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 {
399 struct wait_queue_head *head;
405 struct wait_queue_entry wait;
410 struct file *put_file;
414 struct io_timeout_data {
415 struct io_kiocb *req;
416 struct hrtimer timer;
417 struct timespec64 ts;
418 enum hrtimer_mode mode;
423 struct sockaddr __user *addr;
424 int __user *addr_len;
426 unsigned long nofile;
446 struct list_head list;
449 struct io_timeout_rem {
455 /* NOTE: kiocb has the file as the first member, so don't do it here */
463 struct sockaddr __user *addr;
470 struct user_msghdr __user *umsg;
476 struct io_buffer *kbuf;
482 bool ignore_nonblock;
483 struct filename *filename;
485 unsigned long nofile;
488 struct io_files_update {
514 struct epoll_event event;
518 struct file *file_out;
519 struct file *file_in;
526 struct io_provide_buf {
540 const char __user *filename;
541 struct statx __user *buffer;
544 struct io_completion {
546 struct list_head list;
550 struct io_async_connect {
551 struct sockaddr_storage address;
554 struct io_async_msghdr {
555 struct iovec fast_iov[UIO_FASTIOV];
557 struct sockaddr __user *uaddr;
559 struct sockaddr_storage addr;
563 struct iovec fast_iov[UIO_FASTIOV];
564 const struct iovec *free_iovec;
565 struct iov_iter iter;
567 struct wait_page_queue wpq;
571 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
572 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
573 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
574 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
575 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
576 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
583 REQ_F_LINK_TIMEOUT_BIT,
585 REQ_F_NEED_CLEANUP_BIT,
587 REQ_F_BUFFER_SELECTED_BIT,
588 REQ_F_NO_FILE_TABLE_BIT,
589 REQ_F_WORK_INITIALIZED_BIT,
590 REQ_F_LTIMEOUT_ACTIVE_BIT,
592 /* not a real bit, just to check we're not overflowing the space */
598 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
599 /* drain existing IO first */
600 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
602 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
603 /* doesn't sever on completion < 0 */
604 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
606 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
607 /* IOSQE_BUFFER_SELECT */
608 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
611 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
612 /* fail rest of links */
613 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
614 /* on inflight list */
615 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
616 /* read/write uses file position */
617 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
618 /* must not punt to workers */
619 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
620 /* has or had linked timeout */
621 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
623 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
625 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
626 /* already went through poll handler */
627 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
628 /* buffer already selected */
629 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
630 /* doesn't need file table for this request */
631 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
632 /* io_wq_work is initialized */
633 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
634 /* linked timeout is active, i.e. prepared by link's head */
635 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
639 struct io_poll_iocb poll;
640 struct io_poll_iocb *double_poll;
644 * NOTE! Each of the iocb union members has the file pointer
645 * as the first entry in their struct definition. So you can
646 * access the file pointer through any of the sub-structs,
647 * or directly as just 'ki_filp' in this struct.
653 struct io_poll_iocb poll;
654 struct io_accept accept;
656 struct io_cancel cancel;
657 struct io_timeout timeout;
658 struct io_timeout_rem timeout_rem;
659 struct io_connect connect;
660 struct io_sr_msg sr_msg;
662 struct io_close close;
663 struct io_files_update files_update;
664 struct io_fadvise fadvise;
665 struct io_madvise madvise;
666 struct io_epoll epoll;
667 struct io_splice splice;
668 struct io_provide_buf pbuf;
669 struct io_statx statx;
670 /* use only after cleaning per-op data, see io_clean_op() */
671 struct io_completion compl;
674 /* opcode allocated if it needs to store data for async defer */
677 /* polled IO has completed */
683 struct io_ring_ctx *ctx;
686 struct task_struct *task;
689 struct list_head link_list;
692 * 1. used with ctx->iopoll_list with reads/writes
693 * 2. to track reqs with ->files (see io_op_def::file_table)
695 struct list_head inflight_entry;
697 struct percpu_ref *fixed_file_refs;
698 struct callback_head task_work;
699 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
700 struct hlist_node hash_node;
701 struct async_poll *apoll;
702 struct io_wq_work work;
705 struct io_defer_entry {
706 struct list_head list;
707 struct io_kiocb *req;
711 #define IO_IOPOLL_BATCH 8
713 struct io_comp_state {
715 struct list_head list;
716 struct io_ring_ctx *ctx;
719 struct io_submit_state {
720 struct blk_plug plug;
723 * io_kiocb alloc cache
725 void *reqs[IO_IOPOLL_BATCH];
726 unsigned int free_reqs;
729 * Batch completion logic
731 struct io_comp_state comp;
734 * File reference cache
738 unsigned int has_refs;
739 unsigned int ios_left;
743 /* needs req->file assigned */
744 unsigned needs_file : 1;
745 /* don't fail if file grab fails */
746 unsigned needs_file_no_error : 1;
747 /* hash wq insertion if file is a regular file */
748 unsigned hash_reg_file : 1;
749 /* unbound wq insertion if file is a non-regular file */
750 unsigned unbound_nonreg_file : 1;
751 /* opcode is not supported by this kernel */
752 unsigned not_supported : 1;
753 /* set if opcode supports polled "wait" */
755 unsigned pollout : 1;
756 /* op supports buffer selection */
757 unsigned buffer_select : 1;
758 /* must always have async data allocated */
759 unsigned needs_async_data : 1;
760 /* size of async data needed, if any */
761 unsigned short async_size;
765 static const struct io_op_def io_op_defs[] = {
766 [IORING_OP_NOP] = {},
767 [IORING_OP_READV] = {
769 .unbound_nonreg_file = 1,
772 .needs_async_data = 1,
773 .async_size = sizeof(struct io_async_rw),
774 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
776 [IORING_OP_WRITEV] = {
779 .unbound_nonreg_file = 1,
781 .needs_async_data = 1,
782 .async_size = sizeof(struct io_async_rw),
783 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
786 [IORING_OP_FSYNC] = {
788 .work_flags = IO_WQ_WORK_BLKCG,
790 [IORING_OP_READ_FIXED] = {
792 .unbound_nonreg_file = 1,
794 .async_size = sizeof(struct io_async_rw),
795 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
797 [IORING_OP_WRITE_FIXED] = {
800 .unbound_nonreg_file = 1,
802 .async_size = sizeof(struct io_async_rw),
803 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
806 [IORING_OP_POLL_ADD] = {
808 .unbound_nonreg_file = 1,
810 [IORING_OP_POLL_REMOVE] = {},
811 [IORING_OP_SYNC_FILE_RANGE] = {
813 .work_flags = IO_WQ_WORK_BLKCG,
815 [IORING_OP_SENDMSG] = {
817 .unbound_nonreg_file = 1,
819 .needs_async_data = 1,
820 .async_size = sizeof(struct io_async_msghdr),
821 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
824 [IORING_OP_RECVMSG] = {
826 .unbound_nonreg_file = 1,
829 .needs_async_data = 1,
830 .async_size = sizeof(struct io_async_msghdr),
831 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
834 [IORING_OP_TIMEOUT] = {
835 .needs_async_data = 1,
836 .async_size = sizeof(struct io_timeout_data),
837 .work_flags = IO_WQ_WORK_MM,
839 [IORING_OP_TIMEOUT_REMOVE] = {},
840 [IORING_OP_ACCEPT] = {
842 .unbound_nonreg_file = 1,
844 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
846 [IORING_OP_ASYNC_CANCEL] = {},
847 [IORING_OP_LINK_TIMEOUT] = {
848 .needs_async_data = 1,
849 .async_size = sizeof(struct io_timeout_data),
850 .work_flags = IO_WQ_WORK_MM,
852 [IORING_OP_CONNECT] = {
854 .unbound_nonreg_file = 1,
856 .needs_async_data = 1,
857 .async_size = sizeof(struct io_async_connect),
858 .work_flags = IO_WQ_WORK_MM,
860 [IORING_OP_FALLOCATE] = {
862 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
864 [IORING_OP_OPENAT] = {
865 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
868 [IORING_OP_CLOSE] = {
870 .needs_file_no_error = 1,
871 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
873 [IORING_OP_FILES_UPDATE] = {
874 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
876 [IORING_OP_STATX] = {
877 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
878 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
882 .unbound_nonreg_file = 1,
885 .async_size = sizeof(struct io_async_rw),
886 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
888 [IORING_OP_WRITE] = {
890 .unbound_nonreg_file = 1,
892 .async_size = sizeof(struct io_async_rw),
893 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
896 [IORING_OP_FADVISE] = {
898 .work_flags = IO_WQ_WORK_BLKCG,
900 [IORING_OP_MADVISE] = {
901 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
905 .unbound_nonreg_file = 1,
907 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
911 .unbound_nonreg_file = 1,
914 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
916 [IORING_OP_OPENAT2] = {
917 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
920 [IORING_OP_EPOLL_CTL] = {
921 .unbound_nonreg_file = 1,
922 .work_flags = IO_WQ_WORK_FILES,
924 [IORING_OP_SPLICE] = {
927 .unbound_nonreg_file = 1,
928 .work_flags = IO_WQ_WORK_BLKCG,
930 [IORING_OP_PROVIDE_BUFFERS] = {},
931 [IORING_OP_REMOVE_BUFFERS] = {},
935 .unbound_nonreg_file = 1,
939 enum io_mem_account {
944 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
945 struct io_comp_state *cs);
946 static void io_cqring_fill_event(struct io_kiocb *req, long res);
947 static void io_put_req(struct io_kiocb *req);
948 static void io_put_req_deferred(struct io_kiocb *req, int nr);
949 static void io_double_put_req(struct io_kiocb *req);
950 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
951 static void __io_queue_linked_timeout(struct io_kiocb *req);
952 static void io_queue_linked_timeout(struct io_kiocb *req);
953 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
954 struct io_uring_files_update *ip,
956 static void __io_clean_op(struct io_kiocb *req);
957 static struct file *io_file_get(struct io_submit_state *state,
958 struct io_kiocb *req, int fd, bool fixed);
959 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
960 static void io_file_put_work(struct work_struct *work);
962 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
963 struct iovec **iovec, struct iov_iter *iter,
965 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
966 const struct iovec *fast_iov,
967 struct iov_iter *iter, bool force);
969 static struct kmem_cache *req_cachep;
971 static const struct file_operations io_uring_fops;
973 struct sock *io_uring_get_socket(struct file *file)
975 #if defined(CONFIG_UNIX)
976 if (file->f_op == &io_uring_fops) {
977 struct io_ring_ctx *ctx = file->private_data;
979 return ctx->ring_sock->sk;
984 EXPORT_SYMBOL(io_uring_get_socket);
986 static inline void io_clean_op(struct io_kiocb *req)
988 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
993 static void io_sq_thread_drop_mm(void)
995 struct mm_struct *mm = current->mm;
998 kthread_unuse_mm(mm);
1004 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1006 struct mm_struct *mm;
1011 /* Should never happen */
1012 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL)))
1015 task_lock(ctx->sqo_task);
1016 mm = ctx->sqo_task->mm;
1017 if (unlikely(!mm || !mmget_not_zero(mm)))
1019 task_unlock(ctx->sqo_task);
1029 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1030 struct io_kiocb *req)
1032 if (!(io_op_defs[req->opcode].work_flags & IO_WQ_WORK_MM))
1034 return __io_sq_thread_acquire_mm(ctx);
1037 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1038 struct cgroup_subsys_state **cur_css)
1041 #ifdef CONFIG_BLK_CGROUP
1042 /* puts the old one when swapping */
1043 if (*cur_css != ctx->sqo_blkcg_css) {
1044 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1045 *cur_css = ctx->sqo_blkcg_css;
1050 static void io_sq_thread_unassociate_blkcg(void)
1052 #ifdef CONFIG_BLK_CGROUP
1053 kthread_associate_blkcg(NULL);
1057 static inline void req_set_fail_links(struct io_kiocb *req)
1059 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1060 req->flags |= REQ_F_FAIL_LINK;
1064 * None of these are dereferenced, they are simply used to check if any of
1065 * them have changed. If we're under current and check they are still the
1066 * same, we're fine to grab references to them for actual out-of-line use.
1068 static void io_init_identity(struct io_identity *id)
1070 id->files = current->files;
1071 id->mm = current->mm;
1072 #ifdef CONFIG_BLK_CGROUP
1074 id->blkcg_css = blkcg_css();
1077 id->creds = current_cred();
1078 id->nsproxy = current->nsproxy;
1079 id->fs = current->fs;
1080 id->fsize = rlimit(RLIMIT_FSIZE);
1082 id->loginuid = current->loginuid;
1083 id->sessionid = current->sessionid;
1085 refcount_set(&id->count, 1);
1088 static inline void __io_req_init_async(struct io_kiocb *req)
1090 memset(&req->work, 0, sizeof(req->work));
1091 req->flags |= REQ_F_WORK_INITIALIZED;
1095 * Note: must call io_req_init_async() for the first time you
1096 * touch any members of io_wq_work.
1098 static inline void io_req_init_async(struct io_kiocb *req)
1100 struct io_uring_task *tctx = current->io_uring;
1102 if (req->flags & REQ_F_WORK_INITIALIZED)
1105 __io_req_init_async(req);
1107 /* Grab a ref if this isn't our static identity */
1108 req->work.identity = tctx->identity;
1109 if (tctx->identity != &tctx->__identity)
1110 refcount_inc(&req->work.identity->count);
1113 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1115 return ctx->flags & IORING_SETUP_SQPOLL;
1118 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1120 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1122 complete(&ctx->ref_comp);
1125 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1127 return !req->timeout.off;
1130 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1132 struct io_ring_ctx *ctx;
1135 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1139 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1140 if (!ctx->fallback_req)
1144 * Use 5 bits less than the max cq entries, that should give us around
1145 * 32 entries per hash list if totally full and uniformly spread.
1147 hash_bits = ilog2(p->cq_entries);
1151 ctx->cancel_hash_bits = hash_bits;
1152 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1154 if (!ctx->cancel_hash)
1156 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1158 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1159 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1162 ctx->flags = p->flags;
1163 init_waitqueue_head(&ctx->sqo_sq_wait);
1164 INIT_LIST_HEAD(&ctx->sqd_list);
1165 init_waitqueue_head(&ctx->cq_wait);
1166 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1167 init_completion(&ctx->ref_comp);
1168 init_completion(&ctx->sq_thread_comp);
1169 idr_init(&ctx->io_buffer_idr);
1170 idr_init(&ctx->personality_idr);
1171 mutex_init(&ctx->uring_lock);
1172 init_waitqueue_head(&ctx->wait);
1173 spin_lock_init(&ctx->completion_lock);
1174 INIT_LIST_HEAD(&ctx->iopoll_list);
1175 INIT_LIST_HEAD(&ctx->defer_list);
1176 INIT_LIST_HEAD(&ctx->timeout_list);
1177 init_waitqueue_head(&ctx->inflight_wait);
1178 spin_lock_init(&ctx->inflight_lock);
1179 INIT_LIST_HEAD(&ctx->inflight_list);
1180 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1181 init_llist_head(&ctx->file_put_llist);
1184 if (ctx->fallback_req)
1185 kmem_cache_free(req_cachep, ctx->fallback_req);
1186 kfree(ctx->cancel_hash);
1191 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1193 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1194 struct io_ring_ctx *ctx = req->ctx;
1196 return seq != ctx->cached_cq_tail
1197 + READ_ONCE(ctx->cached_cq_overflow);
1203 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1205 struct io_rings *rings = ctx->rings;
1207 /* order cqe stores with ring update */
1208 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1210 if (wq_has_sleeper(&ctx->cq_wait)) {
1211 wake_up_interruptible(&ctx->cq_wait);
1212 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1216 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1218 if (req->work.identity == &tctx->__identity)
1220 if (refcount_dec_and_test(&req->work.identity->count))
1221 kfree(req->work.identity);
1224 static void io_req_clean_work(struct io_kiocb *req)
1226 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1229 req->flags &= ~REQ_F_WORK_INITIALIZED;
1231 if (req->work.flags & IO_WQ_WORK_MM) {
1232 mmdrop(req->work.identity->mm);
1233 req->work.flags &= ~IO_WQ_WORK_MM;
1235 #ifdef CONFIG_BLK_CGROUP
1236 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1237 css_put(req->work.identity->blkcg_css);
1238 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1241 if (req->work.flags & IO_WQ_WORK_CREDS) {
1242 put_cred(req->work.identity->creds);
1243 req->work.flags &= ~IO_WQ_WORK_CREDS;
1245 if (req->work.flags & IO_WQ_WORK_FS) {
1246 struct fs_struct *fs = req->work.identity->fs;
1248 spin_lock(&req->work.identity->fs->lock);
1251 spin_unlock(&req->work.identity->fs->lock);
1254 req->work.flags &= ~IO_WQ_WORK_FS;
1257 io_put_identity(req->task->io_uring, req);
1261 * Create a private copy of io_identity, since some fields don't match
1262 * the current context.
1264 static bool io_identity_cow(struct io_kiocb *req)
1266 struct io_uring_task *tctx = current->io_uring;
1267 const struct cred *creds = NULL;
1268 struct io_identity *id;
1270 if (req->work.flags & IO_WQ_WORK_CREDS)
1271 creds = req->work.identity->creds;
1273 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1274 if (unlikely(!id)) {
1275 req->work.flags |= IO_WQ_WORK_CANCEL;
1280 * We can safely just re-init the creds we copied Either the field
1281 * matches the current one, or we haven't grabbed it yet. The only
1282 * exception is ->creds, through registered personalities, so handle
1283 * that one separately.
1285 io_init_identity(id);
1287 req->work.identity->creds = creds;
1289 /* add one for this request */
1290 refcount_inc(&id->count);
1292 /* drop tctx and req identity references, if needed */
1293 if (tctx->identity != &tctx->__identity &&
1294 refcount_dec_and_test(&tctx->identity->count))
1295 kfree(tctx->identity);
1296 if (req->work.identity != &tctx->__identity &&
1297 refcount_dec_and_test(&req->work.identity->count))
1298 kfree(req->work.identity);
1300 req->work.identity = id;
1301 tctx->identity = id;
1305 static bool io_grab_identity(struct io_kiocb *req)
1307 const struct io_op_def *def = &io_op_defs[req->opcode];
1308 struct io_identity *id = req->work.identity;
1309 struct io_ring_ctx *ctx = req->ctx;
1311 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1312 if (id->fsize != rlimit(RLIMIT_FSIZE))
1314 req->work.flags |= IO_WQ_WORK_FSIZE;
1317 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1318 (def->work_flags & IO_WQ_WORK_FILES) &&
1319 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1320 if (id->files != current->files ||
1321 id->nsproxy != current->nsproxy)
1323 atomic_inc(&id->files->count);
1324 get_nsproxy(id->nsproxy);
1325 req->flags |= REQ_F_INFLIGHT;
1327 spin_lock_irq(&ctx->inflight_lock);
1328 list_add(&req->inflight_entry, &ctx->inflight_list);
1329 spin_unlock_irq(&ctx->inflight_lock);
1330 req->work.flags |= IO_WQ_WORK_FILES;
1332 #ifdef CONFIG_BLK_CGROUP
1333 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1334 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1336 if (id->blkcg_css != blkcg_css()) {
1341 * This should be rare, either the cgroup is dying or the task
1342 * is moving cgroups. Just punt to root for the handful of ios.
1344 if (css_tryget_online(id->blkcg_css))
1345 req->work.flags |= IO_WQ_WORK_BLKCG;
1349 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1350 if (id->creds != current_cred())
1352 get_cred(id->creds);
1353 req->work.flags |= IO_WQ_WORK_CREDS;
1356 if (!uid_eq(current->loginuid, id->loginuid) ||
1357 current->sessionid != id->sessionid)
1360 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1361 (def->work_flags & IO_WQ_WORK_FS)) {
1362 if (current->fs != id->fs)
1364 spin_lock(&id->fs->lock);
1365 if (!id->fs->in_exec) {
1367 req->work.flags |= IO_WQ_WORK_FS;
1369 req->work.flags |= IO_WQ_WORK_CANCEL;
1371 spin_unlock(¤t->fs->lock);
1377 static void io_prep_async_work(struct io_kiocb *req)
1379 const struct io_op_def *def = &io_op_defs[req->opcode];
1380 struct io_ring_ctx *ctx = req->ctx;
1381 struct io_identity *id;
1383 io_req_init_async(req);
1384 id = req->work.identity;
1386 if (req->flags & REQ_F_FORCE_ASYNC)
1387 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1389 if (req->flags & REQ_F_ISREG) {
1390 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1391 io_wq_hash_work(&req->work, file_inode(req->file));
1393 if (def->unbound_nonreg_file)
1394 req->work.flags |= IO_WQ_WORK_UNBOUND;
1397 /* ->mm can never change on us */
1398 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1399 (def->work_flags & IO_WQ_WORK_MM)) {
1401 req->work.flags |= IO_WQ_WORK_MM;
1404 /* if we fail grabbing identity, we must COW, regrab, and retry */
1405 if (io_grab_identity(req))
1408 if (!io_identity_cow(req))
1411 /* can't fail at this point */
1412 if (!io_grab_identity(req))
1416 static void io_prep_async_link(struct io_kiocb *req)
1418 struct io_kiocb *cur;
1420 io_prep_async_work(req);
1421 if (req->flags & REQ_F_LINK_HEAD)
1422 list_for_each_entry(cur, &req->link_list, link_list)
1423 io_prep_async_work(cur);
1426 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1428 struct io_ring_ctx *ctx = req->ctx;
1429 struct io_kiocb *link = io_prep_linked_timeout(req);
1431 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1432 &req->work, req->flags);
1433 io_wq_enqueue(ctx->io_wq, &req->work);
1437 static void io_queue_async_work(struct io_kiocb *req)
1439 struct io_kiocb *link;
1441 /* init ->work of the whole link before punting */
1442 io_prep_async_link(req);
1443 link = __io_queue_async_work(req);
1446 io_queue_linked_timeout(link);
1449 static void io_kill_timeout(struct io_kiocb *req)
1451 struct io_timeout_data *io = req->async_data;
1454 ret = hrtimer_try_to_cancel(&io->timer);
1456 atomic_set(&req->ctx->cq_timeouts,
1457 atomic_read(&req->ctx->cq_timeouts) + 1);
1458 list_del_init(&req->timeout.list);
1459 io_cqring_fill_event(req, 0);
1460 io_put_req_deferred(req, 1);
1464 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1466 struct io_ring_ctx *ctx = req->ctx;
1468 if (!tsk || req->task == tsk)
1470 if (ctx->flags & IORING_SETUP_SQPOLL) {
1471 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1478 * Returns true if we found and killed one or more timeouts
1480 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1482 struct io_kiocb *req, *tmp;
1485 spin_lock_irq(&ctx->completion_lock);
1486 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1487 if (io_task_match(req, tsk)) {
1488 io_kill_timeout(req);
1492 spin_unlock_irq(&ctx->completion_lock);
1493 return canceled != 0;
1496 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1499 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1500 struct io_defer_entry, list);
1501 struct io_kiocb *link;
1503 if (req_need_defer(de->req, de->seq))
1505 list_del_init(&de->list);
1506 /* punt-init is done before queueing for defer */
1507 link = __io_queue_async_work(de->req);
1509 __io_queue_linked_timeout(link);
1510 /* drop submission reference */
1511 io_put_req_deferred(link, 1);
1514 } while (!list_empty(&ctx->defer_list));
1517 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1519 while (!list_empty(&ctx->timeout_list)) {
1520 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1521 struct io_kiocb, timeout.list);
1523 if (io_is_timeout_noseq(req))
1525 if (req->timeout.target_seq != ctx->cached_cq_tail
1526 - atomic_read(&ctx->cq_timeouts))
1529 list_del_init(&req->timeout.list);
1530 io_kill_timeout(req);
1534 static void io_commit_cqring(struct io_ring_ctx *ctx)
1536 io_flush_timeouts(ctx);
1537 __io_commit_cqring(ctx);
1539 if (unlikely(!list_empty(&ctx->defer_list)))
1540 __io_queue_deferred(ctx);
1543 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1545 struct io_rings *r = ctx->rings;
1547 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1550 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1552 struct io_rings *rings = ctx->rings;
1555 tail = ctx->cached_cq_tail;
1557 * writes to the cq entry need to come after reading head; the
1558 * control dependency is enough as we're using WRITE_ONCE to
1561 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1564 ctx->cached_cq_tail++;
1565 return &rings->cqes[tail & ctx->cq_mask];
1568 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1572 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1574 if (!ctx->eventfd_async)
1576 return io_wq_current_is_worker();
1579 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1581 if (waitqueue_active(&ctx->wait))
1582 wake_up(&ctx->wait);
1583 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1584 wake_up(&ctx->sq_data->wait);
1585 if (io_should_trigger_evfd(ctx))
1586 eventfd_signal(ctx->cq_ev_fd, 1);
1589 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1591 if (list_empty(&ctx->cq_overflow_list)) {
1592 clear_bit(0, &ctx->sq_check_overflow);
1593 clear_bit(0, &ctx->cq_check_overflow);
1594 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1598 static inline bool __io_match_files(struct io_kiocb *req,
1599 struct files_struct *files)
1601 return ((req->flags & REQ_F_WORK_INITIALIZED) &&
1602 (req->work.flags & IO_WQ_WORK_FILES)) &&
1603 req->work.identity->files == files;
1606 static bool io_match_files(struct io_kiocb *req,
1607 struct files_struct *files)
1609 struct io_kiocb *link;
1613 if (__io_match_files(req, files))
1615 if (req->flags & REQ_F_LINK_HEAD) {
1616 list_for_each_entry(link, &req->link_list, link_list) {
1617 if (__io_match_files(link, files))
1624 /* Returns true if there are no backlogged entries after the flush */
1625 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1626 struct task_struct *tsk,
1627 struct files_struct *files)
1629 struct io_rings *rings = ctx->rings;
1630 struct io_kiocb *req, *tmp;
1631 struct io_uring_cqe *cqe;
1632 unsigned long flags;
1636 if (list_empty_careful(&ctx->cq_overflow_list))
1638 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1639 rings->cq_ring_entries))
1643 spin_lock_irqsave(&ctx->completion_lock, flags);
1645 /* if force is set, the ring is going away. always drop after that */
1647 ctx->cq_overflow_flushed = 1;
1650 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1651 if (tsk && req->task != tsk)
1653 if (!io_match_files(req, files))
1656 cqe = io_get_cqring(ctx);
1660 list_move(&req->compl.list, &list);
1662 WRITE_ONCE(cqe->user_data, req->user_data);
1663 WRITE_ONCE(cqe->res, req->result);
1664 WRITE_ONCE(cqe->flags, req->compl.cflags);
1666 ctx->cached_cq_overflow++;
1667 WRITE_ONCE(ctx->rings->cq_overflow,
1668 ctx->cached_cq_overflow);
1672 io_commit_cqring(ctx);
1673 io_cqring_mark_overflow(ctx);
1675 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1676 io_cqring_ev_posted(ctx);
1678 while (!list_empty(&list)) {
1679 req = list_first_entry(&list, struct io_kiocb, compl.list);
1680 list_del(&req->compl.list);
1687 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1689 struct io_ring_ctx *ctx = req->ctx;
1690 struct io_uring_cqe *cqe;
1692 trace_io_uring_complete(ctx, req->user_data, res);
1695 * If we can't get a cq entry, userspace overflowed the
1696 * submission (by quite a lot). Increment the overflow count in
1699 cqe = io_get_cqring(ctx);
1701 WRITE_ONCE(cqe->user_data, req->user_data);
1702 WRITE_ONCE(cqe->res, res);
1703 WRITE_ONCE(cqe->flags, cflags);
1704 } else if (ctx->cq_overflow_flushed ||
1705 atomic_read(&req->task->io_uring->in_idle)) {
1707 * If we're in ring overflow flush mode, or in task cancel mode,
1708 * then we cannot store the request for later flushing, we need
1709 * to drop it on the floor.
1711 ctx->cached_cq_overflow++;
1712 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1714 if (list_empty(&ctx->cq_overflow_list)) {
1715 set_bit(0, &ctx->sq_check_overflow);
1716 set_bit(0, &ctx->cq_check_overflow);
1717 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1721 req->compl.cflags = cflags;
1722 refcount_inc(&req->refs);
1723 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1727 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1729 __io_cqring_fill_event(req, res, 0);
1732 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1734 struct io_ring_ctx *ctx = req->ctx;
1735 unsigned long flags;
1737 spin_lock_irqsave(&ctx->completion_lock, flags);
1738 __io_cqring_fill_event(req, res, cflags);
1739 io_commit_cqring(ctx);
1740 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1742 io_cqring_ev_posted(ctx);
1745 static void io_submit_flush_completions(struct io_comp_state *cs)
1747 struct io_ring_ctx *ctx = cs->ctx;
1749 spin_lock_irq(&ctx->completion_lock);
1750 while (!list_empty(&cs->list)) {
1751 struct io_kiocb *req;
1753 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1754 list_del(&req->compl.list);
1755 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1758 * io_free_req() doesn't care about completion_lock unless one
1759 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1760 * because of a potential deadlock with req->work.fs->lock
1762 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1763 |REQ_F_WORK_INITIALIZED)) {
1764 spin_unlock_irq(&ctx->completion_lock);
1766 spin_lock_irq(&ctx->completion_lock);
1771 io_commit_cqring(ctx);
1772 spin_unlock_irq(&ctx->completion_lock);
1774 io_cqring_ev_posted(ctx);
1778 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1779 struct io_comp_state *cs)
1782 io_cqring_add_event(req, res, cflags);
1787 req->compl.cflags = cflags;
1788 list_add_tail(&req->compl.list, &cs->list);
1790 io_submit_flush_completions(cs);
1794 static void io_req_complete(struct io_kiocb *req, long res)
1796 __io_req_complete(req, res, 0, NULL);
1799 static inline bool io_is_fallback_req(struct io_kiocb *req)
1801 return req == (struct io_kiocb *)
1802 ((unsigned long) req->ctx->fallback_req & ~1UL);
1805 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1807 struct io_kiocb *req;
1809 req = ctx->fallback_req;
1810 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1816 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1817 struct io_submit_state *state)
1819 if (!state->free_reqs) {
1820 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1824 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1825 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1828 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1829 * retry single alloc to be on the safe side.
1831 if (unlikely(ret <= 0)) {
1832 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1833 if (!state->reqs[0])
1837 state->free_reqs = ret;
1841 return state->reqs[state->free_reqs];
1843 return io_get_fallback_req(ctx);
1846 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1850 percpu_ref_put(req->fixed_file_refs);
1855 static void io_dismantle_req(struct io_kiocb *req)
1859 if (req->async_data)
1860 kfree(req->async_data);
1862 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1864 io_req_clean_work(req);
1867 static void __io_free_req(struct io_kiocb *req)
1869 struct io_uring_task *tctx = req->task->io_uring;
1870 struct io_ring_ctx *ctx = req->ctx;
1872 io_dismantle_req(req);
1874 percpu_counter_dec(&tctx->inflight);
1875 if (atomic_read(&tctx->in_idle))
1876 wake_up(&tctx->wait);
1877 put_task_struct(req->task);
1879 if (likely(!io_is_fallback_req(req)))
1880 kmem_cache_free(req_cachep, req);
1882 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1883 percpu_ref_put(&ctx->refs);
1886 static void io_kill_linked_timeout(struct io_kiocb *req)
1888 struct io_ring_ctx *ctx = req->ctx;
1889 struct io_kiocb *link;
1890 bool cancelled = false;
1891 unsigned long flags;
1893 spin_lock_irqsave(&ctx->completion_lock, flags);
1894 link = list_first_entry_or_null(&req->link_list, struct io_kiocb,
1897 * Can happen if a linked timeout fired and link had been like
1898 * req -> link t-out -> link t-out [-> ...]
1900 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1901 struct io_timeout_data *io = link->async_data;
1904 list_del_init(&link->link_list);
1905 ret = hrtimer_try_to_cancel(&io->timer);
1907 io_cqring_fill_event(link, -ECANCELED);
1908 io_commit_cqring(ctx);
1912 req->flags &= ~REQ_F_LINK_TIMEOUT;
1913 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1916 io_cqring_ev_posted(ctx);
1921 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1923 struct io_kiocb *nxt;
1926 * The list should never be empty when we are called here. But could
1927 * potentially happen if the chain is messed up, check to be on the
1930 if (unlikely(list_empty(&req->link_list)))
1933 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1934 list_del_init(&req->link_list);
1935 if (!list_empty(&nxt->link_list))
1936 nxt->flags |= REQ_F_LINK_HEAD;
1941 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1943 static void io_fail_links(struct io_kiocb *req)
1945 struct io_ring_ctx *ctx = req->ctx;
1946 unsigned long flags;
1948 spin_lock_irqsave(&ctx->completion_lock, flags);
1949 while (!list_empty(&req->link_list)) {
1950 struct io_kiocb *link = list_first_entry(&req->link_list,
1951 struct io_kiocb, link_list);
1953 list_del_init(&link->link_list);
1954 trace_io_uring_fail_link(req, link);
1956 io_cqring_fill_event(link, -ECANCELED);
1959 * It's ok to free under spinlock as they're not linked anymore,
1960 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1963 if (link->flags & REQ_F_WORK_INITIALIZED)
1964 io_put_req_deferred(link, 2);
1966 io_double_put_req(link);
1969 io_commit_cqring(ctx);
1970 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1972 io_cqring_ev_posted(ctx);
1975 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1977 req->flags &= ~REQ_F_LINK_HEAD;
1978 if (req->flags & REQ_F_LINK_TIMEOUT)
1979 io_kill_linked_timeout(req);
1982 * If LINK is set, we have dependent requests in this chain. If we
1983 * didn't fail this request, queue the first one up, moving any other
1984 * dependencies to the next request. In case of failure, fail the rest
1987 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1988 return io_req_link_next(req);
1993 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1995 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1997 return __io_req_find_next(req);
2000 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
2002 struct task_struct *tsk = req->task;
2003 struct io_ring_ctx *ctx = req->ctx;
2004 enum task_work_notify_mode notify;
2007 if (tsk->flags & PF_EXITING)
2011 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2012 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2013 * processing task_work. There's no reliable way to tell if TWA_RESUME
2017 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
2018 notify = TWA_SIGNAL;
2020 ret = task_work_add(tsk, &req->task_work, notify);
2022 wake_up_process(tsk);
2027 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2029 struct io_ring_ctx *ctx = req->ctx;
2031 spin_lock_irq(&ctx->completion_lock);
2032 io_cqring_fill_event(req, error);
2033 io_commit_cqring(ctx);
2034 spin_unlock_irq(&ctx->completion_lock);
2036 io_cqring_ev_posted(ctx);
2037 req_set_fail_links(req);
2038 io_double_put_req(req);
2041 static void io_req_task_cancel(struct callback_head *cb)
2043 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2044 struct io_ring_ctx *ctx = req->ctx;
2046 __io_req_task_cancel(req, -ECANCELED);
2047 percpu_ref_put(&ctx->refs);
2050 static void __io_req_task_submit(struct io_kiocb *req)
2052 struct io_ring_ctx *ctx = req->ctx;
2054 if (!__io_sq_thread_acquire_mm(ctx)) {
2055 mutex_lock(&ctx->uring_lock);
2056 __io_queue_sqe(req, NULL);
2057 mutex_unlock(&ctx->uring_lock);
2059 __io_req_task_cancel(req, -EFAULT);
2063 static void io_req_task_submit(struct callback_head *cb)
2065 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2066 struct io_ring_ctx *ctx = req->ctx;
2068 __io_req_task_submit(req);
2069 percpu_ref_put(&ctx->refs);
2072 static void io_req_task_queue(struct io_kiocb *req)
2076 init_task_work(&req->task_work, io_req_task_submit);
2077 percpu_ref_get(&req->ctx->refs);
2079 ret = io_req_task_work_add(req, true);
2080 if (unlikely(ret)) {
2081 struct task_struct *tsk;
2083 init_task_work(&req->task_work, io_req_task_cancel);
2084 tsk = io_wq_get_task(req->ctx->io_wq);
2085 task_work_add(tsk, &req->task_work, TWA_NONE);
2086 wake_up_process(tsk);
2090 static void io_queue_next(struct io_kiocb *req)
2092 struct io_kiocb *nxt = io_req_find_next(req);
2095 io_req_task_queue(nxt);
2098 static void io_free_req(struct io_kiocb *req)
2105 void *reqs[IO_IOPOLL_BATCH];
2108 struct task_struct *task;
2112 static inline void io_init_req_batch(struct req_batch *rb)
2119 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2120 struct req_batch *rb)
2122 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2123 percpu_ref_put_many(&ctx->refs, rb->to_free);
2127 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2128 struct req_batch *rb)
2131 __io_req_free_batch_flush(ctx, rb);
2133 struct io_uring_task *tctx = rb->task->io_uring;
2135 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2136 put_task_struct_many(rb->task, rb->task_refs);
2141 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2143 if (unlikely(io_is_fallback_req(req))) {
2147 if (req->flags & REQ_F_LINK_HEAD)
2150 if (req->task != rb->task) {
2152 struct io_uring_task *tctx = rb->task->io_uring;
2154 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2155 put_task_struct_many(rb->task, rb->task_refs);
2157 rb->task = req->task;
2162 io_dismantle_req(req);
2163 rb->reqs[rb->to_free++] = req;
2164 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2165 __io_req_free_batch_flush(req->ctx, rb);
2169 * Drop reference to request, return next in chain (if there is one) if this
2170 * was the last reference to this request.
2172 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2174 struct io_kiocb *nxt = NULL;
2176 if (refcount_dec_and_test(&req->refs)) {
2177 nxt = io_req_find_next(req);
2183 static void io_put_req(struct io_kiocb *req)
2185 if (refcount_dec_and_test(&req->refs))
2189 static void io_put_req_deferred_cb(struct callback_head *cb)
2191 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2196 static void io_free_req_deferred(struct io_kiocb *req)
2200 init_task_work(&req->task_work, io_put_req_deferred_cb);
2201 ret = io_req_task_work_add(req, true);
2202 if (unlikely(ret)) {
2203 struct task_struct *tsk;
2205 tsk = io_wq_get_task(req->ctx->io_wq);
2206 task_work_add(tsk, &req->task_work, TWA_NONE);
2207 wake_up_process(tsk);
2211 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2213 if (refcount_sub_and_test(refs, &req->refs))
2214 io_free_req_deferred(req);
2217 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2219 struct io_kiocb *nxt;
2222 * A ref is owned by io-wq in which context we're. So, if that's the
2223 * last one, it's safe to steal next work. False negatives are Ok,
2224 * it just will be re-punted async in io_put_work()
2226 if (refcount_read(&req->refs) != 1)
2229 nxt = io_req_find_next(req);
2230 return nxt ? &nxt->work : NULL;
2233 static void io_double_put_req(struct io_kiocb *req)
2235 /* drop both submit and complete references */
2236 if (refcount_sub_and_test(2, &req->refs))
2240 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2242 struct io_rings *rings = ctx->rings;
2244 if (test_bit(0, &ctx->cq_check_overflow)) {
2246 * noflush == true is from the waitqueue handler, just ensure
2247 * we wake up the task, and the next invocation will flush the
2248 * entries. We cannot safely to it from here.
2250 if (noflush && !list_empty(&ctx->cq_overflow_list))
2253 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2256 /* See comment at the top of this file */
2258 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2261 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2263 struct io_rings *rings = ctx->rings;
2265 /* make sure SQ entry isn't read before tail */
2266 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2269 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2271 unsigned int cflags;
2273 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2274 cflags |= IORING_CQE_F_BUFFER;
2275 req->flags &= ~REQ_F_BUFFER_SELECTED;
2280 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2282 struct io_buffer *kbuf;
2284 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2285 return io_put_kbuf(req, kbuf);
2288 static inline bool io_run_task_work(void)
2291 * Not safe to run on exiting task, and the task_work handling will
2292 * not add work to such a task.
2294 if (unlikely(current->flags & PF_EXITING))
2296 if (current->task_works) {
2297 __set_current_state(TASK_RUNNING);
2305 static void io_iopoll_queue(struct list_head *again)
2307 struct io_kiocb *req;
2310 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2311 list_del(&req->inflight_entry);
2312 __io_complete_rw(req, -EAGAIN, 0, NULL);
2313 } while (!list_empty(again));
2317 * Find and free completed poll iocbs
2319 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2320 struct list_head *done)
2322 struct req_batch rb;
2323 struct io_kiocb *req;
2326 /* order with ->result store in io_complete_rw_iopoll() */
2329 io_init_req_batch(&rb);
2330 while (!list_empty(done)) {
2333 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2334 if (READ_ONCE(req->result) == -EAGAIN) {
2336 req->iopoll_completed = 0;
2337 list_move_tail(&req->inflight_entry, &again);
2340 list_del(&req->inflight_entry);
2342 if (req->flags & REQ_F_BUFFER_SELECTED)
2343 cflags = io_put_rw_kbuf(req);
2345 __io_cqring_fill_event(req, req->result, cflags);
2348 if (refcount_dec_and_test(&req->refs))
2349 io_req_free_batch(&rb, req);
2352 io_commit_cqring(ctx);
2353 if (ctx->flags & IORING_SETUP_SQPOLL)
2354 io_cqring_ev_posted(ctx);
2355 io_req_free_batch_finish(ctx, &rb);
2357 if (!list_empty(&again))
2358 io_iopoll_queue(&again);
2361 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2364 struct io_kiocb *req, *tmp;
2370 * Only spin for completions if we don't have multiple devices hanging
2371 * off our complete list, and we're under the requested amount.
2373 spin = !ctx->poll_multi_file && *nr_events < min;
2376 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2377 struct kiocb *kiocb = &req->rw.kiocb;
2380 * Move completed and retryable entries to our local lists.
2381 * If we find a request that requires polling, break out
2382 * and complete those lists first, if we have entries there.
2384 if (READ_ONCE(req->iopoll_completed)) {
2385 list_move_tail(&req->inflight_entry, &done);
2388 if (!list_empty(&done))
2391 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2395 /* iopoll may have completed current req */
2396 if (READ_ONCE(req->iopoll_completed))
2397 list_move_tail(&req->inflight_entry, &done);
2404 if (!list_empty(&done))
2405 io_iopoll_complete(ctx, nr_events, &done);
2411 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2412 * non-spinning poll check - we'll still enter the driver poll loop, but only
2413 * as a non-spinning completion check.
2415 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2418 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2421 ret = io_do_iopoll(ctx, nr_events, min);
2424 if (*nr_events >= min)
2432 * We can't just wait for polled events to come to us, we have to actively
2433 * find and complete them.
2435 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2437 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2440 mutex_lock(&ctx->uring_lock);
2441 while (!list_empty(&ctx->iopoll_list)) {
2442 unsigned int nr_events = 0;
2444 io_do_iopoll(ctx, &nr_events, 0);
2446 /* let it sleep and repeat later if can't complete a request */
2450 * Ensure we allow local-to-the-cpu processing to take place,
2451 * in this case we need to ensure that we reap all events.
2452 * Also let task_work, etc. to progress by releasing the mutex
2454 if (need_resched()) {
2455 mutex_unlock(&ctx->uring_lock);
2457 mutex_lock(&ctx->uring_lock);
2460 mutex_unlock(&ctx->uring_lock);
2463 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2465 unsigned int nr_events = 0;
2466 int iters = 0, ret = 0;
2469 * We disallow the app entering submit/complete with polling, but we
2470 * still need to lock the ring to prevent racing with polled issue
2471 * that got punted to a workqueue.
2473 mutex_lock(&ctx->uring_lock);
2476 * Don't enter poll loop if we already have events pending.
2477 * If we do, we can potentially be spinning for commands that
2478 * already triggered a CQE (eg in error).
2480 if (io_cqring_events(ctx, false))
2484 * If a submit got punted to a workqueue, we can have the
2485 * application entering polling for a command before it gets
2486 * issued. That app will hold the uring_lock for the duration
2487 * of the poll right here, so we need to take a breather every
2488 * now and then to ensure that the issue has a chance to add
2489 * the poll to the issued list. Otherwise we can spin here
2490 * forever, while the workqueue is stuck trying to acquire the
2493 if (!(++iters & 7)) {
2494 mutex_unlock(&ctx->uring_lock);
2496 mutex_lock(&ctx->uring_lock);
2499 ret = io_iopoll_getevents(ctx, &nr_events, min);
2503 } while (min && !nr_events && !need_resched());
2505 mutex_unlock(&ctx->uring_lock);
2509 static void kiocb_end_write(struct io_kiocb *req)
2512 * Tell lockdep we inherited freeze protection from submission
2515 if (req->flags & REQ_F_ISREG) {
2516 struct inode *inode = file_inode(req->file);
2518 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2520 file_end_write(req->file);
2523 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2524 struct io_comp_state *cs)
2526 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2529 if (kiocb->ki_flags & IOCB_WRITE)
2530 kiocb_end_write(req);
2532 if (res != req->result)
2533 req_set_fail_links(req);
2534 if (req->flags & REQ_F_BUFFER_SELECTED)
2535 cflags = io_put_rw_kbuf(req);
2536 __io_req_complete(req, res, cflags, cs);
2540 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2542 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2543 ssize_t ret = -ECANCELED;
2544 struct iov_iter iter;
2552 switch (req->opcode) {
2553 case IORING_OP_READV:
2554 case IORING_OP_READ_FIXED:
2555 case IORING_OP_READ:
2558 case IORING_OP_WRITEV:
2559 case IORING_OP_WRITE_FIXED:
2560 case IORING_OP_WRITE:
2564 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2569 if (!req->async_data) {
2570 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2573 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2581 req_set_fail_links(req);
2586 static bool io_rw_reissue(struct io_kiocb *req, long res)
2589 umode_t mode = file_inode(req->file)->i_mode;
2592 if (!S_ISBLK(mode) && !S_ISREG(mode))
2594 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2597 ret = io_sq_thread_acquire_mm(req->ctx, req);
2599 if (io_resubmit_prep(req, ret)) {
2600 refcount_inc(&req->refs);
2601 io_queue_async_work(req);
2609 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2610 struct io_comp_state *cs)
2612 if (!io_rw_reissue(req, res))
2613 io_complete_rw_common(&req->rw.kiocb, res, cs);
2616 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2618 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2620 __io_complete_rw(req, res, res2, NULL);
2623 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2625 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2627 if (kiocb->ki_flags & IOCB_WRITE)
2628 kiocb_end_write(req);
2630 if (res != -EAGAIN && res != req->result)
2631 req_set_fail_links(req);
2633 WRITE_ONCE(req->result, res);
2634 /* order with io_poll_complete() checking ->result */
2636 WRITE_ONCE(req->iopoll_completed, 1);
2640 * After the iocb has been issued, it's safe to be found on the poll list.
2641 * Adding the kiocb to the list AFTER submission ensures that we don't
2642 * find it from a io_iopoll_getevents() thread before the issuer is done
2643 * accessing the kiocb cookie.
2645 static void io_iopoll_req_issued(struct io_kiocb *req)
2647 struct io_ring_ctx *ctx = req->ctx;
2650 * Track whether we have multiple files in our lists. This will impact
2651 * how we do polling eventually, not spinning if we're on potentially
2652 * different devices.
2654 if (list_empty(&ctx->iopoll_list)) {
2655 ctx->poll_multi_file = false;
2656 } else if (!ctx->poll_multi_file) {
2657 struct io_kiocb *list_req;
2659 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2661 if (list_req->file != req->file)
2662 ctx->poll_multi_file = true;
2666 * For fast devices, IO may have already completed. If it has, add
2667 * it to the front so we find it first.
2669 if (READ_ONCE(req->iopoll_completed))
2670 list_add(&req->inflight_entry, &ctx->iopoll_list);
2672 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2674 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2675 wq_has_sleeper(&ctx->sq_data->wait))
2676 wake_up(&ctx->sq_data->wait);
2679 static void __io_state_file_put(struct io_submit_state *state)
2681 if (state->has_refs)
2682 fput_many(state->file, state->has_refs);
2686 static inline void io_state_file_put(struct io_submit_state *state)
2689 __io_state_file_put(state);
2693 * Get as many references to a file as we have IOs left in this submission,
2694 * assuming most submissions are for one file, or at least that each file
2695 * has more than one submission.
2697 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2703 if (state->fd == fd) {
2707 __io_state_file_put(state);
2709 state->file = fget_many(fd, state->ios_left);
2714 state->has_refs = state->ios_left - 1;
2718 static bool io_bdev_nowait(struct block_device *bdev)
2721 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2728 * If we tracked the file through the SCM inflight mechanism, we could support
2729 * any file. For now, just ensure that anything potentially problematic is done
2732 static bool io_file_supports_async(struct file *file, int rw)
2734 umode_t mode = file_inode(file)->i_mode;
2736 if (S_ISBLK(mode)) {
2737 if (io_bdev_nowait(file->f_inode->i_bdev))
2741 if (S_ISCHR(mode) || S_ISSOCK(mode))
2743 if (S_ISREG(mode)) {
2744 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2745 file->f_op != &io_uring_fops)
2750 /* any ->read/write should understand O_NONBLOCK */
2751 if (file->f_flags & O_NONBLOCK)
2754 if (!(file->f_mode & FMODE_NOWAIT))
2758 return file->f_op->read_iter != NULL;
2760 return file->f_op->write_iter != NULL;
2763 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2765 struct io_ring_ctx *ctx = req->ctx;
2766 struct kiocb *kiocb = &req->rw.kiocb;
2770 if (S_ISREG(file_inode(req->file)->i_mode))
2771 req->flags |= REQ_F_ISREG;
2773 kiocb->ki_pos = READ_ONCE(sqe->off);
2774 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2775 req->flags |= REQ_F_CUR_POS;
2776 kiocb->ki_pos = req->file->f_pos;
2778 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2779 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2780 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2784 ioprio = READ_ONCE(sqe->ioprio);
2786 ret = ioprio_check_cap(ioprio);
2790 kiocb->ki_ioprio = ioprio;
2792 kiocb->ki_ioprio = get_current_ioprio();
2794 /* don't allow async punt if RWF_NOWAIT was requested */
2795 if (kiocb->ki_flags & IOCB_NOWAIT)
2796 req->flags |= REQ_F_NOWAIT;
2798 if (ctx->flags & IORING_SETUP_IOPOLL) {
2799 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2800 !kiocb->ki_filp->f_op->iopoll)
2803 kiocb->ki_flags |= IOCB_HIPRI;
2804 kiocb->ki_complete = io_complete_rw_iopoll;
2805 req->iopoll_completed = 0;
2807 if (kiocb->ki_flags & IOCB_HIPRI)
2809 kiocb->ki_complete = io_complete_rw;
2812 req->rw.addr = READ_ONCE(sqe->addr);
2813 req->rw.len = READ_ONCE(sqe->len);
2814 req->buf_index = READ_ONCE(sqe->buf_index);
2818 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2824 case -ERESTARTNOINTR:
2825 case -ERESTARTNOHAND:
2826 case -ERESTART_RESTARTBLOCK:
2828 * We can't just restart the syscall, since previously
2829 * submitted sqes may already be in progress. Just fail this
2835 kiocb->ki_complete(kiocb, ret, 0);
2839 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2840 struct io_comp_state *cs)
2842 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2843 struct io_async_rw *io = req->async_data;
2845 /* add previously done IO, if any */
2846 if (io && io->bytes_done > 0) {
2848 ret = io->bytes_done;
2850 ret += io->bytes_done;
2853 if (req->flags & REQ_F_CUR_POS)
2854 req->file->f_pos = kiocb->ki_pos;
2855 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2856 __io_complete_rw(req, ret, 0, cs);
2858 io_rw_done(kiocb, ret);
2861 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2862 struct iov_iter *iter)
2864 struct io_ring_ctx *ctx = req->ctx;
2865 size_t len = req->rw.len;
2866 struct io_mapped_ubuf *imu;
2867 u16 index, buf_index = req->buf_index;
2871 if (unlikely(buf_index >= ctx->nr_user_bufs))
2873 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2874 imu = &ctx->user_bufs[index];
2875 buf_addr = req->rw.addr;
2878 if (buf_addr + len < buf_addr)
2880 /* not inside the mapped region */
2881 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2885 * May not be a start of buffer, set size appropriately
2886 * and advance us to the beginning.
2888 offset = buf_addr - imu->ubuf;
2889 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2893 * Don't use iov_iter_advance() here, as it's really slow for
2894 * using the latter parts of a big fixed buffer - it iterates
2895 * over each segment manually. We can cheat a bit here, because
2898 * 1) it's a BVEC iter, we set it up
2899 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2900 * first and last bvec
2902 * So just find our index, and adjust the iterator afterwards.
2903 * If the offset is within the first bvec (or the whole first
2904 * bvec, just use iov_iter_advance(). This makes it easier
2905 * since we can just skip the first segment, which may not
2906 * be PAGE_SIZE aligned.
2908 const struct bio_vec *bvec = imu->bvec;
2910 if (offset <= bvec->bv_len) {
2911 iov_iter_advance(iter, offset);
2913 unsigned long seg_skip;
2915 /* skip first vec */
2916 offset -= bvec->bv_len;
2917 seg_skip = 1 + (offset >> PAGE_SHIFT);
2919 iter->bvec = bvec + seg_skip;
2920 iter->nr_segs -= seg_skip;
2921 iter->count -= bvec->bv_len + offset;
2922 iter->iov_offset = offset & ~PAGE_MASK;
2929 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2932 mutex_unlock(&ctx->uring_lock);
2935 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2938 * "Normal" inline submissions always hold the uring_lock, since we
2939 * grab it from the system call. Same is true for the SQPOLL offload.
2940 * The only exception is when we've detached the request and issue it
2941 * from an async worker thread, grab the lock for that case.
2944 mutex_lock(&ctx->uring_lock);
2947 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2948 int bgid, struct io_buffer *kbuf,
2951 struct io_buffer *head;
2953 if (req->flags & REQ_F_BUFFER_SELECTED)
2956 io_ring_submit_lock(req->ctx, needs_lock);
2958 lockdep_assert_held(&req->ctx->uring_lock);
2960 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2962 if (!list_empty(&head->list)) {
2963 kbuf = list_last_entry(&head->list, struct io_buffer,
2965 list_del(&kbuf->list);
2968 idr_remove(&req->ctx->io_buffer_idr, bgid);
2970 if (*len > kbuf->len)
2973 kbuf = ERR_PTR(-ENOBUFS);
2976 io_ring_submit_unlock(req->ctx, needs_lock);
2981 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2984 struct io_buffer *kbuf;
2987 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2988 bgid = req->buf_index;
2989 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2992 req->rw.addr = (u64) (unsigned long) kbuf;
2993 req->flags |= REQ_F_BUFFER_SELECTED;
2994 return u64_to_user_ptr(kbuf->addr);
2997 #ifdef CONFIG_COMPAT
2998 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3001 struct compat_iovec __user *uiov;
3002 compat_ssize_t clen;
3006 uiov = u64_to_user_ptr(req->rw.addr);
3007 if (!access_ok(uiov, sizeof(*uiov)))
3009 if (__get_user(clen, &uiov->iov_len))
3015 buf = io_rw_buffer_select(req, &len, needs_lock);
3017 return PTR_ERR(buf);
3018 iov[0].iov_base = buf;
3019 iov[0].iov_len = (compat_size_t) len;
3024 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3027 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3031 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3034 len = iov[0].iov_len;
3037 buf = io_rw_buffer_select(req, &len, needs_lock);
3039 return PTR_ERR(buf);
3040 iov[0].iov_base = buf;
3041 iov[0].iov_len = len;
3045 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3048 if (req->flags & REQ_F_BUFFER_SELECTED) {
3049 struct io_buffer *kbuf;
3051 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3052 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3053 iov[0].iov_len = kbuf->len;
3058 else if (req->rw.len > 1)
3061 #ifdef CONFIG_COMPAT
3062 if (req->ctx->compat)
3063 return io_compat_import(req, iov, needs_lock);
3066 return __io_iov_buffer_select(req, iov, needs_lock);
3069 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3070 struct iovec **iovec, struct iov_iter *iter,
3073 void __user *buf = u64_to_user_ptr(req->rw.addr);
3074 size_t sqe_len = req->rw.len;
3078 opcode = req->opcode;
3079 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3081 return io_import_fixed(req, rw, iter);
3084 /* buffer index only valid with fixed read/write, or buffer select */
3085 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3088 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3089 if (req->flags & REQ_F_BUFFER_SELECT) {
3090 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3092 return PTR_ERR(buf);
3093 req->rw.len = sqe_len;
3096 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3098 return ret < 0 ? ret : sqe_len;
3101 if (req->flags & REQ_F_BUFFER_SELECT) {
3102 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3104 ret = (*iovec)->iov_len;
3105 iov_iter_init(iter, rw, *iovec, 1, ret);
3111 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3115 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3116 struct iovec **iovec, struct iov_iter *iter,
3119 struct io_async_rw *iorw = req->async_data;
3122 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3124 return iov_iter_count(&iorw->iter);
3127 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3129 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3133 * For files that don't have ->read_iter() and ->write_iter(), handle them
3134 * by looping over ->read() or ->write() manually.
3136 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3138 struct kiocb *kiocb = &req->rw.kiocb;
3139 struct file *file = req->file;
3143 * Don't support polled IO through this interface, and we can't
3144 * support non-blocking either. For the latter, this just causes
3145 * the kiocb to be handled from an async context.
3147 if (kiocb->ki_flags & IOCB_HIPRI)
3149 if (kiocb->ki_flags & IOCB_NOWAIT)
3152 while (iov_iter_count(iter)) {
3156 if (!iov_iter_is_bvec(iter)) {
3157 iovec = iov_iter_iovec(iter);
3159 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3160 iovec.iov_len = req->rw.len;
3164 nr = file->f_op->read(file, iovec.iov_base,
3165 iovec.iov_len, io_kiocb_ppos(kiocb));
3167 nr = file->f_op->write(file, iovec.iov_base,
3168 iovec.iov_len, io_kiocb_ppos(kiocb));
3177 if (nr != iovec.iov_len)
3181 iov_iter_advance(iter, nr);
3187 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3188 const struct iovec *fast_iov, struct iov_iter *iter)
3190 struct io_async_rw *rw = req->async_data;
3192 memcpy(&rw->iter, iter, sizeof(*iter));
3193 rw->free_iovec = iovec;
3195 /* can only be fixed buffers, no need to do anything */
3196 if (iter->type == ITER_BVEC)
3199 unsigned iov_off = 0;
3201 rw->iter.iov = rw->fast_iov;
3202 if (iter->iov != fast_iov) {
3203 iov_off = iter->iov - fast_iov;
3204 rw->iter.iov += iov_off;
3206 if (rw->fast_iov != fast_iov)
3207 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3208 sizeof(struct iovec) * iter->nr_segs);
3210 req->flags |= REQ_F_NEED_CLEANUP;
3214 static inline int __io_alloc_async_data(struct io_kiocb *req)
3216 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3217 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3218 return req->async_data == NULL;
3221 static int io_alloc_async_data(struct io_kiocb *req)
3223 if (!io_op_defs[req->opcode].needs_async_data)
3226 return __io_alloc_async_data(req);
3229 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3230 const struct iovec *fast_iov,
3231 struct iov_iter *iter, bool force)
3233 if (!force && !io_op_defs[req->opcode].needs_async_data)
3235 if (!req->async_data) {
3236 if (__io_alloc_async_data(req))
3239 io_req_map_rw(req, iovec, fast_iov, iter);
3244 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3246 struct io_async_rw *iorw = req->async_data;
3247 struct iovec *iov = iorw->fast_iov;
3250 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3251 if (unlikely(ret < 0))
3254 iorw->bytes_done = 0;
3255 iorw->free_iovec = iov;
3257 req->flags |= REQ_F_NEED_CLEANUP;
3261 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3265 ret = io_prep_rw(req, sqe);
3269 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3272 /* either don't need iovec imported or already have it */
3273 if (!req->async_data)
3275 return io_rw_prep_async(req, READ);
3279 * This is our waitqueue callback handler, registered through lock_page_async()
3280 * when we initially tried to do the IO with the iocb armed our waitqueue.
3281 * This gets called when the page is unlocked, and we generally expect that to
3282 * happen when the page IO is completed and the page is now uptodate. This will
3283 * queue a task_work based retry of the operation, attempting to copy the data
3284 * again. If the latter fails because the page was NOT uptodate, then we will
3285 * do a thread based blocking retry of the operation. That's the unexpected
3288 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3289 int sync, void *arg)
3291 struct wait_page_queue *wpq;
3292 struct io_kiocb *req = wait->private;
3293 struct wait_page_key *key = arg;
3296 wpq = container_of(wait, struct wait_page_queue, wait);
3298 if (!wake_page_match(wpq, key))
3301 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3302 list_del_init(&wait->entry);
3304 init_task_work(&req->task_work, io_req_task_submit);
3305 percpu_ref_get(&req->ctx->refs);
3307 /* submit ref gets dropped, acquire a new one */
3308 refcount_inc(&req->refs);
3309 ret = io_req_task_work_add(req, true);
3310 if (unlikely(ret)) {
3311 struct task_struct *tsk;
3313 /* queue just for cancelation */
3314 init_task_work(&req->task_work, io_req_task_cancel);
3315 tsk = io_wq_get_task(req->ctx->io_wq);
3316 task_work_add(tsk, &req->task_work, TWA_NONE);
3317 wake_up_process(tsk);
3323 * This controls whether a given IO request should be armed for async page
3324 * based retry. If we return false here, the request is handed to the async
3325 * worker threads for retry. If we're doing buffered reads on a regular file,
3326 * we prepare a private wait_page_queue entry and retry the operation. This
3327 * will either succeed because the page is now uptodate and unlocked, or it
3328 * will register a callback when the page is unlocked at IO completion. Through
3329 * that callback, io_uring uses task_work to setup a retry of the operation.
3330 * That retry will attempt the buffered read again. The retry will generally
3331 * succeed, or in rare cases where it fails, we then fall back to using the
3332 * async worker threads for a blocking retry.
3334 static bool io_rw_should_retry(struct io_kiocb *req)
3336 struct io_async_rw *rw = req->async_data;
3337 struct wait_page_queue *wait = &rw->wpq;
3338 struct kiocb *kiocb = &req->rw.kiocb;
3340 /* never retry for NOWAIT, we just complete with -EAGAIN */
3341 if (req->flags & REQ_F_NOWAIT)
3344 /* Only for buffered IO */
3345 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3349 * just use poll if we can, and don't attempt if the fs doesn't
3350 * support callback based unlocks
3352 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3355 wait->wait.func = io_async_buf_func;
3356 wait->wait.private = req;
3357 wait->wait.flags = 0;
3358 INIT_LIST_HEAD(&wait->wait.entry);
3359 kiocb->ki_flags |= IOCB_WAITQ;
3360 kiocb->ki_flags &= ~IOCB_NOWAIT;
3361 kiocb->ki_waitq = wait;
3365 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3367 if (req->file->f_op->read_iter)
3368 return call_read_iter(req->file, &req->rw.kiocb, iter);
3369 else if (req->file->f_op->read)
3370 return loop_rw_iter(READ, req, iter);
3375 static int io_read(struct io_kiocb *req, bool force_nonblock,
3376 struct io_comp_state *cs)
3378 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3379 struct kiocb *kiocb = &req->rw.kiocb;
3380 struct iov_iter __iter, *iter = &__iter;
3381 struct io_async_rw *rw = req->async_data;
3382 ssize_t io_size, ret, ret2;
3389 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3392 iov_count = iov_iter_count(iter);
3394 req->result = io_size;
3397 /* Ensure we clear previously set non-block flag */
3398 if (!force_nonblock)
3399 kiocb->ki_flags &= ~IOCB_NOWAIT;
3401 kiocb->ki_flags |= IOCB_NOWAIT;
3404 /* If the file doesn't support async, just async punt */
3405 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3409 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3413 ret = io_iter_do_read(req, iter);
3417 } else if (ret == -EIOCBQUEUED) {
3420 } else if (ret == -EAGAIN) {
3421 /* IOPOLL retry should happen for io-wq threads */
3422 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3424 /* no retry on NONBLOCK marked file */
3425 if (req->file->f_flags & O_NONBLOCK)
3427 /* some cases will consume bytes even on error returns */
3428 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3431 } else if (ret < 0) {
3432 /* make sure -ERESTARTSYS -> -EINTR is done */
3436 /* read it all, or we did blocking attempt. no retry. */
3437 if (!iov_iter_count(iter) || !force_nonblock ||
3438 (req->file->f_flags & O_NONBLOCK))
3443 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3450 rw = req->async_data;
3451 /* it's copied and will be cleaned with ->io */
3453 /* now use our persistent iterator, if we aren't already */
3456 rw->bytes_done += ret;
3457 /* if we can retry, do so with the callbacks armed */
3458 if (!io_rw_should_retry(req)) {
3459 kiocb->ki_flags &= ~IOCB_WAITQ;
3464 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3465 * get -EIOCBQUEUED, then we'll get a notification when the desired
3466 * page gets unlocked. We can also get a partial read here, and if we
3467 * do, then just retry at the new offset.
3469 ret = io_iter_do_read(req, iter);
3470 if (ret == -EIOCBQUEUED) {
3473 } else if (ret > 0 && ret < io_size) {
3474 /* we got some bytes, but not all. retry. */
3478 kiocb_done(kiocb, ret, cs);
3481 /* it's reportedly faster than delegating the null check to kfree() */
3487 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3491 ret = io_prep_rw(req, sqe);
3495 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3498 /* either don't need iovec imported or already have it */
3499 if (!req->async_data)
3501 return io_rw_prep_async(req, WRITE);
3504 static int io_write(struct io_kiocb *req, bool force_nonblock,
3505 struct io_comp_state *cs)
3507 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3508 struct kiocb *kiocb = &req->rw.kiocb;
3509 struct iov_iter __iter, *iter = &__iter;
3510 struct io_async_rw *rw = req->async_data;
3512 ssize_t ret, ret2, io_size;
3517 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3520 iov_count = iov_iter_count(iter);
3522 req->result = io_size;
3524 /* Ensure we clear previously set non-block flag */
3525 if (!force_nonblock)
3526 kiocb->ki_flags &= ~IOCB_NOWAIT;
3528 kiocb->ki_flags |= IOCB_NOWAIT;
3530 /* If the file doesn't support async, just async punt */
3531 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3534 /* file path doesn't support NOWAIT for non-direct_IO */
3535 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3536 (req->flags & REQ_F_ISREG))
3539 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3544 * Open-code file_start_write here to grab freeze protection,
3545 * which will be released by another thread in
3546 * io_complete_rw(). Fool lockdep by telling it the lock got
3547 * released so that it doesn't complain about the held lock when
3548 * we return to userspace.
3550 if (req->flags & REQ_F_ISREG) {
3551 __sb_start_write(file_inode(req->file)->i_sb,
3552 SB_FREEZE_WRITE, true);
3553 __sb_writers_release(file_inode(req->file)->i_sb,
3556 kiocb->ki_flags |= IOCB_WRITE;
3558 if (req->file->f_op->write_iter)
3559 ret2 = call_write_iter(req->file, kiocb, iter);
3560 else if (req->file->f_op->write)
3561 ret2 = loop_rw_iter(WRITE, req, iter);
3566 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3567 * retry them without IOCB_NOWAIT.
3569 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3571 /* no retry on NONBLOCK marked file */
3572 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3574 if (!force_nonblock || ret2 != -EAGAIN) {
3575 /* IOPOLL retry should happen for io-wq threads */
3576 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3579 kiocb_done(kiocb, ret2, cs);
3582 /* some cases will consume bytes even on error returns */
3583 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3584 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3589 /* it's reportedly faster than delegating the null check to kfree() */
3595 static int __io_splice_prep(struct io_kiocb *req,
3596 const struct io_uring_sqe *sqe)
3598 struct io_splice* sp = &req->splice;
3599 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3601 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3605 sp->len = READ_ONCE(sqe->len);
3606 sp->flags = READ_ONCE(sqe->splice_flags);
3608 if (unlikely(sp->flags & ~valid_flags))
3611 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3612 (sp->flags & SPLICE_F_FD_IN_FIXED));
3615 req->flags |= REQ_F_NEED_CLEANUP;
3617 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3619 * Splice operation will be punted aync, and here need to
3620 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3622 io_req_init_async(req);
3623 req->work.flags |= IO_WQ_WORK_UNBOUND;
3629 static int io_tee_prep(struct io_kiocb *req,
3630 const struct io_uring_sqe *sqe)
3632 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3634 return __io_splice_prep(req, sqe);
3637 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3639 struct io_splice *sp = &req->splice;
3640 struct file *in = sp->file_in;
3641 struct file *out = sp->file_out;
3642 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3648 ret = do_tee(in, out, sp->len, flags);
3650 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3651 req->flags &= ~REQ_F_NEED_CLEANUP;
3654 req_set_fail_links(req);
3655 io_req_complete(req, ret);
3659 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3661 struct io_splice* sp = &req->splice;
3663 sp->off_in = READ_ONCE(sqe->splice_off_in);
3664 sp->off_out = READ_ONCE(sqe->off);
3665 return __io_splice_prep(req, sqe);
3668 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3670 struct io_splice *sp = &req->splice;
3671 struct file *in = sp->file_in;
3672 struct file *out = sp->file_out;
3673 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3674 loff_t *poff_in, *poff_out;
3680 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3681 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3684 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3686 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3687 req->flags &= ~REQ_F_NEED_CLEANUP;
3690 req_set_fail_links(req);
3691 io_req_complete(req, ret);
3696 * IORING_OP_NOP just posts a completion event, nothing else.
3698 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3700 struct io_ring_ctx *ctx = req->ctx;
3702 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3705 __io_req_complete(req, 0, 0, cs);
3709 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3711 struct io_ring_ctx *ctx = req->ctx;
3716 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3718 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3721 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3722 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3725 req->sync.off = READ_ONCE(sqe->off);
3726 req->sync.len = READ_ONCE(sqe->len);
3730 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3732 loff_t end = req->sync.off + req->sync.len;
3735 /* fsync always requires a blocking context */
3739 ret = vfs_fsync_range(req->file, req->sync.off,
3740 end > 0 ? end : LLONG_MAX,
3741 req->sync.flags & IORING_FSYNC_DATASYNC);
3743 req_set_fail_links(req);
3744 io_req_complete(req, ret);
3748 static int io_fallocate_prep(struct io_kiocb *req,
3749 const struct io_uring_sqe *sqe)
3751 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3753 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3756 req->sync.off = READ_ONCE(sqe->off);
3757 req->sync.len = READ_ONCE(sqe->addr);
3758 req->sync.mode = READ_ONCE(sqe->len);
3762 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3766 /* fallocate always requiring blocking context */
3769 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3772 req_set_fail_links(req);
3773 io_req_complete(req, ret);
3777 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3779 const char __user *fname;
3782 if (unlikely(sqe->ioprio || sqe->buf_index))
3784 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3787 /* open.how should be already initialised */
3788 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3789 req->open.how.flags |= O_LARGEFILE;
3791 req->open.dfd = READ_ONCE(sqe->fd);
3792 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3793 req->open.filename = getname(fname);
3794 if (IS_ERR(req->open.filename)) {
3795 ret = PTR_ERR(req->open.filename);
3796 req->open.filename = NULL;
3799 req->open.nofile = rlimit(RLIMIT_NOFILE);
3800 req->open.ignore_nonblock = false;
3801 req->flags |= REQ_F_NEED_CLEANUP;
3805 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3809 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3811 mode = READ_ONCE(sqe->len);
3812 flags = READ_ONCE(sqe->open_flags);
3813 req->open.how = build_open_how(flags, mode);
3814 return __io_openat_prep(req, sqe);
3817 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3819 struct open_how __user *how;
3823 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3825 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3826 len = READ_ONCE(sqe->len);
3827 if (len < OPEN_HOW_SIZE_VER0)
3830 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3835 return __io_openat_prep(req, sqe);
3838 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3840 struct open_flags op;
3844 if (force_nonblock && !req->open.ignore_nonblock)
3847 ret = build_open_flags(&req->open.how, &op);
3851 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3855 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3858 ret = PTR_ERR(file);
3860 * A work-around to ensure that /proc/self works that way
3861 * that it should - if we get -EOPNOTSUPP back, then assume
3862 * that proc_self_get_link() failed us because we're in async
3863 * context. We should be safe to retry this from the task
3864 * itself with force_nonblock == false set, as it should not
3865 * block on lookup. Would be nice to know this upfront and
3866 * avoid the async dance, but doesn't seem feasible.
3868 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
3869 req->open.ignore_nonblock = true;
3870 refcount_inc(&req->refs);
3871 io_req_task_queue(req);
3875 fsnotify_open(file);
3876 fd_install(ret, file);
3879 putname(req->open.filename);
3880 req->flags &= ~REQ_F_NEED_CLEANUP;
3882 req_set_fail_links(req);
3883 io_req_complete(req, ret);
3887 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3889 return io_openat2(req, force_nonblock);
3892 static int io_remove_buffers_prep(struct io_kiocb *req,
3893 const struct io_uring_sqe *sqe)
3895 struct io_provide_buf *p = &req->pbuf;
3898 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3901 tmp = READ_ONCE(sqe->fd);
3902 if (!tmp || tmp > USHRT_MAX)
3905 memset(p, 0, sizeof(*p));
3907 p->bgid = READ_ONCE(sqe->buf_group);
3911 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3912 int bgid, unsigned nbufs)
3916 /* shouldn't happen */
3920 /* the head kbuf is the list itself */
3921 while (!list_empty(&buf->list)) {
3922 struct io_buffer *nxt;
3924 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3925 list_del(&nxt->list);
3932 idr_remove(&ctx->io_buffer_idr, bgid);
3937 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3938 struct io_comp_state *cs)
3940 struct io_provide_buf *p = &req->pbuf;
3941 struct io_ring_ctx *ctx = req->ctx;
3942 struct io_buffer *head;
3945 io_ring_submit_lock(ctx, !force_nonblock);
3947 lockdep_assert_held(&ctx->uring_lock);
3950 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3952 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3954 io_ring_submit_lock(ctx, !force_nonblock);
3956 req_set_fail_links(req);
3957 __io_req_complete(req, ret, 0, cs);
3961 static int io_provide_buffers_prep(struct io_kiocb *req,
3962 const struct io_uring_sqe *sqe)
3964 struct io_provide_buf *p = &req->pbuf;
3967 if (sqe->ioprio || sqe->rw_flags)
3970 tmp = READ_ONCE(sqe->fd);
3971 if (!tmp || tmp > USHRT_MAX)
3974 p->addr = READ_ONCE(sqe->addr);
3975 p->len = READ_ONCE(sqe->len);
3977 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3980 p->bgid = READ_ONCE(sqe->buf_group);
3981 tmp = READ_ONCE(sqe->off);
3982 if (tmp > USHRT_MAX)
3988 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3990 struct io_buffer *buf;
3991 u64 addr = pbuf->addr;
3992 int i, bid = pbuf->bid;
3994 for (i = 0; i < pbuf->nbufs; i++) {
3995 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4000 buf->len = pbuf->len;
4005 INIT_LIST_HEAD(&buf->list);
4008 list_add_tail(&buf->list, &(*head)->list);
4012 return i ? i : -ENOMEM;
4015 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4016 struct io_comp_state *cs)
4018 struct io_provide_buf *p = &req->pbuf;
4019 struct io_ring_ctx *ctx = req->ctx;
4020 struct io_buffer *head, *list;
4023 io_ring_submit_lock(ctx, !force_nonblock);
4025 lockdep_assert_held(&ctx->uring_lock);
4027 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4029 ret = io_add_buffers(p, &head);
4034 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4037 __io_remove_buffers(ctx, head, p->bgid, -1U);
4042 io_ring_submit_unlock(ctx, !force_nonblock);
4044 req_set_fail_links(req);
4045 __io_req_complete(req, ret, 0, cs);
4049 static int io_epoll_ctl_prep(struct io_kiocb *req,
4050 const struct io_uring_sqe *sqe)
4052 #if defined(CONFIG_EPOLL)
4053 if (sqe->ioprio || sqe->buf_index)
4055 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4058 req->epoll.epfd = READ_ONCE(sqe->fd);
4059 req->epoll.op = READ_ONCE(sqe->len);
4060 req->epoll.fd = READ_ONCE(sqe->off);
4062 if (ep_op_has_event(req->epoll.op)) {
4063 struct epoll_event __user *ev;
4065 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4066 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4076 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4077 struct io_comp_state *cs)
4079 #if defined(CONFIG_EPOLL)
4080 struct io_epoll *ie = &req->epoll;
4083 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4084 if (force_nonblock && ret == -EAGAIN)
4088 req_set_fail_links(req);
4089 __io_req_complete(req, ret, 0, cs);
4096 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4098 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4099 if (sqe->ioprio || sqe->buf_index || sqe->off)
4101 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4104 req->madvise.addr = READ_ONCE(sqe->addr);
4105 req->madvise.len = READ_ONCE(sqe->len);
4106 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4113 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4115 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4116 struct io_madvise *ma = &req->madvise;
4122 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4124 req_set_fail_links(req);
4125 io_req_complete(req, ret);
4132 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4134 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4136 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4139 req->fadvise.offset = READ_ONCE(sqe->off);
4140 req->fadvise.len = READ_ONCE(sqe->len);
4141 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4145 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4147 struct io_fadvise *fa = &req->fadvise;
4150 if (force_nonblock) {
4151 switch (fa->advice) {
4152 case POSIX_FADV_NORMAL:
4153 case POSIX_FADV_RANDOM:
4154 case POSIX_FADV_SEQUENTIAL:
4161 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4163 req_set_fail_links(req);
4164 io_req_complete(req, ret);
4168 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4170 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4172 if (sqe->ioprio || sqe->buf_index)
4174 if (req->flags & REQ_F_FIXED_FILE)
4177 req->statx.dfd = READ_ONCE(sqe->fd);
4178 req->statx.mask = READ_ONCE(sqe->len);
4179 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4180 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4181 req->statx.flags = READ_ONCE(sqe->statx_flags);
4186 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4188 struct io_statx *ctx = &req->statx;
4191 if (force_nonblock) {
4192 /* only need file table for an actual valid fd */
4193 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4194 req->flags |= REQ_F_NO_FILE_TABLE;
4198 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4202 req_set_fail_links(req);
4203 io_req_complete(req, ret);
4207 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4210 * If we queue this for async, it must not be cancellable. That would
4211 * leave the 'file' in an undeterminate state, and here need to modify
4212 * io_wq_work.flags, so initialize io_wq_work firstly.
4214 io_req_init_async(req);
4215 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4217 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4219 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4220 sqe->rw_flags || sqe->buf_index)
4222 if (req->flags & REQ_F_FIXED_FILE)
4225 req->close.fd = READ_ONCE(sqe->fd);
4226 if ((req->file && req->file->f_op == &io_uring_fops))
4229 req->close.put_file = NULL;
4233 static int io_close(struct io_kiocb *req, bool force_nonblock,
4234 struct io_comp_state *cs)
4236 struct io_close *close = &req->close;
4239 /* might be already done during nonblock submission */
4240 if (!close->put_file) {
4241 ret = __close_fd_get_file(close->fd, &close->put_file);
4243 return (ret == -ENOENT) ? -EBADF : ret;
4246 /* if the file has a flush method, be safe and punt to async */
4247 if (close->put_file->f_op->flush && force_nonblock) {
4248 /* was never set, but play safe */
4249 req->flags &= ~REQ_F_NOWAIT;
4250 /* avoid grabbing files - we don't need the files */
4251 req->flags |= REQ_F_NO_FILE_TABLE;
4255 /* No ->flush() or already async, safely close from here */
4256 ret = filp_close(close->put_file, req->work.identity->files);
4258 req_set_fail_links(req);
4259 fput(close->put_file);
4260 close->put_file = NULL;
4261 __io_req_complete(req, ret, 0, cs);
4265 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4267 struct io_ring_ctx *ctx = req->ctx;
4272 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4274 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4277 req->sync.off = READ_ONCE(sqe->off);
4278 req->sync.len = READ_ONCE(sqe->len);
4279 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4283 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4287 /* sync_file_range always requires a blocking context */
4291 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4294 req_set_fail_links(req);
4295 io_req_complete(req, ret);
4299 #if defined(CONFIG_NET)
4300 static int io_setup_async_msg(struct io_kiocb *req,
4301 struct io_async_msghdr *kmsg)
4303 struct io_async_msghdr *async_msg = req->async_data;
4307 if (io_alloc_async_data(req)) {
4308 if (kmsg->iov != kmsg->fast_iov)
4312 async_msg = req->async_data;
4313 req->flags |= REQ_F_NEED_CLEANUP;
4314 memcpy(async_msg, kmsg, sizeof(*kmsg));
4318 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4319 struct io_async_msghdr *iomsg)
4321 iomsg->iov = iomsg->fast_iov;
4322 iomsg->msg.msg_name = &iomsg->addr;
4323 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4324 req->sr_msg.msg_flags, &iomsg->iov);
4327 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4329 struct io_async_msghdr *async_msg = req->async_data;
4330 struct io_sr_msg *sr = &req->sr_msg;
4333 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4336 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4337 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4338 sr->len = READ_ONCE(sqe->len);
4340 #ifdef CONFIG_COMPAT
4341 if (req->ctx->compat)
4342 sr->msg_flags |= MSG_CMSG_COMPAT;
4345 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4347 ret = io_sendmsg_copy_hdr(req, async_msg);
4349 req->flags |= REQ_F_NEED_CLEANUP;
4353 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4354 struct io_comp_state *cs)
4356 struct io_async_msghdr iomsg, *kmsg;
4357 struct socket *sock;
4361 sock = sock_from_file(req->file, &ret);
4362 if (unlikely(!sock))
4365 if (req->async_data) {
4366 kmsg = req->async_data;
4367 kmsg->msg.msg_name = &kmsg->addr;
4368 /* if iov is set, it's allocated already */
4370 kmsg->iov = kmsg->fast_iov;
4371 kmsg->msg.msg_iter.iov = kmsg->iov;
4373 ret = io_sendmsg_copy_hdr(req, &iomsg);
4379 flags = req->sr_msg.msg_flags;
4380 if (flags & MSG_DONTWAIT)
4381 req->flags |= REQ_F_NOWAIT;
4382 else if (force_nonblock)
4383 flags |= MSG_DONTWAIT;
4385 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4386 if (force_nonblock && ret == -EAGAIN)
4387 return io_setup_async_msg(req, kmsg);
4388 if (ret == -ERESTARTSYS)
4391 if (kmsg->iov != kmsg->fast_iov)
4393 req->flags &= ~REQ_F_NEED_CLEANUP;
4395 req_set_fail_links(req);
4396 __io_req_complete(req, ret, 0, cs);
4400 static int io_send(struct io_kiocb *req, bool force_nonblock,
4401 struct io_comp_state *cs)
4403 struct io_sr_msg *sr = &req->sr_msg;
4406 struct socket *sock;
4410 sock = sock_from_file(req->file, &ret);
4411 if (unlikely(!sock))
4414 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4418 msg.msg_name = NULL;
4419 msg.msg_control = NULL;
4420 msg.msg_controllen = 0;
4421 msg.msg_namelen = 0;
4423 flags = req->sr_msg.msg_flags;
4424 if (flags & MSG_DONTWAIT)
4425 req->flags |= REQ_F_NOWAIT;
4426 else if (force_nonblock)
4427 flags |= MSG_DONTWAIT;
4429 msg.msg_flags = flags;
4430 ret = sock_sendmsg(sock, &msg);
4431 if (force_nonblock && ret == -EAGAIN)
4433 if (ret == -ERESTARTSYS)
4437 req_set_fail_links(req);
4438 __io_req_complete(req, ret, 0, cs);
4442 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4443 struct io_async_msghdr *iomsg)
4445 struct io_sr_msg *sr = &req->sr_msg;
4446 struct iovec __user *uiov;
4450 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4451 &iomsg->uaddr, &uiov, &iov_len);
4455 if (req->flags & REQ_F_BUFFER_SELECT) {
4458 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4460 sr->len = iomsg->iov[0].iov_len;
4461 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4465 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4466 &iomsg->iov, &iomsg->msg.msg_iter,
4475 #ifdef CONFIG_COMPAT
4476 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4477 struct io_async_msghdr *iomsg)
4479 struct compat_msghdr __user *msg_compat;
4480 struct io_sr_msg *sr = &req->sr_msg;
4481 struct compat_iovec __user *uiov;
4486 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4487 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4492 uiov = compat_ptr(ptr);
4493 if (req->flags & REQ_F_BUFFER_SELECT) {
4494 compat_ssize_t clen;
4498 if (!access_ok(uiov, sizeof(*uiov)))
4500 if (__get_user(clen, &uiov->iov_len))
4504 sr->len = iomsg->iov[0].iov_len;
4507 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4508 UIO_FASTIOV, &iomsg->iov,
4509 &iomsg->msg.msg_iter, true);
4518 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4519 struct io_async_msghdr *iomsg)
4521 iomsg->msg.msg_name = &iomsg->addr;
4522 iomsg->iov = iomsg->fast_iov;
4524 #ifdef CONFIG_COMPAT
4525 if (req->ctx->compat)
4526 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4529 return __io_recvmsg_copy_hdr(req, iomsg);
4532 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4535 struct io_sr_msg *sr = &req->sr_msg;
4536 struct io_buffer *kbuf;
4538 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4543 req->flags |= REQ_F_BUFFER_SELECTED;
4547 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4549 return io_put_kbuf(req, req->sr_msg.kbuf);
4552 static int io_recvmsg_prep(struct io_kiocb *req,
4553 const struct io_uring_sqe *sqe)
4555 struct io_async_msghdr *async_msg = req->async_data;
4556 struct io_sr_msg *sr = &req->sr_msg;
4559 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4562 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4563 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4564 sr->len = READ_ONCE(sqe->len);
4565 sr->bgid = READ_ONCE(sqe->buf_group);
4567 #ifdef CONFIG_COMPAT
4568 if (req->ctx->compat)
4569 sr->msg_flags |= MSG_CMSG_COMPAT;
4572 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4574 ret = io_recvmsg_copy_hdr(req, async_msg);
4576 req->flags |= REQ_F_NEED_CLEANUP;
4580 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4581 struct io_comp_state *cs)
4583 struct io_async_msghdr iomsg, *kmsg;
4584 struct socket *sock;
4585 struct io_buffer *kbuf;
4587 int ret, cflags = 0;
4589 sock = sock_from_file(req->file, &ret);
4590 if (unlikely(!sock))
4593 if (req->async_data) {
4594 kmsg = req->async_data;
4595 kmsg->msg.msg_name = &kmsg->addr;
4596 /* if iov is set, it's allocated already */
4598 kmsg->iov = kmsg->fast_iov;
4599 kmsg->msg.msg_iter.iov = kmsg->iov;
4601 ret = io_recvmsg_copy_hdr(req, &iomsg);
4607 if (req->flags & REQ_F_BUFFER_SELECT) {
4608 kbuf = io_recv_buffer_select(req, !force_nonblock);
4610 return PTR_ERR(kbuf);
4611 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4612 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4613 1, req->sr_msg.len);
4616 flags = req->sr_msg.msg_flags;
4617 if (flags & MSG_DONTWAIT)
4618 req->flags |= REQ_F_NOWAIT;
4619 else if (force_nonblock)
4620 flags |= MSG_DONTWAIT;
4622 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4623 kmsg->uaddr, flags);
4624 if (force_nonblock && ret == -EAGAIN)
4625 return io_setup_async_msg(req, kmsg);
4626 if (ret == -ERESTARTSYS)
4629 if (req->flags & REQ_F_BUFFER_SELECTED)
4630 cflags = io_put_recv_kbuf(req);
4631 if (kmsg->iov != kmsg->fast_iov)
4633 req->flags &= ~REQ_F_NEED_CLEANUP;
4635 req_set_fail_links(req);
4636 __io_req_complete(req, ret, cflags, cs);
4640 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4641 struct io_comp_state *cs)
4643 struct io_buffer *kbuf;
4644 struct io_sr_msg *sr = &req->sr_msg;
4646 void __user *buf = sr->buf;
4647 struct socket *sock;
4650 int ret, cflags = 0;
4652 sock = sock_from_file(req->file, &ret);
4653 if (unlikely(!sock))
4656 if (req->flags & REQ_F_BUFFER_SELECT) {
4657 kbuf = io_recv_buffer_select(req, !force_nonblock);
4659 return PTR_ERR(kbuf);
4660 buf = u64_to_user_ptr(kbuf->addr);
4663 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4667 msg.msg_name = NULL;
4668 msg.msg_control = NULL;
4669 msg.msg_controllen = 0;
4670 msg.msg_namelen = 0;
4671 msg.msg_iocb = NULL;
4674 flags = req->sr_msg.msg_flags;
4675 if (flags & MSG_DONTWAIT)
4676 req->flags |= REQ_F_NOWAIT;
4677 else if (force_nonblock)
4678 flags |= MSG_DONTWAIT;
4680 ret = sock_recvmsg(sock, &msg, flags);
4681 if (force_nonblock && ret == -EAGAIN)
4683 if (ret == -ERESTARTSYS)
4686 if (req->flags & REQ_F_BUFFER_SELECTED)
4687 cflags = io_put_recv_kbuf(req);
4689 req_set_fail_links(req);
4690 __io_req_complete(req, ret, cflags, cs);
4694 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4696 struct io_accept *accept = &req->accept;
4698 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4700 if (sqe->ioprio || sqe->len || sqe->buf_index)
4703 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4704 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4705 accept->flags = READ_ONCE(sqe->accept_flags);
4706 accept->nofile = rlimit(RLIMIT_NOFILE);
4710 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4711 struct io_comp_state *cs)
4713 struct io_accept *accept = &req->accept;
4714 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4717 if (req->file->f_flags & O_NONBLOCK)
4718 req->flags |= REQ_F_NOWAIT;
4720 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4721 accept->addr_len, accept->flags,
4723 if (ret == -EAGAIN && force_nonblock)
4726 if (ret == -ERESTARTSYS)
4728 req_set_fail_links(req);
4730 __io_req_complete(req, ret, 0, cs);
4734 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4736 struct io_connect *conn = &req->connect;
4737 struct io_async_connect *io = req->async_data;
4739 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4741 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4744 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4745 conn->addr_len = READ_ONCE(sqe->addr2);
4750 return move_addr_to_kernel(conn->addr, conn->addr_len,
4754 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4755 struct io_comp_state *cs)
4757 struct io_async_connect __io, *io;
4758 unsigned file_flags;
4761 if (req->async_data) {
4762 io = req->async_data;
4764 ret = move_addr_to_kernel(req->connect.addr,
4765 req->connect.addr_len,
4772 file_flags = force_nonblock ? O_NONBLOCK : 0;
4774 ret = __sys_connect_file(req->file, &io->address,
4775 req->connect.addr_len, file_flags);
4776 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4777 if (req->async_data)
4779 if (io_alloc_async_data(req)) {
4783 io = req->async_data;
4784 memcpy(req->async_data, &__io, sizeof(__io));
4787 if (ret == -ERESTARTSYS)
4791 req_set_fail_links(req);
4792 __io_req_complete(req, ret, 0, cs);
4795 #else /* !CONFIG_NET */
4796 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4801 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4802 struct io_comp_state *cs)
4807 static int io_send(struct io_kiocb *req, bool force_nonblock,
4808 struct io_comp_state *cs)
4813 static int io_recvmsg_prep(struct io_kiocb *req,
4814 const struct io_uring_sqe *sqe)
4819 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4820 struct io_comp_state *cs)
4825 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4826 struct io_comp_state *cs)
4831 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4836 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4837 struct io_comp_state *cs)
4842 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4847 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4848 struct io_comp_state *cs)
4852 #endif /* CONFIG_NET */
4854 struct io_poll_table {
4855 struct poll_table_struct pt;
4856 struct io_kiocb *req;
4860 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4861 __poll_t mask, task_work_func_t func)
4866 /* for instances that support it check for an event match first: */
4867 if (mask && !(mask & poll->events))
4870 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4872 list_del_init(&poll->wait.entry);
4875 init_task_work(&req->task_work, func);
4876 percpu_ref_get(&req->ctx->refs);
4879 * If we using the signalfd wait_queue_head for this wakeup, then
4880 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4881 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4882 * either, as the normal wakeup will suffice.
4884 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4887 * If this fails, then the task is exiting. When a task exits, the
4888 * work gets canceled, so just cancel this request as well instead
4889 * of executing it. We can't safely execute it anyway, as we may not
4890 * have the needed state needed for it anyway.
4892 ret = io_req_task_work_add(req, twa_signal_ok);
4893 if (unlikely(ret)) {
4894 struct task_struct *tsk;
4896 WRITE_ONCE(poll->canceled, true);
4897 tsk = io_wq_get_task(req->ctx->io_wq);
4898 task_work_add(tsk, &req->task_work, TWA_NONE);
4899 wake_up_process(tsk);
4904 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4905 __acquires(&req->ctx->completion_lock)
4907 struct io_ring_ctx *ctx = req->ctx;
4909 if (!req->result && !READ_ONCE(poll->canceled)) {
4910 struct poll_table_struct pt = { ._key = poll->events };
4912 req->result = vfs_poll(req->file, &pt) & poll->events;
4915 spin_lock_irq(&ctx->completion_lock);
4916 if (!req->result && !READ_ONCE(poll->canceled)) {
4917 add_wait_queue(poll->head, &poll->wait);
4924 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4926 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4927 if (req->opcode == IORING_OP_POLL_ADD)
4928 return req->async_data;
4929 return req->apoll->double_poll;
4932 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4934 if (req->opcode == IORING_OP_POLL_ADD)
4936 return &req->apoll->poll;
4939 static void io_poll_remove_double(struct io_kiocb *req)
4941 struct io_poll_iocb *poll = io_poll_get_double(req);
4943 lockdep_assert_held(&req->ctx->completion_lock);
4945 if (poll && poll->head) {
4946 struct wait_queue_head *head = poll->head;
4948 spin_lock(&head->lock);
4949 list_del_init(&poll->wait.entry);
4950 if (poll->wait.private)
4951 refcount_dec(&req->refs);
4953 spin_unlock(&head->lock);
4957 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4959 struct io_ring_ctx *ctx = req->ctx;
4961 io_poll_remove_double(req);
4962 req->poll.done = true;
4963 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4964 io_commit_cqring(ctx);
4967 static void io_poll_task_func(struct callback_head *cb)
4969 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4970 struct io_ring_ctx *ctx = req->ctx;
4971 struct io_kiocb *nxt;
4973 if (io_poll_rewait(req, &req->poll)) {
4974 spin_unlock_irq(&ctx->completion_lock);
4976 hash_del(&req->hash_node);
4977 io_poll_complete(req, req->result, 0);
4978 spin_unlock_irq(&ctx->completion_lock);
4980 nxt = io_put_req_find_next(req);
4981 io_cqring_ev_posted(ctx);
4983 __io_req_task_submit(nxt);
4986 percpu_ref_put(&ctx->refs);
4989 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4990 int sync, void *key)
4992 struct io_kiocb *req = wait->private;
4993 struct io_poll_iocb *poll = io_poll_get_single(req);
4994 __poll_t mask = key_to_poll(key);
4996 /* for instances that support it check for an event match first: */
4997 if (mask && !(mask & poll->events))
5000 list_del_init(&wait->entry);
5002 if (poll && poll->head) {
5005 spin_lock(&poll->head->lock);
5006 done = list_empty(&poll->wait.entry);
5008 list_del_init(&poll->wait.entry);
5009 /* make sure double remove sees this as being gone */
5010 wait->private = NULL;
5011 spin_unlock(&poll->head->lock);
5013 /* use wait func handler, so it matches the rq type */
5014 poll->wait.func(&poll->wait, mode, sync, key);
5017 refcount_dec(&req->refs);
5021 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5022 wait_queue_func_t wake_func)
5026 poll->canceled = false;
5027 poll->events = events;
5028 INIT_LIST_HEAD(&poll->wait.entry);
5029 init_waitqueue_func_entry(&poll->wait, wake_func);
5032 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5033 struct wait_queue_head *head,
5034 struct io_poll_iocb **poll_ptr)
5036 struct io_kiocb *req = pt->req;
5039 * If poll->head is already set, it's because the file being polled
5040 * uses multiple waitqueues for poll handling (eg one for read, one
5041 * for write). Setup a separate io_poll_iocb if this happens.
5043 if (unlikely(poll->head)) {
5044 struct io_poll_iocb *poll_one = poll;
5046 /* already have a 2nd entry, fail a third attempt */
5048 pt->error = -EINVAL;
5051 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5053 pt->error = -ENOMEM;
5056 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5057 refcount_inc(&req->refs);
5058 poll->wait.private = req;
5065 if (poll->events & EPOLLEXCLUSIVE)
5066 add_wait_queue_exclusive(head, &poll->wait);
5068 add_wait_queue(head, &poll->wait);
5071 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5072 struct poll_table_struct *p)
5074 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5075 struct async_poll *apoll = pt->req->apoll;
5077 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5080 static void io_async_task_func(struct callback_head *cb)
5082 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5083 struct async_poll *apoll = req->apoll;
5084 struct io_ring_ctx *ctx = req->ctx;
5086 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5088 if (io_poll_rewait(req, &apoll->poll)) {
5089 spin_unlock_irq(&ctx->completion_lock);
5090 percpu_ref_put(&ctx->refs);
5094 /* If req is still hashed, it cannot have been canceled. Don't check. */
5095 if (hash_hashed(&req->hash_node))
5096 hash_del(&req->hash_node);
5098 io_poll_remove_double(req);
5099 spin_unlock_irq(&ctx->completion_lock);
5101 if (!READ_ONCE(apoll->poll.canceled))
5102 __io_req_task_submit(req);
5104 __io_req_task_cancel(req, -ECANCELED);
5106 percpu_ref_put(&ctx->refs);
5107 kfree(apoll->double_poll);
5111 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5114 struct io_kiocb *req = wait->private;
5115 struct io_poll_iocb *poll = &req->apoll->poll;
5117 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5120 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5123 static void io_poll_req_insert(struct io_kiocb *req)
5125 struct io_ring_ctx *ctx = req->ctx;
5126 struct hlist_head *list;
5128 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5129 hlist_add_head(&req->hash_node, list);
5132 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5133 struct io_poll_iocb *poll,
5134 struct io_poll_table *ipt, __poll_t mask,
5135 wait_queue_func_t wake_func)
5136 __acquires(&ctx->completion_lock)
5138 struct io_ring_ctx *ctx = req->ctx;
5139 bool cancel = false;
5141 INIT_HLIST_NODE(&req->hash_node);
5142 io_init_poll_iocb(poll, mask, wake_func);
5143 poll->file = req->file;
5144 poll->wait.private = req;
5146 ipt->pt._key = mask;
5148 ipt->error = -EINVAL;
5150 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5152 spin_lock_irq(&ctx->completion_lock);
5153 if (likely(poll->head)) {
5154 spin_lock(&poll->head->lock);
5155 if (unlikely(list_empty(&poll->wait.entry))) {
5161 if (mask || ipt->error)
5162 list_del_init(&poll->wait.entry);
5164 WRITE_ONCE(poll->canceled, true);
5165 else if (!poll->done) /* actually waiting for an event */
5166 io_poll_req_insert(req);
5167 spin_unlock(&poll->head->lock);
5173 static bool io_arm_poll_handler(struct io_kiocb *req)
5175 const struct io_op_def *def = &io_op_defs[req->opcode];
5176 struct io_ring_ctx *ctx = req->ctx;
5177 struct async_poll *apoll;
5178 struct io_poll_table ipt;
5182 if (!req->file || !file_can_poll(req->file))
5184 if (req->flags & REQ_F_POLLED)
5188 else if (def->pollout)
5192 /* if we can't nonblock try, then no point in arming a poll handler */
5193 if (!io_file_supports_async(req->file, rw))
5196 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5197 if (unlikely(!apoll))
5199 apoll->double_poll = NULL;
5201 req->flags |= REQ_F_POLLED;
5206 mask |= POLLIN | POLLRDNORM;
5208 mask |= POLLOUT | POLLWRNORM;
5210 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5211 if ((req->opcode == IORING_OP_RECVMSG) &&
5212 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5215 mask |= POLLERR | POLLPRI;
5217 ipt.pt._qproc = io_async_queue_proc;
5219 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5221 if (ret || ipt.error) {
5222 io_poll_remove_double(req);
5223 spin_unlock_irq(&ctx->completion_lock);
5224 kfree(apoll->double_poll);
5228 spin_unlock_irq(&ctx->completion_lock);
5229 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5230 apoll->poll.events);
5234 static bool __io_poll_remove_one(struct io_kiocb *req,
5235 struct io_poll_iocb *poll)
5237 bool do_complete = false;
5239 spin_lock(&poll->head->lock);
5240 WRITE_ONCE(poll->canceled, true);
5241 if (!list_empty(&poll->wait.entry)) {
5242 list_del_init(&poll->wait.entry);
5245 spin_unlock(&poll->head->lock);
5246 hash_del(&req->hash_node);
5250 static bool io_poll_remove_one(struct io_kiocb *req)
5254 io_poll_remove_double(req);
5256 if (req->opcode == IORING_OP_POLL_ADD) {
5257 do_complete = __io_poll_remove_one(req, &req->poll);
5259 struct async_poll *apoll = req->apoll;
5261 /* non-poll requests have submit ref still */
5262 do_complete = __io_poll_remove_one(req, &apoll->poll);
5265 kfree(apoll->double_poll);
5271 io_cqring_fill_event(req, -ECANCELED);
5272 io_commit_cqring(req->ctx);
5273 req_set_fail_links(req);
5274 io_put_req_deferred(req, 1);
5281 * Returns true if we found and killed one or more poll requests
5283 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5285 struct hlist_node *tmp;
5286 struct io_kiocb *req;
5289 spin_lock_irq(&ctx->completion_lock);
5290 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5291 struct hlist_head *list;
5293 list = &ctx->cancel_hash[i];
5294 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5295 if (io_task_match(req, tsk))
5296 posted += io_poll_remove_one(req);
5299 spin_unlock_irq(&ctx->completion_lock);
5302 io_cqring_ev_posted(ctx);
5307 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5309 struct hlist_head *list;
5310 struct io_kiocb *req;
5312 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5313 hlist_for_each_entry(req, list, hash_node) {
5314 if (sqe_addr != req->user_data)
5316 if (io_poll_remove_one(req))
5324 static int io_poll_remove_prep(struct io_kiocb *req,
5325 const struct io_uring_sqe *sqe)
5327 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5329 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5333 req->poll.addr = READ_ONCE(sqe->addr);
5338 * Find a running poll command that matches one specified in sqe->addr,
5339 * and remove it if found.
5341 static int io_poll_remove(struct io_kiocb *req)
5343 struct io_ring_ctx *ctx = req->ctx;
5347 addr = req->poll.addr;
5348 spin_lock_irq(&ctx->completion_lock);
5349 ret = io_poll_cancel(ctx, addr);
5350 spin_unlock_irq(&ctx->completion_lock);
5353 req_set_fail_links(req);
5354 io_req_complete(req, ret);
5358 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5361 struct io_kiocb *req = wait->private;
5362 struct io_poll_iocb *poll = &req->poll;
5364 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5367 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5368 struct poll_table_struct *p)
5370 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5372 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5375 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5377 struct io_poll_iocb *poll = &req->poll;
5380 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5382 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5385 events = READ_ONCE(sqe->poll32_events);
5387 events = swahw32(events);
5389 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5390 (events & EPOLLEXCLUSIVE);
5394 static int io_poll_add(struct io_kiocb *req)
5396 struct io_poll_iocb *poll = &req->poll;
5397 struct io_ring_ctx *ctx = req->ctx;
5398 struct io_poll_table ipt;
5401 ipt.pt._qproc = io_poll_queue_proc;
5403 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5406 if (mask) { /* no async, we'd stolen it */
5408 io_poll_complete(req, mask, 0);
5410 spin_unlock_irq(&ctx->completion_lock);
5413 io_cqring_ev_posted(ctx);
5419 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5421 struct io_timeout_data *data = container_of(timer,
5422 struct io_timeout_data, timer);
5423 struct io_kiocb *req = data->req;
5424 struct io_ring_ctx *ctx = req->ctx;
5425 unsigned long flags;
5427 spin_lock_irqsave(&ctx->completion_lock, flags);
5428 list_del_init(&req->timeout.list);
5429 atomic_set(&req->ctx->cq_timeouts,
5430 atomic_read(&req->ctx->cq_timeouts) + 1);
5432 io_cqring_fill_event(req, -ETIME);
5433 io_commit_cqring(ctx);
5434 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5436 io_cqring_ev_posted(ctx);
5437 req_set_fail_links(req);
5439 return HRTIMER_NORESTART;
5442 static int __io_timeout_cancel(struct io_kiocb *req)
5444 struct io_timeout_data *io = req->async_data;
5447 ret = hrtimer_try_to_cancel(&io->timer);
5450 list_del_init(&req->timeout.list);
5452 req_set_fail_links(req);
5453 io_cqring_fill_event(req, -ECANCELED);
5454 io_put_req_deferred(req, 1);
5458 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5460 struct io_kiocb *req;
5463 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5464 if (user_data == req->user_data) {
5473 return __io_timeout_cancel(req);
5476 static int io_timeout_remove_prep(struct io_kiocb *req,
5477 const struct io_uring_sqe *sqe)
5479 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5481 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5483 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5486 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5491 * Remove or update an existing timeout command
5493 static int io_timeout_remove(struct io_kiocb *req)
5495 struct io_ring_ctx *ctx = req->ctx;
5498 spin_lock_irq(&ctx->completion_lock);
5499 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5501 io_cqring_fill_event(req, ret);
5502 io_commit_cqring(ctx);
5503 spin_unlock_irq(&ctx->completion_lock);
5504 io_cqring_ev_posted(ctx);
5506 req_set_fail_links(req);
5511 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5512 bool is_timeout_link)
5514 struct io_timeout_data *data;
5516 u32 off = READ_ONCE(sqe->off);
5518 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5520 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5522 if (off && is_timeout_link)
5524 flags = READ_ONCE(sqe->timeout_flags);
5525 if (flags & ~IORING_TIMEOUT_ABS)
5528 req->timeout.off = off;
5530 if (!req->async_data && io_alloc_async_data(req))
5533 data = req->async_data;
5536 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5539 if (flags & IORING_TIMEOUT_ABS)
5540 data->mode = HRTIMER_MODE_ABS;
5542 data->mode = HRTIMER_MODE_REL;
5544 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5548 static int io_timeout(struct io_kiocb *req)
5550 struct io_ring_ctx *ctx = req->ctx;
5551 struct io_timeout_data *data = req->async_data;
5552 struct list_head *entry;
5553 u32 tail, off = req->timeout.off;
5555 spin_lock_irq(&ctx->completion_lock);
5558 * sqe->off holds how many events that need to occur for this
5559 * timeout event to be satisfied. If it isn't set, then this is
5560 * a pure timeout request, sequence isn't used.
5562 if (io_is_timeout_noseq(req)) {
5563 entry = ctx->timeout_list.prev;
5567 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5568 req->timeout.target_seq = tail + off;
5571 * Insertion sort, ensuring the first entry in the list is always
5572 * the one we need first.
5574 list_for_each_prev(entry, &ctx->timeout_list) {
5575 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5578 if (io_is_timeout_noseq(nxt))
5580 /* nxt.seq is behind @tail, otherwise would've been completed */
5581 if (off >= nxt->timeout.target_seq - tail)
5585 list_add(&req->timeout.list, entry);
5586 data->timer.function = io_timeout_fn;
5587 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5588 spin_unlock_irq(&ctx->completion_lock);
5592 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5594 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5596 return req->user_data == (unsigned long) data;
5599 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5601 enum io_wq_cancel cancel_ret;
5604 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5605 switch (cancel_ret) {
5606 case IO_WQ_CANCEL_OK:
5609 case IO_WQ_CANCEL_RUNNING:
5612 case IO_WQ_CANCEL_NOTFOUND:
5620 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5621 struct io_kiocb *req, __u64 sqe_addr,
5624 unsigned long flags;
5627 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5628 if (ret != -ENOENT) {
5629 spin_lock_irqsave(&ctx->completion_lock, flags);
5633 spin_lock_irqsave(&ctx->completion_lock, flags);
5634 ret = io_timeout_cancel(ctx, sqe_addr);
5637 ret = io_poll_cancel(ctx, sqe_addr);
5641 io_cqring_fill_event(req, ret);
5642 io_commit_cqring(ctx);
5643 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5644 io_cqring_ev_posted(ctx);
5647 req_set_fail_links(req);
5651 static int io_async_cancel_prep(struct io_kiocb *req,
5652 const struct io_uring_sqe *sqe)
5654 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5656 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5658 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5661 req->cancel.addr = READ_ONCE(sqe->addr);
5665 static int io_async_cancel(struct io_kiocb *req)
5667 struct io_ring_ctx *ctx = req->ctx;
5669 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5673 static int io_files_update_prep(struct io_kiocb *req,
5674 const struct io_uring_sqe *sqe)
5676 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5678 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5680 if (sqe->ioprio || sqe->rw_flags)
5683 req->files_update.offset = READ_ONCE(sqe->off);
5684 req->files_update.nr_args = READ_ONCE(sqe->len);
5685 if (!req->files_update.nr_args)
5687 req->files_update.arg = READ_ONCE(sqe->addr);
5691 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5692 struct io_comp_state *cs)
5694 struct io_ring_ctx *ctx = req->ctx;
5695 struct io_uring_files_update up;
5701 up.offset = req->files_update.offset;
5702 up.fds = req->files_update.arg;
5704 mutex_lock(&ctx->uring_lock);
5705 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5706 mutex_unlock(&ctx->uring_lock);
5709 req_set_fail_links(req);
5710 __io_req_complete(req, ret, 0, cs);
5714 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5716 switch (req->opcode) {
5719 case IORING_OP_READV:
5720 case IORING_OP_READ_FIXED:
5721 case IORING_OP_READ:
5722 return io_read_prep(req, sqe);
5723 case IORING_OP_WRITEV:
5724 case IORING_OP_WRITE_FIXED:
5725 case IORING_OP_WRITE:
5726 return io_write_prep(req, sqe);
5727 case IORING_OP_POLL_ADD:
5728 return io_poll_add_prep(req, sqe);
5729 case IORING_OP_POLL_REMOVE:
5730 return io_poll_remove_prep(req, sqe);
5731 case IORING_OP_FSYNC:
5732 return io_prep_fsync(req, sqe);
5733 case IORING_OP_SYNC_FILE_RANGE:
5734 return io_prep_sfr(req, sqe);
5735 case IORING_OP_SENDMSG:
5736 case IORING_OP_SEND:
5737 return io_sendmsg_prep(req, sqe);
5738 case IORING_OP_RECVMSG:
5739 case IORING_OP_RECV:
5740 return io_recvmsg_prep(req, sqe);
5741 case IORING_OP_CONNECT:
5742 return io_connect_prep(req, sqe);
5743 case IORING_OP_TIMEOUT:
5744 return io_timeout_prep(req, sqe, false);
5745 case IORING_OP_TIMEOUT_REMOVE:
5746 return io_timeout_remove_prep(req, sqe);
5747 case IORING_OP_ASYNC_CANCEL:
5748 return io_async_cancel_prep(req, sqe);
5749 case IORING_OP_LINK_TIMEOUT:
5750 return io_timeout_prep(req, sqe, true);
5751 case IORING_OP_ACCEPT:
5752 return io_accept_prep(req, sqe);
5753 case IORING_OP_FALLOCATE:
5754 return io_fallocate_prep(req, sqe);
5755 case IORING_OP_OPENAT:
5756 return io_openat_prep(req, sqe);
5757 case IORING_OP_CLOSE:
5758 return io_close_prep(req, sqe);
5759 case IORING_OP_FILES_UPDATE:
5760 return io_files_update_prep(req, sqe);
5761 case IORING_OP_STATX:
5762 return io_statx_prep(req, sqe);
5763 case IORING_OP_FADVISE:
5764 return io_fadvise_prep(req, sqe);
5765 case IORING_OP_MADVISE:
5766 return io_madvise_prep(req, sqe);
5767 case IORING_OP_OPENAT2:
5768 return io_openat2_prep(req, sqe);
5769 case IORING_OP_EPOLL_CTL:
5770 return io_epoll_ctl_prep(req, sqe);
5771 case IORING_OP_SPLICE:
5772 return io_splice_prep(req, sqe);
5773 case IORING_OP_PROVIDE_BUFFERS:
5774 return io_provide_buffers_prep(req, sqe);
5775 case IORING_OP_REMOVE_BUFFERS:
5776 return io_remove_buffers_prep(req, sqe);
5778 return io_tee_prep(req, sqe);
5781 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5786 static int io_req_defer_prep(struct io_kiocb *req,
5787 const struct io_uring_sqe *sqe)
5791 if (io_alloc_async_data(req))
5793 return io_req_prep(req, sqe);
5796 static u32 io_get_sequence(struct io_kiocb *req)
5798 struct io_kiocb *pos;
5799 struct io_ring_ctx *ctx = req->ctx;
5800 u32 total_submitted, nr_reqs = 1;
5802 if (req->flags & REQ_F_LINK_HEAD)
5803 list_for_each_entry(pos, &req->link_list, link_list)
5806 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5807 return total_submitted - nr_reqs;
5810 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5812 struct io_ring_ctx *ctx = req->ctx;
5813 struct io_defer_entry *de;
5817 /* Still need defer if there is pending req in defer list. */
5818 if (likely(list_empty_careful(&ctx->defer_list) &&
5819 !(req->flags & REQ_F_IO_DRAIN)))
5822 seq = io_get_sequence(req);
5823 /* Still a chance to pass the sequence check */
5824 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5827 if (!req->async_data) {
5828 ret = io_req_defer_prep(req, sqe);
5832 io_prep_async_link(req);
5833 de = kmalloc(sizeof(*de), GFP_KERNEL);
5837 spin_lock_irq(&ctx->completion_lock);
5838 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5839 spin_unlock_irq(&ctx->completion_lock);
5841 io_queue_async_work(req);
5842 return -EIOCBQUEUED;
5845 trace_io_uring_defer(ctx, req, req->user_data);
5848 list_add_tail(&de->list, &ctx->defer_list);
5849 spin_unlock_irq(&ctx->completion_lock);
5850 return -EIOCBQUEUED;
5853 static void io_req_drop_files(struct io_kiocb *req)
5855 struct io_ring_ctx *ctx = req->ctx;
5856 unsigned long flags;
5858 spin_lock_irqsave(&ctx->inflight_lock, flags);
5859 list_del(&req->inflight_entry);
5860 if (waitqueue_active(&ctx->inflight_wait))
5861 wake_up(&ctx->inflight_wait);
5862 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5863 req->flags &= ~REQ_F_INFLIGHT;
5864 put_files_struct(req->work.identity->files);
5865 put_nsproxy(req->work.identity->nsproxy);
5866 req->work.flags &= ~IO_WQ_WORK_FILES;
5869 static void __io_clean_op(struct io_kiocb *req)
5871 if (req->flags & REQ_F_BUFFER_SELECTED) {
5872 switch (req->opcode) {
5873 case IORING_OP_READV:
5874 case IORING_OP_READ_FIXED:
5875 case IORING_OP_READ:
5876 kfree((void *)(unsigned long)req->rw.addr);
5878 case IORING_OP_RECVMSG:
5879 case IORING_OP_RECV:
5880 kfree(req->sr_msg.kbuf);
5883 req->flags &= ~REQ_F_BUFFER_SELECTED;
5886 if (req->flags & REQ_F_NEED_CLEANUP) {
5887 switch (req->opcode) {
5888 case IORING_OP_READV:
5889 case IORING_OP_READ_FIXED:
5890 case IORING_OP_READ:
5891 case IORING_OP_WRITEV:
5892 case IORING_OP_WRITE_FIXED:
5893 case IORING_OP_WRITE: {
5894 struct io_async_rw *io = req->async_data;
5896 kfree(io->free_iovec);
5899 case IORING_OP_RECVMSG:
5900 case IORING_OP_SENDMSG: {
5901 struct io_async_msghdr *io = req->async_data;
5902 if (io->iov != io->fast_iov)
5906 case IORING_OP_SPLICE:
5908 io_put_file(req, req->splice.file_in,
5909 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5911 case IORING_OP_OPENAT:
5912 case IORING_OP_OPENAT2:
5913 if (req->open.filename)
5914 putname(req->open.filename);
5917 req->flags &= ~REQ_F_NEED_CLEANUP;
5920 if (req->flags & REQ_F_INFLIGHT)
5921 io_req_drop_files(req);
5924 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5925 struct io_comp_state *cs)
5927 struct io_ring_ctx *ctx = req->ctx;
5930 switch (req->opcode) {
5932 ret = io_nop(req, cs);
5934 case IORING_OP_READV:
5935 case IORING_OP_READ_FIXED:
5936 case IORING_OP_READ:
5937 ret = io_read(req, force_nonblock, cs);
5939 case IORING_OP_WRITEV:
5940 case IORING_OP_WRITE_FIXED:
5941 case IORING_OP_WRITE:
5942 ret = io_write(req, force_nonblock, cs);
5944 case IORING_OP_FSYNC:
5945 ret = io_fsync(req, force_nonblock);
5947 case IORING_OP_POLL_ADD:
5948 ret = io_poll_add(req);
5950 case IORING_OP_POLL_REMOVE:
5951 ret = io_poll_remove(req);
5953 case IORING_OP_SYNC_FILE_RANGE:
5954 ret = io_sync_file_range(req, force_nonblock);
5956 case IORING_OP_SENDMSG:
5957 ret = io_sendmsg(req, force_nonblock, cs);
5959 case IORING_OP_SEND:
5960 ret = io_send(req, force_nonblock, cs);
5962 case IORING_OP_RECVMSG:
5963 ret = io_recvmsg(req, force_nonblock, cs);
5965 case IORING_OP_RECV:
5966 ret = io_recv(req, force_nonblock, cs);
5968 case IORING_OP_TIMEOUT:
5969 ret = io_timeout(req);
5971 case IORING_OP_TIMEOUT_REMOVE:
5972 ret = io_timeout_remove(req);
5974 case IORING_OP_ACCEPT:
5975 ret = io_accept(req, force_nonblock, cs);
5977 case IORING_OP_CONNECT:
5978 ret = io_connect(req, force_nonblock, cs);
5980 case IORING_OP_ASYNC_CANCEL:
5981 ret = io_async_cancel(req);
5983 case IORING_OP_FALLOCATE:
5984 ret = io_fallocate(req, force_nonblock);
5986 case IORING_OP_OPENAT:
5987 ret = io_openat(req, force_nonblock);
5989 case IORING_OP_CLOSE:
5990 ret = io_close(req, force_nonblock, cs);
5992 case IORING_OP_FILES_UPDATE:
5993 ret = io_files_update(req, force_nonblock, cs);
5995 case IORING_OP_STATX:
5996 ret = io_statx(req, force_nonblock);
5998 case IORING_OP_FADVISE:
5999 ret = io_fadvise(req, force_nonblock);
6001 case IORING_OP_MADVISE:
6002 ret = io_madvise(req, force_nonblock);
6004 case IORING_OP_OPENAT2:
6005 ret = io_openat2(req, force_nonblock);
6007 case IORING_OP_EPOLL_CTL:
6008 ret = io_epoll_ctl(req, force_nonblock, cs);
6010 case IORING_OP_SPLICE:
6011 ret = io_splice(req, force_nonblock);
6013 case IORING_OP_PROVIDE_BUFFERS:
6014 ret = io_provide_buffers(req, force_nonblock, cs);
6016 case IORING_OP_REMOVE_BUFFERS:
6017 ret = io_remove_buffers(req, force_nonblock, cs);
6020 ret = io_tee(req, force_nonblock);
6030 /* If the op doesn't have a file, we're not polling for it */
6031 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6032 const bool in_async = io_wq_current_is_worker();
6034 /* workqueue context doesn't hold uring_lock, grab it now */
6036 mutex_lock(&ctx->uring_lock);
6038 io_iopoll_req_issued(req);
6041 mutex_unlock(&ctx->uring_lock);
6047 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6049 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6050 struct io_kiocb *timeout;
6053 timeout = io_prep_linked_timeout(req);
6055 io_queue_linked_timeout(timeout);
6057 /* if NO_CANCEL is set, we must still run the work */
6058 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6059 IO_WQ_WORK_CANCEL) {
6065 ret = io_issue_sqe(req, false, NULL);
6067 * We can get EAGAIN for polled IO even though we're
6068 * forcing a sync submission from here, since we can't
6069 * wait for request slots on the block side.
6078 req_set_fail_links(req);
6079 io_req_complete(req, ret);
6082 return io_steal_work(req);
6085 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6088 struct fixed_file_table *table;
6090 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6091 return table->files[index & IORING_FILE_TABLE_MASK];
6094 static struct file *io_file_get(struct io_submit_state *state,
6095 struct io_kiocb *req, int fd, bool fixed)
6097 struct io_ring_ctx *ctx = req->ctx;
6101 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6103 fd = array_index_nospec(fd, ctx->nr_user_files);
6104 file = io_file_from_index(ctx, fd);
6106 req->fixed_file_refs = &ctx->file_data->node->refs;
6107 percpu_ref_get(req->fixed_file_refs);
6110 trace_io_uring_file_get(ctx, fd);
6111 file = __io_file_get(state, fd);
6117 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6122 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6123 if (unlikely(!fixed && io_async_submit(req->ctx)))
6126 req->file = io_file_get(state, req, fd, fixed);
6127 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6132 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6134 struct io_timeout_data *data = container_of(timer,
6135 struct io_timeout_data, timer);
6136 struct io_kiocb *req = data->req;
6137 struct io_ring_ctx *ctx = req->ctx;
6138 struct io_kiocb *prev = NULL;
6139 unsigned long flags;
6141 spin_lock_irqsave(&ctx->completion_lock, flags);
6144 * We don't expect the list to be empty, that will only happen if we
6145 * race with the completion of the linked work.
6147 if (!list_empty(&req->link_list)) {
6148 prev = list_entry(req->link_list.prev, struct io_kiocb,
6150 if (refcount_inc_not_zero(&prev->refs))
6151 list_del_init(&req->link_list);
6156 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6159 req_set_fail_links(prev);
6160 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6163 io_req_complete(req, -ETIME);
6165 return HRTIMER_NORESTART;
6168 static void __io_queue_linked_timeout(struct io_kiocb *req)
6171 * If the list is now empty, then our linked request finished before
6172 * we got a chance to setup the timer
6174 if (!list_empty(&req->link_list)) {
6175 struct io_timeout_data *data = req->async_data;
6177 data->timer.function = io_link_timeout_fn;
6178 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6183 static void io_queue_linked_timeout(struct io_kiocb *req)
6185 struct io_ring_ctx *ctx = req->ctx;
6187 spin_lock_irq(&ctx->completion_lock);
6188 __io_queue_linked_timeout(req);
6189 spin_unlock_irq(&ctx->completion_lock);
6191 /* drop submission reference */
6195 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6197 struct io_kiocb *nxt;
6199 if (!(req->flags & REQ_F_LINK_HEAD))
6201 if (req->flags & REQ_F_LINK_TIMEOUT)
6204 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6206 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6209 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6210 req->flags |= REQ_F_LINK_TIMEOUT;
6214 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6216 struct io_kiocb *linked_timeout;
6217 const struct cred *old_creds = NULL;
6221 linked_timeout = io_prep_linked_timeout(req);
6223 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6224 (req->work.flags & IO_WQ_WORK_CREDS) &&
6225 req->work.identity->creds != current_cred()) {
6227 revert_creds(old_creds);
6228 if (old_creds == req->work.identity->creds)
6229 old_creds = NULL; /* restored original creds */
6231 old_creds = override_creds(req->work.identity->creds);
6234 ret = io_issue_sqe(req, true, cs);
6237 * We async punt it if the file wasn't marked NOWAIT, or if the file
6238 * doesn't support non-blocking read/write attempts
6240 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6241 if (!io_arm_poll_handler(req)) {
6243 * Queued up for async execution, worker will release
6244 * submit reference when the iocb is actually submitted.
6246 io_queue_async_work(req);
6250 io_queue_linked_timeout(linked_timeout);
6251 } else if (likely(!ret)) {
6252 /* drop submission reference */
6253 req = io_put_req_find_next(req);
6255 io_queue_linked_timeout(linked_timeout);
6258 if (!(req->flags & REQ_F_FORCE_ASYNC))
6260 io_queue_async_work(req);
6263 /* un-prep timeout, so it'll be killed as any other linked */
6264 req->flags &= ~REQ_F_LINK_TIMEOUT;
6265 req_set_fail_links(req);
6267 io_req_complete(req, ret);
6271 revert_creds(old_creds);
6274 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6275 struct io_comp_state *cs)
6279 ret = io_req_defer(req, sqe);
6281 if (ret != -EIOCBQUEUED) {
6283 req_set_fail_links(req);
6285 io_req_complete(req, ret);
6287 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6288 if (!req->async_data) {
6289 ret = io_req_defer_prep(req, sqe);
6293 io_queue_async_work(req);
6296 ret = io_req_prep(req, sqe);
6300 __io_queue_sqe(req, cs);
6304 static inline void io_queue_link_head(struct io_kiocb *req,
6305 struct io_comp_state *cs)
6307 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6309 io_req_complete(req, -ECANCELED);
6311 io_queue_sqe(req, NULL, cs);
6314 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6315 struct io_kiocb **link, struct io_comp_state *cs)
6317 struct io_ring_ctx *ctx = req->ctx;
6321 * If we already have a head request, queue this one for async
6322 * submittal once the head completes. If we don't have a head but
6323 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6324 * submitted sync once the chain is complete. If none of those
6325 * conditions are true (normal request), then just queue it.
6328 struct io_kiocb *head = *link;
6331 * Taking sequential execution of a link, draining both sides
6332 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6333 * requests in the link. So, it drains the head and the
6334 * next after the link request. The last one is done via
6335 * drain_next flag to persist the effect across calls.
6337 if (req->flags & REQ_F_IO_DRAIN) {
6338 head->flags |= REQ_F_IO_DRAIN;
6339 ctx->drain_next = 1;
6341 ret = io_req_defer_prep(req, sqe);
6342 if (unlikely(ret)) {
6343 /* fail even hard links since we don't submit */
6344 head->flags |= REQ_F_FAIL_LINK;
6347 trace_io_uring_link(ctx, req, head);
6348 list_add_tail(&req->link_list, &head->link_list);
6350 /* last request of a link, enqueue the link */
6351 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6352 io_queue_link_head(head, cs);
6356 if (unlikely(ctx->drain_next)) {
6357 req->flags |= REQ_F_IO_DRAIN;
6358 ctx->drain_next = 0;
6360 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6361 req->flags |= REQ_F_LINK_HEAD;
6362 INIT_LIST_HEAD(&req->link_list);
6364 ret = io_req_defer_prep(req, sqe);
6366 req->flags |= REQ_F_FAIL_LINK;
6369 io_queue_sqe(req, sqe, cs);
6377 * Batched submission is done, ensure local IO is flushed out.
6379 static void io_submit_state_end(struct io_submit_state *state)
6381 if (!list_empty(&state->comp.list))
6382 io_submit_flush_completions(&state->comp);
6383 blk_finish_plug(&state->plug);
6384 io_state_file_put(state);
6385 if (state->free_reqs)
6386 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6390 * Start submission side cache.
6392 static void io_submit_state_start(struct io_submit_state *state,
6393 struct io_ring_ctx *ctx, unsigned int max_ios)
6395 blk_start_plug(&state->plug);
6397 INIT_LIST_HEAD(&state->comp.list);
6398 state->comp.ctx = ctx;
6399 state->free_reqs = 0;
6401 state->ios_left = max_ios;
6404 static void io_commit_sqring(struct io_ring_ctx *ctx)
6406 struct io_rings *rings = ctx->rings;
6409 * Ensure any loads from the SQEs are done at this point,
6410 * since once we write the new head, the application could
6411 * write new data to them.
6413 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6417 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6418 * that is mapped by userspace. This means that care needs to be taken to
6419 * ensure that reads are stable, as we cannot rely on userspace always
6420 * being a good citizen. If members of the sqe are validated and then later
6421 * used, it's important that those reads are done through READ_ONCE() to
6422 * prevent a re-load down the line.
6424 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6426 u32 *sq_array = ctx->sq_array;
6430 * The cached sq head (or cq tail) serves two purposes:
6432 * 1) allows us to batch the cost of updating the user visible
6434 * 2) allows the kernel side to track the head on its own, even
6435 * though the application is the one updating it.
6437 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6438 if (likely(head < ctx->sq_entries))
6439 return &ctx->sq_sqes[head];
6441 /* drop invalid entries */
6442 ctx->cached_sq_dropped++;
6443 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6447 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6449 ctx->cached_sq_head++;
6453 * Check SQE restrictions (opcode and flags).
6455 * Returns 'true' if SQE is allowed, 'false' otherwise.
6457 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6458 struct io_kiocb *req,
6459 unsigned int sqe_flags)
6461 if (!ctx->restricted)
6464 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6467 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6468 ctx->restrictions.sqe_flags_required)
6471 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6472 ctx->restrictions.sqe_flags_required))
6478 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6479 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6480 IOSQE_BUFFER_SELECT)
6482 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6483 const struct io_uring_sqe *sqe,
6484 struct io_submit_state *state)
6486 unsigned int sqe_flags;
6489 req->opcode = READ_ONCE(sqe->opcode);
6490 req->user_data = READ_ONCE(sqe->user_data);
6491 req->async_data = NULL;
6495 /* one is dropped after submission, the other at completion */
6496 refcount_set(&req->refs, 2);
6497 req->task = current;
6500 if (unlikely(req->opcode >= IORING_OP_LAST))
6503 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6506 sqe_flags = READ_ONCE(sqe->flags);
6507 /* enforce forwards compatibility on users */
6508 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6511 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6514 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6515 !io_op_defs[req->opcode].buffer_select)
6518 id = READ_ONCE(sqe->personality);
6520 struct io_identity *iod;
6522 iod = idr_find(&ctx->personality_idr, id);
6525 refcount_inc(&iod->count);
6527 __io_req_init_async(req);
6528 get_cred(iod->creds);
6529 req->work.identity = iod;
6530 req->work.flags |= IO_WQ_WORK_CREDS;
6533 /* same numerical values with corresponding REQ_F_*, safe to copy */
6534 req->flags |= sqe_flags;
6536 if (!io_op_defs[req->opcode].needs_file)
6539 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6544 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6546 struct io_submit_state state;
6547 struct io_kiocb *link = NULL;
6548 int i, submitted = 0;
6550 /* if we have a backlog and couldn't flush it all, return BUSY */
6551 if (test_bit(0, &ctx->sq_check_overflow)) {
6552 if (!list_empty(&ctx->cq_overflow_list) &&
6553 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6557 /* make sure SQ entry isn't read before tail */
6558 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6560 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6563 percpu_counter_add(¤t->io_uring->inflight, nr);
6564 refcount_add(nr, ¤t->usage);
6566 io_submit_state_start(&state, ctx, nr);
6568 for (i = 0; i < nr; i++) {
6569 const struct io_uring_sqe *sqe;
6570 struct io_kiocb *req;
6573 sqe = io_get_sqe(ctx);
6574 if (unlikely(!sqe)) {
6575 io_consume_sqe(ctx);
6578 req = io_alloc_req(ctx, &state);
6579 if (unlikely(!req)) {
6581 submitted = -EAGAIN;
6584 io_consume_sqe(ctx);
6585 /* will complete beyond this point, count as submitted */
6588 err = io_init_req(ctx, req, sqe, &state);
6589 if (unlikely(err)) {
6592 io_req_complete(req, err);
6596 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6597 true, io_async_submit(ctx));
6598 err = io_submit_sqe(req, sqe, &link, &state.comp);
6603 if (unlikely(submitted != nr)) {
6604 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6605 struct io_uring_task *tctx = current->io_uring;
6606 int unused = nr - ref_used;
6608 percpu_ref_put_many(&ctx->refs, unused);
6609 percpu_counter_sub(&tctx->inflight, unused);
6610 put_task_struct_many(current, unused);
6613 io_queue_link_head(link, &state.comp);
6614 io_submit_state_end(&state);
6616 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6617 io_commit_sqring(ctx);
6622 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6624 /* Tell userspace we may need a wakeup call */
6625 spin_lock_irq(&ctx->completion_lock);
6626 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6627 spin_unlock_irq(&ctx->completion_lock);
6630 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6632 spin_lock_irq(&ctx->completion_lock);
6633 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6634 spin_unlock_irq(&ctx->completion_lock);
6637 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6638 int sync, void *key)
6640 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6643 ret = autoremove_wake_function(wqe, mode, sync, key);
6645 unsigned long flags;
6647 spin_lock_irqsave(&ctx->completion_lock, flags);
6648 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6649 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6660 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6661 unsigned long start_jiffies, bool cap_entries)
6663 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6664 struct io_sq_data *sqd = ctx->sq_data;
6665 unsigned int to_submit;
6669 if (!list_empty(&ctx->iopoll_list)) {
6670 unsigned nr_events = 0;
6672 mutex_lock(&ctx->uring_lock);
6673 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6674 io_do_iopoll(ctx, &nr_events, 0);
6675 mutex_unlock(&ctx->uring_lock);
6678 to_submit = io_sqring_entries(ctx);
6681 * If submit got -EBUSY, flag us as needing the application
6682 * to enter the kernel to reap and flush events.
6684 if (!to_submit || ret == -EBUSY || need_resched()) {
6686 * Drop cur_mm before scheduling, we can't hold it for
6687 * long periods (or over schedule()). Do this before
6688 * adding ourselves to the waitqueue, as the unuse/drop
6691 io_sq_thread_drop_mm();
6694 * We're polling. If we're within the defined idle
6695 * period, then let us spin without work before going
6696 * to sleep. The exception is if we got EBUSY doing
6697 * more IO, we should wait for the application to
6698 * reap events and wake us up.
6700 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6701 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6702 !percpu_ref_is_dying(&ctx->refs)))
6705 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6706 TASK_INTERRUPTIBLE);
6709 * While doing polled IO, before going to sleep, we need
6710 * to check if there are new reqs added to iopoll_list,
6711 * it is because reqs may have been punted to io worker
6712 * and will be added to iopoll_list later, hence check
6713 * the iopoll_list again.
6715 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6716 !list_empty_careful(&ctx->iopoll_list)) {
6717 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6721 to_submit = io_sqring_entries(ctx);
6722 if (!to_submit || ret == -EBUSY)
6726 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6727 io_ring_clear_wakeup_flag(ctx);
6729 /* if we're handling multiple rings, cap submit size for fairness */
6730 if (cap_entries && to_submit > 8)
6733 mutex_lock(&ctx->uring_lock);
6734 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6735 ret = io_submit_sqes(ctx, to_submit);
6736 mutex_unlock(&ctx->uring_lock);
6738 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6739 wake_up(&ctx->sqo_sq_wait);
6741 return SQT_DID_WORK;
6744 static void io_sqd_init_new(struct io_sq_data *sqd)
6746 struct io_ring_ctx *ctx;
6748 while (!list_empty(&sqd->ctx_new_list)) {
6749 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6750 init_wait(&ctx->sqo_wait_entry);
6751 ctx->sqo_wait_entry.func = io_sq_wake_function;
6752 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6753 complete(&ctx->sq_thread_comp);
6757 static int io_sq_thread(void *data)
6759 struct cgroup_subsys_state *cur_css = NULL;
6760 const struct cred *old_cred = NULL;
6761 struct io_sq_data *sqd = data;
6762 struct io_ring_ctx *ctx;
6763 unsigned long start_jiffies;
6765 start_jiffies = jiffies;
6766 while (!kthread_should_stop()) {
6767 enum sq_ret ret = 0;
6771 * Any changes to the sqd lists are synchronized through the
6772 * kthread parking. This synchronizes the thread vs users,
6773 * the users are synchronized on the sqd->ctx_lock.
6775 if (kthread_should_park())
6778 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6779 io_sqd_init_new(sqd);
6781 cap_entries = !list_is_singular(&sqd->ctx_list);
6783 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6784 if (current->cred != ctx->creds) {
6786 revert_creds(old_cred);
6787 old_cred = override_creds(ctx->creds);
6789 io_sq_thread_associate_blkcg(ctx, &cur_css);
6791 current->loginuid = ctx->loginuid;
6792 current->sessionid = ctx->sessionid;
6795 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6797 io_sq_thread_drop_mm();
6800 if (ret & SQT_SPIN) {
6803 } else if (ret == SQT_IDLE) {
6804 if (kthread_should_park())
6806 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6807 io_ring_set_wakeup_flag(ctx);
6809 start_jiffies = jiffies;
6810 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6811 io_ring_clear_wakeup_flag(ctx);
6818 io_sq_thread_unassociate_blkcg();
6820 revert_creds(old_cred);
6827 struct io_wait_queue {
6828 struct wait_queue_entry wq;
6829 struct io_ring_ctx *ctx;
6831 unsigned nr_timeouts;
6834 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6836 struct io_ring_ctx *ctx = iowq->ctx;
6839 * Wake up if we have enough events, or if a timeout occurred since we
6840 * started waiting. For timeouts, we always want to return to userspace,
6841 * regardless of event count.
6843 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6844 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6847 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6848 int wake_flags, void *key)
6850 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6853 /* use noflush == true, as we can't safely rely on locking context */
6854 if (!io_should_wake(iowq, true))
6857 return autoremove_wake_function(curr, mode, wake_flags, key);
6860 static int io_run_task_work_sig(void)
6862 if (io_run_task_work())
6864 if (!signal_pending(current))
6866 if (current->jobctl & JOBCTL_TASK_WORK) {
6867 spin_lock_irq(¤t->sighand->siglock);
6868 current->jobctl &= ~JOBCTL_TASK_WORK;
6869 recalc_sigpending();
6870 spin_unlock_irq(¤t->sighand->siglock);
6877 * Wait until events become available, if we don't already have some. The
6878 * application must reap them itself, as they reside on the shared cq ring.
6880 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6881 const sigset_t __user *sig, size_t sigsz)
6883 struct io_wait_queue iowq = {
6886 .func = io_wake_function,
6887 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6890 .to_wait = min_events,
6892 struct io_rings *rings = ctx->rings;
6896 if (io_cqring_events(ctx, false) >= min_events)
6898 if (!io_run_task_work())
6903 #ifdef CONFIG_COMPAT
6904 if (in_compat_syscall())
6905 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6909 ret = set_user_sigmask(sig, sigsz);
6915 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6916 trace_io_uring_cqring_wait(ctx, min_events);
6918 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6919 TASK_INTERRUPTIBLE);
6920 /* make sure we run task_work before checking for signals */
6921 ret = io_run_task_work_sig();
6926 if (io_should_wake(&iowq, false))
6930 finish_wait(&ctx->wait, &iowq.wq);
6932 restore_saved_sigmask_unless(ret == -EINTR);
6934 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6937 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6939 #if defined(CONFIG_UNIX)
6940 if (ctx->ring_sock) {
6941 struct sock *sock = ctx->ring_sock->sk;
6942 struct sk_buff *skb;
6944 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6950 for (i = 0; i < ctx->nr_user_files; i++) {
6953 file = io_file_from_index(ctx, i);
6960 static void io_file_ref_kill(struct percpu_ref *ref)
6962 struct fixed_file_data *data;
6964 data = container_of(ref, struct fixed_file_data, refs);
6965 complete(&data->done);
6968 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6970 struct fixed_file_data *data = ctx->file_data;
6971 struct fixed_file_ref_node *ref_node = NULL;
6972 unsigned nr_tables, i;
6977 spin_lock(&data->lock);
6978 ref_node = data->node;
6979 spin_unlock(&data->lock);
6981 percpu_ref_kill(&ref_node->refs);
6983 percpu_ref_kill(&data->refs);
6985 /* wait for all refs nodes to complete */
6986 flush_delayed_work(&ctx->file_put_work);
6987 wait_for_completion(&data->done);
6989 __io_sqe_files_unregister(ctx);
6990 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6991 for (i = 0; i < nr_tables; i++)
6992 kfree(data->table[i].files);
6994 percpu_ref_exit(&data->refs);
6996 ctx->file_data = NULL;
6997 ctx->nr_user_files = 0;
7001 static void io_put_sq_data(struct io_sq_data *sqd)
7003 if (refcount_dec_and_test(&sqd->refs)) {
7005 * The park is a bit of a work-around, without it we get
7006 * warning spews on shutdown with SQPOLL set and affinity
7007 * set to a single CPU.
7010 kthread_park(sqd->thread);
7011 kthread_stop(sqd->thread);
7018 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7020 struct io_ring_ctx *ctx_attach;
7021 struct io_sq_data *sqd;
7024 f = fdget(p->wq_fd);
7026 return ERR_PTR(-ENXIO);
7027 if (f.file->f_op != &io_uring_fops) {
7029 return ERR_PTR(-EINVAL);
7032 ctx_attach = f.file->private_data;
7033 sqd = ctx_attach->sq_data;
7036 return ERR_PTR(-EINVAL);
7039 refcount_inc(&sqd->refs);
7044 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7046 struct io_sq_data *sqd;
7048 if (p->flags & IORING_SETUP_ATTACH_WQ)
7049 return io_attach_sq_data(p);
7051 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7053 return ERR_PTR(-ENOMEM);
7055 refcount_set(&sqd->refs, 1);
7056 INIT_LIST_HEAD(&sqd->ctx_list);
7057 INIT_LIST_HEAD(&sqd->ctx_new_list);
7058 mutex_init(&sqd->ctx_lock);
7059 mutex_init(&sqd->lock);
7060 init_waitqueue_head(&sqd->wait);
7064 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7065 __releases(&sqd->lock)
7069 kthread_unpark(sqd->thread);
7070 mutex_unlock(&sqd->lock);
7073 static void io_sq_thread_park(struct io_sq_data *sqd)
7074 __acquires(&sqd->lock)
7078 mutex_lock(&sqd->lock);
7079 kthread_park(sqd->thread);
7082 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7084 struct io_sq_data *sqd = ctx->sq_data;
7089 * We may arrive here from the error branch in
7090 * io_sq_offload_create() where the kthread is created
7091 * without being waked up, thus wake it up now to make
7092 * sure the wait will complete.
7094 wake_up_process(sqd->thread);
7095 wait_for_completion(&ctx->sq_thread_comp);
7097 io_sq_thread_park(sqd);
7100 mutex_lock(&sqd->ctx_lock);
7101 list_del(&ctx->sqd_list);
7102 mutex_unlock(&sqd->ctx_lock);
7105 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7106 io_sq_thread_unpark(sqd);
7109 io_put_sq_data(sqd);
7110 ctx->sq_data = NULL;
7114 static void io_finish_async(struct io_ring_ctx *ctx)
7116 io_sq_thread_stop(ctx);
7119 io_wq_destroy(ctx->io_wq);
7124 #if defined(CONFIG_UNIX)
7126 * Ensure the UNIX gc is aware of our file set, so we are certain that
7127 * the io_uring can be safely unregistered on process exit, even if we have
7128 * loops in the file referencing.
7130 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7132 struct sock *sk = ctx->ring_sock->sk;
7133 struct scm_fp_list *fpl;
7134 struct sk_buff *skb;
7137 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7141 skb = alloc_skb(0, GFP_KERNEL);
7150 fpl->user = get_uid(ctx->user);
7151 for (i = 0; i < nr; i++) {
7152 struct file *file = io_file_from_index(ctx, i + offset);
7156 fpl->fp[nr_files] = get_file(file);
7157 unix_inflight(fpl->user, fpl->fp[nr_files]);
7162 fpl->max = SCM_MAX_FD;
7163 fpl->count = nr_files;
7164 UNIXCB(skb).fp = fpl;
7165 skb->destructor = unix_destruct_scm;
7166 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7167 skb_queue_head(&sk->sk_receive_queue, skb);
7169 for (i = 0; i < nr_files; i++)
7180 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7181 * causes regular reference counting to break down. We rely on the UNIX
7182 * garbage collection to take care of this problem for us.
7184 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7186 unsigned left, total;
7190 left = ctx->nr_user_files;
7192 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7194 ret = __io_sqe_files_scm(ctx, this_files, total);
7198 total += this_files;
7204 while (total < ctx->nr_user_files) {
7205 struct file *file = io_file_from_index(ctx, total);
7215 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7221 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7222 unsigned nr_tables, unsigned nr_files)
7226 for (i = 0; i < nr_tables; i++) {
7227 struct fixed_file_table *table = &file_data->table[i];
7228 unsigned this_files;
7230 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7231 table->files = kcalloc(this_files, sizeof(struct file *),
7235 nr_files -= this_files;
7241 for (i = 0; i < nr_tables; i++) {
7242 struct fixed_file_table *table = &file_data->table[i];
7243 kfree(table->files);
7248 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7250 #if defined(CONFIG_UNIX)
7251 struct sock *sock = ctx->ring_sock->sk;
7252 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7253 struct sk_buff *skb;
7256 __skb_queue_head_init(&list);
7259 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7260 * remove this entry and rearrange the file array.
7262 skb = skb_dequeue(head);
7264 struct scm_fp_list *fp;
7266 fp = UNIXCB(skb).fp;
7267 for (i = 0; i < fp->count; i++) {
7270 if (fp->fp[i] != file)
7273 unix_notinflight(fp->user, fp->fp[i]);
7274 left = fp->count - 1 - i;
7276 memmove(&fp->fp[i], &fp->fp[i + 1],
7277 left * sizeof(struct file *));
7284 __skb_queue_tail(&list, skb);
7294 __skb_queue_tail(&list, skb);
7296 skb = skb_dequeue(head);
7299 if (skb_peek(&list)) {
7300 spin_lock_irq(&head->lock);
7301 while ((skb = __skb_dequeue(&list)) != NULL)
7302 __skb_queue_tail(head, skb);
7303 spin_unlock_irq(&head->lock);
7310 struct io_file_put {
7311 struct list_head list;
7315 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7317 struct fixed_file_data *file_data = ref_node->file_data;
7318 struct io_ring_ctx *ctx = file_data->ctx;
7319 struct io_file_put *pfile, *tmp;
7321 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7322 list_del(&pfile->list);
7323 io_ring_file_put(ctx, pfile->file);
7327 percpu_ref_exit(&ref_node->refs);
7329 percpu_ref_put(&file_data->refs);
7332 static void io_file_put_work(struct work_struct *work)
7334 struct io_ring_ctx *ctx;
7335 struct llist_node *node;
7337 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7338 node = llist_del_all(&ctx->file_put_llist);
7341 struct fixed_file_ref_node *ref_node;
7342 struct llist_node *next = node->next;
7344 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7345 __io_file_put_work(ref_node);
7350 static void io_file_data_ref_zero(struct percpu_ref *ref)
7352 struct fixed_file_ref_node *ref_node;
7353 struct fixed_file_data *data;
7354 struct io_ring_ctx *ctx;
7355 bool first_add = false;
7358 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7359 data = ref_node->file_data;
7362 spin_lock(&data->lock);
7363 ref_node->done = true;
7365 while (!list_empty(&data->ref_list)) {
7366 ref_node = list_first_entry(&data->ref_list,
7367 struct fixed_file_ref_node, node);
7368 /* recycle ref nodes in order */
7369 if (!ref_node->done)
7371 list_del(&ref_node->node);
7372 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7374 spin_unlock(&data->lock);
7376 if (percpu_ref_is_dying(&data->refs))
7380 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7382 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7385 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7386 struct io_ring_ctx *ctx)
7388 struct fixed_file_ref_node *ref_node;
7390 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7392 return ERR_PTR(-ENOMEM);
7394 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7397 return ERR_PTR(-ENOMEM);
7399 INIT_LIST_HEAD(&ref_node->node);
7400 INIT_LIST_HEAD(&ref_node->file_list);
7401 ref_node->file_data = ctx->file_data;
7402 ref_node->done = false;
7406 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7408 percpu_ref_exit(&ref_node->refs);
7412 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7415 __s32 __user *fds = (__s32 __user *) arg;
7416 unsigned nr_tables, i;
7418 int fd, ret = -ENOMEM;
7419 struct fixed_file_ref_node *ref_node;
7420 struct fixed_file_data *file_data;
7426 if (nr_args > IORING_MAX_FIXED_FILES)
7429 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7432 file_data->ctx = ctx;
7433 init_completion(&file_data->done);
7434 INIT_LIST_HEAD(&file_data->ref_list);
7435 spin_lock_init(&file_data->lock);
7437 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7438 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7440 if (!file_data->table)
7443 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7444 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7447 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7449 ctx->file_data = file_data;
7451 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7452 struct fixed_file_table *table;
7455 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7459 /* allow sparse sets */
7469 * Don't allow io_uring instances to be registered. If UNIX
7470 * isn't enabled, then this causes a reference cycle and this
7471 * instance can never get freed. If UNIX is enabled we'll
7472 * handle it just fine, but there's still no point in allowing
7473 * a ring fd as it doesn't support regular read/write anyway.
7475 if (file->f_op == &io_uring_fops) {
7479 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7480 index = i & IORING_FILE_TABLE_MASK;
7481 table->files[index] = file;
7484 ret = io_sqe_files_scm(ctx);
7486 io_sqe_files_unregister(ctx);
7490 ref_node = alloc_fixed_file_ref_node(ctx);
7491 if (IS_ERR(ref_node)) {
7492 io_sqe_files_unregister(ctx);
7493 return PTR_ERR(ref_node);
7496 file_data->node = ref_node;
7497 spin_lock(&file_data->lock);
7498 list_add_tail(&ref_node->node, &file_data->ref_list);
7499 spin_unlock(&file_data->lock);
7500 percpu_ref_get(&file_data->refs);
7503 for (i = 0; i < ctx->nr_user_files; i++) {
7504 file = io_file_from_index(ctx, i);
7508 for (i = 0; i < nr_tables; i++)
7509 kfree(file_data->table[i].files);
7510 ctx->nr_user_files = 0;
7512 percpu_ref_exit(&file_data->refs);
7514 kfree(file_data->table);
7516 ctx->file_data = NULL;
7520 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7523 #if defined(CONFIG_UNIX)
7524 struct sock *sock = ctx->ring_sock->sk;
7525 struct sk_buff_head *head = &sock->sk_receive_queue;
7526 struct sk_buff *skb;
7529 * See if we can merge this file into an existing skb SCM_RIGHTS
7530 * file set. If there's no room, fall back to allocating a new skb
7531 * and filling it in.
7533 spin_lock_irq(&head->lock);
7534 skb = skb_peek(head);
7536 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7538 if (fpl->count < SCM_MAX_FD) {
7539 __skb_unlink(skb, head);
7540 spin_unlock_irq(&head->lock);
7541 fpl->fp[fpl->count] = get_file(file);
7542 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7544 spin_lock_irq(&head->lock);
7545 __skb_queue_head(head, skb);
7550 spin_unlock_irq(&head->lock);
7557 return __io_sqe_files_scm(ctx, 1, index);
7563 static int io_queue_file_removal(struct fixed_file_data *data,
7566 struct io_file_put *pfile;
7567 struct fixed_file_ref_node *ref_node = data->node;
7569 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7574 list_add(&pfile->list, &ref_node->file_list);
7579 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7580 struct io_uring_files_update *up,
7583 struct fixed_file_data *data = ctx->file_data;
7584 struct fixed_file_ref_node *ref_node;
7589 bool needs_switch = false;
7591 if (check_add_overflow(up->offset, nr_args, &done))
7593 if (done > ctx->nr_user_files)
7596 ref_node = alloc_fixed_file_ref_node(ctx);
7597 if (IS_ERR(ref_node))
7598 return PTR_ERR(ref_node);
7601 fds = u64_to_user_ptr(up->fds);
7603 struct fixed_file_table *table;
7607 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7611 i = array_index_nospec(up->offset, ctx->nr_user_files);
7612 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7613 index = i & IORING_FILE_TABLE_MASK;
7614 if (table->files[index]) {
7615 file = table->files[index];
7616 err = io_queue_file_removal(data, file);
7619 table->files[index] = NULL;
7620 needs_switch = true;
7629 * Don't allow io_uring instances to be registered. If
7630 * UNIX isn't enabled, then this causes a reference
7631 * cycle and this instance can never get freed. If UNIX
7632 * is enabled we'll handle it just fine, but there's
7633 * still no point in allowing a ring fd as it doesn't
7634 * support regular read/write anyway.
7636 if (file->f_op == &io_uring_fops) {
7641 table->files[index] = file;
7642 err = io_sqe_file_register(ctx, file, i);
7644 table->files[index] = NULL;
7655 percpu_ref_kill(&data->node->refs);
7656 spin_lock(&data->lock);
7657 list_add_tail(&ref_node->node, &data->ref_list);
7658 data->node = ref_node;
7659 spin_unlock(&data->lock);
7660 percpu_ref_get(&ctx->file_data->refs);
7662 destroy_fixed_file_ref_node(ref_node);
7664 return done ? done : err;
7667 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7670 struct io_uring_files_update up;
7672 if (!ctx->file_data)
7676 if (copy_from_user(&up, arg, sizeof(up)))
7681 return __io_sqe_files_update(ctx, &up, nr_args);
7684 static void io_free_work(struct io_wq_work *work)
7686 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7688 /* Consider that io_steal_work() relies on this ref */
7692 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7693 struct io_uring_params *p)
7695 struct io_wq_data data;
7697 struct io_ring_ctx *ctx_attach;
7698 unsigned int concurrency;
7701 data.user = ctx->user;
7702 data.free_work = io_free_work;
7703 data.do_work = io_wq_submit_work;
7705 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7706 /* Do QD, or 4 * CPUS, whatever is smallest */
7707 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7709 ctx->io_wq = io_wq_create(concurrency, &data);
7710 if (IS_ERR(ctx->io_wq)) {
7711 ret = PTR_ERR(ctx->io_wq);
7717 f = fdget(p->wq_fd);
7721 if (f.file->f_op != &io_uring_fops) {
7726 ctx_attach = f.file->private_data;
7727 /* @io_wq is protected by holding the fd */
7728 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7733 ctx->io_wq = ctx_attach->io_wq;
7739 static int io_uring_alloc_task_context(struct task_struct *task)
7741 struct io_uring_task *tctx;
7744 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7745 if (unlikely(!tctx))
7748 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7749 if (unlikely(ret)) {
7755 init_waitqueue_head(&tctx->wait);
7757 atomic_set(&tctx->in_idle, 0);
7758 tctx->sqpoll = false;
7759 io_init_identity(&tctx->__identity);
7760 tctx->identity = &tctx->__identity;
7761 task->io_uring = tctx;
7765 void __io_uring_free(struct task_struct *tsk)
7767 struct io_uring_task *tctx = tsk->io_uring;
7769 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7770 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7771 if (tctx->identity != &tctx->__identity)
7772 kfree(tctx->identity);
7773 percpu_counter_destroy(&tctx->inflight);
7775 tsk->io_uring = NULL;
7778 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7779 struct io_uring_params *p)
7783 if (ctx->flags & IORING_SETUP_SQPOLL) {
7784 struct io_sq_data *sqd;
7787 if (!capable(CAP_SYS_ADMIN))
7790 sqd = io_get_sq_data(p);
7797 io_sq_thread_park(sqd);
7798 mutex_lock(&sqd->ctx_lock);
7799 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7800 mutex_unlock(&sqd->ctx_lock);
7801 io_sq_thread_unpark(sqd);
7803 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7804 if (!ctx->sq_thread_idle)
7805 ctx->sq_thread_idle = HZ;
7810 if (p->flags & IORING_SETUP_SQ_AFF) {
7811 int cpu = p->sq_thread_cpu;
7814 if (cpu >= nr_cpu_ids)
7816 if (!cpu_online(cpu))
7819 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7820 cpu, "io_uring-sq");
7822 sqd->thread = kthread_create(io_sq_thread, sqd,
7825 if (IS_ERR(sqd->thread)) {
7826 ret = PTR_ERR(sqd->thread);
7830 ret = io_uring_alloc_task_context(sqd->thread);
7833 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7834 /* Can't have SQ_AFF without SQPOLL */
7840 ret = io_init_wq_offload(ctx, p);
7846 io_finish_async(ctx);
7850 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7852 struct io_sq_data *sqd = ctx->sq_data;
7854 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7855 wake_up_process(sqd->thread);
7858 static inline void __io_unaccount_mem(struct user_struct *user,
7859 unsigned long nr_pages)
7861 atomic_long_sub(nr_pages, &user->locked_vm);
7864 static inline int __io_account_mem(struct user_struct *user,
7865 unsigned long nr_pages)
7867 unsigned long page_limit, cur_pages, new_pages;
7869 /* Don't allow more pages than we can safely lock */
7870 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7873 cur_pages = atomic_long_read(&user->locked_vm);
7874 new_pages = cur_pages + nr_pages;
7875 if (new_pages > page_limit)
7877 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7878 new_pages) != cur_pages);
7883 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7884 enum io_mem_account acct)
7887 __io_unaccount_mem(ctx->user, nr_pages);
7889 if (ctx->mm_account) {
7890 if (acct == ACCT_LOCKED)
7891 ctx->mm_account->locked_vm -= nr_pages;
7892 else if (acct == ACCT_PINNED)
7893 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7897 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7898 enum io_mem_account acct)
7902 if (ctx->limit_mem) {
7903 ret = __io_account_mem(ctx->user, nr_pages);
7908 if (ctx->mm_account) {
7909 if (acct == ACCT_LOCKED)
7910 ctx->mm_account->locked_vm += nr_pages;
7911 else if (acct == ACCT_PINNED)
7912 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7918 static void io_mem_free(void *ptr)
7925 page = virt_to_head_page(ptr);
7926 if (put_page_testzero(page))
7927 free_compound_page(page);
7930 static void *io_mem_alloc(size_t size)
7932 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7935 return (void *) __get_free_pages(gfp_flags, get_order(size));
7938 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7941 struct io_rings *rings;
7942 size_t off, sq_array_size;
7944 off = struct_size(rings, cqes, cq_entries);
7945 if (off == SIZE_MAX)
7949 off = ALIGN(off, SMP_CACHE_BYTES);
7957 sq_array_size = array_size(sizeof(u32), sq_entries);
7958 if (sq_array_size == SIZE_MAX)
7961 if (check_add_overflow(off, sq_array_size, &off))
7967 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7971 pages = (size_t)1 << get_order(
7972 rings_size(sq_entries, cq_entries, NULL));
7973 pages += (size_t)1 << get_order(
7974 array_size(sizeof(struct io_uring_sqe), sq_entries));
7979 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7983 if (!ctx->user_bufs)
7986 for (i = 0; i < ctx->nr_user_bufs; i++) {
7987 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7989 for (j = 0; j < imu->nr_bvecs; j++)
7990 unpin_user_page(imu->bvec[j].bv_page);
7992 if (imu->acct_pages)
7993 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7998 kfree(ctx->user_bufs);
7999 ctx->user_bufs = NULL;
8000 ctx->nr_user_bufs = 0;
8004 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8005 void __user *arg, unsigned index)
8007 struct iovec __user *src;
8009 #ifdef CONFIG_COMPAT
8011 struct compat_iovec __user *ciovs;
8012 struct compat_iovec ciov;
8014 ciovs = (struct compat_iovec __user *) arg;
8015 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8018 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8019 dst->iov_len = ciov.iov_len;
8023 src = (struct iovec __user *) arg;
8024 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8030 * Not super efficient, but this is just a registration time. And we do cache
8031 * the last compound head, so generally we'll only do a full search if we don't
8034 * We check if the given compound head page has already been accounted, to
8035 * avoid double accounting it. This allows us to account the full size of the
8036 * page, not just the constituent pages of a huge page.
8038 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8039 int nr_pages, struct page *hpage)
8043 /* check current page array */
8044 for (i = 0; i < nr_pages; i++) {
8045 if (!PageCompound(pages[i]))
8047 if (compound_head(pages[i]) == hpage)
8051 /* check previously registered pages */
8052 for (i = 0; i < ctx->nr_user_bufs; i++) {
8053 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8055 for (j = 0; j < imu->nr_bvecs; j++) {
8056 if (!PageCompound(imu->bvec[j].bv_page))
8058 if (compound_head(imu->bvec[j].bv_page) == hpage)
8066 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8067 int nr_pages, struct io_mapped_ubuf *imu,
8068 struct page **last_hpage)
8072 for (i = 0; i < nr_pages; i++) {
8073 if (!PageCompound(pages[i])) {
8078 hpage = compound_head(pages[i]);
8079 if (hpage == *last_hpage)
8081 *last_hpage = hpage;
8082 if (headpage_already_acct(ctx, pages, i, hpage))
8084 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8088 if (!imu->acct_pages)
8091 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8093 imu->acct_pages = 0;
8097 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8100 struct vm_area_struct **vmas = NULL;
8101 struct page **pages = NULL;
8102 struct page *last_hpage = NULL;
8103 int i, j, got_pages = 0;
8108 if (!nr_args || nr_args > UIO_MAXIOV)
8111 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8113 if (!ctx->user_bufs)
8116 for (i = 0; i < nr_args; i++) {
8117 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8118 unsigned long off, start, end, ubuf;
8123 ret = io_copy_iov(ctx, &iov, arg, i);
8128 * Don't impose further limits on the size and buffer
8129 * constraints here, we'll -EINVAL later when IO is
8130 * submitted if they are wrong.
8133 if (!iov.iov_base || !iov.iov_len)
8136 /* arbitrary limit, but we need something */
8137 if (iov.iov_len > SZ_1G)
8140 ubuf = (unsigned long) iov.iov_base;
8141 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8142 start = ubuf >> PAGE_SHIFT;
8143 nr_pages = end - start;
8146 if (!pages || nr_pages > got_pages) {
8149 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8151 vmas = kvmalloc_array(nr_pages,
8152 sizeof(struct vm_area_struct *),
8154 if (!pages || !vmas) {
8158 got_pages = nr_pages;
8161 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8168 mmap_read_lock(current->mm);
8169 pret = pin_user_pages(ubuf, nr_pages,
8170 FOLL_WRITE | FOLL_LONGTERM,
8172 if (pret == nr_pages) {
8173 /* don't support file backed memory */
8174 for (j = 0; j < nr_pages; j++) {
8175 struct vm_area_struct *vma = vmas[j];
8178 !is_file_hugepages(vma->vm_file)) {
8184 ret = pret < 0 ? pret : -EFAULT;
8186 mmap_read_unlock(current->mm);
8189 * if we did partial map, or found file backed vmas,
8190 * release any pages we did get
8193 unpin_user_pages(pages, pret);
8198 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8200 unpin_user_pages(pages, pret);
8205 off = ubuf & ~PAGE_MASK;
8207 for (j = 0; j < nr_pages; j++) {
8210 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8211 imu->bvec[j].bv_page = pages[j];
8212 imu->bvec[j].bv_len = vec_len;
8213 imu->bvec[j].bv_offset = off;
8217 /* store original address for later verification */
8219 imu->len = iov.iov_len;
8220 imu->nr_bvecs = nr_pages;
8222 ctx->nr_user_bufs++;
8230 io_sqe_buffer_unregister(ctx);
8234 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8236 __s32 __user *fds = arg;
8242 if (copy_from_user(&fd, fds, sizeof(*fds)))
8245 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8246 if (IS_ERR(ctx->cq_ev_fd)) {
8247 int ret = PTR_ERR(ctx->cq_ev_fd);
8248 ctx->cq_ev_fd = NULL;
8255 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8257 if (ctx->cq_ev_fd) {
8258 eventfd_ctx_put(ctx->cq_ev_fd);
8259 ctx->cq_ev_fd = NULL;
8266 static int __io_destroy_buffers(int id, void *p, void *data)
8268 struct io_ring_ctx *ctx = data;
8269 struct io_buffer *buf = p;
8271 __io_remove_buffers(ctx, buf, id, -1U);
8275 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8277 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8278 idr_destroy(&ctx->io_buffer_idr);
8281 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8283 io_finish_async(ctx);
8284 io_sqe_buffer_unregister(ctx);
8286 if (ctx->sqo_task) {
8287 put_task_struct(ctx->sqo_task);
8288 ctx->sqo_task = NULL;
8289 mmdrop(ctx->mm_account);
8290 ctx->mm_account = NULL;
8293 #ifdef CONFIG_BLK_CGROUP
8294 if (ctx->sqo_blkcg_css)
8295 css_put(ctx->sqo_blkcg_css);
8298 io_sqe_files_unregister(ctx);
8299 io_eventfd_unregister(ctx);
8300 io_destroy_buffers(ctx);
8301 idr_destroy(&ctx->personality_idr);
8303 #if defined(CONFIG_UNIX)
8304 if (ctx->ring_sock) {
8305 ctx->ring_sock->file = NULL; /* so that iput() is called */
8306 sock_release(ctx->ring_sock);
8310 io_mem_free(ctx->rings);
8311 io_mem_free(ctx->sq_sqes);
8313 percpu_ref_exit(&ctx->refs);
8314 free_uid(ctx->user);
8315 put_cred(ctx->creds);
8316 kfree(ctx->cancel_hash);
8317 kmem_cache_free(req_cachep, ctx->fallback_req);
8321 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8323 struct io_ring_ctx *ctx = file->private_data;
8326 poll_wait(file, &ctx->cq_wait, wait);
8328 * synchronizes with barrier from wq_has_sleeper call in
8332 if (!io_sqring_full(ctx))
8333 mask |= EPOLLOUT | EPOLLWRNORM;
8334 if (io_cqring_events(ctx, false))
8335 mask |= EPOLLIN | EPOLLRDNORM;
8340 static int io_uring_fasync(int fd, struct file *file, int on)
8342 struct io_ring_ctx *ctx = file->private_data;
8344 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8347 static int io_remove_personalities(int id, void *p, void *data)
8349 struct io_ring_ctx *ctx = data;
8350 struct io_identity *iod;
8352 iod = idr_remove(&ctx->personality_idr, id);
8354 put_cred(iod->creds);
8355 if (refcount_dec_and_test(&iod->count))
8361 static void io_ring_exit_work(struct work_struct *work)
8363 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8367 * If we're doing polled IO and end up having requests being
8368 * submitted async (out-of-line), then completions can come in while
8369 * we're waiting for refs to drop. We need to reap these manually,
8370 * as nobody else will be looking for them.
8374 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8375 io_iopoll_try_reap_events(ctx);
8376 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8377 io_ring_ctx_free(ctx);
8380 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8382 mutex_lock(&ctx->uring_lock);
8383 percpu_ref_kill(&ctx->refs);
8384 mutex_unlock(&ctx->uring_lock);
8386 io_kill_timeouts(ctx, NULL);
8387 io_poll_remove_all(ctx, NULL);
8390 io_wq_cancel_all(ctx->io_wq);
8392 /* if we failed setting up the ctx, we might not have any rings */
8394 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8395 io_iopoll_try_reap_events(ctx);
8396 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8399 * Do this upfront, so we won't have a grace period where the ring
8400 * is closed but resources aren't reaped yet. This can cause
8401 * spurious failure in setting up a new ring.
8403 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8406 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8408 * Use system_unbound_wq to avoid spawning tons of event kworkers
8409 * if we're exiting a ton of rings at the same time. It just adds
8410 * noise and overhead, there's no discernable change in runtime
8411 * over using system_wq.
8413 queue_work(system_unbound_wq, &ctx->exit_work);
8416 static int io_uring_release(struct inode *inode, struct file *file)
8418 struct io_ring_ctx *ctx = file->private_data;
8420 file->private_data = NULL;
8421 io_ring_ctx_wait_and_kill(ctx);
8425 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8427 struct files_struct *files = data;
8429 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8430 work->identity->files == files);
8434 * Returns true if 'preq' is the link parent of 'req'
8436 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8438 struct io_kiocb *link;
8440 if (!(preq->flags & REQ_F_LINK_HEAD))
8443 list_for_each_entry(link, &preq->link_list, link_list) {
8452 * We're looking to cancel 'req' because it's holding on to our files, but
8453 * 'req' could be a link to another request. See if it is, and cancel that
8454 * parent request if so.
8456 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8458 struct hlist_node *tmp;
8459 struct io_kiocb *preq;
8463 spin_lock_irq(&ctx->completion_lock);
8464 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8465 struct hlist_head *list;
8467 list = &ctx->cancel_hash[i];
8468 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8469 found = io_match_link(preq, req);
8471 io_poll_remove_one(preq);
8476 spin_unlock_irq(&ctx->completion_lock);
8480 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8481 struct io_kiocb *req)
8483 struct io_kiocb *preq;
8486 spin_lock_irq(&ctx->completion_lock);
8487 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8488 found = io_match_link(preq, req);
8490 __io_timeout_cancel(preq);
8494 spin_unlock_irq(&ctx->completion_lock);
8498 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8500 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8503 if (req->flags & REQ_F_LINK_TIMEOUT) {
8504 unsigned long flags;
8505 struct io_ring_ctx *ctx = req->ctx;
8507 /* protect against races with linked timeouts */
8508 spin_lock_irqsave(&ctx->completion_lock, flags);
8509 ret = io_match_link(req, data);
8510 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8512 ret = io_match_link(req, data);
8517 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8519 enum io_wq_cancel cret;
8521 /* cancel this particular work, if it's running */
8522 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8523 if (cret != IO_WQ_CANCEL_NOTFOUND)
8526 /* find links that hold this pending, cancel those */
8527 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8528 if (cret != IO_WQ_CANCEL_NOTFOUND)
8531 /* if we have a poll link holding this pending, cancel that */
8532 if (io_poll_remove_link(ctx, req))
8535 /* final option, timeout link is holding this req pending */
8536 io_timeout_remove_link(ctx, req);
8539 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8540 struct task_struct *task,
8541 struct files_struct *files)
8543 struct io_defer_entry *de = NULL;
8546 spin_lock_irq(&ctx->completion_lock);
8547 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8548 if (io_task_match(de->req, task) &&
8549 io_match_files(de->req, files)) {
8550 list_cut_position(&list, &ctx->defer_list, &de->list);
8554 spin_unlock_irq(&ctx->completion_lock);
8556 while (!list_empty(&list)) {
8557 de = list_first_entry(&list, struct io_defer_entry, list);
8558 list_del_init(&de->list);
8559 req_set_fail_links(de->req);
8560 io_put_req(de->req);
8561 io_req_complete(de->req, -ECANCELED);
8567 * Returns true if we found and killed one or more files pinning requests
8569 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8570 struct files_struct *files)
8572 if (list_empty_careful(&ctx->inflight_list))
8575 /* cancel all at once, should be faster than doing it one by one*/
8576 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8578 while (!list_empty_careful(&ctx->inflight_list)) {
8579 struct io_kiocb *cancel_req = NULL, *req;
8582 spin_lock_irq(&ctx->inflight_lock);
8583 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8584 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8585 req->work.identity->files != files)
8587 /* req is being completed, ignore */
8588 if (!refcount_inc_not_zero(&req->refs))
8594 prepare_to_wait(&ctx->inflight_wait, &wait,
8595 TASK_UNINTERRUPTIBLE);
8596 spin_unlock_irq(&ctx->inflight_lock);
8598 /* We need to keep going until we don't find a matching req */
8601 /* cancel this request, or head link requests */
8602 io_attempt_cancel(ctx, cancel_req);
8603 io_put_req(cancel_req);
8604 /* cancellations _may_ trigger task work */
8607 finish_wait(&ctx->inflight_wait, &wait);
8613 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8615 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8616 struct task_struct *task = data;
8618 return io_task_match(req, task);
8621 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8622 struct task_struct *task,
8623 struct files_struct *files)
8627 ret = io_uring_cancel_files(ctx, files);
8629 enum io_wq_cancel cret;
8631 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8632 if (cret != IO_WQ_CANCEL_NOTFOUND)
8635 /* SQPOLL thread does its own polling */
8636 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8637 while (!list_empty_careful(&ctx->iopoll_list)) {
8638 io_iopoll_try_reap_events(ctx);
8643 ret |= io_poll_remove_all(ctx, task);
8644 ret |= io_kill_timeouts(ctx, task);
8651 * We need to iteratively cancel requests, in case a request has dependent
8652 * hard links. These persist even for failure of cancelations, hence keep
8653 * looping until none are found.
8655 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8656 struct files_struct *files)
8658 struct task_struct *task = current;
8660 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8661 task = ctx->sq_data->thread;
8662 atomic_inc(&task->io_uring->in_idle);
8663 io_sq_thread_park(ctx->sq_data);
8667 io_cancel_defer_files(ctx, NULL, files);
8669 io_cancel_defer_files(ctx, task, NULL);
8671 io_cqring_overflow_flush(ctx, true, task, files);
8673 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8678 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8679 atomic_dec(&task->io_uring->in_idle);
8681 * If the files that are going away are the ones in the thread
8682 * identity, clear them out.
8684 if (task->io_uring->identity->files == files)
8685 task->io_uring->identity->files = NULL;
8686 io_sq_thread_unpark(ctx->sq_data);
8691 * Note that this task has used io_uring. We use it for cancelation purposes.
8693 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8695 struct io_uring_task *tctx = current->io_uring;
8697 if (unlikely(!tctx)) {
8700 ret = io_uring_alloc_task_context(current);
8703 tctx = current->io_uring;
8705 if (tctx->last != file) {
8706 void *old = xa_load(&tctx->xa, (unsigned long)file);
8710 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8716 * This is race safe in that the task itself is doing this, hence it
8717 * cannot be going through the exit/cancel paths at the same time.
8718 * This cannot be modified while exit/cancel is running.
8720 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8721 tctx->sqpoll = true;
8727 * Remove this io_uring_file -> task mapping.
8729 static void io_uring_del_task_file(struct file *file)
8731 struct io_uring_task *tctx = current->io_uring;
8733 if (tctx->last == file)
8735 file = xa_erase(&tctx->xa, (unsigned long)file);
8741 * Drop task note for this file if we're the only ones that hold it after
8744 static void io_uring_attempt_task_drop(struct file *file)
8746 if (!current->io_uring)
8749 * fput() is pending, will be 2 if the only other ref is our potential
8750 * task file note. If the task is exiting, drop regardless of count.
8752 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8753 atomic_long_read(&file->f_count) == 2)
8754 io_uring_del_task_file(file);
8757 void __io_uring_files_cancel(struct files_struct *files)
8759 struct io_uring_task *tctx = current->io_uring;
8761 unsigned long index;
8763 /* make sure overflow events are dropped */
8764 atomic_inc(&tctx->in_idle);
8766 xa_for_each(&tctx->xa, index, file) {
8767 struct io_ring_ctx *ctx = file->private_data;
8769 io_uring_cancel_task_requests(ctx, files);
8771 io_uring_del_task_file(file);
8774 atomic_dec(&tctx->in_idle);
8777 static s64 tctx_inflight(struct io_uring_task *tctx)
8779 unsigned long index;
8783 inflight = percpu_counter_sum(&tctx->inflight);
8788 * If we have SQPOLL rings, then we need to iterate and find them, and
8789 * add the pending count for those.
8791 xa_for_each(&tctx->xa, index, file) {
8792 struct io_ring_ctx *ctx = file->private_data;
8794 if (ctx->flags & IORING_SETUP_SQPOLL) {
8795 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
8797 inflight += percpu_counter_sum(&__tctx->inflight);
8805 * Find any io_uring fd that this task has registered or done IO on, and cancel
8808 void __io_uring_task_cancel(void)
8810 struct io_uring_task *tctx = current->io_uring;
8814 /* make sure overflow events are dropped */
8815 atomic_inc(&tctx->in_idle);
8818 /* read completions before cancelations */
8819 inflight = tctx_inflight(tctx);
8822 __io_uring_files_cancel(NULL);
8824 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8827 * If we've seen completions, retry. This avoids a race where
8828 * a completion comes in before we did prepare_to_wait().
8830 if (inflight != tctx_inflight(tctx))
8835 finish_wait(&tctx->wait, &wait);
8836 atomic_dec(&tctx->in_idle);
8839 static int io_uring_flush(struct file *file, void *data)
8841 io_uring_attempt_task_drop(file);
8845 static void *io_uring_validate_mmap_request(struct file *file,
8846 loff_t pgoff, size_t sz)
8848 struct io_ring_ctx *ctx = file->private_data;
8849 loff_t offset = pgoff << PAGE_SHIFT;
8854 case IORING_OFF_SQ_RING:
8855 case IORING_OFF_CQ_RING:
8858 case IORING_OFF_SQES:
8862 return ERR_PTR(-EINVAL);
8865 page = virt_to_head_page(ptr);
8866 if (sz > page_size(page))
8867 return ERR_PTR(-EINVAL);
8874 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8876 size_t sz = vma->vm_end - vma->vm_start;
8880 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8882 return PTR_ERR(ptr);
8884 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8885 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8888 #else /* !CONFIG_MMU */
8890 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8892 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8895 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8897 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8900 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8901 unsigned long addr, unsigned long len,
8902 unsigned long pgoff, unsigned long flags)
8906 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8908 return PTR_ERR(ptr);
8910 return (unsigned long) ptr;
8913 #endif /* !CONFIG_MMU */
8915 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8920 if (!io_sqring_full(ctx))
8923 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8925 if (!io_sqring_full(ctx))
8929 } while (!signal_pending(current));
8931 finish_wait(&ctx->sqo_sq_wait, &wait);
8934 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8935 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8938 struct io_ring_ctx *ctx;
8945 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8946 IORING_ENTER_SQ_WAIT))
8954 if (f.file->f_op != &io_uring_fops)
8958 ctx = f.file->private_data;
8959 if (!percpu_ref_tryget(&ctx->refs))
8963 if (ctx->flags & IORING_SETUP_R_DISABLED)
8967 * For SQ polling, the thread will do all submissions and completions.
8968 * Just return the requested submit count, and wake the thread if
8972 if (ctx->flags & IORING_SETUP_SQPOLL) {
8973 if (!list_empty_careful(&ctx->cq_overflow_list))
8974 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8975 if (flags & IORING_ENTER_SQ_WAKEUP)
8976 wake_up(&ctx->sq_data->wait);
8977 if (flags & IORING_ENTER_SQ_WAIT)
8978 io_sqpoll_wait_sq(ctx);
8979 submitted = to_submit;
8980 } else if (to_submit) {
8981 ret = io_uring_add_task_file(ctx, f.file);
8984 mutex_lock(&ctx->uring_lock);
8985 submitted = io_submit_sqes(ctx, to_submit);
8986 mutex_unlock(&ctx->uring_lock);
8988 if (submitted != to_submit)
8991 if (flags & IORING_ENTER_GETEVENTS) {
8992 min_complete = min(min_complete, ctx->cq_entries);
8995 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8996 * space applications don't need to do io completion events
8997 * polling again, they can rely on io_sq_thread to do polling
8998 * work, which can reduce cpu usage and uring_lock contention.
9000 if (ctx->flags & IORING_SETUP_IOPOLL &&
9001 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9002 ret = io_iopoll_check(ctx, min_complete);
9004 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
9009 percpu_ref_put(&ctx->refs);
9012 return submitted ? submitted : ret;
9015 #ifdef CONFIG_PROC_FS
9016 static int io_uring_show_cred(int id, void *p, void *data)
9018 struct io_identity *iod = p;
9019 const struct cred *cred = iod->creds;
9020 struct seq_file *m = data;
9021 struct user_namespace *uns = seq_user_ns(m);
9022 struct group_info *gi;
9027 seq_printf(m, "%5d\n", id);
9028 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9029 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9030 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9031 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9032 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9033 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9034 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9035 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9036 seq_puts(m, "\n\tGroups:\t");
9037 gi = cred->group_info;
9038 for (g = 0; g < gi->ngroups; g++) {
9039 seq_put_decimal_ull(m, g ? " " : "",
9040 from_kgid_munged(uns, gi->gid[g]));
9042 seq_puts(m, "\n\tCapEff:\t");
9043 cap = cred->cap_effective;
9044 CAP_FOR_EACH_U32(__capi)
9045 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9050 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9052 struct io_sq_data *sq = NULL;
9057 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9058 * since fdinfo case grabs it in the opposite direction of normal use
9059 * cases. If we fail to get the lock, we just don't iterate any
9060 * structures that could be going away outside the io_uring mutex.
9062 has_lock = mutex_trylock(&ctx->uring_lock);
9064 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9067 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9068 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9069 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9070 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9071 struct fixed_file_table *table;
9074 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9075 f = table->files[i & IORING_FILE_TABLE_MASK];
9077 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9079 seq_printf(m, "%5u: <none>\n", i);
9081 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9082 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9083 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9085 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9086 (unsigned int) buf->len);
9088 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9089 seq_printf(m, "Personalities:\n");
9090 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9092 seq_printf(m, "PollList:\n");
9093 spin_lock_irq(&ctx->completion_lock);
9094 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9095 struct hlist_head *list = &ctx->cancel_hash[i];
9096 struct io_kiocb *req;
9098 hlist_for_each_entry(req, list, hash_node)
9099 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9100 req->task->task_works != NULL);
9102 spin_unlock_irq(&ctx->completion_lock);
9104 mutex_unlock(&ctx->uring_lock);
9107 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9109 struct io_ring_ctx *ctx = f->private_data;
9111 if (percpu_ref_tryget(&ctx->refs)) {
9112 __io_uring_show_fdinfo(ctx, m);
9113 percpu_ref_put(&ctx->refs);
9118 static const struct file_operations io_uring_fops = {
9119 .release = io_uring_release,
9120 .flush = io_uring_flush,
9121 .mmap = io_uring_mmap,
9123 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9124 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9126 .poll = io_uring_poll,
9127 .fasync = io_uring_fasync,
9128 #ifdef CONFIG_PROC_FS
9129 .show_fdinfo = io_uring_show_fdinfo,
9133 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9134 struct io_uring_params *p)
9136 struct io_rings *rings;
9137 size_t size, sq_array_offset;
9139 /* make sure these are sane, as we already accounted them */
9140 ctx->sq_entries = p->sq_entries;
9141 ctx->cq_entries = p->cq_entries;
9143 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9144 if (size == SIZE_MAX)
9147 rings = io_mem_alloc(size);
9152 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9153 rings->sq_ring_mask = p->sq_entries - 1;
9154 rings->cq_ring_mask = p->cq_entries - 1;
9155 rings->sq_ring_entries = p->sq_entries;
9156 rings->cq_ring_entries = p->cq_entries;
9157 ctx->sq_mask = rings->sq_ring_mask;
9158 ctx->cq_mask = rings->cq_ring_mask;
9160 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9161 if (size == SIZE_MAX) {
9162 io_mem_free(ctx->rings);
9167 ctx->sq_sqes = io_mem_alloc(size);
9168 if (!ctx->sq_sqes) {
9169 io_mem_free(ctx->rings);
9178 * Allocate an anonymous fd, this is what constitutes the application
9179 * visible backing of an io_uring instance. The application mmaps this
9180 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9181 * we have to tie this fd to a socket for file garbage collection purposes.
9183 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9188 #if defined(CONFIG_UNIX)
9189 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9195 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9199 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9200 O_RDWR | O_CLOEXEC);
9204 ret = PTR_ERR(file);
9208 #if defined(CONFIG_UNIX)
9209 ctx->ring_sock->file = file;
9211 if (unlikely(io_uring_add_task_file(ctx, file))) {
9212 file = ERR_PTR(-ENOMEM);
9215 fd_install(ret, file);
9218 #if defined(CONFIG_UNIX)
9219 sock_release(ctx->ring_sock);
9220 ctx->ring_sock = NULL;
9225 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9226 struct io_uring_params __user *params)
9228 struct user_struct *user = NULL;
9229 struct io_ring_ctx *ctx;
9235 if (entries > IORING_MAX_ENTRIES) {
9236 if (!(p->flags & IORING_SETUP_CLAMP))
9238 entries = IORING_MAX_ENTRIES;
9242 * Use twice as many entries for the CQ ring. It's possible for the
9243 * application to drive a higher depth than the size of the SQ ring,
9244 * since the sqes are only used at submission time. This allows for
9245 * some flexibility in overcommitting a bit. If the application has
9246 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9247 * of CQ ring entries manually.
9249 p->sq_entries = roundup_pow_of_two(entries);
9250 if (p->flags & IORING_SETUP_CQSIZE) {
9252 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9253 * to a power-of-two, if it isn't already. We do NOT impose
9254 * any cq vs sq ring sizing.
9256 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9257 if (p->cq_entries < p->sq_entries)
9259 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9260 if (!(p->flags & IORING_SETUP_CLAMP))
9262 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9265 p->cq_entries = 2 * p->sq_entries;
9268 user = get_uid(current_user());
9269 limit_mem = !capable(CAP_IPC_LOCK);
9272 ret = __io_account_mem(user,
9273 ring_pages(p->sq_entries, p->cq_entries));
9280 ctx = io_ring_ctx_alloc(p);
9283 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9288 ctx->compat = in_compat_syscall();
9290 ctx->creds = get_current_cred();
9292 ctx->loginuid = current->loginuid;
9293 ctx->sessionid = current->sessionid;
9295 ctx->sqo_task = get_task_struct(current);
9298 * This is just grabbed for accounting purposes. When a process exits,
9299 * the mm is exited and dropped before the files, hence we need to hang
9300 * on to this mm purely for the purposes of being able to unaccount
9301 * memory (locked/pinned vm). It's not used for anything else.
9303 mmgrab(current->mm);
9304 ctx->mm_account = current->mm;
9306 #ifdef CONFIG_BLK_CGROUP
9308 * The sq thread will belong to the original cgroup it was inited in.
9309 * If the cgroup goes offline (e.g. disabling the io controller), then
9310 * issued bios will be associated with the closest cgroup later in the
9314 ctx->sqo_blkcg_css = blkcg_css();
9315 ret = css_tryget_online(ctx->sqo_blkcg_css);
9318 /* don't init against a dying cgroup, have the user try again */
9319 ctx->sqo_blkcg_css = NULL;
9326 * Account memory _before_ installing the file descriptor. Once
9327 * the descriptor is installed, it can get closed at any time. Also
9328 * do this before hitting the general error path, as ring freeing
9329 * will un-account as well.
9331 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9333 ctx->limit_mem = limit_mem;
9335 ret = io_allocate_scq_urings(ctx, p);
9339 ret = io_sq_offload_create(ctx, p);
9343 if (!(p->flags & IORING_SETUP_R_DISABLED))
9344 io_sq_offload_start(ctx);
9346 memset(&p->sq_off, 0, sizeof(p->sq_off));
9347 p->sq_off.head = offsetof(struct io_rings, sq.head);
9348 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9349 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9350 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9351 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9352 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9353 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9355 memset(&p->cq_off, 0, sizeof(p->cq_off));
9356 p->cq_off.head = offsetof(struct io_rings, cq.head);
9357 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9358 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9359 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9360 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9361 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9362 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9364 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9365 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9366 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9367 IORING_FEAT_POLL_32BITS;
9369 if (copy_to_user(params, p, sizeof(*p))) {
9375 * Install ring fd as the very last thing, so we don't risk someone
9376 * having closed it before we finish setup
9378 ret = io_uring_get_fd(ctx);
9382 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9385 io_ring_ctx_wait_and_kill(ctx);
9390 * Sets up an aio uring context, and returns the fd. Applications asks for a
9391 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9392 * params structure passed in.
9394 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9396 struct io_uring_params p;
9399 if (copy_from_user(&p, params, sizeof(p)))
9401 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9406 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9407 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9408 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9409 IORING_SETUP_R_DISABLED))
9412 return io_uring_create(entries, &p, params);
9415 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9416 struct io_uring_params __user *, params)
9418 return io_uring_setup(entries, params);
9421 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9423 struct io_uring_probe *p;
9427 size = struct_size(p, ops, nr_args);
9428 if (size == SIZE_MAX)
9430 p = kzalloc(size, GFP_KERNEL);
9435 if (copy_from_user(p, arg, size))
9438 if (memchr_inv(p, 0, size))
9441 p->last_op = IORING_OP_LAST - 1;
9442 if (nr_args > IORING_OP_LAST)
9443 nr_args = IORING_OP_LAST;
9445 for (i = 0; i < nr_args; i++) {
9447 if (!io_op_defs[i].not_supported)
9448 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9453 if (copy_to_user(arg, p, size))
9460 static int io_register_personality(struct io_ring_ctx *ctx)
9462 struct io_identity *id;
9465 id = kmalloc(sizeof(*id), GFP_KERNEL);
9469 io_init_identity(id);
9470 id->creds = get_current_cred();
9472 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9474 put_cred(id->creds);
9480 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9482 struct io_identity *iod;
9484 iod = idr_remove(&ctx->personality_idr, id);
9486 put_cred(iod->creds);
9487 if (refcount_dec_and_test(&iod->count))
9495 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9496 unsigned int nr_args)
9498 struct io_uring_restriction *res;
9502 /* Restrictions allowed only if rings started disabled */
9503 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9506 /* We allow only a single restrictions registration */
9507 if (ctx->restrictions.registered)
9510 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9513 size = array_size(nr_args, sizeof(*res));
9514 if (size == SIZE_MAX)
9517 res = memdup_user(arg, size);
9519 return PTR_ERR(res);
9523 for (i = 0; i < nr_args; i++) {
9524 switch (res[i].opcode) {
9525 case IORING_RESTRICTION_REGISTER_OP:
9526 if (res[i].register_op >= IORING_REGISTER_LAST) {
9531 __set_bit(res[i].register_op,
9532 ctx->restrictions.register_op);
9534 case IORING_RESTRICTION_SQE_OP:
9535 if (res[i].sqe_op >= IORING_OP_LAST) {
9540 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9542 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9543 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9545 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9546 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9555 /* Reset all restrictions if an error happened */
9557 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9559 ctx->restrictions.registered = true;
9565 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9567 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9570 if (ctx->restrictions.registered)
9571 ctx->restricted = 1;
9573 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9575 io_sq_offload_start(ctx);
9580 static bool io_register_op_must_quiesce(int op)
9583 case IORING_UNREGISTER_FILES:
9584 case IORING_REGISTER_FILES_UPDATE:
9585 case IORING_REGISTER_PROBE:
9586 case IORING_REGISTER_PERSONALITY:
9587 case IORING_UNREGISTER_PERSONALITY:
9594 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9595 void __user *arg, unsigned nr_args)
9596 __releases(ctx->uring_lock)
9597 __acquires(ctx->uring_lock)
9602 * We're inside the ring mutex, if the ref is already dying, then
9603 * someone else killed the ctx or is already going through
9604 * io_uring_register().
9606 if (percpu_ref_is_dying(&ctx->refs))
9609 if (io_register_op_must_quiesce(opcode)) {
9610 percpu_ref_kill(&ctx->refs);
9613 * Drop uring mutex before waiting for references to exit. If
9614 * another thread is currently inside io_uring_enter() it might
9615 * need to grab the uring_lock to make progress. If we hold it
9616 * here across the drain wait, then we can deadlock. It's safe
9617 * to drop the mutex here, since no new references will come in
9618 * after we've killed the percpu ref.
9620 mutex_unlock(&ctx->uring_lock);
9622 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9625 ret = io_run_task_work_sig();
9630 mutex_lock(&ctx->uring_lock);
9633 percpu_ref_resurrect(&ctx->refs);
9638 if (ctx->restricted) {
9639 if (opcode >= IORING_REGISTER_LAST) {
9644 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9651 case IORING_REGISTER_BUFFERS:
9652 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9654 case IORING_UNREGISTER_BUFFERS:
9658 ret = io_sqe_buffer_unregister(ctx);
9660 case IORING_REGISTER_FILES:
9661 ret = io_sqe_files_register(ctx, arg, nr_args);
9663 case IORING_UNREGISTER_FILES:
9667 ret = io_sqe_files_unregister(ctx);
9669 case IORING_REGISTER_FILES_UPDATE:
9670 ret = io_sqe_files_update(ctx, arg, nr_args);
9672 case IORING_REGISTER_EVENTFD:
9673 case IORING_REGISTER_EVENTFD_ASYNC:
9677 ret = io_eventfd_register(ctx, arg);
9680 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9681 ctx->eventfd_async = 1;
9683 ctx->eventfd_async = 0;
9685 case IORING_UNREGISTER_EVENTFD:
9689 ret = io_eventfd_unregister(ctx);
9691 case IORING_REGISTER_PROBE:
9693 if (!arg || nr_args > 256)
9695 ret = io_probe(ctx, arg, nr_args);
9697 case IORING_REGISTER_PERSONALITY:
9701 ret = io_register_personality(ctx);
9703 case IORING_UNREGISTER_PERSONALITY:
9707 ret = io_unregister_personality(ctx, nr_args);
9709 case IORING_REGISTER_ENABLE_RINGS:
9713 ret = io_register_enable_rings(ctx);
9715 case IORING_REGISTER_RESTRICTIONS:
9716 ret = io_register_restrictions(ctx, arg, nr_args);
9724 if (io_register_op_must_quiesce(opcode)) {
9725 /* bring the ctx back to life */
9726 percpu_ref_reinit(&ctx->refs);
9728 reinit_completion(&ctx->ref_comp);
9733 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9734 void __user *, arg, unsigned int, nr_args)
9736 struct io_ring_ctx *ctx;
9745 if (f.file->f_op != &io_uring_fops)
9748 ctx = f.file->private_data;
9750 mutex_lock(&ctx->uring_lock);
9751 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9752 mutex_unlock(&ctx->uring_lock);
9753 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9754 ctx->cq_ev_fd != NULL, ret);
9760 static int __init io_uring_init(void)
9762 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9763 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9764 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9767 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9768 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9769 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9770 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9771 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9772 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9773 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9774 BUILD_BUG_SQE_ELEM(8, __u64, off);
9775 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9776 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9777 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9778 BUILD_BUG_SQE_ELEM(24, __u32, len);
9779 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9780 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9781 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9782 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9783 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9784 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9785 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9786 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9787 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9788 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9789 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9790 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9791 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9792 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9793 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9794 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9795 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9796 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9797 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9799 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9800 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9801 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9804 __initcall(io_uring_init);