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/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81 #include <linux/freezer.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
101 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
102 IORING_REGISTER_LAST + IORING_OP_LAST)
104 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
105 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 u32 head ____cacheline_aligned_in_smp;
110 u32 tail ____cacheline_aligned_in_smp;
114 * This data is shared with the application through the mmap at offsets
115 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
117 * The offsets to the member fields are published through struct
118 * io_sqring_offsets when calling io_uring_setup.
122 * Head and tail offsets into the ring; the offsets need to be
123 * masked to get valid indices.
125 * The kernel controls head of the sq ring and the tail of the cq ring,
126 * and the application controls tail of the sq ring and the head of the
129 struct io_uring sq, cq;
131 * Bitmasks to apply to head and tail offsets (constant, equals
134 u32 sq_ring_mask, cq_ring_mask;
135 /* Ring sizes (constant, power of 2) */
136 u32 sq_ring_entries, cq_ring_entries;
138 * Number of invalid entries dropped by the kernel due to
139 * invalid index stored in array
141 * Written by the kernel, shouldn't be modified by the
142 * application (i.e. get number of "new events" by comparing to
145 * After a new SQ head value was read by the application this
146 * counter includes all submissions that were dropped reaching
147 * the new SQ head (and possibly more).
153 * Written by the kernel, shouldn't be modified by the
156 * The application needs a full memory barrier before checking
157 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
163 * Written by the application, shouldn't be modified by the
168 * Number of completion events lost because the queue was full;
169 * this should be avoided by the application by making sure
170 * there are not more requests pending than there is space in
171 * the completion queue.
173 * Written by the kernel, shouldn't be modified by the
174 * application (i.e. get number of "new events" by comparing to
177 * As completion events come in out of order this counter is not
178 * ordered with any other data.
182 * Ring buffer of completion events.
184 * The kernel writes completion events fresh every time they are
185 * produced, so the application is allowed to modify pending
188 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
191 enum io_uring_cmd_flags {
192 IO_URING_F_NONBLOCK = 1,
193 IO_URING_F_COMPLETE_DEFER = 2,
196 struct io_mapped_ubuf {
199 struct bio_vec *bvec;
200 unsigned int nr_bvecs;
201 unsigned long acct_pages;
207 struct list_head list;
214 struct fixed_rsrc_table {
218 struct fixed_rsrc_ref_node {
219 struct percpu_ref refs;
220 struct list_head node;
221 struct list_head rsrc_list;
222 struct fixed_rsrc_data *rsrc_data;
223 void (*rsrc_put)(struct io_ring_ctx *ctx,
224 struct io_rsrc_put *prsrc);
225 struct llist_node llist;
229 struct fixed_rsrc_data {
230 struct fixed_rsrc_table *table;
231 struct io_ring_ctx *ctx;
233 struct fixed_rsrc_ref_node *node;
234 struct percpu_ref refs;
235 struct completion done;
240 struct list_head list;
246 struct io_restriction {
247 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
248 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
249 u8 sqe_flags_allowed;
250 u8 sqe_flags_required;
255 IO_SQ_THREAD_SHOULD_STOP = 0,
256 IO_SQ_THREAD_SHOULD_PARK,
261 struct rw_semaphore rw_lock;
263 /* ctx's that are using this sqd */
264 struct list_head ctx_list;
266 struct task_struct *thread;
267 struct wait_queue_head wait;
269 unsigned sq_thread_idle;
275 struct completion exited;
278 #define IO_IOPOLL_BATCH 8
279 #define IO_COMPL_BATCH 32
280 #define IO_REQ_CACHE_SIZE 32
281 #define IO_REQ_ALLOC_BATCH 8
283 struct io_comp_state {
284 struct io_kiocb *reqs[IO_COMPL_BATCH];
286 unsigned int locked_free_nr;
287 /* inline/task_work completion list, under ->uring_lock */
288 struct list_head free_list;
289 /* IRQ completion list, under ->completion_lock */
290 struct list_head locked_free_list;
293 struct io_submit_link {
294 struct io_kiocb *head;
295 struct io_kiocb *last;
298 struct io_submit_state {
299 struct blk_plug plug;
300 struct io_submit_link link;
303 * io_kiocb alloc cache
305 void *reqs[IO_REQ_CACHE_SIZE];
306 unsigned int free_reqs;
311 * Batch completion logic
313 struct io_comp_state comp;
316 * File reference cache
320 unsigned int file_refs;
321 unsigned int ios_left;
326 struct percpu_ref refs;
327 } ____cacheline_aligned_in_smp;
331 unsigned int compat: 1;
332 unsigned int cq_overflow_flushed: 1;
333 unsigned int drain_next: 1;
334 unsigned int eventfd_async: 1;
335 unsigned int restricted: 1;
338 * Ring buffer of indices into array of io_uring_sqe, which is
339 * mmapped by the application using the IORING_OFF_SQES offset.
341 * This indirection could e.g. be used to assign fixed
342 * io_uring_sqe entries to operations and only submit them to
343 * the queue when needed.
345 * The kernel modifies neither the indices array nor the entries
349 unsigned cached_sq_head;
352 unsigned sq_thread_idle;
353 unsigned cached_sq_dropped;
354 unsigned cached_cq_overflow;
355 unsigned long sq_check_overflow;
357 /* hashed buffered write serialization */
358 struct io_wq_hash *hash_map;
360 struct list_head defer_list;
361 struct list_head timeout_list;
362 struct list_head cq_overflow_list;
364 struct io_uring_sqe *sq_sqes;
365 } ____cacheline_aligned_in_smp;
368 struct mutex uring_lock;
369 wait_queue_head_t wait;
370 } ____cacheline_aligned_in_smp;
372 struct io_submit_state submit_state;
374 struct io_rings *rings;
376 /* Only used for accounting purposes */
377 struct mm_struct *mm_account;
379 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
380 struct io_sq_data *sq_data; /* if using sq thread polling */
382 struct wait_queue_head sqo_sq_wait;
383 struct list_head sqd_list;
386 * If used, fixed file set. Writers must ensure that ->refs is dead,
387 * readers must ensure that ->refs is alive as long as the file* is
388 * used. Only updated through io_uring_register(2).
390 struct fixed_rsrc_data *file_data;
391 unsigned nr_user_files;
393 /* if used, fixed mapped user buffers */
394 unsigned nr_user_bufs;
395 struct io_mapped_ubuf *user_bufs;
397 struct user_struct *user;
399 struct completion ref_comp;
401 #if defined(CONFIG_UNIX)
402 struct socket *ring_sock;
405 struct xarray io_buffers;
407 struct xarray personalities;
411 unsigned cached_cq_tail;
414 atomic_t cq_timeouts;
415 unsigned cq_last_tm_flush;
416 unsigned long cq_check_overflow;
417 struct wait_queue_head cq_wait;
418 struct fasync_struct *cq_fasync;
419 struct eventfd_ctx *cq_ev_fd;
420 } ____cacheline_aligned_in_smp;
423 spinlock_t completion_lock;
426 * ->iopoll_list is protected by the ctx->uring_lock for
427 * io_uring instances that don't use IORING_SETUP_SQPOLL.
428 * For SQPOLL, only the single threaded io_sq_thread() will
429 * manipulate the list, hence no extra locking is needed there.
431 struct list_head iopoll_list;
432 struct hlist_head *cancel_hash;
433 unsigned cancel_hash_bits;
434 bool poll_multi_file;
436 spinlock_t inflight_lock;
437 struct list_head inflight_list;
438 } ____cacheline_aligned_in_smp;
440 struct delayed_work rsrc_put_work;
441 struct llist_head rsrc_put_llist;
442 struct list_head rsrc_ref_list;
443 spinlock_t rsrc_ref_lock;
445 struct io_restriction restrictions;
448 struct callback_head *exit_task_work;
450 struct wait_queue_head hash_wait;
452 /* Keep this last, we don't need it for the fast path */
453 struct work_struct exit_work;
454 struct list_head tctx_list;
458 * First field must be the file pointer in all the
459 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
461 struct io_poll_iocb {
463 struct wait_queue_head *head;
467 struct wait_queue_entry wait;
470 struct io_poll_remove {
480 struct io_timeout_data {
481 struct io_kiocb *req;
482 struct hrtimer timer;
483 struct timespec64 ts;
484 enum hrtimer_mode mode;
489 struct sockaddr __user *addr;
490 int __user *addr_len;
492 unsigned long nofile;
512 struct list_head list;
513 /* head of the link, used by linked timeouts only */
514 struct io_kiocb *head;
517 struct io_timeout_rem {
522 struct timespec64 ts;
527 /* NOTE: kiocb has the file as the first member, so don't do it here */
535 struct sockaddr __user *addr;
542 struct user_msghdr __user *umsg;
548 struct io_buffer *kbuf;
554 struct filename *filename;
556 unsigned long nofile;
559 struct io_rsrc_update {
585 struct epoll_event event;
589 struct file *file_out;
590 struct file *file_in;
597 struct io_provide_buf {
611 const char __user *filename;
612 struct statx __user *buffer;
624 struct filename *oldpath;
625 struct filename *newpath;
633 struct filename *filename;
636 struct io_completion {
638 struct list_head list;
642 struct io_async_connect {
643 struct sockaddr_storage address;
646 struct io_async_msghdr {
647 struct iovec fast_iov[UIO_FASTIOV];
648 /* points to an allocated iov, if NULL we use fast_iov instead */
649 struct iovec *free_iov;
650 struct sockaddr __user *uaddr;
652 struct sockaddr_storage addr;
656 struct iovec fast_iov[UIO_FASTIOV];
657 const struct iovec *free_iovec;
658 struct iov_iter iter;
660 struct wait_page_queue wpq;
664 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
665 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
666 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
667 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
668 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
669 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
675 REQ_F_LINK_TIMEOUT_BIT,
677 REQ_F_NEED_CLEANUP_BIT,
679 REQ_F_BUFFER_SELECTED_BIT,
680 REQ_F_NO_FILE_TABLE_BIT,
681 REQ_F_LTIMEOUT_ACTIVE_BIT,
682 REQ_F_COMPLETE_INLINE_BIT,
684 /* not a real bit, just to check we're not overflowing the space */
690 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
691 /* drain existing IO first */
692 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
694 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
695 /* doesn't sever on completion < 0 */
696 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
698 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
699 /* IOSQE_BUFFER_SELECT */
700 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
702 /* fail rest of links */
703 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
704 /* on inflight list, should be cancelled and waited on exit reliably */
705 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
706 /* read/write uses file position */
707 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
708 /* must not punt to workers */
709 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
710 /* has or had linked timeout */
711 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
713 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
715 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
716 /* already went through poll handler */
717 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
718 /* buffer already selected */
719 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
720 /* doesn't need file table for this request */
721 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
722 /* linked timeout is active, i.e. prepared by link's head */
723 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
724 /* completion is deferred through io_comp_state */
725 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
729 struct io_poll_iocb poll;
730 struct io_poll_iocb *double_poll;
733 struct io_task_work {
734 struct io_wq_work_node node;
735 task_work_func_t func;
739 * NOTE! Each of the iocb union members has the file pointer
740 * as the first entry in their struct definition. So you can
741 * access the file pointer through any of the sub-structs,
742 * or directly as just 'ki_filp' in this struct.
748 struct io_poll_iocb poll;
749 struct io_poll_remove poll_remove;
750 struct io_accept accept;
752 struct io_cancel cancel;
753 struct io_timeout timeout;
754 struct io_timeout_rem timeout_rem;
755 struct io_connect connect;
756 struct io_sr_msg sr_msg;
758 struct io_close close;
759 struct io_rsrc_update rsrc_update;
760 struct io_fadvise fadvise;
761 struct io_madvise madvise;
762 struct io_epoll epoll;
763 struct io_splice splice;
764 struct io_provide_buf pbuf;
765 struct io_statx statx;
766 struct io_shutdown shutdown;
767 struct io_rename rename;
768 struct io_unlink unlink;
769 /* use only after cleaning per-op data, see io_clean_op() */
770 struct io_completion compl;
773 /* opcode allocated if it needs to store data for async defer */
776 /* polled IO has completed */
782 struct io_ring_ctx *ctx;
785 struct task_struct *task;
788 struct io_kiocb *link;
789 struct percpu_ref *fixed_rsrc_refs;
792 * 1. used with ctx->iopoll_list with reads/writes
793 * 2. to track reqs with ->files (see io_op_def::file_table)
795 struct list_head inflight_entry;
797 struct io_task_work io_task_work;
798 struct callback_head task_work;
800 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
801 struct hlist_node hash_node;
802 struct async_poll *apoll;
803 struct io_wq_work work;
806 struct io_tctx_node {
807 struct list_head ctx_node;
808 struct task_struct *task;
809 struct io_ring_ctx *ctx;
812 struct io_defer_entry {
813 struct list_head list;
814 struct io_kiocb *req;
819 /* needs req->file assigned */
820 unsigned needs_file : 1;
821 /* hash wq insertion if file is a regular file */
822 unsigned hash_reg_file : 1;
823 /* unbound wq insertion if file is a non-regular file */
824 unsigned unbound_nonreg_file : 1;
825 /* opcode is not supported by this kernel */
826 unsigned not_supported : 1;
827 /* set if opcode supports polled "wait" */
829 unsigned pollout : 1;
830 /* op supports buffer selection */
831 unsigned buffer_select : 1;
832 /* must always have async data allocated */
833 unsigned needs_async_data : 1;
834 /* should block plug */
836 /* size of async data needed, if any */
837 unsigned short async_size;
840 static const struct io_op_def io_op_defs[] = {
841 [IORING_OP_NOP] = {},
842 [IORING_OP_READV] = {
844 .unbound_nonreg_file = 1,
847 .needs_async_data = 1,
849 .async_size = sizeof(struct io_async_rw),
851 [IORING_OP_WRITEV] = {
854 .unbound_nonreg_file = 1,
856 .needs_async_data = 1,
858 .async_size = sizeof(struct io_async_rw),
860 [IORING_OP_FSYNC] = {
863 [IORING_OP_READ_FIXED] = {
865 .unbound_nonreg_file = 1,
868 .async_size = sizeof(struct io_async_rw),
870 [IORING_OP_WRITE_FIXED] = {
873 .unbound_nonreg_file = 1,
876 .async_size = sizeof(struct io_async_rw),
878 [IORING_OP_POLL_ADD] = {
880 .unbound_nonreg_file = 1,
882 [IORING_OP_POLL_REMOVE] = {},
883 [IORING_OP_SYNC_FILE_RANGE] = {
886 [IORING_OP_SENDMSG] = {
888 .unbound_nonreg_file = 1,
890 .needs_async_data = 1,
891 .async_size = sizeof(struct io_async_msghdr),
893 [IORING_OP_RECVMSG] = {
895 .unbound_nonreg_file = 1,
898 .needs_async_data = 1,
899 .async_size = sizeof(struct io_async_msghdr),
901 [IORING_OP_TIMEOUT] = {
902 .needs_async_data = 1,
903 .async_size = sizeof(struct io_timeout_data),
905 [IORING_OP_TIMEOUT_REMOVE] = {
906 /* used by timeout updates' prep() */
908 [IORING_OP_ACCEPT] = {
910 .unbound_nonreg_file = 1,
913 [IORING_OP_ASYNC_CANCEL] = {},
914 [IORING_OP_LINK_TIMEOUT] = {
915 .needs_async_data = 1,
916 .async_size = sizeof(struct io_timeout_data),
918 [IORING_OP_CONNECT] = {
920 .unbound_nonreg_file = 1,
922 .needs_async_data = 1,
923 .async_size = sizeof(struct io_async_connect),
925 [IORING_OP_FALLOCATE] = {
928 [IORING_OP_OPENAT] = {},
929 [IORING_OP_CLOSE] = {},
930 [IORING_OP_FILES_UPDATE] = {},
931 [IORING_OP_STATX] = {},
934 .unbound_nonreg_file = 1,
938 .async_size = sizeof(struct io_async_rw),
940 [IORING_OP_WRITE] = {
942 .unbound_nonreg_file = 1,
945 .async_size = sizeof(struct io_async_rw),
947 [IORING_OP_FADVISE] = {
950 [IORING_OP_MADVISE] = {},
953 .unbound_nonreg_file = 1,
958 .unbound_nonreg_file = 1,
962 [IORING_OP_OPENAT2] = {
964 [IORING_OP_EPOLL_CTL] = {
965 .unbound_nonreg_file = 1,
967 [IORING_OP_SPLICE] = {
970 .unbound_nonreg_file = 1,
972 [IORING_OP_PROVIDE_BUFFERS] = {},
973 [IORING_OP_REMOVE_BUFFERS] = {},
977 .unbound_nonreg_file = 1,
979 [IORING_OP_SHUTDOWN] = {
982 [IORING_OP_RENAMEAT] = {},
983 [IORING_OP_UNLINKAT] = {},
986 static bool io_disarm_next(struct io_kiocb *req);
987 static void io_uring_del_task_file(unsigned long index);
988 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
989 struct task_struct *task,
990 struct files_struct *files);
991 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
992 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
993 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
994 struct io_ring_ctx *ctx);
995 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
997 static bool io_rw_reissue(struct io_kiocb *req);
998 static void io_cqring_fill_event(struct io_kiocb *req, long res);
999 static void io_put_req(struct io_kiocb *req);
1000 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1001 static void io_double_put_req(struct io_kiocb *req);
1002 static void io_dismantle_req(struct io_kiocb *req);
1003 static void io_put_task(struct task_struct *task, int nr);
1004 static void io_queue_next(struct io_kiocb *req);
1005 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1006 static void __io_queue_linked_timeout(struct io_kiocb *req);
1007 static void io_queue_linked_timeout(struct io_kiocb *req);
1008 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1009 struct io_uring_rsrc_update *ip,
1011 static void __io_clean_op(struct io_kiocb *req);
1012 static struct file *io_file_get(struct io_submit_state *state,
1013 struct io_kiocb *req, int fd, bool fixed);
1014 static void __io_queue_sqe(struct io_kiocb *req);
1015 static void io_rsrc_put_work(struct work_struct *work);
1017 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1018 struct iov_iter *iter, bool needs_lock);
1019 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1020 const struct iovec *fast_iov,
1021 struct iov_iter *iter, bool force);
1022 static void io_req_task_queue(struct io_kiocb *req);
1023 static void io_submit_flush_completions(struct io_comp_state *cs,
1024 struct io_ring_ctx *ctx);
1026 static struct kmem_cache *req_cachep;
1028 static const struct file_operations io_uring_fops;
1030 struct sock *io_uring_get_socket(struct file *file)
1032 #if defined(CONFIG_UNIX)
1033 if (file->f_op == &io_uring_fops) {
1034 struct io_ring_ctx *ctx = file->private_data;
1036 return ctx->ring_sock->sk;
1041 EXPORT_SYMBOL(io_uring_get_socket);
1043 #define io_for_each_link(pos, head) \
1044 for (pos = (head); pos; pos = pos->link)
1046 static inline void io_clean_op(struct io_kiocb *req)
1048 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1052 static inline void io_set_resource_node(struct io_kiocb *req)
1054 struct io_ring_ctx *ctx = req->ctx;
1056 if (!req->fixed_rsrc_refs) {
1057 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1058 percpu_ref_get(req->fixed_rsrc_refs);
1062 static bool io_match_task(struct io_kiocb *head,
1063 struct task_struct *task,
1064 struct files_struct *files)
1066 struct io_kiocb *req;
1068 if (task && head->task != task) {
1069 /* in terms of cancelation, always match if req task is dead */
1070 if (head->task->flags & PF_EXITING)
1077 io_for_each_link(req, head) {
1078 if (req->flags & REQ_F_INFLIGHT)
1080 if (req->task->files == files)
1086 static inline void req_set_fail_links(struct io_kiocb *req)
1088 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1089 req->flags |= REQ_F_FAIL_LINK;
1092 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1094 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1096 complete(&ctx->ref_comp);
1099 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1101 return !req->timeout.off;
1104 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1106 struct io_ring_ctx *ctx;
1109 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1114 * Use 5 bits less than the max cq entries, that should give us around
1115 * 32 entries per hash list if totally full and uniformly spread.
1117 hash_bits = ilog2(p->cq_entries);
1121 ctx->cancel_hash_bits = hash_bits;
1122 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1124 if (!ctx->cancel_hash)
1126 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1128 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1129 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1132 ctx->flags = p->flags;
1133 init_waitqueue_head(&ctx->sqo_sq_wait);
1134 INIT_LIST_HEAD(&ctx->sqd_list);
1135 init_waitqueue_head(&ctx->cq_wait);
1136 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1137 init_completion(&ctx->ref_comp);
1138 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1139 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1140 mutex_init(&ctx->uring_lock);
1141 init_waitqueue_head(&ctx->wait);
1142 spin_lock_init(&ctx->completion_lock);
1143 INIT_LIST_HEAD(&ctx->iopoll_list);
1144 INIT_LIST_HEAD(&ctx->defer_list);
1145 INIT_LIST_HEAD(&ctx->timeout_list);
1146 spin_lock_init(&ctx->inflight_lock);
1147 INIT_LIST_HEAD(&ctx->inflight_list);
1148 spin_lock_init(&ctx->rsrc_ref_lock);
1149 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1150 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1151 init_llist_head(&ctx->rsrc_put_llist);
1152 INIT_LIST_HEAD(&ctx->tctx_list);
1153 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1154 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1157 kfree(ctx->cancel_hash);
1162 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1164 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1165 struct io_ring_ctx *ctx = req->ctx;
1167 return seq != ctx->cached_cq_tail
1168 + READ_ONCE(ctx->cached_cq_overflow);
1174 static void io_req_track_inflight(struct io_kiocb *req)
1176 struct io_ring_ctx *ctx = req->ctx;
1178 if (!(req->flags & REQ_F_INFLIGHT)) {
1179 req->flags |= REQ_F_INFLIGHT;
1181 spin_lock_irq(&ctx->inflight_lock);
1182 list_add(&req->inflight_entry, &ctx->inflight_list);
1183 spin_unlock_irq(&ctx->inflight_lock);
1187 static void io_prep_async_work(struct io_kiocb *req)
1189 const struct io_op_def *def = &io_op_defs[req->opcode];
1190 struct io_ring_ctx *ctx = req->ctx;
1192 if (!req->work.creds)
1193 req->work.creds = get_current_cred();
1195 if (req->flags & REQ_F_FORCE_ASYNC)
1196 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1198 if (req->flags & REQ_F_ISREG) {
1199 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1200 io_wq_hash_work(&req->work, file_inode(req->file));
1202 if (def->unbound_nonreg_file)
1203 req->work.flags |= IO_WQ_WORK_UNBOUND;
1207 static void io_prep_async_link(struct io_kiocb *req)
1209 struct io_kiocb *cur;
1211 io_for_each_link(cur, req)
1212 io_prep_async_work(cur);
1215 static void io_queue_async_work(struct io_kiocb *req)
1217 struct io_ring_ctx *ctx = req->ctx;
1218 struct io_kiocb *link = io_prep_linked_timeout(req);
1219 struct io_uring_task *tctx = req->task->io_uring;
1222 BUG_ON(!tctx->io_wq);
1224 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1225 &req->work, req->flags);
1226 /* init ->work of the whole link before punting */
1227 io_prep_async_link(req);
1228 io_wq_enqueue(tctx->io_wq, &req->work);
1230 io_queue_linked_timeout(link);
1233 static void io_kill_timeout(struct io_kiocb *req)
1235 struct io_timeout_data *io = req->async_data;
1238 ret = hrtimer_try_to_cancel(&io->timer);
1240 atomic_set(&req->ctx->cq_timeouts,
1241 atomic_read(&req->ctx->cq_timeouts) + 1);
1242 list_del_init(&req->timeout.list);
1243 io_cqring_fill_event(req, 0);
1244 io_put_req_deferred(req, 1);
1249 * Returns true if we found and killed one or more timeouts
1251 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1252 struct files_struct *files)
1254 struct io_kiocb *req, *tmp;
1257 spin_lock_irq(&ctx->completion_lock);
1258 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1259 if (io_match_task(req, tsk, files)) {
1260 io_kill_timeout(req);
1264 spin_unlock_irq(&ctx->completion_lock);
1265 return canceled != 0;
1268 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1271 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1272 struct io_defer_entry, list);
1274 if (req_need_defer(de->req, de->seq))
1276 list_del_init(&de->list);
1277 io_req_task_queue(de->req);
1279 } while (!list_empty(&ctx->defer_list));
1282 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1286 if (list_empty(&ctx->timeout_list))
1289 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1292 u32 events_needed, events_got;
1293 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1294 struct io_kiocb, timeout.list);
1296 if (io_is_timeout_noseq(req))
1300 * Since seq can easily wrap around over time, subtract
1301 * the last seq at which timeouts were flushed before comparing.
1302 * Assuming not more than 2^31-1 events have happened since,
1303 * these subtractions won't have wrapped, so we can check if
1304 * target is in [last_seq, current_seq] by comparing the two.
1306 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1307 events_got = seq - ctx->cq_last_tm_flush;
1308 if (events_got < events_needed)
1311 list_del_init(&req->timeout.list);
1312 io_kill_timeout(req);
1313 } while (!list_empty(&ctx->timeout_list));
1315 ctx->cq_last_tm_flush = seq;
1318 static void io_commit_cqring(struct io_ring_ctx *ctx)
1320 io_flush_timeouts(ctx);
1322 /* order cqe stores with ring update */
1323 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1325 if (unlikely(!list_empty(&ctx->defer_list)))
1326 __io_queue_deferred(ctx);
1329 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1331 struct io_rings *r = ctx->rings;
1333 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1336 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1338 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1341 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1343 struct io_rings *rings = ctx->rings;
1347 * writes to the cq entry need to come after reading head; the
1348 * control dependency is enough as we're using WRITE_ONCE to
1351 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1354 tail = ctx->cached_cq_tail++;
1355 return &rings->cqes[tail & ctx->cq_mask];
1358 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1362 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1364 if (!ctx->eventfd_async)
1366 return io_wq_current_is_worker();
1369 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1371 /* see waitqueue_active() comment */
1374 if (waitqueue_active(&ctx->wait))
1375 wake_up(&ctx->wait);
1376 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1377 wake_up(&ctx->sq_data->wait);
1378 if (io_should_trigger_evfd(ctx))
1379 eventfd_signal(ctx->cq_ev_fd, 1);
1380 if (waitqueue_active(&ctx->cq_wait)) {
1381 wake_up_interruptible(&ctx->cq_wait);
1382 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1386 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1388 /* see waitqueue_active() comment */
1391 if (ctx->flags & IORING_SETUP_SQPOLL) {
1392 if (waitqueue_active(&ctx->wait))
1393 wake_up(&ctx->wait);
1395 if (io_should_trigger_evfd(ctx))
1396 eventfd_signal(ctx->cq_ev_fd, 1);
1397 if (waitqueue_active(&ctx->cq_wait)) {
1398 wake_up_interruptible(&ctx->cq_wait);
1399 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1403 /* Returns true if there are no backlogged entries after the flush */
1404 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1405 struct task_struct *tsk,
1406 struct files_struct *files)
1408 struct io_rings *rings = ctx->rings;
1409 struct io_kiocb *req, *tmp;
1410 struct io_uring_cqe *cqe;
1411 unsigned long flags;
1412 bool all_flushed, posted;
1415 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1419 spin_lock_irqsave(&ctx->completion_lock, flags);
1420 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1421 if (!io_match_task(req, tsk, files))
1424 cqe = io_get_cqring(ctx);
1428 list_move(&req->compl.list, &list);
1430 WRITE_ONCE(cqe->user_data, req->user_data);
1431 WRITE_ONCE(cqe->res, req->result);
1432 WRITE_ONCE(cqe->flags, req->compl.cflags);
1434 ctx->cached_cq_overflow++;
1435 WRITE_ONCE(ctx->rings->cq_overflow,
1436 ctx->cached_cq_overflow);
1441 all_flushed = list_empty(&ctx->cq_overflow_list);
1443 clear_bit(0, &ctx->sq_check_overflow);
1444 clear_bit(0, &ctx->cq_check_overflow);
1445 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1449 io_commit_cqring(ctx);
1450 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1452 io_cqring_ev_posted(ctx);
1454 while (!list_empty(&list)) {
1455 req = list_first_entry(&list, struct io_kiocb, compl.list);
1456 list_del(&req->compl.list);
1463 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1464 struct task_struct *tsk,
1465 struct files_struct *files)
1469 if (test_bit(0, &ctx->cq_check_overflow)) {
1470 /* iopoll syncs against uring_lock, not completion_lock */
1471 if (ctx->flags & IORING_SETUP_IOPOLL)
1472 mutex_lock(&ctx->uring_lock);
1473 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1474 if (ctx->flags & IORING_SETUP_IOPOLL)
1475 mutex_unlock(&ctx->uring_lock);
1481 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1483 struct io_ring_ctx *ctx = req->ctx;
1484 struct io_uring_cqe *cqe;
1486 trace_io_uring_complete(ctx, req->user_data, res);
1489 * If we can't get a cq entry, userspace overflowed the
1490 * submission (by quite a lot). Increment the overflow count in
1493 cqe = io_get_cqring(ctx);
1495 WRITE_ONCE(cqe->user_data, req->user_data);
1496 WRITE_ONCE(cqe->res, res);
1497 WRITE_ONCE(cqe->flags, cflags);
1498 } else if (ctx->cq_overflow_flushed ||
1499 atomic_read(&req->task->io_uring->in_idle)) {
1501 * If we're in ring overflow flush mode, or in task cancel mode,
1502 * then we cannot store the request for later flushing, we need
1503 * to drop it on the floor.
1505 ctx->cached_cq_overflow++;
1506 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1508 if (list_empty(&ctx->cq_overflow_list)) {
1509 set_bit(0, &ctx->sq_check_overflow);
1510 set_bit(0, &ctx->cq_check_overflow);
1511 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1515 req->compl.cflags = cflags;
1516 refcount_inc(&req->refs);
1517 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1521 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1523 __io_cqring_fill_event(req, res, 0);
1526 static void io_req_complete_post(struct io_kiocb *req, long res,
1527 unsigned int cflags)
1529 struct io_ring_ctx *ctx = req->ctx;
1530 unsigned long flags;
1532 spin_lock_irqsave(&ctx->completion_lock, flags);
1533 __io_cqring_fill_event(req, res, cflags);
1535 * If we're the last reference to this request, add to our locked
1538 if (refcount_dec_and_test(&req->refs)) {
1539 struct io_comp_state *cs = &ctx->submit_state.comp;
1541 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1542 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1543 io_disarm_next(req);
1545 io_req_task_queue(req->link);
1549 io_dismantle_req(req);
1550 io_put_task(req->task, 1);
1551 list_add(&req->compl.list, &cs->locked_free_list);
1552 cs->locked_free_nr++;
1555 io_commit_cqring(ctx);
1556 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1557 io_cqring_ev_posted(ctx);
1560 percpu_ref_put(&ctx->refs);
1563 static void io_req_complete_state(struct io_kiocb *req, long res,
1564 unsigned int cflags)
1568 req->compl.cflags = cflags;
1569 req->flags |= REQ_F_COMPLETE_INLINE;
1572 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1573 long res, unsigned cflags)
1575 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1576 io_req_complete_state(req, res, cflags);
1578 io_req_complete_post(req, res, cflags);
1581 static inline void io_req_complete(struct io_kiocb *req, long res)
1583 __io_req_complete(req, 0, res, 0);
1586 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1588 struct io_submit_state *state = &ctx->submit_state;
1589 struct io_comp_state *cs = &state->comp;
1590 struct io_kiocb *req = NULL;
1593 * If we have more than a batch's worth of requests in our IRQ side
1594 * locked cache, grab the lock and move them over to our submission
1597 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1598 spin_lock_irq(&ctx->completion_lock);
1599 list_splice_init(&cs->locked_free_list, &cs->free_list);
1600 cs->locked_free_nr = 0;
1601 spin_unlock_irq(&ctx->completion_lock);
1604 while (!list_empty(&cs->free_list)) {
1605 req = list_first_entry(&cs->free_list, struct io_kiocb,
1607 list_del(&req->compl.list);
1608 state->reqs[state->free_reqs++] = req;
1609 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1616 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1618 struct io_submit_state *state = &ctx->submit_state;
1620 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1622 if (!state->free_reqs) {
1623 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1626 if (io_flush_cached_reqs(ctx))
1629 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1633 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1634 * retry single alloc to be on the safe side.
1636 if (unlikely(ret <= 0)) {
1637 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1638 if (!state->reqs[0])
1642 state->free_reqs = ret;
1646 return state->reqs[state->free_reqs];
1649 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1656 static void io_dismantle_req(struct io_kiocb *req)
1660 if (req->async_data)
1661 kfree(req->async_data);
1663 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1664 if (req->fixed_rsrc_refs)
1665 percpu_ref_put(req->fixed_rsrc_refs);
1666 if (req->work.creds) {
1667 put_cred(req->work.creds);
1668 req->work.creds = NULL;
1671 if (req->flags & REQ_F_INFLIGHT) {
1672 struct io_ring_ctx *ctx = req->ctx;
1673 unsigned long flags;
1675 spin_lock_irqsave(&ctx->inflight_lock, flags);
1676 list_del(&req->inflight_entry);
1677 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1678 req->flags &= ~REQ_F_INFLIGHT;
1682 /* must to be called somewhat shortly after putting a request */
1683 static inline void io_put_task(struct task_struct *task, int nr)
1685 struct io_uring_task *tctx = task->io_uring;
1687 percpu_counter_sub(&tctx->inflight, nr);
1688 if (unlikely(atomic_read(&tctx->in_idle)))
1689 wake_up(&tctx->wait);
1690 put_task_struct_many(task, nr);
1693 static void __io_free_req(struct io_kiocb *req)
1695 struct io_ring_ctx *ctx = req->ctx;
1697 io_dismantle_req(req);
1698 io_put_task(req->task, 1);
1700 kmem_cache_free(req_cachep, req);
1701 percpu_ref_put(&ctx->refs);
1704 static inline void io_remove_next_linked(struct io_kiocb *req)
1706 struct io_kiocb *nxt = req->link;
1708 req->link = nxt->link;
1712 static bool io_kill_linked_timeout(struct io_kiocb *req)
1713 __must_hold(&req->ctx->completion_lock)
1715 struct io_kiocb *link = req->link;
1716 bool cancelled = false;
1719 * Can happen if a linked timeout fired and link had been like
1720 * req -> link t-out -> link t-out [-> ...]
1722 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1723 struct io_timeout_data *io = link->async_data;
1726 io_remove_next_linked(req);
1727 link->timeout.head = NULL;
1728 ret = hrtimer_try_to_cancel(&io->timer);
1730 io_cqring_fill_event(link, -ECANCELED);
1731 io_put_req_deferred(link, 1);
1735 req->flags &= ~REQ_F_LINK_TIMEOUT;
1739 static void io_fail_links(struct io_kiocb *req)
1740 __must_hold(&req->ctx->completion_lock)
1742 struct io_kiocb *nxt, *link = req->link;
1749 trace_io_uring_fail_link(req, link);
1750 io_cqring_fill_event(link, -ECANCELED);
1751 io_put_req_deferred(link, 2);
1756 static bool io_disarm_next(struct io_kiocb *req)
1757 __must_hold(&req->ctx->completion_lock)
1759 bool posted = false;
1761 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1762 posted = io_kill_linked_timeout(req);
1763 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1764 posted |= (req->link != NULL);
1770 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1772 struct io_kiocb *nxt;
1775 * If LINK is set, we have dependent requests in this chain. If we
1776 * didn't fail this request, queue the first one up, moving any other
1777 * dependencies to the next request. In case of failure, fail the rest
1780 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1781 struct io_ring_ctx *ctx = req->ctx;
1782 unsigned long flags;
1785 spin_lock_irqsave(&ctx->completion_lock, flags);
1786 posted = io_disarm_next(req);
1788 io_commit_cqring(req->ctx);
1789 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1791 io_cqring_ev_posted(ctx);
1798 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1800 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1802 return __io_req_find_next(req);
1805 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1809 if (ctx->submit_state.comp.nr) {
1810 mutex_lock(&ctx->uring_lock);
1811 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1812 mutex_unlock(&ctx->uring_lock);
1814 percpu_ref_put(&ctx->refs);
1817 static bool __tctx_task_work(struct io_uring_task *tctx)
1819 struct io_ring_ctx *ctx = NULL;
1820 struct io_wq_work_list list;
1821 struct io_wq_work_node *node;
1823 if (wq_list_empty(&tctx->task_list))
1826 spin_lock_irq(&tctx->task_lock);
1827 list = tctx->task_list;
1828 INIT_WQ_LIST(&tctx->task_list);
1829 spin_unlock_irq(&tctx->task_lock);
1833 struct io_wq_work_node *next = node->next;
1834 struct io_kiocb *req;
1836 req = container_of(node, struct io_kiocb, io_task_work.node);
1837 if (req->ctx != ctx) {
1838 ctx_flush_and_put(ctx);
1840 percpu_ref_get(&ctx->refs);
1843 req->task_work.func(&req->task_work);
1847 ctx_flush_and_put(ctx);
1848 return list.first != NULL;
1851 static void tctx_task_work(struct callback_head *cb)
1853 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1855 clear_bit(0, &tctx->task_state);
1857 while (__tctx_task_work(tctx))
1861 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1862 enum task_work_notify_mode notify)
1864 struct io_uring_task *tctx = tsk->io_uring;
1865 struct io_wq_work_node *node, *prev;
1866 unsigned long flags;
1869 WARN_ON_ONCE(!tctx);
1871 spin_lock_irqsave(&tctx->task_lock, flags);
1872 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1873 spin_unlock_irqrestore(&tctx->task_lock, flags);
1875 /* task_work already pending, we're done */
1876 if (test_bit(0, &tctx->task_state) ||
1877 test_and_set_bit(0, &tctx->task_state))
1880 if (!task_work_add(tsk, &tctx->task_work, notify))
1884 * Slow path - we failed, find and delete work. if the work is not
1885 * in the list, it got run and we're fine.
1888 spin_lock_irqsave(&tctx->task_lock, flags);
1889 wq_list_for_each(node, prev, &tctx->task_list) {
1890 if (&req->io_task_work.node == node) {
1891 wq_list_del(&tctx->task_list, node, prev);
1896 spin_unlock_irqrestore(&tctx->task_lock, flags);
1897 clear_bit(0, &tctx->task_state);
1901 static int io_req_task_work_add(struct io_kiocb *req)
1903 struct task_struct *tsk = req->task;
1904 struct io_ring_ctx *ctx = req->ctx;
1905 enum task_work_notify_mode notify;
1908 if (tsk->flags & PF_EXITING)
1912 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1913 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1914 * processing task_work. There's no reliable way to tell if TWA_RESUME
1918 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1919 notify = TWA_SIGNAL;
1921 ret = io_task_work_add(tsk, req, notify);
1923 wake_up_process(tsk);
1928 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1929 task_work_func_t cb)
1931 struct io_ring_ctx *ctx = req->ctx;
1932 struct callback_head *head;
1934 init_task_work(&req->task_work, cb);
1936 head = READ_ONCE(ctx->exit_task_work);
1937 req->task_work.next = head;
1938 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
1941 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1943 struct io_ring_ctx *ctx = req->ctx;
1945 spin_lock_irq(&ctx->completion_lock);
1946 io_cqring_fill_event(req, error);
1947 io_commit_cqring(ctx);
1948 spin_unlock_irq(&ctx->completion_lock);
1950 io_cqring_ev_posted(ctx);
1951 req_set_fail_links(req);
1952 io_double_put_req(req);
1955 static void io_req_task_cancel(struct callback_head *cb)
1957 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1958 struct io_ring_ctx *ctx = req->ctx;
1960 mutex_lock(&ctx->uring_lock);
1961 __io_req_task_cancel(req, req->result);
1962 mutex_unlock(&ctx->uring_lock);
1963 percpu_ref_put(&ctx->refs);
1966 static void __io_req_task_submit(struct io_kiocb *req)
1968 struct io_ring_ctx *ctx = req->ctx;
1970 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1971 mutex_lock(&ctx->uring_lock);
1972 if (!(current->flags & PF_EXITING) && !current->in_execve)
1973 __io_queue_sqe(req);
1975 __io_req_task_cancel(req, -EFAULT);
1976 mutex_unlock(&ctx->uring_lock);
1979 static void io_req_task_submit(struct callback_head *cb)
1981 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1983 __io_req_task_submit(req);
1986 static void io_req_task_queue(struct io_kiocb *req)
1990 req->task_work.func = io_req_task_submit;
1991 ret = io_req_task_work_add(req);
1992 if (unlikely(ret)) {
1993 req->result = -ECANCELED;
1994 percpu_ref_get(&req->ctx->refs);
1995 io_req_task_work_add_fallback(req, io_req_task_cancel);
1999 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2001 percpu_ref_get(&req->ctx->refs);
2003 req->task_work.func = io_req_task_cancel;
2005 if (unlikely(io_req_task_work_add(req)))
2006 io_req_task_work_add_fallback(req, io_req_task_cancel);
2009 static inline void io_queue_next(struct io_kiocb *req)
2011 struct io_kiocb *nxt = io_req_find_next(req);
2014 io_req_task_queue(nxt);
2017 static void io_free_req(struct io_kiocb *req)
2024 struct task_struct *task;
2029 static inline void io_init_req_batch(struct req_batch *rb)
2036 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2037 struct req_batch *rb)
2040 io_put_task(rb->task, rb->task_refs);
2042 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2045 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2046 struct io_submit_state *state)
2050 if (req->task != rb->task) {
2052 io_put_task(rb->task, rb->task_refs);
2053 rb->task = req->task;
2059 io_dismantle_req(req);
2060 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2061 state->reqs[state->free_reqs++] = req;
2063 list_add(&req->compl.list, &state->comp.free_list);
2066 static void io_submit_flush_completions(struct io_comp_state *cs,
2067 struct io_ring_ctx *ctx)
2070 struct io_kiocb *req;
2071 struct req_batch rb;
2073 io_init_req_batch(&rb);
2074 spin_lock_irq(&ctx->completion_lock);
2075 for (i = 0; i < nr; i++) {
2077 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2079 io_commit_cqring(ctx);
2080 spin_unlock_irq(&ctx->completion_lock);
2082 io_cqring_ev_posted(ctx);
2083 for (i = 0; i < nr; i++) {
2086 /* submission and completion refs */
2087 if (refcount_sub_and_test(2, &req->refs))
2088 io_req_free_batch(&rb, req, &ctx->submit_state);
2091 io_req_free_batch_finish(ctx, &rb);
2096 * Drop reference to request, return next in chain (if there is one) if this
2097 * was the last reference to this request.
2099 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2101 struct io_kiocb *nxt = NULL;
2103 if (refcount_dec_and_test(&req->refs)) {
2104 nxt = io_req_find_next(req);
2110 static void io_put_req(struct io_kiocb *req)
2112 if (refcount_dec_and_test(&req->refs))
2116 static void io_put_req_deferred_cb(struct callback_head *cb)
2118 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2123 static void io_free_req_deferred(struct io_kiocb *req)
2127 req->task_work.func = io_put_req_deferred_cb;
2128 ret = io_req_task_work_add(req);
2130 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2133 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2135 if (refcount_sub_and_test(refs, &req->refs))
2136 io_free_req_deferred(req);
2139 static void io_double_put_req(struct io_kiocb *req)
2141 /* drop both submit and complete references */
2142 if (refcount_sub_and_test(2, &req->refs))
2146 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2148 /* See comment at the top of this file */
2150 return __io_cqring_events(ctx);
2153 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2155 struct io_rings *rings = ctx->rings;
2157 /* make sure SQ entry isn't read before tail */
2158 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2161 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2163 unsigned int cflags;
2165 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2166 cflags |= IORING_CQE_F_BUFFER;
2167 req->flags &= ~REQ_F_BUFFER_SELECTED;
2172 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2174 struct io_buffer *kbuf;
2176 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2177 return io_put_kbuf(req, kbuf);
2180 static inline bool io_run_task_work(void)
2183 * Not safe to run on exiting task, and the task_work handling will
2184 * not add work to such a task.
2186 if (unlikely(current->flags & PF_EXITING))
2188 if (current->task_works) {
2189 __set_current_state(TASK_RUNNING);
2198 * Find and free completed poll iocbs
2200 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2201 struct list_head *done)
2203 struct req_batch rb;
2204 struct io_kiocb *req;
2206 /* order with ->result store in io_complete_rw_iopoll() */
2209 io_init_req_batch(&rb);
2210 while (!list_empty(done)) {
2213 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2214 list_del(&req->inflight_entry);
2216 if (READ_ONCE(req->result) == -EAGAIN) {
2217 req->iopoll_completed = 0;
2218 if (io_rw_reissue(req))
2222 if (req->flags & REQ_F_BUFFER_SELECTED)
2223 cflags = io_put_rw_kbuf(req);
2225 __io_cqring_fill_event(req, req->result, cflags);
2228 if (refcount_dec_and_test(&req->refs))
2229 io_req_free_batch(&rb, req, &ctx->submit_state);
2232 io_commit_cqring(ctx);
2233 io_cqring_ev_posted_iopoll(ctx);
2234 io_req_free_batch_finish(ctx, &rb);
2237 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2240 struct io_kiocb *req, *tmp;
2246 * Only spin for completions if we don't have multiple devices hanging
2247 * off our complete list, and we're under the requested amount.
2249 spin = !ctx->poll_multi_file && *nr_events < min;
2252 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2253 struct kiocb *kiocb = &req->rw.kiocb;
2256 * Move completed and retryable entries to our local lists.
2257 * If we find a request that requires polling, break out
2258 * and complete those lists first, if we have entries there.
2260 if (READ_ONCE(req->iopoll_completed)) {
2261 list_move_tail(&req->inflight_entry, &done);
2264 if (!list_empty(&done))
2267 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2271 /* iopoll may have completed current req */
2272 if (READ_ONCE(req->iopoll_completed))
2273 list_move_tail(&req->inflight_entry, &done);
2280 if (!list_empty(&done))
2281 io_iopoll_complete(ctx, nr_events, &done);
2287 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2288 * non-spinning poll check - we'll still enter the driver poll loop, but only
2289 * as a non-spinning completion check.
2291 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2294 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2297 ret = io_do_iopoll(ctx, nr_events, min);
2300 if (*nr_events >= min)
2308 * We can't just wait for polled events to come to us, we have to actively
2309 * find and complete them.
2311 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2313 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2316 mutex_lock(&ctx->uring_lock);
2317 while (!list_empty(&ctx->iopoll_list)) {
2318 unsigned int nr_events = 0;
2320 io_do_iopoll(ctx, &nr_events, 0);
2322 /* let it sleep and repeat later if can't complete a request */
2326 * Ensure we allow local-to-the-cpu processing to take place,
2327 * in this case we need to ensure that we reap all events.
2328 * Also let task_work, etc. to progress by releasing the mutex
2330 if (need_resched()) {
2331 mutex_unlock(&ctx->uring_lock);
2333 mutex_lock(&ctx->uring_lock);
2336 mutex_unlock(&ctx->uring_lock);
2339 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2341 unsigned int nr_events = 0;
2342 int iters = 0, ret = 0;
2345 * We disallow the app entering submit/complete with polling, but we
2346 * still need to lock the ring to prevent racing with polled issue
2347 * that got punted to a workqueue.
2349 mutex_lock(&ctx->uring_lock);
2352 * Don't enter poll loop if we already have events pending.
2353 * If we do, we can potentially be spinning for commands that
2354 * already triggered a CQE (eg in error).
2356 if (test_bit(0, &ctx->cq_check_overflow))
2357 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2358 if (io_cqring_events(ctx))
2362 * If a submit got punted to a workqueue, we can have the
2363 * application entering polling for a command before it gets
2364 * issued. That app will hold the uring_lock for the duration
2365 * of the poll right here, so we need to take a breather every
2366 * now and then to ensure that the issue has a chance to add
2367 * the poll to the issued list. Otherwise we can spin here
2368 * forever, while the workqueue is stuck trying to acquire the
2371 if (!(++iters & 7)) {
2372 mutex_unlock(&ctx->uring_lock);
2374 mutex_lock(&ctx->uring_lock);
2377 ret = io_iopoll_getevents(ctx, &nr_events, min);
2381 } while (min && !nr_events && !need_resched());
2383 mutex_unlock(&ctx->uring_lock);
2387 static void kiocb_end_write(struct io_kiocb *req)
2390 * Tell lockdep we inherited freeze protection from submission
2393 if (req->flags & REQ_F_ISREG) {
2394 struct inode *inode = file_inode(req->file);
2396 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2398 file_end_write(req->file);
2402 static bool io_resubmit_prep(struct io_kiocb *req)
2404 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2406 struct iov_iter iter;
2408 /* already prepared */
2409 if (req->async_data)
2412 switch (req->opcode) {
2413 case IORING_OP_READV:
2414 case IORING_OP_READ_FIXED:
2415 case IORING_OP_READ:
2418 case IORING_OP_WRITEV:
2419 case IORING_OP_WRITE_FIXED:
2420 case IORING_OP_WRITE:
2424 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2429 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2432 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2435 static bool io_rw_should_reissue(struct io_kiocb *req)
2437 umode_t mode = file_inode(req->file)->i_mode;
2438 struct io_ring_ctx *ctx = req->ctx;
2440 if (!S_ISBLK(mode) && !S_ISREG(mode))
2442 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2443 !(ctx->flags & IORING_SETUP_IOPOLL)))
2446 * If ref is dying, we might be running poll reap from the exit work.
2447 * Don't attempt to reissue from that path, just let it fail with
2450 if (percpu_ref_is_dying(&ctx->refs))
2456 static bool io_rw_reissue(struct io_kiocb *req)
2459 if (!io_rw_should_reissue(req))
2462 lockdep_assert_held(&req->ctx->uring_lock);
2464 if (io_resubmit_prep(req)) {
2465 refcount_inc(&req->refs);
2466 io_queue_async_work(req);
2469 req_set_fail_links(req);
2474 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2475 unsigned int issue_flags)
2479 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2481 if (res != req->result)
2482 req_set_fail_links(req);
2484 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2485 kiocb_end_write(req);
2486 if (req->flags & REQ_F_BUFFER_SELECTED)
2487 cflags = io_put_rw_kbuf(req);
2488 __io_req_complete(req, issue_flags, res, cflags);
2491 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2493 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2495 __io_complete_rw(req, res, res2, 0);
2498 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2500 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2503 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2504 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2505 struct io_async_rw *rw = req->async_data;
2508 iov_iter_revert(&rw->iter,
2509 req->result - iov_iter_count(&rw->iter));
2510 else if (!io_resubmit_prep(req))
2515 if (kiocb->ki_flags & IOCB_WRITE)
2516 kiocb_end_write(req);
2518 if (res != -EAGAIN && res != req->result)
2519 req_set_fail_links(req);
2521 WRITE_ONCE(req->result, res);
2522 /* order with io_poll_complete() checking ->result */
2524 WRITE_ONCE(req->iopoll_completed, 1);
2528 * After the iocb has been issued, it's safe to be found on the poll list.
2529 * Adding the kiocb to the list AFTER submission ensures that we don't
2530 * find it from a io_iopoll_getevents() thread before the issuer is done
2531 * accessing the kiocb cookie.
2533 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2535 struct io_ring_ctx *ctx = req->ctx;
2538 * Track whether we have multiple files in our lists. This will impact
2539 * how we do polling eventually, not spinning if we're on potentially
2540 * different devices.
2542 if (list_empty(&ctx->iopoll_list)) {
2543 ctx->poll_multi_file = false;
2544 } else if (!ctx->poll_multi_file) {
2545 struct io_kiocb *list_req;
2547 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2549 if (list_req->file != req->file)
2550 ctx->poll_multi_file = true;
2554 * For fast devices, IO may have already completed. If it has, add
2555 * it to the front so we find it first.
2557 if (READ_ONCE(req->iopoll_completed))
2558 list_add(&req->inflight_entry, &ctx->iopoll_list);
2560 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2563 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2564 * task context or in io worker task context. If current task context is
2565 * sq thread, we don't need to check whether should wake up sq thread.
2567 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2568 wq_has_sleeper(&ctx->sq_data->wait))
2569 wake_up(&ctx->sq_data->wait);
2572 static inline void io_state_file_put(struct io_submit_state *state)
2574 if (state->file_refs) {
2575 fput_many(state->file, state->file_refs);
2576 state->file_refs = 0;
2581 * Get as many references to a file as we have IOs left in this submission,
2582 * assuming most submissions are for one file, or at least that each file
2583 * has more than one submission.
2585 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2590 if (state->file_refs) {
2591 if (state->fd == fd) {
2595 io_state_file_put(state);
2597 state->file = fget_many(fd, state->ios_left);
2598 if (unlikely(!state->file))
2602 state->file_refs = state->ios_left - 1;
2606 static bool io_bdev_nowait(struct block_device *bdev)
2608 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2612 * If we tracked the file through the SCM inflight mechanism, we could support
2613 * any file. For now, just ensure that anything potentially problematic is done
2616 static bool io_file_supports_async(struct file *file, int rw)
2618 umode_t mode = file_inode(file)->i_mode;
2620 if (S_ISBLK(mode)) {
2621 if (IS_ENABLED(CONFIG_BLOCK) &&
2622 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2626 if (S_ISCHR(mode) || S_ISSOCK(mode))
2628 if (S_ISREG(mode)) {
2629 if (IS_ENABLED(CONFIG_BLOCK) &&
2630 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2631 file->f_op != &io_uring_fops)
2636 /* any ->read/write should understand O_NONBLOCK */
2637 if (file->f_flags & O_NONBLOCK)
2640 if (!(file->f_mode & FMODE_NOWAIT))
2644 return file->f_op->read_iter != NULL;
2646 return file->f_op->write_iter != NULL;
2649 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2651 struct io_ring_ctx *ctx = req->ctx;
2652 struct kiocb *kiocb = &req->rw.kiocb;
2653 struct file *file = req->file;
2657 if (S_ISREG(file_inode(file)->i_mode))
2658 req->flags |= REQ_F_ISREG;
2660 kiocb->ki_pos = READ_ONCE(sqe->off);
2661 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2662 req->flags |= REQ_F_CUR_POS;
2663 kiocb->ki_pos = file->f_pos;
2665 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2666 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2667 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2671 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2672 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2673 req->flags |= REQ_F_NOWAIT;
2675 ioprio = READ_ONCE(sqe->ioprio);
2677 ret = ioprio_check_cap(ioprio);
2681 kiocb->ki_ioprio = ioprio;
2683 kiocb->ki_ioprio = get_current_ioprio();
2685 if (ctx->flags & IORING_SETUP_IOPOLL) {
2686 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2687 !kiocb->ki_filp->f_op->iopoll)
2690 kiocb->ki_flags |= IOCB_HIPRI;
2691 kiocb->ki_complete = io_complete_rw_iopoll;
2692 req->iopoll_completed = 0;
2694 if (kiocb->ki_flags & IOCB_HIPRI)
2696 kiocb->ki_complete = io_complete_rw;
2699 req->rw.addr = READ_ONCE(sqe->addr);
2700 req->rw.len = READ_ONCE(sqe->len);
2701 req->buf_index = READ_ONCE(sqe->buf_index);
2705 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2711 case -ERESTARTNOINTR:
2712 case -ERESTARTNOHAND:
2713 case -ERESTART_RESTARTBLOCK:
2715 * We can't just restart the syscall, since previously
2716 * submitted sqes may already be in progress. Just fail this
2722 kiocb->ki_complete(kiocb, ret, 0);
2726 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2727 unsigned int issue_flags)
2729 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2730 struct io_async_rw *io = req->async_data;
2732 /* add previously done IO, if any */
2733 if (io && io->bytes_done > 0) {
2735 ret = io->bytes_done;
2737 ret += io->bytes_done;
2740 if (req->flags & REQ_F_CUR_POS)
2741 req->file->f_pos = kiocb->ki_pos;
2742 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2743 __io_complete_rw(req, ret, 0, issue_flags);
2745 io_rw_done(kiocb, ret);
2748 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2750 struct io_ring_ctx *ctx = req->ctx;
2751 size_t len = req->rw.len;
2752 struct io_mapped_ubuf *imu;
2753 u16 index, buf_index = req->buf_index;
2757 if (unlikely(buf_index >= ctx->nr_user_bufs))
2759 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2760 imu = &ctx->user_bufs[index];
2761 buf_addr = req->rw.addr;
2764 if (buf_addr + len < buf_addr)
2766 /* not inside the mapped region */
2767 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2771 * May not be a start of buffer, set size appropriately
2772 * and advance us to the beginning.
2774 offset = buf_addr - imu->ubuf;
2775 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2779 * Don't use iov_iter_advance() here, as it's really slow for
2780 * using the latter parts of a big fixed buffer - it iterates
2781 * over each segment manually. We can cheat a bit here, because
2784 * 1) it's a BVEC iter, we set it up
2785 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2786 * first and last bvec
2788 * So just find our index, and adjust the iterator afterwards.
2789 * If the offset is within the first bvec (or the whole first
2790 * bvec, just use iov_iter_advance(). This makes it easier
2791 * since we can just skip the first segment, which may not
2792 * be PAGE_SIZE aligned.
2794 const struct bio_vec *bvec = imu->bvec;
2796 if (offset <= bvec->bv_len) {
2797 iov_iter_advance(iter, offset);
2799 unsigned long seg_skip;
2801 /* skip first vec */
2802 offset -= bvec->bv_len;
2803 seg_skip = 1 + (offset >> PAGE_SHIFT);
2805 iter->bvec = bvec + seg_skip;
2806 iter->nr_segs -= seg_skip;
2807 iter->count -= bvec->bv_len + offset;
2808 iter->iov_offset = offset & ~PAGE_MASK;
2815 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2818 mutex_unlock(&ctx->uring_lock);
2821 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2824 * "Normal" inline submissions always hold the uring_lock, since we
2825 * grab it from the system call. Same is true for the SQPOLL offload.
2826 * The only exception is when we've detached the request and issue it
2827 * from an async worker thread, grab the lock for that case.
2830 mutex_lock(&ctx->uring_lock);
2833 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2834 int bgid, struct io_buffer *kbuf,
2837 struct io_buffer *head;
2839 if (req->flags & REQ_F_BUFFER_SELECTED)
2842 io_ring_submit_lock(req->ctx, needs_lock);
2844 lockdep_assert_held(&req->ctx->uring_lock);
2846 head = xa_load(&req->ctx->io_buffers, bgid);
2848 if (!list_empty(&head->list)) {
2849 kbuf = list_last_entry(&head->list, struct io_buffer,
2851 list_del(&kbuf->list);
2854 xa_erase(&req->ctx->io_buffers, bgid);
2856 if (*len > kbuf->len)
2859 kbuf = ERR_PTR(-ENOBUFS);
2862 io_ring_submit_unlock(req->ctx, needs_lock);
2867 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2870 struct io_buffer *kbuf;
2873 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2874 bgid = req->buf_index;
2875 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2878 req->rw.addr = (u64) (unsigned long) kbuf;
2879 req->flags |= REQ_F_BUFFER_SELECTED;
2880 return u64_to_user_ptr(kbuf->addr);
2883 #ifdef CONFIG_COMPAT
2884 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2887 struct compat_iovec __user *uiov;
2888 compat_ssize_t clen;
2892 uiov = u64_to_user_ptr(req->rw.addr);
2893 if (!access_ok(uiov, sizeof(*uiov)))
2895 if (__get_user(clen, &uiov->iov_len))
2901 buf = io_rw_buffer_select(req, &len, needs_lock);
2903 return PTR_ERR(buf);
2904 iov[0].iov_base = buf;
2905 iov[0].iov_len = (compat_size_t) len;
2910 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2913 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2917 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2920 len = iov[0].iov_len;
2923 buf = io_rw_buffer_select(req, &len, needs_lock);
2925 return PTR_ERR(buf);
2926 iov[0].iov_base = buf;
2927 iov[0].iov_len = len;
2931 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2934 if (req->flags & REQ_F_BUFFER_SELECTED) {
2935 struct io_buffer *kbuf;
2937 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2938 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2939 iov[0].iov_len = kbuf->len;
2942 if (req->rw.len != 1)
2945 #ifdef CONFIG_COMPAT
2946 if (req->ctx->compat)
2947 return io_compat_import(req, iov, needs_lock);
2950 return __io_iov_buffer_select(req, iov, needs_lock);
2953 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2954 struct iov_iter *iter, bool needs_lock)
2956 void __user *buf = u64_to_user_ptr(req->rw.addr);
2957 size_t sqe_len = req->rw.len;
2958 u8 opcode = req->opcode;
2961 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2963 return io_import_fixed(req, rw, iter);
2966 /* buffer index only valid with fixed read/write, or buffer select */
2967 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2970 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2971 if (req->flags & REQ_F_BUFFER_SELECT) {
2972 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2974 return PTR_ERR(buf);
2975 req->rw.len = sqe_len;
2978 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2983 if (req->flags & REQ_F_BUFFER_SELECT) {
2984 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2986 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2991 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2995 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2997 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3001 * For files that don't have ->read_iter() and ->write_iter(), handle them
3002 * by looping over ->read() or ->write() manually.
3004 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3006 struct kiocb *kiocb = &req->rw.kiocb;
3007 struct file *file = req->file;
3011 * Don't support polled IO through this interface, and we can't
3012 * support non-blocking either. For the latter, this just causes
3013 * the kiocb to be handled from an async context.
3015 if (kiocb->ki_flags & IOCB_HIPRI)
3017 if (kiocb->ki_flags & IOCB_NOWAIT)
3020 while (iov_iter_count(iter)) {
3024 if (!iov_iter_is_bvec(iter)) {
3025 iovec = iov_iter_iovec(iter);
3027 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3028 iovec.iov_len = req->rw.len;
3032 nr = file->f_op->read(file, iovec.iov_base,
3033 iovec.iov_len, io_kiocb_ppos(kiocb));
3035 nr = file->f_op->write(file, iovec.iov_base,
3036 iovec.iov_len, io_kiocb_ppos(kiocb));
3045 if (nr != iovec.iov_len)
3049 iov_iter_advance(iter, nr);
3055 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3056 const struct iovec *fast_iov, struct iov_iter *iter)
3058 struct io_async_rw *rw = req->async_data;
3060 memcpy(&rw->iter, iter, sizeof(*iter));
3061 rw->free_iovec = iovec;
3063 /* can only be fixed buffers, no need to do anything */
3064 if (iov_iter_is_bvec(iter))
3067 unsigned iov_off = 0;
3069 rw->iter.iov = rw->fast_iov;
3070 if (iter->iov != fast_iov) {
3071 iov_off = iter->iov - fast_iov;
3072 rw->iter.iov += iov_off;
3074 if (rw->fast_iov != fast_iov)
3075 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3076 sizeof(struct iovec) * iter->nr_segs);
3078 req->flags |= REQ_F_NEED_CLEANUP;
3082 static inline int __io_alloc_async_data(struct io_kiocb *req)
3084 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3085 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3086 return req->async_data == NULL;
3089 static int io_alloc_async_data(struct io_kiocb *req)
3091 if (!io_op_defs[req->opcode].needs_async_data)
3094 return __io_alloc_async_data(req);
3097 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3098 const struct iovec *fast_iov,
3099 struct iov_iter *iter, bool force)
3101 if (!force && !io_op_defs[req->opcode].needs_async_data)
3103 if (!req->async_data) {
3104 if (__io_alloc_async_data(req)) {
3109 io_req_map_rw(req, iovec, fast_iov, iter);
3114 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3116 struct io_async_rw *iorw = req->async_data;
3117 struct iovec *iov = iorw->fast_iov;
3120 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3121 if (unlikely(ret < 0))
3124 iorw->bytes_done = 0;
3125 iorw->free_iovec = iov;
3127 req->flags |= REQ_F_NEED_CLEANUP;
3131 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3133 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3135 return io_prep_rw(req, sqe);
3139 * This is our waitqueue callback handler, registered through lock_page_async()
3140 * when we initially tried to do the IO with the iocb armed our waitqueue.
3141 * This gets called when the page is unlocked, and we generally expect that to
3142 * happen when the page IO is completed and the page is now uptodate. This will
3143 * queue a task_work based retry of the operation, attempting to copy the data
3144 * again. If the latter fails because the page was NOT uptodate, then we will
3145 * do a thread based blocking retry of the operation. That's the unexpected
3148 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3149 int sync, void *arg)
3151 struct wait_page_queue *wpq;
3152 struct io_kiocb *req = wait->private;
3153 struct wait_page_key *key = arg;
3155 wpq = container_of(wait, struct wait_page_queue, wait);
3157 if (!wake_page_match(wpq, key))
3160 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3161 list_del_init(&wait->entry);
3163 /* submit ref gets dropped, acquire a new one */
3164 refcount_inc(&req->refs);
3165 io_req_task_queue(req);
3170 * This controls whether a given IO request should be armed for async page
3171 * based retry. If we return false here, the request is handed to the async
3172 * worker threads for retry. If we're doing buffered reads on a regular file,
3173 * we prepare a private wait_page_queue entry and retry the operation. This
3174 * will either succeed because the page is now uptodate and unlocked, or it
3175 * will register a callback when the page is unlocked at IO completion. Through
3176 * that callback, io_uring uses task_work to setup a retry of the operation.
3177 * That retry will attempt the buffered read again. The retry will generally
3178 * succeed, or in rare cases where it fails, we then fall back to using the
3179 * async worker threads for a blocking retry.
3181 static bool io_rw_should_retry(struct io_kiocb *req)
3183 struct io_async_rw *rw = req->async_data;
3184 struct wait_page_queue *wait = &rw->wpq;
3185 struct kiocb *kiocb = &req->rw.kiocb;
3187 /* never retry for NOWAIT, we just complete with -EAGAIN */
3188 if (req->flags & REQ_F_NOWAIT)
3191 /* Only for buffered IO */
3192 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3196 * just use poll if we can, and don't attempt if the fs doesn't
3197 * support callback based unlocks
3199 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3202 wait->wait.func = io_async_buf_func;
3203 wait->wait.private = req;
3204 wait->wait.flags = 0;
3205 INIT_LIST_HEAD(&wait->wait.entry);
3206 kiocb->ki_flags |= IOCB_WAITQ;
3207 kiocb->ki_flags &= ~IOCB_NOWAIT;
3208 kiocb->ki_waitq = wait;
3212 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3214 if (req->file->f_op->read_iter)
3215 return call_read_iter(req->file, &req->rw.kiocb, iter);
3216 else if (req->file->f_op->read)
3217 return loop_rw_iter(READ, req, iter);
3222 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3224 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3225 struct kiocb *kiocb = &req->rw.kiocb;
3226 struct iov_iter __iter, *iter = &__iter;
3227 struct io_async_rw *rw = req->async_data;
3228 ssize_t io_size, ret, ret2;
3229 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3235 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3239 io_size = iov_iter_count(iter);
3240 req->result = io_size;
3242 /* Ensure we clear previously set non-block flag */
3243 if (!force_nonblock)
3244 kiocb->ki_flags &= ~IOCB_NOWAIT;
3246 kiocb->ki_flags |= IOCB_NOWAIT;
3248 /* If the file doesn't support async, just async punt */
3249 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3250 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3251 return ret ?: -EAGAIN;
3254 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3255 if (unlikely(ret)) {
3260 ret = io_iter_do_read(req, iter);
3262 if (ret == -EIOCBQUEUED) {
3263 if (req->async_data)
3264 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3266 } else if (ret == -EAGAIN) {
3267 /* IOPOLL retry should happen for io-wq threads */
3268 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3270 /* no retry on NONBLOCK nor RWF_NOWAIT */
3271 if (req->flags & REQ_F_NOWAIT)
3273 /* some cases will consume bytes even on error returns */
3274 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3276 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3277 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3278 /* read all, failed, already did sync or don't want to retry */
3282 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3287 rw = req->async_data;
3288 /* now use our persistent iterator, if we aren't already */
3293 rw->bytes_done += ret;
3294 /* if we can retry, do so with the callbacks armed */
3295 if (!io_rw_should_retry(req)) {
3296 kiocb->ki_flags &= ~IOCB_WAITQ;
3301 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3302 * we get -EIOCBQUEUED, then we'll get a notification when the
3303 * desired page gets unlocked. We can also get a partial read
3304 * here, and if we do, then just retry at the new offset.
3306 ret = io_iter_do_read(req, iter);
3307 if (ret == -EIOCBQUEUED)
3309 /* we got some bytes, but not all. retry. */
3310 kiocb->ki_flags &= ~IOCB_WAITQ;
3311 } while (ret > 0 && ret < io_size);
3313 kiocb_done(kiocb, ret, issue_flags);
3315 /* it's faster to check here then delegate to kfree */
3321 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3323 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3325 return io_prep_rw(req, sqe);
3328 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3330 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3331 struct kiocb *kiocb = &req->rw.kiocb;
3332 struct iov_iter __iter, *iter = &__iter;
3333 struct io_async_rw *rw = req->async_data;
3334 ssize_t ret, ret2, io_size;
3335 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3341 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3345 io_size = iov_iter_count(iter);
3346 req->result = io_size;
3348 /* Ensure we clear previously set non-block flag */
3349 if (!force_nonblock)
3350 kiocb->ki_flags &= ~IOCB_NOWAIT;
3352 kiocb->ki_flags |= IOCB_NOWAIT;
3354 /* If the file doesn't support async, just async punt */
3355 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3358 /* file path doesn't support NOWAIT for non-direct_IO */
3359 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3360 (req->flags & REQ_F_ISREG))
3363 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3368 * Open-code file_start_write here to grab freeze protection,
3369 * which will be released by another thread in
3370 * io_complete_rw(). Fool lockdep by telling it the lock got
3371 * released so that it doesn't complain about the held lock when
3372 * we return to userspace.
3374 if (req->flags & REQ_F_ISREG) {
3375 sb_start_write(file_inode(req->file)->i_sb);
3376 __sb_writers_release(file_inode(req->file)->i_sb,
3379 kiocb->ki_flags |= IOCB_WRITE;
3381 if (req->file->f_op->write_iter)
3382 ret2 = call_write_iter(req->file, kiocb, iter);
3383 else if (req->file->f_op->write)
3384 ret2 = loop_rw_iter(WRITE, req, iter);
3389 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3390 * retry them without IOCB_NOWAIT.
3392 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3394 /* no retry on NONBLOCK nor RWF_NOWAIT */
3395 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3397 if (ret2 == -EIOCBQUEUED && req->async_data)
3398 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3399 if (!force_nonblock || ret2 != -EAGAIN) {
3400 /* IOPOLL retry should happen for io-wq threads */
3401 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3404 kiocb_done(kiocb, ret2, issue_flags);
3407 /* some cases will consume bytes even on error returns */
3408 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3409 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3410 return ret ?: -EAGAIN;
3413 /* it's reportedly faster than delegating the null check to kfree() */
3419 static int io_renameat_prep(struct io_kiocb *req,
3420 const struct io_uring_sqe *sqe)
3422 struct io_rename *ren = &req->rename;
3423 const char __user *oldf, *newf;
3425 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3428 ren->old_dfd = READ_ONCE(sqe->fd);
3429 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3430 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3431 ren->new_dfd = READ_ONCE(sqe->len);
3432 ren->flags = READ_ONCE(sqe->rename_flags);
3434 ren->oldpath = getname(oldf);
3435 if (IS_ERR(ren->oldpath))
3436 return PTR_ERR(ren->oldpath);
3438 ren->newpath = getname(newf);
3439 if (IS_ERR(ren->newpath)) {
3440 putname(ren->oldpath);
3441 return PTR_ERR(ren->newpath);
3444 req->flags |= REQ_F_NEED_CLEANUP;
3448 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3450 struct io_rename *ren = &req->rename;
3453 if (issue_flags & IO_URING_F_NONBLOCK)
3456 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3457 ren->newpath, ren->flags);
3459 req->flags &= ~REQ_F_NEED_CLEANUP;
3461 req_set_fail_links(req);
3462 io_req_complete(req, ret);
3466 static int io_unlinkat_prep(struct io_kiocb *req,
3467 const struct io_uring_sqe *sqe)
3469 struct io_unlink *un = &req->unlink;
3470 const char __user *fname;
3472 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3475 un->dfd = READ_ONCE(sqe->fd);
3477 un->flags = READ_ONCE(sqe->unlink_flags);
3478 if (un->flags & ~AT_REMOVEDIR)
3481 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3482 un->filename = getname(fname);
3483 if (IS_ERR(un->filename))
3484 return PTR_ERR(un->filename);
3486 req->flags |= REQ_F_NEED_CLEANUP;
3490 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3492 struct io_unlink *un = &req->unlink;
3495 if (issue_flags & IO_URING_F_NONBLOCK)
3498 if (un->flags & AT_REMOVEDIR)
3499 ret = do_rmdir(un->dfd, un->filename);
3501 ret = do_unlinkat(un->dfd, un->filename);
3503 req->flags &= ~REQ_F_NEED_CLEANUP;
3505 req_set_fail_links(req);
3506 io_req_complete(req, ret);
3510 static int io_shutdown_prep(struct io_kiocb *req,
3511 const struct io_uring_sqe *sqe)
3513 #if defined(CONFIG_NET)
3514 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3516 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3520 req->shutdown.how = READ_ONCE(sqe->len);
3527 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3529 #if defined(CONFIG_NET)
3530 struct socket *sock;
3533 if (issue_flags & IO_URING_F_NONBLOCK)
3536 sock = sock_from_file(req->file);
3537 if (unlikely(!sock))
3540 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3542 req_set_fail_links(req);
3543 io_req_complete(req, ret);
3550 static int __io_splice_prep(struct io_kiocb *req,
3551 const struct io_uring_sqe *sqe)
3553 struct io_splice* sp = &req->splice;
3554 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3556 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3560 sp->len = READ_ONCE(sqe->len);
3561 sp->flags = READ_ONCE(sqe->splice_flags);
3563 if (unlikely(sp->flags & ~valid_flags))
3566 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3567 (sp->flags & SPLICE_F_FD_IN_FIXED));
3570 req->flags |= REQ_F_NEED_CLEANUP;
3572 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3574 * Splice operation will be punted aync, and here need to
3575 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3577 req->work.flags |= IO_WQ_WORK_UNBOUND;
3583 static int io_tee_prep(struct io_kiocb *req,
3584 const struct io_uring_sqe *sqe)
3586 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3588 return __io_splice_prep(req, sqe);
3591 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3593 struct io_splice *sp = &req->splice;
3594 struct file *in = sp->file_in;
3595 struct file *out = sp->file_out;
3596 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3599 if (issue_flags & IO_URING_F_NONBLOCK)
3602 ret = do_tee(in, out, sp->len, flags);
3604 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3605 req->flags &= ~REQ_F_NEED_CLEANUP;
3608 req_set_fail_links(req);
3609 io_req_complete(req, ret);
3613 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3615 struct io_splice* sp = &req->splice;
3617 sp->off_in = READ_ONCE(sqe->splice_off_in);
3618 sp->off_out = READ_ONCE(sqe->off);
3619 return __io_splice_prep(req, sqe);
3622 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3624 struct io_splice *sp = &req->splice;
3625 struct file *in = sp->file_in;
3626 struct file *out = sp->file_out;
3627 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3628 loff_t *poff_in, *poff_out;
3631 if (issue_flags & IO_URING_F_NONBLOCK)
3634 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3635 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3638 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3640 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3641 req->flags &= ~REQ_F_NEED_CLEANUP;
3644 req_set_fail_links(req);
3645 io_req_complete(req, ret);
3650 * IORING_OP_NOP just posts a completion event, nothing else.
3652 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3654 struct io_ring_ctx *ctx = req->ctx;
3656 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3659 __io_req_complete(req, issue_flags, 0, 0);
3663 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3665 struct io_ring_ctx *ctx = req->ctx;
3670 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3672 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3675 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3676 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3679 req->sync.off = READ_ONCE(sqe->off);
3680 req->sync.len = READ_ONCE(sqe->len);
3684 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3686 loff_t end = req->sync.off + req->sync.len;
3689 /* fsync always requires a blocking context */
3690 if (issue_flags & IO_URING_F_NONBLOCK)
3693 ret = vfs_fsync_range(req->file, req->sync.off,
3694 end > 0 ? end : LLONG_MAX,
3695 req->sync.flags & IORING_FSYNC_DATASYNC);
3697 req_set_fail_links(req);
3698 io_req_complete(req, ret);
3702 static int io_fallocate_prep(struct io_kiocb *req,
3703 const struct io_uring_sqe *sqe)
3705 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3707 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3710 req->sync.off = READ_ONCE(sqe->off);
3711 req->sync.len = READ_ONCE(sqe->addr);
3712 req->sync.mode = READ_ONCE(sqe->len);
3716 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3720 /* fallocate always requiring blocking context */
3721 if (issue_flags & IO_URING_F_NONBLOCK)
3723 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3726 req_set_fail_links(req);
3727 io_req_complete(req, ret);
3731 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3733 const char __user *fname;
3736 if (unlikely(sqe->ioprio || sqe->buf_index))
3738 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3741 /* open.how should be already initialised */
3742 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3743 req->open.how.flags |= O_LARGEFILE;
3745 req->open.dfd = READ_ONCE(sqe->fd);
3746 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3747 req->open.filename = getname(fname);
3748 if (IS_ERR(req->open.filename)) {
3749 ret = PTR_ERR(req->open.filename);
3750 req->open.filename = NULL;
3753 req->open.nofile = rlimit(RLIMIT_NOFILE);
3754 req->flags |= REQ_F_NEED_CLEANUP;
3758 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3762 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3764 mode = READ_ONCE(sqe->len);
3765 flags = READ_ONCE(sqe->open_flags);
3766 req->open.how = build_open_how(flags, mode);
3767 return __io_openat_prep(req, sqe);
3770 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3772 struct open_how __user *how;
3776 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3778 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3779 len = READ_ONCE(sqe->len);
3780 if (len < OPEN_HOW_SIZE_VER0)
3783 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3788 return __io_openat_prep(req, sqe);
3791 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3793 struct open_flags op;
3796 bool resolve_nonblock;
3799 ret = build_open_flags(&req->open.how, &op);
3802 nonblock_set = op.open_flag & O_NONBLOCK;
3803 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3804 if (issue_flags & IO_URING_F_NONBLOCK) {
3806 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3807 * it'll always -EAGAIN
3809 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3811 op.lookup_flags |= LOOKUP_CACHED;
3812 op.open_flag |= O_NONBLOCK;
3815 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3819 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3820 /* only retry if RESOLVE_CACHED wasn't already set by application */
3821 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3822 file == ERR_PTR(-EAGAIN)) {
3824 * We could hang on to this 'fd', but seems like marginal
3825 * gain for something that is now known to be a slower path.
3826 * So just put it, and we'll get a new one when we retry.
3834 ret = PTR_ERR(file);
3836 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3837 file->f_flags &= ~O_NONBLOCK;
3838 fsnotify_open(file);
3839 fd_install(ret, file);
3842 putname(req->open.filename);
3843 req->flags &= ~REQ_F_NEED_CLEANUP;
3845 req_set_fail_links(req);
3846 io_req_complete(req, ret);
3850 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3852 return io_openat2(req, issue_flags);
3855 static int io_remove_buffers_prep(struct io_kiocb *req,
3856 const struct io_uring_sqe *sqe)
3858 struct io_provide_buf *p = &req->pbuf;
3861 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3864 tmp = READ_ONCE(sqe->fd);
3865 if (!tmp || tmp > USHRT_MAX)
3868 memset(p, 0, sizeof(*p));
3870 p->bgid = READ_ONCE(sqe->buf_group);
3874 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3875 int bgid, unsigned nbufs)
3879 /* shouldn't happen */
3883 /* the head kbuf is the list itself */
3884 while (!list_empty(&buf->list)) {
3885 struct io_buffer *nxt;
3887 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3888 list_del(&nxt->list);
3895 xa_erase(&ctx->io_buffers, bgid);
3900 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3902 struct io_provide_buf *p = &req->pbuf;
3903 struct io_ring_ctx *ctx = req->ctx;
3904 struct io_buffer *head;
3906 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3908 io_ring_submit_lock(ctx, !force_nonblock);
3910 lockdep_assert_held(&ctx->uring_lock);
3913 head = xa_load(&ctx->io_buffers, p->bgid);
3915 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3917 req_set_fail_links(req);
3919 /* need to hold the lock to complete IOPOLL requests */
3920 if (ctx->flags & IORING_SETUP_IOPOLL) {
3921 __io_req_complete(req, issue_flags, ret, 0);
3922 io_ring_submit_unlock(ctx, !force_nonblock);
3924 io_ring_submit_unlock(ctx, !force_nonblock);
3925 __io_req_complete(req, issue_flags, ret, 0);
3930 static int io_provide_buffers_prep(struct io_kiocb *req,
3931 const struct io_uring_sqe *sqe)
3933 struct io_provide_buf *p = &req->pbuf;
3936 if (sqe->ioprio || sqe->rw_flags)
3939 tmp = READ_ONCE(sqe->fd);
3940 if (!tmp || tmp > USHRT_MAX)
3943 p->addr = READ_ONCE(sqe->addr);
3944 p->len = READ_ONCE(sqe->len);
3946 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3949 p->bgid = READ_ONCE(sqe->buf_group);
3950 tmp = READ_ONCE(sqe->off);
3951 if (tmp > USHRT_MAX)
3957 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3959 struct io_buffer *buf;
3960 u64 addr = pbuf->addr;
3961 int i, bid = pbuf->bid;
3963 for (i = 0; i < pbuf->nbufs; i++) {
3964 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3969 buf->len = pbuf->len;
3974 INIT_LIST_HEAD(&buf->list);
3977 list_add_tail(&buf->list, &(*head)->list);
3981 return i ? i : -ENOMEM;
3984 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3986 struct io_provide_buf *p = &req->pbuf;
3987 struct io_ring_ctx *ctx = req->ctx;
3988 struct io_buffer *head, *list;
3990 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3992 io_ring_submit_lock(ctx, !force_nonblock);
3994 lockdep_assert_held(&ctx->uring_lock);
3996 list = head = xa_load(&ctx->io_buffers, p->bgid);
3998 ret = io_add_buffers(p, &head);
3999 if (ret >= 0 && !list) {
4000 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4002 __io_remove_buffers(ctx, head, p->bgid, -1U);
4005 req_set_fail_links(req);
4007 /* need to hold the lock to complete IOPOLL requests */
4008 if (ctx->flags & IORING_SETUP_IOPOLL) {
4009 __io_req_complete(req, issue_flags, ret, 0);
4010 io_ring_submit_unlock(ctx, !force_nonblock);
4012 io_ring_submit_unlock(ctx, !force_nonblock);
4013 __io_req_complete(req, issue_flags, ret, 0);
4018 static int io_epoll_ctl_prep(struct io_kiocb *req,
4019 const struct io_uring_sqe *sqe)
4021 #if defined(CONFIG_EPOLL)
4022 if (sqe->ioprio || sqe->buf_index)
4024 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4027 req->epoll.epfd = READ_ONCE(sqe->fd);
4028 req->epoll.op = READ_ONCE(sqe->len);
4029 req->epoll.fd = READ_ONCE(sqe->off);
4031 if (ep_op_has_event(req->epoll.op)) {
4032 struct epoll_event __user *ev;
4034 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4035 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4045 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4047 #if defined(CONFIG_EPOLL)
4048 struct io_epoll *ie = &req->epoll;
4050 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4052 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4053 if (force_nonblock && ret == -EAGAIN)
4057 req_set_fail_links(req);
4058 __io_req_complete(req, issue_flags, ret, 0);
4065 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4067 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4068 if (sqe->ioprio || sqe->buf_index || sqe->off)
4070 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4073 req->madvise.addr = READ_ONCE(sqe->addr);
4074 req->madvise.len = READ_ONCE(sqe->len);
4075 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4082 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4084 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4085 struct io_madvise *ma = &req->madvise;
4088 if (issue_flags & IO_URING_F_NONBLOCK)
4091 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4093 req_set_fail_links(req);
4094 io_req_complete(req, ret);
4101 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4103 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4105 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4108 req->fadvise.offset = READ_ONCE(sqe->off);
4109 req->fadvise.len = READ_ONCE(sqe->len);
4110 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4114 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4116 struct io_fadvise *fa = &req->fadvise;
4119 if (issue_flags & IO_URING_F_NONBLOCK) {
4120 switch (fa->advice) {
4121 case POSIX_FADV_NORMAL:
4122 case POSIX_FADV_RANDOM:
4123 case POSIX_FADV_SEQUENTIAL:
4130 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4132 req_set_fail_links(req);
4133 io_req_complete(req, ret);
4137 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4139 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4141 if (sqe->ioprio || sqe->buf_index)
4143 if (req->flags & REQ_F_FIXED_FILE)
4146 req->statx.dfd = READ_ONCE(sqe->fd);
4147 req->statx.mask = READ_ONCE(sqe->len);
4148 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4149 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4150 req->statx.flags = READ_ONCE(sqe->statx_flags);
4155 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4157 struct io_statx *ctx = &req->statx;
4160 if (issue_flags & IO_URING_F_NONBLOCK) {
4161 /* only need file table for an actual valid fd */
4162 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4163 req->flags |= REQ_F_NO_FILE_TABLE;
4167 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4171 req_set_fail_links(req);
4172 io_req_complete(req, ret);
4176 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4178 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4180 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4181 sqe->rw_flags || sqe->buf_index)
4183 if (req->flags & REQ_F_FIXED_FILE)
4186 req->close.fd = READ_ONCE(sqe->fd);
4190 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4192 struct files_struct *files = current->files;
4193 struct io_close *close = &req->close;
4194 struct fdtable *fdt;
4200 spin_lock(&files->file_lock);
4201 fdt = files_fdtable(files);
4202 if (close->fd >= fdt->max_fds) {
4203 spin_unlock(&files->file_lock);
4206 file = fdt->fd[close->fd];
4208 spin_unlock(&files->file_lock);
4212 if (file->f_op == &io_uring_fops) {
4213 spin_unlock(&files->file_lock);
4218 /* if the file has a flush method, be safe and punt to async */
4219 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4220 spin_unlock(&files->file_lock);
4224 ret = __close_fd_get_file(close->fd, &file);
4225 spin_unlock(&files->file_lock);
4232 /* No ->flush() or already async, safely close from here */
4233 ret = filp_close(file, current->files);
4236 req_set_fail_links(req);
4239 __io_req_complete(req, issue_flags, ret, 0);
4243 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4245 struct io_ring_ctx *ctx = req->ctx;
4247 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4249 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4252 req->sync.off = READ_ONCE(sqe->off);
4253 req->sync.len = READ_ONCE(sqe->len);
4254 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4258 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4262 /* sync_file_range always requires a blocking context */
4263 if (issue_flags & IO_URING_F_NONBLOCK)
4266 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4269 req_set_fail_links(req);
4270 io_req_complete(req, ret);
4274 #if defined(CONFIG_NET)
4275 static int io_setup_async_msg(struct io_kiocb *req,
4276 struct io_async_msghdr *kmsg)
4278 struct io_async_msghdr *async_msg = req->async_data;
4282 if (io_alloc_async_data(req)) {
4283 kfree(kmsg->free_iov);
4286 async_msg = req->async_data;
4287 req->flags |= REQ_F_NEED_CLEANUP;
4288 memcpy(async_msg, kmsg, sizeof(*kmsg));
4289 async_msg->msg.msg_name = &async_msg->addr;
4290 /* if were using fast_iov, set it to the new one */
4291 if (!async_msg->free_iov)
4292 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4297 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4298 struct io_async_msghdr *iomsg)
4300 iomsg->msg.msg_name = &iomsg->addr;
4301 iomsg->free_iov = iomsg->fast_iov;
4302 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4303 req->sr_msg.msg_flags, &iomsg->free_iov);
4306 static int io_sendmsg_prep_async(struct io_kiocb *req)
4310 if (!io_op_defs[req->opcode].needs_async_data)
4312 ret = io_sendmsg_copy_hdr(req, req->async_data);
4314 req->flags |= REQ_F_NEED_CLEANUP;
4318 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4320 struct io_sr_msg *sr = &req->sr_msg;
4322 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4325 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4326 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4327 sr->len = READ_ONCE(sqe->len);
4329 #ifdef CONFIG_COMPAT
4330 if (req->ctx->compat)
4331 sr->msg_flags |= MSG_CMSG_COMPAT;
4336 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4338 struct io_async_msghdr iomsg, *kmsg;
4339 struct socket *sock;
4343 sock = sock_from_file(req->file);
4344 if (unlikely(!sock))
4347 kmsg = req->async_data;
4349 ret = io_sendmsg_copy_hdr(req, &iomsg);
4355 flags = req->sr_msg.msg_flags;
4356 if (flags & MSG_DONTWAIT)
4357 req->flags |= REQ_F_NOWAIT;
4358 else if (issue_flags & IO_URING_F_NONBLOCK)
4359 flags |= MSG_DONTWAIT;
4361 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4362 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4363 return io_setup_async_msg(req, kmsg);
4364 if (ret == -ERESTARTSYS)
4367 /* fast path, check for non-NULL to avoid function call */
4369 kfree(kmsg->free_iov);
4370 req->flags &= ~REQ_F_NEED_CLEANUP;
4372 req_set_fail_links(req);
4373 __io_req_complete(req, issue_flags, ret, 0);
4377 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4379 struct io_sr_msg *sr = &req->sr_msg;
4382 struct socket *sock;
4386 sock = sock_from_file(req->file);
4387 if (unlikely(!sock))
4390 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4394 msg.msg_name = NULL;
4395 msg.msg_control = NULL;
4396 msg.msg_controllen = 0;
4397 msg.msg_namelen = 0;
4399 flags = req->sr_msg.msg_flags;
4400 if (flags & MSG_DONTWAIT)
4401 req->flags |= REQ_F_NOWAIT;
4402 else if (issue_flags & IO_URING_F_NONBLOCK)
4403 flags |= MSG_DONTWAIT;
4405 msg.msg_flags = flags;
4406 ret = sock_sendmsg(sock, &msg);
4407 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4409 if (ret == -ERESTARTSYS)
4413 req_set_fail_links(req);
4414 __io_req_complete(req, issue_flags, ret, 0);
4418 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4419 struct io_async_msghdr *iomsg)
4421 struct io_sr_msg *sr = &req->sr_msg;
4422 struct iovec __user *uiov;
4426 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4427 &iomsg->uaddr, &uiov, &iov_len);
4431 if (req->flags & REQ_F_BUFFER_SELECT) {
4434 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4436 sr->len = iomsg->fast_iov[0].iov_len;
4437 iomsg->free_iov = NULL;
4439 iomsg->free_iov = iomsg->fast_iov;
4440 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4441 &iomsg->free_iov, &iomsg->msg.msg_iter,
4450 #ifdef CONFIG_COMPAT
4451 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4452 struct io_async_msghdr *iomsg)
4454 struct compat_msghdr __user *msg_compat;
4455 struct io_sr_msg *sr = &req->sr_msg;
4456 struct compat_iovec __user *uiov;
4461 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4462 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4467 uiov = compat_ptr(ptr);
4468 if (req->flags & REQ_F_BUFFER_SELECT) {
4469 compat_ssize_t clen;
4473 if (!access_ok(uiov, sizeof(*uiov)))
4475 if (__get_user(clen, &uiov->iov_len))
4480 iomsg->free_iov = NULL;
4482 iomsg->free_iov = iomsg->fast_iov;
4483 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4484 UIO_FASTIOV, &iomsg->free_iov,
4485 &iomsg->msg.msg_iter, true);
4494 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4495 struct io_async_msghdr *iomsg)
4497 iomsg->msg.msg_name = &iomsg->addr;
4499 #ifdef CONFIG_COMPAT
4500 if (req->ctx->compat)
4501 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4504 return __io_recvmsg_copy_hdr(req, iomsg);
4507 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4510 struct io_sr_msg *sr = &req->sr_msg;
4511 struct io_buffer *kbuf;
4513 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4518 req->flags |= REQ_F_BUFFER_SELECTED;
4522 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4524 return io_put_kbuf(req, req->sr_msg.kbuf);
4527 static int io_recvmsg_prep_async(struct io_kiocb *req)
4531 if (!io_op_defs[req->opcode].needs_async_data)
4533 ret = io_recvmsg_copy_hdr(req, req->async_data);
4535 req->flags |= REQ_F_NEED_CLEANUP;
4539 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4541 struct io_sr_msg *sr = &req->sr_msg;
4543 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4546 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4547 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4548 sr->len = READ_ONCE(sqe->len);
4549 sr->bgid = READ_ONCE(sqe->buf_group);
4551 #ifdef CONFIG_COMPAT
4552 if (req->ctx->compat)
4553 sr->msg_flags |= MSG_CMSG_COMPAT;
4558 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4560 struct io_async_msghdr iomsg, *kmsg;
4561 struct socket *sock;
4562 struct io_buffer *kbuf;
4564 int ret, cflags = 0;
4565 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4567 sock = sock_from_file(req->file);
4568 if (unlikely(!sock))
4571 kmsg = req->async_data;
4573 ret = io_recvmsg_copy_hdr(req, &iomsg);
4579 if (req->flags & REQ_F_BUFFER_SELECT) {
4580 kbuf = io_recv_buffer_select(req, !force_nonblock);
4582 return PTR_ERR(kbuf);
4583 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4584 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4585 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4586 1, req->sr_msg.len);
4589 flags = req->sr_msg.msg_flags;
4590 if (flags & MSG_DONTWAIT)
4591 req->flags |= REQ_F_NOWAIT;
4592 else if (force_nonblock)
4593 flags |= MSG_DONTWAIT;
4595 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4596 kmsg->uaddr, flags);
4597 if (force_nonblock && ret == -EAGAIN)
4598 return io_setup_async_msg(req, kmsg);
4599 if (ret == -ERESTARTSYS)
4602 if (req->flags & REQ_F_BUFFER_SELECTED)
4603 cflags = io_put_recv_kbuf(req);
4604 /* fast path, check for non-NULL to avoid function call */
4606 kfree(kmsg->free_iov);
4607 req->flags &= ~REQ_F_NEED_CLEANUP;
4609 req_set_fail_links(req);
4610 __io_req_complete(req, issue_flags, ret, cflags);
4614 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4616 struct io_buffer *kbuf;
4617 struct io_sr_msg *sr = &req->sr_msg;
4619 void __user *buf = sr->buf;
4620 struct socket *sock;
4623 int ret, cflags = 0;
4624 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4626 sock = sock_from_file(req->file);
4627 if (unlikely(!sock))
4630 if (req->flags & REQ_F_BUFFER_SELECT) {
4631 kbuf = io_recv_buffer_select(req, !force_nonblock);
4633 return PTR_ERR(kbuf);
4634 buf = u64_to_user_ptr(kbuf->addr);
4637 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4641 msg.msg_name = NULL;
4642 msg.msg_control = NULL;
4643 msg.msg_controllen = 0;
4644 msg.msg_namelen = 0;
4645 msg.msg_iocb = NULL;
4648 flags = req->sr_msg.msg_flags;
4649 if (flags & MSG_DONTWAIT)
4650 req->flags |= REQ_F_NOWAIT;
4651 else if (force_nonblock)
4652 flags |= MSG_DONTWAIT;
4654 ret = sock_recvmsg(sock, &msg, flags);
4655 if (force_nonblock && ret == -EAGAIN)
4657 if (ret == -ERESTARTSYS)
4660 if (req->flags & REQ_F_BUFFER_SELECTED)
4661 cflags = io_put_recv_kbuf(req);
4663 req_set_fail_links(req);
4664 __io_req_complete(req, issue_flags, ret, cflags);
4668 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4670 struct io_accept *accept = &req->accept;
4672 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4674 if (sqe->ioprio || sqe->len || sqe->buf_index)
4677 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4678 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4679 accept->flags = READ_ONCE(sqe->accept_flags);
4680 accept->nofile = rlimit(RLIMIT_NOFILE);
4684 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4686 struct io_accept *accept = &req->accept;
4687 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4688 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4691 if (req->file->f_flags & O_NONBLOCK)
4692 req->flags |= REQ_F_NOWAIT;
4694 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4695 accept->addr_len, accept->flags,
4697 if (ret == -EAGAIN && force_nonblock)
4700 if (ret == -ERESTARTSYS)
4702 req_set_fail_links(req);
4704 __io_req_complete(req, issue_flags, ret, 0);
4708 static int io_connect_prep_async(struct io_kiocb *req)
4710 struct io_async_connect *io = req->async_data;
4711 struct io_connect *conn = &req->connect;
4713 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4716 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4718 struct io_connect *conn = &req->connect;
4720 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4722 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4725 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4726 conn->addr_len = READ_ONCE(sqe->addr2);
4730 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4732 struct io_async_connect __io, *io;
4733 unsigned file_flags;
4735 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4737 if (req->async_data) {
4738 io = req->async_data;
4740 ret = move_addr_to_kernel(req->connect.addr,
4741 req->connect.addr_len,
4748 file_flags = force_nonblock ? O_NONBLOCK : 0;
4750 ret = __sys_connect_file(req->file, &io->address,
4751 req->connect.addr_len, file_flags);
4752 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4753 if (req->async_data)
4755 if (io_alloc_async_data(req)) {
4759 io = req->async_data;
4760 memcpy(req->async_data, &__io, sizeof(__io));
4763 if (ret == -ERESTARTSYS)
4767 req_set_fail_links(req);
4768 __io_req_complete(req, issue_flags, ret, 0);
4771 #else /* !CONFIG_NET */
4772 #define IO_NETOP_FN(op) \
4773 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4775 return -EOPNOTSUPP; \
4778 #define IO_NETOP_PREP(op) \
4780 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4782 return -EOPNOTSUPP; \
4785 #define IO_NETOP_PREP_ASYNC(op) \
4787 static int io_##op##_prep_async(struct io_kiocb *req) \
4789 return -EOPNOTSUPP; \
4792 IO_NETOP_PREP_ASYNC(sendmsg);
4793 IO_NETOP_PREP_ASYNC(recvmsg);
4794 IO_NETOP_PREP_ASYNC(connect);
4795 IO_NETOP_PREP(accept);
4798 #endif /* CONFIG_NET */
4800 struct io_poll_table {
4801 struct poll_table_struct pt;
4802 struct io_kiocb *req;
4806 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4807 __poll_t mask, task_work_func_t func)
4811 /* for instances that support it check for an event match first: */
4812 if (mask && !(mask & poll->events))
4815 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4817 list_del_init(&poll->wait.entry);
4820 req->task_work.func = func;
4821 percpu_ref_get(&req->ctx->refs);
4824 * If this fails, then the task is exiting. When a task exits, the
4825 * work gets canceled, so just cancel this request as well instead
4826 * of executing it. We can't safely execute it anyway, as we may not
4827 * have the needed state needed for it anyway.
4829 ret = io_req_task_work_add(req);
4830 if (unlikely(ret)) {
4831 WRITE_ONCE(poll->canceled, true);
4832 io_req_task_work_add_fallback(req, func);
4837 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4838 __acquires(&req->ctx->completion_lock)
4840 struct io_ring_ctx *ctx = req->ctx;
4842 if (!req->result && !READ_ONCE(poll->canceled)) {
4843 struct poll_table_struct pt = { ._key = poll->events };
4845 req->result = vfs_poll(req->file, &pt) & poll->events;
4848 spin_lock_irq(&ctx->completion_lock);
4849 if (!req->result && !READ_ONCE(poll->canceled)) {
4850 add_wait_queue(poll->head, &poll->wait);
4857 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4859 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4860 if (req->opcode == IORING_OP_POLL_ADD)
4861 return req->async_data;
4862 return req->apoll->double_poll;
4865 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4867 if (req->opcode == IORING_OP_POLL_ADD)
4869 return &req->apoll->poll;
4872 static void io_poll_remove_double(struct io_kiocb *req)
4874 struct io_poll_iocb *poll = io_poll_get_double(req);
4876 lockdep_assert_held(&req->ctx->completion_lock);
4878 if (poll && poll->head) {
4879 struct wait_queue_head *head = poll->head;
4881 spin_lock(&head->lock);
4882 list_del_init(&poll->wait.entry);
4883 if (poll->wait.private)
4884 refcount_dec(&req->refs);
4886 spin_unlock(&head->lock);
4890 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4892 struct io_ring_ctx *ctx = req->ctx;
4894 io_poll_remove_double(req);
4895 req->poll.done = true;
4896 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4897 io_commit_cqring(ctx);
4900 static void io_poll_task_func(struct callback_head *cb)
4902 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4903 struct io_ring_ctx *ctx = req->ctx;
4904 struct io_kiocb *nxt;
4906 if (io_poll_rewait(req, &req->poll)) {
4907 spin_unlock_irq(&ctx->completion_lock);
4909 hash_del(&req->hash_node);
4910 io_poll_complete(req, req->result, 0);
4911 spin_unlock_irq(&ctx->completion_lock);
4913 nxt = io_put_req_find_next(req);
4914 io_cqring_ev_posted(ctx);
4916 __io_req_task_submit(nxt);
4919 percpu_ref_put(&ctx->refs);
4922 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4923 int sync, void *key)
4925 struct io_kiocb *req = wait->private;
4926 struct io_poll_iocb *poll = io_poll_get_single(req);
4927 __poll_t mask = key_to_poll(key);
4929 /* for instances that support it check for an event match first: */
4930 if (mask && !(mask & poll->events))
4933 list_del_init(&wait->entry);
4935 if (poll && poll->head) {
4938 spin_lock(&poll->head->lock);
4939 done = list_empty(&poll->wait.entry);
4941 list_del_init(&poll->wait.entry);
4942 /* make sure double remove sees this as being gone */
4943 wait->private = NULL;
4944 spin_unlock(&poll->head->lock);
4946 /* use wait func handler, so it matches the rq type */
4947 poll->wait.func(&poll->wait, mode, sync, key);
4950 refcount_dec(&req->refs);
4954 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4955 wait_queue_func_t wake_func)
4959 poll->canceled = false;
4960 poll->events = events;
4961 INIT_LIST_HEAD(&poll->wait.entry);
4962 init_waitqueue_func_entry(&poll->wait, wake_func);
4965 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4966 struct wait_queue_head *head,
4967 struct io_poll_iocb **poll_ptr)
4969 struct io_kiocb *req = pt->req;
4972 * If poll->head is already set, it's because the file being polled
4973 * uses multiple waitqueues for poll handling (eg one for read, one
4974 * for write). Setup a separate io_poll_iocb if this happens.
4976 if (unlikely(poll->head)) {
4977 struct io_poll_iocb *poll_one = poll;
4979 /* already have a 2nd entry, fail a third attempt */
4981 pt->error = -EINVAL;
4984 /* double add on the same waitqueue head, ignore */
4985 if (poll->head == head)
4987 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4989 pt->error = -ENOMEM;
4992 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
4993 refcount_inc(&req->refs);
4994 poll->wait.private = req;
5001 if (poll->events & EPOLLEXCLUSIVE)
5002 add_wait_queue_exclusive(head, &poll->wait);
5004 add_wait_queue(head, &poll->wait);
5007 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5008 struct poll_table_struct *p)
5010 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5011 struct async_poll *apoll = pt->req->apoll;
5013 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5016 static void io_async_task_func(struct callback_head *cb)
5018 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5019 struct async_poll *apoll = req->apoll;
5020 struct io_ring_ctx *ctx = req->ctx;
5022 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5024 if (io_poll_rewait(req, &apoll->poll)) {
5025 spin_unlock_irq(&ctx->completion_lock);
5026 percpu_ref_put(&ctx->refs);
5030 /* If req is still hashed, it cannot have been canceled. Don't check. */
5031 if (hash_hashed(&req->hash_node))
5032 hash_del(&req->hash_node);
5034 io_poll_remove_double(req);
5035 spin_unlock_irq(&ctx->completion_lock);
5037 if (!READ_ONCE(apoll->poll.canceled))
5038 __io_req_task_submit(req);
5040 __io_req_task_cancel(req, -ECANCELED);
5042 percpu_ref_put(&ctx->refs);
5043 kfree(apoll->double_poll);
5047 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5050 struct io_kiocb *req = wait->private;
5051 struct io_poll_iocb *poll = &req->apoll->poll;
5053 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5056 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5059 static void io_poll_req_insert(struct io_kiocb *req)
5061 struct io_ring_ctx *ctx = req->ctx;
5062 struct hlist_head *list;
5064 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5065 hlist_add_head(&req->hash_node, list);
5068 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5069 struct io_poll_iocb *poll,
5070 struct io_poll_table *ipt, __poll_t mask,
5071 wait_queue_func_t wake_func)
5072 __acquires(&ctx->completion_lock)
5074 struct io_ring_ctx *ctx = req->ctx;
5075 bool cancel = false;
5077 INIT_HLIST_NODE(&req->hash_node);
5078 io_init_poll_iocb(poll, mask, wake_func);
5079 poll->file = req->file;
5080 poll->wait.private = req;
5082 ipt->pt._key = mask;
5084 ipt->error = -EINVAL;
5086 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5088 spin_lock_irq(&ctx->completion_lock);
5089 if (likely(poll->head)) {
5090 spin_lock(&poll->head->lock);
5091 if (unlikely(list_empty(&poll->wait.entry))) {
5097 if (mask || ipt->error)
5098 list_del_init(&poll->wait.entry);
5100 WRITE_ONCE(poll->canceled, true);
5101 else if (!poll->done) /* actually waiting for an event */
5102 io_poll_req_insert(req);
5103 spin_unlock(&poll->head->lock);
5109 static bool io_arm_poll_handler(struct io_kiocb *req)
5111 const struct io_op_def *def = &io_op_defs[req->opcode];
5112 struct io_ring_ctx *ctx = req->ctx;
5113 struct async_poll *apoll;
5114 struct io_poll_table ipt;
5118 if (!req->file || !file_can_poll(req->file))
5120 if (req->flags & REQ_F_POLLED)
5124 else if (def->pollout)
5128 /* if we can't nonblock try, then no point in arming a poll handler */
5129 if (!io_file_supports_async(req->file, rw))
5132 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5133 if (unlikely(!apoll))
5135 apoll->double_poll = NULL;
5137 req->flags |= REQ_F_POLLED;
5142 mask |= POLLIN | POLLRDNORM;
5144 mask |= POLLOUT | POLLWRNORM;
5146 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5147 if ((req->opcode == IORING_OP_RECVMSG) &&
5148 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5151 mask |= POLLERR | POLLPRI;
5153 ipt.pt._qproc = io_async_queue_proc;
5155 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5157 if (ret || ipt.error) {
5158 io_poll_remove_double(req);
5159 spin_unlock_irq(&ctx->completion_lock);
5160 kfree(apoll->double_poll);
5164 spin_unlock_irq(&ctx->completion_lock);
5165 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5166 apoll->poll.events);
5170 static bool __io_poll_remove_one(struct io_kiocb *req,
5171 struct io_poll_iocb *poll)
5173 bool do_complete = false;
5175 spin_lock(&poll->head->lock);
5176 WRITE_ONCE(poll->canceled, true);
5177 if (!list_empty(&poll->wait.entry)) {
5178 list_del_init(&poll->wait.entry);
5181 spin_unlock(&poll->head->lock);
5182 hash_del(&req->hash_node);
5186 static bool io_poll_remove_one(struct io_kiocb *req)
5190 io_poll_remove_double(req);
5192 if (req->opcode == IORING_OP_POLL_ADD) {
5193 do_complete = __io_poll_remove_one(req, &req->poll);
5195 struct async_poll *apoll = req->apoll;
5197 /* non-poll requests have submit ref still */
5198 do_complete = __io_poll_remove_one(req, &apoll->poll);
5201 kfree(apoll->double_poll);
5207 io_cqring_fill_event(req, -ECANCELED);
5208 io_commit_cqring(req->ctx);
5209 req_set_fail_links(req);
5210 io_put_req_deferred(req, 1);
5217 * Returns true if we found and killed one or more poll requests
5219 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5220 struct files_struct *files)
5222 struct hlist_node *tmp;
5223 struct io_kiocb *req;
5226 spin_lock_irq(&ctx->completion_lock);
5227 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5228 struct hlist_head *list;
5230 list = &ctx->cancel_hash[i];
5231 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5232 if (io_match_task(req, tsk, files))
5233 posted += io_poll_remove_one(req);
5236 spin_unlock_irq(&ctx->completion_lock);
5239 io_cqring_ev_posted(ctx);
5244 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5246 struct hlist_head *list;
5247 struct io_kiocb *req;
5249 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5250 hlist_for_each_entry(req, list, hash_node) {
5251 if (sqe_addr != req->user_data)
5253 if (io_poll_remove_one(req))
5261 static int io_poll_remove_prep(struct io_kiocb *req,
5262 const struct io_uring_sqe *sqe)
5264 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5266 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5270 req->poll_remove.addr = READ_ONCE(sqe->addr);
5275 * Find a running poll command that matches one specified in sqe->addr,
5276 * and remove it if found.
5278 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5280 struct io_ring_ctx *ctx = req->ctx;
5283 spin_lock_irq(&ctx->completion_lock);
5284 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5285 spin_unlock_irq(&ctx->completion_lock);
5288 req_set_fail_links(req);
5289 io_req_complete(req, ret);
5293 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5296 struct io_kiocb *req = wait->private;
5297 struct io_poll_iocb *poll = &req->poll;
5299 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5302 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5303 struct poll_table_struct *p)
5305 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5307 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5310 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5312 struct io_poll_iocb *poll = &req->poll;
5315 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5317 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5320 events = READ_ONCE(sqe->poll32_events);
5322 events = swahw32(events);
5324 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5325 (events & EPOLLEXCLUSIVE);
5329 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5331 struct io_poll_iocb *poll = &req->poll;
5332 struct io_ring_ctx *ctx = req->ctx;
5333 struct io_poll_table ipt;
5336 ipt.pt._qproc = io_poll_queue_proc;
5338 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5341 if (mask) { /* no async, we'd stolen it */
5343 io_poll_complete(req, mask, 0);
5345 spin_unlock_irq(&ctx->completion_lock);
5348 io_cqring_ev_posted(ctx);
5354 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5356 struct io_timeout_data *data = container_of(timer,
5357 struct io_timeout_data, timer);
5358 struct io_kiocb *req = data->req;
5359 struct io_ring_ctx *ctx = req->ctx;
5360 unsigned long flags;
5362 spin_lock_irqsave(&ctx->completion_lock, flags);
5363 list_del_init(&req->timeout.list);
5364 atomic_set(&req->ctx->cq_timeouts,
5365 atomic_read(&req->ctx->cq_timeouts) + 1);
5367 io_cqring_fill_event(req, -ETIME);
5368 io_commit_cqring(ctx);
5369 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5371 io_cqring_ev_posted(ctx);
5372 req_set_fail_links(req);
5374 return HRTIMER_NORESTART;
5377 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5380 struct io_timeout_data *io;
5381 struct io_kiocb *req;
5384 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5385 if (user_data == req->user_data) {
5392 return ERR_PTR(ret);
5394 io = req->async_data;
5395 ret = hrtimer_try_to_cancel(&io->timer);
5397 return ERR_PTR(-EALREADY);
5398 list_del_init(&req->timeout.list);
5402 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5404 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5407 return PTR_ERR(req);
5409 req_set_fail_links(req);
5410 io_cqring_fill_event(req, -ECANCELED);
5411 io_put_req_deferred(req, 1);
5415 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5416 struct timespec64 *ts, enum hrtimer_mode mode)
5418 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5419 struct io_timeout_data *data;
5422 return PTR_ERR(req);
5424 req->timeout.off = 0; /* noseq */
5425 data = req->async_data;
5426 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5427 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5428 data->timer.function = io_timeout_fn;
5429 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5433 static int io_timeout_remove_prep(struct io_kiocb *req,
5434 const struct io_uring_sqe *sqe)
5436 struct io_timeout_rem *tr = &req->timeout_rem;
5438 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5440 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5442 if (sqe->ioprio || sqe->buf_index || sqe->len)
5445 tr->addr = READ_ONCE(sqe->addr);
5446 tr->flags = READ_ONCE(sqe->timeout_flags);
5447 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5448 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5450 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5452 } else if (tr->flags) {
5453 /* timeout removal doesn't support flags */
5460 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5462 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5467 * Remove or update an existing timeout command
5469 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5471 struct io_timeout_rem *tr = &req->timeout_rem;
5472 struct io_ring_ctx *ctx = req->ctx;
5475 spin_lock_irq(&ctx->completion_lock);
5476 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5477 ret = io_timeout_cancel(ctx, tr->addr);
5479 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5480 io_translate_timeout_mode(tr->flags));
5482 io_cqring_fill_event(req, ret);
5483 io_commit_cqring(ctx);
5484 spin_unlock_irq(&ctx->completion_lock);
5485 io_cqring_ev_posted(ctx);
5487 req_set_fail_links(req);
5492 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5493 bool is_timeout_link)
5495 struct io_timeout_data *data;
5497 u32 off = READ_ONCE(sqe->off);
5499 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5501 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5503 if (off && is_timeout_link)
5505 flags = READ_ONCE(sqe->timeout_flags);
5506 if (flags & ~IORING_TIMEOUT_ABS)
5509 req->timeout.off = off;
5511 if (!req->async_data && io_alloc_async_data(req))
5514 data = req->async_data;
5517 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5520 data->mode = io_translate_timeout_mode(flags);
5521 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5522 io_req_track_inflight(req);
5526 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5528 struct io_ring_ctx *ctx = req->ctx;
5529 struct io_timeout_data *data = req->async_data;
5530 struct list_head *entry;
5531 u32 tail, off = req->timeout.off;
5533 spin_lock_irq(&ctx->completion_lock);
5536 * sqe->off holds how many events that need to occur for this
5537 * timeout event to be satisfied. If it isn't set, then this is
5538 * a pure timeout request, sequence isn't used.
5540 if (io_is_timeout_noseq(req)) {
5541 entry = ctx->timeout_list.prev;
5545 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5546 req->timeout.target_seq = tail + off;
5548 /* Update the last seq here in case io_flush_timeouts() hasn't.
5549 * This is safe because ->completion_lock is held, and submissions
5550 * and completions are never mixed in the same ->completion_lock section.
5552 ctx->cq_last_tm_flush = tail;
5555 * Insertion sort, ensuring the first entry in the list is always
5556 * the one we need first.
5558 list_for_each_prev(entry, &ctx->timeout_list) {
5559 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5562 if (io_is_timeout_noseq(nxt))
5564 /* nxt.seq is behind @tail, otherwise would've been completed */
5565 if (off >= nxt->timeout.target_seq - tail)
5569 list_add(&req->timeout.list, entry);
5570 data->timer.function = io_timeout_fn;
5571 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5572 spin_unlock_irq(&ctx->completion_lock);
5576 struct io_cancel_data {
5577 struct io_ring_ctx *ctx;
5581 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5583 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5584 struct io_cancel_data *cd = data;
5586 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5589 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5590 struct io_ring_ctx *ctx)
5592 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5593 enum io_wq_cancel cancel_ret;
5596 if (!tctx || !tctx->io_wq)
5599 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5600 switch (cancel_ret) {
5601 case IO_WQ_CANCEL_OK:
5604 case IO_WQ_CANCEL_RUNNING:
5607 case IO_WQ_CANCEL_NOTFOUND:
5615 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5616 struct io_kiocb *req, __u64 sqe_addr,
5619 unsigned long flags;
5622 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5623 if (ret != -ENOENT) {
5624 spin_lock_irqsave(&ctx->completion_lock, flags);
5628 spin_lock_irqsave(&ctx->completion_lock, flags);
5629 ret = io_timeout_cancel(ctx, sqe_addr);
5632 ret = io_poll_cancel(ctx, sqe_addr);
5636 io_cqring_fill_event(req, ret);
5637 io_commit_cqring(ctx);
5638 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5639 io_cqring_ev_posted(ctx);
5642 req_set_fail_links(req);
5646 static int io_async_cancel_prep(struct io_kiocb *req,
5647 const struct io_uring_sqe *sqe)
5649 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5651 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5653 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5656 req->cancel.addr = READ_ONCE(sqe->addr);
5660 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5662 struct io_ring_ctx *ctx = req->ctx;
5663 u64 sqe_addr = req->cancel.addr;
5664 struct io_tctx_node *node;
5667 /* tasks should wait for their io-wq threads, so safe w/o sync */
5668 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5669 spin_lock_irq(&ctx->completion_lock);
5672 ret = io_timeout_cancel(ctx, sqe_addr);
5675 ret = io_poll_cancel(ctx, sqe_addr);
5678 spin_unlock_irq(&ctx->completion_lock);
5680 /* slow path, try all io-wq's */
5681 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5683 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5684 struct io_uring_task *tctx = node->task->io_uring;
5686 if (!tctx || !tctx->io_wq)
5688 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5692 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5694 spin_lock_irq(&ctx->completion_lock);
5696 io_cqring_fill_event(req, ret);
5697 io_commit_cqring(ctx);
5698 spin_unlock_irq(&ctx->completion_lock);
5699 io_cqring_ev_posted(ctx);
5702 req_set_fail_links(req);
5707 static int io_rsrc_update_prep(struct io_kiocb *req,
5708 const struct io_uring_sqe *sqe)
5710 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5712 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5714 if (sqe->ioprio || sqe->rw_flags)
5717 req->rsrc_update.offset = READ_ONCE(sqe->off);
5718 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5719 if (!req->rsrc_update.nr_args)
5721 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5725 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5727 struct io_ring_ctx *ctx = req->ctx;
5728 struct io_uring_rsrc_update up;
5731 if (issue_flags & IO_URING_F_NONBLOCK)
5734 up.offset = req->rsrc_update.offset;
5735 up.data = req->rsrc_update.arg;
5737 mutex_lock(&ctx->uring_lock);
5738 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5739 mutex_unlock(&ctx->uring_lock);
5742 req_set_fail_links(req);
5743 __io_req_complete(req, issue_flags, ret, 0);
5747 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5749 switch (req->opcode) {
5752 case IORING_OP_READV:
5753 case IORING_OP_READ_FIXED:
5754 case IORING_OP_READ:
5755 return io_read_prep(req, sqe);
5756 case IORING_OP_WRITEV:
5757 case IORING_OP_WRITE_FIXED:
5758 case IORING_OP_WRITE:
5759 return io_write_prep(req, sqe);
5760 case IORING_OP_POLL_ADD:
5761 return io_poll_add_prep(req, sqe);
5762 case IORING_OP_POLL_REMOVE:
5763 return io_poll_remove_prep(req, sqe);
5764 case IORING_OP_FSYNC:
5765 return io_fsync_prep(req, sqe);
5766 case IORING_OP_SYNC_FILE_RANGE:
5767 return io_sfr_prep(req, sqe);
5768 case IORING_OP_SENDMSG:
5769 case IORING_OP_SEND:
5770 return io_sendmsg_prep(req, sqe);
5771 case IORING_OP_RECVMSG:
5772 case IORING_OP_RECV:
5773 return io_recvmsg_prep(req, sqe);
5774 case IORING_OP_CONNECT:
5775 return io_connect_prep(req, sqe);
5776 case IORING_OP_TIMEOUT:
5777 return io_timeout_prep(req, sqe, false);
5778 case IORING_OP_TIMEOUT_REMOVE:
5779 return io_timeout_remove_prep(req, sqe);
5780 case IORING_OP_ASYNC_CANCEL:
5781 return io_async_cancel_prep(req, sqe);
5782 case IORING_OP_LINK_TIMEOUT:
5783 return io_timeout_prep(req, sqe, true);
5784 case IORING_OP_ACCEPT:
5785 return io_accept_prep(req, sqe);
5786 case IORING_OP_FALLOCATE:
5787 return io_fallocate_prep(req, sqe);
5788 case IORING_OP_OPENAT:
5789 return io_openat_prep(req, sqe);
5790 case IORING_OP_CLOSE:
5791 return io_close_prep(req, sqe);
5792 case IORING_OP_FILES_UPDATE:
5793 return io_rsrc_update_prep(req, sqe);
5794 case IORING_OP_STATX:
5795 return io_statx_prep(req, sqe);
5796 case IORING_OP_FADVISE:
5797 return io_fadvise_prep(req, sqe);
5798 case IORING_OP_MADVISE:
5799 return io_madvise_prep(req, sqe);
5800 case IORING_OP_OPENAT2:
5801 return io_openat2_prep(req, sqe);
5802 case IORING_OP_EPOLL_CTL:
5803 return io_epoll_ctl_prep(req, sqe);
5804 case IORING_OP_SPLICE:
5805 return io_splice_prep(req, sqe);
5806 case IORING_OP_PROVIDE_BUFFERS:
5807 return io_provide_buffers_prep(req, sqe);
5808 case IORING_OP_REMOVE_BUFFERS:
5809 return io_remove_buffers_prep(req, sqe);
5811 return io_tee_prep(req, sqe);
5812 case IORING_OP_SHUTDOWN:
5813 return io_shutdown_prep(req, sqe);
5814 case IORING_OP_RENAMEAT:
5815 return io_renameat_prep(req, sqe);
5816 case IORING_OP_UNLINKAT:
5817 return io_unlinkat_prep(req, sqe);
5820 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5825 static int io_req_prep_async(struct io_kiocb *req)
5827 switch (req->opcode) {
5828 case IORING_OP_READV:
5829 case IORING_OP_READ_FIXED:
5830 case IORING_OP_READ:
5831 return io_rw_prep_async(req, READ);
5832 case IORING_OP_WRITEV:
5833 case IORING_OP_WRITE_FIXED:
5834 case IORING_OP_WRITE:
5835 return io_rw_prep_async(req, WRITE);
5836 case IORING_OP_SENDMSG:
5837 case IORING_OP_SEND:
5838 return io_sendmsg_prep_async(req);
5839 case IORING_OP_RECVMSG:
5840 case IORING_OP_RECV:
5841 return io_recvmsg_prep_async(req);
5842 case IORING_OP_CONNECT:
5843 return io_connect_prep_async(req);
5848 static int io_req_defer_prep(struct io_kiocb *req)
5850 if (!io_op_defs[req->opcode].needs_async_data)
5852 /* some opcodes init it during the inital prep */
5853 if (req->async_data)
5855 if (__io_alloc_async_data(req))
5857 return io_req_prep_async(req);
5860 static u32 io_get_sequence(struct io_kiocb *req)
5862 struct io_kiocb *pos;
5863 struct io_ring_ctx *ctx = req->ctx;
5864 u32 total_submitted, nr_reqs = 0;
5866 io_for_each_link(pos, req)
5869 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5870 return total_submitted - nr_reqs;
5873 static int io_req_defer(struct io_kiocb *req)
5875 struct io_ring_ctx *ctx = req->ctx;
5876 struct io_defer_entry *de;
5880 /* Still need defer if there is pending req in defer list. */
5881 if (likely(list_empty_careful(&ctx->defer_list) &&
5882 !(req->flags & REQ_F_IO_DRAIN)))
5885 seq = io_get_sequence(req);
5886 /* Still a chance to pass the sequence check */
5887 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5890 ret = io_req_defer_prep(req);
5893 io_prep_async_link(req);
5894 de = kmalloc(sizeof(*de), GFP_KERNEL);
5898 spin_lock_irq(&ctx->completion_lock);
5899 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5900 spin_unlock_irq(&ctx->completion_lock);
5902 io_queue_async_work(req);
5903 return -EIOCBQUEUED;
5906 trace_io_uring_defer(ctx, req, req->user_data);
5909 list_add_tail(&de->list, &ctx->defer_list);
5910 spin_unlock_irq(&ctx->completion_lock);
5911 return -EIOCBQUEUED;
5914 static void __io_clean_op(struct io_kiocb *req)
5916 if (req->flags & REQ_F_BUFFER_SELECTED) {
5917 switch (req->opcode) {
5918 case IORING_OP_READV:
5919 case IORING_OP_READ_FIXED:
5920 case IORING_OP_READ:
5921 kfree((void *)(unsigned long)req->rw.addr);
5923 case IORING_OP_RECVMSG:
5924 case IORING_OP_RECV:
5925 kfree(req->sr_msg.kbuf);
5928 req->flags &= ~REQ_F_BUFFER_SELECTED;
5931 if (req->flags & REQ_F_NEED_CLEANUP) {
5932 switch (req->opcode) {
5933 case IORING_OP_READV:
5934 case IORING_OP_READ_FIXED:
5935 case IORING_OP_READ:
5936 case IORING_OP_WRITEV:
5937 case IORING_OP_WRITE_FIXED:
5938 case IORING_OP_WRITE: {
5939 struct io_async_rw *io = req->async_data;
5941 kfree(io->free_iovec);
5944 case IORING_OP_RECVMSG:
5945 case IORING_OP_SENDMSG: {
5946 struct io_async_msghdr *io = req->async_data;
5948 kfree(io->free_iov);
5951 case IORING_OP_SPLICE:
5953 io_put_file(req, req->splice.file_in,
5954 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5956 case IORING_OP_OPENAT:
5957 case IORING_OP_OPENAT2:
5958 if (req->open.filename)
5959 putname(req->open.filename);
5961 case IORING_OP_RENAMEAT:
5962 putname(req->rename.oldpath);
5963 putname(req->rename.newpath);
5965 case IORING_OP_UNLINKAT:
5966 putname(req->unlink.filename);
5969 req->flags &= ~REQ_F_NEED_CLEANUP;
5973 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5975 struct io_ring_ctx *ctx = req->ctx;
5976 const struct cred *creds = NULL;
5979 if (req->work.creds && req->work.creds != current_cred())
5980 creds = override_creds(req->work.creds);
5982 switch (req->opcode) {
5984 ret = io_nop(req, issue_flags);
5986 case IORING_OP_READV:
5987 case IORING_OP_READ_FIXED:
5988 case IORING_OP_READ:
5989 ret = io_read(req, issue_flags);
5991 case IORING_OP_WRITEV:
5992 case IORING_OP_WRITE_FIXED:
5993 case IORING_OP_WRITE:
5994 ret = io_write(req, issue_flags);
5996 case IORING_OP_FSYNC:
5997 ret = io_fsync(req, issue_flags);
5999 case IORING_OP_POLL_ADD:
6000 ret = io_poll_add(req, issue_flags);
6002 case IORING_OP_POLL_REMOVE:
6003 ret = io_poll_remove(req, issue_flags);
6005 case IORING_OP_SYNC_FILE_RANGE:
6006 ret = io_sync_file_range(req, issue_flags);
6008 case IORING_OP_SENDMSG:
6009 ret = io_sendmsg(req, issue_flags);
6011 case IORING_OP_SEND:
6012 ret = io_send(req, issue_flags);
6014 case IORING_OP_RECVMSG:
6015 ret = io_recvmsg(req, issue_flags);
6017 case IORING_OP_RECV:
6018 ret = io_recv(req, issue_flags);
6020 case IORING_OP_TIMEOUT:
6021 ret = io_timeout(req, issue_flags);
6023 case IORING_OP_TIMEOUT_REMOVE:
6024 ret = io_timeout_remove(req, issue_flags);
6026 case IORING_OP_ACCEPT:
6027 ret = io_accept(req, issue_flags);
6029 case IORING_OP_CONNECT:
6030 ret = io_connect(req, issue_flags);
6032 case IORING_OP_ASYNC_CANCEL:
6033 ret = io_async_cancel(req, issue_flags);
6035 case IORING_OP_FALLOCATE:
6036 ret = io_fallocate(req, issue_flags);
6038 case IORING_OP_OPENAT:
6039 ret = io_openat(req, issue_flags);
6041 case IORING_OP_CLOSE:
6042 ret = io_close(req, issue_flags);
6044 case IORING_OP_FILES_UPDATE:
6045 ret = io_files_update(req, issue_flags);
6047 case IORING_OP_STATX:
6048 ret = io_statx(req, issue_flags);
6050 case IORING_OP_FADVISE:
6051 ret = io_fadvise(req, issue_flags);
6053 case IORING_OP_MADVISE:
6054 ret = io_madvise(req, issue_flags);
6056 case IORING_OP_OPENAT2:
6057 ret = io_openat2(req, issue_flags);
6059 case IORING_OP_EPOLL_CTL:
6060 ret = io_epoll_ctl(req, issue_flags);
6062 case IORING_OP_SPLICE:
6063 ret = io_splice(req, issue_flags);
6065 case IORING_OP_PROVIDE_BUFFERS:
6066 ret = io_provide_buffers(req, issue_flags);
6068 case IORING_OP_REMOVE_BUFFERS:
6069 ret = io_remove_buffers(req, issue_flags);
6072 ret = io_tee(req, issue_flags);
6074 case IORING_OP_SHUTDOWN:
6075 ret = io_shutdown(req, issue_flags);
6077 case IORING_OP_RENAMEAT:
6078 ret = io_renameat(req, issue_flags);
6080 case IORING_OP_UNLINKAT:
6081 ret = io_unlinkat(req, issue_flags);
6089 revert_creds(creds);
6094 /* If the op doesn't have a file, we're not polling for it */
6095 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6096 const bool in_async = io_wq_current_is_worker();
6098 /* workqueue context doesn't hold uring_lock, grab it now */
6100 mutex_lock(&ctx->uring_lock);
6102 io_iopoll_req_issued(req, in_async);
6105 mutex_unlock(&ctx->uring_lock);
6111 static void io_wq_submit_work(struct io_wq_work *work)
6113 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6114 struct io_kiocb *timeout;
6117 timeout = io_prep_linked_timeout(req);
6119 io_queue_linked_timeout(timeout);
6121 if (work->flags & IO_WQ_WORK_CANCEL)
6126 ret = io_issue_sqe(req, 0);
6128 * We can get EAGAIN for polled IO even though we're
6129 * forcing a sync submission from here, since we can't
6130 * wait for request slots on the block side.
6138 /* avoid locking problems by failing it from a clean context */
6140 /* io-wq is going to take one down */
6141 refcount_inc(&req->refs);
6142 io_req_task_queue_fail(req, ret);
6146 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6149 struct fixed_rsrc_table *table;
6151 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6152 return table->files[index & IORING_FILE_TABLE_MASK];
6155 static struct file *io_file_get(struct io_submit_state *state,
6156 struct io_kiocb *req, int fd, bool fixed)
6158 struct io_ring_ctx *ctx = req->ctx;
6162 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6164 fd = array_index_nospec(fd, ctx->nr_user_files);
6165 file = io_file_from_index(ctx, fd);
6166 io_set_resource_node(req);
6168 trace_io_uring_file_get(ctx, fd);
6169 file = __io_file_get(state, fd);
6172 if (file && unlikely(file->f_op == &io_uring_fops))
6173 io_req_track_inflight(req);
6177 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6179 struct io_timeout_data *data = container_of(timer,
6180 struct io_timeout_data, timer);
6181 struct io_kiocb *prev, *req = data->req;
6182 struct io_ring_ctx *ctx = req->ctx;
6183 unsigned long flags;
6185 spin_lock_irqsave(&ctx->completion_lock, flags);
6186 prev = req->timeout.head;
6187 req->timeout.head = NULL;
6190 * We don't expect the list to be empty, that will only happen if we
6191 * race with the completion of the linked work.
6193 if (prev && refcount_inc_not_zero(&prev->refs))
6194 io_remove_next_linked(prev);
6197 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6200 req_set_fail_links(prev);
6201 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6202 io_put_req_deferred(prev, 1);
6204 io_req_complete_post(req, -ETIME, 0);
6205 io_put_req_deferred(req, 1);
6207 return HRTIMER_NORESTART;
6210 static void __io_queue_linked_timeout(struct io_kiocb *req)
6213 * If the back reference is NULL, then our linked request finished
6214 * before we got a chance to setup the timer
6216 if (req->timeout.head) {
6217 struct io_timeout_data *data = req->async_data;
6219 data->timer.function = io_link_timeout_fn;
6220 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6225 static void io_queue_linked_timeout(struct io_kiocb *req)
6227 struct io_ring_ctx *ctx = req->ctx;
6229 spin_lock_irq(&ctx->completion_lock);
6230 __io_queue_linked_timeout(req);
6231 spin_unlock_irq(&ctx->completion_lock);
6233 /* drop submission reference */
6237 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6239 struct io_kiocb *nxt = req->link;
6241 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6242 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6245 nxt->timeout.head = req;
6246 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6247 req->flags |= REQ_F_LINK_TIMEOUT;
6251 static void __io_queue_sqe(struct io_kiocb *req)
6253 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6256 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6259 * We async punt it if the file wasn't marked NOWAIT, or if the file
6260 * doesn't support non-blocking read/write attempts
6262 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6263 if (!io_arm_poll_handler(req)) {
6265 * Queued up for async execution, worker will release
6266 * submit reference when the iocb is actually submitted.
6268 io_queue_async_work(req);
6270 } else if (likely(!ret)) {
6271 /* drop submission reference */
6272 if (req->flags & REQ_F_COMPLETE_INLINE) {
6273 struct io_ring_ctx *ctx = req->ctx;
6274 struct io_comp_state *cs = &ctx->submit_state.comp;
6276 cs->reqs[cs->nr++] = req;
6277 if (cs->nr == ARRAY_SIZE(cs->reqs))
6278 io_submit_flush_completions(cs, ctx);
6283 req_set_fail_links(req);
6285 io_req_complete(req, ret);
6288 io_queue_linked_timeout(linked_timeout);
6291 static void io_queue_sqe(struct io_kiocb *req)
6295 ret = io_req_defer(req);
6297 if (ret != -EIOCBQUEUED) {
6299 req_set_fail_links(req);
6301 io_req_complete(req, ret);
6303 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6304 ret = io_req_defer_prep(req);
6307 io_queue_async_work(req);
6309 __io_queue_sqe(req);
6314 * Check SQE restrictions (opcode and flags).
6316 * Returns 'true' if SQE is allowed, 'false' otherwise.
6318 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6319 struct io_kiocb *req,
6320 unsigned int sqe_flags)
6322 if (!ctx->restricted)
6325 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6328 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6329 ctx->restrictions.sqe_flags_required)
6332 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6333 ctx->restrictions.sqe_flags_required))
6339 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6340 const struct io_uring_sqe *sqe)
6342 struct io_submit_state *state;
6343 unsigned int sqe_flags;
6344 int personality, ret = 0;
6346 req->opcode = READ_ONCE(sqe->opcode);
6347 /* same numerical values with corresponding REQ_F_*, safe to copy */
6348 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6349 req->user_data = READ_ONCE(sqe->user_data);
6350 req->async_data = NULL;
6354 req->fixed_rsrc_refs = NULL;
6355 /* one is dropped after submission, the other at completion */
6356 refcount_set(&req->refs, 2);
6357 req->task = current;
6359 req->work.list.next = NULL;
6360 req->work.creds = NULL;
6361 req->work.flags = 0;
6363 /* enforce forwards compatibility on users */
6364 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6369 if (unlikely(req->opcode >= IORING_OP_LAST))
6372 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6375 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6376 !io_op_defs[req->opcode].buffer_select)
6379 personality = READ_ONCE(sqe->personality);
6381 req->work.creds = xa_load(&ctx->personalities, personality);
6382 if (!req->work.creds)
6384 get_cred(req->work.creds);
6386 state = &ctx->submit_state;
6389 * Plug now if we have more than 1 IO left after this, and the target
6390 * is potentially a read/write to block based storage.
6392 if (!state->plug_started && state->ios_left > 1 &&
6393 io_op_defs[req->opcode].plug) {
6394 blk_start_plug(&state->plug);
6395 state->plug_started = true;
6398 if (io_op_defs[req->opcode].needs_file) {
6399 bool fixed = req->flags & REQ_F_FIXED_FILE;
6401 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6402 if (unlikely(!req->file))
6410 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6411 const struct io_uring_sqe *sqe)
6413 struct io_submit_link *link = &ctx->submit_state.link;
6416 ret = io_init_req(ctx, req, sqe);
6417 if (unlikely(ret)) {
6420 io_req_complete(req, ret);
6422 /* fail even hard links since we don't submit */
6423 link->head->flags |= REQ_F_FAIL_LINK;
6424 io_put_req(link->head);
6425 io_req_complete(link->head, -ECANCELED);
6430 ret = io_req_prep(req, sqe);
6434 /* don't need @sqe from now on */
6435 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6436 true, ctx->flags & IORING_SETUP_SQPOLL);
6439 * If we already have a head request, queue this one for async
6440 * submittal once the head completes. If we don't have a head but
6441 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6442 * submitted sync once the chain is complete. If none of those
6443 * conditions are true (normal request), then just queue it.
6446 struct io_kiocb *head = link->head;
6449 * Taking sequential execution of a link, draining both sides
6450 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6451 * requests in the link. So, it drains the head and the
6452 * next after the link request. The last one is done via
6453 * drain_next flag to persist the effect across calls.
6455 if (req->flags & REQ_F_IO_DRAIN) {
6456 head->flags |= REQ_F_IO_DRAIN;
6457 ctx->drain_next = 1;
6459 ret = io_req_defer_prep(req);
6462 trace_io_uring_link(ctx, req, head);
6463 link->last->link = req;
6466 /* last request of a link, enqueue the link */
6467 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6472 if (unlikely(ctx->drain_next)) {
6473 req->flags |= REQ_F_IO_DRAIN;
6474 ctx->drain_next = 0;
6476 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6488 * Batched submission is done, ensure local IO is flushed out.
6490 static void io_submit_state_end(struct io_submit_state *state,
6491 struct io_ring_ctx *ctx)
6493 if (state->link.head)
6494 io_queue_sqe(state->link.head);
6496 io_submit_flush_completions(&state->comp, ctx);
6497 if (state->plug_started)
6498 blk_finish_plug(&state->plug);
6499 io_state_file_put(state);
6503 * Start submission side cache.
6505 static void io_submit_state_start(struct io_submit_state *state,
6506 unsigned int max_ios)
6508 state->plug_started = false;
6509 state->ios_left = max_ios;
6510 /* set only head, no need to init link_last in advance */
6511 state->link.head = NULL;
6514 static void io_commit_sqring(struct io_ring_ctx *ctx)
6516 struct io_rings *rings = ctx->rings;
6519 * Ensure any loads from the SQEs are done at this point,
6520 * since once we write the new head, the application could
6521 * write new data to them.
6523 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6527 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6528 * that is mapped by userspace. This means that care needs to be taken to
6529 * ensure that reads are stable, as we cannot rely on userspace always
6530 * being a good citizen. If members of the sqe are validated and then later
6531 * used, it's important that those reads are done through READ_ONCE() to
6532 * prevent a re-load down the line.
6534 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6536 u32 *sq_array = ctx->sq_array;
6540 * The cached sq head (or cq tail) serves two purposes:
6542 * 1) allows us to batch the cost of updating the user visible
6544 * 2) allows the kernel side to track the head on its own, even
6545 * though the application is the one updating it.
6547 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6548 if (likely(head < ctx->sq_entries))
6549 return &ctx->sq_sqes[head];
6551 /* drop invalid entries */
6552 ctx->cached_sq_dropped++;
6553 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6557 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6561 /* if we have a backlog and couldn't flush it all, return BUSY */
6562 if (test_bit(0, &ctx->sq_check_overflow)) {
6563 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6567 /* make sure SQ entry isn't read before tail */
6568 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6570 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6573 percpu_counter_add(¤t->io_uring->inflight, nr);
6574 refcount_add(nr, ¤t->usage);
6575 io_submit_state_start(&ctx->submit_state, nr);
6577 while (submitted < nr) {
6578 const struct io_uring_sqe *sqe;
6579 struct io_kiocb *req;
6581 req = io_alloc_req(ctx);
6582 if (unlikely(!req)) {
6584 submitted = -EAGAIN;
6587 sqe = io_get_sqe(ctx);
6588 if (unlikely(!sqe)) {
6589 kmem_cache_free(req_cachep, req);
6592 /* will complete beyond this point, count as submitted */
6594 if (io_submit_sqe(ctx, req, sqe))
6598 if (unlikely(submitted != nr)) {
6599 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6600 struct io_uring_task *tctx = current->io_uring;
6601 int unused = nr - ref_used;
6603 percpu_ref_put_many(&ctx->refs, unused);
6604 percpu_counter_sub(&tctx->inflight, unused);
6605 put_task_struct_many(current, unused);
6608 io_submit_state_end(&ctx->submit_state, ctx);
6609 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6610 io_commit_sqring(ctx);
6615 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6617 /* Tell userspace we may need a wakeup call */
6618 spin_lock_irq(&ctx->completion_lock);
6619 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6620 spin_unlock_irq(&ctx->completion_lock);
6623 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6625 spin_lock_irq(&ctx->completion_lock);
6626 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6627 spin_unlock_irq(&ctx->completion_lock);
6630 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6632 unsigned int to_submit;
6635 to_submit = io_sqring_entries(ctx);
6636 /* if we're handling multiple rings, cap submit size for fairness */
6637 if (cap_entries && to_submit > 8)
6640 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6641 unsigned nr_events = 0;
6643 mutex_lock(&ctx->uring_lock);
6644 if (!list_empty(&ctx->iopoll_list))
6645 io_do_iopoll(ctx, &nr_events, 0);
6647 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6648 !(ctx->flags & IORING_SETUP_R_DISABLED))
6649 ret = io_submit_sqes(ctx, to_submit);
6650 mutex_unlock(&ctx->uring_lock);
6653 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6654 wake_up(&ctx->sqo_sq_wait);
6659 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6661 struct io_ring_ctx *ctx;
6662 unsigned sq_thread_idle = 0;
6664 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6665 if (sq_thread_idle < ctx->sq_thread_idle)
6666 sq_thread_idle = ctx->sq_thread_idle;
6669 sqd->sq_thread_idle = sq_thread_idle;
6672 static int io_sq_thread(void *data)
6674 struct io_sq_data *sqd = data;
6675 struct io_ring_ctx *ctx;
6676 unsigned long timeout = 0;
6677 char buf[TASK_COMM_LEN];
6680 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6681 set_task_comm(current, buf);
6682 current->pf_io_worker = NULL;
6684 if (sqd->sq_cpu != -1)
6685 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6687 set_cpus_allowed_ptr(current, cpu_online_mask);
6688 current->flags |= PF_NO_SETAFFINITY;
6690 down_read(&sqd->rw_lock);
6692 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6694 bool cap_entries, sqt_spin, needs_sched;
6696 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6697 up_read(&sqd->rw_lock);
6699 down_read(&sqd->rw_lock);
6701 timeout = jiffies + sqd->sq_thread_idle;
6704 if (fatal_signal_pending(current))
6707 cap_entries = !list_is_singular(&sqd->ctx_list);
6708 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6709 const struct cred *creds = NULL;
6711 if (ctx->sq_creds != current_cred())
6712 creds = override_creds(ctx->sq_creds);
6713 ret = __io_sq_thread(ctx, cap_entries);
6715 revert_creds(creds);
6716 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6720 if (sqt_spin || !time_after(jiffies, timeout)) {
6724 timeout = jiffies + sqd->sq_thread_idle;
6729 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6730 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6731 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6732 !list_empty_careful(&ctx->iopoll_list)) {
6733 needs_sched = false;
6736 if (io_sqring_entries(ctx)) {
6737 needs_sched = false;
6742 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6743 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6744 io_ring_set_wakeup_flag(ctx);
6746 up_read(&sqd->rw_lock);
6749 down_read(&sqd->rw_lock);
6750 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6751 io_ring_clear_wakeup_flag(ctx);
6754 finish_wait(&sqd->wait, &wait);
6755 timeout = jiffies + sqd->sq_thread_idle;
6757 up_read(&sqd->rw_lock);
6758 down_write(&sqd->rw_lock);
6760 * someone may have parked and added a cancellation task_work, run
6761 * it first because we don't want it in io_uring_cancel_sqpoll()
6765 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6766 io_uring_cancel_sqpoll(ctx);
6768 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6769 io_ring_set_wakeup_flag(ctx);
6770 up_write(&sqd->rw_lock);
6773 complete(&sqd->exited);
6777 struct io_wait_queue {
6778 struct wait_queue_entry wq;
6779 struct io_ring_ctx *ctx;
6781 unsigned nr_timeouts;
6784 static inline bool io_should_wake(struct io_wait_queue *iowq)
6786 struct io_ring_ctx *ctx = iowq->ctx;
6789 * Wake up if we have enough events, or if a timeout occurred since we
6790 * started waiting. For timeouts, we always want to return to userspace,
6791 * regardless of event count.
6793 return io_cqring_events(ctx) >= iowq->to_wait ||
6794 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6797 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6798 int wake_flags, void *key)
6800 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6804 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6805 * the task, and the next invocation will do it.
6807 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6808 return autoremove_wake_function(curr, mode, wake_flags, key);
6812 static int io_run_task_work_sig(void)
6814 if (io_run_task_work())
6816 if (!signal_pending(current))
6818 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6819 return -ERESTARTSYS;
6823 /* when returns >0, the caller should retry */
6824 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6825 struct io_wait_queue *iowq,
6826 signed long *timeout)
6830 /* make sure we run task_work before checking for signals */
6831 ret = io_run_task_work_sig();
6832 if (ret || io_should_wake(iowq))
6834 /* let the caller flush overflows, retry */
6835 if (test_bit(0, &ctx->cq_check_overflow))
6838 *timeout = schedule_timeout(*timeout);
6839 return !*timeout ? -ETIME : 1;
6843 * Wait until events become available, if we don't already have some. The
6844 * application must reap them itself, as they reside on the shared cq ring.
6846 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6847 const sigset_t __user *sig, size_t sigsz,
6848 struct __kernel_timespec __user *uts)
6850 struct io_wait_queue iowq = {
6853 .func = io_wake_function,
6854 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6857 .to_wait = min_events,
6859 struct io_rings *rings = ctx->rings;
6860 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6864 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6865 if (io_cqring_events(ctx) >= min_events)
6867 if (!io_run_task_work())
6872 #ifdef CONFIG_COMPAT
6873 if (in_compat_syscall())
6874 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6878 ret = set_user_sigmask(sig, sigsz);
6885 struct timespec64 ts;
6887 if (get_timespec64(&ts, uts))
6889 timeout = timespec64_to_jiffies(&ts);
6892 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6893 trace_io_uring_cqring_wait(ctx, min_events);
6895 /* if we can't even flush overflow, don't wait for more */
6896 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6900 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6901 TASK_INTERRUPTIBLE);
6902 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6903 finish_wait(&ctx->wait, &iowq.wq);
6907 restore_saved_sigmask_unless(ret == -EINTR);
6909 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6912 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6914 #if defined(CONFIG_UNIX)
6915 if (ctx->ring_sock) {
6916 struct sock *sock = ctx->ring_sock->sk;
6917 struct sk_buff *skb;
6919 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6925 for (i = 0; i < ctx->nr_user_files; i++) {
6928 file = io_file_from_index(ctx, i);
6935 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6937 struct fixed_rsrc_data *data;
6939 data = container_of(ref, struct fixed_rsrc_data, refs);
6940 complete(&data->done);
6943 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6945 spin_lock_bh(&ctx->rsrc_ref_lock);
6948 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6950 spin_unlock_bh(&ctx->rsrc_ref_lock);
6953 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6954 struct fixed_rsrc_data *rsrc_data,
6955 struct fixed_rsrc_ref_node *ref_node)
6957 io_rsrc_ref_lock(ctx);
6958 rsrc_data->node = ref_node;
6959 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6960 io_rsrc_ref_unlock(ctx);
6961 percpu_ref_get(&rsrc_data->refs);
6964 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6966 struct fixed_rsrc_ref_node *ref_node = NULL;
6968 io_rsrc_ref_lock(ctx);
6969 ref_node = data->node;
6971 io_rsrc_ref_unlock(ctx);
6973 percpu_ref_kill(&ref_node->refs);
6976 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6977 struct io_ring_ctx *ctx,
6978 void (*rsrc_put)(struct io_ring_ctx *ctx,
6979 struct io_rsrc_put *prsrc))
6981 struct fixed_rsrc_ref_node *backup_node;
6987 data->quiesce = true;
6990 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6993 backup_node->rsrc_data = data;
6994 backup_node->rsrc_put = rsrc_put;
6996 io_sqe_rsrc_kill_node(ctx, data);
6997 percpu_ref_kill(&data->refs);
6998 flush_delayed_work(&ctx->rsrc_put_work);
7000 ret = wait_for_completion_interruptible(&data->done);
7004 percpu_ref_resurrect(&data->refs);
7005 io_sqe_rsrc_set_node(ctx, data, backup_node);
7007 reinit_completion(&data->done);
7008 mutex_unlock(&ctx->uring_lock);
7009 ret = io_run_task_work_sig();
7010 mutex_lock(&ctx->uring_lock);
7012 data->quiesce = false;
7015 destroy_fixed_rsrc_ref_node(backup_node);
7019 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7021 struct fixed_rsrc_data *data;
7023 data = kzalloc(sizeof(*data), GFP_KERNEL);
7027 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7028 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7033 init_completion(&data->done);
7037 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7039 percpu_ref_exit(&data->refs);
7044 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7046 struct fixed_rsrc_data *data = ctx->file_data;
7047 unsigned nr_tables, i;
7051 * percpu_ref_is_dying() is to stop parallel files unregister
7052 * Since we possibly drop uring lock later in this function to
7055 if (!data || percpu_ref_is_dying(&data->refs))
7057 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7061 __io_sqe_files_unregister(ctx);
7062 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7063 for (i = 0; i < nr_tables; i++)
7064 kfree(data->table[i].files);
7065 free_fixed_rsrc_data(data);
7066 ctx->file_data = NULL;
7067 ctx->nr_user_files = 0;
7071 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7072 __releases(&sqd->rw_lock)
7074 WARN_ON_ONCE(sqd->thread == current);
7076 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7077 up_write(&sqd->rw_lock);
7080 static void io_sq_thread_park(struct io_sq_data *sqd)
7081 __acquires(&sqd->rw_lock)
7083 WARN_ON_ONCE(sqd->thread == current);
7085 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7086 down_write(&sqd->rw_lock);
7087 /* set again for consistency, in case concurrent parks are happening */
7088 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7090 wake_up_process(sqd->thread);
7093 static void io_sq_thread_stop(struct io_sq_data *sqd)
7095 WARN_ON_ONCE(sqd->thread == current);
7097 down_write(&sqd->rw_lock);
7098 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7100 wake_up_process(sqd->thread);
7101 up_write(&sqd->rw_lock);
7102 wait_for_completion(&sqd->exited);
7105 static void io_put_sq_data(struct io_sq_data *sqd)
7107 if (refcount_dec_and_test(&sqd->refs)) {
7108 io_sq_thread_stop(sqd);
7113 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7115 struct io_sq_data *sqd = ctx->sq_data;
7118 io_sq_thread_park(sqd);
7119 list_del_init(&ctx->sqd_list);
7120 io_sqd_update_thread_idle(sqd);
7121 io_sq_thread_unpark(sqd);
7123 io_put_sq_data(sqd);
7124 ctx->sq_data = NULL;
7126 put_cred(ctx->sq_creds);
7130 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7132 struct io_ring_ctx *ctx_attach;
7133 struct io_sq_data *sqd;
7136 f = fdget(p->wq_fd);
7138 return ERR_PTR(-ENXIO);
7139 if (f.file->f_op != &io_uring_fops) {
7141 return ERR_PTR(-EINVAL);
7144 ctx_attach = f.file->private_data;
7145 sqd = ctx_attach->sq_data;
7148 return ERR_PTR(-EINVAL);
7150 if (sqd->task_tgid != current->tgid) {
7152 return ERR_PTR(-EPERM);
7155 refcount_inc(&sqd->refs);
7160 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7163 struct io_sq_data *sqd;
7166 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7167 sqd = io_attach_sq_data(p);
7172 /* fall through for EPERM case, setup new sqd/task */
7173 if (PTR_ERR(sqd) != -EPERM)
7177 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7179 return ERR_PTR(-ENOMEM);
7181 refcount_set(&sqd->refs, 1);
7182 INIT_LIST_HEAD(&sqd->ctx_list);
7183 init_rwsem(&sqd->rw_lock);
7184 init_waitqueue_head(&sqd->wait);
7185 init_completion(&sqd->exited);
7189 #if defined(CONFIG_UNIX)
7191 * Ensure the UNIX gc is aware of our file set, so we are certain that
7192 * the io_uring can be safely unregistered on process exit, even if we have
7193 * loops in the file referencing.
7195 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7197 struct sock *sk = ctx->ring_sock->sk;
7198 struct scm_fp_list *fpl;
7199 struct sk_buff *skb;
7202 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7206 skb = alloc_skb(0, GFP_KERNEL);
7215 fpl->user = get_uid(current_user());
7216 for (i = 0; i < nr; i++) {
7217 struct file *file = io_file_from_index(ctx, i + offset);
7221 fpl->fp[nr_files] = get_file(file);
7222 unix_inflight(fpl->user, fpl->fp[nr_files]);
7227 fpl->max = SCM_MAX_FD;
7228 fpl->count = nr_files;
7229 UNIXCB(skb).fp = fpl;
7230 skb->destructor = unix_destruct_scm;
7231 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7232 skb_queue_head(&sk->sk_receive_queue, skb);
7234 for (i = 0; i < nr_files; i++)
7245 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7246 * causes regular reference counting to break down. We rely on the UNIX
7247 * garbage collection to take care of this problem for us.
7249 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7251 unsigned left, total;
7255 left = ctx->nr_user_files;
7257 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7259 ret = __io_sqe_files_scm(ctx, this_files, total);
7263 total += this_files;
7269 while (total < ctx->nr_user_files) {
7270 struct file *file = io_file_from_index(ctx, total);
7280 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7286 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7287 unsigned nr_tables, unsigned nr_files)
7291 for (i = 0; i < nr_tables; i++) {
7292 struct fixed_rsrc_table *table = &file_data->table[i];
7293 unsigned this_files;
7295 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7296 table->files = kcalloc(this_files, sizeof(struct file *),
7300 nr_files -= this_files;
7306 for (i = 0; i < nr_tables; i++) {
7307 struct fixed_rsrc_table *table = &file_data->table[i];
7308 kfree(table->files);
7313 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7315 struct file *file = prsrc->file;
7316 #if defined(CONFIG_UNIX)
7317 struct sock *sock = ctx->ring_sock->sk;
7318 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7319 struct sk_buff *skb;
7322 __skb_queue_head_init(&list);
7325 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7326 * remove this entry and rearrange the file array.
7328 skb = skb_dequeue(head);
7330 struct scm_fp_list *fp;
7332 fp = UNIXCB(skb).fp;
7333 for (i = 0; i < fp->count; i++) {
7336 if (fp->fp[i] != file)
7339 unix_notinflight(fp->user, fp->fp[i]);
7340 left = fp->count - 1 - i;
7342 memmove(&fp->fp[i], &fp->fp[i + 1],
7343 left * sizeof(struct file *));
7350 __skb_queue_tail(&list, skb);
7360 __skb_queue_tail(&list, skb);
7362 skb = skb_dequeue(head);
7365 if (skb_peek(&list)) {
7366 spin_lock_irq(&head->lock);
7367 while ((skb = __skb_dequeue(&list)) != NULL)
7368 __skb_queue_tail(head, skb);
7369 spin_unlock_irq(&head->lock);
7376 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7378 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7379 struct io_ring_ctx *ctx = rsrc_data->ctx;
7380 struct io_rsrc_put *prsrc, *tmp;
7382 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7383 list_del(&prsrc->list);
7384 ref_node->rsrc_put(ctx, prsrc);
7388 percpu_ref_exit(&ref_node->refs);
7390 percpu_ref_put(&rsrc_data->refs);
7393 static void io_rsrc_put_work(struct work_struct *work)
7395 struct io_ring_ctx *ctx;
7396 struct llist_node *node;
7398 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7399 node = llist_del_all(&ctx->rsrc_put_llist);
7402 struct fixed_rsrc_ref_node *ref_node;
7403 struct llist_node *next = node->next;
7405 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7406 __io_rsrc_put_work(ref_node);
7411 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7414 struct fixed_rsrc_table *table;
7416 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7417 return &table->files[i & IORING_FILE_TABLE_MASK];
7420 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7422 struct fixed_rsrc_ref_node *ref_node;
7423 struct fixed_rsrc_data *data;
7424 struct io_ring_ctx *ctx;
7425 bool first_add = false;
7428 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7429 data = ref_node->rsrc_data;
7432 io_rsrc_ref_lock(ctx);
7433 ref_node->done = true;
7435 while (!list_empty(&ctx->rsrc_ref_list)) {
7436 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7437 struct fixed_rsrc_ref_node, node);
7438 /* recycle ref nodes in order */
7439 if (!ref_node->done)
7441 list_del(&ref_node->node);
7442 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7444 io_rsrc_ref_unlock(ctx);
7446 if (percpu_ref_is_dying(&data->refs))
7450 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7452 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7455 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7456 struct io_ring_ctx *ctx)
7458 struct fixed_rsrc_ref_node *ref_node;
7460 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7464 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7469 INIT_LIST_HEAD(&ref_node->node);
7470 INIT_LIST_HEAD(&ref_node->rsrc_list);
7471 ref_node->done = false;
7475 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7476 struct fixed_rsrc_ref_node *ref_node)
7478 ref_node->rsrc_data = ctx->file_data;
7479 ref_node->rsrc_put = io_ring_file_put;
7482 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7484 percpu_ref_exit(&ref_node->refs);
7489 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7492 __s32 __user *fds = (__s32 __user *) arg;
7493 unsigned nr_tables, i;
7495 int fd, ret = -ENOMEM;
7496 struct fixed_rsrc_ref_node *ref_node;
7497 struct fixed_rsrc_data *file_data;
7503 if (nr_args > IORING_MAX_FIXED_FILES)
7506 file_data = alloc_fixed_rsrc_data(ctx);
7509 ctx->file_data = file_data;
7511 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7512 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7514 if (!file_data->table)
7517 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7520 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7521 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7525 /* allow sparse sets */
7535 * Don't allow io_uring instances to be registered. If UNIX
7536 * isn't enabled, then this causes a reference cycle and this
7537 * instance can never get freed. If UNIX is enabled we'll
7538 * handle it just fine, but there's still no point in allowing
7539 * a ring fd as it doesn't support regular read/write anyway.
7541 if (file->f_op == &io_uring_fops) {
7545 *io_fixed_file_slot(file_data, i) = file;
7548 ret = io_sqe_files_scm(ctx);
7550 io_sqe_files_unregister(ctx);
7554 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7556 io_sqe_files_unregister(ctx);
7559 init_fixed_file_ref_node(ctx, ref_node);
7561 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7564 for (i = 0; i < ctx->nr_user_files; i++) {
7565 file = io_file_from_index(ctx, i);
7569 for (i = 0; i < nr_tables; i++)
7570 kfree(file_data->table[i].files);
7571 ctx->nr_user_files = 0;
7573 free_fixed_rsrc_data(ctx->file_data);
7574 ctx->file_data = NULL;
7578 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7581 #if defined(CONFIG_UNIX)
7582 struct sock *sock = ctx->ring_sock->sk;
7583 struct sk_buff_head *head = &sock->sk_receive_queue;
7584 struct sk_buff *skb;
7587 * See if we can merge this file into an existing skb SCM_RIGHTS
7588 * file set. If there's no room, fall back to allocating a new skb
7589 * and filling it in.
7591 spin_lock_irq(&head->lock);
7592 skb = skb_peek(head);
7594 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7596 if (fpl->count < SCM_MAX_FD) {
7597 __skb_unlink(skb, head);
7598 spin_unlock_irq(&head->lock);
7599 fpl->fp[fpl->count] = get_file(file);
7600 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7602 spin_lock_irq(&head->lock);
7603 __skb_queue_head(head, skb);
7608 spin_unlock_irq(&head->lock);
7615 return __io_sqe_files_scm(ctx, 1, index);
7621 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7623 struct io_rsrc_put *prsrc;
7624 struct fixed_rsrc_ref_node *ref_node = data->node;
7626 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7631 list_add(&prsrc->list, &ref_node->rsrc_list);
7636 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7639 return io_queue_rsrc_removal(data, (void *)file);
7642 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7643 struct io_uring_rsrc_update *up,
7646 struct fixed_rsrc_data *data = ctx->file_data;
7647 struct fixed_rsrc_ref_node *ref_node;
7648 struct file *file, **file_slot;
7652 bool needs_switch = false;
7654 if (check_add_overflow(up->offset, nr_args, &done))
7656 if (done > ctx->nr_user_files)
7659 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7662 init_fixed_file_ref_node(ctx, ref_node);
7664 fds = u64_to_user_ptr(up->data);
7665 for (done = 0; done < nr_args; done++) {
7667 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7671 if (fd == IORING_REGISTER_FILES_SKIP)
7674 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7675 file_slot = io_fixed_file_slot(ctx->file_data, i);
7678 err = io_queue_file_removal(data, *file_slot);
7682 needs_switch = true;
7691 * Don't allow io_uring instances to be registered. If
7692 * UNIX isn't enabled, then this causes a reference
7693 * cycle and this instance can never get freed. If UNIX
7694 * is enabled we'll handle it just fine, but there's
7695 * still no point in allowing a ring fd as it doesn't
7696 * support regular read/write anyway.
7698 if (file->f_op == &io_uring_fops) {
7704 err = io_sqe_file_register(ctx, file, i);
7714 percpu_ref_kill(&data->node->refs);
7715 io_sqe_rsrc_set_node(ctx, data, ref_node);
7717 destroy_fixed_rsrc_ref_node(ref_node);
7719 return done ? done : err;
7722 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7725 struct io_uring_rsrc_update up;
7727 if (!ctx->file_data)
7731 if (copy_from_user(&up, arg, sizeof(up)))
7736 return __io_sqe_files_update(ctx, &up, nr_args);
7739 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7741 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7743 req = io_put_req_find_next(req);
7744 return req ? &req->work : NULL;
7747 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7749 struct io_wq_hash *hash;
7750 struct io_wq_data data;
7751 unsigned int concurrency;
7753 hash = ctx->hash_map;
7755 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7757 return ERR_PTR(-ENOMEM);
7758 refcount_set(&hash->refs, 1);
7759 init_waitqueue_head(&hash->wait);
7760 ctx->hash_map = hash;
7764 data.free_work = io_free_work;
7765 data.do_work = io_wq_submit_work;
7767 /* Do QD, or 4 * CPUS, whatever is smallest */
7768 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7770 return io_wq_create(concurrency, &data);
7773 static int io_uring_alloc_task_context(struct task_struct *task,
7774 struct io_ring_ctx *ctx)
7776 struct io_uring_task *tctx;
7779 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7780 if (unlikely(!tctx))
7783 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7784 if (unlikely(ret)) {
7789 tctx->io_wq = io_init_wq_offload(ctx);
7790 if (IS_ERR(tctx->io_wq)) {
7791 ret = PTR_ERR(tctx->io_wq);
7792 percpu_counter_destroy(&tctx->inflight);
7798 init_waitqueue_head(&tctx->wait);
7800 atomic_set(&tctx->in_idle, 0);
7801 task->io_uring = tctx;
7802 spin_lock_init(&tctx->task_lock);
7803 INIT_WQ_LIST(&tctx->task_list);
7804 tctx->task_state = 0;
7805 init_task_work(&tctx->task_work, tctx_task_work);
7809 void __io_uring_free(struct task_struct *tsk)
7811 struct io_uring_task *tctx = tsk->io_uring;
7813 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7814 WARN_ON_ONCE(tctx->io_wq);
7816 percpu_counter_destroy(&tctx->inflight);
7818 tsk->io_uring = NULL;
7821 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7822 struct io_uring_params *p)
7826 /* Retain compatibility with failing for an invalid attach attempt */
7827 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7828 IORING_SETUP_ATTACH_WQ) {
7831 f = fdget(p->wq_fd);
7834 if (f.file->f_op != &io_uring_fops) {
7840 if (ctx->flags & IORING_SETUP_SQPOLL) {
7841 struct task_struct *tsk;
7842 struct io_sq_data *sqd;
7846 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7849 sqd = io_get_sq_data(p, &attached);
7855 ctx->sq_creds = get_current_cred();
7857 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7858 if (!ctx->sq_thread_idle)
7859 ctx->sq_thread_idle = HZ;
7862 io_sq_thread_park(sqd);
7863 /* don't attach to a dying SQPOLL thread, would be racy */
7864 if (attached && !sqd->thread) {
7867 list_add(&ctx->sqd_list, &sqd->ctx_list);
7868 io_sqd_update_thread_idle(sqd);
7870 io_sq_thread_unpark(sqd);
7873 io_put_sq_data(sqd);
7874 ctx->sq_data = NULL;
7876 } else if (attached) {
7880 if (p->flags & IORING_SETUP_SQ_AFF) {
7881 int cpu = p->sq_thread_cpu;
7884 if (cpu >= nr_cpu_ids)
7886 if (!cpu_online(cpu))
7894 sqd->task_pid = current->pid;
7895 sqd->task_tgid = current->tgid;
7896 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7903 ret = io_uring_alloc_task_context(tsk, ctx);
7904 wake_up_new_task(tsk);
7907 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7908 /* Can't have SQ_AFF without SQPOLL */
7915 io_sq_thread_finish(ctx);
7918 complete(&ctx->sq_data->exited);
7922 static inline void __io_unaccount_mem(struct user_struct *user,
7923 unsigned long nr_pages)
7925 atomic_long_sub(nr_pages, &user->locked_vm);
7928 static inline int __io_account_mem(struct user_struct *user,
7929 unsigned long nr_pages)
7931 unsigned long page_limit, cur_pages, new_pages;
7933 /* Don't allow more pages than we can safely lock */
7934 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7937 cur_pages = atomic_long_read(&user->locked_vm);
7938 new_pages = cur_pages + nr_pages;
7939 if (new_pages > page_limit)
7941 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7942 new_pages) != cur_pages);
7947 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7950 __io_unaccount_mem(ctx->user, nr_pages);
7952 if (ctx->mm_account)
7953 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7956 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7961 ret = __io_account_mem(ctx->user, nr_pages);
7966 if (ctx->mm_account)
7967 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7972 static void io_mem_free(void *ptr)
7979 page = virt_to_head_page(ptr);
7980 if (put_page_testzero(page))
7981 free_compound_page(page);
7984 static void *io_mem_alloc(size_t size)
7986 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7987 __GFP_NORETRY | __GFP_ACCOUNT;
7989 return (void *) __get_free_pages(gfp_flags, get_order(size));
7992 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7995 struct io_rings *rings;
7996 size_t off, sq_array_size;
7998 off = struct_size(rings, cqes, cq_entries);
7999 if (off == SIZE_MAX)
8003 off = ALIGN(off, SMP_CACHE_BYTES);
8011 sq_array_size = array_size(sizeof(u32), sq_entries);
8012 if (sq_array_size == SIZE_MAX)
8015 if (check_add_overflow(off, sq_array_size, &off))
8021 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8025 if (!ctx->user_bufs)
8028 for (i = 0; i < ctx->nr_user_bufs; i++) {
8029 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8031 for (j = 0; j < imu->nr_bvecs; j++)
8032 unpin_user_page(imu->bvec[j].bv_page);
8034 if (imu->acct_pages)
8035 io_unaccount_mem(ctx, imu->acct_pages);
8040 kfree(ctx->user_bufs);
8041 ctx->user_bufs = NULL;
8042 ctx->nr_user_bufs = 0;
8046 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8047 void __user *arg, unsigned index)
8049 struct iovec __user *src;
8051 #ifdef CONFIG_COMPAT
8053 struct compat_iovec __user *ciovs;
8054 struct compat_iovec ciov;
8056 ciovs = (struct compat_iovec __user *) arg;
8057 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8060 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8061 dst->iov_len = ciov.iov_len;
8065 src = (struct iovec __user *) arg;
8066 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8072 * Not super efficient, but this is just a registration time. And we do cache
8073 * the last compound head, so generally we'll only do a full search if we don't
8076 * We check if the given compound head page has already been accounted, to
8077 * avoid double accounting it. This allows us to account the full size of the
8078 * page, not just the constituent pages of a huge page.
8080 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8081 int nr_pages, struct page *hpage)
8085 /* check current page array */
8086 for (i = 0; i < nr_pages; i++) {
8087 if (!PageCompound(pages[i]))
8089 if (compound_head(pages[i]) == hpage)
8093 /* check previously registered pages */
8094 for (i = 0; i < ctx->nr_user_bufs; i++) {
8095 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8097 for (j = 0; j < imu->nr_bvecs; j++) {
8098 if (!PageCompound(imu->bvec[j].bv_page))
8100 if (compound_head(imu->bvec[j].bv_page) == hpage)
8108 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8109 int nr_pages, struct io_mapped_ubuf *imu,
8110 struct page **last_hpage)
8114 for (i = 0; i < nr_pages; i++) {
8115 if (!PageCompound(pages[i])) {
8120 hpage = compound_head(pages[i]);
8121 if (hpage == *last_hpage)
8123 *last_hpage = hpage;
8124 if (headpage_already_acct(ctx, pages, i, hpage))
8126 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8130 if (!imu->acct_pages)
8133 ret = io_account_mem(ctx, imu->acct_pages);
8135 imu->acct_pages = 0;
8139 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8140 struct io_mapped_ubuf *imu,
8141 struct page **last_hpage)
8143 struct vm_area_struct **vmas = NULL;
8144 struct page **pages = NULL;
8145 unsigned long off, start, end, ubuf;
8147 int ret, pret, nr_pages, i;
8149 ubuf = (unsigned long) iov->iov_base;
8150 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8151 start = ubuf >> PAGE_SHIFT;
8152 nr_pages = end - start;
8156 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8160 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8165 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8171 mmap_read_lock(current->mm);
8172 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8174 if (pret == nr_pages) {
8175 /* don't support file backed memory */
8176 for (i = 0; i < nr_pages; i++) {
8177 struct vm_area_struct *vma = vmas[i];
8180 !is_file_hugepages(vma->vm_file)) {
8186 ret = pret < 0 ? pret : -EFAULT;
8188 mmap_read_unlock(current->mm);
8191 * if we did partial map, or found file backed vmas,
8192 * release any pages we did get
8195 unpin_user_pages(pages, pret);
8200 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8202 unpin_user_pages(pages, pret);
8207 off = ubuf & ~PAGE_MASK;
8208 size = iov->iov_len;
8209 for (i = 0; i < nr_pages; i++) {
8212 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8213 imu->bvec[i].bv_page = pages[i];
8214 imu->bvec[i].bv_len = vec_len;
8215 imu->bvec[i].bv_offset = off;
8219 /* store original address for later verification */
8221 imu->len = iov->iov_len;
8222 imu->nr_bvecs = nr_pages;
8230 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8234 if (!nr_args || nr_args > UIO_MAXIOV)
8237 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8239 if (!ctx->user_bufs)
8245 static int io_buffer_validate(struct iovec *iov)
8248 * Don't impose further limits on the size and buffer
8249 * constraints here, we'll -EINVAL later when IO is
8250 * submitted if they are wrong.
8252 if (!iov->iov_base || !iov->iov_len)
8255 /* arbitrary limit, but we need something */
8256 if (iov->iov_len > SZ_1G)
8262 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8263 unsigned int nr_args)
8267 struct page *last_hpage = NULL;
8269 ret = io_buffers_map_alloc(ctx, nr_args);
8273 for (i = 0; i < nr_args; i++) {
8274 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8276 ret = io_copy_iov(ctx, &iov, arg, i);
8280 ret = io_buffer_validate(&iov);
8284 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8288 ctx->nr_user_bufs++;
8292 io_sqe_buffers_unregister(ctx);
8297 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8299 __s32 __user *fds = arg;
8305 if (copy_from_user(&fd, fds, sizeof(*fds)))
8308 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8309 if (IS_ERR(ctx->cq_ev_fd)) {
8310 int ret = PTR_ERR(ctx->cq_ev_fd);
8311 ctx->cq_ev_fd = NULL;
8318 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8320 if (ctx->cq_ev_fd) {
8321 eventfd_ctx_put(ctx->cq_ev_fd);
8322 ctx->cq_ev_fd = NULL;
8329 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8331 struct io_buffer *buf;
8332 unsigned long index;
8334 xa_for_each(&ctx->io_buffers, index, buf)
8335 __io_remove_buffers(ctx, buf, index, -1U);
8338 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8340 struct io_kiocb *req, *nxt;
8342 list_for_each_entry_safe(req, nxt, list, compl.list) {
8343 if (tsk && req->task != tsk)
8345 list_del(&req->compl.list);
8346 kmem_cache_free(req_cachep, req);
8350 static void io_req_caches_free(struct io_ring_ctx *ctx)
8352 struct io_submit_state *submit_state = &ctx->submit_state;
8353 struct io_comp_state *cs = &ctx->submit_state.comp;
8355 mutex_lock(&ctx->uring_lock);
8357 if (submit_state->free_reqs) {
8358 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8359 submit_state->reqs);
8360 submit_state->free_reqs = 0;
8363 spin_lock_irq(&ctx->completion_lock);
8364 list_splice_init(&cs->locked_free_list, &cs->free_list);
8365 cs->locked_free_nr = 0;
8366 spin_unlock_irq(&ctx->completion_lock);
8368 io_req_cache_free(&cs->free_list, NULL);
8370 mutex_unlock(&ctx->uring_lock);
8373 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8376 * Some may use context even when all refs and requests have been put,
8377 * and they are free to do so while still holding uring_lock, see
8378 * __io_req_task_submit(). Wait for them to finish.
8380 mutex_lock(&ctx->uring_lock);
8381 mutex_unlock(&ctx->uring_lock);
8383 io_sq_thread_finish(ctx);
8384 io_sqe_buffers_unregister(ctx);
8386 if (ctx->mm_account) {
8387 mmdrop(ctx->mm_account);
8388 ctx->mm_account = NULL;
8391 mutex_lock(&ctx->uring_lock);
8392 io_sqe_files_unregister(ctx);
8393 mutex_unlock(&ctx->uring_lock);
8394 io_eventfd_unregister(ctx);
8395 io_destroy_buffers(ctx);
8397 #if defined(CONFIG_UNIX)
8398 if (ctx->ring_sock) {
8399 ctx->ring_sock->file = NULL; /* so that iput() is called */
8400 sock_release(ctx->ring_sock);
8404 io_mem_free(ctx->rings);
8405 io_mem_free(ctx->sq_sqes);
8407 percpu_ref_exit(&ctx->refs);
8408 free_uid(ctx->user);
8409 io_req_caches_free(ctx);
8411 io_wq_put_hash(ctx->hash_map);
8412 kfree(ctx->cancel_hash);
8416 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8418 struct io_ring_ctx *ctx = file->private_data;
8421 poll_wait(file, &ctx->cq_wait, wait);
8423 * synchronizes with barrier from wq_has_sleeper call in
8427 if (!io_sqring_full(ctx))
8428 mask |= EPOLLOUT | EPOLLWRNORM;
8431 * Don't flush cqring overflow list here, just do a simple check.
8432 * Otherwise there could possible be ABBA deadlock:
8435 * lock(&ctx->uring_lock);
8437 * lock(&ctx->uring_lock);
8440 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8441 * pushs them to do the flush.
8443 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8444 mask |= EPOLLIN | EPOLLRDNORM;
8449 static int io_uring_fasync(int fd, struct file *file, int on)
8451 struct io_ring_ctx *ctx = file->private_data;
8453 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8456 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8458 const struct cred *creds;
8460 creds = xa_erase(&ctx->personalities, id);
8469 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8471 struct callback_head *work, *next;
8472 bool executed = false;
8475 work = xchg(&ctx->exit_task_work, NULL);
8491 struct io_tctx_exit {
8492 struct callback_head task_work;
8493 struct completion completion;
8494 struct io_ring_ctx *ctx;
8497 static void io_tctx_exit_cb(struct callback_head *cb)
8499 struct io_uring_task *tctx = current->io_uring;
8500 struct io_tctx_exit *work;
8502 work = container_of(cb, struct io_tctx_exit, task_work);
8504 * When @in_idle, we're in cancellation and it's racy to remove the
8505 * node. It'll be removed by the end of cancellation, just ignore it.
8507 if (!atomic_read(&tctx->in_idle))
8508 io_uring_del_task_file((unsigned long)work->ctx);
8509 complete(&work->completion);
8512 static void io_ring_exit_work(struct work_struct *work)
8514 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8515 unsigned long timeout = jiffies + HZ * 60 * 5;
8516 struct io_tctx_exit exit;
8517 struct io_tctx_node *node;
8521 * If we're doing polled IO and end up having requests being
8522 * submitted async (out-of-line), then completions can come in while
8523 * we're waiting for refs to drop. We need to reap these manually,
8524 * as nobody else will be looking for them.
8527 io_uring_try_cancel_requests(ctx, NULL, NULL);
8529 WARN_ON_ONCE(time_after(jiffies, timeout));
8530 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8532 mutex_lock(&ctx->uring_lock);
8533 while (!list_empty(&ctx->tctx_list)) {
8534 WARN_ON_ONCE(time_after(jiffies, timeout));
8536 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8539 init_completion(&exit.completion);
8540 init_task_work(&exit.task_work, io_tctx_exit_cb);
8541 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8542 if (WARN_ON_ONCE(ret))
8544 wake_up_process(node->task);
8546 mutex_unlock(&ctx->uring_lock);
8547 wait_for_completion(&exit.completion);
8549 mutex_lock(&ctx->uring_lock);
8551 mutex_unlock(&ctx->uring_lock);
8553 io_ring_ctx_free(ctx);
8556 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8558 unsigned long index;
8559 struct creds *creds;
8561 mutex_lock(&ctx->uring_lock);
8562 percpu_ref_kill(&ctx->refs);
8563 /* if force is set, the ring is going away. always drop after that */
8564 ctx->cq_overflow_flushed = 1;
8566 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8567 xa_for_each(&ctx->personalities, index, creds)
8568 io_unregister_personality(ctx, index);
8569 mutex_unlock(&ctx->uring_lock);
8571 io_kill_timeouts(ctx, NULL, NULL);
8572 io_poll_remove_all(ctx, NULL, NULL);
8574 /* if we failed setting up the ctx, we might not have any rings */
8575 io_iopoll_try_reap_events(ctx);
8577 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8579 * Use system_unbound_wq to avoid spawning tons of event kworkers
8580 * if we're exiting a ton of rings at the same time. It just adds
8581 * noise and overhead, there's no discernable change in runtime
8582 * over using system_wq.
8584 queue_work(system_unbound_wq, &ctx->exit_work);
8587 static int io_uring_release(struct inode *inode, struct file *file)
8589 struct io_ring_ctx *ctx = file->private_data;
8591 file->private_data = NULL;
8592 io_ring_ctx_wait_and_kill(ctx);
8596 struct io_task_cancel {
8597 struct task_struct *task;
8598 struct files_struct *files;
8601 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8603 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8604 struct io_task_cancel *cancel = data;
8607 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8608 unsigned long flags;
8609 struct io_ring_ctx *ctx = req->ctx;
8611 /* protect against races with linked timeouts */
8612 spin_lock_irqsave(&ctx->completion_lock, flags);
8613 ret = io_match_task(req, cancel->task, cancel->files);
8614 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8616 ret = io_match_task(req, cancel->task, cancel->files);
8621 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8622 struct task_struct *task,
8623 struct files_struct *files)
8625 struct io_defer_entry *de;
8628 spin_lock_irq(&ctx->completion_lock);
8629 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8630 if (io_match_task(de->req, task, files)) {
8631 list_cut_position(&list, &ctx->defer_list, &de->list);
8635 spin_unlock_irq(&ctx->completion_lock);
8636 if (list_empty(&list))
8639 while (!list_empty(&list)) {
8640 de = list_first_entry(&list, struct io_defer_entry, list);
8641 list_del_init(&de->list);
8642 req_set_fail_links(de->req);
8643 io_put_req(de->req);
8644 io_req_complete(de->req, -ECANCELED);
8650 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8652 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8654 return req->ctx == data;
8657 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8659 struct io_tctx_node *node;
8660 enum io_wq_cancel cret;
8663 mutex_lock(&ctx->uring_lock);
8664 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8665 struct io_uring_task *tctx = node->task->io_uring;
8668 * io_wq will stay alive while we hold uring_lock, because it's
8669 * killed after ctx nodes, which requires to take the lock.
8671 if (!tctx || !tctx->io_wq)
8673 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8674 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8676 mutex_unlock(&ctx->uring_lock);
8681 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8682 struct task_struct *task,
8683 struct files_struct *files)
8685 struct io_task_cancel cancel = { .task = task, .files = files, };
8686 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8689 enum io_wq_cancel cret;
8693 ret |= io_uring_try_cancel_iowq(ctx);
8694 } else if (tctx && tctx->io_wq) {
8696 * Cancels requests of all rings, not only @ctx, but
8697 * it's fine as the task is in exit/exec.
8699 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8701 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8704 /* SQPOLL thread does its own polling */
8705 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8706 (ctx->sq_data && ctx->sq_data->thread == current)) {
8707 while (!list_empty_careful(&ctx->iopoll_list)) {
8708 io_iopoll_try_reap_events(ctx);
8713 ret |= io_cancel_defer_files(ctx, task, files);
8714 ret |= io_poll_remove_all(ctx, task, files);
8715 ret |= io_kill_timeouts(ctx, task, files);
8716 ret |= io_run_task_work();
8717 ret |= io_run_ctx_fallback(ctx);
8718 io_cqring_overflow_flush(ctx, true, task, files);
8725 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8726 struct task_struct *task,
8727 struct files_struct *files)
8729 struct io_kiocb *req;
8732 spin_lock_irq(&ctx->inflight_lock);
8733 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8734 cnt += io_match_task(req, task, files);
8735 spin_unlock_irq(&ctx->inflight_lock);
8739 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8740 struct task_struct *task,
8741 struct files_struct *files)
8743 while (!list_empty_careful(&ctx->inflight_list)) {
8747 inflight = io_uring_count_inflight(ctx, task, files);
8751 io_uring_try_cancel_requests(ctx, task, files);
8753 prepare_to_wait(&task->io_uring->wait, &wait,
8754 TASK_UNINTERRUPTIBLE);
8755 if (inflight == io_uring_count_inflight(ctx, task, files))
8757 finish_wait(&task->io_uring->wait, &wait);
8762 * Note that this task has used io_uring. We use it for cancelation purposes.
8764 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8766 struct io_uring_task *tctx = current->io_uring;
8767 struct io_tctx_node *node;
8770 if (unlikely(!tctx)) {
8771 ret = io_uring_alloc_task_context(current, ctx);
8774 tctx = current->io_uring;
8776 if (tctx->last != ctx) {
8777 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8780 node = kmalloc(sizeof(*node), GFP_KERNEL);
8784 node->task = current;
8786 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8793 mutex_lock(&ctx->uring_lock);
8794 list_add(&node->ctx_node, &ctx->tctx_list);
8795 mutex_unlock(&ctx->uring_lock);
8803 * Remove this io_uring_file -> task mapping.
8805 static void io_uring_del_task_file(unsigned long index)
8807 struct io_uring_task *tctx = current->io_uring;
8808 struct io_tctx_node *node;
8812 node = xa_erase(&tctx->xa, index);
8816 WARN_ON_ONCE(current != node->task);
8817 WARN_ON_ONCE(list_empty(&node->ctx_node));
8819 mutex_lock(&node->ctx->uring_lock);
8820 list_del(&node->ctx_node);
8821 mutex_unlock(&node->ctx->uring_lock);
8823 if (tctx->last == node->ctx)
8828 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8830 struct io_tctx_node *node;
8831 unsigned long index;
8833 xa_for_each(&tctx->xa, index, node)
8834 io_uring_del_task_file(index);
8836 io_wq_put_and_exit(tctx->io_wq);
8841 static s64 tctx_inflight(struct io_uring_task *tctx)
8843 return percpu_counter_sum(&tctx->inflight);
8846 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8848 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8849 struct io_ring_ctx *ctx = work->ctx;
8850 struct io_sq_data *sqd = ctx->sq_data;
8853 io_uring_cancel_sqpoll(ctx);
8854 complete(&work->completion);
8857 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8859 struct io_sq_data *sqd = ctx->sq_data;
8860 struct io_tctx_exit work = { .ctx = ctx, };
8861 struct task_struct *task;
8863 io_sq_thread_park(sqd);
8864 list_del_init(&ctx->sqd_list);
8865 io_sqd_update_thread_idle(sqd);
8868 init_completion(&work.completion);
8869 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8870 WARN_ON_ONCE(task_work_add(task, &work.task_work, TWA_SIGNAL));
8871 wake_up_process(task);
8873 io_sq_thread_unpark(sqd);
8876 wait_for_completion(&work.completion);
8879 void __io_uring_files_cancel(struct files_struct *files)
8881 struct io_uring_task *tctx = current->io_uring;
8882 struct io_tctx_node *node;
8883 unsigned long index;
8885 /* make sure overflow events are dropped */
8886 atomic_inc(&tctx->in_idle);
8887 xa_for_each(&tctx->xa, index, node) {
8888 struct io_ring_ctx *ctx = node->ctx;
8891 io_sqpoll_cancel_sync(ctx);
8894 io_uring_cancel_files(ctx, current, files);
8896 io_uring_try_cancel_requests(ctx, current, NULL);
8898 atomic_dec(&tctx->in_idle);
8901 io_uring_clean_tctx(tctx);
8904 /* should only be called by SQPOLL task */
8905 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8907 struct io_sq_data *sqd = ctx->sq_data;
8908 struct io_uring_task *tctx = current->io_uring;
8912 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
8914 atomic_inc(&tctx->in_idle);
8916 /* read completions before cancelations */
8917 inflight = tctx_inflight(tctx);
8920 io_uring_try_cancel_requests(ctx, current, NULL);
8922 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8924 * If we've seen completions, retry without waiting. This
8925 * avoids a race where a completion comes in before we did
8926 * prepare_to_wait().
8928 if (inflight == tctx_inflight(tctx))
8930 finish_wait(&tctx->wait, &wait);
8932 atomic_dec(&tctx->in_idle);
8936 * Find any io_uring fd that this task has registered or done IO on, and cancel
8939 void __io_uring_task_cancel(void)
8941 struct io_uring_task *tctx = current->io_uring;
8945 /* make sure overflow events are dropped */
8946 atomic_inc(&tctx->in_idle);
8948 /* read completions before cancelations */
8949 inflight = tctx_inflight(tctx);
8952 __io_uring_files_cancel(NULL);
8954 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8957 * If we've seen completions, retry without waiting. This
8958 * avoids a race where a completion comes in before we did
8959 * prepare_to_wait().
8961 if (inflight == tctx_inflight(tctx))
8963 finish_wait(&tctx->wait, &wait);
8966 atomic_dec(&tctx->in_idle);
8968 io_uring_clean_tctx(tctx);
8969 /* all current's requests should be gone, we can kill tctx */
8970 __io_uring_free(current);
8973 static void *io_uring_validate_mmap_request(struct file *file,
8974 loff_t pgoff, size_t sz)
8976 struct io_ring_ctx *ctx = file->private_data;
8977 loff_t offset = pgoff << PAGE_SHIFT;
8982 case IORING_OFF_SQ_RING:
8983 case IORING_OFF_CQ_RING:
8986 case IORING_OFF_SQES:
8990 return ERR_PTR(-EINVAL);
8993 page = virt_to_head_page(ptr);
8994 if (sz > page_size(page))
8995 return ERR_PTR(-EINVAL);
9002 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9004 size_t sz = vma->vm_end - vma->vm_start;
9008 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9010 return PTR_ERR(ptr);
9012 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9013 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9016 #else /* !CONFIG_MMU */
9018 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9020 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9023 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9025 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9028 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9029 unsigned long addr, unsigned long len,
9030 unsigned long pgoff, unsigned long flags)
9034 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9036 return PTR_ERR(ptr);
9038 return (unsigned long) ptr;
9041 #endif /* !CONFIG_MMU */
9043 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9048 if (!io_sqring_full(ctx))
9050 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9052 if (!io_sqring_full(ctx))
9055 } while (!signal_pending(current));
9057 finish_wait(&ctx->sqo_sq_wait, &wait);
9061 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9062 struct __kernel_timespec __user **ts,
9063 const sigset_t __user **sig)
9065 struct io_uring_getevents_arg arg;
9068 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9069 * is just a pointer to the sigset_t.
9071 if (!(flags & IORING_ENTER_EXT_ARG)) {
9072 *sig = (const sigset_t __user *) argp;
9078 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9079 * timespec and sigset_t pointers if good.
9081 if (*argsz != sizeof(arg))
9083 if (copy_from_user(&arg, argp, sizeof(arg)))
9085 *sig = u64_to_user_ptr(arg.sigmask);
9086 *argsz = arg.sigmask_sz;
9087 *ts = u64_to_user_ptr(arg.ts);
9091 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9092 u32, min_complete, u32, flags, const void __user *, argp,
9095 struct io_ring_ctx *ctx;
9102 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9103 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9111 if (f.file->f_op != &io_uring_fops)
9115 ctx = f.file->private_data;
9116 if (!percpu_ref_tryget(&ctx->refs))
9120 if (ctx->flags & IORING_SETUP_R_DISABLED)
9124 * For SQ polling, the thread will do all submissions and completions.
9125 * Just return the requested submit count, and wake the thread if
9129 if (ctx->flags & IORING_SETUP_SQPOLL) {
9130 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9133 if (unlikely(ctx->sq_data->thread == NULL)) {
9136 if (flags & IORING_ENTER_SQ_WAKEUP)
9137 wake_up(&ctx->sq_data->wait);
9138 if (flags & IORING_ENTER_SQ_WAIT) {
9139 ret = io_sqpoll_wait_sq(ctx);
9143 submitted = to_submit;
9144 } else if (to_submit) {
9145 ret = io_uring_add_task_file(ctx);
9148 mutex_lock(&ctx->uring_lock);
9149 submitted = io_submit_sqes(ctx, to_submit);
9150 mutex_unlock(&ctx->uring_lock);
9152 if (submitted != to_submit)
9155 if (flags & IORING_ENTER_GETEVENTS) {
9156 const sigset_t __user *sig;
9157 struct __kernel_timespec __user *ts;
9159 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9163 min_complete = min(min_complete, ctx->cq_entries);
9166 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9167 * space applications don't need to do io completion events
9168 * polling again, they can rely on io_sq_thread to do polling
9169 * work, which can reduce cpu usage and uring_lock contention.
9171 if (ctx->flags & IORING_SETUP_IOPOLL &&
9172 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9173 ret = io_iopoll_check(ctx, min_complete);
9175 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9180 percpu_ref_put(&ctx->refs);
9183 return submitted ? submitted : ret;
9186 #ifdef CONFIG_PROC_FS
9187 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9188 const struct cred *cred)
9190 struct user_namespace *uns = seq_user_ns(m);
9191 struct group_info *gi;
9196 seq_printf(m, "%5d\n", id);
9197 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9198 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9199 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9200 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9201 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9202 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9203 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9204 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9205 seq_puts(m, "\n\tGroups:\t");
9206 gi = cred->group_info;
9207 for (g = 0; g < gi->ngroups; g++) {
9208 seq_put_decimal_ull(m, g ? " " : "",
9209 from_kgid_munged(uns, gi->gid[g]));
9211 seq_puts(m, "\n\tCapEff:\t");
9212 cap = cred->cap_effective;
9213 CAP_FOR_EACH_U32(__capi)
9214 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9219 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9221 struct io_sq_data *sq = NULL;
9226 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9227 * since fdinfo case grabs it in the opposite direction of normal use
9228 * cases. If we fail to get the lock, we just don't iterate any
9229 * structures that could be going away outside the io_uring mutex.
9231 has_lock = mutex_trylock(&ctx->uring_lock);
9233 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9239 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9240 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9241 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9242 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9243 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9246 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9248 seq_printf(m, "%5u: <none>\n", i);
9250 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9251 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9252 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9254 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9255 (unsigned int) buf->len);
9257 if (has_lock && !xa_empty(&ctx->personalities)) {
9258 unsigned long index;
9259 const struct cred *cred;
9261 seq_printf(m, "Personalities:\n");
9262 xa_for_each(&ctx->personalities, index, cred)
9263 io_uring_show_cred(m, index, cred);
9265 seq_printf(m, "PollList:\n");
9266 spin_lock_irq(&ctx->completion_lock);
9267 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9268 struct hlist_head *list = &ctx->cancel_hash[i];
9269 struct io_kiocb *req;
9271 hlist_for_each_entry(req, list, hash_node)
9272 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9273 req->task->task_works != NULL);
9275 spin_unlock_irq(&ctx->completion_lock);
9277 mutex_unlock(&ctx->uring_lock);
9280 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9282 struct io_ring_ctx *ctx = f->private_data;
9284 if (percpu_ref_tryget(&ctx->refs)) {
9285 __io_uring_show_fdinfo(ctx, m);
9286 percpu_ref_put(&ctx->refs);
9291 static const struct file_operations io_uring_fops = {
9292 .release = io_uring_release,
9293 .mmap = io_uring_mmap,
9295 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9296 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9298 .poll = io_uring_poll,
9299 .fasync = io_uring_fasync,
9300 #ifdef CONFIG_PROC_FS
9301 .show_fdinfo = io_uring_show_fdinfo,
9305 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9306 struct io_uring_params *p)
9308 struct io_rings *rings;
9309 size_t size, sq_array_offset;
9311 /* make sure these are sane, as we already accounted them */
9312 ctx->sq_entries = p->sq_entries;
9313 ctx->cq_entries = p->cq_entries;
9315 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9316 if (size == SIZE_MAX)
9319 rings = io_mem_alloc(size);
9324 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9325 rings->sq_ring_mask = p->sq_entries - 1;
9326 rings->cq_ring_mask = p->cq_entries - 1;
9327 rings->sq_ring_entries = p->sq_entries;
9328 rings->cq_ring_entries = p->cq_entries;
9329 ctx->sq_mask = rings->sq_ring_mask;
9330 ctx->cq_mask = rings->cq_ring_mask;
9332 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9333 if (size == SIZE_MAX) {
9334 io_mem_free(ctx->rings);
9339 ctx->sq_sqes = io_mem_alloc(size);
9340 if (!ctx->sq_sqes) {
9341 io_mem_free(ctx->rings);
9349 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9353 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9357 ret = io_uring_add_task_file(ctx);
9362 fd_install(fd, file);
9367 * Allocate an anonymous fd, this is what constitutes the application
9368 * visible backing of an io_uring instance. The application mmaps this
9369 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9370 * we have to tie this fd to a socket for file garbage collection purposes.
9372 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9375 #if defined(CONFIG_UNIX)
9378 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9381 return ERR_PTR(ret);
9384 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9385 O_RDWR | O_CLOEXEC);
9386 #if defined(CONFIG_UNIX)
9388 sock_release(ctx->ring_sock);
9389 ctx->ring_sock = NULL;
9391 ctx->ring_sock->file = file;
9397 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9398 struct io_uring_params __user *params)
9400 struct io_ring_ctx *ctx;
9406 if (entries > IORING_MAX_ENTRIES) {
9407 if (!(p->flags & IORING_SETUP_CLAMP))
9409 entries = IORING_MAX_ENTRIES;
9413 * Use twice as many entries for the CQ ring. It's possible for the
9414 * application to drive a higher depth than the size of the SQ ring,
9415 * since the sqes are only used at submission time. This allows for
9416 * some flexibility in overcommitting a bit. If the application has
9417 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9418 * of CQ ring entries manually.
9420 p->sq_entries = roundup_pow_of_two(entries);
9421 if (p->flags & IORING_SETUP_CQSIZE) {
9423 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9424 * to a power-of-two, if it isn't already. We do NOT impose
9425 * any cq vs sq ring sizing.
9429 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9430 if (!(p->flags & IORING_SETUP_CLAMP))
9432 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9434 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9435 if (p->cq_entries < p->sq_entries)
9438 p->cq_entries = 2 * p->sq_entries;
9441 ctx = io_ring_ctx_alloc(p);
9444 ctx->compat = in_compat_syscall();
9445 if (!capable(CAP_IPC_LOCK))
9446 ctx->user = get_uid(current_user());
9449 * This is just grabbed for accounting purposes. When a process exits,
9450 * the mm is exited and dropped before the files, hence we need to hang
9451 * on to this mm purely for the purposes of being able to unaccount
9452 * memory (locked/pinned vm). It's not used for anything else.
9454 mmgrab(current->mm);
9455 ctx->mm_account = current->mm;
9457 ret = io_allocate_scq_urings(ctx, p);
9461 ret = io_sq_offload_create(ctx, p);
9465 memset(&p->sq_off, 0, sizeof(p->sq_off));
9466 p->sq_off.head = offsetof(struct io_rings, sq.head);
9467 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9468 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9469 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9470 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9471 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9472 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9474 memset(&p->cq_off, 0, sizeof(p->cq_off));
9475 p->cq_off.head = offsetof(struct io_rings, cq.head);
9476 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9477 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9478 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9479 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9480 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9481 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9483 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9484 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9485 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9486 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9487 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9489 if (copy_to_user(params, p, sizeof(*p))) {
9494 file = io_uring_get_file(ctx);
9496 ret = PTR_ERR(file);
9501 * Install ring fd as the very last thing, so we don't risk someone
9502 * having closed it before we finish setup
9504 ret = io_uring_install_fd(ctx, file);
9506 /* fput will clean it up */
9511 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9514 io_ring_ctx_wait_and_kill(ctx);
9519 * Sets up an aio uring context, and returns the fd. Applications asks for a
9520 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9521 * params structure passed in.
9523 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9525 struct io_uring_params p;
9528 if (copy_from_user(&p, params, sizeof(p)))
9530 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9535 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9536 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9537 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9538 IORING_SETUP_R_DISABLED))
9541 return io_uring_create(entries, &p, params);
9544 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9545 struct io_uring_params __user *, params)
9547 return io_uring_setup(entries, params);
9550 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9552 struct io_uring_probe *p;
9556 size = struct_size(p, ops, nr_args);
9557 if (size == SIZE_MAX)
9559 p = kzalloc(size, GFP_KERNEL);
9564 if (copy_from_user(p, arg, size))
9567 if (memchr_inv(p, 0, size))
9570 p->last_op = IORING_OP_LAST - 1;
9571 if (nr_args > IORING_OP_LAST)
9572 nr_args = IORING_OP_LAST;
9574 for (i = 0; i < nr_args; i++) {
9576 if (!io_op_defs[i].not_supported)
9577 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9582 if (copy_to_user(arg, p, size))
9589 static int io_register_personality(struct io_ring_ctx *ctx)
9591 const struct cred *creds;
9595 creds = get_current_cred();
9597 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9598 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9605 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9606 unsigned int nr_args)
9608 struct io_uring_restriction *res;
9612 /* Restrictions allowed only if rings started disabled */
9613 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9616 /* We allow only a single restrictions registration */
9617 if (ctx->restrictions.registered)
9620 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9623 size = array_size(nr_args, sizeof(*res));
9624 if (size == SIZE_MAX)
9627 res = memdup_user(arg, size);
9629 return PTR_ERR(res);
9633 for (i = 0; i < nr_args; i++) {
9634 switch (res[i].opcode) {
9635 case IORING_RESTRICTION_REGISTER_OP:
9636 if (res[i].register_op >= IORING_REGISTER_LAST) {
9641 __set_bit(res[i].register_op,
9642 ctx->restrictions.register_op);
9644 case IORING_RESTRICTION_SQE_OP:
9645 if (res[i].sqe_op >= IORING_OP_LAST) {
9650 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9652 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9653 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9655 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9656 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9665 /* Reset all restrictions if an error happened */
9667 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9669 ctx->restrictions.registered = true;
9675 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9677 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9680 if (ctx->restrictions.registered)
9681 ctx->restricted = 1;
9683 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9684 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9685 wake_up(&ctx->sq_data->wait);
9689 static bool io_register_op_must_quiesce(int op)
9692 case IORING_UNREGISTER_FILES:
9693 case IORING_REGISTER_FILES_UPDATE:
9694 case IORING_REGISTER_PROBE:
9695 case IORING_REGISTER_PERSONALITY:
9696 case IORING_UNREGISTER_PERSONALITY:
9703 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9704 void __user *arg, unsigned nr_args)
9705 __releases(ctx->uring_lock)
9706 __acquires(ctx->uring_lock)
9711 * We're inside the ring mutex, if the ref is already dying, then
9712 * someone else killed the ctx or is already going through
9713 * io_uring_register().
9715 if (percpu_ref_is_dying(&ctx->refs))
9718 if (io_register_op_must_quiesce(opcode)) {
9719 percpu_ref_kill(&ctx->refs);
9722 * Drop uring mutex before waiting for references to exit. If
9723 * another thread is currently inside io_uring_enter() it might
9724 * need to grab the uring_lock to make progress. If we hold it
9725 * here across the drain wait, then we can deadlock. It's safe
9726 * to drop the mutex here, since no new references will come in
9727 * after we've killed the percpu ref.
9729 mutex_unlock(&ctx->uring_lock);
9731 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9734 ret = io_run_task_work_sig();
9739 mutex_lock(&ctx->uring_lock);
9742 percpu_ref_resurrect(&ctx->refs);
9747 if (ctx->restricted) {
9748 if (opcode >= IORING_REGISTER_LAST) {
9753 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9760 case IORING_REGISTER_BUFFERS:
9761 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9763 case IORING_UNREGISTER_BUFFERS:
9767 ret = io_sqe_buffers_unregister(ctx);
9769 case IORING_REGISTER_FILES:
9770 ret = io_sqe_files_register(ctx, arg, nr_args);
9772 case IORING_UNREGISTER_FILES:
9776 ret = io_sqe_files_unregister(ctx);
9778 case IORING_REGISTER_FILES_UPDATE:
9779 ret = io_sqe_files_update(ctx, arg, nr_args);
9781 case IORING_REGISTER_EVENTFD:
9782 case IORING_REGISTER_EVENTFD_ASYNC:
9786 ret = io_eventfd_register(ctx, arg);
9789 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9790 ctx->eventfd_async = 1;
9792 ctx->eventfd_async = 0;
9794 case IORING_UNREGISTER_EVENTFD:
9798 ret = io_eventfd_unregister(ctx);
9800 case IORING_REGISTER_PROBE:
9802 if (!arg || nr_args > 256)
9804 ret = io_probe(ctx, arg, nr_args);
9806 case IORING_REGISTER_PERSONALITY:
9810 ret = io_register_personality(ctx);
9812 case IORING_UNREGISTER_PERSONALITY:
9816 ret = io_unregister_personality(ctx, nr_args);
9818 case IORING_REGISTER_ENABLE_RINGS:
9822 ret = io_register_enable_rings(ctx);
9824 case IORING_REGISTER_RESTRICTIONS:
9825 ret = io_register_restrictions(ctx, arg, nr_args);
9833 if (io_register_op_must_quiesce(opcode)) {
9834 /* bring the ctx back to life */
9835 percpu_ref_reinit(&ctx->refs);
9837 reinit_completion(&ctx->ref_comp);
9842 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9843 void __user *, arg, unsigned int, nr_args)
9845 struct io_ring_ctx *ctx;
9854 if (f.file->f_op != &io_uring_fops)
9857 ctx = f.file->private_data;
9861 mutex_lock(&ctx->uring_lock);
9862 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9863 mutex_unlock(&ctx->uring_lock);
9864 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9865 ctx->cq_ev_fd != NULL, ret);
9871 static int __init io_uring_init(void)
9873 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9874 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9875 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9878 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9879 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9880 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9881 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9882 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9883 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9884 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9885 BUILD_BUG_SQE_ELEM(8, __u64, off);
9886 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9887 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9888 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9889 BUILD_BUG_SQE_ELEM(24, __u32, len);
9890 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9891 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9892 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9893 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9894 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9895 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9896 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9897 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9898 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9899 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9900 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9901 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9902 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9903 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9904 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9905 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9906 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9907 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9908 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9910 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9911 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9912 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9916 __initcall(io_uring_init);