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;
1316 #ifdef CONFIG_BLK_CGROUP
1317 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1318 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1320 if (id->blkcg_css != blkcg_css()) {
1325 * This should be rare, either the cgroup is dying or the task
1326 * is moving cgroups. Just punt to root for the handful of ios.
1328 if (css_tryget_online(id->blkcg_css))
1329 req->work.flags |= IO_WQ_WORK_BLKCG;
1333 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1334 if (id->creds != current_cred())
1336 get_cred(id->creds);
1337 req->work.flags |= IO_WQ_WORK_CREDS;
1340 if (!uid_eq(current->loginuid, id->loginuid) ||
1341 current->sessionid != id->sessionid)
1344 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1345 (def->work_flags & IO_WQ_WORK_FS)) {
1346 if (current->fs != id->fs)
1348 spin_lock(&id->fs->lock);
1349 if (!id->fs->in_exec) {
1351 req->work.flags |= IO_WQ_WORK_FS;
1353 req->work.flags |= IO_WQ_WORK_CANCEL;
1355 spin_unlock(¤t->fs->lock);
1357 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1358 (def->work_flags & IO_WQ_WORK_FILES) &&
1359 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1360 if (id->files != current->files ||
1361 id->nsproxy != current->nsproxy)
1363 atomic_inc(&id->files->count);
1364 get_nsproxy(id->nsproxy);
1365 req->flags |= REQ_F_INFLIGHT;
1367 spin_lock_irq(&ctx->inflight_lock);
1368 list_add(&req->inflight_entry, &ctx->inflight_list);
1369 spin_unlock_irq(&ctx->inflight_lock);
1370 req->work.flags |= IO_WQ_WORK_FILES;
1376 static void io_prep_async_work(struct io_kiocb *req)
1378 const struct io_op_def *def = &io_op_defs[req->opcode];
1379 struct io_ring_ctx *ctx = req->ctx;
1380 struct io_identity *id;
1382 io_req_init_async(req);
1383 id = req->work.identity;
1385 if (req->flags & REQ_F_FORCE_ASYNC)
1386 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1388 if (req->flags & REQ_F_ISREG) {
1389 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1390 io_wq_hash_work(&req->work, file_inode(req->file));
1392 if (def->unbound_nonreg_file)
1393 req->work.flags |= IO_WQ_WORK_UNBOUND;
1396 /* ->mm can never change on us */
1397 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1398 (def->work_flags & IO_WQ_WORK_MM)) {
1400 req->work.flags |= IO_WQ_WORK_MM;
1403 /* if we fail grabbing identity, we must COW, regrab, and retry */
1404 if (io_grab_identity(req))
1407 if (!io_identity_cow(req))
1410 /* can't fail at this point */
1411 if (!io_grab_identity(req))
1415 static void io_prep_async_link(struct io_kiocb *req)
1417 struct io_kiocb *cur;
1419 io_prep_async_work(req);
1420 if (req->flags & REQ_F_LINK_HEAD)
1421 list_for_each_entry(cur, &req->link_list, link_list)
1422 io_prep_async_work(cur);
1425 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1427 struct io_ring_ctx *ctx = req->ctx;
1428 struct io_kiocb *link = io_prep_linked_timeout(req);
1430 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1431 &req->work, req->flags);
1432 io_wq_enqueue(ctx->io_wq, &req->work);
1436 static void io_queue_async_work(struct io_kiocb *req)
1438 struct io_kiocb *link;
1440 /* init ->work of the whole link before punting */
1441 io_prep_async_link(req);
1442 link = __io_queue_async_work(req);
1445 io_queue_linked_timeout(link);
1448 static void io_kill_timeout(struct io_kiocb *req)
1450 struct io_timeout_data *io = req->async_data;
1453 ret = hrtimer_try_to_cancel(&io->timer);
1455 atomic_set(&req->ctx->cq_timeouts,
1456 atomic_read(&req->ctx->cq_timeouts) + 1);
1457 list_del_init(&req->timeout.list);
1458 io_cqring_fill_event(req, 0);
1459 io_put_req_deferred(req, 1);
1463 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1465 struct io_ring_ctx *ctx = req->ctx;
1467 if (!tsk || req->task == tsk)
1469 if (ctx->flags & IORING_SETUP_SQPOLL) {
1470 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1477 * Returns true if we found and killed one or more timeouts
1479 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1481 struct io_kiocb *req, *tmp;
1484 spin_lock_irq(&ctx->completion_lock);
1485 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1486 if (io_task_match(req, tsk)) {
1487 io_kill_timeout(req);
1491 spin_unlock_irq(&ctx->completion_lock);
1492 return canceled != 0;
1495 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1498 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1499 struct io_defer_entry, list);
1500 struct io_kiocb *link;
1502 if (req_need_defer(de->req, de->seq))
1504 list_del_init(&de->list);
1505 /* punt-init is done before queueing for defer */
1506 link = __io_queue_async_work(de->req);
1508 __io_queue_linked_timeout(link);
1509 /* drop submission reference */
1510 io_put_req_deferred(link, 1);
1513 } while (!list_empty(&ctx->defer_list));
1516 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1518 while (!list_empty(&ctx->timeout_list)) {
1519 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1520 struct io_kiocb, timeout.list);
1522 if (io_is_timeout_noseq(req))
1524 if (req->timeout.target_seq != ctx->cached_cq_tail
1525 - atomic_read(&ctx->cq_timeouts))
1528 list_del_init(&req->timeout.list);
1529 io_kill_timeout(req);
1533 static void io_commit_cqring(struct io_ring_ctx *ctx)
1535 io_flush_timeouts(ctx);
1536 __io_commit_cqring(ctx);
1538 if (unlikely(!list_empty(&ctx->defer_list)))
1539 __io_queue_deferred(ctx);
1542 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1544 struct io_rings *r = ctx->rings;
1546 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1549 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1551 struct io_rings *rings = ctx->rings;
1554 tail = ctx->cached_cq_tail;
1556 * writes to the cq entry need to come after reading head; the
1557 * control dependency is enough as we're using WRITE_ONCE to
1560 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1563 ctx->cached_cq_tail++;
1564 return &rings->cqes[tail & ctx->cq_mask];
1567 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1571 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1573 if (!ctx->eventfd_async)
1575 return io_wq_current_is_worker();
1578 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1580 if (waitqueue_active(&ctx->wait))
1581 wake_up(&ctx->wait);
1582 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1583 wake_up(&ctx->sq_data->wait);
1584 if (io_should_trigger_evfd(ctx))
1585 eventfd_signal(ctx->cq_ev_fd, 1);
1588 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1590 if (list_empty(&ctx->cq_overflow_list)) {
1591 clear_bit(0, &ctx->sq_check_overflow);
1592 clear_bit(0, &ctx->cq_check_overflow);
1593 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1597 static inline bool __io_match_files(struct io_kiocb *req,
1598 struct files_struct *files)
1600 return ((req->flags & REQ_F_WORK_INITIALIZED) &&
1601 (req->work.flags & IO_WQ_WORK_FILES)) &&
1602 req->work.identity->files == files;
1605 static bool io_match_files(struct io_kiocb *req,
1606 struct files_struct *files)
1608 struct io_kiocb *link;
1612 if (__io_match_files(req, files))
1614 if (req->flags & REQ_F_LINK_HEAD) {
1615 list_for_each_entry(link, &req->link_list, link_list) {
1616 if (__io_match_files(link, files))
1623 /* Returns true if there are no backlogged entries after the flush */
1624 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1625 struct task_struct *tsk,
1626 struct files_struct *files)
1628 struct io_rings *rings = ctx->rings;
1629 struct io_kiocb *req, *tmp;
1630 struct io_uring_cqe *cqe;
1631 unsigned long flags;
1635 if (list_empty_careful(&ctx->cq_overflow_list))
1637 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1638 rings->cq_ring_entries))
1642 spin_lock_irqsave(&ctx->completion_lock, flags);
1644 /* if force is set, the ring is going away. always drop after that */
1646 ctx->cq_overflow_flushed = 1;
1649 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1650 if (tsk && req->task != tsk)
1652 if (!io_match_files(req, files))
1655 cqe = io_get_cqring(ctx);
1659 list_move(&req->compl.list, &list);
1661 WRITE_ONCE(cqe->user_data, req->user_data);
1662 WRITE_ONCE(cqe->res, req->result);
1663 WRITE_ONCE(cqe->flags, req->compl.cflags);
1665 ctx->cached_cq_overflow++;
1666 WRITE_ONCE(ctx->rings->cq_overflow,
1667 ctx->cached_cq_overflow);
1671 io_commit_cqring(ctx);
1672 io_cqring_mark_overflow(ctx);
1674 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1675 io_cqring_ev_posted(ctx);
1677 while (!list_empty(&list)) {
1678 req = list_first_entry(&list, struct io_kiocb, compl.list);
1679 list_del(&req->compl.list);
1686 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1688 struct io_ring_ctx *ctx = req->ctx;
1689 struct io_uring_cqe *cqe;
1691 trace_io_uring_complete(ctx, req->user_data, res);
1694 * If we can't get a cq entry, userspace overflowed the
1695 * submission (by quite a lot). Increment the overflow count in
1698 cqe = io_get_cqring(ctx);
1700 WRITE_ONCE(cqe->user_data, req->user_data);
1701 WRITE_ONCE(cqe->res, res);
1702 WRITE_ONCE(cqe->flags, cflags);
1703 } else if (ctx->cq_overflow_flushed ||
1704 atomic_read(&req->task->io_uring->in_idle)) {
1706 * If we're in ring overflow flush mode, or in task cancel mode,
1707 * then we cannot store the request for later flushing, we need
1708 * to drop it on the floor.
1710 ctx->cached_cq_overflow++;
1711 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1713 if (list_empty(&ctx->cq_overflow_list)) {
1714 set_bit(0, &ctx->sq_check_overflow);
1715 set_bit(0, &ctx->cq_check_overflow);
1716 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1720 req->compl.cflags = cflags;
1721 refcount_inc(&req->refs);
1722 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1726 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1728 __io_cqring_fill_event(req, res, 0);
1731 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1733 struct io_ring_ctx *ctx = req->ctx;
1734 unsigned long flags;
1736 spin_lock_irqsave(&ctx->completion_lock, flags);
1737 __io_cqring_fill_event(req, res, cflags);
1738 io_commit_cqring(ctx);
1739 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1741 io_cqring_ev_posted(ctx);
1744 static void io_submit_flush_completions(struct io_comp_state *cs)
1746 struct io_ring_ctx *ctx = cs->ctx;
1748 spin_lock_irq(&ctx->completion_lock);
1749 while (!list_empty(&cs->list)) {
1750 struct io_kiocb *req;
1752 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1753 list_del(&req->compl.list);
1754 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1757 * io_free_req() doesn't care about completion_lock unless one
1758 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1759 * because of a potential deadlock with req->work.fs->lock
1761 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1762 |REQ_F_WORK_INITIALIZED)) {
1763 spin_unlock_irq(&ctx->completion_lock);
1765 spin_lock_irq(&ctx->completion_lock);
1770 io_commit_cqring(ctx);
1771 spin_unlock_irq(&ctx->completion_lock);
1773 io_cqring_ev_posted(ctx);
1777 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1778 struct io_comp_state *cs)
1781 io_cqring_add_event(req, res, cflags);
1786 req->compl.cflags = cflags;
1787 list_add_tail(&req->compl.list, &cs->list);
1789 io_submit_flush_completions(cs);
1793 static void io_req_complete(struct io_kiocb *req, long res)
1795 __io_req_complete(req, res, 0, NULL);
1798 static inline bool io_is_fallback_req(struct io_kiocb *req)
1800 return req == (struct io_kiocb *)
1801 ((unsigned long) req->ctx->fallback_req & ~1UL);
1804 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1806 struct io_kiocb *req;
1808 req = ctx->fallback_req;
1809 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1815 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1816 struct io_submit_state *state)
1818 if (!state->free_reqs) {
1819 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1823 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1824 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1827 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1828 * retry single alloc to be on the safe side.
1830 if (unlikely(ret <= 0)) {
1831 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1832 if (!state->reqs[0])
1836 state->free_reqs = ret;
1840 return state->reqs[state->free_reqs];
1842 return io_get_fallback_req(ctx);
1845 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1849 percpu_ref_put(req->fixed_file_refs);
1854 static void io_dismantle_req(struct io_kiocb *req)
1858 if (req->async_data)
1859 kfree(req->async_data);
1861 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1863 io_req_clean_work(req);
1866 static void __io_free_req(struct io_kiocb *req)
1868 struct io_uring_task *tctx = req->task->io_uring;
1869 struct io_ring_ctx *ctx = req->ctx;
1871 io_dismantle_req(req);
1873 percpu_counter_dec(&tctx->inflight);
1874 if (atomic_read(&tctx->in_idle))
1875 wake_up(&tctx->wait);
1876 put_task_struct(req->task);
1878 if (likely(!io_is_fallback_req(req)))
1879 kmem_cache_free(req_cachep, req);
1881 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1882 percpu_ref_put(&ctx->refs);
1885 static void io_kill_linked_timeout(struct io_kiocb *req)
1887 struct io_ring_ctx *ctx = req->ctx;
1888 struct io_kiocb *link;
1889 bool cancelled = false;
1890 unsigned long flags;
1892 spin_lock_irqsave(&ctx->completion_lock, flags);
1893 link = list_first_entry_or_null(&req->link_list, struct io_kiocb,
1896 * Can happen if a linked timeout fired and link had been like
1897 * req -> link t-out -> link t-out [-> ...]
1899 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1900 struct io_timeout_data *io = link->async_data;
1903 list_del_init(&link->link_list);
1904 ret = hrtimer_try_to_cancel(&io->timer);
1906 io_cqring_fill_event(link, -ECANCELED);
1907 io_commit_cqring(ctx);
1911 req->flags &= ~REQ_F_LINK_TIMEOUT;
1912 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1915 io_cqring_ev_posted(ctx);
1920 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1922 struct io_kiocb *nxt;
1925 * The list should never be empty when we are called here. But could
1926 * potentially happen if the chain is messed up, check to be on the
1929 if (unlikely(list_empty(&req->link_list)))
1932 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1933 list_del_init(&req->link_list);
1934 if (!list_empty(&nxt->link_list))
1935 nxt->flags |= REQ_F_LINK_HEAD;
1940 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1942 static void io_fail_links(struct io_kiocb *req)
1944 struct io_ring_ctx *ctx = req->ctx;
1945 unsigned long flags;
1947 spin_lock_irqsave(&ctx->completion_lock, flags);
1948 while (!list_empty(&req->link_list)) {
1949 struct io_kiocb *link = list_first_entry(&req->link_list,
1950 struct io_kiocb, link_list);
1952 list_del_init(&link->link_list);
1953 trace_io_uring_fail_link(req, link);
1955 io_cqring_fill_event(link, -ECANCELED);
1958 * It's ok to free under spinlock as they're not linked anymore,
1959 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1962 if (link->flags & REQ_F_WORK_INITIALIZED)
1963 io_put_req_deferred(link, 2);
1965 io_double_put_req(link);
1968 io_commit_cqring(ctx);
1969 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1971 io_cqring_ev_posted(ctx);
1974 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1976 req->flags &= ~REQ_F_LINK_HEAD;
1977 if (req->flags & REQ_F_LINK_TIMEOUT)
1978 io_kill_linked_timeout(req);
1981 * If LINK is set, we have dependent requests in this chain. If we
1982 * didn't fail this request, queue the first one up, moving any other
1983 * dependencies to the next request. In case of failure, fail the rest
1986 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1987 return io_req_link_next(req);
1992 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1994 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1996 return __io_req_find_next(req);
1999 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
2001 struct task_struct *tsk = req->task;
2002 struct io_ring_ctx *ctx = req->ctx;
2003 enum task_work_notify_mode notify;
2006 if (tsk->flags & PF_EXITING)
2010 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2011 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2012 * processing task_work. There's no reliable way to tell if TWA_RESUME
2016 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
2017 notify = TWA_SIGNAL;
2019 ret = task_work_add(tsk, &req->task_work, notify);
2021 wake_up_process(tsk);
2026 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2028 struct io_ring_ctx *ctx = req->ctx;
2030 spin_lock_irq(&ctx->completion_lock);
2031 io_cqring_fill_event(req, error);
2032 io_commit_cqring(ctx);
2033 spin_unlock_irq(&ctx->completion_lock);
2035 io_cqring_ev_posted(ctx);
2036 req_set_fail_links(req);
2037 io_double_put_req(req);
2040 static void io_req_task_cancel(struct callback_head *cb)
2042 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2043 struct io_ring_ctx *ctx = req->ctx;
2045 __io_req_task_cancel(req, -ECANCELED);
2046 percpu_ref_put(&ctx->refs);
2049 static void __io_req_task_submit(struct io_kiocb *req)
2051 struct io_ring_ctx *ctx = req->ctx;
2053 if (!__io_sq_thread_acquire_mm(ctx)) {
2054 mutex_lock(&ctx->uring_lock);
2055 __io_queue_sqe(req, NULL);
2056 mutex_unlock(&ctx->uring_lock);
2058 __io_req_task_cancel(req, -EFAULT);
2062 static void io_req_task_submit(struct callback_head *cb)
2064 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2065 struct io_ring_ctx *ctx = req->ctx;
2067 __io_req_task_submit(req);
2068 percpu_ref_put(&ctx->refs);
2071 static void io_req_task_queue(struct io_kiocb *req)
2075 init_task_work(&req->task_work, io_req_task_submit);
2076 percpu_ref_get(&req->ctx->refs);
2078 ret = io_req_task_work_add(req, true);
2079 if (unlikely(ret)) {
2080 struct task_struct *tsk;
2082 init_task_work(&req->task_work, io_req_task_cancel);
2083 tsk = io_wq_get_task(req->ctx->io_wq);
2084 task_work_add(tsk, &req->task_work, TWA_NONE);
2085 wake_up_process(tsk);
2089 static void io_queue_next(struct io_kiocb *req)
2091 struct io_kiocb *nxt = io_req_find_next(req);
2094 io_req_task_queue(nxt);
2097 static void io_free_req(struct io_kiocb *req)
2104 void *reqs[IO_IOPOLL_BATCH];
2107 struct task_struct *task;
2111 static inline void io_init_req_batch(struct req_batch *rb)
2118 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2119 struct req_batch *rb)
2121 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2122 percpu_ref_put_many(&ctx->refs, rb->to_free);
2126 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2127 struct req_batch *rb)
2130 __io_req_free_batch_flush(ctx, rb);
2132 struct io_uring_task *tctx = rb->task->io_uring;
2134 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2135 put_task_struct_many(rb->task, rb->task_refs);
2140 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2142 if (unlikely(io_is_fallback_req(req))) {
2146 if (req->flags & REQ_F_LINK_HEAD)
2149 if (req->task != rb->task) {
2151 struct io_uring_task *tctx = rb->task->io_uring;
2153 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2154 put_task_struct_many(rb->task, rb->task_refs);
2156 rb->task = req->task;
2161 io_dismantle_req(req);
2162 rb->reqs[rb->to_free++] = req;
2163 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2164 __io_req_free_batch_flush(req->ctx, rb);
2168 * Drop reference to request, return next in chain (if there is one) if this
2169 * was the last reference to this request.
2171 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2173 struct io_kiocb *nxt = NULL;
2175 if (refcount_dec_and_test(&req->refs)) {
2176 nxt = io_req_find_next(req);
2182 static void io_put_req(struct io_kiocb *req)
2184 if (refcount_dec_and_test(&req->refs))
2188 static void io_put_req_deferred_cb(struct callback_head *cb)
2190 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2195 static void io_free_req_deferred(struct io_kiocb *req)
2199 init_task_work(&req->task_work, io_put_req_deferred_cb);
2200 ret = io_req_task_work_add(req, true);
2201 if (unlikely(ret)) {
2202 struct task_struct *tsk;
2204 tsk = io_wq_get_task(req->ctx->io_wq);
2205 task_work_add(tsk, &req->task_work, TWA_NONE);
2206 wake_up_process(tsk);
2210 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2212 if (refcount_sub_and_test(refs, &req->refs))
2213 io_free_req_deferred(req);
2216 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2218 struct io_kiocb *nxt;
2221 * A ref is owned by io-wq in which context we're. So, if that's the
2222 * last one, it's safe to steal next work. False negatives are Ok,
2223 * it just will be re-punted async in io_put_work()
2225 if (refcount_read(&req->refs) != 1)
2228 nxt = io_req_find_next(req);
2229 return nxt ? &nxt->work : NULL;
2232 static void io_double_put_req(struct io_kiocb *req)
2234 /* drop both submit and complete references */
2235 if (refcount_sub_and_test(2, &req->refs))
2239 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2241 struct io_rings *rings = ctx->rings;
2243 if (test_bit(0, &ctx->cq_check_overflow)) {
2245 * noflush == true is from the waitqueue handler, just ensure
2246 * we wake up the task, and the next invocation will flush the
2247 * entries. We cannot safely to it from here.
2249 if (noflush && !list_empty(&ctx->cq_overflow_list))
2252 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2255 /* See comment at the top of this file */
2257 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2260 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2262 struct io_rings *rings = ctx->rings;
2264 /* make sure SQ entry isn't read before tail */
2265 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2268 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2270 unsigned int cflags;
2272 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2273 cflags |= IORING_CQE_F_BUFFER;
2274 req->flags &= ~REQ_F_BUFFER_SELECTED;
2279 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2281 struct io_buffer *kbuf;
2283 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2284 return io_put_kbuf(req, kbuf);
2287 static inline bool io_run_task_work(void)
2290 * Not safe to run on exiting task, and the task_work handling will
2291 * not add work to such a task.
2293 if (unlikely(current->flags & PF_EXITING))
2295 if (current->task_works) {
2296 __set_current_state(TASK_RUNNING);
2304 static void io_iopoll_queue(struct list_head *again)
2306 struct io_kiocb *req;
2309 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2310 list_del(&req->inflight_entry);
2311 __io_complete_rw(req, -EAGAIN, 0, NULL);
2312 } while (!list_empty(again));
2316 * Find and free completed poll iocbs
2318 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2319 struct list_head *done)
2321 struct req_batch rb;
2322 struct io_kiocb *req;
2325 /* order with ->result store in io_complete_rw_iopoll() */
2328 io_init_req_batch(&rb);
2329 while (!list_empty(done)) {
2332 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2333 if (READ_ONCE(req->result) == -EAGAIN) {
2335 req->iopoll_completed = 0;
2336 list_move_tail(&req->inflight_entry, &again);
2339 list_del(&req->inflight_entry);
2341 if (req->flags & REQ_F_BUFFER_SELECTED)
2342 cflags = io_put_rw_kbuf(req);
2344 __io_cqring_fill_event(req, req->result, cflags);
2347 if (refcount_dec_and_test(&req->refs))
2348 io_req_free_batch(&rb, req);
2351 io_commit_cqring(ctx);
2352 if (ctx->flags & IORING_SETUP_SQPOLL)
2353 io_cqring_ev_posted(ctx);
2354 io_req_free_batch_finish(ctx, &rb);
2356 if (!list_empty(&again))
2357 io_iopoll_queue(&again);
2360 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2363 struct io_kiocb *req, *tmp;
2369 * Only spin for completions if we don't have multiple devices hanging
2370 * off our complete list, and we're under the requested amount.
2372 spin = !ctx->poll_multi_file && *nr_events < min;
2375 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2376 struct kiocb *kiocb = &req->rw.kiocb;
2379 * Move completed and retryable entries to our local lists.
2380 * If we find a request that requires polling, break out
2381 * and complete those lists first, if we have entries there.
2383 if (READ_ONCE(req->iopoll_completed)) {
2384 list_move_tail(&req->inflight_entry, &done);
2387 if (!list_empty(&done))
2390 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2394 /* iopoll may have completed current req */
2395 if (READ_ONCE(req->iopoll_completed))
2396 list_move_tail(&req->inflight_entry, &done);
2403 if (!list_empty(&done))
2404 io_iopoll_complete(ctx, nr_events, &done);
2410 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2411 * non-spinning poll check - we'll still enter the driver poll loop, but only
2412 * as a non-spinning completion check.
2414 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2417 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2420 ret = io_do_iopoll(ctx, nr_events, min);
2423 if (*nr_events >= min)
2431 * We can't just wait for polled events to come to us, we have to actively
2432 * find and complete them.
2434 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2436 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2439 mutex_lock(&ctx->uring_lock);
2440 while (!list_empty(&ctx->iopoll_list)) {
2441 unsigned int nr_events = 0;
2443 io_do_iopoll(ctx, &nr_events, 0);
2445 /* let it sleep and repeat later if can't complete a request */
2449 * Ensure we allow local-to-the-cpu processing to take place,
2450 * in this case we need to ensure that we reap all events.
2451 * Also let task_work, etc. to progress by releasing the mutex
2453 if (need_resched()) {
2454 mutex_unlock(&ctx->uring_lock);
2456 mutex_lock(&ctx->uring_lock);
2459 mutex_unlock(&ctx->uring_lock);
2462 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2464 unsigned int nr_events = 0;
2465 int iters = 0, ret = 0;
2468 * We disallow the app entering submit/complete with polling, but we
2469 * still need to lock the ring to prevent racing with polled issue
2470 * that got punted to a workqueue.
2472 mutex_lock(&ctx->uring_lock);
2475 * Don't enter poll loop if we already have events pending.
2476 * If we do, we can potentially be spinning for commands that
2477 * already triggered a CQE (eg in error).
2479 if (io_cqring_events(ctx, false))
2483 * If a submit got punted to a workqueue, we can have the
2484 * application entering polling for a command before it gets
2485 * issued. That app will hold the uring_lock for the duration
2486 * of the poll right here, so we need to take a breather every
2487 * now and then to ensure that the issue has a chance to add
2488 * the poll to the issued list. Otherwise we can spin here
2489 * forever, while the workqueue is stuck trying to acquire the
2492 if (!(++iters & 7)) {
2493 mutex_unlock(&ctx->uring_lock);
2495 mutex_lock(&ctx->uring_lock);
2498 ret = io_iopoll_getevents(ctx, &nr_events, min);
2502 } while (min && !nr_events && !need_resched());
2504 mutex_unlock(&ctx->uring_lock);
2508 static void kiocb_end_write(struct io_kiocb *req)
2511 * Tell lockdep we inherited freeze protection from submission
2514 if (req->flags & REQ_F_ISREG) {
2515 struct inode *inode = file_inode(req->file);
2517 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2519 file_end_write(req->file);
2522 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2523 struct io_comp_state *cs)
2525 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2528 if (kiocb->ki_flags & IOCB_WRITE)
2529 kiocb_end_write(req);
2531 if (res != req->result)
2532 req_set_fail_links(req);
2533 if (req->flags & REQ_F_BUFFER_SELECTED)
2534 cflags = io_put_rw_kbuf(req);
2535 __io_req_complete(req, res, cflags, cs);
2539 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2541 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2542 ssize_t ret = -ECANCELED;
2543 struct iov_iter iter;
2551 switch (req->opcode) {
2552 case IORING_OP_READV:
2553 case IORING_OP_READ_FIXED:
2554 case IORING_OP_READ:
2557 case IORING_OP_WRITEV:
2558 case IORING_OP_WRITE_FIXED:
2559 case IORING_OP_WRITE:
2563 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2568 if (!req->async_data) {
2569 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2572 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2580 req_set_fail_links(req);
2585 static bool io_rw_reissue(struct io_kiocb *req, long res)
2588 umode_t mode = file_inode(req->file)->i_mode;
2591 if (!S_ISBLK(mode) && !S_ISREG(mode))
2593 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2596 ret = io_sq_thread_acquire_mm(req->ctx, req);
2598 if (io_resubmit_prep(req, ret)) {
2599 refcount_inc(&req->refs);
2600 io_queue_async_work(req);
2608 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2609 struct io_comp_state *cs)
2611 if (!io_rw_reissue(req, res))
2612 io_complete_rw_common(&req->rw.kiocb, res, cs);
2615 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2617 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2619 __io_complete_rw(req, res, res2, NULL);
2622 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2624 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2626 if (kiocb->ki_flags & IOCB_WRITE)
2627 kiocb_end_write(req);
2629 if (res != -EAGAIN && res != req->result)
2630 req_set_fail_links(req);
2632 WRITE_ONCE(req->result, res);
2633 /* order with io_poll_complete() checking ->result */
2635 WRITE_ONCE(req->iopoll_completed, 1);
2639 * After the iocb has been issued, it's safe to be found on the poll list.
2640 * Adding the kiocb to the list AFTER submission ensures that we don't
2641 * find it from a io_iopoll_getevents() thread before the issuer is done
2642 * accessing the kiocb cookie.
2644 static void io_iopoll_req_issued(struct io_kiocb *req)
2646 struct io_ring_ctx *ctx = req->ctx;
2649 * Track whether we have multiple files in our lists. This will impact
2650 * how we do polling eventually, not spinning if we're on potentially
2651 * different devices.
2653 if (list_empty(&ctx->iopoll_list)) {
2654 ctx->poll_multi_file = false;
2655 } else if (!ctx->poll_multi_file) {
2656 struct io_kiocb *list_req;
2658 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2660 if (list_req->file != req->file)
2661 ctx->poll_multi_file = true;
2665 * For fast devices, IO may have already completed. If it has, add
2666 * it to the front so we find it first.
2668 if (READ_ONCE(req->iopoll_completed))
2669 list_add(&req->inflight_entry, &ctx->iopoll_list);
2671 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2673 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2674 wq_has_sleeper(&ctx->sq_data->wait))
2675 wake_up(&ctx->sq_data->wait);
2678 static void __io_state_file_put(struct io_submit_state *state)
2680 if (state->has_refs)
2681 fput_many(state->file, state->has_refs);
2685 static inline void io_state_file_put(struct io_submit_state *state)
2688 __io_state_file_put(state);
2692 * Get as many references to a file as we have IOs left in this submission,
2693 * assuming most submissions are for one file, or at least that each file
2694 * has more than one submission.
2696 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2702 if (state->fd == fd) {
2706 __io_state_file_put(state);
2708 state->file = fget_many(fd, state->ios_left);
2713 state->has_refs = state->ios_left - 1;
2717 static bool io_bdev_nowait(struct block_device *bdev)
2720 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2727 * If we tracked the file through the SCM inflight mechanism, we could support
2728 * any file. For now, just ensure that anything potentially problematic is done
2731 static bool io_file_supports_async(struct file *file, int rw)
2733 umode_t mode = file_inode(file)->i_mode;
2735 if (S_ISBLK(mode)) {
2736 if (io_bdev_nowait(file->f_inode->i_bdev))
2740 if (S_ISCHR(mode) || S_ISSOCK(mode))
2742 if (S_ISREG(mode)) {
2743 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2744 file->f_op != &io_uring_fops)
2749 /* any ->read/write should understand O_NONBLOCK */
2750 if (file->f_flags & O_NONBLOCK)
2753 if (!(file->f_mode & FMODE_NOWAIT))
2757 return file->f_op->read_iter != NULL;
2759 return file->f_op->write_iter != NULL;
2762 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2764 struct io_ring_ctx *ctx = req->ctx;
2765 struct kiocb *kiocb = &req->rw.kiocb;
2769 if (S_ISREG(file_inode(req->file)->i_mode))
2770 req->flags |= REQ_F_ISREG;
2772 kiocb->ki_pos = READ_ONCE(sqe->off);
2773 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2774 req->flags |= REQ_F_CUR_POS;
2775 kiocb->ki_pos = req->file->f_pos;
2777 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2778 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2779 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2783 ioprio = READ_ONCE(sqe->ioprio);
2785 ret = ioprio_check_cap(ioprio);
2789 kiocb->ki_ioprio = ioprio;
2791 kiocb->ki_ioprio = get_current_ioprio();
2793 /* don't allow async punt if RWF_NOWAIT was requested */
2794 if (kiocb->ki_flags & IOCB_NOWAIT)
2795 req->flags |= REQ_F_NOWAIT;
2797 if (ctx->flags & IORING_SETUP_IOPOLL) {
2798 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2799 !kiocb->ki_filp->f_op->iopoll)
2802 kiocb->ki_flags |= IOCB_HIPRI;
2803 kiocb->ki_complete = io_complete_rw_iopoll;
2804 req->iopoll_completed = 0;
2806 if (kiocb->ki_flags & IOCB_HIPRI)
2808 kiocb->ki_complete = io_complete_rw;
2811 req->rw.addr = READ_ONCE(sqe->addr);
2812 req->rw.len = READ_ONCE(sqe->len);
2813 req->buf_index = READ_ONCE(sqe->buf_index);
2817 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2823 case -ERESTARTNOINTR:
2824 case -ERESTARTNOHAND:
2825 case -ERESTART_RESTARTBLOCK:
2827 * We can't just restart the syscall, since previously
2828 * submitted sqes may already be in progress. Just fail this
2834 kiocb->ki_complete(kiocb, ret, 0);
2838 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2839 struct io_comp_state *cs)
2841 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2842 struct io_async_rw *io = req->async_data;
2844 /* add previously done IO, if any */
2845 if (io && io->bytes_done > 0) {
2847 ret = io->bytes_done;
2849 ret += io->bytes_done;
2852 if (req->flags & REQ_F_CUR_POS)
2853 req->file->f_pos = kiocb->ki_pos;
2854 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2855 __io_complete_rw(req, ret, 0, cs);
2857 io_rw_done(kiocb, ret);
2860 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2861 struct iov_iter *iter)
2863 struct io_ring_ctx *ctx = req->ctx;
2864 size_t len = req->rw.len;
2865 struct io_mapped_ubuf *imu;
2866 u16 index, buf_index = req->buf_index;
2870 if (unlikely(buf_index >= ctx->nr_user_bufs))
2872 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2873 imu = &ctx->user_bufs[index];
2874 buf_addr = req->rw.addr;
2877 if (buf_addr + len < buf_addr)
2879 /* not inside the mapped region */
2880 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2884 * May not be a start of buffer, set size appropriately
2885 * and advance us to the beginning.
2887 offset = buf_addr - imu->ubuf;
2888 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2892 * Don't use iov_iter_advance() here, as it's really slow for
2893 * using the latter parts of a big fixed buffer - it iterates
2894 * over each segment manually. We can cheat a bit here, because
2897 * 1) it's a BVEC iter, we set it up
2898 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2899 * first and last bvec
2901 * So just find our index, and adjust the iterator afterwards.
2902 * If the offset is within the first bvec (or the whole first
2903 * bvec, just use iov_iter_advance(). This makes it easier
2904 * since we can just skip the first segment, which may not
2905 * be PAGE_SIZE aligned.
2907 const struct bio_vec *bvec = imu->bvec;
2909 if (offset <= bvec->bv_len) {
2910 iov_iter_advance(iter, offset);
2912 unsigned long seg_skip;
2914 /* skip first vec */
2915 offset -= bvec->bv_len;
2916 seg_skip = 1 + (offset >> PAGE_SHIFT);
2918 iter->bvec = bvec + seg_skip;
2919 iter->nr_segs -= seg_skip;
2920 iter->count -= bvec->bv_len + offset;
2921 iter->iov_offset = offset & ~PAGE_MASK;
2928 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2931 mutex_unlock(&ctx->uring_lock);
2934 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2937 * "Normal" inline submissions always hold the uring_lock, since we
2938 * grab it from the system call. Same is true for the SQPOLL offload.
2939 * The only exception is when we've detached the request and issue it
2940 * from an async worker thread, grab the lock for that case.
2943 mutex_lock(&ctx->uring_lock);
2946 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2947 int bgid, struct io_buffer *kbuf,
2950 struct io_buffer *head;
2952 if (req->flags & REQ_F_BUFFER_SELECTED)
2955 io_ring_submit_lock(req->ctx, needs_lock);
2957 lockdep_assert_held(&req->ctx->uring_lock);
2959 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2961 if (!list_empty(&head->list)) {
2962 kbuf = list_last_entry(&head->list, struct io_buffer,
2964 list_del(&kbuf->list);
2967 idr_remove(&req->ctx->io_buffer_idr, bgid);
2969 if (*len > kbuf->len)
2972 kbuf = ERR_PTR(-ENOBUFS);
2975 io_ring_submit_unlock(req->ctx, needs_lock);
2980 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2983 struct io_buffer *kbuf;
2986 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2987 bgid = req->buf_index;
2988 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2991 req->rw.addr = (u64) (unsigned long) kbuf;
2992 req->flags |= REQ_F_BUFFER_SELECTED;
2993 return u64_to_user_ptr(kbuf->addr);
2996 #ifdef CONFIG_COMPAT
2997 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3000 struct compat_iovec __user *uiov;
3001 compat_ssize_t clen;
3005 uiov = u64_to_user_ptr(req->rw.addr);
3006 if (!access_ok(uiov, sizeof(*uiov)))
3008 if (__get_user(clen, &uiov->iov_len))
3014 buf = io_rw_buffer_select(req, &len, needs_lock);
3016 return PTR_ERR(buf);
3017 iov[0].iov_base = buf;
3018 iov[0].iov_len = (compat_size_t) len;
3023 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3026 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3030 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3033 len = iov[0].iov_len;
3036 buf = io_rw_buffer_select(req, &len, needs_lock);
3038 return PTR_ERR(buf);
3039 iov[0].iov_base = buf;
3040 iov[0].iov_len = len;
3044 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3047 if (req->flags & REQ_F_BUFFER_SELECTED) {
3048 struct io_buffer *kbuf;
3050 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3051 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3052 iov[0].iov_len = kbuf->len;
3057 else if (req->rw.len > 1)
3060 #ifdef CONFIG_COMPAT
3061 if (req->ctx->compat)
3062 return io_compat_import(req, iov, needs_lock);
3065 return __io_iov_buffer_select(req, iov, needs_lock);
3068 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3069 struct iovec **iovec, struct iov_iter *iter,
3072 void __user *buf = u64_to_user_ptr(req->rw.addr);
3073 size_t sqe_len = req->rw.len;
3077 opcode = req->opcode;
3078 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3080 return io_import_fixed(req, rw, iter);
3083 /* buffer index only valid with fixed read/write, or buffer select */
3084 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3087 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3088 if (req->flags & REQ_F_BUFFER_SELECT) {
3089 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3091 return PTR_ERR(buf);
3092 req->rw.len = sqe_len;
3095 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3097 return ret < 0 ? ret : sqe_len;
3100 if (req->flags & REQ_F_BUFFER_SELECT) {
3101 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3103 ret = (*iovec)->iov_len;
3104 iov_iter_init(iter, rw, *iovec, 1, ret);
3110 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3114 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3115 struct iovec **iovec, struct iov_iter *iter,
3118 struct io_async_rw *iorw = req->async_data;
3121 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3123 return iov_iter_count(&iorw->iter);
3126 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3128 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3132 * For files that don't have ->read_iter() and ->write_iter(), handle them
3133 * by looping over ->read() or ->write() manually.
3135 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3137 struct kiocb *kiocb = &req->rw.kiocb;
3138 struct file *file = req->file;
3142 * Don't support polled IO through this interface, and we can't
3143 * support non-blocking either. For the latter, this just causes
3144 * the kiocb to be handled from an async context.
3146 if (kiocb->ki_flags & IOCB_HIPRI)
3148 if (kiocb->ki_flags & IOCB_NOWAIT)
3151 while (iov_iter_count(iter)) {
3155 if (!iov_iter_is_bvec(iter)) {
3156 iovec = iov_iter_iovec(iter);
3158 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3159 iovec.iov_len = req->rw.len;
3163 nr = file->f_op->read(file, iovec.iov_base,
3164 iovec.iov_len, io_kiocb_ppos(kiocb));
3166 nr = file->f_op->write(file, iovec.iov_base,
3167 iovec.iov_len, io_kiocb_ppos(kiocb));
3176 if (nr != iovec.iov_len)
3180 iov_iter_advance(iter, nr);
3186 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3187 const struct iovec *fast_iov, struct iov_iter *iter)
3189 struct io_async_rw *rw = req->async_data;
3191 memcpy(&rw->iter, iter, sizeof(*iter));
3192 rw->free_iovec = iovec;
3194 /* can only be fixed buffers, no need to do anything */
3195 if (iov_iter_is_bvec(iter))
3198 unsigned iov_off = 0;
3200 rw->iter.iov = rw->fast_iov;
3201 if (iter->iov != fast_iov) {
3202 iov_off = iter->iov - fast_iov;
3203 rw->iter.iov += iov_off;
3205 if (rw->fast_iov != fast_iov)
3206 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3207 sizeof(struct iovec) * iter->nr_segs);
3209 req->flags |= REQ_F_NEED_CLEANUP;
3213 static inline int __io_alloc_async_data(struct io_kiocb *req)
3215 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3216 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3217 return req->async_data == NULL;
3220 static int io_alloc_async_data(struct io_kiocb *req)
3222 if (!io_op_defs[req->opcode].needs_async_data)
3225 return __io_alloc_async_data(req);
3228 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3229 const struct iovec *fast_iov,
3230 struct iov_iter *iter, bool force)
3232 if (!force && !io_op_defs[req->opcode].needs_async_data)
3234 if (!req->async_data) {
3235 if (__io_alloc_async_data(req))
3238 io_req_map_rw(req, iovec, fast_iov, iter);
3243 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3245 struct io_async_rw *iorw = req->async_data;
3246 struct iovec *iov = iorw->fast_iov;
3249 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3250 if (unlikely(ret < 0))
3253 iorw->bytes_done = 0;
3254 iorw->free_iovec = iov;
3256 req->flags |= REQ_F_NEED_CLEANUP;
3260 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3264 ret = io_prep_rw(req, sqe);
3268 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3271 /* either don't need iovec imported or already have it */
3272 if (!req->async_data)
3274 return io_rw_prep_async(req, READ);
3278 * This is our waitqueue callback handler, registered through lock_page_async()
3279 * when we initially tried to do the IO with the iocb armed our waitqueue.
3280 * This gets called when the page is unlocked, and we generally expect that to
3281 * happen when the page IO is completed and the page is now uptodate. This will
3282 * queue a task_work based retry of the operation, attempting to copy the data
3283 * again. If the latter fails because the page was NOT uptodate, then we will
3284 * do a thread based blocking retry of the operation. That's the unexpected
3287 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3288 int sync, void *arg)
3290 struct wait_page_queue *wpq;
3291 struct io_kiocb *req = wait->private;
3292 struct wait_page_key *key = arg;
3295 wpq = container_of(wait, struct wait_page_queue, wait);
3297 if (!wake_page_match(wpq, key))
3300 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3301 list_del_init(&wait->entry);
3303 init_task_work(&req->task_work, io_req_task_submit);
3304 percpu_ref_get(&req->ctx->refs);
3306 /* submit ref gets dropped, acquire a new one */
3307 refcount_inc(&req->refs);
3308 ret = io_req_task_work_add(req, true);
3309 if (unlikely(ret)) {
3310 struct task_struct *tsk;
3312 /* queue just for cancelation */
3313 init_task_work(&req->task_work, io_req_task_cancel);
3314 tsk = io_wq_get_task(req->ctx->io_wq);
3315 task_work_add(tsk, &req->task_work, TWA_NONE);
3316 wake_up_process(tsk);
3322 * This controls whether a given IO request should be armed for async page
3323 * based retry. If we return false here, the request is handed to the async
3324 * worker threads for retry. If we're doing buffered reads on a regular file,
3325 * we prepare a private wait_page_queue entry and retry the operation. This
3326 * will either succeed because the page is now uptodate and unlocked, or it
3327 * will register a callback when the page is unlocked at IO completion. Through
3328 * that callback, io_uring uses task_work to setup a retry of the operation.
3329 * That retry will attempt the buffered read again. The retry will generally
3330 * succeed, or in rare cases where it fails, we then fall back to using the
3331 * async worker threads for a blocking retry.
3333 static bool io_rw_should_retry(struct io_kiocb *req)
3335 struct io_async_rw *rw = req->async_data;
3336 struct wait_page_queue *wait = &rw->wpq;
3337 struct kiocb *kiocb = &req->rw.kiocb;
3339 /* never retry for NOWAIT, we just complete with -EAGAIN */
3340 if (req->flags & REQ_F_NOWAIT)
3343 /* Only for buffered IO */
3344 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3348 * just use poll if we can, and don't attempt if the fs doesn't
3349 * support callback based unlocks
3351 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3354 wait->wait.func = io_async_buf_func;
3355 wait->wait.private = req;
3356 wait->wait.flags = 0;
3357 INIT_LIST_HEAD(&wait->wait.entry);
3358 kiocb->ki_flags |= IOCB_WAITQ;
3359 kiocb->ki_flags &= ~IOCB_NOWAIT;
3360 kiocb->ki_waitq = wait;
3364 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3366 if (req->file->f_op->read_iter)
3367 return call_read_iter(req->file, &req->rw.kiocb, iter);
3368 else if (req->file->f_op->read)
3369 return loop_rw_iter(READ, req, iter);
3374 static int io_read(struct io_kiocb *req, bool force_nonblock,
3375 struct io_comp_state *cs)
3377 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3378 struct kiocb *kiocb = &req->rw.kiocb;
3379 struct iov_iter __iter, *iter = &__iter;
3380 struct io_async_rw *rw = req->async_data;
3381 ssize_t io_size, ret, ret2;
3388 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3391 iov_count = iov_iter_count(iter);
3393 req->result = io_size;
3396 /* Ensure we clear previously set non-block flag */
3397 if (!force_nonblock)
3398 kiocb->ki_flags &= ~IOCB_NOWAIT;
3400 kiocb->ki_flags |= IOCB_NOWAIT;
3403 /* If the file doesn't support async, just async punt */
3404 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3408 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3412 ret = io_iter_do_read(req, iter);
3416 } else if (ret == -EIOCBQUEUED) {
3419 } else if (ret == -EAGAIN) {
3420 /* IOPOLL retry should happen for io-wq threads */
3421 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3423 /* no retry on NONBLOCK marked file */
3424 if (req->file->f_flags & O_NONBLOCK)
3426 /* some cases will consume bytes even on error returns */
3427 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3430 } else if (ret < 0) {
3431 /* make sure -ERESTARTSYS -> -EINTR is done */
3435 /* read it all, or we did blocking attempt. no retry. */
3436 if (!iov_iter_count(iter) || !force_nonblock ||
3437 (req->file->f_flags & O_NONBLOCK))
3442 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3449 rw = req->async_data;
3450 /* it's copied and will be cleaned with ->io */
3452 /* now use our persistent iterator, if we aren't already */
3455 rw->bytes_done += ret;
3456 /* if we can retry, do so with the callbacks armed */
3457 if (!io_rw_should_retry(req)) {
3458 kiocb->ki_flags &= ~IOCB_WAITQ;
3463 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3464 * get -EIOCBQUEUED, then we'll get a notification when the desired
3465 * page gets unlocked. We can also get a partial read here, and if we
3466 * do, then just retry at the new offset.
3468 ret = io_iter_do_read(req, iter);
3469 if (ret == -EIOCBQUEUED) {
3472 } else if (ret > 0 && ret < io_size) {
3473 /* we got some bytes, but not all. retry. */
3477 kiocb_done(kiocb, ret, cs);
3480 /* it's reportedly faster than delegating the null check to kfree() */
3486 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3490 ret = io_prep_rw(req, sqe);
3494 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3497 /* either don't need iovec imported or already have it */
3498 if (!req->async_data)
3500 return io_rw_prep_async(req, WRITE);
3503 static int io_write(struct io_kiocb *req, bool force_nonblock,
3504 struct io_comp_state *cs)
3506 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3507 struct kiocb *kiocb = &req->rw.kiocb;
3508 struct iov_iter __iter, *iter = &__iter;
3509 struct io_async_rw *rw = req->async_data;
3511 ssize_t ret, ret2, io_size;
3516 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3519 iov_count = iov_iter_count(iter);
3521 req->result = io_size;
3523 /* Ensure we clear previously set non-block flag */
3524 if (!force_nonblock)
3525 kiocb->ki_flags &= ~IOCB_NOWAIT;
3527 kiocb->ki_flags |= IOCB_NOWAIT;
3529 /* If the file doesn't support async, just async punt */
3530 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3533 /* file path doesn't support NOWAIT for non-direct_IO */
3534 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3535 (req->flags & REQ_F_ISREG))
3538 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3543 * Open-code file_start_write here to grab freeze protection,
3544 * which will be released by another thread in
3545 * io_complete_rw(). Fool lockdep by telling it the lock got
3546 * released so that it doesn't complain about the held lock when
3547 * we return to userspace.
3549 if (req->flags & REQ_F_ISREG) {
3550 sb_start_write(file_inode(req->file)->i_sb);
3551 __sb_writers_release(file_inode(req->file)->i_sb,
3554 kiocb->ki_flags |= IOCB_WRITE;
3556 if (req->file->f_op->write_iter)
3557 ret2 = call_write_iter(req->file, kiocb, iter);
3558 else if (req->file->f_op->write)
3559 ret2 = loop_rw_iter(WRITE, req, iter);
3564 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3565 * retry them without IOCB_NOWAIT.
3567 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3569 /* no retry on NONBLOCK marked file */
3570 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3572 if (!force_nonblock || ret2 != -EAGAIN) {
3573 /* IOPOLL retry should happen for io-wq threads */
3574 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3577 kiocb_done(kiocb, ret2, cs);
3580 /* some cases will consume bytes even on error returns */
3581 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3582 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3587 /* it's reportedly faster than delegating the null check to kfree() */
3593 static int __io_splice_prep(struct io_kiocb *req,
3594 const struct io_uring_sqe *sqe)
3596 struct io_splice* sp = &req->splice;
3597 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3599 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3603 sp->len = READ_ONCE(sqe->len);
3604 sp->flags = READ_ONCE(sqe->splice_flags);
3606 if (unlikely(sp->flags & ~valid_flags))
3609 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3610 (sp->flags & SPLICE_F_FD_IN_FIXED));
3613 req->flags |= REQ_F_NEED_CLEANUP;
3615 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3617 * Splice operation will be punted aync, and here need to
3618 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3620 io_req_init_async(req);
3621 req->work.flags |= IO_WQ_WORK_UNBOUND;
3627 static int io_tee_prep(struct io_kiocb *req,
3628 const struct io_uring_sqe *sqe)
3630 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3632 return __io_splice_prep(req, sqe);
3635 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3637 struct io_splice *sp = &req->splice;
3638 struct file *in = sp->file_in;
3639 struct file *out = sp->file_out;
3640 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3646 ret = do_tee(in, out, sp->len, flags);
3648 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3649 req->flags &= ~REQ_F_NEED_CLEANUP;
3652 req_set_fail_links(req);
3653 io_req_complete(req, ret);
3657 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3659 struct io_splice* sp = &req->splice;
3661 sp->off_in = READ_ONCE(sqe->splice_off_in);
3662 sp->off_out = READ_ONCE(sqe->off);
3663 return __io_splice_prep(req, sqe);
3666 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3668 struct io_splice *sp = &req->splice;
3669 struct file *in = sp->file_in;
3670 struct file *out = sp->file_out;
3671 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3672 loff_t *poff_in, *poff_out;
3678 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3679 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3682 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3684 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3685 req->flags &= ~REQ_F_NEED_CLEANUP;
3688 req_set_fail_links(req);
3689 io_req_complete(req, ret);
3694 * IORING_OP_NOP just posts a completion event, nothing else.
3696 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3698 struct io_ring_ctx *ctx = req->ctx;
3700 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3703 __io_req_complete(req, 0, 0, cs);
3707 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3709 struct io_ring_ctx *ctx = req->ctx;
3714 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3716 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3719 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3720 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3723 req->sync.off = READ_ONCE(sqe->off);
3724 req->sync.len = READ_ONCE(sqe->len);
3728 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3730 loff_t end = req->sync.off + req->sync.len;
3733 /* fsync always requires a blocking context */
3737 ret = vfs_fsync_range(req->file, req->sync.off,
3738 end > 0 ? end : LLONG_MAX,
3739 req->sync.flags & IORING_FSYNC_DATASYNC);
3741 req_set_fail_links(req);
3742 io_req_complete(req, ret);
3746 static int io_fallocate_prep(struct io_kiocb *req,
3747 const struct io_uring_sqe *sqe)
3749 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3751 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3754 req->sync.off = READ_ONCE(sqe->off);
3755 req->sync.len = READ_ONCE(sqe->addr);
3756 req->sync.mode = READ_ONCE(sqe->len);
3760 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3764 /* fallocate always requiring blocking context */
3767 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3770 req_set_fail_links(req);
3771 io_req_complete(req, ret);
3775 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3777 const char __user *fname;
3780 if (unlikely(sqe->ioprio || sqe->buf_index))
3782 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3785 /* open.how should be already initialised */
3786 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3787 req->open.how.flags |= O_LARGEFILE;
3789 req->open.dfd = READ_ONCE(sqe->fd);
3790 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3791 req->open.filename = getname(fname);
3792 if (IS_ERR(req->open.filename)) {
3793 ret = PTR_ERR(req->open.filename);
3794 req->open.filename = NULL;
3797 req->open.nofile = rlimit(RLIMIT_NOFILE);
3798 req->open.ignore_nonblock = false;
3799 req->flags |= REQ_F_NEED_CLEANUP;
3803 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3807 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3809 mode = READ_ONCE(sqe->len);
3810 flags = READ_ONCE(sqe->open_flags);
3811 req->open.how = build_open_how(flags, mode);
3812 return __io_openat_prep(req, sqe);
3815 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3817 struct open_how __user *how;
3821 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3823 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3824 len = READ_ONCE(sqe->len);
3825 if (len < OPEN_HOW_SIZE_VER0)
3828 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3833 return __io_openat_prep(req, sqe);
3836 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3838 struct open_flags op;
3842 if (force_nonblock && !req->open.ignore_nonblock)
3845 ret = build_open_flags(&req->open.how, &op);
3849 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3853 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3856 ret = PTR_ERR(file);
3858 * A work-around to ensure that /proc/self works that way
3859 * that it should - if we get -EOPNOTSUPP back, then assume
3860 * that proc_self_get_link() failed us because we're in async
3861 * context. We should be safe to retry this from the task
3862 * itself with force_nonblock == false set, as it should not
3863 * block on lookup. Would be nice to know this upfront and
3864 * avoid the async dance, but doesn't seem feasible.
3866 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
3867 req->open.ignore_nonblock = true;
3868 refcount_inc(&req->refs);
3869 io_req_task_queue(req);
3873 fsnotify_open(file);
3874 fd_install(ret, file);
3877 putname(req->open.filename);
3878 req->flags &= ~REQ_F_NEED_CLEANUP;
3880 req_set_fail_links(req);
3881 io_req_complete(req, ret);
3885 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3887 return io_openat2(req, force_nonblock);
3890 static int io_remove_buffers_prep(struct io_kiocb *req,
3891 const struct io_uring_sqe *sqe)
3893 struct io_provide_buf *p = &req->pbuf;
3896 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3899 tmp = READ_ONCE(sqe->fd);
3900 if (!tmp || tmp > USHRT_MAX)
3903 memset(p, 0, sizeof(*p));
3905 p->bgid = READ_ONCE(sqe->buf_group);
3909 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3910 int bgid, unsigned nbufs)
3914 /* shouldn't happen */
3918 /* the head kbuf is the list itself */
3919 while (!list_empty(&buf->list)) {
3920 struct io_buffer *nxt;
3922 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3923 list_del(&nxt->list);
3930 idr_remove(&ctx->io_buffer_idr, bgid);
3935 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3936 struct io_comp_state *cs)
3938 struct io_provide_buf *p = &req->pbuf;
3939 struct io_ring_ctx *ctx = req->ctx;
3940 struct io_buffer *head;
3943 io_ring_submit_lock(ctx, !force_nonblock);
3945 lockdep_assert_held(&ctx->uring_lock);
3948 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3950 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3952 io_ring_submit_lock(ctx, !force_nonblock);
3954 req_set_fail_links(req);
3955 __io_req_complete(req, ret, 0, cs);
3959 static int io_provide_buffers_prep(struct io_kiocb *req,
3960 const struct io_uring_sqe *sqe)
3962 struct io_provide_buf *p = &req->pbuf;
3965 if (sqe->ioprio || sqe->rw_flags)
3968 tmp = READ_ONCE(sqe->fd);
3969 if (!tmp || tmp > USHRT_MAX)
3972 p->addr = READ_ONCE(sqe->addr);
3973 p->len = READ_ONCE(sqe->len);
3975 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3978 p->bgid = READ_ONCE(sqe->buf_group);
3979 tmp = READ_ONCE(sqe->off);
3980 if (tmp > USHRT_MAX)
3986 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3988 struct io_buffer *buf;
3989 u64 addr = pbuf->addr;
3990 int i, bid = pbuf->bid;
3992 for (i = 0; i < pbuf->nbufs; i++) {
3993 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3998 buf->len = pbuf->len;
4003 INIT_LIST_HEAD(&buf->list);
4006 list_add_tail(&buf->list, &(*head)->list);
4010 return i ? i : -ENOMEM;
4013 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4014 struct io_comp_state *cs)
4016 struct io_provide_buf *p = &req->pbuf;
4017 struct io_ring_ctx *ctx = req->ctx;
4018 struct io_buffer *head, *list;
4021 io_ring_submit_lock(ctx, !force_nonblock);
4023 lockdep_assert_held(&ctx->uring_lock);
4025 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4027 ret = io_add_buffers(p, &head);
4032 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4035 __io_remove_buffers(ctx, head, p->bgid, -1U);
4040 io_ring_submit_unlock(ctx, !force_nonblock);
4042 req_set_fail_links(req);
4043 __io_req_complete(req, ret, 0, cs);
4047 static int io_epoll_ctl_prep(struct io_kiocb *req,
4048 const struct io_uring_sqe *sqe)
4050 #if defined(CONFIG_EPOLL)
4051 if (sqe->ioprio || sqe->buf_index)
4053 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4056 req->epoll.epfd = READ_ONCE(sqe->fd);
4057 req->epoll.op = READ_ONCE(sqe->len);
4058 req->epoll.fd = READ_ONCE(sqe->off);
4060 if (ep_op_has_event(req->epoll.op)) {
4061 struct epoll_event __user *ev;
4063 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4064 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4074 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4075 struct io_comp_state *cs)
4077 #if defined(CONFIG_EPOLL)
4078 struct io_epoll *ie = &req->epoll;
4081 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4082 if (force_nonblock && ret == -EAGAIN)
4086 req_set_fail_links(req);
4087 __io_req_complete(req, ret, 0, cs);
4094 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4096 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4097 if (sqe->ioprio || sqe->buf_index || sqe->off)
4099 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4102 req->madvise.addr = READ_ONCE(sqe->addr);
4103 req->madvise.len = READ_ONCE(sqe->len);
4104 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4111 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4113 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4114 struct io_madvise *ma = &req->madvise;
4120 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4122 req_set_fail_links(req);
4123 io_req_complete(req, ret);
4130 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4132 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4134 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4137 req->fadvise.offset = READ_ONCE(sqe->off);
4138 req->fadvise.len = READ_ONCE(sqe->len);
4139 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4143 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4145 struct io_fadvise *fa = &req->fadvise;
4148 if (force_nonblock) {
4149 switch (fa->advice) {
4150 case POSIX_FADV_NORMAL:
4151 case POSIX_FADV_RANDOM:
4152 case POSIX_FADV_SEQUENTIAL:
4159 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4161 req_set_fail_links(req);
4162 io_req_complete(req, ret);
4166 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4168 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4170 if (sqe->ioprio || sqe->buf_index)
4172 if (req->flags & REQ_F_FIXED_FILE)
4175 req->statx.dfd = READ_ONCE(sqe->fd);
4176 req->statx.mask = READ_ONCE(sqe->len);
4177 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4178 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4179 req->statx.flags = READ_ONCE(sqe->statx_flags);
4184 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4186 struct io_statx *ctx = &req->statx;
4189 if (force_nonblock) {
4190 /* only need file table for an actual valid fd */
4191 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4192 req->flags |= REQ_F_NO_FILE_TABLE;
4196 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4200 req_set_fail_links(req);
4201 io_req_complete(req, ret);
4205 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4208 * If we queue this for async, it must not be cancellable. That would
4209 * leave the 'file' in an undeterminate state, and here need to modify
4210 * io_wq_work.flags, so initialize io_wq_work firstly.
4212 io_req_init_async(req);
4213 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4215 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4217 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4218 sqe->rw_flags || sqe->buf_index)
4220 if (req->flags & REQ_F_FIXED_FILE)
4223 req->close.fd = READ_ONCE(sqe->fd);
4224 if ((req->file && req->file->f_op == &io_uring_fops))
4227 req->close.put_file = NULL;
4231 static int io_close(struct io_kiocb *req, bool force_nonblock,
4232 struct io_comp_state *cs)
4234 struct io_close *close = &req->close;
4237 /* might be already done during nonblock submission */
4238 if (!close->put_file) {
4239 ret = __close_fd_get_file(close->fd, &close->put_file);
4241 return (ret == -ENOENT) ? -EBADF : ret;
4244 /* if the file has a flush method, be safe and punt to async */
4245 if (close->put_file->f_op->flush && force_nonblock) {
4246 /* was never set, but play safe */
4247 req->flags &= ~REQ_F_NOWAIT;
4248 /* avoid grabbing files - we don't need the files */
4249 req->flags |= REQ_F_NO_FILE_TABLE;
4253 /* No ->flush() or already async, safely close from here */
4254 ret = filp_close(close->put_file, req->work.identity->files);
4256 req_set_fail_links(req);
4257 fput(close->put_file);
4258 close->put_file = NULL;
4259 __io_req_complete(req, ret, 0, cs);
4263 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4265 struct io_ring_ctx *ctx = req->ctx;
4270 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4272 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4275 req->sync.off = READ_ONCE(sqe->off);
4276 req->sync.len = READ_ONCE(sqe->len);
4277 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4281 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4285 /* sync_file_range always requires a blocking context */
4289 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4292 req_set_fail_links(req);
4293 io_req_complete(req, ret);
4297 #if defined(CONFIG_NET)
4298 static int io_setup_async_msg(struct io_kiocb *req,
4299 struct io_async_msghdr *kmsg)
4301 struct io_async_msghdr *async_msg = req->async_data;
4305 if (io_alloc_async_data(req)) {
4306 if (kmsg->iov != kmsg->fast_iov)
4310 async_msg = req->async_data;
4311 req->flags |= REQ_F_NEED_CLEANUP;
4312 memcpy(async_msg, kmsg, sizeof(*kmsg));
4316 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4317 struct io_async_msghdr *iomsg)
4319 iomsg->iov = iomsg->fast_iov;
4320 iomsg->msg.msg_name = &iomsg->addr;
4321 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4322 req->sr_msg.msg_flags, &iomsg->iov);
4325 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4327 struct io_async_msghdr *async_msg = req->async_data;
4328 struct io_sr_msg *sr = &req->sr_msg;
4331 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4334 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4335 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4336 sr->len = READ_ONCE(sqe->len);
4338 #ifdef CONFIG_COMPAT
4339 if (req->ctx->compat)
4340 sr->msg_flags |= MSG_CMSG_COMPAT;
4343 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4345 ret = io_sendmsg_copy_hdr(req, async_msg);
4347 req->flags |= REQ_F_NEED_CLEANUP;
4351 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4352 struct io_comp_state *cs)
4354 struct io_async_msghdr iomsg, *kmsg;
4355 struct socket *sock;
4359 sock = sock_from_file(req->file, &ret);
4360 if (unlikely(!sock))
4363 if (req->async_data) {
4364 kmsg = req->async_data;
4365 kmsg->msg.msg_name = &kmsg->addr;
4366 /* if iov is set, it's allocated already */
4368 kmsg->iov = kmsg->fast_iov;
4369 kmsg->msg.msg_iter.iov = kmsg->iov;
4371 ret = io_sendmsg_copy_hdr(req, &iomsg);
4377 flags = req->sr_msg.msg_flags;
4378 if (flags & MSG_DONTWAIT)
4379 req->flags |= REQ_F_NOWAIT;
4380 else if (force_nonblock)
4381 flags |= MSG_DONTWAIT;
4383 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4384 if (force_nonblock && ret == -EAGAIN)
4385 return io_setup_async_msg(req, kmsg);
4386 if (ret == -ERESTARTSYS)
4389 if (kmsg->iov != kmsg->fast_iov)
4391 req->flags &= ~REQ_F_NEED_CLEANUP;
4393 req_set_fail_links(req);
4394 __io_req_complete(req, ret, 0, cs);
4398 static int io_send(struct io_kiocb *req, bool force_nonblock,
4399 struct io_comp_state *cs)
4401 struct io_sr_msg *sr = &req->sr_msg;
4404 struct socket *sock;
4408 sock = sock_from_file(req->file, &ret);
4409 if (unlikely(!sock))
4412 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4416 msg.msg_name = NULL;
4417 msg.msg_control = NULL;
4418 msg.msg_controllen = 0;
4419 msg.msg_namelen = 0;
4421 flags = req->sr_msg.msg_flags;
4422 if (flags & MSG_DONTWAIT)
4423 req->flags |= REQ_F_NOWAIT;
4424 else if (force_nonblock)
4425 flags |= MSG_DONTWAIT;
4427 msg.msg_flags = flags;
4428 ret = sock_sendmsg(sock, &msg);
4429 if (force_nonblock && ret == -EAGAIN)
4431 if (ret == -ERESTARTSYS)
4435 req_set_fail_links(req);
4436 __io_req_complete(req, ret, 0, cs);
4440 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4441 struct io_async_msghdr *iomsg)
4443 struct io_sr_msg *sr = &req->sr_msg;
4444 struct iovec __user *uiov;
4448 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4449 &iomsg->uaddr, &uiov, &iov_len);
4453 if (req->flags & REQ_F_BUFFER_SELECT) {
4456 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4458 sr->len = iomsg->iov[0].iov_len;
4459 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4463 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4464 &iomsg->iov, &iomsg->msg.msg_iter,
4473 #ifdef CONFIG_COMPAT
4474 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4475 struct io_async_msghdr *iomsg)
4477 struct compat_msghdr __user *msg_compat;
4478 struct io_sr_msg *sr = &req->sr_msg;
4479 struct compat_iovec __user *uiov;
4484 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4485 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4490 uiov = compat_ptr(ptr);
4491 if (req->flags & REQ_F_BUFFER_SELECT) {
4492 compat_ssize_t clen;
4496 if (!access_ok(uiov, sizeof(*uiov)))
4498 if (__get_user(clen, &uiov->iov_len))
4502 sr->len = iomsg->iov[0].iov_len;
4505 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4506 UIO_FASTIOV, &iomsg->iov,
4507 &iomsg->msg.msg_iter, true);
4516 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4517 struct io_async_msghdr *iomsg)
4519 iomsg->msg.msg_name = &iomsg->addr;
4520 iomsg->iov = iomsg->fast_iov;
4522 #ifdef CONFIG_COMPAT
4523 if (req->ctx->compat)
4524 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4527 return __io_recvmsg_copy_hdr(req, iomsg);
4530 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4533 struct io_sr_msg *sr = &req->sr_msg;
4534 struct io_buffer *kbuf;
4536 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4541 req->flags |= REQ_F_BUFFER_SELECTED;
4545 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4547 return io_put_kbuf(req, req->sr_msg.kbuf);
4550 static int io_recvmsg_prep(struct io_kiocb *req,
4551 const struct io_uring_sqe *sqe)
4553 struct io_async_msghdr *async_msg = req->async_data;
4554 struct io_sr_msg *sr = &req->sr_msg;
4557 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4560 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4561 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4562 sr->len = READ_ONCE(sqe->len);
4563 sr->bgid = READ_ONCE(sqe->buf_group);
4565 #ifdef CONFIG_COMPAT
4566 if (req->ctx->compat)
4567 sr->msg_flags |= MSG_CMSG_COMPAT;
4570 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4572 ret = io_recvmsg_copy_hdr(req, async_msg);
4574 req->flags |= REQ_F_NEED_CLEANUP;
4578 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4579 struct io_comp_state *cs)
4581 struct io_async_msghdr iomsg, *kmsg;
4582 struct socket *sock;
4583 struct io_buffer *kbuf;
4585 int ret, cflags = 0;
4587 sock = sock_from_file(req->file, &ret);
4588 if (unlikely(!sock))
4591 if (req->async_data) {
4592 kmsg = req->async_data;
4593 kmsg->msg.msg_name = &kmsg->addr;
4594 /* if iov is set, it's allocated already */
4596 kmsg->iov = kmsg->fast_iov;
4597 kmsg->msg.msg_iter.iov = kmsg->iov;
4599 ret = io_recvmsg_copy_hdr(req, &iomsg);
4605 if (req->flags & REQ_F_BUFFER_SELECT) {
4606 kbuf = io_recv_buffer_select(req, !force_nonblock);
4608 return PTR_ERR(kbuf);
4609 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4610 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4611 1, req->sr_msg.len);
4614 flags = req->sr_msg.msg_flags;
4615 if (flags & MSG_DONTWAIT)
4616 req->flags |= REQ_F_NOWAIT;
4617 else if (force_nonblock)
4618 flags |= MSG_DONTWAIT;
4620 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4621 kmsg->uaddr, flags);
4622 if (force_nonblock && ret == -EAGAIN)
4623 return io_setup_async_msg(req, kmsg);
4624 if (ret == -ERESTARTSYS)
4627 if (req->flags & REQ_F_BUFFER_SELECTED)
4628 cflags = io_put_recv_kbuf(req);
4629 if (kmsg->iov != kmsg->fast_iov)
4631 req->flags &= ~REQ_F_NEED_CLEANUP;
4633 req_set_fail_links(req);
4634 __io_req_complete(req, ret, cflags, cs);
4638 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4639 struct io_comp_state *cs)
4641 struct io_buffer *kbuf;
4642 struct io_sr_msg *sr = &req->sr_msg;
4644 void __user *buf = sr->buf;
4645 struct socket *sock;
4648 int ret, cflags = 0;
4650 sock = sock_from_file(req->file, &ret);
4651 if (unlikely(!sock))
4654 if (req->flags & REQ_F_BUFFER_SELECT) {
4655 kbuf = io_recv_buffer_select(req, !force_nonblock);
4657 return PTR_ERR(kbuf);
4658 buf = u64_to_user_ptr(kbuf->addr);
4661 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4665 msg.msg_name = NULL;
4666 msg.msg_control = NULL;
4667 msg.msg_controllen = 0;
4668 msg.msg_namelen = 0;
4669 msg.msg_iocb = NULL;
4672 flags = req->sr_msg.msg_flags;
4673 if (flags & MSG_DONTWAIT)
4674 req->flags |= REQ_F_NOWAIT;
4675 else if (force_nonblock)
4676 flags |= MSG_DONTWAIT;
4678 ret = sock_recvmsg(sock, &msg, flags);
4679 if (force_nonblock && ret == -EAGAIN)
4681 if (ret == -ERESTARTSYS)
4684 if (req->flags & REQ_F_BUFFER_SELECTED)
4685 cflags = io_put_recv_kbuf(req);
4687 req_set_fail_links(req);
4688 __io_req_complete(req, ret, cflags, cs);
4692 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4694 struct io_accept *accept = &req->accept;
4696 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4698 if (sqe->ioprio || sqe->len || sqe->buf_index)
4701 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4702 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4703 accept->flags = READ_ONCE(sqe->accept_flags);
4704 accept->nofile = rlimit(RLIMIT_NOFILE);
4708 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4709 struct io_comp_state *cs)
4711 struct io_accept *accept = &req->accept;
4712 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4715 if (req->file->f_flags & O_NONBLOCK)
4716 req->flags |= REQ_F_NOWAIT;
4718 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4719 accept->addr_len, accept->flags,
4721 if (ret == -EAGAIN && force_nonblock)
4724 if (ret == -ERESTARTSYS)
4726 req_set_fail_links(req);
4728 __io_req_complete(req, ret, 0, cs);
4732 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4734 struct io_connect *conn = &req->connect;
4735 struct io_async_connect *io = req->async_data;
4737 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4739 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4742 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4743 conn->addr_len = READ_ONCE(sqe->addr2);
4748 return move_addr_to_kernel(conn->addr, conn->addr_len,
4752 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4753 struct io_comp_state *cs)
4755 struct io_async_connect __io, *io;
4756 unsigned file_flags;
4759 if (req->async_data) {
4760 io = req->async_data;
4762 ret = move_addr_to_kernel(req->connect.addr,
4763 req->connect.addr_len,
4770 file_flags = force_nonblock ? O_NONBLOCK : 0;
4772 ret = __sys_connect_file(req->file, &io->address,
4773 req->connect.addr_len, file_flags);
4774 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4775 if (req->async_data)
4777 if (io_alloc_async_data(req)) {
4781 io = req->async_data;
4782 memcpy(req->async_data, &__io, sizeof(__io));
4785 if (ret == -ERESTARTSYS)
4789 req_set_fail_links(req);
4790 __io_req_complete(req, ret, 0, cs);
4793 #else /* !CONFIG_NET */
4794 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4799 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4800 struct io_comp_state *cs)
4805 static int io_send(struct io_kiocb *req, bool force_nonblock,
4806 struct io_comp_state *cs)
4811 static int io_recvmsg_prep(struct io_kiocb *req,
4812 const struct io_uring_sqe *sqe)
4817 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4818 struct io_comp_state *cs)
4823 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4824 struct io_comp_state *cs)
4829 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4834 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4835 struct io_comp_state *cs)
4840 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4845 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4846 struct io_comp_state *cs)
4850 #endif /* CONFIG_NET */
4852 struct io_poll_table {
4853 struct poll_table_struct pt;
4854 struct io_kiocb *req;
4858 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4859 __poll_t mask, task_work_func_t func)
4864 /* for instances that support it check for an event match first: */
4865 if (mask && !(mask & poll->events))
4868 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4870 list_del_init(&poll->wait.entry);
4873 init_task_work(&req->task_work, func);
4874 percpu_ref_get(&req->ctx->refs);
4877 * If we using the signalfd wait_queue_head for this wakeup, then
4878 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4879 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4880 * either, as the normal wakeup will suffice.
4882 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4885 * If this fails, then the task is exiting. When a task exits, the
4886 * work gets canceled, so just cancel this request as well instead
4887 * of executing it. We can't safely execute it anyway, as we may not
4888 * have the needed state needed for it anyway.
4890 ret = io_req_task_work_add(req, twa_signal_ok);
4891 if (unlikely(ret)) {
4892 struct task_struct *tsk;
4894 WRITE_ONCE(poll->canceled, true);
4895 tsk = io_wq_get_task(req->ctx->io_wq);
4896 task_work_add(tsk, &req->task_work, TWA_NONE);
4897 wake_up_process(tsk);
4902 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4903 __acquires(&req->ctx->completion_lock)
4905 struct io_ring_ctx *ctx = req->ctx;
4907 if (!req->result && !READ_ONCE(poll->canceled)) {
4908 struct poll_table_struct pt = { ._key = poll->events };
4910 req->result = vfs_poll(req->file, &pt) & poll->events;
4913 spin_lock_irq(&ctx->completion_lock);
4914 if (!req->result && !READ_ONCE(poll->canceled)) {
4915 add_wait_queue(poll->head, &poll->wait);
4922 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4924 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4925 if (req->opcode == IORING_OP_POLL_ADD)
4926 return req->async_data;
4927 return req->apoll->double_poll;
4930 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4932 if (req->opcode == IORING_OP_POLL_ADD)
4934 return &req->apoll->poll;
4937 static void io_poll_remove_double(struct io_kiocb *req)
4939 struct io_poll_iocb *poll = io_poll_get_double(req);
4941 lockdep_assert_held(&req->ctx->completion_lock);
4943 if (poll && poll->head) {
4944 struct wait_queue_head *head = poll->head;
4946 spin_lock(&head->lock);
4947 list_del_init(&poll->wait.entry);
4948 if (poll->wait.private)
4949 refcount_dec(&req->refs);
4951 spin_unlock(&head->lock);
4955 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4957 struct io_ring_ctx *ctx = req->ctx;
4959 io_poll_remove_double(req);
4960 req->poll.done = true;
4961 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4962 io_commit_cqring(ctx);
4965 static void io_poll_task_func(struct callback_head *cb)
4967 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4968 struct io_ring_ctx *ctx = req->ctx;
4969 struct io_kiocb *nxt;
4971 if (io_poll_rewait(req, &req->poll)) {
4972 spin_unlock_irq(&ctx->completion_lock);
4974 hash_del(&req->hash_node);
4975 io_poll_complete(req, req->result, 0);
4976 spin_unlock_irq(&ctx->completion_lock);
4978 nxt = io_put_req_find_next(req);
4979 io_cqring_ev_posted(ctx);
4981 __io_req_task_submit(nxt);
4984 percpu_ref_put(&ctx->refs);
4987 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4988 int sync, void *key)
4990 struct io_kiocb *req = wait->private;
4991 struct io_poll_iocb *poll = io_poll_get_single(req);
4992 __poll_t mask = key_to_poll(key);
4994 /* for instances that support it check for an event match first: */
4995 if (mask && !(mask & poll->events))
4998 list_del_init(&wait->entry);
5000 if (poll && poll->head) {
5003 spin_lock(&poll->head->lock);
5004 done = list_empty(&poll->wait.entry);
5006 list_del_init(&poll->wait.entry);
5007 /* make sure double remove sees this as being gone */
5008 wait->private = NULL;
5009 spin_unlock(&poll->head->lock);
5011 /* use wait func handler, so it matches the rq type */
5012 poll->wait.func(&poll->wait, mode, sync, key);
5015 refcount_dec(&req->refs);
5019 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5020 wait_queue_func_t wake_func)
5024 poll->canceled = false;
5025 poll->events = events;
5026 INIT_LIST_HEAD(&poll->wait.entry);
5027 init_waitqueue_func_entry(&poll->wait, wake_func);
5030 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5031 struct wait_queue_head *head,
5032 struct io_poll_iocb **poll_ptr)
5034 struct io_kiocb *req = pt->req;
5037 * If poll->head is already set, it's because the file being polled
5038 * uses multiple waitqueues for poll handling (eg one for read, one
5039 * for write). Setup a separate io_poll_iocb if this happens.
5041 if (unlikely(poll->head)) {
5042 struct io_poll_iocb *poll_one = poll;
5044 /* already have a 2nd entry, fail a third attempt */
5046 pt->error = -EINVAL;
5049 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5051 pt->error = -ENOMEM;
5054 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5055 refcount_inc(&req->refs);
5056 poll->wait.private = req;
5063 if (poll->events & EPOLLEXCLUSIVE)
5064 add_wait_queue_exclusive(head, &poll->wait);
5066 add_wait_queue(head, &poll->wait);
5069 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5070 struct poll_table_struct *p)
5072 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5073 struct async_poll *apoll = pt->req->apoll;
5075 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5078 static void io_async_task_func(struct callback_head *cb)
5080 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5081 struct async_poll *apoll = req->apoll;
5082 struct io_ring_ctx *ctx = req->ctx;
5084 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5086 if (io_poll_rewait(req, &apoll->poll)) {
5087 spin_unlock_irq(&ctx->completion_lock);
5088 percpu_ref_put(&ctx->refs);
5092 /* If req is still hashed, it cannot have been canceled. Don't check. */
5093 if (hash_hashed(&req->hash_node))
5094 hash_del(&req->hash_node);
5096 io_poll_remove_double(req);
5097 spin_unlock_irq(&ctx->completion_lock);
5099 if (!READ_ONCE(apoll->poll.canceled))
5100 __io_req_task_submit(req);
5102 __io_req_task_cancel(req, -ECANCELED);
5104 percpu_ref_put(&ctx->refs);
5105 kfree(apoll->double_poll);
5109 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5112 struct io_kiocb *req = wait->private;
5113 struct io_poll_iocb *poll = &req->apoll->poll;
5115 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5118 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5121 static void io_poll_req_insert(struct io_kiocb *req)
5123 struct io_ring_ctx *ctx = req->ctx;
5124 struct hlist_head *list;
5126 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5127 hlist_add_head(&req->hash_node, list);
5130 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5131 struct io_poll_iocb *poll,
5132 struct io_poll_table *ipt, __poll_t mask,
5133 wait_queue_func_t wake_func)
5134 __acquires(&ctx->completion_lock)
5136 struct io_ring_ctx *ctx = req->ctx;
5137 bool cancel = false;
5139 INIT_HLIST_NODE(&req->hash_node);
5140 io_init_poll_iocb(poll, mask, wake_func);
5141 poll->file = req->file;
5142 poll->wait.private = req;
5144 ipt->pt._key = mask;
5146 ipt->error = -EINVAL;
5148 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5150 spin_lock_irq(&ctx->completion_lock);
5151 if (likely(poll->head)) {
5152 spin_lock(&poll->head->lock);
5153 if (unlikely(list_empty(&poll->wait.entry))) {
5159 if (mask || ipt->error)
5160 list_del_init(&poll->wait.entry);
5162 WRITE_ONCE(poll->canceled, true);
5163 else if (!poll->done) /* actually waiting for an event */
5164 io_poll_req_insert(req);
5165 spin_unlock(&poll->head->lock);
5171 static bool io_arm_poll_handler(struct io_kiocb *req)
5173 const struct io_op_def *def = &io_op_defs[req->opcode];
5174 struct io_ring_ctx *ctx = req->ctx;
5175 struct async_poll *apoll;
5176 struct io_poll_table ipt;
5180 if (!req->file || !file_can_poll(req->file))
5182 if (req->flags & REQ_F_POLLED)
5186 else if (def->pollout)
5190 /* if we can't nonblock try, then no point in arming a poll handler */
5191 if (!io_file_supports_async(req->file, rw))
5194 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5195 if (unlikely(!apoll))
5197 apoll->double_poll = NULL;
5199 req->flags |= REQ_F_POLLED;
5204 mask |= POLLIN | POLLRDNORM;
5206 mask |= POLLOUT | POLLWRNORM;
5208 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5209 if ((req->opcode == IORING_OP_RECVMSG) &&
5210 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5213 mask |= POLLERR | POLLPRI;
5215 ipt.pt._qproc = io_async_queue_proc;
5217 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5219 if (ret || ipt.error) {
5220 io_poll_remove_double(req);
5221 spin_unlock_irq(&ctx->completion_lock);
5222 kfree(apoll->double_poll);
5226 spin_unlock_irq(&ctx->completion_lock);
5227 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5228 apoll->poll.events);
5232 static bool __io_poll_remove_one(struct io_kiocb *req,
5233 struct io_poll_iocb *poll)
5235 bool do_complete = false;
5237 spin_lock(&poll->head->lock);
5238 WRITE_ONCE(poll->canceled, true);
5239 if (!list_empty(&poll->wait.entry)) {
5240 list_del_init(&poll->wait.entry);
5243 spin_unlock(&poll->head->lock);
5244 hash_del(&req->hash_node);
5248 static bool io_poll_remove_one(struct io_kiocb *req)
5252 io_poll_remove_double(req);
5254 if (req->opcode == IORING_OP_POLL_ADD) {
5255 do_complete = __io_poll_remove_one(req, &req->poll);
5257 struct async_poll *apoll = req->apoll;
5259 /* non-poll requests have submit ref still */
5260 do_complete = __io_poll_remove_one(req, &apoll->poll);
5263 kfree(apoll->double_poll);
5269 io_cqring_fill_event(req, -ECANCELED);
5270 io_commit_cqring(req->ctx);
5271 req_set_fail_links(req);
5272 io_put_req_deferred(req, 1);
5279 * Returns true if we found and killed one or more poll requests
5281 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5283 struct hlist_node *tmp;
5284 struct io_kiocb *req;
5287 spin_lock_irq(&ctx->completion_lock);
5288 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5289 struct hlist_head *list;
5291 list = &ctx->cancel_hash[i];
5292 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5293 if (io_task_match(req, tsk))
5294 posted += io_poll_remove_one(req);
5297 spin_unlock_irq(&ctx->completion_lock);
5300 io_cqring_ev_posted(ctx);
5305 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5307 struct hlist_head *list;
5308 struct io_kiocb *req;
5310 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5311 hlist_for_each_entry(req, list, hash_node) {
5312 if (sqe_addr != req->user_data)
5314 if (io_poll_remove_one(req))
5322 static int io_poll_remove_prep(struct io_kiocb *req,
5323 const struct io_uring_sqe *sqe)
5325 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5327 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5331 req->poll.addr = READ_ONCE(sqe->addr);
5336 * Find a running poll command that matches one specified in sqe->addr,
5337 * and remove it if found.
5339 static int io_poll_remove(struct io_kiocb *req)
5341 struct io_ring_ctx *ctx = req->ctx;
5345 addr = req->poll.addr;
5346 spin_lock_irq(&ctx->completion_lock);
5347 ret = io_poll_cancel(ctx, addr);
5348 spin_unlock_irq(&ctx->completion_lock);
5351 req_set_fail_links(req);
5352 io_req_complete(req, ret);
5356 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5359 struct io_kiocb *req = wait->private;
5360 struct io_poll_iocb *poll = &req->poll;
5362 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5365 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5366 struct poll_table_struct *p)
5368 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5370 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5373 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5375 struct io_poll_iocb *poll = &req->poll;
5378 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5380 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5383 events = READ_ONCE(sqe->poll32_events);
5385 events = swahw32(events);
5387 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5388 (events & EPOLLEXCLUSIVE);
5392 static int io_poll_add(struct io_kiocb *req)
5394 struct io_poll_iocb *poll = &req->poll;
5395 struct io_ring_ctx *ctx = req->ctx;
5396 struct io_poll_table ipt;
5399 ipt.pt._qproc = io_poll_queue_proc;
5401 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5404 if (mask) { /* no async, we'd stolen it */
5406 io_poll_complete(req, mask, 0);
5408 spin_unlock_irq(&ctx->completion_lock);
5411 io_cqring_ev_posted(ctx);
5417 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5419 struct io_timeout_data *data = container_of(timer,
5420 struct io_timeout_data, timer);
5421 struct io_kiocb *req = data->req;
5422 struct io_ring_ctx *ctx = req->ctx;
5423 unsigned long flags;
5425 spin_lock_irqsave(&ctx->completion_lock, flags);
5426 list_del_init(&req->timeout.list);
5427 atomic_set(&req->ctx->cq_timeouts,
5428 atomic_read(&req->ctx->cq_timeouts) + 1);
5430 io_cqring_fill_event(req, -ETIME);
5431 io_commit_cqring(ctx);
5432 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5434 io_cqring_ev_posted(ctx);
5435 req_set_fail_links(req);
5437 return HRTIMER_NORESTART;
5440 static int __io_timeout_cancel(struct io_kiocb *req)
5442 struct io_timeout_data *io = req->async_data;
5445 ret = hrtimer_try_to_cancel(&io->timer);
5448 list_del_init(&req->timeout.list);
5450 req_set_fail_links(req);
5451 io_cqring_fill_event(req, -ECANCELED);
5452 io_put_req_deferred(req, 1);
5456 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5458 struct io_kiocb *req;
5461 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5462 if (user_data == req->user_data) {
5471 return __io_timeout_cancel(req);
5474 static int io_timeout_remove_prep(struct io_kiocb *req,
5475 const struct io_uring_sqe *sqe)
5477 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5479 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5481 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5484 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5489 * Remove or update an existing timeout command
5491 static int io_timeout_remove(struct io_kiocb *req)
5493 struct io_ring_ctx *ctx = req->ctx;
5496 spin_lock_irq(&ctx->completion_lock);
5497 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5499 io_cqring_fill_event(req, ret);
5500 io_commit_cqring(ctx);
5501 spin_unlock_irq(&ctx->completion_lock);
5502 io_cqring_ev_posted(ctx);
5504 req_set_fail_links(req);
5509 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5510 bool is_timeout_link)
5512 struct io_timeout_data *data;
5514 u32 off = READ_ONCE(sqe->off);
5516 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5518 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5520 if (off && is_timeout_link)
5522 flags = READ_ONCE(sqe->timeout_flags);
5523 if (flags & ~IORING_TIMEOUT_ABS)
5526 req->timeout.off = off;
5528 if (!req->async_data && io_alloc_async_data(req))
5531 data = req->async_data;
5534 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5537 if (flags & IORING_TIMEOUT_ABS)
5538 data->mode = HRTIMER_MODE_ABS;
5540 data->mode = HRTIMER_MODE_REL;
5542 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5546 static int io_timeout(struct io_kiocb *req)
5548 struct io_ring_ctx *ctx = req->ctx;
5549 struct io_timeout_data *data = req->async_data;
5550 struct list_head *entry;
5551 u32 tail, off = req->timeout.off;
5553 spin_lock_irq(&ctx->completion_lock);
5556 * sqe->off holds how many events that need to occur for this
5557 * timeout event to be satisfied. If it isn't set, then this is
5558 * a pure timeout request, sequence isn't used.
5560 if (io_is_timeout_noseq(req)) {
5561 entry = ctx->timeout_list.prev;
5565 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5566 req->timeout.target_seq = tail + off;
5569 * Insertion sort, ensuring the first entry in the list is always
5570 * the one we need first.
5572 list_for_each_prev(entry, &ctx->timeout_list) {
5573 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5576 if (io_is_timeout_noseq(nxt))
5578 /* nxt.seq is behind @tail, otherwise would've been completed */
5579 if (off >= nxt->timeout.target_seq - tail)
5583 list_add(&req->timeout.list, entry);
5584 data->timer.function = io_timeout_fn;
5585 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5586 spin_unlock_irq(&ctx->completion_lock);
5590 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5592 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5594 return req->user_data == (unsigned long) data;
5597 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5599 enum io_wq_cancel cancel_ret;
5602 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5603 switch (cancel_ret) {
5604 case IO_WQ_CANCEL_OK:
5607 case IO_WQ_CANCEL_RUNNING:
5610 case IO_WQ_CANCEL_NOTFOUND:
5618 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5619 struct io_kiocb *req, __u64 sqe_addr,
5622 unsigned long flags;
5625 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5626 if (ret != -ENOENT) {
5627 spin_lock_irqsave(&ctx->completion_lock, flags);
5631 spin_lock_irqsave(&ctx->completion_lock, flags);
5632 ret = io_timeout_cancel(ctx, sqe_addr);
5635 ret = io_poll_cancel(ctx, sqe_addr);
5639 io_cqring_fill_event(req, ret);
5640 io_commit_cqring(ctx);
5641 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5642 io_cqring_ev_posted(ctx);
5645 req_set_fail_links(req);
5649 static int io_async_cancel_prep(struct io_kiocb *req,
5650 const struct io_uring_sqe *sqe)
5652 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5654 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5656 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5659 req->cancel.addr = READ_ONCE(sqe->addr);
5663 static int io_async_cancel(struct io_kiocb *req)
5665 struct io_ring_ctx *ctx = req->ctx;
5667 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5671 static int io_files_update_prep(struct io_kiocb *req,
5672 const struct io_uring_sqe *sqe)
5674 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5676 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5678 if (sqe->ioprio || sqe->rw_flags)
5681 req->files_update.offset = READ_ONCE(sqe->off);
5682 req->files_update.nr_args = READ_ONCE(sqe->len);
5683 if (!req->files_update.nr_args)
5685 req->files_update.arg = READ_ONCE(sqe->addr);
5689 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5690 struct io_comp_state *cs)
5692 struct io_ring_ctx *ctx = req->ctx;
5693 struct io_uring_files_update up;
5699 up.offset = req->files_update.offset;
5700 up.fds = req->files_update.arg;
5702 mutex_lock(&ctx->uring_lock);
5703 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5704 mutex_unlock(&ctx->uring_lock);
5707 req_set_fail_links(req);
5708 __io_req_complete(req, ret, 0, cs);
5712 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5714 switch (req->opcode) {
5717 case IORING_OP_READV:
5718 case IORING_OP_READ_FIXED:
5719 case IORING_OP_READ:
5720 return io_read_prep(req, sqe);
5721 case IORING_OP_WRITEV:
5722 case IORING_OP_WRITE_FIXED:
5723 case IORING_OP_WRITE:
5724 return io_write_prep(req, sqe);
5725 case IORING_OP_POLL_ADD:
5726 return io_poll_add_prep(req, sqe);
5727 case IORING_OP_POLL_REMOVE:
5728 return io_poll_remove_prep(req, sqe);
5729 case IORING_OP_FSYNC:
5730 return io_prep_fsync(req, sqe);
5731 case IORING_OP_SYNC_FILE_RANGE:
5732 return io_prep_sfr(req, sqe);
5733 case IORING_OP_SENDMSG:
5734 case IORING_OP_SEND:
5735 return io_sendmsg_prep(req, sqe);
5736 case IORING_OP_RECVMSG:
5737 case IORING_OP_RECV:
5738 return io_recvmsg_prep(req, sqe);
5739 case IORING_OP_CONNECT:
5740 return io_connect_prep(req, sqe);
5741 case IORING_OP_TIMEOUT:
5742 return io_timeout_prep(req, sqe, false);
5743 case IORING_OP_TIMEOUT_REMOVE:
5744 return io_timeout_remove_prep(req, sqe);
5745 case IORING_OP_ASYNC_CANCEL:
5746 return io_async_cancel_prep(req, sqe);
5747 case IORING_OP_LINK_TIMEOUT:
5748 return io_timeout_prep(req, sqe, true);
5749 case IORING_OP_ACCEPT:
5750 return io_accept_prep(req, sqe);
5751 case IORING_OP_FALLOCATE:
5752 return io_fallocate_prep(req, sqe);
5753 case IORING_OP_OPENAT:
5754 return io_openat_prep(req, sqe);
5755 case IORING_OP_CLOSE:
5756 return io_close_prep(req, sqe);
5757 case IORING_OP_FILES_UPDATE:
5758 return io_files_update_prep(req, sqe);
5759 case IORING_OP_STATX:
5760 return io_statx_prep(req, sqe);
5761 case IORING_OP_FADVISE:
5762 return io_fadvise_prep(req, sqe);
5763 case IORING_OP_MADVISE:
5764 return io_madvise_prep(req, sqe);
5765 case IORING_OP_OPENAT2:
5766 return io_openat2_prep(req, sqe);
5767 case IORING_OP_EPOLL_CTL:
5768 return io_epoll_ctl_prep(req, sqe);
5769 case IORING_OP_SPLICE:
5770 return io_splice_prep(req, sqe);
5771 case IORING_OP_PROVIDE_BUFFERS:
5772 return io_provide_buffers_prep(req, sqe);
5773 case IORING_OP_REMOVE_BUFFERS:
5774 return io_remove_buffers_prep(req, sqe);
5776 return io_tee_prep(req, sqe);
5779 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5784 static int io_req_defer_prep(struct io_kiocb *req,
5785 const struct io_uring_sqe *sqe)
5789 if (io_alloc_async_data(req))
5791 return io_req_prep(req, sqe);
5794 static u32 io_get_sequence(struct io_kiocb *req)
5796 struct io_kiocb *pos;
5797 struct io_ring_ctx *ctx = req->ctx;
5798 u32 total_submitted, nr_reqs = 1;
5800 if (req->flags & REQ_F_LINK_HEAD)
5801 list_for_each_entry(pos, &req->link_list, link_list)
5804 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5805 return total_submitted - nr_reqs;
5808 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5810 struct io_ring_ctx *ctx = req->ctx;
5811 struct io_defer_entry *de;
5815 /* Still need defer if there is pending req in defer list. */
5816 if (likely(list_empty_careful(&ctx->defer_list) &&
5817 !(req->flags & REQ_F_IO_DRAIN)))
5820 seq = io_get_sequence(req);
5821 /* Still a chance to pass the sequence check */
5822 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5825 if (!req->async_data) {
5826 ret = io_req_defer_prep(req, sqe);
5830 io_prep_async_link(req);
5831 de = kmalloc(sizeof(*de), GFP_KERNEL);
5835 spin_lock_irq(&ctx->completion_lock);
5836 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5837 spin_unlock_irq(&ctx->completion_lock);
5839 io_queue_async_work(req);
5840 return -EIOCBQUEUED;
5843 trace_io_uring_defer(ctx, req, req->user_data);
5846 list_add_tail(&de->list, &ctx->defer_list);
5847 spin_unlock_irq(&ctx->completion_lock);
5848 return -EIOCBQUEUED;
5851 static void io_req_drop_files(struct io_kiocb *req)
5853 struct io_ring_ctx *ctx = req->ctx;
5854 unsigned long flags;
5856 spin_lock_irqsave(&ctx->inflight_lock, flags);
5857 list_del(&req->inflight_entry);
5858 if (waitqueue_active(&ctx->inflight_wait))
5859 wake_up(&ctx->inflight_wait);
5860 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5861 req->flags &= ~REQ_F_INFLIGHT;
5862 put_files_struct(req->work.identity->files);
5863 put_nsproxy(req->work.identity->nsproxy);
5864 req->work.flags &= ~IO_WQ_WORK_FILES;
5867 static void __io_clean_op(struct io_kiocb *req)
5869 if (req->flags & REQ_F_BUFFER_SELECTED) {
5870 switch (req->opcode) {
5871 case IORING_OP_READV:
5872 case IORING_OP_READ_FIXED:
5873 case IORING_OP_READ:
5874 kfree((void *)(unsigned long)req->rw.addr);
5876 case IORING_OP_RECVMSG:
5877 case IORING_OP_RECV:
5878 kfree(req->sr_msg.kbuf);
5881 req->flags &= ~REQ_F_BUFFER_SELECTED;
5884 if (req->flags & REQ_F_NEED_CLEANUP) {
5885 switch (req->opcode) {
5886 case IORING_OP_READV:
5887 case IORING_OP_READ_FIXED:
5888 case IORING_OP_READ:
5889 case IORING_OP_WRITEV:
5890 case IORING_OP_WRITE_FIXED:
5891 case IORING_OP_WRITE: {
5892 struct io_async_rw *io = req->async_data;
5894 kfree(io->free_iovec);
5897 case IORING_OP_RECVMSG:
5898 case IORING_OP_SENDMSG: {
5899 struct io_async_msghdr *io = req->async_data;
5900 if (io->iov != io->fast_iov)
5904 case IORING_OP_SPLICE:
5906 io_put_file(req, req->splice.file_in,
5907 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5909 case IORING_OP_OPENAT:
5910 case IORING_OP_OPENAT2:
5911 if (req->open.filename)
5912 putname(req->open.filename);
5915 req->flags &= ~REQ_F_NEED_CLEANUP;
5918 if (req->flags & REQ_F_INFLIGHT)
5919 io_req_drop_files(req);
5922 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5923 struct io_comp_state *cs)
5925 struct io_ring_ctx *ctx = req->ctx;
5928 switch (req->opcode) {
5930 ret = io_nop(req, cs);
5932 case IORING_OP_READV:
5933 case IORING_OP_READ_FIXED:
5934 case IORING_OP_READ:
5935 ret = io_read(req, force_nonblock, cs);
5937 case IORING_OP_WRITEV:
5938 case IORING_OP_WRITE_FIXED:
5939 case IORING_OP_WRITE:
5940 ret = io_write(req, force_nonblock, cs);
5942 case IORING_OP_FSYNC:
5943 ret = io_fsync(req, force_nonblock);
5945 case IORING_OP_POLL_ADD:
5946 ret = io_poll_add(req);
5948 case IORING_OP_POLL_REMOVE:
5949 ret = io_poll_remove(req);
5951 case IORING_OP_SYNC_FILE_RANGE:
5952 ret = io_sync_file_range(req, force_nonblock);
5954 case IORING_OP_SENDMSG:
5955 ret = io_sendmsg(req, force_nonblock, cs);
5957 case IORING_OP_SEND:
5958 ret = io_send(req, force_nonblock, cs);
5960 case IORING_OP_RECVMSG:
5961 ret = io_recvmsg(req, force_nonblock, cs);
5963 case IORING_OP_RECV:
5964 ret = io_recv(req, force_nonblock, cs);
5966 case IORING_OP_TIMEOUT:
5967 ret = io_timeout(req);
5969 case IORING_OP_TIMEOUT_REMOVE:
5970 ret = io_timeout_remove(req);
5972 case IORING_OP_ACCEPT:
5973 ret = io_accept(req, force_nonblock, cs);
5975 case IORING_OP_CONNECT:
5976 ret = io_connect(req, force_nonblock, cs);
5978 case IORING_OP_ASYNC_CANCEL:
5979 ret = io_async_cancel(req);
5981 case IORING_OP_FALLOCATE:
5982 ret = io_fallocate(req, force_nonblock);
5984 case IORING_OP_OPENAT:
5985 ret = io_openat(req, force_nonblock);
5987 case IORING_OP_CLOSE:
5988 ret = io_close(req, force_nonblock, cs);
5990 case IORING_OP_FILES_UPDATE:
5991 ret = io_files_update(req, force_nonblock, cs);
5993 case IORING_OP_STATX:
5994 ret = io_statx(req, force_nonblock);
5996 case IORING_OP_FADVISE:
5997 ret = io_fadvise(req, force_nonblock);
5999 case IORING_OP_MADVISE:
6000 ret = io_madvise(req, force_nonblock);
6002 case IORING_OP_OPENAT2:
6003 ret = io_openat2(req, force_nonblock);
6005 case IORING_OP_EPOLL_CTL:
6006 ret = io_epoll_ctl(req, force_nonblock, cs);
6008 case IORING_OP_SPLICE:
6009 ret = io_splice(req, force_nonblock);
6011 case IORING_OP_PROVIDE_BUFFERS:
6012 ret = io_provide_buffers(req, force_nonblock, cs);
6014 case IORING_OP_REMOVE_BUFFERS:
6015 ret = io_remove_buffers(req, force_nonblock, cs);
6018 ret = io_tee(req, force_nonblock);
6028 /* If the op doesn't have a file, we're not polling for it */
6029 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6030 const bool in_async = io_wq_current_is_worker();
6032 /* workqueue context doesn't hold uring_lock, grab it now */
6034 mutex_lock(&ctx->uring_lock);
6036 io_iopoll_req_issued(req);
6039 mutex_unlock(&ctx->uring_lock);
6045 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6047 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6048 struct io_kiocb *timeout;
6051 timeout = io_prep_linked_timeout(req);
6053 io_queue_linked_timeout(timeout);
6055 /* if NO_CANCEL is set, we must still run the work */
6056 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6057 IO_WQ_WORK_CANCEL) {
6063 ret = io_issue_sqe(req, false, NULL);
6065 * We can get EAGAIN for polled IO even though we're
6066 * forcing a sync submission from here, since we can't
6067 * wait for request slots on the block side.
6076 req_set_fail_links(req);
6077 io_req_complete(req, ret);
6080 return io_steal_work(req);
6083 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6086 struct fixed_file_table *table;
6088 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6089 return table->files[index & IORING_FILE_TABLE_MASK];
6092 static struct file *io_file_get(struct io_submit_state *state,
6093 struct io_kiocb *req, int fd, bool fixed)
6095 struct io_ring_ctx *ctx = req->ctx;
6099 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6101 fd = array_index_nospec(fd, ctx->nr_user_files);
6102 file = io_file_from_index(ctx, fd);
6104 req->fixed_file_refs = &ctx->file_data->node->refs;
6105 percpu_ref_get(req->fixed_file_refs);
6108 trace_io_uring_file_get(ctx, fd);
6109 file = __io_file_get(state, fd);
6115 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6120 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6121 if (unlikely(!fixed && io_async_submit(req->ctx)))
6124 req->file = io_file_get(state, req, fd, fixed);
6125 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6130 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6132 struct io_timeout_data *data = container_of(timer,
6133 struct io_timeout_data, timer);
6134 struct io_kiocb *req = data->req;
6135 struct io_ring_ctx *ctx = req->ctx;
6136 struct io_kiocb *prev = NULL;
6137 unsigned long flags;
6139 spin_lock_irqsave(&ctx->completion_lock, flags);
6142 * We don't expect the list to be empty, that will only happen if we
6143 * race with the completion of the linked work.
6145 if (!list_empty(&req->link_list)) {
6146 prev = list_entry(req->link_list.prev, struct io_kiocb,
6148 if (refcount_inc_not_zero(&prev->refs))
6149 list_del_init(&req->link_list);
6154 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6157 req_set_fail_links(prev);
6158 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6161 io_req_complete(req, -ETIME);
6163 return HRTIMER_NORESTART;
6166 static void __io_queue_linked_timeout(struct io_kiocb *req)
6169 * If the list is now empty, then our linked request finished before
6170 * we got a chance to setup the timer
6172 if (!list_empty(&req->link_list)) {
6173 struct io_timeout_data *data = req->async_data;
6175 data->timer.function = io_link_timeout_fn;
6176 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6181 static void io_queue_linked_timeout(struct io_kiocb *req)
6183 struct io_ring_ctx *ctx = req->ctx;
6185 spin_lock_irq(&ctx->completion_lock);
6186 __io_queue_linked_timeout(req);
6187 spin_unlock_irq(&ctx->completion_lock);
6189 /* drop submission reference */
6193 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6195 struct io_kiocb *nxt;
6197 if (!(req->flags & REQ_F_LINK_HEAD))
6199 if (req->flags & REQ_F_LINK_TIMEOUT)
6202 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6204 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6207 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6208 req->flags |= REQ_F_LINK_TIMEOUT;
6212 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6214 struct io_kiocb *linked_timeout;
6215 const struct cred *old_creds = NULL;
6219 linked_timeout = io_prep_linked_timeout(req);
6221 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6222 (req->work.flags & IO_WQ_WORK_CREDS) &&
6223 req->work.identity->creds != current_cred()) {
6225 revert_creds(old_creds);
6226 if (old_creds == req->work.identity->creds)
6227 old_creds = NULL; /* restored original creds */
6229 old_creds = override_creds(req->work.identity->creds);
6232 ret = io_issue_sqe(req, true, cs);
6235 * We async punt it if the file wasn't marked NOWAIT, or if the file
6236 * doesn't support non-blocking read/write attempts
6238 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6239 if (!io_arm_poll_handler(req)) {
6241 * Queued up for async execution, worker will release
6242 * submit reference when the iocb is actually submitted.
6244 io_queue_async_work(req);
6248 io_queue_linked_timeout(linked_timeout);
6249 } else if (likely(!ret)) {
6250 /* drop submission reference */
6251 req = io_put_req_find_next(req);
6253 io_queue_linked_timeout(linked_timeout);
6256 if (!(req->flags & REQ_F_FORCE_ASYNC))
6258 io_queue_async_work(req);
6261 /* un-prep timeout, so it'll be killed as any other linked */
6262 req->flags &= ~REQ_F_LINK_TIMEOUT;
6263 req_set_fail_links(req);
6265 io_req_complete(req, ret);
6269 revert_creds(old_creds);
6272 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6273 struct io_comp_state *cs)
6277 ret = io_req_defer(req, sqe);
6279 if (ret != -EIOCBQUEUED) {
6281 req_set_fail_links(req);
6283 io_req_complete(req, ret);
6285 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6286 if (!req->async_data) {
6287 ret = io_req_defer_prep(req, sqe);
6291 io_queue_async_work(req);
6294 ret = io_req_prep(req, sqe);
6298 __io_queue_sqe(req, cs);
6302 static inline void io_queue_link_head(struct io_kiocb *req,
6303 struct io_comp_state *cs)
6305 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6307 io_req_complete(req, -ECANCELED);
6309 io_queue_sqe(req, NULL, cs);
6312 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6313 struct io_kiocb **link, struct io_comp_state *cs)
6315 struct io_ring_ctx *ctx = req->ctx;
6319 * If we already have a head request, queue this one for async
6320 * submittal once the head completes. If we don't have a head but
6321 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6322 * submitted sync once the chain is complete. If none of those
6323 * conditions are true (normal request), then just queue it.
6326 struct io_kiocb *head = *link;
6329 * Taking sequential execution of a link, draining both sides
6330 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6331 * requests in the link. So, it drains the head and the
6332 * next after the link request. The last one is done via
6333 * drain_next flag to persist the effect across calls.
6335 if (req->flags & REQ_F_IO_DRAIN) {
6336 head->flags |= REQ_F_IO_DRAIN;
6337 ctx->drain_next = 1;
6339 ret = io_req_defer_prep(req, sqe);
6340 if (unlikely(ret)) {
6341 /* fail even hard links since we don't submit */
6342 head->flags |= REQ_F_FAIL_LINK;
6345 trace_io_uring_link(ctx, req, head);
6346 list_add_tail(&req->link_list, &head->link_list);
6348 /* last request of a link, enqueue the link */
6349 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6350 io_queue_link_head(head, cs);
6354 if (unlikely(ctx->drain_next)) {
6355 req->flags |= REQ_F_IO_DRAIN;
6356 ctx->drain_next = 0;
6358 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6359 req->flags |= REQ_F_LINK_HEAD;
6360 INIT_LIST_HEAD(&req->link_list);
6362 ret = io_req_defer_prep(req, sqe);
6364 req->flags |= REQ_F_FAIL_LINK;
6367 io_queue_sqe(req, sqe, cs);
6375 * Batched submission is done, ensure local IO is flushed out.
6377 static void io_submit_state_end(struct io_submit_state *state)
6379 if (!list_empty(&state->comp.list))
6380 io_submit_flush_completions(&state->comp);
6381 blk_finish_plug(&state->plug);
6382 io_state_file_put(state);
6383 if (state->free_reqs)
6384 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6388 * Start submission side cache.
6390 static void io_submit_state_start(struct io_submit_state *state,
6391 struct io_ring_ctx *ctx, unsigned int max_ios)
6393 blk_start_plug(&state->plug);
6395 INIT_LIST_HEAD(&state->comp.list);
6396 state->comp.ctx = ctx;
6397 state->free_reqs = 0;
6399 state->ios_left = max_ios;
6402 static void io_commit_sqring(struct io_ring_ctx *ctx)
6404 struct io_rings *rings = ctx->rings;
6407 * Ensure any loads from the SQEs are done at this point,
6408 * since once we write the new head, the application could
6409 * write new data to them.
6411 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6415 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6416 * that is mapped by userspace. This means that care needs to be taken to
6417 * ensure that reads are stable, as we cannot rely on userspace always
6418 * being a good citizen. If members of the sqe are validated and then later
6419 * used, it's important that those reads are done through READ_ONCE() to
6420 * prevent a re-load down the line.
6422 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6424 u32 *sq_array = ctx->sq_array;
6428 * The cached sq head (or cq tail) serves two purposes:
6430 * 1) allows us to batch the cost of updating the user visible
6432 * 2) allows the kernel side to track the head on its own, even
6433 * though the application is the one updating it.
6435 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6436 if (likely(head < ctx->sq_entries))
6437 return &ctx->sq_sqes[head];
6439 /* drop invalid entries */
6440 ctx->cached_sq_dropped++;
6441 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6445 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6447 ctx->cached_sq_head++;
6451 * Check SQE restrictions (opcode and flags).
6453 * Returns 'true' if SQE is allowed, 'false' otherwise.
6455 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6456 struct io_kiocb *req,
6457 unsigned int sqe_flags)
6459 if (!ctx->restricted)
6462 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6465 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6466 ctx->restrictions.sqe_flags_required)
6469 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6470 ctx->restrictions.sqe_flags_required))
6476 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6477 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6478 IOSQE_BUFFER_SELECT)
6480 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6481 const struct io_uring_sqe *sqe,
6482 struct io_submit_state *state)
6484 unsigned int sqe_flags;
6487 req->opcode = READ_ONCE(sqe->opcode);
6488 req->user_data = READ_ONCE(sqe->user_data);
6489 req->async_data = NULL;
6493 /* one is dropped after submission, the other at completion */
6494 refcount_set(&req->refs, 2);
6495 req->task = current;
6498 if (unlikely(req->opcode >= IORING_OP_LAST))
6501 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6504 sqe_flags = READ_ONCE(sqe->flags);
6505 /* enforce forwards compatibility on users */
6506 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6509 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6512 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6513 !io_op_defs[req->opcode].buffer_select)
6516 id = READ_ONCE(sqe->personality);
6518 struct io_identity *iod;
6520 iod = idr_find(&ctx->personality_idr, id);
6523 refcount_inc(&iod->count);
6525 __io_req_init_async(req);
6526 get_cred(iod->creds);
6527 req->work.identity = iod;
6528 req->work.flags |= IO_WQ_WORK_CREDS;
6531 /* same numerical values with corresponding REQ_F_*, safe to copy */
6532 req->flags |= sqe_flags;
6534 if (!io_op_defs[req->opcode].needs_file)
6537 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6542 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6544 struct io_submit_state state;
6545 struct io_kiocb *link = NULL;
6546 int i, submitted = 0;
6548 /* if we have a backlog and couldn't flush it all, return BUSY */
6549 if (test_bit(0, &ctx->sq_check_overflow)) {
6550 if (!list_empty(&ctx->cq_overflow_list) &&
6551 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6555 /* make sure SQ entry isn't read before tail */
6556 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6558 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6561 percpu_counter_add(¤t->io_uring->inflight, nr);
6562 refcount_add(nr, ¤t->usage);
6564 io_submit_state_start(&state, ctx, nr);
6566 for (i = 0; i < nr; i++) {
6567 const struct io_uring_sqe *sqe;
6568 struct io_kiocb *req;
6571 sqe = io_get_sqe(ctx);
6572 if (unlikely(!sqe)) {
6573 io_consume_sqe(ctx);
6576 req = io_alloc_req(ctx, &state);
6577 if (unlikely(!req)) {
6579 submitted = -EAGAIN;
6582 io_consume_sqe(ctx);
6583 /* will complete beyond this point, count as submitted */
6586 err = io_init_req(ctx, req, sqe, &state);
6587 if (unlikely(err)) {
6590 io_req_complete(req, err);
6594 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6595 true, io_async_submit(ctx));
6596 err = io_submit_sqe(req, sqe, &link, &state.comp);
6601 if (unlikely(submitted != nr)) {
6602 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6603 struct io_uring_task *tctx = current->io_uring;
6604 int unused = nr - ref_used;
6606 percpu_ref_put_many(&ctx->refs, unused);
6607 percpu_counter_sub(&tctx->inflight, unused);
6608 put_task_struct_many(current, unused);
6611 io_queue_link_head(link, &state.comp);
6612 io_submit_state_end(&state);
6614 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6615 io_commit_sqring(ctx);
6620 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6622 /* Tell userspace we may need a wakeup call */
6623 spin_lock_irq(&ctx->completion_lock);
6624 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6625 spin_unlock_irq(&ctx->completion_lock);
6628 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6630 spin_lock_irq(&ctx->completion_lock);
6631 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6632 spin_unlock_irq(&ctx->completion_lock);
6635 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6636 int sync, void *key)
6638 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6641 ret = autoremove_wake_function(wqe, mode, sync, key);
6643 unsigned long flags;
6645 spin_lock_irqsave(&ctx->completion_lock, flags);
6646 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6647 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6658 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6659 unsigned long start_jiffies, bool cap_entries)
6661 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6662 struct io_sq_data *sqd = ctx->sq_data;
6663 unsigned int to_submit;
6667 if (!list_empty(&ctx->iopoll_list)) {
6668 unsigned nr_events = 0;
6670 mutex_lock(&ctx->uring_lock);
6671 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6672 io_do_iopoll(ctx, &nr_events, 0);
6673 mutex_unlock(&ctx->uring_lock);
6676 to_submit = io_sqring_entries(ctx);
6679 * If submit got -EBUSY, flag us as needing the application
6680 * to enter the kernel to reap and flush events.
6682 if (!to_submit || ret == -EBUSY || need_resched()) {
6684 * Drop cur_mm before scheduling, we can't hold it for
6685 * long periods (or over schedule()). Do this before
6686 * adding ourselves to the waitqueue, as the unuse/drop
6689 io_sq_thread_drop_mm();
6692 * We're polling. If we're within the defined idle
6693 * period, then let us spin without work before going
6694 * to sleep. The exception is if we got EBUSY doing
6695 * more IO, we should wait for the application to
6696 * reap events and wake us up.
6698 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6699 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6700 !percpu_ref_is_dying(&ctx->refs)))
6703 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6704 TASK_INTERRUPTIBLE);
6707 * While doing polled IO, before going to sleep, we need
6708 * to check if there are new reqs added to iopoll_list,
6709 * it is because reqs may have been punted to io worker
6710 * and will be added to iopoll_list later, hence check
6711 * the iopoll_list again.
6713 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6714 !list_empty_careful(&ctx->iopoll_list)) {
6715 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6719 to_submit = io_sqring_entries(ctx);
6720 if (!to_submit || ret == -EBUSY)
6724 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6725 io_ring_clear_wakeup_flag(ctx);
6727 /* if we're handling multiple rings, cap submit size for fairness */
6728 if (cap_entries && to_submit > 8)
6731 mutex_lock(&ctx->uring_lock);
6732 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6733 ret = io_submit_sqes(ctx, to_submit);
6734 mutex_unlock(&ctx->uring_lock);
6736 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6737 wake_up(&ctx->sqo_sq_wait);
6739 return SQT_DID_WORK;
6742 static void io_sqd_init_new(struct io_sq_data *sqd)
6744 struct io_ring_ctx *ctx;
6746 while (!list_empty(&sqd->ctx_new_list)) {
6747 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6748 init_wait(&ctx->sqo_wait_entry);
6749 ctx->sqo_wait_entry.func = io_sq_wake_function;
6750 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6751 complete(&ctx->sq_thread_comp);
6755 static int io_sq_thread(void *data)
6757 struct cgroup_subsys_state *cur_css = NULL;
6758 const struct cred *old_cred = NULL;
6759 struct io_sq_data *sqd = data;
6760 struct io_ring_ctx *ctx;
6761 unsigned long start_jiffies;
6763 start_jiffies = jiffies;
6764 while (!kthread_should_stop()) {
6765 enum sq_ret ret = 0;
6769 * Any changes to the sqd lists are synchronized through the
6770 * kthread parking. This synchronizes the thread vs users,
6771 * the users are synchronized on the sqd->ctx_lock.
6773 if (kthread_should_park())
6776 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6777 io_sqd_init_new(sqd);
6779 cap_entries = !list_is_singular(&sqd->ctx_list);
6781 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6782 if (current->cred != ctx->creds) {
6784 revert_creds(old_cred);
6785 old_cred = override_creds(ctx->creds);
6787 io_sq_thread_associate_blkcg(ctx, &cur_css);
6789 current->loginuid = ctx->loginuid;
6790 current->sessionid = ctx->sessionid;
6793 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6795 io_sq_thread_drop_mm();
6798 if (ret & SQT_SPIN) {
6801 } else if (ret == SQT_IDLE) {
6802 if (kthread_should_park())
6804 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6805 io_ring_set_wakeup_flag(ctx);
6807 start_jiffies = jiffies;
6808 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6809 io_ring_clear_wakeup_flag(ctx);
6816 io_sq_thread_unassociate_blkcg();
6818 revert_creds(old_cred);
6825 struct io_wait_queue {
6826 struct wait_queue_entry wq;
6827 struct io_ring_ctx *ctx;
6829 unsigned nr_timeouts;
6832 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6834 struct io_ring_ctx *ctx = iowq->ctx;
6837 * Wake up if we have enough events, or if a timeout occurred since we
6838 * started waiting. For timeouts, we always want to return to userspace,
6839 * regardless of event count.
6841 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6842 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6845 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6846 int wake_flags, void *key)
6848 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6851 /* use noflush == true, as we can't safely rely on locking context */
6852 if (!io_should_wake(iowq, true))
6855 return autoremove_wake_function(curr, mode, wake_flags, key);
6858 static int io_run_task_work_sig(void)
6860 if (io_run_task_work())
6862 if (!signal_pending(current))
6864 if (current->jobctl & JOBCTL_TASK_WORK) {
6865 spin_lock_irq(¤t->sighand->siglock);
6866 current->jobctl &= ~JOBCTL_TASK_WORK;
6867 recalc_sigpending();
6868 spin_unlock_irq(¤t->sighand->siglock);
6875 * Wait until events become available, if we don't already have some. The
6876 * application must reap them itself, as they reside on the shared cq ring.
6878 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6879 const sigset_t __user *sig, size_t sigsz)
6881 struct io_wait_queue iowq = {
6884 .func = io_wake_function,
6885 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6888 .to_wait = min_events,
6890 struct io_rings *rings = ctx->rings;
6894 if (io_cqring_events(ctx, false) >= min_events)
6896 if (!io_run_task_work())
6901 #ifdef CONFIG_COMPAT
6902 if (in_compat_syscall())
6903 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6907 ret = set_user_sigmask(sig, sigsz);
6913 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6914 trace_io_uring_cqring_wait(ctx, min_events);
6916 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6917 TASK_INTERRUPTIBLE);
6918 /* make sure we run task_work before checking for signals */
6919 ret = io_run_task_work_sig();
6924 if (io_should_wake(&iowq, false))
6928 finish_wait(&ctx->wait, &iowq.wq);
6930 restore_saved_sigmask_unless(ret == -EINTR);
6932 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6935 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6937 #if defined(CONFIG_UNIX)
6938 if (ctx->ring_sock) {
6939 struct sock *sock = ctx->ring_sock->sk;
6940 struct sk_buff *skb;
6942 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6948 for (i = 0; i < ctx->nr_user_files; i++) {
6951 file = io_file_from_index(ctx, i);
6958 static void io_file_ref_kill(struct percpu_ref *ref)
6960 struct fixed_file_data *data;
6962 data = container_of(ref, struct fixed_file_data, refs);
6963 complete(&data->done);
6966 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6968 struct fixed_file_data *data = ctx->file_data;
6969 struct fixed_file_ref_node *ref_node = NULL;
6970 unsigned nr_tables, i;
6975 spin_lock(&data->lock);
6976 ref_node = data->node;
6977 spin_unlock(&data->lock);
6979 percpu_ref_kill(&ref_node->refs);
6981 percpu_ref_kill(&data->refs);
6983 /* wait for all refs nodes to complete */
6984 flush_delayed_work(&ctx->file_put_work);
6985 wait_for_completion(&data->done);
6987 __io_sqe_files_unregister(ctx);
6988 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6989 for (i = 0; i < nr_tables; i++)
6990 kfree(data->table[i].files);
6992 percpu_ref_exit(&data->refs);
6994 ctx->file_data = NULL;
6995 ctx->nr_user_files = 0;
6999 static void io_put_sq_data(struct io_sq_data *sqd)
7001 if (refcount_dec_and_test(&sqd->refs)) {
7003 * The park is a bit of a work-around, without it we get
7004 * warning spews on shutdown with SQPOLL set and affinity
7005 * set to a single CPU.
7008 kthread_park(sqd->thread);
7009 kthread_stop(sqd->thread);
7016 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7018 struct io_ring_ctx *ctx_attach;
7019 struct io_sq_data *sqd;
7022 f = fdget(p->wq_fd);
7024 return ERR_PTR(-ENXIO);
7025 if (f.file->f_op != &io_uring_fops) {
7027 return ERR_PTR(-EINVAL);
7030 ctx_attach = f.file->private_data;
7031 sqd = ctx_attach->sq_data;
7034 return ERR_PTR(-EINVAL);
7037 refcount_inc(&sqd->refs);
7042 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7044 struct io_sq_data *sqd;
7046 if (p->flags & IORING_SETUP_ATTACH_WQ)
7047 return io_attach_sq_data(p);
7049 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7051 return ERR_PTR(-ENOMEM);
7053 refcount_set(&sqd->refs, 1);
7054 INIT_LIST_HEAD(&sqd->ctx_list);
7055 INIT_LIST_HEAD(&sqd->ctx_new_list);
7056 mutex_init(&sqd->ctx_lock);
7057 mutex_init(&sqd->lock);
7058 init_waitqueue_head(&sqd->wait);
7062 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7063 __releases(&sqd->lock)
7067 kthread_unpark(sqd->thread);
7068 mutex_unlock(&sqd->lock);
7071 static void io_sq_thread_park(struct io_sq_data *sqd)
7072 __acquires(&sqd->lock)
7076 mutex_lock(&sqd->lock);
7077 kthread_park(sqd->thread);
7080 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7082 struct io_sq_data *sqd = ctx->sq_data;
7087 * We may arrive here from the error branch in
7088 * io_sq_offload_create() where the kthread is created
7089 * without being waked up, thus wake it up now to make
7090 * sure the wait will complete.
7092 wake_up_process(sqd->thread);
7093 wait_for_completion(&ctx->sq_thread_comp);
7095 io_sq_thread_park(sqd);
7098 mutex_lock(&sqd->ctx_lock);
7099 list_del(&ctx->sqd_list);
7100 mutex_unlock(&sqd->ctx_lock);
7103 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7104 io_sq_thread_unpark(sqd);
7107 io_put_sq_data(sqd);
7108 ctx->sq_data = NULL;
7112 static void io_finish_async(struct io_ring_ctx *ctx)
7114 io_sq_thread_stop(ctx);
7117 io_wq_destroy(ctx->io_wq);
7122 #if defined(CONFIG_UNIX)
7124 * Ensure the UNIX gc is aware of our file set, so we are certain that
7125 * the io_uring can be safely unregistered on process exit, even if we have
7126 * loops in the file referencing.
7128 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7130 struct sock *sk = ctx->ring_sock->sk;
7131 struct scm_fp_list *fpl;
7132 struct sk_buff *skb;
7135 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7139 skb = alloc_skb(0, GFP_KERNEL);
7148 fpl->user = get_uid(ctx->user);
7149 for (i = 0; i < nr; i++) {
7150 struct file *file = io_file_from_index(ctx, i + offset);
7154 fpl->fp[nr_files] = get_file(file);
7155 unix_inflight(fpl->user, fpl->fp[nr_files]);
7160 fpl->max = SCM_MAX_FD;
7161 fpl->count = nr_files;
7162 UNIXCB(skb).fp = fpl;
7163 skb->destructor = unix_destruct_scm;
7164 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7165 skb_queue_head(&sk->sk_receive_queue, skb);
7167 for (i = 0; i < nr_files; i++)
7178 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7179 * causes regular reference counting to break down. We rely on the UNIX
7180 * garbage collection to take care of this problem for us.
7182 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7184 unsigned left, total;
7188 left = ctx->nr_user_files;
7190 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7192 ret = __io_sqe_files_scm(ctx, this_files, total);
7196 total += this_files;
7202 while (total < ctx->nr_user_files) {
7203 struct file *file = io_file_from_index(ctx, total);
7213 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7219 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7220 unsigned nr_tables, unsigned nr_files)
7224 for (i = 0; i < nr_tables; i++) {
7225 struct fixed_file_table *table = &file_data->table[i];
7226 unsigned this_files;
7228 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7229 table->files = kcalloc(this_files, sizeof(struct file *),
7233 nr_files -= this_files;
7239 for (i = 0; i < nr_tables; i++) {
7240 struct fixed_file_table *table = &file_data->table[i];
7241 kfree(table->files);
7246 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7248 #if defined(CONFIG_UNIX)
7249 struct sock *sock = ctx->ring_sock->sk;
7250 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7251 struct sk_buff *skb;
7254 __skb_queue_head_init(&list);
7257 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7258 * remove this entry and rearrange the file array.
7260 skb = skb_dequeue(head);
7262 struct scm_fp_list *fp;
7264 fp = UNIXCB(skb).fp;
7265 for (i = 0; i < fp->count; i++) {
7268 if (fp->fp[i] != file)
7271 unix_notinflight(fp->user, fp->fp[i]);
7272 left = fp->count - 1 - i;
7274 memmove(&fp->fp[i], &fp->fp[i + 1],
7275 left * sizeof(struct file *));
7282 __skb_queue_tail(&list, skb);
7292 __skb_queue_tail(&list, skb);
7294 skb = skb_dequeue(head);
7297 if (skb_peek(&list)) {
7298 spin_lock_irq(&head->lock);
7299 while ((skb = __skb_dequeue(&list)) != NULL)
7300 __skb_queue_tail(head, skb);
7301 spin_unlock_irq(&head->lock);
7308 struct io_file_put {
7309 struct list_head list;
7313 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7315 struct fixed_file_data *file_data = ref_node->file_data;
7316 struct io_ring_ctx *ctx = file_data->ctx;
7317 struct io_file_put *pfile, *tmp;
7319 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7320 list_del(&pfile->list);
7321 io_ring_file_put(ctx, pfile->file);
7325 percpu_ref_exit(&ref_node->refs);
7327 percpu_ref_put(&file_data->refs);
7330 static void io_file_put_work(struct work_struct *work)
7332 struct io_ring_ctx *ctx;
7333 struct llist_node *node;
7335 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7336 node = llist_del_all(&ctx->file_put_llist);
7339 struct fixed_file_ref_node *ref_node;
7340 struct llist_node *next = node->next;
7342 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7343 __io_file_put_work(ref_node);
7348 static void io_file_data_ref_zero(struct percpu_ref *ref)
7350 struct fixed_file_ref_node *ref_node;
7351 struct fixed_file_data *data;
7352 struct io_ring_ctx *ctx;
7353 bool first_add = false;
7356 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7357 data = ref_node->file_data;
7360 spin_lock(&data->lock);
7361 ref_node->done = true;
7363 while (!list_empty(&data->ref_list)) {
7364 ref_node = list_first_entry(&data->ref_list,
7365 struct fixed_file_ref_node, node);
7366 /* recycle ref nodes in order */
7367 if (!ref_node->done)
7369 list_del(&ref_node->node);
7370 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7372 spin_unlock(&data->lock);
7374 if (percpu_ref_is_dying(&data->refs))
7378 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7380 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7383 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7384 struct io_ring_ctx *ctx)
7386 struct fixed_file_ref_node *ref_node;
7388 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7390 return ERR_PTR(-ENOMEM);
7392 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7395 return ERR_PTR(-ENOMEM);
7397 INIT_LIST_HEAD(&ref_node->node);
7398 INIT_LIST_HEAD(&ref_node->file_list);
7399 ref_node->file_data = ctx->file_data;
7400 ref_node->done = false;
7404 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7406 percpu_ref_exit(&ref_node->refs);
7410 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7413 __s32 __user *fds = (__s32 __user *) arg;
7414 unsigned nr_tables, i;
7416 int fd, ret = -ENOMEM;
7417 struct fixed_file_ref_node *ref_node;
7418 struct fixed_file_data *file_data;
7424 if (nr_args > IORING_MAX_FIXED_FILES)
7427 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7430 file_data->ctx = ctx;
7431 init_completion(&file_data->done);
7432 INIT_LIST_HEAD(&file_data->ref_list);
7433 spin_lock_init(&file_data->lock);
7435 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7436 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7438 if (!file_data->table)
7441 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7442 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7445 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7447 ctx->file_data = file_data;
7449 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7450 struct fixed_file_table *table;
7453 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7457 /* allow sparse sets */
7467 * Don't allow io_uring instances to be registered. If UNIX
7468 * isn't enabled, then this causes a reference cycle and this
7469 * instance can never get freed. If UNIX is enabled we'll
7470 * handle it just fine, but there's still no point in allowing
7471 * a ring fd as it doesn't support regular read/write anyway.
7473 if (file->f_op == &io_uring_fops) {
7477 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7478 index = i & IORING_FILE_TABLE_MASK;
7479 table->files[index] = file;
7482 ret = io_sqe_files_scm(ctx);
7484 io_sqe_files_unregister(ctx);
7488 ref_node = alloc_fixed_file_ref_node(ctx);
7489 if (IS_ERR(ref_node)) {
7490 io_sqe_files_unregister(ctx);
7491 return PTR_ERR(ref_node);
7494 file_data->node = ref_node;
7495 spin_lock(&file_data->lock);
7496 list_add_tail(&ref_node->node, &file_data->ref_list);
7497 spin_unlock(&file_data->lock);
7498 percpu_ref_get(&file_data->refs);
7501 for (i = 0; i < ctx->nr_user_files; i++) {
7502 file = io_file_from_index(ctx, i);
7506 for (i = 0; i < nr_tables; i++)
7507 kfree(file_data->table[i].files);
7508 ctx->nr_user_files = 0;
7510 percpu_ref_exit(&file_data->refs);
7512 kfree(file_data->table);
7514 ctx->file_data = NULL;
7518 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7521 #if defined(CONFIG_UNIX)
7522 struct sock *sock = ctx->ring_sock->sk;
7523 struct sk_buff_head *head = &sock->sk_receive_queue;
7524 struct sk_buff *skb;
7527 * See if we can merge this file into an existing skb SCM_RIGHTS
7528 * file set. If there's no room, fall back to allocating a new skb
7529 * and filling it in.
7531 spin_lock_irq(&head->lock);
7532 skb = skb_peek(head);
7534 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7536 if (fpl->count < SCM_MAX_FD) {
7537 __skb_unlink(skb, head);
7538 spin_unlock_irq(&head->lock);
7539 fpl->fp[fpl->count] = get_file(file);
7540 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7542 spin_lock_irq(&head->lock);
7543 __skb_queue_head(head, skb);
7548 spin_unlock_irq(&head->lock);
7555 return __io_sqe_files_scm(ctx, 1, index);
7561 static int io_queue_file_removal(struct fixed_file_data *data,
7564 struct io_file_put *pfile;
7565 struct fixed_file_ref_node *ref_node = data->node;
7567 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7572 list_add(&pfile->list, &ref_node->file_list);
7577 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7578 struct io_uring_files_update *up,
7581 struct fixed_file_data *data = ctx->file_data;
7582 struct fixed_file_ref_node *ref_node;
7587 bool needs_switch = false;
7589 if (check_add_overflow(up->offset, nr_args, &done))
7591 if (done > ctx->nr_user_files)
7594 ref_node = alloc_fixed_file_ref_node(ctx);
7595 if (IS_ERR(ref_node))
7596 return PTR_ERR(ref_node);
7599 fds = u64_to_user_ptr(up->fds);
7601 struct fixed_file_table *table;
7605 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7609 i = array_index_nospec(up->offset, ctx->nr_user_files);
7610 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7611 index = i & IORING_FILE_TABLE_MASK;
7612 if (table->files[index]) {
7613 file = table->files[index];
7614 err = io_queue_file_removal(data, file);
7617 table->files[index] = NULL;
7618 needs_switch = true;
7627 * Don't allow io_uring instances to be registered. If
7628 * UNIX isn't enabled, then this causes a reference
7629 * cycle and this instance can never get freed. If UNIX
7630 * is enabled we'll handle it just fine, but there's
7631 * still no point in allowing a ring fd as it doesn't
7632 * support regular read/write anyway.
7634 if (file->f_op == &io_uring_fops) {
7639 table->files[index] = file;
7640 err = io_sqe_file_register(ctx, file, i);
7642 table->files[index] = NULL;
7653 percpu_ref_kill(&data->node->refs);
7654 spin_lock(&data->lock);
7655 list_add_tail(&ref_node->node, &data->ref_list);
7656 data->node = ref_node;
7657 spin_unlock(&data->lock);
7658 percpu_ref_get(&ctx->file_data->refs);
7660 destroy_fixed_file_ref_node(ref_node);
7662 return done ? done : err;
7665 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7668 struct io_uring_files_update up;
7670 if (!ctx->file_data)
7674 if (copy_from_user(&up, arg, sizeof(up)))
7679 return __io_sqe_files_update(ctx, &up, nr_args);
7682 static void io_free_work(struct io_wq_work *work)
7684 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7686 /* Consider that io_steal_work() relies on this ref */
7690 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7691 struct io_uring_params *p)
7693 struct io_wq_data data;
7695 struct io_ring_ctx *ctx_attach;
7696 unsigned int concurrency;
7699 data.user = ctx->user;
7700 data.free_work = io_free_work;
7701 data.do_work = io_wq_submit_work;
7703 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7704 /* Do QD, or 4 * CPUS, whatever is smallest */
7705 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7707 ctx->io_wq = io_wq_create(concurrency, &data);
7708 if (IS_ERR(ctx->io_wq)) {
7709 ret = PTR_ERR(ctx->io_wq);
7715 f = fdget(p->wq_fd);
7719 if (f.file->f_op != &io_uring_fops) {
7724 ctx_attach = f.file->private_data;
7725 /* @io_wq is protected by holding the fd */
7726 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7731 ctx->io_wq = ctx_attach->io_wq;
7737 static int io_uring_alloc_task_context(struct task_struct *task)
7739 struct io_uring_task *tctx;
7742 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7743 if (unlikely(!tctx))
7746 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7747 if (unlikely(ret)) {
7753 init_waitqueue_head(&tctx->wait);
7755 atomic_set(&tctx->in_idle, 0);
7756 tctx->sqpoll = false;
7757 io_init_identity(&tctx->__identity);
7758 tctx->identity = &tctx->__identity;
7759 task->io_uring = tctx;
7763 void __io_uring_free(struct task_struct *tsk)
7765 struct io_uring_task *tctx = tsk->io_uring;
7767 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7768 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7769 if (tctx->identity != &tctx->__identity)
7770 kfree(tctx->identity);
7771 percpu_counter_destroy(&tctx->inflight);
7773 tsk->io_uring = NULL;
7776 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7777 struct io_uring_params *p)
7781 if (ctx->flags & IORING_SETUP_SQPOLL) {
7782 struct io_sq_data *sqd;
7785 if (!capable(CAP_SYS_ADMIN))
7788 sqd = io_get_sq_data(p);
7795 io_sq_thread_park(sqd);
7796 mutex_lock(&sqd->ctx_lock);
7797 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7798 mutex_unlock(&sqd->ctx_lock);
7799 io_sq_thread_unpark(sqd);
7801 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7802 if (!ctx->sq_thread_idle)
7803 ctx->sq_thread_idle = HZ;
7808 if (p->flags & IORING_SETUP_SQ_AFF) {
7809 int cpu = p->sq_thread_cpu;
7812 if (cpu >= nr_cpu_ids)
7814 if (!cpu_online(cpu))
7817 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7818 cpu, "io_uring-sq");
7820 sqd->thread = kthread_create(io_sq_thread, sqd,
7823 if (IS_ERR(sqd->thread)) {
7824 ret = PTR_ERR(sqd->thread);
7828 ret = io_uring_alloc_task_context(sqd->thread);
7831 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7832 /* Can't have SQ_AFF without SQPOLL */
7838 ret = io_init_wq_offload(ctx, p);
7844 io_finish_async(ctx);
7848 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7850 struct io_sq_data *sqd = ctx->sq_data;
7852 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7853 wake_up_process(sqd->thread);
7856 static inline void __io_unaccount_mem(struct user_struct *user,
7857 unsigned long nr_pages)
7859 atomic_long_sub(nr_pages, &user->locked_vm);
7862 static inline int __io_account_mem(struct user_struct *user,
7863 unsigned long nr_pages)
7865 unsigned long page_limit, cur_pages, new_pages;
7867 /* Don't allow more pages than we can safely lock */
7868 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7871 cur_pages = atomic_long_read(&user->locked_vm);
7872 new_pages = cur_pages + nr_pages;
7873 if (new_pages > page_limit)
7875 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7876 new_pages) != cur_pages);
7881 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7882 enum io_mem_account acct)
7885 __io_unaccount_mem(ctx->user, nr_pages);
7887 if (ctx->mm_account) {
7888 if (acct == ACCT_LOCKED)
7889 ctx->mm_account->locked_vm -= nr_pages;
7890 else if (acct == ACCT_PINNED)
7891 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7895 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7896 enum io_mem_account acct)
7900 if (ctx->limit_mem) {
7901 ret = __io_account_mem(ctx->user, nr_pages);
7906 if (ctx->mm_account) {
7907 if (acct == ACCT_LOCKED)
7908 ctx->mm_account->locked_vm += nr_pages;
7909 else if (acct == ACCT_PINNED)
7910 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7916 static void io_mem_free(void *ptr)
7923 page = virt_to_head_page(ptr);
7924 if (put_page_testzero(page))
7925 free_compound_page(page);
7928 static void *io_mem_alloc(size_t size)
7930 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7933 return (void *) __get_free_pages(gfp_flags, get_order(size));
7936 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7939 struct io_rings *rings;
7940 size_t off, sq_array_size;
7942 off = struct_size(rings, cqes, cq_entries);
7943 if (off == SIZE_MAX)
7947 off = ALIGN(off, SMP_CACHE_BYTES);
7955 sq_array_size = array_size(sizeof(u32), sq_entries);
7956 if (sq_array_size == SIZE_MAX)
7959 if (check_add_overflow(off, sq_array_size, &off))
7965 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7969 pages = (size_t)1 << get_order(
7970 rings_size(sq_entries, cq_entries, NULL));
7971 pages += (size_t)1 << get_order(
7972 array_size(sizeof(struct io_uring_sqe), sq_entries));
7977 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7981 if (!ctx->user_bufs)
7984 for (i = 0; i < ctx->nr_user_bufs; i++) {
7985 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7987 for (j = 0; j < imu->nr_bvecs; j++)
7988 unpin_user_page(imu->bvec[j].bv_page);
7990 if (imu->acct_pages)
7991 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7996 kfree(ctx->user_bufs);
7997 ctx->user_bufs = NULL;
7998 ctx->nr_user_bufs = 0;
8002 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8003 void __user *arg, unsigned index)
8005 struct iovec __user *src;
8007 #ifdef CONFIG_COMPAT
8009 struct compat_iovec __user *ciovs;
8010 struct compat_iovec ciov;
8012 ciovs = (struct compat_iovec __user *) arg;
8013 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8016 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8017 dst->iov_len = ciov.iov_len;
8021 src = (struct iovec __user *) arg;
8022 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8028 * Not super efficient, but this is just a registration time. And we do cache
8029 * the last compound head, so generally we'll only do a full search if we don't
8032 * We check if the given compound head page has already been accounted, to
8033 * avoid double accounting it. This allows us to account the full size of the
8034 * page, not just the constituent pages of a huge page.
8036 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8037 int nr_pages, struct page *hpage)
8041 /* check current page array */
8042 for (i = 0; i < nr_pages; i++) {
8043 if (!PageCompound(pages[i]))
8045 if (compound_head(pages[i]) == hpage)
8049 /* check previously registered pages */
8050 for (i = 0; i < ctx->nr_user_bufs; i++) {
8051 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8053 for (j = 0; j < imu->nr_bvecs; j++) {
8054 if (!PageCompound(imu->bvec[j].bv_page))
8056 if (compound_head(imu->bvec[j].bv_page) == hpage)
8064 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8065 int nr_pages, struct io_mapped_ubuf *imu,
8066 struct page **last_hpage)
8070 for (i = 0; i < nr_pages; i++) {
8071 if (!PageCompound(pages[i])) {
8076 hpage = compound_head(pages[i]);
8077 if (hpage == *last_hpage)
8079 *last_hpage = hpage;
8080 if (headpage_already_acct(ctx, pages, i, hpage))
8082 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8086 if (!imu->acct_pages)
8089 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8091 imu->acct_pages = 0;
8095 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8098 struct vm_area_struct **vmas = NULL;
8099 struct page **pages = NULL;
8100 struct page *last_hpage = NULL;
8101 int i, j, got_pages = 0;
8106 if (!nr_args || nr_args > UIO_MAXIOV)
8109 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8111 if (!ctx->user_bufs)
8114 for (i = 0; i < nr_args; i++) {
8115 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8116 unsigned long off, start, end, ubuf;
8121 ret = io_copy_iov(ctx, &iov, arg, i);
8126 * Don't impose further limits on the size and buffer
8127 * constraints here, we'll -EINVAL later when IO is
8128 * submitted if they are wrong.
8131 if (!iov.iov_base || !iov.iov_len)
8134 /* arbitrary limit, but we need something */
8135 if (iov.iov_len > SZ_1G)
8138 ubuf = (unsigned long) iov.iov_base;
8139 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8140 start = ubuf >> PAGE_SHIFT;
8141 nr_pages = end - start;
8144 if (!pages || nr_pages > got_pages) {
8147 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8149 vmas = kvmalloc_array(nr_pages,
8150 sizeof(struct vm_area_struct *),
8152 if (!pages || !vmas) {
8156 got_pages = nr_pages;
8159 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8166 mmap_read_lock(current->mm);
8167 pret = pin_user_pages(ubuf, nr_pages,
8168 FOLL_WRITE | FOLL_LONGTERM,
8170 if (pret == nr_pages) {
8171 /* don't support file backed memory */
8172 for (j = 0; j < nr_pages; j++) {
8173 struct vm_area_struct *vma = vmas[j];
8176 !is_file_hugepages(vma->vm_file)) {
8182 ret = pret < 0 ? pret : -EFAULT;
8184 mmap_read_unlock(current->mm);
8187 * if we did partial map, or found file backed vmas,
8188 * release any pages we did get
8191 unpin_user_pages(pages, pret);
8196 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8198 unpin_user_pages(pages, pret);
8203 off = ubuf & ~PAGE_MASK;
8205 for (j = 0; j < nr_pages; j++) {
8208 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8209 imu->bvec[j].bv_page = pages[j];
8210 imu->bvec[j].bv_len = vec_len;
8211 imu->bvec[j].bv_offset = off;
8215 /* store original address for later verification */
8217 imu->len = iov.iov_len;
8218 imu->nr_bvecs = nr_pages;
8220 ctx->nr_user_bufs++;
8228 io_sqe_buffer_unregister(ctx);
8232 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8234 __s32 __user *fds = arg;
8240 if (copy_from_user(&fd, fds, sizeof(*fds)))
8243 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8244 if (IS_ERR(ctx->cq_ev_fd)) {
8245 int ret = PTR_ERR(ctx->cq_ev_fd);
8246 ctx->cq_ev_fd = NULL;
8253 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8255 if (ctx->cq_ev_fd) {
8256 eventfd_ctx_put(ctx->cq_ev_fd);
8257 ctx->cq_ev_fd = NULL;
8264 static int __io_destroy_buffers(int id, void *p, void *data)
8266 struct io_ring_ctx *ctx = data;
8267 struct io_buffer *buf = p;
8269 __io_remove_buffers(ctx, buf, id, -1U);
8273 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8275 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8276 idr_destroy(&ctx->io_buffer_idr);
8279 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8281 io_finish_async(ctx);
8282 io_sqe_buffer_unregister(ctx);
8284 if (ctx->sqo_task) {
8285 put_task_struct(ctx->sqo_task);
8286 ctx->sqo_task = NULL;
8287 mmdrop(ctx->mm_account);
8288 ctx->mm_account = NULL;
8291 #ifdef CONFIG_BLK_CGROUP
8292 if (ctx->sqo_blkcg_css)
8293 css_put(ctx->sqo_blkcg_css);
8296 io_sqe_files_unregister(ctx);
8297 io_eventfd_unregister(ctx);
8298 io_destroy_buffers(ctx);
8299 idr_destroy(&ctx->personality_idr);
8301 #if defined(CONFIG_UNIX)
8302 if (ctx->ring_sock) {
8303 ctx->ring_sock->file = NULL; /* so that iput() is called */
8304 sock_release(ctx->ring_sock);
8308 io_mem_free(ctx->rings);
8309 io_mem_free(ctx->sq_sqes);
8311 percpu_ref_exit(&ctx->refs);
8312 free_uid(ctx->user);
8313 put_cred(ctx->creds);
8314 kfree(ctx->cancel_hash);
8315 kmem_cache_free(req_cachep, ctx->fallback_req);
8319 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8321 struct io_ring_ctx *ctx = file->private_data;
8324 poll_wait(file, &ctx->cq_wait, wait);
8326 * synchronizes with barrier from wq_has_sleeper call in
8330 if (!io_sqring_full(ctx))
8331 mask |= EPOLLOUT | EPOLLWRNORM;
8332 if (io_cqring_events(ctx, false))
8333 mask |= EPOLLIN | EPOLLRDNORM;
8338 static int io_uring_fasync(int fd, struct file *file, int on)
8340 struct io_ring_ctx *ctx = file->private_data;
8342 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8345 static int io_remove_personalities(int id, void *p, void *data)
8347 struct io_ring_ctx *ctx = data;
8348 struct io_identity *iod;
8350 iod = idr_remove(&ctx->personality_idr, id);
8352 put_cred(iod->creds);
8353 if (refcount_dec_and_test(&iod->count))
8359 static void io_ring_exit_work(struct work_struct *work)
8361 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8365 * If we're doing polled IO and end up having requests being
8366 * submitted async (out-of-line), then completions can come in while
8367 * we're waiting for refs to drop. We need to reap these manually,
8368 * as nobody else will be looking for them.
8372 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8373 io_iopoll_try_reap_events(ctx);
8374 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8375 io_ring_ctx_free(ctx);
8378 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8380 mutex_lock(&ctx->uring_lock);
8381 percpu_ref_kill(&ctx->refs);
8382 mutex_unlock(&ctx->uring_lock);
8384 io_kill_timeouts(ctx, NULL);
8385 io_poll_remove_all(ctx, NULL);
8388 io_wq_cancel_all(ctx->io_wq);
8390 /* if we failed setting up the ctx, we might not have any rings */
8392 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8393 io_iopoll_try_reap_events(ctx);
8394 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8397 * Do this upfront, so we won't have a grace period where the ring
8398 * is closed but resources aren't reaped yet. This can cause
8399 * spurious failure in setting up a new ring.
8401 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8404 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8406 * Use system_unbound_wq to avoid spawning tons of event kworkers
8407 * if we're exiting a ton of rings at the same time. It just adds
8408 * noise and overhead, there's no discernable change in runtime
8409 * over using system_wq.
8411 queue_work(system_unbound_wq, &ctx->exit_work);
8414 static int io_uring_release(struct inode *inode, struct file *file)
8416 struct io_ring_ctx *ctx = file->private_data;
8418 file->private_data = NULL;
8419 io_ring_ctx_wait_and_kill(ctx);
8423 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8425 struct files_struct *files = data;
8427 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8428 work->identity->files == files);
8432 * Returns true if 'preq' is the link parent of 'req'
8434 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8436 struct io_kiocb *link;
8438 if (!(preq->flags & REQ_F_LINK_HEAD))
8441 list_for_each_entry(link, &preq->link_list, link_list) {
8450 * We're looking to cancel 'req' because it's holding on to our files, but
8451 * 'req' could be a link to another request. See if it is, and cancel that
8452 * parent request if so.
8454 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8456 struct hlist_node *tmp;
8457 struct io_kiocb *preq;
8461 spin_lock_irq(&ctx->completion_lock);
8462 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8463 struct hlist_head *list;
8465 list = &ctx->cancel_hash[i];
8466 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8467 found = io_match_link(preq, req);
8469 io_poll_remove_one(preq);
8474 spin_unlock_irq(&ctx->completion_lock);
8478 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8479 struct io_kiocb *req)
8481 struct io_kiocb *preq;
8484 spin_lock_irq(&ctx->completion_lock);
8485 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8486 found = io_match_link(preq, req);
8488 __io_timeout_cancel(preq);
8492 spin_unlock_irq(&ctx->completion_lock);
8496 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8498 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8501 if (req->flags & REQ_F_LINK_TIMEOUT) {
8502 unsigned long flags;
8503 struct io_ring_ctx *ctx = req->ctx;
8505 /* protect against races with linked timeouts */
8506 spin_lock_irqsave(&ctx->completion_lock, flags);
8507 ret = io_match_link(req, data);
8508 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8510 ret = io_match_link(req, data);
8515 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8517 enum io_wq_cancel cret;
8519 /* cancel this particular work, if it's running */
8520 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8521 if (cret != IO_WQ_CANCEL_NOTFOUND)
8524 /* find links that hold this pending, cancel those */
8525 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8526 if (cret != IO_WQ_CANCEL_NOTFOUND)
8529 /* if we have a poll link holding this pending, cancel that */
8530 if (io_poll_remove_link(ctx, req))
8533 /* final option, timeout link is holding this req pending */
8534 io_timeout_remove_link(ctx, req);
8537 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8538 struct task_struct *task,
8539 struct files_struct *files)
8541 struct io_defer_entry *de = NULL;
8544 spin_lock_irq(&ctx->completion_lock);
8545 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8546 if (io_task_match(de->req, task) &&
8547 io_match_files(de->req, files)) {
8548 list_cut_position(&list, &ctx->defer_list, &de->list);
8552 spin_unlock_irq(&ctx->completion_lock);
8554 while (!list_empty(&list)) {
8555 de = list_first_entry(&list, struct io_defer_entry, list);
8556 list_del_init(&de->list);
8557 req_set_fail_links(de->req);
8558 io_put_req(de->req);
8559 io_req_complete(de->req, -ECANCELED);
8565 * Returns true if we found and killed one or more files pinning requests
8567 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8568 struct files_struct *files)
8570 if (list_empty_careful(&ctx->inflight_list))
8573 /* cancel all at once, should be faster than doing it one by one*/
8574 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8576 while (!list_empty_careful(&ctx->inflight_list)) {
8577 struct io_kiocb *cancel_req = NULL, *req;
8580 spin_lock_irq(&ctx->inflight_lock);
8581 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8582 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8583 req->work.identity->files != files)
8585 /* req is being completed, ignore */
8586 if (!refcount_inc_not_zero(&req->refs))
8592 prepare_to_wait(&ctx->inflight_wait, &wait,
8593 TASK_UNINTERRUPTIBLE);
8594 spin_unlock_irq(&ctx->inflight_lock);
8596 /* We need to keep going until we don't find a matching req */
8599 /* cancel this request, or head link requests */
8600 io_attempt_cancel(ctx, cancel_req);
8601 io_put_req(cancel_req);
8602 /* cancellations _may_ trigger task work */
8605 finish_wait(&ctx->inflight_wait, &wait);
8611 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8613 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8614 struct task_struct *task = data;
8616 return io_task_match(req, task);
8619 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8620 struct task_struct *task,
8621 struct files_struct *files)
8625 ret = io_uring_cancel_files(ctx, files);
8627 enum io_wq_cancel cret;
8629 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8630 if (cret != IO_WQ_CANCEL_NOTFOUND)
8633 /* SQPOLL thread does its own polling */
8634 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8635 while (!list_empty_careful(&ctx->iopoll_list)) {
8636 io_iopoll_try_reap_events(ctx);
8641 ret |= io_poll_remove_all(ctx, task);
8642 ret |= io_kill_timeouts(ctx, task);
8649 * We need to iteratively cancel requests, in case a request has dependent
8650 * hard links. These persist even for failure of cancelations, hence keep
8651 * looping until none are found.
8653 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8654 struct files_struct *files)
8656 struct task_struct *task = current;
8658 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8659 task = ctx->sq_data->thread;
8660 atomic_inc(&task->io_uring->in_idle);
8661 io_sq_thread_park(ctx->sq_data);
8665 io_cancel_defer_files(ctx, NULL, files);
8667 io_cancel_defer_files(ctx, task, NULL);
8669 io_cqring_overflow_flush(ctx, true, task, files);
8671 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8676 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8677 atomic_dec(&task->io_uring->in_idle);
8679 * If the files that are going away are the ones in the thread
8680 * identity, clear them out.
8682 if (task->io_uring->identity->files == files)
8683 task->io_uring->identity->files = NULL;
8684 io_sq_thread_unpark(ctx->sq_data);
8689 * Note that this task has used io_uring. We use it for cancelation purposes.
8691 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8693 struct io_uring_task *tctx = current->io_uring;
8695 if (unlikely(!tctx)) {
8698 ret = io_uring_alloc_task_context(current);
8701 tctx = current->io_uring;
8703 if (tctx->last != file) {
8704 void *old = xa_load(&tctx->xa, (unsigned long)file);
8708 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8714 * This is race safe in that the task itself is doing this, hence it
8715 * cannot be going through the exit/cancel paths at the same time.
8716 * This cannot be modified while exit/cancel is running.
8718 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8719 tctx->sqpoll = true;
8725 * Remove this io_uring_file -> task mapping.
8727 static void io_uring_del_task_file(struct file *file)
8729 struct io_uring_task *tctx = current->io_uring;
8731 if (tctx->last == file)
8733 file = xa_erase(&tctx->xa, (unsigned long)file);
8739 * Drop task note for this file if we're the only ones that hold it after
8742 static void io_uring_attempt_task_drop(struct file *file)
8744 if (!current->io_uring)
8747 * fput() is pending, will be 2 if the only other ref is our potential
8748 * task file note. If the task is exiting, drop regardless of count.
8750 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8751 atomic_long_read(&file->f_count) == 2)
8752 io_uring_del_task_file(file);
8755 void __io_uring_files_cancel(struct files_struct *files)
8757 struct io_uring_task *tctx = current->io_uring;
8759 unsigned long index;
8761 /* make sure overflow events are dropped */
8762 atomic_inc(&tctx->in_idle);
8764 xa_for_each(&tctx->xa, index, file) {
8765 struct io_ring_ctx *ctx = file->private_data;
8767 io_uring_cancel_task_requests(ctx, files);
8769 io_uring_del_task_file(file);
8772 atomic_dec(&tctx->in_idle);
8775 static s64 tctx_inflight(struct io_uring_task *tctx)
8777 unsigned long index;
8781 inflight = percpu_counter_sum(&tctx->inflight);
8786 * If we have SQPOLL rings, then we need to iterate and find them, and
8787 * add the pending count for those.
8789 xa_for_each(&tctx->xa, index, file) {
8790 struct io_ring_ctx *ctx = file->private_data;
8792 if (ctx->flags & IORING_SETUP_SQPOLL) {
8793 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
8795 inflight += percpu_counter_sum(&__tctx->inflight);
8803 * Find any io_uring fd that this task has registered or done IO on, and cancel
8806 void __io_uring_task_cancel(void)
8808 struct io_uring_task *tctx = current->io_uring;
8812 /* make sure overflow events are dropped */
8813 atomic_inc(&tctx->in_idle);
8816 /* read completions before cancelations */
8817 inflight = tctx_inflight(tctx);
8820 __io_uring_files_cancel(NULL);
8822 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8825 * If we've seen completions, retry. This avoids a race where
8826 * a completion comes in before we did prepare_to_wait().
8828 if (inflight != tctx_inflight(tctx))
8833 finish_wait(&tctx->wait, &wait);
8834 atomic_dec(&tctx->in_idle);
8837 static int io_uring_flush(struct file *file, void *data)
8839 io_uring_attempt_task_drop(file);
8843 static void *io_uring_validate_mmap_request(struct file *file,
8844 loff_t pgoff, size_t sz)
8846 struct io_ring_ctx *ctx = file->private_data;
8847 loff_t offset = pgoff << PAGE_SHIFT;
8852 case IORING_OFF_SQ_RING:
8853 case IORING_OFF_CQ_RING:
8856 case IORING_OFF_SQES:
8860 return ERR_PTR(-EINVAL);
8863 page = virt_to_head_page(ptr);
8864 if (sz > page_size(page))
8865 return ERR_PTR(-EINVAL);
8872 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8874 size_t sz = vma->vm_end - vma->vm_start;
8878 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8880 return PTR_ERR(ptr);
8882 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8883 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8886 #else /* !CONFIG_MMU */
8888 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8890 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8893 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8895 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8898 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8899 unsigned long addr, unsigned long len,
8900 unsigned long pgoff, unsigned long flags)
8904 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8906 return PTR_ERR(ptr);
8908 return (unsigned long) ptr;
8911 #endif /* !CONFIG_MMU */
8913 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8918 if (!io_sqring_full(ctx))
8921 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8923 if (!io_sqring_full(ctx))
8927 } while (!signal_pending(current));
8929 finish_wait(&ctx->sqo_sq_wait, &wait);
8932 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8933 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8936 struct io_ring_ctx *ctx;
8943 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8944 IORING_ENTER_SQ_WAIT))
8952 if (f.file->f_op != &io_uring_fops)
8956 ctx = f.file->private_data;
8957 if (!percpu_ref_tryget(&ctx->refs))
8961 if (ctx->flags & IORING_SETUP_R_DISABLED)
8965 * For SQ polling, the thread will do all submissions and completions.
8966 * Just return the requested submit count, and wake the thread if
8970 if (ctx->flags & IORING_SETUP_SQPOLL) {
8971 if (!list_empty_careful(&ctx->cq_overflow_list))
8972 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8973 if (flags & IORING_ENTER_SQ_WAKEUP)
8974 wake_up(&ctx->sq_data->wait);
8975 if (flags & IORING_ENTER_SQ_WAIT)
8976 io_sqpoll_wait_sq(ctx);
8977 submitted = to_submit;
8978 } else if (to_submit) {
8979 ret = io_uring_add_task_file(ctx, f.file);
8982 mutex_lock(&ctx->uring_lock);
8983 submitted = io_submit_sqes(ctx, to_submit);
8984 mutex_unlock(&ctx->uring_lock);
8986 if (submitted != to_submit)
8989 if (flags & IORING_ENTER_GETEVENTS) {
8990 min_complete = min(min_complete, ctx->cq_entries);
8993 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8994 * space applications don't need to do io completion events
8995 * polling again, they can rely on io_sq_thread to do polling
8996 * work, which can reduce cpu usage and uring_lock contention.
8998 if (ctx->flags & IORING_SETUP_IOPOLL &&
8999 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9000 ret = io_iopoll_check(ctx, min_complete);
9002 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
9007 percpu_ref_put(&ctx->refs);
9010 return submitted ? submitted : ret;
9013 #ifdef CONFIG_PROC_FS
9014 static int io_uring_show_cred(int id, void *p, void *data)
9016 struct io_identity *iod = p;
9017 const struct cred *cred = iod->creds;
9018 struct seq_file *m = data;
9019 struct user_namespace *uns = seq_user_ns(m);
9020 struct group_info *gi;
9025 seq_printf(m, "%5d\n", id);
9026 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9027 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9028 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9029 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9030 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9031 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9032 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9033 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9034 seq_puts(m, "\n\tGroups:\t");
9035 gi = cred->group_info;
9036 for (g = 0; g < gi->ngroups; g++) {
9037 seq_put_decimal_ull(m, g ? " " : "",
9038 from_kgid_munged(uns, gi->gid[g]));
9040 seq_puts(m, "\n\tCapEff:\t");
9041 cap = cred->cap_effective;
9042 CAP_FOR_EACH_U32(__capi)
9043 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9048 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9050 struct io_sq_data *sq = NULL;
9055 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9056 * since fdinfo case grabs it in the opposite direction of normal use
9057 * cases. If we fail to get the lock, we just don't iterate any
9058 * structures that could be going away outside the io_uring mutex.
9060 has_lock = mutex_trylock(&ctx->uring_lock);
9062 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9065 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9066 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9067 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9068 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9069 struct fixed_file_table *table;
9072 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9073 f = table->files[i & IORING_FILE_TABLE_MASK];
9075 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9077 seq_printf(m, "%5u: <none>\n", i);
9079 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9080 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9081 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9083 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9084 (unsigned int) buf->len);
9086 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9087 seq_printf(m, "Personalities:\n");
9088 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9090 seq_printf(m, "PollList:\n");
9091 spin_lock_irq(&ctx->completion_lock);
9092 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9093 struct hlist_head *list = &ctx->cancel_hash[i];
9094 struct io_kiocb *req;
9096 hlist_for_each_entry(req, list, hash_node)
9097 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9098 req->task->task_works != NULL);
9100 spin_unlock_irq(&ctx->completion_lock);
9102 mutex_unlock(&ctx->uring_lock);
9105 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9107 struct io_ring_ctx *ctx = f->private_data;
9109 if (percpu_ref_tryget(&ctx->refs)) {
9110 __io_uring_show_fdinfo(ctx, m);
9111 percpu_ref_put(&ctx->refs);
9116 static const struct file_operations io_uring_fops = {
9117 .release = io_uring_release,
9118 .flush = io_uring_flush,
9119 .mmap = io_uring_mmap,
9121 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9122 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9124 .poll = io_uring_poll,
9125 .fasync = io_uring_fasync,
9126 #ifdef CONFIG_PROC_FS
9127 .show_fdinfo = io_uring_show_fdinfo,
9131 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9132 struct io_uring_params *p)
9134 struct io_rings *rings;
9135 size_t size, sq_array_offset;
9137 /* make sure these are sane, as we already accounted them */
9138 ctx->sq_entries = p->sq_entries;
9139 ctx->cq_entries = p->cq_entries;
9141 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9142 if (size == SIZE_MAX)
9145 rings = io_mem_alloc(size);
9150 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9151 rings->sq_ring_mask = p->sq_entries - 1;
9152 rings->cq_ring_mask = p->cq_entries - 1;
9153 rings->sq_ring_entries = p->sq_entries;
9154 rings->cq_ring_entries = p->cq_entries;
9155 ctx->sq_mask = rings->sq_ring_mask;
9156 ctx->cq_mask = rings->cq_ring_mask;
9158 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9159 if (size == SIZE_MAX) {
9160 io_mem_free(ctx->rings);
9165 ctx->sq_sqes = io_mem_alloc(size);
9166 if (!ctx->sq_sqes) {
9167 io_mem_free(ctx->rings);
9176 * Allocate an anonymous fd, this is what constitutes the application
9177 * visible backing of an io_uring instance. The application mmaps this
9178 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9179 * we have to tie this fd to a socket for file garbage collection purposes.
9181 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9186 #if defined(CONFIG_UNIX)
9187 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9193 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9197 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9198 O_RDWR | O_CLOEXEC);
9202 ret = PTR_ERR(file);
9206 #if defined(CONFIG_UNIX)
9207 ctx->ring_sock->file = file;
9209 if (unlikely(io_uring_add_task_file(ctx, file))) {
9210 file = ERR_PTR(-ENOMEM);
9213 fd_install(ret, file);
9216 #if defined(CONFIG_UNIX)
9217 sock_release(ctx->ring_sock);
9218 ctx->ring_sock = NULL;
9223 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9224 struct io_uring_params __user *params)
9226 struct user_struct *user = NULL;
9227 struct io_ring_ctx *ctx;
9233 if (entries > IORING_MAX_ENTRIES) {
9234 if (!(p->flags & IORING_SETUP_CLAMP))
9236 entries = IORING_MAX_ENTRIES;
9240 * Use twice as many entries for the CQ ring. It's possible for the
9241 * application to drive a higher depth than the size of the SQ ring,
9242 * since the sqes are only used at submission time. This allows for
9243 * some flexibility in overcommitting a bit. If the application has
9244 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9245 * of CQ ring entries manually.
9247 p->sq_entries = roundup_pow_of_two(entries);
9248 if (p->flags & IORING_SETUP_CQSIZE) {
9250 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9251 * to a power-of-two, if it isn't already. We do NOT impose
9252 * any cq vs sq ring sizing.
9256 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9257 if (!(p->flags & IORING_SETUP_CLAMP))
9259 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9261 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9262 if (p->cq_entries < p->sq_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);