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>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define IORING_MAX_REG_BUFFERS (1U << 14)
105 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
106 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
110 u32 head ____cacheline_aligned_in_smp;
111 u32 tail ____cacheline_aligned_in_smp;
115 * This data is shared with the application through the mmap at offsets
116 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
118 * The offsets to the member fields are published through struct
119 * io_sqring_offsets when calling io_uring_setup.
123 * Head and tail offsets into the ring; the offsets need to be
124 * masked to get valid indices.
126 * The kernel controls head of the sq ring and the tail of the cq ring,
127 * and the application controls tail of the sq ring and the head of the
130 struct io_uring sq, cq;
132 * Bitmasks to apply to head and tail offsets (constant, equals
135 u32 sq_ring_mask, cq_ring_mask;
136 /* Ring sizes (constant, power of 2) */
137 u32 sq_ring_entries, cq_ring_entries;
139 * Number of invalid entries dropped by the kernel due to
140 * invalid index stored in array
142 * Written by the kernel, shouldn't be modified by the
143 * application (i.e. get number of "new events" by comparing to
146 * After a new SQ head value was read by the application this
147 * counter includes all submissions that were dropped reaching
148 * the new SQ head (and possibly more).
154 * Written by the kernel, shouldn't be modified by the
157 * The application needs a full memory barrier before checking
158 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
164 * Written by the application, shouldn't be modified by the
169 * Number of completion events lost because the queue was full;
170 * this should be avoided by the application by making sure
171 * there are not more requests pending than there is space in
172 * the completion queue.
174 * Written by the kernel, shouldn't be modified by the
175 * application (i.e. get number of "new events" by comparing to
178 * As completion events come in out of order this counter is not
179 * ordered with any other data.
183 * Ring buffer of completion events.
185 * The kernel writes completion events fresh every time they are
186 * produced, so the application is allowed to modify pending
189 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
192 enum io_uring_cmd_flags {
193 IO_URING_F_NONBLOCK = 1,
194 IO_URING_F_COMPLETE_DEFER = 2,
197 struct io_mapped_ubuf {
200 unsigned int nr_bvecs;
201 unsigned long acct_pages;
202 struct bio_vec bvec[];
207 struct io_overflow_cqe {
208 struct io_uring_cqe cqe;
209 struct list_head list;
212 struct io_fixed_file {
213 /* file * with additional FFS_* flags */
214 unsigned long file_ptr;
218 struct list_head list;
223 struct io_mapped_ubuf *buf;
227 struct io_file_table {
228 /* two level table */
229 struct io_fixed_file **files;
232 struct io_rsrc_node {
233 struct percpu_ref refs;
234 struct list_head node;
235 struct list_head rsrc_list;
236 struct io_rsrc_data *rsrc_data;
237 struct llist_node llist;
241 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
243 struct io_rsrc_data {
244 struct io_ring_ctx *ctx;
249 struct completion done;
254 struct list_head list;
260 struct io_restriction {
261 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
262 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
263 u8 sqe_flags_allowed;
264 u8 sqe_flags_required;
269 IO_SQ_THREAD_SHOULD_STOP = 0,
270 IO_SQ_THREAD_SHOULD_PARK,
275 atomic_t park_pending;
278 /* ctx's that are using this sqd */
279 struct list_head ctx_list;
281 struct task_struct *thread;
282 struct wait_queue_head wait;
284 unsigned sq_thread_idle;
290 struct completion exited;
291 struct callback_head *park_task_work;
294 #define IO_IOPOLL_BATCH 8
295 #define IO_COMPL_BATCH 32
296 #define IO_REQ_CACHE_SIZE 32
297 #define IO_REQ_ALLOC_BATCH 8
299 struct io_comp_state {
300 struct io_kiocb *reqs[IO_COMPL_BATCH];
302 unsigned int locked_free_nr;
303 /* inline/task_work completion list, under ->uring_lock */
304 struct list_head free_list;
305 /* IRQ completion list, under ->completion_lock */
306 struct list_head locked_free_list;
309 struct io_submit_link {
310 struct io_kiocb *head;
311 struct io_kiocb *last;
314 struct io_submit_state {
315 struct blk_plug plug;
316 struct io_submit_link link;
319 * io_kiocb alloc cache
321 void *reqs[IO_REQ_CACHE_SIZE];
322 unsigned int free_reqs;
327 * Batch completion logic
329 struct io_comp_state comp;
332 * File reference cache
336 unsigned int file_refs;
337 unsigned int ios_left;
342 struct percpu_ref refs;
343 } ____cacheline_aligned_in_smp;
347 unsigned int compat: 1;
348 unsigned int drain_next: 1;
349 unsigned int eventfd_async: 1;
350 unsigned int restricted: 1;
353 * Ring buffer of indices into array of io_uring_sqe, which is
354 * mmapped by the application using the IORING_OFF_SQES offset.
356 * This indirection could e.g. be used to assign fixed
357 * io_uring_sqe entries to operations and only submit them to
358 * the queue when needed.
360 * The kernel modifies neither the indices array nor the entries
364 unsigned cached_sq_head;
367 unsigned sq_thread_idle;
368 unsigned cached_sq_dropped;
369 unsigned cached_cq_overflow;
370 unsigned long sq_check_overflow;
372 /* hashed buffered write serialization */
373 struct io_wq_hash *hash_map;
375 struct list_head defer_list;
376 struct list_head timeout_list;
377 struct list_head cq_overflow_list;
379 struct io_uring_sqe *sq_sqes;
380 } ____cacheline_aligned_in_smp;
383 struct mutex uring_lock;
384 wait_queue_head_t wait;
385 } ____cacheline_aligned_in_smp;
387 struct io_submit_state submit_state;
389 struct io_rings *rings;
391 /* Only used for accounting purposes */
392 struct mm_struct *mm_account;
394 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
395 struct io_sq_data *sq_data; /* if using sq thread polling */
397 struct wait_queue_head sqo_sq_wait;
398 struct list_head sqd_list;
401 * If used, fixed file set. Writers must ensure that ->refs is dead,
402 * readers must ensure that ->refs is alive as long as the file* is
403 * used. Only updated through io_uring_register(2).
405 struct io_rsrc_data *file_data;
406 struct io_file_table file_table;
407 unsigned nr_user_files;
409 /* if used, fixed mapped user buffers */
410 struct io_rsrc_data *buf_data;
411 unsigned nr_user_bufs;
412 struct io_mapped_ubuf **user_bufs;
414 struct user_struct *user;
416 struct completion ref_comp;
418 #if defined(CONFIG_UNIX)
419 struct socket *ring_sock;
422 struct xarray io_buffers;
424 struct xarray personalities;
428 unsigned cached_cq_tail;
431 atomic_t cq_timeouts;
432 unsigned cq_last_tm_flush;
434 unsigned long cq_check_overflow;
435 struct wait_queue_head cq_wait;
436 struct fasync_struct *cq_fasync;
437 struct eventfd_ctx *cq_ev_fd;
438 } ____cacheline_aligned_in_smp;
441 spinlock_t completion_lock;
444 * ->iopoll_list is protected by the ctx->uring_lock for
445 * io_uring instances that don't use IORING_SETUP_SQPOLL.
446 * For SQPOLL, only the single threaded io_sq_thread() will
447 * manipulate the list, hence no extra locking is needed there.
449 struct list_head iopoll_list;
450 struct hlist_head *cancel_hash;
451 unsigned cancel_hash_bits;
452 bool poll_multi_file;
453 } ____cacheline_aligned_in_smp;
455 struct delayed_work rsrc_put_work;
456 struct llist_head rsrc_put_llist;
457 struct list_head rsrc_ref_list;
458 spinlock_t rsrc_ref_lock;
459 struct io_rsrc_node *rsrc_node;
460 struct io_rsrc_node *rsrc_backup_node;
461 struct io_mapped_ubuf *dummy_ubuf;
463 struct io_restriction restrictions;
466 struct callback_head *exit_task_work;
468 /* Keep this last, we don't need it for the fast path */
469 struct work_struct exit_work;
470 struct list_head tctx_list;
473 struct io_uring_task {
474 /* submission side */
476 struct wait_queue_head wait;
477 const struct io_ring_ctx *last;
479 struct percpu_counter inflight;
480 atomic_t inflight_tracked;
483 spinlock_t task_lock;
484 struct io_wq_work_list task_list;
485 unsigned long task_state;
486 struct callback_head task_work;
490 * First field must be the file pointer in all the
491 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
493 struct io_poll_iocb {
495 struct wait_queue_head *head;
499 struct wait_queue_entry wait;
502 struct io_poll_update {
508 bool update_user_data;
516 struct io_timeout_data {
517 struct io_kiocb *req;
518 struct hrtimer timer;
519 struct timespec64 ts;
520 enum hrtimer_mode mode;
525 struct sockaddr __user *addr;
526 int __user *addr_len;
528 unsigned long nofile;
548 struct list_head list;
549 /* head of the link, used by linked timeouts only */
550 struct io_kiocb *head;
553 struct io_timeout_rem {
558 struct timespec64 ts;
563 /* NOTE: kiocb has the file as the first member, so don't do it here */
571 struct sockaddr __user *addr;
578 struct compat_msghdr __user *umsg_compat;
579 struct user_msghdr __user *umsg;
585 struct io_buffer *kbuf;
591 struct filename *filename;
593 unsigned long nofile;
596 struct io_rsrc_update {
622 struct epoll_event event;
626 struct file *file_out;
627 struct file *file_in;
634 struct io_provide_buf {
648 const char __user *filename;
649 struct statx __user *buffer;
661 struct filename *oldpath;
662 struct filename *newpath;
670 struct filename *filename;
673 struct io_completion {
675 struct list_head list;
679 struct io_async_connect {
680 struct sockaddr_storage address;
683 struct io_async_msghdr {
684 struct iovec fast_iov[UIO_FASTIOV];
685 /* points to an allocated iov, if NULL we use fast_iov instead */
686 struct iovec *free_iov;
687 struct sockaddr __user *uaddr;
689 struct sockaddr_storage addr;
693 struct iovec fast_iov[UIO_FASTIOV];
694 const struct iovec *free_iovec;
695 struct iov_iter iter;
697 struct wait_page_queue wpq;
701 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
702 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
703 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
704 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
705 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
706 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
708 /* first byte is taken by user flags, shift it to not overlap */
709 REQ_F_FAIL_LINK_BIT = 8,
713 REQ_F_LINK_TIMEOUT_BIT,
714 REQ_F_NEED_CLEANUP_BIT,
716 REQ_F_BUFFER_SELECTED_BIT,
717 REQ_F_LTIMEOUT_ACTIVE_BIT,
718 REQ_F_COMPLETE_INLINE_BIT,
720 REQ_F_DONT_REISSUE_BIT,
721 /* keep async read/write and isreg together and in order */
722 REQ_F_ASYNC_READ_BIT,
723 REQ_F_ASYNC_WRITE_BIT,
726 /* not a real bit, just to check we're not overflowing the space */
732 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
733 /* drain existing IO first */
734 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
736 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
737 /* doesn't sever on completion < 0 */
738 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
740 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
741 /* IOSQE_BUFFER_SELECT */
742 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
744 /* fail rest of links */
745 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
746 /* on inflight list, should be cancelled and waited on exit reliably */
747 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
748 /* read/write uses file position */
749 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
750 /* must not punt to workers */
751 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
752 /* has or had linked timeout */
753 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
755 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
756 /* already went through poll handler */
757 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
758 /* buffer already selected */
759 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
760 /* linked timeout is active, i.e. prepared by link's head */
761 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
762 /* completion is deferred through io_comp_state */
763 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
764 /* caller should reissue async */
765 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
766 /* don't attempt request reissue, see io_rw_reissue() */
767 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
768 /* supports async reads */
769 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
770 /* supports async writes */
771 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
773 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
777 struct io_poll_iocb poll;
778 struct io_poll_iocb *double_poll;
781 struct io_task_work {
782 struct io_wq_work_node node;
783 task_work_func_t func;
787 IORING_RSRC_FILE = 0,
788 IORING_RSRC_BUFFER = 1,
792 * NOTE! Each of the iocb union members has the file pointer
793 * as the first entry in their struct definition. So you can
794 * access the file pointer through any of the sub-structs,
795 * or directly as just 'ki_filp' in this struct.
801 struct io_poll_iocb poll;
802 struct io_poll_update poll_update;
803 struct io_accept accept;
805 struct io_cancel cancel;
806 struct io_timeout timeout;
807 struct io_timeout_rem timeout_rem;
808 struct io_connect connect;
809 struct io_sr_msg sr_msg;
811 struct io_close close;
812 struct io_rsrc_update rsrc_update;
813 struct io_fadvise fadvise;
814 struct io_madvise madvise;
815 struct io_epoll epoll;
816 struct io_splice splice;
817 struct io_provide_buf pbuf;
818 struct io_statx statx;
819 struct io_shutdown shutdown;
820 struct io_rename rename;
821 struct io_unlink unlink;
822 /* use only after cleaning per-op data, see io_clean_op() */
823 struct io_completion compl;
826 /* opcode allocated if it needs to store data for async defer */
829 /* polled IO has completed */
835 struct io_ring_ctx *ctx;
838 struct task_struct *task;
841 struct io_kiocb *link;
842 struct percpu_ref *fixed_rsrc_refs;
844 /* used with ctx->iopoll_list with reads/writes */
845 struct list_head inflight_entry;
847 struct io_task_work io_task_work;
848 struct callback_head task_work;
850 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
851 struct hlist_node hash_node;
852 struct async_poll *apoll;
853 struct io_wq_work work;
854 /* store used ubuf, so we can prevent reloading */
855 struct io_mapped_ubuf *imu;
858 struct io_tctx_node {
859 struct list_head ctx_node;
860 struct task_struct *task;
861 struct io_ring_ctx *ctx;
864 struct io_defer_entry {
865 struct list_head list;
866 struct io_kiocb *req;
871 /* needs req->file assigned */
872 unsigned needs_file : 1;
873 /* hash wq insertion if file is a regular file */
874 unsigned hash_reg_file : 1;
875 /* unbound wq insertion if file is a non-regular file */
876 unsigned unbound_nonreg_file : 1;
877 /* opcode is not supported by this kernel */
878 unsigned not_supported : 1;
879 /* set if opcode supports polled "wait" */
881 unsigned pollout : 1;
882 /* op supports buffer selection */
883 unsigned buffer_select : 1;
884 /* do prep async if is going to be punted */
885 unsigned needs_async_setup : 1;
886 /* should block plug */
888 /* size of async data needed, if any */
889 unsigned short async_size;
892 static const struct io_op_def io_op_defs[] = {
893 [IORING_OP_NOP] = {},
894 [IORING_OP_READV] = {
896 .unbound_nonreg_file = 1,
899 .needs_async_setup = 1,
901 .async_size = sizeof(struct io_async_rw),
903 [IORING_OP_WRITEV] = {
906 .unbound_nonreg_file = 1,
908 .needs_async_setup = 1,
910 .async_size = sizeof(struct io_async_rw),
912 [IORING_OP_FSYNC] = {
915 [IORING_OP_READ_FIXED] = {
917 .unbound_nonreg_file = 1,
920 .async_size = sizeof(struct io_async_rw),
922 [IORING_OP_WRITE_FIXED] = {
925 .unbound_nonreg_file = 1,
928 .async_size = sizeof(struct io_async_rw),
930 [IORING_OP_POLL_ADD] = {
932 .unbound_nonreg_file = 1,
934 [IORING_OP_POLL_REMOVE] = {},
935 [IORING_OP_SYNC_FILE_RANGE] = {
938 [IORING_OP_SENDMSG] = {
940 .unbound_nonreg_file = 1,
942 .needs_async_setup = 1,
943 .async_size = sizeof(struct io_async_msghdr),
945 [IORING_OP_RECVMSG] = {
947 .unbound_nonreg_file = 1,
950 .needs_async_setup = 1,
951 .async_size = sizeof(struct io_async_msghdr),
953 [IORING_OP_TIMEOUT] = {
954 .async_size = sizeof(struct io_timeout_data),
956 [IORING_OP_TIMEOUT_REMOVE] = {
957 /* used by timeout updates' prep() */
959 [IORING_OP_ACCEPT] = {
961 .unbound_nonreg_file = 1,
964 [IORING_OP_ASYNC_CANCEL] = {},
965 [IORING_OP_LINK_TIMEOUT] = {
966 .async_size = sizeof(struct io_timeout_data),
968 [IORING_OP_CONNECT] = {
970 .unbound_nonreg_file = 1,
972 .needs_async_setup = 1,
973 .async_size = sizeof(struct io_async_connect),
975 [IORING_OP_FALLOCATE] = {
978 [IORING_OP_OPENAT] = {},
979 [IORING_OP_CLOSE] = {},
980 [IORING_OP_FILES_UPDATE] = {},
981 [IORING_OP_STATX] = {},
984 .unbound_nonreg_file = 1,
988 .async_size = sizeof(struct io_async_rw),
990 [IORING_OP_WRITE] = {
992 .unbound_nonreg_file = 1,
995 .async_size = sizeof(struct io_async_rw),
997 [IORING_OP_FADVISE] = {
1000 [IORING_OP_MADVISE] = {},
1001 [IORING_OP_SEND] = {
1003 .unbound_nonreg_file = 1,
1006 [IORING_OP_RECV] = {
1008 .unbound_nonreg_file = 1,
1012 [IORING_OP_OPENAT2] = {
1014 [IORING_OP_EPOLL_CTL] = {
1015 .unbound_nonreg_file = 1,
1017 [IORING_OP_SPLICE] = {
1020 .unbound_nonreg_file = 1,
1022 [IORING_OP_PROVIDE_BUFFERS] = {},
1023 [IORING_OP_REMOVE_BUFFERS] = {},
1027 .unbound_nonreg_file = 1,
1029 [IORING_OP_SHUTDOWN] = {
1032 [IORING_OP_RENAMEAT] = {},
1033 [IORING_OP_UNLINKAT] = {},
1036 static bool io_disarm_next(struct io_kiocb *req);
1037 static void io_uring_del_task_file(unsigned long index);
1038 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1039 struct task_struct *task,
1040 struct files_struct *files);
1041 static void io_uring_cancel_sqpoll(struct io_sq_data *sqd);
1042 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1044 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1045 long res, unsigned int cflags);
1046 static void io_put_req(struct io_kiocb *req);
1047 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1048 static void io_dismantle_req(struct io_kiocb *req);
1049 static void io_put_task(struct task_struct *task, int nr);
1050 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1051 static void io_queue_linked_timeout(struct io_kiocb *req);
1052 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1053 struct io_uring_rsrc_update2 *up,
1055 static void io_clean_op(struct io_kiocb *req);
1056 static struct file *io_file_get(struct io_submit_state *state,
1057 struct io_kiocb *req, int fd, bool fixed);
1058 static void __io_queue_sqe(struct io_kiocb *req);
1059 static void io_rsrc_put_work(struct work_struct *work);
1061 static void io_req_task_queue(struct io_kiocb *req);
1062 static void io_submit_flush_completions(struct io_comp_state *cs,
1063 struct io_ring_ctx *ctx);
1064 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1065 static int io_req_prep_async(struct io_kiocb *req);
1067 static struct kmem_cache *req_cachep;
1069 static const struct file_operations io_uring_fops;
1071 struct sock *io_uring_get_socket(struct file *file)
1073 #if defined(CONFIG_UNIX)
1074 if (file->f_op == &io_uring_fops) {
1075 struct io_ring_ctx *ctx = file->private_data;
1077 return ctx->ring_sock->sk;
1082 EXPORT_SYMBOL(io_uring_get_socket);
1084 #define io_for_each_link(pos, head) \
1085 for (pos = (head); pos; pos = pos->link)
1087 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1089 struct io_ring_ctx *ctx = req->ctx;
1091 if (!req->fixed_rsrc_refs) {
1092 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1093 percpu_ref_get(req->fixed_rsrc_refs);
1097 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1099 bool got = percpu_ref_tryget(ref);
1101 /* already at zero, wait for ->release() */
1103 wait_for_completion(compl);
1104 percpu_ref_resurrect(ref);
1106 percpu_ref_put(ref);
1109 static bool io_match_task(struct io_kiocb *head,
1110 struct task_struct *task,
1111 struct files_struct *files)
1113 struct io_kiocb *req;
1115 if (task && head->task != task)
1120 io_for_each_link(req, head) {
1121 if (req->flags & REQ_F_INFLIGHT)
1127 static inline void req_set_fail_links(struct io_kiocb *req)
1129 if (req->flags & REQ_F_LINK)
1130 req->flags |= REQ_F_FAIL_LINK;
1133 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1135 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1137 complete(&ctx->ref_comp);
1140 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1142 return !req->timeout.off;
1145 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1147 struct io_ring_ctx *ctx;
1150 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1155 * Use 5 bits less than the max cq entries, that should give us around
1156 * 32 entries per hash list if totally full and uniformly spread.
1158 hash_bits = ilog2(p->cq_entries);
1162 ctx->cancel_hash_bits = hash_bits;
1163 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1165 if (!ctx->cancel_hash)
1167 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1169 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1170 if (!ctx->dummy_ubuf)
1172 /* set invalid range, so io_import_fixed() fails meeting it */
1173 ctx->dummy_ubuf->ubuf = -1UL;
1175 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1176 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1179 ctx->flags = p->flags;
1180 init_waitqueue_head(&ctx->sqo_sq_wait);
1181 INIT_LIST_HEAD(&ctx->sqd_list);
1182 init_waitqueue_head(&ctx->cq_wait);
1183 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1184 init_completion(&ctx->ref_comp);
1185 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1186 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1187 mutex_init(&ctx->uring_lock);
1188 init_waitqueue_head(&ctx->wait);
1189 spin_lock_init(&ctx->completion_lock);
1190 INIT_LIST_HEAD(&ctx->iopoll_list);
1191 INIT_LIST_HEAD(&ctx->defer_list);
1192 INIT_LIST_HEAD(&ctx->timeout_list);
1193 spin_lock_init(&ctx->rsrc_ref_lock);
1194 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1195 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1196 init_llist_head(&ctx->rsrc_put_llist);
1197 INIT_LIST_HEAD(&ctx->tctx_list);
1198 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1199 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1202 kfree(ctx->dummy_ubuf);
1203 kfree(ctx->cancel_hash);
1208 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1210 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1211 struct io_ring_ctx *ctx = req->ctx;
1213 return seq + ctx->cq_extra != ctx->cached_cq_tail
1214 + READ_ONCE(ctx->cached_cq_overflow);
1220 static void io_req_track_inflight(struct io_kiocb *req)
1222 if (!(req->flags & REQ_F_INFLIGHT)) {
1223 req->flags |= REQ_F_INFLIGHT;
1224 atomic_inc(¤t->io_uring->inflight_tracked);
1228 static void io_prep_async_work(struct io_kiocb *req)
1230 const struct io_op_def *def = &io_op_defs[req->opcode];
1231 struct io_ring_ctx *ctx = req->ctx;
1233 if (!req->work.creds)
1234 req->work.creds = get_current_cred();
1236 req->work.list.next = NULL;
1237 req->work.flags = 0;
1238 if (req->flags & REQ_F_FORCE_ASYNC)
1239 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1241 if (req->flags & REQ_F_ISREG) {
1242 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1243 io_wq_hash_work(&req->work, file_inode(req->file));
1244 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1245 if (def->unbound_nonreg_file)
1246 req->work.flags |= IO_WQ_WORK_UNBOUND;
1249 switch (req->opcode) {
1250 case IORING_OP_SPLICE:
1252 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1253 req->work.flags |= IO_WQ_WORK_UNBOUND;
1258 static void io_prep_async_link(struct io_kiocb *req)
1260 struct io_kiocb *cur;
1262 io_for_each_link(cur, req)
1263 io_prep_async_work(cur);
1266 static void io_queue_async_work(struct io_kiocb *req)
1268 struct io_ring_ctx *ctx = req->ctx;
1269 struct io_kiocb *link = io_prep_linked_timeout(req);
1270 struct io_uring_task *tctx = req->task->io_uring;
1273 BUG_ON(!tctx->io_wq);
1275 /* init ->work of the whole link before punting */
1276 io_prep_async_link(req);
1277 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1278 &req->work, req->flags);
1279 io_wq_enqueue(tctx->io_wq, &req->work);
1281 io_queue_linked_timeout(link);
1284 static void io_kill_timeout(struct io_kiocb *req, int status)
1285 __must_hold(&req->ctx->completion_lock)
1287 struct io_timeout_data *io = req->async_data;
1289 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1290 atomic_set(&req->ctx->cq_timeouts,
1291 atomic_read(&req->ctx->cq_timeouts) + 1);
1292 list_del_init(&req->timeout.list);
1293 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1294 io_put_req_deferred(req, 1);
1298 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1301 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1302 struct io_defer_entry, list);
1304 if (req_need_defer(de->req, de->seq))
1306 list_del_init(&de->list);
1307 io_req_task_queue(de->req);
1309 } while (!list_empty(&ctx->defer_list));
1312 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1316 if (list_empty(&ctx->timeout_list))
1319 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1322 u32 events_needed, events_got;
1323 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1324 struct io_kiocb, timeout.list);
1326 if (io_is_timeout_noseq(req))
1330 * Since seq can easily wrap around over time, subtract
1331 * the last seq at which timeouts were flushed before comparing.
1332 * Assuming not more than 2^31-1 events have happened since,
1333 * these subtractions won't have wrapped, so we can check if
1334 * target is in [last_seq, current_seq] by comparing the two.
1336 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1337 events_got = seq - ctx->cq_last_tm_flush;
1338 if (events_got < events_needed)
1341 list_del_init(&req->timeout.list);
1342 io_kill_timeout(req, 0);
1343 } while (!list_empty(&ctx->timeout_list));
1345 ctx->cq_last_tm_flush = seq;
1348 static void io_commit_cqring(struct io_ring_ctx *ctx)
1350 io_flush_timeouts(ctx);
1352 /* order cqe stores with ring update */
1353 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1355 if (unlikely(!list_empty(&ctx->defer_list)))
1356 __io_queue_deferred(ctx);
1359 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1361 struct io_rings *r = ctx->rings;
1363 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1366 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1368 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1371 static inline struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1373 struct io_rings *rings = ctx->rings;
1377 * writes to the cq entry need to come after reading head; the
1378 * control dependency is enough as we're using WRITE_ONCE to
1381 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1384 tail = ctx->cached_cq_tail++;
1385 return &rings->cqes[tail & ctx->cq_mask];
1388 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1390 if (likely(!ctx->cq_ev_fd))
1392 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1394 return !ctx->eventfd_async || io_wq_current_is_worker();
1397 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1399 /* see waitqueue_active() comment */
1402 if (waitqueue_active(&ctx->wait))
1403 wake_up(&ctx->wait);
1404 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1405 wake_up(&ctx->sq_data->wait);
1406 if (io_should_trigger_evfd(ctx))
1407 eventfd_signal(ctx->cq_ev_fd, 1);
1408 if (waitqueue_active(&ctx->cq_wait)) {
1409 wake_up_interruptible(&ctx->cq_wait);
1410 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1414 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1416 /* see waitqueue_active() comment */
1419 if (ctx->flags & IORING_SETUP_SQPOLL) {
1420 if (waitqueue_active(&ctx->wait))
1421 wake_up(&ctx->wait);
1423 if (io_should_trigger_evfd(ctx))
1424 eventfd_signal(ctx->cq_ev_fd, 1);
1425 if (waitqueue_active(&ctx->cq_wait)) {
1426 wake_up_interruptible(&ctx->cq_wait);
1427 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1431 /* Returns true if there are no backlogged entries after the flush */
1432 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1434 struct io_rings *rings = ctx->rings;
1435 unsigned long flags;
1436 bool all_flushed, posted;
1438 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1442 spin_lock_irqsave(&ctx->completion_lock, flags);
1443 while (!list_empty(&ctx->cq_overflow_list)) {
1444 struct io_uring_cqe *cqe = io_get_cqring(ctx);
1445 struct io_overflow_cqe *ocqe;
1449 ocqe = list_first_entry(&ctx->cq_overflow_list,
1450 struct io_overflow_cqe, list);
1452 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1454 WRITE_ONCE(ctx->rings->cq_overflow,
1455 ++ctx->cached_cq_overflow);
1457 list_del(&ocqe->list);
1461 all_flushed = list_empty(&ctx->cq_overflow_list);
1463 clear_bit(0, &ctx->sq_check_overflow);
1464 clear_bit(0, &ctx->cq_check_overflow);
1465 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1469 io_commit_cqring(ctx);
1470 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1472 io_cqring_ev_posted(ctx);
1476 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1480 if (test_bit(0, &ctx->cq_check_overflow)) {
1481 /* iopoll syncs against uring_lock, not completion_lock */
1482 if (ctx->flags & IORING_SETUP_IOPOLL)
1483 mutex_lock(&ctx->uring_lock);
1484 ret = __io_cqring_overflow_flush(ctx, force);
1485 if (ctx->flags & IORING_SETUP_IOPOLL)
1486 mutex_unlock(&ctx->uring_lock);
1493 * Shamelessly stolen from the mm implementation of page reference checking,
1494 * see commit f958d7b528b1 for details.
1496 #define req_ref_zero_or_close_to_overflow(req) \
1497 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1499 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1501 return atomic_inc_not_zero(&req->refs);
1504 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1506 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1507 return atomic_sub_and_test(refs, &req->refs);
1510 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1512 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1513 return atomic_dec_and_test(&req->refs);
1516 static inline void req_ref_put(struct io_kiocb *req)
1518 WARN_ON_ONCE(req_ref_put_and_test(req));
1521 static inline void req_ref_get(struct io_kiocb *req)
1523 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1524 atomic_inc(&req->refs);
1527 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1528 long res, unsigned int cflags)
1530 struct io_overflow_cqe *ocqe;
1532 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1535 * If we're in ring overflow flush mode, or in task cancel mode,
1536 * or cannot allocate an overflow entry, then we need to drop it
1539 WRITE_ONCE(ctx->rings->cq_overflow, ++ctx->cached_cq_overflow);
1542 if (list_empty(&ctx->cq_overflow_list)) {
1543 set_bit(0, &ctx->sq_check_overflow);
1544 set_bit(0, &ctx->cq_check_overflow);
1545 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1547 ocqe->cqe.user_data = user_data;
1548 ocqe->cqe.res = res;
1549 ocqe->cqe.flags = cflags;
1550 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1554 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1555 long res, unsigned int cflags)
1557 struct io_uring_cqe *cqe;
1559 trace_io_uring_complete(ctx, user_data, res, cflags);
1562 * If we can't get a cq entry, userspace overflowed the
1563 * submission (by quite a lot). Increment the overflow count in
1566 cqe = io_get_cqring(ctx);
1568 WRITE_ONCE(cqe->user_data, user_data);
1569 WRITE_ONCE(cqe->res, res);
1570 WRITE_ONCE(cqe->flags, cflags);
1573 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1576 /* not as hot to bloat with inlining */
1577 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1578 long res, unsigned int cflags)
1580 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1583 static void io_req_complete_post(struct io_kiocb *req, long res,
1584 unsigned int cflags)
1586 struct io_ring_ctx *ctx = req->ctx;
1587 unsigned long flags;
1589 spin_lock_irqsave(&ctx->completion_lock, flags);
1590 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1592 * If we're the last reference to this request, add to our locked
1595 if (req_ref_put_and_test(req)) {
1596 struct io_comp_state *cs = &ctx->submit_state.comp;
1598 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1599 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1600 io_disarm_next(req);
1602 io_req_task_queue(req->link);
1606 io_dismantle_req(req);
1607 io_put_task(req->task, 1);
1608 list_add(&req->compl.list, &cs->locked_free_list);
1609 cs->locked_free_nr++;
1611 if (!percpu_ref_tryget(&ctx->refs))
1614 io_commit_cqring(ctx);
1615 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1618 io_cqring_ev_posted(ctx);
1619 percpu_ref_put(&ctx->refs);
1623 static inline bool io_req_needs_clean(struct io_kiocb *req)
1625 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1626 REQ_F_POLLED | REQ_F_INFLIGHT);
1629 static void io_req_complete_state(struct io_kiocb *req, long res,
1630 unsigned int cflags)
1632 if (io_req_needs_clean(req))
1635 req->compl.cflags = cflags;
1636 req->flags |= REQ_F_COMPLETE_INLINE;
1639 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1640 long res, unsigned cflags)
1642 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1643 io_req_complete_state(req, res, cflags);
1645 io_req_complete_post(req, res, cflags);
1648 static inline void io_req_complete(struct io_kiocb *req, long res)
1650 __io_req_complete(req, 0, res, 0);
1653 static void io_req_complete_failed(struct io_kiocb *req, long res)
1655 req_set_fail_links(req);
1657 io_req_complete_post(req, res, 0);
1660 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1661 struct io_comp_state *cs)
1663 spin_lock_irq(&ctx->completion_lock);
1664 list_splice_init(&cs->locked_free_list, &cs->free_list);
1665 cs->locked_free_nr = 0;
1666 spin_unlock_irq(&ctx->completion_lock);
1669 /* Returns true IFF there are requests in the cache */
1670 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1672 struct io_submit_state *state = &ctx->submit_state;
1673 struct io_comp_state *cs = &state->comp;
1677 * If we have more than a batch's worth of requests in our IRQ side
1678 * locked cache, grab the lock and move them over to our submission
1681 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH)
1682 io_flush_cached_locked_reqs(ctx, cs);
1684 nr = state->free_reqs;
1685 while (!list_empty(&cs->free_list)) {
1686 struct io_kiocb *req = list_first_entry(&cs->free_list,
1687 struct io_kiocb, compl.list);
1689 list_del(&req->compl.list);
1690 state->reqs[nr++] = req;
1691 if (nr == ARRAY_SIZE(state->reqs))
1695 state->free_reqs = nr;
1699 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1701 struct io_submit_state *state = &ctx->submit_state;
1703 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1705 if (!state->free_reqs) {
1706 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1709 if (io_flush_cached_reqs(ctx))
1712 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1716 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1717 * retry single alloc to be on the safe side.
1719 if (unlikely(ret <= 0)) {
1720 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1721 if (!state->reqs[0])
1725 state->free_reqs = ret;
1729 return state->reqs[state->free_reqs];
1732 static inline void io_put_file(struct file *file)
1738 static void io_dismantle_req(struct io_kiocb *req)
1740 unsigned int flags = req->flags;
1742 if (io_req_needs_clean(req))
1744 if (!(flags & REQ_F_FIXED_FILE))
1745 io_put_file(req->file);
1746 if (req->fixed_rsrc_refs)
1747 percpu_ref_put(req->fixed_rsrc_refs);
1748 if (req->async_data)
1749 kfree(req->async_data);
1750 if (req->work.creds) {
1751 put_cred(req->work.creds);
1752 req->work.creds = NULL;
1756 /* must to be called somewhat shortly after putting a request */
1757 static inline void io_put_task(struct task_struct *task, int nr)
1759 struct io_uring_task *tctx = task->io_uring;
1761 percpu_counter_sub(&tctx->inflight, nr);
1762 if (unlikely(atomic_read(&tctx->in_idle)))
1763 wake_up(&tctx->wait);
1764 put_task_struct_many(task, nr);
1767 static void __io_free_req(struct io_kiocb *req)
1769 struct io_ring_ctx *ctx = req->ctx;
1771 io_dismantle_req(req);
1772 io_put_task(req->task, 1);
1774 kmem_cache_free(req_cachep, req);
1775 percpu_ref_put(&ctx->refs);
1778 static inline void io_remove_next_linked(struct io_kiocb *req)
1780 struct io_kiocb *nxt = req->link;
1782 req->link = nxt->link;
1786 static bool io_kill_linked_timeout(struct io_kiocb *req)
1787 __must_hold(&req->ctx->completion_lock)
1789 struct io_kiocb *link = req->link;
1792 * Can happen if a linked timeout fired and link had been like
1793 * req -> link t-out -> link t-out [-> ...]
1795 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1796 struct io_timeout_data *io = link->async_data;
1798 io_remove_next_linked(req);
1799 link->timeout.head = NULL;
1800 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1801 io_cqring_fill_event(link->ctx, link->user_data,
1803 io_put_req_deferred(link, 1);
1810 static void io_fail_links(struct io_kiocb *req)
1811 __must_hold(&req->ctx->completion_lock)
1813 struct io_kiocb *nxt, *link = req->link;
1820 trace_io_uring_fail_link(req, link);
1821 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1822 io_put_req_deferred(link, 2);
1827 static bool io_disarm_next(struct io_kiocb *req)
1828 __must_hold(&req->ctx->completion_lock)
1830 bool posted = false;
1832 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1833 posted = io_kill_linked_timeout(req);
1834 if (unlikely((req->flags & REQ_F_FAIL_LINK) &&
1835 !(req->flags & REQ_F_HARDLINK))) {
1836 posted |= (req->link != NULL);
1842 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1844 struct io_kiocb *nxt;
1847 * If LINK is set, we have dependent requests in this chain. If we
1848 * didn't fail this request, queue the first one up, moving any other
1849 * dependencies to the next request. In case of failure, fail the rest
1852 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1853 struct io_ring_ctx *ctx = req->ctx;
1854 unsigned long flags;
1857 spin_lock_irqsave(&ctx->completion_lock, flags);
1858 posted = io_disarm_next(req);
1860 io_commit_cqring(req->ctx);
1861 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1863 io_cqring_ev_posted(ctx);
1870 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1872 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1874 return __io_req_find_next(req);
1877 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1881 if (ctx->submit_state.comp.nr) {
1882 mutex_lock(&ctx->uring_lock);
1883 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1884 mutex_unlock(&ctx->uring_lock);
1886 percpu_ref_put(&ctx->refs);
1889 static bool __tctx_task_work(struct io_uring_task *tctx)
1891 struct io_ring_ctx *ctx = NULL;
1892 struct io_wq_work_list list;
1893 struct io_wq_work_node *node;
1895 if (wq_list_empty(&tctx->task_list))
1898 spin_lock_irq(&tctx->task_lock);
1899 list = tctx->task_list;
1900 INIT_WQ_LIST(&tctx->task_list);
1901 spin_unlock_irq(&tctx->task_lock);
1905 struct io_wq_work_node *next = node->next;
1906 struct io_kiocb *req;
1908 req = container_of(node, struct io_kiocb, io_task_work.node);
1909 if (req->ctx != ctx) {
1910 ctx_flush_and_put(ctx);
1912 percpu_ref_get(&ctx->refs);
1915 req->task_work.func(&req->task_work);
1919 ctx_flush_and_put(ctx);
1920 return list.first != NULL;
1923 static void tctx_task_work(struct callback_head *cb)
1925 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1927 clear_bit(0, &tctx->task_state);
1929 while (__tctx_task_work(tctx))
1933 static int io_req_task_work_add(struct io_kiocb *req)
1935 struct task_struct *tsk = req->task;
1936 struct io_uring_task *tctx = tsk->io_uring;
1937 enum task_work_notify_mode notify;
1938 struct io_wq_work_node *node, *prev;
1939 unsigned long flags;
1942 if (unlikely(tsk->flags & PF_EXITING))
1945 WARN_ON_ONCE(!tctx);
1947 spin_lock_irqsave(&tctx->task_lock, flags);
1948 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1949 spin_unlock_irqrestore(&tctx->task_lock, flags);
1951 /* task_work already pending, we're done */
1952 if (test_bit(0, &tctx->task_state) ||
1953 test_and_set_bit(0, &tctx->task_state))
1957 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1958 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1959 * processing task_work. There's no reliable way to tell if TWA_RESUME
1962 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1964 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1965 wake_up_process(tsk);
1970 * Slow path - we failed, find and delete work. if the work is not
1971 * in the list, it got run and we're fine.
1973 spin_lock_irqsave(&tctx->task_lock, flags);
1974 wq_list_for_each(node, prev, &tctx->task_list) {
1975 if (&req->io_task_work.node == node) {
1976 wq_list_del(&tctx->task_list, node, prev);
1981 spin_unlock_irqrestore(&tctx->task_lock, flags);
1982 clear_bit(0, &tctx->task_state);
1986 static bool io_run_task_work_head(struct callback_head **work_head)
1988 struct callback_head *work, *next;
1989 bool executed = false;
1992 work = xchg(work_head, NULL);
2008 static void io_task_work_add_head(struct callback_head **work_head,
2009 struct callback_head *task_work)
2011 struct callback_head *head;
2014 head = READ_ONCE(*work_head);
2015 task_work->next = head;
2016 } while (cmpxchg(work_head, head, task_work) != head);
2019 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2020 task_work_func_t cb)
2022 init_task_work(&req->task_work, cb);
2023 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2026 static void io_req_task_cancel(struct callback_head *cb)
2028 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2029 struct io_ring_ctx *ctx = req->ctx;
2031 /* ctx is guaranteed to stay alive while we hold uring_lock */
2032 mutex_lock(&ctx->uring_lock);
2033 io_req_complete_failed(req, req->result);
2034 mutex_unlock(&ctx->uring_lock);
2037 static void __io_req_task_submit(struct io_kiocb *req)
2039 struct io_ring_ctx *ctx = req->ctx;
2041 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2042 mutex_lock(&ctx->uring_lock);
2043 if (!(current->flags & PF_EXITING) && !current->in_execve)
2044 __io_queue_sqe(req);
2046 io_req_complete_failed(req, -EFAULT);
2047 mutex_unlock(&ctx->uring_lock);
2050 static void io_req_task_submit(struct callback_head *cb)
2052 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2054 __io_req_task_submit(req);
2057 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2060 req->task_work.func = io_req_task_cancel;
2062 if (unlikely(io_req_task_work_add(req)))
2063 io_req_task_work_add_fallback(req, io_req_task_cancel);
2066 static void io_req_task_queue(struct io_kiocb *req)
2068 req->task_work.func = io_req_task_submit;
2070 if (unlikely(io_req_task_work_add(req)))
2071 io_req_task_queue_fail(req, -ECANCELED);
2074 static inline void io_queue_next(struct io_kiocb *req)
2076 struct io_kiocb *nxt = io_req_find_next(req);
2079 io_req_task_queue(nxt);
2082 static void io_free_req(struct io_kiocb *req)
2089 struct task_struct *task;
2094 static inline void io_init_req_batch(struct req_batch *rb)
2101 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2102 struct req_batch *rb)
2105 io_put_task(rb->task, rb->task_refs);
2107 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2110 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2111 struct io_submit_state *state)
2114 io_dismantle_req(req);
2116 if (req->task != rb->task) {
2118 io_put_task(rb->task, rb->task_refs);
2119 rb->task = req->task;
2125 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2126 state->reqs[state->free_reqs++] = req;
2128 list_add(&req->compl.list, &state->comp.free_list);
2131 static void io_submit_flush_completions(struct io_comp_state *cs,
2132 struct io_ring_ctx *ctx)
2135 struct io_kiocb *req;
2136 struct req_batch rb;
2138 io_init_req_batch(&rb);
2139 spin_lock_irq(&ctx->completion_lock);
2140 for (i = 0; i < nr; i++) {
2142 __io_cqring_fill_event(ctx, req->user_data, req->result,
2145 io_commit_cqring(ctx);
2146 spin_unlock_irq(&ctx->completion_lock);
2148 io_cqring_ev_posted(ctx);
2149 for (i = 0; i < nr; i++) {
2152 /* submission and completion refs */
2153 if (req_ref_sub_and_test(req, 2))
2154 io_req_free_batch(&rb, req, &ctx->submit_state);
2157 io_req_free_batch_finish(ctx, &rb);
2162 * Drop reference to request, return next in chain (if there is one) if this
2163 * was the last reference to this request.
2165 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2167 struct io_kiocb *nxt = NULL;
2169 if (req_ref_put_and_test(req)) {
2170 nxt = io_req_find_next(req);
2176 static inline void io_put_req(struct io_kiocb *req)
2178 if (req_ref_put_and_test(req))
2182 static void io_put_req_deferred_cb(struct callback_head *cb)
2184 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2189 static void io_free_req_deferred(struct io_kiocb *req)
2191 req->task_work.func = io_put_req_deferred_cb;
2192 if (unlikely(io_req_task_work_add(req)))
2193 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2196 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2198 if (req_ref_sub_and_test(req, refs))
2199 io_free_req_deferred(req);
2202 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2204 /* See comment at the top of this file */
2206 return __io_cqring_events(ctx);
2209 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2211 struct io_rings *rings = ctx->rings;
2213 /* make sure SQ entry isn't read before tail */
2214 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2217 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2219 unsigned int cflags;
2221 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2222 cflags |= IORING_CQE_F_BUFFER;
2223 req->flags &= ~REQ_F_BUFFER_SELECTED;
2228 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2230 struct io_buffer *kbuf;
2232 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2233 return io_put_kbuf(req, kbuf);
2236 static inline bool io_run_task_work(void)
2239 * Not safe to run on exiting task, and the task_work handling will
2240 * not add work to such a task.
2242 if (unlikely(current->flags & PF_EXITING))
2244 if (current->task_works) {
2245 __set_current_state(TASK_RUNNING);
2254 * Find and free completed poll iocbs
2256 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2257 struct list_head *done)
2259 struct req_batch rb;
2260 struct io_kiocb *req;
2262 /* order with ->result store in io_complete_rw_iopoll() */
2265 io_init_req_batch(&rb);
2266 while (!list_empty(done)) {
2269 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2270 list_del(&req->inflight_entry);
2272 if (READ_ONCE(req->result) == -EAGAIN &&
2273 !(req->flags & REQ_F_DONT_REISSUE)) {
2274 req->iopoll_completed = 0;
2276 io_queue_async_work(req);
2280 if (req->flags & REQ_F_BUFFER_SELECTED)
2281 cflags = io_put_rw_kbuf(req);
2283 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2286 if (req_ref_put_and_test(req))
2287 io_req_free_batch(&rb, req, &ctx->submit_state);
2290 io_commit_cqring(ctx);
2291 io_cqring_ev_posted_iopoll(ctx);
2292 io_req_free_batch_finish(ctx, &rb);
2295 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2298 struct io_kiocb *req, *tmp;
2304 * Only spin for completions if we don't have multiple devices hanging
2305 * off our complete list, and we're under the requested amount.
2307 spin = !ctx->poll_multi_file && *nr_events < min;
2310 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2311 struct kiocb *kiocb = &req->rw.kiocb;
2314 * Move completed and retryable entries to our local lists.
2315 * If we find a request that requires polling, break out
2316 * and complete those lists first, if we have entries there.
2318 if (READ_ONCE(req->iopoll_completed)) {
2319 list_move_tail(&req->inflight_entry, &done);
2322 if (!list_empty(&done))
2325 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2329 /* iopoll may have completed current req */
2330 if (READ_ONCE(req->iopoll_completed))
2331 list_move_tail(&req->inflight_entry, &done);
2338 if (!list_empty(&done))
2339 io_iopoll_complete(ctx, nr_events, &done);
2345 * We can't just wait for polled events to come to us, we have to actively
2346 * find and complete them.
2348 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2350 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2353 mutex_lock(&ctx->uring_lock);
2354 while (!list_empty(&ctx->iopoll_list)) {
2355 unsigned int nr_events = 0;
2357 io_do_iopoll(ctx, &nr_events, 0);
2359 /* let it sleep and repeat later if can't complete a request */
2363 * Ensure we allow local-to-the-cpu processing to take place,
2364 * in this case we need to ensure that we reap all events.
2365 * Also let task_work, etc. to progress by releasing the mutex
2367 if (need_resched()) {
2368 mutex_unlock(&ctx->uring_lock);
2370 mutex_lock(&ctx->uring_lock);
2373 mutex_unlock(&ctx->uring_lock);
2376 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2378 unsigned int nr_events = 0;
2382 * We disallow the app entering submit/complete with polling, but we
2383 * still need to lock the ring to prevent racing with polled issue
2384 * that got punted to a workqueue.
2386 mutex_lock(&ctx->uring_lock);
2388 * Don't enter poll loop if we already have events pending.
2389 * If we do, we can potentially be spinning for commands that
2390 * already triggered a CQE (eg in error).
2392 if (test_bit(0, &ctx->cq_check_overflow))
2393 __io_cqring_overflow_flush(ctx, false);
2394 if (io_cqring_events(ctx))
2398 * If a submit got punted to a workqueue, we can have the
2399 * application entering polling for a command before it gets
2400 * issued. That app will hold the uring_lock for the duration
2401 * of the poll right here, so we need to take a breather every
2402 * now and then to ensure that the issue has a chance to add
2403 * the poll to the issued list. Otherwise we can spin here
2404 * forever, while the workqueue is stuck trying to acquire the
2407 if (list_empty(&ctx->iopoll_list)) {
2408 mutex_unlock(&ctx->uring_lock);
2410 mutex_lock(&ctx->uring_lock);
2412 if (list_empty(&ctx->iopoll_list))
2415 ret = io_do_iopoll(ctx, &nr_events, min);
2416 } while (!ret && nr_events < min && !need_resched());
2418 mutex_unlock(&ctx->uring_lock);
2422 static void kiocb_end_write(struct io_kiocb *req)
2425 * Tell lockdep we inherited freeze protection from submission
2428 if (req->flags & REQ_F_ISREG) {
2429 struct super_block *sb = file_inode(req->file)->i_sb;
2431 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2437 static bool io_resubmit_prep(struct io_kiocb *req)
2439 struct io_async_rw *rw = req->async_data;
2442 return !io_req_prep_async(req);
2443 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2444 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2448 static bool io_rw_should_reissue(struct io_kiocb *req)
2450 umode_t mode = file_inode(req->file)->i_mode;
2451 struct io_ring_ctx *ctx = req->ctx;
2453 if (!S_ISBLK(mode) && !S_ISREG(mode))
2455 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2456 !(ctx->flags & IORING_SETUP_IOPOLL)))
2459 * If ref is dying, we might be running poll reap from the exit work.
2460 * Don't attempt to reissue from that path, just let it fail with
2463 if (percpu_ref_is_dying(&ctx->refs))
2468 static bool io_resubmit_prep(struct io_kiocb *req)
2472 static bool io_rw_should_reissue(struct io_kiocb *req)
2478 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2479 unsigned int issue_flags)
2483 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2484 kiocb_end_write(req);
2485 if (res != req->result) {
2486 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2487 io_rw_should_reissue(req)) {
2488 req->flags |= REQ_F_REISSUE;
2491 req_set_fail_links(req);
2493 if (req->flags & REQ_F_BUFFER_SELECTED)
2494 cflags = io_put_rw_kbuf(req);
2495 __io_req_complete(req, issue_flags, res, cflags);
2498 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2500 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2502 __io_complete_rw(req, res, res2, 0);
2505 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2507 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2509 if (kiocb->ki_flags & IOCB_WRITE)
2510 kiocb_end_write(req);
2511 if (unlikely(res != req->result)) {
2512 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2513 io_resubmit_prep(req))) {
2514 req_set_fail_links(req);
2515 req->flags |= REQ_F_DONT_REISSUE;
2519 WRITE_ONCE(req->result, res);
2520 /* order with io_iopoll_complete() checking ->result */
2522 WRITE_ONCE(req->iopoll_completed, 1);
2526 * After the iocb has been issued, it's safe to be found on the poll list.
2527 * Adding the kiocb to the list AFTER submission ensures that we don't
2528 * find it from a io_do_iopoll() thread before the issuer is done
2529 * accessing the kiocb cookie.
2531 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2533 struct io_ring_ctx *ctx = req->ctx;
2536 * Track whether we have multiple files in our lists. This will impact
2537 * how we do polling eventually, not spinning if we're on potentially
2538 * different devices.
2540 if (list_empty(&ctx->iopoll_list)) {
2541 ctx->poll_multi_file = false;
2542 } else if (!ctx->poll_multi_file) {
2543 struct io_kiocb *list_req;
2545 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2547 if (list_req->file != req->file)
2548 ctx->poll_multi_file = true;
2552 * For fast devices, IO may have already completed. If it has, add
2553 * it to the front so we find it first.
2555 if (READ_ONCE(req->iopoll_completed))
2556 list_add(&req->inflight_entry, &ctx->iopoll_list);
2558 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2561 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2562 * task context or in io worker task context. If current task context is
2563 * sq thread, we don't need to check whether should wake up sq thread.
2565 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2566 wq_has_sleeper(&ctx->sq_data->wait))
2567 wake_up(&ctx->sq_data->wait);
2570 static inline void io_state_file_put(struct io_submit_state *state)
2572 if (state->file_refs) {
2573 fput_many(state->file, state->file_refs);
2574 state->file_refs = 0;
2579 * Get as many references to a file as we have IOs left in this submission,
2580 * assuming most submissions are for one file, or at least that each file
2581 * has more than one submission.
2583 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2588 if (state->file_refs) {
2589 if (state->fd == fd) {
2593 io_state_file_put(state);
2595 state->file = fget_many(fd, state->ios_left);
2596 if (unlikely(!state->file))
2600 state->file_refs = state->ios_left - 1;
2604 static bool io_bdev_nowait(struct block_device *bdev)
2606 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2610 * If we tracked the file through the SCM inflight mechanism, we could support
2611 * any file. For now, just ensure that anything potentially problematic is done
2614 static bool __io_file_supports_async(struct file *file, int rw)
2616 umode_t mode = file_inode(file)->i_mode;
2618 if (S_ISBLK(mode)) {
2619 if (IS_ENABLED(CONFIG_BLOCK) &&
2620 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2624 if (S_ISCHR(mode) || S_ISSOCK(mode))
2626 if (S_ISREG(mode)) {
2627 if (IS_ENABLED(CONFIG_BLOCK) &&
2628 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2629 file->f_op != &io_uring_fops)
2634 /* any ->read/write should understand O_NONBLOCK */
2635 if (file->f_flags & O_NONBLOCK)
2638 if (!(file->f_mode & FMODE_NOWAIT))
2642 return file->f_op->read_iter != NULL;
2644 return file->f_op->write_iter != NULL;
2647 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2649 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2651 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2654 return __io_file_supports_async(req->file, rw);
2657 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2659 struct io_ring_ctx *ctx = req->ctx;
2660 struct kiocb *kiocb = &req->rw.kiocb;
2661 struct file *file = req->file;
2665 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2666 req->flags |= REQ_F_ISREG;
2668 kiocb->ki_pos = READ_ONCE(sqe->off);
2669 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2670 req->flags |= REQ_F_CUR_POS;
2671 kiocb->ki_pos = file->f_pos;
2673 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2674 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2675 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2679 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2680 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2681 req->flags |= REQ_F_NOWAIT;
2683 ioprio = READ_ONCE(sqe->ioprio);
2685 ret = ioprio_check_cap(ioprio);
2689 kiocb->ki_ioprio = ioprio;
2691 kiocb->ki_ioprio = get_current_ioprio();
2693 if (ctx->flags & IORING_SETUP_IOPOLL) {
2694 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2695 !kiocb->ki_filp->f_op->iopoll)
2698 kiocb->ki_flags |= IOCB_HIPRI;
2699 kiocb->ki_complete = io_complete_rw_iopoll;
2700 req->iopoll_completed = 0;
2702 if (kiocb->ki_flags & IOCB_HIPRI)
2704 kiocb->ki_complete = io_complete_rw;
2707 if (req->opcode == IORING_OP_READ_FIXED ||
2708 req->opcode == IORING_OP_WRITE_FIXED) {
2710 io_req_set_rsrc_node(req);
2713 req->rw.addr = READ_ONCE(sqe->addr);
2714 req->rw.len = READ_ONCE(sqe->len);
2715 req->buf_index = READ_ONCE(sqe->buf_index);
2719 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2725 case -ERESTARTNOINTR:
2726 case -ERESTARTNOHAND:
2727 case -ERESTART_RESTARTBLOCK:
2729 * We can't just restart the syscall, since previously
2730 * submitted sqes may already be in progress. Just fail this
2736 kiocb->ki_complete(kiocb, ret, 0);
2740 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2741 unsigned int issue_flags)
2743 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2744 struct io_async_rw *io = req->async_data;
2745 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2747 /* add previously done IO, if any */
2748 if (io && io->bytes_done > 0) {
2750 ret = io->bytes_done;
2752 ret += io->bytes_done;
2755 if (req->flags & REQ_F_CUR_POS)
2756 req->file->f_pos = kiocb->ki_pos;
2757 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2758 __io_complete_rw(req, ret, 0, issue_flags);
2760 io_rw_done(kiocb, ret);
2762 if (check_reissue && req->flags & REQ_F_REISSUE) {
2763 req->flags &= ~REQ_F_REISSUE;
2764 if (io_resubmit_prep(req)) {
2766 io_queue_async_work(req);
2770 req_set_fail_links(req);
2771 if (req->flags & REQ_F_BUFFER_SELECTED)
2772 cflags = io_put_rw_kbuf(req);
2773 __io_req_complete(req, issue_flags, ret, cflags);
2778 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2779 struct io_mapped_ubuf *imu)
2781 size_t len = req->rw.len;
2782 u64 buf_end, buf_addr = req->rw.addr;
2785 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2787 /* not inside the mapped region */
2788 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2792 * May not be a start of buffer, set size appropriately
2793 * and advance us to the beginning.
2795 offset = buf_addr - imu->ubuf;
2796 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2800 * Don't use iov_iter_advance() here, as it's really slow for
2801 * using the latter parts of a big fixed buffer - it iterates
2802 * over each segment manually. We can cheat a bit here, because
2805 * 1) it's a BVEC iter, we set it up
2806 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2807 * first and last bvec
2809 * So just find our index, and adjust the iterator afterwards.
2810 * If the offset is within the first bvec (or the whole first
2811 * bvec, just use iov_iter_advance(). This makes it easier
2812 * since we can just skip the first segment, which may not
2813 * be PAGE_SIZE aligned.
2815 const struct bio_vec *bvec = imu->bvec;
2817 if (offset <= bvec->bv_len) {
2818 iov_iter_advance(iter, offset);
2820 unsigned long seg_skip;
2822 /* skip first vec */
2823 offset -= bvec->bv_len;
2824 seg_skip = 1 + (offset >> PAGE_SHIFT);
2826 iter->bvec = bvec + seg_skip;
2827 iter->nr_segs -= seg_skip;
2828 iter->count -= bvec->bv_len + offset;
2829 iter->iov_offset = offset & ~PAGE_MASK;
2836 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2838 struct io_ring_ctx *ctx = req->ctx;
2839 struct io_mapped_ubuf *imu = req->imu;
2840 u16 index, buf_index = req->buf_index;
2843 if (unlikely(buf_index >= ctx->nr_user_bufs))
2845 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2846 imu = READ_ONCE(ctx->user_bufs[index]);
2849 return __io_import_fixed(req, rw, iter, imu);
2852 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2855 mutex_unlock(&ctx->uring_lock);
2858 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2861 * "Normal" inline submissions always hold the uring_lock, since we
2862 * grab it from the system call. Same is true for the SQPOLL offload.
2863 * The only exception is when we've detached the request and issue it
2864 * from an async worker thread, grab the lock for that case.
2867 mutex_lock(&ctx->uring_lock);
2870 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2871 int bgid, struct io_buffer *kbuf,
2874 struct io_buffer *head;
2876 if (req->flags & REQ_F_BUFFER_SELECTED)
2879 io_ring_submit_lock(req->ctx, needs_lock);
2881 lockdep_assert_held(&req->ctx->uring_lock);
2883 head = xa_load(&req->ctx->io_buffers, bgid);
2885 if (!list_empty(&head->list)) {
2886 kbuf = list_last_entry(&head->list, struct io_buffer,
2888 list_del(&kbuf->list);
2891 xa_erase(&req->ctx->io_buffers, bgid);
2893 if (*len > kbuf->len)
2896 kbuf = ERR_PTR(-ENOBUFS);
2899 io_ring_submit_unlock(req->ctx, needs_lock);
2904 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2907 struct io_buffer *kbuf;
2910 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2911 bgid = req->buf_index;
2912 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2915 req->rw.addr = (u64) (unsigned long) kbuf;
2916 req->flags |= REQ_F_BUFFER_SELECTED;
2917 return u64_to_user_ptr(kbuf->addr);
2920 #ifdef CONFIG_COMPAT
2921 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2924 struct compat_iovec __user *uiov;
2925 compat_ssize_t clen;
2929 uiov = u64_to_user_ptr(req->rw.addr);
2930 if (!access_ok(uiov, sizeof(*uiov)))
2932 if (__get_user(clen, &uiov->iov_len))
2938 buf = io_rw_buffer_select(req, &len, needs_lock);
2940 return PTR_ERR(buf);
2941 iov[0].iov_base = buf;
2942 iov[0].iov_len = (compat_size_t) len;
2947 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2950 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2954 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2957 len = iov[0].iov_len;
2960 buf = io_rw_buffer_select(req, &len, needs_lock);
2962 return PTR_ERR(buf);
2963 iov[0].iov_base = buf;
2964 iov[0].iov_len = len;
2968 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2971 if (req->flags & REQ_F_BUFFER_SELECTED) {
2972 struct io_buffer *kbuf;
2974 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2975 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2976 iov[0].iov_len = kbuf->len;
2979 if (req->rw.len != 1)
2982 #ifdef CONFIG_COMPAT
2983 if (req->ctx->compat)
2984 return io_compat_import(req, iov, needs_lock);
2987 return __io_iov_buffer_select(req, iov, needs_lock);
2990 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2991 struct iov_iter *iter, bool needs_lock)
2993 void __user *buf = u64_to_user_ptr(req->rw.addr);
2994 size_t sqe_len = req->rw.len;
2995 u8 opcode = req->opcode;
2998 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3000 return io_import_fixed(req, rw, iter);
3003 /* buffer index only valid with fixed read/write, or buffer select */
3004 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3007 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3008 if (req->flags & REQ_F_BUFFER_SELECT) {
3009 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3011 return PTR_ERR(buf);
3012 req->rw.len = sqe_len;
3015 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3020 if (req->flags & REQ_F_BUFFER_SELECT) {
3021 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3023 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3028 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3032 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3034 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3038 * For files that don't have ->read_iter() and ->write_iter(), handle them
3039 * by looping over ->read() or ->write() manually.
3041 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3043 struct kiocb *kiocb = &req->rw.kiocb;
3044 struct file *file = req->file;
3048 * Don't support polled IO through this interface, and we can't
3049 * support non-blocking either. For the latter, this just causes
3050 * the kiocb to be handled from an async context.
3052 if (kiocb->ki_flags & IOCB_HIPRI)
3054 if (kiocb->ki_flags & IOCB_NOWAIT)
3057 while (iov_iter_count(iter)) {
3061 if (!iov_iter_is_bvec(iter)) {
3062 iovec = iov_iter_iovec(iter);
3064 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3065 iovec.iov_len = req->rw.len;
3069 nr = file->f_op->read(file, iovec.iov_base,
3070 iovec.iov_len, io_kiocb_ppos(kiocb));
3072 nr = file->f_op->write(file, iovec.iov_base,
3073 iovec.iov_len, io_kiocb_ppos(kiocb));
3082 if (nr != iovec.iov_len)
3086 iov_iter_advance(iter, nr);
3092 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3093 const struct iovec *fast_iov, struct iov_iter *iter)
3095 struct io_async_rw *rw = req->async_data;
3097 memcpy(&rw->iter, iter, sizeof(*iter));
3098 rw->free_iovec = iovec;
3100 /* can only be fixed buffers, no need to do anything */
3101 if (iov_iter_is_bvec(iter))
3104 unsigned iov_off = 0;
3106 rw->iter.iov = rw->fast_iov;
3107 if (iter->iov != fast_iov) {
3108 iov_off = iter->iov - fast_iov;
3109 rw->iter.iov += iov_off;
3111 if (rw->fast_iov != fast_iov)
3112 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3113 sizeof(struct iovec) * iter->nr_segs);
3115 req->flags |= REQ_F_NEED_CLEANUP;
3119 static inline int io_alloc_async_data(struct io_kiocb *req)
3121 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3122 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3123 return req->async_data == NULL;
3126 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3127 const struct iovec *fast_iov,
3128 struct iov_iter *iter, bool force)
3130 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3132 if (!req->async_data) {
3133 if (io_alloc_async_data(req)) {
3138 io_req_map_rw(req, iovec, fast_iov, iter);
3143 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3145 struct io_async_rw *iorw = req->async_data;
3146 struct iovec *iov = iorw->fast_iov;
3149 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3150 if (unlikely(ret < 0))
3153 iorw->bytes_done = 0;
3154 iorw->free_iovec = iov;
3156 req->flags |= REQ_F_NEED_CLEANUP;
3160 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3162 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3164 return io_prep_rw(req, sqe);
3168 * This is our waitqueue callback handler, registered through lock_page_async()
3169 * when we initially tried to do the IO with the iocb armed our waitqueue.
3170 * This gets called when the page is unlocked, and we generally expect that to
3171 * happen when the page IO is completed and the page is now uptodate. This will
3172 * queue a task_work based retry of the operation, attempting to copy the data
3173 * again. If the latter fails because the page was NOT uptodate, then we will
3174 * do a thread based blocking retry of the operation. That's the unexpected
3177 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3178 int sync, void *arg)
3180 struct wait_page_queue *wpq;
3181 struct io_kiocb *req = wait->private;
3182 struct wait_page_key *key = arg;
3184 wpq = container_of(wait, struct wait_page_queue, wait);
3186 if (!wake_page_match(wpq, key))
3189 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3190 list_del_init(&wait->entry);
3192 /* submit ref gets dropped, acquire a new one */
3194 io_req_task_queue(req);
3199 * This controls whether a given IO request should be armed for async page
3200 * based retry. If we return false here, the request is handed to the async
3201 * worker threads for retry. If we're doing buffered reads on a regular file,
3202 * we prepare a private wait_page_queue entry and retry the operation. This
3203 * will either succeed because the page is now uptodate and unlocked, or it
3204 * will register a callback when the page is unlocked at IO completion. Through
3205 * that callback, io_uring uses task_work to setup a retry of the operation.
3206 * That retry will attempt the buffered read again. The retry will generally
3207 * succeed, or in rare cases where it fails, we then fall back to using the
3208 * async worker threads for a blocking retry.
3210 static bool io_rw_should_retry(struct io_kiocb *req)
3212 struct io_async_rw *rw = req->async_data;
3213 struct wait_page_queue *wait = &rw->wpq;
3214 struct kiocb *kiocb = &req->rw.kiocb;
3216 /* never retry for NOWAIT, we just complete with -EAGAIN */
3217 if (req->flags & REQ_F_NOWAIT)
3220 /* Only for buffered IO */
3221 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3225 * just use poll if we can, and don't attempt if the fs doesn't
3226 * support callback based unlocks
3228 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3231 wait->wait.func = io_async_buf_func;
3232 wait->wait.private = req;
3233 wait->wait.flags = 0;
3234 INIT_LIST_HEAD(&wait->wait.entry);
3235 kiocb->ki_flags |= IOCB_WAITQ;
3236 kiocb->ki_flags &= ~IOCB_NOWAIT;
3237 kiocb->ki_waitq = wait;
3241 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3243 if (req->file->f_op->read_iter)
3244 return call_read_iter(req->file, &req->rw.kiocb, iter);
3245 else if (req->file->f_op->read)
3246 return loop_rw_iter(READ, req, iter);
3251 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3253 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3254 struct kiocb *kiocb = &req->rw.kiocb;
3255 struct iov_iter __iter, *iter = &__iter;
3256 struct io_async_rw *rw = req->async_data;
3257 ssize_t io_size, ret, ret2;
3258 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3264 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3268 io_size = iov_iter_count(iter);
3269 req->result = io_size;
3271 /* Ensure we clear previously set non-block flag */
3272 if (!force_nonblock)
3273 kiocb->ki_flags &= ~IOCB_NOWAIT;
3275 kiocb->ki_flags |= IOCB_NOWAIT;
3277 /* If the file doesn't support async, just async punt */
3278 if (force_nonblock && !io_file_supports_async(req, READ)) {
3279 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3280 return ret ?: -EAGAIN;
3283 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3284 if (unlikely(ret)) {
3289 ret = io_iter_do_read(req, iter);
3291 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3292 req->flags &= ~REQ_F_REISSUE;
3293 /* IOPOLL retry should happen for io-wq threads */
3294 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3296 /* no retry on NONBLOCK nor RWF_NOWAIT */
3297 if (req->flags & REQ_F_NOWAIT)
3299 /* some cases will consume bytes even on error returns */
3300 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3302 } else if (ret == -EIOCBQUEUED) {
3304 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3305 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3306 /* read all, failed, already did sync or don't want to retry */
3310 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3315 rw = req->async_data;
3316 /* now use our persistent iterator, if we aren't already */
3321 rw->bytes_done += ret;
3322 /* if we can retry, do so with the callbacks armed */
3323 if (!io_rw_should_retry(req)) {
3324 kiocb->ki_flags &= ~IOCB_WAITQ;
3329 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3330 * we get -EIOCBQUEUED, then we'll get a notification when the
3331 * desired page gets unlocked. We can also get a partial read
3332 * here, and if we do, then just retry at the new offset.
3334 ret = io_iter_do_read(req, iter);
3335 if (ret == -EIOCBQUEUED)
3337 /* we got some bytes, but not all. retry. */
3338 kiocb->ki_flags &= ~IOCB_WAITQ;
3339 } while (ret > 0 && ret < io_size);
3341 kiocb_done(kiocb, ret, issue_flags);
3343 /* it's faster to check here then delegate to kfree */
3349 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3351 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3353 return io_prep_rw(req, sqe);
3356 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3358 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3359 struct kiocb *kiocb = &req->rw.kiocb;
3360 struct iov_iter __iter, *iter = &__iter;
3361 struct io_async_rw *rw = req->async_data;
3362 ssize_t ret, ret2, io_size;
3363 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3369 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3373 io_size = iov_iter_count(iter);
3374 req->result = io_size;
3376 /* Ensure we clear previously set non-block flag */
3377 if (!force_nonblock)
3378 kiocb->ki_flags &= ~IOCB_NOWAIT;
3380 kiocb->ki_flags |= IOCB_NOWAIT;
3382 /* If the file doesn't support async, just async punt */
3383 if (force_nonblock && !io_file_supports_async(req, WRITE))
3386 /* file path doesn't support NOWAIT for non-direct_IO */
3387 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3388 (req->flags & REQ_F_ISREG))
3391 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3396 * Open-code file_start_write here to grab freeze protection,
3397 * which will be released by another thread in
3398 * io_complete_rw(). Fool lockdep by telling it the lock got
3399 * released so that it doesn't complain about the held lock when
3400 * we return to userspace.
3402 if (req->flags & REQ_F_ISREG) {
3403 sb_start_write(file_inode(req->file)->i_sb);
3404 __sb_writers_release(file_inode(req->file)->i_sb,
3407 kiocb->ki_flags |= IOCB_WRITE;
3409 if (req->file->f_op->write_iter)
3410 ret2 = call_write_iter(req->file, kiocb, iter);
3411 else if (req->file->f_op->write)
3412 ret2 = loop_rw_iter(WRITE, req, iter);
3416 if (req->flags & REQ_F_REISSUE) {
3417 req->flags &= ~REQ_F_REISSUE;
3422 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3423 * retry them without IOCB_NOWAIT.
3425 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3427 /* no retry on NONBLOCK nor RWF_NOWAIT */
3428 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3430 if (!force_nonblock || ret2 != -EAGAIN) {
3431 /* IOPOLL retry should happen for io-wq threads */
3432 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3435 kiocb_done(kiocb, ret2, issue_flags);
3438 /* some cases will consume bytes even on error returns */
3439 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3440 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3441 return ret ?: -EAGAIN;
3444 /* it's reportedly faster than delegating the null check to kfree() */
3450 static int io_renameat_prep(struct io_kiocb *req,
3451 const struct io_uring_sqe *sqe)
3453 struct io_rename *ren = &req->rename;
3454 const char __user *oldf, *newf;
3456 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3459 ren->old_dfd = READ_ONCE(sqe->fd);
3460 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3461 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3462 ren->new_dfd = READ_ONCE(sqe->len);
3463 ren->flags = READ_ONCE(sqe->rename_flags);
3465 ren->oldpath = getname(oldf);
3466 if (IS_ERR(ren->oldpath))
3467 return PTR_ERR(ren->oldpath);
3469 ren->newpath = getname(newf);
3470 if (IS_ERR(ren->newpath)) {
3471 putname(ren->oldpath);
3472 return PTR_ERR(ren->newpath);
3475 req->flags |= REQ_F_NEED_CLEANUP;
3479 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3481 struct io_rename *ren = &req->rename;
3484 if (issue_flags & IO_URING_F_NONBLOCK)
3487 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3488 ren->newpath, ren->flags);
3490 req->flags &= ~REQ_F_NEED_CLEANUP;
3492 req_set_fail_links(req);
3493 io_req_complete(req, ret);
3497 static int io_unlinkat_prep(struct io_kiocb *req,
3498 const struct io_uring_sqe *sqe)
3500 struct io_unlink *un = &req->unlink;
3501 const char __user *fname;
3503 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3506 un->dfd = READ_ONCE(sqe->fd);
3508 un->flags = READ_ONCE(sqe->unlink_flags);
3509 if (un->flags & ~AT_REMOVEDIR)
3512 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3513 un->filename = getname(fname);
3514 if (IS_ERR(un->filename))
3515 return PTR_ERR(un->filename);
3517 req->flags |= REQ_F_NEED_CLEANUP;
3521 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3523 struct io_unlink *un = &req->unlink;
3526 if (issue_flags & IO_URING_F_NONBLOCK)
3529 if (un->flags & AT_REMOVEDIR)
3530 ret = do_rmdir(un->dfd, un->filename);
3532 ret = do_unlinkat(un->dfd, un->filename);
3534 req->flags &= ~REQ_F_NEED_CLEANUP;
3536 req_set_fail_links(req);
3537 io_req_complete(req, ret);
3541 static int io_shutdown_prep(struct io_kiocb *req,
3542 const struct io_uring_sqe *sqe)
3544 #if defined(CONFIG_NET)
3545 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3547 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3551 req->shutdown.how = READ_ONCE(sqe->len);
3558 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3560 #if defined(CONFIG_NET)
3561 struct socket *sock;
3564 if (issue_flags & IO_URING_F_NONBLOCK)
3567 sock = sock_from_file(req->file);
3568 if (unlikely(!sock))
3571 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3573 req_set_fail_links(req);
3574 io_req_complete(req, ret);
3581 static int __io_splice_prep(struct io_kiocb *req,
3582 const struct io_uring_sqe *sqe)
3584 struct io_splice* sp = &req->splice;
3585 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3587 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3591 sp->len = READ_ONCE(sqe->len);
3592 sp->flags = READ_ONCE(sqe->splice_flags);
3594 if (unlikely(sp->flags & ~valid_flags))
3597 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3598 (sp->flags & SPLICE_F_FD_IN_FIXED));
3601 req->flags |= REQ_F_NEED_CLEANUP;
3605 static int io_tee_prep(struct io_kiocb *req,
3606 const struct io_uring_sqe *sqe)
3608 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3610 return __io_splice_prep(req, sqe);
3613 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3615 struct io_splice *sp = &req->splice;
3616 struct file *in = sp->file_in;
3617 struct file *out = sp->file_out;
3618 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3621 if (issue_flags & IO_URING_F_NONBLOCK)
3624 ret = do_tee(in, out, sp->len, flags);
3626 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3628 req->flags &= ~REQ_F_NEED_CLEANUP;
3631 req_set_fail_links(req);
3632 io_req_complete(req, ret);
3636 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3638 struct io_splice* sp = &req->splice;
3640 sp->off_in = READ_ONCE(sqe->splice_off_in);
3641 sp->off_out = READ_ONCE(sqe->off);
3642 return __io_splice_prep(req, sqe);
3645 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3647 struct io_splice *sp = &req->splice;
3648 struct file *in = sp->file_in;
3649 struct file *out = sp->file_out;
3650 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3651 loff_t *poff_in, *poff_out;
3654 if (issue_flags & IO_URING_F_NONBLOCK)
3657 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3658 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3661 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3663 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3665 req->flags &= ~REQ_F_NEED_CLEANUP;
3668 req_set_fail_links(req);
3669 io_req_complete(req, ret);
3674 * IORING_OP_NOP just posts a completion event, nothing else.
3676 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3678 struct io_ring_ctx *ctx = req->ctx;
3680 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3683 __io_req_complete(req, issue_flags, 0, 0);
3687 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3689 struct io_ring_ctx *ctx = req->ctx;
3694 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3696 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3699 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3700 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3703 req->sync.off = READ_ONCE(sqe->off);
3704 req->sync.len = READ_ONCE(sqe->len);
3708 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3710 loff_t end = req->sync.off + req->sync.len;
3713 /* fsync always requires a blocking context */
3714 if (issue_flags & IO_URING_F_NONBLOCK)
3717 ret = vfs_fsync_range(req->file, req->sync.off,
3718 end > 0 ? end : LLONG_MAX,
3719 req->sync.flags & IORING_FSYNC_DATASYNC);
3721 req_set_fail_links(req);
3722 io_req_complete(req, ret);
3726 static int io_fallocate_prep(struct io_kiocb *req,
3727 const struct io_uring_sqe *sqe)
3729 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3731 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3734 req->sync.off = READ_ONCE(sqe->off);
3735 req->sync.len = READ_ONCE(sqe->addr);
3736 req->sync.mode = READ_ONCE(sqe->len);
3740 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3744 /* fallocate always requiring blocking context */
3745 if (issue_flags & IO_URING_F_NONBLOCK)
3747 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3750 req_set_fail_links(req);
3751 io_req_complete(req, ret);
3755 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3757 const char __user *fname;
3760 if (unlikely(sqe->ioprio || sqe->buf_index))
3762 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3765 /* open.how should be already initialised */
3766 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3767 req->open.how.flags |= O_LARGEFILE;
3769 req->open.dfd = READ_ONCE(sqe->fd);
3770 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3771 req->open.filename = getname(fname);
3772 if (IS_ERR(req->open.filename)) {
3773 ret = PTR_ERR(req->open.filename);
3774 req->open.filename = NULL;
3777 req->open.nofile = rlimit(RLIMIT_NOFILE);
3778 req->flags |= REQ_F_NEED_CLEANUP;
3782 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3786 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3788 mode = READ_ONCE(sqe->len);
3789 flags = READ_ONCE(sqe->open_flags);
3790 req->open.how = build_open_how(flags, mode);
3791 return __io_openat_prep(req, sqe);
3794 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3796 struct open_how __user *how;
3800 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3802 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3803 len = READ_ONCE(sqe->len);
3804 if (len < OPEN_HOW_SIZE_VER0)
3807 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3812 return __io_openat_prep(req, sqe);
3815 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3817 struct open_flags op;
3820 bool resolve_nonblock;
3823 ret = build_open_flags(&req->open.how, &op);
3826 nonblock_set = op.open_flag & O_NONBLOCK;
3827 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3828 if (issue_flags & IO_URING_F_NONBLOCK) {
3830 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3831 * it'll always -EAGAIN
3833 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3835 op.lookup_flags |= LOOKUP_CACHED;
3836 op.open_flag |= O_NONBLOCK;
3839 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3843 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3844 /* only retry if RESOLVE_CACHED wasn't already set by application */
3845 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3846 file == ERR_PTR(-EAGAIN)) {
3848 * We could hang on to this 'fd', but seems like marginal
3849 * gain for something that is now known to be a slower path.
3850 * So just put it, and we'll get a new one when we retry.
3858 ret = PTR_ERR(file);
3860 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3861 file->f_flags &= ~O_NONBLOCK;
3862 fsnotify_open(file);
3863 fd_install(ret, file);
3866 putname(req->open.filename);
3867 req->flags &= ~REQ_F_NEED_CLEANUP;
3869 req_set_fail_links(req);
3870 __io_req_complete(req, issue_flags, ret, 0);
3874 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3876 return io_openat2(req, issue_flags);
3879 static int io_remove_buffers_prep(struct io_kiocb *req,
3880 const struct io_uring_sqe *sqe)
3882 struct io_provide_buf *p = &req->pbuf;
3885 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3888 tmp = READ_ONCE(sqe->fd);
3889 if (!tmp || tmp > USHRT_MAX)
3892 memset(p, 0, sizeof(*p));
3894 p->bgid = READ_ONCE(sqe->buf_group);
3898 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3899 int bgid, unsigned nbufs)
3903 /* shouldn't happen */
3907 /* the head kbuf is the list itself */
3908 while (!list_empty(&buf->list)) {
3909 struct io_buffer *nxt;
3911 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3912 list_del(&nxt->list);
3919 xa_erase(&ctx->io_buffers, bgid);
3924 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3926 struct io_provide_buf *p = &req->pbuf;
3927 struct io_ring_ctx *ctx = req->ctx;
3928 struct io_buffer *head;
3930 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3932 io_ring_submit_lock(ctx, !force_nonblock);
3934 lockdep_assert_held(&ctx->uring_lock);
3937 head = xa_load(&ctx->io_buffers, p->bgid);
3939 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3941 req_set_fail_links(req);
3943 /* complete before unlock, IOPOLL may need the lock */
3944 __io_req_complete(req, issue_flags, ret, 0);
3945 io_ring_submit_unlock(ctx, !force_nonblock);
3949 static int io_provide_buffers_prep(struct io_kiocb *req,
3950 const struct io_uring_sqe *sqe)
3952 unsigned long size, tmp_check;
3953 struct io_provide_buf *p = &req->pbuf;
3956 if (sqe->ioprio || sqe->rw_flags)
3959 tmp = READ_ONCE(sqe->fd);
3960 if (!tmp || tmp > USHRT_MAX)
3963 p->addr = READ_ONCE(sqe->addr);
3964 p->len = READ_ONCE(sqe->len);
3966 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3969 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3972 size = (unsigned long)p->len * p->nbufs;
3973 if (!access_ok(u64_to_user_ptr(p->addr), size))
3976 p->bgid = READ_ONCE(sqe->buf_group);
3977 tmp = READ_ONCE(sqe->off);
3978 if (tmp > USHRT_MAX)
3984 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3986 struct io_buffer *buf;
3987 u64 addr = pbuf->addr;
3988 int i, bid = pbuf->bid;
3990 for (i = 0; i < pbuf->nbufs; i++) {
3991 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3996 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4001 INIT_LIST_HEAD(&buf->list);
4004 list_add_tail(&buf->list, &(*head)->list);
4008 return i ? i : -ENOMEM;
4011 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4013 struct io_provide_buf *p = &req->pbuf;
4014 struct io_ring_ctx *ctx = req->ctx;
4015 struct io_buffer *head, *list;
4017 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4019 io_ring_submit_lock(ctx, !force_nonblock);
4021 lockdep_assert_held(&ctx->uring_lock);
4023 list = head = xa_load(&ctx->io_buffers, p->bgid);
4025 ret = io_add_buffers(p, &head);
4026 if (ret >= 0 && !list) {
4027 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4029 __io_remove_buffers(ctx, head, p->bgid, -1U);
4032 req_set_fail_links(req);
4033 /* complete before unlock, IOPOLL may need the lock */
4034 __io_req_complete(req, issue_flags, ret, 0);
4035 io_ring_submit_unlock(ctx, !force_nonblock);
4039 static int io_epoll_ctl_prep(struct io_kiocb *req,
4040 const struct io_uring_sqe *sqe)
4042 #if defined(CONFIG_EPOLL)
4043 if (sqe->ioprio || sqe->buf_index)
4045 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4048 req->epoll.epfd = READ_ONCE(sqe->fd);
4049 req->epoll.op = READ_ONCE(sqe->len);
4050 req->epoll.fd = READ_ONCE(sqe->off);
4052 if (ep_op_has_event(req->epoll.op)) {
4053 struct epoll_event __user *ev;
4055 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4056 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4066 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4068 #if defined(CONFIG_EPOLL)
4069 struct io_epoll *ie = &req->epoll;
4071 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4073 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4074 if (force_nonblock && ret == -EAGAIN)
4078 req_set_fail_links(req);
4079 __io_req_complete(req, issue_flags, ret, 0);
4086 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4088 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4089 if (sqe->ioprio || sqe->buf_index || sqe->off)
4091 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4094 req->madvise.addr = READ_ONCE(sqe->addr);
4095 req->madvise.len = READ_ONCE(sqe->len);
4096 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4103 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4105 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4106 struct io_madvise *ma = &req->madvise;
4109 if (issue_flags & IO_URING_F_NONBLOCK)
4112 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4114 req_set_fail_links(req);
4115 io_req_complete(req, ret);
4122 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4124 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4126 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4129 req->fadvise.offset = READ_ONCE(sqe->off);
4130 req->fadvise.len = READ_ONCE(sqe->len);
4131 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4135 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4137 struct io_fadvise *fa = &req->fadvise;
4140 if (issue_flags & IO_URING_F_NONBLOCK) {
4141 switch (fa->advice) {
4142 case POSIX_FADV_NORMAL:
4143 case POSIX_FADV_RANDOM:
4144 case POSIX_FADV_SEQUENTIAL:
4151 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4153 req_set_fail_links(req);
4154 __io_req_complete(req, issue_flags, ret, 0);
4158 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4160 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4162 if (sqe->ioprio || sqe->buf_index)
4164 if (req->flags & REQ_F_FIXED_FILE)
4167 req->statx.dfd = READ_ONCE(sqe->fd);
4168 req->statx.mask = READ_ONCE(sqe->len);
4169 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4170 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4171 req->statx.flags = READ_ONCE(sqe->statx_flags);
4176 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4178 struct io_statx *ctx = &req->statx;
4181 if (issue_flags & IO_URING_F_NONBLOCK)
4184 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4188 req_set_fail_links(req);
4189 io_req_complete(req, ret);
4193 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4195 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4197 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4198 sqe->rw_flags || sqe->buf_index)
4200 if (req->flags & REQ_F_FIXED_FILE)
4203 req->close.fd = READ_ONCE(sqe->fd);
4207 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4209 struct files_struct *files = current->files;
4210 struct io_close *close = &req->close;
4211 struct fdtable *fdt;
4212 struct file *file = NULL;
4215 spin_lock(&files->file_lock);
4216 fdt = files_fdtable(files);
4217 if (close->fd >= fdt->max_fds) {
4218 spin_unlock(&files->file_lock);
4221 file = fdt->fd[close->fd];
4222 if (!file || file->f_op == &io_uring_fops) {
4223 spin_unlock(&files->file_lock);
4228 /* if the file has a flush method, be safe and punt to async */
4229 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4230 spin_unlock(&files->file_lock);
4234 ret = __close_fd_get_file(close->fd, &file);
4235 spin_unlock(&files->file_lock);
4242 /* No ->flush() or already async, safely close from here */
4243 ret = filp_close(file, current->files);
4246 req_set_fail_links(req);
4249 __io_req_complete(req, issue_flags, ret, 0);
4253 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4255 struct io_ring_ctx *ctx = req->ctx;
4257 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4259 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4262 req->sync.off = READ_ONCE(sqe->off);
4263 req->sync.len = READ_ONCE(sqe->len);
4264 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4268 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4272 /* sync_file_range always requires a blocking context */
4273 if (issue_flags & IO_URING_F_NONBLOCK)
4276 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4279 req_set_fail_links(req);
4280 io_req_complete(req, ret);
4284 #if defined(CONFIG_NET)
4285 static int io_setup_async_msg(struct io_kiocb *req,
4286 struct io_async_msghdr *kmsg)
4288 struct io_async_msghdr *async_msg = req->async_data;
4292 if (io_alloc_async_data(req)) {
4293 kfree(kmsg->free_iov);
4296 async_msg = req->async_data;
4297 req->flags |= REQ_F_NEED_CLEANUP;
4298 memcpy(async_msg, kmsg, sizeof(*kmsg));
4299 async_msg->msg.msg_name = &async_msg->addr;
4300 /* if were using fast_iov, set it to the new one */
4301 if (!async_msg->free_iov)
4302 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4307 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4308 struct io_async_msghdr *iomsg)
4310 iomsg->msg.msg_name = &iomsg->addr;
4311 iomsg->free_iov = iomsg->fast_iov;
4312 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4313 req->sr_msg.msg_flags, &iomsg->free_iov);
4316 static int io_sendmsg_prep_async(struct io_kiocb *req)
4320 ret = io_sendmsg_copy_hdr(req, req->async_data);
4322 req->flags |= REQ_F_NEED_CLEANUP;
4326 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4328 struct io_sr_msg *sr = &req->sr_msg;
4330 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4333 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4334 sr->len = READ_ONCE(sqe->len);
4335 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4336 if (sr->msg_flags & MSG_DONTWAIT)
4337 req->flags |= REQ_F_NOWAIT;
4339 #ifdef CONFIG_COMPAT
4340 if (req->ctx->compat)
4341 sr->msg_flags |= MSG_CMSG_COMPAT;
4346 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4348 struct io_async_msghdr iomsg, *kmsg;
4349 struct socket *sock;
4354 sock = sock_from_file(req->file);
4355 if (unlikely(!sock))
4358 kmsg = req->async_data;
4360 ret = io_sendmsg_copy_hdr(req, &iomsg);
4366 flags = req->sr_msg.msg_flags;
4367 if (issue_flags & IO_URING_F_NONBLOCK)
4368 flags |= MSG_DONTWAIT;
4369 if (flags & MSG_WAITALL)
4370 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4372 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4373 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4374 return io_setup_async_msg(req, kmsg);
4375 if (ret == -ERESTARTSYS)
4378 /* fast path, check for non-NULL to avoid function call */
4380 kfree(kmsg->free_iov);
4381 req->flags &= ~REQ_F_NEED_CLEANUP;
4383 req_set_fail_links(req);
4384 __io_req_complete(req, issue_flags, ret, 0);
4388 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4390 struct io_sr_msg *sr = &req->sr_msg;
4393 struct socket *sock;
4398 sock = sock_from_file(req->file);
4399 if (unlikely(!sock))
4402 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4406 msg.msg_name = NULL;
4407 msg.msg_control = NULL;
4408 msg.msg_controllen = 0;
4409 msg.msg_namelen = 0;
4411 flags = req->sr_msg.msg_flags;
4412 if (issue_flags & IO_URING_F_NONBLOCK)
4413 flags |= MSG_DONTWAIT;
4414 if (flags & MSG_WAITALL)
4415 min_ret = iov_iter_count(&msg.msg_iter);
4417 msg.msg_flags = flags;
4418 ret = sock_sendmsg(sock, &msg);
4419 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4421 if (ret == -ERESTARTSYS)
4425 req_set_fail_links(req);
4426 __io_req_complete(req, issue_flags, ret, 0);
4430 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4431 struct io_async_msghdr *iomsg)
4433 struct io_sr_msg *sr = &req->sr_msg;
4434 struct iovec __user *uiov;
4438 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4439 &iomsg->uaddr, &uiov, &iov_len);
4443 if (req->flags & REQ_F_BUFFER_SELECT) {
4446 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4448 sr->len = iomsg->fast_iov[0].iov_len;
4449 iomsg->free_iov = NULL;
4451 iomsg->free_iov = iomsg->fast_iov;
4452 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4453 &iomsg->free_iov, &iomsg->msg.msg_iter,
4462 #ifdef CONFIG_COMPAT
4463 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4464 struct io_async_msghdr *iomsg)
4466 struct io_sr_msg *sr = &req->sr_msg;
4467 struct compat_iovec __user *uiov;
4472 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4477 uiov = compat_ptr(ptr);
4478 if (req->flags & REQ_F_BUFFER_SELECT) {
4479 compat_ssize_t clen;
4483 if (!access_ok(uiov, sizeof(*uiov)))
4485 if (__get_user(clen, &uiov->iov_len))
4490 iomsg->free_iov = NULL;
4492 iomsg->free_iov = iomsg->fast_iov;
4493 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4494 UIO_FASTIOV, &iomsg->free_iov,
4495 &iomsg->msg.msg_iter, true);
4504 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4505 struct io_async_msghdr *iomsg)
4507 iomsg->msg.msg_name = &iomsg->addr;
4509 #ifdef CONFIG_COMPAT
4510 if (req->ctx->compat)
4511 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4514 return __io_recvmsg_copy_hdr(req, iomsg);
4517 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4520 struct io_sr_msg *sr = &req->sr_msg;
4521 struct io_buffer *kbuf;
4523 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4528 req->flags |= REQ_F_BUFFER_SELECTED;
4532 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4534 return io_put_kbuf(req, req->sr_msg.kbuf);
4537 static int io_recvmsg_prep_async(struct io_kiocb *req)
4541 ret = io_recvmsg_copy_hdr(req, req->async_data);
4543 req->flags |= REQ_F_NEED_CLEANUP;
4547 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4549 struct io_sr_msg *sr = &req->sr_msg;
4551 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
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);
4557 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4558 if (sr->msg_flags & MSG_DONTWAIT)
4559 req->flags |= REQ_F_NOWAIT;
4561 #ifdef CONFIG_COMPAT
4562 if (req->ctx->compat)
4563 sr->msg_flags |= MSG_CMSG_COMPAT;
4568 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4570 struct io_async_msghdr iomsg, *kmsg;
4571 struct socket *sock;
4572 struct io_buffer *kbuf;
4575 int ret, cflags = 0;
4576 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4578 sock = sock_from_file(req->file);
4579 if (unlikely(!sock))
4582 kmsg = req->async_data;
4584 ret = io_recvmsg_copy_hdr(req, &iomsg);
4590 if (req->flags & REQ_F_BUFFER_SELECT) {
4591 kbuf = io_recv_buffer_select(req, !force_nonblock);
4593 return PTR_ERR(kbuf);
4594 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4595 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4596 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4597 1, req->sr_msg.len);
4600 flags = req->sr_msg.msg_flags;
4602 flags |= MSG_DONTWAIT;
4603 if (flags & MSG_WAITALL)
4604 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4606 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4607 kmsg->uaddr, flags);
4608 if (force_nonblock && ret == -EAGAIN)
4609 return io_setup_async_msg(req, kmsg);
4610 if (ret == -ERESTARTSYS)
4613 if (req->flags & REQ_F_BUFFER_SELECTED)
4614 cflags = io_put_recv_kbuf(req);
4615 /* fast path, check for non-NULL to avoid function call */
4617 kfree(kmsg->free_iov);
4618 req->flags &= ~REQ_F_NEED_CLEANUP;
4619 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4620 req_set_fail_links(req);
4621 __io_req_complete(req, issue_flags, ret, cflags);
4625 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4627 struct io_buffer *kbuf;
4628 struct io_sr_msg *sr = &req->sr_msg;
4630 void __user *buf = sr->buf;
4631 struct socket *sock;
4635 int ret, cflags = 0;
4636 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4638 sock = sock_from_file(req->file);
4639 if (unlikely(!sock))
4642 if (req->flags & REQ_F_BUFFER_SELECT) {
4643 kbuf = io_recv_buffer_select(req, !force_nonblock);
4645 return PTR_ERR(kbuf);
4646 buf = u64_to_user_ptr(kbuf->addr);
4649 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4653 msg.msg_name = NULL;
4654 msg.msg_control = NULL;
4655 msg.msg_controllen = 0;
4656 msg.msg_namelen = 0;
4657 msg.msg_iocb = NULL;
4660 flags = req->sr_msg.msg_flags;
4662 flags |= MSG_DONTWAIT;
4663 if (flags & MSG_WAITALL)
4664 min_ret = iov_iter_count(&msg.msg_iter);
4666 ret = sock_recvmsg(sock, &msg, flags);
4667 if (force_nonblock && ret == -EAGAIN)
4669 if (ret == -ERESTARTSYS)
4672 if (req->flags & REQ_F_BUFFER_SELECTED)
4673 cflags = io_put_recv_kbuf(req);
4674 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4675 req_set_fail_links(req);
4676 __io_req_complete(req, issue_flags, ret, cflags);
4680 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4682 struct io_accept *accept = &req->accept;
4684 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4686 if (sqe->ioprio || sqe->len || sqe->buf_index)
4689 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4690 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4691 accept->flags = READ_ONCE(sqe->accept_flags);
4692 accept->nofile = rlimit(RLIMIT_NOFILE);
4696 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4698 struct io_accept *accept = &req->accept;
4699 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4700 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4703 if (req->file->f_flags & O_NONBLOCK)
4704 req->flags |= REQ_F_NOWAIT;
4706 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4707 accept->addr_len, accept->flags,
4709 if (ret == -EAGAIN && force_nonblock)
4712 if (ret == -ERESTARTSYS)
4714 req_set_fail_links(req);
4716 __io_req_complete(req, issue_flags, ret, 0);
4720 static int io_connect_prep_async(struct io_kiocb *req)
4722 struct io_async_connect *io = req->async_data;
4723 struct io_connect *conn = &req->connect;
4725 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4728 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4730 struct io_connect *conn = &req->connect;
4732 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4734 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4737 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4738 conn->addr_len = READ_ONCE(sqe->addr2);
4742 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4744 struct io_async_connect __io, *io;
4745 unsigned file_flags;
4747 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4749 if (req->async_data) {
4750 io = req->async_data;
4752 ret = move_addr_to_kernel(req->connect.addr,
4753 req->connect.addr_len,
4760 file_flags = force_nonblock ? O_NONBLOCK : 0;
4762 ret = __sys_connect_file(req->file, &io->address,
4763 req->connect.addr_len, file_flags);
4764 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4765 if (req->async_data)
4767 if (io_alloc_async_data(req)) {
4771 memcpy(req->async_data, &__io, sizeof(__io));
4774 if (ret == -ERESTARTSYS)
4778 req_set_fail_links(req);
4779 __io_req_complete(req, issue_flags, ret, 0);
4782 #else /* !CONFIG_NET */
4783 #define IO_NETOP_FN(op) \
4784 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4786 return -EOPNOTSUPP; \
4789 #define IO_NETOP_PREP(op) \
4791 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4793 return -EOPNOTSUPP; \
4796 #define IO_NETOP_PREP_ASYNC(op) \
4798 static int io_##op##_prep_async(struct io_kiocb *req) \
4800 return -EOPNOTSUPP; \
4803 IO_NETOP_PREP_ASYNC(sendmsg);
4804 IO_NETOP_PREP_ASYNC(recvmsg);
4805 IO_NETOP_PREP_ASYNC(connect);
4806 IO_NETOP_PREP(accept);
4809 #endif /* CONFIG_NET */
4811 struct io_poll_table {
4812 struct poll_table_struct pt;
4813 struct io_kiocb *req;
4817 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4818 __poll_t mask, task_work_func_t func)
4822 /* for instances that support it check for an event match first: */
4823 if (mask && !(mask & poll->events))
4826 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4828 list_del_init(&poll->wait.entry);
4831 req->task_work.func = func;
4834 * If this fails, then the task is exiting. When a task exits, the
4835 * work gets canceled, so just cancel this request as well instead
4836 * of executing it. We can't safely execute it anyway, as we may not
4837 * have the needed state needed for it anyway.
4839 ret = io_req_task_work_add(req);
4840 if (unlikely(ret)) {
4841 WRITE_ONCE(poll->canceled, true);
4842 io_req_task_work_add_fallback(req, func);
4847 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4848 __acquires(&req->ctx->completion_lock)
4850 struct io_ring_ctx *ctx = req->ctx;
4852 if (!req->result && !READ_ONCE(poll->canceled)) {
4853 struct poll_table_struct pt = { ._key = poll->events };
4855 req->result = vfs_poll(req->file, &pt) & poll->events;
4858 spin_lock_irq(&ctx->completion_lock);
4859 if (!req->result && !READ_ONCE(poll->canceled)) {
4860 add_wait_queue(poll->head, &poll->wait);
4867 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4869 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4870 if (req->opcode == IORING_OP_POLL_ADD)
4871 return req->async_data;
4872 return req->apoll->double_poll;
4875 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4877 if (req->opcode == IORING_OP_POLL_ADD)
4879 return &req->apoll->poll;
4882 static void io_poll_remove_double(struct io_kiocb *req)
4883 __must_hold(&req->ctx->completion_lock)
4885 struct io_poll_iocb *poll = io_poll_get_double(req);
4887 lockdep_assert_held(&req->ctx->completion_lock);
4889 if (poll && poll->head) {
4890 struct wait_queue_head *head = poll->head;
4892 spin_lock(&head->lock);
4893 list_del_init(&poll->wait.entry);
4894 if (poll->wait.private)
4897 spin_unlock(&head->lock);
4901 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4902 __must_hold(&req->ctx->completion_lock)
4904 struct io_ring_ctx *ctx = req->ctx;
4905 unsigned flags = IORING_CQE_F_MORE;
4908 if (READ_ONCE(req->poll.canceled)) {
4910 req->poll.events |= EPOLLONESHOT;
4912 error = mangle_poll(mask);
4914 if (req->poll.events & EPOLLONESHOT)
4916 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4917 io_poll_remove_waitqs(req);
4918 req->poll.done = true;
4921 if (flags & IORING_CQE_F_MORE)
4924 io_commit_cqring(ctx);
4925 return !(flags & IORING_CQE_F_MORE);
4928 static void io_poll_task_func(struct callback_head *cb)
4930 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4931 struct io_ring_ctx *ctx = req->ctx;
4932 struct io_kiocb *nxt;
4934 if (io_poll_rewait(req, &req->poll)) {
4935 spin_unlock_irq(&ctx->completion_lock);
4939 done = io_poll_complete(req, req->result);
4941 hash_del(&req->hash_node);
4944 add_wait_queue(req->poll.head, &req->poll.wait);
4946 spin_unlock_irq(&ctx->completion_lock);
4947 io_cqring_ev_posted(ctx);
4950 nxt = io_put_req_find_next(req);
4952 __io_req_task_submit(nxt);
4957 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4958 int sync, void *key)
4960 struct io_kiocb *req = wait->private;
4961 struct io_poll_iocb *poll = io_poll_get_single(req);
4962 __poll_t mask = key_to_poll(key);
4964 /* for instances that support it check for an event match first: */
4965 if (mask && !(mask & poll->events))
4967 if (!(poll->events & EPOLLONESHOT))
4968 return poll->wait.func(&poll->wait, mode, sync, key);
4970 list_del_init(&wait->entry);
4972 if (poll && poll->head) {
4975 spin_lock(&poll->head->lock);
4976 done = list_empty(&poll->wait.entry);
4978 list_del_init(&poll->wait.entry);
4979 /* make sure double remove sees this as being gone */
4980 wait->private = NULL;
4981 spin_unlock(&poll->head->lock);
4983 /* use wait func handler, so it matches the rq type */
4984 poll->wait.func(&poll->wait, mode, sync, key);
4991 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4992 wait_queue_func_t wake_func)
4996 poll->canceled = false;
4997 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
4998 /* mask in events that we always want/need */
4999 poll->events = events | IO_POLL_UNMASK;
5000 INIT_LIST_HEAD(&poll->wait.entry);
5001 init_waitqueue_func_entry(&poll->wait, wake_func);
5004 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5005 struct wait_queue_head *head,
5006 struct io_poll_iocb **poll_ptr)
5008 struct io_kiocb *req = pt->req;
5011 * If poll->head is already set, it's because the file being polled
5012 * uses multiple waitqueues for poll handling (eg one for read, one
5013 * for write). Setup a separate io_poll_iocb if this happens.
5015 if (unlikely(poll->head)) {
5016 struct io_poll_iocb *poll_one = poll;
5018 /* already have a 2nd entry, fail a third attempt */
5020 pt->error = -EINVAL;
5024 * Can't handle multishot for double wait for now, turn it
5025 * into one-shot mode.
5027 if (!(poll_one->events & EPOLLONESHOT))
5028 poll_one->events |= EPOLLONESHOT;
5029 /* double add on the same waitqueue head, ignore */
5030 if (poll_one->head == head)
5032 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5034 pt->error = -ENOMEM;
5037 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5039 poll->wait.private = req;
5046 if (poll->events & EPOLLEXCLUSIVE)
5047 add_wait_queue_exclusive(head, &poll->wait);
5049 add_wait_queue(head, &poll->wait);
5052 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5053 struct poll_table_struct *p)
5055 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5056 struct async_poll *apoll = pt->req->apoll;
5058 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5061 static void io_async_task_func(struct callback_head *cb)
5063 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5064 struct async_poll *apoll = req->apoll;
5065 struct io_ring_ctx *ctx = req->ctx;
5067 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5069 if (io_poll_rewait(req, &apoll->poll)) {
5070 spin_unlock_irq(&ctx->completion_lock);
5074 hash_del(&req->hash_node);
5075 io_poll_remove_double(req);
5076 spin_unlock_irq(&ctx->completion_lock);
5078 if (!READ_ONCE(apoll->poll.canceled))
5079 __io_req_task_submit(req);
5081 io_req_complete_failed(req, -ECANCELED);
5084 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5087 struct io_kiocb *req = wait->private;
5088 struct io_poll_iocb *poll = &req->apoll->poll;
5090 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5093 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5096 static void io_poll_req_insert(struct io_kiocb *req)
5098 struct io_ring_ctx *ctx = req->ctx;
5099 struct hlist_head *list;
5101 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5102 hlist_add_head(&req->hash_node, list);
5105 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5106 struct io_poll_iocb *poll,
5107 struct io_poll_table *ipt, __poll_t mask,
5108 wait_queue_func_t wake_func)
5109 __acquires(&ctx->completion_lock)
5111 struct io_ring_ctx *ctx = req->ctx;
5112 bool cancel = false;
5114 INIT_HLIST_NODE(&req->hash_node);
5115 io_init_poll_iocb(poll, mask, wake_func);
5116 poll->file = req->file;
5117 poll->wait.private = req;
5119 ipt->pt._key = mask;
5121 ipt->error = -EINVAL;
5123 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5125 spin_lock_irq(&ctx->completion_lock);
5126 if (likely(poll->head)) {
5127 spin_lock(&poll->head->lock);
5128 if (unlikely(list_empty(&poll->wait.entry))) {
5134 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5135 list_del_init(&poll->wait.entry);
5137 WRITE_ONCE(poll->canceled, true);
5138 else if (!poll->done) /* actually waiting for an event */
5139 io_poll_req_insert(req);
5140 spin_unlock(&poll->head->lock);
5146 static bool io_arm_poll_handler(struct io_kiocb *req)
5148 const struct io_op_def *def = &io_op_defs[req->opcode];
5149 struct io_ring_ctx *ctx = req->ctx;
5150 struct async_poll *apoll;
5151 struct io_poll_table ipt;
5155 if (!req->file || !file_can_poll(req->file))
5157 if (req->flags & REQ_F_POLLED)
5161 else if (def->pollout)
5165 /* if we can't nonblock try, then no point in arming a poll handler */
5166 if (!io_file_supports_async(req, rw))
5169 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5170 if (unlikely(!apoll))
5172 apoll->double_poll = NULL;
5174 req->flags |= REQ_F_POLLED;
5177 mask = EPOLLONESHOT;
5179 mask |= POLLIN | POLLRDNORM;
5181 mask |= POLLOUT | POLLWRNORM;
5183 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5184 if ((req->opcode == IORING_OP_RECVMSG) &&
5185 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5188 mask |= POLLERR | POLLPRI;
5190 ipt.pt._qproc = io_async_queue_proc;
5192 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5194 if (ret || ipt.error) {
5195 io_poll_remove_double(req);
5196 spin_unlock_irq(&ctx->completion_lock);
5199 spin_unlock_irq(&ctx->completion_lock);
5200 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5201 apoll->poll.events);
5205 static bool __io_poll_remove_one(struct io_kiocb *req,
5206 struct io_poll_iocb *poll, bool do_cancel)
5207 __must_hold(&req->ctx->completion_lock)
5209 bool do_complete = false;
5213 spin_lock(&poll->head->lock);
5215 WRITE_ONCE(poll->canceled, true);
5216 if (!list_empty(&poll->wait.entry)) {
5217 list_del_init(&poll->wait.entry);
5220 spin_unlock(&poll->head->lock);
5221 hash_del(&req->hash_node);
5225 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5226 __must_hold(&req->ctx->completion_lock)
5230 io_poll_remove_double(req);
5231 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5233 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5234 /* non-poll requests have submit ref still */
5240 static bool io_poll_remove_one(struct io_kiocb *req)
5241 __must_hold(&req->ctx->completion_lock)
5245 do_complete = io_poll_remove_waitqs(req);
5247 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5248 io_commit_cqring(req->ctx);
5249 req_set_fail_links(req);
5250 io_put_req_deferred(req, 1);
5257 * Returns true if we found and killed one or more poll requests
5259 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5260 struct files_struct *files)
5262 struct hlist_node *tmp;
5263 struct io_kiocb *req;
5266 spin_lock_irq(&ctx->completion_lock);
5267 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5268 struct hlist_head *list;
5270 list = &ctx->cancel_hash[i];
5271 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5272 if (io_match_task(req, tsk, files))
5273 posted += io_poll_remove_one(req);
5276 spin_unlock_irq(&ctx->completion_lock);
5279 io_cqring_ev_posted(ctx);
5284 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5286 __must_hold(&ctx->completion_lock)
5288 struct hlist_head *list;
5289 struct io_kiocb *req;
5291 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5292 hlist_for_each_entry(req, list, hash_node) {
5293 if (sqe_addr != req->user_data)
5295 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5302 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5304 __must_hold(&ctx->completion_lock)
5306 struct io_kiocb *req;
5308 req = io_poll_find(ctx, sqe_addr, poll_only);
5311 if (io_poll_remove_one(req))
5317 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5322 events = READ_ONCE(sqe->poll32_events);
5324 events = swahw32(events);
5326 if (!(flags & IORING_POLL_ADD_MULTI))
5327 events |= EPOLLONESHOT;
5328 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5331 static int io_poll_update_prep(struct io_kiocb *req,
5332 const struct io_uring_sqe *sqe)
5334 struct io_poll_update *upd = &req->poll_update;
5337 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5339 if (sqe->ioprio || sqe->buf_index)
5341 flags = READ_ONCE(sqe->len);
5342 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5343 IORING_POLL_ADD_MULTI))
5345 /* meaningless without update */
5346 if (flags == IORING_POLL_ADD_MULTI)
5349 upd->old_user_data = READ_ONCE(sqe->addr);
5350 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5351 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5353 upd->new_user_data = READ_ONCE(sqe->off);
5354 if (!upd->update_user_data && upd->new_user_data)
5356 if (upd->update_events)
5357 upd->events = io_poll_parse_events(sqe, flags);
5358 else if (sqe->poll32_events)
5364 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5367 struct io_kiocb *req = wait->private;
5368 struct io_poll_iocb *poll = &req->poll;
5370 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5373 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5374 struct poll_table_struct *p)
5376 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5378 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5381 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5383 struct io_poll_iocb *poll = &req->poll;
5386 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5388 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5390 flags = READ_ONCE(sqe->len);
5391 if (flags & ~IORING_POLL_ADD_MULTI)
5394 poll->events = io_poll_parse_events(sqe, flags);
5398 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5400 struct io_poll_iocb *poll = &req->poll;
5401 struct io_ring_ctx *ctx = req->ctx;
5402 struct io_poll_table ipt;
5405 ipt.pt._qproc = io_poll_queue_proc;
5407 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5410 if (mask) { /* no async, we'd stolen it */
5412 io_poll_complete(req, mask);
5414 spin_unlock_irq(&ctx->completion_lock);
5417 io_cqring_ev_posted(ctx);
5418 if (poll->events & EPOLLONESHOT)
5424 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5426 struct io_ring_ctx *ctx = req->ctx;
5427 struct io_kiocb *preq;
5431 spin_lock_irq(&ctx->completion_lock);
5432 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5438 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5440 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5445 * Don't allow racy completion with singleshot, as we cannot safely
5446 * update those. For multishot, if we're racing with completion, just
5447 * let completion re-add it.
5449 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5450 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5454 /* we now have a detached poll request. reissue. */
5458 spin_unlock_irq(&ctx->completion_lock);
5459 req_set_fail_links(req);
5460 io_req_complete(req, ret);
5463 /* only mask one event flags, keep behavior flags */
5464 if (req->poll_update.update_events) {
5465 preq->poll.events &= ~0xffff;
5466 preq->poll.events |= req->poll_update.events & 0xffff;
5467 preq->poll.events |= IO_POLL_UNMASK;
5469 if (req->poll_update.update_user_data)
5470 preq->user_data = req->poll_update.new_user_data;
5471 spin_unlock_irq(&ctx->completion_lock);
5473 /* complete update request, we're done with it */
5474 io_req_complete(req, ret);
5477 ret = io_poll_add(preq, issue_flags);
5479 req_set_fail_links(preq);
5480 io_req_complete(preq, ret);
5486 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5488 struct io_timeout_data *data = container_of(timer,
5489 struct io_timeout_data, timer);
5490 struct io_kiocb *req = data->req;
5491 struct io_ring_ctx *ctx = req->ctx;
5492 unsigned long flags;
5494 spin_lock_irqsave(&ctx->completion_lock, flags);
5495 list_del_init(&req->timeout.list);
5496 atomic_set(&req->ctx->cq_timeouts,
5497 atomic_read(&req->ctx->cq_timeouts) + 1);
5499 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5500 io_commit_cqring(ctx);
5501 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5503 io_cqring_ev_posted(ctx);
5504 req_set_fail_links(req);
5506 return HRTIMER_NORESTART;
5509 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5511 __must_hold(&ctx->completion_lock)
5513 struct io_timeout_data *io;
5514 struct io_kiocb *req;
5517 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5518 found = user_data == req->user_data;
5523 return ERR_PTR(-ENOENT);
5525 io = req->async_data;
5526 if (hrtimer_try_to_cancel(&io->timer) == -1)
5527 return ERR_PTR(-EALREADY);
5528 list_del_init(&req->timeout.list);
5532 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5533 __must_hold(&ctx->completion_lock)
5535 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5538 return PTR_ERR(req);
5540 req_set_fail_links(req);
5541 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5542 io_put_req_deferred(req, 1);
5546 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5547 struct timespec64 *ts, enum hrtimer_mode mode)
5548 __must_hold(&ctx->completion_lock)
5550 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5551 struct io_timeout_data *data;
5554 return PTR_ERR(req);
5556 req->timeout.off = 0; /* noseq */
5557 data = req->async_data;
5558 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5559 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5560 data->timer.function = io_timeout_fn;
5561 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5565 static int io_timeout_remove_prep(struct io_kiocb *req,
5566 const struct io_uring_sqe *sqe)
5568 struct io_timeout_rem *tr = &req->timeout_rem;
5570 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5572 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5574 if (sqe->ioprio || sqe->buf_index || sqe->len)
5577 tr->addr = READ_ONCE(sqe->addr);
5578 tr->flags = READ_ONCE(sqe->timeout_flags);
5579 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5580 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5582 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5584 } else if (tr->flags) {
5585 /* timeout removal doesn't support flags */
5592 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5594 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5599 * Remove or update an existing timeout command
5601 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5603 struct io_timeout_rem *tr = &req->timeout_rem;
5604 struct io_ring_ctx *ctx = req->ctx;
5607 spin_lock_irq(&ctx->completion_lock);
5608 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5609 ret = io_timeout_cancel(ctx, tr->addr);
5611 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5612 io_translate_timeout_mode(tr->flags));
5614 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5615 io_commit_cqring(ctx);
5616 spin_unlock_irq(&ctx->completion_lock);
5617 io_cqring_ev_posted(ctx);
5619 req_set_fail_links(req);
5624 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5625 bool is_timeout_link)
5627 struct io_timeout_data *data;
5629 u32 off = READ_ONCE(sqe->off);
5631 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5633 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5635 if (off && is_timeout_link)
5637 flags = READ_ONCE(sqe->timeout_flags);
5638 if (flags & ~IORING_TIMEOUT_ABS)
5641 req->timeout.off = off;
5643 if (!req->async_data && io_alloc_async_data(req))
5646 data = req->async_data;
5649 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5652 data->mode = io_translate_timeout_mode(flags);
5653 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5654 if (is_timeout_link)
5655 io_req_track_inflight(req);
5659 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5661 struct io_ring_ctx *ctx = req->ctx;
5662 struct io_timeout_data *data = req->async_data;
5663 struct list_head *entry;
5664 u32 tail, off = req->timeout.off;
5666 spin_lock_irq(&ctx->completion_lock);
5669 * sqe->off holds how many events that need to occur for this
5670 * timeout event to be satisfied. If it isn't set, then this is
5671 * a pure timeout request, sequence isn't used.
5673 if (io_is_timeout_noseq(req)) {
5674 entry = ctx->timeout_list.prev;
5678 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5679 req->timeout.target_seq = tail + off;
5681 /* Update the last seq here in case io_flush_timeouts() hasn't.
5682 * This is safe because ->completion_lock is held, and submissions
5683 * and completions are never mixed in the same ->completion_lock section.
5685 ctx->cq_last_tm_flush = tail;
5688 * Insertion sort, ensuring the first entry in the list is always
5689 * the one we need first.
5691 list_for_each_prev(entry, &ctx->timeout_list) {
5692 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5695 if (io_is_timeout_noseq(nxt))
5697 /* nxt.seq is behind @tail, otherwise would've been completed */
5698 if (off >= nxt->timeout.target_seq - tail)
5702 list_add(&req->timeout.list, entry);
5703 data->timer.function = io_timeout_fn;
5704 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5705 spin_unlock_irq(&ctx->completion_lock);
5709 struct io_cancel_data {
5710 struct io_ring_ctx *ctx;
5714 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5716 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5717 struct io_cancel_data *cd = data;
5719 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5722 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5723 struct io_ring_ctx *ctx)
5725 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5726 enum io_wq_cancel cancel_ret;
5729 if (!tctx || !tctx->io_wq)
5732 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5733 switch (cancel_ret) {
5734 case IO_WQ_CANCEL_OK:
5737 case IO_WQ_CANCEL_RUNNING:
5740 case IO_WQ_CANCEL_NOTFOUND:
5748 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5749 struct io_kiocb *req, __u64 sqe_addr,
5752 unsigned long flags;
5755 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5756 spin_lock_irqsave(&ctx->completion_lock, flags);
5759 ret = io_timeout_cancel(ctx, sqe_addr);
5762 ret = io_poll_cancel(ctx, sqe_addr, false);
5766 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5767 io_commit_cqring(ctx);
5768 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5769 io_cqring_ev_posted(ctx);
5772 req_set_fail_links(req);
5775 static int io_async_cancel_prep(struct io_kiocb *req,
5776 const struct io_uring_sqe *sqe)
5778 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5780 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5782 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5785 req->cancel.addr = READ_ONCE(sqe->addr);
5789 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5791 struct io_ring_ctx *ctx = req->ctx;
5792 u64 sqe_addr = req->cancel.addr;
5793 struct io_tctx_node *node;
5796 /* tasks should wait for their io-wq threads, so safe w/o sync */
5797 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5798 spin_lock_irq(&ctx->completion_lock);
5801 ret = io_timeout_cancel(ctx, sqe_addr);
5804 ret = io_poll_cancel(ctx, sqe_addr, false);
5807 spin_unlock_irq(&ctx->completion_lock);
5809 /* slow path, try all io-wq's */
5810 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5812 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5813 struct io_uring_task *tctx = node->task->io_uring;
5815 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5819 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5821 spin_lock_irq(&ctx->completion_lock);
5823 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5824 io_commit_cqring(ctx);
5825 spin_unlock_irq(&ctx->completion_lock);
5826 io_cqring_ev_posted(ctx);
5829 req_set_fail_links(req);
5834 static int io_rsrc_update_prep(struct io_kiocb *req,
5835 const struct io_uring_sqe *sqe)
5837 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5839 if (sqe->ioprio || sqe->rw_flags)
5842 req->rsrc_update.offset = READ_ONCE(sqe->off);
5843 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5844 if (!req->rsrc_update.nr_args)
5846 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5850 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5852 struct io_ring_ctx *ctx = req->ctx;
5853 struct io_uring_rsrc_update2 up;
5856 if (issue_flags & IO_URING_F_NONBLOCK)
5859 up.offset = req->rsrc_update.offset;
5860 up.data = req->rsrc_update.arg;
5865 mutex_lock(&ctx->uring_lock);
5866 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5867 &up, req->rsrc_update.nr_args);
5868 mutex_unlock(&ctx->uring_lock);
5871 req_set_fail_links(req);
5872 __io_req_complete(req, issue_flags, ret, 0);
5876 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5878 switch (req->opcode) {
5881 case IORING_OP_READV:
5882 case IORING_OP_READ_FIXED:
5883 case IORING_OP_READ:
5884 return io_read_prep(req, sqe);
5885 case IORING_OP_WRITEV:
5886 case IORING_OP_WRITE_FIXED:
5887 case IORING_OP_WRITE:
5888 return io_write_prep(req, sqe);
5889 case IORING_OP_POLL_ADD:
5890 return io_poll_add_prep(req, sqe);
5891 case IORING_OP_POLL_REMOVE:
5892 return io_poll_update_prep(req, sqe);
5893 case IORING_OP_FSYNC:
5894 return io_fsync_prep(req, sqe);
5895 case IORING_OP_SYNC_FILE_RANGE:
5896 return io_sfr_prep(req, sqe);
5897 case IORING_OP_SENDMSG:
5898 case IORING_OP_SEND:
5899 return io_sendmsg_prep(req, sqe);
5900 case IORING_OP_RECVMSG:
5901 case IORING_OP_RECV:
5902 return io_recvmsg_prep(req, sqe);
5903 case IORING_OP_CONNECT:
5904 return io_connect_prep(req, sqe);
5905 case IORING_OP_TIMEOUT:
5906 return io_timeout_prep(req, sqe, false);
5907 case IORING_OP_TIMEOUT_REMOVE:
5908 return io_timeout_remove_prep(req, sqe);
5909 case IORING_OP_ASYNC_CANCEL:
5910 return io_async_cancel_prep(req, sqe);
5911 case IORING_OP_LINK_TIMEOUT:
5912 return io_timeout_prep(req, sqe, true);
5913 case IORING_OP_ACCEPT:
5914 return io_accept_prep(req, sqe);
5915 case IORING_OP_FALLOCATE:
5916 return io_fallocate_prep(req, sqe);
5917 case IORING_OP_OPENAT:
5918 return io_openat_prep(req, sqe);
5919 case IORING_OP_CLOSE:
5920 return io_close_prep(req, sqe);
5921 case IORING_OP_FILES_UPDATE:
5922 return io_rsrc_update_prep(req, sqe);
5923 case IORING_OP_STATX:
5924 return io_statx_prep(req, sqe);
5925 case IORING_OP_FADVISE:
5926 return io_fadvise_prep(req, sqe);
5927 case IORING_OP_MADVISE:
5928 return io_madvise_prep(req, sqe);
5929 case IORING_OP_OPENAT2:
5930 return io_openat2_prep(req, sqe);
5931 case IORING_OP_EPOLL_CTL:
5932 return io_epoll_ctl_prep(req, sqe);
5933 case IORING_OP_SPLICE:
5934 return io_splice_prep(req, sqe);
5935 case IORING_OP_PROVIDE_BUFFERS:
5936 return io_provide_buffers_prep(req, sqe);
5937 case IORING_OP_REMOVE_BUFFERS:
5938 return io_remove_buffers_prep(req, sqe);
5940 return io_tee_prep(req, sqe);
5941 case IORING_OP_SHUTDOWN:
5942 return io_shutdown_prep(req, sqe);
5943 case IORING_OP_RENAMEAT:
5944 return io_renameat_prep(req, sqe);
5945 case IORING_OP_UNLINKAT:
5946 return io_unlinkat_prep(req, sqe);
5949 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5954 static int io_req_prep_async(struct io_kiocb *req)
5956 if (!io_op_defs[req->opcode].needs_async_setup)
5958 if (WARN_ON_ONCE(req->async_data))
5960 if (io_alloc_async_data(req))
5963 switch (req->opcode) {
5964 case IORING_OP_READV:
5965 return io_rw_prep_async(req, READ);
5966 case IORING_OP_WRITEV:
5967 return io_rw_prep_async(req, WRITE);
5968 case IORING_OP_SENDMSG:
5969 return io_sendmsg_prep_async(req);
5970 case IORING_OP_RECVMSG:
5971 return io_recvmsg_prep_async(req);
5972 case IORING_OP_CONNECT:
5973 return io_connect_prep_async(req);
5975 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5980 static u32 io_get_sequence(struct io_kiocb *req)
5982 struct io_kiocb *pos;
5983 struct io_ring_ctx *ctx = req->ctx;
5984 u32 total_submitted, nr_reqs = 0;
5986 io_for_each_link(pos, req)
5989 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5990 return total_submitted - nr_reqs;
5993 static int io_req_defer(struct io_kiocb *req)
5995 struct io_ring_ctx *ctx = req->ctx;
5996 struct io_defer_entry *de;
6000 /* Still need defer if there is pending req in defer list. */
6001 if (likely(list_empty_careful(&ctx->defer_list) &&
6002 !(req->flags & REQ_F_IO_DRAIN)))
6005 seq = io_get_sequence(req);
6006 /* Still a chance to pass the sequence check */
6007 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6010 ret = io_req_prep_async(req);
6013 io_prep_async_link(req);
6014 de = kmalloc(sizeof(*de), GFP_KERNEL);
6018 spin_lock_irq(&ctx->completion_lock);
6019 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6020 spin_unlock_irq(&ctx->completion_lock);
6022 io_queue_async_work(req);
6023 return -EIOCBQUEUED;
6026 trace_io_uring_defer(ctx, req, req->user_data);
6029 list_add_tail(&de->list, &ctx->defer_list);
6030 spin_unlock_irq(&ctx->completion_lock);
6031 return -EIOCBQUEUED;
6034 static void io_clean_op(struct io_kiocb *req)
6036 if (req->flags & REQ_F_BUFFER_SELECTED) {
6037 switch (req->opcode) {
6038 case IORING_OP_READV:
6039 case IORING_OP_READ_FIXED:
6040 case IORING_OP_READ:
6041 kfree((void *)(unsigned long)req->rw.addr);
6043 case IORING_OP_RECVMSG:
6044 case IORING_OP_RECV:
6045 kfree(req->sr_msg.kbuf);
6048 req->flags &= ~REQ_F_BUFFER_SELECTED;
6051 if (req->flags & REQ_F_NEED_CLEANUP) {
6052 switch (req->opcode) {
6053 case IORING_OP_READV:
6054 case IORING_OP_READ_FIXED:
6055 case IORING_OP_READ:
6056 case IORING_OP_WRITEV:
6057 case IORING_OP_WRITE_FIXED:
6058 case IORING_OP_WRITE: {
6059 struct io_async_rw *io = req->async_data;
6061 kfree(io->free_iovec);
6064 case IORING_OP_RECVMSG:
6065 case IORING_OP_SENDMSG: {
6066 struct io_async_msghdr *io = req->async_data;
6068 kfree(io->free_iov);
6071 case IORING_OP_SPLICE:
6073 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6074 io_put_file(req->splice.file_in);
6076 case IORING_OP_OPENAT:
6077 case IORING_OP_OPENAT2:
6078 if (req->open.filename)
6079 putname(req->open.filename);
6081 case IORING_OP_RENAMEAT:
6082 putname(req->rename.oldpath);
6083 putname(req->rename.newpath);
6085 case IORING_OP_UNLINKAT:
6086 putname(req->unlink.filename);
6089 req->flags &= ~REQ_F_NEED_CLEANUP;
6091 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6092 kfree(req->apoll->double_poll);
6096 if (req->flags & REQ_F_INFLIGHT) {
6097 struct io_uring_task *tctx = req->task->io_uring;
6099 atomic_dec(&tctx->inflight_tracked);
6100 req->flags &= ~REQ_F_INFLIGHT;
6104 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6106 struct io_ring_ctx *ctx = req->ctx;
6107 const struct cred *creds = NULL;
6110 if (req->work.creds && req->work.creds != current_cred())
6111 creds = override_creds(req->work.creds);
6113 switch (req->opcode) {
6115 ret = io_nop(req, issue_flags);
6117 case IORING_OP_READV:
6118 case IORING_OP_READ_FIXED:
6119 case IORING_OP_READ:
6120 ret = io_read(req, issue_flags);
6122 case IORING_OP_WRITEV:
6123 case IORING_OP_WRITE_FIXED:
6124 case IORING_OP_WRITE:
6125 ret = io_write(req, issue_flags);
6127 case IORING_OP_FSYNC:
6128 ret = io_fsync(req, issue_flags);
6130 case IORING_OP_POLL_ADD:
6131 ret = io_poll_add(req, issue_flags);
6133 case IORING_OP_POLL_REMOVE:
6134 ret = io_poll_update(req, issue_flags);
6136 case IORING_OP_SYNC_FILE_RANGE:
6137 ret = io_sync_file_range(req, issue_flags);
6139 case IORING_OP_SENDMSG:
6140 ret = io_sendmsg(req, issue_flags);
6142 case IORING_OP_SEND:
6143 ret = io_send(req, issue_flags);
6145 case IORING_OP_RECVMSG:
6146 ret = io_recvmsg(req, issue_flags);
6148 case IORING_OP_RECV:
6149 ret = io_recv(req, issue_flags);
6151 case IORING_OP_TIMEOUT:
6152 ret = io_timeout(req, issue_flags);
6154 case IORING_OP_TIMEOUT_REMOVE:
6155 ret = io_timeout_remove(req, issue_flags);
6157 case IORING_OP_ACCEPT:
6158 ret = io_accept(req, issue_flags);
6160 case IORING_OP_CONNECT:
6161 ret = io_connect(req, issue_flags);
6163 case IORING_OP_ASYNC_CANCEL:
6164 ret = io_async_cancel(req, issue_flags);
6166 case IORING_OP_FALLOCATE:
6167 ret = io_fallocate(req, issue_flags);
6169 case IORING_OP_OPENAT:
6170 ret = io_openat(req, issue_flags);
6172 case IORING_OP_CLOSE:
6173 ret = io_close(req, issue_flags);
6175 case IORING_OP_FILES_UPDATE:
6176 ret = io_files_update(req, issue_flags);
6178 case IORING_OP_STATX:
6179 ret = io_statx(req, issue_flags);
6181 case IORING_OP_FADVISE:
6182 ret = io_fadvise(req, issue_flags);
6184 case IORING_OP_MADVISE:
6185 ret = io_madvise(req, issue_flags);
6187 case IORING_OP_OPENAT2:
6188 ret = io_openat2(req, issue_flags);
6190 case IORING_OP_EPOLL_CTL:
6191 ret = io_epoll_ctl(req, issue_flags);
6193 case IORING_OP_SPLICE:
6194 ret = io_splice(req, issue_flags);
6196 case IORING_OP_PROVIDE_BUFFERS:
6197 ret = io_provide_buffers(req, issue_flags);
6199 case IORING_OP_REMOVE_BUFFERS:
6200 ret = io_remove_buffers(req, issue_flags);
6203 ret = io_tee(req, issue_flags);
6205 case IORING_OP_SHUTDOWN:
6206 ret = io_shutdown(req, issue_flags);
6208 case IORING_OP_RENAMEAT:
6209 ret = io_renameat(req, issue_flags);
6211 case IORING_OP_UNLINKAT:
6212 ret = io_unlinkat(req, issue_flags);
6220 revert_creds(creds);
6225 /* If the op doesn't have a file, we're not polling for it */
6226 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6227 const bool in_async = io_wq_current_is_worker();
6229 /* workqueue context doesn't hold uring_lock, grab it now */
6231 mutex_lock(&ctx->uring_lock);
6233 io_iopoll_req_issued(req, in_async);
6236 mutex_unlock(&ctx->uring_lock);
6242 static void io_wq_submit_work(struct io_wq_work *work)
6244 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6245 struct io_kiocb *timeout;
6248 timeout = io_prep_linked_timeout(req);
6250 io_queue_linked_timeout(timeout);
6252 if (work->flags & IO_WQ_WORK_CANCEL)
6257 ret = io_issue_sqe(req, 0);
6259 * We can get EAGAIN for polled IO even though we're
6260 * forcing a sync submission from here, since we can't
6261 * wait for request slots on the block side.
6269 /* avoid locking problems by failing it from a clean context */
6271 /* io-wq is going to take one down */
6273 io_req_task_queue_fail(req, ret);
6277 #define FFS_ASYNC_READ 0x1UL
6278 #define FFS_ASYNC_WRITE 0x2UL
6280 #define FFS_ISREG 0x4UL
6282 #define FFS_ISREG 0x0UL
6284 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6286 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6289 struct io_fixed_file *table_l2;
6291 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6292 return &table_l2[i & IORING_FILE_TABLE_MASK];
6295 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6298 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6300 return (struct file *) (slot->file_ptr & FFS_MASK);
6303 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6305 unsigned long file_ptr = (unsigned long) file;
6307 if (__io_file_supports_async(file, READ))
6308 file_ptr |= FFS_ASYNC_READ;
6309 if (__io_file_supports_async(file, WRITE))
6310 file_ptr |= FFS_ASYNC_WRITE;
6311 if (S_ISREG(file_inode(file)->i_mode))
6312 file_ptr |= FFS_ISREG;
6313 file_slot->file_ptr = file_ptr;
6316 static struct file *io_file_get(struct io_submit_state *state,
6317 struct io_kiocb *req, int fd, bool fixed)
6319 struct io_ring_ctx *ctx = req->ctx;
6323 unsigned long file_ptr;
6325 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6327 fd = array_index_nospec(fd, ctx->nr_user_files);
6328 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6329 file = (struct file *) (file_ptr & FFS_MASK);
6330 file_ptr &= ~FFS_MASK;
6331 /* mask in overlapping REQ_F and FFS bits */
6332 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6333 io_req_set_rsrc_node(req);
6335 trace_io_uring_file_get(ctx, fd);
6336 file = __io_file_get(state, fd);
6338 /* we don't allow fixed io_uring files */
6339 if (file && unlikely(file->f_op == &io_uring_fops))
6340 io_req_track_inflight(req);
6346 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6348 struct io_timeout_data *data = container_of(timer,
6349 struct io_timeout_data, timer);
6350 struct io_kiocb *prev, *req = data->req;
6351 struct io_ring_ctx *ctx = req->ctx;
6352 unsigned long flags;
6354 spin_lock_irqsave(&ctx->completion_lock, flags);
6355 prev = req->timeout.head;
6356 req->timeout.head = NULL;
6359 * We don't expect the list to be empty, that will only happen if we
6360 * race with the completion of the linked work.
6363 io_remove_next_linked(prev);
6364 if (!req_ref_inc_not_zero(prev))
6367 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6370 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6371 io_put_req_deferred(prev, 1);
6372 io_put_req_deferred(req, 1);
6374 io_req_complete_post(req, -ETIME, 0);
6376 return HRTIMER_NORESTART;
6379 static void io_queue_linked_timeout(struct io_kiocb *req)
6381 struct io_ring_ctx *ctx = req->ctx;
6383 spin_lock_irq(&ctx->completion_lock);
6385 * If the back reference is NULL, then our linked request finished
6386 * before we got a chance to setup the timer
6388 if (req->timeout.head) {
6389 struct io_timeout_data *data = req->async_data;
6391 data->timer.function = io_link_timeout_fn;
6392 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6395 spin_unlock_irq(&ctx->completion_lock);
6396 /* drop submission reference */
6400 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6402 struct io_kiocb *nxt = req->link;
6404 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6405 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6408 nxt->timeout.head = req;
6409 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6410 req->flags |= REQ_F_LINK_TIMEOUT;
6414 static void __io_queue_sqe(struct io_kiocb *req)
6416 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6419 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6422 * We async punt it if the file wasn't marked NOWAIT, or if the file
6423 * doesn't support non-blocking read/write attempts
6426 /* drop submission reference */
6427 if (req->flags & REQ_F_COMPLETE_INLINE) {
6428 struct io_ring_ctx *ctx = req->ctx;
6429 struct io_comp_state *cs = &ctx->submit_state.comp;
6431 cs->reqs[cs->nr++] = req;
6432 if (cs->nr == ARRAY_SIZE(cs->reqs))
6433 io_submit_flush_completions(cs, ctx);
6437 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6438 if (!io_arm_poll_handler(req)) {
6440 * Queued up for async execution, worker will release
6441 * submit reference when the iocb is actually submitted.
6443 io_queue_async_work(req);
6446 io_req_complete_failed(req, ret);
6449 io_queue_linked_timeout(linked_timeout);
6452 static void io_queue_sqe(struct io_kiocb *req)
6456 ret = io_req_defer(req);
6458 if (ret != -EIOCBQUEUED) {
6460 io_req_complete_failed(req, ret);
6462 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6463 ret = io_req_prep_async(req);
6466 io_queue_async_work(req);
6468 __io_queue_sqe(req);
6473 * Check SQE restrictions (opcode and flags).
6475 * Returns 'true' if SQE is allowed, 'false' otherwise.
6477 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6478 struct io_kiocb *req,
6479 unsigned int sqe_flags)
6481 if (!ctx->restricted)
6484 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6487 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6488 ctx->restrictions.sqe_flags_required)
6491 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6492 ctx->restrictions.sqe_flags_required))
6498 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6499 const struct io_uring_sqe *sqe)
6501 struct io_submit_state *state;
6502 unsigned int sqe_flags;
6503 int personality, ret = 0;
6505 req->opcode = READ_ONCE(sqe->opcode);
6506 /* same numerical values with corresponding REQ_F_*, safe to copy */
6507 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6508 req->user_data = READ_ONCE(sqe->user_data);
6509 req->async_data = NULL;
6513 req->fixed_rsrc_refs = NULL;
6514 /* one is dropped after submission, the other at completion */
6515 atomic_set(&req->refs, 2);
6516 req->task = current;
6518 req->work.creds = NULL;
6520 /* enforce forwards compatibility on users */
6521 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6523 if (unlikely(req->opcode >= IORING_OP_LAST))
6525 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6528 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6529 !io_op_defs[req->opcode].buffer_select)
6532 personality = READ_ONCE(sqe->personality);
6534 req->work.creds = xa_load(&ctx->personalities, personality);
6535 if (!req->work.creds)
6537 get_cred(req->work.creds);
6539 state = &ctx->submit_state;
6542 * Plug now if we have more than 1 IO left after this, and the target
6543 * is potentially a read/write to block based storage.
6545 if (!state->plug_started && state->ios_left > 1 &&
6546 io_op_defs[req->opcode].plug) {
6547 blk_start_plug(&state->plug);
6548 state->plug_started = true;
6551 if (io_op_defs[req->opcode].needs_file) {
6552 bool fixed = req->flags & REQ_F_FIXED_FILE;
6554 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6555 if (unlikely(!req->file))
6563 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6564 const struct io_uring_sqe *sqe)
6566 struct io_submit_link *link = &ctx->submit_state.link;
6569 ret = io_init_req(ctx, req, sqe);
6570 if (unlikely(ret)) {
6573 /* fail even hard links since we don't submit */
6574 link->head->flags |= REQ_F_FAIL_LINK;
6575 io_req_complete_failed(link->head, -ECANCELED);
6578 io_req_complete_failed(req, ret);
6581 ret = io_req_prep(req, sqe);
6585 /* don't need @sqe from now on */
6586 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6587 true, ctx->flags & IORING_SETUP_SQPOLL);
6590 * If we already have a head request, queue this one for async
6591 * submittal once the head completes. If we don't have a head but
6592 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6593 * submitted sync once the chain is complete. If none of those
6594 * conditions are true (normal request), then just queue it.
6597 struct io_kiocb *head = link->head;
6600 * Taking sequential execution of a link, draining both sides
6601 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6602 * requests in the link. So, it drains the head and the
6603 * next after the link request. The last one is done via
6604 * drain_next flag to persist the effect across calls.
6606 if (req->flags & REQ_F_IO_DRAIN) {
6607 head->flags |= REQ_F_IO_DRAIN;
6608 ctx->drain_next = 1;
6610 ret = io_req_prep_async(req);
6613 trace_io_uring_link(ctx, req, head);
6614 link->last->link = req;
6617 /* last request of a link, enqueue the link */
6618 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6623 if (unlikely(ctx->drain_next)) {
6624 req->flags |= REQ_F_IO_DRAIN;
6625 ctx->drain_next = 0;
6627 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6639 * Batched submission is done, ensure local IO is flushed out.
6641 static void io_submit_state_end(struct io_submit_state *state,
6642 struct io_ring_ctx *ctx)
6644 if (state->link.head)
6645 io_queue_sqe(state->link.head);
6647 io_submit_flush_completions(&state->comp, ctx);
6648 if (state->plug_started)
6649 blk_finish_plug(&state->plug);
6650 io_state_file_put(state);
6654 * Start submission side cache.
6656 static void io_submit_state_start(struct io_submit_state *state,
6657 unsigned int max_ios)
6659 state->plug_started = false;
6660 state->ios_left = max_ios;
6661 /* set only head, no need to init link_last in advance */
6662 state->link.head = NULL;
6665 static void io_commit_sqring(struct io_ring_ctx *ctx)
6667 struct io_rings *rings = ctx->rings;
6670 * Ensure any loads from the SQEs are done at this point,
6671 * since once we write the new head, the application could
6672 * write new data to them.
6674 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6678 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6679 * that is mapped by userspace. This means that care needs to be taken to
6680 * ensure that reads are stable, as we cannot rely on userspace always
6681 * being a good citizen. If members of the sqe are validated and then later
6682 * used, it's important that those reads are done through READ_ONCE() to
6683 * prevent a re-load down the line.
6685 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6687 u32 *sq_array = ctx->sq_array;
6691 * The cached sq head (or cq tail) serves two purposes:
6693 * 1) allows us to batch the cost of updating the user visible
6695 * 2) allows the kernel side to track the head on its own, even
6696 * though the application is the one updating it.
6698 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6699 if (likely(head < ctx->sq_entries))
6700 return &ctx->sq_sqes[head];
6702 /* drop invalid entries */
6703 ctx->cached_sq_dropped++;
6704 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6708 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6712 /* make sure SQ entry isn't read before tail */
6713 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6715 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6718 percpu_counter_add(¤t->io_uring->inflight, nr);
6719 refcount_add(nr, ¤t->usage);
6720 io_submit_state_start(&ctx->submit_state, nr);
6722 while (submitted < nr) {
6723 const struct io_uring_sqe *sqe;
6724 struct io_kiocb *req;
6726 req = io_alloc_req(ctx);
6727 if (unlikely(!req)) {
6729 submitted = -EAGAIN;
6732 sqe = io_get_sqe(ctx);
6733 if (unlikely(!sqe)) {
6734 kmem_cache_free(req_cachep, req);
6737 /* will complete beyond this point, count as submitted */
6739 if (io_submit_sqe(ctx, req, sqe))
6743 if (unlikely(submitted != nr)) {
6744 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6745 struct io_uring_task *tctx = current->io_uring;
6746 int unused = nr - ref_used;
6748 percpu_ref_put_many(&ctx->refs, unused);
6749 percpu_counter_sub(&tctx->inflight, unused);
6750 put_task_struct_many(current, unused);
6753 io_submit_state_end(&ctx->submit_state, ctx);
6754 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6755 io_commit_sqring(ctx);
6760 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6762 /* Tell userspace we may need a wakeup call */
6763 spin_lock_irq(&ctx->completion_lock);
6764 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6765 spin_unlock_irq(&ctx->completion_lock);
6768 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6770 spin_lock_irq(&ctx->completion_lock);
6771 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6772 spin_unlock_irq(&ctx->completion_lock);
6775 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6777 unsigned int to_submit;
6780 to_submit = io_sqring_entries(ctx);
6781 /* if we're handling multiple rings, cap submit size for fairness */
6782 if (cap_entries && to_submit > 8)
6785 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6786 unsigned nr_events = 0;
6788 mutex_lock(&ctx->uring_lock);
6789 if (!list_empty(&ctx->iopoll_list))
6790 io_do_iopoll(ctx, &nr_events, 0);
6793 * Don't submit if refs are dying, good for io_uring_register(),
6794 * but also it is relied upon by io_ring_exit_work()
6796 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6797 !(ctx->flags & IORING_SETUP_R_DISABLED))
6798 ret = io_submit_sqes(ctx, to_submit);
6799 mutex_unlock(&ctx->uring_lock);
6802 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6803 wake_up(&ctx->sqo_sq_wait);
6808 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6810 struct io_ring_ctx *ctx;
6811 unsigned sq_thread_idle = 0;
6813 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6814 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6815 sqd->sq_thread_idle = sq_thread_idle;
6818 static int io_sq_thread(void *data)
6820 struct io_sq_data *sqd = data;
6821 struct io_ring_ctx *ctx;
6822 unsigned long timeout = 0;
6823 char buf[TASK_COMM_LEN];
6826 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6827 set_task_comm(current, buf);
6829 if (sqd->sq_cpu != -1)
6830 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6832 set_cpus_allowed_ptr(current, cpu_online_mask);
6833 current->flags |= PF_NO_SETAFFINITY;
6835 mutex_lock(&sqd->lock);
6836 /* a user may had exited before the thread started */
6837 io_run_task_work_head(&sqd->park_task_work);
6839 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6841 bool cap_entries, sqt_spin, needs_sched;
6843 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6844 signal_pending(current)) {
6845 bool did_sig = false;
6847 mutex_unlock(&sqd->lock);
6848 if (signal_pending(current)) {
6849 struct ksignal ksig;
6851 did_sig = get_signal(&ksig);
6854 mutex_lock(&sqd->lock);
6856 io_run_task_work_head(&sqd->park_task_work);
6859 timeout = jiffies + sqd->sq_thread_idle;
6863 cap_entries = !list_is_singular(&sqd->ctx_list);
6864 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6865 const struct cred *creds = NULL;
6867 if (ctx->sq_creds != current_cred())
6868 creds = override_creds(ctx->sq_creds);
6869 ret = __io_sq_thread(ctx, cap_entries);
6871 revert_creds(creds);
6872 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6876 if (sqt_spin || !time_after(jiffies, timeout)) {
6880 timeout = jiffies + sqd->sq_thread_idle;
6884 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6885 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6886 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6887 io_ring_set_wakeup_flag(ctx);
6890 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6891 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6892 !list_empty_careful(&ctx->iopoll_list)) {
6893 needs_sched = false;
6896 if (io_sqring_entries(ctx)) {
6897 needs_sched = false;
6903 mutex_unlock(&sqd->lock);
6905 mutex_lock(&sqd->lock);
6907 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6908 io_ring_clear_wakeup_flag(ctx);
6911 finish_wait(&sqd->wait, &wait);
6912 io_run_task_work_head(&sqd->park_task_work);
6913 timeout = jiffies + sqd->sq_thread_idle;
6916 io_uring_cancel_sqpoll(sqd);
6918 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6919 io_ring_set_wakeup_flag(ctx);
6921 io_run_task_work_head(&sqd->park_task_work);
6922 mutex_unlock(&sqd->lock);
6924 complete(&sqd->exited);
6928 struct io_wait_queue {
6929 struct wait_queue_entry wq;
6930 struct io_ring_ctx *ctx;
6932 unsigned nr_timeouts;
6935 static inline bool io_should_wake(struct io_wait_queue *iowq)
6937 struct io_ring_ctx *ctx = iowq->ctx;
6940 * Wake up if we have enough events, or if a timeout occurred since we
6941 * started waiting. For timeouts, we always want to return to userspace,
6942 * regardless of event count.
6944 return io_cqring_events(ctx) >= iowq->to_wait ||
6945 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6948 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6949 int wake_flags, void *key)
6951 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6955 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6956 * the task, and the next invocation will do it.
6958 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6959 return autoremove_wake_function(curr, mode, wake_flags, key);
6963 static int io_run_task_work_sig(void)
6965 if (io_run_task_work())
6967 if (!signal_pending(current))
6969 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6970 return -ERESTARTSYS;
6974 /* when returns >0, the caller should retry */
6975 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6976 struct io_wait_queue *iowq,
6977 signed long *timeout)
6981 /* make sure we run task_work before checking for signals */
6982 ret = io_run_task_work_sig();
6983 if (ret || io_should_wake(iowq))
6985 /* let the caller flush overflows, retry */
6986 if (test_bit(0, &ctx->cq_check_overflow))
6989 *timeout = schedule_timeout(*timeout);
6990 return !*timeout ? -ETIME : 1;
6994 * Wait until events become available, if we don't already have some. The
6995 * application must reap them itself, as they reside on the shared cq ring.
6997 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6998 const sigset_t __user *sig, size_t sigsz,
6999 struct __kernel_timespec __user *uts)
7001 struct io_wait_queue iowq = {
7004 .func = io_wake_function,
7005 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7008 .to_wait = min_events,
7010 struct io_rings *rings = ctx->rings;
7011 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7015 io_cqring_overflow_flush(ctx, false);
7016 if (io_cqring_events(ctx) >= min_events)
7018 if (!io_run_task_work())
7023 #ifdef CONFIG_COMPAT
7024 if (in_compat_syscall())
7025 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7029 ret = set_user_sigmask(sig, sigsz);
7036 struct timespec64 ts;
7038 if (get_timespec64(&ts, uts))
7040 timeout = timespec64_to_jiffies(&ts);
7043 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7044 trace_io_uring_cqring_wait(ctx, min_events);
7046 /* if we can't even flush overflow, don't wait for more */
7047 if (!io_cqring_overflow_flush(ctx, false)) {
7051 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7052 TASK_INTERRUPTIBLE);
7053 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7054 finish_wait(&ctx->wait, &iowq.wq);
7058 restore_saved_sigmask_unless(ret == -EINTR);
7060 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7063 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7065 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7067 for (i = 0; i < nr_tables; i++)
7068 kfree(table->files[i]);
7069 kfree(table->files);
7070 table->files = NULL;
7073 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7075 spin_lock_bh(&ctx->rsrc_ref_lock);
7078 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7080 spin_unlock_bh(&ctx->rsrc_ref_lock);
7083 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7085 percpu_ref_exit(&ref_node->refs);
7089 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7090 struct io_rsrc_data *data_to_kill)
7092 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7093 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7096 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7098 rsrc_node->rsrc_data = data_to_kill;
7099 io_rsrc_ref_lock(ctx);
7100 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7101 io_rsrc_ref_unlock(ctx);
7103 atomic_inc(&data_to_kill->refs);
7104 percpu_ref_kill(&rsrc_node->refs);
7105 ctx->rsrc_node = NULL;
7108 if (!ctx->rsrc_node) {
7109 ctx->rsrc_node = ctx->rsrc_backup_node;
7110 ctx->rsrc_backup_node = NULL;
7114 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7116 if (ctx->rsrc_backup_node)
7118 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7119 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7122 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7126 /* As we may drop ->uring_lock, other task may have started quiesce */
7130 data->quiesce = true;
7132 ret = io_rsrc_node_switch_start(ctx);
7135 io_rsrc_node_switch(ctx, data);
7137 /* kill initial ref, already quiesced if zero */
7138 if (atomic_dec_and_test(&data->refs))
7140 flush_delayed_work(&ctx->rsrc_put_work);
7141 ret = wait_for_completion_interruptible(&data->done);
7145 atomic_inc(&data->refs);
7146 /* wait for all works potentially completing data->done */
7147 flush_delayed_work(&ctx->rsrc_put_work);
7148 reinit_completion(&data->done);
7150 mutex_unlock(&ctx->uring_lock);
7151 ret = io_run_task_work_sig();
7152 mutex_lock(&ctx->uring_lock);
7154 data->quiesce = false;
7159 static void io_rsrc_data_free(struct io_rsrc_data *data)
7165 static struct io_rsrc_data *io_rsrc_data_alloc(struct io_ring_ctx *ctx,
7166 rsrc_put_fn *do_put,
7169 struct io_rsrc_data *data;
7171 data = kzalloc(sizeof(*data), GFP_KERNEL);
7175 data->tags = kvcalloc(nr, sizeof(*data->tags), GFP_KERNEL);
7181 atomic_set(&data->refs, 1);
7183 data->do_put = do_put;
7184 init_completion(&data->done);
7188 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7190 #if defined(CONFIG_UNIX)
7191 if (ctx->ring_sock) {
7192 struct sock *sock = ctx->ring_sock->sk;
7193 struct sk_buff *skb;
7195 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7201 for (i = 0; i < ctx->nr_user_files; i++) {
7204 file = io_file_from_index(ctx, i);
7209 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7210 io_rsrc_data_free(ctx->file_data);
7211 ctx->file_data = NULL;
7212 ctx->nr_user_files = 0;
7215 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7219 if (!ctx->file_data)
7221 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7223 __io_sqe_files_unregister(ctx);
7227 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7228 __releases(&sqd->lock)
7230 WARN_ON_ONCE(sqd->thread == current);
7233 * Do the dance but not conditional clear_bit() because it'd race with
7234 * other threads incrementing park_pending and setting the bit.
7236 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7237 if (atomic_dec_return(&sqd->park_pending))
7238 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7239 mutex_unlock(&sqd->lock);
7242 static void io_sq_thread_park(struct io_sq_data *sqd)
7243 __acquires(&sqd->lock)
7245 WARN_ON_ONCE(sqd->thread == current);
7247 atomic_inc(&sqd->park_pending);
7248 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7249 mutex_lock(&sqd->lock);
7251 wake_up_process(sqd->thread);
7254 static void io_sq_thread_stop(struct io_sq_data *sqd)
7256 WARN_ON_ONCE(sqd->thread == current);
7257 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7259 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7260 mutex_lock(&sqd->lock);
7262 wake_up_process(sqd->thread);
7263 mutex_unlock(&sqd->lock);
7264 wait_for_completion(&sqd->exited);
7267 static void io_put_sq_data(struct io_sq_data *sqd)
7269 if (refcount_dec_and_test(&sqd->refs)) {
7270 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7272 io_sq_thread_stop(sqd);
7277 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7279 struct io_sq_data *sqd = ctx->sq_data;
7282 io_sq_thread_park(sqd);
7283 list_del_init(&ctx->sqd_list);
7284 io_sqd_update_thread_idle(sqd);
7285 io_sq_thread_unpark(sqd);
7287 io_put_sq_data(sqd);
7288 ctx->sq_data = NULL;
7292 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7294 struct io_ring_ctx *ctx_attach;
7295 struct io_sq_data *sqd;
7298 f = fdget(p->wq_fd);
7300 return ERR_PTR(-ENXIO);
7301 if (f.file->f_op != &io_uring_fops) {
7303 return ERR_PTR(-EINVAL);
7306 ctx_attach = f.file->private_data;
7307 sqd = ctx_attach->sq_data;
7310 return ERR_PTR(-EINVAL);
7312 if (sqd->task_tgid != current->tgid) {
7314 return ERR_PTR(-EPERM);
7317 refcount_inc(&sqd->refs);
7322 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7325 struct io_sq_data *sqd;
7328 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7329 sqd = io_attach_sq_data(p);
7334 /* fall through for EPERM case, setup new sqd/task */
7335 if (PTR_ERR(sqd) != -EPERM)
7339 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7341 return ERR_PTR(-ENOMEM);
7343 atomic_set(&sqd->park_pending, 0);
7344 refcount_set(&sqd->refs, 1);
7345 INIT_LIST_HEAD(&sqd->ctx_list);
7346 mutex_init(&sqd->lock);
7347 init_waitqueue_head(&sqd->wait);
7348 init_completion(&sqd->exited);
7352 #if defined(CONFIG_UNIX)
7354 * Ensure the UNIX gc is aware of our file set, so we are certain that
7355 * the io_uring can be safely unregistered on process exit, even if we have
7356 * loops in the file referencing.
7358 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7360 struct sock *sk = ctx->ring_sock->sk;
7361 struct scm_fp_list *fpl;
7362 struct sk_buff *skb;
7365 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7369 skb = alloc_skb(0, GFP_KERNEL);
7378 fpl->user = get_uid(current_user());
7379 for (i = 0; i < nr; i++) {
7380 struct file *file = io_file_from_index(ctx, i + offset);
7384 fpl->fp[nr_files] = get_file(file);
7385 unix_inflight(fpl->user, fpl->fp[nr_files]);
7390 fpl->max = SCM_MAX_FD;
7391 fpl->count = nr_files;
7392 UNIXCB(skb).fp = fpl;
7393 skb->destructor = unix_destruct_scm;
7394 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7395 skb_queue_head(&sk->sk_receive_queue, skb);
7397 for (i = 0; i < nr_files; i++)
7408 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7409 * causes regular reference counting to break down. We rely on the UNIX
7410 * garbage collection to take care of this problem for us.
7412 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7414 unsigned left, total;
7418 left = ctx->nr_user_files;
7420 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7422 ret = __io_sqe_files_scm(ctx, this_files, total);
7426 total += this_files;
7432 while (total < ctx->nr_user_files) {
7433 struct file *file = io_file_from_index(ctx, total);
7443 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7449 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7451 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7453 table->files = kcalloc(nr_tables, sizeof(*table->files), GFP_KERNEL);
7457 for (i = 0; i < nr_tables; i++) {
7458 unsigned int this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7460 table->files[i] = kcalloc(this_files, sizeof(*table->files[i]),
7462 if (!table->files[i])
7464 nr_files -= this_files;
7470 io_free_file_tables(table, nr_tables * IORING_MAX_FILES_TABLE);
7474 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7476 struct file *file = prsrc->file;
7477 #if defined(CONFIG_UNIX)
7478 struct sock *sock = ctx->ring_sock->sk;
7479 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7480 struct sk_buff *skb;
7483 __skb_queue_head_init(&list);
7486 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7487 * remove this entry and rearrange the file array.
7489 skb = skb_dequeue(head);
7491 struct scm_fp_list *fp;
7493 fp = UNIXCB(skb).fp;
7494 for (i = 0; i < fp->count; i++) {
7497 if (fp->fp[i] != file)
7500 unix_notinflight(fp->user, fp->fp[i]);
7501 left = fp->count - 1 - i;
7503 memmove(&fp->fp[i], &fp->fp[i + 1],
7504 left * sizeof(struct file *));
7511 __skb_queue_tail(&list, skb);
7521 __skb_queue_tail(&list, skb);
7523 skb = skb_dequeue(head);
7526 if (skb_peek(&list)) {
7527 spin_lock_irq(&head->lock);
7528 while ((skb = __skb_dequeue(&list)) != NULL)
7529 __skb_queue_tail(head, skb);
7530 spin_unlock_irq(&head->lock);
7537 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7539 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7540 struct io_ring_ctx *ctx = rsrc_data->ctx;
7541 struct io_rsrc_put *prsrc, *tmp;
7543 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7544 list_del(&prsrc->list);
7547 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7548 unsigned long flags;
7550 io_ring_submit_lock(ctx, lock_ring);
7551 spin_lock_irqsave(&ctx->completion_lock, flags);
7552 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7554 io_commit_cqring(ctx);
7555 spin_unlock_irqrestore(&ctx->completion_lock, flags);
7556 io_cqring_ev_posted(ctx);
7557 io_ring_submit_unlock(ctx, lock_ring);
7560 rsrc_data->do_put(ctx, prsrc);
7564 io_rsrc_node_destroy(ref_node);
7565 if (atomic_dec_and_test(&rsrc_data->refs))
7566 complete(&rsrc_data->done);
7569 static void io_rsrc_put_work(struct work_struct *work)
7571 struct io_ring_ctx *ctx;
7572 struct llist_node *node;
7574 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7575 node = llist_del_all(&ctx->rsrc_put_llist);
7578 struct io_rsrc_node *ref_node;
7579 struct llist_node *next = node->next;
7581 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7582 __io_rsrc_put_work(ref_node);
7587 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7589 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7590 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7591 bool first_add = false;
7593 io_rsrc_ref_lock(ctx);
7596 while (!list_empty(&ctx->rsrc_ref_list)) {
7597 node = list_first_entry(&ctx->rsrc_ref_list,
7598 struct io_rsrc_node, node);
7599 /* recycle ref nodes in order */
7602 list_del(&node->node);
7603 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7605 io_rsrc_ref_unlock(ctx);
7608 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7611 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7613 struct io_rsrc_node *ref_node;
7615 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7619 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7624 INIT_LIST_HEAD(&ref_node->node);
7625 INIT_LIST_HEAD(&ref_node->rsrc_list);
7626 ref_node->done = false;
7630 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7631 unsigned nr_args, u64 __user *tags)
7633 __s32 __user *fds = (__s32 __user *) arg;
7637 struct io_rsrc_data *file_data;
7643 if (nr_args > IORING_MAX_FIXED_FILES)
7645 ret = io_rsrc_node_switch_start(ctx);
7649 file_data = io_rsrc_data_alloc(ctx, io_rsrc_file_put, nr_args);
7652 ctx->file_data = file_data;
7654 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7657 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7660 if ((tags && copy_from_user(&tag, &tags[i], sizeof(tag))) ||
7661 copy_from_user(&fd, &fds[i], sizeof(fd))) {
7665 /* allow sparse sets */
7675 if (unlikely(!file))
7679 * Don't allow io_uring instances to be registered. If UNIX
7680 * isn't enabled, then this causes a reference cycle and this
7681 * instance can never get freed. If UNIX is enabled we'll
7682 * handle it just fine, but there's still no point in allowing
7683 * a ring fd as it doesn't support regular read/write anyway.
7685 if (file->f_op == &io_uring_fops) {
7689 ctx->file_data->tags[i] = tag;
7690 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7693 ret = io_sqe_files_scm(ctx);
7695 __io_sqe_files_unregister(ctx);
7699 io_rsrc_node_switch(ctx, NULL);
7702 for (i = 0; i < ctx->nr_user_files; i++) {
7703 file = io_file_from_index(ctx, i);
7707 io_free_file_tables(&ctx->file_table, nr_args);
7708 ctx->nr_user_files = 0;
7710 io_rsrc_data_free(ctx->file_data);
7711 ctx->file_data = NULL;
7715 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7718 #if defined(CONFIG_UNIX)
7719 struct sock *sock = ctx->ring_sock->sk;
7720 struct sk_buff_head *head = &sock->sk_receive_queue;
7721 struct sk_buff *skb;
7724 * See if we can merge this file into an existing skb SCM_RIGHTS
7725 * file set. If there's no room, fall back to allocating a new skb
7726 * and filling it in.
7728 spin_lock_irq(&head->lock);
7729 skb = skb_peek(head);
7731 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7733 if (fpl->count < SCM_MAX_FD) {
7734 __skb_unlink(skb, head);
7735 spin_unlock_irq(&head->lock);
7736 fpl->fp[fpl->count] = get_file(file);
7737 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7739 spin_lock_irq(&head->lock);
7740 __skb_queue_head(head, skb);
7745 spin_unlock_irq(&head->lock);
7752 return __io_sqe_files_scm(ctx, 1, index);
7758 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7759 struct io_rsrc_node *node, void *rsrc)
7761 struct io_rsrc_put *prsrc;
7763 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7767 prsrc->tag = data->tags[idx];
7769 list_add(&prsrc->list, &node->rsrc_list);
7773 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7774 struct io_uring_rsrc_update2 *up,
7777 u64 __user *tags = u64_to_user_ptr(up->tags);
7778 __s32 __user *fds = u64_to_user_ptr(up->data);
7779 struct io_rsrc_data *data = ctx->file_data;
7780 struct io_fixed_file *file_slot;
7784 bool needs_switch = false;
7786 if (!ctx->file_data)
7788 if (up->offset + nr_args > ctx->nr_user_files)
7791 for (done = 0; done < nr_args; done++) {
7794 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7795 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7799 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7803 if (fd == IORING_REGISTER_FILES_SKIP)
7806 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7807 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7809 if (file_slot->file_ptr) {
7810 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7811 err = io_queue_rsrc_removal(data, up->offset + done,
7812 ctx->rsrc_node, file);
7815 file_slot->file_ptr = 0;
7816 needs_switch = true;
7825 * Don't allow io_uring instances to be registered. If
7826 * UNIX isn't enabled, then this causes a reference
7827 * cycle and this instance can never get freed. If UNIX
7828 * is enabled we'll handle it just fine, but there's
7829 * still no point in allowing a ring fd as it doesn't
7830 * support regular read/write anyway.
7832 if (file->f_op == &io_uring_fops) {
7837 data->tags[up->offset + done] = tag;
7838 io_fixed_file_set(file_slot, file);
7839 err = io_sqe_file_register(ctx, file, i);
7841 file_slot->file_ptr = 0;
7849 io_rsrc_node_switch(ctx, data);
7850 return done ? done : err;
7853 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7855 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7857 req = io_put_req_find_next(req);
7858 return req ? &req->work : NULL;
7861 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7862 struct task_struct *task)
7864 struct io_wq_hash *hash;
7865 struct io_wq_data data;
7866 unsigned int concurrency;
7868 hash = ctx->hash_map;
7870 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7872 return ERR_PTR(-ENOMEM);
7873 refcount_set(&hash->refs, 1);
7874 init_waitqueue_head(&hash->wait);
7875 ctx->hash_map = hash;
7880 data.free_work = io_free_work;
7881 data.do_work = io_wq_submit_work;
7883 /* Do QD, or 4 * CPUS, whatever is smallest */
7884 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7886 return io_wq_create(concurrency, &data);
7889 static int io_uring_alloc_task_context(struct task_struct *task,
7890 struct io_ring_ctx *ctx)
7892 struct io_uring_task *tctx;
7895 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7896 if (unlikely(!tctx))
7899 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7900 if (unlikely(ret)) {
7905 tctx->io_wq = io_init_wq_offload(ctx, task);
7906 if (IS_ERR(tctx->io_wq)) {
7907 ret = PTR_ERR(tctx->io_wq);
7908 percpu_counter_destroy(&tctx->inflight);
7914 init_waitqueue_head(&tctx->wait);
7916 atomic_set(&tctx->in_idle, 0);
7917 atomic_set(&tctx->inflight_tracked, 0);
7918 task->io_uring = tctx;
7919 spin_lock_init(&tctx->task_lock);
7920 INIT_WQ_LIST(&tctx->task_list);
7921 tctx->task_state = 0;
7922 init_task_work(&tctx->task_work, tctx_task_work);
7926 void __io_uring_free(struct task_struct *tsk)
7928 struct io_uring_task *tctx = tsk->io_uring;
7930 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7931 WARN_ON_ONCE(tctx->io_wq);
7933 percpu_counter_destroy(&tctx->inflight);
7935 tsk->io_uring = NULL;
7938 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7939 struct io_uring_params *p)
7943 /* Retain compatibility with failing for an invalid attach attempt */
7944 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7945 IORING_SETUP_ATTACH_WQ) {
7948 f = fdget(p->wq_fd);
7952 if (f.file->f_op != &io_uring_fops)
7955 if (ctx->flags & IORING_SETUP_SQPOLL) {
7956 struct task_struct *tsk;
7957 struct io_sq_data *sqd;
7960 sqd = io_get_sq_data(p, &attached);
7966 ctx->sq_creds = get_current_cred();
7968 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7969 if (!ctx->sq_thread_idle)
7970 ctx->sq_thread_idle = HZ;
7972 io_sq_thread_park(sqd);
7973 list_add(&ctx->sqd_list, &sqd->ctx_list);
7974 io_sqd_update_thread_idle(sqd);
7975 /* don't attach to a dying SQPOLL thread, would be racy */
7976 ret = (attached && !sqd->thread) ? -ENXIO : 0;
7977 io_sq_thread_unpark(sqd);
7984 if (p->flags & IORING_SETUP_SQ_AFF) {
7985 int cpu = p->sq_thread_cpu;
7988 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
7995 sqd->task_pid = current->pid;
7996 sqd->task_tgid = current->tgid;
7997 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8004 ret = io_uring_alloc_task_context(tsk, ctx);
8005 wake_up_new_task(tsk);
8008 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8009 /* Can't have SQ_AFF without SQPOLL */
8016 complete(&ctx->sq_data->exited);
8018 io_sq_thread_finish(ctx);
8022 static inline void __io_unaccount_mem(struct user_struct *user,
8023 unsigned long nr_pages)
8025 atomic_long_sub(nr_pages, &user->locked_vm);
8028 static inline int __io_account_mem(struct user_struct *user,
8029 unsigned long nr_pages)
8031 unsigned long page_limit, cur_pages, new_pages;
8033 /* Don't allow more pages than we can safely lock */
8034 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8037 cur_pages = atomic_long_read(&user->locked_vm);
8038 new_pages = cur_pages + nr_pages;
8039 if (new_pages > page_limit)
8041 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8042 new_pages) != cur_pages);
8047 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8050 __io_unaccount_mem(ctx->user, nr_pages);
8052 if (ctx->mm_account)
8053 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8056 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8061 ret = __io_account_mem(ctx->user, nr_pages);
8066 if (ctx->mm_account)
8067 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8072 static void io_mem_free(void *ptr)
8079 page = virt_to_head_page(ptr);
8080 if (put_page_testzero(page))
8081 free_compound_page(page);
8084 static void *io_mem_alloc(size_t size)
8086 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8087 __GFP_NORETRY | __GFP_ACCOUNT;
8089 return (void *) __get_free_pages(gfp_flags, get_order(size));
8092 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8095 struct io_rings *rings;
8096 size_t off, sq_array_size;
8098 off = struct_size(rings, cqes, cq_entries);
8099 if (off == SIZE_MAX)
8103 off = ALIGN(off, SMP_CACHE_BYTES);
8111 sq_array_size = array_size(sizeof(u32), sq_entries);
8112 if (sq_array_size == SIZE_MAX)
8115 if (check_add_overflow(off, sq_array_size, &off))
8121 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8123 struct io_mapped_ubuf *imu = *slot;
8126 if (imu != ctx->dummy_ubuf) {
8127 for (i = 0; i < imu->nr_bvecs; i++)
8128 unpin_user_page(imu->bvec[i].bv_page);
8129 if (imu->acct_pages)
8130 io_unaccount_mem(ctx, imu->acct_pages);
8136 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8138 io_buffer_unmap(ctx, &prsrc->buf);
8142 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8146 for (i = 0; i < ctx->nr_user_bufs; i++)
8147 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8148 kfree(ctx->user_bufs);
8149 io_rsrc_data_free(ctx->buf_data);
8150 ctx->user_bufs = NULL;
8151 ctx->buf_data = NULL;
8152 ctx->nr_user_bufs = 0;
8155 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8162 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8164 __io_sqe_buffers_unregister(ctx);
8168 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8169 void __user *arg, unsigned index)
8171 struct iovec __user *src;
8173 #ifdef CONFIG_COMPAT
8175 struct compat_iovec __user *ciovs;
8176 struct compat_iovec ciov;
8178 ciovs = (struct compat_iovec __user *) arg;
8179 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8182 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8183 dst->iov_len = ciov.iov_len;
8187 src = (struct iovec __user *) arg;
8188 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8194 * Not super efficient, but this is just a registration time. And we do cache
8195 * the last compound head, so generally we'll only do a full search if we don't
8198 * We check if the given compound head page has already been accounted, to
8199 * avoid double accounting it. This allows us to account the full size of the
8200 * page, not just the constituent pages of a huge page.
8202 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8203 int nr_pages, struct page *hpage)
8207 /* check current page array */
8208 for (i = 0; i < nr_pages; i++) {
8209 if (!PageCompound(pages[i]))
8211 if (compound_head(pages[i]) == hpage)
8215 /* check previously registered pages */
8216 for (i = 0; i < ctx->nr_user_bufs; i++) {
8217 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8219 for (j = 0; j < imu->nr_bvecs; j++) {
8220 if (!PageCompound(imu->bvec[j].bv_page))
8222 if (compound_head(imu->bvec[j].bv_page) == hpage)
8230 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8231 int nr_pages, struct io_mapped_ubuf *imu,
8232 struct page **last_hpage)
8236 imu->acct_pages = 0;
8237 for (i = 0; i < nr_pages; i++) {
8238 if (!PageCompound(pages[i])) {
8243 hpage = compound_head(pages[i]);
8244 if (hpage == *last_hpage)
8246 *last_hpage = hpage;
8247 if (headpage_already_acct(ctx, pages, i, hpage))
8249 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8253 if (!imu->acct_pages)
8256 ret = io_account_mem(ctx, imu->acct_pages);
8258 imu->acct_pages = 0;
8262 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8263 struct io_mapped_ubuf **pimu,
8264 struct page **last_hpage)
8266 struct io_mapped_ubuf *imu = NULL;
8267 struct vm_area_struct **vmas = NULL;
8268 struct page **pages = NULL;
8269 unsigned long off, start, end, ubuf;
8271 int ret, pret, nr_pages, i;
8273 if (!iov->iov_base) {
8274 *pimu = ctx->dummy_ubuf;
8278 ubuf = (unsigned long) iov->iov_base;
8279 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8280 start = ubuf >> PAGE_SHIFT;
8281 nr_pages = end - start;
8286 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8290 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8295 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8300 mmap_read_lock(current->mm);
8301 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8303 if (pret == nr_pages) {
8304 /* don't support file backed memory */
8305 for (i = 0; i < nr_pages; i++) {
8306 struct vm_area_struct *vma = vmas[i];
8309 !is_file_hugepages(vma->vm_file)) {
8315 ret = pret < 0 ? pret : -EFAULT;
8317 mmap_read_unlock(current->mm);
8320 * if we did partial map, or found file backed vmas,
8321 * release any pages we did get
8324 unpin_user_pages(pages, pret);
8328 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8330 unpin_user_pages(pages, pret);
8334 off = ubuf & ~PAGE_MASK;
8335 size = iov->iov_len;
8336 for (i = 0; i < nr_pages; i++) {
8339 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8340 imu->bvec[i].bv_page = pages[i];
8341 imu->bvec[i].bv_len = vec_len;
8342 imu->bvec[i].bv_offset = off;
8346 /* store original address for later verification */
8348 imu->ubuf_end = ubuf + iov->iov_len;
8349 imu->nr_bvecs = nr_pages;
8360 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8362 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8363 return ctx->user_bufs ? 0 : -ENOMEM;
8366 static int io_buffer_validate(struct iovec *iov)
8368 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8371 * Don't impose further limits on the size and buffer
8372 * constraints here, we'll -EINVAL later when IO is
8373 * submitted if they are wrong.
8376 return iov->iov_len ? -EFAULT : 0;
8380 /* arbitrary limit, but we need something */
8381 if (iov->iov_len > SZ_1G)
8384 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8390 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8391 unsigned int nr_args, u64 __user *tags)
8393 struct page *last_hpage = NULL;
8394 struct io_rsrc_data *data;
8400 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8402 ret = io_rsrc_node_switch_start(ctx);
8405 data = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, nr_args);
8408 ret = io_buffers_map_alloc(ctx, nr_args);
8410 io_rsrc_data_free(data);
8414 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8417 if (tags && copy_from_user(&tag, &tags[i], sizeof(tag))) {
8421 ret = io_copy_iov(ctx, &iov, arg, i);
8424 ret = io_buffer_validate(&iov);
8427 if (!iov.iov_base && tag) {
8432 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8436 data->tags[i] = tag;
8439 WARN_ON_ONCE(ctx->buf_data);
8441 ctx->buf_data = data;
8443 __io_sqe_buffers_unregister(ctx);
8445 io_rsrc_node_switch(ctx, NULL);
8449 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8450 struct io_uring_rsrc_update2 *up,
8451 unsigned int nr_args)
8453 u64 __user *tags = u64_to_user_ptr(up->tags);
8454 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8455 struct page *last_hpage = NULL;
8456 bool needs_switch = false;
8462 if (up->offset + nr_args > ctx->nr_user_bufs)
8465 for (done = 0; done < nr_args; done++) {
8466 struct io_mapped_ubuf *imu;
8467 int offset = up->offset + done;
8470 err = io_copy_iov(ctx, &iov, iovs, done);
8473 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8477 err = io_buffer_validate(&iov);
8480 if (!iov.iov_base && tag) {
8484 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8488 i = array_index_nospec(offset, ctx->nr_user_bufs);
8489 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8490 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8491 ctx->rsrc_node, ctx->user_bufs[i]);
8492 if (unlikely(err)) {
8493 io_buffer_unmap(ctx, &imu);
8496 ctx->user_bufs[i] = NULL;
8497 needs_switch = true;
8500 ctx->user_bufs[i] = imu;
8501 ctx->buf_data->tags[offset] = tag;
8505 io_rsrc_node_switch(ctx, ctx->buf_data);
8506 return done ? done : err;
8509 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8511 __s32 __user *fds = arg;
8517 if (copy_from_user(&fd, fds, sizeof(*fds)))
8520 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8521 if (IS_ERR(ctx->cq_ev_fd)) {
8522 int ret = PTR_ERR(ctx->cq_ev_fd);
8523 ctx->cq_ev_fd = NULL;
8530 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8532 if (ctx->cq_ev_fd) {
8533 eventfd_ctx_put(ctx->cq_ev_fd);
8534 ctx->cq_ev_fd = NULL;
8541 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8543 struct io_buffer *buf;
8544 unsigned long index;
8546 xa_for_each(&ctx->io_buffers, index, buf)
8547 __io_remove_buffers(ctx, buf, index, -1U);
8550 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8552 struct io_kiocb *req, *nxt;
8554 list_for_each_entry_safe(req, nxt, list, compl.list) {
8555 if (tsk && req->task != tsk)
8557 list_del(&req->compl.list);
8558 kmem_cache_free(req_cachep, req);
8562 static void io_req_caches_free(struct io_ring_ctx *ctx)
8564 struct io_submit_state *submit_state = &ctx->submit_state;
8565 struct io_comp_state *cs = &ctx->submit_state.comp;
8567 mutex_lock(&ctx->uring_lock);
8569 if (submit_state->free_reqs) {
8570 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8571 submit_state->reqs);
8572 submit_state->free_reqs = 0;
8575 io_flush_cached_locked_reqs(ctx, cs);
8576 io_req_cache_free(&cs->free_list, NULL);
8577 mutex_unlock(&ctx->uring_lock);
8580 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8584 if (!atomic_dec_and_test(&data->refs))
8585 wait_for_completion(&data->done);
8589 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8591 io_sq_thread_finish(ctx);
8593 if (ctx->mm_account) {
8594 mmdrop(ctx->mm_account);
8595 ctx->mm_account = NULL;
8598 mutex_lock(&ctx->uring_lock);
8599 if (io_wait_rsrc_data(ctx->buf_data))
8600 __io_sqe_buffers_unregister(ctx);
8601 if (io_wait_rsrc_data(ctx->file_data))
8602 __io_sqe_files_unregister(ctx);
8604 __io_cqring_overflow_flush(ctx, true);
8605 mutex_unlock(&ctx->uring_lock);
8606 io_eventfd_unregister(ctx);
8607 io_destroy_buffers(ctx);
8609 put_cred(ctx->sq_creds);
8611 /* there are no registered resources left, nobody uses it */
8613 io_rsrc_node_destroy(ctx->rsrc_node);
8614 if (ctx->rsrc_backup_node)
8615 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8616 flush_delayed_work(&ctx->rsrc_put_work);
8618 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8619 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8621 #if defined(CONFIG_UNIX)
8622 if (ctx->ring_sock) {
8623 ctx->ring_sock->file = NULL; /* so that iput() is called */
8624 sock_release(ctx->ring_sock);
8628 io_mem_free(ctx->rings);
8629 io_mem_free(ctx->sq_sqes);
8631 percpu_ref_exit(&ctx->refs);
8632 free_uid(ctx->user);
8633 io_req_caches_free(ctx);
8635 io_wq_put_hash(ctx->hash_map);
8636 kfree(ctx->cancel_hash);
8637 kfree(ctx->dummy_ubuf);
8641 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8643 struct io_ring_ctx *ctx = file->private_data;
8646 poll_wait(file, &ctx->cq_wait, wait);
8648 * synchronizes with barrier from wq_has_sleeper call in
8652 if (!io_sqring_full(ctx))
8653 mask |= EPOLLOUT | EPOLLWRNORM;
8656 * Don't flush cqring overflow list here, just do a simple check.
8657 * Otherwise there could possible be ABBA deadlock:
8660 * lock(&ctx->uring_lock);
8662 * lock(&ctx->uring_lock);
8665 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8666 * pushs them to do the flush.
8668 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8669 mask |= EPOLLIN | EPOLLRDNORM;
8674 static int io_uring_fasync(int fd, struct file *file, int on)
8676 struct io_ring_ctx *ctx = file->private_data;
8678 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8681 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8683 const struct cred *creds;
8685 creds = xa_erase(&ctx->personalities, id);
8694 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8696 return io_run_task_work_head(&ctx->exit_task_work);
8699 struct io_tctx_exit {
8700 struct callback_head task_work;
8701 struct completion completion;
8702 struct io_ring_ctx *ctx;
8705 static void io_tctx_exit_cb(struct callback_head *cb)
8707 struct io_uring_task *tctx = current->io_uring;
8708 struct io_tctx_exit *work;
8710 work = container_of(cb, struct io_tctx_exit, task_work);
8712 * When @in_idle, we're in cancellation and it's racy to remove the
8713 * node. It'll be removed by the end of cancellation, just ignore it.
8715 if (!atomic_read(&tctx->in_idle))
8716 io_uring_del_task_file((unsigned long)work->ctx);
8717 complete(&work->completion);
8720 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8722 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8724 return req->ctx == data;
8727 static void io_ring_exit_work(struct work_struct *work)
8729 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8730 unsigned long timeout = jiffies + HZ * 60 * 5;
8731 struct io_tctx_exit exit;
8732 struct io_tctx_node *node;
8736 * If we're doing polled IO and end up having requests being
8737 * submitted async (out-of-line), then completions can come in while
8738 * we're waiting for refs to drop. We need to reap these manually,
8739 * as nobody else will be looking for them.
8742 io_uring_try_cancel_requests(ctx, NULL, NULL);
8744 struct io_sq_data *sqd = ctx->sq_data;
8745 struct task_struct *tsk;
8747 io_sq_thread_park(sqd);
8749 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8750 io_wq_cancel_cb(tsk->io_uring->io_wq,
8751 io_cancel_ctx_cb, ctx, true);
8752 io_sq_thread_unpark(sqd);
8755 WARN_ON_ONCE(time_after(jiffies, timeout));
8756 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8758 init_completion(&exit.completion);
8759 init_task_work(&exit.task_work, io_tctx_exit_cb);
8762 * Some may use context even when all refs and requests have been put,
8763 * and they are free to do so while still holding uring_lock or
8764 * completion_lock, see __io_req_task_submit(). Apart from other work,
8765 * this lock/unlock section also waits them to finish.
8767 mutex_lock(&ctx->uring_lock);
8768 while (!list_empty(&ctx->tctx_list)) {
8769 WARN_ON_ONCE(time_after(jiffies, timeout));
8771 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8773 /* don't spin on a single task if cancellation failed */
8774 list_rotate_left(&ctx->tctx_list);
8775 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8776 if (WARN_ON_ONCE(ret))
8778 wake_up_process(node->task);
8780 mutex_unlock(&ctx->uring_lock);
8781 wait_for_completion(&exit.completion);
8782 mutex_lock(&ctx->uring_lock);
8784 mutex_unlock(&ctx->uring_lock);
8785 spin_lock_irq(&ctx->completion_lock);
8786 spin_unlock_irq(&ctx->completion_lock);
8788 io_ring_ctx_free(ctx);
8791 /* Returns true if we found and killed one or more timeouts */
8792 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8793 struct files_struct *files)
8795 struct io_kiocb *req, *tmp;
8798 spin_lock_irq(&ctx->completion_lock);
8799 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8800 if (io_match_task(req, tsk, files)) {
8801 io_kill_timeout(req, -ECANCELED);
8806 io_commit_cqring(ctx);
8807 spin_unlock_irq(&ctx->completion_lock);
8809 io_cqring_ev_posted(ctx);
8810 return canceled != 0;
8813 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8815 unsigned long index;
8816 struct creds *creds;
8818 mutex_lock(&ctx->uring_lock);
8819 percpu_ref_kill(&ctx->refs);
8821 __io_cqring_overflow_flush(ctx, true);
8822 xa_for_each(&ctx->personalities, index, creds)
8823 io_unregister_personality(ctx, index);
8824 mutex_unlock(&ctx->uring_lock);
8826 io_kill_timeouts(ctx, NULL, NULL);
8827 io_poll_remove_all(ctx, NULL, NULL);
8829 /* if we failed setting up the ctx, we might not have any rings */
8830 io_iopoll_try_reap_events(ctx);
8832 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8834 * Use system_unbound_wq to avoid spawning tons of event kworkers
8835 * if we're exiting a ton of rings at the same time. It just adds
8836 * noise and overhead, there's no discernable change in runtime
8837 * over using system_wq.
8839 queue_work(system_unbound_wq, &ctx->exit_work);
8842 static int io_uring_release(struct inode *inode, struct file *file)
8844 struct io_ring_ctx *ctx = file->private_data;
8846 file->private_data = NULL;
8847 io_ring_ctx_wait_and_kill(ctx);
8851 struct io_task_cancel {
8852 struct task_struct *task;
8853 struct files_struct *files;
8856 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8858 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8859 struct io_task_cancel *cancel = data;
8862 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8863 unsigned long flags;
8864 struct io_ring_ctx *ctx = req->ctx;
8866 /* protect against races with linked timeouts */
8867 spin_lock_irqsave(&ctx->completion_lock, flags);
8868 ret = io_match_task(req, cancel->task, cancel->files);
8869 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8871 ret = io_match_task(req, cancel->task, cancel->files);
8876 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8877 struct task_struct *task,
8878 struct files_struct *files)
8880 struct io_defer_entry *de;
8883 spin_lock_irq(&ctx->completion_lock);
8884 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8885 if (io_match_task(de->req, task, files)) {
8886 list_cut_position(&list, &ctx->defer_list, &de->list);
8890 spin_unlock_irq(&ctx->completion_lock);
8891 if (list_empty(&list))
8894 while (!list_empty(&list)) {
8895 de = list_first_entry(&list, struct io_defer_entry, list);
8896 list_del_init(&de->list);
8897 io_req_complete_failed(de->req, -ECANCELED);
8903 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8905 struct io_tctx_node *node;
8906 enum io_wq_cancel cret;
8909 mutex_lock(&ctx->uring_lock);
8910 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8911 struct io_uring_task *tctx = node->task->io_uring;
8914 * io_wq will stay alive while we hold uring_lock, because it's
8915 * killed after ctx nodes, which requires to take the lock.
8917 if (!tctx || !tctx->io_wq)
8919 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8920 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8922 mutex_unlock(&ctx->uring_lock);
8927 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8928 struct task_struct *task,
8929 struct files_struct *files)
8931 struct io_task_cancel cancel = { .task = task, .files = files, };
8932 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8935 enum io_wq_cancel cret;
8939 ret |= io_uring_try_cancel_iowq(ctx);
8940 } else if (tctx && tctx->io_wq) {
8942 * Cancels requests of all rings, not only @ctx, but
8943 * it's fine as the task is in exit/exec.
8945 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8947 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8950 /* SQPOLL thread does its own polling */
8951 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8952 (ctx->sq_data && ctx->sq_data->thread == current)) {
8953 while (!list_empty_careful(&ctx->iopoll_list)) {
8954 io_iopoll_try_reap_events(ctx);
8959 ret |= io_cancel_defer_files(ctx, task, files);
8960 ret |= io_poll_remove_all(ctx, task, files);
8961 ret |= io_kill_timeouts(ctx, task, files);
8962 ret |= io_run_task_work();
8963 ret |= io_run_ctx_fallback(ctx);
8970 static int __io_uring_add_task_file(struct io_ring_ctx *ctx)
8972 struct io_uring_task *tctx = current->io_uring;
8973 struct io_tctx_node *node;
8976 if (unlikely(!tctx)) {
8977 ret = io_uring_alloc_task_context(current, ctx);
8980 tctx = current->io_uring;
8982 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
8983 node = kmalloc(sizeof(*node), GFP_KERNEL);
8987 node->task = current;
8989 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8996 mutex_lock(&ctx->uring_lock);
8997 list_add(&node->ctx_node, &ctx->tctx_list);
8998 mutex_unlock(&ctx->uring_lock);
9005 * Note that this task has used io_uring. We use it for cancelation purposes.
9007 static inline int io_uring_add_task_file(struct io_ring_ctx *ctx)
9009 struct io_uring_task *tctx = current->io_uring;
9011 if (likely(tctx && tctx->last == ctx))
9013 return __io_uring_add_task_file(ctx);
9017 * Remove this io_uring_file -> task mapping.
9019 static void io_uring_del_task_file(unsigned long index)
9021 struct io_uring_task *tctx = current->io_uring;
9022 struct io_tctx_node *node;
9026 node = xa_erase(&tctx->xa, index);
9030 WARN_ON_ONCE(current != node->task);
9031 WARN_ON_ONCE(list_empty(&node->ctx_node));
9033 mutex_lock(&node->ctx->uring_lock);
9034 list_del(&node->ctx_node);
9035 mutex_unlock(&node->ctx->uring_lock);
9037 if (tctx->last == node->ctx)
9042 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9044 struct io_wq *wq = tctx->io_wq;
9045 struct io_tctx_node *node;
9046 unsigned long index;
9048 xa_for_each(&tctx->xa, index, node)
9049 io_uring_del_task_file(index);
9052 * Must be after io_uring_del_task_file() (removes nodes under
9053 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9056 io_wq_put_and_exit(wq);
9060 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9063 return atomic_read(&tctx->inflight_tracked);
9064 return percpu_counter_sum(&tctx->inflight);
9067 static void io_uring_try_cancel(struct files_struct *files)
9069 struct io_uring_task *tctx = current->io_uring;
9070 struct io_tctx_node *node;
9071 unsigned long index;
9073 xa_for_each(&tctx->xa, index, node) {
9074 struct io_ring_ctx *ctx = node->ctx;
9076 /* sqpoll task will cancel all its requests */
9078 io_uring_try_cancel_requests(ctx, current, files);
9082 /* should only be called by SQPOLL task */
9083 static void io_uring_cancel_sqpoll(struct io_sq_data *sqd)
9085 struct io_uring_task *tctx = current->io_uring;
9086 struct io_ring_ctx *ctx;
9090 if (!current->io_uring)
9093 io_wq_exit_start(tctx->io_wq);
9095 WARN_ON_ONCE(!sqd || sqd->thread != current);
9097 atomic_inc(&tctx->in_idle);
9099 /* read completions before cancelations */
9100 inflight = tctx_inflight(tctx, false);
9103 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9104 io_uring_try_cancel_requests(ctx, current, NULL);
9106 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9108 * If we've seen completions, retry without waiting. This
9109 * avoids a race where a completion comes in before we did
9110 * prepare_to_wait().
9112 if (inflight == tctx_inflight(tctx, false))
9114 finish_wait(&tctx->wait, &wait);
9116 atomic_dec(&tctx->in_idle);
9120 * Find any io_uring fd that this task has registered or done IO on, and cancel
9123 void __io_uring_cancel(struct files_struct *files)
9125 struct io_uring_task *tctx = current->io_uring;
9130 io_wq_exit_start(tctx->io_wq);
9132 /* make sure overflow events are dropped */
9133 atomic_inc(&tctx->in_idle);
9135 /* read completions before cancelations */
9136 inflight = tctx_inflight(tctx, !!files);
9139 io_uring_try_cancel(files);
9140 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9143 * If we've seen completions, retry without waiting. This
9144 * avoids a race where a completion comes in before we did
9145 * prepare_to_wait().
9147 if (inflight == tctx_inflight(tctx, !!files))
9149 finish_wait(&tctx->wait, &wait);
9151 atomic_dec(&tctx->in_idle);
9153 io_uring_clean_tctx(tctx);
9155 /* for exec all current's requests should be gone, kill tctx */
9156 __io_uring_free(current);
9160 static void *io_uring_validate_mmap_request(struct file *file,
9161 loff_t pgoff, size_t sz)
9163 struct io_ring_ctx *ctx = file->private_data;
9164 loff_t offset = pgoff << PAGE_SHIFT;
9169 case IORING_OFF_SQ_RING:
9170 case IORING_OFF_CQ_RING:
9173 case IORING_OFF_SQES:
9177 return ERR_PTR(-EINVAL);
9180 page = virt_to_head_page(ptr);
9181 if (sz > page_size(page))
9182 return ERR_PTR(-EINVAL);
9189 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9191 size_t sz = vma->vm_end - vma->vm_start;
9195 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9197 return PTR_ERR(ptr);
9199 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9200 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9203 #else /* !CONFIG_MMU */
9205 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9207 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9210 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9212 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9215 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9216 unsigned long addr, unsigned long len,
9217 unsigned long pgoff, unsigned long flags)
9221 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9223 return PTR_ERR(ptr);
9225 return (unsigned long) ptr;
9228 #endif /* !CONFIG_MMU */
9230 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9235 if (!io_sqring_full(ctx))
9237 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9239 if (!io_sqring_full(ctx))
9242 } while (!signal_pending(current));
9244 finish_wait(&ctx->sqo_sq_wait, &wait);
9248 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9249 struct __kernel_timespec __user **ts,
9250 const sigset_t __user **sig)
9252 struct io_uring_getevents_arg arg;
9255 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9256 * is just a pointer to the sigset_t.
9258 if (!(flags & IORING_ENTER_EXT_ARG)) {
9259 *sig = (const sigset_t __user *) argp;
9265 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9266 * timespec and sigset_t pointers if good.
9268 if (*argsz != sizeof(arg))
9270 if (copy_from_user(&arg, argp, sizeof(arg)))
9272 *sig = u64_to_user_ptr(arg.sigmask);
9273 *argsz = arg.sigmask_sz;
9274 *ts = u64_to_user_ptr(arg.ts);
9278 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9279 u32, min_complete, u32, flags, const void __user *, argp,
9282 struct io_ring_ctx *ctx;
9289 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9290 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9294 if (unlikely(!f.file))
9298 if (unlikely(f.file->f_op != &io_uring_fops))
9302 ctx = f.file->private_data;
9303 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9307 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9311 * For SQ polling, the thread will do all submissions and completions.
9312 * Just return the requested submit count, and wake the thread if
9316 if (ctx->flags & IORING_SETUP_SQPOLL) {
9317 io_cqring_overflow_flush(ctx, false);
9320 if (unlikely(ctx->sq_data->thread == NULL)) {
9323 if (flags & IORING_ENTER_SQ_WAKEUP)
9324 wake_up(&ctx->sq_data->wait);
9325 if (flags & IORING_ENTER_SQ_WAIT) {
9326 ret = io_sqpoll_wait_sq(ctx);
9330 submitted = to_submit;
9331 } else if (to_submit) {
9332 ret = io_uring_add_task_file(ctx);
9335 mutex_lock(&ctx->uring_lock);
9336 submitted = io_submit_sqes(ctx, to_submit);
9337 mutex_unlock(&ctx->uring_lock);
9339 if (submitted != to_submit)
9342 if (flags & IORING_ENTER_GETEVENTS) {
9343 const sigset_t __user *sig;
9344 struct __kernel_timespec __user *ts;
9346 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9350 min_complete = min(min_complete, ctx->cq_entries);
9353 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9354 * space applications don't need to do io completion events
9355 * polling again, they can rely on io_sq_thread to do polling
9356 * work, which can reduce cpu usage and uring_lock contention.
9358 if (ctx->flags & IORING_SETUP_IOPOLL &&
9359 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9360 ret = io_iopoll_check(ctx, min_complete);
9362 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9367 percpu_ref_put(&ctx->refs);
9370 return submitted ? submitted : ret;
9373 #ifdef CONFIG_PROC_FS
9374 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9375 const struct cred *cred)
9377 struct user_namespace *uns = seq_user_ns(m);
9378 struct group_info *gi;
9383 seq_printf(m, "%5d\n", id);
9384 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9385 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9386 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9387 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9388 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9389 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9390 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9391 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9392 seq_puts(m, "\n\tGroups:\t");
9393 gi = cred->group_info;
9394 for (g = 0; g < gi->ngroups; g++) {
9395 seq_put_decimal_ull(m, g ? " " : "",
9396 from_kgid_munged(uns, gi->gid[g]));
9398 seq_puts(m, "\n\tCapEff:\t");
9399 cap = cred->cap_effective;
9400 CAP_FOR_EACH_U32(__capi)
9401 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9406 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9408 struct io_sq_data *sq = NULL;
9413 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9414 * since fdinfo case grabs it in the opposite direction of normal use
9415 * cases. If we fail to get the lock, we just don't iterate any
9416 * structures that could be going away outside the io_uring mutex.
9418 has_lock = mutex_trylock(&ctx->uring_lock);
9420 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9426 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9427 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9428 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9429 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9430 struct file *f = io_file_from_index(ctx, i);
9433 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9435 seq_printf(m, "%5u: <none>\n", i);
9437 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9438 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9439 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9440 unsigned int len = buf->ubuf_end - buf->ubuf;
9442 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9444 if (has_lock && !xa_empty(&ctx->personalities)) {
9445 unsigned long index;
9446 const struct cred *cred;
9448 seq_printf(m, "Personalities:\n");
9449 xa_for_each(&ctx->personalities, index, cred)
9450 io_uring_show_cred(m, index, cred);
9452 seq_printf(m, "PollList:\n");
9453 spin_lock_irq(&ctx->completion_lock);
9454 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9455 struct hlist_head *list = &ctx->cancel_hash[i];
9456 struct io_kiocb *req;
9458 hlist_for_each_entry(req, list, hash_node)
9459 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9460 req->task->task_works != NULL);
9462 spin_unlock_irq(&ctx->completion_lock);
9464 mutex_unlock(&ctx->uring_lock);
9467 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9469 struct io_ring_ctx *ctx = f->private_data;
9471 if (percpu_ref_tryget(&ctx->refs)) {
9472 __io_uring_show_fdinfo(ctx, m);
9473 percpu_ref_put(&ctx->refs);
9478 static const struct file_operations io_uring_fops = {
9479 .release = io_uring_release,
9480 .mmap = io_uring_mmap,
9482 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9483 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9485 .poll = io_uring_poll,
9486 .fasync = io_uring_fasync,
9487 #ifdef CONFIG_PROC_FS
9488 .show_fdinfo = io_uring_show_fdinfo,
9492 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9493 struct io_uring_params *p)
9495 struct io_rings *rings;
9496 size_t size, sq_array_offset;
9498 /* make sure these are sane, as we already accounted them */
9499 ctx->sq_entries = p->sq_entries;
9500 ctx->cq_entries = p->cq_entries;
9502 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9503 if (size == SIZE_MAX)
9506 rings = io_mem_alloc(size);
9511 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9512 rings->sq_ring_mask = p->sq_entries - 1;
9513 rings->cq_ring_mask = p->cq_entries - 1;
9514 rings->sq_ring_entries = p->sq_entries;
9515 rings->cq_ring_entries = p->cq_entries;
9516 ctx->sq_mask = rings->sq_ring_mask;
9517 ctx->cq_mask = rings->cq_ring_mask;
9519 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9520 if (size == SIZE_MAX) {
9521 io_mem_free(ctx->rings);
9526 ctx->sq_sqes = io_mem_alloc(size);
9527 if (!ctx->sq_sqes) {
9528 io_mem_free(ctx->rings);
9536 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9540 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9544 ret = io_uring_add_task_file(ctx);
9549 fd_install(fd, file);
9554 * Allocate an anonymous fd, this is what constitutes the application
9555 * visible backing of an io_uring instance. The application mmaps this
9556 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9557 * we have to tie this fd to a socket for file garbage collection purposes.
9559 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9562 #if defined(CONFIG_UNIX)
9565 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9568 return ERR_PTR(ret);
9571 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9572 O_RDWR | O_CLOEXEC);
9573 #if defined(CONFIG_UNIX)
9575 sock_release(ctx->ring_sock);
9576 ctx->ring_sock = NULL;
9578 ctx->ring_sock->file = file;
9584 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9585 struct io_uring_params __user *params)
9587 struct io_ring_ctx *ctx;
9593 if (entries > IORING_MAX_ENTRIES) {
9594 if (!(p->flags & IORING_SETUP_CLAMP))
9596 entries = IORING_MAX_ENTRIES;
9600 * Use twice as many entries for the CQ ring. It's possible for the
9601 * application to drive a higher depth than the size of the SQ ring,
9602 * since the sqes are only used at submission time. This allows for
9603 * some flexibility in overcommitting a bit. If the application has
9604 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9605 * of CQ ring entries manually.
9607 p->sq_entries = roundup_pow_of_two(entries);
9608 if (p->flags & IORING_SETUP_CQSIZE) {
9610 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9611 * to a power-of-two, if it isn't already. We do NOT impose
9612 * any cq vs sq ring sizing.
9616 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9617 if (!(p->flags & IORING_SETUP_CLAMP))
9619 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9621 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9622 if (p->cq_entries < p->sq_entries)
9625 p->cq_entries = 2 * p->sq_entries;
9628 ctx = io_ring_ctx_alloc(p);
9631 ctx->compat = in_compat_syscall();
9632 if (!capable(CAP_IPC_LOCK))
9633 ctx->user = get_uid(current_user());
9636 * This is just grabbed for accounting purposes. When a process exits,
9637 * the mm is exited and dropped before the files, hence we need to hang
9638 * on to this mm purely for the purposes of being able to unaccount
9639 * memory (locked/pinned vm). It's not used for anything else.
9641 mmgrab(current->mm);
9642 ctx->mm_account = current->mm;
9644 ret = io_allocate_scq_urings(ctx, p);
9648 ret = io_sq_offload_create(ctx, p);
9651 /* always set a rsrc node */
9652 ret = io_rsrc_node_switch_start(ctx);
9655 io_rsrc_node_switch(ctx, NULL);
9657 memset(&p->sq_off, 0, sizeof(p->sq_off));
9658 p->sq_off.head = offsetof(struct io_rings, sq.head);
9659 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9660 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9661 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9662 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9663 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9664 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9666 memset(&p->cq_off, 0, sizeof(p->cq_off));
9667 p->cq_off.head = offsetof(struct io_rings, cq.head);
9668 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9669 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9670 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9671 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9672 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9673 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9675 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9676 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9677 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9678 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9679 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9680 IORING_FEAT_RSRC_TAGS;
9682 if (copy_to_user(params, p, sizeof(*p))) {
9687 file = io_uring_get_file(ctx);
9689 ret = PTR_ERR(file);
9694 * Install ring fd as the very last thing, so we don't risk someone
9695 * having closed it before we finish setup
9697 ret = io_uring_install_fd(ctx, file);
9699 /* fput will clean it up */
9704 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9707 io_ring_ctx_wait_and_kill(ctx);
9712 * Sets up an aio uring context, and returns the fd. Applications asks for a
9713 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9714 * params structure passed in.
9716 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9718 struct io_uring_params p;
9721 if (copy_from_user(&p, params, sizeof(p)))
9723 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9728 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9729 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9730 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9731 IORING_SETUP_R_DISABLED))
9734 return io_uring_create(entries, &p, params);
9737 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9738 struct io_uring_params __user *, params)
9740 return io_uring_setup(entries, params);
9743 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9745 struct io_uring_probe *p;
9749 size = struct_size(p, ops, nr_args);
9750 if (size == SIZE_MAX)
9752 p = kzalloc(size, GFP_KERNEL);
9757 if (copy_from_user(p, arg, size))
9760 if (memchr_inv(p, 0, size))
9763 p->last_op = IORING_OP_LAST - 1;
9764 if (nr_args > IORING_OP_LAST)
9765 nr_args = IORING_OP_LAST;
9767 for (i = 0; i < nr_args; i++) {
9769 if (!io_op_defs[i].not_supported)
9770 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9775 if (copy_to_user(arg, p, size))
9782 static int io_register_personality(struct io_ring_ctx *ctx)
9784 const struct cred *creds;
9788 creds = get_current_cred();
9790 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9791 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9798 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9799 unsigned int nr_args)
9801 struct io_uring_restriction *res;
9805 /* Restrictions allowed only if rings started disabled */
9806 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9809 /* We allow only a single restrictions registration */
9810 if (ctx->restrictions.registered)
9813 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9816 size = array_size(nr_args, sizeof(*res));
9817 if (size == SIZE_MAX)
9820 res = memdup_user(arg, size);
9822 return PTR_ERR(res);
9826 for (i = 0; i < nr_args; i++) {
9827 switch (res[i].opcode) {
9828 case IORING_RESTRICTION_REGISTER_OP:
9829 if (res[i].register_op >= IORING_REGISTER_LAST) {
9834 __set_bit(res[i].register_op,
9835 ctx->restrictions.register_op);
9837 case IORING_RESTRICTION_SQE_OP:
9838 if (res[i].sqe_op >= IORING_OP_LAST) {
9843 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9845 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9846 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9848 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9849 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9858 /* Reset all restrictions if an error happened */
9860 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9862 ctx->restrictions.registered = true;
9868 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9870 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9873 if (ctx->restrictions.registered)
9874 ctx->restricted = 1;
9876 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9877 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9878 wake_up(&ctx->sq_data->wait);
9882 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9883 struct io_uring_rsrc_update2 *up,
9891 if (check_add_overflow(up->offset, nr_args, &tmp))
9893 err = io_rsrc_node_switch_start(ctx);
9898 case IORING_RSRC_FILE:
9899 return __io_sqe_files_update(ctx, up, nr_args);
9900 case IORING_RSRC_BUFFER:
9901 return __io_sqe_buffers_update(ctx, up, nr_args);
9906 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9909 struct io_uring_rsrc_update2 up;
9913 memset(&up, 0, sizeof(up));
9914 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9916 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9919 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9920 unsigned size, unsigned type)
9922 struct io_uring_rsrc_update2 up;
9924 if (size != sizeof(up))
9926 if (copy_from_user(&up, arg, sizeof(up)))
9928 if (!up.nr || up.resv)
9930 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9933 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9934 unsigned int size, unsigned int type)
9936 struct io_uring_rsrc_register rr;
9938 /* keep it extendible */
9939 if (size != sizeof(rr))
9942 memset(&rr, 0, sizeof(rr));
9943 if (copy_from_user(&rr, arg, size))
9945 if (!rr.nr || rr.resv || rr.resv2)
9949 case IORING_RSRC_FILE:
9950 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9951 rr.nr, u64_to_user_ptr(rr.tags));
9952 case IORING_RSRC_BUFFER:
9953 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9954 rr.nr, u64_to_user_ptr(rr.tags));
9959 static bool io_register_op_must_quiesce(int op)
9962 case IORING_REGISTER_BUFFERS:
9963 case IORING_UNREGISTER_BUFFERS:
9964 case IORING_REGISTER_FILES:
9965 case IORING_UNREGISTER_FILES:
9966 case IORING_REGISTER_FILES_UPDATE:
9967 case IORING_REGISTER_PROBE:
9968 case IORING_REGISTER_PERSONALITY:
9969 case IORING_UNREGISTER_PERSONALITY:
9970 case IORING_REGISTER_FILES2:
9971 case IORING_REGISTER_FILES_UPDATE2:
9972 case IORING_REGISTER_BUFFERS2:
9973 case IORING_REGISTER_BUFFERS_UPDATE:
9980 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9981 void __user *arg, unsigned nr_args)
9982 __releases(ctx->uring_lock)
9983 __acquires(ctx->uring_lock)
9988 * We're inside the ring mutex, if the ref is already dying, then
9989 * someone else killed the ctx or is already going through
9990 * io_uring_register().
9992 if (percpu_ref_is_dying(&ctx->refs))
9995 if (ctx->restricted) {
9996 if (opcode >= IORING_REGISTER_LAST)
9998 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
9999 if (!test_bit(opcode, ctx->restrictions.register_op))
10003 if (io_register_op_must_quiesce(opcode)) {
10004 percpu_ref_kill(&ctx->refs);
10007 * Drop uring mutex before waiting for references to exit. If
10008 * another thread is currently inside io_uring_enter() it might
10009 * need to grab the uring_lock to make progress. If we hold it
10010 * here across the drain wait, then we can deadlock. It's safe
10011 * to drop the mutex here, since no new references will come in
10012 * after we've killed the percpu ref.
10014 mutex_unlock(&ctx->uring_lock);
10016 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10019 ret = io_run_task_work_sig();
10023 mutex_lock(&ctx->uring_lock);
10026 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10032 case IORING_REGISTER_BUFFERS:
10033 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10035 case IORING_UNREGISTER_BUFFERS:
10037 if (arg || nr_args)
10039 ret = io_sqe_buffers_unregister(ctx);
10041 case IORING_REGISTER_FILES:
10042 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10044 case IORING_UNREGISTER_FILES:
10046 if (arg || nr_args)
10048 ret = io_sqe_files_unregister(ctx);
10050 case IORING_REGISTER_FILES_UPDATE:
10051 ret = io_register_files_update(ctx, arg, nr_args);
10053 case IORING_REGISTER_EVENTFD:
10054 case IORING_REGISTER_EVENTFD_ASYNC:
10058 ret = io_eventfd_register(ctx, arg);
10061 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10062 ctx->eventfd_async = 1;
10064 ctx->eventfd_async = 0;
10066 case IORING_UNREGISTER_EVENTFD:
10068 if (arg || nr_args)
10070 ret = io_eventfd_unregister(ctx);
10072 case IORING_REGISTER_PROBE:
10074 if (!arg || nr_args > 256)
10076 ret = io_probe(ctx, arg, nr_args);
10078 case IORING_REGISTER_PERSONALITY:
10080 if (arg || nr_args)
10082 ret = io_register_personality(ctx);
10084 case IORING_UNREGISTER_PERSONALITY:
10088 ret = io_unregister_personality(ctx, nr_args);
10090 case IORING_REGISTER_ENABLE_RINGS:
10092 if (arg || nr_args)
10094 ret = io_register_enable_rings(ctx);
10096 case IORING_REGISTER_RESTRICTIONS:
10097 ret = io_register_restrictions(ctx, arg, nr_args);
10099 case IORING_REGISTER_FILES2:
10100 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10102 case IORING_REGISTER_FILES_UPDATE2:
10103 ret = io_register_rsrc_update(ctx, arg, nr_args,
10106 case IORING_REGISTER_BUFFERS2:
10107 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10109 case IORING_REGISTER_BUFFERS_UPDATE:
10110 ret = io_register_rsrc_update(ctx, arg, nr_args,
10111 IORING_RSRC_BUFFER);
10118 if (io_register_op_must_quiesce(opcode)) {
10119 /* bring the ctx back to life */
10120 percpu_ref_reinit(&ctx->refs);
10121 reinit_completion(&ctx->ref_comp);
10126 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10127 void __user *, arg, unsigned int, nr_args)
10129 struct io_ring_ctx *ctx;
10138 if (f.file->f_op != &io_uring_fops)
10141 ctx = f.file->private_data;
10143 io_run_task_work();
10145 mutex_lock(&ctx->uring_lock);
10146 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10147 mutex_unlock(&ctx->uring_lock);
10148 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10149 ctx->cq_ev_fd != NULL, ret);
10155 static int __init io_uring_init(void)
10157 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10158 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10159 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10162 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10163 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10164 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10165 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10166 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10167 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10168 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10169 BUILD_BUG_SQE_ELEM(8, __u64, off);
10170 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10171 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10172 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10173 BUILD_BUG_SQE_ELEM(24, __u32, len);
10174 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10175 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10176 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10177 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10178 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10179 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10180 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10181 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10182 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10183 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10184 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10185 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10186 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10187 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10188 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10189 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10190 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10191 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10192 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10194 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10195 sizeof(struct io_uring_rsrc_update));
10196 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10197 sizeof(struct io_uring_rsrc_update2));
10198 /* should fit into one byte */
10199 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10201 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10202 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10203 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10207 __initcall(io_uring_init);