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_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81 #include <linux/tracehook.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp;
116 u32 tail ____cacheline_aligned_in_smp;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq, cq;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask, cq_ring_mask;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries, cq_ring_entries;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
197 enum io_uring_cmd_flags {
198 IO_URING_F_NONBLOCK = 1,
199 IO_URING_F_COMPLETE_DEFER = 2,
202 struct io_mapped_ubuf {
205 unsigned int nr_bvecs;
206 unsigned long acct_pages;
207 struct bio_vec bvec[];
212 struct io_overflow_cqe {
213 struct io_uring_cqe cqe;
214 struct list_head list;
217 struct io_fixed_file {
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr;
223 struct list_head list;
228 struct io_mapped_ubuf *buf;
232 struct io_file_table {
233 struct io_fixed_file *files;
236 struct io_rsrc_node {
237 struct percpu_ref refs;
238 struct list_head node;
239 struct list_head rsrc_list;
240 struct io_rsrc_data *rsrc_data;
241 struct llist_node llist;
245 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
247 struct io_rsrc_data {
248 struct io_ring_ctx *ctx;
254 struct completion done;
259 struct list_head list;
265 struct io_restriction {
266 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268 u8 sqe_flags_allowed;
269 u8 sqe_flags_required;
274 IO_SQ_THREAD_SHOULD_STOP = 0,
275 IO_SQ_THREAD_SHOULD_PARK,
280 atomic_t park_pending;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list;
286 struct task_struct *thread;
287 struct wait_queue_head wait;
289 unsigned sq_thread_idle;
295 struct completion exited;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_submit_link {
303 struct io_kiocb *head;
304 struct io_kiocb *last;
307 struct io_submit_state {
308 struct blk_plug plug;
309 struct io_submit_link link;
312 * io_kiocb alloc cache
314 void *reqs[IO_REQ_CACHE_SIZE];
315 unsigned int free_reqs;
320 * Batch completion logic
322 struct io_kiocb *compl_reqs[IO_COMPL_BATCH];
323 unsigned int compl_nr;
324 /* inline/task_work completion list, under ->uring_lock */
325 struct list_head free_list;
327 unsigned int ios_left;
331 /* const or read-mostly hot data */
333 struct percpu_ref refs;
335 struct io_rings *rings;
337 unsigned int compat: 1;
338 unsigned int drain_next: 1;
339 unsigned int eventfd_async: 1;
340 unsigned int restricted: 1;
341 unsigned int off_timeout_used: 1;
342 unsigned int drain_active: 1;
343 } ____cacheline_aligned_in_smp;
345 /* submission data */
347 struct mutex uring_lock;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 struct io_uring_sqe *sq_sqes;
362 unsigned cached_sq_head;
364 struct list_head defer_list;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node *rsrc_node;
371 struct io_file_table file_table;
372 unsigned nr_user_files;
373 unsigned nr_user_bufs;
374 struct io_mapped_ubuf **user_bufs;
376 struct io_submit_state submit_state;
377 struct list_head timeout_list;
378 struct list_head ltimeout_list;
379 struct list_head cq_overflow_list;
380 struct xarray io_buffers;
381 struct xarray personalities;
383 unsigned sq_thread_idle;
384 } ____cacheline_aligned_in_smp;
386 /* IRQ completion list, under ->completion_lock */
387 struct list_head locked_free_list;
388 unsigned int locked_free_nr;
390 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
391 struct io_sq_data *sq_data; /* if using sq thread polling */
393 struct wait_queue_head sqo_sq_wait;
394 struct list_head sqd_list;
396 unsigned long check_cq_overflow;
399 unsigned cached_cq_tail;
401 struct eventfd_ctx *cq_ev_fd;
402 struct wait_queue_head poll_wait;
403 struct wait_queue_head cq_wait;
405 atomic_t cq_timeouts;
406 struct fasync_struct *cq_fasync;
407 unsigned cq_last_tm_flush;
408 } ____cacheline_aligned_in_smp;
411 spinlock_t completion_lock;
413 spinlock_t timeout_lock;
416 * ->iopoll_list is protected by the ctx->uring_lock for
417 * io_uring instances that don't use IORING_SETUP_SQPOLL.
418 * For SQPOLL, only the single threaded io_sq_thread() will
419 * manipulate the list, hence no extra locking is needed there.
421 struct list_head iopoll_list;
422 struct hlist_head *cancel_hash;
423 unsigned cancel_hash_bits;
424 bool poll_multi_queue;
425 } ____cacheline_aligned_in_smp;
427 struct io_restriction restrictions;
429 /* slow path rsrc auxilary data, used by update/register */
431 struct io_rsrc_node *rsrc_backup_node;
432 struct io_mapped_ubuf *dummy_ubuf;
433 struct io_rsrc_data *file_data;
434 struct io_rsrc_data *buf_data;
436 struct delayed_work rsrc_put_work;
437 struct llist_head rsrc_put_llist;
438 struct list_head rsrc_ref_list;
439 spinlock_t rsrc_ref_lock;
442 /* Keep this last, we don't need it for the fast path */
444 #if defined(CONFIG_UNIX)
445 struct socket *ring_sock;
447 /* hashed buffered write serialization */
448 struct io_wq_hash *hash_map;
450 /* Only used for accounting purposes */
451 struct user_struct *user;
452 struct mm_struct *mm_account;
454 /* ctx exit and cancelation */
455 struct llist_head fallback_llist;
456 struct delayed_work fallback_work;
457 struct work_struct exit_work;
458 struct list_head tctx_list;
459 struct completion ref_comp;
463 struct io_uring_task {
464 /* submission side */
467 struct wait_queue_head wait;
468 const struct io_ring_ctx *last;
470 struct percpu_counter inflight;
471 atomic_t inflight_tracked;
474 spinlock_t task_lock;
475 struct io_wq_work_list task_list;
476 struct callback_head task_work;
481 * First field must be the file pointer in all the
482 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
484 struct io_poll_iocb {
486 struct wait_queue_head *head;
490 struct wait_queue_entry wait;
493 struct io_poll_update {
499 bool update_user_data;
508 struct io_timeout_data {
509 struct io_kiocb *req;
510 struct hrtimer timer;
511 struct timespec64 ts;
512 enum hrtimer_mode mode;
518 struct sockaddr __user *addr;
519 int __user *addr_len;
522 unsigned long nofile;
542 struct list_head list;
543 /* head of the link, used by linked timeouts only */
544 struct io_kiocb *head;
545 /* for linked completions */
546 struct io_kiocb *prev;
549 struct io_timeout_rem {
554 struct timespec64 ts;
560 /* NOTE: kiocb has the file as the first member, so don't do it here */
568 struct sockaddr __user *addr;
575 struct compat_msghdr __user *umsg_compat;
576 struct user_msghdr __user *umsg;
582 struct io_buffer *kbuf;
589 struct filename *filename;
591 unsigned long nofile;
594 struct io_rsrc_update {
620 struct epoll_event event;
624 struct file *file_out;
625 struct file *file_in;
632 struct io_provide_buf {
646 const char __user *filename;
647 struct statx __user *buffer;
659 struct filename *oldpath;
660 struct filename *newpath;
668 struct filename *filename;
675 struct filename *filename;
681 struct filename *oldpath;
682 struct filename *newpath;
689 struct filename *oldpath;
690 struct filename *newpath;
694 struct io_completion {
699 struct io_async_connect {
700 struct sockaddr_storage address;
703 struct io_async_msghdr {
704 struct iovec fast_iov[UIO_FASTIOV];
705 /* points to an allocated iov, if NULL we use fast_iov instead */
706 struct iovec *free_iov;
707 struct sockaddr __user *uaddr;
709 struct sockaddr_storage addr;
713 struct iovec fast_iov[UIO_FASTIOV];
714 const struct iovec *free_iovec;
715 struct iov_iter iter;
716 struct iov_iter_state iter_state;
718 struct wait_page_queue wpq;
722 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
723 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
724 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
725 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
726 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
727 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
729 /* first byte is taken by user flags, shift it to not overlap */
734 REQ_F_LINK_TIMEOUT_BIT,
735 REQ_F_NEED_CLEANUP_BIT,
737 REQ_F_BUFFER_SELECTED_BIT,
738 REQ_F_COMPLETE_INLINE_BIT,
742 REQ_F_ARM_LTIMEOUT_BIT,
743 /* keep async read/write and isreg together and in order */
744 REQ_F_NOWAIT_READ_BIT,
745 REQ_F_NOWAIT_WRITE_BIT,
748 /* not a real bit, just to check we're not overflowing the space */
754 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
755 /* drain existing IO first */
756 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
758 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
759 /* doesn't sever on completion < 0 */
760 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
762 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
763 /* IOSQE_BUFFER_SELECT */
764 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
766 /* fail rest of links */
767 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
768 /* on inflight list, should be cancelled and waited on exit reliably */
769 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
770 /* read/write uses file position */
771 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
772 /* must not punt to workers */
773 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
774 /* has or had linked timeout */
775 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
777 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
778 /* already went through poll handler */
779 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
780 /* buffer already selected */
781 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
782 /* completion is deferred through io_comp_state */
783 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
784 /* caller should reissue async */
785 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
786 /* supports async reads */
787 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
788 /* supports async writes */
789 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
791 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
792 /* has creds assigned */
793 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
794 /* skip refcounting if not set */
795 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
796 /* there is a linked timeout that has to be armed */
797 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
801 struct io_poll_iocb poll;
802 struct io_poll_iocb *double_poll;
805 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
807 struct io_task_work {
809 struct io_wq_work_node node;
810 struct llist_node fallback_node;
812 io_req_tw_func_t func;
816 IORING_RSRC_FILE = 0,
817 IORING_RSRC_BUFFER = 1,
821 * NOTE! Each of the iocb union members has the file pointer
822 * as the first entry in their struct definition. So you can
823 * access the file pointer through any of the sub-structs,
824 * or directly as just 'ki_filp' in this struct.
830 struct io_poll_iocb poll;
831 struct io_poll_update poll_update;
832 struct io_accept accept;
834 struct io_cancel cancel;
835 struct io_timeout timeout;
836 struct io_timeout_rem timeout_rem;
837 struct io_connect connect;
838 struct io_sr_msg sr_msg;
840 struct io_close close;
841 struct io_rsrc_update rsrc_update;
842 struct io_fadvise fadvise;
843 struct io_madvise madvise;
844 struct io_epoll epoll;
845 struct io_splice splice;
846 struct io_provide_buf pbuf;
847 struct io_statx statx;
848 struct io_shutdown shutdown;
849 struct io_rename rename;
850 struct io_unlink unlink;
851 struct io_mkdir mkdir;
852 struct io_symlink symlink;
853 struct io_hardlink hardlink;
854 /* use only after cleaning per-op data, see io_clean_op() */
855 struct io_completion compl;
858 /* opcode allocated if it needs to store data for async defer */
861 /* polled IO has completed */
867 struct io_ring_ctx *ctx;
870 struct task_struct *task;
873 struct io_kiocb *link;
874 struct percpu_ref *fixed_rsrc_refs;
876 /* used with ctx->iopoll_list with reads/writes */
877 struct list_head inflight_entry;
878 struct io_task_work io_task_work;
879 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
880 struct hlist_node hash_node;
881 struct async_poll *apoll;
882 struct io_wq_work work;
883 const struct cred *creds;
885 /* store used ubuf, so we can prevent reloading */
886 struct io_mapped_ubuf *imu;
889 struct io_tctx_node {
890 struct list_head ctx_node;
891 struct task_struct *task;
892 struct io_ring_ctx *ctx;
895 struct io_defer_entry {
896 struct list_head list;
897 struct io_kiocb *req;
902 /* needs req->file assigned */
903 unsigned needs_file : 1;
904 /* hash wq insertion if file is a regular file */
905 unsigned hash_reg_file : 1;
906 /* unbound wq insertion if file is a non-regular file */
907 unsigned unbound_nonreg_file : 1;
908 /* opcode is not supported by this kernel */
909 unsigned not_supported : 1;
910 /* set if opcode supports polled "wait" */
912 unsigned pollout : 1;
913 /* op supports buffer selection */
914 unsigned buffer_select : 1;
915 /* do prep async if is going to be punted */
916 unsigned needs_async_setup : 1;
917 /* should block plug */
919 /* size of async data needed, if any */
920 unsigned short async_size;
923 static const struct io_op_def io_op_defs[] = {
924 [IORING_OP_NOP] = {},
925 [IORING_OP_READV] = {
927 .unbound_nonreg_file = 1,
930 .needs_async_setup = 1,
932 .async_size = sizeof(struct io_async_rw),
934 [IORING_OP_WRITEV] = {
937 .unbound_nonreg_file = 1,
939 .needs_async_setup = 1,
941 .async_size = sizeof(struct io_async_rw),
943 [IORING_OP_FSYNC] = {
946 [IORING_OP_READ_FIXED] = {
948 .unbound_nonreg_file = 1,
951 .async_size = sizeof(struct io_async_rw),
953 [IORING_OP_WRITE_FIXED] = {
956 .unbound_nonreg_file = 1,
959 .async_size = sizeof(struct io_async_rw),
961 [IORING_OP_POLL_ADD] = {
963 .unbound_nonreg_file = 1,
965 [IORING_OP_POLL_REMOVE] = {},
966 [IORING_OP_SYNC_FILE_RANGE] = {
969 [IORING_OP_SENDMSG] = {
971 .unbound_nonreg_file = 1,
973 .needs_async_setup = 1,
974 .async_size = sizeof(struct io_async_msghdr),
976 [IORING_OP_RECVMSG] = {
978 .unbound_nonreg_file = 1,
981 .needs_async_setup = 1,
982 .async_size = sizeof(struct io_async_msghdr),
984 [IORING_OP_TIMEOUT] = {
985 .async_size = sizeof(struct io_timeout_data),
987 [IORING_OP_TIMEOUT_REMOVE] = {
988 /* used by timeout updates' prep() */
990 [IORING_OP_ACCEPT] = {
992 .unbound_nonreg_file = 1,
995 [IORING_OP_ASYNC_CANCEL] = {},
996 [IORING_OP_LINK_TIMEOUT] = {
997 .async_size = sizeof(struct io_timeout_data),
999 [IORING_OP_CONNECT] = {
1001 .unbound_nonreg_file = 1,
1003 .needs_async_setup = 1,
1004 .async_size = sizeof(struct io_async_connect),
1006 [IORING_OP_FALLOCATE] = {
1009 [IORING_OP_OPENAT] = {},
1010 [IORING_OP_CLOSE] = {},
1011 [IORING_OP_FILES_UPDATE] = {},
1012 [IORING_OP_STATX] = {},
1013 [IORING_OP_READ] = {
1015 .unbound_nonreg_file = 1,
1019 .async_size = sizeof(struct io_async_rw),
1021 [IORING_OP_WRITE] = {
1024 .unbound_nonreg_file = 1,
1027 .async_size = sizeof(struct io_async_rw),
1029 [IORING_OP_FADVISE] = {
1032 [IORING_OP_MADVISE] = {},
1033 [IORING_OP_SEND] = {
1035 .unbound_nonreg_file = 1,
1038 [IORING_OP_RECV] = {
1040 .unbound_nonreg_file = 1,
1044 [IORING_OP_OPENAT2] = {
1046 [IORING_OP_EPOLL_CTL] = {
1047 .unbound_nonreg_file = 1,
1049 [IORING_OP_SPLICE] = {
1052 .unbound_nonreg_file = 1,
1054 [IORING_OP_PROVIDE_BUFFERS] = {},
1055 [IORING_OP_REMOVE_BUFFERS] = {},
1059 .unbound_nonreg_file = 1,
1061 [IORING_OP_SHUTDOWN] = {
1064 [IORING_OP_RENAMEAT] = {},
1065 [IORING_OP_UNLINKAT] = {},
1066 [IORING_OP_MKDIRAT] = {},
1067 [IORING_OP_SYMLINKAT] = {},
1068 [IORING_OP_LINKAT] = {},
1071 /* requests with any of those set should undergo io_disarm_next() */
1072 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1074 static bool io_disarm_next(struct io_kiocb *req);
1075 static void io_uring_del_tctx_node(unsigned long index);
1076 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1077 struct task_struct *task,
1079 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1081 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1082 long res, unsigned int cflags);
1083 static void io_put_req(struct io_kiocb *req);
1084 static void io_put_req_deferred(struct io_kiocb *req);
1085 static void io_dismantle_req(struct io_kiocb *req);
1086 static void io_queue_linked_timeout(struct io_kiocb *req);
1087 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1088 struct io_uring_rsrc_update2 *up,
1090 static void io_clean_op(struct io_kiocb *req);
1091 static struct file *io_file_get(struct io_ring_ctx *ctx,
1092 struct io_kiocb *req, int fd, bool fixed);
1093 static void __io_queue_sqe(struct io_kiocb *req);
1094 static void io_rsrc_put_work(struct work_struct *work);
1096 static void io_req_task_queue(struct io_kiocb *req);
1097 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1098 static int io_req_prep_async(struct io_kiocb *req);
1100 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1101 unsigned int issue_flags, u32 slot_index);
1102 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
1104 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
1106 static struct kmem_cache *req_cachep;
1108 static const struct file_operations io_uring_fops;
1110 struct sock *io_uring_get_socket(struct file *file)
1112 #if defined(CONFIG_UNIX)
1113 if (file->f_op == &io_uring_fops) {
1114 struct io_ring_ctx *ctx = file->private_data;
1116 return ctx->ring_sock->sk;
1121 EXPORT_SYMBOL(io_uring_get_socket);
1123 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1126 mutex_lock(&ctx->uring_lock);
1131 #define io_for_each_link(pos, head) \
1132 for (pos = (head); pos; pos = pos->link)
1135 * Shamelessly stolen from the mm implementation of page reference checking,
1136 * see commit f958d7b528b1 for details.
1138 #define req_ref_zero_or_close_to_overflow(req) \
1139 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1141 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1143 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1144 return atomic_inc_not_zero(&req->refs);
1147 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1149 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1152 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1153 return atomic_dec_and_test(&req->refs);
1156 static inline void req_ref_put(struct io_kiocb *req)
1158 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1159 WARN_ON_ONCE(req_ref_put_and_test(req));
1162 static inline void req_ref_get(struct io_kiocb *req)
1164 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1165 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1166 atomic_inc(&req->refs);
1169 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1171 if (!(req->flags & REQ_F_REFCOUNT)) {
1172 req->flags |= REQ_F_REFCOUNT;
1173 atomic_set(&req->refs, nr);
1177 static inline void io_req_set_refcount(struct io_kiocb *req)
1179 __io_req_set_refcount(req, 1);
1182 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1184 struct io_ring_ctx *ctx = req->ctx;
1186 if (!req->fixed_rsrc_refs) {
1187 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1188 percpu_ref_get(req->fixed_rsrc_refs);
1192 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1194 bool got = percpu_ref_tryget(ref);
1196 /* already at zero, wait for ->release() */
1198 wait_for_completion(compl);
1199 percpu_ref_resurrect(ref);
1201 percpu_ref_put(ref);
1204 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1207 struct io_kiocb *req;
1209 if (task && head->task != task)
1214 io_for_each_link(req, head) {
1215 if (req->flags & REQ_F_INFLIGHT)
1221 static inline void req_set_fail(struct io_kiocb *req)
1223 req->flags |= REQ_F_FAIL;
1226 static inline void req_fail_link_node(struct io_kiocb *req, int res)
1232 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1234 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1236 complete(&ctx->ref_comp);
1239 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1241 return !req->timeout.off;
1244 static void io_fallback_req_func(struct work_struct *work)
1246 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1247 fallback_work.work);
1248 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1249 struct io_kiocb *req, *tmp;
1250 bool locked = false;
1252 percpu_ref_get(&ctx->refs);
1253 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1254 req->io_task_work.func(req, &locked);
1257 if (ctx->submit_state.compl_nr)
1258 io_submit_flush_completions(ctx);
1259 mutex_unlock(&ctx->uring_lock);
1261 percpu_ref_put(&ctx->refs);
1265 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1267 struct io_ring_ctx *ctx;
1270 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1275 * Use 5 bits less than the max cq entries, that should give us around
1276 * 32 entries per hash list if totally full and uniformly spread.
1278 hash_bits = ilog2(p->cq_entries);
1282 ctx->cancel_hash_bits = hash_bits;
1283 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1285 if (!ctx->cancel_hash)
1287 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1289 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1290 if (!ctx->dummy_ubuf)
1292 /* set invalid range, so io_import_fixed() fails meeting it */
1293 ctx->dummy_ubuf->ubuf = -1UL;
1295 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1296 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1299 ctx->flags = p->flags;
1300 init_waitqueue_head(&ctx->sqo_sq_wait);
1301 INIT_LIST_HEAD(&ctx->sqd_list);
1302 init_waitqueue_head(&ctx->poll_wait);
1303 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1304 init_completion(&ctx->ref_comp);
1305 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1306 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1307 mutex_init(&ctx->uring_lock);
1308 init_waitqueue_head(&ctx->cq_wait);
1309 spin_lock_init(&ctx->completion_lock);
1310 spin_lock_init(&ctx->timeout_lock);
1311 INIT_LIST_HEAD(&ctx->iopoll_list);
1312 INIT_LIST_HEAD(&ctx->defer_list);
1313 INIT_LIST_HEAD(&ctx->timeout_list);
1314 INIT_LIST_HEAD(&ctx->ltimeout_list);
1315 spin_lock_init(&ctx->rsrc_ref_lock);
1316 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1317 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1318 init_llist_head(&ctx->rsrc_put_llist);
1319 INIT_LIST_HEAD(&ctx->tctx_list);
1320 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1321 INIT_LIST_HEAD(&ctx->locked_free_list);
1322 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1325 kfree(ctx->dummy_ubuf);
1326 kfree(ctx->cancel_hash);
1331 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1333 struct io_rings *r = ctx->rings;
1335 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1339 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1341 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1342 struct io_ring_ctx *ctx = req->ctx;
1344 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1350 #define FFS_ASYNC_READ 0x1UL
1351 #define FFS_ASYNC_WRITE 0x2UL
1353 #define FFS_ISREG 0x4UL
1355 #define FFS_ISREG 0x0UL
1357 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1359 static inline bool io_req_ffs_set(struct io_kiocb *req)
1361 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1364 static void io_req_track_inflight(struct io_kiocb *req)
1366 if (!(req->flags & REQ_F_INFLIGHT)) {
1367 req->flags |= REQ_F_INFLIGHT;
1368 atomic_inc(¤t->io_uring->inflight_tracked);
1372 static inline void io_unprep_linked_timeout(struct io_kiocb *req)
1374 req->flags &= ~REQ_F_LINK_TIMEOUT;
1377 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1379 if (WARN_ON_ONCE(!req->link))
1382 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1383 req->flags |= REQ_F_LINK_TIMEOUT;
1385 /* linked timeouts should have two refs once prep'ed */
1386 io_req_set_refcount(req);
1387 __io_req_set_refcount(req->link, 2);
1391 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1393 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1395 return __io_prep_linked_timeout(req);
1398 static void io_prep_async_work(struct io_kiocb *req)
1400 const struct io_op_def *def = &io_op_defs[req->opcode];
1401 struct io_ring_ctx *ctx = req->ctx;
1403 if (!(req->flags & REQ_F_CREDS)) {
1404 req->flags |= REQ_F_CREDS;
1405 req->creds = get_current_cred();
1408 req->work.list.next = NULL;
1409 req->work.flags = 0;
1410 if (req->flags & REQ_F_FORCE_ASYNC)
1411 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1413 if (req->flags & REQ_F_ISREG) {
1414 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1415 io_wq_hash_work(&req->work, file_inode(req->file));
1416 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1417 if (def->unbound_nonreg_file)
1418 req->work.flags |= IO_WQ_WORK_UNBOUND;
1421 switch (req->opcode) {
1422 case IORING_OP_SPLICE:
1424 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1425 req->work.flags |= IO_WQ_WORK_UNBOUND;
1430 static void io_prep_async_link(struct io_kiocb *req)
1432 struct io_kiocb *cur;
1434 if (req->flags & REQ_F_LINK_TIMEOUT) {
1435 struct io_ring_ctx *ctx = req->ctx;
1437 spin_lock(&ctx->completion_lock);
1438 io_for_each_link(cur, req)
1439 io_prep_async_work(cur);
1440 spin_unlock(&ctx->completion_lock);
1442 io_for_each_link(cur, req)
1443 io_prep_async_work(cur);
1447 static void io_queue_async_work(struct io_kiocb *req, bool *locked)
1449 struct io_ring_ctx *ctx = req->ctx;
1450 struct io_kiocb *link = io_prep_linked_timeout(req);
1451 struct io_uring_task *tctx = req->task->io_uring;
1453 /* must not take the lock, NULL it as a precaution */
1457 BUG_ON(!tctx->io_wq);
1459 /* init ->work of the whole link before punting */
1460 io_prep_async_link(req);
1463 * Not expected to happen, but if we do have a bug where this _can_
1464 * happen, catch it here and ensure the request is marked as
1465 * canceled. That will make io-wq go through the usual work cancel
1466 * procedure rather than attempt to run this request (or create a new
1469 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1470 req->work.flags |= IO_WQ_WORK_CANCEL;
1472 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1473 &req->work, req->flags);
1474 io_wq_enqueue(tctx->io_wq, &req->work);
1476 io_queue_linked_timeout(link);
1479 static void io_kill_timeout(struct io_kiocb *req, int status)
1480 __must_hold(&req->ctx->completion_lock)
1481 __must_hold(&req->ctx->timeout_lock)
1483 struct io_timeout_data *io = req->async_data;
1485 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1488 atomic_set(&req->ctx->cq_timeouts,
1489 atomic_read(&req->ctx->cq_timeouts) + 1);
1490 list_del_init(&req->timeout.list);
1491 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1492 io_put_req_deferred(req);
1496 static void io_queue_deferred(struct io_ring_ctx *ctx)
1498 while (!list_empty(&ctx->defer_list)) {
1499 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1500 struct io_defer_entry, list);
1502 if (req_need_defer(de->req, de->seq))
1504 list_del_init(&de->list);
1505 io_req_task_queue(de->req);
1510 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1511 __must_hold(&ctx->completion_lock)
1513 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1515 spin_lock_irq(&ctx->timeout_lock);
1516 while (!list_empty(&ctx->timeout_list)) {
1517 u32 events_needed, events_got;
1518 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1519 struct io_kiocb, timeout.list);
1521 if (io_is_timeout_noseq(req))
1525 * Since seq can easily wrap around over time, subtract
1526 * the last seq at which timeouts were flushed before comparing.
1527 * Assuming not more than 2^31-1 events have happened since,
1528 * these subtractions won't have wrapped, so we can check if
1529 * target is in [last_seq, current_seq] by comparing the two.
1531 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1532 events_got = seq - ctx->cq_last_tm_flush;
1533 if (events_got < events_needed)
1536 list_del_init(&req->timeout.list);
1537 io_kill_timeout(req, 0);
1539 ctx->cq_last_tm_flush = seq;
1540 spin_unlock_irq(&ctx->timeout_lock);
1543 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1545 if (ctx->off_timeout_used)
1546 io_flush_timeouts(ctx);
1547 if (ctx->drain_active)
1548 io_queue_deferred(ctx);
1551 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1553 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1554 __io_commit_cqring_flush(ctx);
1555 /* order cqe stores with ring update */
1556 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1559 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1561 struct io_rings *r = ctx->rings;
1563 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1566 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1568 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1571 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1573 struct io_rings *rings = ctx->rings;
1574 unsigned tail, mask = ctx->cq_entries - 1;
1577 * writes to the cq entry need to come after reading head; the
1578 * control dependency is enough as we're using WRITE_ONCE to
1581 if (__io_cqring_events(ctx) == ctx->cq_entries)
1584 tail = ctx->cached_cq_tail++;
1585 return &rings->cqes[tail & mask];
1588 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1590 if (likely(!ctx->cq_ev_fd))
1592 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1594 return !ctx->eventfd_async || io_wq_current_is_worker();
1598 * This should only get called when at least one event has been posted.
1599 * Some applications rely on the eventfd notification count only changing
1600 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1601 * 1:1 relationship between how many times this function is called (and
1602 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1604 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1607 * wake_up_all() may seem excessive, but io_wake_function() and
1608 * io_should_wake() handle the termination of the loop and only
1609 * wake as many waiters as we need to.
1611 if (wq_has_sleeper(&ctx->cq_wait))
1612 wake_up_all(&ctx->cq_wait);
1613 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1614 wake_up(&ctx->sq_data->wait);
1615 if (io_should_trigger_evfd(ctx))
1616 eventfd_signal(ctx->cq_ev_fd, 1);
1617 if (waitqueue_active(&ctx->poll_wait)) {
1618 wake_up_interruptible(&ctx->poll_wait);
1619 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1623 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1625 /* see waitqueue_active() comment */
1628 if (ctx->flags & IORING_SETUP_SQPOLL) {
1629 if (waitqueue_active(&ctx->cq_wait))
1630 wake_up_all(&ctx->cq_wait);
1632 if (io_should_trigger_evfd(ctx))
1633 eventfd_signal(ctx->cq_ev_fd, 1);
1634 if (waitqueue_active(&ctx->poll_wait)) {
1635 wake_up_interruptible(&ctx->poll_wait);
1636 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1640 /* Returns true if there are no backlogged entries after the flush */
1641 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1643 bool all_flushed, posted;
1645 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1649 spin_lock(&ctx->completion_lock);
1650 while (!list_empty(&ctx->cq_overflow_list)) {
1651 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1652 struct io_overflow_cqe *ocqe;
1656 ocqe = list_first_entry(&ctx->cq_overflow_list,
1657 struct io_overflow_cqe, list);
1659 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1661 io_account_cq_overflow(ctx);
1664 list_del(&ocqe->list);
1668 all_flushed = list_empty(&ctx->cq_overflow_list);
1670 clear_bit(0, &ctx->check_cq_overflow);
1671 WRITE_ONCE(ctx->rings->sq_flags,
1672 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1676 io_commit_cqring(ctx);
1677 spin_unlock(&ctx->completion_lock);
1679 io_cqring_ev_posted(ctx);
1683 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1687 if (test_bit(0, &ctx->check_cq_overflow)) {
1688 /* iopoll syncs against uring_lock, not completion_lock */
1689 if (ctx->flags & IORING_SETUP_IOPOLL)
1690 mutex_lock(&ctx->uring_lock);
1691 ret = __io_cqring_overflow_flush(ctx, false);
1692 if (ctx->flags & IORING_SETUP_IOPOLL)
1693 mutex_unlock(&ctx->uring_lock);
1699 /* must to be called somewhat shortly after putting a request */
1700 static inline void io_put_task(struct task_struct *task, int nr)
1702 struct io_uring_task *tctx = task->io_uring;
1704 if (likely(task == current)) {
1705 tctx->cached_refs += nr;
1707 percpu_counter_sub(&tctx->inflight, nr);
1708 if (unlikely(atomic_read(&tctx->in_idle)))
1709 wake_up(&tctx->wait);
1710 put_task_struct_many(task, nr);
1714 static void io_task_refs_refill(struct io_uring_task *tctx)
1716 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1718 percpu_counter_add(&tctx->inflight, refill);
1719 refcount_add(refill, ¤t->usage);
1720 tctx->cached_refs += refill;
1723 static inline void io_get_task_refs(int nr)
1725 struct io_uring_task *tctx = current->io_uring;
1727 tctx->cached_refs -= nr;
1728 if (unlikely(tctx->cached_refs < 0))
1729 io_task_refs_refill(tctx);
1732 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1733 long res, unsigned int cflags)
1735 struct io_overflow_cqe *ocqe;
1737 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1740 * If we're in ring overflow flush mode, or in task cancel mode,
1741 * or cannot allocate an overflow entry, then we need to drop it
1744 io_account_cq_overflow(ctx);
1747 if (list_empty(&ctx->cq_overflow_list)) {
1748 set_bit(0, &ctx->check_cq_overflow);
1749 WRITE_ONCE(ctx->rings->sq_flags,
1750 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1753 ocqe->cqe.user_data = user_data;
1754 ocqe->cqe.res = res;
1755 ocqe->cqe.flags = cflags;
1756 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1760 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1761 long res, unsigned int cflags)
1763 struct io_uring_cqe *cqe;
1765 trace_io_uring_complete(ctx, user_data, res, cflags);
1768 * If we can't get a cq entry, userspace overflowed the
1769 * submission (by quite a lot). Increment the overflow count in
1772 cqe = io_get_cqe(ctx);
1774 WRITE_ONCE(cqe->user_data, user_data);
1775 WRITE_ONCE(cqe->res, res);
1776 WRITE_ONCE(cqe->flags, cflags);
1779 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1782 /* not as hot to bloat with inlining */
1783 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1784 long res, unsigned int cflags)
1786 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1789 static void io_req_complete_post(struct io_kiocb *req, long res,
1790 unsigned int cflags)
1792 struct io_ring_ctx *ctx = req->ctx;
1794 spin_lock(&ctx->completion_lock);
1795 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1797 * If we're the last reference to this request, add to our locked
1800 if (req_ref_put_and_test(req)) {
1801 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1802 if (req->flags & IO_DISARM_MASK)
1803 io_disarm_next(req);
1805 io_req_task_queue(req->link);
1809 io_dismantle_req(req);
1810 io_put_task(req->task, 1);
1811 list_add(&req->inflight_entry, &ctx->locked_free_list);
1812 ctx->locked_free_nr++;
1814 if (!percpu_ref_tryget(&ctx->refs))
1817 io_commit_cqring(ctx);
1818 spin_unlock(&ctx->completion_lock);
1821 io_cqring_ev_posted(ctx);
1822 percpu_ref_put(&ctx->refs);
1826 static inline bool io_req_needs_clean(struct io_kiocb *req)
1828 return req->flags & IO_REQ_CLEAN_FLAGS;
1831 static void io_req_complete_state(struct io_kiocb *req, long res,
1832 unsigned int cflags)
1834 if (io_req_needs_clean(req))
1837 req->compl.cflags = cflags;
1838 req->flags |= REQ_F_COMPLETE_INLINE;
1841 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1842 long res, unsigned cflags)
1844 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1845 io_req_complete_state(req, res, cflags);
1847 io_req_complete_post(req, res, cflags);
1850 static inline void io_req_complete(struct io_kiocb *req, long res)
1852 __io_req_complete(req, 0, res, 0);
1855 static void io_req_complete_failed(struct io_kiocb *req, long res)
1858 io_req_complete_post(req, res, 0);
1861 static void io_req_complete_fail_submit(struct io_kiocb *req)
1864 * We don't submit, fail them all, for that replace hardlinks with
1865 * normal links. Extra REQ_F_LINK is tolerated.
1867 req->flags &= ~REQ_F_HARDLINK;
1868 req->flags |= REQ_F_LINK;
1869 io_req_complete_failed(req, req->result);
1873 * Don't initialise the fields below on every allocation, but do that in
1874 * advance and keep them valid across allocations.
1876 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1880 req->async_data = NULL;
1881 /* not necessary, but safer to zero */
1885 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1886 struct io_submit_state *state)
1888 spin_lock(&ctx->completion_lock);
1889 list_splice_init(&ctx->locked_free_list, &state->free_list);
1890 ctx->locked_free_nr = 0;
1891 spin_unlock(&ctx->completion_lock);
1894 /* Returns true IFF there are requests in the cache */
1895 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1897 struct io_submit_state *state = &ctx->submit_state;
1901 * If we have more than a batch's worth of requests in our IRQ side
1902 * locked cache, grab the lock and move them over to our submission
1905 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1906 io_flush_cached_locked_reqs(ctx, state);
1908 nr = state->free_reqs;
1909 while (!list_empty(&state->free_list)) {
1910 struct io_kiocb *req = list_first_entry(&state->free_list,
1911 struct io_kiocb, inflight_entry);
1913 list_del(&req->inflight_entry);
1914 state->reqs[nr++] = req;
1915 if (nr == ARRAY_SIZE(state->reqs))
1919 state->free_reqs = nr;
1924 * A request might get retired back into the request caches even before opcode
1925 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1926 * Because of that, io_alloc_req() should be called only under ->uring_lock
1927 * and with extra caution to not get a request that is still worked on.
1929 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1930 __must_hold(&ctx->uring_lock)
1932 struct io_submit_state *state = &ctx->submit_state;
1933 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1936 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1938 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1941 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1945 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1946 * retry single alloc to be on the safe side.
1948 if (unlikely(ret <= 0)) {
1949 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1950 if (!state->reqs[0])
1955 for (i = 0; i < ret; i++)
1956 io_preinit_req(state->reqs[i], ctx);
1957 state->free_reqs = ret;
1960 return state->reqs[state->free_reqs];
1963 static inline void io_put_file(struct file *file)
1969 static void io_dismantle_req(struct io_kiocb *req)
1971 unsigned int flags = req->flags;
1973 if (io_req_needs_clean(req))
1975 if (!(flags & REQ_F_FIXED_FILE))
1976 io_put_file(req->file);
1977 if (req->fixed_rsrc_refs)
1978 percpu_ref_put(req->fixed_rsrc_refs);
1979 if (req->async_data) {
1980 kfree(req->async_data);
1981 req->async_data = NULL;
1985 static void __io_free_req(struct io_kiocb *req)
1987 struct io_ring_ctx *ctx = req->ctx;
1989 io_dismantle_req(req);
1990 io_put_task(req->task, 1);
1992 spin_lock(&ctx->completion_lock);
1993 list_add(&req->inflight_entry, &ctx->locked_free_list);
1994 ctx->locked_free_nr++;
1995 spin_unlock(&ctx->completion_lock);
1997 percpu_ref_put(&ctx->refs);
2000 static inline void io_remove_next_linked(struct io_kiocb *req)
2002 struct io_kiocb *nxt = req->link;
2004 req->link = nxt->link;
2008 static bool io_kill_linked_timeout(struct io_kiocb *req)
2009 __must_hold(&req->ctx->completion_lock)
2010 __must_hold(&req->ctx->timeout_lock)
2012 struct io_kiocb *link = req->link;
2014 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2015 struct io_timeout_data *io = link->async_data;
2017 io_remove_next_linked(req);
2018 link->timeout.head = NULL;
2019 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2020 list_del(&link->timeout.list);
2021 io_cqring_fill_event(link->ctx, link->user_data,
2023 io_put_req_deferred(link);
2030 static void io_fail_links(struct io_kiocb *req)
2031 __must_hold(&req->ctx->completion_lock)
2033 struct io_kiocb *nxt, *link = req->link;
2037 long res = -ECANCELED;
2039 if (link->flags & REQ_F_FAIL)
2045 trace_io_uring_fail_link(req, link);
2046 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2047 io_put_req_deferred(link);
2052 static bool io_disarm_next(struct io_kiocb *req)
2053 __must_hold(&req->ctx->completion_lock)
2055 bool posted = false;
2057 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2058 struct io_kiocb *link = req->link;
2060 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2061 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2062 io_remove_next_linked(req);
2063 io_cqring_fill_event(link->ctx, link->user_data,
2065 io_put_req_deferred(link);
2068 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2069 struct io_ring_ctx *ctx = req->ctx;
2071 spin_lock_irq(&ctx->timeout_lock);
2072 posted = io_kill_linked_timeout(req);
2073 spin_unlock_irq(&ctx->timeout_lock);
2075 if (unlikely((req->flags & REQ_F_FAIL) &&
2076 !(req->flags & REQ_F_HARDLINK))) {
2077 posted |= (req->link != NULL);
2083 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2085 struct io_kiocb *nxt;
2088 * If LINK is set, we have dependent requests in this chain. If we
2089 * didn't fail this request, queue the first one up, moving any other
2090 * dependencies to the next request. In case of failure, fail the rest
2093 if (req->flags & IO_DISARM_MASK) {
2094 struct io_ring_ctx *ctx = req->ctx;
2097 spin_lock(&ctx->completion_lock);
2098 posted = io_disarm_next(req);
2100 io_commit_cqring(req->ctx);
2101 spin_unlock(&ctx->completion_lock);
2103 io_cqring_ev_posted(ctx);
2110 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2112 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2114 return __io_req_find_next(req);
2117 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2122 if (ctx->submit_state.compl_nr)
2123 io_submit_flush_completions(ctx);
2124 mutex_unlock(&ctx->uring_lock);
2127 percpu_ref_put(&ctx->refs);
2130 static void tctx_task_work(struct callback_head *cb)
2132 bool locked = false;
2133 struct io_ring_ctx *ctx = NULL;
2134 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2138 struct io_wq_work_node *node;
2140 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2141 io_submit_flush_completions(ctx);
2143 spin_lock_irq(&tctx->task_lock);
2144 node = tctx->task_list.first;
2145 INIT_WQ_LIST(&tctx->task_list);
2147 tctx->task_running = false;
2148 spin_unlock_irq(&tctx->task_lock);
2153 struct io_wq_work_node *next = node->next;
2154 struct io_kiocb *req = container_of(node, struct io_kiocb,
2157 if (req->ctx != ctx) {
2158 ctx_flush_and_put(ctx, &locked);
2160 /* if not contended, grab and improve batching */
2161 locked = mutex_trylock(&ctx->uring_lock);
2162 percpu_ref_get(&ctx->refs);
2164 req->io_task_work.func(req, &locked);
2171 ctx_flush_and_put(ctx, &locked);
2174 static void io_req_task_work_add(struct io_kiocb *req)
2176 struct task_struct *tsk = req->task;
2177 struct io_uring_task *tctx = tsk->io_uring;
2178 enum task_work_notify_mode notify;
2179 struct io_wq_work_node *node;
2180 unsigned long flags;
2183 WARN_ON_ONCE(!tctx);
2185 spin_lock_irqsave(&tctx->task_lock, flags);
2186 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2187 running = tctx->task_running;
2189 tctx->task_running = true;
2190 spin_unlock_irqrestore(&tctx->task_lock, flags);
2192 /* task_work already pending, we're done */
2197 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2198 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2199 * processing task_work. There's no reliable way to tell if TWA_RESUME
2202 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2203 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2204 wake_up_process(tsk);
2208 spin_lock_irqsave(&tctx->task_lock, flags);
2209 tctx->task_running = false;
2210 node = tctx->task_list.first;
2211 INIT_WQ_LIST(&tctx->task_list);
2212 spin_unlock_irqrestore(&tctx->task_lock, flags);
2215 req = container_of(node, struct io_kiocb, io_task_work.node);
2217 if (llist_add(&req->io_task_work.fallback_node,
2218 &req->ctx->fallback_llist))
2219 schedule_delayed_work(&req->ctx->fallback_work, 1);
2223 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2225 struct io_ring_ctx *ctx = req->ctx;
2227 /* not needed for normal modes, but SQPOLL depends on it */
2228 io_tw_lock(ctx, locked);
2229 io_req_complete_failed(req, req->result);
2232 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2234 struct io_ring_ctx *ctx = req->ctx;
2236 io_tw_lock(ctx, locked);
2237 /* req->task == current here, checking PF_EXITING is safe */
2238 if (likely(!(req->task->flags & PF_EXITING)))
2239 __io_queue_sqe(req);
2241 io_req_complete_failed(req, -EFAULT);
2244 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2247 req->io_task_work.func = io_req_task_cancel;
2248 io_req_task_work_add(req);
2251 static void io_req_task_queue(struct io_kiocb *req)
2253 req->io_task_work.func = io_req_task_submit;
2254 io_req_task_work_add(req);
2257 static void io_req_task_queue_reissue(struct io_kiocb *req)
2259 req->io_task_work.func = io_queue_async_work;
2260 io_req_task_work_add(req);
2263 static inline void io_queue_next(struct io_kiocb *req)
2265 struct io_kiocb *nxt = io_req_find_next(req);
2268 io_req_task_queue(nxt);
2271 static void io_free_req(struct io_kiocb *req)
2277 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2283 struct task_struct *task;
2288 static inline void io_init_req_batch(struct req_batch *rb)
2295 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2296 struct req_batch *rb)
2299 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2301 io_put_task(rb->task, rb->task_refs);
2304 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2305 struct io_submit_state *state)
2308 io_dismantle_req(req);
2310 if (req->task != rb->task) {
2312 io_put_task(rb->task, rb->task_refs);
2313 rb->task = req->task;
2319 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2320 state->reqs[state->free_reqs++] = req;
2322 list_add(&req->inflight_entry, &state->free_list);
2325 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2326 __must_hold(&ctx->uring_lock)
2328 struct io_submit_state *state = &ctx->submit_state;
2329 int i, nr = state->compl_nr;
2330 struct req_batch rb;
2332 spin_lock(&ctx->completion_lock);
2333 for (i = 0; i < nr; i++) {
2334 struct io_kiocb *req = state->compl_reqs[i];
2336 __io_cqring_fill_event(ctx, req->user_data, req->result,
2339 io_commit_cqring(ctx);
2340 spin_unlock(&ctx->completion_lock);
2341 io_cqring_ev_posted(ctx);
2343 io_init_req_batch(&rb);
2344 for (i = 0; i < nr; i++) {
2345 struct io_kiocb *req = state->compl_reqs[i];
2347 if (req_ref_put_and_test(req))
2348 io_req_free_batch(&rb, req, &ctx->submit_state);
2351 io_req_free_batch_finish(ctx, &rb);
2352 state->compl_nr = 0;
2356 * Drop reference to request, return next in chain (if there is one) if this
2357 * was the last reference to this request.
2359 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2361 struct io_kiocb *nxt = NULL;
2363 if (req_ref_put_and_test(req)) {
2364 nxt = io_req_find_next(req);
2370 static inline void io_put_req(struct io_kiocb *req)
2372 if (req_ref_put_and_test(req))
2376 static inline void io_put_req_deferred(struct io_kiocb *req)
2378 if (req_ref_put_and_test(req)) {
2379 req->io_task_work.func = io_free_req_work;
2380 io_req_task_work_add(req);
2384 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2386 /* See comment at the top of this file */
2388 return __io_cqring_events(ctx);
2391 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2393 struct io_rings *rings = ctx->rings;
2395 /* make sure SQ entry isn't read before tail */
2396 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2399 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2401 unsigned int cflags;
2403 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2404 cflags |= IORING_CQE_F_BUFFER;
2405 req->flags &= ~REQ_F_BUFFER_SELECTED;
2410 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2412 struct io_buffer *kbuf;
2414 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2416 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2417 return io_put_kbuf(req, kbuf);
2420 static inline bool io_run_task_work(void)
2422 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2423 __set_current_state(TASK_RUNNING);
2424 tracehook_notify_signal();
2432 * Find and free completed poll iocbs
2434 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2435 struct list_head *done)
2437 struct req_batch rb;
2438 struct io_kiocb *req;
2440 /* order with ->result store in io_complete_rw_iopoll() */
2443 io_init_req_batch(&rb);
2444 while (!list_empty(done)) {
2445 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2446 list_del(&req->inflight_entry);
2448 __io_cqring_fill_event(ctx, req->user_data, req->result,
2449 io_put_rw_kbuf(req));
2452 if (req_ref_put_and_test(req))
2453 io_req_free_batch(&rb, req, &ctx->submit_state);
2456 io_commit_cqring(ctx);
2457 io_cqring_ev_posted_iopoll(ctx);
2458 io_req_free_batch_finish(ctx, &rb);
2461 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2464 struct io_kiocb *req, *tmp;
2469 * Only spin for completions if we don't have multiple devices hanging
2470 * off our complete list, and we're under the requested amount.
2472 spin = !ctx->poll_multi_queue && *nr_events < min;
2474 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2475 struct kiocb *kiocb = &req->rw.kiocb;
2479 * Move completed and retryable entries to our local lists.
2480 * If we find a request that requires polling, break out
2481 * and complete those lists first, if we have entries there.
2483 if (READ_ONCE(req->iopoll_completed)) {
2484 list_move_tail(&req->inflight_entry, &done);
2487 if (!list_empty(&done))
2490 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2491 if (unlikely(ret < 0))
2496 /* iopoll may have completed current req */
2497 if (READ_ONCE(req->iopoll_completed))
2498 list_move_tail(&req->inflight_entry, &done);
2501 if (!list_empty(&done))
2502 io_iopoll_complete(ctx, nr_events, &done);
2508 * We can't just wait for polled events to come to us, we have to actively
2509 * find and complete them.
2511 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2513 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2516 mutex_lock(&ctx->uring_lock);
2517 while (!list_empty(&ctx->iopoll_list)) {
2518 unsigned int nr_events = 0;
2520 io_do_iopoll(ctx, &nr_events, 0);
2522 /* let it sleep and repeat later if can't complete a request */
2526 * Ensure we allow local-to-the-cpu processing to take place,
2527 * in this case we need to ensure that we reap all events.
2528 * Also let task_work, etc. to progress by releasing the mutex
2530 if (need_resched()) {
2531 mutex_unlock(&ctx->uring_lock);
2533 mutex_lock(&ctx->uring_lock);
2536 mutex_unlock(&ctx->uring_lock);
2539 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2541 unsigned int nr_events = 0;
2545 * We disallow the app entering submit/complete with polling, but we
2546 * still need to lock the ring to prevent racing with polled issue
2547 * that got punted to a workqueue.
2549 mutex_lock(&ctx->uring_lock);
2551 * Don't enter poll loop if we already have events pending.
2552 * If we do, we can potentially be spinning for commands that
2553 * already triggered a CQE (eg in error).
2555 if (test_bit(0, &ctx->check_cq_overflow))
2556 __io_cqring_overflow_flush(ctx, false);
2557 if (io_cqring_events(ctx))
2561 * If a submit got punted to a workqueue, we can have the
2562 * application entering polling for a command before it gets
2563 * issued. That app will hold the uring_lock for the duration
2564 * of the poll right here, so we need to take a breather every
2565 * now and then to ensure that the issue has a chance to add
2566 * the poll to the issued list. Otherwise we can spin here
2567 * forever, while the workqueue is stuck trying to acquire the
2570 if (list_empty(&ctx->iopoll_list)) {
2571 u32 tail = ctx->cached_cq_tail;
2573 mutex_unlock(&ctx->uring_lock);
2575 mutex_lock(&ctx->uring_lock);
2577 /* some requests don't go through iopoll_list */
2578 if (tail != ctx->cached_cq_tail ||
2579 list_empty(&ctx->iopoll_list))
2582 ret = io_do_iopoll(ctx, &nr_events, min);
2583 } while (!ret && nr_events < min && !need_resched());
2585 mutex_unlock(&ctx->uring_lock);
2589 static void kiocb_end_write(struct io_kiocb *req)
2592 * Tell lockdep we inherited freeze protection from submission
2595 if (req->flags & REQ_F_ISREG) {
2596 struct super_block *sb = file_inode(req->file)->i_sb;
2598 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2604 static bool io_resubmit_prep(struct io_kiocb *req)
2606 struct io_async_rw *rw = req->async_data;
2609 return !io_req_prep_async(req);
2610 iov_iter_restore(&rw->iter, &rw->iter_state);
2614 static bool io_rw_should_reissue(struct io_kiocb *req)
2616 umode_t mode = file_inode(req->file)->i_mode;
2617 struct io_ring_ctx *ctx = req->ctx;
2619 if (!S_ISBLK(mode) && !S_ISREG(mode))
2621 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2622 !(ctx->flags & IORING_SETUP_IOPOLL)))
2625 * If ref is dying, we might be running poll reap from the exit work.
2626 * Don't attempt to reissue from that path, just let it fail with
2629 if (percpu_ref_is_dying(&ctx->refs))
2632 * Play it safe and assume not safe to re-import and reissue if we're
2633 * not in the original thread group (or in task context).
2635 if (!same_thread_group(req->task, current) || !in_task())
2640 static bool io_resubmit_prep(struct io_kiocb *req)
2644 static bool io_rw_should_reissue(struct io_kiocb *req)
2650 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2652 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2653 kiocb_end_write(req);
2654 if (res != req->result) {
2655 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2656 io_rw_should_reissue(req)) {
2657 req->flags |= REQ_F_REISSUE;
2666 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2668 unsigned int cflags = io_put_rw_kbuf(req);
2669 long res = req->result;
2672 struct io_ring_ctx *ctx = req->ctx;
2673 struct io_submit_state *state = &ctx->submit_state;
2675 io_req_complete_state(req, res, cflags);
2676 state->compl_reqs[state->compl_nr++] = req;
2677 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2678 io_submit_flush_completions(ctx);
2680 io_req_complete_post(req, res, cflags);
2684 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2685 unsigned int issue_flags)
2687 if (__io_complete_rw_common(req, res))
2689 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2692 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2694 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2696 if (__io_complete_rw_common(req, res))
2699 req->io_task_work.func = io_req_task_complete;
2700 io_req_task_work_add(req);
2703 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2705 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2707 if (kiocb->ki_flags & IOCB_WRITE)
2708 kiocb_end_write(req);
2709 if (unlikely(res != req->result)) {
2710 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2711 req->flags |= REQ_F_REISSUE;
2716 WRITE_ONCE(req->result, res);
2717 /* order with io_iopoll_complete() checking ->result */
2719 WRITE_ONCE(req->iopoll_completed, 1);
2723 * After the iocb has been issued, it's safe to be found on the poll list.
2724 * Adding the kiocb to the list AFTER submission ensures that we don't
2725 * find it from a io_do_iopoll() thread before the issuer is done
2726 * accessing the kiocb cookie.
2728 static void io_iopoll_req_issued(struct io_kiocb *req)
2730 struct io_ring_ctx *ctx = req->ctx;
2731 const bool in_async = io_wq_current_is_worker();
2733 /* workqueue context doesn't hold uring_lock, grab it now */
2734 if (unlikely(in_async))
2735 mutex_lock(&ctx->uring_lock);
2738 * Track whether we have multiple files in our lists. This will impact
2739 * how we do polling eventually, not spinning if we're on potentially
2740 * different devices.
2742 if (list_empty(&ctx->iopoll_list)) {
2743 ctx->poll_multi_queue = false;
2744 } else if (!ctx->poll_multi_queue) {
2745 struct io_kiocb *list_req;
2746 unsigned int queue_num0, queue_num1;
2748 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2751 if (list_req->file != req->file) {
2752 ctx->poll_multi_queue = true;
2754 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2755 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2756 if (queue_num0 != queue_num1)
2757 ctx->poll_multi_queue = true;
2762 * For fast devices, IO may have already completed. If it has, add
2763 * it to the front so we find it first.
2765 if (READ_ONCE(req->iopoll_completed))
2766 list_add(&req->inflight_entry, &ctx->iopoll_list);
2768 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2770 if (unlikely(in_async)) {
2772 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2773 * in sq thread task context or in io worker task context. If
2774 * current task context is sq thread, we don't need to check
2775 * whether should wake up sq thread.
2777 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2778 wq_has_sleeper(&ctx->sq_data->wait))
2779 wake_up(&ctx->sq_data->wait);
2781 mutex_unlock(&ctx->uring_lock);
2785 static bool io_bdev_nowait(struct block_device *bdev)
2787 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2791 * If we tracked the file through the SCM inflight mechanism, we could support
2792 * any file. For now, just ensure that anything potentially problematic is done
2795 static bool __io_file_supports_nowait(struct file *file, int rw)
2797 umode_t mode = file_inode(file)->i_mode;
2799 if (S_ISBLK(mode)) {
2800 if (IS_ENABLED(CONFIG_BLOCK) &&
2801 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2807 if (S_ISREG(mode)) {
2808 if (IS_ENABLED(CONFIG_BLOCK) &&
2809 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2810 file->f_op != &io_uring_fops)
2815 /* any ->read/write should understand O_NONBLOCK */
2816 if (file->f_flags & O_NONBLOCK)
2819 if (!(file->f_mode & FMODE_NOWAIT))
2823 return file->f_op->read_iter != NULL;
2825 return file->f_op->write_iter != NULL;
2828 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2830 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2832 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2835 return __io_file_supports_nowait(req->file, rw);
2838 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2841 struct io_ring_ctx *ctx = req->ctx;
2842 struct kiocb *kiocb = &req->rw.kiocb;
2843 struct file *file = req->file;
2847 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2848 req->flags |= REQ_F_ISREG;
2850 kiocb->ki_pos = READ_ONCE(sqe->off);
2851 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2852 req->flags |= REQ_F_CUR_POS;
2853 kiocb->ki_pos = file->f_pos;
2855 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2856 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2857 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2862 * If the file is marked O_NONBLOCK, still allow retry for it if it
2863 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2864 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2866 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2867 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2868 req->flags |= REQ_F_NOWAIT;
2870 ioprio = READ_ONCE(sqe->ioprio);
2872 ret = ioprio_check_cap(ioprio);
2876 kiocb->ki_ioprio = ioprio;
2878 kiocb->ki_ioprio = get_current_ioprio();
2880 if (ctx->flags & IORING_SETUP_IOPOLL) {
2881 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2882 !kiocb->ki_filp->f_op->iopoll)
2885 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2886 kiocb->ki_complete = io_complete_rw_iopoll;
2887 req->iopoll_completed = 0;
2889 if (kiocb->ki_flags & IOCB_HIPRI)
2891 kiocb->ki_complete = io_complete_rw;
2894 if (req->opcode == IORING_OP_READ_FIXED ||
2895 req->opcode == IORING_OP_WRITE_FIXED) {
2897 io_req_set_rsrc_node(req);
2900 req->rw.addr = READ_ONCE(sqe->addr);
2901 req->rw.len = READ_ONCE(sqe->len);
2902 req->buf_index = READ_ONCE(sqe->buf_index);
2906 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2912 case -ERESTARTNOINTR:
2913 case -ERESTARTNOHAND:
2914 case -ERESTART_RESTARTBLOCK:
2916 * We can't just restart the syscall, since previously
2917 * submitted sqes may already be in progress. Just fail this
2923 kiocb->ki_complete(kiocb, ret, 0);
2927 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2928 unsigned int issue_flags)
2930 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2931 struct io_async_rw *io = req->async_data;
2933 /* add previously done IO, if any */
2934 if (io && io->bytes_done > 0) {
2936 ret = io->bytes_done;
2938 ret += io->bytes_done;
2941 if (req->flags & REQ_F_CUR_POS)
2942 req->file->f_pos = kiocb->ki_pos;
2943 if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
2944 __io_complete_rw(req, ret, 0, issue_flags);
2946 io_rw_done(kiocb, ret);
2948 if (req->flags & REQ_F_REISSUE) {
2949 req->flags &= ~REQ_F_REISSUE;
2950 if (io_resubmit_prep(req)) {
2951 io_req_task_queue_reissue(req);
2953 unsigned int cflags = io_put_rw_kbuf(req);
2954 struct io_ring_ctx *ctx = req->ctx;
2957 if (issue_flags & IO_URING_F_NONBLOCK) {
2958 mutex_lock(&ctx->uring_lock);
2959 __io_req_complete(req, issue_flags, ret, cflags);
2960 mutex_unlock(&ctx->uring_lock);
2962 __io_req_complete(req, issue_flags, ret, cflags);
2968 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2969 struct io_mapped_ubuf *imu)
2971 size_t len = req->rw.len;
2972 u64 buf_end, buf_addr = req->rw.addr;
2975 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2977 /* not inside the mapped region */
2978 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2982 * May not be a start of buffer, set size appropriately
2983 * and advance us to the beginning.
2985 offset = buf_addr - imu->ubuf;
2986 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2990 * Don't use iov_iter_advance() here, as it's really slow for
2991 * using the latter parts of a big fixed buffer - it iterates
2992 * over each segment manually. We can cheat a bit here, because
2995 * 1) it's a BVEC iter, we set it up
2996 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2997 * first and last bvec
2999 * So just find our index, and adjust the iterator afterwards.
3000 * If the offset is within the first bvec (or the whole first
3001 * bvec, just use iov_iter_advance(). This makes it easier
3002 * since we can just skip the first segment, which may not
3003 * be PAGE_SIZE aligned.
3005 const struct bio_vec *bvec = imu->bvec;
3007 if (offset <= bvec->bv_len) {
3008 iov_iter_advance(iter, offset);
3010 unsigned long seg_skip;
3012 /* skip first vec */
3013 offset -= bvec->bv_len;
3014 seg_skip = 1 + (offset >> PAGE_SHIFT);
3016 iter->bvec = bvec + seg_skip;
3017 iter->nr_segs -= seg_skip;
3018 iter->count -= bvec->bv_len + offset;
3019 iter->iov_offset = offset & ~PAGE_MASK;
3026 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3028 struct io_ring_ctx *ctx = req->ctx;
3029 struct io_mapped_ubuf *imu = req->imu;
3030 u16 index, buf_index = req->buf_index;
3033 if (unlikely(buf_index >= ctx->nr_user_bufs))
3035 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3036 imu = READ_ONCE(ctx->user_bufs[index]);
3039 return __io_import_fixed(req, rw, iter, imu);
3042 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3045 mutex_unlock(&ctx->uring_lock);
3048 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3051 * "Normal" inline submissions always hold the uring_lock, since we
3052 * grab it from the system call. Same is true for the SQPOLL offload.
3053 * The only exception is when we've detached the request and issue it
3054 * from an async worker thread, grab the lock for that case.
3057 mutex_lock(&ctx->uring_lock);
3060 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3061 int bgid, struct io_buffer *kbuf,
3064 struct io_buffer *head;
3066 if (req->flags & REQ_F_BUFFER_SELECTED)
3069 io_ring_submit_lock(req->ctx, needs_lock);
3071 lockdep_assert_held(&req->ctx->uring_lock);
3073 head = xa_load(&req->ctx->io_buffers, bgid);
3075 if (!list_empty(&head->list)) {
3076 kbuf = list_last_entry(&head->list, struct io_buffer,
3078 list_del(&kbuf->list);
3081 xa_erase(&req->ctx->io_buffers, bgid);
3083 if (*len > kbuf->len)
3086 kbuf = ERR_PTR(-ENOBUFS);
3089 io_ring_submit_unlock(req->ctx, needs_lock);
3094 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3097 struct io_buffer *kbuf;
3100 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3101 bgid = req->buf_index;
3102 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3105 req->rw.addr = (u64) (unsigned long) kbuf;
3106 req->flags |= REQ_F_BUFFER_SELECTED;
3107 return u64_to_user_ptr(kbuf->addr);
3110 #ifdef CONFIG_COMPAT
3111 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3114 struct compat_iovec __user *uiov;
3115 compat_ssize_t clen;
3119 uiov = u64_to_user_ptr(req->rw.addr);
3120 if (!access_ok(uiov, sizeof(*uiov)))
3122 if (__get_user(clen, &uiov->iov_len))
3128 buf = io_rw_buffer_select(req, &len, needs_lock);
3130 return PTR_ERR(buf);
3131 iov[0].iov_base = buf;
3132 iov[0].iov_len = (compat_size_t) len;
3137 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3140 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3144 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3147 len = iov[0].iov_len;
3150 buf = io_rw_buffer_select(req, &len, needs_lock);
3152 return PTR_ERR(buf);
3153 iov[0].iov_base = buf;
3154 iov[0].iov_len = len;
3158 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3161 if (req->flags & REQ_F_BUFFER_SELECTED) {
3162 struct io_buffer *kbuf;
3164 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3165 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3166 iov[0].iov_len = kbuf->len;
3169 if (req->rw.len != 1)
3172 #ifdef CONFIG_COMPAT
3173 if (req->ctx->compat)
3174 return io_compat_import(req, iov, needs_lock);
3177 return __io_iov_buffer_select(req, iov, needs_lock);
3180 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3181 struct iov_iter *iter, bool needs_lock)
3183 void __user *buf = u64_to_user_ptr(req->rw.addr);
3184 size_t sqe_len = req->rw.len;
3185 u8 opcode = req->opcode;
3188 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3190 return io_import_fixed(req, rw, iter);
3193 /* buffer index only valid with fixed read/write, or buffer select */
3194 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3197 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3198 if (req->flags & REQ_F_BUFFER_SELECT) {
3199 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3201 return PTR_ERR(buf);
3202 req->rw.len = sqe_len;
3205 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3210 if (req->flags & REQ_F_BUFFER_SELECT) {
3211 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3213 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3218 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3222 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3224 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3228 * For files that don't have ->read_iter() and ->write_iter(), handle them
3229 * by looping over ->read() or ->write() manually.
3231 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3233 struct kiocb *kiocb = &req->rw.kiocb;
3234 struct file *file = req->file;
3238 * Don't support polled IO through this interface, and we can't
3239 * support non-blocking either. For the latter, this just causes
3240 * the kiocb to be handled from an async context.
3242 if (kiocb->ki_flags & IOCB_HIPRI)
3244 if (kiocb->ki_flags & IOCB_NOWAIT)
3247 while (iov_iter_count(iter)) {
3251 if (!iov_iter_is_bvec(iter)) {
3252 iovec = iov_iter_iovec(iter);
3254 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3255 iovec.iov_len = req->rw.len;
3259 nr = file->f_op->read(file, iovec.iov_base,
3260 iovec.iov_len, io_kiocb_ppos(kiocb));
3262 nr = file->f_op->write(file, iovec.iov_base,
3263 iovec.iov_len, io_kiocb_ppos(kiocb));
3271 if (!iov_iter_is_bvec(iter)) {
3272 iov_iter_advance(iter, nr);
3278 if (nr != iovec.iov_len)
3285 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3286 const struct iovec *fast_iov, struct iov_iter *iter)
3288 struct io_async_rw *rw = req->async_data;
3290 memcpy(&rw->iter, iter, sizeof(*iter));
3291 rw->free_iovec = iovec;
3293 /* can only be fixed buffers, no need to do anything */
3294 if (iov_iter_is_bvec(iter))
3297 unsigned iov_off = 0;
3299 rw->iter.iov = rw->fast_iov;
3300 if (iter->iov != fast_iov) {
3301 iov_off = iter->iov - fast_iov;
3302 rw->iter.iov += iov_off;
3304 if (rw->fast_iov != fast_iov)
3305 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3306 sizeof(struct iovec) * iter->nr_segs);
3308 req->flags |= REQ_F_NEED_CLEANUP;
3312 static inline int io_alloc_async_data(struct io_kiocb *req)
3314 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3315 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3316 return req->async_data == NULL;
3319 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3320 const struct iovec *fast_iov,
3321 struct iov_iter *iter, bool force)
3323 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3325 if (!req->async_data) {
3326 struct io_async_rw *iorw;
3328 if (io_alloc_async_data(req)) {
3333 io_req_map_rw(req, iovec, fast_iov, iter);
3334 iorw = req->async_data;
3335 /* we've copied and mapped the iter, ensure state is saved */
3336 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3341 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3343 struct io_async_rw *iorw = req->async_data;
3344 struct iovec *iov = iorw->fast_iov;
3347 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3348 if (unlikely(ret < 0))
3351 iorw->bytes_done = 0;
3352 iorw->free_iovec = iov;
3354 req->flags |= REQ_F_NEED_CLEANUP;
3355 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3359 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3361 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3363 return io_prep_rw(req, sqe, READ);
3367 * This is our waitqueue callback handler, registered through lock_page_async()
3368 * when we initially tried to do the IO with the iocb armed our waitqueue.
3369 * This gets called when the page is unlocked, and we generally expect that to
3370 * happen when the page IO is completed and the page is now uptodate. This will
3371 * queue a task_work based retry of the operation, attempting to copy the data
3372 * again. If the latter fails because the page was NOT uptodate, then we will
3373 * do a thread based blocking retry of the operation. That's the unexpected
3376 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3377 int sync, void *arg)
3379 struct wait_page_queue *wpq;
3380 struct io_kiocb *req = wait->private;
3381 struct wait_page_key *key = arg;
3383 wpq = container_of(wait, struct wait_page_queue, wait);
3385 if (!wake_page_match(wpq, key))
3388 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3389 list_del_init(&wait->entry);
3390 io_req_task_queue(req);
3395 * This controls whether a given IO request should be armed for async page
3396 * based retry. If we return false here, the request is handed to the async
3397 * worker threads for retry. If we're doing buffered reads on a regular file,
3398 * we prepare a private wait_page_queue entry and retry the operation. This
3399 * will either succeed because the page is now uptodate and unlocked, or it
3400 * will register a callback when the page is unlocked at IO completion. Through
3401 * that callback, io_uring uses task_work to setup a retry of the operation.
3402 * That retry will attempt the buffered read again. The retry will generally
3403 * succeed, or in rare cases where it fails, we then fall back to using the
3404 * async worker threads for a blocking retry.
3406 static bool io_rw_should_retry(struct io_kiocb *req)
3408 struct io_async_rw *rw = req->async_data;
3409 struct wait_page_queue *wait = &rw->wpq;
3410 struct kiocb *kiocb = &req->rw.kiocb;
3412 /* never retry for NOWAIT, we just complete with -EAGAIN */
3413 if (req->flags & REQ_F_NOWAIT)
3416 /* Only for buffered IO */
3417 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3421 * just use poll if we can, and don't attempt if the fs doesn't
3422 * support callback based unlocks
3424 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3427 wait->wait.func = io_async_buf_func;
3428 wait->wait.private = req;
3429 wait->wait.flags = 0;
3430 INIT_LIST_HEAD(&wait->wait.entry);
3431 kiocb->ki_flags |= IOCB_WAITQ;
3432 kiocb->ki_flags &= ~IOCB_NOWAIT;
3433 kiocb->ki_waitq = wait;
3437 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3439 if (req->file->f_op->read_iter)
3440 return call_read_iter(req->file, &req->rw.kiocb, iter);
3441 else if (req->file->f_op->read)
3442 return loop_rw_iter(READ, req, iter);
3447 static bool need_read_all(struct io_kiocb *req)
3449 return req->flags & REQ_F_ISREG ||
3450 S_ISBLK(file_inode(req->file)->i_mode);
3453 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3455 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3456 struct kiocb *kiocb = &req->rw.kiocb;
3457 struct iov_iter __iter, *iter = &__iter;
3458 struct io_async_rw *rw = req->async_data;
3459 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3460 struct iov_iter_state __state, *state;
3465 state = &rw->iter_state;
3467 * We come here from an earlier attempt, restore our state to
3468 * match in case it doesn't. It's cheap enough that we don't
3469 * need to make this conditional.
3471 iov_iter_restore(iter, state);
3474 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3478 iov_iter_save_state(iter, state);
3480 req->result = iov_iter_count(iter);
3482 /* Ensure we clear previously set non-block flag */
3483 if (!force_nonblock)
3484 kiocb->ki_flags &= ~IOCB_NOWAIT;
3486 kiocb->ki_flags |= IOCB_NOWAIT;
3488 /* If the file doesn't support async, just async punt */
3489 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3490 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3491 return ret ?: -EAGAIN;
3494 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3495 if (unlikely(ret)) {
3500 ret = io_iter_do_read(req, iter);
3502 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3503 req->flags &= ~REQ_F_REISSUE;
3504 /* IOPOLL retry should happen for io-wq threads */
3505 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3507 /* no retry on NONBLOCK nor RWF_NOWAIT */
3508 if (req->flags & REQ_F_NOWAIT)
3511 } else if (ret == -EIOCBQUEUED) {
3513 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3514 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3515 /* read all, failed, already did sync or don't want to retry */
3520 * Don't depend on the iter state matching what was consumed, or being
3521 * untouched in case of error. Restore it and we'll advance it
3522 * manually if we need to.
3524 iov_iter_restore(iter, state);
3526 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3531 rw = req->async_data;
3533 * Now use our persistent iterator and state, if we aren't already.
3534 * We've restored and mapped the iter to match.
3536 if (iter != &rw->iter) {
3538 state = &rw->iter_state;
3543 * We end up here because of a partial read, either from
3544 * above or inside this loop. Advance the iter by the bytes
3545 * that were consumed.
3547 iov_iter_advance(iter, ret);
3548 if (!iov_iter_count(iter))
3550 rw->bytes_done += ret;
3551 iov_iter_save_state(iter, state);
3553 /* if we can retry, do so with the callbacks armed */
3554 if (!io_rw_should_retry(req)) {
3555 kiocb->ki_flags &= ~IOCB_WAITQ;
3560 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3561 * we get -EIOCBQUEUED, then we'll get a notification when the
3562 * desired page gets unlocked. We can also get a partial read
3563 * here, and if we do, then just retry at the new offset.
3565 ret = io_iter_do_read(req, iter);
3566 if (ret == -EIOCBQUEUED)
3568 /* we got some bytes, but not all. retry. */
3569 kiocb->ki_flags &= ~IOCB_WAITQ;
3570 iov_iter_restore(iter, state);
3573 kiocb_done(kiocb, ret, issue_flags);
3575 /* it's faster to check here then delegate to kfree */
3581 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3583 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3585 return io_prep_rw(req, sqe, WRITE);
3588 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3590 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3591 struct kiocb *kiocb = &req->rw.kiocb;
3592 struct iov_iter __iter, *iter = &__iter;
3593 struct io_async_rw *rw = req->async_data;
3594 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3595 struct iov_iter_state __state, *state;
3600 state = &rw->iter_state;
3601 iov_iter_restore(iter, state);
3604 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3608 iov_iter_save_state(iter, state);
3610 req->result = iov_iter_count(iter);
3612 /* Ensure we clear previously set non-block flag */
3613 if (!force_nonblock)
3614 kiocb->ki_flags &= ~IOCB_NOWAIT;
3616 kiocb->ki_flags |= IOCB_NOWAIT;
3618 /* If the file doesn't support async, just async punt */
3619 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3622 /* file path doesn't support NOWAIT for non-direct_IO */
3623 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3624 (req->flags & REQ_F_ISREG))
3627 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3632 * Open-code file_start_write here to grab freeze protection,
3633 * which will be released by another thread in
3634 * io_complete_rw(). Fool lockdep by telling it the lock got
3635 * released so that it doesn't complain about the held lock when
3636 * we return to userspace.
3638 if (req->flags & REQ_F_ISREG) {
3639 sb_start_write(file_inode(req->file)->i_sb);
3640 __sb_writers_release(file_inode(req->file)->i_sb,
3643 kiocb->ki_flags |= IOCB_WRITE;
3645 if (req->file->f_op->write_iter)
3646 ret2 = call_write_iter(req->file, kiocb, iter);
3647 else if (req->file->f_op->write)
3648 ret2 = loop_rw_iter(WRITE, req, iter);
3652 if (req->flags & REQ_F_REISSUE) {
3653 req->flags &= ~REQ_F_REISSUE;
3658 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3659 * retry them without IOCB_NOWAIT.
3661 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3663 /* no retry on NONBLOCK nor RWF_NOWAIT */
3664 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3666 if (!force_nonblock || ret2 != -EAGAIN) {
3667 /* IOPOLL retry should happen for io-wq threads */
3668 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3671 kiocb_done(kiocb, ret2, issue_flags);
3674 iov_iter_restore(iter, state);
3675 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3676 return ret ?: -EAGAIN;
3679 /* it's reportedly faster than delegating the null check to kfree() */
3685 static int io_renameat_prep(struct io_kiocb *req,
3686 const struct io_uring_sqe *sqe)
3688 struct io_rename *ren = &req->rename;
3689 const char __user *oldf, *newf;
3691 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3693 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3695 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3698 ren->old_dfd = READ_ONCE(sqe->fd);
3699 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3700 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3701 ren->new_dfd = READ_ONCE(sqe->len);
3702 ren->flags = READ_ONCE(sqe->rename_flags);
3704 ren->oldpath = getname(oldf);
3705 if (IS_ERR(ren->oldpath))
3706 return PTR_ERR(ren->oldpath);
3708 ren->newpath = getname(newf);
3709 if (IS_ERR(ren->newpath)) {
3710 putname(ren->oldpath);
3711 return PTR_ERR(ren->newpath);
3714 req->flags |= REQ_F_NEED_CLEANUP;
3718 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3720 struct io_rename *ren = &req->rename;
3723 if (issue_flags & IO_URING_F_NONBLOCK)
3726 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3727 ren->newpath, ren->flags);
3729 req->flags &= ~REQ_F_NEED_CLEANUP;
3732 io_req_complete(req, ret);
3736 static int io_unlinkat_prep(struct io_kiocb *req,
3737 const struct io_uring_sqe *sqe)
3739 struct io_unlink *un = &req->unlink;
3740 const char __user *fname;
3742 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3744 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3747 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3750 un->dfd = READ_ONCE(sqe->fd);
3752 un->flags = READ_ONCE(sqe->unlink_flags);
3753 if (un->flags & ~AT_REMOVEDIR)
3756 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3757 un->filename = getname(fname);
3758 if (IS_ERR(un->filename))
3759 return PTR_ERR(un->filename);
3761 req->flags |= REQ_F_NEED_CLEANUP;
3765 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3767 struct io_unlink *un = &req->unlink;
3770 if (issue_flags & IO_URING_F_NONBLOCK)
3773 if (un->flags & AT_REMOVEDIR)
3774 ret = do_rmdir(un->dfd, un->filename);
3776 ret = do_unlinkat(un->dfd, un->filename);
3778 req->flags &= ~REQ_F_NEED_CLEANUP;
3781 io_req_complete(req, ret);
3785 static int io_mkdirat_prep(struct io_kiocb *req,
3786 const struct io_uring_sqe *sqe)
3788 struct io_mkdir *mkd = &req->mkdir;
3789 const char __user *fname;
3791 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3793 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3796 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3799 mkd->dfd = READ_ONCE(sqe->fd);
3800 mkd->mode = READ_ONCE(sqe->len);
3802 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3803 mkd->filename = getname(fname);
3804 if (IS_ERR(mkd->filename))
3805 return PTR_ERR(mkd->filename);
3807 req->flags |= REQ_F_NEED_CLEANUP;
3811 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3813 struct io_mkdir *mkd = &req->mkdir;
3816 if (issue_flags & IO_URING_F_NONBLOCK)
3819 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3821 req->flags &= ~REQ_F_NEED_CLEANUP;
3824 io_req_complete(req, ret);
3828 static int io_symlinkat_prep(struct io_kiocb *req,
3829 const struct io_uring_sqe *sqe)
3831 struct io_symlink *sl = &req->symlink;
3832 const char __user *oldpath, *newpath;
3834 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3836 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3839 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3842 sl->new_dfd = READ_ONCE(sqe->fd);
3843 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3844 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3846 sl->oldpath = getname(oldpath);
3847 if (IS_ERR(sl->oldpath))
3848 return PTR_ERR(sl->oldpath);
3850 sl->newpath = getname(newpath);
3851 if (IS_ERR(sl->newpath)) {
3852 putname(sl->oldpath);
3853 return PTR_ERR(sl->newpath);
3856 req->flags |= REQ_F_NEED_CLEANUP;
3860 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3862 struct io_symlink *sl = &req->symlink;
3865 if (issue_flags & IO_URING_F_NONBLOCK)
3868 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3870 req->flags &= ~REQ_F_NEED_CLEANUP;
3873 io_req_complete(req, ret);
3877 static int io_linkat_prep(struct io_kiocb *req,
3878 const struct io_uring_sqe *sqe)
3880 struct io_hardlink *lnk = &req->hardlink;
3881 const char __user *oldf, *newf;
3883 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3885 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3887 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3890 lnk->old_dfd = READ_ONCE(sqe->fd);
3891 lnk->new_dfd = READ_ONCE(sqe->len);
3892 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3893 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3894 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3896 lnk->oldpath = getname(oldf);
3897 if (IS_ERR(lnk->oldpath))
3898 return PTR_ERR(lnk->oldpath);
3900 lnk->newpath = getname(newf);
3901 if (IS_ERR(lnk->newpath)) {
3902 putname(lnk->oldpath);
3903 return PTR_ERR(lnk->newpath);
3906 req->flags |= REQ_F_NEED_CLEANUP;
3910 static int io_linkat(struct io_kiocb *req, int issue_flags)
3912 struct io_hardlink *lnk = &req->hardlink;
3915 if (issue_flags & IO_URING_F_NONBLOCK)
3918 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3919 lnk->newpath, lnk->flags);
3921 req->flags &= ~REQ_F_NEED_CLEANUP;
3924 io_req_complete(req, ret);
3928 static int io_shutdown_prep(struct io_kiocb *req,
3929 const struct io_uring_sqe *sqe)
3931 #if defined(CONFIG_NET)
3932 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3934 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3935 sqe->buf_index || sqe->splice_fd_in))
3938 req->shutdown.how = READ_ONCE(sqe->len);
3945 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3947 #if defined(CONFIG_NET)
3948 struct socket *sock;
3951 if (issue_flags & IO_URING_F_NONBLOCK)
3954 sock = sock_from_file(req->file);
3955 if (unlikely(!sock))
3958 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3961 io_req_complete(req, ret);
3968 static int __io_splice_prep(struct io_kiocb *req,
3969 const struct io_uring_sqe *sqe)
3971 struct io_splice *sp = &req->splice;
3972 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3974 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3978 sp->len = READ_ONCE(sqe->len);
3979 sp->flags = READ_ONCE(sqe->splice_flags);
3981 if (unlikely(sp->flags & ~valid_flags))
3984 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3985 (sp->flags & SPLICE_F_FD_IN_FIXED));
3988 req->flags |= REQ_F_NEED_CLEANUP;
3992 static int io_tee_prep(struct io_kiocb *req,
3993 const struct io_uring_sqe *sqe)
3995 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3997 return __io_splice_prep(req, sqe);
4000 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
4002 struct io_splice *sp = &req->splice;
4003 struct file *in = sp->file_in;
4004 struct file *out = sp->file_out;
4005 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4008 if (issue_flags & IO_URING_F_NONBLOCK)
4011 ret = do_tee(in, out, sp->len, flags);
4013 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4015 req->flags &= ~REQ_F_NEED_CLEANUP;
4019 io_req_complete(req, ret);
4023 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4025 struct io_splice *sp = &req->splice;
4027 sp->off_in = READ_ONCE(sqe->splice_off_in);
4028 sp->off_out = READ_ONCE(sqe->off);
4029 return __io_splice_prep(req, sqe);
4032 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4034 struct io_splice *sp = &req->splice;
4035 struct file *in = sp->file_in;
4036 struct file *out = sp->file_out;
4037 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4038 loff_t *poff_in, *poff_out;
4041 if (issue_flags & IO_URING_F_NONBLOCK)
4044 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4045 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4048 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4050 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4052 req->flags &= ~REQ_F_NEED_CLEANUP;
4056 io_req_complete(req, ret);
4061 * IORING_OP_NOP just posts a completion event, nothing else.
4063 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4065 struct io_ring_ctx *ctx = req->ctx;
4067 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4070 __io_req_complete(req, issue_flags, 0, 0);
4074 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4076 struct io_ring_ctx *ctx = req->ctx;
4081 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4083 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4087 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4088 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4091 req->sync.off = READ_ONCE(sqe->off);
4092 req->sync.len = READ_ONCE(sqe->len);
4096 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4098 loff_t end = req->sync.off + req->sync.len;
4101 /* fsync always requires a blocking context */
4102 if (issue_flags & IO_URING_F_NONBLOCK)
4105 ret = vfs_fsync_range(req->file, req->sync.off,
4106 end > 0 ? end : LLONG_MAX,
4107 req->sync.flags & IORING_FSYNC_DATASYNC);
4110 io_req_complete(req, ret);
4114 static int io_fallocate_prep(struct io_kiocb *req,
4115 const struct io_uring_sqe *sqe)
4117 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4120 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4123 req->sync.off = READ_ONCE(sqe->off);
4124 req->sync.len = READ_ONCE(sqe->addr);
4125 req->sync.mode = READ_ONCE(sqe->len);
4129 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4133 /* fallocate always requiring blocking context */
4134 if (issue_flags & IO_URING_F_NONBLOCK)
4136 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4140 io_req_complete(req, ret);
4144 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4146 const char __user *fname;
4149 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4151 if (unlikely(sqe->ioprio || sqe->buf_index))
4153 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4156 /* open.how should be already initialised */
4157 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4158 req->open.how.flags |= O_LARGEFILE;
4160 req->open.dfd = READ_ONCE(sqe->fd);
4161 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4162 req->open.filename = getname(fname);
4163 if (IS_ERR(req->open.filename)) {
4164 ret = PTR_ERR(req->open.filename);
4165 req->open.filename = NULL;
4169 req->open.file_slot = READ_ONCE(sqe->file_index);
4170 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4173 req->open.nofile = rlimit(RLIMIT_NOFILE);
4174 req->flags |= REQ_F_NEED_CLEANUP;
4178 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4180 u64 mode = READ_ONCE(sqe->len);
4181 u64 flags = READ_ONCE(sqe->open_flags);
4183 req->open.how = build_open_how(flags, mode);
4184 return __io_openat_prep(req, sqe);
4187 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4189 struct open_how __user *how;
4193 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4194 len = READ_ONCE(sqe->len);
4195 if (len < OPEN_HOW_SIZE_VER0)
4198 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4203 return __io_openat_prep(req, sqe);
4206 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4208 struct open_flags op;
4210 bool resolve_nonblock, nonblock_set;
4211 bool fixed = !!req->open.file_slot;
4214 ret = build_open_flags(&req->open.how, &op);
4217 nonblock_set = op.open_flag & O_NONBLOCK;
4218 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4219 if (issue_flags & IO_URING_F_NONBLOCK) {
4221 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4222 * it'll always -EAGAIN
4224 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4226 op.lookup_flags |= LOOKUP_CACHED;
4227 op.open_flag |= O_NONBLOCK;
4231 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4236 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4239 * We could hang on to this 'fd' on retrying, but seems like
4240 * marginal gain for something that is now known to be a slower
4241 * path. So just put it, and we'll get a new one when we retry.
4246 ret = PTR_ERR(file);
4247 /* only retry if RESOLVE_CACHED wasn't already set by application */
4248 if (ret == -EAGAIN &&
4249 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4254 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4255 file->f_flags &= ~O_NONBLOCK;
4256 fsnotify_open(file);
4259 fd_install(ret, file);
4261 ret = io_install_fixed_file(req, file, issue_flags,
4262 req->open.file_slot - 1);
4264 putname(req->open.filename);
4265 req->flags &= ~REQ_F_NEED_CLEANUP;
4268 __io_req_complete(req, issue_flags, ret, 0);
4272 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4274 return io_openat2(req, issue_flags);
4277 static int io_remove_buffers_prep(struct io_kiocb *req,
4278 const struct io_uring_sqe *sqe)
4280 struct io_provide_buf *p = &req->pbuf;
4283 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4287 tmp = READ_ONCE(sqe->fd);
4288 if (!tmp || tmp > USHRT_MAX)
4291 memset(p, 0, sizeof(*p));
4293 p->bgid = READ_ONCE(sqe->buf_group);
4297 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4298 int bgid, unsigned nbufs)
4302 /* shouldn't happen */
4306 /* the head kbuf is the list itself */
4307 while (!list_empty(&buf->list)) {
4308 struct io_buffer *nxt;
4310 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4311 list_del(&nxt->list);
4318 xa_erase(&ctx->io_buffers, bgid);
4323 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4325 struct io_provide_buf *p = &req->pbuf;
4326 struct io_ring_ctx *ctx = req->ctx;
4327 struct io_buffer *head;
4329 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4331 io_ring_submit_lock(ctx, !force_nonblock);
4333 lockdep_assert_held(&ctx->uring_lock);
4336 head = xa_load(&ctx->io_buffers, p->bgid);
4338 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4342 /* complete before unlock, IOPOLL may need the lock */
4343 __io_req_complete(req, issue_flags, ret, 0);
4344 io_ring_submit_unlock(ctx, !force_nonblock);
4348 static int io_provide_buffers_prep(struct io_kiocb *req,
4349 const struct io_uring_sqe *sqe)
4351 unsigned long size, tmp_check;
4352 struct io_provide_buf *p = &req->pbuf;
4355 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4358 tmp = READ_ONCE(sqe->fd);
4359 if (!tmp || tmp > USHRT_MAX)
4362 p->addr = READ_ONCE(sqe->addr);
4363 p->len = READ_ONCE(sqe->len);
4365 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4368 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4371 size = (unsigned long)p->len * p->nbufs;
4372 if (!access_ok(u64_to_user_ptr(p->addr), size))
4375 p->bgid = READ_ONCE(sqe->buf_group);
4376 tmp = READ_ONCE(sqe->off);
4377 if (tmp > USHRT_MAX)
4383 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4385 struct io_buffer *buf;
4386 u64 addr = pbuf->addr;
4387 int i, bid = pbuf->bid;
4389 for (i = 0; i < pbuf->nbufs; i++) {
4390 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4395 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4400 INIT_LIST_HEAD(&buf->list);
4403 list_add_tail(&buf->list, &(*head)->list);
4407 return i ? i : -ENOMEM;
4410 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4412 struct io_provide_buf *p = &req->pbuf;
4413 struct io_ring_ctx *ctx = req->ctx;
4414 struct io_buffer *head, *list;
4416 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4418 io_ring_submit_lock(ctx, !force_nonblock);
4420 lockdep_assert_held(&ctx->uring_lock);
4422 list = head = xa_load(&ctx->io_buffers, p->bgid);
4424 ret = io_add_buffers(p, &head);
4425 if (ret >= 0 && !list) {
4426 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4428 __io_remove_buffers(ctx, head, p->bgid, -1U);
4432 /* complete before unlock, IOPOLL may need the lock */
4433 __io_req_complete(req, issue_flags, ret, 0);
4434 io_ring_submit_unlock(ctx, !force_nonblock);
4438 static int io_epoll_ctl_prep(struct io_kiocb *req,
4439 const struct io_uring_sqe *sqe)
4441 #if defined(CONFIG_EPOLL)
4442 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4444 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4447 req->epoll.epfd = READ_ONCE(sqe->fd);
4448 req->epoll.op = READ_ONCE(sqe->len);
4449 req->epoll.fd = READ_ONCE(sqe->off);
4451 if (ep_op_has_event(req->epoll.op)) {
4452 struct epoll_event __user *ev;
4454 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4455 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4465 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4467 #if defined(CONFIG_EPOLL)
4468 struct io_epoll *ie = &req->epoll;
4470 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4472 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4473 if (force_nonblock && ret == -EAGAIN)
4478 __io_req_complete(req, issue_flags, ret, 0);
4485 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4487 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4488 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4490 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4493 req->madvise.addr = READ_ONCE(sqe->addr);
4494 req->madvise.len = READ_ONCE(sqe->len);
4495 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4502 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4504 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4505 struct io_madvise *ma = &req->madvise;
4508 if (issue_flags & IO_URING_F_NONBLOCK)
4511 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4514 io_req_complete(req, ret);
4521 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4523 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4525 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4528 req->fadvise.offset = READ_ONCE(sqe->off);
4529 req->fadvise.len = READ_ONCE(sqe->len);
4530 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4534 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4536 struct io_fadvise *fa = &req->fadvise;
4539 if (issue_flags & IO_URING_F_NONBLOCK) {
4540 switch (fa->advice) {
4541 case POSIX_FADV_NORMAL:
4542 case POSIX_FADV_RANDOM:
4543 case POSIX_FADV_SEQUENTIAL:
4550 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4553 __io_req_complete(req, issue_flags, ret, 0);
4557 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4559 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4561 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4563 if (req->flags & REQ_F_FIXED_FILE)
4566 req->statx.dfd = READ_ONCE(sqe->fd);
4567 req->statx.mask = READ_ONCE(sqe->len);
4568 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4569 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4570 req->statx.flags = READ_ONCE(sqe->statx_flags);
4575 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4577 struct io_statx *ctx = &req->statx;
4580 if (issue_flags & IO_URING_F_NONBLOCK)
4583 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4588 io_req_complete(req, ret);
4592 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4594 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4596 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4597 sqe->rw_flags || sqe->buf_index)
4599 if (req->flags & REQ_F_FIXED_FILE)
4602 req->close.fd = READ_ONCE(sqe->fd);
4603 req->close.file_slot = READ_ONCE(sqe->file_index);
4604 if (req->close.file_slot && req->close.fd)
4610 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4612 struct files_struct *files = current->files;
4613 struct io_close *close = &req->close;
4614 struct fdtable *fdt;
4615 struct file *file = NULL;
4618 if (req->close.file_slot) {
4619 ret = io_close_fixed(req, issue_flags);
4623 spin_lock(&files->file_lock);
4624 fdt = files_fdtable(files);
4625 if (close->fd >= fdt->max_fds) {
4626 spin_unlock(&files->file_lock);
4629 file = fdt->fd[close->fd];
4630 if (!file || file->f_op == &io_uring_fops) {
4631 spin_unlock(&files->file_lock);
4636 /* if the file has a flush method, be safe and punt to async */
4637 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4638 spin_unlock(&files->file_lock);
4642 ret = __close_fd_get_file(close->fd, &file);
4643 spin_unlock(&files->file_lock);
4650 /* No ->flush() or already async, safely close from here */
4651 ret = filp_close(file, current->files);
4657 __io_req_complete(req, issue_flags, ret, 0);
4661 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4663 struct io_ring_ctx *ctx = req->ctx;
4665 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4667 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4671 req->sync.off = READ_ONCE(sqe->off);
4672 req->sync.len = READ_ONCE(sqe->len);
4673 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4677 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4681 /* sync_file_range always requires a blocking context */
4682 if (issue_flags & IO_URING_F_NONBLOCK)
4685 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4689 io_req_complete(req, ret);
4693 #if defined(CONFIG_NET)
4694 static int io_setup_async_msg(struct io_kiocb *req,
4695 struct io_async_msghdr *kmsg)
4697 struct io_async_msghdr *async_msg = req->async_data;
4701 if (io_alloc_async_data(req)) {
4702 kfree(kmsg->free_iov);
4705 async_msg = req->async_data;
4706 req->flags |= REQ_F_NEED_CLEANUP;
4707 memcpy(async_msg, kmsg, sizeof(*kmsg));
4708 async_msg->msg.msg_name = &async_msg->addr;
4709 /* if were using fast_iov, set it to the new one */
4710 if (!async_msg->free_iov)
4711 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4716 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4717 struct io_async_msghdr *iomsg)
4719 iomsg->msg.msg_name = &iomsg->addr;
4720 iomsg->free_iov = iomsg->fast_iov;
4721 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4722 req->sr_msg.msg_flags, &iomsg->free_iov);
4725 static int io_sendmsg_prep_async(struct io_kiocb *req)
4729 ret = io_sendmsg_copy_hdr(req, req->async_data);
4731 req->flags |= REQ_F_NEED_CLEANUP;
4735 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4737 struct io_sr_msg *sr = &req->sr_msg;
4739 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4742 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4743 sr->len = READ_ONCE(sqe->len);
4744 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4745 if (sr->msg_flags & MSG_DONTWAIT)
4746 req->flags |= REQ_F_NOWAIT;
4748 #ifdef CONFIG_COMPAT
4749 if (req->ctx->compat)
4750 sr->msg_flags |= MSG_CMSG_COMPAT;
4755 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4757 struct io_async_msghdr iomsg, *kmsg;
4758 struct socket *sock;
4763 sock = sock_from_file(req->file);
4764 if (unlikely(!sock))
4767 kmsg = req->async_data;
4769 ret = io_sendmsg_copy_hdr(req, &iomsg);
4775 flags = req->sr_msg.msg_flags;
4776 if (issue_flags & IO_URING_F_NONBLOCK)
4777 flags |= MSG_DONTWAIT;
4778 if (flags & MSG_WAITALL)
4779 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4781 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4782 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4783 return io_setup_async_msg(req, kmsg);
4784 if (ret == -ERESTARTSYS)
4787 /* fast path, check for non-NULL to avoid function call */
4789 kfree(kmsg->free_iov);
4790 req->flags &= ~REQ_F_NEED_CLEANUP;
4793 __io_req_complete(req, issue_flags, ret, 0);
4797 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4799 struct io_sr_msg *sr = &req->sr_msg;
4802 struct socket *sock;
4807 sock = sock_from_file(req->file);
4808 if (unlikely(!sock))
4811 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4815 msg.msg_name = NULL;
4816 msg.msg_control = NULL;
4817 msg.msg_controllen = 0;
4818 msg.msg_namelen = 0;
4820 flags = req->sr_msg.msg_flags;
4821 if (issue_flags & IO_URING_F_NONBLOCK)
4822 flags |= MSG_DONTWAIT;
4823 if (flags & MSG_WAITALL)
4824 min_ret = iov_iter_count(&msg.msg_iter);
4826 msg.msg_flags = flags;
4827 ret = sock_sendmsg(sock, &msg);
4828 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4830 if (ret == -ERESTARTSYS)
4835 __io_req_complete(req, issue_flags, ret, 0);
4839 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4840 struct io_async_msghdr *iomsg)
4842 struct io_sr_msg *sr = &req->sr_msg;
4843 struct iovec __user *uiov;
4847 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4848 &iomsg->uaddr, &uiov, &iov_len);
4852 if (req->flags & REQ_F_BUFFER_SELECT) {
4855 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4857 sr->len = iomsg->fast_iov[0].iov_len;
4858 iomsg->free_iov = NULL;
4860 iomsg->free_iov = iomsg->fast_iov;
4861 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4862 &iomsg->free_iov, &iomsg->msg.msg_iter,
4871 #ifdef CONFIG_COMPAT
4872 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4873 struct io_async_msghdr *iomsg)
4875 struct io_sr_msg *sr = &req->sr_msg;
4876 struct compat_iovec __user *uiov;
4881 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4886 uiov = compat_ptr(ptr);
4887 if (req->flags & REQ_F_BUFFER_SELECT) {
4888 compat_ssize_t clen;
4892 if (!access_ok(uiov, sizeof(*uiov)))
4894 if (__get_user(clen, &uiov->iov_len))
4899 iomsg->free_iov = NULL;
4901 iomsg->free_iov = iomsg->fast_iov;
4902 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4903 UIO_FASTIOV, &iomsg->free_iov,
4904 &iomsg->msg.msg_iter, true);
4913 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4914 struct io_async_msghdr *iomsg)
4916 iomsg->msg.msg_name = &iomsg->addr;
4918 #ifdef CONFIG_COMPAT
4919 if (req->ctx->compat)
4920 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4923 return __io_recvmsg_copy_hdr(req, iomsg);
4926 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4929 struct io_sr_msg *sr = &req->sr_msg;
4930 struct io_buffer *kbuf;
4932 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4937 req->flags |= REQ_F_BUFFER_SELECTED;
4941 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4943 return io_put_kbuf(req, req->sr_msg.kbuf);
4946 static int io_recvmsg_prep_async(struct io_kiocb *req)
4950 ret = io_recvmsg_copy_hdr(req, req->async_data);
4952 req->flags |= REQ_F_NEED_CLEANUP;
4956 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4958 struct io_sr_msg *sr = &req->sr_msg;
4960 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4963 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4964 sr->len = READ_ONCE(sqe->len);
4965 sr->bgid = READ_ONCE(sqe->buf_group);
4966 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4967 if (sr->msg_flags & MSG_DONTWAIT)
4968 req->flags |= REQ_F_NOWAIT;
4970 #ifdef CONFIG_COMPAT
4971 if (req->ctx->compat)
4972 sr->msg_flags |= MSG_CMSG_COMPAT;
4977 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4979 struct io_async_msghdr iomsg, *kmsg;
4980 struct socket *sock;
4981 struct io_buffer *kbuf;
4984 int ret, cflags = 0;
4985 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4987 sock = sock_from_file(req->file);
4988 if (unlikely(!sock))
4991 kmsg = req->async_data;
4993 ret = io_recvmsg_copy_hdr(req, &iomsg);
4999 if (req->flags & REQ_F_BUFFER_SELECT) {
5000 kbuf = io_recv_buffer_select(req, !force_nonblock);
5002 return PTR_ERR(kbuf);
5003 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5004 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5005 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5006 1, req->sr_msg.len);
5009 flags = req->sr_msg.msg_flags;
5011 flags |= MSG_DONTWAIT;
5012 if (flags & MSG_WAITALL)
5013 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5015 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5016 kmsg->uaddr, flags);
5017 if (force_nonblock && ret == -EAGAIN)
5018 return io_setup_async_msg(req, kmsg);
5019 if (ret == -ERESTARTSYS)
5022 if (req->flags & REQ_F_BUFFER_SELECTED)
5023 cflags = io_put_recv_kbuf(req);
5024 /* fast path, check for non-NULL to avoid function call */
5026 kfree(kmsg->free_iov);
5027 req->flags &= ~REQ_F_NEED_CLEANUP;
5028 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5030 __io_req_complete(req, issue_flags, ret, cflags);
5034 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5036 struct io_buffer *kbuf;
5037 struct io_sr_msg *sr = &req->sr_msg;
5039 void __user *buf = sr->buf;
5040 struct socket *sock;
5044 int ret, cflags = 0;
5045 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5047 sock = sock_from_file(req->file);
5048 if (unlikely(!sock))
5051 if (req->flags & REQ_F_BUFFER_SELECT) {
5052 kbuf = io_recv_buffer_select(req, !force_nonblock);
5054 return PTR_ERR(kbuf);
5055 buf = u64_to_user_ptr(kbuf->addr);
5058 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5062 msg.msg_name = NULL;
5063 msg.msg_control = NULL;
5064 msg.msg_controllen = 0;
5065 msg.msg_namelen = 0;
5066 msg.msg_iocb = NULL;
5069 flags = req->sr_msg.msg_flags;
5071 flags |= MSG_DONTWAIT;
5072 if (flags & MSG_WAITALL)
5073 min_ret = iov_iter_count(&msg.msg_iter);
5075 ret = sock_recvmsg(sock, &msg, flags);
5076 if (force_nonblock && ret == -EAGAIN)
5078 if (ret == -ERESTARTSYS)
5081 if (req->flags & REQ_F_BUFFER_SELECTED)
5082 cflags = io_put_recv_kbuf(req);
5083 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5085 __io_req_complete(req, issue_flags, ret, cflags);
5089 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5091 struct io_accept *accept = &req->accept;
5093 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5095 if (sqe->ioprio || sqe->len || sqe->buf_index)
5098 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5099 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5100 accept->flags = READ_ONCE(sqe->accept_flags);
5101 accept->nofile = rlimit(RLIMIT_NOFILE);
5103 accept->file_slot = READ_ONCE(sqe->file_index);
5104 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5105 (accept->flags & SOCK_CLOEXEC)))
5107 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5109 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5110 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5114 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5116 struct io_accept *accept = &req->accept;
5117 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5118 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5119 bool fixed = !!accept->file_slot;
5123 if (req->file->f_flags & O_NONBLOCK)
5124 req->flags |= REQ_F_NOWAIT;
5127 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5128 if (unlikely(fd < 0))
5131 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5136 ret = PTR_ERR(file);
5137 if (ret == -EAGAIN && force_nonblock)
5139 if (ret == -ERESTARTSYS)
5142 } else if (!fixed) {
5143 fd_install(fd, file);
5146 ret = io_install_fixed_file(req, file, issue_flags,
5147 accept->file_slot - 1);
5149 __io_req_complete(req, issue_flags, ret, 0);
5153 static int io_connect_prep_async(struct io_kiocb *req)
5155 struct io_async_connect *io = req->async_data;
5156 struct io_connect *conn = &req->connect;
5158 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5161 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5163 struct io_connect *conn = &req->connect;
5165 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5167 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5171 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5172 conn->addr_len = READ_ONCE(sqe->addr2);
5176 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5178 struct io_async_connect __io, *io;
5179 unsigned file_flags;
5181 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5183 if (req->async_data) {
5184 io = req->async_data;
5186 ret = move_addr_to_kernel(req->connect.addr,
5187 req->connect.addr_len,
5194 file_flags = force_nonblock ? O_NONBLOCK : 0;
5196 ret = __sys_connect_file(req->file, &io->address,
5197 req->connect.addr_len, file_flags);
5198 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5199 if (req->async_data)
5201 if (io_alloc_async_data(req)) {
5205 memcpy(req->async_data, &__io, sizeof(__io));
5208 if (ret == -ERESTARTSYS)
5213 __io_req_complete(req, issue_flags, ret, 0);
5216 #else /* !CONFIG_NET */
5217 #define IO_NETOP_FN(op) \
5218 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5220 return -EOPNOTSUPP; \
5223 #define IO_NETOP_PREP(op) \
5225 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5227 return -EOPNOTSUPP; \
5230 #define IO_NETOP_PREP_ASYNC(op) \
5232 static int io_##op##_prep_async(struct io_kiocb *req) \
5234 return -EOPNOTSUPP; \
5237 IO_NETOP_PREP_ASYNC(sendmsg);
5238 IO_NETOP_PREP_ASYNC(recvmsg);
5239 IO_NETOP_PREP_ASYNC(connect);
5240 IO_NETOP_PREP(accept);
5243 #endif /* CONFIG_NET */
5245 struct io_poll_table {
5246 struct poll_table_struct pt;
5247 struct io_kiocb *req;
5252 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5253 __poll_t mask, io_req_tw_func_t func)
5255 /* for instances that support it check for an event match first: */
5256 if (mask && !(mask & poll->events))
5259 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5261 list_del_init(&poll->wait.entry);
5264 req->io_task_work.func = func;
5267 * If this fails, then the task is exiting. When a task exits, the
5268 * work gets canceled, so just cancel this request as well instead
5269 * of executing it. We can't safely execute it anyway, as we may not
5270 * have the needed state needed for it anyway.
5272 io_req_task_work_add(req);
5276 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5277 __acquires(&req->ctx->completion_lock)
5279 struct io_ring_ctx *ctx = req->ctx;
5281 /* req->task == current here, checking PF_EXITING is safe */
5282 if (unlikely(req->task->flags & PF_EXITING))
5283 WRITE_ONCE(poll->canceled, true);
5285 if (!req->result && !READ_ONCE(poll->canceled)) {
5286 struct poll_table_struct pt = { ._key = poll->events };
5288 req->result = vfs_poll(req->file, &pt) & poll->events;
5291 spin_lock(&ctx->completion_lock);
5292 if (!req->result && !READ_ONCE(poll->canceled)) {
5293 add_wait_queue(poll->head, &poll->wait);
5300 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5302 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5303 if (req->opcode == IORING_OP_POLL_ADD)
5304 return req->async_data;
5305 return req->apoll->double_poll;
5308 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5310 if (req->opcode == IORING_OP_POLL_ADD)
5312 return &req->apoll->poll;
5315 static void io_poll_remove_double(struct io_kiocb *req)
5316 __must_hold(&req->ctx->completion_lock)
5318 struct io_poll_iocb *poll = io_poll_get_double(req);
5320 lockdep_assert_held(&req->ctx->completion_lock);
5322 if (poll && poll->head) {
5323 struct wait_queue_head *head = poll->head;
5325 spin_lock_irq(&head->lock);
5326 list_del_init(&poll->wait.entry);
5327 if (poll->wait.private)
5330 spin_unlock_irq(&head->lock);
5334 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5335 __must_hold(&req->ctx->completion_lock)
5337 struct io_ring_ctx *ctx = req->ctx;
5338 unsigned flags = IORING_CQE_F_MORE;
5341 if (READ_ONCE(req->poll.canceled)) {
5343 req->poll.events |= EPOLLONESHOT;
5345 error = mangle_poll(mask);
5347 if (req->poll.events & EPOLLONESHOT)
5349 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5350 req->poll.events |= EPOLLONESHOT;
5353 if (flags & IORING_CQE_F_MORE)
5356 return !(flags & IORING_CQE_F_MORE);
5359 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5360 __must_hold(&req->ctx->completion_lock)
5364 done = __io_poll_complete(req, mask);
5365 io_commit_cqring(req->ctx);
5369 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5371 struct io_ring_ctx *ctx = req->ctx;
5372 struct io_kiocb *nxt;
5374 if (io_poll_rewait(req, &req->poll)) {
5375 spin_unlock(&ctx->completion_lock);
5379 if (req->poll.done) {
5380 spin_unlock(&ctx->completion_lock);
5383 done = __io_poll_complete(req, req->result);
5385 io_poll_remove_double(req);
5386 hash_del(&req->hash_node);
5387 req->poll.done = true;
5390 add_wait_queue(req->poll.head, &req->poll.wait);
5392 io_commit_cqring(ctx);
5393 spin_unlock(&ctx->completion_lock);
5394 io_cqring_ev_posted(ctx);
5397 nxt = io_put_req_find_next(req);
5399 io_req_task_submit(nxt, locked);
5404 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5405 int sync, void *key)
5407 struct io_kiocb *req = wait->private;
5408 struct io_poll_iocb *poll = io_poll_get_single(req);
5409 __poll_t mask = key_to_poll(key);
5410 unsigned long flags;
5412 /* for instances that support it check for an event match first: */
5413 if (mask && !(mask & poll->events))
5415 if (!(poll->events & EPOLLONESHOT))
5416 return poll->wait.func(&poll->wait, mode, sync, key);
5418 list_del_init(&wait->entry);
5423 spin_lock_irqsave(&poll->head->lock, flags);
5424 done = list_empty(&poll->wait.entry);
5426 list_del_init(&poll->wait.entry);
5427 /* make sure double remove sees this as being gone */
5428 wait->private = NULL;
5429 spin_unlock_irqrestore(&poll->head->lock, flags);
5431 /* use wait func handler, so it matches the rq type */
5432 poll->wait.func(&poll->wait, mode, sync, key);
5439 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5440 wait_queue_func_t wake_func)
5444 poll->canceled = false;
5445 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5446 /* mask in events that we always want/need */
5447 poll->events = events | IO_POLL_UNMASK;
5448 INIT_LIST_HEAD(&poll->wait.entry);
5449 init_waitqueue_func_entry(&poll->wait, wake_func);
5452 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5453 struct wait_queue_head *head,
5454 struct io_poll_iocb **poll_ptr)
5456 struct io_kiocb *req = pt->req;
5459 * The file being polled uses multiple waitqueues for poll handling
5460 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5463 if (unlikely(pt->nr_entries)) {
5464 struct io_poll_iocb *poll_one = poll;
5466 /* double add on the same waitqueue head, ignore */
5467 if (poll_one->head == head)
5469 /* already have a 2nd entry, fail a third attempt */
5471 if ((*poll_ptr)->head == head)
5473 pt->error = -EINVAL;
5477 * Can't handle multishot for double wait for now, turn it
5478 * into one-shot mode.
5480 if (!(poll_one->events & EPOLLONESHOT))
5481 poll_one->events |= EPOLLONESHOT;
5482 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5484 pt->error = -ENOMEM;
5487 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5489 poll->wait.private = req;
5496 if (poll->events & EPOLLEXCLUSIVE)
5497 add_wait_queue_exclusive(head, &poll->wait);
5499 add_wait_queue(head, &poll->wait);
5502 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5503 struct poll_table_struct *p)
5505 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5506 struct async_poll *apoll = pt->req->apoll;
5508 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5511 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5513 struct async_poll *apoll = req->apoll;
5514 struct io_ring_ctx *ctx = req->ctx;
5516 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5518 if (io_poll_rewait(req, &apoll->poll)) {
5519 spin_unlock(&ctx->completion_lock);
5523 hash_del(&req->hash_node);
5524 io_poll_remove_double(req);
5525 apoll->poll.done = true;
5526 spin_unlock(&ctx->completion_lock);
5528 if (!READ_ONCE(apoll->poll.canceled))
5529 io_req_task_submit(req, locked);
5531 io_req_complete_failed(req, -ECANCELED);
5534 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5537 struct io_kiocb *req = wait->private;
5538 struct io_poll_iocb *poll = &req->apoll->poll;
5540 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5543 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5546 static void io_poll_req_insert(struct io_kiocb *req)
5548 struct io_ring_ctx *ctx = req->ctx;
5549 struct hlist_head *list;
5551 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5552 hlist_add_head(&req->hash_node, list);
5555 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5556 struct io_poll_iocb *poll,
5557 struct io_poll_table *ipt, __poll_t mask,
5558 wait_queue_func_t wake_func)
5559 __acquires(&ctx->completion_lock)
5561 struct io_ring_ctx *ctx = req->ctx;
5562 bool cancel = false;
5564 INIT_HLIST_NODE(&req->hash_node);
5565 io_init_poll_iocb(poll, mask, wake_func);
5566 poll->file = req->file;
5567 poll->wait.private = req;
5569 ipt->pt._key = mask;
5572 ipt->nr_entries = 0;
5574 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5575 if (unlikely(!ipt->nr_entries) && !ipt->error)
5576 ipt->error = -EINVAL;
5578 spin_lock(&ctx->completion_lock);
5579 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5580 io_poll_remove_double(req);
5581 if (likely(poll->head)) {
5582 spin_lock_irq(&poll->head->lock);
5583 if (unlikely(list_empty(&poll->wait.entry))) {
5589 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5590 list_del_init(&poll->wait.entry);
5592 WRITE_ONCE(poll->canceled, true);
5593 else if (!poll->done) /* actually waiting for an event */
5594 io_poll_req_insert(req);
5595 spin_unlock_irq(&poll->head->lock);
5607 static int io_arm_poll_handler(struct io_kiocb *req)
5609 const struct io_op_def *def = &io_op_defs[req->opcode];
5610 struct io_ring_ctx *ctx = req->ctx;
5611 struct async_poll *apoll;
5612 struct io_poll_table ipt;
5613 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5616 if (!req->file || !file_can_poll(req->file))
5617 return IO_APOLL_ABORTED;
5618 if (req->flags & REQ_F_POLLED)
5619 return IO_APOLL_ABORTED;
5620 if (!def->pollin && !def->pollout)
5621 return IO_APOLL_ABORTED;
5625 mask |= POLLIN | POLLRDNORM;
5627 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5628 if ((req->opcode == IORING_OP_RECVMSG) &&
5629 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5633 mask |= POLLOUT | POLLWRNORM;
5636 /* if we can't nonblock try, then no point in arming a poll handler */
5637 if (!io_file_supports_nowait(req, rw))
5638 return IO_APOLL_ABORTED;
5640 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5641 if (unlikely(!apoll))
5642 return IO_APOLL_ABORTED;
5643 apoll->double_poll = NULL;
5645 req->flags |= REQ_F_POLLED;
5646 ipt.pt._qproc = io_async_queue_proc;
5647 io_req_set_refcount(req);
5649 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5651 spin_unlock(&ctx->completion_lock);
5652 if (ret || ipt.error)
5653 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5655 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5656 mask, apoll->poll.events);
5660 static bool __io_poll_remove_one(struct io_kiocb *req,
5661 struct io_poll_iocb *poll, bool do_cancel)
5662 __must_hold(&req->ctx->completion_lock)
5664 bool do_complete = false;
5668 spin_lock_irq(&poll->head->lock);
5670 WRITE_ONCE(poll->canceled, true);
5671 if (!list_empty(&poll->wait.entry)) {
5672 list_del_init(&poll->wait.entry);
5675 spin_unlock_irq(&poll->head->lock);
5676 hash_del(&req->hash_node);
5680 static bool io_poll_remove_one(struct io_kiocb *req)
5681 __must_hold(&req->ctx->completion_lock)
5685 io_poll_remove_double(req);
5686 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5689 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5690 io_commit_cqring(req->ctx);
5692 io_put_req_deferred(req);
5698 * Returns true if we found and killed one or more poll requests
5700 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5703 struct hlist_node *tmp;
5704 struct io_kiocb *req;
5707 spin_lock(&ctx->completion_lock);
5708 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5709 struct hlist_head *list;
5711 list = &ctx->cancel_hash[i];
5712 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5713 if (io_match_task(req, tsk, cancel_all))
5714 posted += io_poll_remove_one(req);
5717 spin_unlock(&ctx->completion_lock);
5720 io_cqring_ev_posted(ctx);
5725 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5727 __must_hold(&ctx->completion_lock)
5729 struct hlist_head *list;
5730 struct io_kiocb *req;
5732 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5733 hlist_for_each_entry(req, list, hash_node) {
5734 if (sqe_addr != req->user_data)
5736 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5743 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5745 __must_hold(&ctx->completion_lock)
5747 struct io_kiocb *req;
5749 req = io_poll_find(ctx, sqe_addr, poll_only);
5752 if (io_poll_remove_one(req))
5758 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5763 events = READ_ONCE(sqe->poll32_events);
5765 events = swahw32(events);
5767 if (!(flags & IORING_POLL_ADD_MULTI))
5768 events |= EPOLLONESHOT;
5769 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5772 static int io_poll_update_prep(struct io_kiocb *req,
5773 const struct io_uring_sqe *sqe)
5775 struct io_poll_update *upd = &req->poll_update;
5778 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5780 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5782 flags = READ_ONCE(sqe->len);
5783 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5784 IORING_POLL_ADD_MULTI))
5786 /* meaningless without update */
5787 if (flags == IORING_POLL_ADD_MULTI)
5790 upd->old_user_data = READ_ONCE(sqe->addr);
5791 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5792 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5794 upd->new_user_data = READ_ONCE(sqe->off);
5795 if (!upd->update_user_data && upd->new_user_data)
5797 if (upd->update_events)
5798 upd->events = io_poll_parse_events(sqe, flags);
5799 else if (sqe->poll32_events)
5805 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5808 struct io_kiocb *req = wait->private;
5809 struct io_poll_iocb *poll = &req->poll;
5811 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5814 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5815 struct poll_table_struct *p)
5817 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5819 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5822 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5824 struct io_poll_iocb *poll = &req->poll;
5827 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5829 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5831 flags = READ_ONCE(sqe->len);
5832 if (flags & ~IORING_POLL_ADD_MULTI)
5835 io_req_set_refcount(req);
5836 poll->events = io_poll_parse_events(sqe, flags);
5840 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5842 struct io_poll_iocb *poll = &req->poll;
5843 struct io_ring_ctx *ctx = req->ctx;
5844 struct io_poll_table ipt;
5848 ipt.pt._qproc = io_poll_queue_proc;
5850 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5853 if (mask) { /* no async, we'd stolen it */
5855 done = io_poll_complete(req, mask);
5857 spin_unlock(&ctx->completion_lock);
5860 io_cqring_ev_posted(ctx);
5867 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5869 struct io_ring_ctx *ctx = req->ctx;
5870 struct io_kiocb *preq;
5874 spin_lock(&ctx->completion_lock);
5875 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5881 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5883 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5888 * Don't allow racy completion with singleshot, as we cannot safely
5889 * update those. For multishot, if we're racing with completion, just
5890 * let completion re-add it.
5892 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5893 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5897 /* we now have a detached poll request. reissue. */
5901 spin_unlock(&ctx->completion_lock);
5903 io_req_complete(req, ret);
5906 /* only mask one event flags, keep behavior flags */
5907 if (req->poll_update.update_events) {
5908 preq->poll.events &= ~0xffff;
5909 preq->poll.events |= req->poll_update.events & 0xffff;
5910 preq->poll.events |= IO_POLL_UNMASK;
5912 if (req->poll_update.update_user_data)
5913 preq->user_data = req->poll_update.new_user_data;
5914 spin_unlock(&ctx->completion_lock);
5916 /* complete update request, we're done with it */
5917 io_req_complete(req, ret);
5920 ret = io_poll_add(preq, issue_flags);
5923 io_req_complete(preq, ret);
5929 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5932 io_req_complete_post(req, -ETIME, 0);
5935 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5937 struct io_timeout_data *data = container_of(timer,
5938 struct io_timeout_data, timer);
5939 struct io_kiocb *req = data->req;
5940 struct io_ring_ctx *ctx = req->ctx;
5941 unsigned long flags;
5943 spin_lock_irqsave(&ctx->timeout_lock, flags);
5944 list_del_init(&req->timeout.list);
5945 atomic_set(&req->ctx->cq_timeouts,
5946 atomic_read(&req->ctx->cq_timeouts) + 1);
5947 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5949 req->io_task_work.func = io_req_task_timeout;
5950 io_req_task_work_add(req);
5951 return HRTIMER_NORESTART;
5954 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5956 __must_hold(&ctx->timeout_lock)
5958 struct io_timeout_data *io;
5959 struct io_kiocb *req;
5962 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5963 found = user_data == req->user_data;
5968 return ERR_PTR(-ENOENT);
5970 io = req->async_data;
5971 if (hrtimer_try_to_cancel(&io->timer) == -1)
5972 return ERR_PTR(-EALREADY);
5973 list_del_init(&req->timeout.list);
5977 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5978 __must_hold(&ctx->completion_lock)
5979 __must_hold(&ctx->timeout_lock)
5981 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5984 return PTR_ERR(req);
5987 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5988 io_put_req_deferred(req);
5992 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
5994 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
5995 case IORING_TIMEOUT_BOOTTIME:
5996 return CLOCK_BOOTTIME;
5997 case IORING_TIMEOUT_REALTIME:
5998 return CLOCK_REALTIME;
6000 /* can't happen, vetted at prep time */
6004 return CLOCK_MONOTONIC;
6008 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6009 struct timespec64 *ts, enum hrtimer_mode mode)
6010 __must_hold(&ctx->timeout_lock)
6012 struct io_timeout_data *io;
6013 struct io_kiocb *req;
6016 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6017 found = user_data == req->user_data;
6024 io = req->async_data;
6025 if (hrtimer_try_to_cancel(&io->timer) == -1)
6027 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6028 io->timer.function = io_link_timeout_fn;
6029 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6033 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6034 struct timespec64 *ts, enum hrtimer_mode mode)
6035 __must_hold(&ctx->timeout_lock)
6037 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6038 struct io_timeout_data *data;
6041 return PTR_ERR(req);
6043 req->timeout.off = 0; /* noseq */
6044 data = req->async_data;
6045 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6046 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6047 data->timer.function = io_timeout_fn;
6048 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6052 static int io_timeout_remove_prep(struct io_kiocb *req,
6053 const struct io_uring_sqe *sqe)
6055 struct io_timeout_rem *tr = &req->timeout_rem;
6057 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6059 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6061 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6064 tr->ltimeout = false;
6065 tr->addr = READ_ONCE(sqe->addr);
6066 tr->flags = READ_ONCE(sqe->timeout_flags);
6067 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6068 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6070 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6071 tr->ltimeout = true;
6072 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6074 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6076 } else if (tr->flags) {
6077 /* timeout removal doesn't support flags */
6084 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6086 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6091 * Remove or update an existing timeout command
6093 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6095 struct io_timeout_rem *tr = &req->timeout_rem;
6096 struct io_ring_ctx *ctx = req->ctx;
6099 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6100 spin_lock(&ctx->completion_lock);
6101 spin_lock_irq(&ctx->timeout_lock);
6102 ret = io_timeout_cancel(ctx, tr->addr);
6103 spin_unlock_irq(&ctx->timeout_lock);
6104 spin_unlock(&ctx->completion_lock);
6106 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6108 spin_lock_irq(&ctx->timeout_lock);
6110 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6112 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6113 spin_unlock_irq(&ctx->timeout_lock);
6118 io_req_complete_post(req, ret, 0);
6122 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6123 bool is_timeout_link)
6125 struct io_timeout_data *data;
6127 u32 off = READ_ONCE(sqe->off);
6129 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6131 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6134 if (off && is_timeout_link)
6136 flags = READ_ONCE(sqe->timeout_flags);
6137 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6139 /* more than one clock specified is invalid, obviously */
6140 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6143 INIT_LIST_HEAD(&req->timeout.list);
6144 req->timeout.off = off;
6145 if (unlikely(off && !req->ctx->off_timeout_used))
6146 req->ctx->off_timeout_used = true;
6148 if (!req->async_data && io_alloc_async_data(req))
6151 data = req->async_data;
6153 data->flags = flags;
6155 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6158 data->mode = io_translate_timeout_mode(flags);
6159 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6161 if (is_timeout_link) {
6162 struct io_submit_link *link = &req->ctx->submit_state.link;
6166 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6168 req->timeout.head = link->last;
6169 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6174 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6176 struct io_ring_ctx *ctx = req->ctx;
6177 struct io_timeout_data *data = req->async_data;
6178 struct list_head *entry;
6179 u32 tail, off = req->timeout.off;
6181 spin_lock_irq(&ctx->timeout_lock);
6184 * sqe->off holds how many events that need to occur for this
6185 * timeout event to be satisfied. If it isn't set, then this is
6186 * a pure timeout request, sequence isn't used.
6188 if (io_is_timeout_noseq(req)) {
6189 entry = ctx->timeout_list.prev;
6193 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6194 req->timeout.target_seq = tail + off;
6196 /* Update the last seq here in case io_flush_timeouts() hasn't.
6197 * This is safe because ->completion_lock is held, and submissions
6198 * and completions are never mixed in the same ->completion_lock section.
6200 ctx->cq_last_tm_flush = tail;
6203 * Insertion sort, ensuring the first entry in the list is always
6204 * the one we need first.
6206 list_for_each_prev(entry, &ctx->timeout_list) {
6207 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6210 if (io_is_timeout_noseq(nxt))
6212 /* nxt.seq is behind @tail, otherwise would've been completed */
6213 if (off >= nxt->timeout.target_seq - tail)
6217 list_add(&req->timeout.list, entry);
6218 data->timer.function = io_timeout_fn;
6219 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6220 spin_unlock_irq(&ctx->timeout_lock);
6224 struct io_cancel_data {
6225 struct io_ring_ctx *ctx;
6229 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6231 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6232 struct io_cancel_data *cd = data;
6234 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6237 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6238 struct io_ring_ctx *ctx)
6240 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6241 enum io_wq_cancel cancel_ret;
6244 if (!tctx || !tctx->io_wq)
6247 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6248 switch (cancel_ret) {
6249 case IO_WQ_CANCEL_OK:
6252 case IO_WQ_CANCEL_RUNNING:
6255 case IO_WQ_CANCEL_NOTFOUND:
6263 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6265 struct io_ring_ctx *ctx = req->ctx;
6268 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6270 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6274 spin_lock(&ctx->completion_lock);
6275 spin_lock_irq(&ctx->timeout_lock);
6276 ret = io_timeout_cancel(ctx, sqe_addr);
6277 spin_unlock_irq(&ctx->timeout_lock);
6280 ret = io_poll_cancel(ctx, sqe_addr, false);
6282 spin_unlock(&ctx->completion_lock);
6286 static int io_async_cancel_prep(struct io_kiocb *req,
6287 const struct io_uring_sqe *sqe)
6289 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6291 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6293 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6297 req->cancel.addr = READ_ONCE(sqe->addr);
6301 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6303 struct io_ring_ctx *ctx = req->ctx;
6304 u64 sqe_addr = req->cancel.addr;
6305 struct io_tctx_node *node;
6308 ret = io_try_cancel_userdata(req, sqe_addr);
6312 /* slow path, try all io-wq's */
6313 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6315 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6316 struct io_uring_task *tctx = node->task->io_uring;
6318 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6322 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6326 io_req_complete_post(req, ret, 0);
6330 static int io_rsrc_update_prep(struct io_kiocb *req,
6331 const struct io_uring_sqe *sqe)
6333 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6335 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6338 req->rsrc_update.offset = READ_ONCE(sqe->off);
6339 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6340 if (!req->rsrc_update.nr_args)
6342 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6346 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6348 struct io_ring_ctx *ctx = req->ctx;
6349 struct io_uring_rsrc_update2 up;
6352 up.offset = req->rsrc_update.offset;
6353 up.data = req->rsrc_update.arg;
6358 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6359 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6360 &up, req->rsrc_update.nr_args);
6361 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6365 __io_req_complete(req, issue_flags, ret, 0);
6369 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6371 switch (req->opcode) {
6374 case IORING_OP_READV:
6375 case IORING_OP_READ_FIXED:
6376 case IORING_OP_READ:
6377 return io_read_prep(req, sqe);
6378 case IORING_OP_WRITEV:
6379 case IORING_OP_WRITE_FIXED:
6380 case IORING_OP_WRITE:
6381 return io_write_prep(req, sqe);
6382 case IORING_OP_POLL_ADD:
6383 return io_poll_add_prep(req, sqe);
6384 case IORING_OP_POLL_REMOVE:
6385 return io_poll_update_prep(req, sqe);
6386 case IORING_OP_FSYNC:
6387 return io_fsync_prep(req, sqe);
6388 case IORING_OP_SYNC_FILE_RANGE:
6389 return io_sfr_prep(req, sqe);
6390 case IORING_OP_SENDMSG:
6391 case IORING_OP_SEND:
6392 return io_sendmsg_prep(req, sqe);
6393 case IORING_OP_RECVMSG:
6394 case IORING_OP_RECV:
6395 return io_recvmsg_prep(req, sqe);
6396 case IORING_OP_CONNECT:
6397 return io_connect_prep(req, sqe);
6398 case IORING_OP_TIMEOUT:
6399 return io_timeout_prep(req, sqe, false);
6400 case IORING_OP_TIMEOUT_REMOVE:
6401 return io_timeout_remove_prep(req, sqe);
6402 case IORING_OP_ASYNC_CANCEL:
6403 return io_async_cancel_prep(req, sqe);
6404 case IORING_OP_LINK_TIMEOUT:
6405 return io_timeout_prep(req, sqe, true);
6406 case IORING_OP_ACCEPT:
6407 return io_accept_prep(req, sqe);
6408 case IORING_OP_FALLOCATE:
6409 return io_fallocate_prep(req, sqe);
6410 case IORING_OP_OPENAT:
6411 return io_openat_prep(req, sqe);
6412 case IORING_OP_CLOSE:
6413 return io_close_prep(req, sqe);
6414 case IORING_OP_FILES_UPDATE:
6415 return io_rsrc_update_prep(req, sqe);
6416 case IORING_OP_STATX:
6417 return io_statx_prep(req, sqe);
6418 case IORING_OP_FADVISE:
6419 return io_fadvise_prep(req, sqe);
6420 case IORING_OP_MADVISE:
6421 return io_madvise_prep(req, sqe);
6422 case IORING_OP_OPENAT2:
6423 return io_openat2_prep(req, sqe);
6424 case IORING_OP_EPOLL_CTL:
6425 return io_epoll_ctl_prep(req, sqe);
6426 case IORING_OP_SPLICE:
6427 return io_splice_prep(req, sqe);
6428 case IORING_OP_PROVIDE_BUFFERS:
6429 return io_provide_buffers_prep(req, sqe);
6430 case IORING_OP_REMOVE_BUFFERS:
6431 return io_remove_buffers_prep(req, sqe);
6433 return io_tee_prep(req, sqe);
6434 case IORING_OP_SHUTDOWN:
6435 return io_shutdown_prep(req, sqe);
6436 case IORING_OP_RENAMEAT:
6437 return io_renameat_prep(req, sqe);
6438 case IORING_OP_UNLINKAT:
6439 return io_unlinkat_prep(req, sqe);
6440 case IORING_OP_MKDIRAT:
6441 return io_mkdirat_prep(req, sqe);
6442 case IORING_OP_SYMLINKAT:
6443 return io_symlinkat_prep(req, sqe);
6444 case IORING_OP_LINKAT:
6445 return io_linkat_prep(req, sqe);
6448 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6453 static int io_req_prep_async(struct io_kiocb *req)
6455 if (!io_op_defs[req->opcode].needs_async_setup)
6457 if (WARN_ON_ONCE(req->async_data))
6459 if (io_alloc_async_data(req))
6462 switch (req->opcode) {
6463 case IORING_OP_READV:
6464 return io_rw_prep_async(req, READ);
6465 case IORING_OP_WRITEV:
6466 return io_rw_prep_async(req, WRITE);
6467 case IORING_OP_SENDMSG:
6468 return io_sendmsg_prep_async(req);
6469 case IORING_OP_RECVMSG:
6470 return io_recvmsg_prep_async(req);
6471 case IORING_OP_CONNECT:
6472 return io_connect_prep_async(req);
6474 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6479 static u32 io_get_sequence(struct io_kiocb *req)
6481 u32 seq = req->ctx->cached_sq_head;
6483 /* need original cached_sq_head, but it was increased for each req */
6484 io_for_each_link(req, req)
6489 static bool io_drain_req(struct io_kiocb *req)
6491 struct io_kiocb *pos;
6492 struct io_ring_ctx *ctx = req->ctx;
6493 struct io_defer_entry *de;
6497 if (req->flags & REQ_F_FAIL) {
6498 io_req_complete_fail_submit(req);
6503 * If we need to drain a request in the middle of a link, drain the
6504 * head request and the next request/link after the current link.
6505 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6506 * maintained for every request of our link.
6508 if (ctx->drain_next) {
6509 req->flags |= REQ_F_IO_DRAIN;
6510 ctx->drain_next = false;
6512 /* not interested in head, start from the first linked */
6513 io_for_each_link(pos, req->link) {
6514 if (pos->flags & REQ_F_IO_DRAIN) {
6515 ctx->drain_next = true;
6516 req->flags |= REQ_F_IO_DRAIN;
6521 /* Still need defer if there is pending req in defer list. */
6522 if (likely(list_empty_careful(&ctx->defer_list) &&
6523 !(req->flags & REQ_F_IO_DRAIN))) {
6524 ctx->drain_active = false;
6528 seq = io_get_sequence(req);
6529 /* Still a chance to pass the sequence check */
6530 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6533 ret = io_req_prep_async(req);
6536 io_prep_async_link(req);
6537 de = kmalloc(sizeof(*de), GFP_KERNEL);
6541 io_req_complete_failed(req, ret);
6545 spin_lock(&ctx->completion_lock);
6546 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6547 spin_unlock(&ctx->completion_lock);
6549 io_queue_async_work(req, NULL);
6553 trace_io_uring_defer(ctx, req, req->user_data);
6556 list_add_tail(&de->list, &ctx->defer_list);
6557 spin_unlock(&ctx->completion_lock);
6561 static void io_clean_op(struct io_kiocb *req)
6563 if (req->flags & REQ_F_BUFFER_SELECTED) {
6564 switch (req->opcode) {
6565 case IORING_OP_READV:
6566 case IORING_OP_READ_FIXED:
6567 case IORING_OP_READ:
6568 kfree((void *)(unsigned long)req->rw.addr);
6570 case IORING_OP_RECVMSG:
6571 case IORING_OP_RECV:
6572 kfree(req->sr_msg.kbuf);
6577 if (req->flags & REQ_F_NEED_CLEANUP) {
6578 switch (req->opcode) {
6579 case IORING_OP_READV:
6580 case IORING_OP_READ_FIXED:
6581 case IORING_OP_READ:
6582 case IORING_OP_WRITEV:
6583 case IORING_OP_WRITE_FIXED:
6584 case IORING_OP_WRITE: {
6585 struct io_async_rw *io = req->async_data;
6587 kfree(io->free_iovec);
6590 case IORING_OP_RECVMSG:
6591 case IORING_OP_SENDMSG: {
6592 struct io_async_msghdr *io = req->async_data;
6594 kfree(io->free_iov);
6597 case IORING_OP_SPLICE:
6599 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6600 io_put_file(req->splice.file_in);
6602 case IORING_OP_OPENAT:
6603 case IORING_OP_OPENAT2:
6604 if (req->open.filename)
6605 putname(req->open.filename);
6607 case IORING_OP_RENAMEAT:
6608 putname(req->rename.oldpath);
6609 putname(req->rename.newpath);
6611 case IORING_OP_UNLINKAT:
6612 putname(req->unlink.filename);
6614 case IORING_OP_MKDIRAT:
6615 putname(req->mkdir.filename);
6617 case IORING_OP_SYMLINKAT:
6618 putname(req->symlink.oldpath);
6619 putname(req->symlink.newpath);
6621 case IORING_OP_LINKAT:
6622 putname(req->hardlink.oldpath);
6623 putname(req->hardlink.newpath);
6627 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6628 kfree(req->apoll->double_poll);
6632 if (req->flags & REQ_F_INFLIGHT) {
6633 struct io_uring_task *tctx = req->task->io_uring;
6635 atomic_dec(&tctx->inflight_tracked);
6637 if (req->flags & REQ_F_CREDS)
6638 put_cred(req->creds);
6640 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6643 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6645 struct io_ring_ctx *ctx = req->ctx;
6646 const struct cred *creds = NULL;
6649 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6650 creds = override_creds(req->creds);
6652 switch (req->opcode) {
6654 ret = io_nop(req, issue_flags);
6656 case IORING_OP_READV:
6657 case IORING_OP_READ_FIXED:
6658 case IORING_OP_READ:
6659 ret = io_read(req, issue_flags);
6661 case IORING_OP_WRITEV:
6662 case IORING_OP_WRITE_FIXED:
6663 case IORING_OP_WRITE:
6664 ret = io_write(req, issue_flags);
6666 case IORING_OP_FSYNC:
6667 ret = io_fsync(req, issue_flags);
6669 case IORING_OP_POLL_ADD:
6670 ret = io_poll_add(req, issue_flags);
6672 case IORING_OP_POLL_REMOVE:
6673 ret = io_poll_update(req, issue_flags);
6675 case IORING_OP_SYNC_FILE_RANGE:
6676 ret = io_sync_file_range(req, issue_flags);
6678 case IORING_OP_SENDMSG:
6679 ret = io_sendmsg(req, issue_flags);
6681 case IORING_OP_SEND:
6682 ret = io_send(req, issue_flags);
6684 case IORING_OP_RECVMSG:
6685 ret = io_recvmsg(req, issue_flags);
6687 case IORING_OP_RECV:
6688 ret = io_recv(req, issue_flags);
6690 case IORING_OP_TIMEOUT:
6691 ret = io_timeout(req, issue_flags);
6693 case IORING_OP_TIMEOUT_REMOVE:
6694 ret = io_timeout_remove(req, issue_flags);
6696 case IORING_OP_ACCEPT:
6697 ret = io_accept(req, issue_flags);
6699 case IORING_OP_CONNECT:
6700 ret = io_connect(req, issue_flags);
6702 case IORING_OP_ASYNC_CANCEL:
6703 ret = io_async_cancel(req, issue_flags);
6705 case IORING_OP_FALLOCATE:
6706 ret = io_fallocate(req, issue_flags);
6708 case IORING_OP_OPENAT:
6709 ret = io_openat(req, issue_flags);
6711 case IORING_OP_CLOSE:
6712 ret = io_close(req, issue_flags);
6714 case IORING_OP_FILES_UPDATE:
6715 ret = io_files_update(req, issue_flags);
6717 case IORING_OP_STATX:
6718 ret = io_statx(req, issue_flags);
6720 case IORING_OP_FADVISE:
6721 ret = io_fadvise(req, issue_flags);
6723 case IORING_OP_MADVISE:
6724 ret = io_madvise(req, issue_flags);
6726 case IORING_OP_OPENAT2:
6727 ret = io_openat2(req, issue_flags);
6729 case IORING_OP_EPOLL_CTL:
6730 ret = io_epoll_ctl(req, issue_flags);
6732 case IORING_OP_SPLICE:
6733 ret = io_splice(req, issue_flags);
6735 case IORING_OP_PROVIDE_BUFFERS:
6736 ret = io_provide_buffers(req, issue_flags);
6738 case IORING_OP_REMOVE_BUFFERS:
6739 ret = io_remove_buffers(req, issue_flags);
6742 ret = io_tee(req, issue_flags);
6744 case IORING_OP_SHUTDOWN:
6745 ret = io_shutdown(req, issue_flags);
6747 case IORING_OP_RENAMEAT:
6748 ret = io_renameat(req, issue_flags);
6750 case IORING_OP_UNLINKAT:
6751 ret = io_unlinkat(req, issue_flags);
6753 case IORING_OP_MKDIRAT:
6754 ret = io_mkdirat(req, issue_flags);
6756 case IORING_OP_SYMLINKAT:
6757 ret = io_symlinkat(req, issue_flags);
6759 case IORING_OP_LINKAT:
6760 ret = io_linkat(req, issue_flags);
6768 revert_creds(creds);
6771 /* If the op doesn't have a file, we're not polling for it */
6772 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6773 io_iopoll_req_issued(req);
6778 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6780 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6782 req = io_put_req_find_next(req);
6783 return req ? &req->work : NULL;
6786 static void io_wq_submit_work(struct io_wq_work *work)
6788 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6789 struct io_kiocb *timeout;
6792 /* one will be dropped by ->io_free_work() after returning to io-wq */
6793 if (!(req->flags & REQ_F_REFCOUNT))
6794 __io_req_set_refcount(req, 2);
6798 timeout = io_prep_linked_timeout(req);
6800 io_queue_linked_timeout(timeout);
6802 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6803 if (work->flags & IO_WQ_WORK_CANCEL)
6808 ret = io_issue_sqe(req, 0);
6810 * We can get EAGAIN for polled IO even though we're
6811 * forcing a sync submission from here, since we can't
6812 * wait for request slots on the block side.
6820 /* avoid locking problems by failing it from a clean context */
6822 io_req_task_queue_fail(req, ret);
6825 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6828 return &table->files[i];
6831 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6834 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6836 return (struct file *) (slot->file_ptr & FFS_MASK);
6839 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6841 unsigned long file_ptr = (unsigned long) file;
6843 if (__io_file_supports_nowait(file, READ))
6844 file_ptr |= FFS_ASYNC_READ;
6845 if (__io_file_supports_nowait(file, WRITE))
6846 file_ptr |= FFS_ASYNC_WRITE;
6847 if (S_ISREG(file_inode(file)->i_mode))
6848 file_ptr |= FFS_ISREG;
6849 file_slot->file_ptr = file_ptr;
6852 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6853 struct io_kiocb *req, int fd)
6856 unsigned long file_ptr;
6858 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6860 fd = array_index_nospec(fd, ctx->nr_user_files);
6861 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6862 file = (struct file *) (file_ptr & FFS_MASK);
6863 file_ptr &= ~FFS_MASK;
6864 /* mask in overlapping REQ_F and FFS bits */
6865 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6866 io_req_set_rsrc_node(req);
6870 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6871 struct io_kiocb *req, int fd)
6873 struct file *file = fget(fd);
6875 trace_io_uring_file_get(ctx, fd);
6877 /* we don't allow fixed io_uring files */
6878 if (file && unlikely(file->f_op == &io_uring_fops))
6879 io_req_track_inflight(req);
6883 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6884 struct io_kiocb *req, int fd, bool fixed)
6887 return io_file_get_fixed(ctx, req, fd);
6889 return io_file_get_normal(ctx, req, fd);
6892 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6894 struct io_kiocb *prev = req->timeout.prev;
6898 ret = io_try_cancel_userdata(req, prev->user_data);
6899 io_req_complete_post(req, ret ?: -ETIME, 0);
6902 io_req_complete_post(req, -ETIME, 0);
6906 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6908 struct io_timeout_data *data = container_of(timer,
6909 struct io_timeout_data, timer);
6910 struct io_kiocb *prev, *req = data->req;
6911 struct io_ring_ctx *ctx = req->ctx;
6912 unsigned long flags;
6914 spin_lock_irqsave(&ctx->timeout_lock, flags);
6915 prev = req->timeout.head;
6916 req->timeout.head = NULL;
6919 * We don't expect the list to be empty, that will only happen if we
6920 * race with the completion of the linked work.
6923 io_remove_next_linked(prev);
6924 if (!req_ref_inc_not_zero(prev))
6927 list_del(&req->timeout.list);
6928 req->timeout.prev = prev;
6929 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6931 req->io_task_work.func = io_req_task_link_timeout;
6932 io_req_task_work_add(req);
6933 return HRTIMER_NORESTART;
6936 static void io_queue_linked_timeout(struct io_kiocb *req)
6938 struct io_ring_ctx *ctx = req->ctx;
6940 spin_lock_irq(&ctx->timeout_lock);
6942 * If the back reference is NULL, then our linked request finished
6943 * before we got a chance to setup the timer
6945 if (req->timeout.head) {
6946 struct io_timeout_data *data = req->async_data;
6948 data->timer.function = io_link_timeout_fn;
6949 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6951 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6953 spin_unlock_irq(&ctx->timeout_lock);
6954 /* drop submission reference */
6958 static void __io_queue_sqe(struct io_kiocb *req)
6959 __must_hold(&req->ctx->uring_lock)
6961 struct io_kiocb *linked_timeout;
6965 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6968 * We async punt it if the file wasn't marked NOWAIT, or if the file
6969 * doesn't support non-blocking read/write attempts
6972 if (req->flags & REQ_F_COMPLETE_INLINE) {
6973 struct io_ring_ctx *ctx = req->ctx;
6974 struct io_submit_state *state = &ctx->submit_state;
6976 state->compl_reqs[state->compl_nr++] = req;
6977 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6978 io_submit_flush_completions(ctx);
6982 linked_timeout = io_prep_linked_timeout(req);
6984 io_queue_linked_timeout(linked_timeout);
6985 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6986 linked_timeout = io_prep_linked_timeout(req);
6988 switch (io_arm_poll_handler(req)) {
6989 case IO_APOLL_READY:
6991 io_unprep_linked_timeout(req);
6993 case IO_APOLL_ABORTED:
6995 * Queued up for async execution, worker will release
6996 * submit reference when the iocb is actually submitted.
6998 io_queue_async_work(req, NULL);
7003 io_queue_linked_timeout(linked_timeout);
7005 io_req_complete_failed(req, ret);
7009 static inline void io_queue_sqe(struct io_kiocb *req)
7010 __must_hold(&req->ctx->uring_lock)
7012 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7015 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7016 __io_queue_sqe(req);
7017 } else if (req->flags & REQ_F_FAIL) {
7018 io_req_complete_fail_submit(req);
7020 int ret = io_req_prep_async(req);
7023 io_req_complete_failed(req, ret);
7025 io_queue_async_work(req, NULL);
7030 * Check SQE restrictions (opcode and flags).
7032 * Returns 'true' if SQE is allowed, 'false' otherwise.
7034 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7035 struct io_kiocb *req,
7036 unsigned int sqe_flags)
7038 if (likely(!ctx->restricted))
7041 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7044 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7045 ctx->restrictions.sqe_flags_required)
7048 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7049 ctx->restrictions.sqe_flags_required))
7055 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7056 const struct io_uring_sqe *sqe)
7057 __must_hold(&ctx->uring_lock)
7059 struct io_submit_state *state;
7060 unsigned int sqe_flags;
7061 int personality, ret = 0;
7063 /* req is partially pre-initialised, see io_preinit_req() */
7064 req->opcode = READ_ONCE(sqe->opcode);
7065 /* same numerical values with corresponding REQ_F_*, safe to copy */
7066 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7067 req->user_data = READ_ONCE(sqe->user_data);
7069 req->fixed_rsrc_refs = NULL;
7070 req->task = current;
7072 /* enforce forwards compatibility on users */
7073 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7075 if (unlikely(req->opcode >= IORING_OP_LAST))
7077 if (!io_check_restriction(ctx, req, sqe_flags))
7080 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7081 !io_op_defs[req->opcode].buffer_select)
7083 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7084 ctx->drain_active = true;
7086 personality = READ_ONCE(sqe->personality);
7088 req->creds = xa_load(&ctx->personalities, personality);
7091 get_cred(req->creds);
7092 req->flags |= REQ_F_CREDS;
7094 state = &ctx->submit_state;
7097 * Plug now if we have more than 1 IO left after this, and the target
7098 * is potentially a read/write to block based storage.
7100 if (!state->plug_started && state->ios_left > 1 &&
7101 io_op_defs[req->opcode].plug) {
7102 blk_start_plug(&state->plug);
7103 state->plug_started = true;
7106 if (io_op_defs[req->opcode].needs_file) {
7107 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7108 (sqe_flags & IOSQE_FIXED_FILE));
7109 if (unlikely(!req->file))
7117 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7118 const struct io_uring_sqe *sqe)
7119 __must_hold(&ctx->uring_lock)
7121 struct io_submit_link *link = &ctx->submit_state.link;
7124 ret = io_init_req(ctx, req, sqe);
7125 if (unlikely(ret)) {
7127 /* fail even hard links since we don't submit */
7130 * we can judge a link req is failed or cancelled by if
7131 * REQ_F_FAIL is set, but the head is an exception since
7132 * it may be set REQ_F_FAIL because of other req's failure
7133 * so let's leverage req->result to distinguish if a head
7134 * is set REQ_F_FAIL because of its failure or other req's
7135 * failure so that we can set the correct ret code for it.
7136 * init result here to avoid affecting the normal path.
7138 if (!(link->head->flags & REQ_F_FAIL))
7139 req_fail_link_node(link->head, -ECANCELED);
7140 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7142 * the current req is a normal req, we should return
7143 * error and thus break the submittion loop.
7145 io_req_complete_failed(req, ret);
7148 req_fail_link_node(req, ret);
7150 ret = io_req_prep(req, sqe);
7155 /* don't need @sqe from now on */
7156 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7158 ctx->flags & IORING_SETUP_SQPOLL);
7161 * If we already have a head request, queue this one for async
7162 * submittal once the head completes. If we don't have a head but
7163 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7164 * submitted sync once the chain is complete. If none of those
7165 * conditions are true (normal request), then just queue it.
7168 struct io_kiocb *head = link->head;
7170 if (!(req->flags & REQ_F_FAIL)) {
7171 ret = io_req_prep_async(req);
7172 if (unlikely(ret)) {
7173 req_fail_link_node(req, ret);
7174 if (!(head->flags & REQ_F_FAIL))
7175 req_fail_link_node(head, -ECANCELED);
7178 trace_io_uring_link(ctx, req, head);
7179 link->last->link = req;
7182 /* last request of a link, enqueue the link */
7183 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7188 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7200 * Batched submission is done, ensure local IO is flushed out.
7202 static void io_submit_state_end(struct io_submit_state *state,
7203 struct io_ring_ctx *ctx)
7205 if (state->link.head)
7206 io_queue_sqe(state->link.head);
7207 if (state->compl_nr)
7208 io_submit_flush_completions(ctx);
7209 if (state->plug_started)
7210 blk_finish_plug(&state->plug);
7214 * Start submission side cache.
7216 static void io_submit_state_start(struct io_submit_state *state,
7217 unsigned int max_ios)
7219 state->plug_started = false;
7220 state->ios_left = max_ios;
7221 /* set only head, no need to init link_last in advance */
7222 state->link.head = NULL;
7225 static void io_commit_sqring(struct io_ring_ctx *ctx)
7227 struct io_rings *rings = ctx->rings;
7230 * Ensure any loads from the SQEs are done at this point,
7231 * since once we write the new head, the application could
7232 * write new data to them.
7234 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7238 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7239 * that is mapped by userspace. This means that care needs to be taken to
7240 * ensure that reads are stable, as we cannot rely on userspace always
7241 * being a good citizen. If members of the sqe are validated and then later
7242 * used, it's important that those reads are done through READ_ONCE() to
7243 * prevent a re-load down the line.
7245 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7247 unsigned head, mask = ctx->sq_entries - 1;
7248 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7251 * The cached sq head (or cq tail) serves two purposes:
7253 * 1) allows us to batch the cost of updating the user visible
7255 * 2) allows the kernel side to track the head on its own, even
7256 * though the application is the one updating it.
7258 head = READ_ONCE(ctx->sq_array[sq_idx]);
7259 if (likely(head < ctx->sq_entries))
7260 return &ctx->sq_sqes[head];
7262 /* drop invalid entries */
7264 WRITE_ONCE(ctx->rings->sq_dropped,
7265 READ_ONCE(ctx->rings->sq_dropped) + 1);
7269 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7270 __must_hold(&ctx->uring_lock)
7274 /* make sure SQ entry isn't read before tail */
7275 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7276 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7278 io_get_task_refs(nr);
7280 io_submit_state_start(&ctx->submit_state, nr);
7281 while (submitted < nr) {
7282 const struct io_uring_sqe *sqe;
7283 struct io_kiocb *req;
7285 req = io_alloc_req(ctx);
7286 if (unlikely(!req)) {
7288 submitted = -EAGAIN;
7291 sqe = io_get_sqe(ctx);
7292 if (unlikely(!sqe)) {
7293 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7296 /* will complete beyond this point, count as submitted */
7298 if (io_submit_sqe(ctx, req, sqe))
7302 if (unlikely(submitted != nr)) {
7303 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7304 int unused = nr - ref_used;
7306 current->io_uring->cached_refs += unused;
7307 percpu_ref_put_many(&ctx->refs, unused);
7310 io_submit_state_end(&ctx->submit_state, ctx);
7311 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7312 io_commit_sqring(ctx);
7317 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7319 return READ_ONCE(sqd->state);
7322 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7324 /* Tell userspace we may need a wakeup call */
7325 spin_lock(&ctx->completion_lock);
7326 WRITE_ONCE(ctx->rings->sq_flags,
7327 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7328 spin_unlock(&ctx->completion_lock);
7331 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7333 spin_lock(&ctx->completion_lock);
7334 WRITE_ONCE(ctx->rings->sq_flags,
7335 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7336 spin_unlock(&ctx->completion_lock);
7339 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7341 unsigned int to_submit;
7344 to_submit = io_sqring_entries(ctx);
7345 /* if we're handling multiple rings, cap submit size for fairness */
7346 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7347 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7349 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7350 unsigned nr_events = 0;
7351 const struct cred *creds = NULL;
7353 if (ctx->sq_creds != current_cred())
7354 creds = override_creds(ctx->sq_creds);
7356 mutex_lock(&ctx->uring_lock);
7357 if (!list_empty(&ctx->iopoll_list))
7358 io_do_iopoll(ctx, &nr_events, 0);
7361 * Don't submit if refs are dying, good for io_uring_register(),
7362 * but also it is relied upon by io_ring_exit_work()
7364 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7365 !(ctx->flags & IORING_SETUP_R_DISABLED))
7366 ret = io_submit_sqes(ctx, to_submit);
7367 mutex_unlock(&ctx->uring_lock);
7369 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7370 wake_up(&ctx->sqo_sq_wait);
7372 revert_creds(creds);
7378 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7380 struct io_ring_ctx *ctx;
7381 unsigned sq_thread_idle = 0;
7383 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7384 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7385 sqd->sq_thread_idle = sq_thread_idle;
7388 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7390 bool did_sig = false;
7391 struct ksignal ksig;
7393 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7394 signal_pending(current)) {
7395 mutex_unlock(&sqd->lock);
7396 if (signal_pending(current))
7397 did_sig = get_signal(&ksig);
7399 mutex_lock(&sqd->lock);
7401 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7404 static int io_sq_thread(void *data)
7406 struct io_sq_data *sqd = data;
7407 struct io_ring_ctx *ctx;
7408 unsigned long timeout = 0;
7409 char buf[TASK_COMM_LEN];
7412 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7413 set_task_comm(current, buf);
7415 if (sqd->sq_cpu != -1)
7416 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7418 set_cpus_allowed_ptr(current, cpu_online_mask);
7419 current->flags |= PF_NO_SETAFFINITY;
7421 mutex_lock(&sqd->lock);
7423 bool cap_entries, sqt_spin = false;
7425 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7426 if (io_sqd_handle_event(sqd))
7428 timeout = jiffies + sqd->sq_thread_idle;
7431 cap_entries = !list_is_singular(&sqd->ctx_list);
7432 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7433 int ret = __io_sq_thread(ctx, cap_entries);
7435 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7438 if (io_run_task_work())
7441 if (sqt_spin || !time_after(jiffies, timeout)) {
7444 timeout = jiffies + sqd->sq_thread_idle;
7448 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7449 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7450 bool needs_sched = true;
7452 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7453 io_ring_set_wakeup_flag(ctx);
7455 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7456 !list_empty_careful(&ctx->iopoll_list)) {
7457 needs_sched = false;
7460 if (io_sqring_entries(ctx)) {
7461 needs_sched = false;
7467 mutex_unlock(&sqd->lock);
7469 mutex_lock(&sqd->lock);
7471 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7472 io_ring_clear_wakeup_flag(ctx);
7475 finish_wait(&sqd->wait, &wait);
7476 timeout = jiffies + sqd->sq_thread_idle;
7479 io_uring_cancel_generic(true, sqd);
7481 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7482 io_ring_set_wakeup_flag(ctx);
7484 mutex_unlock(&sqd->lock);
7486 complete(&sqd->exited);
7490 struct io_wait_queue {
7491 struct wait_queue_entry wq;
7492 struct io_ring_ctx *ctx;
7494 unsigned nr_timeouts;
7497 static inline bool io_should_wake(struct io_wait_queue *iowq)
7499 struct io_ring_ctx *ctx = iowq->ctx;
7500 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7503 * Wake up if we have enough events, or if a timeout occurred since we
7504 * started waiting. For timeouts, we always want to return to userspace,
7505 * regardless of event count.
7507 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7510 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7511 int wake_flags, void *key)
7513 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7517 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7518 * the task, and the next invocation will do it.
7520 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7521 return autoremove_wake_function(curr, mode, wake_flags, key);
7525 static int io_run_task_work_sig(void)
7527 if (io_run_task_work())
7529 if (!signal_pending(current))
7531 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7532 return -ERESTARTSYS;
7536 /* when returns >0, the caller should retry */
7537 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7538 struct io_wait_queue *iowq,
7539 signed long *timeout)
7543 /* make sure we run task_work before checking for signals */
7544 ret = io_run_task_work_sig();
7545 if (ret || io_should_wake(iowq))
7547 /* let the caller flush overflows, retry */
7548 if (test_bit(0, &ctx->check_cq_overflow))
7551 *timeout = schedule_timeout(*timeout);
7552 return !*timeout ? -ETIME : 1;
7556 * Wait until events become available, if we don't already have some. The
7557 * application must reap them itself, as they reside on the shared cq ring.
7559 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7560 const sigset_t __user *sig, size_t sigsz,
7561 struct __kernel_timespec __user *uts)
7563 struct io_wait_queue iowq;
7564 struct io_rings *rings = ctx->rings;
7565 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7569 io_cqring_overflow_flush(ctx);
7570 if (io_cqring_events(ctx) >= min_events)
7572 if (!io_run_task_work())
7577 struct timespec64 ts;
7579 if (get_timespec64(&ts, uts))
7581 timeout = timespec64_to_jiffies(&ts);
7585 #ifdef CONFIG_COMPAT
7586 if (in_compat_syscall())
7587 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7591 ret = set_user_sigmask(sig, sigsz);
7597 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7598 iowq.wq.private = current;
7599 INIT_LIST_HEAD(&iowq.wq.entry);
7601 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7602 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7604 trace_io_uring_cqring_wait(ctx, min_events);
7606 /* if we can't even flush overflow, don't wait for more */
7607 if (!io_cqring_overflow_flush(ctx)) {
7611 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7612 TASK_INTERRUPTIBLE);
7613 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7614 finish_wait(&ctx->cq_wait, &iowq.wq);
7618 restore_saved_sigmask_unless(ret == -EINTR);
7620 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7623 static void io_free_page_table(void **table, size_t size)
7625 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7627 for (i = 0; i < nr_tables; i++)
7632 static void **io_alloc_page_table(size_t size)
7634 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7635 size_t init_size = size;
7638 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7642 for (i = 0; i < nr_tables; i++) {
7643 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7645 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7647 io_free_page_table(table, init_size);
7655 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7657 percpu_ref_exit(&ref_node->refs);
7661 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7663 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7664 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7665 unsigned long flags;
7666 bool first_add = false;
7668 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7671 while (!list_empty(&ctx->rsrc_ref_list)) {
7672 node = list_first_entry(&ctx->rsrc_ref_list,
7673 struct io_rsrc_node, node);
7674 /* recycle ref nodes in order */
7677 list_del(&node->node);
7678 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7680 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7683 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7686 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7688 struct io_rsrc_node *ref_node;
7690 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7694 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7699 INIT_LIST_HEAD(&ref_node->node);
7700 INIT_LIST_HEAD(&ref_node->rsrc_list);
7701 ref_node->done = false;
7705 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7706 struct io_rsrc_data *data_to_kill)
7708 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7709 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7712 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7714 rsrc_node->rsrc_data = data_to_kill;
7715 spin_lock_irq(&ctx->rsrc_ref_lock);
7716 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7717 spin_unlock_irq(&ctx->rsrc_ref_lock);
7719 atomic_inc(&data_to_kill->refs);
7720 percpu_ref_kill(&rsrc_node->refs);
7721 ctx->rsrc_node = NULL;
7724 if (!ctx->rsrc_node) {
7725 ctx->rsrc_node = ctx->rsrc_backup_node;
7726 ctx->rsrc_backup_node = NULL;
7730 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7732 if (ctx->rsrc_backup_node)
7734 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7735 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7738 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7742 /* As we may drop ->uring_lock, other task may have started quiesce */
7746 data->quiesce = true;
7748 ret = io_rsrc_node_switch_start(ctx);
7751 io_rsrc_node_switch(ctx, data);
7753 /* kill initial ref, already quiesced if zero */
7754 if (atomic_dec_and_test(&data->refs))
7756 mutex_unlock(&ctx->uring_lock);
7757 flush_delayed_work(&ctx->rsrc_put_work);
7758 ret = wait_for_completion_interruptible(&data->done);
7760 mutex_lock(&ctx->uring_lock);
7764 atomic_inc(&data->refs);
7765 /* wait for all works potentially completing data->done */
7766 flush_delayed_work(&ctx->rsrc_put_work);
7767 reinit_completion(&data->done);
7769 ret = io_run_task_work_sig();
7770 mutex_lock(&ctx->uring_lock);
7772 data->quiesce = false;
7777 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7779 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7780 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7782 return &data->tags[table_idx][off];
7785 static void io_rsrc_data_free(struct io_rsrc_data *data)
7787 size_t size = data->nr * sizeof(data->tags[0][0]);
7790 io_free_page_table((void **)data->tags, size);
7794 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7795 u64 __user *utags, unsigned nr,
7796 struct io_rsrc_data **pdata)
7798 struct io_rsrc_data *data;
7802 data = kzalloc(sizeof(*data), GFP_KERNEL);
7805 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7813 data->do_put = do_put;
7816 for (i = 0; i < nr; i++) {
7817 u64 *tag_slot = io_get_tag_slot(data, i);
7819 if (copy_from_user(tag_slot, &utags[i],
7825 atomic_set(&data->refs, 1);
7826 init_completion(&data->done);
7830 io_rsrc_data_free(data);
7834 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7836 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7837 GFP_KERNEL_ACCOUNT);
7838 return !!table->files;
7841 static void io_free_file_tables(struct io_file_table *table)
7843 kvfree(table->files);
7844 table->files = NULL;
7847 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7849 #if defined(CONFIG_UNIX)
7850 if (ctx->ring_sock) {
7851 struct sock *sock = ctx->ring_sock->sk;
7852 struct sk_buff *skb;
7854 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7860 for (i = 0; i < ctx->nr_user_files; i++) {
7863 file = io_file_from_index(ctx, i);
7868 io_free_file_tables(&ctx->file_table);
7869 io_rsrc_data_free(ctx->file_data);
7870 ctx->file_data = NULL;
7871 ctx->nr_user_files = 0;
7874 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7878 if (!ctx->file_data)
7880 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7882 __io_sqe_files_unregister(ctx);
7886 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7887 __releases(&sqd->lock)
7889 WARN_ON_ONCE(sqd->thread == current);
7892 * Do the dance but not conditional clear_bit() because it'd race with
7893 * other threads incrementing park_pending and setting the bit.
7895 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7896 if (atomic_dec_return(&sqd->park_pending))
7897 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7898 mutex_unlock(&sqd->lock);
7901 static void io_sq_thread_park(struct io_sq_data *sqd)
7902 __acquires(&sqd->lock)
7904 WARN_ON_ONCE(sqd->thread == current);
7906 atomic_inc(&sqd->park_pending);
7907 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7908 mutex_lock(&sqd->lock);
7910 wake_up_process(sqd->thread);
7913 static void io_sq_thread_stop(struct io_sq_data *sqd)
7915 WARN_ON_ONCE(sqd->thread == current);
7916 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7918 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7919 mutex_lock(&sqd->lock);
7921 wake_up_process(sqd->thread);
7922 mutex_unlock(&sqd->lock);
7923 wait_for_completion(&sqd->exited);
7926 static void io_put_sq_data(struct io_sq_data *sqd)
7928 if (refcount_dec_and_test(&sqd->refs)) {
7929 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7931 io_sq_thread_stop(sqd);
7936 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7938 struct io_sq_data *sqd = ctx->sq_data;
7941 io_sq_thread_park(sqd);
7942 list_del_init(&ctx->sqd_list);
7943 io_sqd_update_thread_idle(sqd);
7944 io_sq_thread_unpark(sqd);
7946 io_put_sq_data(sqd);
7947 ctx->sq_data = NULL;
7951 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7953 struct io_ring_ctx *ctx_attach;
7954 struct io_sq_data *sqd;
7957 f = fdget(p->wq_fd);
7959 return ERR_PTR(-ENXIO);
7960 if (f.file->f_op != &io_uring_fops) {
7962 return ERR_PTR(-EINVAL);
7965 ctx_attach = f.file->private_data;
7966 sqd = ctx_attach->sq_data;
7969 return ERR_PTR(-EINVAL);
7971 if (sqd->task_tgid != current->tgid) {
7973 return ERR_PTR(-EPERM);
7976 refcount_inc(&sqd->refs);
7981 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7984 struct io_sq_data *sqd;
7987 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7988 sqd = io_attach_sq_data(p);
7993 /* fall through for EPERM case, setup new sqd/task */
7994 if (PTR_ERR(sqd) != -EPERM)
7998 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
8000 return ERR_PTR(-ENOMEM);
8002 atomic_set(&sqd->park_pending, 0);
8003 refcount_set(&sqd->refs, 1);
8004 INIT_LIST_HEAD(&sqd->ctx_list);
8005 mutex_init(&sqd->lock);
8006 init_waitqueue_head(&sqd->wait);
8007 init_completion(&sqd->exited);
8011 #if defined(CONFIG_UNIX)
8013 * Ensure the UNIX gc is aware of our file set, so we are certain that
8014 * the io_uring can be safely unregistered on process exit, even if we have
8015 * loops in the file referencing.
8017 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8019 struct sock *sk = ctx->ring_sock->sk;
8020 struct scm_fp_list *fpl;
8021 struct sk_buff *skb;
8024 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8028 skb = alloc_skb(0, GFP_KERNEL);
8037 fpl->user = get_uid(current_user());
8038 for (i = 0; i < nr; i++) {
8039 struct file *file = io_file_from_index(ctx, i + offset);
8043 fpl->fp[nr_files] = get_file(file);
8044 unix_inflight(fpl->user, fpl->fp[nr_files]);
8049 fpl->max = SCM_MAX_FD;
8050 fpl->count = nr_files;
8051 UNIXCB(skb).fp = fpl;
8052 skb->destructor = unix_destruct_scm;
8053 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8054 skb_queue_head(&sk->sk_receive_queue, skb);
8056 for (i = 0; i < nr_files; i++)
8067 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8068 * causes regular reference counting to break down. We rely on the UNIX
8069 * garbage collection to take care of this problem for us.
8071 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8073 unsigned left, total;
8077 left = ctx->nr_user_files;
8079 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8081 ret = __io_sqe_files_scm(ctx, this_files, total);
8085 total += this_files;
8091 while (total < ctx->nr_user_files) {
8092 struct file *file = io_file_from_index(ctx, total);
8102 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8108 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8110 struct file *file = prsrc->file;
8111 #if defined(CONFIG_UNIX)
8112 struct sock *sock = ctx->ring_sock->sk;
8113 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8114 struct sk_buff *skb;
8117 __skb_queue_head_init(&list);
8120 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8121 * remove this entry and rearrange the file array.
8123 skb = skb_dequeue(head);
8125 struct scm_fp_list *fp;
8127 fp = UNIXCB(skb).fp;
8128 for (i = 0; i < fp->count; i++) {
8131 if (fp->fp[i] != file)
8134 unix_notinflight(fp->user, fp->fp[i]);
8135 left = fp->count - 1 - i;
8137 memmove(&fp->fp[i], &fp->fp[i + 1],
8138 left * sizeof(struct file *));
8145 __skb_queue_tail(&list, skb);
8155 __skb_queue_tail(&list, skb);
8157 skb = skb_dequeue(head);
8160 if (skb_peek(&list)) {
8161 spin_lock_irq(&head->lock);
8162 while ((skb = __skb_dequeue(&list)) != NULL)
8163 __skb_queue_tail(head, skb);
8164 spin_unlock_irq(&head->lock);
8171 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8173 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8174 struct io_ring_ctx *ctx = rsrc_data->ctx;
8175 struct io_rsrc_put *prsrc, *tmp;
8177 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8178 list_del(&prsrc->list);
8181 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8183 io_ring_submit_lock(ctx, lock_ring);
8184 spin_lock(&ctx->completion_lock);
8185 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8187 io_commit_cqring(ctx);
8188 spin_unlock(&ctx->completion_lock);
8189 io_cqring_ev_posted(ctx);
8190 io_ring_submit_unlock(ctx, lock_ring);
8193 rsrc_data->do_put(ctx, prsrc);
8197 io_rsrc_node_destroy(ref_node);
8198 if (atomic_dec_and_test(&rsrc_data->refs))
8199 complete(&rsrc_data->done);
8202 static void io_rsrc_put_work(struct work_struct *work)
8204 struct io_ring_ctx *ctx;
8205 struct llist_node *node;
8207 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8208 node = llist_del_all(&ctx->rsrc_put_llist);
8211 struct io_rsrc_node *ref_node;
8212 struct llist_node *next = node->next;
8214 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8215 __io_rsrc_put_work(ref_node);
8220 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8221 unsigned nr_args, u64 __user *tags)
8223 __s32 __user *fds = (__s32 __user *) arg;
8232 if (nr_args > IORING_MAX_FIXED_FILES)
8234 if (nr_args > rlimit(RLIMIT_NOFILE))
8236 ret = io_rsrc_node_switch_start(ctx);
8239 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8245 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8248 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8249 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8253 /* allow sparse sets */
8256 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8263 if (unlikely(!file))
8267 * Don't allow io_uring instances to be registered. If UNIX
8268 * isn't enabled, then this causes a reference cycle and this
8269 * instance can never get freed. If UNIX is enabled we'll
8270 * handle it just fine, but there's still no point in allowing
8271 * a ring fd as it doesn't support regular read/write anyway.
8273 if (file->f_op == &io_uring_fops) {
8277 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8280 ret = io_sqe_files_scm(ctx);
8282 __io_sqe_files_unregister(ctx);
8286 io_rsrc_node_switch(ctx, NULL);
8289 for (i = 0; i < ctx->nr_user_files; i++) {
8290 file = io_file_from_index(ctx, i);
8294 io_free_file_tables(&ctx->file_table);
8295 ctx->nr_user_files = 0;
8297 io_rsrc_data_free(ctx->file_data);
8298 ctx->file_data = NULL;
8302 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8305 #if defined(CONFIG_UNIX)
8306 struct sock *sock = ctx->ring_sock->sk;
8307 struct sk_buff_head *head = &sock->sk_receive_queue;
8308 struct sk_buff *skb;
8311 * See if we can merge this file into an existing skb SCM_RIGHTS
8312 * file set. If there's no room, fall back to allocating a new skb
8313 * and filling it in.
8315 spin_lock_irq(&head->lock);
8316 skb = skb_peek(head);
8318 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8320 if (fpl->count < SCM_MAX_FD) {
8321 __skb_unlink(skb, head);
8322 spin_unlock_irq(&head->lock);
8323 fpl->fp[fpl->count] = get_file(file);
8324 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8326 spin_lock_irq(&head->lock);
8327 __skb_queue_head(head, skb);
8332 spin_unlock_irq(&head->lock);
8339 return __io_sqe_files_scm(ctx, 1, index);
8345 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8346 struct io_rsrc_node *node, void *rsrc)
8348 struct io_rsrc_put *prsrc;
8350 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8354 prsrc->tag = *io_get_tag_slot(data, idx);
8356 list_add(&prsrc->list, &node->rsrc_list);
8360 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8361 unsigned int issue_flags, u32 slot_index)
8363 struct io_ring_ctx *ctx = req->ctx;
8364 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8365 bool needs_switch = false;
8366 struct io_fixed_file *file_slot;
8369 io_ring_submit_lock(ctx, !force_nonblock);
8370 if (file->f_op == &io_uring_fops)
8373 if (!ctx->file_data)
8376 if (slot_index >= ctx->nr_user_files)
8379 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8380 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8382 if (file_slot->file_ptr) {
8383 struct file *old_file;
8385 ret = io_rsrc_node_switch_start(ctx);
8389 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8390 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8391 ctx->rsrc_node, old_file);
8394 file_slot->file_ptr = 0;
8395 needs_switch = true;
8398 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8399 io_fixed_file_set(file_slot, file);
8400 ret = io_sqe_file_register(ctx, file, slot_index);
8402 file_slot->file_ptr = 0;
8409 io_rsrc_node_switch(ctx, ctx->file_data);
8410 io_ring_submit_unlock(ctx, !force_nonblock);
8416 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8418 unsigned int offset = req->close.file_slot - 1;
8419 struct io_ring_ctx *ctx = req->ctx;
8420 struct io_fixed_file *file_slot;
8424 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8426 if (unlikely(!ctx->file_data))
8429 if (offset >= ctx->nr_user_files)
8431 ret = io_rsrc_node_switch_start(ctx);
8435 i = array_index_nospec(offset, ctx->nr_user_files);
8436 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8438 if (!file_slot->file_ptr)
8441 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8442 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8446 file_slot->file_ptr = 0;
8447 io_rsrc_node_switch(ctx, ctx->file_data);
8450 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8454 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8455 struct io_uring_rsrc_update2 *up,
8458 u64 __user *tags = u64_to_user_ptr(up->tags);
8459 __s32 __user *fds = u64_to_user_ptr(up->data);
8460 struct io_rsrc_data *data = ctx->file_data;
8461 struct io_fixed_file *file_slot;
8465 bool needs_switch = false;
8467 if (!ctx->file_data)
8469 if (up->offset + nr_args > ctx->nr_user_files)
8472 for (done = 0; done < nr_args; done++) {
8475 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8476 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8480 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8484 if (fd == IORING_REGISTER_FILES_SKIP)
8487 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8488 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8490 if (file_slot->file_ptr) {
8491 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8492 err = io_queue_rsrc_removal(data, up->offset + done,
8493 ctx->rsrc_node, file);
8496 file_slot->file_ptr = 0;
8497 needs_switch = true;
8506 * Don't allow io_uring instances to be registered. If
8507 * UNIX isn't enabled, then this causes a reference
8508 * cycle and this instance can never get freed. If UNIX
8509 * is enabled we'll handle it just fine, but there's
8510 * still no point in allowing a ring fd as it doesn't
8511 * support regular read/write anyway.
8513 if (file->f_op == &io_uring_fops) {
8518 *io_get_tag_slot(data, up->offset + done) = tag;
8519 io_fixed_file_set(file_slot, file);
8520 err = io_sqe_file_register(ctx, file, i);
8522 file_slot->file_ptr = 0;
8530 io_rsrc_node_switch(ctx, data);
8531 return done ? done : err;
8534 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8535 struct task_struct *task)
8537 struct io_wq_hash *hash;
8538 struct io_wq_data data;
8539 unsigned int concurrency;
8541 mutex_lock(&ctx->uring_lock);
8542 hash = ctx->hash_map;
8544 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8546 mutex_unlock(&ctx->uring_lock);
8547 return ERR_PTR(-ENOMEM);
8549 refcount_set(&hash->refs, 1);
8550 init_waitqueue_head(&hash->wait);
8551 ctx->hash_map = hash;
8553 mutex_unlock(&ctx->uring_lock);
8557 data.free_work = io_wq_free_work;
8558 data.do_work = io_wq_submit_work;
8560 /* Do QD, or 4 * CPUS, whatever is smallest */
8561 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8563 return io_wq_create(concurrency, &data);
8566 static int io_uring_alloc_task_context(struct task_struct *task,
8567 struct io_ring_ctx *ctx)
8569 struct io_uring_task *tctx;
8572 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8573 if (unlikely(!tctx))
8576 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8577 if (unlikely(ret)) {
8582 tctx->io_wq = io_init_wq_offload(ctx, task);
8583 if (IS_ERR(tctx->io_wq)) {
8584 ret = PTR_ERR(tctx->io_wq);
8585 percpu_counter_destroy(&tctx->inflight);
8591 init_waitqueue_head(&tctx->wait);
8592 atomic_set(&tctx->in_idle, 0);
8593 atomic_set(&tctx->inflight_tracked, 0);
8594 task->io_uring = tctx;
8595 spin_lock_init(&tctx->task_lock);
8596 INIT_WQ_LIST(&tctx->task_list);
8597 init_task_work(&tctx->task_work, tctx_task_work);
8601 void __io_uring_free(struct task_struct *tsk)
8603 struct io_uring_task *tctx = tsk->io_uring;
8605 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8606 WARN_ON_ONCE(tctx->io_wq);
8607 WARN_ON_ONCE(tctx->cached_refs);
8609 percpu_counter_destroy(&tctx->inflight);
8611 tsk->io_uring = NULL;
8614 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8615 struct io_uring_params *p)
8619 /* Retain compatibility with failing for an invalid attach attempt */
8620 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8621 IORING_SETUP_ATTACH_WQ) {
8624 f = fdget(p->wq_fd);
8627 if (f.file->f_op != &io_uring_fops) {
8633 if (ctx->flags & IORING_SETUP_SQPOLL) {
8634 struct task_struct *tsk;
8635 struct io_sq_data *sqd;
8638 sqd = io_get_sq_data(p, &attached);
8644 ctx->sq_creds = get_current_cred();
8646 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8647 if (!ctx->sq_thread_idle)
8648 ctx->sq_thread_idle = HZ;
8650 io_sq_thread_park(sqd);
8651 list_add(&ctx->sqd_list, &sqd->ctx_list);
8652 io_sqd_update_thread_idle(sqd);
8653 /* don't attach to a dying SQPOLL thread, would be racy */
8654 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8655 io_sq_thread_unpark(sqd);
8662 if (p->flags & IORING_SETUP_SQ_AFF) {
8663 int cpu = p->sq_thread_cpu;
8666 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8673 sqd->task_pid = current->pid;
8674 sqd->task_tgid = current->tgid;
8675 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8682 ret = io_uring_alloc_task_context(tsk, ctx);
8683 wake_up_new_task(tsk);
8686 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8687 /* Can't have SQ_AFF without SQPOLL */
8694 complete(&ctx->sq_data->exited);
8696 io_sq_thread_finish(ctx);
8700 static inline void __io_unaccount_mem(struct user_struct *user,
8701 unsigned long nr_pages)
8703 atomic_long_sub(nr_pages, &user->locked_vm);
8706 static inline int __io_account_mem(struct user_struct *user,
8707 unsigned long nr_pages)
8709 unsigned long page_limit, cur_pages, new_pages;
8711 /* Don't allow more pages than we can safely lock */
8712 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8715 cur_pages = atomic_long_read(&user->locked_vm);
8716 new_pages = cur_pages + nr_pages;
8717 if (new_pages > page_limit)
8719 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8720 new_pages) != cur_pages);
8725 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8728 __io_unaccount_mem(ctx->user, nr_pages);
8730 if (ctx->mm_account)
8731 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8734 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8739 ret = __io_account_mem(ctx->user, nr_pages);
8744 if (ctx->mm_account)
8745 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8750 static void io_mem_free(void *ptr)
8757 page = virt_to_head_page(ptr);
8758 if (put_page_testzero(page))
8759 free_compound_page(page);
8762 static void *io_mem_alloc(size_t size)
8764 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8765 __GFP_NORETRY | __GFP_ACCOUNT;
8767 return (void *) __get_free_pages(gfp_flags, get_order(size));
8770 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8773 struct io_rings *rings;
8774 size_t off, sq_array_size;
8776 off = struct_size(rings, cqes, cq_entries);
8777 if (off == SIZE_MAX)
8781 off = ALIGN(off, SMP_CACHE_BYTES);
8789 sq_array_size = array_size(sizeof(u32), sq_entries);
8790 if (sq_array_size == SIZE_MAX)
8793 if (check_add_overflow(off, sq_array_size, &off))
8799 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8801 struct io_mapped_ubuf *imu = *slot;
8804 if (imu != ctx->dummy_ubuf) {
8805 for (i = 0; i < imu->nr_bvecs; i++)
8806 unpin_user_page(imu->bvec[i].bv_page);
8807 if (imu->acct_pages)
8808 io_unaccount_mem(ctx, imu->acct_pages);
8814 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8816 io_buffer_unmap(ctx, &prsrc->buf);
8820 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8824 for (i = 0; i < ctx->nr_user_bufs; i++)
8825 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8826 kfree(ctx->user_bufs);
8827 io_rsrc_data_free(ctx->buf_data);
8828 ctx->user_bufs = NULL;
8829 ctx->buf_data = NULL;
8830 ctx->nr_user_bufs = 0;
8833 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8840 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8842 __io_sqe_buffers_unregister(ctx);
8846 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8847 void __user *arg, unsigned index)
8849 struct iovec __user *src;
8851 #ifdef CONFIG_COMPAT
8853 struct compat_iovec __user *ciovs;
8854 struct compat_iovec ciov;
8856 ciovs = (struct compat_iovec __user *) arg;
8857 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8860 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8861 dst->iov_len = ciov.iov_len;
8865 src = (struct iovec __user *) arg;
8866 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8872 * Not super efficient, but this is just a registration time. And we do cache
8873 * the last compound head, so generally we'll only do a full search if we don't
8876 * We check if the given compound head page has already been accounted, to
8877 * avoid double accounting it. This allows us to account the full size of the
8878 * page, not just the constituent pages of a huge page.
8880 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8881 int nr_pages, struct page *hpage)
8885 /* check current page array */
8886 for (i = 0; i < nr_pages; i++) {
8887 if (!PageCompound(pages[i]))
8889 if (compound_head(pages[i]) == hpage)
8893 /* check previously registered pages */
8894 for (i = 0; i < ctx->nr_user_bufs; i++) {
8895 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8897 for (j = 0; j < imu->nr_bvecs; j++) {
8898 if (!PageCompound(imu->bvec[j].bv_page))
8900 if (compound_head(imu->bvec[j].bv_page) == hpage)
8908 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8909 int nr_pages, struct io_mapped_ubuf *imu,
8910 struct page **last_hpage)
8914 imu->acct_pages = 0;
8915 for (i = 0; i < nr_pages; i++) {
8916 if (!PageCompound(pages[i])) {
8921 hpage = compound_head(pages[i]);
8922 if (hpage == *last_hpage)
8924 *last_hpage = hpage;
8925 if (headpage_already_acct(ctx, pages, i, hpage))
8927 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8931 if (!imu->acct_pages)
8934 ret = io_account_mem(ctx, imu->acct_pages);
8936 imu->acct_pages = 0;
8940 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8941 struct io_mapped_ubuf **pimu,
8942 struct page **last_hpage)
8944 struct io_mapped_ubuf *imu = NULL;
8945 struct vm_area_struct **vmas = NULL;
8946 struct page **pages = NULL;
8947 unsigned long off, start, end, ubuf;
8949 int ret, pret, nr_pages, i;
8951 if (!iov->iov_base) {
8952 *pimu = ctx->dummy_ubuf;
8956 ubuf = (unsigned long) iov->iov_base;
8957 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8958 start = ubuf >> PAGE_SHIFT;
8959 nr_pages = end - start;
8964 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8968 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8973 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8978 mmap_read_lock(current->mm);
8979 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8981 if (pret == nr_pages) {
8982 /* don't support file backed memory */
8983 for (i = 0; i < nr_pages; i++) {
8984 struct vm_area_struct *vma = vmas[i];
8986 if (vma_is_shmem(vma))
8989 !is_file_hugepages(vma->vm_file)) {
8995 ret = pret < 0 ? pret : -EFAULT;
8997 mmap_read_unlock(current->mm);
9000 * if we did partial map, or found file backed vmas,
9001 * release any pages we did get
9004 unpin_user_pages(pages, pret);
9008 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9010 unpin_user_pages(pages, pret);
9014 off = ubuf & ~PAGE_MASK;
9015 size = iov->iov_len;
9016 for (i = 0; i < nr_pages; i++) {
9019 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9020 imu->bvec[i].bv_page = pages[i];
9021 imu->bvec[i].bv_len = vec_len;
9022 imu->bvec[i].bv_offset = off;
9026 /* store original address for later verification */
9028 imu->ubuf_end = ubuf + iov->iov_len;
9029 imu->nr_bvecs = nr_pages;
9040 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9042 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9043 return ctx->user_bufs ? 0 : -ENOMEM;
9046 static int io_buffer_validate(struct iovec *iov)
9048 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9051 * Don't impose further limits on the size and buffer
9052 * constraints here, we'll -EINVAL later when IO is
9053 * submitted if they are wrong.
9056 return iov->iov_len ? -EFAULT : 0;
9060 /* arbitrary limit, but we need something */
9061 if (iov->iov_len > SZ_1G)
9064 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9070 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9071 unsigned int nr_args, u64 __user *tags)
9073 struct page *last_hpage = NULL;
9074 struct io_rsrc_data *data;
9080 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9082 ret = io_rsrc_node_switch_start(ctx);
9085 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9088 ret = io_buffers_map_alloc(ctx, nr_args);
9090 io_rsrc_data_free(data);
9094 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9095 ret = io_copy_iov(ctx, &iov, arg, i);
9098 ret = io_buffer_validate(&iov);
9101 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9106 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9112 WARN_ON_ONCE(ctx->buf_data);
9114 ctx->buf_data = data;
9116 __io_sqe_buffers_unregister(ctx);
9118 io_rsrc_node_switch(ctx, NULL);
9122 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9123 struct io_uring_rsrc_update2 *up,
9124 unsigned int nr_args)
9126 u64 __user *tags = u64_to_user_ptr(up->tags);
9127 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9128 struct page *last_hpage = NULL;
9129 bool needs_switch = false;
9135 if (up->offset + nr_args > ctx->nr_user_bufs)
9138 for (done = 0; done < nr_args; done++) {
9139 struct io_mapped_ubuf *imu;
9140 int offset = up->offset + done;
9143 err = io_copy_iov(ctx, &iov, iovs, done);
9146 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9150 err = io_buffer_validate(&iov);
9153 if (!iov.iov_base && tag) {
9157 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9161 i = array_index_nospec(offset, ctx->nr_user_bufs);
9162 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9163 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9164 ctx->rsrc_node, ctx->user_bufs[i]);
9165 if (unlikely(err)) {
9166 io_buffer_unmap(ctx, &imu);
9169 ctx->user_bufs[i] = NULL;
9170 needs_switch = true;
9173 ctx->user_bufs[i] = imu;
9174 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9178 io_rsrc_node_switch(ctx, ctx->buf_data);
9179 return done ? done : err;
9182 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9184 __s32 __user *fds = arg;
9190 if (copy_from_user(&fd, fds, sizeof(*fds)))
9193 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9194 if (IS_ERR(ctx->cq_ev_fd)) {
9195 int ret = PTR_ERR(ctx->cq_ev_fd);
9197 ctx->cq_ev_fd = NULL;
9204 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9206 if (ctx->cq_ev_fd) {
9207 eventfd_ctx_put(ctx->cq_ev_fd);
9208 ctx->cq_ev_fd = NULL;
9215 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9217 struct io_buffer *buf;
9218 unsigned long index;
9220 xa_for_each(&ctx->io_buffers, index, buf) {
9221 __io_remove_buffers(ctx, buf, index, -1U);
9226 static void io_req_cache_free(struct list_head *list)
9228 struct io_kiocb *req, *nxt;
9230 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9231 list_del(&req->inflight_entry);
9232 kmem_cache_free(req_cachep, req);
9236 static void io_req_caches_free(struct io_ring_ctx *ctx)
9238 struct io_submit_state *state = &ctx->submit_state;
9240 mutex_lock(&ctx->uring_lock);
9242 if (state->free_reqs) {
9243 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9244 state->free_reqs = 0;
9247 io_flush_cached_locked_reqs(ctx, state);
9248 io_req_cache_free(&state->free_list);
9249 mutex_unlock(&ctx->uring_lock);
9252 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9254 if (data && !atomic_dec_and_test(&data->refs))
9255 wait_for_completion(&data->done);
9258 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9260 io_sq_thread_finish(ctx);
9262 if (ctx->mm_account) {
9263 mmdrop(ctx->mm_account);
9264 ctx->mm_account = NULL;
9267 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9268 io_wait_rsrc_data(ctx->buf_data);
9269 io_wait_rsrc_data(ctx->file_data);
9271 mutex_lock(&ctx->uring_lock);
9273 __io_sqe_buffers_unregister(ctx);
9275 __io_sqe_files_unregister(ctx);
9277 __io_cqring_overflow_flush(ctx, true);
9278 mutex_unlock(&ctx->uring_lock);
9279 io_eventfd_unregister(ctx);
9280 io_destroy_buffers(ctx);
9282 put_cred(ctx->sq_creds);
9284 /* there are no registered resources left, nobody uses it */
9286 io_rsrc_node_destroy(ctx->rsrc_node);
9287 if (ctx->rsrc_backup_node)
9288 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9289 flush_delayed_work(&ctx->rsrc_put_work);
9291 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9292 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9294 #if defined(CONFIG_UNIX)
9295 if (ctx->ring_sock) {
9296 ctx->ring_sock->file = NULL; /* so that iput() is called */
9297 sock_release(ctx->ring_sock);
9300 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9302 io_mem_free(ctx->rings);
9303 io_mem_free(ctx->sq_sqes);
9305 percpu_ref_exit(&ctx->refs);
9306 free_uid(ctx->user);
9307 io_req_caches_free(ctx);
9309 io_wq_put_hash(ctx->hash_map);
9310 kfree(ctx->cancel_hash);
9311 kfree(ctx->dummy_ubuf);
9315 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9317 struct io_ring_ctx *ctx = file->private_data;
9320 poll_wait(file, &ctx->poll_wait, wait);
9322 * synchronizes with barrier from wq_has_sleeper call in
9326 if (!io_sqring_full(ctx))
9327 mask |= EPOLLOUT | EPOLLWRNORM;
9330 * Don't flush cqring overflow list here, just do a simple check.
9331 * Otherwise there could possible be ABBA deadlock:
9334 * lock(&ctx->uring_lock);
9336 * lock(&ctx->uring_lock);
9339 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9340 * pushs them to do the flush.
9342 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9343 mask |= EPOLLIN | EPOLLRDNORM;
9348 static int io_uring_fasync(int fd, struct file *file, int on)
9350 struct io_ring_ctx *ctx = file->private_data;
9352 return fasync_helper(fd, file, on, &ctx->cq_fasync);
9355 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9357 const struct cred *creds;
9359 creds = xa_erase(&ctx->personalities, id);
9368 struct io_tctx_exit {
9369 struct callback_head task_work;
9370 struct completion completion;
9371 struct io_ring_ctx *ctx;
9374 static void io_tctx_exit_cb(struct callback_head *cb)
9376 struct io_uring_task *tctx = current->io_uring;
9377 struct io_tctx_exit *work;
9379 work = container_of(cb, struct io_tctx_exit, task_work);
9381 * When @in_idle, we're in cancellation and it's racy to remove the
9382 * node. It'll be removed by the end of cancellation, just ignore it.
9384 if (!atomic_read(&tctx->in_idle))
9385 io_uring_del_tctx_node((unsigned long)work->ctx);
9386 complete(&work->completion);
9389 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9391 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9393 return req->ctx == data;
9396 static void io_ring_exit_work(struct work_struct *work)
9398 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9399 unsigned long timeout = jiffies + HZ * 60 * 5;
9400 unsigned long interval = HZ / 20;
9401 struct io_tctx_exit exit;
9402 struct io_tctx_node *node;
9406 * If we're doing polled IO and end up having requests being
9407 * submitted async (out-of-line), then completions can come in while
9408 * we're waiting for refs to drop. We need to reap these manually,
9409 * as nobody else will be looking for them.
9412 io_uring_try_cancel_requests(ctx, NULL, true);
9414 struct io_sq_data *sqd = ctx->sq_data;
9415 struct task_struct *tsk;
9417 io_sq_thread_park(sqd);
9419 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9420 io_wq_cancel_cb(tsk->io_uring->io_wq,
9421 io_cancel_ctx_cb, ctx, true);
9422 io_sq_thread_unpark(sqd);
9425 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9426 /* there is little hope left, don't run it too often */
9429 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9431 init_completion(&exit.completion);
9432 init_task_work(&exit.task_work, io_tctx_exit_cb);
9435 * Some may use context even when all refs and requests have been put,
9436 * and they are free to do so while still holding uring_lock or
9437 * completion_lock, see io_req_task_submit(). Apart from other work,
9438 * this lock/unlock section also waits them to finish.
9440 mutex_lock(&ctx->uring_lock);
9441 while (!list_empty(&ctx->tctx_list)) {
9442 WARN_ON_ONCE(time_after(jiffies, timeout));
9444 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9446 /* don't spin on a single task if cancellation failed */
9447 list_rotate_left(&ctx->tctx_list);
9448 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9449 if (WARN_ON_ONCE(ret))
9451 wake_up_process(node->task);
9453 mutex_unlock(&ctx->uring_lock);
9454 wait_for_completion(&exit.completion);
9455 mutex_lock(&ctx->uring_lock);
9457 mutex_unlock(&ctx->uring_lock);
9458 spin_lock(&ctx->completion_lock);
9459 spin_unlock(&ctx->completion_lock);
9461 io_ring_ctx_free(ctx);
9464 /* Returns true if we found and killed one or more timeouts */
9465 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9468 struct io_kiocb *req, *tmp;
9471 spin_lock(&ctx->completion_lock);
9472 spin_lock_irq(&ctx->timeout_lock);
9473 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9474 if (io_match_task(req, tsk, cancel_all)) {
9475 io_kill_timeout(req, -ECANCELED);
9479 spin_unlock_irq(&ctx->timeout_lock);
9481 io_commit_cqring(ctx);
9482 spin_unlock(&ctx->completion_lock);
9484 io_cqring_ev_posted(ctx);
9485 return canceled != 0;
9488 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9490 unsigned long index;
9491 struct creds *creds;
9493 mutex_lock(&ctx->uring_lock);
9494 percpu_ref_kill(&ctx->refs);
9496 __io_cqring_overflow_flush(ctx, true);
9497 xa_for_each(&ctx->personalities, index, creds)
9498 io_unregister_personality(ctx, index);
9499 mutex_unlock(&ctx->uring_lock);
9501 io_kill_timeouts(ctx, NULL, true);
9502 io_poll_remove_all(ctx, NULL, true);
9504 /* if we failed setting up the ctx, we might not have any rings */
9505 io_iopoll_try_reap_events(ctx);
9507 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9509 * Use system_unbound_wq to avoid spawning tons of event kworkers
9510 * if we're exiting a ton of rings at the same time. It just adds
9511 * noise and overhead, there's no discernable change in runtime
9512 * over using system_wq.
9514 queue_work(system_unbound_wq, &ctx->exit_work);
9517 static int io_uring_release(struct inode *inode, struct file *file)
9519 struct io_ring_ctx *ctx = file->private_data;
9521 file->private_data = NULL;
9522 io_ring_ctx_wait_and_kill(ctx);
9526 struct io_task_cancel {
9527 struct task_struct *task;
9531 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9533 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9534 struct io_task_cancel *cancel = data;
9537 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9538 struct io_ring_ctx *ctx = req->ctx;
9540 /* protect against races with linked timeouts */
9541 spin_lock(&ctx->completion_lock);
9542 ret = io_match_task(req, cancel->task, cancel->all);
9543 spin_unlock(&ctx->completion_lock);
9545 ret = io_match_task(req, cancel->task, cancel->all);
9550 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9551 struct task_struct *task, bool cancel_all)
9553 struct io_defer_entry *de;
9556 spin_lock(&ctx->completion_lock);
9557 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9558 if (io_match_task(de->req, task, cancel_all)) {
9559 list_cut_position(&list, &ctx->defer_list, &de->list);
9563 spin_unlock(&ctx->completion_lock);
9564 if (list_empty(&list))
9567 while (!list_empty(&list)) {
9568 de = list_first_entry(&list, struct io_defer_entry, list);
9569 list_del_init(&de->list);
9570 io_req_complete_failed(de->req, -ECANCELED);
9576 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9578 struct io_tctx_node *node;
9579 enum io_wq_cancel cret;
9582 mutex_lock(&ctx->uring_lock);
9583 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9584 struct io_uring_task *tctx = node->task->io_uring;
9587 * io_wq will stay alive while we hold uring_lock, because it's
9588 * killed after ctx nodes, which requires to take the lock.
9590 if (!tctx || !tctx->io_wq)
9592 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9593 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9595 mutex_unlock(&ctx->uring_lock);
9600 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9601 struct task_struct *task,
9604 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9605 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9608 enum io_wq_cancel cret;
9612 ret |= io_uring_try_cancel_iowq(ctx);
9613 } else if (tctx && tctx->io_wq) {
9615 * Cancels requests of all rings, not only @ctx, but
9616 * it's fine as the task is in exit/exec.
9618 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9620 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9623 /* SQPOLL thread does its own polling */
9624 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9625 (ctx->sq_data && ctx->sq_data->thread == current)) {
9626 while (!list_empty_careful(&ctx->iopoll_list)) {
9627 io_iopoll_try_reap_events(ctx);
9632 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9633 ret |= io_poll_remove_all(ctx, task, cancel_all);
9634 ret |= io_kill_timeouts(ctx, task, cancel_all);
9636 ret |= io_run_task_work();
9643 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9645 struct io_uring_task *tctx = current->io_uring;
9646 struct io_tctx_node *node;
9649 if (unlikely(!tctx)) {
9650 ret = io_uring_alloc_task_context(current, ctx);
9653 tctx = current->io_uring;
9655 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9656 node = kmalloc(sizeof(*node), GFP_KERNEL);
9660 node->task = current;
9662 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9669 mutex_lock(&ctx->uring_lock);
9670 list_add(&node->ctx_node, &ctx->tctx_list);
9671 mutex_unlock(&ctx->uring_lock);
9678 * Note that this task has used io_uring. We use it for cancelation purposes.
9680 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9682 struct io_uring_task *tctx = current->io_uring;
9684 if (likely(tctx && tctx->last == ctx))
9686 return __io_uring_add_tctx_node(ctx);
9690 * Remove this io_uring_file -> task mapping.
9692 static void io_uring_del_tctx_node(unsigned long index)
9694 struct io_uring_task *tctx = current->io_uring;
9695 struct io_tctx_node *node;
9699 node = xa_erase(&tctx->xa, index);
9703 WARN_ON_ONCE(current != node->task);
9704 WARN_ON_ONCE(list_empty(&node->ctx_node));
9706 mutex_lock(&node->ctx->uring_lock);
9707 list_del(&node->ctx_node);
9708 mutex_unlock(&node->ctx->uring_lock);
9710 if (tctx->last == node->ctx)
9715 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9717 struct io_wq *wq = tctx->io_wq;
9718 struct io_tctx_node *node;
9719 unsigned long index;
9721 xa_for_each(&tctx->xa, index, node) {
9722 io_uring_del_tctx_node(index);
9727 * Must be after io_uring_del_task_file() (removes nodes under
9728 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9730 io_wq_put_and_exit(wq);
9735 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9738 return atomic_read(&tctx->inflight_tracked);
9739 return percpu_counter_sum(&tctx->inflight);
9742 static void io_uring_drop_tctx_refs(struct task_struct *task)
9744 struct io_uring_task *tctx = task->io_uring;
9745 unsigned int refs = tctx->cached_refs;
9748 tctx->cached_refs = 0;
9749 percpu_counter_sub(&tctx->inflight, refs);
9750 put_task_struct_many(task, refs);
9755 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9756 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9758 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9760 struct io_uring_task *tctx = current->io_uring;
9761 struct io_ring_ctx *ctx;
9765 WARN_ON_ONCE(sqd && sqd->thread != current);
9767 if (!current->io_uring)
9770 io_wq_exit_start(tctx->io_wq);
9772 atomic_inc(&tctx->in_idle);
9774 io_uring_drop_tctx_refs(current);
9775 /* read completions before cancelations */
9776 inflight = tctx_inflight(tctx, !cancel_all);
9781 struct io_tctx_node *node;
9782 unsigned long index;
9784 xa_for_each(&tctx->xa, index, node) {
9785 /* sqpoll task will cancel all its requests */
9786 if (node->ctx->sq_data)
9788 io_uring_try_cancel_requests(node->ctx, current,
9792 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9793 io_uring_try_cancel_requests(ctx, current,
9797 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9798 io_uring_drop_tctx_refs(current);
9800 * If we've seen completions, retry without waiting. This
9801 * avoids a race where a completion comes in before we did
9802 * prepare_to_wait().
9804 if (inflight == tctx_inflight(tctx, !cancel_all))
9806 finish_wait(&tctx->wait, &wait);
9808 atomic_dec(&tctx->in_idle);
9810 io_uring_clean_tctx(tctx);
9812 /* for exec all current's requests should be gone, kill tctx */
9813 __io_uring_free(current);
9817 void __io_uring_cancel(bool cancel_all)
9819 io_uring_cancel_generic(cancel_all, NULL);
9822 static void *io_uring_validate_mmap_request(struct file *file,
9823 loff_t pgoff, size_t sz)
9825 struct io_ring_ctx *ctx = file->private_data;
9826 loff_t offset = pgoff << PAGE_SHIFT;
9831 case IORING_OFF_SQ_RING:
9832 case IORING_OFF_CQ_RING:
9835 case IORING_OFF_SQES:
9839 return ERR_PTR(-EINVAL);
9842 page = virt_to_head_page(ptr);
9843 if (sz > page_size(page))
9844 return ERR_PTR(-EINVAL);
9851 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9853 size_t sz = vma->vm_end - vma->vm_start;
9857 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9859 return PTR_ERR(ptr);
9861 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9862 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9865 #else /* !CONFIG_MMU */
9867 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9869 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9872 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9874 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9877 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9878 unsigned long addr, unsigned long len,
9879 unsigned long pgoff, unsigned long flags)
9883 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9885 return PTR_ERR(ptr);
9887 return (unsigned long) ptr;
9890 #endif /* !CONFIG_MMU */
9892 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9897 if (!io_sqring_full(ctx))
9899 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9901 if (!io_sqring_full(ctx))
9904 } while (!signal_pending(current));
9906 finish_wait(&ctx->sqo_sq_wait, &wait);
9910 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9911 struct __kernel_timespec __user **ts,
9912 const sigset_t __user **sig)
9914 struct io_uring_getevents_arg arg;
9917 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9918 * is just a pointer to the sigset_t.
9920 if (!(flags & IORING_ENTER_EXT_ARG)) {
9921 *sig = (const sigset_t __user *) argp;
9927 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9928 * timespec and sigset_t pointers if good.
9930 if (*argsz != sizeof(arg))
9932 if (copy_from_user(&arg, argp, sizeof(arg)))
9934 *sig = u64_to_user_ptr(arg.sigmask);
9935 *argsz = arg.sigmask_sz;
9936 *ts = u64_to_user_ptr(arg.ts);
9940 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9941 u32, min_complete, u32, flags, const void __user *, argp,
9944 struct io_ring_ctx *ctx;
9951 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9952 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9956 if (unlikely(!f.file))
9960 if (unlikely(f.file->f_op != &io_uring_fops))
9964 ctx = f.file->private_data;
9965 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9969 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9973 * For SQ polling, the thread will do all submissions and completions.
9974 * Just return the requested submit count, and wake the thread if
9978 if (ctx->flags & IORING_SETUP_SQPOLL) {
9979 io_cqring_overflow_flush(ctx);
9981 if (unlikely(ctx->sq_data->thread == NULL)) {
9985 if (flags & IORING_ENTER_SQ_WAKEUP)
9986 wake_up(&ctx->sq_data->wait);
9987 if (flags & IORING_ENTER_SQ_WAIT) {
9988 ret = io_sqpoll_wait_sq(ctx);
9992 submitted = to_submit;
9993 } else if (to_submit) {
9994 ret = io_uring_add_tctx_node(ctx);
9997 mutex_lock(&ctx->uring_lock);
9998 submitted = io_submit_sqes(ctx, to_submit);
9999 mutex_unlock(&ctx->uring_lock);
10001 if (submitted != to_submit)
10004 if (flags & IORING_ENTER_GETEVENTS) {
10005 const sigset_t __user *sig;
10006 struct __kernel_timespec __user *ts;
10008 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10012 min_complete = min(min_complete, ctx->cq_entries);
10015 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10016 * space applications don't need to do io completion events
10017 * polling again, they can rely on io_sq_thread to do polling
10018 * work, which can reduce cpu usage and uring_lock contention.
10020 if (ctx->flags & IORING_SETUP_IOPOLL &&
10021 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10022 ret = io_iopoll_check(ctx, min_complete);
10024 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10029 percpu_ref_put(&ctx->refs);
10032 return submitted ? submitted : ret;
10035 #ifdef CONFIG_PROC_FS
10036 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10037 const struct cred *cred)
10039 struct user_namespace *uns = seq_user_ns(m);
10040 struct group_info *gi;
10045 seq_printf(m, "%5d\n", id);
10046 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10047 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10048 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10049 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10050 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10051 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10052 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10053 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10054 seq_puts(m, "\n\tGroups:\t");
10055 gi = cred->group_info;
10056 for (g = 0; g < gi->ngroups; g++) {
10057 seq_put_decimal_ull(m, g ? " " : "",
10058 from_kgid_munged(uns, gi->gid[g]));
10060 seq_puts(m, "\n\tCapEff:\t");
10061 cap = cred->cap_effective;
10062 CAP_FOR_EACH_U32(__capi)
10063 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10068 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10070 struct io_sq_data *sq = NULL;
10075 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10076 * since fdinfo case grabs it in the opposite direction of normal use
10077 * cases. If we fail to get the lock, we just don't iterate any
10078 * structures that could be going away outside the io_uring mutex.
10080 has_lock = mutex_trylock(&ctx->uring_lock);
10082 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10088 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10089 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10090 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10091 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10092 struct file *f = io_file_from_index(ctx, i);
10095 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10097 seq_printf(m, "%5u: <none>\n", i);
10099 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10100 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10101 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10102 unsigned int len = buf->ubuf_end - buf->ubuf;
10104 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10106 if (has_lock && !xa_empty(&ctx->personalities)) {
10107 unsigned long index;
10108 const struct cred *cred;
10110 seq_printf(m, "Personalities:\n");
10111 xa_for_each(&ctx->personalities, index, cred)
10112 io_uring_show_cred(m, index, cred);
10114 seq_printf(m, "PollList:\n");
10115 spin_lock(&ctx->completion_lock);
10116 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10117 struct hlist_head *list = &ctx->cancel_hash[i];
10118 struct io_kiocb *req;
10120 hlist_for_each_entry(req, list, hash_node)
10121 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10122 req->task->task_works != NULL);
10124 spin_unlock(&ctx->completion_lock);
10126 mutex_unlock(&ctx->uring_lock);
10129 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10131 struct io_ring_ctx *ctx = f->private_data;
10133 if (percpu_ref_tryget(&ctx->refs)) {
10134 __io_uring_show_fdinfo(ctx, m);
10135 percpu_ref_put(&ctx->refs);
10140 static const struct file_operations io_uring_fops = {
10141 .release = io_uring_release,
10142 .mmap = io_uring_mmap,
10144 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10145 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10147 .poll = io_uring_poll,
10148 .fasync = io_uring_fasync,
10149 #ifdef CONFIG_PROC_FS
10150 .show_fdinfo = io_uring_show_fdinfo,
10154 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10155 struct io_uring_params *p)
10157 struct io_rings *rings;
10158 size_t size, sq_array_offset;
10160 /* make sure these are sane, as we already accounted them */
10161 ctx->sq_entries = p->sq_entries;
10162 ctx->cq_entries = p->cq_entries;
10164 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10165 if (size == SIZE_MAX)
10168 rings = io_mem_alloc(size);
10172 ctx->rings = rings;
10173 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10174 rings->sq_ring_mask = p->sq_entries - 1;
10175 rings->cq_ring_mask = p->cq_entries - 1;
10176 rings->sq_ring_entries = p->sq_entries;
10177 rings->cq_ring_entries = p->cq_entries;
10179 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10180 if (size == SIZE_MAX) {
10181 io_mem_free(ctx->rings);
10186 ctx->sq_sqes = io_mem_alloc(size);
10187 if (!ctx->sq_sqes) {
10188 io_mem_free(ctx->rings);
10196 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10200 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10204 ret = io_uring_add_tctx_node(ctx);
10209 fd_install(fd, file);
10214 * Allocate an anonymous fd, this is what constitutes the application
10215 * visible backing of an io_uring instance. The application mmaps this
10216 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10217 * we have to tie this fd to a socket for file garbage collection purposes.
10219 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10222 #if defined(CONFIG_UNIX)
10225 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10228 return ERR_PTR(ret);
10231 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10232 O_RDWR | O_CLOEXEC);
10233 #if defined(CONFIG_UNIX)
10234 if (IS_ERR(file)) {
10235 sock_release(ctx->ring_sock);
10236 ctx->ring_sock = NULL;
10238 ctx->ring_sock->file = file;
10244 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10245 struct io_uring_params __user *params)
10247 struct io_ring_ctx *ctx;
10253 if (entries > IORING_MAX_ENTRIES) {
10254 if (!(p->flags & IORING_SETUP_CLAMP))
10256 entries = IORING_MAX_ENTRIES;
10260 * Use twice as many entries for the CQ ring. It's possible for the
10261 * application to drive a higher depth than the size of the SQ ring,
10262 * since the sqes are only used at submission time. This allows for
10263 * some flexibility in overcommitting a bit. If the application has
10264 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10265 * of CQ ring entries manually.
10267 p->sq_entries = roundup_pow_of_two(entries);
10268 if (p->flags & IORING_SETUP_CQSIZE) {
10270 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10271 * to a power-of-two, if it isn't already. We do NOT impose
10272 * any cq vs sq ring sizing.
10274 if (!p->cq_entries)
10276 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10277 if (!(p->flags & IORING_SETUP_CLAMP))
10279 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10281 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10282 if (p->cq_entries < p->sq_entries)
10285 p->cq_entries = 2 * p->sq_entries;
10288 ctx = io_ring_ctx_alloc(p);
10291 ctx->compat = in_compat_syscall();
10292 if (!capable(CAP_IPC_LOCK))
10293 ctx->user = get_uid(current_user());
10296 * This is just grabbed for accounting purposes. When a process exits,
10297 * the mm is exited and dropped before the files, hence we need to hang
10298 * on to this mm purely for the purposes of being able to unaccount
10299 * memory (locked/pinned vm). It's not used for anything else.
10301 mmgrab(current->mm);
10302 ctx->mm_account = current->mm;
10304 ret = io_allocate_scq_urings(ctx, p);
10308 ret = io_sq_offload_create(ctx, p);
10311 /* always set a rsrc node */
10312 ret = io_rsrc_node_switch_start(ctx);
10315 io_rsrc_node_switch(ctx, NULL);
10317 memset(&p->sq_off, 0, sizeof(p->sq_off));
10318 p->sq_off.head = offsetof(struct io_rings, sq.head);
10319 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10320 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10321 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10322 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10323 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10324 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10326 memset(&p->cq_off, 0, sizeof(p->cq_off));
10327 p->cq_off.head = offsetof(struct io_rings, cq.head);
10328 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10329 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10330 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10331 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10332 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10333 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10335 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10336 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10337 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10338 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10339 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10340 IORING_FEAT_RSRC_TAGS;
10342 if (copy_to_user(params, p, sizeof(*p))) {
10347 file = io_uring_get_file(ctx);
10348 if (IS_ERR(file)) {
10349 ret = PTR_ERR(file);
10354 * Install ring fd as the very last thing, so we don't risk someone
10355 * having closed it before we finish setup
10357 ret = io_uring_install_fd(ctx, file);
10359 /* fput will clean it up */
10364 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10367 io_ring_ctx_wait_and_kill(ctx);
10372 * Sets up an aio uring context, and returns the fd. Applications asks for a
10373 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10374 * params structure passed in.
10376 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10378 struct io_uring_params p;
10381 if (copy_from_user(&p, params, sizeof(p)))
10383 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10388 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10389 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10390 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10391 IORING_SETUP_R_DISABLED))
10394 return io_uring_create(entries, &p, params);
10397 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10398 struct io_uring_params __user *, params)
10400 return io_uring_setup(entries, params);
10403 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10405 struct io_uring_probe *p;
10409 size = struct_size(p, ops, nr_args);
10410 if (size == SIZE_MAX)
10412 p = kzalloc(size, GFP_KERNEL);
10417 if (copy_from_user(p, arg, size))
10420 if (memchr_inv(p, 0, size))
10423 p->last_op = IORING_OP_LAST - 1;
10424 if (nr_args > IORING_OP_LAST)
10425 nr_args = IORING_OP_LAST;
10427 for (i = 0; i < nr_args; i++) {
10429 if (!io_op_defs[i].not_supported)
10430 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10435 if (copy_to_user(arg, p, size))
10442 static int io_register_personality(struct io_ring_ctx *ctx)
10444 const struct cred *creds;
10448 creds = get_current_cred();
10450 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10451 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10459 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10460 unsigned int nr_args)
10462 struct io_uring_restriction *res;
10466 /* Restrictions allowed only if rings started disabled */
10467 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10470 /* We allow only a single restrictions registration */
10471 if (ctx->restrictions.registered)
10474 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10477 size = array_size(nr_args, sizeof(*res));
10478 if (size == SIZE_MAX)
10481 res = memdup_user(arg, size);
10483 return PTR_ERR(res);
10487 for (i = 0; i < nr_args; i++) {
10488 switch (res[i].opcode) {
10489 case IORING_RESTRICTION_REGISTER_OP:
10490 if (res[i].register_op >= IORING_REGISTER_LAST) {
10495 __set_bit(res[i].register_op,
10496 ctx->restrictions.register_op);
10498 case IORING_RESTRICTION_SQE_OP:
10499 if (res[i].sqe_op >= IORING_OP_LAST) {
10504 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10506 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10507 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10509 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10510 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10519 /* Reset all restrictions if an error happened */
10521 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10523 ctx->restrictions.registered = true;
10529 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10531 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10534 if (ctx->restrictions.registered)
10535 ctx->restricted = 1;
10537 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10538 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10539 wake_up(&ctx->sq_data->wait);
10543 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10544 struct io_uring_rsrc_update2 *up,
10552 if (check_add_overflow(up->offset, nr_args, &tmp))
10554 err = io_rsrc_node_switch_start(ctx);
10559 case IORING_RSRC_FILE:
10560 return __io_sqe_files_update(ctx, up, nr_args);
10561 case IORING_RSRC_BUFFER:
10562 return __io_sqe_buffers_update(ctx, up, nr_args);
10567 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10570 struct io_uring_rsrc_update2 up;
10574 memset(&up, 0, sizeof(up));
10575 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10577 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10580 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10581 unsigned size, unsigned type)
10583 struct io_uring_rsrc_update2 up;
10585 if (size != sizeof(up))
10587 if (copy_from_user(&up, arg, sizeof(up)))
10589 if (!up.nr || up.resv)
10591 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10594 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10595 unsigned int size, unsigned int type)
10597 struct io_uring_rsrc_register rr;
10599 /* keep it extendible */
10600 if (size != sizeof(rr))
10603 memset(&rr, 0, sizeof(rr));
10604 if (copy_from_user(&rr, arg, size))
10606 if (!rr.nr || rr.resv || rr.resv2)
10610 case IORING_RSRC_FILE:
10611 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10612 rr.nr, u64_to_user_ptr(rr.tags));
10613 case IORING_RSRC_BUFFER:
10614 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10615 rr.nr, u64_to_user_ptr(rr.tags));
10620 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10623 struct io_uring_task *tctx = current->io_uring;
10624 cpumask_var_t new_mask;
10627 if (!tctx || !tctx->io_wq)
10630 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10633 cpumask_clear(new_mask);
10634 if (len > cpumask_size())
10635 len = cpumask_size();
10637 if (copy_from_user(new_mask, arg, len)) {
10638 free_cpumask_var(new_mask);
10642 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10643 free_cpumask_var(new_mask);
10647 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10649 struct io_uring_task *tctx = current->io_uring;
10651 if (!tctx || !tctx->io_wq)
10654 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10657 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10660 struct io_uring_task *tctx = NULL;
10661 struct io_sq_data *sqd = NULL;
10662 __u32 new_count[2];
10665 if (copy_from_user(new_count, arg, sizeof(new_count)))
10667 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10668 if (new_count[i] > INT_MAX)
10671 if (ctx->flags & IORING_SETUP_SQPOLL) {
10672 sqd = ctx->sq_data;
10675 * Observe the correct sqd->lock -> ctx->uring_lock
10676 * ordering. Fine to drop uring_lock here, we hold
10677 * a ref to the ctx.
10679 refcount_inc(&sqd->refs);
10680 mutex_unlock(&ctx->uring_lock);
10681 mutex_lock(&sqd->lock);
10682 mutex_lock(&ctx->uring_lock);
10684 tctx = sqd->thread->io_uring;
10687 tctx = current->io_uring;
10691 if (!tctx || !tctx->io_wq)
10694 ret = io_wq_max_workers(tctx->io_wq, new_count);
10699 mutex_unlock(&sqd->lock);
10700 io_put_sq_data(sqd);
10703 if (copy_to_user(arg, new_count, sizeof(new_count)))
10709 mutex_unlock(&sqd->lock);
10710 io_put_sq_data(sqd);
10715 static bool io_register_op_must_quiesce(int op)
10718 case IORING_REGISTER_BUFFERS:
10719 case IORING_UNREGISTER_BUFFERS:
10720 case IORING_REGISTER_FILES:
10721 case IORING_UNREGISTER_FILES:
10722 case IORING_REGISTER_FILES_UPDATE:
10723 case IORING_REGISTER_PROBE:
10724 case IORING_REGISTER_PERSONALITY:
10725 case IORING_UNREGISTER_PERSONALITY:
10726 case IORING_REGISTER_FILES2:
10727 case IORING_REGISTER_FILES_UPDATE2:
10728 case IORING_REGISTER_BUFFERS2:
10729 case IORING_REGISTER_BUFFERS_UPDATE:
10730 case IORING_REGISTER_IOWQ_AFF:
10731 case IORING_UNREGISTER_IOWQ_AFF:
10732 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10739 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10743 percpu_ref_kill(&ctx->refs);
10746 * Drop uring mutex before waiting for references to exit. If another
10747 * thread is currently inside io_uring_enter() it might need to grab the
10748 * uring_lock to make progress. If we hold it here across the drain
10749 * wait, then we can deadlock. It's safe to drop the mutex here, since
10750 * no new references will come in after we've killed the percpu ref.
10752 mutex_unlock(&ctx->uring_lock);
10754 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10757 ret = io_run_task_work_sig();
10758 } while (ret >= 0);
10759 mutex_lock(&ctx->uring_lock);
10762 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10766 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10767 void __user *arg, unsigned nr_args)
10768 __releases(ctx->uring_lock)
10769 __acquires(ctx->uring_lock)
10774 * We're inside the ring mutex, if the ref is already dying, then
10775 * someone else killed the ctx or is already going through
10776 * io_uring_register().
10778 if (percpu_ref_is_dying(&ctx->refs))
10781 if (ctx->restricted) {
10782 if (opcode >= IORING_REGISTER_LAST)
10784 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10785 if (!test_bit(opcode, ctx->restrictions.register_op))
10789 if (io_register_op_must_quiesce(opcode)) {
10790 ret = io_ctx_quiesce(ctx);
10796 case IORING_REGISTER_BUFFERS:
10797 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10799 case IORING_UNREGISTER_BUFFERS:
10801 if (arg || nr_args)
10803 ret = io_sqe_buffers_unregister(ctx);
10805 case IORING_REGISTER_FILES:
10806 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10808 case IORING_UNREGISTER_FILES:
10810 if (arg || nr_args)
10812 ret = io_sqe_files_unregister(ctx);
10814 case IORING_REGISTER_FILES_UPDATE:
10815 ret = io_register_files_update(ctx, arg, nr_args);
10817 case IORING_REGISTER_EVENTFD:
10818 case IORING_REGISTER_EVENTFD_ASYNC:
10822 ret = io_eventfd_register(ctx, arg);
10825 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10826 ctx->eventfd_async = 1;
10828 ctx->eventfd_async = 0;
10830 case IORING_UNREGISTER_EVENTFD:
10832 if (arg || nr_args)
10834 ret = io_eventfd_unregister(ctx);
10836 case IORING_REGISTER_PROBE:
10838 if (!arg || nr_args > 256)
10840 ret = io_probe(ctx, arg, nr_args);
10842 case IORING_REGISTER_PERSONALITY:
10844 if (arg || nr_args)
10846 ret = io_register_personality(ctx);
10848 case IORING_UNREGISTER_PERSONALITY:
10852 ret = io_unregister_personality(ctx, nr_args);
10854 case IORING_REGISTER_ENABLE_RINGS:
10856 if (arg || nr_args)
10858 ret = io_register_enable_rings(ctx);
10860 case IORING_REGISTER_RESTRICTIONS:
10861 ret = io_register_restrictions(ctx, arg, nr_args);
10863 case IORING_REGISTER_FILES2:
10864 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10866 case IORING_REGISTER_FILES_UPDATE2:
10867 ret = io_register_rsrc_update(ctx, arg, nr_args,
10870 case IORING_REGISTER_BUFFERS2:
10871 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10873 case IORING_REGISTER_BUFFERS_UPDATE:
10874 ret = io_register_rsrc_update(ctx, arg, nr_args,
10875 IORING_RSRC_BUFFER);
10877 case IORING_REGISTER_IOWQ_AFF:
10879 if (!arg || !nr_args)
10881 ret = io_register_iowq_aff(ctx, arg, nr_args);
10883 case IORING_UNREGISTER_IOWQ_AFF:
10885 if (arg || nr_args)
10887 ret = io_unregister_iowq_aff(ctx);
10889 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10891 if (!arg || nr_args != 2)
10893 ret = io_register_iowq_max_workers(ctx, arg);
10900 if (io_register_op_must_quiesce(opcode)) {
10901 /* bring the ctx back to life */
10902 percpu_ref_reinit(&ctx->refs);
10903 reinit_completion(&ctx->ref_comp);
10908 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10909 void __user *, arg, unsigned int, nr_args)
10911 struct io_ring_ctx *ctx;
10920 if (f.file->f_op != &io_uring_fops)
10923 ctx = f.file->private_data;
10925 io_run_task_work();
10927 mutex_lock(&ctx->uring_lock);
10928 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10929 mutex_unlock(&ctx->uring_lock);
10930 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10931 ctx->cq_ev_fd != NULL, ret);
10937 static int __init io_uring_init(void)
10939 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10940 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10941 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10944 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10945 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10946 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10947 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10948 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10949 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10950 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10951 BUILD_BUG_SQE_ELEM(8, __u64, off);
10952 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10953 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10954 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10955 BUILD_BUG_SQE_ELEM(24, __u32, len);
10956 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10957 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10958 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10959 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10960 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10961 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10962 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10963 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10964 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10965 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10966 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10967 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10968 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10969 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10970 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10971 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10972 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10973 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10974 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10975 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10976 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
10978 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10979 sizeof(struct io_uring_rsrc_update));
10980 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10981 sizeof(struct io_uring_rsrc_update2));
10983 /* ->buf_index is u16 */
10984 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
10986 /* should fit into one byte */
10987 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10989 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10990 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
10992 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10996 __initcall(io_uring_init);