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;
265 * Ring buffer of indices into array of io_uring_sqe, which is
266 * mmapped by the application using the IORING_OFF_SQES offset.
268 * This indirection could e.g. be used to assign fixed
269 * io_uring_sqe entries to operations and only submit them to
270 * the queue when needed.
272 * The kernel modifies neither the indices array nor the entries
276 unsigned cached_sq_head;
279 unsigned sq_thread_idle;
280 unsigned cached_sq_dropped;
281 unsigned cached_cq_overflow;
282 unsigned long sq_check_overflow;
284 struct list_head defer_list;
285 struct list_head timeout_list;
286 struct list_head cq_overflow_list;
288 wait_queue_head_t inflight_wait;
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 idr io_buffer_idr;
349 struct idr personality_idr;
352 unsigned cached_cq_tail;
355 atomic_t cq_timeouts;
356 unsigned long cq_check_overflow;
357 struct wait_queue_head cq_wait;
358 struct fasync_struct *cq_fasync;
359 struct eventfd_ctx *cq_ev_fd;
360 } ____cacheline_aligned_in_smp;
363 struct mutex uring_lock;
364 wait_queue_head_t wait;
365 } ____cacheline_aligned_in_smp;
368 spinlock_t completion_lock;
371 * ->iopoll_list is protected by the ctx->uring_lock for
372 * io_uring instances that don't use IORING_SETUP_SQPOLL.
373 * For SQPOLL, only the single threaded io_sq_thread() will
374 * manipulate the list, hence no extra locking is needed there.
376 struct list_head iopoll_list;
377 struct hlist_head *cancel_hash;
378 unsigned cancel_hash_bits;
379 bool poll_multi_file;
381 spinlock_t inflight_lock;
382 struct list_head inflight_list;
383 } ____cacheline_aligned_in_smp;
385 struct delayed_work file_put_work;
386 struct llist_head file_put_llist;
388 struct work_struct exit_work;
389 struct io_restriction restrictions;
393 * First field must be the file pointer in all the
394 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
396 struct io_poll_iocb {
398 struct wait_queue_head *head;
402 struct wait_queue_entry wait;
405 struct io_poll_remove {
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;
555 struct filename *oldpath;
556 struct filename *newpath;
564 struct filename *filename;
567 struct io_completion {
569 struct list_head list;
573 struct io_async_connect {
574 struct sockaddr_storage address;
577 struct io_async_msghdr {
578 struct iovec fast_iov[UIO_FASTIOV];
580 struct sockaddr __user *uaddr;
582 struct sockaddr_storage addr;
586 struct iovec fast_iov[UIO_FASTIOV];
587 const struct iovec *free_iovec;
588 struct iov_iter iter;
590 struct wait_page_queue wpq;
594 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
595 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
596 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
597 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
598 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
599 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
606 REQ_F_LINK_TIMEOUT_BIT,
608 REQ_F_NEED_CLEANUP_BIT,
610 REQ_F_BUFFER_SELECTED_BIT,
611 REQ_F_NO_FILE_TABLE_BIT,
612 REQ_F_WORK_INITIALIZED_BIT,
613 REQ_F_LTIMEOUT_ACTIVE_BIT,
615 /* not a real bit, just to check we're not overflowing the space */
621 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
622 /* drain existing IO first */
623 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
625 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
626 /* doesn't sever on completion < 0 */
627 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
629 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
630 /* IOSQE_BUFFER_SELECT */
631 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
634 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
635 /* fail rest of links */
636 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
637 /* on inflight list */
638 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
639 /* read/write uses file position */
640 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
641 /* must not punt to workers */
642 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
643 /* has or had linked timeout */
644 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
646 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
648 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
649 /* already went through poll handler */
650 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
651 /* buffer already selected */
652 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
653 /* doesn't need file table for this request */
654 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
655 /* io_wq_work is initialized */
656 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
657 /* linked timeout is active, i.e. prepared by link's head */
658 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
662 struct io_poll_iocb poll;
663 struct io_poll_iocb *double_poll;
667 * NOTE! Each of the iocb union members has the file pointer
668 * as the first entry in their struct definition. So you can
669 * access the file pointer through any of the sub-structs,
670 * or directly as just 'ki_filp' in this struct.
676 struct io_poll_iocb poll;
677 struct io_poll_remove poll_remove;
678 struct io_accept accept;
680 struct io_cancel cancel;
681 struct io_timeout timeout;
682 struct io_timeout_rem timeout_rem;
683 struct io_connect connect;
684 struct io_sr_msg sr_msg;
686 struct io_close close;
687 struct io_files_update files_update;
688 struct io_fadvise fadvise;
689 struct io_madvise madvise;
690 struct io_epoll epoll;
691 struct io_splice splice;
692 struct io_provide_buf pbuf;
693 struct io_statx statx;
694 struct io_shutdown shutdown;
695 struct io_rename rename;
696 struct io_unlink unlink;
697 /* use only after cleaning per-op data, see io_clean_op() */
698 struct io_completion compl;
701 /* opcode allocated if it needs to store data for async defer */
704 /* polled IO has completed */
710 struct io_ring_ctx *ctx;
713 struct task_struct *task;
716 struct list_head link_list;
719 * 1. used with ctx->iopoll_list with reads/writes
720 * 2. to track reqs with ->files (see io_op_def::file_table)
722 struct list_head inflight_entry;
724 struct percpu_ref *fixed_file_refs;
725 struct callback_head task_work;
726 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
727 struct hlist_node hash_node;
728 struct async_poll *apoll;
729 struct io_wq_work work;
732 struct io_defer_entry {
733 struct list_head list;
734 struct io_kiocb *req;
738 #define IO_IOPOLL_BATCH 8
740 struct io_comp_state {
742 struct list_head list;
743 struct io_ring_ctx *ctx;
746 struct io_submit_state {
747 struct blk_plug plug;
750 * io_kiocb alloc cache
752 void *reqs[IO_IOPOLL_BATCH];
753 unsigned int free_reqs;
756 * Batch completion logic
758 struct io_comp_state comp;
761 * File reference cache
765 unsigned int has_refs;
766 unsigned int ios_left;
770 /* needs req->file assigned */
771 unsigned needs_file : 1;
772 /* don't fail if file grab fails */
773 unsigned needs_file_no_error : 1;
774 /* hash wq insertion if file is a regular file */
775 unsigned hash_reg_file : 1;
776 /* unbound wq insertion if file is a non-regular file */
777 unsigned unbound_nonreg_file : 1;
778 /* opcode is not supported by this kernel */
779 unsigned not_supported : 1;
780 /* set if opcode supports polled "wait" */
782 unsigned pollout : 1;
783 /* op supports buffer selection */
784 unsigned buffer_select : 1;
785 /* must always have async data allocated */
786 unsigned needs_async_data : 1;
787 /* size of async data needed, if any */
788 unsigned short async_size;
792 static const struct io_op_def io_op_defs[] = {
793 [IORING_OP_NOP] = {},
794 [IORING_OP_READV] = {
796 .unbound_nonreg_file = 1,
799 .needs_async_data = 1,
800 .async_size = sizeof(struct io_async_rw),
801 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
803 [IORING_OP_WRITEV] = {
806 .unbound_nonreg_file = 1,
808 .needs_async_data = 1,
809 .async_size = sizeof(struct io_async_rw),
810 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
813 [IORING_OP_FSYNC] = {
815 .work_flags = IO_WQ_WORK_BLKCG,
817 [IORING_OP_READ_FIXED] = {
819 .unbound_nonreg_file = 1,
821 .async_size = sizeof(struct io_async_rw),
822 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
824 [IORING_OP_WRITE_FIXED] = {
827 .unbound_nonreg_file = 1,
829 .async_size = sizeof(struct io_async_rw),
830 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
833 [IORING_OP_POLL_ADD] = {
835 .unbound_nonreg_file = 1,
837 [IORING_OP_POLL_REMOVE] = {},
838 [IORING_OP_SYNC_FILE_RANGE] = {
840 .work_flags = IO_WQ_WORK_BLKCG,
842 [IORING_OP_SENDMSG] = {
844 .unbound_nonreg_file = 1,
846 .needs_async_data = 1,
847 .async_size = sizeof(struct io_async_msghdr),
848 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
851 [IORING_OP_RECVMSG] = {
853 .unbound_nonreg_file = 1,
856 .needs_async_data = 1,
857 .async_size = sizeof(struct io_async_msghdr),
858 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
861 [IORING_OP_TIMEOUT] = {
862 .needs_async_data = 1,
863 .async_size = sizeof(struct io_timeout_data),
864 .work_flags = IO_WQ_WORK_MM,
866 [IORING_OP_TIMEOUT_REMOVE] = {},
867 [IORING_OP_ACCEPT] = {
869 .unbound_nonreg_file = 1,
871 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
873 [IORING_OP_ASYNC_CANCEL] = {},
874 [IORING_OP_LINK_TIMEOUT] = {
875 .needs_async_data = 1,
876 .async_size = sizeof(struct io_timeout_data),
877 .work_flags = IO_WQ_WORK_MM,
879 [IORING_OP_CONNECT] = {
881 .unbound_nonreg_file = 1,
883 .needs_async_data = 1,
884 .async_size = sizeof(struct io_async_connect),
885 .work_flags = IO_WQ_WORK_MM,
887 [IORING_OP_FALLOCATE] = {
889 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
891 [IORING_OP_OPENAT] = {
892 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
893 IO_WQ_WORK_FS | IO_WQ_WORK_MM,
895 [IORING_OP_CLOSE] = {
897 .needs_file_no_error = 1,
898 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
900 [IORING_OP_FILES_UPDATE] = {
901 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
903 [IORING_OP_STATX] = {
904 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
905 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
909 .unbound_nonreg_file = 1,
912 .async_size = sizeof(struct io_async_rw),
913 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
915 [IORING_OP_WRITE] = {
917 .unbound_nonreg_file = 1,
919 .async_size = sizeof(struct io_async_rw),
920 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
923 [IORING_OP_FADVISE] = {
925 .work_flags = IO_WQ_WORK_BLKCG,
927 [IORING_OP_MADVISE] = {
928 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
932 .unbound_nonreg_file = 1,
934 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
938 .unbound_nonreg_file = 1,
941 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
943 [IORING_OP_OPENAT2] = {
944 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
945 IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
947 [IORING_OP_EPOLL_CTL] = {
948 .unbound_nonreg_file = 1,
949 .work_flags = IO_WQ_WORK_FILES,
951 [IORING_OP_SPLICE] = {
954 .unbound_nonreg_file = 1,
955 .work_flags = IO_WQ_WORK_BLKCG,
957 [IORING_OP_PROVIDE_BUFFERS] = {},
958 [IORING_OP_REMOVE_BUFFERS] = {},
962 .unbound_nonreg_file = 1,
964 [IORING_OP_SHUTDOWN] = {
967 [IORING_OP_RENAMEAT] = {
968 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
969 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
971 [IORING_OP_UNLINKAT] = {
972 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
973 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
977 enum io_mem_account {
982 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
983 struct io_comp_state *cs);
984 static void io_cqring_fill_event(struct io_kiocb *req, long res);
985 static void io_put_req(struct io_kiocb *req);
986 static void io_put_req_deferred(struct io_kiocb *req, int nr);
987 static void io_double_put_req(struct io_kiocb *req);
988 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
989 static void __io_queue_linked_timeout(struct io_kiocb *req);
990 static void io_queue_linked_timeout(struct io_kiocb *req);
991 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
992 struct io_uring_files_update *ip,
994 static void __io_clean_op(struct io_kiocb *req);
995 static struct file *io_file_get(struct io_submit_state *state,
996 struct io_kiocb *req, int fd, bool fixed);
997 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
998 static void io_file_put_work(struct work_struct *work);
1000 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1001 struct iovec **iovec, struct iov_iter *iter,
1003 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1004 const struct iovec *fast_iov,
1005 struct iov_iter *iter, bool force);
1007 static struct kmem_cache *req_cachep;
1009 static const struct file_operations io_uring_fops;
1011 struct sock *io_uring_get_socket(struct file *file)
1013 #if defined(CONFIG_UNIX)
1014 if (file->f_op == &io_uring_fops) {
1015 struct io_ring_ctx *ctx = file->private_data;
1017 return ctx->ring_sock->sk;
1022 EXPORT_SYMBOL(io_uring_get_socket);
1024 static inline void io_clean_op(struct io_kiocb *req)
1026 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
1031 static void io_sq_thread_drop_mm_files(void)
1033 struct files_struct *files = current->files;
1034 struct mm_struct *mm = current->mm;
1037 kthread_unuse_mm(mm);
1042 struct nsproxy *nsproxy = current->nsproxy;
1045 current->files = NULL;
1046 current->nsproxy = NULL;
1047 task_unlock(current);
1048 put_files_struct(files);
1049 put_nsproxy(nsproxy);
1053 static void __io_sq_thread_acquire_files(struct io_ring_ctx *ctx)
1055 if (!current->files) {
1056 struct files_struct *files;
1057 struct nsproxy *nsproxy;
1059 task_lock(ctx->sqo_task);
1060 files = ctx->sqo_task->files;
1062 task_unlock(ctx->sqo_task);
1065 atomic_inc(&files->count);
1066 get_nsproxy(ctx->sqo_task->nsproxy);
1067 nsproxy = ctx->sqo_task->nsproxy;
1068 task_unlock(ctx->sqo_task);
1071 current->files = files;
1072 current->nsproxy = nsproxy;
1073 task_unlock(current);
1077 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1079 struct mm_struct *mm;
1084 /* Should never happen */
1085 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL)))
1088 task_lock(ctx->sqo_task);
1089 mm = ctx->sqo_task->mm;
1090 if (unlikely(!mm || !mmget_not_zero(mm)))
1092 task_unlock(ctx->sqo_task);
1102 static int io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx,
1103 struct io_kiocb *req)
1105 const struct io_op_def *def = &io_op_defs[req->opcode];
1107 if (def->work_flags & IO_WQ_WORK_MM) {
1108 int ret = __io_sq_thread_acquire_mm(ctx);
1113 if (def->needs_file || (def->work_flags & IO_WQ_WORK_FILES))
1114 __io_sq_thread_acquire_files(ctx);
1119 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1120 struct cgroup_subsys_state **cur_css)
1123 #ifdef CONFIG_BLK_CGROUP
1124 /* puts the old one when swapping */
1125 if (*cur_css != ctx->sqo_blkcg_css) {
1126 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1127 *cur_css = ctx->sqo_blkcg_css;
1132 static void io_sq_thread_unassociate_blkcg(void)
1134 #ifdef CONFIG_BLK_CGROUP
1135 kthread_associate_blkcg(NULL);
1139 static inline void req_set_fail_links(struct io_kiocb *req)
1141 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1142 req->flags |= REQ_F_FAIL_LINK;
1146 * None of these are dereferenced, they are simply used to check if any of
1147 * them have changed. If we're under current and check they are still the
1148 * same, we're fine to grab references to them for actual out-of-line use.
1150 static void io_init_identity(struct io_identity *id)
1152 id->files = current->files;
1153 id->mm = current->mm;
1154 #ifdef CONFIG_BLK_CGROUP
1156 id->blkcg_css = blkcg_css();
1159 id->creds = current_cred();
1160 id->nsproxy = current->nsproxy;
1161 id->fs = current->fs;
1162 id->fsize = rlimit(RLIMIT_FSIZE);
1164 id->loginuid = current->loginuid;
1165 id->sessionid = current->sessionid;
1167 refcount_set(&id->count, 1);
1170 static inline void __io_req_init_async(struct io_kiocb *req)
1172 memset(&req->work, 0, sizeof(req->work));
1173 req->flags |= REQ_F_WORK_INITIALIZED;
1177 * Note: must call io_req_init_async() for the first time you
1178 * touch any members of io_wq_work.
1180 static inline void io_req_init_async(struct io_kiocb *req)
1182 struct io_uring_task *tctx = current->io_uring;
1184 if (req->flags & REQ_F_WORK_INITIALIZED)
1187 __io_req_init_async(req);
1189 /* Grab a ref if this isn't our static identity */
1190 req->work.identity = tctx->identity;
1191 if (tctx->identity != &tctx->__identity)
1192 refcount_inc(&req->work.identity->count);
1195 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1197 return ctx->flags & IORING_SETUP_SQPOLL;
1200 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1202 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1204 complete(&ctx->ref_comp);
1207 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1209 return !req->timeout.off;
1212 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1214 struct io_ring_ctx *ctx;
1217 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1221 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1222 if (!ctx->fallback_req)
1226 * Use 5 bits less than the max cq entries, that should give us around
1227 * 32 entries per hash list if totally full and uniformly spread.
1229 hash_bits = ilog2(p->cq_entries);
1233 ctx->cancel_hash_bits = hash_bits;
1234 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1236 if (!ctx->cancel_hash)
1238 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1240 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1241 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1244 ctx->flags = p->flags;
1245 init_waitqueue_head(&ctx->sqo_sq_wait);
1246 INIT_LIST_HEAD(&ctx->sqd_list);
1247 init_waitqueue_head(&ctx->cq_wait);
1248 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1249 init_completion(&ctx->ref_comp);
1250 init_completion(&ctx->sq_thread_comp);
1251 idr_init(&ctx->io_buffer_idr);
1252 idr_init(&ctx->personality_idr);
1253 mutex_init(&ctx->uring_lock);
1254 init_waitqueue_head(&ctx->wait);
1255 spin_lock_init(&ctx->completion_lock);
1256 INIT_LIST_HEAD(&ctx->iopoll_list);
1257 INIT_LIST_HEAD(&ctx->defer_list);
1258 INIT_LIST_HEAD(&ctx->timeout_list);
1259 init_waitqueue_head(&ctx->inflight_wait);
1260 spin_lock_init(&ctx->inflight_lock);
1261 INIT_LIST_HEAD(&ctx->inflight_list);
1262 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1263 init_llist_head(&ctx->file_put_llist);
1266 if (ctx->fallback_req)
1267 kmem_cache_free(req_cachep, ctx->fallback_req);
1268 kfree(ctx->cancel_hash);
1273 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1275 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1276 struct io_ring_ctx *ctx = req->ctx;
1278 return seq != ctx->cached_cq_tail
1279 + READ_ONCE(ctx->cached_cq_overflow);
1285 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1287 struct io_rings *rings = ctx->rings;
1289 /* order cqe stores with ring update */
1290 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1292 if (wq_has_sleeper(&ctx->cq_wait)) {
1293 wake_up_interruptible(&ctx->cq_wait);
1294 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1298 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1300 if (req->work.identity == &tctx->__identity)
1302 if (refcount_dec_and_test(&req->work.identity->count))
1303 kfree(req->work.identity);
1306 static void io_req_clean_work(struct io_kiocb *req)
1308 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1311 req->flags &= ~REQ_F_WORK_INITIALIZED;
1313 if (req->work.flags & IO_WQ_WORK_MM) {
1314 mmdrop(req->work.identity->mm);
1315 req->work.flags &= ~IO_WQ_WORK_MM;
1317 #ifdef CONFIG_BLK_CGROUP
1318 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1319 css_put(req->work.identity->blkcg_css);
1320 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1323 if (req->work.flags & IO_WQ_WORK_CREDS) {
1324 put_cred(req->work.identity->creds);
1325 req->work.flags &= ~IO_WQ_WORK_CREDS;
1327 if (req->work.flags & IO_WQ_WORK_FS) {
1328 struct fs_struct *fs = req->work.identity->fs;
1330 spin_lock(&req->work.identity->fs->lock);
1333 spin_unlock(&req->work.identity->fs->lock);
1336 req->work.flags &= ~IO_WQ_WORK_FS;
1339 io_put_identity(req->task->io_uring, req);
1343 * Create a private copy of io_identity, since some fields don't match
1344 * the current context.
1346 static bool io_identity_cow(struct io_kiocb *req)
1348 struct io_uring_task *tctx = current->io_uring;
1349 const struct cred *creds = NULL;
1350 struct io_identity *id;
1352 if (req->work.flags & IO_WQ_WORK_CREDS)
1353 creds = req->work.identity->creds;
1355 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1356 if (unlikely(!id)) {
1357 req->work.flags |= IO_WQ_WORK_CANCEL;
1362 * We can safely just re-init the creds we copied Either the field
1363 * matches the current one, or we haven't grabbed it yet. The only
1364 * exception is ->creds, through registered personalities, so handle
1365 * that one separately.
1367 io_init_identity(id);
1369 req->work.identity->creds = creds;
1371 /* add one for this request */
1372 refcount_inc(&id->count);
1374 /* drop tctx and req identity references, if needed */
1375 if (tctx->identity != &tctx->__identity &&
1376 refcount_dec_and_test(&tctx->identity->count))
1377 kfree(tctx->identity);
1378 if (req->work.identity != &tctx->__identity &&
1379 refcount_dec_and_test(&req->work.identity->count))
1380 kfree(req->work.identity);
1382 req->work.identity = id;
1383 tctx->identity = id;
1387 static bool io_grab_identity(struct io_kiocb *req)
1389 const struct io_op_def *def = &io_op_defs[req->opcode];
1390 struct io_identity *id = req->work.identity;
1391 struct io_ring_ctx *ctx = req->ctx;
1393 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1394 if (id->fsize != rlimit(RLIMIT_FSIZE))
1396 req->work.flags |= IO_WQ_WORK_FSIZE;
1399 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1400 (def->work_flags & IO_WQ_WORK_FILES) &&
1401 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1402 if (id->files != current->files ||
1403 id->nsproxy != current->nsproxy)
1405 atomic_inc(&id->files->count);
1406 get_nsproxy(id->nsproxy);
1407 req->flags |= REQ_F_INFLIGHT;
1409 spin_lock_irq(&ctx->inflight_lock);
1410 list_add(&req->inflight_entry, &ctx->inflight_list);
1411 spin_unlock_irq(&ctx->inflight_lock);
1412 req->work.flags |= IO_WQ_WORK_FILES;
1414 #ifdef CONFIG_BLK_CGROUP
1415 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1416 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1418 if (id->blkcg_css != blkcg_css()) {
1423 * This should be rare, either the cgroup is dying or the task
1424 * is moving cgroups. Just punt to root for the handful of ios.
1426 if (css_tryget_online(id->blkcg_css))
1427 req->work.flags |= IO_WQ_WORK_BLKCG;
1431 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1432 if (id->creds != current_cred())
1434 get_cred(id->creds);
1435 req->work.flags |= IO_WQ_WORK_CREDS;
1438 if (!uid_eq(current->loginuid, id->loginuid) ||
1439 current->sessionid != id->sessionid)
1442 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1443 (def->work_flags & IO_WQ_WORK_FS)) {
1444 if (current->fs != id->fs)
1446 spin_lock(&id->fs->lock);
1447 if (!id->fs->in_exec) {
1449 req->work.flags |= IO_WQ_WORK_FS;
1451 req->work.flags |= IO_WQ_WORK_CANCEL;
1453 spin_unlock(¤t->fs->lock);
1459 static void io_prep_async_work(struct io_kiocb *req)
1461 const struct io_op_def *def = &io_op_defs[req->opcode];
1462 struct io_ring_ctx *ctx = req->ctx;
1463 struct io_identity *id;
1465 io_req_init_async(req);
1466 id = req->work.identity;
1468 if (req->flags & REQ_F_FORCE_ASYNC)
1469 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1471 if (req->flags & REQ_F_ISREG) {
1472 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1473 io_wq_hash_work(&req->work, file_inode(req->file));
1475 if (def->unbound_nonreg_file)
1476 req->work.flags |= IO_WQ_WORK_UNBOUND;
1479 /* ->mm can never change on us */
1480 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1481 (def->work_flags & IO_WQ_WORK_MM)) {
1483 req->work.flags |= IO_WQ_WORK_MM;
1486 /* if we fail grabbing identity, we must COW, regrab, and retry */
1487 if (io_grab_identity(req))
1490 if (!io_identity_cow(req))
1493 /* can't fail at this point */
1494 if (!io_grab_identity(req))
1498 static void io_prep_async_link(struct io_kiocb *req)
1500 struct io_kiocb *cur;
1502 io_prep_async_work(req);
1503 if (req->flags & REQ_F_LINK_HEAD)
1504 list_for_each_entry(cur, &req->link_list, link_list)
1505 io_prep_async_work(cur);
1508 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1510 struct io_ring_ctx *ctx = req->ctx;
1511 struct io_kiocb *link = io_prep_linked_timeout(req);
1513 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1514 &req->work, req->flags);
1515 io_wq_enqueue(ctx->io_wq, &req->work);
1519 static void io_queue_async_work(struct io_kiocb *req)
1521 struct io_kiocb *link;
1523 /* init ->work of the whole link before punting */
1524 io_prep_async_link(req);
1525 link = __io_queue_async_work(req);
1528 io_queue_linked_timeout(link);
1531 static void io_kill_timeout(struct io_kiocb *req)
1533 struct io_timeout_data *io = req->async_data;
1536 ret = hrtimer_try_to_cancel(&io->timer);
1538 atomic_set(&req->ctx->cq_timeouts,
1539 atomic_read(&req->ctx->cq_timeouts) + 1);
1540 list_del_init(&req->timeout.list);
1541 io_cqring_fill_event(req, 0);
1542 io_put_req_deferred(req, 1);
1546 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1548 struct io_ring_ctx *ctx = req->ctx;
1550 if (!tsk || req->task == tsk)
1552 if (ctx->flags & IORING_SETUP_SQPOLL) {
1553 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1560 * Returns true if we found and killed one or more timeouts
1562 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1564 struct io_kiocb *req, *tmp;
1567 spin_lock_irq(&ctx->completion_lock);
1568 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1569 if (io_task_match(req, tsk)) {
1570 io_kill_timeout(req);
1574 spin_unlock_irq(&ctx->completion_lock);
1575 return canceled != 0;
1578 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1581 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1582 struct io_defer_entry, list);
1583 struct io_kiocb *link;
1585 if (req_need_defer(de->req, de->seq))
1587 list_del_init(&de->list);
1588 /* punt-init is done before queueing for defer */
1589 link = __io_queue_async_work(de->req);
1591 __io_queue_linked_timeout(link);
1592 /* drop submission reference */
1593 io_put_req_deferred(link, 1);
1596 } while (!list_empty(&ctx->defer_list));
1599 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1601 while (!list_empty(&ctx->timeout_list)) {
1602 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1603 struct io_kiocb, timeout.list);
1605 if (io_is_timeout_noseq(req))
1607 if (req->timeout.target_seq != ctx->cached_cq_tail
1608 - atomic_read(&ctx->cq_timeouts))
1611 list_del_init(&req->timeout.list);
1612 io_kill_timeout(req);
1616 static void io_commit_cqring(struct io_ring_ctx *ctx)
1618 io_flush_timeouts(ctx);
1619 __io_commit_cqring(ctx);
1621 if (unlikely(!list_empty(&ctx->defer_list)))
1622 __io_queue_deferred(ctx);
1625 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1627 struct io_rings *r = ctx->rings;
1629 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1632 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1634 struct io_rings *rings = ctx->rings;
1637 tail = ctx->cached_cq_tail;
1639 * writes to the cq entry need to come after reading head; the
1640 * control dependency is enough as we're using WRITE_ONCE to
1643 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1646 ctx->cached_cq_tail++;
1647 return &rings->cqes[tail & ctx->cq_mask];
1650 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1654 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1656 if (!ctx->eventfd_async)
1658 return io_wq_current_is_worker();
1661 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1663 if (waitqueue_active(&ctx->wait))
1664 wake_up(&ctx->wait);
1665 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1666 wake_up(&ctx->sq_data->wait);
1667 if (io_should_trigger_evfd(ctx))
1668 eventfd_signal(ctx->cq_ev_fd, 1);
1671 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1673 if (list_empty(&ctx->cq_overflow_list)) {
1674 clear_bit(0, &ctx->sq_check_overflow);
1675 clear_bit(0, &ctx->cq_check_overflow);
1676 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1680 static inline bool __io_match_files(struct io_kiocb *req,
1681 struct files_struct *files)
1683 return ((req->flags & REQ_F_WORK_INITIALIZED) &&
1684 (req->work.flags & IO_WQ_WORK_FILES)) &&
1685 req->work.identity->files == files;
1688 static bool io_match_files(struct io_kiocb *req,
1689 struct files_struct *files)
1691 struct io_kiocb *link;
1695 if (__io_match_files(req, files))
1697 if (req->flags & REQ_F_LINK_HEAD) {
1698 list_for_each_entry(link, &req->link_list, link_list) {
1699 if (__io_match_files(link, files))
1706 /* Returns true if there are no backlogged entries after the flush */
1707 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1708 struct task_struct *tsk,
1709 struct files_struct *files)
1711 struct io_rings *rings = ctx->rings;
1712 struct io_kiocb *req, *tmp;
1713 struct io_uring_cqe *cqe;
1714 unsigned long flags;
1718 if (list_empty_careful(&ctx->cq_overflow_list))
1720 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1721 rings->cq_ring_entries))
1725 spin_lock_irqsave(&ctx->completion_lock, flags);
1727 /* if force is set, the ring is going away. always drop after that */
1729 ctx->cq_overflow_flushed = 1;
1732 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1733 if (tsk && req->task != tsk)
1735 if (!io_match_files(req, files))
1738 cqe = io_get_cqring(ctx);
1742 list_move(&req->compl.list, &list);
1744 WRITE_ONCE(cqe->user_data, req->user_data);
1745 WRITE_ONCE(cqe->res, req->result);
1746 WRITE_ONCE(cqe->flags, req->compl.cflags);
1748 ctx->cached_cq_overflow++;
1749 WRITE_ONCE(ctx->rings->cq_overflow,
1750 ctx->cached_cq_overflow);
1754 io_commit_cqring(ctx);
1755 io_cqring_mark_overflow(ctx);
1757 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1758 io_cqring_ev_posted(ctx);
1760 while (!list_empty(&list)) {
1761 req = list_first_entry(&list, struct io_kiocb, compl.list);
1762 list_del(&req->compl.list);
1769 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1771 struct io_ring_ctx *ctx = req->ctx;
1772 struct io_uring_cqe *cqe;
1774 trace_io_uring_complete(ctx, req->user_data, res);
1777 * If we can't get a cq entry, userspace overflowed the
1778 * submission (by quite a lot). Increment the overflow count in
1781 cqe = io_get_cqring(ctx);
1783 WRITE_ONCE(cqe->user_data, req->user_data);
1784 WRITE_ONCE(cqe->res, res);
1785 WRITE_ONCE(cqe->flags, cflags);
1786 } else if (ctx->cq_overflow_flushed ||
1787 atomic_read(&req->task->io_uring->in_idle)) {
1789 * If we're in ring overflow flush mode, or in task cancel mode,
1790 * then we cannot store the request for later flushing, we need
1791 * to drop it on the floor.
1793 ctx->cached_cq_overflow++;
1794 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1796 if (list_empty(&ctx->cq_overflow_list)) {
1797 set_bit(0, &ctx->sq_check_overflow);
1798 set_bit(0, &ctx->cq_check_overflow);
1799 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1803 req->compl.cflags = cflags;
1804 refcount_inc(&req->refs);
1805 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1809 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1811 __io_cqring_fill_event(req, res, 0);
1814 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1816 struct io_ring_ctx *ctx = req->ctx;
1817 unsigned long flags;
1819 spin_lock_irqsave(&ctx->completion_lock, flags);
1820 __io_cqring_fill_event(req, res, cflags);
1821 io_commit_cqring(ctx);
1822 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1824 io_cqring_ev_posted(ctx);
1827 static void io_submit_flush_completions(struct io_comp_state *cs)
1829 struct io_ring_ctx *ctx = cs->ctx;
1831 spin_lock_irq(&ctx->completion_lock);
1832 while (!list_empty(&cs->list)) {
1833 struct io_kiocb *req;
1835 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1836 list_del(&req->compl.list);
1837 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1840 * io_free_req() doesn't care about completion_lock unless one
1841 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1842 * because of a potential deadlock with req->work.fs->lock
1844 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1845 |REQ_F_WORK_INITIALIZED)) {
1846 spin_unlock_irq(&ctx->completion_lock);
1848 spin_lock_irq(&ctx->completion_lock);
1853 io_commit_cqring(ctx);
1854 spin_unlock_irq(&ctx->completion_lock);
1856 io_cqring_ev_posted(ctx);
1860 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1861 struct io_comp_state *cs)
1864 io_cqring_add_event(req, res, cflags);
1869 req->compl.cflags = cflags;
1870 list_add_tail(&req->compl.list, &cs->list);
1872 io_submit_flush_completions(cs);
1876 static void io_req_complete(struct io_kiocb *req, long res)
1878 __io_req_complete(req, res, 0, NULL);
1881 static inline bool io_is_fallback_req(struct io_kiocb *req)
1883 return req == (struct io_kiocb *)
1884 ((unsigned long) req->ctx->fallback_req & ~1UL);
1887 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1889 struct io_kiocb *req;
1891 req = ctx->fallback_req;
1892 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1898 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1899 struct io_submit_state *state)
1901 if (!state->free_reqs) {
1902 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1906 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1907 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1910 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1911 * retry single alloc to be on the safe side.
1913 if (unlikely(ret <= 0)) {
1914 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1915 if (!state->reqs[0])
1919 state->free_reqs = ret;
1923 return state->reqs[state->free_reqs];
1925 return io_get_fallback_req(ctx);
1928 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1932 percpu_ref_put(req->fixed_file_refs);
1937 static void io_dismantle_req(struct io_kiocb *req)
1941 if (req->async_data)
1942 kfree(req->async_data);
1944 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1946 io_req_clean_work(req);
1949 static void __io_free_req(struct io_kiocb *req)
1951 struct io_uring_task *tctx = req->task->io_uring;
1952 struct io_ring_ctx *ctx = req->ctx;
1954 io_dismantle_req(req);
1956 percpu_counter_dec(&tctx->inflight);
1957 if (atomic_read(&tctx->in_idle))
1958 wake_up(&tctx->wait);
1959 put_task_struct(req->task);
1961 if (likely(!io_is_fallback_req(req)))
1962 kmem_cache_free(req_cachep, req);
1964 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1965 percpu_ref_put(&ctx->refs);
1968 static void io_kill_linked_timeout(struct io_kiocb *req)
1970 struct io_ring_ctx *ctx = req->ctx;
1971 struct io_kiocb *link;
1972 bool cancelled = false;
1973 unsigned long flags;
1975 spin_lock_irqsave(&ctx->completion_lock, flags);
1976 link = list_first_entry_or_null(&req->link_list, struct io_kiocb,
1979 * Can happen if a linked timeout fired and link had been like
1980 * req -> link t-out -> link t-out [-> ...]
1982 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1983 struct io_timeout_data *io = link->async_data;
1986 list_del_init(&link->link_list);
1987 ret = hrtimer_try_to_cancel(&io->timer);
1989 io_cqring_fill_event(link, -ECANCELED);
1990 io_commit_cqring(ctx);
1994 req->flags &= ~REQ_F_LINK_TIMEOUT;
1995 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1998 io_cqring_ev_posted(ctx);
2003 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
2005 struct io_kiocb *nxt;
2008 * The list should never be empty when we are called here. But could
2009 * potentially happen if the chain is messed up, check to be on the
2012 if (unlikely(list_empty(&req->link_list)))
2015 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
2016 list_del_init(&req->link_list);
2017 if (!list_empty(&nxt->link_list))
2018 nxt->flags |= REQ_F_LINK_HEAD;
2023 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
2025 static void io_fail_links(struct io_kiocb *req)
2027 struct io_ring_ctx *ctx = req->ctx;
2028 unsigned long flags;
2030 spin_lock_irqsave(&ctx->completion_lock, flags);
2031 while (!list_empty(&req->link_list)) {
2032 struct io_kiocb *link = list_first_entry(&req->link_list,
2033 struct io_kiocb, link_list);
2035 list_del_init(&link->link_list);
2036 trace_io_uring_fail_link(req, link);
2038 io_cqring_fill_event(link, -ECANCELED);
2041 * It's ok to free under spinlock as they're not linked anymore,
2042 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2045 if (link->flags & REQ_F_WORK_INITIALIZED)
2046 io_put_req_deferred(link, 2);
2048 io_double_put_req(link);
2051 io_commit_cqring(ctx);
2052 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2054 io_cqring_ev_posted(ctx);
2057 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2059 req->flags &= ~REQ_F_LINK_HEAD;
2060 if (req->flags & REQ_F_LINK_TIMEOUT)
2061 io_kill_linked_timeout(req);
2064 * If LINK is set, we have dependent requests in this chain. If we
2065 * didn't fail this request, queue the first one up, moving any other
2066 * dependencies to the next request. In case of failure, fail the rest
2069 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
2070 return io_req_link_next(req);
2075 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2077 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
2079 return __io_req_find_next(req);
2082 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
2084 struct task_struct *tsk = req->task;
2085 struct io_ring_ctx *ctx = req->ctx;
2086 enum task_work_notify_mode notify;
2089 if (tsk->flags & PF_EXITING)
2093 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2094 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2095 * processing task_work. There's no reliable way to tell if TWA_RESUME
2099 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
2100 notify = TWA_SIGNAL;
2102 ret = task_work_add(tsk, &req->task_work, notify);
2104 wake_up_process(tsk);
2109 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2111 struct io_ring_ctx *ctx = req->ctx;
2113 spin_lock_irq(&ctx->completion_lock);
2114 io_cqring_fill_event(req, error);
2115 io_commit_cqring(ctx);
2116 spin_unlock_irq(&ctx->completion_lock);
2118 io_cqring_ev_posted(ctx);
2119 req_set_fail_links(req);
2120 io_double_put_req(req);
2123 static void io_req_task_cancel(struct callback_head *cb)
2125 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2126 struct io_ring_ctx *ctx = req->ctx;
2128 __io_req_task_cancel(req, -ECANCELED);
2129 percpu_ref_put(&ctx->refs);
2132 static void __io_req_task_submit(struct io_kiocb *req)
2134 struct io_ring_ctx *ctx = req->ctx;
2136 if (!__io_sq_thread_acquire_mm(ctx)) {
2137 __io_sq_thread_acquire_files(ctx);
2138 mutex_lock(&ctx->uring_lock);
2139 __io_queue_sqe(req, NULL);
2140 mutex_unlock(&ctx->uring_lock);
2142 __io_req_task_cancel(req, -EFAULT);
2146 static void io_req_task_submit(struct callback_head *cb)
2148 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2149 struct io_ring_ctx *ctx = req->ctx;
2151 __io_req_task_submit(req);
2152 percpu_ref_put(&ctx->refs);
2155 static void io_req_task_queue(struct io_kiocb *req)
2159 init_task_work(&req->task_work, io_req_task_submit);
2160 percpu_ref_get(&req->ctx->refs);
2162 ret = io_req_task_work_add(req, true);
2163 if (unlikely(ret)) {
2164 struct task_struct *tsk;
2166 init_task_work(&req->task_work, io_req_task_cancel);
2167 tsk = io_wq_get_task(req->ctx->io_wq);
2168 task_work_add(tsk, &req->task_work, TWA_NONE);
2169 wake_up_process(tsk);
2173 static void io_queue_next(struct io_kiocb *req)
2175 struct io_kiocb *nxt = io_req_find_next(req);
2178 io_req_task_queue(nxt);
2181 static void io_free_req(struct io_kiocb *req)
2188 void *reqs[IO_IOPOLL_BATCH];
2191 struct task_struct *task;
2195 static inline void io_init_req_batch(struct req_batch *rb)
2202 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2203 struct req_batch *rb)
2205 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2206 percpu_ref_put_many(&ctx->refs, rb->to_free);
2210 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2211 struct req_batch *rb)
2214 __io_req_free_batch_flush(ctx, rb);
2216 struct io_uring_task *tctx = rb->task->io_uring;
2218 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2219 put_task_struct_many(rb->task, rb->task_refs);
2224 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2226 if (unlikely(io_is_fallback_req(req))) {
2230 if (req->flags & REQ_F_LINK_HEAD)
2233 if (req->task != rb->task) {
2235 struct io_uring_task *tctx = rb->task->io_uring;
2237 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2238 put_task_struct_many(rb->task, rb->task_refs);
2240 rb->task = req->task;
2245 io_dismantle_req(req);
2246 rb->reqs[rb->to_free++] = req;
2247 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2248 __io_req_free_batch_flush(req->ctx, rb);
2252 * Drop reference to request, return next in chain (if there is one) if this
2253 * was the last reference to this request.
2255 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2257 struct io_kiocb *nxt = NULL;
2259 if (refcount_dec_and_test(&req->refs)) {
2260 nxt = io_req_find_next(req);
2266 static void io_put_req(struct io_kiocb *req)
2268 if (refcount_dec_and_test(&req->refs))
2272 static void io_put_req_deferred_cb(struct callback_head *cb)
2274 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2279 static void io_free_req_deferred(struct io_kiocb *req)
2283 init_task_work(&req->task_work, io_put_req_deferred_cb);
2284 ret = io_req_task_work_add(req, true);
2285 if (unlikely(ret)) {
2286 struct task_struct *tsk;
2288 tsk = io_wq_get_task(req->ctx->io_wq);
2289 task_work_add(tsk, &req->task_work, TWA_NONE);
2290 wake_up_process(tsk);
2294 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2296 if (refcount_sub_and_test(refs, &req->refs))
2297 io_free_req_deferred(req);
2300 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2302 struct io_kiocb *nxt;
2305 * A ref is owned by io-wq in which context we're. So, if that's the
2306 * last one, it's safe to steal next work. False negatives are Ok,
2307 * it just will be re-punted async in io_put_work()
2309 if (refcount_read(&req->refs) != 1)
2312 nxt = io_req_find_next(req);
2313 return nxt ? &nxt->work : NULL;
2316 static void io_double_put_req(struct io_kiocb *req)
2318 /* drop both submit and complete references */
2319 if (refcount_sub_and_test(2, &req->refs))
2323 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2325 struct io_rings *rings = ctx->rings;
2327 if (test_bit(0, &ctx->cq_check_overflow)) {
2329 * noflush == true is from the waitqueue handler, just ensure
2330 * we wake up the task, and the next invocation will flush the
2331 * entries. We cannot safely to it from here.
2333 if (noflush && !list_empty(&ctx->cq_overflow_list))
2336 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2339 /* See comment at the top of this file */
2341 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2344 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2346 struct io_rings *rings = ctx->rings;
2348 /* make sure SQ entry isn't read before tail */
2349 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2352 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2354 unsigned int cflags;
2356 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2357 cflags |= IORING_CQE_F_BUFFER;
2358 req->flags &= ~REQ_F_BUFFER_SELECTED;
2363 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2365 struct io_buffer *kbuf;
2367 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2368 return io_put_kbuf(req, kbuf);
2371 static inline bool io_run_task_work(void)
2374 * Not safe to run on exiting task, and the task_work handling will
2375 * not add work to such a task.
2377 if (unlikely(current->flags & PF_EXITING))
2379 if (current->task_works) {
2380 __set_current_state(TASK_RUNNING);
2388 static void io_iopoll_queue(struct list_head *again)
2390 struct io_kiocb *req;
2393 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2394 list_del(&req->inflight_entry);
2395 __io_complete_rw(req, -EAGAIN, 0, NULL);
2396 } while (!list_empty(again));
2400 * Find and free completed poll iocbs
2402 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2403 struct list_head *done)
2405 struct req_batch rb;
2406 struct io_kiocb *req;
2409 /* order with ->result store in io_complete_rw_iopoll() */
2412 io_init_req_batch(&rb);
2413 while (!list_empty(done)) {
2416 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2417 if (READ_ONCE(req->result) == -EAGAIN) {
2419 req->iopoll_completed = 0;
2420 list_move_tail(&req->inflight_entry, &again);
2423 list_del(&req->inflight_entry);
2425 if (req->flags & REQ_F_BUFFER_SELECTED)
2426 cflags = io_put_rw_kbuf(req);
2428 __io_cqring_fill_event(req, req->result, cflags);
2431 if (refcount_dec_and_test(&req->refs))
2432 io_req_free_batch(&rb, req);
2435 io_commit_cqring(ctx);
2436 if (ctx->flags & IORING_SETUP_SQPOLL)
2437 io_cqring_ev_posted(ctx);
2438 io_req_free_batch_finish(ctx, &rb);
2440 if (!list_empty(&again))
2441 io_iopoll_queue(&again);
2444 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2447 struct io_kiocb *req, *tmp;
2453 * Only spin for completions if we don't have multiple devices hanging
2454 * off our complete list, and we're under the requested amount.
2456 spin = !ctx->poll_multi_file && *nr_events < min;
2459 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2460 struct kiocb *kiocb = &req->rw.kiocb;
2463 * Move completed and retryable entries to our local lists.
2464 * If we find a request that requires polling, break out
2465 * and complete those lists first, if we have entries there.
2467 if (READ_ONCE(req->iopoll_completed)) {
2468 list_move_tail(&req->inflight_entry, &done);
2471 if (!list_empty(&done))
2474 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2478 /* iopoll may have completed current req */
2479 if (READ_ONCE(req->iopoll_completed))
2480 list_move_tail(&req->inflight_entry, &done);
2487 if (!list_empty(&done))
2488 io_iopoll_complete(ctx, nr_events, &done);
2494 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2495 * non-spinning poll check - we'll still enter the driver poll loop, but only
2496 * as a non-spinning completion check.
2498 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2501 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2504 ret = io_do_iopoll(ctx, nr_events, min);
2507 if (*nr_events >= min)
2515 * We can't just wait for polled events to come to us, we have to actively
2516 * find and complete them.
2518 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2520 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2523 mutex_lock(&ctx->uring_lock);
2524 while (!list_empty(&ctx->iopoll_list)) {
2525 unsigned int nr_events = 0;
2527 io_do_iopoll(ctx, &nr_events, 0);
2529 /* let it sleep and repeat later if can't complete a request */
2533 * Ensure we allow local-to-the-cpu processing to take place,
2534 * in this case we need to ensure that we reap all events.
2535 * Also let task_work, etc. to progress by releasing the mutex
2537 if (need_resched()) {
2538 mutex_unlock(&ctx->uring_lock);
2540 mutex_lock(&ctx->uring_lock);
2543 mutex_unlock(&ctx->uring_lock);
2546 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2548 unsigned int nr_events = 0;
2549 int iters = 0, ret = 0;
2552 * We disallow the app entering submit/complete with polling, but we
2553 * still need to lock the ring to prevent racing with polled issue
2554 * that got punted to a workqueue.
2556 mutex_lock(&ctx->uring_lock);
2559 * Don't enter poll loop if we already have events pending.
2560 * If we do, we can potentially be spinning for commands that
2561 * already triggered a CQE (eg in error).
2563 if (io_cqring_events(ctx, false))
2567 * If a submit got punted to a workqueue, we can have the
2568 * application entering polling for a command before it gets
2569 * issued. That app will hold the uring_lock for the duration
2570 * of the poll right here, so we need to take a breather every
2571 * now and then to ensure that the issue has a chance to add
2572 * the poll to the issued list. Otherwise we can spin here
2573 * forever, while the workqueue is stuck trying to acquire the
2576 if (!(++iters & 7)) {
2577 mutex_unlock(&ctx->uring_lock);
2579 mutex_lock(&ctx->uring_lock);
2582 ret = io_iopoll_getevents(ctx, &nr_events, min);
2586 } while (min && !nr_events && !need_resched());
2588 mutex_unlock(&ctx->uring_lock);
2592 static void kiocb_end_write(struct io_kiocb *req)
2595 * Tell lockdep we inherited freeze protection from submission
2598 if (req->flags & REQ_F_ISREG) {
2599 struct inode *inode = file_inode(req->file);
2601 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2603 file_end_write(req->file);
2606 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2607 struct io_comp_state *cs)
2609 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2612 if (kiocb->ki_flags & IOCB_WRITE)
2613 kiocb_end_write(req);
2615 if (res != req->result)
2616 req_set_fail_links(req);
2617 if (req->flags & REQ_F_BUFFER_SELECTED)
2618 cflags = io_put_rw_kbuf(req);
2619 __io_req_complete(req, res, cflags, cs);
2623 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2625 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2626 ssize_t ret = -ECANCELED;
2627 struct iov_iter iter;
2635 switch (req->opcode) {
2636 case IORING_OP_READV:
2637 case IORING_OP_READ_FIXED:
2638 case IORING_OP_READ:
2641 case IORING_OP_WRITEV:
2642 case IORING_OP_WRITE_FIXED:
2643 case IORING_OP_WRITE:
2647 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2652 if (!req->async_data) {
2653 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2656 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2664 req_set_fail_links(req);
2669 static bool io_rw_reissue(struct io_kiocb *req, long res)
2672 umode_t mode = file_inode(req->file)->i_mode;
2675 if (!S_ISBLK(mode) && !S_ISREG(mode))
2677 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2680 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2682 if (io_resubmit_prep(req, ret)) {
2683 refcount_inc(&req->refs);
2684 io_queue_async_work(req);
2692 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2693 struct io_comp_state *cs)
2695 if (!io_rw_reissue(req, res))
2696 io_complete_rw_common(&req->rw.kiocb, res, cs);
2699 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2701 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2703 __io_complete_rw(req, res, res2, NULL);
2706 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2708 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2710 if (kiocb->ki_flags & IOCB_WRITE)
2711 kiocb_end_write(req);
2713 if (res != -EAGAIN && res != req->result)
2714 req_set_fail_links(req);
2716 WRITE_ONCE(req->result, res);
2717 /* order with io_poll_complete() checking ->result */
2719 WRITE_ONCE(req->iopoll_completed, 1);
2723 * After the iocb has been issued, it's safe to be found on the poll list.
2724 * Adding the kiocb to the list AFTER submission ensures that we don't
2725 * find it from a io_iopoll_getevents() thread before the issuer is done
2726 * accessing the kiocb cookie.
2728 static void io_iopoll_req_issued(struct io_kiocb *req)
2730 struct io_ring_ctx *ctx = req->ctx;
2733 * Track whether we have multiple files in our lists. This will impact
2734 * how we do polling eventually, not spinning if we're on potentially
2735 * different devices.
2737 if (list_empty(&ctx->iopoll_list)) {
2738 ctx->poll_multi_file = false;
2739 } else if (!ctx->poll_multi_file) {
2740 struct io_kiocb *list_req;
2742 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2744 if (list_req->file != req->file)
2745 ctx->poll_multi_file = true;
2749 * For fast devices, IO may have already completed. If it has, add
2750 * it to the front so we find it first.
2752 if (READ_ONCE(req->iopoll_completed))
2753 list_add(&req->inflight_entry, &ctx->iopoll_list);
2755 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2757 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2758 wq_has_sleeper(&ctx->sq_data->wait))
2759 wake_up(&ctx->sq_data->wait);
2762 static void __io_state_file_put(struct io_submit_state *state)
2764 if (state->has_refs)
2765 fput_many(state->file, state->has_refs);
2769 static inline void io_state_file_put(struct io_submit_state *state)
2772 __io_state_file_put(state);
2776 * Get as many references to a file as we have IOs left in this submission,
2777 * assuming most submissions are for one file, or at least that each file
2778 * has more than one submission.
2780 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2786 if (state->fd == fd) {
2790 __io_state_file_put(state);
2792 state->file = fget_many(fd, state->ios_left);
2797 state->has_refs = state->ios_left - 1;
2801 static bool io_bdev_nowait(struct block_device *bdev)
2804 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2811 * If we tracked the file through the SCM inflight mechanism, we could support
2812 * any file. For now, just ensure that anything potentially problematic is done
2815 static bool io_file_supports_async(struct file *file, int rw)
2817 umode_t mode = file_inode(file)->i_mode;
2819 if (S_ISBLK(mode)) {
2820 if (io_bdev_nowait(file->f_inode->i_bdev))
2824 if (S_ISCHR(mode) || S_ISSOCK(mode))
2826 if (S_ISREG(mode)) {
2827 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2828 file->f_op != &io_uring_fops)
2833 /* any ->read/write should understand O_NONBLOCK */
2834 if (file->f_flags & O_NONBLOCK)
2837 if (!(file->f_mode & FMODE_NOWAIT))
2841 return file->f_op->read_iter != NULL;
2843 return file->f_op->write_iter != NULL;
2846 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2848 struct io_ring_ctx *ctx = req->ctx;
2849 struct kiocb *kiocb = &req->rw.kiocb;
2853 if (S_ISREG(file_inode(req->file)->i_mode))
2854 req->flags |= REQ_F_ISREG;
2856 kiocb->ki_pos = READ_ONCE(sqe->off);
2857 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2858 req->flags |= REQ_F_CUR_POS;
2859 kiocb->ki_pos = req->file->f_pos;
2861 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2862 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2863 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2867 ioprio = READ_ONCE(sqe->ioprio);
2869 ret = ioprio_check_cap(ioprio);
2873 kiocb->ki_ioprio = ioprio;
2875 kiocb->ki_ioprio = get_current_ioprio();
2877 /* don't allow async punt if RWF_NOWAIT was requested */
2878 if (kiocb->ki_flags & IOCB_NOWAIT)
2879 req->flags |= REQ_F_NOWAIT;
2881 if (ctx->flags & IORING_SETUP_IOPOLL) {
2882 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2883 !kiocb->ki_filp->f_op->iopoll)
2886 kiocb->ki_flags |= IOCB_HIPRI;
2887 kiocb->ki_complete = io_complete_rw_iopoll;
2888 req->iopoll_completed = 0;
2890 if (kiocb->ki_flags & IOCB_HIPRI)
2892 kiocb->ki_complete = io_complete_rw;
2895 req->rw.addr = READ_ONCE(sqe->addr);
2896 req->rw.len = READ_ONCE(sqe->len);
2897 req->buf_index = READ_ONCE(sqe->buf_index);
2901 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2907 case -ERESTARTNOINTR:
2908 case -ERESTARTNOHAND:
2909 case -ERESTART_RESTARTBLOCK:
2911 * We can't just restart the syscall, since previously
2912 * submitted sqes may already be in progress. Just fail this
2918 kiocb->ki_complete(kiocb, ret, 0);
2922 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2923 struct io_comp_state *cs)
2925 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2926 struct io_async_rw *io = req->async_data;
2928 /* add previously done IO, if any */
2929 if (io && io->bytes_done > 0) {
2931 ret = io->bytes_done;
2933 ret += io->bytes_done;
2936 if (req->flags & REQ_F_CUR_POS)
2937 req->file->f_pos = kiocb->ki_pos;
2938 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2939 __io_complete_rw(req, ret, 0, cs);
2941 io_rw_done(kiocb, ret);
2944 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2945 struct iov_iter *iter)
2947 struct io_ring_ctx *ctx = req->ctx;
2948 size_t len = req->rw.len;
2949 struct io_mapped_ubuf *imu;
2950 u16 index, buf_index = req->buf_index;
2954 if (unlikely(buf_index >= ctx->nr_user_bufs))
2956 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2957 imu = &ctx->user_bufs[index];
2958 buf_addr = req->rw.addr;
2961 if (buf_addr + len < buf_addr)
2963 /* not inside the mapped region */
2964 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2968 * May not be a start of buffer, set size appropriately
2969 * and advance us to the beginning.
2971 offset = buf_addr - imu->ubuf;
2972 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2976 * Don't use iov_iter_advance() here, as it's really slow for
2977 * using the latter parts of a big fixed buffer - it iterates
2978 * over each segment manually. We can cheat a bit here, because
2981 * 1) it's a BVEC iter, we set it up
2982 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2983 * first and last bvec
2985 * So just find our index, and adjust the iterator afterwards.
2986 * If the offset is within the first bvec (or the whole first
2987 * bvec, just use iov_iter_advance(). This makes it easier
2988 * since we can just skip the first segment, which may not
2989 * be PAGE_SIZE aligned.
2991 const struct bio_vec *bvec = imu->bvec;
2993 if (offset <= bvec->bv_len) {
2994 iov_iter_advance(iter, offset);
2996 unsigned long seg_skip;
2998 /* skip first vec */
2999 offset -= bvec->bv_len;
3000 seg_skip = 1 + (offset >> PAGE_SHIFT);
3002 iter->bvec = bvec + seg_skip;
3003 iter->nr_segs -= seg_skip;
3004 iter->count -= bvec->bv_len + offset;
3005 iter->iov_offset = offset & ~PAGE_MASK;
3012 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3015 mutex_unlock(&ctx->uring_lock);
3018 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3021 * "Normal" inline submissions always hold the uring_lock, since we
3022 * grab it from the system call. Same is true for the SQPOLL offload.
3023 * The only exception is when we've detached the request and issue it
3024 * from an async worker thread, grab the lock for that case.
3027 mutex_lock(&ctx->uring_lock);
3030 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3031 int bgid, struct io_buffer *kbuf,
3034 struct io_buffer *head;
3036 if (req->flags & REQ_F_BUFFER_SELECTED)
3039 io_ring_submit_lock(req->ctx, needs_lock);
3041 lockdep_assert_held(&req->ctx->uring_lock);
3043 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3045 if (!list_empty(&head->list)) {
3046 kbuf = list_last_entry(&head->list, struct io_buffer,
3048 list_del(&kbuf->list);
3051 idr_remove(&req->ctx->io_buffer_idr, bgid);
3053 if (*len > kbuf->len)
3056 kbuf = ERR_PTR(-ENOBUFS);
3059 io_ring_submit_unlock(req->ctx, needs_lock);
3064 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3067 struct io_buffer *kbuf;
3070 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3071 bgid = req->buf_index;
3072 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3075 req->rw.addr = (u64) (unsigned long) kbuf;
3076 req->flags |= REQ_F_BUFFER_SELECTED;
3077 return u64_to_user_ptr(kbuf->addr);
3080 #ifdef CONFIG_COMPAT
3081 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3084 struct compat_iovec __user *uiov;
3085 compat_ssize_t clen;
3089 uiov = u64_to_user_ptr(req->rw.addr);
3090 if (!access_ok(uiov, sizeof(*uiov)))
3092 if (__get_user(clen, &uiov->iov_len))
3098 buf = io_rw_buffer_select(req, &len, needs_lock);
3100 return PTR_ERR(buf);
3101 iov[0].iov_base = buf;
3102 iov[0].iov_len = (compat_size_t) len;
3107 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3110 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3114 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3117 len = iov[0].iov_len;
3120 buf = io_rw_buffer_select(req, &len, needs_lock);
3122 return PTR_ERR(buf);
3123 iov[0].iov_base = buf;
3124 iov[0].iov_len = len;
3128 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3131 if (req->flags & REQ_F_BUFFER_SELECTED) {
3132 struct io_buffer *kbuf;
3134 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3135 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3136 iov[0].iov_len = kbuf->len;
3141 else if (req->rw.len > 1)
3144 #ifdef CONFIG_COMPAT
3145 if (req->ctx->compat)
3146 return io_compat_import(req, iov, needs_lock);
3149 return __io_iov_buffer_select(req, iov, needs_lock);
3152 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3153 struct iovec **iovec, struct iov_iter *iter,
3156 void __user *buf = u64_to_user_ptr(req->rw.addr);
3157 size_t sqe_len = req->rw.len;
3161 opcode = req->opcode;
3162 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3164 return io_import_fixed(req, rw, iter);
3167 /* buffer index only valid with fixed read/write, or buffer select */
3168 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3171 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3172 if (req->flags & REQ_F_BUFFER_SELECT) {
3173 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3175 return PTR_ERR(buf);
3176 req->rw.len = sqe_len;
3179 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3181 return ret < 0 ? ret : sqe_len;
3184 if (req->flags & REQ_F_BUFFER_SELECT) {
3185 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3187 ret = (*iovec)->iov_len;
3188 iov_iter_init(iter, rw, *iovec, 1, ret);
3194 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3198 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3199 struct iovec **iovec, struct iov_iter *iter,
3202 struct io_async_rw *iorw = req->async_data;
3205 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3207 return iov_iter_count(&iorw->iter);
3210 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3212 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3216 * For files that don't have ->read_iter() and ->write_iter(), handle them
3217 * by looping over ->read() or ->write() manually.
3219 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3221 struct kiocb *kiocb = &req->rw.kiocb;
3222 struct file *file = req->file;
3226 * Don't support polled IO through this interface, and we can't
3227 * support non-blocking either. For the latter, this just causes
3228 * the kiocb to be handled from an async context.
3230 if (kiocb->ki_flags & IOCB_HIPRI)
3232 if (kiocb->ki_flags & IOCB_NOWAIT)
3235 while (iov_iter_count(iter)) {
3239 if (!iov_iter_is_bvec(iter)) {
3240 iovec = iov_iter_iovec(iter);
3242 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3243 iovec.iov_len = req->rw.len;
3247 nr = file->f_op->read(file, iovec.iov_base,
3248 iovec.iov_len, io_kiocb_ppos(kiocb));
3250 nr = file->f_op->write(file, iovec.iov_base,
3251 iovec.iov_len, io_kiocb_ppos(kiocb));
3260 if (nr != iovec.iov_len)
3264 iov_iter_advance(iter, nr);
3270 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3271 const struct iovec *fast_iov, struct iov_iter *iter)
3273 struct io_async_rw *rw = req->async_data;
3275 memcpy(&rw->iter, iter, sizeof(*iter));
3276 rw->free_iovec = iovec;
3278 /* can only be fixed buffers, no need to do anything */
3279 if (iter->type == ITER_BVEC)
3282 unsigned iov_off = 0;
3284 rw->iter.iov = rw->fast_iov;
3285 if (iter->iov != fast_iov) {
3286 iov_off = iter->iov - fast_iov;
3287 rw->iter.iov += iov_off;
3289 if (rw->fast_iov != fast_iov)
3290 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3291 sizeof(struct iovec) * iter->nr_segs);
3293 req->flags |= REQ_F_NEED_CLEANUP;
3297 static inline int __io_alloc_async_data(struct io_kiocb *req)
3299 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3300 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3301 return req->async_data == NULL;
3304 static int io_alloc_async_data(struct io_kiocb *req)
3306 if (!io_op_defs[req->opcode].needs_async_data)
3309 return __io_alloc_async_data(req);
3312 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3313 const struct iovec *fast_iov,
3314 struct iov_iter *iter, bool force)
3316 if (!force && !io_op_defs[req->opcode].needs_async_data)
3318 if (!req->async_data) {
3319 if (__io_alloc_async_data(req))
3322 io_req_map_rw(req, iovec, fast_iov, iter);
3327 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3329 struct io_async_rw *iorw = req->async_data;
3330 struct iovec *iov = iorw->fast_iov;
3333 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3334 if (unlikely(ret < 0))
3337 iorw->bytes_done = 0;
3338 iorw->free_iovec = iov;
3340 req->flags |= REQ_F_NEED_CLEANUP;
3344 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3348 ret = io_prep_rw(req, sqe);
3352 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3355 /* either don't need iovec imported or already have it */
3356 if (!req->async_data)
3358 return io_rw_prep_async(req, READ);
3362 * This is our waitqueue callback handler, registered through lock_page_async()
3363 * when we initially tried to do the IO with the iocb armed our waitqueue.
3364 * This gets called when the page is unlocked, and we generally expect that to
3365 * happen when the page IO is completed and the page is now uptodate. This will
3366 * queue a task_work based retry of the operation, attempting to copy the data
3367 * again. If the latter fails because the page was NOT uptodate, then we will
3368 * do a thread based blocking retry of the operation. That's the unexpected
3371 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3372 int sync, void *arg)
3374 struct wait_page_queue *wpq;
3375 struct io_kiocb *req = wait->private;
3376 struct wait_page_key *key = arg;
3379 wpq = container_of(wait, struct wait_page_queue, wait);
3381 if (!wake_page_match(wpq, key))
3384 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3385 list_del_init(&wait->entry);
3387 init_task_work(&req->task_work, io_req_task_submit);
3388 percpu_ref_get(&req->ctx->refs);
3390 /* submit ref gets dropped, acquire a new one */
3391 refcount_inc(&req->refs);
3392 ret = io_req_task_work_add(req, true);
3393 if (unlikely(ret)) {
3394 struct task_struct *tsk;
3396 /* queue just for cancelation */
3397 init_task_work(&req->task_work, io_req_task_cancel);
3398 tsk = io_wq_get_task(req->ctx->io_wq);
3399 task_work_add(tsk, &req->task_work, TWA_NONE);
3400 wake_up_process(tsk);
3406 * This controls whether a given IO request should be armed for async page
3407 * based retry. If we return false here, the request is handed to the async
3408 * worker threads for retry. If we're doing buffered reads on a regular file,
3409 * we prepare a private wait_page_queue entry and retry the operation. This
3410 * will either succeed because the page is now uptodate and unlocked, or it
3411 * will register a callback when the page is unlocked at IO completion. Through
3412 * that callback, io_uring uses task_work to setup a retry of the operation.
3413 * That retry will attempt the buffered read again. The retry will generally
3414 * succeed, or in rare cases where it fails, we then fall back to using the
3415 * async worker threads for a blocking retry.
3417 static bool io_rw_should_retry(struct io_kiocb *req)
3419 struct io_async_rw *rw = req->async_data;
3420 struct wait_page_queue *wait = &rw->wpq;
3421 struct kiocb *kiocb = &req->rw.kiocb;
3423 /* never retry for NOWAIT, we just complete with -EAGAIN */
3424 if (req->flags & REQ_F_NOWAIT)
3427 /* Only for buffered IO */
3428 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3432 * just use poll if we can, and don't attempt if the fs doesn't
3433 * support callback based unlocks
3435 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3438 wait->wait.func = io_async_buf_func;
3439 wait->wait.private = req;
3440 wait->wait.flags = 0;
3441 INIT_LIST_HEAD(&wait->wait.entry);
3442 kiocb->ki_flags |= IOCB_WAITQ;
3443 kiocb->ki_flags &= ~IOCB_NOWAIT;
3444 kiocb->ki_waitq = wait;
3448 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3450 if (req->file->f_op->read_iter)
3451 return call_read_iter(req->file, &req->rw.kiocb, iter);
3452 else if (req->file->f_op->read)
3453 return loop_rw_iter(READ, req, iter);
3458 static int io_read(struct io_kiocb *req, bool force_nonblock,
3459 struct io_comp_state *cs)
3461 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3462 struct kiocb *kiocb = &req->rw.kiocb;
3463 struct iov_iter __iter, *iter = &__iter;
3464 struct io_async_rw *rw = req->async_data;
3465 ssize_t io_size, ret, ret2;
3472 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3475 iov_count = iov_iter_count(iter);
3477 req->result = io_size;
3480 /* Ensure we clear previously set non-block flag */
3481 if (!force_nonblock)
3482 kiocb->ki_flags &= ~IOCB_NOWAIT;
3484 kiocb->ki_flags |= IOCB_NOWAIT;
3487 /* If the file doesn't support async, just async punt */
3488 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3492 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3496 ret = io_iter_do_read(req, iter);
3500 } else if (ret == -EIOCBQUEUED) {
3503 } else if (ret == -EAGAIN) {
3504 /* IOPOLL retry should happen for io-wq threads */
3505 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3507 /* no retry on NONBLOCK marked file */
3508 if (req->file->f_flags & O_NONBLOCK)
3510 /* some cases will consume bytes even on error returns */
3511 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3514 } else if (ret < 0) {
3515 /* make sure -ERESTARTSYS -> -EINTR is done */
3519 /* read it all, or we did blocking attempt. no retry. */
3520 if (!iov_iter_count(iter) || !force_nonblock ||
3521 (req->file->f_flags & O_NONBLOCK))
3526 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3533 rw = req->async_data;
3534 /* it's copied and will be cleaned with ->io */
3536 /* now use our persistent iterator, if we aren't already */
3539 rw->bytes_done += ret;
3540 /* if we can retry, do so with the callbacks armed */
3541 if (!io_rw_should_retry(req)) {
3542 kiocb->ki_flags &= ~IOCB_WAITQ;
3547 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3548 * get -EIOCBQUEUED, then we'll get a notification when the desired
3549 * page gets unlocked. We can also get a partial read here, and if we
3550 * do, then just retry at the new offset.
3552 ret = io_iter_do_read(req, iter);
3553 if (ret == -EIOCBQUEUED) {
3556 } else if (ret > 0 && ret < io_size) {
3557 /* we got some bytes, but not all. retry. */
3561 kiocb_done(kiocb, ret, cs);
3564 /* it's reportedly faster than delegating the null check to kfree() */
3570 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3574 ret = io_prep_rw(req, sqe);
3578 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3581 /* either don't need iovec imported or already have it */
3582 if (!req->async_data)
3584 return io_rw_prep_async(req, WRITE);
3587 static int io_write(struct io_kiocb *req, bool force_nonblock,
3588 struct io_comp_state *cs)
3590 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3591 struct kiocb *kiocb = &req->rw.kiocb;
3592 struct iov_iter __iter, *iter = &__iter;
3593 struct io_async_rw *rw = req->async_data;
3595 ssize_t ret, ret2, io_size;
3600 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3603 iov_count = iov_iter_count(iter);
3605 req->result = io_size;
3607 /* Ensure we clear previously set non-block flag */
3608 if (!force_nonblock)
3609 kiocb->ki_flags &= ~IOCB_NOWAIT;
3611 kiocb->ki_flags |= IOCB_NOWAIT;
3613 /* If the file doesn't support async, just async punt */
3614 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3617 /* file path doesn't support NOWAIT for non-direct_IO */
3618 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3619 (req->flags & REQ_F_ISREG))
3622 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3627 * Open-code file_start_write here to grab freeze protection,
3628 * which will be released by another thread in
3629 * io_complete_rw(). Fool lockdep by telling it the lock got
3630 * released so that it doesn't complain about the held lock when
3631 * we return to userspace.
3633 if (req->flags & REQ_F_ISREG) {
3634 sb_start_write(file_inode(req->file)->i_sb);
3635 __sb_writers_release(file_inode(req->file)->i_sb,
3638 kiocb->ki_flags |= IOCB_WRITE;
3640 if (req->file->f_op->write_iter)
3641 ret2 = call_write_iter(req->file, kiocb, iter);
3642 else if (req->file->f_op->write)
3643 ret2 = loop_rw_iter(WRITE, req, iter);
3648 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3649 * retry them without IOCB_NOWAIT.
3651 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3653 /* no retry on NONBLOCK marked file */
3654 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3656 if (!force_nonblock || ret2 != -EAGAIN) {
3657 /* IOPOLL retry should happen for io-wq threads */
3658 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3661 kiocb_done(kiocb, ret2, cs);
3664 /* some cases will consume bytes even on error returns */
3665 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3666 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3671 /* it's reportedly faster than delegating the null check to kfree() */
3677 static int io_renameat_prep(struct io_kiocb *req,
3678 const struct io_uring_sqe *sqe)
3680 struct io_rename *ren = &req->rename;
3681 const char __user *oldf, *newf;
3683 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3686 ren->old_dfd = READ_ONCE(sqe->fd);
3687 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3688 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3689 ren->new_dfd = READ_ONCE(sqe->len);
3690 ren->flags = READ_ONCE(sqe->rename_flags);
3692 ren->oldpath = getname(oldf);
3693 if (IS_ERR(ren->oldpath))
3694 return PTR_ERR(ren->oldpath);
3696 ren->newpath = getname(newf);
3697 if (IS_ERR(ren->newpath)) {
3698 putname(ren->oldpath);
3699 return PTR_ERR(ren->newpath);
3702 req->flags |= REQ_F_NEED_CLEANUP;
3706 static int io_renameat(struct io_kiocb *req, bool force_nonblock)
3708 struct io_rename *ren = &req->rename;
3714 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3715 ren->newpath, ren->flags);
3717 req->flags &= ~REQ_F_NEED_CLEANUP;
3719 req_set_fail_links(req);
3720 io_req_complete(req, ret);
3724 static int io_unlinkat_prep(struct io_kiocb *req,
3725 const struct io_uring_sqe *sqe)
3727 struct io_unlink *un = &req->unlink;
3728 const char __user *fname;
3730 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3733 un->dfd = READ_ONCE(sqe->fd);
3735 un->flags = READ_ONCE(sqe->unlink_flags);
3736 if (un->flags & ~AT_REMOVEDIR)
3739 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3740 un->filename = getname(fname);
3741 if (IS_ERR(un->filename))
3742 return PTR_ERR(un->filename);
3744 req->flags |= REQ_F_NEED_CLEANUP;
3748 static int io_unlinkat(struct io_kiocb *req, bool force_nonblock)
3750 struct io_unlink *un = &req->unlink;
3756 if (un->flags & AT_REMOVEDIR)
3757 ret = do_rmdir(un->dfd, un->filename);
3759 ret = do_unlinkat(un->dfd, un->filename);
3761 req->flags &= ~REQ_F_NEED_CLEANUP;
3763 req_set_fail_links(req);
3764 io_req_complete(req, ret);
3768 static int io_shutdown_prep(struct io_kiocb *req,
3769 const struct io_uring_sqe *sqe)
3771 #if defined(CONFIG_NET)
3772 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3774 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3778 req->shutdown.how = READ_ONCE(sqe->len);
3785 static int io_shutdown(struct io_kiocb *req, bool force_nonblock)
3787 #if defined(CONFIG_NET)
3788 struct socket *sock;
3794 sock = sock_from_file(req->file, &ret);
3795 if (unlikely(!sock))
3798 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3799 io_req_complete(req, ret);
3806 static int __io_splice_prep(struct io_kiocb *req,
3807 const struct io_uring_sqe *sqe)
3809 struct io_splice* sp = &req->splice;
3810 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3812 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3816 sp->len = READ_ONCE(sqe->len);
3817 sp->flags = READ_ONCE(sqe->splice_flags);
3819 if (unlikely(sp->flags & ~valid_flags))
3822 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3823 (sp->flags & SPLICE_F_FD_IN_FIXED));
3826 req->flags |= REQ_F_NEED_CLEANUP;
3828 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3830 * Splice operation will be punted aync, and here need to
3831 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3833 io_req_init_async(req);
3834 req->work.flags |= IO_WQ_WORK_UNBOUND;
3840 static int io_tee_prep(struct io_kiocb *req,
3841 const struct io_uring_sqe *sqe)
3843 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3845 return __io_splice_prep(req, sqe);
3848 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3850 struct io_splice *sp = &req->splice;
3851 struct file *in = sp->file_in;
3852 struct file *out = sp->file_out;
3853 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3859 ret = do_tee(in, out, sp->len, flags);
3861 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3862 req->flags &= ~REQ_F_NEED_CLEANUP;
3865 req_set_fail_links(req);
3866 io_req_complete(req, ret);
3870 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3872 struct io_splice* sp = &req->splice;
3874 sp->off_in = READ_ONCE(sqe->splice_off_in);
3875 sp->off_out = READ_ONCE(sqe->off);
3876 return __io_splice_prep(req, sqe);
3879 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3881 struct io_splice *sp = &req->splice;
3882 struct file *in = sp->file_in;
3883 struct file *out = sp->file_out;
3884 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3885 loff_t *poff_in, *poff_out;
3891 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3892 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3895 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3897 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3898 req->flags &= ~REQ_F_NEED_CLEANUP;
3901 req_set_fail_links(req);
3902 io_req_complete(req, ret);
3907 * IORING_OP_NOP just posts a completion event, nothing else.
3909 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3911 struct io_ring_ctx *ctx = req->ctx;
3913 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3916 __io_req_complete(req, 0, 0, cs);
3920 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3922 struct io_ring_ctx *ctx = req->ctx;
3927 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3929 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3932 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3933 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3936 req->sync.off = READ_ONCE(sqe->off);
3937 req->sync.len = READ_ONCE(sqe->len);
3941 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3943 loff_t end = req->sync.off + req->sync.len;
3946 /* fsync always requires a blocking context */
3950 ret = vfs_fsync_range(req->file, req->sync.off,
3951 end > 0 ? end : LLONG_MAX,
3952 req->sync.flags & IORING_FSYNC_DATASYNC);
3954 req_set_fail_links(req);
3955 io_req_complete(req, ret);
3959 static int io_fallocate_prep(struct io_kiocb *req,
3960 const struct io_uring_sqe *sqe)
3962 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3964 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3967 req->sync.off = READ_ONCE(sqe->off);
3968 req->sync.len = READ_ONCE(sqe->addr);
3969 req->sync.mode = READ_ONCE(sqe->len);
3973 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3977 /* fallocate always requiring blocking context */
3980 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3983 req_set_fail_links(req);
3984 io_req_complete(req, ret);
3988 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3990 const char __user *fname;
3993 if (unlikely(sqe->ioprio || sqe->buf_index))
3995 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3998 /* open.how should be already initialised */
3999 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4000 req->open.how.flags |= O_LARGEFILE;
4002 req->open.dfd = READ_ONCE(sqe->fd);
4003 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4004 req->open.filename = getname(fname);
4005 if (IS_ERR(req->open.filename)) {
4006 ret = PTR_ERR(req->open.filename);
4007 req->open.filename = NULL;
4010 req->open.nofile = rlimit(RLIMIT_NOFILE);
4011 req->open.ignore_nonblock = false;
4012 req->flags |= REQ_F_NEED_CLEANUP;
4016 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4020 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4022 mode = READ_ONCE(sqe->len);
4023 flags = READ_ONCE(sqe->open_flags);
4024 req->open.how = build_open_how(flags, mode);
4025 return __io_openat_prep(req, sqe);
4028 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4030 struct open_how __user *how;
4034 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4036 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4037 len = READ_ONCE(sqe->len);
4038 if (len < OPEN_HOW_SIZE_VER0)
4041 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4046 return __io_openat_prep(req, sqe);
4049 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
4051 struct open_flags op;
4055 if (force_nonblock && !req->open.ignore_nonblock)
4058 ret = build_open_flags(&req->open.how, &op);
4062 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4066 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4069 ret = PTR_ERR(file);
4071 * A work-around to ensure that /proc/self works that way
4072 * that it should - if we get -EOPNOTSUPP back, then assume
4073 * that proc_self_get_link() failed us because we're in async
4074 * context. We should be safe to retry this from the task
4075 * itself with force_nonblock == false set, as it should not
4076 * block on lookup. Would be nice to know this upfront and
4077 * avoid the async dance, but doesn't seem feasible.
4079 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
4080 req->open.ignore_nonblock = true;
4081 refcount_inc(&req->refs);
4082 io_req_task_queue(req);
4086 fsnotify_open(file);
4087 fd_install(ret, file);
4090 putname(req->open.filename);
4091 req->flags &= ~REQ_F_NEED_CLEANUP;
4093 req_set_fail_links(req);
4094 io_req_complete(req, ret);
4098 static int io_openat(struct io_kiocb *req, bool force_nonblock)
4100 return io_openat2(req, force_nonblock);
4103 static int io_remove_buffers_prep(struct io_kiocb *req,
4104 const struct io_uring_sqe *sqe)
4106 struct io_provide_buf *p = &req->pbuf;
4109 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4112 tmp = READ_ONCE(sqe->fd);
4113 if (!tmp || tmp > USHRT_MAX)
4116 memset(p, 0, sizeof(*p));
4118 p->bgid = READ_ONCE(sqe->buf_group);
4122 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4123 int bgid, unsigned nbufs)
4127 /* shouldn't happen */
4131 /* the head kbuf is the list itself */
4132 while (!list_empty(&buf->list)) {
4133 struct io_buffer *nxt;
4135 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4136 list_del(&nxt->list);
4143 idr_remove(&ctx->io_buffer_idr, bgid);
4148 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
4149 struct io_comp_state *cs)
4151 struct io_provide_buf *p = &req->pbuf;
4152 struct io_ring_ctx *ctx = req->ctx;
4153 struct io_buffer *head;
4156 io_ring_submit_lock(ctx, !force_nonblock);
4158 lockdep_assert_held(&ctx->uring_lock);
4161 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4163 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4165 io_ring_submit_lock(ctx, !force_nonblock);
4167 req_set_fail_links(req);
4168 __io_req_complete(req, ret, 0, cs);
4172 static int io_provide_buffers_prep(struct io_kiocb *req,
4173 const struct io_uring_sqe *sqe)
4175 struct io_provide_buf *p = &req->pbuf;
4178 if (sqe->ioprio || sqe->rw_flags)
4181 tmp = READ_ONCE(sqe->fd);
4182 if (!tmp || tmp > USHRT_MAX)
4185 p->addr = READ_ONCE(sqe->addr);
4186 p->len = READ_ONCE(sqe->len);
4188 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4191 p->bgid = READ_ONCE(sqe->buf_group);
4192 tmp = READ_ONCE(sqe->off);
4193 if (tmp > USHRT_MAX)
4199 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4201 struct io_buffer *buf;
4202 u64 addr = pbuf->addr;
4203 int i, bid = pbuf->bid;
4205 for (i = 0; i < pbuf->nbufs; i++) {
4206 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4211 buf->len = pbuf->len;
4216 INIT_LIST_HEAD(&buf->list);
4219 list_add_tail(&buf->list, &(*head)->list);
4223 return i ? i : -ENOMEM;
4226 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4227 struct io_comp_state *cs)
4229 struct io_provide_buf *p = &req->pbuf;
4230 struct io_ring_ctx *ctx = req->ctx;
4231 struct io_buffer *head, *list;
4234 io_ring_submit_lock(ctx, !force_nonblock);
4236 lockdep_assert_held(&ctx->uring_lock);
4238 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4240 ret = io_add_buffers(p, &head);
4245 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4248 __io_remove_buffers(ctx, head, p->bgid, -1U);
4253 io_ring_submit_unlock(ctx, !force_nonblock);
4255 req_set_fail_links(req);
4256 __io_req_complete(req, ret, 0, cs);
4260 static int io_epoll_ctl_prep(struct io_kiocb *req,
4261 const struct io_uring_sqe *sqe)
4263 #if defined(CONFIG_EPOLL)
4264 if (sqe->ioprio || sqe->buf_index)
4266 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4269 req->epoll.epfd = READ_ONCE(sqe->fd);
4270 req->epoll.op = READ_ONCE(sqe->len);
4271 req->epoll.fd = READ_ONCE(sqe->off);
4273 if (ep_op_has_event(req->epoll.op)) {
4274 struct epoll_event __user *ev;
4276 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4277 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4287 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4288 struct io_comp_state *cs)
4290 #if defined(CONFIG_EPOLL)
4291 struct io_epoll *ie = &req->epoll;
4294 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4295 if (force_nonblock && ret == -EAGAIN)
4299 req_set_fail_links(req);
4300 __io_req_complete(req, ret, 0, cs);
4307 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4309 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4310 if (sqe->ioprio || sqe->buf_index || sqe->off)
4312 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4315 req->madvise.addr = READ_ONCE(sqe->addr);
4316 req->madvise.len = READ_ONCE(sqe->len);
4317 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4324 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4326 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4327 struct io_madvise *ma = &req->madvise;
4333 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4335 req_set_fail_links(req);
4336 io_req_complete(req, ret);
4343 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4345 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4347 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4350 req->fadvise.offset = READ_ONCE(sqe->off);
4351 req->fadvise.len = READ_ONCE(sqe->len);
4352 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4356 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4358 struct io_fadvise *fa = &req->fadvise;
4361 if (force_nonblock) {
4362 switch (fa->advice) {
4363 case POSIX_FADV_NORMAL:
4364 case POSIX_FADV_RANDOM:
4365 case POSIX_FADV_SEQUENTIAL:
4372 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4374 req_set_fail_links(req);
4375 io_req_complete(req, ret);
4379 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4381 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4383 if (sqe->ioprio || sqe->buf_index)
4385 if (req->flags & REQ_F_FIXED_FILE)
4388 req->statx.dfd = READ_ONCE(sqe->fd);
4389 req->statx.mask = READ_ONCE(sqe->len);
4390 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4391 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4392 req->statx.flags = READ_ONCE(sqe->statx_flags);
4397 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4399 struct io_statx *ctx = &req->statx;
4402 if (force_nonblock) {
4403 /* only need file table for an actual valid fd */
4404 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4405 req->flags |= REQ_F_NO_FILE_TABLE;
4409 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4413 req_set_fail_links(req);
4414 io_req_complete(req, ret);
4418 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4421 * If we queue this for async, it must not be cancellable. That would
4422 * leave the 'file' in an undeterminate state, and here need to modify
4423 * io_wq_work.flags, so initialize io_wq_work firstly.
4425 io_req_init_async(req);
4426 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4428 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4430 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4431 sqe->rw_flags || sqe->buf_index)
4433 if (req->flags & REQ_F_FIXED_FILE)
4436 req->close.fd = READ_ONCE(sqe->fd);
4437 if ((req->file && req->file->f_op == &io_uring_fops))
4440 req->close.put_file = NULL;
4444 static int io_close(struct io_kiocb *req, bool force_nonblock,
4445 struct io_comp_state *cs)
4447 struct io_close *close = &req->close;
4450 /* might be already done during nonblock submission */
4451 if (!close->put_file) {
4452 ret = __close_fd_get_file(close->fd, &close->put_file);
4454 return (ret == -ENOENT) ? -EBADF : ret;
4457 /* if the file has a flush method, be safe and punt to async */
4458 if (close->put_file->f_op->flush && force_nonblock) {
4459 /* was never set, but play safe */
4460 req->flags &= ~REQ_F_NOWAIT;
4461 /* avoid grabbing files - we don't need the files */
4462 req->flags |= REQ_F_NO_FILE_TABLE;
4466 /* No ->flush() or already async, safely close from here */
4467 ret = filp_close(close->put_file, req->work.identity->files);
4469 req_set_fail_links(req);
4470 fput(close->put_file);
4471 close->put_file = NULL;
4472 __io_req_complete(req, ret, 0, cs);
4476 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4478 struct io_ring_ctx *ctx = req->ctx;
4483 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4485 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4488 req->sync.off = READ_ONCE(sqe->off);
4489 req->sync.len = READ_ONCE(sqe->len);
4490 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4494 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4498 /* sync_file_range always requires a blocking context */
4502 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4505 req_set_fail_links(req);
4506 io_req_complete(req, ret);
4510 #if defined(CONFIG_NET)
4511 static int io_setup_async_msg(struct io_kiocb *req,
4512 struct io_async_msghdr *kmsg)
4514 struct io_async_msghdr *async_msg = req->async_data;
4518 if (io_alloc_async_data(req)) {
4519 if (kmsg->iov != kmsg->fast_iov)
4523 async_msg = req->async_data;
4524 req->flags |= REQ_F_NEED_CLEANUP;
4525 memcpy(async_msg, kmsg, sizeof(*kmsg));
4529 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4530 struct io_async_msghdr *iomsg)
4532 iomsg->iov = iomsg->fast_iov;
4533 iomsg->msg.msg_name = &iomsg->addr;
4534 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4535 req->sr_msg.msg_flags, &iomsg->iov);
4538 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4540 struct io_async_msghdr *async_msg = req->async_data;
4541 struct io_sr_msg *sr = &req->sr_msg;
4544 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4547 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4548 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4549 sr->len = READ_ONCE(sqe->len);
4551 #ifdef CONFIG_COMPAT
4552 if (req->ctx->compat)
4553 sr->msg_flags |= MSG_CMSG_COMPAT;
4556 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4558 ret = io_sendmsg_copy_hdr(req, async_msg);
4560 req->flags |= REQ_F_NEED_CLEANUP;
4564 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4565 struct io_comp_state *cs)
4567 struct io_async_msghdr iomsg, *kmsg;
4568 struct socket *sock;
4572 sock = sock_from_file(req->file, &ret);
4573 if (unlikely(!sock))
4576 if (req->async_data) {
4577 kmsg = req->async_data;
4578 kmsg->msg.msg_name = &kmsg->addr;
4579 /* if iov is set, it's allocated already */
4581 kmsg->iov = kmsg->fast_iov;
4582 kmsg->msg.msg_iter.iov = kmsg->iov;
4584 ret = io_sendmsg_copy_hdr(req, &iomsg);
4590 flags = req->sr_msg.msg_flags;
4591 if (flags & MSG_DONTWAIT)
4592 req->flags |= REQ_F_NOWAIT;
4593 else if (force_nonblock)
4594 flags |= MSG_DONTWAIT;
4596 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4597 if (force_nonblock && ret == -EAGAIN)
4598 return io_setup_async_msg(req, kmsg);
4599 if (ret == -ERESTARTSYS)
4602 if (kmsg->iov != kmsg->fast_iov)
4604 req->flags &= ~REQ_F_NEED_CLEANUP;
4606 req_set_fail_links(req);
4607 __io_req_complete(req, ret, 0, cs);
4611 static int io_send(struct io_kiocb *req, bool force_nonblock,
4612 struct io_comp_state *cs)
4614 struct io_sr_msg *sr = &req->sr_msg;
4617 struct socket *sock;
4621 sock = sock_from_file(req->file, &ret);
4622 if (unlikely(!sock))
4625 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4629 msg.msg_name = NULL;
4630 msg.msg_control = NULL;
4631 msg.msg_controllen = 0;
4632 msg.msg_namelen = 0;
4634 flags = req->sr_msg.msg_flags;
4635 if (flags & MSG_DONTWAIT)
4636 req->flags |= REQ_F_NOWAIT;
4637 else if (force_nonblock)
4638 flags |= MSG_DONTWAIT;
4640 msg.msg_flags = flags;
4641 ret = sock_sendmsg(sock, &msg);
4642 if (force_nonblock && ret == -EAGAIN)
4644 if (ret == -ERESTARTSYS)
4648 req_set_fail_links(req);
4649 __io_req_complete(req, ret, 0, cs);
4653 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4654 struct io_async_msghdr *iomsg)
4656 struct io_sr_msg *sr = &req->sr_msg;
4657 struct iovec __user *uiov;
4661 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4662 &iomsg->uaddr, &uiov, &iov_len);
4666 if (req->flags & REQ_F_BUFFER_SELECT) {
4669 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4671 sr->len = iomsg->iov[0].iov_len;
4672 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4676 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4677 &iomsg->iov, &iomsg->msg.msg_iter,
4686 #ifdef CONFIG_COMPAT
4687 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4688 struct io_async_msghdr *iomsg)
4690 struct compat_msghdr __user *msg_compat;
4691 struct io_sr_msg *sr = &req->sr_msg;
4692 struct compat_iovec __user *uiov;
4697 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4698 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4703 uiov = compat_ptr(ptr);
4704 if (req->flags & REQ_F_BUFFER_SELECT) {
4705 compat_ssize_t clen;
4709 if (!access_ok(uiov, sizeof(*uiov)))
4711 if (__get_user(clen, &uiov->iov_len))
4715 sr->len = iomsg->iov[0].iov_len;
4718 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4719 UIO_FASTIOV, &iomsg->iov,
4720 &iomsg->msg.msg_iter, true);
4729 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4730 struct io_async_msghdr *iomsg)
4732 iomsg->msg.msg_name = &iomsg->addr;
4733 iomsg->iov = iomsg->fast_iov;
4735 #ifdef CONFIG_COMPAT
4736 if (req->ctx->compat)
4737 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4740 return __io_recvmsg_copy_hdr(req, iomsg);
4743 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4746 struct io_sr_msg *sr = &req->sr_msg;
4747 struct io_buffer *kbuf;
4749 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4754 req->flags |= REQ_F_BUFFER_SELECTED;
4758 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4760 return io_put_kbuf(req, req->sr_msg.kbuf);
4763 static int io_recvmsg_prep(struct io_kiocb *req,
4764 const struct io_uring_sqe *sqe)
4766 struct io_async_msghdr *async_msg = req->async_data;
4767 struct io_sr_msg *sr = &req->sr_msg;
4770 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4773 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4774 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4775 sr->len = READ_ONCE(sqe->len);
4776 sr->bgid = READ_ONCE(sqe->buf_group);
4778 #ifdef CONFIG_COMPAT
4779 if (req->ctx->compat)
4780 sr->msg_flags |= MSG_CMSG_COMPAT;
4783 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4785 ret = io_recvmsg_copy_hdr(req, async_msg);
4787 req->flags |= REQ_F_NEED_CLEANUP;
4791 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4792 struct io_comp_state *cs)
4794 struct io_async_msghdr iomsg, *kmsg;
4795 struct socket *sock;
4796 struct io_buffer *kbuf;
4798 int ret, cflags = 0;
4800 sock = sock_from_file(req->file, &ret);
4801 if (unlikely(!sock))
4804 if (req->async_data) {
4805 kmsg = req->async_data;
4806 kmsg->msg.msg_name = &kmsg->addr;
4807 /* if iov is set, it's allocated already */
4809 kmsg->iov = kmsg->fast_iov;
4810 kmsg->msg.msg_iter.iov = kmsg->iov;
4812 ret = io_recvmsg_copy_hdr(req, &iomsg);
4818 if (req->flags & REQ_F_BUFFER_SELECT) {
4819 kbuf = io_recv_buffer_select(req, !force_nonblock);
4821 return PTR_ERR(kbuf);
4822 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4823 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4824 1, req->sr_msg.len);
4827 flags = req->sr_msg.msg_flags;
4828 if (flags & MSG_DONTWAIT)
4829 req->flags |= REQ_F_NOWAIT;
4830 else if (force_nonblock)
4831 flags |= MSG_DONTWAIT;
4833 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4834 kmsg->uaddr, flags);
4835 if (force_nonblock && ret == -EAGAIN)
4836 return io_setup_async_msg(req, kmsg);
4837 if (ret == -ERESTARTSYS)
4840 if (req->flags & REQ_F_BUFFER_SELECTED)
4841 cflags = io_put_recv_kbuf(req);
4842 if (kmsg->iov != kmsg->fast_iov)
4844 req->flags &= ~REQ_F_NEED_CLEANUP;
4846 req_set_fail_links(req);
4847 __io_req_complete(req, ret, cflags, cs);
4851 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4852 struct io_comp_state *cs)
4854 struct io_buffer *kbuf;
4855 struct io_sr_msg *sr = &req->sr_msg;
4857 void __user *buf = sr->buf;
4858 struct socket *sock;
4861 int ret, cflags = 0;
4863 sock = sock_from_file(req->file, &ret);
4864 if (unlikely(!sock))
4867 if (req->flags & REQ_F_BUFFER_SELECT) {
4868 kbuf = io_recv_buffer_select(req, !force_nonblock);
4870 return PTR_ERR(kbuf);
4871 buf = u64_to_user_ptr(kbuf->addr);
4874 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4878 msg.msg_name = NULL;
4879 msg.msg_control = NULL;
4880 msg.msg_controllen = 0;
4881 msg.msg_namelen = 0;
4882 msg.msg_iocb = NULL;
4885 flags = req->sr_msg.msg_flags;
4886 if (flags & MSG_DONTWAIT)
4887 req->flags |= REQ_F_NOWAIT;
4888 else if (force_nonblock)
4889 flags |= MSG_DONTWAIT;
4891 ret = sock_recvmsg(sock, &msg, flags);
4892 if (force_nonblock && ret == -EAGAIN)
4894 if (ret == -ERESTARTSYS)
4897 if (req->flags & REQ_F_BUFFER_SELECTED)
4898 cflags = io_put_recv_kbuf(req);
4900 req_set_fail_links(req);
4901 __io_req_complete(req, ret, cflags, cs);
4905 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4907 struct io_accept *accept = &req->accept;
4909 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4911 if (sqe->ioprio || sqe->len || sqe->buf_index)
4914 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4915 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4916 accept->flags = READ_ONCE(sqe->accept_flags);
4917 accept->nofile = rlimit(RLIMIT_NOFILE);
4921 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4922 struct io_comp_state *cs)
4924 struct io_accept *accept = &req->accept;
4925 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4928 if (req->file->f_flags & O_NONBLOCK)
4929 req->flags |= REQ_F_NOWAIT;
4931 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4932 accept->addr_len, accept->flags,
4934 if (ret == -EAGAIN && force_nonblock)
4937 if (ret == -ERESTARTSYS)
4939 req_set_fail_links(req);
4941 __io_req_complete(req, ret, 0, cs);
4945 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4947 struct io_connect *conn = &req->connect;
4948 struct io_async_connect *io = req->async_data;
4950 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4952 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4955 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4956 conn->addr_len = READ_ONCE(sqe->addr2);
4961 return move_addr_to_kernel(conn->addr, conn->addr_len,
4965 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4966 struct io_comp_state *cs)
4968 struct io_async_connect __io, *io;
4969 unsigned file_flags;
4972 if (req->async_data) {
4973 io = req->async_data;
4975 ret = move_addr_to_kernel(req->connect.addr,
4976 req->connect.addr_len,
4983 file_flags = force_nonblock ? O_NONBLOCK : 0;
4985 ret = __sys_connect_file(req->file, &io->address,
4986 req->connect.addr_len, file_flags);
4987 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4988 if (req->async_data)
4990 if (io_alloc_async_data(req)) {
4994 io = req->async_data;
4995 memcpy(req->async_data, &__io, sizeof(__io));
4998 if (ret == -ERESTARTSYS)
5002 req_set_fail_links(req);
5003 __io_req_complete(req, ret, 0, cs);
5006 #else /* !CONFIG_NET */
5007 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5012 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
5013 struct io_comp_state *cs)
5018 static int io_send(struct io_kiocb *req, bool force_nonblock,
5019 struct io_comp_state *cs)
5024 static int io_recvmsg_prep(struct io_kiocb *req,
5025 const struct io_uring_sqe *sqe)
5030 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
5031 struct io_comp_state *cs)
5036 static int io_recv(struct io_kiocb *req, bool force_nonblock,
5037 struct io_comp_state *cs)
5042 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5047 static int io_accept(struct io_kiocb *req, bool force_nonblock,
5048 struct io_comp_state *cs)
5053 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5058 static int io_connect(struct io_kiocb *req, bool force_nonblock,
5059 struct io_comp_state *cs)
5063 #endif /* CONFIG_NET */
5065 struct io_poll_table {
5066 struct poll_table_struct pt;
5067 struct io_kiocb *req;
5071 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5072 __poll_t mask, task_work_func_t func)
5077 /* for instances that support it check for an event match first: */
5078 if (mask && !(mask & poll->events))
5081 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5083 list_del_init(&poll->wait.entry);
5086 init_task_work(&req->task_work, func);
5087 percpu_ref_get(&req->ctx->refs);
5090 * If we using the signalfd wait_queue_head for this wakeup, then
5091 * it's not safe to use TWA_SIGNAL as we could be recursing on the
5092 * tsk->sighand->siglock on doing the wakeup. Should not be needed
5093 * either, as the normal wakeup will suffice.
5095 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
5098 * If this fails, then the task is exiting. When a task exits, the
5099 * work gets canceled, so just cancel this request as well instead
5100 * of executing it. We can't safely execute it anyway, as we may not
5101 * have the needed state needed for it anyway.
5103 ret = io_req_task_work_add(req, twa_signal_ok);
5104 if (unlikely(ret)) {
5105 struct task_struct *tsk;
5107 WRITE_ONCE(poll->canceled, true);
5108 tsk = io_wq_get_task(req->ctx->io_wq);
5109 task_work_add(tsk, &req->task_work, TWA_NONE);
5110 wake_up_process(tsk);
5115 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5116 __acquires(&req->ctx->completion_lock)
5118 struct io_ring_ctx *ctx = req->ctx;
5120 if (!req->result && !READ_ONCE(poll->canceled)) {
5121 struct poll_table_struct pt = { ._key = poll->events };
5123 req->result = vfs_poll(req->file, &pt) & poll->events;
5126 spin_lock_irq(&ctx->completion_lock);
5127 if (!req->result && !READ_ONCE(poll->canceled)) {
5128 add_wait_queue(poll->head, &poll->wait);
5135 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5137 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5138 if (req->opcode == IORING_OP_POLL_ADD)
5139 return req->async_data;
5140 return req->apoll->double_poll;
5143 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5145 if (req->opcode == IORING_OP_POLL_ADD)
5147 return &req->apoll->poll;
5150 static void io_poll_remove_double(struct io_kiocb *req)
5152 struct io_poll_iocb *poll = io_poll_get_double(req);
5154 lockdep_assert_held(&req->ctx->completion_lock);
5156 if (poll && poll->head) {
5157 struct wait_queue_head *head = poll->head;
5159 spin_lock(&head->lock);
5160 list_del_init(&poll->wait.entry);
5161 if (poll->wait.private)
5162 refcount_dec(&req->refs);
5164 spin_unlock(&head->lock);
5168 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5170 struct io_ring_ctx *ctx = req->ctx;
5172 io_poll_remove_double(req);
5173 req->poll.done = true;
5174 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5175 io_commit_cqring(ctx);
5178 static void io_poll_task_func(struct callback_head *cb)
5180 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5181 struct io_ring_ctx *ctx = req->ctx;
5182 struct io_kiocb *nxt;
5184 if (io_poll_rewait(req, &req->poll)) {
5185 spin_unlock_irq(&ctx->completion_lock);
5187 hash_del(&req->hash_node);
5188 io_poll_complete(req, req->result, 0);
5189 spin_unlock_irq(&ctx->completion_lock);
5191 nxt = io_put_req_find_next(req);
5192 io_cqring_ev_posted(ctx);
5194 __io_req_task_submit(nxt);
5197 percpu_ref_put(&ctx->refs);
5200 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5201 int sync, void *key)
5203 struct io_kiocb *req = wait->private;
5204 struct io_poll_iocb *poll = io_poll_get_single(req);
5205 __poll_t mask = key_to_poll(key);
5207 /* for instances that support it check for an event match first: */
5208 if (mask && !(mask & poll->events))
5211 list_del_init(&wait->entry);
5213 if (poll && poll->head) {
5216 spin_lock(&poll->head->lock);
5217 done = list_empty(&poll->wait.entry);
5219 list_del_init(&poll->wait.entry);
5220 /* make sure double remove sees this as being gone */
5221 wait->private = NULL;
5222 spin_unlock(&poll->head->lock);
5224 /* use wait func handler, so it matches the rq type */
5225 poll->wait.func(&poll->wait, mode, sync, key);
5228 refcount_dec(&req->refs);
5232 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5233 wait_queue_func_t wake_func)
5237 poll->canceled = false;
5238 poll->events = events;
5239 INIT_LIST_HEAD(&poll->wait.entry);
5240 init_waitqueue_func_entry(&poll->wait, wake_func);
5243 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5244 struct wait_queue_head *head,
5245 struct io_poll_iocb **poll_ptr)
5247 struct io_kiocb *req = pt->req;
5250 * If poll->head is already set, it's because the file being polled
5251 * uses multiple waitqueues for poll handling (eg one for read, one
5252 * for write). Setup a separate io_poll_iocb if this happens.
5254 if (unlikely(poll->head)) {
5255 struct io_poll_iocb *poll_one = poll;
5257 /* already have a 2nd entry, fail a third attempt */
5259 pt->error = -EINVAL;
5262 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5264 pt->error = -ENOMEM;
5267 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5268 refcount_inc(&req->refs);
5269 poll->wait.private = req;
5276 if (poll->events & EPOLLEXCLUSIVE)
5277 add_wait_queue_exclusive(head, &poll->wait);
5279 add_wait_queue(head, &poll->wait);
5282 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5283 struct poll_table_struct *p)
5285 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5286 struct async_poll *apoll = pt->req->apoll;
5288 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5291 static void io_async_task_func(struct callback_head *cb)
5293 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5294 struct async_poll *apoll = req->apoll;
5295 struct io_ring_ctx *ctx = req->ctx;
5297 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5299 if (io_poll_rewait(req, &apoll->poll)) {
5300 spin_unlock_irq(&ctx->completion_lock);
5301 percpu_ref_put(&ctx->refs);
5305 /* If req is still hashed, it cannot have been canceled. Don't check. */
5306 if (hash_hashed(&req->hash_node))
5307 hash_del(&req->hash_node);
5309 io_poll_remove_double(req);
5310 spin_unlock_irq(&ctx->completion_lock);
5312 if (!READ_ONCE(apoll->poll.canceled))
5313 __io_req_task_submit(req);
5315 __io_req_task_cancel(req, -ECANCELED);
5317 percpu_ref_put(&ctx->refs);
5318 kfree(apoll->double_poll);
5322 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5325 struct io_kiocb *req = wait->private;
5326 struct io_poll_iocb *poll = &req->apoll->poll;
5328 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5331 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5334 static void io_poll_req_insert(struct io_kiocb *req)
5336 struct io_ring_ctx *ctx = req->ctx;
5337 struct hlist_head *list;
5339 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5340 hlist_add_head(&req->hash_node, list);
5343 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5344 struct io_poll_iocb *poll,
5345 struct io_poll_table *ipt, __poll_t mask,
5346 wait_queue_func_t wake_func)
5347 __acquires(&ctx->completion_lock)
5349 struct io_ring_ctx *ctx = req->ctx;
5350 bool cancel = false;
5352 INIT_HLIST_NODE(&req->hash_node);
5353 io_init_poll_iocb(poll, mask, wake_func);
5354 poll->file = req->file;
5355 poll->wait.private = req;
5357 ipt->pt._key = mask;
5359 ipt->error = -EINVAL;
5361 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5363 spin_lock_irq(&ctx->completion_lock);
5364 if (likely(poll->head)) {
5365 spin_lock(&poll->head->lock);
5366 if (unlikely(list_empty(&poll->wait.entry))) {
5372 if (mask || ipt->error)
5373 list_del_init(&poll->wait.entry);
5375 WRITE_ONCE(poll->canceled, true);
5376 else if (!poll->done) /* actually waiting for an event */
5377 io_poll_req_insert(req);
5378 spin_unlock(&poll->head->lock);
5384 static bool io_arm_poll_handler(struct io_kiocb *req)
5386 const struct io_op_def *def = &io_op_defs[req->opcode];
5387 struct io_ring_ctx *ctx = req->ctx;
5388 struct async_poll *apoll;
5389 struct io_poll_table ipt;
5393 if (!req->file || !file_can_poll(req->file))
5395 if (req->flags & REQ_F_POLLED)
5399 else if (def->pollout)
5403 /* if we can't nonblock try, then no point in arming a poll handler */
5404 if (!io_file_supports_async(req->file, rw))
5407 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5408 if (unlikely(!apoll))
5410 apoll->double_poll = NULL;
5412 req->flags |= REQ_F_POLLED;
5417 mask |= POLLIN | POLLRDNORM;
5419 mask |= POLLOUT | POLLWRNORM;
5421 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5422 if ((req->opcode == IORING_OP_RECVMSG) &&
5423 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5426 mask |= POLLERR | POLLPRI;
5428 ipt.pt._qproc = io_async_queue_proc;
5430 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5432 if (ret || ipt.error) {
5433 io_poll_remove_double(req);
5434 spin_unlock_irq(&ctx->completion_lock);
5435 kfree(apoll->double_poll);
5439 spin_unlock_irq(&ctx->completion_lock);
5440 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5441 apoll->poll.events);
5445 static bool __io_poll_remove_one(struct io_kiocb *req,
5446 struct io_poll_iocb *poll)
5448 bool do_complete = false;
5450 spin_lock(&poll->head->lock);
5451 WRITE_ONCE(poll->canceled, true);
5452 if (!list_empty(&poll->wait.entry)) {
5453 list_del_init(&poll->wait.entry);
5456 spin_unlock(&poll->head->lock);
5457 hash_del(&req->hash_node);
5461 static bool io_poll_remove_one(struct io_kiocb *req)
5465 io_poll_remove_double(req);
5467 if (req->opcode == IORING_OP_POLL_ADD) {
5468 do_complete = __io_poll_remove_one(req, &req->poll);
5470 struct async_poll *apoll = req->apoll;
5472 /* non-poll requests have submit ref still */
5473 do_complete = __io_poll_remove_one(req, &apoll->poll);
5476 kfree(apoll->double_poll);
5482 io_cqring_fill_event(req, -ECANCELED);
5483 io_commit_cqring(req->ctx);
5484 req_set_fail_links(req);
5485 io_put_req_deferred(req, 1);
5492 * Returns true if we found and killed one or more poll requests
5494 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5496 struct hlist_node *tmp;
5497 struct io_kiocb *req;
5500 spin_lock_irq(&ctx->completion_lock);
5501 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5502 struct hlist_head *list;
5504 list = &ctx->cancel_hash[i];
5505 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5506 if (io_task_match(req, tsk))
5507 posted += io_poll_remove_one(req);
5510 spin_unlock_irq(&ctx->completion_lock);
5513 io_cqring_ev_posted(ctx);
5518 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5520 struct hlist_head *list;
5521 struct io_kiocb *req;
5523 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5524 hlist_for_each_entry(req, list, hash_node) {
5525 if (sqe_addr != req->user_data)
5527 if (io_poll_remove_one(req))
5535 static int io_poll_remove_prep(struct io_kiocb *req,
5536 const struct io_uring_sqe *sqe)
5538 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5540 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5544 req->poll_remove.addr = READ_ONCE(sqe->addr);
5549 * Find a running poll command that matches one specified in sqe->addr,
5550 * and remove it if found.
5552 static int io_poll_remove(struct io_kiocb *req)
5554 struct io_ring_ctx *ctx = req->ctx;
5557 spin_lock_irq(&ctx->completion_lock);
5558 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5559 spin_unlock_irq(&ctx->completion_lock);
5562 req_set_fail_links(req);
5563 io_req_complete(req, ret);
5567 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5570 struct io_kiocb *req = wait->private;
5571 struct io_poll_iocb *poll = &req->poll;
5573 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5576 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5577 struct poll_table_struct *p)
5579 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5581 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5584 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5586 struct io_poll_iocb *poll = &req->poll;
5589 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5591 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5594 events = READ_ONCE(sqe->poll32_events);
5596 events = swahw32(events);
5598 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5599 (events & EPOLLEXCLUSIVE);
5603 static int io_poll_add(struct io_kiocb *req)
5605 struct io_poll_iocb *poll = &req->poll;
5606 struct io_ring_ctx *ctx = req->ctx;
5607 struct io_poll_table ipt;
5610 ipt.pt._qproc = io_poll_queue_proc;
5612 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5615 if (mask) { /* no async, we'd stolen it */
5617 io_poll_complete(req, mask, 0);
5619 spin_unlock_irq(&ctx->completion_lock);
5622 io_cqring_ev_posted(ctx);
5628 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5630 struct io_timeout_data *data = container_of(timer,
5631 struct io_timeout_data, timer);
5632 struct io_kiocb *req = data->req;
5633 struct io_ring_ctx *ctx = req->ctx;
5634 unsigned long flags;
5636 spin_lock_irqsave(&ctx->completion_lock, flags);
5637 list_del_init(&req->timeout.list);
5638 atomic_set(&req->ctx->cq_timeouts,
5639 atomic_read(&req->ctx->cq_timeouts) + 1);
5641 io_cqring_fill_event(req, -ETIME);
5642 io_commit_cqring(ctx);
5643 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5645 io_cqring_ev_posted(ctx);
5646 req_set_fail_links(req);
5648 return HRTIMER_NORESTART;
5651 static int __io_timeout_cancel(struct io_kiocb *req)
5653 struct io_timeout_data *io = req->async_data;
5656 ret = hrtimer_try_to_cancel(&io->timer);
5659 list_del_init(&req->timeout.list);
5661 req_set_fail_links(req);
5662 io_cqring_fill_event(req, -ECANCELED);
5663 io_put_req_deferred(req, 1);
5667 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5669 struct io_kiocb *req;
5672 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5673 if (user_data == req->user_data) {
5682 return __io_timeout_cancel(req);
5685 static int io_timeout_remove_prep(struct io_kiocb *req,
5686 const struct io_uring_sqe *sqe)
5688 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5690 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5692 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5695 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5700 * Remove or update an existing timeout command
5702 static int io_timeout_remove(struct io_kiocb *req)
5704 struct io_ring_ctx *ctx = req->ctx;
5707 spin_lock_irq(&ctx->completion_lock);
5708 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5710 io_cqring_fill_event(req, ret);
5711 io_commit_cqring(ctx);
5712 spin_unlock_irq(&ctx->completion_lock);
5713 io_cqring_ev_posted(ctx);
5715 req_set_fail_links(req);
5720 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5721 bool is_timeout_link)
5723 struct io_timeout_data *data;
5725 u32 off = READ_ONCE(sqe->off);
5727 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5729 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5731 if (off && is_timeout_link)
5733 flags = READ_ONCE(sqe->timeout_flags);
5734 if (flags & ~IORING_TIMEOUT_ABS)
5737 req->timeout.off = off;
5739 if (!req->async_data && io_alloc_async_data(req))
5742 data = req->async_data;
5745 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5748 if (flags & IORING_TIMEOUT_ABS)
5749 data->mode = HRTIMER_MODE_ABS;
5751 data->mode = HRTIMER_MODE_REL;
5753 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5757 static int io_timeout(struct io_kiocb *req)
5759 struct io_ring_ctx *ctx = req->ctx;
5760 struct io_timeout_data *data = req->async_data;
5761 struct list_head *entry;
5762 u32 tail, off = req->timeout.off;
5764 spin_lock_irq(&ctx->completion_lock);
5767 * sqe->off holds how many events that need to occur for this
5768 * timeout event to be satisfied. If it isn't set, then this is
5769 * a pure timeout request, sequence isn't used.
5771 if (io_is_timeout_noseq(req)) {
5772 entry = ctx->timeout_list.prev;
5776 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5777 req->timeout.target_seq = tail + off;
5780 * Insertion sort, ensuring the first entry in the list is always
5781 * the one we need first.
5783 list_for_each_prev(entry, &ctx->timeout_list) {
5784 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5787 if (io_is_timeout_noseq(nxt))
5789 /* nxt.seq is behind @tail, otherwise would've been completed */
5790 if (off >= nxt->timeout.target_seq - tail)
5794 list_add(&req->timeout.list, entry);
5795 data->timer.function = io_timeout_fn;
5796 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5797 spin_unlock_irq(&ctx->completion_lock);
5801 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5803 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5805 return req->user_data == (unsigned long) data;
5808 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5810 enum io_wq_cancel cancel_ret;
5813 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5814 switch (cancel_ret) {
5815 case IO_WQ_CANCEL_OK:
5818 case IO_WQ_CANCEL_RUNNING:
5821 case IO_WQ_CANCEL_NOTFOUND:
5829 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5830 struct io_kiocb *req, __u64 sqe_addr,
5833 unsigned long flags;
5836 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5837 if (ret != -ENOENT) {
5838 spin_lock_irqsave(&ctx->completion_lock, flags);
5842 spin_lock_irqsave(&ctx->completion_lock, flags);
5843 ret = io_timeout_cancel(ctx, sqe_addr);
5846 ret = io_poll_cancel(ctx, sqe_addr);
5850 io_cqring_fill_event(req, ret);
5851 io_commit_cqring(ctx);
5852 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5853 io_cqring_ev_posted(ctx);
5856 req_set_fail_links(req);
5860 static int io_async_cancel_prep(struct io_kiocb *req,
5861 const struct io_uring_sqe *sqe)
5863 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5865 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5867 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5870 req->cancel.addr = READ_ONCE(sqe->addr);
5874 static int io_async_cancel(struct io_kiocb *req)
5876 struct io_ring_ctx *ctx = req->ctx;
5878 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5882 static int io_files_update_prep(struct io_kiocb *req,
5883 const struct io_uring_sqe *sqe)
5885 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5887 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5889 if (sqe->ioprio || sqe->rw_flags)
5892 req->files_update.offset = READ_ONCE(sqe->off);
5893 req->files_update.nr_args = READ_ONCE(sqe->len);
5894 if (!req->files_update.nr_args)
5896 req->files_update.arg = READ_ONCE(sqe->addr);
5900 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5901 struct io_comp_state *cs)
5903 struct io_ring_ctx *ctx = req->ctx;
5904 struct io_uring_files_update up;
5910 up.offset = req->files_update.offset;
5911 up.fds = req->files_update.arg;
5913 mutex_lock(&ctx->uring_lock);
5914 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5915 mutex_unlock(&ctx->uring_lock);
5918 req_set_fail_links(req);
5919 __io_req_complete(req, ret, 0, cs);
5923 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5925 switch (req->opcode) {
5928 case IORING_OP_READV:
5929 case IORING_OP_READ_FIXED:
5930 case IORING_OP_READ:
5931 return io_read_prep(req, sqe);
5932 case IORING_OP_WRITEV:
5933 case IORING_OP_WRITE_FIXED:
5934 case IORING_OP_WRITE:
5935 return io_write_prep(req, sqe);
5936 case IORING_OP_POLL_ADD:
5937 return io_poll_add_prep(req, sqe);
5938 case IORING_OP_POLL_REMOVE:
5939 return io_poll_remove_prep(req, sqe);
5940 case IORING_OP_FSYNC:
5941 return io_prep_fsync(req, sqe);
5942 case IORING_OP_SYNC_FILE_RANGE:
5943 return io_prep_sfr(req, sqe);
5944 case IORING_OP_SENDMSG:
5945 case IORING_OP_SEND:
5946 return io_sendmsg_prep(req, sqe);
5947 case IORING_OP_RECVMSG:
5948 case IORING_OP_RECV:
5949 return io_recvmsg_prep(req, sqe);
5950 case IORING_OP_CONNECT:
5951 return io_connect_prep(req, sqe);
5952 case IORING_OP_TIMEOUT:
5953 return io_timeout_prep(req, sqe, false);
5954 case IORING_OP_TIMEOUT_REMOVE:
5955 return io_timeout_remove_prep(req, sqe);
5956 case IORING_OP_ASYNC_CANCEL:
5957 return io_async_cancel_prep(req, sqe);
5958 case IORING_OP_LINK_TIMEOUT:
5959 return io_timeout_prep(req, sqe, true);
5960 case IORING_OP_ACCEPT:
5961 return io_accept_prep(req, sqe);
5962 case IORING_OP_FALLOCATE:
5963 return io_fallocate_prep(req, sqe);
5964 case IORING_OP_OPENAT:
5965 return io_openat_prep(req, sqe);
5966 case IORING_OP_CLOSE:
5967 return io_close_prep(req, sqe);
5968 case IORING_OP_FILES_UPDATE:
5969 return io_files_update_prep(req, sqe);
5970 case IORING_OP_STATX:
5971 return io_statx_prep(req, sqe);
5972 case IORING_OP_FADVISE:
5973 return io_fadvise_prep(req, sqe);
5974 case IORING_OP_MADVISE:
5975 return io_madvise_prep(req, sqe);
5976 case IORING_OP_OPENAT2:
5977 return io_openat2_prep(req, sqe);
5978 case IORING_OP_EPOLL_CTL:
5979 return io_epoll_ctl_prep(req, sqe);
5980 case IORING_OP_SPLICE:
5981 return io_splice_prep(req, sqe);
5982 case IORING_OP_PROVIDE_BUFFERS:
5983 return io_provide_buffers_prep(req, sqe);
5984 case IORING_OP_REMOVE_BUFFERS:
5985 return io_remove_buffers_prep(req, sqe);
5987 return io_tee_prep(req, sqe);
5988 case IORING_OP_SHUTDOWN:
5989 return io_shutdown_prep(req, sqe);
5990 case IORING_OP_RENAMEAT:
5991 return io_renameat_prep(req, sqe);
5992 case IORING_OP_UNLINKAT:
5993 return io_unlinkat_prep(req, sqe);
5996 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6001 static int io_req_defer_prep(struct io_kiocb *req,
6002 const struct io_uring_sqe *sqe)
6006 if (io_alloc_async_data(req))
6008 return io_req_prep(req, sqe);
6011 static u32 io_get_sequence(struct io_kiocb *req)
6013 struct io_kiocb *pos;
6014 struct io_ring_ctx *ctx = req->ctx;
6015 u32 total_submitted, nr_reqs = 1;
6017 if (req->flags & REQ_F_LINK_HEAD)
6018 list_for_each_entry(pos, &req->link_list, link_list)
6021 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6022 return total_submitted - nr_reqs;
6025 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6027 struct io_ring_ctx *ctx = req->ctx;
6028 struct io_defer_entry *de;
6032 /* Still need defer if there is pending req in defer list. */
6033 if (likely(list_empty_careful(&ctx->defer_list) &&
6034 !(req->flags & REQ_F_IO_DRAIN)))
6037 seq = io_get_sequence(req);
6038 /* Still a chance to pass the sequence check */
6039 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6042 if (!req->async_data) {
6043 ret = io_req_defer_prep(req, sqe);
6047 io_prep_async_link(req);
6048 de = kmalloc(sizeof(*de), GFP_KERNEL);
6052 spin_lock_irq(&ctx->completion_lock);
6053 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6054 spin_unlock_irq(&ctx->completion_lock);
6056 io_queue_async_work(req);
6057 return -EIOCBQUEUED;
6060 trace_io_uring_defer(ctx, req, req->user_data);
6063 list_add_tail(&de->list, &ctx->defer_list);
6064 spin_unlock_irq(&ctx->completion_lock);
6065 return -EIOCBQUEUED;
6068 static void io_req_drop_files(struct io_kiocb *req)
6070 struct io_ring_ctx *ctx = req->ctx;
6071 unsigned long flags;
6073 spin_lock_irqsave(&ctx->inflight_lock, flags);
6074 list_del(&req->inflight_entry);
6075 if (waitqueue_active(&ctx->inflight_wait))
6076 wake_up(&ctx->inflight_wait);
6077 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
6078 req->flags &= ~REQ_F_INFLIGHT;
6079 put_files_struct(req->work.identity->files);
6080 put_nsproxy(req->work.identity->nsproxy);
6081 req->work.flags &= ~IO_WQ_WORK_FILES;
6084 static void __io_clean_op(struct io_kiocb *req)
6086 if (req->flags & REQ_F_BUFFER_SELECTED) {
6087 switch (req->opcode) {
6088 case IORING_OP_READV:
6089 case IORING_OP_READ_FIXED:
6090 case IORING_OP_READ:
6091 kfree((void *)(unsigned long)req->rw.addr);
6093 case IORING_OP_RECVMSG:
6094 case IORING_OP_RECV:
6095 kfree(req->sr_msg.kbuf);
6098 req->flags &= ~REQ_F_BUFFER_SELECTED;
6101 if (req->flags & REQ_F_NEED_CLEANUP) {
6102 switch (req->opcode) {
6103 case IORING_OP_READV:
6104 case IORING_OP_READ_FIXED:
6105 case IORING_OP_READ:
6106 case IORING_OP_WRITEV:
6107 case IORING_OP_WRITE_FIXED:
6108 case IORING_OP_WRITE: {
6109 struct io_async_rw *io = req->async_data;
6111 kfree(io->free_iovec);
6114 case IORING_OP_RECVMSG:
6115 case IORING_OP_SENDMSG: {
6116 struct io_async_msghdr *io = req->async_data;
6117 if (io->iov != io->fast_iov)
6121 case IORING_OP_SPLICE:
6123 io_put_file(req, req->splice.file_in,
6124 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6126 case IORING_OP_OPENAT:
6127 case IORING_OP_OPENAT2:
6128 if (req->open.filename)
6129 putname(req->open.filename);
6131 case IORING_OP_RENAMEAT:
6132 putname(req->rename.oldpath);
6133 putname(req->rename.newpath);
6135 case IORING_OP_UNLINKAT:
6136 putname(req->unlink.filename);
6139 req->flags &= ~REQ_F_NEED_CLEANUP;
6142 if (req->flags & REQ_F_INFLIGHT)
6143 io_req_drop_files(req);
6146 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
6147 struct io_comp_state *cs)
6149 struct io_ring_ctx *ctx = req->ctx;
6152 switch (req->opcode) {
6154 ret = io_nop(req, cs);
6156 case IORING_OP_READV:
6157 case IORING_OP_READ_FIXED:
6158 case IORING_OP_READ:
6159 ret = io_read(req, force_nonblock, cs);
6161 case IORING_OP_WRITEV:
6162 case IORING_OP_WRITE_FIXED:
6163 case IORING_OP_WRITE:
6164 ret = io_write(req, force_nonblock, cs);
6166 case IORING_OP_FSYNC:
6167 ret = io_fsync(req, force_nonblock);
6169 case IORING_OP_POLL_ADD:
6170 ret = io_poll_add(req);
6172 case IORING_OP_POLL_REMOVE:
6173 ret = io_poll_remove(req);
6175 case IORING_OP_SYNC_FILE_RANGE:
6176 ret = io_sync_file_range(req, force_nonblock);
6178 case IORING_OP_SENDMSG:
6179 ret = io_sendmsg(req, force_nonblock, cs);
6181 case IORING_OP_SEND:
6182 ret = io_send(req, force_nonblock, cs);
6184 case IORING_OP_RECVMSG:
6185 ret = io_recvmsg(req, force_nonblock, cs);
6187 case IORING_OP_RECV:
6188 ret = io_recv(req, force_nonblock, cs);
6190 case IORING_OP_TIMEOUT:
6191 ret = io_timeout(req);
6193 case IORING_OP_TIMEOUT_REMOVE:
6194 ret = io_timeout_remove(req);
6196 case IORING_OP_ACCEPT:
6197 ret = io_accept(req, force_nonblock, cs);
6199 case IORING_OP_CONNECT:
6200 ret = io_connect(req, force_nonblock, cs);
6202 case IORING_OP_ASYNC_CANCEL:
6203 ret = io_async_cancel(req);
6205 case IORING_OP_FALLOCATE:
6206 ret = io_fallocate(req, force_nonblock);
6208 case IORING_OP_OPENAT:
6209 ret = io_openat(req, force_nonblock);
6211 case IORING_OP_CLOSE:
6212 ret = io_close(req, force_nonblock, cs);
6214 case IORING_OP_FILES_UPDATE:
6215 ret = io_files_update(req, force_nonblock, cs);
6217 case IORING_OP_STATX:
6218 ret = io_statx(req, force_nonblock);
6220 case IORING_OP_FADVISE:
6221 ret = io_fadvise(req, force_nonblock);
6223 case IORING_OP_MADVISE:
6224 ret = io_madvise(req, force_nonblock);
6226 case IORING_OP_OPENAT2:
6227 ret = io_openat2(req, force_nonblock);
6229 case IORING_OP_EPOLL_CTL:
6230 ret = io_epoll_ctl(req, force_nonblock, cs);
6232 case IORING_OP_SPLICE:
6233 ret = io_splice(req, force_nonblock);
6235 case IORING_OP_PROVIDE_BUFFERS:
6236 ret = io_provide_buffers(req, force_nonblock, cs);
6238 case IORING_OP_REMOVE_BUFFERS:
6239 ret = io_remove_buffers(req, force_nonblock, cs);
6242 ret = io_tee(req, force_nonblock);
6244 case IORING_OP_SHUTDOWN:
6245 ret = io_shutdown(req, force_nonblock);
6247 case IORING_OP_RENAMEAT:
6248 ret = io_renameat(req, force_nonblock);
6250 case IORING_OP_UNLINKAT:
6251 ret = io_unlinkat(req, force_nonblock);
6261 /* If the op doesn't have a file, we're not polling for it */
6262 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6263 const bool in_async = io_wq_current_is_worker();
6265 /* workqueue context doesn't hold uring_lock, grab it now */
6267 mutex_lock(&ctx->uring_lock);
6269 io_iopoll_req_issued(req);
6272 mutex_unlock(&ctx->uring_lock);
6278 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6280 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6281 struct io_kiocb *timeout;
6284 timeout = io_prep_linked_timeout(req);
6286 io_queue_linked_timeout(timeout);
6288 /* if NO_CANCEL is set, we must still run the work */
6289 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6290 IO_WQ_WORK_CANCEL) {
6296 ret = io_issue_sqe(req, false, NULL);
6298 * We can get EAGAIN for polled IO even though we're
6299 * forcing a sync submission from here, since we can't
6300 * wait for request slots on the block side.
6309 req_set_fail_links(req);
6310 io_req_complete(req, ret);
6313 return io_steal_work(req);
6316 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6319 struct fixed_file_table *table;
6321 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6322 return table->files[index & IORING_FILE_TABLE_MASK];
6325 static struct file *io_file_get(struct io_submit_state *state,
6326 struct io_kiocb *req, int fd, bool fixed)
6328 struct io_ring_ctx *ctx = req->ctx;
6332 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6334 fd = array_index_nospec(fd, ctx->nr_user_files);
6335 file = io_file_from_index(ctx, fd);
6337 req->fixed_file_refs = &ctx->file_data->node->refs;
6338 percpu_ref_get(req->fixed_file_refs);
6341 trace_io_uring_file_get(ctx, fd);
6342 file = __io_file_get(state, fd);
6348 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6351 req->file = io_file_get(state, req, fd, req->flags & REQ_F_FIXED_FILE);
6352 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6357 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6359 struct io_timeout_data *data = container_of(timer,
6360 struct io_timeout_data, timer);
6361 struct io_kiocb *req = data->req;
6362 struct io_ring_ctx *ctx = req->ctx;
6363 struct io_kiocb *prev = NULL;
6364 unsigned long flags;
6366 spin_lock_irqsave(&ctx->completion_lock, flags);
6369 * We don't expect the list to be empty, that will only happen if we
6370 * race with the completion of the linked work.
6372 if (!list_empty(&req->link_list)) {
6373 prev = list_entry(req->link_list.prev, struct io_kiocb,
6375 if (refcount_inc_not_zero(&prev->refs))
6376 list_del_init(&req->link_list);
6381 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6384 req_set_fail_links(prev);
6385 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6388 io_req_complete(req, -ETIME);
6390 return HRTIMER_NORESTART;
6393 static void __io_queue_linked_timeout(struct io_kiocb *req)
6396 * If the list is now empty, then our linked request finished before
6397 * we got a chance to setup the timer
6399 if (!list_empty(&req->link_list)) {
6400 struct io_timeout_data *data = req->async_data;
6402 data->timer.function = io_link_timeout_fn;
6403 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6408 static void io_queue_linked_timeout(struct io_kiocb *req)
6410 struct io_ring_ctx *ctx = req->ctx;
6412 spin_lock_irq(&ctx->completion_lock);
6413 __io_queue_linked_timeout(req);
6414 spin_unlock_irq(&ctx->completion_lock);
6416 /* drop submission reference */
6420 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6422 struct io_kiocb *nxt;
6424 if (!(req->flags & REQ_F_LINK_HEAD))
6426 if (req->flags & REQ_F_LINK_TIMEOUT)
6429 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6431 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6434 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6435 req->flags |= REQ_F_LINK_TIMEOUT;
6439 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6441 struct io_kiocb *linked_timeout;
6442 const struct cred *old_creds = NULL;
6446 linked_timeout = io_prep_linked_timeout(req);
6448 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6449 (req->work.flags & IO_WQ_WORK_CREDS) &&
6450 req->work.identity->creds != current_cred()) {
6452 revert_creds(old_creds);
6453 if (old_creds == req->work.identity->creds)
6454 old_creds = NULL; /* restored original creds */
6456 old_creds = override_creds(req->work.identity->creds);
6459 ret = io_issue_sqe(req, true, cs);
6462 * We async punt it if the file wasn't marked NOWAIT, or if the file
6463 * doesn't support non-blocking read/write attempts
6465 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6466 if (!io_arm_poll_handler(req)) {
6468 * Queued up for async execution, worker will release
6469 * submit reference when the iocb is actually submitted.
6471 io_queue_async_work(req);
6475 io_queue_linked_timeout(linked_timeout);
6476 } else if (likely(!ret)) {
6477 /* drop submission reference */
6478 req = io_put_req_find_next(req);
6480 io_queue_linked_timeout(linked_timeout);
6483 if (!(req->flags & REQ_F_FORCE_ASYNC))
6485 io_queue_async_work(req);
6488 /* un-prep timeout, so it'll be killed as any other linked */
6489 req->flags &= ~REQ_F_LINK_TIMEOUT;
6490 req_set_fail_links(req);
6492 io_req_complete(req, ret);
6496 revert_creds(old_creds);
6499 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6500 struct io_comp_state *cs)
6504 ret = io_req_defer(req, sqe);
6506 if (ret != -EIOCBQUEUED) {
6508 req_set_fail_links(req);
6510 io_req_complete(req, ret);
6512 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6513 if (!req->async_data) {
6514 ret = io_req_defer_prep(req, sqe);
6518 io_queue_async_work(req);
6521 ret = io_req_prep(req, sqe);
6525 __io_queue_sqe(req, cs);
6529 static inline void io_queue_link_head(struct io_kiocb *req,
6530 struct io_comp_state *cs)
6532 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6534 io_req_complete(req, -ECANCELED);
6536 io_queue_sqe(req, NULL, cs);
6539 struct io_submit_link {
6540 struct io_kiocb *head;
6541 struct io_kiocb *last;
6544 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6545 struct io_submit_link *link, struct io_comp_state *cs)
6547 struct io_ring_ctx *ctx = req->ctx;
6551 * If we already have a head request, queue this one for async
6552 * submittal once the head completes. If we don't have a head but
6553 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6554 * submitted sync once the chain is complete. If none of those
6555 * conditions are true (normal request), then just queue it.
6558 struct io_kiocb *head = link->head;
6561 * Taking sequential execution of a link, draining both sides
6562 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6563 * requests in the link. So, it drains the head and the
6564 * next after the link request. The last one is done via
6565 * drain_next flag to persist the effect across calls.
6567 if (req->flags & REQ_F_IO_DRAIN) {
6568 head->flags |= REQ_F_IO_DRAIN;
6569 ctx->drain_next = 1;
6571 ret = io_req_defer_prep(req, sqe);
6572 if (unlikely(ret)) {
6573 /* fail even hard links since we don't submit */
6574 head->flags |= REQ_F_FAIL_LINK;
6577 trace_io_uring_link(ctx, req, head);
6578 list_add_tail(&req->link_list, &head->link_list);
6581 /* last request of a link, enqueue the link */
6582 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6583 io_queue_link_head(head, cs);
6587 if (unlikely(ctx->drain_next)) {
6588 req->flags |= REQ_F_IO_DRAIN;
6589 ctx->drain_next = 0;
6591 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6592 req->flags |= REQ_F_LINK_HEAD;
6593 INIT_LIST_HEAD(&req->link_list);
6595 ret = io_req_defer_prep(req, sqe);
6597 req->flags |= REQ_F_FAIL_LINK;
6601 io_queue_sqe(req, sqe, cs);
6609 * Batched submission is done, ensure local IO is flushed out.
6611 static void io_submit_state_end(struct io_submit_state *state)
6613 if (!list_empty(&state->comp.list))
6614 io_submit_flush_completions(&state->comp);
6615 blk_finish_plug(&state->plug);
6616 io_state_file_put(state);
6617 if (state->free_reqs)
6618 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6622 * Start submission side cache.
6624 static void io_submit_state_start(struct io_submit_state *state,
6625 struct io_ring_ctx *ctx, unsigned int max_ios)
6627 blk_start_plug(&state->plug);
6629 INIT_LIST_HEAD(&state->comp.list);
6630 state->comp.ctx = ctx;
6631 state->free_reqs = 0;
6633 state->ios_left = max_ios;
6636 static void io_commit_sqring(struct io_ring_ctx *ctx)
6638 struct io_rings *rings = ctx->rings;
6641 * Ensure any loads from the SQEs are done at this point,
6642 * since once we write the new head, the application could
6643 * write new data to them.
6645 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6649 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6650 * that is mapped by userspace. This means that care needs to be taken to
6651 * ensure that reads are stable, as we cannot rely on userspace always
6652 * being a good citizen. If members of the sqe are validated and then later
6653 * used, it's important that those reads are done through READ_ONCE() to
6654 * prevent a re-load down the line.
6656 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6658 u32 *sq_array = ctx->sq_array;
6662 * The cached sq head (or cq tail) serves two purposes:
6664 * 1) allows us to batch the cost of updating the user visible
6666 * 2) allows the kernel side to track the head on its own, even
6667 * though the application is the one updating it.
6669 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6670 if (likely(head < ctx->sq_entries))
6671 return &ctx->sq_sqes[head];
6673 /* drop invalid entries */
6674 ctx->cached_sq_dropped++;
6675 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6679 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6681 ctx->cached_sq_head++;
6685 * Check SQE restrictions (opcode and flags).
6687 * Returns 'true' if SQE is allowed, 'false' otherwise.
6689 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6690 struct io_kiocb *req,
6691 unsigned int sqe_flags)
6693 if (!ctx->restricted)
6696 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6699 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6700 ctx->restrictions.sqe_flags_required)
6703 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6704 ctx->restrictions.sqe_flags_required))
6710 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6711 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6712 IOSQE_BUFFER_SELECT)
6714 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6715 const struct io_uring_sqe *sqe,
6716 struct io_submit_state *state)
6718 unsigned int sqe_flags;
6721 req->opcode = READ_ONCE(sqe->opcode);
6722 req->user_data = READ_ONCE(sqe->user_data);
6723 req->async_data = NULL;
6727 /* one is dropped after submission, the other at completion */
6728 refcount_set(&req->refs, 2);
6729 req->task = current;
6732 if (unlikely(req->opcode >= IORING_OP_LAST))
6735 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6738 sqe_flags = READ_ONCE(sqe->flags);
6739 /* enforce forwards compatibility on users */
6740 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6743 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6746 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6747 !io_op_defs[req->opcode].buffer_select)
6750 id = READ_ONCE(sqe->personality);
6752 struct io_identity *iod;
6754 iod = idr_find(&ctx->personality_idr, id);
6757 refcount_inc(&iod->count);
6759 __io_req_init_async(req);
6760 get_cred(iod->creds);
6761 req->work.identity = iod;
6762 req->work.flags |= IO_WQ_WORK_CREDS;
6765 /* same numerical values with corresponding REQ_F_*, safe to copy */
6766 req->flags |= sqe_flags;
6768 if (!io_op_defs[req->opcode].needs_file)
6771 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6776 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6778 struct io_submit_state state;
6779 struct io_submit_link link;
6780 int i, submitted = 0;
6782 /* if we have a backlog and couldn't flush it all, return BUSY */
6783 if (test_bit(0, &ctx->sq_check_overflow)) {
6784 if (!list_empty(&ctx->cq_overflow_list) &&
6785 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6789 /* make sure SQ entry isn't read before tail */
6790 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6792 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6795 percpu_counter_add(¤t->io_uring->inflight, nr);
6796 refcount_add(nr, ¤t->usage);
6798 io_submit_state_start(&state, ctx, nr);
6801 for (i = 0; i < nr; i++) {
6802 const struct io_uring_sqe *sqe;
6803 struct io_kiocb *req;
6806 sqe = io_get_sqe(ctx);
6807 if (unlikely(!sqe)) {
6808 io_consume_sqe(ctx);
6811 req = io_alloc_req(ctx, &state);
6812 if (unlikely(!req)) {
6814 submitted = -EAGAIN;
6817 io_consume_sqe(ctx);
6818 /* will complete beyond this point, count as submitted */
6821 err = io_init_req(ctx, req, sqe, &state);
6822 if (unlikely(err)) {
6825 io_req_complete(req, err);
6829 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6830 true, io_async_submit(ctx));
6831 err = io_submit_sqe(req, sqe, &link, &state.comp);
6836 if (unlikely(submitted != nr)) {
6837 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6838 struct io_uring_task *tctx = current->io_uring;
6839 int unused = nr - ref_used;
6841 percpu_ref_put_many(&ctx->refs, unused);
6842 percpu_counter_sub(&tctx->inflight, unused);
6843 put_task_struct_many(current, unused);
6846 io_queue_link_head(link.head, &state.comp);
6847 io_submit_state_end(&state);
6849 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6850 io_commit_sqring(ctx);
6855 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6857 /* Tell userspace we may need a wakeup call */
6858 spin_lock_irq(&ctx->completion_lock);
6859 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6860 spin_unlock_irq(&ctx->completion_lock);
6863 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6865 spin_lock_irq(&ctx->completion_lock);
6866 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6867 spin_unlock_irq(&ctx->completion_lock);
6870 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6871 int sync, void *key)
6873 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6876 ret = autoremove_wake_function(wqe, mode, sync, key);
6878 unsigned long flags;
6880 spin_lock_irqsave(&ctx->completion_lock, flags);
6881 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6882 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6893 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6894 unsigned long start_jiffies, bool cap_entries)
6896 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6897 struct io_sq_data *sqd = ctx->sq_data;
6898 unsigned int to_submit;
6902 if (!list_empty(&ctx->iopoll_list)) {
6903 unsigned nr_events = 0;
6905 mutex_lock(&ctx->uring_lock);
6906 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6907 io_do_iopoll(ctx, &nr_events, 0);
6908 mutex_unlock(&ctx->uring_lock);
6911 to_submit = io_sqring_entries(ctx);
6914 * If submit got -EBUSY, flag us as needing the application
6915 * to enter the kernel to reap and flush events.
6917 if (!to_submit || ret == -EBUSY || need_resched()) {
6919 * Drop cur_mm before scheduling, we can't hold it for
6920 * long periods (or over schedule()). Do this before
6921 * adding ourselves to the waitqueue, as the unuse/drop
6924 io_sq_thread_drop_mm_files();
6927 * We're polling. If we're within the defined idle
6928 * period, then let us spin without work before going
6929 * to sleep. The exception is if we got EBUSY doing
6930 * more IO, we should wait for the application to
6931 * reap events and wake us up.
6933 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6934 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6935 !percpu_ref_is_dying(&ctx->refs)))
6938 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6939 TASK_INTERRUPTIBLE);
6942 * While doing polled IO, before going to sleep, we need
6943 * to check if there are new reqs added to iopoll_list,
6944 * it is because reqs may have been punted to io worker
6945 * and will be added to iopoll_list later, hence check
6946 * the iopoll_list again.
6948 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6949 !list_empty_careful(&ctx->iopoll_list)) {
6950 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6954 to_submit = io_sqring_entries(ctx);
6955 if (!to_submit || ret == -EBUSY)
6959 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6960 io_ring_clear_wakeup_flag(ctx);
6962 /* if we're handling multiple rings, cap submit size for fairness */
6963 if (cap_entries && to_submit > 8)
6966 mutex_lock(&ctx->uring_lock);
6967 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6968 ret = io_submit_sqes(ctx, to_submit);
6969 mutex_unlock(&ctx->uring_lock);
6971 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6972 wake_up(&ctx->sqo_sq_wait);
6974 return SQT_DID_WORK;
6977 static void io_sqd_init_new(struct io_sq_data *sqd)
6979 struct io_ring_ctx *ctx;
6981 while (!list_empty(&sqd->ctx_new_list)) {
6982 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6983 init_wait(&ctx->sqo_wait_entry);
6984 ctx->sqo_wait_entry.func = io_sq_wake_function;
6985 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6986 complete(&ctx->sq_thread_comp);
6990 static int io_sq_thread(void *data)
6992 struct cgroup_subsys_state *cur_css = NULL;
6993 struct files_struct *old_files = current->files;
6994 struct nsproxy *old_nsproxy = current->nsproxy;
6995 const struct cred *old_cred = NULL;
6996 struct io_sq_data *sqd = data;
6997 struct io_ring_ctx *ctx;
6998 unsigned long start_jiffies;
7001 current->files = NULL;
7002 current->nsproxy = NULL;
7003 task_unlock(current);
7005 start_jiffies = jiffies;
7006 while (!kthread_should_stop()) {
7007 enum sq_ret ret = 0;
7011 * Any changes to the sqd lists are synchronized through the
7012 * kthread parking. This synchronizes the thread vs users,
7013 * the users are synchronized on the sqd->ctx_lock.
7015 if (kthread_should_park())
7018 if (unlikely(!list_empty(&sqd->ctx_new_list)))
7019 io_sqd_init_new(sqd);
7021 cap_entries = !list_is_singular(&sqd->ctx_list);
7023 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7024 if (current->cred != ctx->creds) {
7026 revert_creds(old_cred);
7027 old_cred = override_creds(ctx->creds);
7029 io_sq_thread_associate_blkcg(ctx, &cur_css);
7031 current->loginuid = ctx->loginuid;
7032 current->sessionid = ctx->sessionid;
7035 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
7037 io_sq_thread_drop_mm_files();
7040 if (ret & SQT_SPIN) {
7043 } else if (ret == SQT_IDLE) {
7044 if (kthread_should_park())
7046 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7047 io_ring_set_wakeup_flag(ctx);
7049 start_jiffies = jiffies;
7050 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7051 io_ring_clear_wakeup_flag(ctx);
7058 io_sq_thread_unassociate_blkcg();
7060 revert_creds(old_cred);
7063 current->files = old_files;
7064 current->nsproxy = old_nsproxy;
7065 task_unlock(current);
7072 struct io_wait_queue {
7073 struct wait_queue_entry wq;
7074 struct io_ring_ctx *ctx;
7076 unsigned nr_timeouts;
7079 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
7081 struct io_ring_ctx *ctx = iowq->ctx;
7084 * Wake up if we have enough events, or if a timeout occurred since we
7085 * started waiting. For timeouts, we always want to return to userspace,
7086 * regardless of event count.
7088 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
7089 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7092 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7093 int wake_flags, void *key)
7095 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7098 /* use noflush == true, as we can't safely rely on locking context */
7099 if (!io_should_wake(iowq, true))
7102 return autoremove_wake_function(curr, mode, wake_flags, key);
7105 static int io_run_task_work_sig(void)
7107 if (io_run_task_work())
7109 if (!signal_pending(current))
7111 if (current->jobctl & JOBCTL_TASK_WORK) {
7112 spin_lock_irq(¤t->sighand->siglock);
7113 current->jobctl &= ~JOBCTL_TASK_WORK;
7114 recalc_sigpending();
7115 spin_unlock_irq(¤t->sighand->siglock);
7122 * Wait until events become available, if we don't already have some. The
7123 * application must reap them itself, as they reside on the shared cq ring.
7125 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7126 const sigset_t __user *sig, size_t sigsz)
7128 struct io_wait_queue iowq = {
7131 .func = io_wake_function,
7132 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7135 .to_wait = min_events,
7137 struct io_rings *rings = ctx->rings;
7141 if (io_cqring_events(ctx, false) >= min_events)
7143 if (!io_run_task_work())
7148 #ifdef CONFIG_COMPAT
7149 if (in_compat_syscall())
7150 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7154 ret = set_user_sigmask(sig, sigsz);
7160 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7161 trace_io_uring_cqring_wait(ctx, min_events);
7163 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7164 TASK_INTERRUPTIBLE);
7165 /* make sure we run task_work before checking for signals */
7166 ret = io_run_task_work_sig();
7171 if (io_should_wake(&iowq, false))
7175 finish_wait(&ctx->wait, &iowq.wq);
7177 restore_saved_sigmask_unless(ret == -EINTR);
7179 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7182 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7184 #if defined(CONFIG_UNIX)
7185 if (ctx->ring_sock) {
7186 struct sock *sock = ctx->ring_sock->sk;
7187 struct sk_buff *skb;
7189 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7195 for (i = 0; i < ctx->nr_user_files; i++) {
7198 file = io_file_from_index(ctx, i);
7205 static void io_file_ref_kill(struct percpu_ref *ref)
7207 struct fixed_file_data *data;
7209 data = container_of(ref, struct fixed_file_data, refs);
7210 complete(&data->done);
7213 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7215 struct fixed_file_data *data = ctx->file_data;
7216 struct fixed_file_ref_node *ref_node = NULL;
7217 unsigned nr_tables, i;
7222 spin_lock(&data->lock);
7223 ref_node = data->node;
7224 spin_unlock(&data->lock);
7226 percpu_ref_kill(&ref_node->refs);
7228 percpu_ref_kill(&data->refs);
7230 /* wait for all refs nodes to complete */
7231 flush_delayed_work(&ctx->file_put_work);
7232 wait_for_completion(&data->done);
7234 __io_sqe_files_unregister(ctx);
7235 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7236 for (i = 0; i < nr_tables; i++)
7237 kfree(data->table[i].files);
7239 percpu_ref_exit(&data->refs);
7241 ctx->file_data = NULL;
7242 ctx->nr_user_files = 0;
7246 static void io_put_sq_data(struct io_sq_data *sqd)
7248 if (refcount_dec_and_test(&sqd->refs)) {
7250 * The park is a bit of a work-around, without it we get
7251 * warning spews on shutdown with SQPOLL set and affinity
7252 * set to a single CPU.
7255 kthread_park(sqd->thread);
7256 kthread_stop(sqd->thread);
7263 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7265 struct io_ring_ctx *ctx_attach;
7266 struct io_sq_data *sqd;
7269 f = fdget(p->wq_fd);
7271 return ERR_PTR(-ENXIO);
7272 if (f.file->f_op != &io_uring_fops) {
7274 return ERR_PTR(-EINVAL);
7277 ctx_attach = f.file->private_data;
7278 sqd = ctx_attach->sq_data;
7281 return ERR_PTR(-EINVAL);
7284 refcount_inc(&sqd->refs);
7289 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7291 struct io_sq_data *sqd;
7293 if (p->flags & IORING_SETUP_ATTACH_WQ)
7294 return io_attach_sq_data(p);
7296 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7298 return ERR_PTR(-ENOMEM);
7300 refcount_set(&sqd->refs, 1);
7301 INIT_LIST_HEAD(&sqd->ctx_list);
7302 INIT_LIST_HEAD(&sqd->ctx_new_list);
7303 mutex_init(&sqd->ctx_lock);
7304 mutex_init(&sqd->lock);
7305 init_waitqueue_head(&sqd->wait);
7309 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7310 __releases(&sqd->lock)
7314 kthread_unpark(sqd->thread);
7315 mutex_unlock(&sqd->lock);
7318 static void io_sq_thread_park(struct io_sq_data *sqd)
7319 __acquires(&sqd->lock)
7323 mutex_lock(&sqd->lock);
7324 kthread_park(sqd->thread);
7327 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7329 struct io_sq_data *sqd = ctx->sq_data;
7334 * We may arrive here from the error branch in
7335 * io_sq_offload_create() where the kthread is created
7336 * without being waked up, thus wake it up now to make
7337 * sure the wait will complete.
7339 wake_up_process(sqd->thread);
7340 wait_for_completion(&ctx->sq_thread_comp);
7342 io_sq_thread_park(sqd);
7345 mutex_lock(&sqd->ctx_lock);
7346 list_del(&ctx->sqd_list);
7347 mutex_unlock(&sqd->ctx_lock);
7350 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7351 io_sq_thread_unpark(sqd);
7354 io_put_sq_data(sqd);
7355 ctx->sq_data = NULL;
7359 static void io_finish_async(struct io_ring_ctx *ctx)
7361 io_sq_thread_stop(ctx);
7364 io_wq_destroy(ctx->io_wq);
7369 #if defined(CONFIG_UNIX)
7371 * Ensure the UNIX gc is aware of our file set, so we are certain that
7372 * the io_uring can be safely unregistered on process exit, even if we have
7373 * loops in the file referencing.
7375 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7377 struct sock *sk = ctx->ring_sock->sk;
7378 struct scm_fp_list *fpl;
7379 struct sk_buff *skb;
7382 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7386 skb = alloc_skb(0, GFP_KERNEL);
7395 fpl->user = get_uid(ctx->user);
7396 for (i = 0; i < nr; i++) {
7397 struct file *file = io_file_from_index(ctx, i + offset);
7401 fpl->fp[nr_files] = get_file(file);
7402 unix_inflight(fpl->user, fpl->fp[nr_files]);
7407 fpl->max = SCM_MAX_FD;
7408 fpl->count = nr_files;
7409 UNIXCB(skb).fp = fpl;
7410 skb->destructor = unix_destruct_scm;
7411 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7412 skb_queue_head(&sk->sk_receive_queue, skb);
7414 for (i = 0; i < nr_files; i++)
7425 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7426 * causes regular reference counting to break down. We rely on the UNIX
7427 * garbage collection to take care of this problem for us.
7429 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7431 unsigned left, total;
7435 left = ctx->nr_user_files;
7437 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7439 ret = __io_sqe_files_scm(ctx, this_files, total);
7443 total += this_files;
7449 while (total < ctx->nr_user_files) {
7450 struct file *file = io_file_from_index(ctx, total);
7460 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7466 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7467 unsigned nr_tables, unsigned nr_files)
7471 for (i = 0; i < nr_tables; i++) {
7472 struct fixed_file_table *table = &file_data->table[i];
7473 unsigned this_files;
7475 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7476 table->files = kcalloc(this_files, sizeof(struct file *),
7480 nr_files -= this_files;
7486 for (i = 0; i < nr_tables; i++) {
7487 struct fixed_file_table *table = &file_data->table[i];
7488 kfree(table->files);
7493 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7495 #if defined(CONFIG_UNIX)
7496 struct sock *sock = ctx->ring_sock->sk;
7497 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7498 struct sk_buff *skb;
7501 __skb_queue_head_init(&list);
7504 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7505 * remove this entry and rearrange the file array.
7507 skb = skb_dequeue(head);
7509 struct scm_fp_list *fp;
7511 fp = UNIXCB(skb).fp;
7512 for (i = 0; i < fp->count; i++) {
7515 if (fp->fp[i] != file)
7518 unix_notinflight(fp->user, fp->fp[i]);
7519 left = fp->count - 1 - i;
7521 memmove(&fp->fp[i], &fp->fp[i + 1],
7522 left * sizeof(struct file *));
7529 __skb_queue_tail(&list, skb);
7539 __skb_queue_tail(&list, skb);
7541 skb = skb_dequeue(head);
7544 if (skb_peek(&list)) {
7545 spin_lock_irq(&head->lock);
7546 while ((skb = __skb_dequeue(&list)) != NULL)
7547 __skb_queue_tail(head, skb);
7548 spin_unlock_irq(&head->lock);
7555 struct io_file_put {
7556 struct list_head list;
7560 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7562 struct fixed_file_data *file_data = ref_node->file_data;
7563 struct io_ring_ctx *ctx = file_data->ctx;
7564 struct io_file_put *pfile, *tmp;
7566 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7567 list_del(&pfile->list);
7568 io_ring_file_put(ctx, pfile->file);
7572 percpu_ref_exit(&ref_node->refs);
7574 percpu_ref_put(&file_data->refs);
7577 static void io_file_put_work(struct work_struct *work)
7579 struct io_ring_ctx *ctx;
7580 struct llist_node *node;
7582 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7583 node = llist_del_all(&ctx->file_put_llist);
7586 struct fixed_file_ref_node *ref_node;
7587 struct llist_node *next = node->next;
7589 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7590 __io_file_put_work(ref_node);
7595 static void io_file_data_ref_zero(struct percpu_ref *ref)
7597 struct fixed_file_ref_node *ref_node;
7598 struct fixed_file_data *data;
7599 struct io_ring_ctx *ctx;
7600 bool first_add = false;
7603 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7604 data = ref_node->file_data;
7607 spin_lock(&data->lock);
7608 ref_node->done = true;
7610 while (!list_empty(&data->ref_list)) {
7611 ref_node = list_first_entry(&data->ref_list,
7612 struct fixed_file_ref_node, node);
7613 /* recycle ref nodes in order */
7614 if (!ref_node->done)
7616 list_del(&ref_node->node);
7617 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7619 spin_unlock(&data->lock);
7621 if (percpu_ref_is_dying(&data->refs))
7625 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7627 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7630 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7631 struct io_ring_ctx *ctx)
7633 struct fixed_file_ref_node *ref_node;
7635 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7637 return ERR_PTR(-ENOMEM);
7639 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7642 return ERR_PTR(-ENOMEM);
7644 INIT_LIST_HEAD(&ref_node->node);
7645 INIT_LIST_HEAD(&ref_node->file_list);
7646 ref_node->file_data = ctx->file_data;
7647 ref_node->done = false;
7651 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7653 percpu_ref_exit(&ref_node->refs);
7657 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7660 __s32 __user *fds = (__s32 __user *) arg;
7661 unsigned nr_tables, i;
7663 int fd, ret = -ENOMEM;
7664 struct fixed_file_ref_node *ref_node;
7665 struct fixed_file_data *file_data;
7671 if (nr_args > IORING_MAX_FIXED_FILES)
7674 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7677 file_data->ctx = ctx;
7678 init_completion(&file_data->done);
7679 INIT_LIST_HEAD(&file_data->ref_list);
7680 spin_lock_init(&file_data->lock);
7682 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7683 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7685 if (!file_data->table)
7688 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7689 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7692 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7694 ctx->file_data = file_data;
7696 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7697 struct fixed_file_table *table;
7700 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7704 /* allow sparse sets */
7714 * Don't allow io_uring instances to be registered. If UNIX
7715 * isn't enabled, then this causes a reference cycle and this
7716 * instance can never get freed. If UNIX is enabled we'll
7717 * handle it just fine, but there's still no point in allowing
7718 * a ring fd as it doesn't support regular read/write anyway.
7720 if (file->f_op == &io_uring_fops) {
7724 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7725 index = i & IORING_FILE_TABLE_MASK;
7726 table->files[index] = file;
7729 ret = io_sqe_files_scm(ctx);
7731 io_sqe_files_unregister(ctx);
7735 ref_node = alloc_fixed_file_ref_node(ctx);
7736 if (IS_ERR(ref_node)) {
7737 io_sqe_files_unregister(ctx);
7738 return PTR_ERR(ref_node);
7741 file_data->node = ref_node;
7742 spin_lock(&file_data->lock);
7743 list_add_tail(&ref_node->node, &file_data->ref_list);
7744 spin_unlock(&file_data->lock);
7745 percpu_ref_get(&file_data->refs);
7748 for (i = 0; i < ctx->nr_user_files; i++) {
7749 file = io_file_from_index(ctx, i);
7753 for (i = 0; i < nr_tables; i++)
7754 kfree(file_data->table[i].files);
7755 ctx->nr_user_files = 0;
7757 percpu_ref_exit(&file_data->refs);
7759 kfree(file_data->table);
7761 ctx->file_data = NULL;
7765 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7768 #if defined(CONFIG_UNIX)
7769 struct sock *sock = ctx->ring_sock->sk;
7770 struct sk_buff_head *head = &sock->sk_receive_queue;
7771 struct sk_buff *skb;
7774 * See if we can merge this file into an existing skb SCM_RIGHTS
7775 * file set. If there's no room, fall back to allocating a new skb
7776 * and filling it in.
7778 spin_lock_irq(&head->lock);
7779 skb = skb_peek(head);
7781 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7783 if (fpl->count < SCM_MAX_FD) {
7784 __skb_unlink(skb, head);
7785 spin_unlock_irq(&head->lock);
7786 fpl->fp[fpl->count] = get_file(file);
7787 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7789 spin_lock_irq(&head->lock);
7790 __skb_queue_head(head, skb);
7795 spin_unlock_irq(&head->lock);
7802 return __io_sqe_files_scm(ctx, 1, index);
7808 static int io_queue_file_removal(struct fixed_file_data *data,
7811 struct io_file_put *pfile;
7812 struct fixed_file_ref_node *ref_node = data->node;
7814 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7819 list_add(&pfile->list, &ref_node->file_list);
7824 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7825 struct io_uring_files_update *up,
7828 struct fixed_file_data *data = ctx->file_data;
7829 struct fixed_file_ref_node *ref_node;
7834 bool needs_switch = false;
7836 if (check_add_overflow(up->offset, nr_args, &done))
7838 if (done > ctx->nr_user_files)
7841 ref_node = alloc_fixed_file_ref_node(ctx);
7842 if (IS_ERR(ref_node))
7843 return PTR_ERR(ref_node);
7846 fds = u64_to_user_ptr(up->fds);
7848 struct fixed_file_table *table;
7852 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7856 i = array_index_nospec(up->offset, ctx->nr_user_files);
7857 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7858 index = i & IORING_FILE_TABLE_MASK;
7859 if (table->files[index]) {
7860 file = table->files[index];
7861 err = io_queue_file_removal(data, file);
7864 table->files[index] = NULL;
7865 needs_switch = true;
7874 * Don't allow io_uring instances to be registered. If
7875 * UNIX isn't enabled, then this causes a reference
7876 * cycle and this instance can never get freed. If UNIX
7877 * is enabled we'll handle it just fine, but there's
7878 * still no point in allowing a ring fd as it doesn't
7879 * support regular read/write anyway.
7881 if (file->f_op == &io_uring_fops) {
7886 table->files[index] = file;
7887 err = io_sqe_file_register(ctx, file, i);
7889 table->files[index] = NULL;
7900 percpu_ref_kill(&data->node->refs);
7901 spin_lock(&data->lock);
7902 list_add_tail(&ref_node->node, &data->ref_list);
7903 data->node = ref_node;
7904 spin_unlock(&data->lock);
7905 percpu_ref_get(&ctx->file_data->refs);
7907 destroy_fixed_file_ref_node(ref_node);
7909 return done ? done : err;
7912 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7915 struct io_uring_files_update up;
7917 if (!ctx->file_data)
7921 if (copy_from_user(&up, arg, sizeof(up)))
7926 return __io_sqe_files_update(ctx, &up, nr_args);
7929 static void io_free_work(struct io_wq_work *work)
7931 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7933 /* Consider that io_steal_work() relies on this ref */
7937 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7938 struct io_uring_params *p)
7940 struct io_wq_data data;
7942 struct io_ring_ctx *ctx_attach;
7943 unsigned int concurrency;
7946 data.user = ctx->user;
7947 data.free_work = io_free_work;
7948 data.do_work = io_wq_submit_work;
7950 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7951 /* Do QD, or 4 * CPUS, whatever is smallest */
7952 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7954 ctx->io_wq = io_wq_create(concurrency, &data);
7955 if (IS_ERR(ctx->io_wq)) {
7956 ret = PTR_ERR(ctx->io_wq);
7962 f = fdget(p->wq_fd);
7966 if (f.file->f_op != &io_uring_fops) {
7971 ctx_attach = f.file->private_data;
7972 /* @io_wq is protected by holding the fd */
7973 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7978 ctx->io_wq = ctx_attach->io_wq;
7984 static int io_uring_alloc_task_context(struct task_struct *task)
7986 struct io_uring_task *tctx;
7989 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7990 if (unlikely(!tctx))
7993 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7994 if (unlikely(ret)) {
8000 init_waitqueue_head(&tctx->wait);
8002 atomic_set(&tctx->in_idle, 0);
8003 tctx->sqpoll = false;
8004 io_init_identity(&tctx->__identity);
8005 tctx->identity = &tctx->__identity;
8006 task->io_uring = tctx;
8010 void __io_uring_free(struct task_struct *tsk)
8012 struct io_uring_task *tctx = tsk->io_uring;
8014 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8015 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8016 if (tctx->identity != &tctx->__identity)
8017 kfree(tctx->identity);
8018 percpu_counter_destroy(&tctx->inflight);
8020 tsk->io_uring = NULL;
8023 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8024 struct io_uring_params *p)
8028 if (ctx->flags & IORING_SETUP_SQPOLL) {
8029 struct io_sq_data *sqd;
8032 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8035 sqd = io_get_sq_data(p);
8042 io_sq_thread_park(sqd);
8043 mutex_lock(&sqd->ctx_lock);
8044 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8045 mutex_unlock(&sqd->ctx_lock);
8046 io_sq_thread_unpark(sqd);
8048 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8049 if (!ctx->sq_thread_idle)
8050 ctx->sq_thread_idle = HZ;
8055 if (p->flags & IORING_SETUP_SQ_AFF) {
8056 int cpu = p->sq_thread_cpu;
8059 if (cpu >= nr_cpu_ids)
8061 if (!cpu_online(cpu))
8064 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8065 cpu, "io_uring-sq");
8067 sqd->thread = kthread_create(io_sq_thread, sqd,
8070 if (IS_ERR(sqd->thread)) {
8071 ret = PTR_ERR(sqd->thread);
8075 ret = io_uring_alloc_task_context(sqd->thread);
8078 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8079 /* Can't have SQ_AFF without SQPOLL */
8085 ret = io_init_wq_offload(ctx, p);
8091 io_finish_async(ctx);
8095 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8097 struct io_sq_data *sqd = ctx->sq_data;
8099 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8100 wake_up_process(sqd->thread);
8103 static inline void __io_unaccount_mem(struct user_struct *user,
8104 unsigned long nr_pages)
8106 atomic_long_sub(nr_pages, &user->locked_vm);
8109 static inline int __io_account_mem(struct user_struct *user,
8110 unsigned long nr_pages)
8112 unsigned long page_limit, cur_pages, new_pages;
8114 /* Don't allow more pages than we can safely lock */
8115 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8118 cur_pages = atomic_long_read(&user->locked_vm);
8119 new_pages = cur_pages + nr_pages;
8120 if (new_pages > page_limit)
8122 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8123 new_pages) != cur_pages);
8128 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8129 enum io_mem_account acct)
8132 __io_unaccount_mem(ctx->user, nr_pages);
8134 if (ctx->mm_account) {
8135 if (acct == ACCT_LOCKED)
8136 ctx->mm_account->locked_vm -= nr_pages;
8137 else if (acct == ACCT_PINNED)
8138 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8142 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8143 enum io_mem_account acct)
8147 if (ctx->limit_mem) {
8148 ret = __io_account_mem(ctx->user, nr_pages);
8153 if (ctx->mm_account) {
8154 if (acct == ACCT_LOCKED)
8155 ctx->mm_account->locked_vm += nr_pages;
8156 else if (acct == ACCT_PINNED)
8157 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8163 static void io_mem_free(void *ptr)
8170 page = virt_to_head_page(ptr);
8171 if (put_page_testzero(page))
8172 free_compound_page(page);
8175 static void *io_mem_alloc(size_t size)
8177 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8180 return (void *) __get_free_pages(gfp_flags, get_order(size));
8183 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8186 struct io_rings *rings;
8187 size_t off, sq_array_size;
8189 off = struct_size(rings, cqes, cq_entries);
8190 if (off == SIZE_MAX)
8194 off = ALIGN(off, SMP_CACHE_BYTES);
8202 sq_array_size = array_size(sizeof(u32), sq_entries);
8203 if (sq_array_size == SIZE_MAX)
8206 if (check_add_overflow(off, sq_array_size, &off))
8212 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8216 pages = (size_t)1 << get_order(
8217 rings_size(sq_entries, cq_entries, NULL));
8218 pages += (size_t)1 << get_order(
8219 array_size(sizeof(struct io_uring_sqe), sq_entries));
8224 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8228 if (!ctx->user_bufs)
8231 for (i = 0; i < ctx->nr_user_bufs; i++) {
8232 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8234 for (j = 0; j < imu->nr_bvecs; j++)
8235 unpin_user_page(imu->bvec[j].bv_page);
8237 if (imu->acct_pages)
8238 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8243 kfree(ctx->user_bufs);
8244 ctx->user_bufs = NULL;
8245 ctx->nr_user_bufs = 0;
8249 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8250 void __user *arg, unsigned index)
8252 struct iovec __user *src;
8254 #ifdef CONFIG_COMPAT
8256 struct compat_iovec __user *ciovs;
8257 struct compat_iovec ciov;
8259 ciovs = (struct compat_iovec __user *) arg;
8260 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8263 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8264 dst->iov_len = ciov.iov_len;
8268 src = (struct iovec __user *) arg;
8269 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8275 * Not super efficient, but this is just a registration time. And we do cache
8276 * the last compound head, so generally we'll only do a full search if we don't
8279 * We check if the given compound head page has already been accounted, to
8280 * avoid double accounting it. This allows us to account the full size of the
8281 * page, not just the constituent pages of a huge page.
8283 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8284 int nr_pages, struct page *hpage)
8288 /* check current page array */
8289 for (i = 0; i < nr_pages; i++) {
8290 if (!PageCompound(pages[i]))
8292 if (compound_head(pages[i]) == hpage)
8296 /* check previously registered pages */
8297 for (i = 0; i < ctx->nr_user_bufs; i++) {
8298 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8300 for (j = 0; j < imu->nr_bvecs; j++) {
8301 if (!PageCompound(imu->bvec[j].bv_page))
8303 if (compound_head(imu->bvec[j].bv_page) == hpage)
8311 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8312 int nr_pages, struct io_mapped_ubuf *imu,
8313 struct page **last_hpage)
8317 for (i = 0; i < nr_pages; i++) {
8318 if (!PageCompound(pages[i])) {
8323 hpage = compound_head(pages[i]);
8324 if (hpage == *last_hpage)
8326 *last_hpage = hpage;
8327 if (headpage_already_acct(ctx, pages, i, hpage))
8329 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8333 if (!imu->acct_pages)
8336 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8338 imu->acct_pages = 0;
8342 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8345 struct vm_area_struct **vmas = NULL;
8346 struct page **pages = NULL;
8347 struct page *last_hpage = NULL;
8348 int i, j, got_pages = 0;
8353 if (!nr_args || nr_args > UIO_MAXIOV)
8356 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8358 if (!ctx->user_bufs)
8361 for (i = 0; i < nr_args; i++) {
8362 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8363 unsigned long off, start, end, ubuf;
8368 ret = io_copy_iov(ctx, &iov, arg, i);
8373 * Don't impose further limits on the size and buffer
8374 * constraints here, we'll -EINVAL later when IO is
8375 * submitted if they are wrong.
8378 if (!iov.iov_base || !iov.iov_len)
8381 /* arbitrary limit, but we need something */
8382 if (iov.iov_len > SZ_1G)
8385 ubuf = (unsigned long) iov.iov_base;
8386 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8387 start = ubuf >> PAGE_SHIFT;
8388 nr_pages = end - start;
8391 if (!pages || nr_pages > got_pages) {
8394 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8396 vmas = kvmalloc_array(nr_pages,
8397 sizeof(struct vm_area_struct *),
8399 if (!pages || !vmas) {
8403 got_pages = nr_pages;
8406 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8413 mmap_read_lock(current->mm);
8414 pret = pin_user_pages(ubuf, nr_pages,
8415 FOLL_WRITE | FOLL_LONGTERM,
8417 if (pret == nr_pages) {
8418 /* don't support file backed memory */
8419 for (j = 0; j < nr_pages; j++) {
8420 struct vm_area_struct *vma = vmas[j];
8423 !is_file_hugepages(vma->vm_file)) {
8429 ret = pret < 0 ? pret : -EFAULT;
8431 mmap_read_unlock(current->mm);
8434 * if we did partial map, or found file backed vmas,
8435 * release any pages we did get
8438 unpin_user_pages(pages, pret);
8443 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8445 unpin_user_pages(pages, pret);
8450 off = ubuf & ~PAGE_MASK;
8452 for (j = 0; j < nr_pages; j++) {
8455 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8456 imu->bvec[j].bv_page = pages[j];
8457 imu->bvec[j].bv_len = vec_len;
8458 imu->bvec[j].bv_offset = off;
8462 /* store original address for later verification */
8464 imu->len = iov.iov_len;
8465 imu->nr_bvecs = nr_pages;
8467 ctx->nr_user_bufs++;
8475 io_sqe_buffer_unregister(ctx);
8479 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8481 __s32 __user *fds = arg;
8487 if (copy_from_user(&fd, fds, sizeof(*fds)))
8490 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8491 if (IS_ERR(ctx->cq_ev_fd)) {
8492 int ret = PTR_ERR(ctx->cq_ev_fd);
8493 ctx->cq_ev_fd = NULL;
8500 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8502 if (ctx->cq_ev_fd) {
8503 eventfd_ctx_put(ctx->cq_ev_fd);
8504 ctx->cq_ev_fd = NULL;
8511 static int __io_destroy_buffers(int id, void *p, void *data)
8513 struct io_ring_ctx *ctx = data;
8514 struct io_buffer *buf = p;
8516 __io_remove_buffers(ctx, buf, id, -1U);
8520 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8522 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8523 idr_destroy(&ctx->io_buffer_idr);
8526 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8528 io_finish_async(ctx);
8529 io_sqe_buffer_unregister(ctx);
8531 if (ctx->sqo_task) {
8532 put_task_struct(ctx->sqo_task);
8533 ctx->sqo_task = NULL;
8534 mmdrop(ctx->mm_account);
8535 ctx->mm_account = NULL;
8538 #ifdef CONFIG_BLK_CGROUP
8539 if (ctx->sqo_blkcg_css)
8540 css_put(ctx->sqo_blkcg_css);
8543 io_sqe_files_unregister(ctx);
8544 io_eventfd_unregister(ctx);
8545 io_destroy_buffers(ctx);
8546 idr_destroy(&ctx->personality_idr);
8548 #if defined(CONFIG_UNIX)
8549 if (ctx->ring_sock) {
8550 ctx->ring_sock->file = NULL; /* so that iput() is called */
8551 sock_release(ctx->ring_sock);
8555 io_mem_free(ctx->rings);
8556 io_mem_free(ctx->sq_sqes);
8558 percpu_ref_exit(&ctx->refs);
8559 free_uid(ctx->user);
8560 put_cred(ctx->creds);
8561 kfree(ctx->cancel_hash);
8562 kmem_cache_free(req_cachep, ctx->fallback_req);
8566 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8568 struct io_ring_ctx *ctx = file->private_data;
8571 poll_wait(file, &ctx->cq_wait, wait);
8573 * synchronizes with barrier from wq_has_sleeper call in
8577 if (!io_sqring_full(ctx))
8578 mask |= EPOLLOUT | EPOLLWRNORM;
8579 if (io_cqring_events(ctx, false))
8580 mask |= EPOLLIN | EPOLLRDNORM;
8585 static int io_uring_fasync(int fd, struct file *file, int on)
8587 struct io_ring_ctx *ctx = file->private_data;
8589 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8592 static int io_remove_personalities(int id, void *p, void *data)
8594 struct io_ring_ctx *ctx = data;
8595 struct io_identity *iod;
8597 iod = idr_remove(&ctx->personality_idr, id);
8599 put_cred(iod->creds);
8600 if (refcount_dec_and_test(&iod->count))
8606 static void io_ring_exit_work(struct work_struct *work)
8608 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8612 * If we're doing polled IO and end up having requests being
8613 * submitted async (out-of-line), then completions can come in while
8614 * we're waiting for refs to drop. We need to reap these manually,
8615 * as nobody else will be looking for them.
8619 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8620 io_iopoll_try_reap_events(ctx);
8621 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8622 io_ring_ctx_free(ctx);
8625 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8627 mutex_lock(&ctx->uring_lock);
8628 percpu_ref_kill(&ctx->refs);
8629 mutex_unlock(&ctx->uring_lock);
8631 io_kill_timeouts(ctx, NULL);
8632 io_poll_remove_all(ctx, NULL);
8635 io_wq_cancel_all(ctx->io_wq);
8637 /* if we failed setting up the ctx, we might not have any rings */
8639 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8640 io_iopoll_try_reap_events(ctx);
8641 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8644 * Do this upfront, so we won't have a grace period where the ring
8645 * is closed but resources aren't reaped yet. This can cause
8646 * spurious failure in setting up a new ring.
8648 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8651 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8653 * Use system_unbound_wq to avoid spawning tons of event kworkers
8654 * if we're exiting a ton of rings at the same time. It just adds
8655 * noise and overhead, there's no discernable change in runtime
8656 * over using system_wq.
8658 queue_work(system_unbound_wq, &ctx->exit_work);
8661 static int io_uring_release(struct inode *inode, struct file *file)
8663 struct io_ring_ctx *ctx = file->private_data;
8665 file->private_data = NULL;
8666 io_ring_ctx_wait_and_kill(ctx);
8670 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8672 struct files_struct *files = data;
8674 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8675 work->identity->files == files);
8679 * Returns true if 'preq' is the link parent of 'req'
8681 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8683 struct io_kiocb *link;
8685 if (!(preq->flags & REQ_F_LINK_HEAD))
8688 list_for_each_entry(link, &preq->link_list, link_list) {
8697 * We're looking to cancel 'req' because it's holding on to our files, but
8698 * 'req' could be a link to another request. See if it is, and cancel that
8699 * parent request if so.
8701 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8703 struct hlist_node *tmp;
8704 struct io_kiocb *preq;
8708 spin_lock_irq(&ctx->completion_lock);
8709 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8710 struct hlist_head *list;
8712 list = &ctx->cancel_hash[i];
8713 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8714 found = io_match_link(preq, req);
8716 io_poll_remove_one(preq);
8721 spin_unlock_irq(&ctx->completion_lock);
8725 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8726 struct io_kiocb *req)
8728 struct io_kiocb *preq;
8731 spin_lock_irq(&ctx->completion_lock);
8732 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8733 found = io_match_link(preq, req);
8735 __io_timeout_cancel(preq);
8739 spin_unlock_irq(&ctx->completion_lock);
8743 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8745 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8748 if (req->flags & REQ_F_LINK_TIMEOUT) {
8749 unsigned long flags;
8750 struct io_ring_ctx *ctx = req->ctx;
8752 /* protect against races with linked timeouts */
8753 spin_lock_irqsave(&ctx->completion_lock, flags);
8754 ret = io_match_link(req, data);
8755 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8757 ret = io_match_link(req, data);
8762 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8764 enum io_wq_cancel cret;
8766 /* cancel this particular work, if it's running */
8767 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8768 if (cret != IO_WQ_CANCEL_NOTFOUND)
8771 /* find links that hold this pending, cancel those */
8772 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8773 if (cret != IO_WQ_CANCEL_NOTFOUND)
8776 /* if we have a poll link holding this pending, cancel that */
8777 if (io_poll_remove_link(ctx, req))
8780 /* final option, timeout link is holding this req pending */
8781 io_timeout_remove_link(ctx, req);
8784 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8785 struct task_struct *task,
8786 struct files_struct *files)
8788 struct io_defer_entry *de = NULL;
8791 spin_lock_irq(&ctx->completion_lock);
8792 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8793 if (io_task_match(de->req, task) &&
8794 io_match_files(de->req, files)) {
8795 list_cut_position(&list, &ctx->defer_list, &de->list);
8799 spin_unlock_irq(&ctx->completion_lock);
8801 while (!list_empty(&list)) {
8802 de = list_first_entry(&list, struct io_defer_entry, list);
8803 list_del_init(&de->list);
8804 req_set_fail_links(de->req);
8805 io_put_req(de->req);
8806 io_req_complete(de->req, -ECANCELED);
8812 * Returns true if we found and killed one or more files pinning requests
8814 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8815 struct files_struct *files)
8817 if (list_empty_careful(&ctx->inflight_list))
8820 /* cancel all at once, should be faster than doing it one by one*/
8821 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8823 while (!list_empty_careful(&ctx->inflight_list)) {
8824 struct io_kiocb *cancel_req = NULL, *req;
8827 spin_lock_irq(&ctx->inflight_lock);
8828 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8829 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8830 req->work.identity->files != files)
8832 /* req is being completed, ignore */
8833 if (!refcount_inc_not_zero(&req->refs))
8839 prepare_to_wait(&ctx->inflight_wait, &wait,
8840 TASK_UNINTERRUPTIBLE);
8841 spin_unlock_irq(&ctx->inflight_lock);
8843 /* We need to keep going until we don't find a matching req */
8846 /* cancel this request, or head link requests */
8847 io_attempt_cancel(ctx, cancel_req);
8848 io_put_req(cancel_req);
8849 /* cancellations _may_ trigger task work */
8852 finish_wait(&ctx->inflight_wait, &wait);
8858 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8860 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8861 struct task_struct *task = data;
8863 return io_task_match(req, task);
8866 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8867 struct task_struct *task,
8868 struct files_struct *files)
8872 ret = io_uring_cancel_files(ctx, files);
8874 enum io_wq_cancel cret;
8876 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8877 if (cret != IO_WQ_CANCEL_NOTFOUND)
8880 /* SQPOLL thread does its own polling */
8881 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8882 while (!list_empty_careful(&ctx->iopoll_list)) {
8883 io_iopoll_try_reap_events(ctx);
8888 ret |= io_poll_remove_all(ctx, task);
8889 ret |= io_kill_timeouts(ctx, task);
8896 * We need to iteratively cancel requests, in case a request has dependent
8897 * hard links. These persist even for failure of cancelations, hence keep
8898 * looping until none are found.
8900 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8901 struct files_struct *files)
8903 struct task_struct *task = current;
8905 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8906 task = ctx->sq_data->thread;
8907 atomic_inc(&task->io_uring->in_idle);
8908 io_sq_thread_park(ctx->sq_data);
8912 io_cancel_defer_files(ctx, NULL, files);
8914 io_cancel_defer_files(ctx, task, NULL);
8916 io_cqring_overflow_flush(ctx, true, task, files);
8918 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8923 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8924 atomic_dec(&task->io_uring->in_idle);
8926 * If the files that are going away are the ones in the thread
8927 * identity, clear them out.
8929 if (task->io_uring->identity->files == files)
8930 task->io_uring->identity->files = NULL;
8931 io_sq_thread_unpark(ctx->sq_data);
8936 * Note that this task has used io_uring. We use it for cancelation purposes.
8938 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8940 struct io_uring_task *tctx = current->io_uring;
8942 if (unlikely(!tctx)) {
8945 ret = io_uring_alloc_task_context(current);
8948 tctx = current->io_uring;
8950 if (tctx->last != file) {
8951 void *old = xa_load(&tctx->xa, (unsigned long)file);
8955 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8961 * This is race safe in that the task itself is doing this, hence it
8962 * cannot be going through the exit/cancel paths at the same time.
8963 * This cannot be modified while exit/cancel is running.
8965 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8966 tctx->sqpoll = true;
8972 * Remove this io_uring_file -> task mapping.
8974 static void io_uring_del_task_file(struct file *file)
8976 struct io_uring_task *tctx = current->io_uring;
8978 if (tctx->last == file)
8980 file = xa_erase(&tctx->xa, (unsigned long)file);
8986 * Drop task note for this file if we're the only ones that hold it after
8989 static void io_uring_attempt_task_drop(struct file *file)
8991 if (!current->io_uring)
8994 * fput() is pending, will be 2 if the only other ref is our potential
8995 * task file note. If the task is exiting, drop regardless of count.
8997 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8998 atomic_long_read(&file->f_count) == 2)
8999 io_uring_del_task_file(file);
9002 void __io_uring_files_cancel(struct files_struct *files)
9004 struct io_uring_task *tctx = current->io_uring;
9006 unsigned long index;
9008 /* make sure overflow events are dropped */
9009 atomic_inc(&tctx->in_idle);
9011 xa_for_each(&tctx->xa, index, file) {
9012 struct io_ring_ctx *ctx = file->private_data;
9014 io_uring_cancel_task_requests(ctx, files);
9016 io_uring_del_task_file(file);
9019 atomic_dec(&tctx->in_idle);
9022 static s64 tctx_inflight(struct io_uring_task *tctx)
9024 unsigned long index;
9028 inflight = percpu_counter_sum(&tctx->inflight);
9033 * If we have SQPOLL rings, then we need to iterate and find them, and
9034 * add the pending count for those.
9036 xa_for_each(&tctx->xa, index, file) {
9037 struct io_ring_ctx *ctx = file->private_data;
9039 if (ctx->flags & IORING_SETUP_SQPOLL) {
9040 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
9042 inflight += percpu_counter_sum(&__tctx->inflight);
9050 * Find any io_uring fd that this task has registered or done IO on, and cancel
9053 void __io_uring_task_cancel(void)
9055 struct io_uring_task *tctx = current->io_uring;
9059 /* make sure overflow events are dropped */
9060 atomic_inc(&tctx->in_idle);
9063 /* read completions before cancelations */
9064 inflight = tctx_inflight(tctx);
9067 __io_uring_files_cancel(NULL);
9069 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9072 * If we've seen completions, retry. This avoids a race where
9073 * a completion comes in before we did prepare_to_wait().
9075 if (inflight != tctx_inflight(tctx))
9080 finish_wait(&tctx->wait, &wait);
9081 atomic_dec(&tctx->in_idle);
9084 static int io_uring_flush(struct file *file, void *data)
9086 io_uring_attempt_task_drop(file);
9090 static void *io_uring_validate_mmap_request(struct file *file,
9091 loff_t pgoff, size_t sz)
9093 struct io_ring_ctx *ctx = file->private_data;
9094 loff_t offset = pgoff << PAGE_SHIFT;
9099 case IORING_OFF_SQ_RING:
9100 case IORING_OFF_CQ_RING:
9103 case IORING_OFF_SQES:
9107 return ERR_PTR(-EINVAL);
9110 page = virt_to_head_page(ptr);
9111 if (sz > page_size(page))
9112 return ERR_PTR(-EINVAL);
9119 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9121 size_t sz = vma->vm_end - vma->vm_start;
9125 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9127 return PTR_ERR(ptr);
9129 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9130 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9133 #else /* !CONFIG_MMU */
9135 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9137 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9140 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9142 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9145 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9146 unsigned long addr, unsigned long len,
9147 unsigned long pgoff, unsigned long flags)
9151 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9153 return PTR_ERR(ptr);
9155 return (unsigned long) ptr;
9158 #endif /* !CONFIG_MMU */
9160 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9165 if (!io_sqring_full(ctx))
9168 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9170 if (!io_sqring_full(ctx))
9174 } while (!signal_pending(current));
9176 finish_wait(&ctx->sqo_sq_wait, &wait);
9179 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9180 u32, min_complete, u32, flags, const sigset_t __user *, sig,
9183 struct io_ring_ctx *ctx;
9190 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9191 IORING_ENTER_SQ_WAIT))
9199 if (f.file->f_op != &io_uring_fops)
9203 ctx = f.file->private_data;
9204 if (!percpu_ref_tryget(&ctx->refs))
9208 if (ctx->flags & IORING_SETUP_R_DISABLED)
9212 * For SQ polling, the thread will do all submissions and completions.
9213 * Just return the requested submit count, and wake the thread if
9217 if (ctx->flags & IORING_SETUP_SQPOLL) {
9218 if (!list_empty_careful(&ctx->cq_overflow_list))
9219 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9220 if (flags & IORING_ENTER_SQ_WAKEUP)
9221 wake_up(&ctx->sq_data->wait);
9222 if (flags & IORING_ENTER_SQ_WAIT)
9223 io_sqpoll_wait_sq(ctx);
9224 submitted = to_submit;
9225 } else if (to_submit) {
9226 ret = io_uring_add_task_file(ctx, f.file);
9229 mutex_lock(&ctx->uring_lock);
9230 submitted = io_submit_sqes(ctx, to_submit);
9231 mutex_unlock(&ctx->uring_lock);
9233 if (submitted != to_submit)
9236 if (flags & IORING_ENTER_GETEVENTS) {
9237 min_complete = min(min_complete, ctx->cq_entries);
9240 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9241 * space applications don't need to do io completion events
9242 * polling again, they can rely on io_sq_thread to do polling
9243 * work, which can reduce cpu usage and uring_lock contention.
9245 if (ctx->flags & IORING_SETUP_IOPOLL &&
9246 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9247 ret = io_iopoll_check(ctx, min_complete);
9249 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
9254 percpu_ref_put(&ctx->refs);
9257 return submitted ? submitted : ret;
9260 #ifdef CONFIG_PROC_FS
9261 static int io_uring_show_cred(int id, void *p, void *data)
9263 struct io_identity *iod = p;
9264 const struct cred *cred = iod->creds;
9265 struct seq_file *m = data;
9266 struct user_namespace *uns = seq_user_ns(m);
9267 struct group_info *gi;
9272 seq_printf(m, "%5d\n", id);
9273 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9274 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9275 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9276 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9277 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9278 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9279 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9280 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9281 seq_puts(m, "\n\tGroups:\t");
9282 gi = cred->group_info;
9283 for (g = 0; g < gi->ngroups; g++) {
9284 seq_put_decimal_ull(m, g ? " " : "",
9285 from_kgid_munged(uns, gi->gid[g]));
9287 seq_puts(m, "\n\tCapEff:\t");
9288 cap = cred->cap_effective;
9289 CAP_FOR_EACH_U32(__capi)
9290 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9295 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9297 struct io_sq_data *sq = NULL;
9302 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9303 * since fdinfo case grabs it in the opposite direction of normal use
9304 * cases. If we fail to get the lock, we just don't iterate any
9305 * structures that could be going away outside the io_uring mutex.
9307 has_lock = mutex_trylock(&ctx->uring_lock);
9309 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9312 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9313 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9314 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9315 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9316 struct fixed_file_table *table;
9319 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9320 f = table->files[i & IORING_FILE_TABLE_MASK];
9322 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9324 seq_printf(m, "%5u: <none>\n", i);
9326 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9327 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9328 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9330 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9331 (unsigned int) buf->len);
9333 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9334 seq_printf(m, "Personalities:\n");
9335 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9337 seq_printf(m, "PollList:\n");
9338 spin_lock_irq(&ctx->completion_lock);
9339 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9340 struct hlist_head *list = &ctx->cancel_hash[i];
9341 struct io_kiocb *req;
9343 hlist_for_each_entry(req, list, hash_node)
9344 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9345 req->task->task_works != NULL);
9347 spin_unlock_irq(&ctx->completion_lock);
9349 mutex_unlock(&ctx->uring_lock);
9352 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9354 struct io_ring_ctx *ctx = f->private_data;
9356 if (percpu_ref_tryget(&ctx->refs)) {
9357 __io_uring_show_fdinfo(ctx, m);
9358 percpu_ref_put(&ctx->refs);
9363 static const struct file_operations io_uring_fops = {
9364 .release = io_uring_release,
9365 .flush = io_uring_flush,
9366 .mmap = io_uring_mmap,
9368 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9369 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9371 .poll = io_uring_poll,
9372 .fasync = io_uring_fasync,
9373 #ifdef CONFIG_PROC_FS
9374 .show_fdinfo = io_uring_show_fdinfo,
9378 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9379 struct io_uring_params *p)
9381 struct io_rings *rings;
9382 size_t size, sq_array_offset;
9384 /* make sure these are sane, as we already accounted them */
9385 ctx->sq_entries = p->sq_entries;
9386 ctx->cq_entries = p->cq_entries;
9388 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9389 if (size == SIZE_MAX)
9392 rings = io_mem_alloc(size);
9397 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9398 rings->sq_ring_mask = p->sq_entries - 1;
9399 rings->cq_ring_mask = p->cq_entries - 1;
9400 rings->sq_ring_entries = p->sq_entries;
9401 rings->cq_ring_entries = p->cq_entries;
9402 ctx->sq_mask = rings->sq_ring_mask;
9403 ctx->cq_mask = rings->cq_ring_mask;
9405 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9406 if (size == SIZE_MAX) {
9407 io_mem_free(ctx->rings);
9412 ctx->sq_sqes = io_mem_alloc(size);
9413 if (!ctx->sq_sqes) {
9414 io_mem_free(ctx->rings);
9423 * Allocate an anonymous fd, this is what constitutes the application
9424 * visible backing of an io_uring instance. The application mmaps this
9425 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9426 * we have to tie this fd to a socket for file garbage collection purposes.
9428 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9433 #if defined(CONFIG_UNIX)
9434 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9440 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9444 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9445 O_RDWR | O_CLOEXEC);
9449 ret = PTR_ERR(file);
9453 #if defined(CONFIG_UNIX)
9454 ctx->ring_sock->file = file;
9456 if (unlikely(io_uring_add_task_file(ctx, file))) {
9457 file = ERR_PTR(-ENOMEM);
9460 fd_install(ret, file);
9463 #if defined(CONFIG_UNIX)
9464 sock_release(ctx->ring_sock);
9465 ctx->ring_sock = NULL;
9470 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9471 struct io_uring_params __user *params)
9473 struct user_struct *user = NULL;
9474 struct io_ring_ctx *ctx;
9480 if (entries > IORING_MAX_ENTRIES) {
9481 if (!(p->flags & IORING_SETUP_CLAMP))
9483 entries = IORING_MAX_ENTRIES;
9487 * Use twice as many entries for the CQ ring. It's possible for the
9488 * application to drive a higher depth than the size of the SQ ring,
9489 * since the sqes are only used at submission time. This allows for
9490 * some flexibility in overcommitting a bit. If the application has
9491 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9492 * of CQ ring entries manually.
9494 p->sq_entries = roundup_pow_of_two(entries);
9495 if (p->flags & IORING_SETUP_CQSIZE) {
9497 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9498 * to a power-of-two, if it isn't already. We do NOT impose
9499 * any cq vs sq ring sizing.
9501 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9502 if (p->cq_entries < p->sq_entries)
9504 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9505 if (!(p->flags & IORING_SETUP_CLAMP))
9507 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9510 p->cq_entries = 2 * p->sq_entries;
9513 user = get_uid(current_user());
9514 limit_mem = !capable(CAP_IPC_LOCK);
9517 ret = __io_account_mem(user,
9518 ring_pages(p->sq_entries, p->cq_entries));
9525 ctx = io_ring_ctx_alloc(p);
9528 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9533 ctx->compat = in_compat_syscall();
9535 ctx->creds = get_current_cred();
9537 ctx->loginuid = current->loginuid;
9538 ctx->sessionid = current->sessionid;
9540 ctx->sqo_task = get_task_struct(current);
9543 * This is just grabbed for accounting purposes. When a process exits,
9544 * the mm is exited and dropped before the files, hence we need to hang
9545 * on to this mm purely for the purposes of being able to unaccount
9546 * memory (locked/pinned vm). It's not used for anything else.
9548 mmgrab(current->mm);
9549 ctx->mm_account = current->mm;
9551 #ifdef CONFIG_BLK_CGROUP
9553 * The sq thread will belong to the original cgroup it was inited in.
9554 * If the cgroup goes offline (e.g. disabling the io controller), then
9555 * issued bios will be associated with the closest cgroup later in the
9559 ctx->sqo_blkcg_css = blkcg_css();
9560 ret = css_tryget_online(ctx->sqo_blkcg_css);
9563 /* don't init against a dying cgroup, have the user try again */
9564 ctx->sqo_blkcg_css = NULL;
9571 * Account memory _before_ installing the file descriptor. Once
9572 * the descriptor is installed, it can get closed at any time. Also
9573 * do this before hitting the general error path, as ring freeing
9574 * will un-account as well.
9576 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9578 ctx->limit_mem = limit_mem;
9580 ret = io_allocate_scq_urings(ctx, p);
9584 ret = io_sq_offload_create(ctx, p);
9588 if (!(p->flags & IORING_SETUP_R_DISABLED))
9589 io_sq_offload_start(ctx);
9591 memset(&p->sq_off, 0, sizeof(p->sq_off));
9592 p->sq_off.head = offsetof(struct io_rings, sq.head);
9593 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9594 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9595 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9596 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9597 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9598 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9600 memset(&p->cq_off, 0, sizeof(p->cq_off));
9601 p->cq_off.head = offsetof(struct io_rings, cq.head);
9602 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9603 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9604 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9605 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9606 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9607 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9609 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9610 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9611 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9612 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED;
9614 if (copy_to_user(params, p, sizeof(*p))) {
9620 * Install ring fd as the very last thing, so we don't risk someone
9621 * having closed it before we finish setup
9623 ret = io_uring_get_fd(ctx);
9627 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9630 io_ring_ctx_wait_and_kill(ctx);
9635 * Sets up an aio uring context, and returns the fd. Applications asks for a
9636 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9637 * params structure passed in.
9639 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9641 struct io_uring_params p;
9644 if (copy_from_user(&p, params, sizeof(p)))
9646 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9651 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9652 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9653 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9654 IORING_SETUP_R_DISABLED))
9657 return io_uring_create(entries, &p, params);
9660 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9661 struct io_uring_params __user *, params)
9663 return io_uring_setup(entries, params);
9666 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9668 struct io_uring_probe *p;
9672 size = struct_size(p, ops, nr_args);
9673 if (size == SIZE_MAX)
9675 p = kzalloc(size, GFP_KERNEL);
9680 if (copy_from_user(p, arg, size))
9683 if (memchr_inv(p, 0, size))
9686 p->last_op = IORING_OP_LAST - 1;
9687 if (nr_args > IORING_OP_LAST)
9688 nr_args = IORING_OP_LAST;
9690 for (i = 0; i < nr_args; i++) {
9692 if (!io_op_defs[i].not_supported)
9693 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9698 if (copy_to_user(arg, p, size))
9705 static int io_register_personality(struct io_ring_ctx *ctx)
9707 struct io_identity *id;
9710 id = kmalloc(sizeof(*id), GFP_KERNEL);
9714 io_init_identity(id);
9715 id->creds = get_current_cred();
9717 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9719 put_cred(id->creds);
9725 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9727 struct io_identity *iod;
9729 iod = idr_remove(&ctx->personality_idr, id);
9731 put_cred(iod->creds);
9732 if (refcount_dec_and_test(&iod->count))
9740 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9741 unsigned int nr_args)
9743 struct io_uring_restriction *res;
9747 /* Restrictions allowed only if rings started disabled */
9748 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9751 /* We allow only a single restrictions registration */
9752 if (ctx->restrictions.registered)
9755 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9758 size = array_size(nr_args, sizeof(*res));
9759 if (size == SIZE_MAX)
9762 res = memdup_user(arg, size);
9764 return PTR_ERR(res);
9768 for (i = 0; i < nr_args; i++) {
9769 switch (res[i].opcode) {
9770 case IORING_RESTRICTION_REGISTER_OP:
9771 if (res[i].register_op >= IORING_REGISTER_LAST) {
9776 __set_bit(res[i].register_op,
9777 ctx->restrictions.register_op);
9779 case IORING_RESTRICTION_SQE_OP:
9780 if (res[i].sqe_op >= IORING_OP_LAST) {
9785 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9787 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9788 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9790 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9791 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9800 /* Reset all restrictions if an error happened */
9802 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9804 ctx->restrictions.registered = true;
9810 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9812 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9815 if (ctx->restrictions.registered)
9816 ctx->restricted = 1;
9818 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9820 io_sq_offload_start(ctx);
9825 static bool io_register_op_must_quiesce(int op)
9828 case IORING_UNREGISTER_FILES:
9829 case IORING_REGISTER_FILES_UPDATE:
9830 case IORING_REGISTER_PROBE:
9831 case IORING_REGISTER_PERSONALITY:
9832 case IORING_UNREGISTER_PERSONALITY:
9839 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9840 void __user *arg, unsigned nr_args)
9841 __releases(ctx->uring_lock)
9842 __acquires(ctx->uring_lock)
9847 * We're inside the ring mutex, if the ref is already dying, then
9848 * someone else killed the ctx or is already going through
9849 * io_uring_register().
9851 if (percpu_ref_is_dying(&ctx->refs))
9854 if (io_register_op_must_quiesce(opcode)) {
9855 percpu_ref_kill(&ctx->refs);
9858 * Drop uring mutex before waiting for references to exit. If
9859 * another thread is currently inside io_uring_enter() it might
9860 * need to grab the uring_lock to make progress. If we hold it
9861 * here across the drain wait, then we can deadlock. It's safe
9862 * to drop the mutex here, since no new references will come in
9863 * after we've killed the percpu ref.
9865 mutex_unlock(&ctx->uring_lock);
9867 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9870 ret = io_run_task_work_sig();
9875 mutex_lock(&ctx->uring_lock);
9878 percpu_ref_resurrect(&ctx->refs);
9883 if (ctx->restricted) {
9884 if (opcode >= IORING_REGISTER_LAST) {
9889 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9896 case IORING_REGISTER_BUFFERS:
9897 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9899 case IORING_UNREGISTER_BUFFERS:
9903 ret = io_sqe_buffer_unregister(ctx);
9905 case IORING_REGISTER_FILES:
9906 ret = io_sqe_files_register(ctx, arg, nr_args);
9908 case IORING_UNREGISTER_FILES:
9912 ret = io_sqe_files_unregister(ctx);
9914 case IORING_REGISTER_FILES_UPDATE:
9915 ret = io_sqe_files_update(ctx, arg, nr_args);
9917 case IORING_REGISTER_EVENTFD:
9918 case IORING_REGISTER_EVENTFD_ASYNC:
9922 ret = io_eventfd_register(ctx, arg);
9925 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9926 ctx->eventfd_async = 1;
9928 ctx->eventfd_async = 0;
9930 case IORING_UNREGISTER_EVENTFD:
9934 ret = io_eventfd_unregister(ctx);
9936 case IORING_REGISTER_PROBE:
9938 if (!arg || nr_args > 256)
9940 ret = io_probe(ctx, arg, nr_args);
9942 case IORING_REGISTER_PERSONALITY:
9946 ret = io_register_personality(ctx);
9948 case IORING_UNREGISTER_PERSONALITY:
9952 ret = io_unregister_personality(ctx, nr_args);
9954 case IORING_REGISTER_ENABLE_RINGS:
9958 ret = io_register_enable_rings(ctx);
9960 case IORING_REGISTER_RESTRICTIONS:
9961 ret = io_register_restrictions(ctx, arg, nr_args);
9969 if (io_register_op_must_quiesce(opcode)) {
9970 /* bring the ctx back to life */
9971 percpu_ref_reinit(&ctx->refs);
9973 reinit_completion(&ctx->ref_comp);
9978 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9979 void __user *, arg, unsigned int, nr_args)
9981 struct io_ring_ctx *ctx;
9990 if (f.file->f_op != &io_uring_fops)
9993 ctx = f.file->private_data;
9995 mutex_lock(&ctx->uring_lock);
9996 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9997 mutex_unlock(&ctx->uring_lock);
9998 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9999 ctx->cq_ev_fd != NULL, ret);
10005 static int __init io_uring_init(void)
10007 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10008 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10009 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10012 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10013 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10014 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10015 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10016 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10017 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10018 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10019 BUILD_BUG_SQE_ELEM(8, __u64, off);
10020 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10021 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10022 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10023 BUILD_BUG_SQE_ELEM(24, __u32, len);
10024 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10025 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10026 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10027 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10028 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10029 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10030 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10031 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10032 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10033 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10034 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10035 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10036 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10037 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10038 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10039 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10040 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10041 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10042 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10044 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10045 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10046 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
10049 __initcall(io_uring_init);