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 idr io_buffer_idr;
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 idr_init(&ctx->io_buffer_idr);
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 = idr_find(&req->ctx->io_buffer_idr, bgid);
2848 if (!list_empty(&head->list)) {
2849 kbuf = list_last_entry(&head->list, struct io_buffer,
2851 list_del(&kbuf->list);
2854 idr_remove(&req->ctx->io_buffer_idr, 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 idr_remove(&ctx->io_buffer_idr, 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 = idr_find(&ctx->io_buffer_idr, 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 = idr_find(&ctx->io_buffer_idr, p->bgid);
3998 ret = io_add_buffers(p, &head);
4003 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4006 __io_remove_buffers(ctx, head, p->bgid, -1U);
4012 req_set_fail_links(req);
4014 /* need to hold the lock to complete IOPOLL requests */
4015 if (ctx->flags & IORING_SETUP_IOPOLL) {
4016 __io_req_complete(req, issue_flags, ret, 0);
4017 io_ring_submit_unlock(ctx, !force_nonblock);
4019 io_ring_submit_unlock(ctx, !force_nonblock);
4020 __io_req_complete(req, issue_flags, ret, 0);
4025 static int io_epoll_ctl_prep(struct io_kiocb *req,
4026 const struct io_uring_sqe *sqe)
4028 #if defined(CONFIG_EPOLL)
4029 if (sqe->ioprio || sqe->buf_index)
4031 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4034 req->epoll.epfd = READ_ONCE(sqe->fd);
4035 req->epoll.op = READ_ONCE(sqe->len);
4036 req->epoll.fd = READ_ONCE(sqe->off);
4038 if (ep_op_has_event(req->epoll.op)) {
4039 struct epoll_event __user *ev;
4041 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4042 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4052 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4054 #if defined(CONFIG_EPOLL)
4055 struct io_epoll *ie = &req->epoll;
4057 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4059 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4060 if (force_nonblock && ret == -EAGAIN)
4064 req_set_fail_links(req);
4065 __io_req_complete(req, issue_flags, ret, 0);
4072 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4074 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4075 if (sqe->ioprio || sqe->buf_index || sqe->off)
4077 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4080 req->madvise.addr = READ_ONCE(sqe->addr);
4081 req->madvise.len = READ_ONCE(sqe->len);
4082 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4089 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4091 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4092 struct io_madvise *ma = &req->madvise;
4095 if (issue_flags & IO_URING_F_NONBLOCK)
4098 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4100 req_set_fail_links(req);
4101 io_req_complete(req, ret);
4108 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4110 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4112 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4115 req->fadvise.offset = READ_ONCE(sqe->off);
4116 req->fadvise.len = READ_ONCE(sqe->len);
4117 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4121 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4123 struct io_fadvise *fa = &req->fadvise;
4126 if (issue_flags & IO_URING_F_NONBLOCK) {
4127 switch (fa->advice) {
4128 case POSIX_FADV_NORMAL:
4129 case POSIX_FADV_RANDOM:
4130 case POSIX_FADV_SEQUENTIAL:
4137 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4139 req_set_fail_links(req);
4140 io_req_complete(req, ret);
4144 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4146 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4148 if (sqe->ioprio || sqe->buf_index)
4150 if (req->flags & REQ_F_FIXED_FILE)
4153 req->statx.dfd = READ_ONCE(sqe->fd);
4154 req->statx.mask = READ_ONCE(sqe->len);
4155 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4156 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4157 req->statx.flags = READ_ONCE(sqe->statx_flags);
4162 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4164 struct io_statx *ctx = &req->statx;
4167 if (issue_flags & IO_URING_F_NONBLOCK) {
4168 /* only need file table for an actual valid fd */
4169 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4170 req->flags |= REQ_F_NO_FILE_TABLE;
4174 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4178 req_set_fail_links(req);
4179 io_req_complete(req, ret);
4183 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4185 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4187 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4188 sqe->rw_flags || sqe->buf_index)
4190 if (req->flags & REQ_F_FIXED_FILE)
4193 req->close.fd = READ_ONCE(sqe->fd);
4197 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4199 struct files_struct *files = current->files;
4200 struct io_close *close = &req->close;
4201 struct fdtable *fdt;
4207 spin_lock(&files->file_lock);
4208 fdt = files_fdtable(files);
4209 if (close->fd >= fdt->max_fds) {
4210 spin_unlock(&files->file_lock);
4213 file = fdt->fd[close->fd];
4215 spin_unlock(&files->file_lock);
4219 if (file->f_op == &io_uring_fops) {
4220 spin_unlock(&files->file_lock);
4225 /* if the file has a flush method, be safe and punt to async */
4226 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4227 spin_unlock(&files->file_lock);
4231 ret = __close_fd_get_file(close->fd, &file);
4232 spin_unlock(&files->file_lock);
4239 /* No ->flush() or already async, safely close from here */
4240 ret = filp_close(file, current->files);
4243 req_set_fail_links(req);
4246 __io_req_complete(req, issue_flags, ret, 0);
4250 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4252 struct io_ring_ctx *ctx = req->ctx;
4254 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4256 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4259 req->sync.off = READ_ONCE(sqe->off);
4260 req->sync.len = READ_ONCE(sqe->len);
4261 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4265 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4269 /* sync_file_range always requires a blocking context */
4270 if (issue_flags & IO_URING_F_NONBLOCK)
4273 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4276 req_set_fail_links(req);
4277 io_req_complete(req, ret);
4281 #if defined(CONFIG_NET)
4282 static int io_setup_async_msg(struct io_kiocb *req,
4283 struct io_async_msghdr *kmsg)
4285 struct io_async_msghdr *async_msg = req->async_data;
4289 if (io_alloc_async_data(req)) {
4290 kfree(kmsg->free_iov);
4293 async_msg = req->async_data;
4294 req->flags |= REQ_F_NEED_CLEANUP;
4295 memcpy(async_msg, kmsg, sizeof(*kmsg));
4296 async_msg->msg.msg_name = &async_msg->addr;
4297 /* if were using fast_iov, set it to the new one */
4298 if (!async_msg->free_iov)
4299 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4304 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4305 struct io_async_msghdr *iomsg)
4307 iomsg->msg.msg_name = &iomsg->addr;
4308 iomsg->free_iov = iomsg->fast_iov;
4309 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4310 req->sr_msg.msg_flags, &iomsg->free_iov);
4313 static int io_sendmsg_prep_async(struct io_kiocb *req)
4317 if (!io_op_defs[req->opcode].needs_async_data)
4319 ret = io_sendmsg_copy_hdr(req, req->async_data);
4321 req->flags |= REQ_F_NEED_CLEANUP;
4325 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4327 struct io_sr_msg *sr = &req->sr_msg;
4329 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4332 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4333 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4334 sr->len = READ_ONCE(sqe->len);
4336 #ifdef CONFIG_COMPAT
4337 if (req->ctx->compat)
4338 sr->msg_flags |= MSG_CMSG_COMPAT;
4343 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4345 struct io_async_msghdr iomsg, *kmsg;
4346 struct socket *sock;
4350 sock = sock_from_file(req->file);
4351 if (unlikely(!sock))
4354 kmsg = req->async_data;
4356 ret = io_sendmsg_copy_hdr(req, &iomsg);
4362 flags = req->sr_msg.msg_flags;
4363 if (flags & MSG_DONTWAIT)
4364 req->flags |= REQ_F_NOWAIT;
4365 else if (issue_flags & IO_URING_F_NONBLOCK)
4366 flags |= MSG_DONTWAIT;
4368 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4369 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4370 return io_setup_async_msg(req, kmsg);
4371 if (ret == -ERESTARTSYS)
4374 /* fast path, check for non-NULL to avoid function call */
4376 kfree(kmsg->free_iov);
4377 req->flags &= ~REQ_F_NEED_CLEANUP;
4379 req_set_fail_links(req);
4380 __io_req_complete(req, issue_flags, ret, 0);
4384 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4386 struct io_sr_msg *sr = &req->sr_msg;
4389 struct socket *sock;
4393 sock = sock_from_file(req->file);
4394 if (unlikely(!sock))
4397 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4401 msg.msg_name = NULL;
4402 msg.msg_control = NULL;
4403 msg.msg_controllen = 0;
4404 msg.msg_namelen = 0;
4406 flags = req->sr_msg.msg_flags;
4407 if (flags & MSG_DONTWAIT)
4408 req->flags |= REQ_F_NOWAIT;
4409 else if (issue_flags & IO_URING_F_NONBLOCK)
4410 flags |= MSG_DONTWAIT;
4412 msg.msg_flags = flags;
4413 ret = sock_sendmsg(sock, &msg);
4414 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4416 if (ret == -ERESTARTSYS)
4420 req_set_fail_links(req);
4421 __io_req_complete(req, issue_flags, ret, 0);
4425 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4426 struct io_async_msghdr *iomsg)
4428 struct io_sr_msg *sr = &req->sr_msg;
4429 struct iovec __user *uiov;
4433 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4434 &iomsg->uaddr, &uiov, &iov_len);
4438 if (req->flags & REQ_F_BUFFER_SELECT) {
4441 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4443 sr->len = iomsg->fast_iov[0].iov_len;
4444 iomsg->free_iov = NULL;
4446 iomsg->free_iov = iomsg->fast_iov;
4447 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4448 &iomsg->free_iov, &iomsg->msg.msg_iter,
4457 #ifdef CONFIG_COMPAT
4458 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4459 struct io_async_msghdr *iomsg)
4461 struct compat_msghdr __user *msg_compat;
4462 struct io_sr_msg *sr = &req->sr_msg;
4463 struct compat_iovec __user *uiov;
4468 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4469 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4474 uiov = compat_ptr(ptr);
4475 if (req->flags & REQ_F_BUFFER_SELECT) {
4476 compat_ssize_t clen;
4480 if (!access_ok(uiov, sizeof(*uiov)))
4482 if (__get_user(clen, &uiov->iov_len))
4487 iomsg->free_iov = NULL;
4489 iomsg->free_iov = iomsg->fast_iov;
4490 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4491 UIO_FASTIOV, &iomsg->free_iov,
4492 &iomsg->msg.msg_iter, true);
4501 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4502 struct io_async_msghdr *iomsg)
4504 iomsg->msg.msg_name = &iomsg->addr;
4506 #ifdef CONFIG_COMPAT
4507 if (req->ctx->compat)
4508 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4511 return __io_recvmsg_copy_hdr(req, iomsg);
4514 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4517 struct io_sr_msg *sr = &req->sr_msg;
4518 struct io_buffer *kbuf;
4520 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4525 req->flags |= REQ_F_BUFFER_SELECTED;
4529 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4531 return io_put_kbuf(req, req->sr_msg.kbuf);
4534 static int io_recvmsg_prep_async(struct io_kiocb *req)
4538 if (!io_op_defs[req->opcode].needs_async_data)
4540 ret = io_recvmsg_copy_hdr(req, req->async_data);
4542 req->flags |= REQ_F_NEED_CLEANUP;
4546 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4548 struct io_sr_msg *sr = &req->sr_msg;
4550 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4553 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4554 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4555 sr->len = READ_ONCE(sqe->len);
4556 sr->bgid = READ_ONCE(sqe->buf_group);
4558 #ifdef CONFIG_COMPAT
4559 if (req->ctx->compat)
4560 sr->msg_flags |= MSG_CMSG_COMPAT;
4565 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4567 struct io_async_msghdr iomsg, *kmsg;
4568 struct socket *sock;
4569 struct io_buffer *kbuf;
4571 int ret, cflags = 0;
4572 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4574 sock = sock_from_file(req->file);
4575 if (unlikely(!sock))
4578 kmsg = req->async_data;
4580 ret = io_recvmsg_copy_hdr(req, &iomsg);
4586 if (req->flags & REQ_F_BUFFER_SELECT) {
4587 kbuf = io_recv_buffer_select(req, !force_nonblock);
4589 return PTR_ERR(kbuf);
4590 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4591 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4592 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4593 1, req->sr_msg.len);
4596 flags = req->sr_msg.msg_flags;
4597 if (flags & MSG_DONTWAIT)
4598 req->flags |= REQ_F_NOWAIT;
4599 else if (force_nonblock)
4600 flags |= MSG_DONTWAIT;
4602 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4603 kmsg->uaddr, flags);
4604 if (force_nonblock && ret == -EAGAIN)
4605 return io_setup_async_msg(req, kmsg);
4606 if (ret == -ERESTARTSYS)
4609 if (req->flags & REQ_F_BUFFER_SELECTED)
4610 cflags = io_put_recv_kbuf(req);
4611 /* fast path, check for non-NULL to avoid function call */
4613 kfree(kmsg->free_iov);
4614 req->flags &= ~REQ_F_NEED_CLEANUP;
4616 req_set_fail_links(req);
4617 __io_req_complete(req, issue_flags, ret, cflags);
4621 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4623 struct io_buffer *kbuf;
4624 struct io_sr_msg *sr = &req->sr_msg;
4626 void __user *buf = sr->buf;
4627 struct socket *sock;
4630 int ret, cflags = 0;
4631 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4633 sock = sock_from_file(req->file);
4634 if (unlikely(!sock))
4637 if (req->flags & REQ_F_BUFFER_SELECT) {
4638 kbuf = io_recv_buffer_select(req, !force_nonblock);
4640 return PTR_ERR(kbuf);
4641 buf = u64_to_user_ptr(kbuf->addr);
4644 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4648 msg.msg_name = NULL;
4649 msg.msg_control = NULL;
4650 msg.msg_controllen = 0;
4651 msg.msg_namelen = 0;
4652 msg.msg_iocb = NULL;
4655 flags = req->sr_msg.msg_flags;
4656 if (flags & MSG_DONTWAIT)
4657 req->flags |= REQ_F_NOWAIT;
4658 else if (force_nonblock)
4659 flags |= MSG_DONTWAIT;
4661 ret = sock_recvmsg(sock, &msg, flags);
4662 if (force_nonblock && ret == -EAGAIN)
4664 if (ret == -ERESTARTSYS)
4667 if (req->flags & REQ_F_BUFFER_SELECTED)
4668 cflags = io_put_recv_kbuf(req);
4670 req_set_fail_links(req);
4671 __io_req_complete(req, issue_flags, ret, cflags);
4675 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4677 struct io_accept *accept = &req->accept;
4679 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4681 if (sqe->ioprio || sqe->len || sqe->buf_index)
4684 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4685 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4686 accept->flags = READ_ONCE(sqe->accept_flags);
4687 accept->nofile = rlimit(RLIMIT_NOFILE);
4691 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4693 struct io_accept *accept = &req->accept;
4694 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4695 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4698 if (req->file->f_flags & O_NONBLOCK)
4699 req->flags |= REQ_F_NOWAIT;
4701 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4702 accept->addr_len, accept->flags,
4704 if (ret == -EAGAIN && force_nonblock)
4707 if (ret == -ERESTARTSYS)
4709 req_set_fail_links(req);
4711 __io_req_complete(req, issue_flags, ret, 0);
4715 static int io_connect_prep_async(struct io_kiocb *req)
4717 struct io_async_connect *io = req->async_data;
4718 struct io_connect *conn = &req->connect;
4720 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4723 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4725 struct io_connect *conn = &req->connect;
4727 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4729 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4732 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4733 conn->addr_len = READ_ONCE(sqe->addr2);
4737 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4739 struct io_async_connect __io, *io;
4740 unsigned file_flags;
4742 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4744 if (req->async_data) {
4745 io = req->async_data;
4747 ret = move_addr_to_kernel(req->connect.addr,
4748 req->connect.addr_len,
4755 file_flags = force_nonblock ? O_NONBLOCK : 0;
4757 ret = __sys_connect_file(req->file, &io->address,
4758 req->connect.addr_len, file_flags);
4759 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4760 if (req->async_data)
4762 if (io_alloc_async_data(req)) {
4766 io = req->async_data;
4767 memcpy(req->async_data, &__io, sizeof(__io));
4770 if (ret == -ERESTARTSYS)
4774 req_set_fail_links(req);
4775 __io_req_complete(req, issue_flags, ret, 0);
4778 #else /* !CONFIG_NET */
4779 #define IO_NETOP_FN(op) \
4780 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4782 return -EOPNOTSUPP; \
4785 #define IO_NETOP_PREP(op) \
4787 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4789 return -EOPNOTSUPP; \
4792 #define IO_NETOP_PREP_ASYNC(op) \
4794 static int io_##op##_prep_async(struct io_kiocb *req) \
4796 return -EOPNOTSUPP; \
4799 IO_NETOP_PREP_ASYNC(sendmsg);
4800 IO_NETOP_PREP_ASYNC(recvmsg);
4801 IO_NETOP_PREP_ASYNC(connect);
4802 IO_NETOP_PREP(accept);
4805 #endif /* CONFIG_NET */
4807 struct io_poll_table {
4808 struct poll_table_struct pt;
4809 struct io_kiocb *req;
4813 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4814 __poll_t mask, task_work_func_t func)
4818 /* for instances that support it check for an event match first: */
4819 if (mask && !(mask & poll->events))
4822 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4824 list_del_init(&poll->wait.entry);
4827 req->task_work.func = func;
4828 percpu_ref_get(&req->ctx->refs);
4831 * If this fails, then the task is exiting. When a task exits, the
4832 * work gets canceled, so just cancel this request as well instead
4833 * of executing it. We can't safely execute it anyway, as we may not
4834 * have the needed state needed for it anyway.
4836 ret = io_req_task_work_add(req);
4837 if (unlikely(ret)) {
4838 WRITE_ONCE(poll->canceled, true);
4839 io_req_task_work_add_fallback(req, func);
4844 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4845 __acquires(&req->ctx->completion_lock)
4847 struct io_ring_ctx *ctx = req->ctx;
4849 if (!req->result && !READ_ONCE(poll->canceled)) {
4850 struct poll_table_struct pt = { ._key = poll->events };
4852 req->result = vfs_poll(req->file, &pt) & poll->events;
4855 spin_lock_irq(&ctx->completion_lock);
4856 if (!req->result && !READ_ONCE(poll->canceled)) {
4857 add_wait_queue(poll->head, &poll->wait);
4864 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4866 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4867 if (req->opcode == IORING_OP_POLL_ADD)
4868 return req->async_data;
4869 return req->apoll->double_poll;
4872 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4874 if (req->opcode == IORING_OP_POLL_ADD)
4876 return &req->apoll->poll;
4879 static void io_poll_remove_double(struct io_kiocb *req)
4881 struct io_poll_iocb *poll = io_poll_get_double(req);
4883 lockdep_assert_held(&req->ctx->completion_lock);
4885 if (poll && poll->head) {
4886 struct wait_queue_head *head = poll->head;
4888 spin_lock(&head->lock);
4889 list_del_init(&poll->wait.entry);
4890 if (poll->wait.private)
4891 refcount_dec(&req->refs);
4893 spin_unlock(&head->lock);
4897 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4899 struct io_ring_ctx *ctx = req->ctx;
4901 io_poll_remove_double(req);
4902 req->poll.done = true;
4903 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4904 io_commit_cqring(ctx);
4907 static void io_poll_task_func(struct callback_head *cb)
4909 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4910 struct io_ring_ctx *ctx = req->ctx;
4911 struct io_kiocb *nxt;
4913 if (io_poll_rewait(req, &req->poll)) {
4914 spin_unlock_irq(&ctx->completion_lock);
4916 hash_del(&req->hash_node);
4917 io_poll_complete(req, req->result, 0);
4918 spin_unlock_irq(&ctx->completion_lock);
4920 nxt = io_put_req_find_next(req);
4921 io_cqring_ev_posted(ctx);
4923 __io_req_task_submit(nxt);
4926 percpu_ref_put(&ctx->refs);
4929 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4930 int sync, void *key)
4932 struct io_kiocb *req = wait->private;
4933 struct io_poll_iocb *poll = io_poll_get_single(req);
4934 __poll_t mask = key_to_poll(key);
4936 /* for instances that support it check for an event match first: */
4937 if (mask && !(mask & poll->events))
4940 list_del_init(&wait->entry);
4942 if (poll && poll->head) {
4945 spin_lock(&poll->head->lock);
4946 done = list_empty(&poll->wait.entry);
4948 list_del_init(&poll->wait.entry);
4949 /* make sure double remove sees this as being gone */
4950 wait->private = NULL;
4951 spin_unlock(&poll->head->lock);
4953 /* use wait func handler, so it matches the rq type */
4954 poll->wait.func(&poll->wait, mode, sync, key);
4957 refcount_dec(&req->refs);
4961 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4962 wait_queue_func_t wake_func)
4966 poll->canceled = false;
4967 poll->events = events;
4968 INIT_LIST_HEAD(&poll->wait.entry);
4969 init_waitqueue_func_entry(&poll->wait, wake_func);
4972 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4973 struct wait_queue_head *head,
4974 struct io_poll_iocb **poll_ptr)
4976 struct io_kiocb *req = pt->req;
4979 * If poll->head is already set, it's because the file being polled
4980 * uses multiple waitqueues for poll handling (eg one for read, one
4981 * for write). Setup a separate io_poll_iocb if this happens.
4983 if (unlikely(poll->head)) {
4984 struct io_poll_iocb *poll_one = poll;
4986 /* already have a 2nd entry, fail a third attempt */
4988 pt->error = -EINVAL;
4991 /* double add on the same waitqueue head, ignore */
4992 if (poll->head == head)
4994 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4996 pt->error = -ENOMEM;
4999 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5000 refcount_inc(&req->refs);
5001 poll->wait.private = req;
5008 if (poll->events & EPOLLEXCLUSIVE)
5009 add_wait_queue_exclusive(head, &poll->wait);
5011 add_wait_queue(head, &poll->wait);
5014 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5015 struct poll_table_struct *p)
5017 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5018 struct async_poll *apoll = pt->req->apoll;
5020 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5023 static void io_async_task_func(struct callback_head *cb)
5025 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5026 struct async_poll *apoll = req->apoll;
5027 struct io_ring_ctx *ctx = req->ctx;
5029 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5031 if (io_poll_rewait(req, &apoll->poll)) {
5032 spin_unlock_irq(&ctx->completion_lock);
5033 percpu_ref_put(&ctx->refs);
5037 /* If req is still hashed, it cannot have been canceled. Don't check. */
5038 if (hash_hashed(&req->hash_node))
5039 hash_del(&req->hash_node);
5041 io_poll_remove_double(req);
5042 spin_unlock_irq(&ctx->completion_lock);
5044 if (!READ_ONCE(apoll->poll.canceled))
5045 __io_req_task_submit(req);
5047 __io_req_task_cancel(req, -ECANCELED);
5049 percpu_ref_put(&ctx->refs);
5050 kfree(apoll->double_poll);
5054 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5057 struct io_kiocb *req = wait->private;
5058 struct io_poll_iocb *poll = &req->apoll->poll;
5060 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5063 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5066 static void io_poll_req_insert(struct io_kiocb *req)
5068 struct io_ring_ctx *ctx = req->ctx;
5069 struct hlist_head *list;
5071 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5072 hlist_add_head(&req->hash_node, list);
5075 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5076 struct io_poll_iocb *poll,
5077 struct io_poll_table *ipt, __poll_t mask,
5078 wait_queue_func_t wake_func)
5079 __acquires(&ctx->completion_lock)
5081 struct io_ring_ctx *ctx = req->ctx;
5082 bool cancel = false;
5084 INIT_HLIST_NODE(&req->hash_node);
5085 io_init_poll_iocb(poll, mask, wake_func);
5086 poll->file = req->file;
5087 poll->wait.private = req;
5089 ipt->pt._key = mask;
5091 ipt->error = -EINVAL;
5093 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5095 spin_lock_irq(&ctx->completion_lock);
5096 if (likely(poll->head)) {
5097 spin_lock(&poll->head->lock);
5098 if (unlikely(list_empty(&poll->wait.entry))) {
5104 if (mask || ipt->error)
5105 list_del_init(&poll->wait.entry);
5107 WRITE_ONCE(poll->canceled, true);
5108 else if (!poll->done) /* actually waiting for an event */
5109 io_poll_req_insert(req);
5110 spin_unlock(&poll->head->lock);
5116 static bool io_arm_poll_handler(struct io_kiocb *req)
5118 const struct io_op_def *def = &io_op_defs[req->opcode];
5119 struct io_ring_ctx *ctx = req->ctx;
5120 struct async_poll *apoll;
5121 struct io_poll_table ipt;
5125 if (!req->file || !file_can_poll(req->file))
5127 if (req->flags & REQ_F_POLLED)
5131 else if (def->pollout)
5135 /* if we can't nonblock try, then no point in arming a poll handler */
5136 if (!io_file_supports_async(req->file, rw))
5139 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5140 if (unlikely(!apoll))
5142 apoll->double_poll = NULL;
5144 req->flags |= REQ_F_POLLED;
5149 mask |= POLLIN | POLLRDNORM;
5151 mask |= POLLOUT | POLLWRNORM;
5153 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5154 if ((req->opcode == IORING_OP_RECVMSG) &&
5155 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5158 mask |= POLLERR | POLLPRI;
5160 ipt.pt._qproc = io_async_queue_proc;
5162 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5164 if (ret || ipt.error) {
5165 io_poll_remove_double(req);
5166 spin_unlock_irq(&ctx->completion_lock);
5167 kfree(apoll->double_poll);
5171 spin_unlock_irq(&ctx->completion_lock);
5172 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5173 apoll->poll.events);
5177 static bool __io_poll_remove_one(struct io_kiocb *req,
5178 struct io_poll_iocb *poll)
5180 bool do_complete = false;
5182 spin_lock(&poll->head->lock);
5183 WRITE_ONCE(poll->canceled, true);
5184 if (!list_empty(&poll->wait.entry)) {
5185 list_del_init(&poll->wait.entry);
5188 spin_unlock(&poll->head->lock);
5189 hash_del(&req->hash_node);
5193 static bool io_poll_remove_one(struct io_kiocb *req)
5197 io_poll_remove_double(req);
5199 if (req->opcode == IORING_OP_POLL_ADD) {
5200 do_complete = __io_poll_remove_one(req, &req->poll);
5202 struct async_poll *apoll = req->apoll;
5204 /* non-poll requests have submit ref still */
5205 do_complete = __io_poll_remove_one(req, &apoll->poll);
5208 kfree(apoll->double_poll);
5214 io_cqring_fill_event(req, -ECANCELED);
5215 io_commit_cqring(req->ctx);
5216 req_set_fail_links(req);
5217 io_put_req_deferred(req, 1);
5224 * Returns true if we found and killed one or more poll requests
5226 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5227 struct files_struct *files)
5229 struct hlist_node *tmp;
5230 struct io_kiocb *req;
5233 spin_lock_irq(&ctx->completion_lock);
5234 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5235 struct hlist_head *list;
5237 list = &ctx->cancel_hash[i];
5238 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5239 if (io_match_task(req, tsk, files))
5240 posted += io_poll_remove_one(req);
5243 spin_unlock_irq(&ctx->completion_lock);
5246 io_cqring_ev_posted(ctx);
5251 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5253 struct hlist_head *list;
5254 struct io_kiocb *req;
5256 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5257 hlist_for_each_entry(req, list, hash_node) {
5258 if (sqe_addr != req->user_data)
5260 if (io_poll_remove_one(req))
5268 static int io_poll_remove_prep(struct io_kiocb *req,
5269 const struct io_uring_sqe *sqe)
5271 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5273 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5277 req->poll_remove.addr = READ_ONCE(sqe->addr);
5282 * Find a running poll command that matches one specified in sqe->addr,
5283 * and remove it if found.
5285 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5287 struct io_ring_ctx *ctx = req->ctx;
5290 spin_lock_irq(&ctx->completion_lock);
5291 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5292 spin_unlock_irq(&ctx->completion_lock);
5295 req_set_fail_links(req);
5296 io_req_complete(req, ret);
5300 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5303 struct io_kiocb *req = wait->private;
5304 struct io_poll_iocb *poll = &req->poll;
5306 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5309 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5310 struct poll_table_struct *p)
5312 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5314 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5317 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5319 struct io_poll_iocb *poll = &req->poll;
5322 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5324 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5327 events = READ_ONCE(sqe->poll32_events);
5329 events = swahw32(events);
5331 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5332 (events & EPOLLEXCLUSIVE);
5336 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5338 struct io_poll_iocb *poll = &req->poll;
5339 struct io_ring_ctx *ctx = req->ctx;
5340 struct io_poll_table ipt;
5343 ipt.pt._qproc = io_poll_queue_proc;
5345 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5348 if (mask) { /* no async, we'd stolen it */
5350 io_poll_complete(req, mask, 0);
5352 spin_unlock_irq(&ctx->completion_lock);
5355 io_cqring_ev_posted(ctx);
5361 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5363 struct io_timeout_data *data = container_of(timer,
5364 struct io_timeout_data, timer);
5365 struct io_kiocb *req = data->req;
5366 struct io_ring_ctx *ctx = req->ctx;
5367 unsigned long flags;
5369 spin_lock_irqsave(&ctx->completion_lock, flags);
5370 list_del_init(&req->timeout.list);
5371 atomic_set(&req->ctx->cq_timeouts,
5372 atomic_read(&req->ctx->cq_timeouts) + 1);
5374 io_cqring_fill_event(req, -ETIME);
5375 io_commit_cqring(ctx);
5376 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5378 io_cqring_ev_posted(ctx);
5379 req_set_fail_links(req);
5381 return HRTIMER_NORESTART;
5384 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5387 struct io_timeout_data *io;
5388 struct io_kiocb *req;
5391 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5392 if (user_data == req->user_data) {
5399 return ERR_PTR(ret);
5401 io = req->async_data;
5402 ret = hrtimer_try_to_cancel(&io->timer);
5404 return ERR_PTR(-EALREADY);
5405 list_del_init(&req->timeout.list);
5409 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5411 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5414 return PTR_ERR(req);
5416 req_set_fail_links(req);
5417 io_cqring_fill_event(req, -ECANCELED);
5418 io_put_req_deferred(req, 1);
5422 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5423 struct timespec64 *ts, enum hrtimer_mode mode)
5425 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5426 struct io_timeout_data *data;
5429 return PTR_ERR(req);
5431 req->timeout.off = 0; /* noseq */
5432 data = req->async_data;
5433 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5434 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5435 data->timer.function = io_timeout_fn;
5436 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5440 static int io_timeout_remove_prep(struct io_kiocb *req,
5441 const struct io_uring_sqe *sqe)
5443 struct io_timeout_rem *tr = &req->timeout_rem;
5445 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5447 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5449 if (sqe->ioprio || sqe->buf_index || sqe->len)
5452 tr->addr = READ_ONCE(sqe->addr);
5453 tr->flags = READ_ONCE(sqe->timeout_flags);
5454 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5455 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5457 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5459 } else if (tr->flags) {
5460 /* timeout removal doesn't support flags */
5467 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5469 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5474 * Remove or update an existing timeout command
5476 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5478 struct io_timeout_rem *tr = &req->timeout_rem;
5479 struct io_ring_ctx *ctx = req->ctx;
5482 spin_lock_irq(&ctx->completion_lock);
5483 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5484 ret = io_timeout_cancel(ctx, tr->addr);
5486 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5487 io_translate_timeout_mode(tr->flags));
5489 io_cqring_fill_event(req, ret);
5490 io_commit_cqring(ctx);
5491 spin_unlock_irq(&ctx->completion_lock);
5492 io_cqring_ev_posted(ctx);
5494 req_set_fail_links(req);
5499 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5500 bool is_timeout_link)
5502 struct io_timeout_data *data;
5504 u32 off = READ_ONCE(sqe->off);
5506 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5508 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5510 if (off && is_timeout_link)
5512 flags = READ_ONCE(sqe->timeout_flags);
5513 if (flags & ~IORING_TIMEOUT_ABS)
5516 req->timeout.off = off;
5518 if (!req->async_data && io_alloc_async_data(req))
5521 data = req->async_data;
5524 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5527 data->mode = io_translate_timeout_mode(flags);
5528 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5529 io_req_track_inflight(req);
5533 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5535 struct io_ring_ctx *ctx = req->ctx;
5536 struct io_timeout_data *data = req->async_data;
5537 struct list_head *entry;
5538 u32 tail, off = req->timeout.off;
5540 spin_lock_irq(&ctx->completion_lock);
5543 * sqe->off holds how many events that need to occur for this
5544 * timeout event to be satisfied. If it isn't set, then this is
5545 * a pure timeout request, sequence isn't used.
5547 if (io_is_timeout_noseq(req)) {
5548 entry = ctx->timeout_list.prev;
5552 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5553 req->timeout.target_seq = tail + off;
5555 /* Update the last seq here in case io_flush_timeouts() hasn't.
5556 * This is safe because ->completion_lock is held, and submissions
5557 * and completions are never mixed in the same ->completion_lock section.
5559 ctx->cq_last_tm_flush = tail;
5562 * Insertion sort, ensuring the first entry in the list is always
5563 * the one we need first.
5565 list_for_each_prev(entry, &ctx->timeout_list) {
5566 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5569 if (io_is_timeout_noseq(nxt))
5571 /* nxt.seq is behind @tail, otherwise would've been completed */
5572 if (off >= nxt->timeout.target_seq - tail)
5576 list_add(&req->timeout.list, entry);
5577 data->timer.function = io_timeout_fn;
5578 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5579 spin_unlock_irq(&ctx->completion_lock);
5583 struct io_cancel_data {
5584 struct io_ring_ctx *ctx;
5588 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5590 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5591 struct io_cancel_data *cd = data;
5593 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5596 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5597 struct io_ring_ctx *ctx)
5599 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5600 enum io_wq_cancel cancel_ret;
5603 if (!tctx || !tctx->io_wq)
5606 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5607 switch (cancel_ret) {
5608 case IO_WQ_CANCEL_OK:
5611 case IO_WQ_CANCEL_RUNNING:
5614 case IO_WQ_CANCEL_NOTFOUND:
5622 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5623 struct io_kiocb *req, __u64 sqe_addr,
5626 unsigned long flags;
5629 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5630 if (ret != -ENOENT) {
5631 spin_lock_irqsave(&ctx->completion_lock, flags);
5635 spin_lock_irqsave(&ctx->completion_lock, flags);
5636 ret = io_timeout_cancel(ctx, sqe_addr);
5639 ret = io_poll_cancel(ctx, sqe_addr);
5643 io_cqring_fill_event(req, ret);
5644 io_commit_cqring(ctx);
5645 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5646 io_cqring_ev_posted(ctx);
5649 req_set_fail_links(req);
5653 static int io_async_cancel_prep(struct io_kiocb *req,
5654 const struct io_uring_sqe *sqe)
5656 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5658 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5660 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5663 req->cancel.addr = READ_ONCE(sqe->addr);
5667 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5669 struct io_ring_ctx *ctx = req->ctx;
5671 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5675 static int io_rsrc_update_prep(struct io_kiocb *req,
5676 const struct io_uring_sqe *sqe)
5678 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5680 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5682 if (sqe->ioprio || sqe->rw_flags)
5685 req->rsrc_update.offset = READ_ONCE(sqe->off);
5686 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5687 if (!req->rsrc_update.nr_args)
5689 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5693 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5695 struct io_ring_ctx *ctx = req->ctx;
5696 struct io_uring_rsrc_update up;
5699 if (issue_flags & IO_URING_F_NONBLOCK)
5702 up.offset = req->rsrc_update.offset;
5703 up.data = req->rsrc_update.arg;
5705 mutex_lock(&ctx->uring_lock);
5706 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5707 mutex_unlock(&ctx->uring_lock);
5710 req_set_fail_links(req);
5711 __io_req_complete(req, issue_flags, ret, 0);
5715 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5717 switch (req->opcode) {
5720 case IORING_OP_READV:
5721 case IORING_OP_READ_FIXED:
5722 case IORING_OP_READ:
5723 return io_read_prep(req, sqe);
5724 case IORING_OP_WRITEV:
5725 case IORING_OP_WRITE_FIXED:
5726 case IORING_OP_WRITE:
5727 return io_write_prep(req, sqe);
5728 case IORING_OP_POLL_ADD:
5729 return io_poll_add_prep(req, sqe);
5730 case IORING_OP_POLL_REMOVE:
5731 return io_poll_remove_prep(req, sqe);
5732 case IORING_OP_FSYNC:
5733 return io_fsync_prep(req, sqe);
5734 case IORING_OP_SYNC_FILE_RANGE:
5735 return io_sfr_prep(req, sqe);
5736 case IORING_OP_SENDMSG:
5737 case IORING_OP_SEND:
5738 return io_sendmsg_prep(req, sqe);
5739 case IORING_OP_RECVMSG:
5740 case IORING_OP_RECV:
5741 return io_recvmsg_prep(req, sqe);
5742 case IORING_OP_CONNECT:
5743 return io_connect_prep(req, sqe);
5744 case IORING_OP_TIMEOUT:
5745 return io_timeout_prep(req, sqe, false);
5746 case IORING_OP_TIMEOUT_REMOVE:
5747 return io_timeout_remove_prep(req, sqe);
5748 case IORING_OP_ASYNC_CANCEL:
5749 return io_async_cancel_prep(req, sqe);
5750 case IORING_OP_LINK_TIMEOUT:
5751 return io_timeout_prep(req, sqe, true);
5752 case IORING_OP_ACCEPT:
5753 return io_accept_prep(req, sqe);
5754 case IORING_OP_FALLOCATE:
5755 return io_fallocate_prep(req, sqe);
5756 case IORING_OP_OPENAT:
5757 return io_openat_prep(req, sqe);
5758 case IORING_OP_CLOSE:
5759 return io_close_prep(req, sqe);
5760 case IORING_OP_FILES_UPDATE:
5761 return io_rsrc_update_prep(req, sqe);
5762 case IORING_OP_STATX:
5763 return io_statx_prep(req, sqe);
5764 case IORING_OP_FADVISE:
5765 return io_fadvise_prep(req, sqe);
5766 case IORING_OP_MADVISE:
5767 return io_madvise_prep(req, sqe);
5768 case IORING_OP_OPENAT2:
5769 return io_openat2_prep(req, sqe);
5770 case IORING_OP_EPOLL_CTL:
5771 return io_epoll_ctl_prep(req, sqe);
5772 case IORING_OP_SPLICE:
5773 return io_splice_prep(req, sqe);
5774 case IORING_OP_PROVIDE_BUFFERS:
5775 return io_provide_buffers_prep(req, sqe);
5776 case IORING_OP_REMOVE_BUFFERS:
5777 return io_remove_buffers_prep(req, sqe);
5779 return io_tee_prep(req, sqe);
5780 case IORING_OP_SHUTDOWN:
5781 return io_shutdown_prep(req, sqe);
5782 case IORING_OP_RENAMEAT:
5783 return io_renameat_prep(req, sqe);
5784 case IORING_OP_UNLINKAT:
5785 return io_unlinkat_prep(req, sqe);
5788 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5793 static int io_req_prep_async(struct io_kiocb *req)
5795 switch (req->opcode) {
5796 case IORING_OP_READV:
5797 case IORING_OP_READ_FIXED:
5798 case IORING_OP_READ:
5799 return io_rw_prep_async(req, READ);
5800 case IORING_OP_WRITEV:
5801 case IORING_OP_WRITE_FIXED:
5802 case IORING_OP_WRITE:
5803 return io_rw_prep_async(req, WRITE);
5804 case IORING_OP_SENDMSG:
5805 case IORING_OP_SEND:
5806 return io_sendmsg_prep_async(req);
5807 case IORING_OP_RECVMSG:
5808 case IORING_OP_RECV:
5809 return io_recvmsg_prep_async(req);
5810 case IORING_OP_CONNECT:
5811 return io_connect_prep_async(req);
5816 static int io_req_defer_prep(struct io_kiocb *req)
5818 if (!io_op_defs[req->opcode].needs_async_data)
5820 /* some opcodes init it during the inital prep */
5821 if (req->async_data)
5823 if (__io_alloc_async_data(req))
5825 return io_req_prep_async(req);
5828 static u32 io_get_sequence(struct io_kiocb *req)
5830 struct io_kiocb *pos;
5831 struct io_ring_ctx *ctx = req->ctx;
5832 u32 total_submitted, nr_reqs = 0;
5834 io_for_each_link(pos, req)
5837 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5838 return total_submitted - nr_reqs;
5841 static int io_req_defer(struct io_kiocb *req)
5843 struct io_ring_ctx *ctx = req->ctx;
5844 struct io_defer_entry *de;
5848 /* Still need defer if there is pending req in defer list. */
5849 if (likely(list_empty_careful(&ctx->defer_list) &&
5850 !(req->flags & REQ_F_IO_DRAIN)))
5853 seq = io_get_sequence(req);
5854 /* Still a chance to pass the sequence check */
5855 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5858 ret = io_req_defer_prep(req);
5861 io_prep_async_link(req);
5862 de = kmalloc(sizeof(*de), GFP_KERNEL);
5866 spin_lock_irq(&ctx->completion_lock);
5867 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5868 spin_unlock_irq(&ctx->completion_lock);
5870 io_queue_async_work(req);
5871 return -EIOCBQUEUED;
5874 trace_io_uring_defer(ctx, req, req->user_data);
5877 list_add_tail(&de->list, &ctx->defer_list);
5878 spin_unlock_irq(&ctx->completion_lock);
5879 return -EIOCBQUEUED;
5882 static void __io_clean_op(struct io_kiocb *req)
5884 if (req->flags & REQ_F_BUFFER_SELECTED) {
5885 switch (req->opcode) {
5886 case IORING_OP_READV:
5887 case IORING_OP_READ_FIXED:
5888 case IORING_OP_READ:
5889 kfree((void *)(unsigned long)req->rw.addr);
5891 case IORING_OP_RECVMSG:
5892 case IORING_OP_RECV:
5893 kfree(req->sr_msg.kbuf);
5896 req->flags &= ~REQ_F_BUFFER_SELECTED;
5899 if (req->flags & REQ_F_NEED_CLEANUP) {
5900 switch (req->opcode) {
5901 case IORING_OP_READV:
5902 case IORING_OP_READ_FIXED:
5903 case IORING_OP_READ:
5904 case IORING_OP_WRITEV:
5905 case IORING_OP_WRITE_FIXED:
5906 case IORING_OP_WRITE: {
5907 struct io_async_rw *io = req->async_data;
5909 kfree(io->free_iovec);
5912 case IORING_OP_RECVMSG:
5913 case IORING_OP_SENDMSG: {
5914 struct io_async_msghdr *io = req->async_data;
5916 kfree(io->free_iov);
5919 case IORING_OP_SPLICE:
5921 io_put_file(req, req->splice.file_in,
5922 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5924 case IORING_OP_OPENAT:
5925 case IORING_OP_OPENAT2:
5926 if (req->open.filename)
5927 putname(req->open.filename);
5929 case IORING_OP_RENAMEAT:
5930 putname(req->rename.oldpath);
5931 putname(req->rename.newpath);
5933 case IORING_OP_UNLINKAT:
5934 putname(req->unlink.filename);
5937 req->flags &= ~REQ_F_NEED_CLEANUP;
5941 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5943 struct io_ring_ctx *ctx = req->ctx;
5944 const struct cred *creds = NULL;
5947 if (req->work.creds && req->work.creds != current_cred())
5948 creds = override_creds(req->work.creds);
5950 switch (req->opcode) {
5952 ret = io_nop(req, issue_flags);
5954 case IORING_OP_READV:
5955 case IORING_OP_READ_FIXED:
5956 case IORING_OP_READ:
5957 ret = io_read(req, issue_flags);
5959 case IORING_OP_WRITEV:
5960 case IORING_OP_WRITE_FIXED:
5961 case IORING_OP_WRITE:
5962 ret = io_write(req, issue_flags);
5964 case IORING_OP_FSYNC:
5965 ret = io_fsync(req, issue_flags);
5967 case IORING_OP_POLL_ADD:
5968 ret = io_poll_add(req, issue_flags);
5970 case IORING_OP_POLL_REMOVE:
5971 ret = io_poll_remove(req, issue_flags);
5973 case IORING_OP_SYNC_FILE_RANGE:
5974 ret = io_sync_file_range(req, issue_flags);
5976 case IORING_OP_SENDMSG:
5977 ret = io_sendmsg(req, issue_flags);
5979 case IORING_OP_SEND:
5980 ret = io_send(req, issue_flags);
5982 case IORING_OP_RECVMSG:
5983 ret = io_recvmsg(req, issue_flags);
5985 case IORING_OP_RECV:
5986 ret = io_recv(req, issue_flags);
5988 case IORING_OP_TIMEOUT:
5989 ret = io_timeout(req, issue_flags);
5991 case IORING_OP_TIMEOUT_REMOVE:
5992 ret = io_timeout_remove(req, issue_flags);
5994 case IORING_OP_ACCEPT:
5995 ret = io_accept(req, issue_flags);
5997 case IORING_OP_CONNECT:
5998 ret = io_connect(req, issue_flags);
6000 case IORING_OP_ASYNC_CANCEL:
6001 ret = io_async_cancel(req, issue_flags);
6003 case IORING_OP_FALLOCATE:
6004 ret = io_fallocate(req, issue_flags);
6006 case IORING_OP_OPENAT:
6007 ret = io_openat(req, issue_flags);
6009 case IORING_OP_CLOSE:
6010 ret = io_close(req, issue_flags);
6012 case IORING_OP_FILES_UPDATE:
6013 ret = io_files_update(req, issue_flags);
6015 case IORING_OP_STATX:
6016 ret = io_statx(req, issue_flags);
6018 case IORING_OP_FADVISE:
6019 ret = io_fadvise(req, issue_flags);
6021 case IORING_OP_MADVISE:
6022 ret = io_madvise(req, issue_flags);
6024 case IORING_OP_OPENAT2:
6025 ret = io_openat2(req, issue_flags);
6027 case IORING_OP_EPOLL_CTL:
6028 ret = io_epoll_ctl(req, issue_flags);
6030 case IORING_OP_SPLICE:
6031 ret = io_splice(req, issue_flags);
6033 case IORING_OP_PROVIDE_BUFFERS:
6034 ret = io_provide_buffers(req, issue_flags);
6036 case IORING_OP_REMOVE_BUFFERS:
6037 ret = io_remove_buffers(req, issue_flags);
6040 ret = io_tee(req, issue_flags);
6042 case IORING_OP_SHUTDOWN:
6043 ret = io_shutdown(req, issue_flags);
6045 case IORING_OP_RENAMEAT:
6046 ret = io_renameat(req, issue_flags);
6048 case IORING_OP_UNLINKAT:
6049 ret = io_unlinkat(req, issue_flags);
6057 revert_creds(creds);
6062 /* If the op doesn't have a file, we're not polling for it */
6063 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6064 const bool in_async = io_wq_current_is_worker();
6066 /* workqueue context doesn't hold uring_lock, grab it now */
6068 mutex_lock(&ctx->uring_lock);
6070 io_iopoll_req_issued(req, in_async);
6073 mutex_unlock(&ctx->uring_lock);
6079 static void io_wq_submit_work(struct io_wq_work *work)
6081 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6082 struct io_kiocb *timeout;
6085 timeout = io_prep_linked_timeout(req);
6087 io_queue_linked_timeout(timeout);
6089 if (work->flags & IO_WQ_WORK_CANCEL)
6094 ret = io_issue_sqe(req, 0);
6096 * We can get EAGAIN for polled IO even though we're
6097 * forcing a sync submission from here, since we can't
6098 * wait for request slots on the block side.
6106 /* avoid locking problems by failing it from a clean context */
6108 /* io-wq is going to take one down */
6109 refcount_inc(&req->refs);
6110 io_req_task_queue_fail(req, ret);
6114 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6117 struct fixed_rsrc_table *table;
6119 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6120 return table->files[index & IORING_FILE_TABLE_MASK];
6123 static struct file *io_file_get(struct io_submit_state *state,
6124 struct io_kiocb *req, int fd, bool fixed)
6126 struct io_ring_ctx *ctx = req->ctx;
6130 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6132 fd = array_index_nospec(fd, ctx->nr_user_files);
6133 file = io_file_from_index(ctx, fd);
6134 io_set_resource_node(req);
6136 trace_io_uring_file_get(ctx, fd);
6137 file = __io_file_get(state, fd);
6140 if (file && unlikely(file->f_op == &io_uring_fops))
6141 io_req_track_inflight(req);
6145 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6147 struct io_timeout_data *data = container_of(timer,
6148 struct io_timeout_data, timer);
6149 struct io_kiocb *prev, *req = data->req;
6150 struct io_ring_ctx *ctx = req->ctx;
6151 unsigned long flags;
6153 spin_lock_irqsave(&ctx->completion_lock, flags);
6154 prev = req->timeout.head;
6155 req->timeout.head = NULL;
6158 * We don't expect the list to be empty, that will only happen if we
6159 * race with the completion of the linked work.
6161 if (prev && refcount_inc_not_zero(&prev->refs))
6162 io_remove_next_linked(prev);
6165 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6168 req_set_fail_links(prev);
6169 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6170 io_put_req_deferred(prev, 1);
6172 io_req_complete_post(req, -ETIME, 0);
6173 io_put_req_deferred(req, 1);
6175 return HRTIMER_NORESTART;
6178 static void __io_queue_linked_timeout(struct io_kiocb *req)
6181 * If the back reference is NULL, then our linked request finished
6182 * before we got a chance to setup the timer
6184 if (req->timeout.head) {
6185 struct io_timeout_data *data = req->async_data;
6187 data->timer.function = io_link_timeout_fn;
6188 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6193 static void io_queue_linked_timeout(struct io_kiocb *req)
6195 struct io_ring_ctx *ctx = req->ctx;
6197 spin_lock_irq(&ctx->completion_lock);
6198 __io_queue_linked_timeout(req);
6199 spin_unlock_irq(&ctx->completion_lock);
6201 /* drop submission reference */
6205 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6207 struct io_kiocb *nxt = req->link;
6209 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6210 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6213 nxt->timeout.head = req;
6214 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6215 req->flags |= REQ_F_LINK_TIMEOUT;
6219 static void __io_queue_sqe(struct io_kiocb *req)
6221 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6224 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6227 * We async punt it if the file wasn't marked NOWAIT, or if the file
6228 * doesn't support non-blocking read/write attempts
6230 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6231 if (!io_arm_poll_handler(req)) {
6233 * Queued up for async execution, worker will release
6234 * submit reference when the iocb is actually submitted.
6236 io_queue_async_work(req);
6238 } else if (likely(!ret)) {
6239 /* drop submission reference */
6240 if (req->flags & REQ_F_COMPLETE_INLINE) {
6241 struct io_ring_ctx *ctx = req->ctx;
6242 struct io_comp_state *cs = &ctx->submit_state.comp;
6244 cs->reqs[cs->nr++] = req;
6245 if (cs->nr == ARRAY_SIZE(cs->reqs))
6246 io_submit_flush_completions(cs, ctx);
6251 req_set_fail_links(req);
6253 io_req_complete(req, ret);
6256 io_queue_linked_timeout(linked_timeout);
6259 static void io_queue_sqe(struct io_kiocb *req)
6263 ret = io_req_defer(req);
6265 if (ret != -EIOCBQUEUED) {
6267 req_set_fail_links(req);
6269 io_req_complete(req, ret);
6271 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6272 ret = io_req_defer_prep(req);
6275 io_queue_async_work(req);
6277 __io_queue_sqe(req);
6282 * Check SQE restrictions (opcode and flags).
6284 * Returns 'true' if SQE is allowed, 'false' otherwise.
6286 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6287 struct io_kiocb *req,
6288 unsigned int sqe_flags)
6290 if (!ctx->restricted)
6293 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6296 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6297 ctx->restrictions.sqe_flags_required)
6300 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6301 ctx->restrictions.sqe_flags_required))
6307 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6308 const struct io_uring_sqe *sqe)
6310 struct io_submit_state *state;
6311 unsigned int sqe_flags;
6312 int personality, ret = 0;
6314 req->opcode = READ_ONCE(sqe->opcode);
6315 /* same numerical values with corresponding REQ_F_*, safe to copy */
6316 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6317 req->user_data = READ_ONCE(sqe->user_data);
6318 req->async_data = NULL;
6322 req->fixed_rsrc_refs = NULL;
6323 /* one is dropped after submission, the other at completion */
6324 refcount_set(&req->refs, 2);
6325 req->task = current;
6327 req->work.list.next = NULL;
6328 req->work.creds = NULL;
6329 req->work.flags = 0;
6331 /* enforce forwards compatibility on users */
6332 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6337 if (unlikely(req->opcode >= IORING_OP_LAST))
6340 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6343 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6344 !io_op_defs[req->opcode].buffer_select)
6347 personality = READ_ONCE(sqe->personality);
6349 req->work.creds = xa_load(&ctx->personalities, personality);
6350 if (!req->work.creds)
6352 get_cred(req->work.creds);
6354 state = &ctx->submit_state;
6357 * Plug now if we have more than 1 IO left after this, and the target
6358 * is potentially a read/write to block based storage.
6360 if (!state->plug_started && state->ios_left > 1 &&
6361 io_op_defs[req->opcode].plug) {
6362 blk_start_plug(&state->plug);
6363 state->plug_started = true;
6366 if (io_op_defs[req->opcode].needs_file) {
6367 bool fixed = req->flags & REQ_F_FIXED_FILE;
6369 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6370 if (unlikely(!req->file))
6378 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6379 const struct io_uring_sqe *sqe)
6381 struct io_submit_link *link = &ctx->submit_state.link;
6384 ret = io_init_req(ctx, req, sqe);
6385 if (unlikely(ret)) {
6388 io_req_complete(req, ret);
6390 /* fail even hard links since we don't submit */
6391 link->head->flags |= REQ_F_FAIL_LINK;
6392 io_put_req(link->head);
6393 io_req_complete(link->head, -ECANCELED);
6398 ret = io_req_prep(req, sqe);
6402 /* don't need @sqe from now on */
6403 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6404 true, ctx->flags & IORING_SETUP_SQPOLL);
6407 * If we already have a head request, queue this one for async
6408 * submittal once the head completes. If we don't have a head but
6409 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6410 * submitted sync once the chain is complete. If none of those
6411 * conditions are true (normal request), then just queue it.
6414 struct io_kiocb *head = link->head;
6417 * Taking sequential execution of a link, draining both sides
6418 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6419 * requests in the link. So, it drains the head and the
6420 * next after the link request. The last one is done via
6421 * drain_next flag to persist the effect across calls.
6423 if (req->flags & REQ_F_IO_DRAIN) {
6424 head->flags |= REQ_F_IO_DRAIN;
6425 ctx->drain_next = 1;
6427 ret = io_req_defer_prep(req);
6430 trace_io_uring_link(ctx, req, head);
6431 link->last->link = req;
6434 /* last request of a link, enqueue the link */
6435 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6440 if (unlikely(ctx->drain_next)) {
6441 req->flags |= REQ_F_IO_DRAIN;
6442 ctx->drain_next = 0;
6444 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6456 * Batched submission is done, ensure local IO is flushed out.
6458 static void io_submit_state_end(struct io_submit_state *state,
6459 struct io_ring_ctx *ctx)
6461 if (state->link.head)
6462 io_queue_sqe(state->link.head);
6464 io_submit_flush_completions(&state->comp, ctx);
6465 if (state->plug_started)
6466 blk_finish_plug(&state->plug);
6467 io_state_file_put(state);
6471 * Start submission side cache.
6473 static void io_submit_state_start(struct io_submit_state *state,
6474 unsigned int max_ios)
6476 state->plug_started = false;
6477 state->ios_left = max_ios;
6478 /* set only head, no need to init link_last in advance */
6479 state->link.head = NULL;
6482 static void io_commit_sqring(struct io_ring_ctx *ctx)
6484 struct io_rings *rings = ctx->rings;
6487 * Ensure any loads from the SQEs are done at this point,
6488 * since once we write the new head, the application could
6489 * write new data to them.
6491 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6495 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6496 * that is mapped by userspace. This means that care needs to be taken to
6497 * ensure that reads are stable, as we cannot rely on userspace always
6498 * being a good citizen. If members of the sqe are validated and then later
6499 * used, it's important that those reads are done through READ_ONCE() to
6500 * prevent a re-load down the line.
6502 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6504 u32 *sq_array = ctx->sq_array;
6508 * The cached sq head (or cq tail) serves two purposes:
6510 * 1) allows us to batch the cost of updating the user visible
6512 * 2) allows the kernel side to track the head on its own, even
6513 * though the application is the one updating it.
6515 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6516 if (likely(head < ctx->sq_entries))
6517 return &ctx->sq_sqes[head];
6519 /* drop invalid entries */
6520 ctx->cached_sq_dropped++;
6521 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6525 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6529 /* if we have a backlog and couldn't flush it all, return BUSY */
6530 if (test_bit(0, &ctx->sq_check_overflow)) {
6531 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6535 /* make sure SQ entry isn't read before tail */
6536 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6538 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6541 percpu_counter_add(¤t->io_uring->inflight, nr);
6542 refcount_add(nr, ¤t->usage);
6543 io_submit_state_start(&ctx->submit_state, nr);
6545 while (submitted < nr) {
6546 const struct io_uring_sqe *sqe;
6547 struct io_kiocb *req;
6549 req = io_alloc_req(ctx);
6550 if (unlikely(!req)) {
6552 submitted = -EAGAIN;
6555 sqe = io_get_sqe(ctx);
6556 if (unlikely(!sqe)) {
6557 kmem_cache_free(req_cachep, req);
6560 /* will complete beyond this point, count as submitted */
6562 if (io_submit_sqe(ctx, req, sqe))
6566 if (unlikely(submitted != nr)) {
6567 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6568 struct io_uring_task *tctx = current->io_uring;
6569 int unused = nr - ref_used;
6571 percpu_ref_put_many(&ctx->refs, unused);
6572 percpu_counter_sub(&tctx->inflight, unused);
6573 put_task_struct_many(current, unused);
6576 io_submit_state_end(&ctx->submit_state, ctx);
6577 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6578 io_commit_sqring(ctx);
6583 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6585 /* Tell userspace we may need a wakeup call */
6586 spin_lock_irq(&ctx->completion_lock);
6587 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6588 spin_unlock_irq(&ctx->completion_lock);
6591 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6593 spin_lock_irq(&ctx->completion_lock);
6594 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6595 spin_unlock_irq(&ctx->completion_lock);
6598 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6600 unsigned int to_submit;
6603 to_submit = io_sqring_entries(ctx);
6604 /* if we're handling multiple rings, cap submit size for fairness */
6605 if (cap_entries && to_submit > 8)
6608 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6609 unsigned nr_events = 0;
6611 mutex_lock(&ctx->uring_lock);
6612 if (!list_empty(&ctx->iopoll_list))
6613 io_do_iopoll(ctx, &nr_events, 0);
6615 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6616 !(ctx->flags & IORING_SETUP_R_DISABLED))
6617 ret = io_submit_sqes(ctx, to_submit);
6618 mutex_unlock(&ctx->uring_lock);
6621 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6622 wake_up(&ctx->sqo_sq_wait);
6627 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6629 struct io_ring_ctx *ctx;
6630 unsigned sq_thread_idle = 0;
6632 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6633 if (sq_thread_idle < ctx->sq_thread_idle)
6634 sq_thread_idle = ctx->sq_thread_idle;
6637 sqd->sq_thread_idle = sq_thread_idle;
6640 static int io_sq_thread(void *data)
6642 struct io_sq_data *sqd = data;
6643 struct io_ring_ctx *ctx;
6644 unsigned long timeout = 0;
6645 char buf[TASK_COMM_LEN];
6648 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6649 set_task_comm(current, buf);
6650 current->pf_io_worker = NULL;
6652 if (sqd->sq_cpu != -1)
6653 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6655 set_cpus_allowed_ptr(current, cpu_online_mask);
6656 current->flags |= PF_NO_SETAFFINITY;
6658 down_read(&sqd->rw_lock);
6660 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6662 bool cap_entries, sqt_spin, needs_sched;
6664 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6665 up_read(&sqd->rw_lock);
6667 down_read(&sqd->rw_lock);
6668 timeout = jiffies + sqd->sq_thread_idle;
6671 if (fatal_signal_pending(current))
6674 cap_entries = !list_is_singular(&sqd->ctx_list);
6675 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6676 const struct cred *creds = NULL;
6678 if (ctx->sq_creds != current_cred())
6679 creds = override_creds(ctx->sq_creds);
6680 ret = __io_sq_thread(ctx, cap_entries);
6682 revert_creds(creds);
6683 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6687 if (sqt_spin || !time_after(jiffies, timeout)) {
6691 timeout = jiffies + sqd->sq_thread_idle;
6696 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6697 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6698 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6699 !list_empty_careful(&ctx->iopoll_list)) {
6700 needs_sched = false;
6703 if (io_sqring_entries(ctx)) {
6704 needs_sched = false;
6709 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6710 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6711 io_ring_set_wakeup_flag(ctx);
6713 up_read(&sqd->rw_lock);
6715 down_read(&sqd->rw_lock);
6716 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6717 io_ring_clear_wakeup_flag(ctx);
6720 finish_wait(&sqd->wait, &wait);
6721 timeout = jiffies + sqd->sq_thread_idle;
6724 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6725 io_uring_cancel_sqpoll(ctx);
6726 up_read(&sqd->rw_lock);
6730 down_write(&sqd->rw_lock);
6732 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6733 io_ring_set_wakeup_flag(ctx);
6734 up_write(&sqd->rw_lock);
6735 complete(&sqd->exited);
6739 struct io_wait_queue {
6740 struct wait_queue_entry wq;
6741 struct io_ring_ctx *ctx;
6743 unsigned nr_timeouts;
6746 static inline bool io_should_wake(struct io_wait_queue *iowq)
6748 struct io_ring_ctx *ctx = iowq->ctx;
6751 * Wake up if we have enough events, or if a timeout occurred since we
6752 * started waiting. For timeouts, we always want to return to userspace,
6753 * regardless of event count.
6755 return io_cqring_events(ctx) >= iowq->to_wait ||
6756 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6759 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6760 int wake_flags, void *key)
6762 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6766 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6767 * the task, and the next invocation will do it.
6769 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6770 return autoremove_wake_function(curr, mode, wake_flags, key);
6774 static int io_run_task_work_sig(void)
6776 if (io_run_task_work())
6778 if (!signal_pending(current))
6780 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6781 return -ERESTARTSYS;
6785 /* when returns >0, the caller should retry */
6786 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6787 struct io_wait_queue *iowq,
6788 signed long *timeout)
6792 /* make sure we run task_work before checking for signals */
6793 ret = io_run_task_work_sig();
6794 if (ret || io_should_wake(iowq))
6796 /* let the caller flush overflows, retry */
6797 if (test_bit(0, &ctx->cq_check_overflow))
6800 *timeout = schedule_timeout(*timeout);
6801 return !*timeout ? -ETIME : 1;
6805 * Wait until events become available, if we don't already have some. The
6806 * application must reap them itself, as they reside on the shared cq ring.
6808 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6809 const sigset_t __user *sig, size_t sigsz,
6810 struct __kernel_timespec __user *uts)
6812 struct io_wait_queue iowq = {
6815 .func = io_wake_function,
6816 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6819 .to_wait = min_events,
6821 struct io_rings *rings = ctx->rings;
6822 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6826 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6827 if (io_cqring_events(ctx) >= min_events)
6829 if (!io_run_task_work())
6834 #ifdef CONFIG_COMPAT
6835 if (in_compat_syscall())
6836 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6840 ret = set_user_sigmask(sig, sigsz);
6847 struct timespec64 ts;
6849 if (get_timespec64(&ts, uts))
6851 timeout = timespec64_to_jiffies(&ts);
6854 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6855 trace_io_uring_cqring_wait(ctx, min_events);
6857 /* if we can't even flush overflow, don't wait for more */
6858 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6862 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6863 TASK_INTERRUPTIBLE);
6864 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6865 finish_wait(&ctx->wait, &iowq.wq);
6869 restore_saved_sigmask_unless(ret == -EINTR);
6871 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6874 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6876 #if defined(CONFIG_UNIX)
6877 if (ctx->ring_sock) {
6878 struct sock *sock = ctx->ring_sock->sk;
6879 struct sk_buff *skb;
6881 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6887 for (i = 0; i < ctx->nr_user_files; i++) {
6890 file = io_file_from_index(ctx, i);
6897 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6899 struct fixed_rsrc_data *data;
6901 data = container_of(ref, struct fixed_rsrc_data, refs);
6902 complete(&data->done);
6905 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6907 spin_lock_bh(&ctx->rsrc_ref_lock);
6910 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6912 spin_unlock_bh(&ctx->rsrc_ref_lock);
6915 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6916 struct fixed_rsrc_data *rsrc_data,
6917 struct fixed_rsrc_ref_node *ref_node)
6919 io_rsrc_ref_lock(ctx);
6920 rsrc_data->node = ref_node;
6921 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6922 io_rsrc_ref_unlock(ctx);
6923 percpu_ref_get(&rsrc_data->refs);
6926 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6928 struct fixed_rsrc_ref_node *ref_node = NULL;
6930 io_rsrc_ref_lock(ctx);
6931 ref_node = data->node;
6933 io_rsrc_ref_unlock(ctx);
6935 percpu_ref_kill(&ref_node->refs);
6938 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6939 struct io_ring_ctx *ctx,
6940 void (*rsrc_put)(struct io_ring_ctx *ctx,
6941 struct io_rsrc_put *prsrc))
6943 struct fixed_rsrc_ref_node *backup_node;
6949 data->quiesce = true;
6952 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6955 backup_node->rsrc_data = data;
6956 backup_node->rsrc_put = rsrc_put;
6958 io_sqe_rsrc_kill_node(ctx, data);
6959 percpu_ref_kill(&data->refs);
6960 flush_delayed_work(&ctx->rsrc_put_work);
6962 ret = wait_for_completion_interruptible(&data->done);
6966 percpu_ref_resurrect(&data->refs);
6967 io_sqe_rsrc_set_node(ctx, data, backup_node);
6969 reinit_completion(&data->done);
6970 mutex_unlock(&ctx->uring_lock);
6971 ret = io_run_task_work_sig();
6972 mutex_lock(&ctx->uring_lock);
6974 data->quiesce = false;
6977 destroy_fixed_rsrc_ref_node(backup_node);
6981 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
6983 struct fixed_rsrc_data *data;
6985 data = kzalloc(sizeof(*data), GFP_KERNEL);
6989 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
6990 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6995 init_completion(&data->done);
6999 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7001 percpu_ref_exit(&data->refs);
7006 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7008 struct fixed_rsrc_data *data = ctx->file_data;
7009 unsigned nr_tables, i;
7013 * percpu_ref_is_dying() is to stop parallel files unregister
7014 * Since we possibly drop uring lock later in this function to
7017 if (!data || percpu_ref_is_dying(&data->refs))
7019 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7023 __io_sqe_files_unregister(ctx);
7024 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7025 for (i = 0; i < nr_tables; i++)
7026 kfree(data->table[i].files);
7027 free_fixed_rsrc_data(data);
7028 ctx->file_data = NULL;
7029 ctx->nr_user_files = 0;
7033 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7034 __releases(&sqd->rw_lock)
7036 if (sqd->thread == current)
7038 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7039 up_write(&sqd->rw_lock);
7042 static void io_sq_thread_park(struct io_sq_data *sqd)
7043 __acquires(&sqd->rw_lock)
7045 if (sqd->thread == current)
7047 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7048 down_write(&sqd->rw_lock);
7049 /* set again for consistency, in case concurrent parks are happening */
7050 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7052 wake_up_process(sqd->thread);
7055 static void io_sq_thread_stop(struct io_sq_data *sqd)
7057 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7059 down_write(&sqd->rw_lock);
7060 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7062 wake_up_process(sqd->thread);
7063 up_write(&sqd->rw_lock);
7064 wait_for_completion(&sqd->exited);
7067 static void io_put_sq_data(struct io_sq_data *sqd)
7069 if (refcount_dec_and_test(&sqd->refs)) {
7070 io_sq_thread_stop(sqd);
7075 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7077 struct io_sq_data *sqd = ctx->sq_data;
7080 io_sq_thread_park(sqd);
7081 list_del(&ctx->sqd_list);
7082 io_sqd_update_thread_idle(sqd);
7083 io_sq_thread_unpark(sqd);
7085 io_put_sq_data(sqd);
7086 ctx->sq_data = NULL;
7088 put_cred(ctx->sq_creds);
7092 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7094 struct io_ring_ctx *ctx_attach;
7095 struct io_sq_data *sqd;
7098 f = fdget(p->wq_fd);
7100 return ERR_PTR(-ENXIO);
7101 if (f.file->f_op != &io_uring_fops) {
7103 return ERR_PTR(-EINVAL);
7106 ctx_attach = f.file->private_data;
7107 sqd = ctx_attach->sq_data;
7110 return ERR_PTR(-EINVAL);
7112 if (sqd->task_tgid != current->tgid) {
7114 return ERR_PTR(-EPERM);
7117 refcount_inc(&sqd->refs);
7122 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7125 struct io_sq_data *sqd;
7128 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7129 sqd = io_attach_sq_data(p);
7134 /* fall through for EPERM case, setup new sqd/task */
7135 if (PTR_ERR(sqd) != -EPERM)
7139 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7141 return ERR_PTR(-ENOMEM);
7143 refcount_set(&sqd->refs, 1);
7144 INIT_LIST_HEAD(&sqd->ctx_list);
7145 init_rwsem(&sqd->rw_lock);
7146 init_waitqueue_head(&sqd->wait);
7147 init_completion(&sqd->exited);
7151 #if defined(CONFIG_UNIX)
7153 * Ensure the UNIX gc is aware of our file set, so we are certain that
7154 * the io_uring can be safely unregistered on process exit, even if we have
7155 * loops in the file referencing.
7157 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7159 struct sock *sk = ctx->ring_sock->sk;
7160 struct scm_fp_list *fpl;
7161 struct sk_buff *skb;
7164 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7168 skb = alloc_skb(0, GFP_KERNEL);
7177 fpl->user = get_uid(current_user());
7178 for (i = 0; i < nr; i++) {
7179 struct file *file = io_file_from_index(ctx, i + offset);
7183 fpl->fp[nr_files] = get_file(file);
7184 unix_inflight(fpl->user, fpl->fp[nr_files]);
7189 fpl->max = SCM_MAX_FD;
7190 fpl->count = nr_files;
7191 UNIXCB(skb).fp = fpl;
7192 skb->destructor = unix_destruct_scm;
7193 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7194 skb_queue_head(&sk->sk_receive_queue, skb);
7196 for (i = 0; i < nr_files; i++)
7207 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7208 * causes regular reference counting to break down. We rely on the UNIX
7209 * garbage collection to take care of this problem for us.
7211 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7213 unsigned left, total;
7217 left = ctx->nr_user_files;
7219 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7221 ret = __io_sqe_files_scm(ctx, this_files, total);
7225 total += this_files;
7231 while (total < ctx->nr_user_files) {
7232 struct file *file = io_file_from_index(ctx, total);
7242 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7248 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7249 unsigned nr_tables, unsigned nr_files)
7253 for (i = 0; i < nr_tables; i++) {
7254 struct fixed_rsrc_table *table = &file_data->table[i];
7255 unsigned this_files;
7257 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7258 table->files = kcalloc(this_files, sizeof(struct file *),
7262 nr_files -= this_files;
7268 for (i = 0; i < nr_tables; i++) {
7269 struct fixed_rsrc_table *table = &file_data->table[i];
7270 kfree(table->files);
7275 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7277 struct file *file = prsrc->file;
7278 #if defined(CONFIG_UNIX)
7279 struct sock *sock = ctx->ring_sock->sk;
7280 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7281 struct sk_buff *skb;
7284 __skb_queue_head_init(&list);
7287 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7288 * remove this entry and rearrange the file array.
7290 skb = skb_dequeue(head);
7292 struct scm_fp_list *fp;
7294 fp = UNIXCB(skb).fp;
7295 for (i = 0; i < fp->count; i++) {
7298 if (fp->fp[i] != file)
7301 unix_notinflight(fp->user, fp->fp[i]);
7302 left = fp->count - 1 - i;
7304 memmove(&fp->fp[i], &fp->fp[i + 1],
7305 left * sizeof(struct file *));
7312 __skb_queue_tail(&list, skb);
7322 __skb_queue_tail(&list, skb);
7324 skb = skb_dequeue(head);
7327 if (skb_peek(&list)) {
7328 spin_lock_irq(&head->lock);
7329 while ((skb = __skb_dequeue(&list)) != NULL)
7330 __skb_queue_tail(head, skb);
7331 spin_unlock_irq(&head->lock);
7338 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7340 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7341 struct io_ring_ctx *ctx = rsrc_data->ctx;
7342 struct io_rsrc_put *prsrc, *tmp;
7344 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7345 list_del(&prsrc->list);
7346 ref_node->rsrc_put(ctx, prsrc);
7350 percpu_ref_exit(&ref_node->refs);
7352 percpu_ref_put(&rsrc_data->refs);
7355 static void io_rsrc_put_work(struct work_struct *work)
7357 struct io_ring_ctx *ctx;
7358 struct llist_node *node;
7360 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7361 node = llist_del_all(&ctx->rsrc_put_llist);
7364 struct fixed_rsrc_ref_node *ref_node;
7365 struct llist_node *next = node->next;
7367 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7368 __io_rsrc_put_work(ref_node);
7373 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7376 struct fixed_rsrc_table *table;
7378 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7379 return &table->files[i & IORING_FILE_TABLE_MASK];
7382 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7384 struct fixed_rsrc_ref_node *ref_node;
7385 struct fixed_rsrc_data *data;
7386 struct io_ring_ctx *ctx;
7387 bool first_add = false;
7390 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7391 data = ref_node->rsrc_data;
7394 io_rsrc_ref_lock(ctx);
7395 ref_node->done = true;
7397 while (!list_empty(&ctx->rsrc_ref_list)) {
7398 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7399 struct fixed_rsrc_ref_node, node);
7400 /* recycle ref nodes in order */
7401 if (!ref_node->done)
7403 list_del(&ref_node->node);
7404 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7406 io_rsrc_ref_unlock(ctx);
7408 if (percpu_ref_is_dying(&data->refs))
7412 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7414 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7417 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7418 struct io_ring_ctx *ctx)
7420 struct fixed_rsrc_ref_node *ref_node;
7422 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7426 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7431 INIT_LIST_HEAD(&ref_node->node);
7432 INIT_LIST_HEAD(&ref_node->rsrc_list);
7433 ref_node->done = false;
7437 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7438 struct fixed_rsrc_ref_node *ref_node)
7440 ref_node->rsrc_data = ctx->file_data;
7441 ref_node->rsrc_put = io_ring_file_put;
7444 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7446 percpu_ref_exit(&ref_node->refs);
7451 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7454 __s32 __user *fds = (__s32 __user *) arg;
7455 unsigned nr_tables, i;
7457 int fd, ret = -ENOMEM;
7458 struct fixed_rsrc_ref_node *ref_node;
7459 struct fixed_rsrc_data *file_data;
7465 if (nr_args > IORING_MAX_FIXED_FILES)
7468 file_data = alloc_fixed_rsrc_data(ctx);
7471 ctx->file_data = file_data;
7473 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7474 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7476 if (!file_data->table)
7479 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7482 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7483 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7487 /* allow sparse sets */
7497 * Don't allow io_uring instances to be registered. If UNIX
7498 * isn't enabled, then this causes a reference cycle and this
7499 * instance can never get freed. If UNIX is enabled we'll
7500 * handle it just fine, but there's still no point in allowing
7501 * a ring fd as it doesn't support regular read/write anyway.
7503 if (file->f_op == &io_uring_fops) {
7507 *io_fixed_file_slot(file_data, i) = file;
7510 ret = io_sqe_files_scm(ctx);
7512 io_sqe_files_unregister(ctx);
7516 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7518 io_sqe_files_unregister(ctx);
7521 init_fixed_file_ref_node(ctx, ref_node);
7523 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7526 for (i = 0; i < ctx->nr_user_files; i++) {
7527 file = io_file_from_index(ctx, i);
7531 for (i = 0; i < nr_tables; i++)
7532 kfree(file_data->table[i].files);
7533 ctx->nr_user_files = 0;
7535 free_fixed_rsrc_data(ctx->file_data);
7536 ctx->file_data = NULL;
7540 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7543 #if defined(CONFIG_UNIX)
7544 struct sock *sock = ctx->ring_sock->sk;
7545 struct sk_buff_head *head = &sock->sk_receive_queue;
7546 struct sk_buff *skb;
7549 * See if we can merge this file into an existing skb SCM_RIGHTS
7550 * file set. If there's no room, fall back to allocating a new skb
7551 * and filling it in.
7553 spin_lock_irq(&head->lock);
7554 skb = skb_peek(head);
7556 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7558 if (fpl->count < SCM_MAX_FD) {
7559 __skb_unlink(skb, head);
7560 spin_unlock_irq(&head->lock);
7561 fpl->fp[fpl->count] = get_file(file);
7562 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7564 spin_lock_irq(&head->lock);
7565 __skb_queue_head(head, skb);
7570 spin_unlock_irq(&head->lock);
7577 return __io_sqe_files_scm(ctx, 1, index);
7583 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7585 struct io_rsrc_put *prsrc;
7586 struct fixed_rsrc_ref_node *ref_node = data->node;
7588 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7593 list_add(&prsrc->list, &ref_node->rsrc_list);
7598 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7601 return io_queue_rsrc_removal(data, (void *)file);
7604 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7605 struct io_uring_rsrc_update *up,
7608 struct fixed_rsrc_data *data = ctx->file_data;
7609 struct fixed_rsrc_ref_node *ref_node;
7610 struct file *file, **file_slot;
7614 bool needs_switch = false;
7616 if (check_add_overflow(up->offset, nr_args, &done))
7618 if (done > ctx->nr_user_files)
7621 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7624 init_fixed_file_ref_node(ctx, ref_node);
7626 fds = u64_to_user_ptr(up->data);
7627 for (done = 0; done < nr_args; done++) {
7629 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7633 if (fd == IORING_REGISTER_FILES_SKIP)
7636 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7637 file_slot = io_fixed_file_slot(ctx->file_data, i);
7640 err = io_queue_file_removal(data, *file_slot);
7644 needs_switch = true;
7653 * Don't allow io_uring instances to be registered. If
7654 * UNIX isn't enabled, then this causes a reference
7655 * cycle and this instance can never get freed. If UNIX
7656 * is enabled we'll handle it just fine, but there's
7657 * still no point in allowing a ring fd as it doesn't
7658 * support regular read/write anyway.
7660 if (file->f_op == &io_uring_fops) {
7666 err = io_sqe_file_register(ctx, file, i);
7676 percpu_ref_kill(&data->node->refs);
7677 io_sqe_rsrc_set_node(ctx, data, ref_node);
7679 destroy_fixed_rsrc_ref_node(ref_node);
7681 return done ? done : err;
7684 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7687 struct io_uring_rsrc_update up;
7689 if (!ctx->file_data)
7693 if (copy_from_user(&up, arg, sizeof(up)))
7698 return __io_sqe_files_update(ctx, &up, nr_args);
7701 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7703 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7705 req = io_put_req_find_next(req);
7706 return req ? &req->work : NULL;
7709 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7711 struct io_wq_hash *hash;
7712 struct io_wq_data data;
7713 unsigned int concurrency;
7715 hash = ctx->hash_map;
7717 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7719 return ERR_PTR(-ENOMEM);
7720 refcount_set(&hash->refs, 1);
7721 init_waitqueue_head(&hash->wait);
7722 ctx->hash_map = hash;
7726 data.free_work = io_free_work;
7727 data.do_work = io_wq_submit_work;
7729 /* Do QD, or 4 * CPUS, whatever is smallest */
7730 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7732 return io_wq_create(concurrency, &data);
7735 static int io_uring_alloc_task_context(struct task_struct *task,
7736 struct io_ring_ctx *ctx)
7738 struct io_uring_task *tctx;
7741 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7742 if (unlikely(!tctx))
7745 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7746 if (unlikely(ret)) {
7751 tctx->io_wq = io_init_wq_offload(ctx);
7752 if (IS_ERR(tctx->io_wq)) {
7753 ret = PTR_ERR(tctx->io_wq);
7754 percpu_counter_destroy(&tctx->inflight);
7760 init_waitqueue_head(&tctx->wait);
7762 atomic_set(&tctx->in_idle, 0);
7763 tctx->sqpoll = false;
7764 task->io_uring = tctx;
7765 spin_lock_init(&tctx->task_lock);
7766 INIT_WQ_LIST(&tctx->task_list);
7767 tctx->task_state = 0;
7768 init_task_work(&tctx->task_work, tctx_task_work);
7772 void __io_uring_free(struct task_struct *tsk)
7774 struct io_uring_task *tctx = tsk->io_uring;
7776 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7777 WARN_ON_ONCE(tctx->io_wq);
7779 percpu_counter_destroy(&tctx->inflight);
7781 tsk->io_uring = NULL;
7784 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7785 struct io_uring_params *p)
7789 /* Retain compatibility with failing for an invalid attach attempt */
7790 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7791 IORING_SETUP_ATTACH_WQ) {
7794 f = fdget(p->wq_fd);
7797 if (f.file->f_op != &io_uring_fops) {
7803 if (ctx->flags & IORING_SETUP_SQPOLL) {
7804 struct task_struct *tsk;
7805 struct io_sq_data *sqd;
7809 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7812 sqd = io_get_sq_data(p, &attached);
7818 ctx->sq_creds = get_current_cred();
7820 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7821 if (!ctx->sq_thread_idle)
7822 ctx->sq_thread_idle = HZ;
7825 io_sq_thread_park(sqd);
7826 /* don't attach to a dying SQPOLL thread, would be racy */
7827 if (attached && !sqd->thread) {
7830 list_add(&ctx->sqd_list, &sqd->ctx_list);
7831 io_sqd_update_thread_idle(sqd);
7833 io_sq_thread_unpark(sqd);
7836 io_put_sq_data(sqd);
7837 ctx->sq_data = NULL;
7839 } else if (attached) {
7843 if (p->flags & IORING_SETUP_SQ_AFF) {
7844 int cpu = p->sq_thread_cpu;
7847 if (cpu >= nr_cpu_ids)
7849 if (!cpu_online(cpu))
7857 sqd->task_pid = current->pid;
7858 sqd->task_tgid = current->tgid;
7859 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7866 ret = io_uring_alloc_task_context(tsk, ctx);
7867 wake_up_new_task(tsk);
7870 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7871 /* Can't have SQ_AFF without SQPOLL */
7878 io_sq_thread_finish(ctx);
7881 complete(&ctx->sq_data->exited);
7885 static inline void __io_unaccount_mem(struct user_struct *user,
7886 unsigned long nr_pages)
7888 atomic_long_sub(nr_pages, &user->locked_vm);
7891 static inline int __io_account_mem(struct user_struct *user,
7892 unsigned long nr_pages)
7894 unsigned long page_limit, cur_pages, new_pages;
7896 /* Don't allow more pages than we can safely lock */
7897 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7900 cur_pages = atomic_long_read(&user->locked_vm);
7901 new_pages = cur_pages + nr_pages;
7902 if (new_pages > page_limit)
7904 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7905 new_pages) != cur_pages);
7910 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7913 __io_unaccount_mem(ctx->user, nr_pages);
7915 if (ctx->mm_account)
7916 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7919 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7924 ret = __io_account_mem(ctx->user, nr_pages);
7929 if (ctx->mm_account)
7930 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7935 static void io_mem_free(void *ptr)
7942 page = virt_to_head_page(ptr);
7943 if (put_page_testzero(page))
7944 free_compound_page(page);
7947 static void *io_mem_alloc(size_t size)
7949 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7950 __GFP_NORETRY | __GFP_ACCOUNT;
7952 return (void *) __get_free_pages(gfp_flags, get_order(size));
7955 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7958 struct io_rings *rings;
7959 size_t off, sq_array_size;
7961 off = struct_size(rings, cqes, cq_entries);
7962 if (off == SIZE_MAX)
7966 off = ALIGN(off, SMP_CACHE_BYTES);
7974 sq_array_size = array_size(sizeof(u32), sq_entries);
7975 if (sq_array_size == SIZE_MAX)
7978 if (check_add_overflow(off, sq_array_size, &off))
7984 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
7988 if (!ctx->user_bufs)
7991 for (i = 0; i < ctx->nr_user_bufs; i++) {
7992 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7994 for (j = 0; j < imu->nr_bvecs; j++)
7995 unpin_user_page(imu->bvec[j].bv_page);
7997 if (imu->acct_pages)
7998 io_unaccount_mem(ctx, imu->acct_pages);
8003 kfree(ctx->user_bufs);
8004 ctx->user_bufs = NULL;
8005 ctx->nr_user_bufs = 0;
8009 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8010 void __user *arg, unsigned index)
8012 struct iovec __user *src;
8014 #ifdef CONFIG_COMPAT
8016 struct compat_iovec __user *ciovs;
8017 struct compat_iovec ciov;
8019 ciovs = (struct compat_iovec __user *) arg;
8020 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8023 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8024 dst->iov_len = ciov.iov_len;
8028 src = (struct iovec __user *) arg;
8029 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8035 * Not super efficient, but this is just a registration time. And we do cache
8036 * the last compound head, so generally we'll only do a full search if we don't
8039 * We check if the given compound head page has already been accounted, to
8040 * avoid double accounting it. This allows us to account the full size of the
8041 * page, not just the constituent pages of a huge page.
8043 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8044 int nr_pages, struct page *hpage)
8048 /* check current page array */
8049 for (i = 0; i < nr_pages; i++) {
8050 if (!PageCompound(pages[i]))
8052 if (compound_head(pages[i]) == hpage)
8056 /* check previously registered pages */
8057 for (i = 0; i < ctx->nr_user_bufs; i++) {
8058 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8060 for (j = 0; j < imu->nr_bvecs; j++) {
8061 if (!PageCompound(imu->bvec[j].bv_page))
8063 if (compound_head(imu->bvec[j].bv_page) == hpage)
8071 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8072 int nr_pages, struct io_mapped_ubuf *imu,
8073 struct page **last_hpage)
8077 for (i = 0; i < nr_pages; i++) {
8078 if (!PageCompound(pages[i])) {
8083 hpage = compound_head(pages[i]);
8084 if (hpage == *last_hpage)
8086 *last_hpage = hpage;
8087 if (headpage_already_acct(ctx, pages, i, hpage))
8089 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8093 if (!imu->acct_pages)
8096 ret = io_account_mem(ctx, imu->acct_pages);
8098 imu->acct_pages = 0;
8102 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8103 struct io_mapped_ubuf *imu,
8104 struct page **last_hpage)
8106 struct vm_area_struct **vmas = NULL;
8107 struct page **pages = NULL;
8108 unsigned long off, start, end, ubuf;
8110 int ret, pret, nr_pages, i;
8112 ubuf = (unsigned long) iov->iov_base;
8113 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8114 start = ubuf >> PAGE_SHIFT;
8115 nr_pages = end - start;
8119 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8123 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8128 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8134 mmap_read_lock(current->mm);
8135 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8137 if (pret == nr_pages) {
8138 /* don't support file backed memory */
8139 for (i = 0; i < nr_pages; i++) {
8140 struct vm_area_struct *vma = vmas[i];
8143 !is_file_hugepages(vma->vm_file)) {
8149 ret = pret < 0 ? pret : -EFAULT;
8151 mmap_read_unlock(current->mm);
8154 * if we did partial map, or found file backed vmas,
8155 * release any pages we did get
8158 unpin_user_pages(pages, pret);
8163 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8165 unpin_user_pages(pages, pret);
8170 off = ubuf & ~PAGE_MASK;
8171 size = iov->iov_len;
8172 for (i = 0; i < nr_pages; i++) {
8175 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8176 imu->bvec[i].bv_page = pages[i];
8177 imu->bvec[i].bv_len = vec_len;
8178 imu->bvec[i].bv_offset = off;
8182 /* store original address for later verification */
8184 imu->len = iov->iov_len;
8185 imu->nr_bvecs = nr_pages;
8193 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8197 if (!nr_args || nr_args > UIO_MAXIOV)
8200 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8202 if (!ctx->user_bufs)
8208 static int io_buffer_validate(struct iovec *iov)
8211 * Don't impose further limits on the size and buffer
8212 * constraints here, we'll -EINVAL later when IO is
8213 * submitted if they are wrong.
8215 if (!iov->iov_base || !iov->iov_len)
8218 /* arbitrary limit, but we need something */
8219 if (iov->iov_len > SZ_1G)
8225 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8226 unsigned int nr_args)
8230 struct page *last_hpage = NULL;
8232 ret = io_buffers_map_alloc(ctx, nr_args);
8236 for (i = 0; i < nr_args; i++) {
8237 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8239 ret = io_copy_iov(ctx, &iov, arg, i);
8243 ret = io_buffer_validate(&iov);
8247 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8251 ctx->nr_user_bufs++;
8255 io_sqe_buffers_unregister(ctx);
8260 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8262 __s32 __user *fds = arg;
8268 if (copy_from_user(&fd, fds, sizeof(*fds)))
8271 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8272 if (IS_ERR(ctx->cq_ev_fd)) {
8273 int ret = PTR_ERR(ctx->cq_ev_fd);
8274 ctx->cq_ev_fd = NULL;
8281 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8283 if (ctx->cq_ev_fd) {
8284 eventfd_ctx_put(ctx->cq_ev_fd);
8285 ctx->cq_ev_fd = NULL;
8292 static int __io_destroy_buffers(int id, void *p, void *data)
8294 struct io_ring_ctx *ctx = data;
8295 struct io_buffer *buf = p;
8297 __io_remove_buffers(ctx, buf, id, -1U);
8301 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8303 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8304 idr_destroy(&ctx->io_buffer_idr);
8307 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8309 struct io_kiocb *req, *nxt;
8311 list_for_each_entry_safe(req, nxt, list, compl.list) {
8312 if (tsk && req->task != tsk)
8314 list_del(&req->compl.list);
8315 kmem_cache_free(req_cachep, req);
8319 static void io_req_caches_free(struct io_ring_ctx *ctx)
8321 struct io_submit_state *submit_state = &ctx->submit_state;
8322 struct io_comp_state *cs = &ctx->submit_state.comp;
8324 mutex_lock(&ctx->uring_lock);
8326 if (submit_state->free_reqs) {
8327 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8328 submit_state->reqs);
8329 submit_state->free_reqs = 0;
8332 spin_lock_irq(&ctx->completion_lock);
8333 list_splice_init(&cs->locked_free_list, &cs->free_list);
8334 cs->locked_free_nr = 0;
8335 spin_unlock_irq(&ctx->completion_lock);
8337 io_req_cache_free(&cs->free_list, NULL);
8339 mutex_unlock(&ctx->uring_lock);
8342 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8345 * Some may use context even when all refs and requests have been put,
8346 * and they are free to do so while still holding uring_lock, see
8347 * __io_req_task_submit(). Wait for them to finish.
8349 mutex_lock(&ctx->uring_lock);
8350 mutex_unlock(&ctx->uring_lock);
8352 io_sq_thread_finish(ctx);
8353 io_sqe_buffers_unregister(ctx);
8355 if (ctx->mm_account) {
8356 mmdrop(ctx->mm_account);
8357 ctx->mm_account = NULL;
8360 mutex_lock(&ctx->uring_lock);
8361 io_sqe_files_unregister(ctx);
8362 mutex_unlock(&ctx->uring_lock);
8363 io_eventfd_unregister(ctx);
8364 io_destroy_buffers(ctx);
8366 #if defined(CONFIG_UNIX)
8367 if (ctx->ring_sock) {
8368 ctx->ring_sock->file = NULL; /* so that iput() is called */
8369 sock_release(ctx->ring_sock);
8373 io_mem_free(ctx->rings);
8374 io_mem_free(ctx->sq_sqes);
8376 percpu_ref_exit(&ctx->refs);
8377 free_uid(ctx->user);
8378 io_req_caches_free(ctx);
8380 io_wq_put_hash(ctx->hash_map);
8381 kfree(ctx->cancel_hash);
8385 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8387 struct io_ring_ctx *ctx = file->private_data;
8390 poll_wait(file, &ctx->cq_wait, wait);
8392 * synchronizes with barrier from wq_has_sleeper call in
8396 if (!io_sqring_full(ctx))
8397 mask |= EPOLLOUT | EPOLLWRNORM;
8400 * Don't flush cqring overflow list here, just do a simple check.
8401 * Otherwise there could possible be ABBA deadlock:
8404 * lock(&ctx->uring_lock);
8406 * lock(&ctx->uring_lock);
8409 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8410 * pushs them to do the flush.
8412 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8413 mask |= EPOLLIN | EPOLLRDNORM;
8418 static int io_uring_fasync(int fd, struct file *file, int on)
8420 struct io_ring_ctx *ctx = file->private_data;
8422 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8425 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8427 const struct cred *creds;
8429 creds = xa_erase(&ctx->personalities, id);
8438 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8440 struct callback_head *work, *next;
8441 bool executed = false;
8444 work = xchg(&ctx->exit_task_work, NULL);
8460 struct io_tctx_exit {
8461 struct callback_head task_work;
8462 struct completion completion;
8463 struct io_ring_ctx *ctx;
8466 static void io_tctx_exit_cb(struct callback_head *cb)
8468 struct io_uring_task *tctx = current->io_uring;
8469 struct io_tctx_exit *work;
8471 work = container_of(cb, struct io_tctx_exit, task_work);
8473 * When @in_idle, we're in cancellation and it's racy to remove the
8474 * node. It'll be removed by the end of cancellation, just ignore it.
8476 if (!atomic_read(&tctx->in_idle))
8477 io_uring_del_task_file((unsigned long)work->ctx);
8478 complete(&work->completion);
8481 static void io_ring_exit_work(struct work_struct *work)
8483 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8484 unsigned long timeout = jiffies + HZ * 60 * 5;
8485 struct io_tctx_exit exit;
8486 struct io_tctx_node *node;
8490 * If we're doing polled IO and end up having requests being
8491 * submitted async (out-of-line), then completions can come in while
8492 * we're waiting for refs to drop. We need to reap these manually,
8493 * as nobody else will be looking for them.
8496 io_uring_try_cancel_requests(ctx, NULL, NULL);
8498 WARN_ON_ONCE(time_after(jiffies, timeout));
8499 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8501 mutex_lock(&ctx->uring_lock);
8502 while (!list_empty(&ctx->tctx_list)) {
8503 WARN_ON_ONCE(time_after(jiffies, timeout));
8505 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8508 init_completion(&exit.completion);
8509 init_task_work(&exit.task_work, io_tctx_exit_cb);
8510 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8511 if (WARN_ON_ONCE(ret))
8513 wake_up_process(node->task);
8515 mutex_unlock(&ctx->uring_lock);
8516 wait_for_completion(&exit.completion);
8518 mutex_lock(&ctx->uring_lock);
8520 mutex_unlock(&ctx->uring_lock);
8522 io_ring_ctx_free(ctx);
8525 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8527 unsigned long index;
8528 struct creds *creds;
8530 mutex_lock(&ctx->uring_lock);
8531 percpu_ref_kill(&ctx->refs);
8532 /* if force is set, the ring is going away. always drop after that */
8533 ctx->cq_overflow_flushed = 1;
8535 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8536 xa_for_each(&ctx->personalities, index, creds)
8537 io_unregister_personality(ctx, index);
8538 mutex_unlock(&ctx->uring_lock);
8540 io_kill_timeouts(ctx, NULL, NULL);
8541 io_poll_remove_all(ctx, NULL, NULL);
8543 /* if we failed setting up the ctx, we might not have any rings */
8544 io_iopoll_try_reap_events(ctx);
8546 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8548 * Use system_unbound_wq to avoid spawning tons of event kworkers
8549 * if we're exiting a ton of rings at the same time. It just adds
8550 * noise and overhead, there's no discernable change in runtime
8551 * over using system_wq.
8553 queue_work(system_unbound_wq, &ctx->exit_work);
8556 static int io_uring_release(struct inode *inode, struct file *file)
8558 struct io_ring_ctx *ctx = file->private_data;
8560 file->private_data = NULL;
8561 io_ring_ctx_wait_and_kill(ctx);
8565 struct io_task_cancel {
8566 struct task_struct *task;
8567 struct files_struct *files;
8570 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8572 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8573 struct io_task_cancel *cancel = data;
8576 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8577 unsigned long flags;
8578 struct io_ring_ctx *ctx = req->ctx;
8580 /* protect against races with linked timeouts */
8581 spin_lock_irqsave(&ctx->completion_lock, flags);
8582 ret = io_match_task(req, cancel->task, cancel->files);
8583 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8585 ret = io_match_task(req, cancel->task, cancel->files);
8590 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8591 struct task_struct *task,
8592 struct files_struct *files)
8594 struct io_defer_entry *de;
8597 spin_lock_irq(&ctx->completion_lock);
8598 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8599 if (io_match_task(de->req, task, files)) {
8600 list_cut_position(&list, &ctx->defer_list, &de->list);
8604 spin_unlock_irq(&ctx->completion_lock);
8605 if (list_empty(&list))
8608 while (!list_empty(&list)) {
8609 de = list_first_entry(&list, struct io_defer_entry, list);
8610 list_del_init(&de->list);
8611 req_set_fail_links(de->req);
8612 io_put_req(de->req);
8613 io_req_complete(de->req, -ECANCELED);
8619 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8621 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8623 return req->ctx == data;
8626 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8628 struct io_tctx_node *node;
8629 enum io_wq_cancel cret;
8632 mutex_lock(&ctx->uring_lock);
8633 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8634 struct io_uring_task *tctx = node->task->io_uring;
8637 * io_wq will stay alive while we hold uring_lock, because it's
8638 * killed after ctx nodes, which requires to take the lock.
8640 if (!tctx || !tctx->io_wq)
8642 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8643 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8645 mutex_unlock(&ctx->uring_lock);
8650 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8651 struct task_struct *task,
8652 struct files_struct *files)
8654 struct io_task_cancel cancel = { .task = task, .files = files, };
8655 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8658 enum io_wq_cancel cret;
8662 ret |= io_uring_try_cancel_iowq(ctx);
8663 } else if (tctx && tctx->io_wq) {
8665 * Cancels requests of all rings, not only @ctx, but
8666 * it's fine as the task is in exit/exec.
8668 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8670 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8673 /* SQPOLL thread does its own polling */
8674 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8675 (ctx->sq_data && ctx->sq_data->thread == current)) {
8676 while (!list_empty_careful(&ctx->iopoll_list)) {
8677 io_iopoll_try_reap_events(ctx);
8682 ret |= io_cancel_defer_files(ctx, task, files);
8683 ret |= io_poll_remove_all(ctx, task, files);
8684 ret |= io_kill_timeouts(ctx, task, files);
8685 ret |= io_run_task_work();
8686 ret |= io_run_ctx_fallback(ctx);
8687 io_cqring_overflow_flush(ctx, true, task, files);
8694 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8695 struct task_struct *task,
8696 struct files_struct *files)
8698 struct io_kiocb *req;
8701 spin_lock_irq(&ctx->inflight_lock);
8702 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8703 cnt += io_match_task(req, task, files);
8704 spin_unlock_irq(&ctx->inflight_lock);
8708 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8709 struct task_struct *task,
8710 struct files_struct *files)
8712 while (!list_empty_careful(&ctx->inflight_list)) {
8716 inflight = io_uring_count_inflight(ctx, task, files);
8720 io_uring_try_cancel_requests(ctx, task, files);
8723 io_sq_thread_unpark(ctx->sq_data);
8724 prepare_to_wait(&task->io_uring->wait, &wait,
8725 TASK_UNINTERRUPTIBLE);
8726 if (inflight == io_uring_count_inflight(ctx, task, files))
8728 finish_wait(&task->io_uring->wait, &wait);
8730 io_sq_thread_park(ctx->sq_data);
8735 * We need to iteratively cancel requests, in case a request has dependent
8736 * hard links. These persist even for failure of cancelations, hence keep
8737 * looping until none are found.
8739 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8740 struct files_struct *files)
8742 struct task_struct *task = current;
8744 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8745 io_sq_thread_park(ctx->sq_data);
8746 task = ctx->sq_data->thread;
8748 atomic_inc(&task->io_uring->in_idle);
8751 io_uring_cancel_files(ctx, task, files);
8753 io_uring_try_cancel_requests(ctx, task, NULL);
8756 atomic_dec(&task->io_uring->in_idle);
8758 io_sq_thread_unpark(ctx->sq_data);
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);
8801 * This is race safe in that the task itself is doing this, hence it
8802 * cannot be going through the exit/cancel paths at the same time.
8803 * This cannot be modified while exit/cancel is running.
8805 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8806 tctx->sqpoll = true;
8812 * Remove this io_uring_file -> task mapping.
8814 static void io_uring_del_task_file(unsigned long index)
8816 struct io_uring_task *tctx = current->io_uring;
8817 struct io_tctx_node *node;
8821 node = xa_erase(&tctx->xa, index);
8825 WARN_ON_ONCE(current != node->task);
8826 WARN_ON_ONCE(list_empty(&node->ctx_node));
8828 mutex_lock(&node->ctx->uring_lock);
8829 list_del(&node->ctx_node);
8830 mutex_unlock(&node->ctx->uring_lock);
8832 if (tctx->last == node->ctx)
8837 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8839 struct io_tctx_node *node;
8840 unsigned long index;
8842 xa_for_each(&tctx->xa, index, node)
8843 io_uring_del_task_file(index);
8845 io_wq_put_and_exit(tctx->io_wq);
8850 void __io_uring_files_cancel(struct files_struct *files)
8852 struct io_uring_task *tctx = current->io_uring;
8853 struct io_tctx_node *node;
8854 unsigned long index;
8856 /* make sure overflow events are dropped */
8857 atomic_inc(&tctx->in_idle);
8858 xa_for_each(&tctx->xa, index, node)
8859 io_uring_cancel_task_requests(node->ctx, files);
8860 atomic_dec(&tctx->in_idle);
8863 io_uring_clean_tctx(tctx);
8866 static s64 tctx_inflight(struct io_uring_task *tctx)
8868 return percpu_counter_sum(&tctx->inflight);
8871 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8873 struct io_sq_data *sqd = ctx->sq_data;
8874 struct io_uring_task *tctx;
8880 io_sq_thread_park(sqd);
8881 if (!sqd->thread || !sqd->thread->io_uring) {
8882 io_sq_thread_unpark(sqd);
8885 tctx = ctx->sq_data->thread->io_uring;
8886 atomic_inc(&tctx->in_idle);
8888 /* read completions before cancelations */
8889 inflight = tctx_inflight(tctx);
8892 io_uring_cancel_task_requests(ctx, NULL);
8894 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8896 * If we've seen completions, retry without waiting. This
8897 * avoids a race where a completion comes in before we did
8898 * prepare_to_wait().
8900 if (inflight == tctx_inflight(tctx))
8902 finish_wait(&tctx->wait, &wait);
8904 atomic_dec(&tctx->in_idle);
8905 io_sq_thread_unpark(sqd);
8909 * Find any io_uring fd that this task has registered or done IO on, and cancel
8912 void __io_uring_task_cancel(void)
8914 struct io_uring_task *tctx = current->io_uring;
8918 /* make sure overflow events are dropped */
8919 atomic_inc(&tctx->in_idle);
8922 struct io_tctx_node *node;
8923 unsigned long index;
8925 xa_for_each(&tctx->xa, index, node)
8926 io_uring_cancel_sqpoll(node->ctx);
8930 /* read completions before cancelations */
8931 inflight = tctx_inflight(tctx);
8934 __io_uring_files_cancel(NULL);
8936 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8939 * If we've seen completions, retry without waiting. This
8940 * avoids a race where a completion comes in before we did
8941 * prepare_to_wait().
8943 if (inflight == tctx_inflight(tctx))
8945 finish_wait(&tctx->wait, &wait);
8948 atomic_dec(&tctx->in_idle);
8950 io_uring_clean_tctx(tctx);
8951 /* all current's requests should be gone, we can kill tctx */
8952 __io_uring_free(current);
8955 static void *io_uring_validate_mmap_request(struct file *file,
8956 loff_t pgoff, size_t sz)
8958 struct io_ring_ctx *ctx = file->private_data;
8959 loff_t offset = pgoff << PAGE_SHIFT;
8964 case IORING_OFF_SQ_RING:
8965 case IORING_OFF_CQ_RING:
8968 case IORING_OFF_SQES:
8972 return ERR_PTR(-EINVAL);
8975 page = virt_to_head_page(ptr);
8976 if (sz > page_size(page))
8977 return ERR_PTR(-EINVAL);
8984 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8986 size_t sz = vma->vm_end - vma->vm_start;
8990 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8992 return PTR_ERR(ptr);
8994 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8995 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8998 #else /* !CONFIG_MMU */
9000 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9002 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9005 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9007 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9010 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9011 unsigned long addr, unsigned long len,
9012 unsigned long pgoff, unsigned long flags)
9016 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9018 return PTR_ERR(ptr);
9020 return (unsigned long) ptr;
9023 #endif /* !CONFIG_MMU */
9025 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9030 if (!io_sqring_full(ctx))
9032 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9034 if (!io_sqring_full(ctx))
9037 } while (!signal_pending(current));
9039 finish_wait(&ctx->sqo_sq_wait, &wait);
9043 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9044 struct __kernel_timespec __user **ts,
9045 const sigset_t __user **sig)
9047 struct io_uring_getevents_arg arg;
9050 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9051 * is just a pointer to the sigset_t.
9053 if (!(flags & IORING_ENTER_EXT_ARG)) {
9054 *sig = (const sigset_t __user *) argp;
9060 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9061 * timespec and sigset_t pointers if good.
9063 if (*argsz != sizeof(arg))
9065 if (copy_from_user(&arg, argp, sizeof(arg)))
9067 *sig = u64_to_user_ptr(arg.sigmask);
9068 *argsz = arg.sigmask_sz;
9069 *ts = u64_to_user_ptr(arg.ts);
9073 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9074 u32, min_complete, u32, flags, const void __user *, argp,
9077 struct io_ring_ctx *ctx;
9084 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9085 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9093 if (f.file->f_op != &io_uring_fops)
9097 ctx = f.file->private_data;
9098 if (!percpu_ref_tryget(&ctx->refs))
9102 if (ctx->flags & IORING_SETUP_R_DISABLED)
9106 * For SQ polling, the thread will do all submissions and completions.
9107 * Just return the requested submit count, and wake the thread if
9111 if (ctx->flags & IORING_SETUP_SQPOLL) {
9112 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9115 if (unlikely(ctx->sq_data->thread == NULL)) {
9118 if (flags & IORING_ENTER_SQ_WAKEUP)
9119 wake_up(&ctx->sq_data->wait);
9120 if (flags & IORING_ENTER_SQ_WAIT) {
9121 ret = io_sqpoll_wait_sq(ctx);
9125 submitted = to_submit;
9126 } else if (to_submit) {
9127 ret = io_uring_add_task_file(ctx);
9130 mutex_lock(&ctx->uring_lock);
9131 submitted = io_submit_sqes(ctx, to_submit);
9132 mutex_unlock(&ctx->uring_lock);
9134 if (submitted != to_submit)
9137 if (flags & IORING_ENTER_GETEVENTS) {
9138 const sigset_t __user *sig;
9139 struct __kernel_timespec __user *ts;
9141 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9145 min_complete = min(min_complete, ctx->cq_entries);
9148 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9149 * space applications don't need to do io completion events
9150 * polling again, they can rely on io_sq_thread to do polling
9151 * work, which can reduce cpu usage and uring_lock contention.
9153 if (ctx->flags & IORING_SETUP_IOPOLL &&
9154 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9155 ret = io_iopoll_check(ctx, min_complete);
9157 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9162 percpu_ref_put(&ctx->refs);
9165 return submitted ? submitted : ret;
9168 #ifdef CONFIG_PROC_FS
9169 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9170 const struct cred *cred)
9172 struct user_namespace *uns = seq_user_ns(m);
9173 struct group_info *gi;
9178 seq_printf(m, "%5d\n", id);
9179 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9180 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9181 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9182 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9183 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9184 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9185 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9186 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9187 seq_puts(m, "\n\tGroups:\t");
9188 gi = cred->group_info;
9189 for (g = 0; g < gi->ngroups; g++) {
9190 seq_put_decimal_ull(m, g ? " " : "",
9191 from_kgid_munged(uns, gi->gid[g]));
9193 seq_puts(m, "\n\tCapEff:\t");
9194 cap = cred->cap_effective;
9195 CAP_FOR_EACH_U32(__capi)
9196 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9201 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9203 struct io_sq_data *sq = NULL;
9208 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9209 * since fdinfo case grabs it in the opposite direction of normal use
9210 * cases. If we fail to get the lock, we just don't iterate any
9211 * structures that could be going away outside the io_uring mutex.
9213 has_lock = mutex_trylock(&ctx->uring_lock);
9215 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9221 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9222 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9223 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9224 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9225 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9228 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9230 seq_printf(m, "%5u: <none>\n", i);
9232 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9233 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9234 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9236 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9237 (unsigned int) buf->len);
9239 if (has_lock && !xa_empty(&ctx->personalities)) {
9240 unsigned long index;
9241 const struct cred *cred;
9243 seq_printf(m, "Personalities:\n");
9244 xa_for_each(&ctx->personalities, index, cred)
9245 io_uring_show_cred(m, index, cred);
9247 seq_printf(m, "PollList:\n");
9248 spin_lock_irq(&ctx->completion_lock);
9249 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9250 struct hlist_head *list = &ctx->cancel_hash[i];
9251 struct io_kiocb *req;
9253 hlist_for_each_entry(req, list, hash_node)
9254 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9255 req->task->task_works != NULL);
9257 spin_unlock_irq(&ctx->completion_lock);
9259 mutex_unlock(&ctx->uring_lock);
9262 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9264 struct io_ring_ctx *ctx = f->private_data;
9266 if (percpu_ref_tryget(&ctx->refs)) {
9267 __io_uring_show_fdinfo(ctx, m);
9268 percpu_ref_put(&ctx->refs);
9273 static const struct file_operations io_uring_fops = {
9274 .release = io_uring_release,
9275 .mmap = io_uring_mmap,
9277 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9278 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9280 .poll = io_uring_poll,
9281 .fasync = io_uring_fasync,
9282 #ifdef CONFIG_PROC_FS
9283 .show_fdinfo = io_uring_show_fdinfo,
9287 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9288 struct io_uring_params *p)
9290 struct io_rings *rings;
9291 size_t size, sq_array_offset;
9293 /* make sure these are sane, as we already accounted them */
9294 ctx->sq_entries = p->sq_entries;
9295 ctx->cq_entries = p->cq_entries;
9297 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9298 if (size == SIZE_MAX)
9301 rings = io_mem_alloc(size);
9306 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9307 rings->sq_ring_mask = p->sq_entries - 1;
9308 rings->cq_ring_mask = p->cq_entries - 1;
9309 rings->sq_ring_entries = p->sq_entries;
9310 rings->cq_ring_entries = p->cq_entries;
9311 ctx->sq_mask = rings->sq_ring_mask;
9312 ctx->cq_mask = rings->cq_ring_mask;
9314 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9315 if (size == SIZE_MAX) {
9316 io_mem_free(ctx->rings);
9321 ctx->sq_sqes = io_mem_alloc(size);
9322 if (!ctx->sq_sqes) {
9323 io_mem_free(ctx->rings);
9331 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9335 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9339 ret = io_uring_add_task_file(ctx);
9344 fd_install(fd, file);
9349 * Allocate an anonymous fd, this is what constitutes the application
9350 * visible backing of an io_uring instance. The application mmaps this
9351 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9352 * we have to tie this fd to a socket for file garbage collection purposes.
9354 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9357 #if defined(CONFIG_UNIX)
9360 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9363 return ERR_PTR(ret);
9366 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9367 O_RDWR | O_CLOEXEC);
9368 #if defined(CONFIG_UNIX)
9370 sock_release(ctx->ring_sock);
9371 ctx->ring_sock = NULL;
9373 ctx->ring_sock->file = file;
9379 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9380 struct io_uring_params __user *params)
9382 struct io_ring_ctx *ctx;
9388 if (entries > IORING_MAX_ENTRIES) {
9389 if (!(p->flags & IORING_SETUP_CLAMP))
9391 entries = IORING_MAX_ENTRIES;
9395 * Use twice as many entries for the CQ ring. It's possible for the
9396 * application to drive a higher depth than the size of the SQ ring,
9397 * since the sqes are only used at submission time. This allows for
9398 * some flexibility in overcommitting a bit. If the application has
9399 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9400 * of CQ ring entries manually.
9402 p->sq_entries = roundup_pow_of_two(entries);
9403 if (p->flags & IORING_SETUP_CQSIZE) {
9405 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9406 * to a power-of-two, if it isn't already. We do NOT impose
9407 * any cq vs sq ring sizing.
9411 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9412 if (!(p->flags & IORING_SETUP_CLAMP))
9414 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9416 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9417 if (p->cq_entries < p->sq_entries)
9420 p->cq_entries = 2 * p->sq_entries;
9423 ctx = io_ring_ctx_alloc(p);
9426 ctx->compat = in_compat_syscall();
9427 if (!capable(CAP_IPC_LOCK))
9428 ctx->user = get_uid(current_user());
9431 * This is just grabbed for accounting purposes. When a process exits,
9432 * the mm is exited and dropped before the files, hence we need to hang
9433 * on to this mm purely for the purposes of being able to unaccount
9434 * memory (locked/pinned vm). It's not used for anything else.
9436 mmgrab(current->mm);
9437 ctx->mm_account = current->mm;
9439 ret = io_allocate_scq_urings(ctx, p);
9443 ret = io_sq_offload_create(ctx, p);
9447 memset(&p->sq_off, 0, sizeof(p->sq_off));
9448 p->sq_off.head = offsetof(struct io_rings, sq.head);
9449 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9450 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9451 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9452 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9453 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9454 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9456 memset(&p->cq_off, 0, sizeof(p->cq_off));
9457 p->cq_off.head = offsetof(struct io_rings, cq.head);
9458 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9459 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9460 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9461 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9462 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9463 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9465 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9466 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9467 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9468 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9469 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9471 if (copy_to_user(params, p, sizeof(*p))) {
9476 file = io_uring_get_file(ctx);
9478 ret = PTR_ERR(file);
9483 * Install ring fd as the very last thing, so we don't risk someone
9484 * having closed it before we finish setup
9486 ret = io_uring_install_fd(ctx, file);
9488 /* fput will clean it up */
9493 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9496 io_ring_ctx_wait_and_kill(ctx);
9501 * Sets up an aio uring context, and returns the fd. Applications asks for a
9502 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9503 * params structure passed in.
9505 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9507 struct io_uring_params p;
9510 if (copy_from_user(&p, params, sizeof(p)))
9512 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9517 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9518 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9519 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9520 IORING_SETUP_R_DISABLED))
9523 return io_uring_create(entries, &p, params);
9526 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9527 struct io_uring_params __user *, params)
9529 return io_uring_setup(entries, params);
9532 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9534 struct io_uring_probe *p;
9538 size = struct_size(p, ops, nr_args);
9539 if (size == SIZE_MAX)
9541 p = kzalloc(size, GFP_KERNEL);
9546 if (copy_from_user(p, arg, size))
9549 if (memchr_inv(p, 0, size))
9552 p->last_op = IORING_OP_LAST - 1;
9553 if (nr_args > IORING_OP_LAST)
9554 nr_args = IORING_OP_LAST;
9556 for (i = 0; i < nr_args; i++) {
9558 if (!io_op_defs[i].not_supported)
9559 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9564 if (copy_to_user(arg, p, size))
9571 static int io_register_personality(struct io_ring_ctx *ctx)
9573 const struct cred *creds;
9577 creds = get_current_cred();
9579 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9580 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9587 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9588 unsigned int nr_args)
9590 struct io_uring_restriction *res;
9594 /* Restrictions allowed only if rings started disabled */
9595 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9598 /* We allow only a single restrictions registration */
9599 if (ctx->restrictions.registered)
9602 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9605 size = array_size(nr_args, sizeof(*res));
9606 if (size == SIZE_MAX)
9609 res = memdup_user(arg, size);
9611 return PTR_ERR(res);
9615 for (i = 0; i < nr_args; i++) {
9616 switch (res[i].opcode) {
9617 case IORING_RESTRICTION_REGISTER_OP:
9618 if (res[i].register_op >= IORING_REGISTER_LAST) {
9623 __set_bit(res[i].register_op,
9624 ctx->restrictions.register_op);
9626 case IORING_RESTRICTION_SQE_OP:
9627 if (res[i].sqe_op >= IORING_OP_LAST) {
9632 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9634 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9635 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9637 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9638 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9647 /* Reset all restrictions if an error happened */
9649 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9651 ctx->restrictions.registered = true;
9657 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9659 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9662 if (ctx->restrictions.registered)
9663 ctx->restricted = 1;
9665 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9666 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9667 wake_up(&ctx->sq_data->wait);
9671 static bool io_register_op_must_quiesce(int op)
9674 case IORING_UNREGISTER_FILES:
9675 case IORING_REGISTER_FILES_UPDATE:
9676 case IORING_REGISTER_PROBE:
9677 case IORING_REGISTER_PERSONALITY:
9678 case IORING_UNREGISTER_PERSONALITY:
9685 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9686 void __user *arg, unsigned nr_args)
9687 __releases(ctx->uring_lock)
9688 __acquires(ctx->uring_lock)
9693 * We're inside the ring mutex, if the ref is already dying, then
9694 * someone else killed the ctx or is already going through
9695 * io_uring_register().
9697 if (percpu_ref_is_dying(&ctx->refs))
9700 if (io_register_op_must_quiesce(opcode)) {
9701 percpu_ref_kill(&ctx->refs);
9704 * Drop uring mutex before waiting for references to exit. If
9705 * another thread is currently inside io_uring_enter() it might
9706 * need to grab the uring_lock to make progress. If we hold it
9707 * here across the drain wait, then we can deadlock. It's safe
9708 * to drop the mutex here, since no new references will come in
9709 * after we've killed the percpu ref.
9711 mutex_unlock(&ctx->uring_lock);
9713 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9716 ret = io_run_task_work_sig();
9721 mutex_lock(&ctx->uring_lock);
9724 percpu_ref_resurrect(&ctx->refs);
9729 if (ctx->restricted) {
9730 if (opcode >= IORING_REGISTER_LAST) {
9735 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9742 case IORING_REGISTER_BUFFERS:
9743 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9745 case IORING_UNREGISTER_BUFFERS:
9749 ret = io_sqe_buffers_unregister(ctx);
9751 case IORING_REGISTER_FILES:
9752 ret = io_sqe_files_register(ctx, arg, nr_args);
9754 case IORING_UNREGISTER_FILES:
9758 ret = io_sqe_files_unregister(ctx);
9760 case IORING_REGISTER_FILES_UPDATE:
9761 ret = io_sqe_files_update(ctx, arg, nr_args);
9763 case IORING_REGISTER_EVENTFD:
9764 case IORING_REGISTER_EVENTFD_ASYNC:
9768 ret = io_eventfd_register(ctx, arg);
9771 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9772 ctx->eventfd_async = 1;
9774 ctx->eventfd_async = 0;
9776 case IORING_UNREGISTER_EVENTFD:
9780 ret = io_eventfd_unregister(ctx);
9782 case IORING_REGISTER_PROBE:
9784 if (!arg || nr_args > 256)
9786 ret = io_probe(ctx, arg, nr_args);
9788 case IORING_REGISTER_PERSONALITY:
9792 ret = io_register_personality(ctx);
9794 case IORING_UNREGISTER_PERSONALITY:
9798 ret = io_unregister_personality(ctx, nr_args);
9800 case IORING_REGISTER_ENABLE_RINGS:
9804 ret = io_register_enable_rings(ctx);
9806 case IORING_REGISTER_RESTRICTIONS:
9807 ret = io_register_restrictions(ctx, arg, nr_args);
9815 if (io_register_op_must_quiesce(opcode)) {
9816 /* bring the ctx back to life */
9817 percpu_ref_reinit(&ctx->refs);
9819 reinit_completion(&ctx->ref_comp);
9824 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9825 void __user *, arg, unsigned int, nr_args)
9827 struct io_ring_ctx *ctx;
9836 if (f.file->f_op != &io_uring_fops)
9839 ctx = f.file->private_data;
9843 mutex_lock(&ctx->uring_lock);
9844 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9845 mutex_unlock(&ctx->uring_lock);
9846 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9847 ctx->cq_ev_fd != NULL, ret);
9853 static int __init io_uring_init(void)
9855 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9856 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9857 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9860 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9861 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9862 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9863 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9864 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9865 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9866 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9867 BUILD_BUG_SQE_ELEM(8, __u64, off);
9868 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9869 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9870 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9871 BUILD_BUG_SQE_ELEM(24, __u32, len);
9872 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9873 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9874 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9875 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9876 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9877 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9878 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9879 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9880 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9881 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9882 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9883 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9884 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9885 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9886 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9887 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9888 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9889 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9890 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9892 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9893 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9894 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9898 __initcall(io_uring_init);