (req->file->f_op->write_iter)
3577 ret2 = call_write_iter(req->file, kiocb, iter);
3578 else if (req->file->f_op->write)
3579 ret2 = loop_rw_iter(WRITE, req, iter);
3584 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3585 * retry them without IOCB_NOWAIT.
3587 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3589 /* no retry on NONBLOCK marked file */
3590 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3592 if (!force_nonblock || ret2 != -EAGAIN) {
3593 /* IOPOLL retry should happen for io-wq threads */
3594 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3597 kiocb_done(kiocb, ret2, cs);
3600 /* some cases will consume bytes even on error returns */
3602 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3607 /* it's reportedly faster than delegating the null check to kfree() */
3613 static int __io_splice_prep(struct io_kiocb *req,
3614 const struct io_uring_sqe *sqe)
3616 struct io_splice* sp = &req->splice;
3617 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3619 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3623 sp->len = READ_ONCE(sqe->len);
3624 sp->flags = READ_ONCE(sqe->splice_flags);
3626 if (unlikely(sp->flags & ~valid_flags))
3629 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3630 (sp->flags & SPLICE_F_FD_IN_FIXED));
3633 req->flags |= REQ_F_NEED_CLEANUP;
3635 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3637 * Splice operation will be punted aync, and here need to
3638 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3640 io_req_init_async(req);
3641 req->work.flags |= IO_WQ_WORK_UNBOUND;
3647 static int io_tee_prep(struct io_kiocb *req,
3648 const struct io_uring_sqe *sqe)
3650 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3652 return __io_splice_prep(req, sqe);
3655 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3657 struct io_splice *sp = &req->splice;
3658 struct file *in = sp->file_in;
3659 struct file *out = sp->file_out;
3660 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3666 ret = do_tee(in, out, sp->len, flags);
3668 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3669 req->flags &= ~REQ_F_NEED_CLEANUP;
3672 req_set_fail_links(req);
3673 io_req_complete(req, ret);
3677 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3679 struct io_splice* sp = &req->splice;
3681 sp->off_in = READ_ONCE(sqe->splice_off_in);
3682 sp->off_out = READ_ONCE(sqe->off);
3683 return __io_splice_prep(req, sqe);
3686 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3688 struct io_splice *sp = &req->splice;
3689 struct file *in = sp->file_in;
3690 struct file *out = sp->file_out;
3691 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3692 loff_t *poff_in, *poff_out;
3698 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3699 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3702 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3704 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3705 req->flags &= ~REQ_F_NEED_CLEANUP;
3708 req_set_fail_links(req);
3709 io_req_complete(req, ret);
3714 * IORING_OP_NOP just posts a completion event, nothing else.
3716 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3718 struct io_ring_ctx *ctx = req->ctx;
3720 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3723 __io_req_complete(req, 0, 0, cs);
3727 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3729 struct io_ring_ctx *ctx = req->ctx;
3734 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3736 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
3740 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3741 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3744 req->sync.off = READ_ONCE(sqe->off);
3745 req->sync.len = READ_ONCE(sqe->len);
3749 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3751 loff_t end = req->sync.off + req->sync.len;
3754 /* fsync always requires a blocking context */
3758 ret = vfs_fsync_range(req->file, req->sync.off,
3759 end > 0 ? end : LLONG_MAX,
3760 req->sync.flags & IORING_FSYNC_DATASYNC);
3762 req_set_fail_links(req);
3763 io_req_complete(req, ret);
3767 static int io_fallocate_prep(struct io_kiocb *req,
3768 const struct io_uring_sqe *sqe)
3770 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
3773 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3776 req->sync.off = READ_ONCE(sqe->off);
3777 req->sync.len = READ_ONCE(sqe->addr);
3778 req->sync.mode = READ_ONCE(sqe->len);
3782 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3786 /* fallocate always requiring blocking context */
3789 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3792 req_set_fail_links(req);
3793 io_req_complete(req, ret);
3797 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3799 const char __user *fname;
3802 if (unlikely(sqe->ioprio || sqe->buf_index || sqe->splice_fd_in))
3804 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3807 /* open.how should be already initialised */
3808 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3809 req->open.how.flags |= O_LARGEFILE;
3811 req->open.dfd = READ_ONCE(sqe->fd);
3812 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3813 req->open.filename = getname(fname);
3814 if (IS_ERR(req->open.filename)) {
3815 ret = PTR_ERR(req->open.filename);
3816 req->open.filename = NULL;
3819 req->open.nofile = rlimit(RLIMIT_NOFILE);
3820 req->open.ignore_nonblock = false;
3821 req->flags |= REQ_F_NEED_CLEANUP;
3825 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3829 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3831 mode = READ_ONCE(sqe->len);
3832 flags = READ_ONCE(sqe->open_flags);
3833 req->open.how = build_open_how(flags, mode);
3834 return __io_openat_prep(req, sqe);
3837 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3839 struct open_how __user *how;
3843 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3845 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3846 len = READ_ONCE(sqe->len);
3847 if (len < OPEN_HOW_SIZE_VER0)
3850 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3855 return __io_openat_prep(req, sqe);
3858 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3860 struct open_flags op;
3864 if (force_nonblock && !req->open.ignore_nonblock)
3867 ret = build_open_flags(&req->open.how, &op);
3871 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3875 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3878 ret = PTR_ERR(file);
3880 * A work-around to ensure that /proc/self works that way
3881 * that it should - if we get -EOPNOTSUPP back, then assume
3882 * that proc_self_get_link() failed us because we're in async
3883 * context. We should be safe to retry this from the task
3884 * itself with force_nonblock == false set, as it should not
3885 * block on lookup. Would be nice to know this upfront and
3886 * avoid the async dance, but doesn't seem feasible.
3888 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
3889 req->open.ignore_nonblock = true;
3890 refcount_inc(&req->refs);
3891 io_req_task_queue(req);
3895 fsnotify_open(file);
3896 fd_install(ret, file);
3899 putname(req->open.filename);
3900 req->flags &= ~REQ_F_NEED_CLEANUP;
3902 req_set_fail_links(req);
3903 io_req_complete(req, ret);
3907 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3909 return io_openat2(req, force_nonblock);
3912 static int io_remove_buffers_prep(struct io_kiocb *req,
3913 const struct io_uring_sqe *sqe)
3915 struct io_provide_buf *p = &req->pbuf;
3918 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
3922 tmp = READ_ONCE(sqe->fd);
3923 if (!tmp || tmp > USHRT_MAX)
3926 memset(p, 0, sizeof(*p));
3928 p->bgid = READ_ONCE(sqe->buf_group);
3932 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3933 int bgid, unsigned nbufs)
3937 /* shouldn't happen */
3941 /* the head kbuf is the list itself */
3942 while (!list_empty(&buf->list)) {
3943 struct io_buffer *nxt;
3945 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3946 list_del(&nxt->list);
3953 xa_erase(&ctx->io_buffers, bgid);
3958 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3959 struct io_comp_state *cs)
3961 struct io_provide_buf *p = &req->pbuf;
3962 struct io_ring_ctx *ctx = req->ctx;
3963 struct io_buffer *head;
3966 io_ring_submit_lock(ctx, !force_nonblock);
3968 lockdep_assert_held(&ctx->uring_lock);
3971 head = xa_load(&ctx->io_buffers, p->bgid);
3973 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3975 req_set_fail_links(req);
3977 /* need to hold the lock to complete IOPOLL requests */
3978 if (ctx->flags & IORING_SETUP_IOPOLL) {
3979 __io_req_complete(req, ret, 0, cs);
3980 io_ring_submit_unlock(ctx, !force_nonblock);
3982 io_ring_submit_unlock(ctx, !force_nonblock);
3983 __io_req_complete(req, ret, 0, cs);
3988 static int io_provide_buffers_prep(struct io_kiocb *req,
3989 const struct io_uring_sqe *sqe)
3991 unsigned long size, tmp_check;
3992 struct io_provide_buf *p = &req->pbuf;
3995 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
3998 tmp = READ_ONCE(sqe->fd);
3999 if (!tmp || tmp > USHRT_MAX)
4002 p->addr = READ_ONCE(sqe->addr);
4003 p->len = READ_ONCE(sqe->len);
4005 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4008 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4011 size = (unsigned long)p->len * p->nbufs;
4012 if (!access_ok(u64_to_user_ptr(p->addr), size))
4015 p->bgid = READ_ONCE(sqe->buf_group);
4016 tmp = READ_ONCE(sqe->off);
4017 if (tmp > USHRT_MAX)
4023 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4025 struct io_buffer *buf;
4026 u64 addr = pbuf->addr;
4027 int i, bid = pbuf->bid;
4029 for (i = 0; i < pbuf->nbufs; i++) {
4030 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4035 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4040 INIT_LIST_HEAD(&buf->list);
4043 list_add_tail(&buf->list, &(*head)->list);
4048 return i ? i : -ENOMEM;
4051 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4052 struct io_comp_state *cs)
4054 struct io_provide_buf *p = &req->pbuf;
4055 struct io_ring_ctx *ctx = req->ctx;
4056 struct io_buffer *head, *list;
4059 io_ring_submit_lock(ctx, !force_nonblock);
4061 lockdep_assert_held(&ctx->uring_lock);
4063 list = head = xa_load(&ctx->io_buffers, p->bgid);
4065 ret = io_add_buffers(p, &head);
4066 if (ret >= 0 && !list) {
4067 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4069 __io_remove_buffers(ctx, head, p->bgid, -1U);
4072 req_set_fail_links(req);
4074 /* need to hold the lock to complete IOPOLL requests */
4075 if (ctx->flags & IORING_SETUP_IOPOLL) {
4076 __io_req_complete(req, ret, 0, cs);
4077 io_ring_submit_unlock(ctx, !force_nonblock);
4079 io_ring_submit_unlock(ctx, !force_nonblock);
4080 __io_req_complete(req, ret, 0, cs);
4085 static int io_epoll_ctl_prep(struct io_kiocb *req,
4086 const struct io_uring_sqe *sqe)
4088 #if defined(CONFIG_EPOLL)
4089 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4091 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4094 req->epoll.epfd = READ_ONCE(sqe->fd);
4095 req->epoll.op = READ_ONCE(sqe->len);
4096 req->epoll.fd = READ_ONCE(sqe->off);
4098 if (ep_op_has_event(req->epoll.op)) {
4099 struct epoll_event __user *ev;
4101 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4102 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4112 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4113 struct io_comp_state *cs)
4115 #if defined(CONFIG_EPOLL)
4116 struct io_epoll *ie = &req->epoll;
4119 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4120 if (force_nonblock && ret == -EAGAIN)
4124 req_set_fail_links(req);
4125 __io_req_complete(req, ret, 0, cs);
4132 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4134 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4135 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4137 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4140 req->madvise.addr = READ_ONCE(sqe->addr);
4141 req->madvise.len = READ_ONCE(sqe->len);
4142 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4149 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4151 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4152 struct io_madvise *ma = &req->madvise;
4158 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4160 req_set_fail_links(req);
4161 io_req_complete(req, ret);
4168 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4170 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4172 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4175 req->fadvise.offset = READ_ONCE(sqe->off);
4176 req->fadvise.len = READ_ONCE(sqe->len);
4177 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4181 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4183 struct io_fadvise *fa = &req->fadvise;
4186 if (force_nonblock) {
4187 switch (fa->advice) {
4188 case POSIX_FADV_NORMAL:
4189 case POSIX_FADV_RANDOM:
4190 case POSIX_FADV_SEQUENTIAL:
4197 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4199 req_set_fail_links(req);
4200 io_req_complete(req, ret);
4204 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4206 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4208 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4210 if (req->flags & REQ_F_FIXED_FILE)
4213 req->statx.dfd = READ_ONCE(sqe->fd);
4214 req->statx.mask = READ_ONCE(sqe->len);
4215 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4216 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4217 req->statx.flags = READ_ONCE(sqe->statx_flags);
4222 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4224 struct io_statx *ctx = &req->statx;
4230 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4234 req_set_fail_links(req);
4235 io_req_complete(req, ret);
4239 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4242 * If we queue this for async, it must not be cancellable. That would
4243 * leave the 'file' in an undeterminate state, and here need to modify
4244 * io_wq_work.flags, so initialize io_wq_work firstly.
4246 io_req_init_async(req);
4248 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4250 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4251 sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
4253 if (req->flags & REQ_F_FIXED_FILE)
4256 req->close.fd = READ_ONCE(sqe->fd);
4257 if ((req->file && req->file->f_op == &io_uring_fops))
4260 req->close.put_file = NULL;
4264 static int io_close(struct io_kiocb *req, bool force_nonblock,
4265 struct io_comp_state *cs)
4267 struct io_close *close = &req->close;
4270 /* might be already done during nonblock submission */
4271 if (!close->put_file) {
4272 ret = __close_fd_get_file(close->fd, &close->put_file);
4274 return (ret == -ENOENT) ? -EBADF : ret;
4277 /* if the file has a flush method, be safe and punt to async */
4278 if (close->put_file->f_op->flush && force_nonblock) {
4279 /* not safe to cancel at this point */
4280 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4281 /* was never set, but play safe */
4282 req->flags &= ~REQ_F_NOWAIT;
4283 /* avoid grabbing files - we don't need the files */
4284 req->flags |= REQ_F_NO_FILE_TABLE;
4288 /* No ->flush() or already async, safely close from here */
4289 ret = filp_close(close->put_file, req->work.identity->files);
4291 req_set_fail_links(req);
4292 fput(close->put_file);
4293 close->put_file = NULL;
4294 __io_req_complete(req, ret, 0, cs);
4298 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4300 struct io_ring_ctx *ctx = req->ctx;
4305 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4307 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4311 req->sync.off = READ_ONCE(sqe->off);
4312 req->sync.len = READ_ONCE(sqe->len);
4313 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4317 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4321 /* sync_file_range always requires a blocking context */
4325 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4328 req_set_fail_links(req);
4329 io_req_complete(req, ret);
4333 #if defined(CONFIG_NET)
4334 static int io_setup_async_msg(struct io_kiocb *req,
4335 struct io_async_msghdr *kmsg)
4337 struct io_async_msghdr *async_msg = req->async_data;
4341 if (io_alloc_async_data(req)) {
4342 if (kmsg->iov != kmsg->fast_iov)
4346 async_msg = req->async_data;
4347 req->flags |= REQ_F_NEED_CLEANUP;
4348 memcpy(async_msg, kmsg, sizeof(*kmsg));
4352 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4353 struct io_async_msghdr *iomsg)
4355 iomsg->iov = iomsg->fast_iov;
4356 iomsg->msg.msg_name = &iomsg->addr;
4357 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4358 req->sr_msg.msg_flags, &iomsg->iov);
4361 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4363 struct io_async_msghdr *async_msg = req->async_data;
4364 struct io_sr_msg *sr = &req->sr_msg;
4367 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4369 if (unlikely(sqe->addr2 || sqe->splice_fd_in || sqe->ioprio))
4372 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4373 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4374 sr->len = READ_ONCE(sqe->len);
4376 #ifdef CONFIG_COMPAT
4377 if (req->ctx->compat)
4378 sr->msg_flags |= MSG_CMSG_COMPAT;
4381 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4383 ret = io_sendmsg_copy_hdr(req, async_msg);
4385 req->flags |= REQ_F_NEED_CLEANUP;
4389 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4390 struct io_comp_state *cs)
4392 struct io_async_msghdr iomsg, *kmsg;
4393 struct socket *sock;
4398 sock = sock_from_file(req->file, &ret);
4399 if (unlikely(!sock))
4402 if (req->async_data) {
4403 kmsg = req->async_data;
4404 kmsg->msg.msg_name = &kmsg->addr;
4405 /* if iov is set, it's allocated already */
4407 kmsg->iov = kmsg->fast_iov;
4408 kmsg->msg.msg_iter.iov = kmsg->iov;
4410 ret = io_sendmsg_copy_hdr(req, &iomsg);
4416 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4417 if (flags & MSG_DONTWAIT)
4418 req->flags |= REQ_F_NOWAIT;
4419 else if (force_nonblock)
4420 flags |= MSG_DONTWAIT;
4422 if (flags & MSG_WAITALL)
4423 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4425 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4426 if (force_nonblock && ret == -EAGAIN)
4427 return io_setup_async_msg(req, kmsg);
4428 if (ret == -ERESTARTSYS)
4431 if (kmsg->iov != kmsg->fast_iov)
4433 req->flags &= ~REQ_F_NEED_CLEANUP;
4435 req_set_fail_links(req);
4436 __io_req_complete(req, ret, 0, cs);
4440 static int io_send(struct io_kiocb *req, bool force_nonblock,
4441 struct io_comp_state *cs)
4443 struct io_sr_msg *sr = &req->sr_msg;
4446 struct socket *sock;
4451 sock = sock_from_file(req->file, &ret);
4452 if (unlikely(!sock))
4455 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4459 msg.msg_name = NULL;
4460 msg.msg_control = NULL;
4461 msg.msg_controllen = 0;
4462 msg.msg_namelen = 0;
4464 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4465 if (flags & MSG_DONTWAIT)
4466 req->flags |= REQ_F_NOWAIT;
4467 else if (force_nonblock)
4468 flags |= MSG_DONTWAIT;
4470 if (flags & MSG_WAITALL)
4471 min_ret = iov_iter_count(&msg.msg_iter);
4473 msg.msg_flags = flags;
4474 ret = sock_sendmsg(sock, &msg);
4475 if (force_nonblock && ret == -EAGAIN)
4477 if (ret == -ERESTARTSYS)
4481 req_set_fail_links(req);
4482 __io_req_complete(req, ret, 0, cs);
4486 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4487 struct io_async_msghdr *iomsg)
4489 struct io_sr_msg *sr = &req->sr_msg;
4490 struct iovec __user *uiov;
4494 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4495 &iomsg->uaddr, &uiov, &iov_len);
4499 if (req->flags & REQ_F_BUFFER_SELECT) {
4502 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4504 sr->len = iomsg->iov[0].iov_len;
4505 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4509 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4510 &iomsg->iov, &iomsg->msg.msg_iter,
4519 #ifdef CONFIG_COMPAT
4520 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4521 struct io_async_msghdr *iomsg)
4523 struct compat_msghdr __user *msg_compat;
4524 struct io_sr_msg *sr = &req->sr_msg;
4525 struct compat_iovec __user *uiov;
4530 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4531 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4536 uiov = compat_ptr(ptr);
4537 if (req->flags & REQ_F_BUFFER_SELECT) {
4538 compat_ssize_t clen;
4542 if (!access_ok(uiov, sizeof(*uiov)))
4544 if (__get_user(clen, &uiov->iov_len))
4549 iomsg->iov[0].iov_len = clen;
4552 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4553 UIO_FASTIOV, &iomsg->iov,
4554 &iomsg->msg.msg_iter, true);
4563 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4564 struct io_async_msghdr *iomsg)
4566 iomsg->msg.msg_name = &iomsg->addr;
4567 iomsg->iov = iomsg->fast_iov;
4569 #ifdef CONFIG_COMPAT
4570 if (req->ctx->compat)
4571 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4574 return __io_recvmsg_copy_hdr(req, iomsg);
4577 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4580 struct io_sr_msg *sr = &req->sr_msg;
4581 struct io_buffer *kbuf;
4583 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4588 req->flags |= REQ_F_BUFFER_SELECTED;
4592 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4594 return io_put_kbuf(req, req->sr_msg.kbuf);
4597 static int io_recvmsg_prep(struct io_kiocb *req,
4598 const struct io_uring_sqe *sqe)
4600 struct io_async_msghdr *async_msg = req->async_data;
4601 struct io_sr_msg *sr = &req->sr_msg;
4604 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4606 if (unlikely(sqe->addr2 || sqe->splice_fd_in || sqe->ioprio))
4609 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4610 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4611 sr->len = READ_ONCE(sqe->len);
4612 sr->bgid = READ_ONCE(sqe->buf_group);
4614 #ifdef CONFIG_COMPAT
4615 if (req->ctx->compat)
4616 sr->msg_flags |= MSG_CMSG_COMPAT;
4619 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4621 ret = io_recvmsg_copy_hdr(req, async_msg);
4623 req->flags |= REQ_F_NEED_CLEANUP;
4627 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4628 struct io_comp_state *cs)
4630 struct io_async_msghdr iomsg, *kmsg;
4631 struct socket *sock;
4632 struct io_buffer *kbuf;
4635 int ret, cflags = 0;
4637 sock = sock_from_file(req->file, &ret);
4638 if (unlikely(!sock))
4641 if (req->async_data) {
4642 kmsg = req->async_data;
4643 kmsg->msg.msg_name = &kmsg->addr;
4644 /* if iov is set, it's allocated already */
4646 kmsg->iov = kmsg->fast_iov;
4647 kmsg->msg.msg_iter.iov = kmsg->iov;
4649 ret = io_recvmsg_copy_hdr(req, &iomsg);
4655 if (req->flags & REQ_F_BUFFER_SELECT) {
4656 kbuf = io_recv_buffer_select(req, !force_nonblock);
4658 return PTR_ERR(kbuf);
4659 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4660 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4661 1, req->sr_msg.len);
4664 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4665 if (flags & MSG_DONTWAIT)
4666 req->flags |= REQ_F_NOWAIT;
4667 else if (force_nonblock)
4668 flags |= MSG_DONTWAIT;
4670 if (flags & MSG_WAITALL)
4671 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4673 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4674 kmsg->uaddr, flags);
4675 if (force_nonblock && ret == -EAGAIN)
4676 return io_setup_async_msg(req, kmsg);
4677 if (ret == -ERESTARTSYS)
4680 if (req->flags & REQ_F_BUFFER_SELECTED)
4681 cflags = io_put_recv_kbuf(req);
4682 if (kmsg->iov != kmsg->fast_iov)
4684 req->flags &= ~REQ_F_NEED_CLEANUP;
4685 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4686 req_set_fail_links(req);
4687 __io_req_complete(req, ret, cflags, cs);
4691 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4692 struct io_comp_state *cs)
4694 struct io_buffer *kbuf;
4695 struct io_sr_msg *sr = &req->sr_msg;
4697 void __user *buf = sr->buf;
4698 struct socket *sock;
4702 int ret, cflags = 0;
4704 sock = sock_from_file(req->file, &ret);
4705 if (unlikely(!sock))
4708 if (req->flags & REQ_F_BUFFER_SELECT) {
4709 kbuf = io_recv_buffer_select(req, !force_nonblock);
4711 return PTR_ERR(kbuf);
4712 buf = u64_to_user_ptr(kbuf->addr);
4715 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4719 msg.msg_name = NULL;
4720 msg.msg_control = NULL;
4721 msg.msg_controllen = 0;
4722 msg.msg_namelen = 0;
4723 msg.msg_iocb = NULL;
4726 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4727 if (flags & MSG_DONTWAIT)
4728 req->flags |= REQ_F_NOWAIT;
4729 else if (force_nonblock)
4730 flags |= MSG_DONTWAIT;
4732 if (flags & MSG_WAITALL)
4733 min_ret = iov_iter_count(&msg.msg_iter);
4735 ret = sock_recvmsg(sock, &msg, flags);
4736 if (force_nonblock && ret == -EAGAIN)
4738 if (ret == -ERESTARTSYS)
4741 if (req->flags & REQ_F_BUFFER_SELECTED)
4742 cflags = io_put_recv_kbuf(req);
4743 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4744 req_set_fail_links(req);
4745 __io_req_complete(req, ret, cflags, cs);
4749 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4751 struct io_accept *accept = &req->accept;
4753 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4755 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->splice_fd_in)
4758 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4759 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4760 accept->flags = READ_ONCE(sqe->accept_flags);
4761 accept->nofile = rlimit(RLIMIT_NOFILE);
4765 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4766 struct io_comp_state *cs)
4768 struct io_accept *accept = &req->accept;
4769 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4772 if (req->file->f_flags & O_NONBLOCK)
4773 req->flags |= REQ_F_NOWAIT;
4775 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4776 accept->addr_len, accept->flags,
4778 if (ret == -EAGAIN && force_nonblock)
4781 if (ret == -ERESTARTSYS)
4783 req_set_fail_links(req);
4785 __io_req_complete(req, ret, 0, cs);
4789 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4791 struct io_connect *conn = &req->connect;
4792 struct io_async_connect *io = req->async_data;
4794 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4796 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
4800 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4801 conn->addr_len = READ_ONCE(sqe->addr2);
4806 return move_addr_to_kernel(conn->addr, conn->addr_len,
4810 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4811 struct io_comp_state *cs)
4813 struct io_async_connect __io, *io;
4814 unsigned file_flags;
4817 if (req->async_data) {
4818 io = req->async_data;
4820 ret = move_addr_to_kernel(req->connect.addr,
4821 req->connect.addr_len,
4828 file_flags = force_nonblock ? O_NONBLOCK : 0;
4830 ret = __sys_connect_file(req->file, &io->address,
4831 req->connect.addr_len, file_flags);
4832 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4833 if (req->async_data)
4835 if (io_alloc_async_data(req)) {
4839 io = req->async_data;
4840 memcpy(req->async_data, &__io, sizeof(__io));
4843 if (ret == -ERESTARTSYS)
4847 req_set_fail_links(req);
4848 __io_req_complete(req, ret, 0, cs);
4851 #else /* !CONFIG_NET */
4852 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4857 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4858 struct io_comp_state *cs)
4863 static int io_send(struct io_kiocb *req, bool force_nonblock,
4864 struct io_comp_state *cs)
4869 static int io_recvmsg_prep(struct io_kiocb *req,
4870 const struct io_uring_sqe *sqe)
4875 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4876 struct io_comp_state *cs)
4881 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4882 struct io_comp_state *cs)
4887 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4892 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4893 struct io_comp_state *cs)
4898 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4903 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4904 struct io_comp_state *cs)
4908 #endif /* CONFIG_NET */
4910 struct io_poll_table {
4911 struct poll_table_struct pt;
4912 struct io_kiocb *req;
4917 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4918 __poll_t mask, task_work_func_t func)
4923 /* for instances that support it check for an event match first: */
4924 if (mask && !(mask & poll->events))
4927 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4929 list_del_init(&poll->wait.entry);
4932 init_task_work(&req->task_work, func);
4933 percpu_ref_get(&req->ctx->refs);
4936 * If we using the signalfd wait_queue_head for this wakeup, then
4937 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4938 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4939 * either, as the normal wakeup will suffice.
4941 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4944 * If this fails, then the task is exiting. When a task exits, the
4945 * work gets canceled, so just cancel this request as well instead
4946 * of executing it. We can't safely execute it anyway, as we may not
4947 * have the needed state needed for it anyway.
4949 ret = io_req_task_work_add(req, twa_signal_ok);
4950 if (unlikely(ret)) {
4951 struct task_struct *tsk;
4953 WRITE_ONCE(poll->canceled, true);
4954 tsk = io_wq_get_task(req->ctx->io_wq);
4955 task_work_add(tsk, &req->task_work, TWA_NONE);
4956 wake_up_process(tsk);
4961 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4962 __acquires(&req->ctx->completion_lock)
4964 struct io_ring_ctx *ctx = req->ctx;
4966 if (!req->result && !READ_ONCE(poll->canceled)) {
4967 struct poll_table_struct pt = { ._key = poll->events };
4969 req->result = vfs_poll(req->file, &pt) & poll->events;
4972 spin_lock_irq(&ctx->completion_lock);
4973 if (!req->result && !READ_ONCE(poll->canceled)) {
4974 add_wait_queue(poll->head, &poll->wait);
4981 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4983 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4984 if (req->opcode == IORING_OP_POLL_ADD)
4985 return req->async_data;
4986 return req->apoll->double_poll;
4989 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4991 if (req->opcode == IORING_OP_POLL_ADD)
4993 return &req->apoll->poll;
4996 static void io_poll_remove_double(struct io_kiocb *req)
4998 struct io_poll_iocb *poll = io_poll_get_double(req);
5000 lockdep_assert_held(&req->ctx->completion_lock);
5002 if (poll && poll->head) {
5003 struct wait_queue_head *head = poll->head;
5005 spin_lock(&head->lock);
5006 list_del_init(&poll->wait.entry);
5007 if (poll->wait.private)
5008 refcount_dec(&req->refs);
5010 spin_unlock(&head->lock);
5014 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5016 struct io_ring_ctx *ctx = req->ctx;
5018 io_poll_remove_double(req);
5019 req->poll.done = true;
5020 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5021 io_commit_cqring(ctx);
5024 static void io_poll_task_func(struct callback_head *cb)
5026 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5027 struct io_ring_ctx *ctx = req->ctx;
5028 struct io_kiocb *nxt;
5030 if (io_poll_rewait(req, &req->poll)) {
5031 spin_unlock_irq(&ctx->completion_lock);
5033 hash_del(&req->hash_node);
5034 io_poll_complete(req, req->result, 0);
5035 spin_unlock_irq(&ctx->completion_lock);
5037 nxt = io_put_req_find_next(req);
5038 io_cqring_ev_posted(ctx);
5040 __io_req_task_submit(nxt);
5043 percpu_ref_put(&ctx->refs);
5046 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5047 int sync, void *key)
5049 struct io_kiocb *req = wait->private;
5050 struct io_poll_iocb *poll = io_poll_get_single(req);
5051 __poll_t mask = key_to_poll(key);
5053 /* for instances that support it check for an event match first: */
5054 if (mask && !(mask & poll->events))
5057 list_del_init(&wait->entry);
5059 if (poll && poll->head) {
5062 spin_lock(&poll->head->lock);
5063 done = list_empty(&poll->wait.entry);
5065 list_del_init(&poll->wait.entry);
5066 /* make sure double remove sees this as being gone */
5067 wait->private = NULL;
5068 spin_unlock(&poll->head->lock);
5070 /* use wait func handler, so it matches the rq type */
5071 poll->wait.func(&poll->wait, mode, sync, key);
5074 refcount_dec(&req->refs);
5078 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5079 wait_queue_func_t wake_func)
5083 poll->canceled = false;
5084 poll->events = events;
5085 INIT_LIST_HEAD(&poll->wait.entry);
5086 init_waitqueue_func_entry(&poll->wait, wake_func);
5089 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5090 struct wait_queue_head *head,
5091 struct io_poll_iocb **poll_ptr)
5093 struct io_kiocb *req = pt->req;
5096 * The file being polled uses multiple waitqueues for poll handling
5097 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5100 if (unlikely(pt->nr_entries)) {
5101 struct io_poll_iocb *poll_one = poll;
5103 /* already have a 2nd entry, fail a third attempt */
5105 pt->error = -EINVAL;
5108 /* double add on the same waitqueue head, ignore */
5109 if (poll->head == head)
5111 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5113 pt->error = -ENOMEM;
5116 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5117 refcount_inc(&req->refs);
5118 poll->wait.private = req;
5125 if (poll->events & EPOLLEXCLUSIVE)
5126 add_wait_queue_exclusive(head, &poll->wait);
5128 add_wait_queue(head, &poll->wait);
5131 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5132 struct poll_table_struct *p)
5134 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5135 struct async_poll *apoll = pt->req->apoll;
5137 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5140 static void io_async_task_func(struct callback_head *cb)
5142 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5143 struct async_poll *apoll = req->apoll;
5144 struct io_ring_ctx *ctx = req->ctx;
5146 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5148 if (io_poll_rewait(req, &apoll->poll)) {
5149 spin_unlock_irq(&ctx->completion_lock);
5150 percpu_ref_put(&ctx->refs);
5154 /* If req is still hashed, it cannot have been canceled. Don't check. */
5155 if (hash_hashed(&req->hash_node))
5156 hash_del(&req->hash_node);
5158 io_poll_remove_double(req);
5159 spin_unlock_irq(&ctx->completion_lock);
5161 if (!READ_ONCE(apoll->poll.canceled))
5162 __io_req_task_submit(req);
5164 __io_req_task_cancel(req, -ECANCELED);
5166 percpu_ref_put(&ctx->refs);
5167 kfree(apoll->double_poll);
5171 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5174 struct io_kiocb *req = wait->private;
5175 struct io_poll_iocb *poll = &req->apoll->poll;
5177 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5180 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5183 static void io_poll_req_insert(struct io_kiocb *req)
5185 struct io_ring_ctx *ctx = req->ctx;
5186 struct hlist_head *list;
5188 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5189 hlist_add_head(&req->hash_node, list);
5192 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5193 struct io_poll_iocb *poll,
5194 struct io_poll_table *ipt, __poll_t mask,
5195 wait_queue_func_t wake_func)
5196 __acquires(&ctx->completion_lock)
5198 struct io_ring_ctx *ctx = req->ctx;
5199 bool cancel = false;
5201 if (req->file->f_op->may_pollfree) {
5202 spin_lock_irq(&ctx->completion_lock);
5206 INIT_HLIST_NODE(&req->hash_node);
5207 io_init_poll_iocb(poll, mask, wake_func);
5208 poll->file = req->file;
5209 poll->wait.private = req;
5211 ipt->pt._key = mask;
5214 ipt->nr_entries = 0;
5216 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5217 if (unlikely(!ipt->nr_entries) && !ipt->error)
5218 ipt->error = -EINVAL;
5220 spin_lock_irq(&ctx->completion_lock);
5222 io_poll_remove_double(req);
5223 if (likely(poll->head)) {
5224 spin_lock(&poll->head->lock);
5225 if (unlikely(list_empty(&poll->wait.entry))) {
5231 if (mask || ipt->error)
5232 list_del_init(&poll->wait.entry);
5234 WRITE_ONCE(poll->canceled, true);
5235 else if (!poll->done) /* actually waiting for an event */
5236 io_poll_req_insert(req);
5237 spin_unlock(&poll->head->lock);
5243 static bool io_arm_poll_handler(struct io_kiocb *req)
5245 const struct io_op_def *def = &io_op_defs[req->opcode];
5246 struct io_ring_ctx *ctx = req->ctx;
5247 struct async_poll *apoll;
5248 struct io_poll_table ipt;
5252 if (!req->file || !file_can_poll(req->file))
5254 if (req->flags & REQ_F_POLLED)
5258 else if (def->pollout)
5262 /* if we can't nonblock try, then no point in arming a poll handler */
5263 if (!io_file_supports_async(req->file, rw))
5266 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5267 if (unlikely(!apoll))
5269 apoll->double_poll = NULL;
5271 req->flags |= REQ_F_POLLED;
5276 mask |= POLLIN | POLLRDNORM;
5278 mask |= POLLOUT | POLLWRNORM;
5280 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5281 if ((req->opcode == IORING_OP_RECVMSG) &&
5282 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5285 mask |= POLLERR | POLLPRI;
5287 ipt.pt._qproc = io_async_queue_proc;
5289 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5291 if (ret || ipt.error) {
5292 io_poll_remove_double(req);
5293 spin_unlock_irq(&ctx->completion_lock);
5294 kfree(apoll->double_poll);
5298 spin_unlock_irq(&ctx->completion_lock);
5299 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5300 apoll->poll.events);
5304 static bool __io_poll_remove_one(struct io_kiocb *req,
5305 struct io_poll_iocb *poll)
5307 bool do_complete = false;
5309 spin_lock(&poll->head->lock);
5310 WRITE_ONCE(poll->canceled, true);
5311 if (!list_empty(&poll->wait.entry)) {
5312 list_del_init(&poll->wait.entry);
5315 spin_unlock(&poll->head->lock);
5316 hash_del(&req->hash_node);
5320 static bool io_poll_remove_one(struct io_kiocb *req)
5324 io_poll_remove_double(req);
5326 if (req->opcode == IORING_OP_POLL_ADD) {
5327 do_complete = __io_poll_remove_one(req, &req->poll);
5329 struct async_poll *apoll = req->apoll;
5331 /* non-poll requests have submit ref still */
5332 do_complete = __io_poll_remove_one(req, &apoll->poll);
5335 kfree(apoll->double_poll);
5341 io_cqring_fill_event(req, -ECANCELED);
5342 io_commit_cqring(req->ctx);
5343 req_set_fail_links(req);
5344 io_put_req_deferred(req, 1);
5351 * Returns true if we found and killed one or more poll requests
5353 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5354 struct files_struct *files)
5356 struct hlist_node *tmp;
5357 struct io_kiocb *req;
5360 spin_lock_irq(&ctx->completion_lock);
5361 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5362 struct hlist_head *list;
5364 list = &ctx->cancel_hash[i];
5365 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5366 if (io_match_task(req, tsk, files))
5367 posted += io_poll_remove_one(req);
5370 spin_unlock_irq(&ctx->completion_lock);
5373 io_cqring_ev_posted(ctx);
5378 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5380 struct hlist_head *list;
5381 struct io_kiocb *req;
5383 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5384 hlist_for_each_entry(req, list, hash_node) {
5385 if (sqe_addr != req->user_data)
5387 if (io_poll_remove_one(req))
5395 static int io_poll_remove_prep(struct io_kiocb *req,
5396 const struct io_uring_sqe *sqe)
5398 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5400 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5404 req->poll.addr = READ_ONCE(sqe->addr);
5409 * Find a running poll command that matches one specified in sqe->addr,
5410 * and remove it if found.
5412 static int io_poll_remove(struct io_kiocb *req)
5414 struct io_ring_ctx *ctx = req->ctx;
5418 addr = req->poll.addr;
5419 spin_lock_irq(&ctx->completion_lock);
5420 ret = io_poll_cancel(ctx, addr);
5421 spin_unlock_irq(&ctx->completion_lock);
5424 req_set_fail_links(req);
5425 io_req_complete(req, ret);
5429 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5432 struct io_kiocb *req = wait->private;
5433 struct io_poll_iocb *poll = &req->poll;
5435 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5438 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5439 struct poll_table_struct *p)
5441 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5443 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5446 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5448 struct io_poll_iocb *poll = &req->poll;
5451 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5453 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5456 events = READ_ONCE(sqe->poll32_events);
5458 events = swahw32(events);
5460 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5461 (events & EPOLLEXCLUSIVE);
5465 static int io_poll_add(struct io_kiocb *req)
5467 struct io_poll_iocb *poll = &req->poll;
5468 struct io_ring_ctx *ctx = req->ctx;
5469 struct io_poll_table ipt;
5472 ipt.pt._qproc = io_poll_queue_proc;
5474 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5477 if (mask) { /* no async, we'd stolen it */
5479 io_poll_complete(req, mask, 0);
5481 spin_unlock_irq(&ctx->completion_lock);
5484 io_cqring_ev_posted(ctx);
5490 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5492 struct io_timeout_data *data = container_of(timer,
5493 struct io_timeout_data, timer);
5494 struct io_kiocb *req = data->req;
5495 struct io_ring_ctx *ctx = req->ctx;
5496 unsigned long flags;
5498 spin_lock_irqsave(&ctx->completion_lock, flags);
5499 list_del_init(&req->timeout.list);
5500 atomic_set(&req->ctx->cq_timeouts,
5501 atomic_read(&req->ctx->cq_timeouts) + 1);
5503 io_cqring_fill_event(req, -ETIME);
5504 io_commit_cqring(ctx);
5505 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5507 io_cqring_ev_posted(ctx);
5508 req_set_fail_links(req);
5510 return HRTIMER_NORESTART;
5513 static int __io_timeout_cancel(struct io_kiocb *req)
5515 struct io_timeout_data *io = req->async_data;
5518 ret = hrtimer_try_to_cancel(&io->timer);
5521 list_del_init(&req->timeout.list);
5523 req_set_fail_links(req);
5524 io_cqring_fill_event(req, -ECANCELED);
5525 io_put_req_deferred(req, 1);
5529 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5531 struct io_kiocb *req;
5534 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5535 if (user_data == req->user_data) {
5544 return __io_timeout_cancel(req);
5547 static int io_timeout_remove_prep(struct io_kiocb *req,
5548 const struct io_uring_sqe *sqe)
5550 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5552 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5554 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags ||
5558 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5563 * Remove or update an existing timeout command
5565 static int io_timeout_remove(struct io_kiocb *req)
5567 struct io_ring_ctx *ctx = req->ctx;
5570 spin_lock_irq(&ctx->completion_lock);
5571 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5573 io_cqring_fill_event(req, ret);
5574 io_commit_cqring(ctx);
5575 spin_unlock_irq(&ctx->completion_lock);
5576 io_cqring_ev_posted(ctx);
5578 req_set_fail_links(req);
5583 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5584 bool is_timeout_link)
5586 struct io_timeout_data *data;
5588 u32 off = READ_ONCE(sqe->off);
5590 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5592 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
5595 if (off && is_timeout_link)
5597 flags = READ_ONCE(sqe->timeout_flags);
5598 if (flags & ~IORING_TIMEOUT_ABS)
5601 req->timeout.off = off;
5603 if (!req->async_data && io_alloc_async_data(req))
5606 data = req->async_data;
5609 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5612 if (flags & IORING_TIMEOUT_ABS)
5613 data->mode = HRTIMER_MODE_ABS;
5615 data->mode = HRTIMER_MODE_REL;
5617 INIT_LIST_HEAD(&req->timeout.list);
5618 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5622 static int io_timeout(struct io_kiocb *req)
5624 struct io_ring_ctx *ctx = req->ctx;
5625 struct io_timeout_data *data = req->async_data;
5626 struct list_head *entry;
5627 u32 tail, off = req->timeout.off;
5629 spin_lock_irq(&ctx->completion_lock);
5632 * sqe->off holds how many events that need to occur for this
5633 * timeout event to be satisfied. If it isn't set, then this is
5634 * a pure timeout request, sequence isn't used.
5636 if (io_is_timeout_noseq(req)) {
5637 entry = ctx->timeout_list.prev;
5641 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5642 req->timeout.target_seq = tail + off;
5644 /* Update the last seq here in case io_flush_timeouts() hasn't.
5645 * This is safe because ->completion_lock is held, and submissions
5646 * and completions are never mixed in the same ->completion_lock section.
5648 ctx->cq_last_tm_flush = tail;
5651 * Insertion sort, ensuring the first entry in the list is always
5652 * the one we need first.
5654 list_for_each_prev(entry, &ctx->timeout_list) {
5655 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5658 if (io_is_timeout_noseq(nxt))
5660 /* nxt.seq is behind @tail, otherwise would've been completed */
5661 if (off >= nxt->timeout.target_seq - tail)
5665 list_add(&req->timeout.list, entry);
5666 data->timer.function = io_timeout_fn;
5667 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5668 spin_unlock_irq(&ctx->completion_lock);
5672 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5674 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5676 return req->user_data == (unsigned long) data;
5679 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5681 enum io_wq_cancel cancel_ret;
5684 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5685 switch (cancel_ret) {
5686 case IO_WQ_CANCEL_OK:
5689 case IO_WQ_CANCEL_RUNNING:
5692 case IO_WQ_CANCEL_NOTFOUND:
5700 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5701 struct io_kiocb *req, __u64 sqe_addr,
5704 unsigned long flags;
5707 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5708 if (ret != -ENOENT) {
5709 spin_lock_irqsave(&ctx->completion_lock, flags);
5713 spin_lock_irqsave(&ctx->completion_lock, flags);
5714 ret = io_timeout_cancel(ctx, sqe_addr);
5717 ret = io_poll_cancel(ctx, sqe_addr);
5721 io_cqring_fill_event(req, ret);
5722 io_commit_cqring(ctx);
5723 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5724 io_cqring_ev_posted(ctx);
5727 req_set_fail_links(req);
5731 static int io_async_cancel_prep(struct io_kiocb *req,
5732 const struct io_uring_sqe *sqe)
5734 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5736 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5738 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
5742 req->cancel.addr = READ_ONCE(sqe->addr);
5746 static int io_async_cancel(struct io_kiocb *req)
5748 struct io_ring_ctx *ctx = req->ctx;
5750 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5754 static int io_files_update_prep(struct io_kiocb *req,
5755 const struct io_uring_sqe *sqe)
5757 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5759 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5761 if (sqe->ioprio || sqe->rw_flags)
5764 req->files_update.offset = READ_ONCE(sqe->off);
5765 req->files_update.nr_args = READ_ONCE(sqe->len);
5766 if (!req->files_update.nr_args)
5768 req->files_update.arg = READ_ONCE(sqe->addr);
5772 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5773 struct io_comp_state *cs)
5775 struct io_ring_ctx *ctx = req->ctx;
5776 struct io_uring_files_update up;
5782 up.offset = req->files_update.offset;
5783 up.fds = req->files_update.arg;
5785 mutex_lock(&ctx->uring_lock);
5786 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5787 mutex_unlock(&ctx->uring_lock);
5790 req_set_fail_links(req);
5791 __io_req_complete(req, ret, 0, cs);
5795 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5797 switch (req->opcode) {
5800 case IORING_OP_READV:
5801 case IORING_OP_READ_FIXED:
5802 case IORING_OP_READ:
5803 return io_read_prep(req, sqe);
5804 case IORING_OP_WRITEV:
5805 case IORING_OP_WRITE_FIXED:
5806 case IORING_OP_WRITE:
5807 return io_write_prep(req, sqe);
5808 case IORING_OP_POLL_ADD:
5809 return io_poll_add_prep(req, sqe);
5810 case IORING_OP_POLL_REMOVE:
5811 return io_poll_remove_prep(req, sqe);
5812 case IORING_OP_FSYNC:
5813 return io_prep_fsync(req, sqe);
5814 case IORING_OP_SYNC_FILE_RANGE:
5815 return io_prep_sfr(req, sqe);
5816 case IORING_OP_SENDMSG:
5817 case IORING_OP_SEND:
5818 return io_sendmsg_prep(req, sqe);
5819 case IORING_OP_RECVMSG:
5820 case IORING_OP_RECV:
5821 return io_recvmsg_prep(req, sqe);
5822 case IORING_OP_CONNECT:
5823 return io_connect_prep(req, sqe);
5824 case IORING_OP_TIMEOUT:
5825 return io_timeout_prep(req, sqe, false);
5826 case IORING_OP_TIMEOUT_REMOVE:
5827 return io_timeout_remove_prep(req, sqe);
5828 case IORING_OP_ASYNC_CANCEL:
5829 return io_async_cancel_prep(req, sqe);
5830 case IORING_OP_LINK_TIMEOUT:
5831 return io_timeout_prep(req, sqe, true);
5832 case IORING_OP_ACCEPT:
5833 return io_accept_prep(req, sqe);
5834 case IORING_OP_FALLOCATE:
5835 return io_fallocate_prep(req, sqe);
5836 case IORING_OP_OPENAT:
5837 return io_openat_prep(req, sqe);
5838 case IORING_OP_CLOSE:
5839 return io_close_prep(req, sqe);
5840 case IORING_OP_FILES_UPDATE:
5841 return io_files_update_prep(req, sqe);
5842 case IORING_OP_STATX:
5843 return io_statx_prep(req, sqe);
5844 case IORING_OP_FADVISE:
5845 return io_fadvise_prep(req, sqe);
5846 case IORING_OP_MADVISE:
5847 return io_madvise_prep(req, sqe);
5848 case IORING_OP_OPENAT2:
5849 return io_openat2_prep(req, sqe);
5850 case IORING_OP_EPOLL_CTL:
5851 return io_epoll_ctl_prep(req, sqe);
5852 case IORING_OP_SPLICE:
5853 return io_splice_prep(req, sqe);
5854 case IORING_OP_PROVIDE_BUFFERS:
5855 return io_provide_buffers_prep(req, sqe);
5856 case IORING_OP_REMOVE_BUFFERS:
5857 return io_remove_buffers_prep(req, sqe);
5859 return io_tee_prep(req, sqe);
5862 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5867 static int io_req_defer_prep(struct io_kiocb *req,
5868 const struct io_uring_sqe *sqe)
5872 if (io_alloc_async_data(req))
5874 return io_req_prep(req, sqe);
5877 static u32 io_get_sequence(struct io_kiocb *req)
5879 struct io_kiocb *pos;
5880 struct io_ring_ctx *ctx = req->ctx;
5881 u32 total_submitted, nr_reqs = 1;
5883 if (req->flags & REQ_F_LINK_HEAD)
5884 list_for_each_entry(pos, &req->link_list, link_list)
5887 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5888 return total_submitted - nr_reqs;
5891 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5893 struct io_ring_ctx *ctx = req->ctx;
5894 struct io_defer_entry *de;
5898 /* Still need defer if there is pending req in defer list. */
5899 if (likely(list_empty_careful(&ctx->defer_list) &&
5900 !(req->flags & REQ_F_IO_DRAIN)))
5903 seq = io_get_sequence(req);
5904 /* Still a chance to pass the sequence check */
5905 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5908 if (!req->async_data) {
5909 ret = io_req_defer_prep(req, sqe);
5913 io_prep_async_link(req);
5914 de = kmalloc(sizeof(*de), GFP_KERNEL);
5918 spin_lock_irq(&ctx->completion_lock);
5919 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5920 spin_unlock_irq(&ctx->completion_lock);
5922 io_queue_async_work(req);
5923 return -EIOCBQUEUED;
5926 trace_io_uring_defer(ctx, req, req->user_data);
5929 list_add_tail(&de->list, &ctx->defer_list);
5930 spin_unlock_irq(&ctx->completion_lock);
5931 return -EIOCBQUEUED;
5934 static void io_req_drop_files(struct io_kiocb *req)
5936 struct io_ring_ctx *ctx = req->ctx;
5937 struct io_uring_task *tctx = req->task->io_uring;
5938 unsigned long flags;
5940 if (req->work.flags & IO_WQ_WORK_FILES) {
5941 put_files_struct(req->work.identity->files);
5942 put_nsproxy(req->work.identity->nsproxy);
5944 spin_lock_irqsave(&ctx->inflight_lock, flags);
5945 list_del(&req->inflight_entry);
5946 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5947 req->flags &= ~REQ_F_INFLIGHT;
5948 req->work.flags &= ~IO_WQ_WORK_FILES;
5949 if (atomic_read(&tctx->in_idle))
5950 wake_up(&tctx->wait);
5953 static void __io_clean_op(struct io_kiocb *req)
5955 if (req->flags & REQ_F_BUFFER_SELECTED) {
5956 switch (req->opcode) {
5957 case IORING_OP_READV:
5958 case IORING_OP_READ_FIXED:
5959 case IORING_OP_READ:
5960 kfree((void *)(unsigned long)req->rw.addr);
5962 case IORING_OP_RECVMSG:
5963 case IORING_OP_RECV:
5964 kfree(req->sr_msg.kbuf);
5967 req->flags &= ~REQ_F_BUFFER_SELECTED;
5970 if (req->flags & REQ_F_NEED_CLEANUP) {
5971 switch (req->opcode) {
5972 case IORING_OP_READV:
5973 case IORING_OP_READ_FIXED:
5974 case IORING_OP_READ:
5975 case IORING_OP_WRITEV:
5976 case IORING_OP_WRITE_FIXED:
5977 case IORING_OP_WRITE: {
5978 struct io_async_rw *io = req->async_data;
5980 kfree(io->free_iovec);
5983 case IORING_OP_RECVMSG:
5984 case IORING_OP_SENDMSG: {
5985 struct io_async_msghdr *io = req->async_data;
5986 if (io->iov != io->fast_iov)
5990 case IORING_OP_SPLICE:
5992 io_put_file(req, req->splice.file_in,
5993 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5995 case IORING_OP_OPENAT:
5996 case IORING_OP_OPENAT2:
5997 if (req->open.filename)
5998 putname(req->open.filename);
6001 req->flags &= ~REQ_F_NEED_CLEANUP;
6005 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
6006 struct io_comp_state *cs)
6008 struct io_ring_ctx *ctx = req->ctx;
6011 switch (req->opcode) {
6013 ret = io_nop(req, cs);
6015 case IORING_OP_READV:
6016 case IORING_OP_READ_FIXED:
6017 case IORING_OP_READ:
6018 ret = io_read(req, force_nonblock, cs);
6020 case IORING_OP_WRITEV:
6021 case IORING_OP_WRITE_FIXED:
6022 case IORING_OP_WRITE:
6023 ret = io_write(req, force_nonblock, cs);
6025 case IORING_OP_FSYNC:
6026 ret = io_fsync(req, force_nonblock);
6028 case IORING_OP_POLL_ADD:
6029 ret = io_poll_add(req);
6031 case IORING_OP_POLL_REMOVE:
6032 ret = io_poll_remove(req);
6034 case IORING_OP_SYNC_FILE_RANGE:
6035 ret = io_sync_file_range(req, force_nonblock);
6037 case IORING_OP_SENDMSG:
6038 ret = io_sendmsg(req, force_nonblock, cs);
6040 case IORING_OP_SEND:
6041 ret = io_send(req, force_nonblock, cs);
6043 case IORING_OP_RECVMSG:
6044 ret = io_recvmsg(req, force_nonblock, cs);
6046 case IORING_OP_RECV:
6047 ret = io_recv(req, force_nonblock, cs);
6049 case IORING_OP_TIMEOUT:
6050 ret = io_timeout(req);
6052 case IORING_OP_TIMEOUT_REMOVE:
6053 ret = io_timeout_remove(req);
6055 case IORING_OP_ACCEPT:
6056 ret = io_accept(req, force_nonblock, cs);
6058 case IORING_OP_CONNECT:
6059 ret = io_connect(req, force_nonblock, cs);
6061 case IORING_OP_ASYNC_CANCEL:
6062 ret = io_async_cancel(req);
6064 case IORING_OP_FALLOCATE:
6065 ret = io_fallocate(req, force_nonblock);
6067 case IORING_OP_OPENAT:
6068 ret = io_openat(req, force_nonblock);
6070 case IORING_OP_CLOSE:
6071 ret = io_close(req, force_nonblock, cs);
6073 case IORING_OP_FILES_UPDATE:
6074 ret = io_files_update(req, force_nonblock, cs);
6076 case IORING_OP_STATX:
6077 ret = io_statx(req, force_nonblock);
6079 case IORING_OP_FADVISE:
6080 ret = io_fadvise(req, force_nonblock);
6082 case IORING_OP_MADVISE:
6083 ret = io_madvise(req, force_nonblock);
6085 case IORING_OP_OPENAT2:
6086 ret = io_openat2(req, force_nonblock);
6088 case IORING_OP_EPOLL_CTL:
6089 ret = io_epoll_ctl(req, force_nonblock, cs);
6091 case IORING_OP_SPLICE:
6092 ret = io_splice(req, force_nonblock);
6094 case IORING_OP_PROVIDE_BUFFERS:
6095 ret = io_provide_buffers(req, force_nonblock, cs);
6097 case IORING_OP_REMOVE_BUFFERS:
6098 ret = io_remove_buffers(req, force_nonblock, cs);
6101 ret = io_tee(req, force_nonblock);
6111 /* If the op doesn't have a file, we're not polling for it */
6112 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6113 const bool in_async = io_wq_current_is_worker();
6115 /* workqueue context doesn't hold uring_lock, grab it now */
6117 mutex_lock(&ctx->uring_lock);
6119 io_iopoll_req_issued(req);
6122 mutex_unlock(&ctx->uring_lock);
6128 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6130 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6131 struct io_kiocb *timeout;
6134 timeout = io_prep_linked_timeout(req);
6136 io_queue_linked_timeout(timeout);
6138 /* if NO_CANCEL is set, we must still run the work */
6139 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6140 IO_WQ_WORK_CANCEL) {
6146 ret = io_issue_sqe(req, false, NULL);
6148 * We can get EAGAIN for polled IO even though we're
6149 * forcing a sync submission from here, since we can't
6150 * wait for request slots on the block side.
6159 struct io_ring_ctx *lock_ctx = NULL;
6161 if (req->ctx->flags & IORING_SETUP_IOPOLL)
6162 lock_ctx = req->ctx;
6165 * io_iopoll_complete() does not hold completion_lock to
6166 * complete polled io, so here for polled io, we can not call
6167 * io_req_complete() directly, otherwise there maybe concurrent
6168 * access to cqring, defer_list, etc, which is not safe. Given
6169 * that io_iopoll_complete() is always called under uring_lock,
6170 * so here for polled io, we also get uring_lock to complete
6174 mutex_lock(&lock_ctx->uring_lock);
6176 req_set_fail_links(req);
6177 io_req_complete(req, ret);
6180 mutex_unlock(&lock_ctx->uring_lock);
6183 return io_steal_work(req);
6186 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6189 struct fixed_file_table *table;
6191 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6192 return table->files[index & IORING_FILE_TABLE_MASK];
6195 static struct file *io_file_get(struct io_submit_state *state,
6196 struct io_kiocb *req, int fd, bool fixed)
6198 struct io_ring_ctx *ctx = req->ctx;
6202 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6204 fd = array_index_nospec(fd, ctx->nr_user_files);
6205 file = io_file_from_index(ctx, fd);
6207 req->fixed_file_refs = &ctx->file_data->node->refs;
6208 percpu_ref_get(req->fixed_file_refs);
6211 trace_io_uring_file_get(ctx, fd);
6212 file = __io_file_get(state, fd);
6215 if (file && file->f_op == &io_uring_fops &&
6216 !(req->flags & REQ_F_INFLIGHT)) {
6217 io_req_init_async(req);
6218 req->flags |= REQ_F_INFLIGHT;
6220 spin_lock_irq(&ctx->inflight_lock);
6221 list_add(&req->inflight_entry, &ctx->inflight_list);
6222 spin_unlock_irq(&ctx->inflight_lock);
6228 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6233 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6234 if (unlikely(!fixed && io_async_submit(req->ctx)))
6237 req->file = io_file_get(state, req, fd, fixed);
6238 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6243 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6245 struct io_timeout_data *data = container_of(timer,
6246 struct io_timeout_data, timer);
6247 struct io_kiocb *req = data->req;
6248 struct io_ring_ctx *ctx = req->ctx;
6249 struct io_kiocb *prev = NULL;
6250 unsigned long flags;
6252 spin_lock_irqsave(&ctx->completion_lock, flags);
6255 * We don't expect the list to be empty, that will only happen if we
6256 * race with the completion of the linked work.
6258 if (!list_empty(&req->link_list)) {
6259 prev = list_entry(req->link_list.prev, struct io_kiocb,
6261 list_del_init(&req->link_list);
6262 if (!refcount_inc_not_zero(&prev->refs))
6266 list_del(&req->timeout.list);
6267 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6270 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6271 io_put_req_deferred(prev, 1);
6273 io_cqring_add_event(req, -ETIME, 0);
6274 io_put_req_deferred(req, 1);
6276 return HRTIMER_NORESTART;
6279 static void __io_queue_linked_timeout(struct io_kiocb *req)
6282 * If the list is now empty, then our linked request finished before
6283 * we got a chance to setup the timer
6285 if (!list_empty(&req->link_list)) {
6286 struct io_timeout_data *data = req->async_data;
6288 data->timer.function = io_link_timeout_fn;
6289 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6294 static void io_queue_linked_timeout(struct io_kiocb *req)
6296 struct io_ring_ctx *ctx = req->ctx;
6298 spin_lock_irq(&ctx->completion_lock);
6299 __io_queue_linked_timeout(req);
6300 spin_unlock_irq(&ctx->completion_lock);
6302 /* drop submission reference */
6306 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6308 struct io_kiocb *nxt;
6310 if (!(req->flags & REQ_F_LINK_HEAD))
6312 if (req->flags & REQ_F_LINK_TIMEOUT)
6315 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6317 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6320 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6321 req->flags |= REQ_F_LINK_TIMEOUT;
6325 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6327 struct io_kiocb *linked_timeout;
6328 const struct cred *old_creds = NULL;
6332 linked_timeout = io_prep_linked_timeout(req);
6334 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6335 (req->work.flags & IO_WQ_WORK_CREDS) &&
6336 req->work.identity->creds != current_cred()) {
6338 revert_creds(old_creds);
6339 if (old_creds == req->work.identity->creds)
6340 old_creds = NULL; /* restored original creds */
6342 old_creds = override_creds(req->work.identity->creds);
6345 ret = io_issue_sqe(req, true, cs);
6348 * We async punt it if the file wasn't marked NOWAIT, or if the file
6349 * doesn't support non-blocking read/write attempts
6351 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6352 if (!io_arm_poll_handler(req)) {
6354 * Queued up for async execution, worker will release
6355 * submit reference when the iocb is actually submitted.
6357 io_queue_async_work(req);
6361 io_queue_linked_timeout(linked_timeout);
6362 } else if (likely(!ret)) {
6363 /* drop submission reference */
6364 req = io_put_req_find_next(req);
6366 io_queue_linked_timeout(linked_timeout);
6369 if (!(req->flags & REQ_F_FORCE_ASYNC))
6371 io_queue_async_work(req);
6374 /* un-prep timeout, so it'll be killed as any other linked */
6375 req->flags &= ~REQ_F_LINK_TIMEOUT;
6376 req_set_fail_links(req);
6378 io_req_complete(req, ret);
6382 revert_creds(old_creds);
6385 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6386 struct io_comp_state *cs)
6390 ret = io_req_defer(req, sqe);
6392 if (ret != -EIOCBQUEUED) {
6394 req_set_fail_links(req);
6396 io_req_complete(req, ret);
6398 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6399 if (!req->async_data) {
6400 ret = io_req_defer_prep(req, sqe);
6404 io_queue_async_work(req);
6407 ret = io_req_prep(req, sqe);
6411 __io_queue_sqe(req, cs);
6415 static inline void io_queue_link_head(struct io_kiocb *req,
6416 struct io_comp_state *cs)
6418 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6420 io_req_complete(req, -ECANCELED);
6422 io_queue_sqe(req, NULL, cs);
6425 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6426 struct io_kiocb **link, struct io_comp_state *cs)
6428 struct io_ring_ctx *ctx = req->ctx;
6432 * If we already have a head request, queue this one for async
6433 * submittal once the head completes. If we don't have a head but
6434 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6435 * submitted sync once the chain is complete. If none of those
6436 * conditions are true (normal request), then just queue it.
6439 struct io_kiocb *head = *link;
6442 * Taking sequential execution of a link, draining both sides
6443 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6444 * requests in the link. So, it drains the head and the
6445 * next after the link request. The last one is done via
6446 * drain_next flag to persist the effect across calls.
6448 if (req->flags & REQ_F_IO_DRAIN) {
6449 head->flags |= REQ_F_IO_DRAIN;
6450 ctx->drain_next = 1;
6452 ret = io_req_defer_prep(req, sqe);
6453 if (unlikely(ret)) {
6454 /* fail even hard links since we don't submit */
6455 head->flags |= REQ_F_FAIL_LINK;
6458 trace_io_uring_link(ctx, req, head);
6459 list_add_tail(&req->link_list, &head->link_list);
6461 /* last request of a link, enqueue the link */
6462 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6463 io_queue_link_head(head, cs);
6467 if (unlikely(ctx->drain_next)) {
6468 req->flags |= REQ_F_IO_DRAIN;
6469 ctx->drain_next = 0;
6471 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6472 req->flags |= REQ_F_LINK_HEAD;
6473 INIT_LIST_HEAD(&req->link_list);
6475 ret = io_req_defer_prep(req, sqe);
6477 req->flags |= REQ_F_FAIL_LINK;
6480 io_queue_sqe(req, sqe, cs);
6488 * Batched submission is done, ensure local IO is flushed out.
6490 static void io_submit_state_end(struct io_submit_state *state)
6492 if (!list_empty(&state->comp.list))
6493 io_submit_flush_completions(&state->comp);
6494 blk_finish_plug(&state->plug);
6495 io_state_file_put(state);
6496 if (state->free_reqs)
6497 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6501 * Start submission side cache.
6503 static void io_submit_state_start(struct io_submit_state *state,
6504 struct io_ring_ctx *ctx, unsigned int max_ios)
6506 blk_start_plug(&state->plug);
6508 INIT_LIST_HEAD(&state->comp.list);
6509 state->comp.ctx = ctx;
6510 state->free_reqs = 0;
6512 state->ios_left = max_ios;
6515 static void io_commit_sqring(struct io_ring_ctx *ctx)
6517 struct io_rings *rings = ctx->rings;
6520 * Ensure any loads from the SQEs are done at this point,
6521 * since once we write the new head, the application could
6522 * write new data to them.
6524 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6528 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6529 * that is mapped by userspace. This means that care needs to be taken to
6530 * ensure that reads are stable, as we cannot rely on userspace always
6531 * being a good citizen. If members of the sqe are validated and then later
6532 * used, it's important that those reads are done through READ_ONCE() to
6533 * prevent a re-load down the line.
6535 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6537 u32 *sq_array = ctx->sq_array;
6541 * The cached sq head (or cq tail) serves two purposes:
6543 * 1) allows us to batch the cost of updating the user visible
6545 * 2) allows the kernel side to track the head on its own, even
6546 * though the application is the one updating it.
6548 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6549 if (likely(head < ctx->sq_entries))
6550 return &ctx->sq_sqes[head];
6552 /* drop invalid entries */
6553 ctx->cached_sq_dropped++;
6554 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6558 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6560 ctx->cached_sq_head++;
6564 * Check SQE restrictions (opcode and flags).
6566 * Returns 'true' if SQE is allowed, 'false' otherwise.
6568 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6569 struct io_kiocb *req,
6570 unsigned int sqe_flags)
6572 if (!ctx->restricted)
6575 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6578 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6579 ctx->restrictions.sqe_flags_required)
6582 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6583 ctx->restrictions.sqe_flags_required))
6589 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6590 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6591 IOSQE_BUFFER_SELECT)
6593 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6594 const struct io_uring_sqe *sqe,
6595 struct io_submit_state *state)
6597 unsigned int sqe_flags;
6600 req->opcode = READ_ONCE(sqe->opcode);
6601 req->user_data = READ_ONCE(sqe->user_data);
6602 req->async_data = NULL;
6606 /* one is dropped after submission, the other at completion */
6607 refcount_set(&req->refs, 2);
6608 req->task = current;
6611 if (unlikely(req->opcode >= IORING_OP_LAST))
6614 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6617 sqe_flags = READ_ONCE(sqe->flags);
6618 /* enforce forwards compatibility on users */
6619 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6622 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6625 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6626 !io_op_defs[req->opcode].buffer_select)
6629 id = READ_ONCE(sqe->personality);
6631 struct io_identity *iod;
6633 iod = xa_load(&ctx->personalities, id);
6636 refcount_inc(&iod->count);
6638 __io_req_init_async(req);
6639 get_cred(iod->creds);
6640 req->work.identity = iod;
6641 req->work.flags |= IO_WQ_WORK_CREDS;
6644 /* same numerical values with corresponding REQ_F_*, safe to copy */
6645 req->flags |= sqe_flags;
6647 if (!io_op_defs[req->opcode].needs_file)
6650 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6655 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6657 struct io_submit_state state;
6658 struct io_kiocb *link = NULL;
6659 int i, submitted = 0;
6661 /* if we have a backlog and couldn't flush it all, return BUSY */
6662 if (test_bit(0, &ctx->sq_check_overflow)) {
6663 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6667 /* make sure SQ entry isn't read before tail */
6668 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6670 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6673 percpu_counter_add(¤t->io_uring->inflight, nr);
6674 refcount_add(nr, ¤t->usage);
6676 io_submit_state_start(&state, ctx, nr);
6678 for (i = 0; i < nr; i++) {
6679 const struct io_uring_sqe *sqe;
6680 struct io_kiocb *req;
6683 sqe = io_get_sqe(ctx);
6684 if (unlikely(!sqe)) {
6685 io_consume_sqe(ctx);
6688 req = io_alloc_req(ctx, &state);
6689 if (unlikely(!req)) {
6691 submitted = -EAGAIN;
6694 io_consume_sqe(ctx);
6695 /* will complete beyond this point, count as submitted */
6698 err = io_init_req(ctx, req, sqe, &state);
6699 if (unlikely(err)) {
6702 io_req_complete(req, err);
6706 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6707 true, io_async_submit(ctx));
6708 err = io_submit_sqe(req, sqe, &link, &state.comp);
6713 if (unlikely(submitted != nr)) {
6714 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6715 struct io_uring_task *tctx = current->io_uring;
6716 int unused = nr - ref_used;
6718 percpu_ref_put_many(&ctx->refs, unused);
6719 percpu_counter_sub(&tctx->inflight, unused);
6720 put_task_struct_many(current, unused);
6723 io_queue_link_head(link, &state.comp);
6724 io_submit_state_end(&state);
6726 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6727 io_commit_sqring(ctx);
6732 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6734 /* Tell userspace we may need a wakeup call */
6735 spin_lock_irq(&ctx->completion_lock);
6736 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6737 spin_unlock_irq(&ctx->completion_lock);
6740 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6742 spin_lock_irq(&ctx->completion_lock);
6743 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6744 spin_unlock_irq(&ctx->completion_lock);
6747 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6748 int sync, void *key)
6750 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6753 ret = autoremove_wake_function(wqe, mode, sync, key);
6755 unsigned long flags;
6757 spin_lock_irqsave(&ctx->completion_lock, flags);
6758 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6759 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6770 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6771 unsigned long start_jiffies, bool cap_entries)
6773 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6774 struct io_sq_data *sqd = ctx->sq_data;
6775 unsigned int to_submit;
6779 if (!list_empty(&ctx->iopoll_list)) {
6780 unsigned nr_events = 0;
6782 mutex_lock(&ctx->uring_lock);
6783 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6784 io_do_iopoll(ctx, &nr_events, 0);
6785 mutex_unlock(&ctx->uring_lock);
6788 to_submit = io_sqring_entries(ctx);
6791 * If submit got -EBUSY, flag us as needing the application
6792 * to enter the kernel to reap and flush events.
6794 if (!to_submit || ret == -EBUSY || need_resched()) {
6796 * Drop cur_mm before scheduling, we can't hold it for
6797 * long periods (or over schedule()). Do this before
6798 * adding ourselves to the waitqueue, as the unuse/drop
6801 io_sq_thread_drop_mm();
6804 * We're polling. If we're within the defined idle
6805 * period, then let us spin without work before going
6806 * to sleep. The exception is if we got EBUSY doing
6807 * more IO, we should wait for the application to
6808 * reap events and wake us up.
6810 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6811 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6812 !percpu_ref_is_dying(&ctx->refs)))
6815 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6816 TASK_INTERRUPTIBLE);
6819 * While doing polled IO, before going to sleep, we need
6820 * to check if there are new reqs added to iopoll_list,
6821 * it is because reqs may have been punted to io worker
6822 * and will be added to iopoll_list later, hence check
6823 * the iopoll_list again.
6825 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6826 !list_empty_careful(&ctx->iopoll_list)) {
6827 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6831 to_submit = io_sqring_entries(ctx);
6832 if (!to_submit || ret == -EBUSY)
6836 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6837 io_ring_clear_wakeup_flag(ctx);
6839 /* if we're handling multiple rings, cap submit size for fairness */
6840 if (cap_entries && to_submit > 8)
6843 mutex_lock(&ctx->uring_lock);
6844 if (likely(!percpu_ref_is_dying(&ctx->refs) && !ctx->sqo_dead))
6845 ret = io_submit_sqes(ctx, to_submit);
6846 mutex_unlock(&ctx->uring_lock);
6848 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6849 wake_up(&ctx->sqo_sq_wait);
6851 return SQT_DID_WORK;
6854 static void io_sqd_init_new(struct io_sq_data *sqd)
6856 struct io_ring_ctx *ctx;
6858 while (!list_empty(&sqd->ctx_new_list)) {
6859 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6860 init_wait(&ctx->sqo_wait_entry);
6861 ctx->sqo_wait_entry.func = io_sq_wake_function;
6862 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6863 complete(&ctx->sq_thread_comp);
6867 static int io_sq_thread(void *data)
6869 struct cgroup_subsys_state *cur_css = NULL;
6870 const struct cred *old_cred = NULL;
6871 struct io_sq_data *sqd = data;
6872 struct io_ring_ctx *ctx;
6873 unsigned long start_jiffies;
6875 start_jiffies = jiffies;
6876 while (!kthread_should_stop()) {
6877 enum sq_ret ret = 0;
6881 * Any changes to the sqd lists are synchronized through the
6882 * kthread parking. This synchronizes the thread vs users,
6883 * the users are synchronized on the sqd->ctx_lock.
6885 if (kthread_should_park()) {
6888 * When sq thread is unparked, in case the previous park operation
6889 * comes from io_put_sq_data(), which means that sq thread is going
6890 * to be stopped, so here needs to have a check.
6892 if (kthread_should_stop())
6896 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6897 io_sqd_init_new(sqd);
6899 cap_entries = !list_is_singular(&sqd->ctx_list);
6901 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6902 if (current->cred != ctx->creds) {
6904 revert_creds(old_cred);
6905 old_cred = override_creds(ctx->creds);
6907 io_sq_thread_associate_blkcg(ctx, &cur_css);
6909 current->loginuid = ctx->loginuid;
6910 current->sessionid = ctx->sessionid;
6913 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6915 io_sq_thread_drop_mm();
6918 if (ret & SQT_SPIN) {
6920 io_sq_thread_drop_mm();
6922 } else if (ret == SQT_IDLE) {
6923 if (kthread_should_park())
6925 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6926 io_ring_set_wakeup_flag(ctx);
6928 start_jiffies = jiffies;
6929 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6930 io_ring_clear_wakeup_flag(ctx);
6935 io_sq_thread_drop_mm();
6938 io_sq_thread_unassociate_blkcg();
6940 revert_creds(old_cred);
6947 struct io_wait_queue {
6948 struct wait_queue_entry wq;
6949 struct io_ring_ctx *ctx;
6951 unsigned nr_timeouts;
6954 static inline bool io_should_wake(struct io_wait_queue *iowq)
6956 struct io_ring_ctx *ctx = iowq->ctx;
6959 * Wake up if we have enough events, or if a timeout occurred since we
6960 * started waiting. For timeouts, we always want to return to userspace,
6961 * regardless of event count.
6963 return io_cqring_events(ctx) >= iowq->to_wait ||
6964 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6967 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6968 int wake_flags, void *key)
6970 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6974 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6975 * the task, and the next invocation will do it.
6977 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6978 return autoremove_wake_function(curr, mode, wake_flags, key);
6982 static int io_run_task_work_sig(void)
6984 if (io_run_task_work())
6986 if (!signal_pending(current))
6988 if (current->jobctl & JOBCTL_TASK_WORK) {
6989 spin_lock_irq(¤t->sighand->siglock);
6990 current->jobctl &= ~JOBCTL_TASK_WORK;
6991 recalc_sigpending();
6992 spin_unlock_irq(¤t->sighand->siglock);
6999 * Wait until events become available, if we don't already have some. The
7000 * application must reap them itself, as they reside on the shared cq ring.
7002 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7003 const sigset_t __user *sig, size_t sigsz)
7005 struct io_wait_queue iowq = {
7008 .func = io_wake_function,
7009 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7012 .to_wait = min_events,
7014 struct io_rings *rings = ctx->rings;
7018 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7019 if (io_cqring_events(ctx) >= min_events)
7021 if (!io_run_task_work())
7026 #ifdef CONFIG_COMPAT
7027 if (in_compat_syscall())
7028 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7032 ret = set_user_sigmask(sig, sigsz);
7038 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7039 trace_io_uring_cqring_wait(ctx, min_events);
7041 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7042 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7043 TASK_INTERRUPTIBLE);
7044 /* make sure we run task_work before checking for signals */
7045 ret = io_run_task_work_sig();
7047 finish_wait(&ctx->wait, &iowq.wq);
7052 if (io_should_wake(&iowq))
7054 if (test_bit(0, &ctx->cq_check_overflow)) {
7055 finish_wait(&ctx->wait, &iowq.wq);
7060 finish_wait(&ctx->wait, &iowq.wq);
7062 restore_saved_sigmask_unless(ret == -EINTR);
7064 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7067 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7069 #if defined(CONFIG_UNIX)
7070 if (ctx->ring_sock) {
7071 struct sock *sock = ctx->ring_sock->sk;
7072 struct sk_buff *skb;
7074 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7080 for (i = 0; i < ctx->nr_user_files; i++) {
7083 file = io_file_from_index(ctx, i);
7090 static void io_file_ref_kill(struct percpu_ref *ref)
7092 struct fixed_file_data *data;
7094 data = container_of(ref, struct fixed_file_data, refs);
7095 complete(&data->done);
7098 static void io_sqe_files_set_node(struct fixed_file_data *file_data,
7099 struct fixed_file_ref_node *ref_node)
7101 spin_lock_bh(&file_data->lock);
7102 file_data->node = ref_node;
7103 list_add_tail(&ref_node->node, &file_data->ref_list);
7104 spin_unlock_bh(&file_data->lock);
7105 percpu_ref_get(&file_data->refs);
7108 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7110 struct fixed_file_data *data = ctx->file_data;
7111 struct fixed_file_ref_node *backup_node, *ref_node = NULL;
7112 unsigned nr_tables, i;
7117 backup_node = alloc_fixed_file_ref_node(ctx);
7121 spin_lock_bh(&data->lock);
7122 ref_node = data->node;
7123 spin_unlock_bh(&data->lock);
7125 percpu_ref_kill(&ref_node->refs);
7127 percpu_ref_kill(&data->refs);
7129 /* wait for all refs nodes to complete */
7130 flush_delayed_work(&ctx->file_put_work);
7132 ret = wait_for_completion_interruptible(&data->done);
7135 ret = io_run_task_work_sig();
7137 percpu_ref_resurrect(&data->refs);
7138 reinit_completion(&data->done);
7139 io_sqe_files_set_node(data, backup_node);
7144 __io_sqe_files_unregister(ctx);
7145 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7146 for (i = 0; i < nr_tables; i++)
7147 kfree(data->table[i].files);
7149 percpu_ref_exit(&data->refs);
7151 ctx->file_data = NULL;
7152 ctx->nr_user_files = 0;
7153 destroy_fixed_file_ref_node(backup_node);
7157 static void io_put_sq_data(struct io_sq_data *sqd)
7159 if (refcount_dec_and_test(&sqd->refs)) {
7161 * The park is a bit of a work-around, without it we get
7162 * warning spews on shutdown with SQPOLL set and affinity
7163 * set to a single CPU.
7166 kthread_park(sqd->thread);
7167 kthread_stop(sqd->thread);
7174 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7176 struct io_ring_ctx *ctx_attach;
7177 struct io_sq_data *sqd;
7180 f = fdget(p->wq_fd);
7182 return ERR_PTR(-ENXIO);
7183 if (f.file->f_op != &io_uring_fops) {
7185 return ERR_PTR(-EINVAL);
7188 ctx_attach = f.file->private_data;
7189 sqd = ctx_attach->sq_data;
7192 return ERR_PTR(-EINVAL);
7195 refcount_inc(&sqd->refs);
7200 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7202 struct io_sq_data *sqd;
7204 if (p->flags & IORING_SETUP_ATTACH_WQ)
7205 return io_attach_sq_data(p);
7207 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7209 return ERR_PTR(-ENOMEM);
7211 refcount_set(&sqd->refs, 1);
7212 INIT_LIST_HEAD(&sqd->ctx_list);
7213 INIT_LIST_HEAD(&sqd->ctx_new_list);
7214 mutex_init(&sqd->ctx_lock);
7215 mutex_init(&sqd->lock);
7216 init_waitqueue_head(&sqd->wait);
7220 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7221 __releases(&sqd->lock)
7225 kthread_unpark(sqd->thread);
7226 mutex_unlock(&sqd->lock);
7229 static void io_sq_thread_park(struct io_sq_data *sqd)
7230 __acquires(&sqd->lock)
7234 mutex_lock(&sqd->lock);
7235 kthread_park(sqd->thread);
7238 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7240 struct io_sq_data *sqd = ctx->sq_data;
7245 * We may arrive here from the error branch in
7246 * io_sq_offload_create() where the kthread is created
7247 * without being waked up, thus wake it up now to make
7248 * sure the wait will complete.
7250 wake_up_process(sqd->thread);
7251 wait_for_completion(&ctx->sq_thread_comp);
7253 io_sq_thread_park(sqd);
7256 mutex_lock(&sqd->ctx_lock);
7257 list_del(&ctx->sqd_list);
7258 mutex_unlock(&sqd->ctx_lock);
7261 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7262 io_sq_thread_unpark(sqd);
7265 io_put_sq_data(sqd);
7266 ctx->sq_data = NULL;
7270 static void io_finish_async(struct io_ring_ctx *ctx)
7272 io_sq_thread_stop(ctx);
7275 io_wq_destroy(ctx->io_wq);
7280 #if defined(CONFIG_UNIX)
7282 * Ensure the UNIX gc is aware of our file set, so we are certain that
7283 * the io_uring can be safely unregistered on process exit, even if we have
7284 * loops in the file referencing.
7286 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7288 struct sock *sk = ctx->ring_sock->sk;
7289 struct scm_fp_list *fpl;
7290 struct sk_buff *skb;
7293 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7297 skb = alloc_skb(0, GFP_KERNEL);
7304 skb->scm_io_uring = 1;
7307 fpl->user = get_uid(ctx->user);
7308 for (i = 0; i < nr; i++) {
7309 struct file *file = io_file_from_index(ctx, i + offset);
7313 fpl->fp[nr_files] = get_file(file);
7314 unix_inflight(fpl->user, fpl->fp[nr_files]);
7319 fpl->max = SCM_MAX_FD;
7320 fpl->count = nr_files;
7321 UNIXCB(skb).fp = fpl;
7322 skb->destructor = unix_destruct_scm;
7323 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7324 skb_queue_head(&sk->sk_receive_queue, skb);
7326 for (i = 0; i < nr; i++) {
7327 struct file *file = io_file_from_index(ctx, i + offset);
7334 free_uid(fpl->user);
7342 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7343 * causes regular reference counting to break down. We rely on the UNIX
7344 * garbage collection to take care of this problem for us.
7346 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7348 unsigned left, total;
7352 left = ctx->nr_user_files;
7354 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7356 ret = __io_sqe_files_scm(ctx, this_files, total);
7360 total += this_files;
7366 while (total < ctx->nr_user_files) {
7367 struct file *file = io_file_from_index(ctx, total);
7377 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7383 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7384 unsigned nr_tables, unsigned nr_files)
7388 for (i = 0; i < nr_tables; i++) {
7389 struct fixed_file_table *table = &file_data->table[i];
7390 unsigned this_files;
7392 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7393 table->files = kcalloc(this_files, sizeof(struct file *),
7394 GFP_KERNEL_ACCOUNT);
7397 nr_files -= this_files;
7403 for (i = 0; i < nr_tables; i++) {
7404 struct fixed_file_table *table = &file_data->table[i];
7405 kfree(table->files);
7410 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7412 #if defined(CONFIG_UNIX)
7413 struct sock *sock = ctx->ring_sock->sk;
7414 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7415 struct sk_buff *skb;
7418 __skb_queue_head_init(&list);
7421 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7422 * remove this entry and rearrange the file array.
7424 skb = skb_dequeue(head);
7426 struct scm_fp_list *fp;
7428 fp = UNIXCB(skb).fp;
7429 for (i = 0; i < fp->count; i++) {
7432 if (fp->fp[i] != file)
7435 unix_notinflight(fp->user, fp->fp[i]);
7436 left = fp->count - 1 - i;
7438 memmove(&fp->fp[i], &fp->fp[i + 1],
7439 left * sizeof(struct file *));
7446 __skb_queue_tail(&list, skb);
7456 __skb_queue_tail(&list, skb);
7458 skb = skb_dequeue(head);
7461 if (skb_peek(&list)) {
7462 spin_lock_irq(&head->lock);
7463 while ((skb = __skb_dequeue(&list)) != NULL)
7464 __skb_queue_tail(head, skb);
7465 spin_unlock_irq(&head->lock);
7472 struct io_file_put {
7473 struct list_head list;
7477 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7479 struct fixed_file_data *file_data = ref_node->file_data;
7480 struct io_ring_ctx *ctx = file_data->ctx;
7481 struct io_file_put *pfile, *tmp;
7483 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7484 list_del(&pfile->list);
7485 io_ring_file_put(ctx, pfile->file);
7489 percpu_ref_exit(&ref_node->refs);
7491 percpu_ref_put(&file_data->refs);
7494 static void io_file_put_work(struct work_struct *work)
7496 struct io_ring_ctx *ctx;
7497 struct llist_node *node;
7499 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7500 node = llist_del_all(&ctx->file_put_llist);
7503 struct fixed_file_ref_node *ref_node;
7504 struct llist_node *next = node->next;
7506 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7507 __io_file_put_work(ref_node);
7512 static void io_file_data_ref_zero(struct percpu_ref *ref)
7514 struct fixed_file_ref_node *ref_node;
7515 struct fixed_file_data *data;
7516 struct io_ring_ctx *ctx;
7517 bool first_add = false;
7520 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7521 data = ref_node->file_data;
7524 spin_lock_bh(&data->lock);
7525 ref_node->done = true;
7527 while (!list_empty(&data->ref_list)) {
7528 ref_node = list_first_entry(&data->ref_list,
7529 struct fixed_file_ref_node, node);
7530 /* recycle ref nodes in order */
7531 if (!ref_node->done)
7533 list_del(&ref_node->node);
7534 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7536 spin_unlock_bh(&data->lock);
7538 if (percpu_ref_is_dying(&data->refs))
7542 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7544 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7547 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7548 struct io_ring_ctx *ctx)
7550 struct fixed_file_ref_node *ref_node;
7552 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7556 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7561 INIT_LIST_HEAD(&ref_node->node);
7562 INIT_LIST_HEAD(&ref_node->file_list);
7563 ref_node->file_data = ctx->file_data;
7564 ref_node->done = false;
7568 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7570 percpu_ref_exit(&ref_node->refs);
7574 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7577 __s32 __user *fds = (__s32 __user *) arg;
7578 unsigned nr_tables, i;
7580 int fd, ret = -ENOMEM;
7581 struct fixed_file_ref_node *ref_node;
7582 struct fixed_file_data *file_data;
7588 if (nr_args > IORING_MAX_FIXED_FILES)
7590 if (nr_args > rlimit(RLIMIT_NOFILE))
7593 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL_ACCOUNT);
7596 file_data->ctx = ctx;
7597 init_completion(&file_data->done);
7598 INIT_LIST_HEAD(&file_data->ref_list);
7599 spin_lock_init(&file_data->lock);
7601 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7602 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7603 GFP_KERNEL_ACCOUNT);
7604 if (!file_data->table)
7607 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7608 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7611 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7613 ctx->file_data = file_data;
7615 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7616 struct fixed_file_table *table;
7619 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7623 /* allow sparse sets */
7633 * Don't allow io_uring instances to be registered. If UNIX
7634 * isn't enabled, then this causes a reference cycle and this
7635 * instance can never get freed. If UNIX is enabled we'll
7636 * handle it just fine, but there's still no point in allowing
7637 * a ring fd as it doesn't support regular read/write anyway.
7639 if (file->f_op == &io_uring_fops) {
7643 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7644 index = i & IORING_FILE_TABLE_MASK;
7645 table->files[index] = file;
7648 ret = io_sqe_files_scm(ctx);
7650 io_sqe_files_unregister(ctx);
7654 ref_node = alloc_fixed_file_ref_node(ctx);
7656 io_sqe_files_unregister(ctx);
7660 io_sqe_files_set_node(file_data, ref_node);
7663 for (i = 0; i < ctx->nr_user_files; i++) {
7664 file = io_file_from_index(ctx, i);
7668 for (i = 0; i < nr_tables; i++)
7669 kfree(file_data->table[i].files);
7670 ctx->nr_user_files = 0;
7672 percpu_ref_exit(&file_data->refs);
7674 kfree(file_data->table);
7676 ctx->file_data = NULL;
7680 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7683 #if defined(CONFIG_UNIX)
7684 struct sock *sock = ctx->ring_sock->sk;
7685 struct sk_buff_head *head = &sock->sk_receive_queue;
7686 struct sk_buff *skb;
7689 * See if we can merge this file into an existing skb SCM_RIGHTS
7690 * file set. If there's no room, fall back to allocating a new skb
7691 * and filling it in.
7693 spin_lock_irq(&head->lock);
7694 skb = skb_peek(head);
7696 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7698 if (fpl->count < SCM_MAX_FD) {
7699 __skb_unlink(skb, head);
7700 spin_unlock_irq(&head->lock);
7701 fpl->fp[fpl->count] = get_file(file);
7702 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7704 spin_lock_irq(&head->lock);
7705 __skb_queue_head(head, skb);
7710 spin_unlock_irq(&head->lock);
7717 return __io_sqe_files_scm(ctx, 1, index);
7723 static int io_queue_file_removal(struct fixed_file_data *data,
7726 struct io_file_put *pfile;
7727 struct fixed_file_ref_node *ref_node = data->node;
7729 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7734 list_add(&pfile->list, &ref_node->file_list);
7739 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7740 struct io_uring_files_update *up,
7743 struct fixed_file_data *data = ctx->file_data;
7744 struct fixed_file_ref_node *ref_node;
7749 bool needs_switch = false;
7751 if (check_add_overflow(up->offset, nr_args, &done))
7753 if (done > ctx->nr_user_files)
7756 ref_node = alloc_fixed_file_ref_node(ctx);
7761 fds = u64_to_user_ptr(up->fds);
7763 struct fixed_file_table *table;
7767 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7771 i = array_index_nospec(up->offset, ctx->nr_user_files);
7772 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7773 index = i & IORING_FILE_TABLE_MASK;
7774 if (table->files[index]) {
7775 file = table->files[index];
7776 err = io_queue_file_removal(data, file);
7779 table->files[index] = NULL;
7780 needs_switch = true;
7789 * Don't allow io_uring instances to be registered. If
7790 * UNIX isn't enabled, then this causes a reference
7791 * cycle and this instance can never get freed. If UNIX
7792 * is enabled we'll handle it just fine, but there's
7793 * still no point in allowing a ring fd as it doesn't
7794 * support regular read/write anyway.
7796 if (file->f_op == &io_uring_fops) {
7801 table->files[index] = file;
7802 err = io_sqe_file_register(ctx, file, i);
7804 table->files[index] = NULL;
7815 percpu_ref_kill(&data->node->refs);
7816 io_sqe_files_set_node(data, ref_node);
7818 destroy_fixed_file_ref_node(ref_node);
7820 return done ? done : err;
7823 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7826 struct io_uring_files_update up;
7828 if (!ctx->file_data)
7832 if (copy_from_user(&up, arg, sizeof(up)))
7837 return __io_sqe_files_update(ctx, &up, nr_args);
7840 static void io_free_work(struct io_wq_work *work)
7842 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7844 /* Consider that io_steal_work() relies on this ref */
7848 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7849 struct io_uring_params *p)
7851 struct io_wq_data data;
7853 struct io_ring_ctx *ctx_attach;
7854 unsigned int concurrency;
7857 data.user = ctx->user;
7858 data.free_work = io_free_work;
7859 data.do_work = io_wq_submit_work;
7861 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7862 /* Do QD, or 4 * CPUS, whatever is smallest */
7863 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7865 ctx->io_wq = io_wq_create(concurrency, &data);
7866 if (IS_ERR(ctx->io_wq)) {
7867 ret = PTR_ERR(ctx->io_wq);
7873 f = fdget(p->wq_fd);
7877 if (f.file->f_op != &io_uring_fops) {
7882 ctx_attach = f.file->private_data;
7883 /* @io_wq is protected by holding the fd */
7884 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7889 ctx->io_wq = ctx_attach->io_wq;
7895 static int io_uring_alloc_task_context(struct task_struct *task)
7897 struct io_uring_task *tctx;
7900 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7901 if (unlikely(!tctx))
7904 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7905 if (unlikely(ret)) {
7911 init_waitqueue_head(&tctx->wait);
7913 atomic_set(&tctx->in_idle, 0);
7914 tctx->sqpoll = false;
7915 io_init_identity(&tctx->__identity);
7916 tctx->identity = &tctx->__identity;
7917 task->io_uring = tctx;
7921 void __io_uring_free(struct task_struct *tsk)
7923 struct io_uring_task *tctx = tsk->io_uring;
7925 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7926 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7927 if (tctx->identity != &tctx->__identity)
7928 kfree(tctx->identity);
7929 percpu_counter_destroy(&tctx->inflight);
7931 tsk->io_uring = NULL;
7934 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7935 struct io_uring_params *p)
7939 if (ctx->flags & IORING_SETUP_SQPOLL) {
7940 struct io_sq_data *sqd;
7943 if (!capable(CAP_SYS_ADMIN))
7946 sqd = io_get_sq_data(p);
7953 io_sq_thread_park(sqd);
7954 mutex_lock(&sqd->ctx_lock);
7955 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7956 mutex_unlock(&sqd->ctx_lock);
7957 io_sq_thread_unpark(sqd);
7959 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7960 if (!ctx->sq_thread_idle)
7961 ctx->sq_thread_idle = HZ;
7966 if (p->flags & IORING_SETUP_SQ_AFF) {
7967 int cpu = p->sq_thread_cpu;
7970 if (cpu >= nr_cpu_ids)
7972 if (!cpu_online(cpu))
7975 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7976 cpu, "io_uring-sq");
7978 sqd->thread = kthread_create(io_sq_thread, sqd,
7981 if (IS_ERR(sqd->thread)) {
7982 ret = PTR_ERR(sqd->thread);
7986 ret = io_uring_alloc_task_context(sqd->thread);
7989 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7990 /* Can't have SQ_AFF without SQPOLL */
7996 ret = io_init_wq_offload(ctx, p);
8002 io_finish_async(ctx);
8006 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8008 struct io_sq_data *sqd = ctx->sq_data;
8010 ctx->flags &= ~IORING_SETUP_R_DISABLED;
8011 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd && sqd->thread)
8012 wake_up_process(sqd->thread);
8015 static inline void __io_unaccount_mem(struct user_struct *user,
8016 unsigned long nr_pages)
8018 atomic_long_sub(nr_pages, &user->locked_vm);
8021 static inline int __io_account_mem(struct user_struct *user,
8022 unsigned long nr_pages)
8024 unsigned long page_limit, cur_pages, new_pages;
8026 /* Don't allow more pages than we can safely lock */
8027 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8030 cur_pages = atomic_long_read(&user->locked_vm);
8031 new_pages = cur_pages + nr_pages;
8032 if (new_pages > page_limit)
8034 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8035 new_pages) != cur_pages);
8040 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8041 enum io_mem_account acct)
8044 __io_unaccount_mem(ctx->user, nr_pages);
8046 if (ctx->mm_account) {
8047 if (acct == ACCT_LOCKED)
8048 ctx->mm_account->locked_vm -= nr_pages;
8049 else if (acct == ACCT_PINNED)
8050 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8054 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8055 enum io_mem_account acct)
8059 if (ctx->limit_mem) {
8060 ret = __io_account_mem(ctx->user, nr_pages);
8065 if (ctx->mm_account) {
8066 if (acct == ACCT_LOCKED)
8067 ctx->mm_account->locked_vm += nr_pages;
8068 else if (acct == ACCT_PINNED)
8069 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8075 static void io_mem_free(void *ptr)
8082 page = virt_to_head_page(ptr);
8083 if (put_page_testzero(page))
8084 free_compound_page(page);
8087 static void *io_mem_alloc(size_t size)
8089 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8092 return (void *) __get_free_pages(gfp_flags, get_order(size));
8095 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8098 struct io_rings *rings;
8099 size_t off, sq_array_size;
8101 off = struct_size(rings, cqes, cq_entries);
8102 if (off == SIZE_MAX)
8106 off = ALIGN(off, SMP_CACHE_BYTES);
8114 sq_array_size = array_size(sizeof(u32), sq_entries);
8115 if (sq_array_size == SIZE_MAX)
8118 if (check_add_overflow(off, sq_array_size, &off))
8124 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8128 pages = (size_t)1 << get_order(
8129 rings_size(sq_entries, cq_entries, NULL));
8130 pages += (size_t)1 << get_order(
8131 array_size(sizeof(struct io_uring_sqe), sq_entries));
8136 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8140 if (!ctx->user_bufs)
8143 for (i = 0; i < ctx->nr_user_bufs; i++) {
8144 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8146 for (j = 0; j < imu->nr_bvecs; j++)
8147 unpin_user_page(imu->bvec[j].bv_page);
8149 if (imu->acct_pages)
8150 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8155 kfree(ctx->user_bufs);
8156 ctx->user_bufs = NULL;
8157 ctx->nr_user_bufs = 0;
8161 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8162 void __user *arg, unsigned index)
8164 struct iovec __user *src;
8166 #ifdef CONFIG_COMPAT
8168 struct compat_iovec __user *ciovs;
8169 struct compat_iovec ciov;
8171 ciovs = (struct compat_iovec __user *) arg;
8172 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8175 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8176 dst->iov_len = ciov.iov_len;
8180 src = (struct iovec __user *) arg;
8181 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8187 * Not super efficient, but this is just a registration time. And we do cache
8188 * the last compound head, so generally we'll only do a full search if we don't
8191 * We check if the given compound head page has already been accounted, to
8192 * avoid double accounting it. This allows us to account the full size of the
8193 * page, not just the constituent pages of a huge page.
8195 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8196 int nr_pages, struct page *hpage)
8200 /* check current page array */
8201 for (i = 0; i < nr_pages; i++) {
8202 if (!PageCompound(pages[i]))
8204 if (compound_head(pages[i]) == hpage)
8208 /* check previously registered pages */
8209 for (i = 0; i < ctx->nr_user_bufs; i++) {
8210 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8212 for (j = 0; j < imu->nr_bvecs; j++) {
8213 if (!PageCompound(imu->bvec[j].bv_page))
8215 if (compound_head(imu->bvec[j].bv_page) == hpage)
8223 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8224 int nr_pages, struct io_mapped_ubuf *imu,
8225 struct page **last_hpage)
8229 for (i = 0; i < nr_pages; i++) {
8230 if (!PageCompound(pages[i])) {
8235 hpage = compound_head(pages[i]);
8236 if (hpage == *last_hpage)
8238 *last_hpage = hpage;
8239 if (headpage_already_acct(ctx, pages, i, hpage))
8241 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8245 if (!imu->acct_pages)
8248 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8250 imu->acct_pages = 0;
8254 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8257 struct vm_area_struct **vmas = NULL;
8258 struct page **pages = NULL;
8259 struct page *last_hpage = NULL;
8260 int i, j, got_pages = 0;
8265 if (!nr_args || nr_args > UIO_MAXIOV)
8268 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8270 if (!ctx->user_bufs)
8273 for (i = 0; i < nr_args; i++) {
8274 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8275 unsigned long off, start, end, ubuf;
8280 ret = io_copy_iov(ctx, &iov, arg, i);
8285 * Don't impose further limits on the size and buffer
8286 * constraints here, we'll -EINVAL later when IO is
8287 * submitted if they are wrong.
8290 if (!iov.iov_base || !iov.iov_len)
8293 /* arbitrary limit, but we need something */
8294 if (iov.iov_len > SZ_1G)
8297 ubuf = (unsigned long) iov.iov_base;
8298 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8299 start = ubuf >> PAGE_SHIFT;
8300 nr_pages = end - start;
8303 if (!pages || nr_pages > got_pages) {
8306 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8308 vmas = kvmalloc_array(nr_pages,
8309 sizeof(struct vm_area_struct *),
8311 if (!pages || !vmas) {
8315 got_pages = nr_pages;
8318 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8325 mmap_read_lock(current->mm);
8326 pret = pin_user_pages(ubuf, nr_pages,
8327 FOLL_WRITE | FOLL_LONGTERM,
8329 if (pret == nr_pages) {
8330 /* don't support file backed memory */
8331 for (j = 0; j < nr_pages; j++) {
8332 struct vm_area_struct *vma = vmas[j];
8335 !is_file_hugepages(vma->vm_file)) {
8341 ret = pret < 0 ? pret : -EFAULT;
8343 mmap_read_unlock(current->mm);
8346 * if we did partial map, or found file backed vmas,
8347 * release any pages we did get
8350 unpin_user_pages(pages, pret);
8355 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8357 unpin_user_pages(pages, pret);
8362 off = ubuf & ~PAGE_MASK;
8364 for (j = 0; j < nr_pages; j++) {
8367 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8368 imu->bvec[j].bv_page = pages[j];
8369 imu->bvec[j].bv_len = vec_len;
8370 imu->bvec[j].bv_offset = off;
8374 /* store original address for later verification */
8376 imu->len = iov.iov_len;
8377 imu->nr_bvecs = nr_pages;
8379 ctx->nr_user_bufs++;
8387 io_sqe_buffer_unregister(ctx);
8391 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8393 __s32 __user *fds = arg;
8399 if (copy_from_user(&fd, fds, sizeof(*fds)))
8402 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8403 if (IS_ERR(ctx->cq_ev_fd)) {
8404 int ret = PTR_ERR(ctx->cq_ev_fd);
8405 ctx->cq_ev_fd = NULL;
8412 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8414 if (ctx->cq_ev_fd) {
8415 eventfd_ctx_put(ctx->cq_ev_fd);
8416 ctx->cq_ev_fd = NULL;
8423 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8425 struct io_buffer *buf;
8426 unsigned long index;
8428 xa_for_each(&ctx->io_buffers, index, buf)
8429 __io_remove_buffers(ctx, buf, index, -1U);
8432 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8434 io_finish_async(ctx);
8435 io_sqe_buffer_unregister(ctx);
8437 if (ctx->sqo_task) {
8438 put_task_struct(ctx->sqo_task);
8439 ctx->sqo_task = NULL;
8442 #ifdef CONFIG_BLK_CGROUP
8443 if (ctx->sqo_blkcg_css)
8444 css_put(ctx->sqo_blkcg_css);
8447 io_sqe_files_unregister(ctx);
8448 io_eventfd_unregister(ctx);
8449 io_destroy_buffers(ctx);
8451 #if defined(CONFIG_UNIX)
8452 if (ctx->ring_sock) {
8453 ctx->ring_sock->file = NULL; /* so that iput() is called */
8454 sock_release(ctx->ring_sock);
8458 if (ctx->mm_account) {
8459 mmdrop(ctx->mm_account);
8460 ctx->mm_account = NULL;
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();
9759 mutex_lock(&ctx->uring_lock);
9762 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
9767 if (ctx->restricted) {
9768 if (opcode >= IORING_REGISTER_LAST) {
9773 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9780 case IORING_REGISTER_BUFFERS:
9781 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9783 case IORING_UNREGISTER_BUFFERS:
9787 ret = io_sqe_buffer_unregister(ctx);
9789 case IORING_REGISTER_FILES:
9790 ret = io_sqe_files_register(ctx, arg, nr_args);
9792 case IORING_UNREGISTER_FILES:
9796 ret = io_sqe_files_unregister(ctx);
9798 case IORING_REGISTER_FILES_UPDATE:
9799 ret = io_sqe_files_update(ctx, arg, nr_args);
9801 case IORING_REGISTER_EVENTFD:
9802 case IORING_REGISTER_EVENTFD_ASYNC:
9806 ret = io_eventfd_register(ctx, arg);
9809 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9810 ctx->eventfd_async = 1;
9812 ctx->eventfd_async = 0;
9814 case IORING_UNREGISTER_EVENTFD:
9818 ret = io_eventfd_unregister(ctx);
9820 case IORING_REGISTER_PROBE:
9822 if (!arg || nr_args > 256)
9824 ret = io_probe(ctx, arg, nr_args);
9826 case IORING_REGISTER_PERSONALITY:
9830 ret = io_register_personality(ctx);
9832 case IORING_UNREGISTER_PERSONALITY:
9836 ret = io_unregister_personality(ctx, nr_args);
9838 case IORING_REGISTER_ENABLE_RINGS:
9842 ret = io_register_enable_rings(ctx);
9844 case IORING_REGISTER_RESTRICTIONS:
9845 ret = io_register_restrictions(ctx, arg, nr_args);
9853 if (io_register_op_must_quiesce(opcode)) {
9854 /* bring the ctx back to life */
9855 percpu_ref_reinit(&ctx->refs);
9856 reinit_completion(&ctx->ref_comp);
9861 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9862 void __user *, arg, unsigned int, nr_args)
9864 struct io_ring_ctx *ctx;
9873 if (f.file->f_op != &io_uring_fops)
9876 ctx = f.file->private_data;
9878 mutex_lock(&ctx->uring_lock);
9879 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9880 mutex_unlock(&ctx->uring_lock);
9881 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9882 ctx->cq_ev_fd != NULL, ret);
9888 static int __init io_uring_init(void)
9890 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9891 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9892 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9895 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9896 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9897 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9898 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9899 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9900 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9901 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9902 BUILD_BUG_SQE_ELEM(8, __u64, off);
9903 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9904 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9905 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9906 BUILD_BUG_SQE_ELEM(24, __u32, len);
9907 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9908 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9909 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9910 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9911 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9912 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9913 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9914 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9915 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9916 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9917 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9918 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9919 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9920 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9921 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9922 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9923 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9924 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9925 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9927 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9928 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9929 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9932 __initcall(io_uring_init);