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;
507 struct io_timeout_data {
508 struct io_kiocb *req;
509 struct hrtimer timer;
510 struct timespec64 ts;
511 enum hrtimer_mode mode;
517 struct sockaddr __user *addr;
518 int __user *addr_len;
521 unsigned long nofile;
541 struct list_head list;
542 /* head of the link, used by linked timeouts only */
543 struct io_kiocb *head;
544 /* for linked completions */
545 struct io_kiocb *prev;
548 struct io_timeout_rem {
553 struct timespec64 ts;
559 /* NOTE: kiocb has the file as the first member, so don't do it here */
567 struct sockaddr __user *addr;
574 struct compat_msghdr __user *umsg_compat;
575 struct user_msghdr __user *umsg;
581 struct io_buffer *kbuf;
588 struct filename *filename;
590 unsigned long nofile;
593 struct io_rsrc_update {
619 struct epoll_event event;
623 struct file *file_out;
624 struct file *file_in;
631 struct io_provide_buf {
645 const char __user *filename;
646 struct statx __user *buffer;
658 struct filename *oldpath;
659 struct filename *newpath;
667 struct filename *filename;
674 struct filename *filename;
680 struct filename *oldpath;
681 struct filename *newpath;
688 struct filename *oldpath;
689 struct filename *newpath;
693 struct io_completion {
698 struct io_async_connect {
699 struct sockaddr_storage address;
702 struct io_async_msghdr {
703 struct iovec fast_iov[UIO_FASTIOV];
704 /* points to an allocated iov, if NULL we use fast_iov instead */
705 struct iovec *free_iov;
706 struct sockaddr __user *uaddr;
708 struct sockaddr_storage addr;
712 struct iovec fast_iov[UIO_FASTIOV];
713 const struct iovec *free_iovec;
714 struct iov_iter iter;
715 struct iov_iter_state iter_state;
717 struct wait_page_queue wpq;
721 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
722 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
723 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
724 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
725 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
726 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
728 /* first byte is taken by user flags, shift it to not overlap */
733 REQ_F_LINK_TIMEOUT_BIT,
734 REQ_F_NEED_CLEANUP_BIT,
736 REQ_F_BUFFER_SELECTED_BIT,
737 REQ_F_COMPLETE_INLINE_BIT,
741 REQ_F_ARM_LTIMEOUT_BIT,
742 /* keep async read/write and isreg together and in order */
743 REQ_F_NOWAIT_READ_BIT,
744 REQ_F_NOWAIT_WRITE_BIT,
747 /* not a real bit, just to check we're not overflowing the space */
753 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
754 /* drain existing IO first */
755 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
757 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
758 /* doesn't sever on completion < 0 */
759 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
761 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
762 /* IOSQE_BUFFER_SELECT */
763 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
765 /* fail rest of links */
766 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
767 /* on inflight list, should be cancelled and waited on exit reliably */
768 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
769 /* read/write uses file position */
770 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
771 /* must not punt to workers */
772 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
773 /* has or had linked timeout */
774 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
776 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
777 /* already went through poll handler */
778 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
779 /* buffer already selected */
780 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
781 /* completion is deferred through io_comp_state */
782 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
783 /* caller should reissue async */
784 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
785 /* supports async reads */
786 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
787 /* supports async writes */
788 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
790 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
791 /* has creds assigned */
792 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
793 /* skip refcounting if not set */
794 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
795 /* there is a linked timeout that has to be armed */
796 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
800 struct io_poll_iocb poll;
801 struct io_poll_iocb *double_poll;
804 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
806 struct io_task_work {
808 struct io_wq_work_node node;
809 struct llist_node fallback_node;
811 io_req_tw_func_t func;
815 IORING_RSRC_FILE = 0,
816 IORING_RSRC_BUFFER = 1,
820 * NOTE! Each of the iocb union members has the file pointer
821 * as the first entry in their struct definition. So you can
822 * access the file pointer through any of the sub-structs,
823 * or directly as just 'ki_filp' in this struct.
829 struct io_poll_iocb poll;
830 struct io_poll_update poll_update;
831 struct io_accept accept;
833 struct io_cancel cancel;
834 struct io_timeout timeout;
835 struct io_timeout_rem timeout_rem;
836 struct io_connect connect;
837 struct io_sr_msg sr_msg;
839 struct io_close close;
840 struct io_rsrc_update rsrc_update;
841 struct io_fadvise fadvise;
842 struct io_madvise madvise;
843 struct io_epoll epoll;
844 struct io_splice splice;
845 struct io_provide_buf pbuf;
846 struct io_statx statx;
847 struct io_shutdown shutdown;
848 struct io_rename rename;
849 struct io_unlink unlink;
850 struct io_mkdir mkdir;
851 struct io_symlink symlink;
852 struct io_hardlink hardlink;
853 /* use only after cleaning per-op data, see io_clean_op() */
854 struct io_completion compl;
857 /* opcode allocated if it needs to store data for async defer */
860 /* polled IO has completed */
866 struct io_ring_ctx *ctx;
869 struct task_struct *task;
872 struct io_kiocb *link;
873 struct percpu_ref *fixed_rsrc_refs;
875 /* used with ctx->iopoll_list with reads/writes */
876 struct list_head inflight_entry;
877 struct io_task_work io_task_work;
878 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
879 struct hlist_node hash_node;
880 struct async_poll *apoll;
881 struct io_wq_work work;
882 const struct cred *creds;
884 /* store used ubuf, so we can prevent reloading */
885 struct io_mapped_ubuf *imu;
888 struct io_tctx_node {
889 struct list_head ctx_node;
890 struct task_struct *task;
891 struct io_ring_ctx *ctx;
894 struct io_defer_entry {
895 struct list_head list;
896 struct io_kiocb *req;
901 /* needs req->file assigned */
902 unsigned needs_file : 1;
903 /* hash wq insertion if file is a regular file */
904 unsigned hash_reg_file : 1;
905 /* unbound wq insertion if file is a non-regular file */
906 unsigned unbound_nonreg_file : 1;
907 /* opcode is not supported by this kernel */
908 unsigned not_supported : 1;
909 /* set if opcode supports polled "wait" */
911 unsigned pollout : 1;
912 /* op supports buffer selection */
913 unsigned buffer_select : 1;
914 /* do prep async if is going to be punted */
915 unsigned needs_async_setup : 1;
916 /* should block plug */
918 /* size of async data needed, if any */
919 unsigned short async_size;
922 static const struct io_op_def io_op_defs[] = {
923 [IORING_OP_NOP] = {},
924 [IORING_OP_READV] = {
926 .unbound_nonreg_file = 1,
929 .needs_async_setup = 1,
931 .async_size = sizeof(struct io_async_rw),
933 [IORING_OP_WRITEV] = {
936 .unbound_nonreg_file = 1,
938 .needs_async_setup = 1,
940 .async_size = sizeof(struct io_async_rw),
942 [IORING_OP_FSYNC] = {
945 [IORING_OP_READ_FIXED] = {
947 .unbound_nonreg_file = 1,
950 .async_size = sizeof(struct io_async_rw),
952 [IORING_OP_WRITE_FIXED] = {
955 .unbound_nonreg_file = 1,
958 .async_size = sizeof(struct io_async_rw),
960 [IORING_OP_POLL_ADD] = {
962 .unbound_nonreg_file = 1,
964 [IORING_OP_POLL_REMOVE] = {},
965 [IORING_OP_SYNC_FILE_RANGE] = {
968 [IORING_OP_SENDMSG] = {
970 .unbound_nonreg_file = 1,
972 .needs_async_setup = 1,
973 .async_size = sizeof(struct io_async_msghdr),
975 [IORING_OP_RECVMSG] = {
977 .unbound_nonreg_file = 1,
980 .needs_async_setup = 1,
981 .async_size = sizeof(struct io_async_msghdr),
983 [IORING_OP_TIMEOUT] = {
984 .async_size = sizeof(struct io_timeout_data),
986 [IORING_OP_TIMEOUT_REMOVE] = {
987 /* used by timeout updates' prep() */
989 [IORING_OP_ACCEPT] = {
991 .unbound_nonreg_file = 1,
994 [IORING_OP_ASYNC_CANCEL] = {},
995 [IORING_OP_LINK_TIMEOUT] = {
996 .async_size = sizeof(struct io_timeout_data),
998 [IORING_OP_CONNECT] = {
1000 .unbound_nonreg_file = 1,
1002 .needs_async_setup = 1,
1003 .async_size = sizeof(struct io_async_connect),
1005 [IORING_OP_FALLOCATE] = {
1008 [IORING_OP_OPENAT] = {},
1009 [IORING_OP_CLOSE] = {},
1010 [IORING_OP_FILES_UPDATE] = {},
1011 [IORING_OP_STATX] = {},
1012 [IORING_OP_READ] = {
1014 .unbound_nonreg_file = 1,
1018 .async_size = sizeof(struct io_async_rw),
1020 [IORING_OP_WRITE] = {
1023 .unbound_nonreg_file = 1,
1026 .async_size = sizeof(struct io_async_rw),
1028 [IORING_OP_FADVISE] = {
1031 [IORING_OP_MADVISE] = {},
1032 [IORING_OP_SEND] = {
1034 .unbound_nonreg_file = 1,
1037 [IORING_OP_RECV] = {
1039 .unbound_nonreg_file = 1,
1043 [IORING_OP_OPENAT2] = {
1045 [IORING_OP_EPOLL_CTL] = {
1046 .unbound_nonreg_file = 1,
1048 [IORING_OP_SPLICE] = {
1051 .unbound_nonreg_file = 1,
1053 [IORING_OP_PROVIDE_BUFFERS] = {},
1054 [IORING_OP_REMOVE_BUFFERS] = {},
1058 .unbound_nonreg_file = 1,
1060 [IORING_OP_SHUTDOWN] = {
1063 [IORING_OP_RENAMEAT] = {},
1064 [IORING_OP_UNLINKAT] = {},
1065 [IORING_OP_MKDIRAT] = {},
1066 [IORING_OP_SYMLINKAT] = {},
1067 [IORING_OP_LINKAT] = {},
1070 /* requests with any of those set should undergo io_disarm_next() */
1071 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1073 static bool io_disarm_next(struct io_kiocb *req);
1074 static void io_uring_del_tctx_node(unsigned long index);
1075 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1076 struct task_struct *task,
1078 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1080 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1081 long res, unsigned int cflags);
1082 static void io_put_req(struct io_kiocb *req);
1083 static void io_put_req_deferred(struct io_kiocb *req);
1084 static void io_dismantle_req(struct io_kiocb *req);
1085 static void io_queue_linked_timeout(struct io_kiocb *req);
1086 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1087 struct io_uring_rsrc_update2 *up,
1089 static void io_clean_op(struct io_kiocb *req);
1090 static struct file *io_file_get(struct io_ring_ctx *ctx,
1091 struct io_kiocb *req, int fd, bool fixed);
1092 static void __io_queue_sqe(struct io_kiocb *req);
1093 static void io_rsrc_put_work(struct work_struct *work);
1095 static void io_req_task_queue(struct io_kiocb *req);
1096 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1097 static int io_req_prep_async(struct io_kiocb *req);
1099 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1100 unsigned int issue_flags, u32 slot_index);
1101 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
1103 static struct kmem_cache *req_cachep;
1105 static const struct file_operations io_uring_fops;
1107 struct sock *io_uring_get_socket(struct file *file)
1109 #if defined(CONFIG_UNIX)
1110 if (file->f_op == &io_uring_fops) {
1111 struct io_ring_ctx *ctx = file->private_data;
1113 return ctx->ring_sock->sk;
1118 EXPORT_SYMBOL(io_uring_get_socket);
1120 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1123 mutex_lock(&ctx->uring_lock);
1128 #define io_for_each_link(pos, head) \
1129 for (pos = (head); pos; pos = pos->link)
1132 * Shamelessly stolen from the mm implementation of page reference checking,
1133 * see commit f958d7b528b1 for details.
1135 #define req_ref_zero_or_close_to_overflow(req) \
1136 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1138 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1140 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1141 return atomic_inc_not_zero(&req->refs);
1144 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1146 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1149 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1150 return atomic_dec_and_test(&req->refs);
1153 static inline void req_ref_put(struct io_kiocb *req)
1155 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1156 WARN_ON_ONCE(req_ref_put_and_test(req));
1159 static inline void req_ref_get(struct io_kiocb *req)
1161 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1162 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1163 atomic_inc(&req->refs);
1166 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1168 if (!(req->flags & REQ_F_REFCOUNT)) {
1169 req->flags |= REQ_F_REFCOUNT;
1170 atomic_set(&req->refs, nr);
1174 static inline void io_req_set_refcount(struct io_kiocb *req)
1176 __io_req_set_refcount(req, 1);
1179 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1181 struct io_ring_ctx *ctx = req->ctx;
1183 if (!req->fixed_rsrc_refs) {
1184 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1185 percpu_ref_get(req->fixed_rsrc_refs);
1189 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1191 bool got = percpu_ref_tryget(ref);
1193 /* already at zero, wait for ->release() */
1195 wait_for_completion(compl);
1196 percpu_ref_resurrect(ref);
1198 percpu_ref_put(ref);
1201 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1204 struct io_kiocb *req;
1206 if (task && head->task != task)
1211 io_for_each_link(req, head) {
1212 if (req->flags & REQ_F_INFLIGHT)
1218 static inline void req_set_fail(struct io_kiocb *req)
1220 req->flags |= REQ_F_FAIL;
1223 static inline void req_fail_link_node(struct io_kiocb *req, int res)
1229 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1231 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1233 complete(&ctx->ref_comp);
1236 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1238 return !req->timeout.off;
1241 static void io_fallback_req_func(struct work_struct *work)
1243 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1244 fallback_work.work);
1245 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1246 struct io_kiocb *req, *tmp;
1247 bool locked = false;
1249 percpu_ref_get(&ctx->refs);
1250 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1251 req->io_task_work.func(req, &locked);
1254 if (ctx->submit_state.compl_nr)
1255 io_submit_flush_completions(ctx);
1256 mutex_unlock(&ctx->uring_lock);
1258 percpu_ref_put(&ctx->refs);
1262 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1264 struct io_ring_ctx *ctx;
1267 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1272 * Use 5 bits less than the max cq entries, that should give us around
1273 * 32 entries per hash list if totally full and uniformly spread.
1275 hash_bits = ilog2(p->cq_entries);
1279 ctx->cancel_hash_bits = hash_bits;
1280 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1282 if (!ctx->cancel_hash)
1284 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1286 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1287 if (!ctx->dummy_ubuf)
1289 /* set invalid range, so io_import_fixed() fails meeting it */
1290 ctx->dummy_ubuf->ubuf = -1UL;
1292 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1293 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1296 ctx->flags = p->flags;
1297 init_waitqueue_head(&ctx->sqo_sq_wait);
1298 INIT_LIST_HEAD(&ctx->sqd_list);
1299 init_waitqueue_head(&ctx->poll_wait);
1300 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1301 init_completion(&ctx->ref_comp);
1302 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1303 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1304 mutex_init(&ctx->uring_lock);
1305 init_waitqueue_head(&ctx->cq_wait);
1306 spin_lock_init(&ctx->completion_lock);
1307 spin_lock_init(&ctx->timeout_lock);
1308 INIT_LIST_HEAD(&ctx->iopoll_list);
1309 INIT_LIST_HEAD(&ctx->defer_list);
1310 INIT_LIST_HEAD(&ctx->timeout_list);
1311 INIT_LIST_HEAD(&ctx->ltimeout_list);
1312 spin_lock_init(&ctx->rsrc_ref_lock);
1313 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1314 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1315 init_llist_head(&ctx->rsrc_put_llist);
1316 INIT_LIST_HEAD(&ctx->tctx_list);
1317 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1318 INIT_LIST_HEAD(&ctx->locked_free_list);
1319 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1322 kfree(ctx->dummy_ubuf);
1323 kfree(ctx->cancel_hash);
1328 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1330 struct io_rings *r = ctx->rings;
1332 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1336 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1338 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1339 struct io_ring_ctx *ctx = req->ctx;
1341 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1347 #define FFS_ASYNC_READ 0x1UL
1348 #define FFS_ASYNC_WRITE 0x2UL
1350 #define FFS_ISREG 0x4UL
1352 #define FFS_ISREG 0x0UL
1354 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1356 static inline bool io_req_ffs_set(struct io_kiocb *req)
1358 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1361 static void io_req_track_inflight(struct io_kiocb *req)
1363 if (!(req->flags & REQ_F_INFLIGHT)) {
1364 req->flags |= REQ_F_INFLIGHT;
1365 atomic_inc(¤t->io_uring->inflight_tracked);
1369 static inline void io_unprep_linked_timeout(struct io_kiocb *req)
1371 req->flags &= ~REQ_F_LINK_TIMEOUT;
1374 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1376 if (WARN_ON_ONCE(!req->link))
1379 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1380 req->flags |= REQ_F_LINK_TIMEOUT;
1382 /* linked timeouts should have two refs once prep'ed */
1383 io_req_set_refcount(req);
1384 __io_req_set_refcount(req->link, 2);
1388 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1390 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1392 return __io_prep_linked_timeout(req);
1395 static void io_prep_async_work(struct io_kiocb *req)
1397 const struct io_op_def *def = &io_op_defs[req->opcode];
1398 struct io_ring_ctx *ctx = req->ctx;
1400 if (!(req->flags & REQ_F_CREDS)) {
1401 req->flags |= REQ_F_CREDS;
1402 req->creds = get_current_cred();
1405 req->work.list.next = NULL;
1406 req->work.flags = 0;
1407 if (req->flags & REQ_F_FORCE_ASYNC)
1408 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1410 if (req->flags & REQ_F_ISREG) {
1411 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1412 io_wq_hash_work(&req->work, file_inode(req->file));
1413 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1414 if (def->unbound_nonreg_file)
1415 req->work.flags |= IO_WQ_WORK_UNBOUND;
1418 switch (req->opcode) {
1419 case IORING_OP_SPLICE:
1421 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1422 req->work.flags |= IO_WQ_WORK_UNBOUND;
1427 static void io_prep_async_link(struct io_kiocb *req)
1429 struct io_kiocb *cur;
1431 if (req->flags & REQ_F_LINK_TIMEOUT) {
1432 struct io_ring_ctx *ctx = req->ctx;
1434 spin_lock(&ctx->completion_lock);
1435 io_for_each_link(cur, req)
1436 io_prep_async_work(cur);
1437 spin_unlock(&ctx->completion_lock);
1439 io_for_each_link(cur, req)
1440 io_prep_async_work(cur);
1444 static void io_queue_async_work(struct io_kiocb *req, bool *locked)
1446 struct io_ring_ctx *ctx = req->ctx;
1447 struct io_kiocb *link = io_prep_linked_timeout(req);
1448 struct io_uring_task *tctx = req->task->io_uring;
1450 /* must not take the lock, NULL it as a precaution */
1454 BUG_ON(!tctx->io_wq);
1456 /* init ->work of the whole link before punting */
1457 io_prep_async_link(req);
1460 * Not expected to happen, but if we do have a bug where this _can_
1461 * happen, catch it here and ensure the request is marked as
1462 * canceled. That will make io-wq go through the usual work cancel
1463 * procedure rather than attempt to run this request (or create a new
1466 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1467 req->work.flags |= IO_WQ_WORK_CANCEL;
1469 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1470 &req->work, req->flags);
1471 io_wq_enqueue(tctx->io_wq, &req->work);
1473 io_queue_linked_timeout(link);
1476 static void io_kill_timeout(struct io_kiocb *req, int status)
1477 __must_hold(&req->ctx->completion_lock)
1478 __must_hold(&req->ctx->timeout_lock)
1480 struct io_timeout_data *io = req->async_data;
1482 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1485 atomic_set(&req->ctx->cq_timeouts,
1486 atomic_read(&req->ctx->cq_timeouts) + 1);
1487 list_del_init(&req->timeout.list);
1488 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1489 io_put_req_deferred(req);
1493 static void io_queue_deferred(struct io_ring_ctx *ctx)
1495 while (!list_empty(&ctx->defer_list)) {
1496 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1497 struct io_defer_entry, list);
1499 if (req_need_defer(de->req, de->seq))
1501 list_del_init(&de->list);
1502 io_req_task_queue(de->req);
1507 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1508 __must_hold(&ctx->completion_lock)
1510 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1512 spin_lock_irq(&ctx->timeout_lock);
1513 while (!list_empty(&ctx->timeout_list)) {
1514 u32 events_needed, events_got;
1515 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1516 struct io_kiocb, timeout.list);
1518 if (io_is_timeout_noseq(req))
1522 * Since seq can easily wrap around over time, subtract
1523 * the last seq at which timeouts were flushed before comparing.
1524 * Assuming not more than 2^31-1 events have happened since,
1525 * these subtractions won't have wrapped, so we can check if
1526 * target is in [last_seq, current_seq] by comparing the two.
1528 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1529 events_got = seq - ctx->cq_last_tm_flush;
1530 if (events_got < events_needed)
1533 list_del_init(&req->timeout.list);
1534 io_kill_timeout(req, 0);
1536 ctx->cq_last_tm_flush = seq;
1537 spin_unlock_irq(&ctx->timeout_lock);
1540 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1542 if (ctx->off_timeout_used)
1543 io_flush_timeouts(ctx);
1544 if (ctx->drain_active)
1545 io_queue_deferred(ctx);
1548 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1550 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1551 __io_commit_cqring_flush(ctx);
1552 /* order cqe stores with ring update */
1553 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1556 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1558 struct io_rings *r = ctx->rings;
1560 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1563 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1565 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1568 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1570 struct io_rings *rings = ctx->rings;
1571 unsigned tail, mask = ctx->cq_entries - 1;
1574 * writes to the cq entry need to come after reading head; the
1575 * control dependency is enough as we're using WRITE_ONCE to
1578 if (__io_cqring_events(ctx) == ctx->cq_entries)
1581 tail = ctx->cached_cq_tail++;
1582 return &rings->cqes[tail & mask];
1585 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1587 if (likely(!ctx->cq_ev_fd))
1589 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1591 return !ctx->eventfd_async || io_wq_current_is_worker();
1595 * This should only get called when at least one event has been posted.
1596 * Some applications rely on the eventfd notification count only changing
1597 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1598 * 1:1 relationship between how many times this function is called (and
1599 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1601 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1604 * wake_up_all() may seem excessive, but io_wake_function() and
1605 * io_should_wake() handle the termination of the loop and only
1606 * wake as many waiters as we need to.
1608 if (wq_has_sleeper(&ctx->cq_wait))
1609 wake_up_all(&ctx->cq_wait);
1610 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1611 wake_up(&ctx->sq_data->wait);
1612 if (io_should_trigger_evfd(ctx))
1613 eventfd_signal(ctx->cq_ev_fd, 1);
1614 if (waitqueue_active(&ctx->poll_wait)) {
1615 wake_up_interruptible(&ctx->poll_wait);
1616 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1620 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1622 /* see waitqueue_active() comment */
1625 if (ctx->flags & IORING_SETUP_SQPOLL) {
1626 if (waitqueue_active(&ctx->cq_wait))
1627 wake_up_all(&ctx->cq_wait);
1629 if (io_should_trigger_evfd(ctx))
1630 eventfd_signal(ctx->cq_ev_fd, 1);
1631 if (waitqueue_active(&ctx->poll_wait)) {
1632 wake_up_interruptible(&ctx->poll_wait);
1633 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1637 /* Returns true if there are no backlogged entries after the flush */
1638 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1640 bool all_flushed, posted;
1642 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1646 spin_lock(&ctx->completion_lock);
1647 while (!list_empty(&ctx->cq_overflow_list)) {
1648 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1649 struct io_overflow_cqe *ocqe;
1653 ocqe = list_first_entry(&ctx->cq_overflow_list,
1654 struct io_overflow_cqe, list);
1656 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1658 io_account_cq_overflow(ctx);
1661 list_del(&ocqe->list);
1665 all_flushed = list_empty(&ctx->cq_overflow_list);
1667 clear_bit(0, &ctx->check_cq_overflow);
1668 WRITE_ONCE(ctx->rings->sq_flags,
1669 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1673 io_commit_cqring(ctx);
1674 spin_unlock(&ctx->completion_lock);
1676 io_cqring_ev_posted(ctx);
1680 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1684 if (test_bit(0, &ctx->check_cq_overflow)) {
1685 /* iopoll syncs against uring_lock, not completion_lock */
1686 if (ctx->flags & IORING_SETUP_IOPOLL)
1687 mutex_lock(&ctx->uring_lock);
1688 ret = __io_cqring_overflow_flush(ctx, false);
1689 if (ctx->flags & IORING_SETUP_IOPOLL)
1690 mutex_unlock(&ctx->uring_lock);
1696 /* must to be called somewhat shortly after putting a request */
1697 static inline void io_put_task(struct task_struct *task, int nr)
1699 struct io_uring_task *tctx = task->io_uring;
1701 if (likely(task == current)) {
1702 tctx->cached_refs += nr;
1704 percpu_counter_sub(&tctx->inflight, nr);
1705 if (unlikely(atomic_read(&tctx->in_idle)))
1706 wake_up(&tctx->wait);
1707 put_task_struct_many(task, nr);
1711 static void io_task_refs_refill(struct io_uring_task *tctx)
1713 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1715 percpu_counter_add(&tctx->inflight, refill);
1716 refcount_add(refill, ¤t->usage);
1717 tctx->cached_refs += refill;
1720 static inline void io_get_task_refs(int nr)
1722 struct io_uring_task *tctx = current->io_uring;
1724 tctx->cached_refs -= nr;
1725 if (unlikely(tctx->cached_refs < 0))
1726 io_task_refs_refill(tctx);
1729 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1730 long res, unsigned int cflags)
1732 struct io_overflow_cqe *ocqe;
1734 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1737 * If we're in ring overflow flush mode, or in task cancel mode,
1738 * or cannot allocate an overflow entry, then we need to drop it
1741 io_account_cq_overflow(ctx);
1744 if (list_empty(&ctx->cq_overflow_list)) {
1745 set_bit(0, &ctx->check_cq_overflow);
1746 WRITE_ONCE(ctx->rings->sq_flags,
1747 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1750 ocqe->cqe.user_data = user_data;
1751 ocqe->cqe.res = res;
1752 ocqe->cqe.flags = cflags;
1753 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1757 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1758 long res, unsigned int cflags)
1760 struct io_uring_cqe *cqe;
1762 trace_io_uring_complete(ctx, user_data, res, cflags);
1765 * If we can't get a cq entry, userspace overflowed the
1766 * submission (by quite a lot). Increment the overflow count in
1769 cqe = io_get_cqe(ctx);
1771 WRITE_ONCE(cqe->user_data, user_data);
1772 WRITE_ONCE(cqe->res, res);
1773 WRITE_ONCE(cqe->flags, cflags);
1776 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1779 /* not as hot to bloat with inlining */
1780 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1781 long res, unsigned int cflags)
1783 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1786 static void io_req_complete_post(struct io_kiocb *req, long res,
1787 unsigned int cflags)
1789 struct io_ring_ctx *ctx = req->ctx;
1791 spin_lock(&ctx->completion_lock);
1792 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1794 * If we're the last reference to this request, add to our locked
1797 if (req_ref_put_and_test(req)) {
1798 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1799 if (req->flags & IO_DISARM_MASK)
1800 io_disarm_next(req);
1802 io_req_task_queue(req->link);
1806 io_dismantle_req(req);
1807 io_put_task(req->task, 1);
1808 list_add(&req->inflight_entry, &ctx->locked_free_list);
1809 ctx->locked_free_nr++;
1811 if (!percpu_ref_tryget(&ctx->refs))
1814 io_commit_cqring(ctx);
1815 spin_unlock(&ctx->completion_lock);
1818 io_cqring_ev_posted(ctx);
1819 percpu_ref_put(&ctx->refs);
1823 static inline bool io_req_needs_clean(struct io_kiocb *req)
1825 return req->flags & IO_REQ_CLEAN_FLAGS;
1828 static void io_req_complete_state(struct io_kiocb *req, long res,
1829 unsigned int cflags)
1831 if (io_req_needs_clean(req))
1834 req->compl.cflags = cflags;
1835 req->flags |= REQ_F_COMPLETE_INLINE;
1838 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1839 long res, unsigned cflags)
1841 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1842 io_req_complete_state(req, res, cflags);
1844 io_req_complete_post(req, res, cflags);
1847 static inline void io_req_complete(struct io_kiocb *req, long res)
1849 __io_req_complete(req, 0, res, 0);
1852 static void io_req_complete_failed(struct io_kiocb *req, long res)
1855 io_req_complete_post(req, res, 0);
1858 static void io_req_complete_fail_submit(struct io_kiocb *req)
1861 * We don't submit, fail them all, for that replace hardlinks with
1862 * normal links. Extra REQ_F_LINK is tolerated.
1864 req->flags &= ~REQ_F_HARDLINK;
1865 req->flags |= REQ_F_LINK;
1866 io_req_complete_failed(req, req->result);
1870 * Don't initialise the fields below on every allocation, but do that in
1871 * advance and keep them valid across allocations.
1873 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1877 req->async_data = NULL;
1878 /* not necessary, but safer to zero */
1882 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1883 struct io_submit_state *state)
1885 spin_lock(&ctx->completion_lock);
1886 list_splice_init(&ctx->locked_free_list, &state->free_list);
1887 ctx->locked_free_nr = 0;
1888 spin_unlock(&ctx->completion_lock);
1891 /* Returns true IFF there are requests in the cache */
1892 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1894 struct io_submit_state *state = &ctx->submit_state;
1898 * If we have more than a batch's worth of requests in our IRQ side
1899 * locked cache, grab the lock and move them over to our submission
1902 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1903 io_flush_cached_locked_reqs(ctx, state);
1905 nr = state->free_reqs;
1906 while (!list_empty(&state->free_list)) {
1907 struct io_kiocb *req = list_first_entry(&state->free_list,
1908 struct io_kiocb, inflight_entry);
1910 list_del(&req->inflight_entry);
1911 state->reqs[nr++] = req;
1912 if (nr == ARRAY_SIZE(state->reqs))
1916 state->free_reqs = nr;
1921 * A request might get retired back into the request caches even before opcode
1922 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1923 * Because of that, io_alloc_req() should be called only under ->uring_lock
1924 * and with extra caution to not get a request that is still worked on.
1926 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1927 __must_hold(&ctx->uring_lock)
1929 struct io_submit_state *state = &ctx->submit_state;
1930 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1933 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1935 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1938 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1942 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1943 * retry single alloc to be on the safe side.
1945 if (unlikely(ret <= 0)) {
1946 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1947 if (!state->reqs[0])
1952 for (i = 0; i < ret; i++)
1953 io_preinit_req(state->reqs[i], ctx);
1954 state->free_reqs = ret;
1957 return state->reqs[state->free_reqs];
1960 static inline void io_put_file(struct file *file)
1966 static void io_dismantle_req(struct io_kiocb *req)
1968 unsigned int flags = req->flags;
1970 if (io_req_needs_clean(req))
1972 if (!(flags & REQ_F_FIXED_FILE))
1973 io_put_file(req->file);
1974 if (req->fixed_rsrc_refs)
1975 percpu_ref_put(req->fixed_rsrc_refs);
1976 if (req->async_data) {
1977 kfree(req->async_data);
1978 req->async_data = NULL;
1982 static void __io_free_req(struct io_kiocb *req)
1984 struct io_ring_ctx *ctx = req->ctx;
1986 io_dismantle_req(req);
1987 io_put_task(req->task, 1);
1989 spin_lock(&ctx->completion_lock);
1990 list_add(&req->inflight_entry, &ctx->locked_free_list);
1991 ctx->locked_free_nr++;
1992 spin_unlock(&ctx->completion_lock);
1994 percpu_ref_put(&ctx->refs);
1997 static inline void io_remove_next_linked(struct io_kiocb *req)
1999 struct io_kiocb *nxt = req->link;
2001 req->link = nxt->link;
2005 static bool io_kill_linked_timeout(struct io_kiocb *req)
2006 __must_hold(&req->ctx->completion_lock)
2007 __must_hold(&req->ctx->timeout_lock)
2009 struct io_kiocb *link = req->link;
2011 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2012 struct io_timeout_data *io = link->async_data;
2014 io_remove_next_linked(req);
2015 link->timeout.head = NULL;
2016 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2017 list_del(&link->timeout.list);
2018 io_cqring_fill_event(link->ctx, link->user_data,
2020 io_put_req_deferred(link);
2027 static void io_fail_links(struct io_kiocb *req)
2028 __must_hold(&req->ctx->completion_lock)
2030 struct io_kiocb *nxt, *link = req->link;
2034 long res = -ECANCELED;
2036 if (link->flags & REQ_F_FAIL)
2042 trace_io_uring_fail_link(req, link);
2043 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2044 io_put_req_deferred(link);
2049 static bool io_disarm_next(struct io_kiocb *req)
2050 __must_hold(&req->ctx->completion_lock)
2052 bool posted = false;
2054 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2055 struct io_kiocb *link = req->link;
2057 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2058 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2059 io_remove_next_linked(req);
2060 io_cqring_fill_event(link->ctx, link->user_data,
2062 io_put_req_deferred(link);
2065 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2066 struct io_ring_ctx *ctx = req->ctx;
2068 spin_lock_irq(&ctx->timeout_lock);
2069 posted = io_kill_linked_timeout(req);
2070 spin_unlock_irq(&ctx->timeout_lock);
2072 if (unlikely((req->flags & REQ_F_FAIL) &&
2073 !(req->flags & REQ_F_HARDLINK))) {
2074 posted |= (req->link != NULL);
2080 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2082 struct io_kiocb *nxt;
2085 * If LINK is set, we have dependent requests in this chain. If we
2086 * didn't fail this request, queue the first one up, moving any other
2087 * dependencies to the next request. In case of failure, fail the rest
2090 if (req->flags & IO_DISARM_MASK) {
2091 struct io_ring_ctx *ctx = req->ctx;
2094 spin_lock(&ctx->completion_lock);
2095 posted = io_disarm_next(req);
2097 io_commit_cqring(req->ctx);
2098 spin_unlock(&ctx->completion_lock);
2100 io_cqring_ev_posted(ctx);
2107 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2109 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2111 return __io_req_find_next(req);
2114 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2119 if (ctx->submit_state.compl_nr)
2120 io_submit_flush_completions(ctx);
2121 mutex_unlock(&ctx->uring_lock);
2124 percpu_ref_put(&ctx->refs);
2127 static void tctx_task_work(struct callback_head *cb)
2129 bool locked = false;
2130 struct io_ring_ctx *ctx = NULL;
2131 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2135 struct io_wq_work_node *node;
2137 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2138 io_submit_flush_completions(ctx);
2140 spin_lock_irq(&tctx->task_lock);
2141 node = tctx->task_list.first;
2142 INIT_WQ_LIST(&tctx->task_list);
2144 tctx->task_running = false;
2145 spin_unlock_irq(&tctx->task_lock);
2150 struct io_wq_work_node *next = node->next;
2151 struct io_kiocb *req = container_of(node, struct io_kiocb,
2154 if (req->ctx != ctx) {
2155 ctx_flush_and_put(ctx, &locked);
2157 /* if not contended, grab and improve batching */
2158 locked = mutex_trylock(&ctx->uring_lock);
2159 percpu_ref_get(&ctx->refs);
2161 req->io_task_work.func(req, &locked);
2168 ctx_flush_and_put(ctx, &locked);
2171 static void io_req_task_work_add(struct io_kiocb *req)
2173 struct task_struct *tsk = req->task;
2174 struct io_uring_task *tctx = tsk->io_uring;
2175 enum task_work_notify_mode notify;
2176 struct io_wq_work_node *node;
2177 unsigned long flags;
2180 WARN_ON_ONCE(!tctx);
2182 spin_lock_irqsave(&tctx->task_lock, flags);
2183 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2184 running = tctx->task_running;
2186 tctx->task_running = true;
2187 spin_unlock_irqrestore(&tctx->task_lock, flags);
2189 /* task_work already pending, we're done */
2194 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2195 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2196 * processing task_work. There's no reliable way to tell if TWA_RESUME
2199 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2200 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2201 wake_up_process(tsk);
2205 spin_lock_irqsave(&tctx->task_lock, flags);
2206 tctx->task_running = false;
2207 node = tctx->task_list.first;
2208 INIT_WQ_LIST(&tctx->task_list);
2209 spin_unlock_irqrestore(&tctx->task_lock, flags);
2212 req = container_of(node, struct io_kiocb, io_task_work.node);
2214 if (llist_add(&req->io_task_work.fallback_node,
2215 &req->ctx->fallback_llist))
2216 schedule_delayed_work(&req->ctx->fallback_work, 1);
2220 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2222 struct io_ring_ctx *ctx = req->ctx;
2224 /* not needed for normal modes, but SQPOLL depends on it */
2225 io_tw_lock(ctx, locked);
2226 io_req_complete_failed(req, req->result);
2229 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2231 struct io_ring_ctx *ctx = req->ctx;
2233 io_tw_lock(ctx, locked);
2234 /* req->task == current here, checking PF_EXITING is safe */
2235 if (likely(!(req->task->flags & PF_EXITING)))
2236 __io_queue_sqe(req);
2238 io_req_complete_failed(req, -EFAULT);
2241 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2244 req->io_task_work.func = io_req_task_cancel;
2245 io_req_task_work_add(req);
2248 static void io_req_task_queue(struct io_kiocb *req)
2250 req->io_task_work.func = io_req_task_submit;
2251 io_req_task_work_add(req);
2254 static void io_req_task_queue_reissue(struct io_kiocb *req)
2256 req->io_task_work.func = io_queue_async_work;
2257 io_req_task_work_add(req);
2260 static inline void io_queue_next(struct io_kiocb *req)
2262 struct io_kiocb *nxt = io_req_find_next(req);
2265 io_req_task_queue(nxt);
2268 static void io_free_req(struct io_kiocb *req)
2274 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2280 struct task_struct *task;
2285 static inline void io_init_req_batch(struct req_batch *rb)
2292 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2293 struct req_batch *rb)
2296 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2298 io_put_task(rb->task, rb->task_refs);
2301 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2302 struct io_submit_state *state)
2305 io_dismantle_req(req);
2307 if (req->task != rb->task) {
2309 io_put_task(rb->task, rb->task_refs);
2310 rb->task = req->task;
2316 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2317 state->reqs[state->free_reqs++] = req;
2319 list_add(&req->inflight_entry, &state->free_list);
2322 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2323 __must_hold(&ctx->uring_lock)
2325 struct io_submit_state *state = &ctx->submit_state;
2326 int i, nr = state->compl_nr;
2327 struct req_batch rb;
2329 spin_lock(&ctx->completion_lock);
2330 for (i = 0; i < nr; i++) {
2331 struct io_kiocb *req = state->compl_reqs[i];
2333 __io_cqring_fill_event(ctx, req->user_data, req->result,
2336 io_commit_cqring(ctx);
2337 spin_unlock(&ctx->completion_lock);
2338 io_cqring_ev_posted(ctx);
2340 io_init_req_batch(&rb);
2341 for (i = 0; i < nr; i++) {
2342 struct io_kiocb *req = state->compl_reqs[i];
2344 if (req_ref_put_and_test(req))
2345 io_req_free_batch(&rb, req, &ctx->submit_state);
2348 io_req_free_batch_finish(ctx, &rb);
2349 state->compl_nr = 0;
2353 * Drop reference to request, return next in chain (if there is one) if this
2354 * was the last reference to this request.
2356 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2358 struct io_kiocb *nxt = NULL;
2360 if (req_ref_put_and_test(req)) {
2361 nxt = io_req_find_next(req);
2367 static inline void io_put_req(struct io_kiocb *req)
2369 if (req_ref_put_and_test(req))
2373 static inline void io_put_req_deferred(struct io_kiocb *req)
2375 if (req_ref_put_and_test(req)) {
2376 req->io_task_work.func = io_free_req_work;
2377 io_req_task_work_add(req);
2381 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2383 /* See comment at the top of this file */
2385 return __io_cqring_events(ctx);
2388 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2390 struct io_rings *rings = ctx->rings;
2392 /* make sure SQ entry isn't read before tail */
2393 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2396 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2398 unsigned int cflags;
2400 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2401 cflags |= IORING_CQE_F_BUFFER;
2402 req->flags &= ~REQ_F_BUFFER_SELECTED;
2407 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2409 struct io_buffer *kbuf;
2411 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2413 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2414 return io_put_kbuf(req, kbuf);
2417 static inline bool io_run_task_work(void)
2419 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2420 __set_current_state(TASK_RUNNING);
2421 tracehook_notify_signal();
2429 * Find and free completed poll iocbs
2431 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2432 struct list_head *done)
2434 struct req_batch rb;
2435 struct io_kiocb *req;
2437 /* order with ->result store in io_complete_rw_iopoll() */
2440 io_init_req_batch(&rb);
2441 while (!list_empty(done)) {
2442 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2443 list_del(&req->inflight_entry);
2445 __io_cqring_fill_event(ctx, req->user_data, req->result,
2446 io_put_rw_kbuf(req));
2449 if (req_ref_put_and_test(req))
2450 io_req_free_batch(&rb, req, &ctx->submit_state);
2453 io_commit_cqring(ctx);
2454 io_cqring_ev_posted_iopoll(ctx);
2455 io_req_free_batch_finish(ctx, &rb);
2458 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2461 struct io_kiocb *req, *tmp;
2466 * Only spin for completions if we don't have multiple devices hanging
2467 * off our complete list, and we're under the requested amount.
2469 spin = !ctx->poll_multi_queue && *nr_events < min;
2471 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2472 struct kiocb *kiocb = &req->rw.kiocb;
2476 * Move completed and retryable entries to our local lists.
2477 * If we find a request that requires polling, break out
2478 * and complete those lists first, if we have entries there.
2480 if (READ_ONCE(req->iopoll_completed)) {
2481 list_move_tail(&req->inflight_entry, &done);
2484 if (!list_empty(&done))
2487 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2488 if (unlikely(ret < 0))
2493 /* iopoll may have completed current req */
2494 if (READ_ONCE(req->iopoll_completed))
2495 list_move_tail(&req->inflight_entry, &done);
2498 if (!list_empty(&done))
2499 io_iopoll_complete(ctx, nr_events, &done);
2505 * We can't just wait for polled events to come to us, we have to actively
2506 * find and complete them.
2508 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2510 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2513 mutex_lock(&ctx->uring_lock);
2514 while (!list_empty(&ctx->iopoll_list)) {
2515 unsigned int nr_events = 0;
2517 io_do_iopoll(ctx, &nr_events, 0);
2519 /* let it sleep and repeat later if can't complete a request */
2523 * Ensure we allow local-to-the-cpu processing to take place,
2524 * in this case we need to ensure that we reap all events.
2525 * Also let task_work, etc. to progress by releasing the mutex
2527 if (need_resched()) {
2528 mutex_unlock(&ctx->uring_lock);
2530 mutex_lock(&ctx->uring_lock);
2533 mutex_unlock(&ctx->uring_lock);
2536 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2538 unsigned int nr_events = 0;
2542 * We disallow the app entering submit/complete with polling, but we
2543 * still need to lock the ring to prevent racing with polled issue
2544 * that got punted to a workqueue.
2546 mutex_lock(&ctx->uring_lock);
2548 * Don't enter poll loop if we already have events pending.
2549 * If we do, we can potentially be spinning for commands that
2550 * already triggered a CQE (eg in error).
2552 if (test_bit(0, &ctx->check_cq_overflow))
2553 __io_cqring_overflow_flush(ctx, false);
2554 if (io_cqring_events(ctx))
2558 * If a submit got punted to a workqueue, we can have the
2559 * application entering polling for a command before it gets
2560 * issued. That app will hold the uring_lock for the duration
2561 * of the poll right here, so we need to take a breather every
2562 * now and then to ensure that the issue has a chance to add
2563 * the poll to the issued list. Otherwise we can spin here
2564 * forever, while the workqueue is stuck trying to acquire the
2567 if (list_empty(&ctx->iopoll_list)) {
2568 u32 tail = ctx->cached_cq_tail;
2570 mutex_unlock(&ctx->uring_lock);
2572 mutex_lock(&ctx->uring_lock);
2574 /* some requests don't go through iopoll_list */
2575 if (tail != ctx->cached_cq_tail ||
2576 list_empty(&ctx->iopoll_list))
2579 ret = io_do_iopoll(ctx, &nr_events, min);
2580 } while (!ret && nr_events < min && !need_resched());
2582 mutex_unlock(&ctx->uring_lock);
2586 static void kiocb_end_write(struct io_kiocb *req)
2589 * Tell lockdep we inherited freeze protection from submission
2592 if (req->flags & REQ_F_ISREG) {
2593 struct super_block *sb = file_inode(req->file)->i_sb;
2595 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2601 static bool io_resubmit_prep(struct io_kiocb *req)
2603 struct io_async_rw *rw = req->async_data;
2606 return !io_req_prep_async(req);
2607 iov_iter_restore(&rw->iter, &rw->iter_state);
2611 static bool io_rw_should_reissue(struct io_kiocb *req)
2613 umode_t mode = file_inode(req->file)->i_mode;
2614 struct io_ring_ctx *ctx = req->ctx;
2616 if (!S_ISBLK(mode) && !S_ISREG(mode))
2618 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2619 !(ctx->flags & IORING_SETUP_IOPOLL)))
2622 * If ref is dying, we might be running poll reap from the exit work.
2623 * Don't attempt to reissue from that path, just let it fail with
2626 if (percpu_ref_is_dying(&ctx->refs))
2629 * Play it safe and assume not safe to re-import and reissue if we're
2630 * not in the original thread group (or in task context).
2632 if (!same_thread_group(req->task, current) || !in_task())
2637 static bool io_resubmit_prep(struct io_kiocb *req)
2641 static bool io_rw_should_reissue(struct io_kiocb *req)
2647 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2649 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2650 kiocb_end_write(req);
2651 if (res != req->result) {
2652 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2653 io_rw_should_reissue(req)) {
2654 req->flags |= REQ_F_REISSUE;
2663 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2665 unsigned int cflags = io_put_rw_kbuf(req);
2666 long res = req->result;
2669 struct io_ring_ctx *ctx = req->ctx;
2670 struct io_submit_state *state = &ctx->submit_state;
2672 io_req_complete_state(req, res, cflags);
2673 state->compl_reqs[state->compl_nr++] = req;
2674 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2675 io_submit_flush_completions(ctx);
2677 io_req_complete_post(req, res, cflags);
2681 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2682 unsigned int issue_flags)
2684 if (__io_complete_rw_common(req, res))
2686 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2689 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2691 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2693 if (__io_complete_rw_common(req, res))
2696 req->io_task_work.func = io_req_task_complete;
2697 io_req_task_work_add(req);
2700 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2702 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2704 if (kiocb->ki_flags & IOCB_WRITE)
2705 kiocb_end_write(req);
2706 if (unlikely(res != req->result)) {
2707 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2708 req->flags |= REQ_F_REISSUE;
2713 WRITE_ONCE(req->result, res);
2714 /* order with io_iopoll_complete() checking ->result */
2716 WRITE_ONCE(req->iopoll_completed, 1);
2720 * After the iocb has been issued, it's safe to be found on the poll list.
2721 * Adding the kiocb to the list AFTER submission ensures that we don't
2722 * find it from a io_do_iopoll() thread before the issuer is done
2723 * accessing the kiocb cookie.
2725 static void io_iopoll_req_issued(struct io_kiocb *req)
2727 struct io_ring_ctx *ctx = req->ctx;
2728 const bool in_async = io_wq_current_is_worker();
2730 /* workqueue context doesn't hold uring_lock, grab it now */
2731 if (unlikely(in_async))
2732 mutex_lock(&ctx->uring_lock);
2735 * Track whether we have multiple files in our lists. This will impact
2736 * how we do polling eventually, not spinning if we're on potentially
2737 * different devices.
2739 if (list_empty(&ctx->iopoll_list)) {
2740 ctx->poll_multi_queue = false;
2741 } else if (!ctx->poll_multi_queue) {
2742 struct io_kiocb *list_req;
2743 unsigned int queue_num0, queue_num1;
2745 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2748 if (list_req->file != req->file) {
2749 ctx->poll_multi_queue = true;
2751 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2752 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2753 if (queue_num0 != queue_num1)
2754 ctx->poll_multi_queue = true;
2759 * For fast devices, IO may have already completed. If it has, add
2760 * it to the front so we find it first.
2762 if (READ_ONCE(req->iopoll_completed))
2763 list_add(&req->inflight_entry, &ctx->iopoll_list);
2765 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2767 if (unlikely(in_async)) {
2769 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2770 * in sq thread task context or in io worker task context. If
2771 * current task context is sq thread, we don't need to check
2772 * whether should wake up sq thread.
2774 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2775 wq_has_sleeper(&ctx->sq_data->wait))
2776 wake_up(&ctx->sq_data->wait);
2778 mutex_unlock(&ctx->uring_lock);
2782 static bool io_bdev_nowait(struct block_device *bdev)
2784 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2788 * If we tracked the file through the SCM inflight mechanism, we could support
2789 * any file. For now, just ensure that anything potentially problematic is done
2792 static bool __io_file_supports_nowait(struct file *file, int rw)
2794 umode_t mode = file_inode(file)->i_mode;
2796 if (S_ISBLK(mode)) {
2797 if (IS_ENABLED(CONFIG_BLOCK) &&
2798 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2804 if (S_ISREG(mode)) {
2805 if (IS_ENABLED(CONFIG_BLOCK) &&
2806 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2807 file->f_op != &io_uring_fops)
2812 /* any ->read/write should understand O_NONBLOCK */
2813 if (file->f_flags & O_NONBLOCK)
2816 if (!(file->f_mode & FMODE_NOWAIT))
2820 return file->f_op->read_iter != NULL;
2822 return file->f_op->write_iter != NULL;
2825 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2827 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2829 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2832 return __io_file_supports_nowait(req->file, rw);
2835 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2838 struct io_ring_ctx *ctx = req->ctx;
2839 struct kiocb *kiocb = &req->rw.kiocb;
2840 struct file *file = req->file;
2844 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2845 req->flags |= REQ_F_ISREG;
2847 kiocb->ki_pos = READ_ONCE(sqe->off);
2848 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2849 req->flags |= REQ_F_CUR_POS;
2850 kiocb->ki_pos = file->f_pos;
2852 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2853 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2854 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2859 * If the file is marked O_NONBLOCK, still allow retry for it if it
2860 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2861 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2863 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2864 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2865 req->flags |= REQ_F_NOWAIT;
2867 ioprio = READ_ONCE(sqe->ioprio);
2869 ret = ioprio_check_cap(ioprio);
2873 kiocb->ki_ioprio = ioprio;
2875 kiocb->ki_ioprio = get_current_ioprio();
2877 if (ctx->flags & IORING_SETUP_IOPOLL) {
2878 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2879 !kiocb->ki_filp->f_op->iopoll)
2882 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2883 kiocb->ki_complete = io_complete_rw_iopoll;
2884 req->iopoll_completed = 0;
2886 if (kiocb->ki_flags & IOCB_HIPRI)
2888 kiocb->ki_complete = io_complete_rw;
2891 if (req->opcode == IORING_OP_READ_FIXED ||
2892 req->opcode == IORING_OP_WRITE_FIXED) {
2894 io_req_set_rsrc_node(req);
2897 req->rw.addr = READ_ONCE(sqe->addr);
2898 req->rw.len = READ_ONCE(sqe->len);
2899 req->buf_index = READ_ONCE(sqe->buf_index);
2903 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2909 case -ERESTARTNOINTR:
2910 case -ERESTARTNOHAND:
2911 case -ERESTART_RESTARTBLOCK:
2913 * We can't just restart the syscall, since previously
2914 * submitted sqes may already be in progress. Just fail this
2920 kiocb->ki_complete(kiocb, ret, 0);
2924 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2925 unsigned int issue_flags)
2927 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2928 struct io_async_rw *io = req->async_data;
2930 /* add previously done IO, if any */
2931 if (io && io->bytes_done > 0) {
2933 ret = io->bytes_done;
2935 ret += io->bytes_done;
2938 if (req->flags & REQ_F_CUR_POS)
2939 req->file->f_pos = kiocb->ki_pos;
2940 if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
2941 __io_complete_rw(req, ret, 0, issue_flags);
2943 io_rw_done(kiocb, ret);
2945 if (req->flags & REQ_F_REISSUE) {
2946 req->flags &= ~REQ_F_REISSUE;
2947 if (io_resubmit_prep(req)) {
2948 io_req_task_queue_reissue(req);
2950 unsigned int cflags = io_put_rw_kbuf(req);
2951 struct io_ring_ctx *ctx = req->ctx;
2954 if (issue_flags & IO_URING_F_NONBLOCK) {
2955 mutex_lock(&ctx->uring_lock);
2956 __io_req_complete(req, issue_flags, ret, cflags);
2957 mutex_unlock(&ctx->uring_lock);
2959 __io_req_complete(req, issue_flags, ret, cflags);
2965 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2966 struct io_mapped_ubuf *imu)
2968 size_t len = req->rw.len;
2969 u64 buf_end, buf_addr = req->rw.addr;
2972 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2974 /* not inside the mapped region */
2975 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2979 * May not be a start of buffer, set size appropriately
2980 * and advance us to the beginning.
2982 offset = buf_addr - imu->ubuf;
2983 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2987 * Don't use iov_iter_advance() here, as it's really slow for
2988 * using the latter parts of a big fixed buffer - it iterates
2989 * over each segment manually. We can cheat a bit here, because
2992 * 1) it's a BVEC iter, we set it up
2993 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2994 * first and last bvec
2996 * So just find our index, and adjust the iterator afterwards.
2997 * If the offset is within the first bvec (or the whole first
2998 * bvec, just use iov_iter_advance(). This makes it easier
2999 * since we can just skip the first segment, which may not
3000 * be PAGE_SIZE aligned.
3002 const struct bio_vec *bvec = imu->bvec;
3004 if (offset <= bvec->bv_len) {
3005 iov_iter_advance(iter, offset);
3007 unsigned long seg_skip;
3009 /* skip first vec */
3010 offset -= bvec->bv_len;
3011 seg_skip = 1 + (offset >> PAGE_SHIFT);
3013 iter->bvec = bvec + seg_skip;
3014 iter->nr_segs -= seg_skip;
3015 iter->count -= bvec->bv_len + offset;
3016 iter->iov_offset = offset & ~PAGE_MASK;
3023 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3025 struct io_ring_ctx *ctx = req->ctx;
3026 struct io_mapped_ubuf *imu = req->imu;
3027 u16 index, buf_index = req->buf_index;
3030 if (unlikely(buf_index >= ctx->nr_user_bufs))
3032 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3033 imu = READ_ONCE(ctx->user_bufs[index]);
3036 return __io_import_fixed(req, rw, iter, imu);
3039 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3042 mutex_unlock(&ctx->uring_lock);
3045 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3048 * "Normal" inline submissions always hold the uring_lock, since we
3049 * grab it from the system call. Same is true for the SQPOLL offload.
3050 * The only exception is when we've detached the request and issue it
3051 * from an async worker thread, grab the lock for that case.
3054 mutex_lock(&ctx->uring_lock);
3057 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3058 int bgid, struct io_buffer *kbuf,
3061 struct io_buffer *head;
3063 if (req->flags & REQ_F_BUFFER_SELECTED)
3066 io_ring_submit_lock(req->ctx, needs_lock);
3068 lockdep_assert_held(&req->ctx->uring_lock);
3070 head = xa_load(&req->ctx->io_buffers, bgid);
3072 if (!list_empty(&head->list)) {
3073 kbuf = list_last_entry(&head->list, struct io_buffer,
3075 list_del(&kbuf->list);
3078 xa_erase(&req->ctx->io_buffers, bgid);
3080 if (*len > kbuf->len)
3083 kbuf = ERR_PTR(-ENOBUFS);
3086 io_ring_submit_unlock(req->ctx, needs_lock);
3091 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3094 struct io_buffer *kbuf;
3097 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3098 bgid = req->buf_index;
3099 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3102 req->rw.addr = (u64) (unsigned long) kbuf;
3103 req->flags |= REQ_F_BUFFER_SELECTED;
3104 return u64_to_user_ptr(kbuf->addr);
3107 #ifdef CONFIG_COMPAT
3108 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3111 struct compat_iovec __user *uiov;
3112 compat_ssize_t clen;
3116 uiov = u64_to_user_ptr(req->rw.addr);
3117 if (!access_ok(uiov, sizeof(*uiov)))
3119 if (__get_user(clen, &uiov->iov_len))
3125 buf = io_rw_buffer_select(req, &len, needs_lock);
3127 return PTR_ERR(buf);
3128 iov[0].iov_base = buf;
3129 iov[0].iov_len = (compat_size_t) len;
3134 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3137 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3141 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3144 len = iov[0].iov_len;
3147 buf = io_rw_buffer_select(req, &len, needs_lock);
3149 return PTR_ERR(buf);
3150 iov[0].iov_base = buf;
3151 iov[0].iov_len = len;
3155 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3158 if (req->flags & REQ_F_BUFFER_SELECTED) {
3159 struct io_buffer *kbuf;
3161 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3162 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3163 iov[0].iov_len = kbuf->len;
3166 if (req->rw.len != 1)
3169 #ifdef CONFIG_COMPAT
3170 if (req->ctx->compat)
3171 return io_compat_import(req, iov, needs_lock);
3174 return __io_iov_buffer_select(req, iov, needs_lock);
3177 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3178 struct iov_iter *iter, bool needs_lock)
3180 void __user *buf = u64_to_user_ptr(req->rw.addr);
3181 size_t sqe_len = req->rw.len;
3182 u8 opcode = req->opcode;
3185 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3187 return io_import_fixed(req, rw, iter);
3190 /* buffer index only valid with fixed read/write, or buffer select */
3191 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3194 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3195 if (req->flags & REQ_F_BUFFER_SELECT) {
3196 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3198 return PTR_ERR(buf);
3199 req->rw.len = sqe_len;
3202 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3207 if (req->flags & REQ_F_BUFFER_SELECT) {
3208 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3210 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3215 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3219 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3221 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3225 * For files that don't have ->read_iter() and ->write_iter(), handle them
3226 * by looping over ->read() or ->write() manually.
3228 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3230 struct kiocb *kiocb = &req->rw.kiocb;
3231 struct file *file = req->file;
3235 * Don't support polled IO through this interface, and we can't
3236 * support non-blocking either. For the latter, this just causes
3237 * the kiocb to be handled from an async context.
3239 if (kiocb->ki_flags & IOCB_HIPRI)
3241 if (kiocb->ki_flags & IOCB_NOWAIT)
3244 while (iov_iter_count(iter)) {
3248 if (!iov_iter_is_bvec(iter)) {
3249 iovec = iov_iter_iovec(iter);
3251 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3252 iovec.iov_len = req->rw.len;
3256 nr = file->f_op->read(file, iovec.iov_base,
3257 iovec.iov_len, io_kiocb_ppos(kiocb));
3259 nr = file->f_op->write(file, iovec.iov_base,
3260 iovec.iov_len, io_kiocb_ppos(kiocb));
3268 if (!iov_iter_is_bvec(iter)) {
3269 iov_iter_advance(iter, nr);
3275 if (nr != iovec.iov_len)
3282 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3283 const struct iovec *fast_iov, struct iov_iter *iter)
3285 struct io_async_rw *rw = req->async_data;
3287 memcpy(&rw->iter, iter, sizeof(*iter));
3288 rw->free_iovec = iovec;
3290 /* can only be fixed buffers, no need to do anything */
3291 if (iov_iter_is_bvec(iter))
3294 unsigned iov_off = 0;
3296 rw->iter.iov = rw->fast_iov;
3297 if (iter->iov != fast_iov) {
3298 iov_off = iter->iov - fast_iov;
3299 rw->iter.iov += iov_off;
3301 if (rw->fast_iov != fast_iov)
3302 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3303 sizeof(struct iovec) * iter->nr_segs);
3305 req->flags |= REQ_F_NEED_CLEANUP;
3309 static inline int io_alloc_async_data(struct io_kiocb *req)
3311 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3312 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3313 return req->async_data == NULL;
3316 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3317 const struct iovec *fast_iov,
3318 struct iov_iter *iter, bool force)
3320 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3322 if (!req->async_data) {
3323 struct io_async_rw *iorw;
3325 if (io_alloc_async_data(req)) {
3330 io_req_map_rw(req, iovec, fast_iov, iter);
3331 iorw = req->async_data;
3332 /* we've copied and mapped the iter, ensure state is saved */
3333 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3338 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3340 struct io_async_rw *iorw = req->async_data;
3341 struct iovec *iov = iorw->fast_iov;
3344 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3345 if (unlikely(ret < 0))
3348 iorw->bytes_done = 0;
3349 iorw->free_iovec = iov;
3351 req->flags |= REQ_F_NEED_CLEANUP;
3352 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3356 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3358 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3360 return io_prep_rw(req, sqe, READ);
3364 * This is our waitqueue callback handler, registered through lock_page_async()
3365 * when we initially tried to do the IO with the iocb armed our waitqueue.
3366 * This gets called when the page is unlocked, and we generally expect that to
3367 * happen when the page IO is completed and the page is now uptodate. This will
3368 * queue a task_work based retry of the operation, attempting to copy the data
3369 * again. If the latter fails because the page was NOT uptodate, then we will
3370 * do a thread based blocking retry of the operation. That's the unexpected
3373 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3374 int sync, void *arg)
3376 struct wait_page_queue *wpq;
3377 struct io_kiocb *req = wait->private;
3378 struct wait_page_key *key = arg;
3380 wpq = container_of(wait, struct wait_page_queue, wait);
3382 if (!wake_page_match(wpq, key))
3385 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3386 list_del_init(&wait->entry);
3387 io_req_task_queue(req);
3392 * This controls whether a given IO request should be armed for async page
3393 * based retry. If we return false here, the request is handed to the async
3394 * worker threads for retry. If we're doing buffered reads on a regular file,
3395 * we prepare a private wait_page_queue entry and retry the operation. This
3396 * will either succeed because the page is now uptodate and unlocked, or it
3397 * will register a callback when the page is unlocked at IO completion. Through
3398 * that callback, io_uring uses task_work to setup a retry of the operation.
3399 * That retry will attempt the buffered read again. The retry will generally
3400 * succeed, or in rare cases where it fails, we then fall back to using the
3401 * async worker threads for a blocking retry.
3403 static bool io_rw_should_retry(struct io_kiocb *req)
3405 struct io_async_rw *rw = req->async_data;
3406 struct wait_page_queue *wait = &rw->wpq;
3407 struct kiocb *kiocb = &req->rw.kiocb;
3409 /* never retry for NOWAIT, we just complete with -EAGAIN */
3410 if (req->flags & REQ_F_NOWAIT)
3413 /* Only for buffered IO */
3414 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3418 * just use poll if we can, and don't attempt if the fs doesn't
3419 * support callback based unlocks
3421 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3424 wait->wait.func = io_async_buf_func;
3425 wait->wait.private = req;
3426 wait->wait.flags = 0;
3427 INIT_LIST_HEAD(&wait->wait.entry);
3428 kiocb->ki_flags |= IOCB_WAITQ;
3429 kiocb->ki_flags &= ~IOCB_NOWAIT;
3430 kiocb->ki_waitq = wait;
3434 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3436 if (req->file->f_op->read_iter)
3437 return call_read_iter(req->file, &req->rw.kiocb, iter);
3438 else if (req->file->f_op->read)
3439 return loop_rw_iter(READ, req, iter);
3444 static bool need_read_all(struct io_kiocb *req)
3446 return req->flags & REQ_F_ISREG ||
3447 S_ISBLK(file_inode(req->file)->i_mode);
3450 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3452 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3453 struct kiocb *kiocb = &req->rw.kiocb;
3454 struct iov_iter __iter, *iter = &__iter;
3455 struct io_async_rw *rw = req->async_data;
3456 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3457 struct iov_iter_state __state, *state;
3462 state = &rw->iter_state;
3464 * We come here from an earlier attempt, restore our state to
3465 * match in case it doesn't. It's cheap enough that we don't
3466 * need to make this conditional.
3468 iov_iter_restore(iter, state);
3471 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3475 iov_iter_save_state(iter, state);
3477 req->result = iov_iter_count(iter);
3479 /* Ensure we clear previously set non-block flag */
3480 if (!force_nonblock)
3481 kiocb->ki_flags &= ~IOCB_NOWAIT;
3483 kiocb->ki_flags |= IOCB_NOWAIT;
3485 /* If the file doesn't support async, just async punt */
3486 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3487 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3488 return ret ?: -EAGAIN;
3491 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3492 if (unlikely(ret)) {
3497 ret = io_iter_do_read(req, iter);
3499 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3500 req->flags &= ~REQ_F_REISSUE;
3501 /* IOPOLL retry should happen for io-wq threads */
3502 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3504 /* no retry on NONBLOCK nor RWF_NOWAIT */
3505 if (req->flags & REQ_F_NOWAIT)
3508 } else if (ret == -EIOCBQUEUED) {
3510 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3511 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3512 /* read all, failed, already did sync or don't want to retry */
3517 * Don't depend on the iter state matching what was consumed, or being
3518 * untouched in case of error. Restore it and we'll advance it
3519 * manually if we need to.
3521 iov_iter_restore(iter, state);
3523 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3528 rw = req->async_data;
3530 * Now use our persistent iterator and state, if we aren't already.
3531 * We've restored and mapped the iter to match.
3533 if (iter != &rw->iter) {
3535 state = &rw->iter_state;
3540 * We end up here because of a partial read, either from
3541 * above or inside this loop. Advance the iter by the bytes
3542 * that were consumed.
3544 iov_iter_advance(iter, ret);
3545 if (!iov_iter_count(iter))
3547 rw->bytes_done += ret;
3548 iov_iter_save_state(iter, state);
3550 /* if we can retry, do so with the callbacks armed */
3551 if (!io_rw_should_retry(req)) {
3552 kiocb->ki_flags &= ~IOCB_WAITQ;
3557 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3558 * we get -EIOCBQUEUED, then we'll get a notification when the
3559 * desired page gets unlocked. We can also get a partial read
3560 * here, and if we do, then just retry at the new offset.
3562 ret = io_iter_do_read(req, iter);
3563 if (ret == -EIOCBQUEUED)
3565 /* we got some bytes, but not all. retry. */
3566 kiocb->ki_flags &= ~IOCB_WAITQ;
3567 iov_iter_restore(iter, state);
3570 kiocb_done(kiocb, ret, issue_flags);
3572 /* it's faster to check here then delegate to kfree */
3578 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3580 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3582 return io_prep_rw(req, sqe, WRITE);
3585 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3587 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3588 struct kiocb *kiocb = &req->rw.kiocb;
3589 struct iov_iter __iter, *iter = &__iter;
3590 struct io_async_rw *rw = req->async_data;
3591 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3592 struct iov_iter_state __state, *state;
3597 state = &rw->iter_state;
3598 iov_iter_restore(iter, state);
3601 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3605 iov_iter_save_state(iter, state);
3607 req->result = iov_iter_count(iter);
3610 /* Ensure we clear previously set non-block flag */
3611 if (!force_nonblock)
3612 kiocb->ki_flags &= ~IOCB_NOWAIT;
3614 kiocb->ki_flags |= IOCB_NOWAIT;
3616 /* If the file doesn't support async, just async punt */
3617 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3620 /* file path doesn't support NOWAIT for non-direct_IO */
3621 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3622 (req->flags & REQ_F_ISREG))
3625 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3630 * Open-code file_start_write here to grab freeze protection,
3631 * which will be released by another thread in
3632 * io_complete_rw(). Fool lockdep by telling it the lock got
3633 * released so that it doesn't complain about the held lock when
3634 * we return to userspace.
3636 if (req->flags & REQ_F_ISREG) {
3637 sb_start_write(file_inode(req->file)->i_sb);
3638 __sb_writers_release(file_inode(req->file)->i_sb,
3641 kiocb->ki_flags |= IOCB_WRITE;
3643 if (req->file->f_op->write_iter)
3644 ret2 = call_write_iter(req->file, kiocb, iter);
3645 else if (req->file->f_op->write)
3646 ret2 = loop_rw_iter(WRITE, req, iter);
3650 if (req->flags & REQ_F_REISSUE) {
3651 req->flags &= ~REQ_F_REISSUE;
3656 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3657 * retry them without IOCB_NOWAIT.
3659 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3661 /* no retry on NONBLOCK nor RWF_NOWAIT */
3662 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3664 if (!force_nonblock || ret2 != -EAGAIN) {
3665 /* IOPOLL retry should happen for io-wq threads */
3666 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3669 kiocb_done(kiocb, ret2, issue_flags);
3672 iov_iter_restore(iter, state);
3674 iov_iter_advance(iter, ret2);
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);
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 || sqe->splice_fd_in)
4599 if (req->flags & REQ_F_FIXED_FILE)
4602 req->close.fd = READ_ONCE(sqe->fd);
4606 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4608 struct files_struct *files = current->files;
4609 struct io_close *close = &req->close;
4610 struct fdtable *fdt;
4611 struct file *file = NULL;
4614 spin_lock(&files->file_lock);
4615 fdt = files_fdtable(files);
4616 if (close->fd >= fdt->max_fds) {
4617 spin_unlock(&files->file_lock);
4620 file = fdt->fd[close->fd];
4621 if (!file || file->f_op == &io_uring_fops) {
4622 spin_unlock(&files->file_lock);
4627 /* if the file has a flush method, be safe and punt to async */
4628 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4629 spin_unlock(&files->file_lock);
4633 ret = __close_fd_get_file(close->fd, &file);
4634 spin_unlock(&files->file_lock);
4641 /* No ->flush() or already async, safely close from here */
4642 ret = filp_close(file, current->files);
4648 __io_req_complete(req, issue_flags, ret, 0);
4652 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4654 struct io_ring_ctx *ctx = req->ctx;
4656 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4658 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4662 req->sync.off = READ_ONCE(sqe->off);
4663 req->sync.len = READ_ONCE(sqe->len);
4664 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4668 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4672 /* sync_file_range always requires a blocking context */
4673 if (issue_flags & IO_URING_F_NONBLOCK)
4676 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4680 io_req_complete(req, ret);
4684 #if defined(CONFIG_NET)
4685 static int io_setup_async_msg(struct io_kiocb *req,
4686 struct io_async_msghdr *kmsg)
4688 struct io_async_msghdr *async_msg = req->async_data;
4692 if (io_alloc_async_data(req)) {
4693 kfree(kmsg->free_iov);
4696 async_msg = req->async_data;
4697 req->flags |= REQ_F_NEED_CLEANUP;
4698 memcpy(async_msg, kmsg, sizeof(*kmsg));
4699 async_msg->msg.msg_name = &async_msg->addr;
4700 /* if were using fast_iov, set it to the new one */
4701 if (!async_msg->free_iov)
4702 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4707 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4708 struct io_async_msghdr *iomsg)
4710 iomsg->msg.msg_name = &iomsg->addr;
4711 iomsg->free_iov = iomsg->fast_iov;
4712 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4713 req->sr_msg.msg_flags, &iomsg->free_iov);
4716 static int io_sendmsg_prep_async(struct io_kiocb *req)
4720 ret = io_sendmsg_copy_hdr(req, req->async_data);
4722 req->flags |= REQ_F_NEED_CLEANUP;
4726 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4728 struct io_sr_msg *sr = &req->sr_msg;
4730 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4733 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4734 sr->len = READ_ONCE(sqe->len);
4735 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4736 if (sr->msg_flags & MSG_DONTWAIT)
4737 req->flags |= REQ_F_NOWAIT;
4739 #ifdef CONFIG_COMPAT
4740 if (req->ctx->compat)
4741 sr->msg_flags |= MSG_CMSG_COMPAT;
4746 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4748 struct io_async_msghdr iomsg, *kmsg;
4749 struct socket *sock;
4754 sock = sock_from_file(req->file);
4755 if (unlikely(!sock))
4758 kmsg = req->async_data;
4760 ret = io_sendmsg_copy_hdr(req, &iomsg);
4766 flags = req->sr_msg.msg_flags;
4767 if (issue_flags & IO_URING_F_NONBLOCK)
4768 flags |= MSG_DONTWAIT;
4769 if (flags & MSG_WAITALL)
4770 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4772 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4773 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4774 return io_setup_async_msg(req, kmsg);
4775 if (ret == -ERESTARTSYS)
4778 /* fast path, check for non-NULL to avoid function call */
4780 kfree(kmsg->free_iov);
4781 req->flags &= ~REQ_F_NEED_CLEANUP;
4784 __io_req_complete(req, issue_flags, ret, 0);
4788 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4790 struct io_sr_msg *sr = &req->sr_msg;
4793 struct socket *sock;
4798 sock = sock_from_file(req->file);
4799 if (unlikely(!sock))
4802 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4806 msg.msg_name = NULL;
4807 msg.msg_control = NULL;
4808 msg.msg_controllen = 0;
4809 msg.msg_namelen = 0;
4811 flags = req->sr_msg.msg_flags;
4812 if (issue_flags & IO_URING_F_NONBLOCK)
4813 flags |= MSG_DONTWAIT;
4814 if (flags & MSG_WAITALL)
4815 min_ret = iov_iter_count(&msg.msg_iter);
4817 msg.msg_flags = flags;
4818 ret = sock_sendmsg(sock, &msg);
4819 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4821 if (ret == -ERESTARTSYS)
4826 __io_req_complete(req, issue_flags, ret, 0);
4830 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4831 struct io_async_msghdr *iomsg)
4833 struct io_sr_msg *sr = &req->sr_msg;
4834 struct iovec __user *uiov;
4838 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4839 &iomsg->uaddr, &uiov, &iov_len);
4843 if (req->flags & REQ_F_BUFFER_SELECT) {
4846 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4848 sr->len = iomsg->fast_iov[0].iov_len;
4849 iomsg->free_iov = NULL;
4851 iomsg->free_iov = iomsg->fast_iov;
4852 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4853 &iomsg->free_iov, &iomsg->msg.msg_iter,
4862 #ifdef CONFIG_COMPAT
4863 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4864 struct io_async_msghdr *iomsg)
4866 struct io_sr_msg *sr = &req->sr_msg;
4867 struct compat_iovec __user *uiov;
4872 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4877 uiov = compat_ptr(ptr);
4878 if (req->flags & REQ_F_BUFFER_SELECT) {
4879 compat_ssize_t clen;
4883 if (!access_ok(uiov, sizeof(*uiov)))
4885 if (__get_user(clen, &uiov->iov_len))
4890 iomsg->free_iov = NULL;
4892 iomsg->free_iov = iomsg->fast_iov;
4893 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4894 UIO_FASTIOV, &iomsg->free_iov,
4895 &iomsg->msg.msg_iter, true);
4904 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4905 struct io_async_msghdr *iomsg)
4907 iomsg->msg.msg_name = &iomsg->addr;
4909 #ifdef CONFIG_COMPAT
4910 if (req->ctx->compat)
4911 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4914 return __io_recvmsg_copy_hdr(req, iomsg);
4917 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4920 struct io_sr_msg *sr = &req->sr_msg;
4921 struct io_buffer *kbuf;
4923 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4928 req->flags |= REQ_F_BUFFER_SELECTED;
4932 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4934 return io_put_kbuf(req, req->sr_msg.kbuf);
4937 static int io_recvmsg_prep_async(struct io_kiocb *req)
4941 ret = io_recvmsg_copy_hdr(req, req->async_data);
4943 req->flags |= REQ_F_NEED_CLEANUP;
4947 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4949 struct io_sr_msg *sr = &req->sr_msg;
4951 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4954 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4955 sr->len = READ_ONCE(sqe->len);
4956 sr->bgid = READ_ONCE(sqe->buf_group);
4957 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4958 if (sr->msg_flags & MSG_DONTWAIT)
4959 req->flags |= REQ_F_NOWAIT;
4961 #ifdef CONFIG_COMPAT
4962 if (req->ctx->compat)
4963 sr->msg_flags |= MSG_CMSG_COMPAT;
4968 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4970 struct io_async_msghdr iomsg, *kmsg;
4971 struct socket *sock;
4972 struct io_buffer *kbuf;
4975 int ret, cflags = 0;
4976 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4978 sock = sock_from_file(req->file);
4979 if (unlikely(!sock))
4982 kmsg = req->async_data;
4984 ret = io_recvmsg_copy_hdr(req, &iomsg);
4990 if (req->flags & REQ_F_BUFFER_SELECT) {
4991 kbuf = io_recv_buffer_select(req, !force_nonblock);
4993 return PTR_ERR(kbuf);
4994 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4995 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4996 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4997 1, req->sr_msg.len);
5000 flags = req->sr_msg.msg_flags;
5002 flags |= MSG_DONTWAIT;
5003 if (flags & MSG_WAITALL)
5004 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5006 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5007 kmsg->uaddr, flags);
5008 if (force_nonblock && ret == -EAGAIN)
5009 return io_setup_async_msg(req, kmsg);
5010 if (ret == -ERESTARTSYS)
5013 if (req->flags & REQ_F_BUFFER_SELECTED)
5014 cflags = io_put_recv_kbuf(req);
5015 /* fast path, check for non-NULL to avoid function call */
5017 kfree(kmsg->free_iov);
5018 req->flags &= ~REQ_F_NEED_CLEANUP;
5019 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5021 __io_req_complete(req, issue_flags, ret, cflags);
5025 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5027 struct io_buffer *kbuf;
5028 struct io_sr_msg *sr = &req->sr_msg;
5030 void __user *buf = sr->buf;
5031 struct socket *sock;
5035 int ret, cflags = 0;
5036 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5038 sock = sock_from_file(req->file);
5039 if (unlikely(!sock))
5042 if (req->flags & REQ_F_BUFFER_SELECT) {
5043 kbuf = io_recv_buffer_select(req, !force_nonblock);
5045 return PTR_ERR(kbuf);
5046 buf = u64_to_user_ptr(kbuf->addr);
5049 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5053 msg.msg_name = NULL;
5054 msg.msg_control = NULL;
5055 msg.msg_controllen = 0;
5056 msg.msg_namelen = 0;
5057 msg.msg_iocb = NULL;
5060 flags = req->sr_msg.msg_flags;
5062 flags |= MSG_DONTWAIT;
5063 if (flags & MSG_WAITALL)
5064 min_ret = iov_iter_count(&msg.msg_iter);
5066 ret = sock_recvmsg(sock, &msg, flags);
5067 if (force_nonblock && ret == -EAGAIN)
5069 if (ret == -ERESTARTSYS)
5072 if (req->flags & REQ_F_BUFFER_SELECTED)
5073 cflags = io_put_recv_kbuf(req);
5074 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5076 __io_req_complete(req, issue_flags, ret, cflags);
5080 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5082 struct io_accept *accept = &req->accept;
5084 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5086 if (sqe->ioprio || sqe->len || sqe->buf_index)
5089 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5090 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5091 accept->flags = READ_ONCE(sqe->accept_flags);
5092 accept->nofile = rlimit(RLIMIT_NOFILE);
5094 accept->file_slot = READ_ONCE(sqe->file_index);
5095 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5096 (accept->flags & SOCK_CLOEXEC)))
5098 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5100 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5101 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5105 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5107 struct io_accept *accept = &req->accept;
5108 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5109 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5110 bool fixed = !!accept->file_slot;
5114 if (req->file->f_flags & O_NONBLOCK)
5115 req->flags |= REQ_F_NOWAIT;
5118 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5119 if (unlikely(fd < 0))
5122 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5127 ret = PTR_ERR(file);
5128 if (ret == -EAGAIN && force_nonblock)
5130 if (ret == -ERESTARTSYS)
5133 } else if (!fixed) {
5134 fd_install(fd, file);
5137 ret = io_install_fixed_file(req, file, issue_flags,
5138 accept->file_slot - 1);
5140 __io_req_complete(req, issue_flags, ret, 0);
5144 static int io_connect_prep_async(struct io_kiocb *req)
5146 struct io_async_connect *io = req->async_data;
5147 struct io_connect *conn = &req->connect;
5149 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5152 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5154 struct io_connect *conn = &req->connect;
5156 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5158 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5162 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5163 conn->addr_len = READ_ONCE(sqe->addr2);
5167 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5169 struct io_async_connect __io, *io;
5170 unsigned file_flags;
5172 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5174 if (req->async_data) {
5175 io = req->async_data;
5177 ret = move_addr_to_kernel(req->connect.addr,
5178 req->connect.addr_len,
5185 file_flags = force_nonblock ? O_NONBLOCK : 0;
5187 ret = __sys_connect_file(req->file, &io->address,
5188 req->connect.addr_len, file_flags);
5189 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5190 if (req->async_data)
5192 if (io_alloc_async_data(req)) {
5196 memcpy(req->async_data, &__io, sizeof(__io));
5199 if (ret == -ERESTARTSYS)
5204 __io_req_complete(req, issue_flags, ret, 0);
5207 #else /* !CONFIG_NET */
5208 #define IO_NETOP_FN(op) \
5209 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5211 return -EOPNOTSUPP; \
5214 #define IO_NETOP_PREP(op) \
5216 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5218 return -EOPNOTSUPP; \
5221 #define IO_NETOP_PREP_ASYNC(op) \
5223 static int io_##op##_prep_async(struct io_kiocb *req) \
5225 return -EOPNOTSUPP; \
5228 IO_NETOP_PREP_ASYNC(sendmsg);
5229 IO_NETOP_PREP_ASYNC(recvmsg);
5230 IO_NETOP_PREP_ASYNC(connect);
5231 IO_NETOP_PREP(accept);
5234 #endif /* CONFIG_NET */
5236 struct io_poll_table {
5237 struct poll_table_struct pt;
5238 struct io_kiocb *req;
5243 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5244 __poll_t mask, io_req_tw_func_t func)
5246 /* for instances that support it check for an event match first: */
5247 if (mask && !(mask & poll->events))
5250 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5252 list_del_init(&poll->wait.entry);
5255 req->io_task_work.func = func;
5258 * If this fails, then the task is exiting. When a task exits, the
5259 * work gets canceled, so just cancel this request as well instead
5260 * of executing it. We can't safely execute it anyway, as we may not
5261 * have the needed state needed for it anyway.
5263 io_req_task_work_add(req);
5267 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5268 __acquires(&req->ctx->completion_lock)
5270 struct io_ring_ctx *ctx = req->ctx;
5272 /* req->task == current here, checking PF_EXITING is safe */
5273 if (unlikely(req->task->flags & PF_EXITING))
5274 WRITE_ONCE(poll->canceled, true);
5276 if (!req->result && !READ_ONCE(poll->canceled)) {
5277 struct poll_table_struct pt = { ._key = poll->events };
5279 req->result = vfs_poll(req->file, &pt) & poll->events;
5282 spin_lock(&ctx->completion_lock);
5283 if (!req->result && !READ_ONCE(poll->canceled)) {
5284 add_wait_queue(poll->head, &poll->wait);
5291 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5293 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5294 if (req->opcode == IORING_OP_POLL_ADD)
5295 return req->async_data;
5296 return req->apoll->double_poll;
5299 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5301 if (req->opcode == IORING_OP_POLL_ADD)
5303 return &req->apoll->poll;
5306 static void io_poll_remove_double(struct io_kiocb *req)
5307 __must_hold(&req->ctx->completion_lock)
5309 struct io_poll_iocb *poll = io_poll_get_double(req);
5311 lockdep_assert_held(&req->ctx->completion_lock);
5313 if (poll && poll->head) {
5314 struct wait_queue_head *head = poll->head;
5316 spin_lock_irq(&head->lock);
5317 list_del_init(&poll->wait.entry);
5318 if (poll->wait.private)
5321 spin_unlock_irq(&head->lock);
5325 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5326 __must_hold(&req->ctx->completion_lock)
5328 struct io_ring_ctx *ctx = req->ctx;
5329 unsigned flags = IORING_CQE_F_MORE;
5332 if (READ_ONCE(req->poll.canceled)) {
5334 req->poll.events |= EPOLLONESHOT;
5336 error = mangle_poll(mask);
5338 if (req->poll.events & EPOLLONESHOT)
5340 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5341 req->poll.done = true;
5344 if (flags & IORING_CQE_F_MORE)
5347 return !(flags & IORING_CQE_F_MORE);
5350 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5351 __must_hold(&req->ctx->completion_lock)
5355 done = __io_poll_complete(req, mask);
5356 io_commit_cqring(req->ctx);
5360 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5362 struct io_ring_ctx *ctx = req->ctx;
5363 struct io_kiocb *nxt;
5365 if (io_poll_rewait(req, &req->poll)) {
5366 spin_unlock(&ctx->completion_lock);
5370 done = __io_poll_complete(req, req->result);
5372 io_poll_remove_double(req);
5373 hash_del(&req->hash_node);
5376 add_wait_queue(req->poll.head, &req->poll.wait);
5378 io_commit_cqring(ctx);
5379 spin_unlock(&ctx->completion_lock);
5380 io_cqring_ev_posted(ctx);
5383 nxt = io_put_req_find_next(req);
5385 io_req_task_submit(nxt, locked);
5390 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5391 int sync, void *key)
5393 struct io_kiocb *req = wait->private;
5394 struct io_poll_iocb *poll = io_poll_get_single(req);
5395 __poll_t mask = key_to_poll(key);
5396 unsigned long flags;
5398 /* for instances that support it check for an event match first: */
5399 if (mask && !(mask & poll->events))
5401 if (!(poll->events & EPOLLONESHOT))
5402 return poll->wait.func(&poll->wait, mode, sync, key);
5404 list_del_init(&wait->entry);
5409 spin_lock_irqsave(&poll->head->lock, flags);
5410 done = list_empty(&poll->wait.entry);
5412 list_del_init(&poll->wait.entry);
5413 /* make sure double remove sees this as being gone */
5414 wait->private = NULL;
5415 spin_unlock_irqrestore(&poll->head->lock, flags);
5417 /* use wait func handler, so it matches the rq type */
5418 poll->wait.func(&poll->wait, mode, sync, key);
5425 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5426 wait_queue_func_t wake_func)
5430 poll->canceled = false;
5431 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5432 /* mask in events that we always want/need */
5433 poll->events = events | IO_POLL_UNMASK;
5434 INIT_LIST_HEAD(&poll->wait.entry);
5435 init_waitqueue_func_entry(&poll->wait, wake_func);
5438 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5439 struct wait_queue_head *head,
5440 struct io_poll_iocb **poll_ptr)
5442 struct io_kiocb *req = pt->req;
5445 * The file being polled uses multiple waitqueues for poll handling
5446 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5449 if (unlikely(pt->nr_entries)) {
5450 struct io_poll_iocb *poll_one = poll;
5452 /* double add on the same waitqueue head, ignore */
5453 if (poll_one->head == head)
5455 /* already have a 2nd entry, fail a third attempt */
5457 if ((*poll_ptr)->head == head)
5459 pt->error = -EINVAL;
5463 * Can't handle multishot for double wait for now, turn it
5464 * into one-shot mode.
5466 if (!(poll_one->events & EPOLLONESHOT))
5467 poll_one->events |= EPOLLONESHOT;
5468 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5470 pt->error = -ENOMEM;
5473 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5475 poll->wait.private = req;
5482 if (poll->events & EPOLLEXCLUSIVE)
5483 add_wait_queue_exclusive(head, &poll->wait);
5485 add_wait_queue(head, &poll->wait);
5488 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5489 struct poll_table_struct *p)
5491 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5492 struct async_poll *apoll = pt->req->apoll;
5494 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5497 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5499 struct async_poll *apoll = req->apoll;
5500 struct io_ring_ctx *ctx = req->ctx;
5502 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5504 if (io_poll_rewait(req, &apoll->poll)) {
5505 spin_unlock(&ctx->completion_lock);
5509 hash_del(&req->hash_node);
5510 io_poll_remove_double(req);
5511 spin_unlock(&ctx->completion_lock);
5513 if (!READ_ONCE(apoll->poll.canceled))
5514 io_req_task_submit(req, locked);
5516 io_req_complete_failed(req, -ECANCELED);
5519 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5522 struct io_kiocb *req = wait->private;
5523 struct io_poll_iocb *poll = &req->apoll->poll;
5525 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5528 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5531 static void io_poll_req_insert(struct io_kiocb *req)
5533 struct io_ring_ctx *ctx = req->ctx;
5534 struct hlist_head *list;
5536 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5537 hlist_add_head(&req->hash_node, list);
5540 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5541 struct io_poll_iocb *poll,
5542 struct io_poll_table *ipt, __poll_t mask,
5543 wait_queue_func_t wake_func)
5544 __acquires(&ctx->completion_lock)
5546 struct io_ring_ctx *ctx = req->ctx;
5547 bool cancel = false;
5549 INIT_HLIST_NODE(&req->hash_node);
5550 io_init_poll_iocb(poll, mask, wake_func);
5551 poll->file = req->file;
5552 poll->wait.private = req;
5554 ipt->pt._key = mask;
5557 ipt->nr_entries = 0;
5559 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5560 if (unlikely(!ipt->nr_entries) && !ipt->error)
5561 ipt->error = -EINVAL;
5563 spin_lock(&ctx->completion_lock);
5564 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5565 io_poll_remove_double(req);
5566 if (likely(poll->head)) {
5567 spin_lock_irq(&poll->head->lock);
5568 if (unlikely(list_empty(&poll->wait.entry))) {
5574 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5575 list_del_init(&poll->wait.entry);
5577 WRITE_ONCE(poll->canceled, true);
5578 else if (!poll->done) /* actually waiting for an event */
5579 io_poll_req_insert(req);
5580 spin_unlock_irq(&poll->head->lock);
5592 static int io_arm_poll_handler(struct io_kiocb *req)
5594 const struct io_op_def *def = &io_op_defs[req->opcode];
5595 struct io_ring_ctx *ctx = req->ctx;
5596 struct async_poll *apoll;
5597 struct io_poll_table ipt;
5598 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5601 if (!req->file || !file_can_poll(req->file))
5602 return IO_APOLL_ABORTED;
5603 if (req->flags & REQ_F_POLLED)
5604 return IO_APOLL_ABORTED;
5605 if (!def->pollin && !def->pollout)
5606 return IO_APOLL_ABORTED;
5610 mask |= POLLIN | POLLRDNORM;
5612 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5613 if ((req->opcode == IORING_OP_RECVMSG) &&
5614 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5618 mask |= POLLOUT | POLLWRNORM;
5621 /* if we can't nonblock try, then no point in arming a poll handler */
5622 if (!io_file_supports_nowait(req, rw))
5623 return IO_APOLL_ABORTED;
5625 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5626 if (unlikely(!apoll))
5627 return IO_APOLL_ABORTED;
5628 apoll->double_poll = NULL;
5630 req->flags |= REQ_F_POLLED;
5631 ipt.pt._qproc = io_async_queue_proc;
5632 io_req_set_refcount(req);
5634 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5636 spin_unlock(&ctx->completion_lock);
5637 if (ret || ipt.error)
5638 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5640 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5641 mask, apoll->poll.events);
5645 static bool __io_poll_remove_one(struct io_kiocb *req,
5646 struct io_poll_iocb *poll, bool do_cancel)
5647 __must_hold(&req->ctx->completion_lock)
5649 bool do_complete = false;
5653 spin_lock_irq(&poll->head->lock);
5655 WRITE_ONCE(poll->canceled, true);
5656 if (!list_empty(&poll->wait.entry)) {
5657 list_del_init(&poll->wait.entry);
5660 spin_unlock_irq(&poll->head->lock);
5661 hash_del(&req->hash_node);
5665 static bool io_poll_remove_one(struct io_kiocb *req)
5666 __must_hold(&req->ctx->completion_lock)
5670 io_poll_remove_double(req);
5671 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5674 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5675 io_commit_cqring(req->ctx);
5677 io_put_req_deferred(req);
5683 * Returns true if we found and killed one or more poll requests
5685 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5688 struct hlist_node *tmp;
5689 struct io_kiocb *req;
5692 spin_lock(&ctx->completion_lock);
5693 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5694 struct hlist_head *list;
5696 list = &ctx->cancel_hash[i];
5697 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5698 if (io_match_task(req, tsk, cancel_all))
5699 posted += io_poll_remove_one(req);
5702 spin_unlock(&ctx->completion_lock);
5705 io_cqring_ev_posted(ctx);
5710 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5712 __must_hold(&ctx->completion_lock)
5714 struct hlist_head *list;
5715 struct io_kiocb *req;
5717 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5718 hlist_for_each_entry(req, list, hash_node) {
5719 if (sqe_addr != req->user_data)
5721 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5728 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5730 __must_hold(&ctx->completion_lock)
5732 struct io_kiocb *req;
5734 req = io_poll_find(ctx, sqe_addr, poll_only);
5737 if (io_poll_remove_one(req))
5743 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5748 events = READ_ONCE(sqe->poll32_events);
5750 events = swahw32(events);
5752 if (!(flags & IORING_POLL_ADD_MULTI))
5753 events |= EPOLLONESHOT;
5754 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5757 static int io_poll_update_prep(struct io_kiocb *req,
5758 const struct io_uring_sqe *sqe)
5760 struct io_poll_update *upd = &req->poll_update;
5763 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5765 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5767 flags = READ_ONCE(sqe->len);
5768 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5769 IORING_POLL_ADD_MULTI))
5771 /* meaningless without update */
5772 if (flags == IORING_POLL_ADD_MULTI)
5775 upd->old_user_data = READ_ONCE(sqe->addr);
5776 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5777 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5779 upd->new_user_data = READ_ONCE(sqe->off);
5780 if (!upd->update_user_data && upd->new_user_data)
5782 if (upd->update_events)
5783 upd->events = io_poll_parse_events(sqe, flags);
5784 else if (sqe->poll32_events)
5790 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5793 struct io_kiocb *req = wait->private;
5794 struct io_poll_iocb *poll = &req->poll;
5796 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5799 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5800 struct poll_table_struct *p)
5802 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5804 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5807 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5809 struct io_poll_iocb *poll = &req->poll;
5812 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5814 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5816 flags = READ_ONCE(sqe->len);
5817 if (flags & ~IORING_POLL_ADD_MULTI)
5820 io_req_set_refcount(req);
5821 poll->events = io_poll_parse_events(sqe, flags);
5825 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5827 struct io_poll_iocb *poll = &req->poll;
5828 struct io_ring_ctx *ctx = req->ctx;
5829 struct io_poll_table ipt;
5832 ipt.pt._qproc = io_poll_queue_proc;
5834 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5837 if (mask) { /* no async, we'd stolen it */
5839 io_poll_complete(req, mask);
5841 spin_unlock(&ctx->completion_lock);
5844 io_cqring_ev_posted(ctx);
5845 if (poll->events & EPOLLONESHOT)
5851 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5853 struct io_ring_ctx *ctx = req->ctx;
5854 struct io_kiocb *preq;
5858 spin_lock(&ctx->completion_lock);
5859 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5865 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5867 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5872 * Don't allow racy completion with singleshot, as we cannot safely
5873 * update those. For multishot, if we're racing with completion, just
5874 * let completion re-add it.
5876 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5877 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5881 /* we now have a detached poll request. reissue. */
5885 spin_unlock(&ctx->completion_lock);
5887 io_req_complete(req, ret);
5890 /* only mask one event flags, keep behavior flags */
5891 if (req->poll_update.update_events) {
5892 preq->poll.events &= ~0xffff;
5893 preq->poll.events |= req->poll_update.events & 0xffff;
5894 preq->poll.events |= IO_POLL_UNMASK;
5896 if (req->poll_update.update_user_data)
5897 preq->user_data = req->poll_update.new_user_data;
5898 spin_unlock(&ctx->completion_lock);
5900 /* complete update request, we're done with it */
5901 io_req_complete(req, ret);
5904 ret = io_poll_add(preq, issue_flags);
5907 io_req_complete(preq, ret);
5913 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5916 io_req_complete_post(req, -ETIME, 0);
5919 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5921 struct io_timeout_data *data = container_of(timer,
5922 struct io_timeout_data, timer);
5923 struct io_kiocb *req = data->req;
5924 struct io_ring_ctx *ctx = req->ctx;
5925 unsigned long flags;
5927 spin_lock_irqsave(&ctx->timeout_lock, flags);
5928 list_del_init(&req->timeout.list);
5929 atomic_set(&req->ctx->cq_timeouts,
5930 atomic_read(&req->ctx->cq_timeouts) + 1);
5931 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5933 req->io_task_work.func = io_req_task_timeout;
5934 io_req_task_work_add(req);
5935 return HRTIMER_NORESTART;
5938 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5940 __must_hold(&ctx->timeout_lock)
5942 struct io_timeout_data *io;
5943 struct io_kiocb *req;
5946 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5947 found = user_data == req->user_data;
5952 return ERR_PTR(-ENOENT);
5954 io = req->async_data;
5955 if (hrtimer_try_to_cancel(&io->timer) == -1)
5956 return ERR_PTR(-EALREADY);
5957 list_del_init(&req->timeout.list);
5961 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5962 __must_hold(&ctx->completion_lock)
5963 __must_hold(&ctx->timeout_lock)
5965 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5968 return PTR_ERR(req);
5971 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5972 io_put_req_deferred(req);
5976 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
5978 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
5979 case IORING_TIMEOUT_BOOTTIME:
5980 return CLOCK_BOOTTIME;
5981 case IORING_TIMEOUT_REALTIME:
5982 return CLOCK_REALTIME;
5984 /* can't happen, vetted at prep time */
5988 return CLOCK_MONOTONIC;
5992 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5993 struct timespec64 *ts, enum hrtimer_mode mode)
5994 __must_hold(&ctx->timeout_lock)
5996 struct io_timeout_data *io;
5997 struct io_kiocb *req;
6000 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6001 found = user_data == req->user_data;
6008 io = req->async_data;
6009 if (hrtimer_try_to_cancel(&io->timer) == -1)
6011 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6012 io->timer.function = io_link_timeout_fn;
6013 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6017 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6018 struct timespec64 *ts, enum hrtimer_mode mode)
6019 __must_hold(&ctx->timeout_lock)
6021 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6022 struct io_timeout_data *data;
6025 return PTR_ERR(req);
6027 req->timeout.off = 0; /* noseq */
6028 data = req->async_data;
6029 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6030 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6031 data->timer.function = io_timeout_fn;
6032 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6036 static int io_timeout_remove_prep(struct io_kiocb *req,
6037 const struct io_uring_sqe *sqe)
6039 struct io_timeout_rem *tr = &req->timeout_rem;
6041 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6043 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6045 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6048 tr->ltimeout = false;
6049 tr->addr = READ_ONCE(sqe->addr);
6050 tr->flags = READ_ONCE(sqe->timeout_flags);
6051 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6052 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6054 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6055 tr->ltimeout = true;
6056 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6058 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6060 } else if (tr->flags) {
6061 /* timeout removal doesn't support flags */
6068 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6070 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6075 * Remove or update an existing timeout command
6077 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6079 struct io_timeout_rem *tr = &req->timeout_rem;
6080 struct io_ring_ctx *ctx = req->ctx;
6083 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6084 spin_lock(&ctx->completion_lock);
6085 spin_lock_irq(&ctx->timeout_lock);
6086 ret = io_timeout_cancel(ctx, tr->addr);
6087 spin_unlock_irq(&ctx->timeout_lock);
6088 spin_unlock(&ctx->completion_lock);
6090 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6092 spin_lock_irq(&ctx->timeout_lock);
6094 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6096 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6097 spin_unlock_irq(&ctx->timeout_lock);
6102 io_req_complete_post(req, ret, 0);
6106 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6107 bool is_timeout_link)
6109 struct io_timeout_data *data;
6111 u32 off = READ_ONCE(sqe->off);
6113 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6115 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6118 if (off && is_timeout_link)
6120 flags = READ_ONCE(sqe->timeout_flags);
6121 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6123 /* more than one clock specified is invalid, obviously */
6124 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6127 INIT_LIST_HEAD(&req->timeout.list);
6128 req->timeout.off = off;
6129 if (unlikely(off && !req->ctx->off_timeout_used))
6130 req->ctx->off_timeout_used = true;
6132 if (!req->async_data && io_alloc_async_data(req))
6135 data = req->async_data;
6137 data->flags = flags;
6139 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6142 data->mode = io_translate_timeout_mode(flags);
6143 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6145 if (is_timeout_link) {
6146 struct io_submit_link *link = &req->ctx->submit_state.link;
6150 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6152 req->timeout.head = link->last;
6153 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6158 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6160 struct io_ring_ctx *ctx = req->ctx;
6161 struct io_timeout_data *data = req->async_data;
6162 struct list_head *entry;
6163 u32 tail, off = req->timeout.off;
6165 spin_lock_irq(&ctx->timeout_lock);
6168 * sqe->off holds how many events that need to occur for this
6169 * timeout event to be satisfied. If it isn't set, then this is
6170 * a pure timeout request, sequence isn't used.
6172 if (io_is_timeout_noseq(req)) {
6173 entry = ctx->timeout_list.prev;
6177 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6178 req->timeout.target_seq = tail + off;
6180 /* Update the last seq here in case io_flush_timeouts() hasn't.
6181 * This is safe because ->completion_lock is held, and submissions
6182 * and completions are never mixed in the same ->completion_lock section.
6184 ctx->cq_last_tm_flush = tail;
6187 * Insertion sort, ensuring the first entry in the list is always
6188 * the one we need first.
6190 list_for_each_prev(entry, &ctx->timeout_list) {
6191 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6194 if (io_is_timeout_noseq(nxt))
6196 /* nxt.seq is behind @tail, otherwise would've been completed */
6197 if (off >= nxt->timeout.target_seq - tail)
6201 list_add(&req->timeout.list, entry);
6202 data->timer.function = io_timeout_fn;
6203 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6204 spin_unlock_irq(&ctx->timeout_lock);
6208 struct io_cancel_data {
6209 struct io_ring_ctx *ctx;
6213 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6215 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6216 struct io_cancel_data *cd = data;
6218 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6221 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6222 struct io_ring_ctx *ctx)
6224 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6225 enum io_wq_cancel cancel_ret;
6228 if (!tctx || !tctx->io_wq)
6231 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6232 switch (cancel_ret) {
6233 case IO_WQ_CANCEL_OK:
6236 case IO_WQ_CANCEL_RUNNING:
6239 case IO_WQ_CANCEL_NOTFOUND:
6247 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6249 struct io_ring_ctx *ctx = req->ctx;
6252 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6254 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6258 spin_lock(&ctx->completion_lock);
6259 spin_lock_irq(&ctx->timeout_lock);
6260 ret = io_timeout_cancel(ctx, sqe_addr);
6261 spin_unlock_irq(&ctx->timeout_lock);
6264 ret = io_poll_cancel(ctx, sqe_addr, false);
6266 spin_unlock(&ctx->completion_lock);
6270 static int io_async_cancel_prep(struct io_kiocb *req,
6271 const struct io_uring_sqe *sqe)
6273 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6275 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6277 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6281 req->cancel.addr = READ_ONCE(sqe->addr);
6285 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6287 struct io_ring_ctx *ctx = req->ctx;
6288 u64 sqe_addr = req->cancel.addr;
6289 struct io_tctx_node *node;
6292 ret = io_try_cancel_userdata(req, sqe_addr);
6296 /* slow path, try all io-wq's */
6297 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6299 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6300 struct io_uring_task *tctx = node->task->io_uring;
6302 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6306 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6310 io_req_complete_post(req, ret, 0);
6314 static int io_rsrc_update_prep(struct io_kiocb *req,
6315 const struct io_uring_sqe *sqe)
6317 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6319 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6322 req->rsrc_update.offset = READ_ONCE(sqe->off);
6323 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6324 if (!req->rsrc_update.nr_args)
6326 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6330 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6332 struct io_ring_ctx *ctx = req->ctx;
6333 struct io_uring_rsrc_update2 up;
6336 if (issue_flags & IO_URING_F_NONBLOCK)
6339 up.offset = req->rsrc_update.offset;
6340 up.data = req->rsrc_update.arg;
6345 mutex_lock(&ctx->uring_lock);
6346 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6347 &up, req->rsrc_update.nr_args);
6348 mutex_unlock(&ctx->uring_lock);
6352 __io_req_complete(req, issue_flags, ret, 0);
6356 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6358 switch (req->opcode) {
6361 case IORING_OP_READV:
6362 case IORING_OP_READ_FIXED:
6363 case IORING_OP_READ:
6364 return io_read_prep(req, sqe);
6365 case IORING_OP_WRITEV:
6366 case IORING_OP_WRITE_FIXED:
6367 case IORING_OP_WRITE:
6368 return io_write_prep(req, sqe);
6369 case IORING_OP_POLL_ADD:
6370 return io_poll_add_prep(req, sqe);
6371 case IORING_OP_POLL_REMOVE:
6372 return io_poll_update_prep(req, sqe);
6373 case IORING_OP_FSYNC:
6374 return io_fsync_prep(req, sqe);
6375 case IORING_OP_SYNC_FILE_RANGE:
6376 return io_sfr_prep(req, sqe);
6377 case IORING_OP_SENDMSG:
6378 case IORING_OP_SEND:
6379 return io_sendmsg_prep(req, sqe);
6380 case IORING_OP_RECVMSG:
6381 case IORING_OP_RECV:
6382 return io_recvmsg_prep(req, sqe);
6383 case IORING_OP_CONNECT:
6384 return io_connect_prep(req, sqe);
6385 case IORING_OP_TIMEOUT:
6386 return io_timeout_prep(req, sqe, false);
6387 case IORING_OP_TIMEOUT_REMOVE:
6388 return io_timeout_remove_prep(req, sqe);
6389 case IORING_OP_ASYNC_CANCEL:
6390 return io_async_cancel_prep(req, sqe);
6391 case IORING_OP_LINK_TIMEOUT:
6392 return io_timeout_prep(req, sqe, true);
6393 case IORING_OP_ACCEPT:
6394 return io_accept_prep(req, sqe);
6395 case IORING_OP_FALLOCATE:
6396 return io_fallocate_prep(req, sqe);
6397 case IORING_OP_OPENAT:
6398 return io_openat_prep(req, sqe);
6399 case IORING_OP_CLOSE:
6400 return io_close_prep(req, sqe);
6401 case IORING_OP_FILES_UPDATE:
6402 return io_rsrc_update_prep(req, sqe);
6403 case IORING_OP_STATX:
6404 return io_statx_prep(req, sqe);
6405 case IORING_OP_FADVISE:
6406 return io_fadvise_prep(req, sqe);
6407 case IORING_OP_MADVISE:
6408 return io_madvise_prep(req, sqe);
6409 case IORING_OP_OPENAT2:
6410 return io_openat2_prep(req, sqe);
6411 case IORING_OP_EPOLL_CTL:
6412 return io_epoll_ctl_prep(req, sqe);
6413 case IORING_OP_SPLICE:
6414 return io_splice_prep(req, sqe);
6415 case IORING_OP_PROVIDE_BUFFERS:
6416 return io_provide_buffers_prep(req, sqe);
6417 case IORING_OP_REMOVE_BUFFERS:
6418 return io_remove_buffers_prep(req, sqe);
6420 return io_tee_prep(req, sqe);
6421 case IORING_OP_SHUTDOWN:
6422 return io_shutdown_prep(req, sqe);
6423 case IORING_OP_RENAMEAT:
6424 return io_renameat_prep(req, sqe);
6425 case IORING_OP_UNLINKAT:
6426 return io_unlinkat_prep(req, sqe);
6427 case IORING_OP_MKDIRAT:
6428 return io_mkdirat_prep(req, sqe);
6429 case IORING_OP_SYMLINKAT:
6430 return io_symlinkat_prep(req, sqe);
6431 case IORING_OP_LINKAT:
6432 return io_linkat_prep(req, sqe);
6435 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6440 static int io_req_prep_async(struct io_kiocb *req)
6442 if (!io_op_defs[req->opcode].needs_async_setup)
6444 if (WARN_ON_ONCE(req->async_data))
6446 if (io_alloc_async_data(req))
6449 switch (req->opcode) {
6450 case IORING_OP_READV:
6451 return io_rw_prep_async(req, READ);
6452 case IORING_OP_WRITEV:
6453 return io_rw_prep_async(req, WRITE);
6454 case IORING_OP_SENDMSG:
6455 return io_sendmsg_prep_async(req);
6456 case IORING_OP_RECVMSG:
6457 return io_recvmsg_prep_async(req);
6458 case IORING_OP_CONNECT:
6459 return io_connect_prep_async(req);
6461 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6466 static u32 io_get_sequence(struct io_kiocb *req)
6468 u32 seq = req->ctx->cached_sq_head;
6470 /* need original cached_sq_head, but it was increased for each req */
6471 io_for_each_link(req, req)
6476 static bool io_drain_req(struct io_kiocb *req)
6478 struct io_kiocb *pos;
6479 struct io_ring_ctx *ctx = req->ctx;
6480 struct io_defer_entry *de;
6484 if (req->flags & REQ_F_FAIL) {
6485 io_req_complete_fail_submit(req);
6490 * If we need to drain a request in the middle of a link, drain the
6491 * head request and the next request/link after the current link.
6492 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6493 * maintained for every request of our link.
6495 if (ctx->drain_next) {
6496 req->flags |= REQ_F_IO_DRAIN;
6497 ctx->drain_next = false;
6499 /* not interested in head, start from the first linked */
6500 io_for_each_link(pos, req->link) {
6501 if (pos->flags & REQ_F_IO_DRAIN) {
6502 ctx->drain_next = true;
6503 req->flags |= REQ_F_IO_DRAIN;
6508 /* Still need defer if there is pending req in defer list. */
6509 if (likely(list_empty_careful(&ctx->defer_list) &&
6510 !(req->flags & REQ_F_IO_DRAIN))) {
6511 ctx->drain_active = false;
6515 seq = io_get_sequence(req);
6516 /* Still a chance to pass the sequence check */
6517 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6520 ret = io_req_prep_async(req);
6523 io_prep_async_link(req);
6524 de = kmalloc(sizeof(*de), GFP_KERNEL);
6528 io_req_complete_failed(req, ret);
6532 spin_lock(&ctx->completion_lock);
6533 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6534 spin_unlock(&ctx->completion_lock);
6536 io_queue_async_work(req, NULL);
6540 trace_io_uring_defer(ctx, req, req->user_data);
6543 list_add_tail(&de->list, &ctx->defer_list);
6544 spin_unlock(&ctx->completion_lock);
6548 static void io_clean_op(struct io_kiocb *req)
6550 if (req->flags & REQ_F_BUFFER_SELECTED) {
6551 switch (req->opcode) {
6552 case IORING_OP_READV:
6553 case IORING_OP_READ_FIXED:
6554 case IORING_OP_READ:
6555 kfree((void *)(unsigned long)req->rw.addr);
6557 case IORING_OP_RECVMSG:
6558 case IORING_OP_RECV:
6559 kfree(req->sr_msg.kbuf);
6564 if (req->flags & REQ_F_NEED_CLEANUP) {
6565 switch (req->opcode) {
6566 case IORING_OP_READV:
6567 case IORING_OP_READ_FIXED:
6568 case IORING_OP_READ:
6569 case IORING_OP_WRITEV:
6570 case IORING_OP_WRITE_FIXED:
6571 case IORING_OP_WRITE: {
6572 struct io_async_rw *io = req->async_data;
6574 kfree(io->free_iovec);
6577 case IORING_OP_RECVMSG:
6578 case IORING_OP_SENDMSG: {
6579 struct io_async_msghdr *io = req->async_data;
6581 kfree(io->free_iov);
6584 case IORING_OP_SPLICE:
6586 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6587 io_put_file(req->splice.file_in);
6589 case IORING_OP_OPENAT:
6590 case IORING_OP_OPENAT2:
6591 if (req->open.filename)
6592 putname(req->open.filename);
6594 case IORING_OP_RENAMEAT:
6595 putname(req->rename.oldpath);
6596 putname(req->rename.newpath);
6598 case IORING_OP_UNLINKAT:
6599 putname(req->unlink.filename);
6601 case IORING_OP_MKDIRAT:
6602 putname(req->mkdir.filename);
6604 case IORING_OP_SYMLINKAT:
6605 putname(req->symlink.oldpath);
6606 putname(req->symlink.newpath);
6608 case IORING_OP_LINKAT:
6609 putname(req->hardlink.oldpath);
6610 putname(req->hardlink.newpath);
6614 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6615 kfree(req->apoll->double_poll);
6619 if (req->flags & REQ_F_INFLIGHT) {
6620 struct io_uring_task *tctx = req->task->io_uring;
6622 atomic_dec(&tctx->inflight_tracked);
6624 if (req->flags & REQ_F_CREDS)
6625 put_cred(req->creds);
6627 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6630 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6632 struct io_ring_ctx *ctx = req->ctx;
6633 const struct cred *creds = NULL;
6636 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6637 creds = override_creds(req->creds);
6639 switch (req->opcode) {
6641 ret = io_nop(req, issue_flags);
6643 case IORING_OP_READV:
6644 case IORING_OP_READ_FIXED:
6645 case IORING_OP_READ:
6646 ret = io_read(req, issue_flags);
6648 case IORING_OP_WRITEV:
6649 case IORING_OP_WRITE_FIXED:
6650 case IORING_OP_WRITE:
6651 ret = io_write(req, issue_flags);
6653 case IORING_OP_FSYNC:
6654 ret = io_fsync(req, issue_flags);
6656 case IORING_OP_POLL_ADD:
6657 ret = io_poll_add(req, issue_flags);
6659 case IORING_OP_POLL_REMOVE:
6660 ret = io_poll_update(req, issue_flags);
6662 case IORING_OP_SYNC_FILE_RANGE:
6663 ret = io_sync_file_range(req, issue_flags);
6665 case IORING_OP_SENDMSG:
6666 ret = io_sendmsg(req, issue_flags);
6668 case IORING_OP_SEND:
6669 ret = io_send(req, issue_flags);
6671 case IORING_OP_RECVMSG:
6672 ret = io_recvmsg(req, issue_flags);
6674 case IORING_OP_RECV:
6675 ret = io_recv(req, issue_flags);
6677 case IORING_OP_TIMEOUT:
6678 ret = io_timeout(req, issue_flags);
6680 case IORING_OP_TIMEOUT_REMOVE:
6681 ret = io_timeout_remove(req, issue_flags);
6683 case IORING_OP_ACCEPT:
6684 ret = io_accept(req, issue_flags);
6686 case IORING_OP_CONNECT:
6687 ret = io_connect(req, issue_flags);
6689 case IORING_OP_ASYNC_CANCEL:
6690 ret = io_async_cancel(req, issue_flags);
6692 case IORING_OP_FALLOCATE:
6693 ret = io_fallocate(req, issue_flags);
6695 case IORING_OP_OPENAT:
6696 ret = io_openat(req, issue_flags);
6698 case IORING_OP_CLOSE:
6699 ret = io_close(req, issue_flags);
6701 case IORING_OP_FILES_UPDATE:
6702 ret = io_files_update(req, issue_flags);
6704 case IORING_OP_STATX:
6705 ret = io_statx(req, issue_flags);
6707 case IORING_OP_FADVISE:
6708 ret = io_fadvise(req, issue_flags);
6710 case IORING_OP_MADVISE:
6711 ret = io_madvise(req, issue_flags);
6713 case IORING_OP_OPENAT2:
6714 ret = io_openat2(req, issue_flags);
6716 case IORING_OP_EPOLL_CTL:
6717 ret = io_epoll_ctl(req, issue_flags);
6719 case IORING_OP_SPLICE:
6720 ret = io_splice(req, issue_flags);
6722 case IORING_OP_PROVIDE_BUFFERS:
6723 ret = io_provide_buffers(req, issue_flags);
6725 case IORING_OP_REMOVE_BUFFERS:
6726 ret = io_remove_buffers(req, issue_flags);
6729 ret = io_tee(req, issue_flags);
6731 case IORING_OP_SHUTDOWN:
6732 ret = io_shutdown(req, issue_flags);
6734 case IORING_OP_RENAMEAT:
6735 ret = io_renameat(req, issue_flags);
6737 case IORING_OP_UNLINKAT:
6738 ret = io_unlinkat(req, issue_flags);
6740 case IORING_OP_MKDIRAT:
6741 ret = io_mkdirat(req, issue_flags);
6743 case IORING_OP_SYMLINKAT:
6744 ret = io_symlinkat(req, issue_flags);
6746 case IORING_OP_LINKAT:
6747 ret = io_linkat(req, issue_flags);
6755 revert_creds(creds);
6758 /* If the op doesn't have a file, we're not polling for it */
6759 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6760 io_iopoll_req_issued(req);
6765 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6767 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6769 req = io_put_req_find_next(req);
6770 return req ? &req->work : NULL;
6773 static void io_wq_submit_work(struct io_wq_work *work)
6775 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6776 struct io_kiocb *timeout;
6779 /* one will be dropped by ->io_free_work() after returning to io-wq */
6780 if (!(req->flags & REQ_F_REFCOUNT))
6781 __io_req_set_refcount(req, 2);
6785 timeout = io_prep_linked_timeout(req);
6787 io_queue_linked_timeout(timeout);
6789 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6790 if (work->flags & IO_WQ_WORK_CANCEL)
6795 ret = io_issue_sqe(req, 0);
6797 * We can get EAGAIN for polled IO even though we're
6798 * forcing a sync submission from here, since we can't
6799 * wait for request slots on the block side.
6807 /* avoid locking problems by failing it from a clean context */
6809 io_req_task_queue_fail(req, ret);
6812 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6815 return &table->files[i];
6818 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6821 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6823 return (struct file *) (slot->file_ptr & FFS_MASK);
6826 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6828 unsigned long file_ptr = (unsigned long) file;
6830 if (__io_file_supports_nowait(file, READ))
6831 file_ptr |= FFS_ASYNC_READ;
6832 if (__io_file_supports_nowait(file, WRITE))
6833 file_ptr |= FFS_ASYNC_WRITE;
6834 if (S_ISREG(file_inode(file)->i_mode))
6835 file_ptr |= FFS_ISREG;
6836 file_slot->file_ptr = file_ptr;
6839 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6840 struct io_kiocb *req, int fd)
6843 unsigned long file_ptr;
6845 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6847 fd = array_index_nospec(fd, ctx->nr_user_files);
6848 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6849 file = (struct file *) (file_ptr & FFS_MASK);
6850 file_ptr &= ~FFS_MASK;
6851 /* mask in overlapping REQ_F and FFS bits */
6852 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6853 io_req_set_rsrc_node(req);
6857 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6858 struct io_kiocb *req, int fd)
6860 struct file *file = fget(fd);
6862 trace_io_uring_file_get(ctx, fd);
6864 /* we don't allow fixed io_uring files */
6865 if (file && unlikely(file->f_op == &io_uring_fops))
6866 io_req_track_inflight(req);
6870 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6871 struct io_kiocb *req, int fd, bool fixed)
6874 return io_file_get_fixed(ctx, req, fd);
6876 return io_file_get_normal(ctx, req, fd);
6879 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6881 struct io_kiocb *prev = req->timeout.prev;
6885 ret = io_try_cancel_userdata(req, prev->user_data);
6886 io_req_complete_post(req, ret ?: -ETIME, 0);
6889 io_req_complete_post(req, -ETIME, 0);
6893 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6895 struct io_timeout_data *data = container_of(timer,
6896 struct io_timeout_data, timer);
6897 struct io_kiocb *prev, *req = data->req;
6898 struct io_ring_ctx *ctx = req->ctx;
6899 unsigned long flags;
6901 spin_lock_irqsave(&ctx->timeout_lock, flags);
6902 prev = req->timeout.head;
6903 req->timeout.head = NULL;
6906 * We don't expect the list to be empty, that will only happen if we
6907 * race with the completion of the linked work.
6910 io_remove_next_linked(prev);
6911 if (!req_ref_inc_not_zero(prev))
6914 list_del(&req->timeout.list);
6915 req->timeout.prev = prev;
6916 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6918 req->io_task_work.func = io_req_task_link_timeout;
6919 io_req_task_work_add(req);
6920 return HRTIMER_NORESTART;
6923 static void io_queue_linked_timeout(struct io_kiocb *req)
6925 struct io_ring_ctx *ctx = req->ctx;
6927 spin_lock_irq(&ctx->timeout_lock);
6929 * If the back reference is NULL, then our linked request finished
6930 * before we got a chance to setup the timer
6932 if (req->timeout.head) {
6933 struct io_timeout_data *data = req->async_data;
6935 data->timer.function = io_link_timeout_fn;
6936 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6938 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6940 spin_unlock_irq(&ctx->timeout_lock);
6941 /* drop submission reference */
6945 static void __io_queue_sqe(struct io_kiocb *req)
6946 __must_hold(&req->ctx->uring_lock)
6948 struct io_kiocb *linked_timeout;
6952 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6955 * We async punt it if the file wasn't marked NOWAIT, or if the file
6956 * doesn't support non-blocking read/write attempts
6959 if (req->flags & REQ_F_COMPLETE_INLINE) {
6960 struct io_ring_ctx *ctx = req->ctx;
6961 struct io_submit_state *state = &ctx->submit_state;
6963 state->compl_reqs[state->compl_nr++] = req;
6964 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6965 io_submit_flush_completions(ctx);
6969 linked_timeout = io_prep_linked_timeout(req);
6971 io_queue_linked_timeout(linked_timeout);
6972 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6973 linked_timeout = io_prep_linked_timeout(req);
6975 switch (io_arm_poll_handler(req)) {
6976 case IO_APOLL_READY:
6978 io_unprep_linked_timeout(req);
6980 case IO_APOLL_ABORTED:
6982 * Queued up for async execution, worker will release
6983 * submit reference when the iocb is actually submitted.
6985 io_queue_async_work(req, NULL);
6990 io_queue_linked_timeout(linked_timeout);
6992 io_req_complete_failed(req, ret);
6996 static inline void io_queue_sqe(struct io_kiocb *req)
6997 __must_hold(&req->ctx->uring_lock)
6999 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7002 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7003 __io_queue_sqe(req);
7004 } else if (req->flags & REQ_F_FAIL) {
7005 io_req_complete_fail_submit(req);
7007 int ret = io_req_prep_async(req);
7010 io_req_complete_failed(req, ret);
7012 io_queue_async_work(req, NULL);
7017 * Check SQE restrictions (opcode and flags).
7019 * Returns 'true' if SQE is allowed, 'false' otherwise.
7021 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7022 struct io_kiocb *req,
7023 unsigned int sqe_flags)
7025 if (likely(!ctx->restricted))
7028 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7031 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7032 ctx->restrictions.sqe_flags_required)
7035 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7036 ctx->restrictions.sqe_flags_required))
7042 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7043 const struct io_uring_sqe *sqe)
7044 __must_hold(&ctx->uring_lock)
7046 struct io_submit_state *state;
7047 unsigned int sqe_flags;
7048 int personality, ret = 0;
7050 /* req is partially pre-initialised, see io_preinit_req() */
7051 req->opcode = READ_ONCE(sqe->opcode);
7052 /* same numerical values with corresponding REQ_F_*, safe to copy */
7053 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7054 req->user_data = READ_ONCE(sqe->user_data);
7056 req->fixed_rsrc_refs = NULL;
7057 req->task = current;
7059 /* enforce forwards compatibility on users */
7060 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7062 if (unlikely(req->opcode >= IORING_OP_LAST))
7064 if (!io_check_restriction(ctx, req, sqe_flags))
7067 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7068 !io_op_defs[req->opcode].buffer_select)
7070 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7071 ctx->drain_active = true;
7073 personality = READ_ONCE(sqe->personality);
7075 req->creds = xa_load(&ctx->personalities, personality);
7078 get_cred(req->creds);
7079 req->flags |= REQ_F_CREDS;
7081 state = &ctx->submit_state;
7084 * Plug now if we have more than 1 IO left after this, and the target
7085 * is potentially a read/write to block based storage.
7087 if (!state->plug_started && state->ios_left > 1 &&
7088 io_op_defs[req->opcode].plug) {
7089 blk_start_plug(&state->plug);
7090 state->plug_started = true;
7093 if (io_op_defs[req->opcode].needs_file) {
7094 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7095 (sqe_flags & IOSQE_FIXED_FILE));
7096 if (unlikely(!req->file))
7104 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7105 const struct io_uring_sqe *sqe)
7106 __must_hold(&ctx->uring_lock)
7108 struct io_submit_link *link = &ctx->submit_state.link;
7111 ret = io_init_req(ctx, req, sqe);
7112 if (unlikely(ret)) {
7114 /* fail even hard links since we don't submit */
7117 * we can judge a link req is failed or cancelled by if
7118 * REQ_F_FAIL is set, but the head is an exception since
7119 * it may be set REQ_F_FAIL because of other req's failure
7120 * so let's leverage req->result to distinguish if a head
7121 * is set REQ_F_FAIL because of its failure or other req's
7122 * failure so that we can set the correct ret code for it.
7123 * init result here to avoid affecting the normal path.
7125 if (!(link->head->flags & REQ_F_FAIL))
7126 req_fail_link_node(link->head, -ECANCELED);
7127 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7129 * the current req is a normal req, we should return
7130 * error and thus break the submittion loop.
7132 io_req_complete_failed(req, ret);
7135 req_fail_link_node(req, ret);
7137 ret = io_req_prep(req, sqe);
7142 /* don't need @sqe from now on */
7143 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7145 ctx->flags & IORING_SETUP_SQPOLL);
7148 * If we already have a head request, queue this one for async
7149 * submittal once the head completes. If we don't have a head but
7150 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7151 * submitted sync once the chain is complete. If none of those
7152 * conditions are true (normal request), then just queue it.
7155 struct io_kiocb *head = link->head;
7157 if (!(req->flags & REQ_F_FAIL)) {
7158 ret = io_req_prep_async(req);
7159 if (unlikely(ret)) {
7160 req_fail_link_node(req, ret);
7161 if (!(head->flags & REQ_F_FAIL))
7162 req_fail_link_node(head, -ECANCELED);
7165 trace_io_uring_link(ctx, req, head);
7166 link->last->link = req;
7169 /* last request of a link, enqueue the link */
7170 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7175 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7187 * Batched submission is done, ensure local IO is flushed out.
7189 static void io_submit_state_end(struct io_submit_state *state,
7190 struct io_ring_ctx *ctx)
7192 if (state->link.head)
7193 io_queue_sqe(state->link.head);
7194 if (state->compl_nr)
7195 io_submit_flush_completions(ctx);
7196 if (state->plug_started)
7197 blk_finish_plug(&state->plug);
7201 * Start submission side cache.
7203 static void io_submit_state_start(struct io_submit_state *state,
7204 unsigned int max_ios)
7206 state->plug_started = false;
7207 state->ios_left = max_ios;
7208 /* set only head, no need to init link_last in advance */
7209 state->link.head = NULL;
7212 static void io_commit_sqring(struct io_ring_ctx *ctx)
7214 struct io_rings *rings = ctx->rings;
7217 * Ensure any loads from the SQEs are done at this point,
7218 * since once we write the new head, the application could
7219 * write new data to them.
7221 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7225 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7226 * that is mapped by userspace. This means that care needs to be taken to
7227 * ensure that reads are stable, as we cannot rely on userspace always
7228 * being a good citizen. If members of the sqe are validated and then later
7229 * used, it's important that those reads are done through READ_ONCE() to
7230 * prevent a re-load down the line.
7232 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7234 unsigned head, mask = ctx->sq_entries - 1;
7235 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7238 * The cached sq head (or cq tail) serves two purposes:
7240 * 1) allows us to batch the cost of updating the user visible
7242 * 2) allows the kernel side to track the head on its own, even
7243 * though the application is the one updating it.
7245 head = READ_ONCE(ctx->sq_array[sq_idx]);
7246 if (likely(head < ctx->sq_entries))
7247 return &ctx->sq_sqes[head];
7249 /* drop invalid entries */
7251 WRITE_ONCE(ctx->rings->sq_dropped,
7252 READ_ONCE(ctx->rings->sq_dropped) + 1);
7256 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7257 __must_hold(&ctx->uring_lock)
7261 /* make sure SQ entry isn't read before tail */
7262 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7263 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7265 io_get_task_refs(nr);
7267 io_submit_state_start(&ctx->submit_state, nr);
7268 while (submitted < nr) {
7269 const struct io_uring_sqe *sqe;
7270 struct io_kiocb *req;
7272 req = io_alloc_req(ctx);
7273 if (unlikely(!req)) {
7275 submitted = -EAGAIN;
7278 sqe = io_get_sqe(ctx);
7279 if (unlikely(!sqe)) {
7280 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7283 /* will complete beyond this point, count as submitted */
7285 if (io_submit_sqe(ctx, req, sqe))
7289 if (unlikely(submitted != nr)) {
7290 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7291 int unused = nr - ref_used;
7293 current->io_uring->cached_refs += unused;
7294 percpu_ref_put_many(&ctx->refs, unused);
7297 io_submit_state_end(&ctx->submit_state, ctx);
7298 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7299 io_commit_sqring(ctx);
7304 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7306 return READ_ONCE(sqd->state);
7309 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7311 /* Tell userspace we may need a wakeup call */
7312 spin_lock(&ctx->completion_lock);
7313 WRITE_ONCE(ctx->rings->sq_flags,
7314 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7315 spin_unlock(&ctx->completion_lock);
7318 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7320 spin_lock(&ctx->completion_lock);
7321 WRITE_ONCE(ctx->rings->sq_flags,
7322 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7323 spin_unlock(&ctx->completion_lock);
7326 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7328 unsigned int to_submit;
7331 to_submit = io_sqring_entries(ctx);
7332 /* if we're handling multiple rings, cap submit size for fairness */
7333 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7334 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7336 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7337 unsigned nr_events = 0;
7338 const struct cred *creds = NULL;
7340 if (ctx->sq_creds != current_cred())
7341 creds = override_creds(ctx->sq_creds);
7343 mutex_lock(&ctx->uring_lock);
7344 if (!list_empty(&ctx->iopoll_list))
7345 io_do_iopoll(ctx, &nr_events, 0);
7348 * Don't submit if refs are dying, good for io_uring_register(),
7349 * but also it is relied upon by io_ring_exit_work()
7351 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7352 !(ctx->flags & IORING_SETUP_R_DISABLED))
7353 ret = io_submit_sqes(ctx, to_submit);
7354 mutex_unlock(&ctx->uring_lock);
7356 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7357 wake_up(&ctx->sqo_sq_wait);
7359 revert_creds(creds);
7365 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7367 struct io_ring_ctx *ctx;
7368 unsigned sq_thread_idle = 0;
7370 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7371 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7372 sqd->sq_thread_idle = sq_thread_idle;
7375 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7377 bool did_sig = false;
7378 struct ksignal ksig;
7380 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7381 signal_pending(current)) {
7382 mutex_unlock(&sqd->lock);
7383 if (signal_pending(current))
7384 did_sig = get_signal(&ksig);
7386 mutex_lock(&sqd->lock);
7388 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7391 static int io_sq_thread(void *data)
7393 struct io_sq_data *sqd = data;
7394 struct io_ring_ctx *ctx;
7395 unsigned long timeout = 0;
7396 char buf[TASK_COMM_LEN];
7399 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7400 set_task_comm(current, buf);
7402 if (sqd->sq_cpu != -1)
7403 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7405 set_cpus_allowed_ptr(current, cpu_online_mask);
7406 current->flags |= PF_NO_SETAFFINITY;
7408 mutex_lock(&sqd->lock);
7410 bool cap_entries, sqt_spin = false;
7412 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7413 if (io_sqd_handle_event(sqd))
7415 timeout = jiffies + sqd->sq_thread_idle;
7418 cap_entries = !list_is_singular(&sqd->ctx_list);
7419 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7420 int ret = __io_sq_thread(ctx, cap_entries);
7422 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7425 if (io_run_task_work())
7428 if (sqt_spin || !time_after(jiffies, timeout)) {
7431 timeout = jiffies + sqd->sq_thread_idle;
7435 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7436 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7437 bool needs_sched = true;
7439 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7440 io_ring_set_wakeup_flag(ctx);
7442 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7443 !list_empty_careful(&ctx->iopoll_list)) {
7444 needs_sched = false;
7447 if (io_sqring_entries(ctx)) {
7448 needs_sched = false;
7454 mutex_unlock(&sqd->lock);
7456 mutex_lock(&sqd->lock);
7458 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7459 io_ring_clear_wakeup_flag(ctx);
7462 finish_wait(&sqd->wait, &wait);
7463 timeout = jiffies + sqd->sq_thread_idle;
7466 io_uring_cancel_generic(true, sqd);
7468 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7469 io_ring_set_wakeup_flag(ctx);
7471 mutex_unlock(&sqd->lock);
7473 complete(&sqd->exited);
7477 struct io_wait_queue {
7478 struct wait_queue_entry wq;
7479 struct io_ring_ctx *ctx;
7481 unsigned nr_timeouts;
7484 static inline bool io_should_wake(struct io_wait_queue *iowq)
7486 struct io_ring_ctx *ctx = iowq->ctx;
7487 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7490 * Wake up if we have enough events, or if a timeout occurred since we
7491 * started waiting. For timeouts, we always want to return to userspace,
7492 * regardless of event count.
7494 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7497 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7498 int wake_flags, void *key)
7500 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7504 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7505 * the task, and the next invocation will do it.
7507 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7508 return autoremove_wake_function(curr, mode, wake_flags, key);
7512 static int io_run_task_work_sig(void)
7514 if (io_run_task_work())
7516 if (!signal_pending(current))
7518 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7519 return -ERESTARTSYS;
7523 /* when returns >0, the caller should retry */
7524 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7525 struct io_wait_queue *iowq,
7526 signed long *timeout)
7530 /* make sure we run task_work before checking for signals */
7531 ret = io_run_task_work_sig();
7532 if (ret || io_should_wake(iowq))
7534 /* let the caller flush overflows, retry */
7535 if (test_bit(0, &ctx->check_cq_overflow))
7538 *timeout = schedule_timeout(*timeout);
7539 return !*timeout ? -ETIME : 1;
7543 * Wait until events become available, if we don't already have some. The
7544 * application must reap them itself, as they reside on the shared cq ring.
7546 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7547 const sigset_t __user *sig, size_t sigsz,
7548 struct __kernel_timespec __user *uts)
7550 struct io_wait_queue iowq;
7551 struct io_rings *rings = ctx->rings;
7552 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7556 io_cqring_overflow_flush(ctx);
7557 if (io_cqring_events(ctx) >= min_events)
7559 if (!io_run_task_work())
7564 struct timespec64 ts;
7566 if (get_timespec64(&ts, uts))
7568 timeout = timespec64_to_jiffies(&ts);
7572 #ifdef CONFIG_COMPAT
7573 if (in_compat_syscall())
7574 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7578 ret = set_user_sigmask(sig, sigsz);
7584 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7585 iowq.wq.private = current;
7586 INIT_LIST_HEAD(&iowq.wq.entry);
7588 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7589 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7591 trace_io_uring_cqring_wait(ctx, min_events);
7593 /* if we can't even flush overflow, don't wait for more */
7594 if (!io_cqring_overflow_flush(ctx)) {
7598 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7599 TASK_INTERRUPTIBLE);
7600 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7601 finish_wait(&ctx->cq_wait, &iowq.wq);
7605 restore_saved_sigmask_unless(ret == -EINTR);
7607 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7610 static void io_free_page_table(void **table, size_t size)
7612 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7614 for (i = 0; i < nr_tables; i++)
7619 static void **io_alloc_page_table(size_t size)
7621 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7622 size_t init_size = size;
7625 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7629 for (i = 0; i < nr_tables; i++) {
7630 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7632 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7634 io_free_page_table(table, init_size);
7642 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7644 percpu_ref_exit(&ref_node->refs);
7648 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7650 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7651 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7652 unsigned long flags;
7653 bool first_add = false;
7655 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7658 while (!list_empty(&ctx->rsrc_ref_list)) {
7659 node = list_first_entry(&ctx->rsrc_ref_list,
7660 struct io_rsrc_node, node);
7661 /* recycle ref nodes in order */
7664 list_del(&node->node);
7665 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7667 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7670 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7673 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7675 struct io_rsrc_node *ref_node;
7677 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7681 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7686 INIT_LIST_HEAD(&ref_node->node);
7687 INIT_LIST_HEAD(&ref_node->rsrc_list);
7688 ref_node->done = false;
7692 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7693 struct io_rsrc_data *data_to_kill)
7695 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7696 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7699 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7701 rsrc_node->rsrc_data = data_to_kill;
7702 spin_lock_irq(&ctx->rsrc_ref_lock);
7703 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7704 spin_unlock_irq(&ctx->rsrc_ref_lock);
7706 atomic_inc(&data_to_kill->refs);
7707 percpu_ref_kill(&rsrc_node->refs);
7708 ctx->rsrc_node = NULL;
7711 if (!ctx->rsrc_node) {
7712 ctx->rsrc_node = ctx->rsrc_backup_node;
7713 ctx->rsrc_backup_node = NULL;
7717 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7719 if (ctx->rsrc_backup_node)
7721 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7722 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7725 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7729 /* As we may drop ->uring_lock, other task may have started quiesce */
7733 data->quiesce = true;
7735 ret = io_rsrc_node_switch_start(ctx);
7738 io_rsrc_node_switch(ctx, data);
7740 /* kill initial ref, already quiesced if zero */
7741 if (atomic_dec_and_test(&data->refs))
7743 mutex_unlock(&ctx->uring_lock);
7744 flush_delayed_work(&ctx->rsrc_put_work);
7745 ret = wait_for_completion_interruptible(&data->done);
7747 mutex_lock(&ctx->uring_lock);
7751 atomic_inc(&data->refs);
7752 /* wait for all works potentially completing data->done */
7753 flush_delayed_work(&ctx->rsrc_put_work);
7754 reinit_completion(&data->done);
7756 ret = io_run_task_work_sig();
7757 mutex_lock(&ctx->uring_lock);
7759 data->quiesce = false;
7764 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7766 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7767 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7769 return &data->tags[table_idx][off];
7772 static void io_rsrc_data_free(struct io_rsrc_data *data)
7774 size_t size = data->nr * sizeof(data->tags[0][0]);
7777 io_free_page_table((void **)data->tags, size);
7781 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7782 u64 __user *utags, unsigned nr,
7783 struct io_rsrc_data **pdata)
7785 struct io_rsrc_data *data;
7789 data = kzalloc(sizeof(*data), GFP_KERNEL);
7792 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7800 data->do_put = do_put;
7803 for (i = 0; i < nr; i++) {
7804 u64 *tag_slot = io_get_tag_slot(data, i);
7806 if (copy_from_user(tag_slot, &utags[i],
7812 atomic_set(&data->refs, 1);
7813 init_completion(&data->done);
7817 io_rsrc_data_free(data);
7821 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7823 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7824 GFP_KERNEL_ACCOUNT);
7825 return !!table->files;
7828 static void io_free_file_tables(struct io_file_table *table)
7830 kvfree(table->files);
7831 table->files = NULL;
7834 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7836 #if defined(CONFIG_UNIX)
7837 if (ctx->ring_sock) {
7838 struct sock *sock = ctx->ring_sock->sk;
7839 struct sk_buff *skb;
7841 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7847 for (i = 0; i < ctx->nr_user_files; i++) {
7850 file = io_file_from_index(ctx, i);
7855 io_free_file_tables(&ctx->file_table);
7856 io_rsrc_data_free(ctx->file_data);
7857 ctx->file_data = NULL;
7858 ctx->nr_user_files = 0;
7861 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7865 if (!ctx->file_data)
7867 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7869 __io_sqe_files_unregister(ctx);
7873 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7874 __releases(&sqd->lock)
7876 WARN_ON_ONCE(sqd->thread == current);
7879 * Do the dance but not conditional clear_bit() because it'd race with
7880 * other threads incrementing park_pending and setting the bit.
7882 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7883 if (atomic_dec_return(&sqd->park_pending))
7884 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7885 mutex_unlock(&sqd->lock);
7888 static void io_sq_thread_park(struct io_sq_data *sqd)
7889 __acquires(&sqd->lock)
7891 WARN_ON_ONCE(sqd->thread == current);
7893 atomic_inc(&sqd->park_pending);
7894 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7895 mutex_lock(&sqd->lock);
7897 wake_up_process(sqd->thread);
7900 static void io_sq_thread_stop(struct io_sq_data *sqd)
7902 WARN_ON_ONCE(sqd->thread == current);
7903 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7905 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7906 mutex_lock(&sqd->lock);
7908 wake_up_process(sqd->thread);
7909 mutex_unlock(&sqd->lock);
7910 wait_for_completion(&sqd->exited);
7913 static void io_put_sq_data(struct io_sq_data *sqd)
7915 if (refcount_dec_and_test(&sqd->refs)) {
7916 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7918 io_sq_thread_stop(sqd);
7923 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7925 struct io_sq_data *sqd = ctx->sq_data;
7928 io_sq_thread_park(sqd);
7929 list_del_init(&ctx->sqd_list);
7930 io_sqd_update_thread_idle(sqd);
7931 io_sq_thread_unpark(sqd);
7933 io_put_sq_data(sqd);
7934 ctx->sq_data = NULL;
7938 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7940 struct io_ring_ctx *ctx_attach;
7941 struct io_sq_data *sqd;
7944 f = fdget(p->wq_fd);
7946 return ERR_PTR(-ENXIO);
7947 if (f.file->f_op != &io_uring_fops) {
7949 return ERR_PTR(-EINVAL);
7952 ctx_attach = f.file->private_data;
7953 sqd = ctx_attach->sq_data;
7956 return ERR_PTR(-EINVAL);
7958 if (sqd->task_tgid != current->tgid) {
7960 return ERR_PTR(-EPERM);
7963 refcount_inc(&sqd->refs);
7968 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7971 struct io_sq_data *sqd;
7974 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7975 sqd = io_attach_sq_data(p);
7980 /* fall through for EPERM case, setup new sqd/task */
7981 if (PTR_ERR(sqd) != -EPERM)
7985 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7987 return ERR_PTR(-ENOMEM);
7989 atomic_set(&sqd->park_pending, 0);
7990 refcount_set(&sqd->refs, 1);
7991 INIT_LIST_HEAD(&sqd->ctx_list);
7992 mutex_init(&sqd->lock);
7993 init_waitqueue_head(&sqd->wait);
7994 init_completion(&sqd->exited);
7998 #if defined(CONFIG_UNIX)
8000 * Ensure the UNIX gc is aware of our file set, so we are certain that
8001 * the io_uring can be safely unregistered on process exit, even if we have
8002 * loops in the file referencing.
8004 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8006 struct sock *sk = ctx->ring_sock->sk;
8007 struct scm_fp_list *fpl;
8008 struct sk_buff *skb;
8011 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8015 skb = alloc_skb(0, GFP_KERNEL);
8024 fpl->user = get_uid(current_user());
8025 for (i = 0; i < nr; i++) {
8026 struct file *file = io_file_from_index(ctx, i + offset);
8030 fpl->fp[nr_files] = get_file(file);
8031 unix_inflight(fpl->user, fpl->fp[nr_files]);
8036 fpl->max = SCM_MAX_FD;
8037 fpl->count = nr_files;
8038 UNIXCB(skb).fp = fpl;
8039 skb->destructor = unix_destruct_scm;
8040 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8041 skb_queue_head(&sk->sk_receive_queue, skb);
8043 for (i = 0; i < nr_files; i++)
8054 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8055 * causes regular reference counting to break down. We rely on the UNIX
8056 * garbage collection to take care of this problem for us.
8058 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8060 unsigned left, total;
8064 left = ctx->nr_user_files;
8066 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8068 ret = __io_sqe_files_scm(ctx, this_files, total);
8072 total += this_files;
8078 while (total < ctx->nr_user_files) {
8079 struct file *file = io_file_from_index(ctx, total);
8089 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8095 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8097 struct file *file = prsrc->file;
8098 #if defined(CONFIG_UNIX)
8099 struct sock *sock = ctx->ring_sock->sk;
8100 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8101 struct sk_buff *skb;
8104 __skb_queue_head_init(&list);
8107 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8108 * remove this entry and rearrange the file array.
8110 skb = skb_dequeue(head);
8112 struct scm_fp_list *fp;
8114 fp = UNIXCB(skb).fp;
8115 for (i = 0; i < fp->count; i++) {
8118 if (fp->fp[i] != file)
8121 unix_notinflight(fp->user, fp->fp[i]);
8122 left = fp->count - 1 - i;
8124 memmove(&fp->fp[i], &fp->fp[i + 1],
8125 left * sizeof(struct file *));
8132 __skb_queue_tail(&list, skb);
8142 __skb_queue_tail(&list, skb);
8144 skb = skb_dequeue(head);
8147 if (skb_peek(&list)) {
8148 spin_lock_irq(&head->lock);
8149 while ((skb = __skb_dequeue(&list)) != NULL)
8150 __skb_queue_tail(head, skb);
8151 spin_unlock_irq(&head->lock);
8158 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8160 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8161 struct io_ring_ctx *ctx = rsrc_data->ctx;
8162 struct io_rsrc_put *prsrc, *tmp;
8164 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8165 list_del(&prsrc->list);
8168 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8170 io_ring_submit_lock(ctx, lock_ring);
8171 spin_lock(&ctx->completion_lock);
8172 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8174 io_commit_cqring(ctx);
8175 spin_unlock(&ctx->completion_lock);
8176 io_cqring_ev_posted(ctx);
8177 io_ring_submit_unlock(ctx, lock_ring);
8180 rsrc_data->do_put(ctx, prsrc);
8184 io_rsrc_node_destroy(ref_node);
8185 if (atomic_dec_and_test(&rsrc_data->refs))
8186 complete(&rsrc_data->done);
8189 static void io_rsrc_put_work(struct work_struct *work)
8191 struct io_ring_ctx *ctx;
8192 struct llist_node *node;
8194 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8195 node = llist_del_all(&ctx->rsrc_put_llist);
8198 struct io_rsrc_node *ref_node;
8199 struct llist_node *next = node->next;
8201 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8202 __io_rsrc_put_work(ref_node);
8207 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8208 unsigned nr_args, u64 __user *tags)
8210 __s32 __user *fds = (__s32 __user *) arg;
8219 if (nr_args > IORING_MAX_FIXED_FILES)
8221 if (nr_args > rlimit(RLIMIT_NOFILE))
8223 ret = io_rsrc_node_switch_start(ctx);
8226 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8232 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8235 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8236 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8240 /* allow sparse sets */
8243 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8250 if (unlikely(!file))
8254 * Don't allow io_uring instances to be registered. If UNIX
8255 * isn't enabled, then this causes a reference cycle and this
8256 * instance can never get freed. If UNIX is enabled we'll
8257 * handle it just fine, but there's still no point in allowing
8258 * a ring fd as it doesn't support regular read/write anyway.
8260 if (file->f_op == &io_uring_fops) {
8264 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8267 ret = io_sqe_files_scm(ctx);
8269 __io_sqe_files_unregister(ctx);
8273 io_rsrc_node_switch(ctx, NULL);
8276 for (i = 0; i < ctx->nr_user_files; i++) {
8277 file = io_file_from_index(ctx, i);
8281 io_free_file_tables(&ctx->file_table);
8282 ctx->nr_user_files = 0;
8284 io_rsrc_data_free(ctx->file_data);
8285 ctx->file_data = NULL;
8289 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8292 #if defined(CONFIG_UNIX)
8293 struct sock *sock = ctx->ring_sock->sk;
8294 struct sk_buff_head *head = &sock->sk_receive_queue;
8295 struct sk_buff *skb;
8298 * See if we can merge this file into an existing skb SCM_RIGHTS
8299 * file set. If there's no room, fall back to allocating a new skb
8300 * and filling it in.
8302 spin_lock_irq(&head->lock);
8303 skb = skb_peek(head);
8305 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8307 if (fpl->count < SCM_MAX_FD) {
8308 __skb_unlink(skb, head);
8309 spin_unlock_irq(&head->lock);
8310 fpl->fp[fpl->count] = get_file(file);
8311 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8313 spin_lock_irq(&head->lock);
8314 __skb_queue_head(head, skb);
8319 spin_unlock_irq(&head->lock);
8326 return __io_sqe_files_scm(ctx, 1, index);
8332 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8333 struct io_rsrc_node *node, void *rsrc)
8335 struct io_rsrc_put *prsrc;
8337 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8341 prsrc->tag = *io_get_tag_slot(data, idx);
8343 list_add(&prsrc->list, &node->rsrc_list);
8347 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8348 unsigned int issue_flags, u32 slot_index)
8350 struct io_ring_ctx *ctx = req->ctx;
8351 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8352 bool needs_switch = false;
8353 struct io_fixed_file *file_slot;
8356 io_ring_submit_lock(ctx, !force_nonblock);
8357 if (file->f_op == &io_uring_fops)
8360 if (!ctx->file_data)
8363 if (slot_index >= ctx->nr_user_files)
8366 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8367 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8369 if (file_slot->file_ptr) {
8370 struct file *old_file;
8372 ret = io_rsrc_node_switch_start(ctx);
8376 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8377 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8378 ctx->rsrc_node, old_file);
8381 file_slot->file_ptr = 0;
8382 needs_switch = true;
8385 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8386 io_fixed_file_set(file_slot, file);
8387 ret = io_sqe_file_register(ctx, file, slot_index);
8389 file_slot->file_ptr = 0;
8396 io_rsrc_node_switch(ctx, ctx->file_data);
8397 io_ring_submit_unlock(ctx, !force_nonblock);
8403 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8404 struct io_uring_rsrc_update2 *up,
8407 u64 __user *tags = u64_to_user_ptr(up->tags);
8408 __s32 __user *fds = u64_to_user_ptr(up->data);
8409 struct io_rsrc_data *data = ctx->file_data;
8410 struct io_fixed_file *file_slot;
8414 bool needs_switch = false;
8416 if (!ctx->file_data)
8418 if (up->offset + nr_args > ctx->nr_user_files)
8421 for (done = 0; done < nr_args; done++) {
8424 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8425 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8429 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8433 if (fd == IORING_REGISTER_FILES_SKIP)
8436 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8437 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8439 if (file_slot->file_ptr) {
8440 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8441 err = io_queue_rsrc_removal(data, up->offset + done,
8442 ctx->rsrc_node, file);
8445 file_slot->file_ptr = 0;
8446 needs_switch = true;
8455 * Don't allow io_uring instances to be registered. If
8456 * UNIX isn't enabled, then this causes a reference
8457 * cycle and this instance can never get freed. If UNIX
8458 * is enabled we'll handle it just fine, but there's
8459 * still no point in allowing a ring fd as it doesn't
8460 * support regular read/write anyway.
8462 if (file->f_op == &io_uring_fops) {
8467 *io_get_tag_slot(data, up->offset + done) = tag;
8468 io_fixed_file_set(file_slot, file);
8469 err = io_sqe_file_register(ctx, file, i);
8471 file_slot->file_ptr = 0;
8479 io_rsrc_node_switch(ctx, data);
8480 return done ? done : err;
8483 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8484 struct task_struct *task)
8486 struct io_wq_hash *hash;
8487 struct io_wq_data data;
8488 unsigned int concurrency;
8490 mutex_lock(&ctx->uring_lock);
8491 hash = ctx->hash_map;
8493 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8495 mutex_unlock(&ctx->uring_lock);
8496 return ERR_PTR(-ENOMEM);
8498 refcount_set(&hash->refs, 1);
8499 init_waitqueue_head(&hash->wait);
8500 ctx->hash_map = hash;
8502 mutex_unlock(&ctx->uring_lock);
8506 data.free_work = io_wq_free_work;
8507 data.do_work = io_wq_submit_work;
8509 /* Do QD, or 4 * CPUS, whatever is smallest */
8510 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8512 return io_wq_create(concurrency, &data);
8515 static int io_uring_alloc_task_context(struct task_struct *task,
8516 struct io_ring_ctx *ctx)
8518 struct io_uring_task *tctx;
8521 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8522 if (unlikely(!tctx))
8525 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8526 if (unlikely(ret)) {
8531 tctx->io_wq = io_init_wq_offload(ctx, task);
8532 if (IS_ERR(tctx->io_wq)) {
8533 ret = PTR_ERR(tctx->io_wq);
8534 percpu_counter_destroy(&tctx->inflight);
8540 init_waitqueue_head(&tctx->wait);
8541 atomic_set(&tctx->in_idle, 0);
8542 atomic_set(&tctx->inflight_tracked, 0);
8543 task->io_uring = tctx;
8544 spin_lock_init(&tctx->task_lock);
8545 INIT_WQ_LIST(&tctx->task_list);
8546 init_task_work(&tctx->task_work, tctx_task_work);
8550 void __io_uring_free(struct task_struct *tsk)
8552 struct io_uring_task *tctx = tsk->io_uring;
8554 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8555 WARN_ON_ONCE(tctx->io_wq);
8556 WARN_ON_ONCE(tctx->cached_refs);
8558 percpu_counter_destroy(&tctx->inflight);
8560 tsk->io_uring = NULL;
8563 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8564 struct io_uring_params *p)
8568 /* Retain compatibility with failing for an invalid attach attempt */
8569 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8570 IORING_SETUP_ATTACH_WQ) {
8573 f = fdget(p->wq_fd);
8576 if (f.file->f_op != &io_uring_fops) {
8582 if (ctx->flags & IORING_SETUP_SQPOLL) {
8583 struct task_struct *tsk;
8584 struct io_sq_data *sqd;
8587 sqd = io_get_sq_data(p, &attached);
8593 ctx->sq_creds = get_current_cred();
8595 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8596 if (!ctx->sq_thread_idle)
8597 ctx->sq_thread_idle = HZ;
8599 io_sq_thread_park(sqd);
8600 list_add(&ctx->sqd_list, &sqd->ctx_list);
8601 io_sqd_update_thread_idle(sqd);
8602 /* don't attach to a dying SQPOLL thread, would be racy */
8603 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8604 io_sq_thread_unpark(sqd);
8611 if (p->flags & IORING_SETUP_SQ_AFF) {
8612 int cpu = p->sq_thread_cpu;
8615 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8622 sqd->task_pid = current->pid;
8623 sqd->task_tgid = current->tgid;
8624 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8631 ret = io_uring_alloc_task_context(tsk, ctx);
8632 wake_up_new_task(tsk);
8635 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8636 /* Can't have SQ_AFF without SQPOLL */
8643 complete(&ctx->sq_data->exited);
8645 io_sq_thread_finish(ctx);
8649 static inline void __io_unaccount_mem(struct user_struct *user,
8650 unsigned long nr_pages)
8652 atomic_long_sub(nr_pages, &user->locked_vm);
8655 static inline int __io_account_mem(struct user_struct *user,
8656 unsigned long nr_pages)
8658 unsigned long page_limit, cur_pages, new_pages;
8660 /* Don't allow more pages than we can safely lock */
8661 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8664 cur_pages = atomic_long_read(&user->locked_vm);
8665 new_pages = cur_pages + nr_pages;
8666 if (new_pages > page_limit)
8668 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8669 new_pages) != cur_pages);
8674 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8677 __io_unaccount_mem(ctx->user, nr_pages);
8679 if (ctx->mm_account)
8680 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8683 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8688 ret = __io_account_mem(ctx->user, nr_pages);
8693 if (ctx->mm_account)
8694 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8699 static void io_mem_free(void *ptr)
8706 page = virt_to_head_page(ptr);
8707 if (put_page_testzero(page))
8708 free_compound_page(page);
8711 static void *io_mem_alloc(size_t size)
8713 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8714 __GFP_NORETRY | __GFP_ACCOUNT;
8716 return (void *) __get_free_pages(gfp_flags, get_order(size));
8719 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8722 struct io_rings *rings;
8723 size_t off, sq_array_size;
8725 off = struct_size(rings, cqes, cq_entries);
8726 if (off == SIZE_MAX)
8730 off = ALIGN(off, SMP_CACHE_BYTES);
8738 sq_array_size = array_size(sizeof(u32), sq_entries);
8739 if (sq_array_size == SIZE_MAX)
8742 if (check_add_overflow(off, sq_array_size, &off))
8748 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8750 struct io_mapped_ubuf *imu = *slot;
8753 if (imu != ctx->dummy_ubuf) {
8754 for (i = 0; i < imu->nr_bvecs; i++)
8755 unpin_user_page(imu->bvec[i].bv_page);
8756 if (imu->acct_pages)
8757 io_unaccount_mem(ctx, imu->acct_pages);
8763 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8765 io_buffer_unmap(ctx, &prsrc->buf);
8769 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8773 for (i = 0; i < ctx->nr_user_bufs; i++)
8774 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8775 kfree(ctx->user_bufs);
8776 io_rsrc_data_free(ctx->buf_data);
8777 ctx->user_bufs = NULL;
8778 ctx->buf_data = NULL;
8779 ctx->nr_user_bufs = 0;
8782 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8789 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8791 __io_sqe_buffers_unregister(ctx);
8795 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8796 void __user *arg, unsigned index)
8798 struct iovec __user *src;
8800 #ifdef CONFIG_COMPAT
8802 struct compat_iovec __user *ciovs;
8803 struct compat_iovec ciov;
8805 ciovs = (struct compat_iovec __user *) arg;
8806 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8809 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8810 dst->iov_len = ciov.iov_len;
8814 src = (struct iovec __user *) arg;
8815 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8821 * Not super efficient, but this is just a registration time. And we do cache
8822 * the last compound head, so generally we'll only do a full search if we don't
8825 * We check if the given compound head page has already been accounted, to
8826 * avoid double accounting it. This allows us to account the full size of the
8827 * page, not just the constituent pages of a huge page.
8829 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8830 int nr_pages, struct page *hpage)
8834 /* check current page array */
8835 for (i = 0; i < nr_pages; i++) {
8836 if (!PageCompound(pages[i]))
8838 if (compound_head(pages[i]) == hpage)
8842 /* check previously registered pages */
8843 for (i = 0; i < ctx->nr_user_bufs; i++) {
8844 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8846 for (j = 0; j < imu->nr_bvecs; j++) {
8847 if (!PageCompound(imu->bvec[j].bv_page))
8849 if (compound_head(imu->bvec[j].bv_page) == hpage)
8857 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8858 int nr_pages, struct io_mapped_ubuf *imu,
8859 struct page **last_hpage)
8863 imu->acct_pages = 0;
8864 for (i = 0; i < nr_pages; i++) {
8865 if (!PageCompound(pages[i])) {
8870 hpage = compound_head(pages[i]);
8871 if (hpage == *last_hpage)
8873 *last_hpage = hpage;
8874 if (headpage_already_acct(ctx, pages, i, hpage))
8876 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8880 if (!imu->acct_pages)
8883 ret = io_account_mem(ctx, imu->acct_pages);
8885 imu->acct_pages = 0;
8889 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8890 struct io_mapped_ubuf **pimu,
8891 struct page **last_hpage)
8893 struct io_mapped_ubuf *imu = NULL;
8894 struct vm_area_struct **vmas = NULL;
8895 struct page **pages = NULL;
8896 unsigned long off, start, end, ubuf;
8898 int ret, pret, nr_pages, i;
8900 if (!iov->iov_base) {
8901 *pimu = ctx->dummy_ubuf;
8905 ubuf = (unsigned long) iov->iov_base;
8906 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8907 start = ubuf >> PAGE_SHIFT;
8908 nr_pages = end - start;
8913 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8917 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8922 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8927 mmap_read_lock(current->mm);
8928 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8930 if (pret == nr_pages) {
8931 /* don't support file backed memory */
8932 for (i = 0; i < nr_pages; i++) {
8933 struct vm_area_struct *vma = vmas[i];
8935 if (vma_is_shmem(vma))
8938 !is_file_hugepages(vma->vm_file)) {
8944 ret = pret < 0 ? pret : -EFAULT;
8946 mmap_read_unlock(current->mm);
8949 * if we did partial map, or found file backed vmas,
8950 * release any pages we did get
8953 unpin_user_pages(pages, pret);
8957 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8959 unpin_user_pages(pages, pret);
8963 off = ubuf & ~PAGE_MASK;
8964 size = iov->iov_len;
8965 for (i = 0; i < nr_pages; i++) {
8968 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8969 imu->bvec[i].bv_page = pages[i];
8970 imu->bvec[i].bv_len = vec_len;
8971 imu->bvec[i].bv_offset = off;
8975 /* store original address for later verification */
8977 imu->ubuf_end = ubuf + iov->iov_len;
8978 imu->nr_bvecs = nr_pages;
8989 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8991 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8992 return ctx->user_bufs ? 0 : -ENOMEM;
8995 static int io_buffer_validate(struct iovec *iov)
8997 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9000 * Don't impose further limits on the size and buffer
9001 * constraints here, we'll -EINVAL later when IO is
9002 * submitted if they are wrong.
9005 return iov->iov_len ? -EFAULT : 0;
9009 /* arbitrary limit, but we need something */
9010 if (iov->iov_len > SZ_1G)
9013 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9019 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9020 unsigned int nr_args, u64 __user *tags)
9022 struct page *last_hpage = NULL;
9023 struct io_rsrc_data *data;
9029 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9031 ret = io_rsrc_node_switch_start(ctx);
9034 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9037 ret = io_buffers_map_alloc(ctx, nr_args);
9039 io_rsrc_data_free(data);
9043 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9044 ret = io_copy_iov(ctx, &iov, arg, i);
9047 ret = io_buffer_validate(&iov);
9050 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9055 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9061 WARN_ON_ONCE(ctx->buf_data);
9063 ctx->buf_data = data;
9065 __io_sqe_buffers_unregister(ctx);
9067 io_rsrc_node_switch(ctx, NULL);
9071 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9072 struct io_uring_rsrc_update2 *up,
9073 unsigned int nr_args)
9075 u64 __user *tags = u64_to_user_ptr(up->tags);
9076 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9077 struct page *last_hpage = NULL;
9078 bool needs_switch = false;
9084 if (up->offset + nr_args > ctx->nr_user_bufs)
9087 for (done = 0; done < nr_args; done++) {
9088 struct io_mapped_ubuf *imu;
9089 int offset = up->offset + done;
9092 err = io_copy_iov(ctx, &iov, iovs, done);
9095 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9099 err = io_buffer_validate(&iov);
9102 if (!iov.iov_base && tag) {
9106 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9110 i = array_index_nospec(offset, ctx->nr_user_bufs);
9111 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9112 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9113 ctx->rsrc_node, ctx->user_bufs[i]);
9114 if (unlikely(err)) {
9115 io_buffer_unmap(ctx, &imu);
9118 ctx->user_bufs[i] = NULL;
9119 needs_switch = true;
9122 ctx->user_bufs[i] = imu;
9123 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9127 io_rsrc_node_switch(ctx, ctx->buf_data);
9128 return done ? done : err;
9131 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9133 __s32 __user *fds = arg;
9139 if (copy_from_user(&fd, fds, sizeof(*fds)))
9142 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9143 if (IS_ERR(ctx->cq_ev_fd)) {
9144 int ret = PTR_ERR(ctx->cq_ev_fd);
9146 ctx->cq_ev_fd = NULL;
9153 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9155 if (ctx->cq_ev_fd) {
9156 eventfd_ctx_put(ctx->cq_ev_fd);
9157 ctx->cq_ev_fd = NULL;
9164 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9166 struct io_buffer *buf;
9167 unsigned long index;
9169 xa_for_each(&ctx->io_buffers, index, buf)
9170 __io_remove_buffers(ctx, buf, index, -1U);
9173 static void io_req_cache_free(struct list_head *list)
9175 struct io_kiocb *req, *nxt;
9177 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9178 list_del(&req->inflight_entry);
9179 kmem_cache_free(req_cachep, req);
9183 static void io_req_caches_free(struct io_ring_ctx *ctx)
9185 struct io_submit_state *state = &ctx->submit_state;
9187 mutex_lock(&ctx->uring_lock);
9189 if (state->free_reqs) {
9190 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9191 state->free_reqs = 0;
9194 io_flush_cached_locked_reqs(ctx, state);
9195 io_req_cache_free(&state->free_list);
9196 mutex_unlock(&ctx->uring_lock);
9199 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9201 if (data && !atomic_dec_and_test(&data->refs))
9202 wait_for_completion(&data->done);
9205 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9207 io_sq_thread_finish(ctx);
9209 if (ctx->mm_account) {
9210 mmdrop(ctx->mm_account);
9211 ctx->mm_account = NULL;
9214 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9215 io_wait_rsrc_data(ctx->buf_data);
9216 io_wait_rsrc_data(ctx->file_data);
9218 mutex_lock(&ctx->uring_lock);
9220 __io_sqe_buffers_unregister(ctx);
9222 __io_sqe_files_unregister(ctx);
9224 __io_cqring_overflow_flush(ctx, true);
9225 mutex_unlock(&ctx->uring_lock);
9226 io_eventfd_unregister(ctx);
9227 io_destroy_buffers(ctx);
9229 put_cred(ctx->sq_creds);
9231 /* there are no registered resources left, nobody uses it */
9233 io_rsrc_node_destroy(ctx->rsrc_node);
9234 if (ctx->rsrc_backup_node)
9235 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9236 flush_delayed_work(&ctx->rsrc_put_work);
9238 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9239 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9241 #if defined(CONFIG_UNIX)
9242 if (ctx->ring_sock) {
9243 ctx->ring_sock->file = NULL; /* so that iput() is called */
9244 sock_release(ctx->ring_sock);
9247 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9249 io_mem_free(ctx->rings);
9250 io_mem_free(ctx->sq_sqes);
9252 percpu_ref_exit(&ctx->refs);
9253 free_uid(ctx->user);
9254 io_req_caches_free(ctx);
9256 io_wq_put_hash(ctx->hash_map);
9257 kfree(ctx->cancel_hash);
9258 kfree(ctx->dummy_ubuf);
9262 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9264 struct io_ring_ctx *ctx = file->private_data;
9267 poll_wait(file, &ctx->poll_wait, wait);
9269 * synchronizes with barrier from wq_has_sleeper call in
9273 if (!io_sqring_full(ctx))
9274 mask |= EPOLLOUT | EPOLLWRNORM;
9277 * Don't flush cqring overflow list here, just do a simple check.
9278 * Otherwise there could possible be ABBA deadlock:
9281 * lock(&ctx->uring_lock);
9283 * lock(&ctx->uring_lock);
9286 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9287 * pushs them to do the flush.
9289 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9290 mask |= EPOLLIN | EPOLLRDNORM;
9295 static int io_uring_fasync(int fd, struct file *file, int on)
9297 struct io_ring_ctx *ctx = file->private_data;
9299 return fasync_helper(fd, file, on, &ctx->cq_fasync);
9302 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9304 const struct cred *creds;
9306 creds = xa_erase(&ctx->personalities, id);
9315 struct io_tctx_exit {
9316 struct callback_head task_work;
9317 struct completion completion;
9318 struct io_ring_ctx *ctx;
9321 static void io_tctx_exit_cb(struct callback_head *cb)
9323 struct io_uring_task *tctx = current->io_uring;
9324 struct io_tctx_exit *work;
9326 work = container_of(cb, struct io_tctx_exit, task_work);
9328 * When @in_idle, we're in cancellation and it's racy to remove the
9329 * node. It'll be removed by the end of cancellation, just ignore it.
9331 if (!atomic_read(&tctx->in_idle))
9332 io_uring_del_tctx_node((unsigned long)work->ctx);
9333 complete(&work->completion);
9336 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9338 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9340 return req->ctx == data;
9343 static void io_ring_exit_work(struct work_struct *work)
9345 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9346 unsigned long timeout = jiffies + HZ * 60 * 5;
9347 unsigned long interval = HZ / 20;
9348 struct io_tctx_exit exit;
9349 struct io_tctx_node *node;
9353 * If we're doing polled IO and end up having requests being
9354 * submitted async (out-of-line), then completions can come in while
9355 * we're waiting for refs to drop. We need to reap these manually,
9356 * as nobody else will be looking for them.
9359 io_uring_try_cancel_requests(ctx, NULL, true);
9361 struct io_sq_data *sqd = ctx->sq_data;
9362 struct task_struct *tsk;
9364 io_sq_thread_park(sqd);
9366 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9367 io_wq_cancel_cb(tsk->io_uring->io_wq,
9368 io_cancel_ctx_cb, ctx, true);
9369 io_sq_thread_unpark(sqd);
9372 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9373 /* there is little hope left, don't run it too often */
9376 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9378 init_completion(&exit.completion);
9379 init_task_work(&exit.task_work, io_tctx_exit_cb);
9382 * Some may use context even when all refs and requests have been put,
9383 * and they are free to do so while still holding uring_lock or
9384 * completion_lock, see io_req_task_submit(). Apart from other work,
9385 * this lock/unlock section also waits them to finish.
9387 mutex_lock(&ctx->uring_lock);
9388 while (!list_empty(&ctx->tctx_list)) {
9389 WARN_ON_ONCE(time_after(jiffies, timeout));
9391 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9393 /* don't spin on a single task if cancellation failed */
9394 list_rotate_left(&ctx->tctx_list);
9395 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9396 if (WARN_ON_ONCE(ret))
9398 wake_up_process(node->task);
9400 mutex_unlock(&ctx->uring_lock);
9401 wait_for_completion(&exit.completion);
9402 mutex_lock(&ctx->uring_lock);
9404 mutex_unlock(&ctx->uring_lock);
9405 spin_lock(&ctx->completion_lock);
9406 spin_unlock(&ctx->completion_lock);
9408 io_ring_ctx_free(ctx);
9411 /* Returns true if we found and killed one or more timeouts */
9412 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9415 struct io_kiocb *req, *tmp;
9418 spin_lock(&ctx->completion_lock);
9419 spin_lock_irq(&ctx->timeout_lock);
9420 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9421 if (io_match_task(req, tsk, cancel_all)) {
9422 io_kill_timeout(req, -ECANCELED);
9426 spin_unlock_irq(&ctx->timeout_lock);
9428 io_commit_cqring(ctx);
9429 spin_unlock(&ctx->completion_lock);
9431 io_cqring_ev_posted(ctx);
9432 return canceled != 0;
9435 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9437 unsigned long index;
9438 struct creds *creds;
9440 mutex_lock(&ctx->uring_lock);
9441 percpu_ref_kill(&ctx->refs);
9443 __io_cqring_overflow_flush(ctx, true);
9444 xa_for_each(&ctx->personalities, index, creds)
9445 io_unregister_personality(ctx, index);
9446 mutex_unlock(&ctx->uring_lock);
9448 io_kill_timeouts(ctx, NULL, true);
9449 io_poll_remove_all(ctx, NULL, true);
9451 /* if we failed setting up the ctx, we might not have any rings */
9452 io_iopoll_try_reap_events(ctx);
9454 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9456 * Use system_unbound_wq to avoid spawning tons of event kworkers
9457 * if we're exiting a ton of rings at the same time. It just adds
9458 * noise and overhead, there's no discernable change in runtime
9459 * over using system_wq.
9461 queue_work(system_unbound_wq, &ctx->exit_work);
9464 static int io_uring_release(struct inode *inode, struct file *file)
9466 struct io_ring_ctx *ctx = file->private_data;
9468 file->private_data = NULL;
9469 io_ring_ctx_wait_and_kill(ctx);
9473 struct io_task_cancel {
9474 struct task_struct *task;
9478 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9480 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9481 struct io_task_cancel *cancel = data;
9484 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9485 struct io_ring_ctx *ctx = req->ctx;
9487 /* protect against races with linked timeouts */
9488 spin_lock(&ctx->completion_lock);
9489 ret = io_match_task(req, cancel->task, cancel->all);
9490 spin_unlock(&ctx->completion_lock);
9492 ret = io_match_task(req, cancel->task, cancel->all);
9497 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9498 struct task_struct *task, bool cancel_all)
9500 struct io_defer_entry *de;
9503 spin_lock(&ctx->completion_lock);
9504 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9505 if (io_match_task(de->req, task, cancel_all)) {
9506 list_cut_position(&list, &ctx->defer_list, &de->list);
9510 spin_unlock(&ctx->completion_lock);
9511 if (list_empty(&list))
9514 while (!list_empty(&list)) {
9515 de = list_first_entry(&list, struct io_defer_entry, list);
9516 list_del_init(&de->list);
9517 io_req_complete_failed(de->req, -ECANCELED);
9523 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9525 struct io_tctx_node *node;
9526 enum io_wq_cancel cret;
9529 mutex_lock(&ctx->uring_lock);
9530 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9531 struct io_uring_task *tctx = node->task->io_uring;
9534 * io_wq will stay alive while we hold uring_lock, because it's
9535 * killed after ctx nodes, which requires to take the lock.
9537 if (!tctx || !tctx->io_wq)
9539 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9540 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9542 mutex_unlock(&ctx->uring_lock);
9547 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9548 struct task_struct *task,
9551 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9552 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9555 enum io_wq_cancel cret;
9559 ret |= io_uring_try_cancel_iowq(ctx);
9560 } else if (tctx && tctx->io_wq) {
9562 * Cancels requests of all rings, not only @ctx, but
9563 * it's fine as the task is in exit/exec.
9565 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9567 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9570 /* SQPOLL thread does its own polling */
9571 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9572 (ctx->sq_data && ctx->sq_data->thread == current)) {
9573 while (!list_empty_careful(&ctx->iopoll_list)) {
9574 io_iopoll_try_reap_events(ctx);
9579 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9580 ret |= io_poll_remove_all(ctx, task, cancel_all);
9581 ret |= io_kill_timeouts(ctx, task, cancel_all);
9583 ret |= io_run_task_work();
9590 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9592 struct io_uring_task *tctx = current->io_uring;
9593 struct io_tctx_node *node;
9596 if (unlikely(!tctx)) {
9597 ret = io_uring_alloc_task_context(current, ctx);
9600 tctx = current->io_uring;
9602 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9603 node = kmalloc(sizeof(*node), GFP_KERNEL);
9607 node->task = current;
9609 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9616 mutex_lock(&ctx->uring_lock);
9617 list_add(&node->ctx_node, &ctx->tctx_list);
9618 mutex_unlock(&ctx->uring_lock);
9625 * Note that this task has used io_uring. We use it for cancelation purposes.
9627 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9629 struct io_uring_task *tctx = current->io_uring;
9631 if (likely(tctx && tctx->last == ctx))
9633 return __io_uring_add_tctx_node(ctx);
9637 * Remove this io_uring_file -> task mapping.
9639 static void io_uring_del_tctx_node(unsigned long index)
9641 struct io_uring_task *tctx = current->io_uring;
9642 struct io_tctx_node *node;
9646 node = xa_erase(&tctx->xa, index);
9650 WARN_ON_ONCE(current != node->task);
9651 WARN_ON_ONCE(list_empty(&node->ctx_node));
9653 mutex_lock(&node->ctx->uring_lock);
9654 list_del(&node->ctx_node);
9655 mutex_unlock(&node->ctx->uring_lock);
9657 if (tctx->last == node->ctx)
9662 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9664 struct io_wq *wq = tctx->io_wq;
9665 struct io_tctx_node *node;
9666 unsigned long index;
9668 xa_for_each(&tctx->xa, index, node)
9669 io_uring_del_tctx_node(index);
9672 * Must be after io_uring_del_task_file() (removes nodes under
9673 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9675 io_wq_put_and_exit(wq);
9680 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9683 return atomic_read(&tctx->inflight_tracked);
9684 return percpu_counter_sum(&tctx->inflight);
9687 static void io_uring_drop_tctx_refs(struct task_struct *task)
9689 struct io_uring_task *tctx = task->io_uring;
9690 unsigned int refs = tctx->cached_refs;
9693 tctx->cached_refs = 0;
9694 percpu_counter_sub(&tctx->inflight, refs);
9695 put_task_struct_many(task, refs);
9700 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9701 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9703 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9705 struct io_uring_task *tctx = current->io_uring;
9706 struct io_ring_ctx *ctx;
9710 WARN_ON_ONCE(sqd && sqd->thread != current);
9712 if (!current->io_uring)
9715 io_wq_exit_start(tctx->io_wq);
9717 atomic_inc(&tctx->in_idle);
9719 io_uring_drop_tctx_refs(current);
9720 /* read completions before cancelations */
9721 inflight = tctx_inflight(tctx, !cancel_all);
9726 struct io_tctx_node *node;
9727 unsigned long index;
9729 xa_for_each(&tctx->xa, index, node) {
9730 /* sqpoll task will cancel all its requests */
9731 if (node->ctx->sq_data)
9733 io_uring_try_cancel_requests(node->ctx, current,
9737 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9738 io_uring_try_cancel_requests(ctx, current,
9742 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9743 io_uring_drop_tctx_refs(current);
9745 * If we've seen completions, retry without waiting. This
9746 * avoids a race where a completion comes in before we did
9747 * prepare_to_wait().
9749 if (inflight == tctx_inflight(tctx, !cancel_all))
9751 finish_wait(&tctx->wait, &wait);
9753 atomic_dec(&tctx->in_idle);
9755 io_uring_clean_tctx(tctx);
9757 /* for exec all current's requests should be gone, kill tctx */
9758 __io_uring_free(current);
9762 void __io_uring_cancel(bool cancel_all)
9764 io_uring_cancel_generic(cancel_all, NULL);
9767 static void *io_uring_validate_mmap_request(struct file *file,
9768 loff_t pgoff, size_t sz)
9770 struct io_ring_ctx *ctx = file->private_data;
9771 loff_t offset = pgoff << PAGE_SHIFT;
9776 case IORING_OFF_SQ_RING:
9777 case IORING_OFF_CQ_RING:
9780 case IORING_OFF_SQES:
9784 return ERR_PTR(-EINVAL);
9787 page = virt_to_head_page(ptr);
9788 if (sz > page_size(page))
9789 return ERR_PTR(-EINVAL);
9796 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9798 size_t sz = vma->vm_end - vma->vm_start;
9802 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9804 return PTR_ERR(ptr);
9806 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9807 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9810 #else /* !CONFIG_MMU */
9812 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9814 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9817 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9819 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9822 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9823 unsigned long addr, unsigned long len,
9824 unsigned long pgoff, unsigned long flags)
9828 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9830 return PTR_ERR(ptr);
9832 return (unsigned long) ptr;
9835 #endif /* !CONFIG_MMU */
9837 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9842 if (!io_sqring_full(ctx))
9844 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9846 if (!io_sqring_full(ctx))
9849 } while (!signal_pending(current));
9851 finish_wait(&ctx->sqo_sq_wait, &wait);
9855 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9856 struct __kernel_timespec __user **ts,
9857 const sigset_t __user **sig)
9859 struct io_uring_getevents_arg arg;
9862 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9863 * is just a pointer to the sigset_t.
9865 if (!(flags & IORING_ENTER_EXT_ARG)) {
9866 *sig = (const sigset_t __user *) argp;
9872 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9873 * timespec and sigset_t pointers if good.
9875 if (*argsz != sizeof(arg))
9877 if (copy_from_user(&arg, argp, sizeof(arg)))
9879 *sig = u64_to_user_ptr(arg.sigmask);
9880 *argsz = arg.sigmask_sz;
9881 *ts = u64_to_user_ptr(arg.ts);
9885 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9886 u32, min_complete, u32, flags, const void __user *, argp,
9889 struct io_ring_ctx *ctx;
9896 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9897 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9901 if (unlikely(!f.file))
9905 if (unlikely(f.file->f_op != &io_uring_fops))
9909 ctx = f.file->private_data;
9910 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9914 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9918 * For SQ polling, the thread will do all submissions and completions.
9919 * Just return the requested submit count, and wake the thread if
9923 if (ctx->flags & IORING_SETUP_SQPOLL) {
9924 io_cqring_overflow_flush(ctx);
9926 if (unlikely(ctx->sq_data->thread == NULL)) {
9930 if (flags & IORING_ENTER_SQ_WAKEUP)
9931 wake_up(&ctx->sq_data->wait);
9932 if (flags & IORING_ENTER_SQ_WAIT) {
9933 ret = io_sqpoll_wait_sq(ctx);
9937 submitted = to_submit;
9938 } else if (to_submit) {
9939 ret = io_uring_add_tctx_node(ctx);
9942 mutex_lock(&ctx->uring_lock);
9943 submitted = io_submit_sqes(ctx, to_submit);
9944 mutex_unlock(&ctx->uring_lock);
9946 if (submitted != to_submit)
9949 if (flags & IORING_ENTER_GETEVENTS) {
9950 const sigset_t __user *sig;
9951 struct __kernel_timespec __user *ts;
9953 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9957 min_complete = min(min_complete, ctx->cq_entries);
9960 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9961 * space applications don't need to do io completion events
9962 * polling again, they can rely on io_sq_thread to do polling
9963 * work, which can reduce cpu usage and uring_lock contention.
9965 if (ctx->flags & IORING_SETUP_IOPOLL &&
9966 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9967 ret = io_iopoll_check(ctx, min_complete);
9969 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9974 percpu_ref_put(&ctx->refs);
9977 return submitted ? submitted : ret;
9980 #ifdef CONFIG_PROC_FS
9981 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9982 const struct cred *cred)
9984 struct user_namespace *uns = seq_user_ns(m);
9985 struct group_info *gi;
9990 seq_printf(m, "%5d\n", id);
9991 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9992 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9993 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9994 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9995 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9996 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9997 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9998 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9999 seq_puts(m, "\n\tGroups:\t");
10000 gi = cred->group_info;
10001 for (g = 0; g < gi->ngroups; g++) {
10002 seq_put_decimal_ull(m, g ? " " : "",
10003 from_kgid_munged(uns, gi->gid[g]));
10005 seq_puts(m, "\n\tCapEff:\t");
10006 cap = cred->cap_effective;
10007 CAP_FOR_EACH_U32(__capi)
10008 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10013 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10015 struct io_sq_data *sq = NULL;
10020 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10021 * since fdinfo case grabs it in the opposite direction of normal use
10022 * cases. If we fail to get the lock, we just don't iterate any
10023 * structures that could be going away outside the io_uring mutex.
10025 has_lock = mutex_trylock(&ctx->uring_lock);
10027 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10033 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10034 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10035 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10036 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10037 struct file *f = io_file_from_index(ctx, i);
10040 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10042 seq_printf(m, "%5u: <none>\n", i);
10044 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10045 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10046 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10047 unsigned int len = buf->ubuf_end - buf->ubuf;
10049 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10051 if (has_lock && !xa_empty(&ctx->personalities)) {
10052 unsigned long index;
10053 const struct cred *cred;
10055 seq_printf(m, "Personalities:\n");
10056 xa_for_each(&ctx->personalities, index, cred)
10057 io_uring_show_cred(m, index, cred);
10059 seq_printf(m, "PollList:\n");
10060 spin_lock(&ctx->completion_lock);
10061 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10062 struct hlist_head *list = &ctx->cancel_hash[i];
10063 struct io_kiocb *req;
10065 hlist_for_each_entry(req, list, hash_node)
10066 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10067 req->task->task_works != NULL);
10069 spin_unlock(&ctx->completion_lock);
10071 mutex_unlock(&ctx->uring_lock);
10074 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10076 struct io_ring_ctx *ctx = f->private_data;
10078 if (percpu_ref_tryget(&ctx->refs)) {
10079 __io_uring_show_fdinfo(ctx, m);
10080 percpu_ref_put(&ctx->refs);
10085 static const struct file_operations io_uring_fops = {
10086 .release = io_uring_release,
10087 .mmap = io_uring_mmap,
10089 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10090 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10092 .poll = io_uring_poll,
10093 .fasync = io_uring_fasync,
10094 #ifdef CONFIG_PROC_FS
10095 .show_fdinfo = io_uring_show_fdinfo,
10099 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10100 struct io_uring_params *p)
10102 struct io_rings *rings;
10103 size_t size, sq_array_offset;
10105 /* make sure these are sane, as we already accounted them */
10106 ctx->sq_entries = p->sq_entries;
10107 ctx->cq_entries = p->cq_entries;
10109 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10110 if (size == SIZE_MAX)
10113 rings = io_mem_alloc(size);
10117 ctx->rings = rings;
10118 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10119 rings->sq_ring_mask = p->sq_entries - 1;
10120 rings->cq_ring_mask = p->cq_entries - 1;
10121 rings->sq_ring_entries = p->sq_entries;
10122 rings->cq_ring_entries = p->cq_entries;
10124 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10125 if (size == SIZE_MAX) {
10126 io_mem_free(ctx->rings);
10131 ctx->sq_sqes = io_mem_alloc(size);
10132 if (!ctx->sq_sqes) {
10133 io_mem_free(ctx->rings);
10141 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10145 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10149 ret = io_uring_add_tctx_node(ctx);
10154 fd_install(fd, file);
10159 * Allocate an anonymous fd, this is what constitutes the application
10160 * visible backing of an io_uring instance. The application mmaps this
10161 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10162 * we have to tie this fd to a socket for file garbage collection purposes.
10164 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10167 #if defined(CONFIG_UNIX)
10170 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10173 return ERR_PTR(ret);
10176 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10177 O_RDWR | O_CLOEXEC);
10178 #if defined(CONFIG_UNIX)
10179 if (IS_ERR(file)) {
10180 sock_release(ctx->ring_sock);
10181 ctx->ring_sock = NULL;
10183 ctx->ring_sock->file = file;
10189 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10190 struct io_uring_params __user *params)
10192 struct io_ring_ctx *ctx;
10198 if (entries > IORING_MAX_ENTRIES) {
10199 if (!(p->flags & IORING_SETUP_CLAMP))
10201 entries = IORING_MAX_ENTRIES;
10205 * Use twice as many entries for the CQ ring. It's possible for the
10206 * application to drive a higher depth than the size of the SQ ring,
10207 * since the sqes are only used at submission time. This allows for
10208 * some flexibility in overcommitting a bit. If the application has
10209 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10210 * of CQ ring entries manually.
10212 p->sq_entries = roundup_pow_of_two(entries);
10213 if (p->flags & IORING_SETUP_CQSIZE) {
10215 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10216 * to a power-of-two, if it isn't already. We do NOT impose
10217 * any cq vs sq ring sizing.
10219 if (!p->cq_entries)
10221 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10222 if (!(p->flags & IORING_SETUP_CLAMP))
10224 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10226 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10227 if (p->cq_entries < p->sq_entries)
10230 p->cq_entries = 2 * p->sq_entries;
10233 ctx = io_ring_ctx_alloc(p);
10236 ctx->compat = in_compat_syscall();
10237 if (!capable(CAP_IPC_LOCK))
10238 ctx->user = get_uid(current_user());
10241 * This is just grabbed for accounting purposes. When a process exits,
10242 * the mm is exited and dropped before the files, hence we need to hang
10243 * on to this mm purely for the purposes of being able to unaccount
10244 * memory (locked/pinned vm). It's not used for anything else.
10246 mmgrab(current->mm);
10247 ctx->mm_account = current->mm;
10249 ret = io_allocate_scq_urings(ctx, p);
10253 ret = io_sq_offload_create(ctx, p);
10256 /* always set a rsrc node */
10257 ret = io_rsrc_node_switch_start(ctx);
10260 io_rsrc_node_switch(ctx, NULL);
10262 memset(&p->sq_off, 0, sizeof(p->sq_off));
10263 p->sq_off.head = offsetof(struct io_rings, sq.head);
10264 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10265 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10266 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10267 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10268 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10269 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10271 memset(&p->cq_off, 0, sizeof(p->cq_off));
10272 p->cq_off.head = offsetof(struct io_rings, cq.head);
10273 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10274 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10275 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10276 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10277 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10278 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10280 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10281 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10282 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10283 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10284 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10285 IORING_FEAT_RSRC_TAGS;
10287 if (copy_to_user(params, p, sizeof(*p))) {
10292 file = io_uring_get_file(ctx);
10293 if (IS_ERR(file)) {
10294 ret = PTR_ERR(file);
10299 * Install ring fd as the very last thing, so we don't risk someone
10300 * having closed it before we finish setup
10302 ret = io_uring_install_fd(ctx, file);
10304 /* fput will clean it up */
10309 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10312 io_ring_ctx_wait_and_kill(ctx);
10317 * Sets up an aio uring context, and returns the fd. Applications asks for a
10318 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10319 * params structure passed in.
10321 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10323 struct io_uring_params p;
10326 if (copy_from_user(&p, params, sizeof(p)))
10328 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10333 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10334 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10335 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10336 IORING_SETUP_R_DISABLED))
10339 return io_uring_create(entries, &p, params);
10342 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10343 struct io_uring_params __user *, params)
10345 return io_uring_setup(entries, params);
10348 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10350 struct io_uring_probe *p;
10354 size = struct_size(p, ops, nr_args);
10355 if (size == SIZE_MAX)
10357 p = kzalloc(size, GFP_KERNEL);
10362 if (copy_from_user(p, arg, size))
10365 if (memchr_inv(p, 0, size))
10368 p->last_op = IORING_OP_LAST - 1;
10369 if (nr_args > IORING_OP_LAST)
10370 nr_args = IORING_OP_LAST;
10372 for (i = 0; i < nr_args; i++) {
10374 if (!io_op_defs[i].not_supported)
10375 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10380 if (copy_to_user(arg, p, size))
10387 static int io_register_personality(struct io_ring_ctx *ctx)
10389 const struct cred *creds;
10393 creds = get_current_cred();
10395 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10396 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10404 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10405 unsigned int nr_args)
10407 struct io_uring_restriction *res;
10411 /* Restrictions allowed only if rings started disabled */
10412 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10415 /* We allow only a single restrictions registration */
10416 if (ctx->restrictions.registered)
10419 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10422 size = array_size(nr_args, sizeof(*res));
10423 if (size == SIZE_MAX)
10426 res = memdup_user(arg, size);
10428 return PTR_ERR(res);
10432 for (i = 0; i < nr_args; i++) {
10433 switch (res[i].opcode) {
10434 case IORING_RESTRICTION_REGISTER_OP:
10435 if (res[i].register_op >= IORING_REGISTER_LAST) {
10440 __set_bit(res[i].register_op,
10441 ctx->restrictions.register_op);
10443 case IORING_RESTRICTION_SQE_OP:
10444 if (res[i].sqe_op >= IORING_OP_LAST) {
10449 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10451 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10452 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10454 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10455 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10464 /* Reset all restrictions if an error happened */
10466 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10468 ctx->restrictions.registered = true;
10474 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10476 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10479 if (ctx->restrictions.registered)
10480 ctx->restricted = 1;
10482 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10483 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10484 wake_up(&ctx->sq_data->wait);
10488 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10489 struct io_uring_rsrc_update2 *up,
10497 if (check_add_overflow(up->offset, nr_args, &tmp))
10499 err = io_rsrc_node_switch_start(ctx);
10504 case IORING_RSRC_FILE:
10505 return __io_sqe_files_update(ctx, up, nr_args);
10506 case IORING_RSRC_BUFFER:
10507 return __io_sqe_buffers_update(ctx, up, nr_args);
10512 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10515 struct io_uring_rsrc_update2 up;
10519 memset(&up, 0, sizeof(up));
10520 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10522 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10525 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10526 unsigned size, unsigned type)
10528 struct io_uring_rsrc_update2 up;
10530 if (size != sizeof(up))
10532 if (copy_from_user(&up, arg, sizeof(up)))
10534 if (!up.nr || up.resv)
10536 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10539 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10540 unsigned int size, unsigned int type)
10542 struct io_uring_rsrc_register rr;
10544 /* keep it extendible */
10545 if (size != sizeof(rr))
10548 memset(&rr, 0, sizeof(rr));
10549 if (copy_from_user(&rr, arg, size))
10551 if (!rr.nr || rr.resv || rr.resv2)
10555 case IORING_RSRC_FILE:
10556 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10557 rr.nr, u64_to_user_ptr(rr.tags));
10558 case IORING_RSRC_BUFFER:
10559 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10560 rr.nr, u64_to_user_ptr(rr.tags));
10565 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10568 struct io_uring_task *tctx = current->io_uring;
10569 cpumask_var_t new_mask;
10572 if (!tctx || !tctx->io_wq)
10575 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10578 cpumask_clear(new_mask);
10579 if (len > cpumask_size())
10580 len = cpumask_size();
10582 if (copy_from_user(new_mask, arg, len)) {
10583 free_cpumask_var(new_mask);
10587 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10588 free_cpumask_var(new_mask);
10592 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10594 struct io_uring_task *tctx = current->io_uring;
10596 if (!tctx || !tctx->io_wq)
10599 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10602 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10605 struct io_uring_task *tctx = NULL;
10606 struct io_sq_data *sqd = NULL;
10607 __u32 new_count[2];
10610 if (copy_from_user(new_count, arg, sizeof(new_count)))
10612 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10613 if (new_count[i] > INT_MAX)
10616 if (ctx->flags & IORING_SETUP_SQPOLL) {
10617 sqd = ctx->sq_data;
10620 * Observe the correct sqd->lock -> ctx->uring_lock
10621 * ordering. Fine to drop uring_lock here, we hold
10622 * a ref to the ctx.
10624 refcount_inc(&sqd->refs);
10625 mutex_unlock(&ctx->uring_lock);
10626 mutex_lock(&sqd->lock);
10627 mutex_lock(&ctx->uring_lock);
10629 tctx = sqd->thread->io_uring;
10632 tctx = current->io_uring;
10636 if (!tctx || !tctx->io_wq)
10639 ret = io_wq_max_workers(tctx->io_wq, new_count);
10644 mutex_unlock(&sqd->lock);
10645 io_put_sq_data(sqd);
10648 if (copy_to_user(arg, new_count, sizeof(new_count)))
10654 mutex_unlock(&sqd->lock);
10655 io_put_sq_data(sqd);
10660 static bool io_register_op_must_quiesce(int op)
10663 case IORING_REGISTER_BUFFERS:
10664 case IORING_UNREGISTER_BUFFERS:
10665 case IORING_REGISTER_FILES:
10666 case IORING_UNREGISTER_FILES:
10667 case IORING_REGISTER_FILES_UPDATE:
10668 case IORING_REGISTER_PROBE:
10669 case IORING_REGISTER_PERSONALITY:
10670 case IORING_UNREGISTER_PERSONALITY:
10671 case IORING_REGISTER_FILES2:
10672 case IORING_REGISTER_FILES_UPDATE2:
10673 case IORING_REGISTER_BUFFERS2:
10674 case IORING_REGISTER_BUFFERS_UPDATE:
10675 case IORING_REGISTER_IOWQ_AFF:
10676 case IORING_UNREGISTER_IOWQ_AFF:
10677 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10684 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10688 percpu_ref_kill(&ctx->refs);
10691 * Drop uring mutex before waiting for references to exit. If another
10692 * thread is currently inside io_uring_enter() it might need to grab the
10693 * uring_lock to make progress. If we hold it here across the drain
10694 * wait, then we can deadlock. It's safe to drop the mutex here, since
10695 * no new references will come in after we've killed the percpu ref.
10697 mutex_unlock(&ctx->uring_lock);
10699 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10702 ret = io_run_task_work_sig();
10703 } while (ret >= 0);
10704 mutex_lock(&ctx->uring_lock);
10707 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10711 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10712 void __user *arg, unsigned nr_args)
10713 __releases(ctx->uring_lock)
10714 __acquires(ctx->uring_lock)
10719 * We're inside the ring mutex, if the ref is already dying, then
10720 * someone else killed the ctx or is already going through
10721 * io_uring_register().
10723 if (percpu_ref_is_dying(&ctx->refs))
10726 if (ctx->restricted) {
10727 if (opcode >= IORING_REGISTER_LAST)
10729 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10730 if (!test_bit(opcode, ctx->restrictions.register_op))
10734 if (io_register_op_must_quiesce(opcode)) {
10735 ret = io_ctx_quiesce(ctx);
10741 case IORING_REGISTER_BUFFERS:
10742 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10744 case IORING_UNREGISTER_BUFFERS:
10746 if (arg || nr_args)
10748 ret = io_sqe_buffers_unregister(ctx);
10750 case IORING_REGISTER_FILES:
10751 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10753 case IORING_UNREGISTER_FILES:
10755 if (arg || nr_args)
10757 ret = io_sqe_files_unregister(ctx);
10759 case IORING_REGISTER_FILES_UPDATE:
10760 ret = io_register_files_update(ctx, arg, nr_args);
10762 case IORING_REGISTER_EVENTFD:
10763 case IORING_REGISTER_EVENTFD_ASYNC:
10767 ret = io_eventfd_register(ctx, arg);
10770 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10771 ctx->eventfd_async = 1;
10773 ctx->eventfd_async = 0;
10775 case IORING_UNREGISTER_EVENTFD:
10777 if (arg || nr_args)
10779 ret = io_eventfd_unregister(ctx);
10781 case IORING_REGISTER_PROBE:
10783 if (!arg || nr_args > 256)
10785 ret = io_probe(ctx, arg, nr_args);
10787 case IORING_REGISTER_PERSONALITY:
10789 if (arg || nr_args)
10791 ret = io_register_personality(ctx);
10793 case IORING_UNREGISTER_PERSONALITY:
10797 ret = io_unregister_personality(ctx, nr_args);
10799 case IORING_REGISTER_ENABLE_RINGS:
10801 if (arg || nr_args)
10803 ret = io_register_enable_rings(ctx);
10805 case IORING_REGISTER_RESTRICTIONS:
10806 ret = io_register_restrictions(ctx, arg, nr_args);
10808 case IORING_REGISTER_FILES2:
10809 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10811 case IORING_REGISTER_FILES_UPDATE2:
10812 ret = io_register_rsrc_update(ctx, arg, nr_args,
10815 case IORING_REGISTER_BUFFERS2:
10816 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10818 case IORING_REGISTER_BUFFERS_UPDATE:
10819 ret = io_register_rsrc_update(ctx, arg, nr_args,
10820 IORING_RSRC_BUFFER);
10822 case IORING_REGISTER_IOWQ_AFF:
10824 if (!arg || !nr_args)
10826 ret = io_register_iowq_aff(ctx, arg, nr_args);
10828 case IORING_UNREGISTER_IOWQ_AFF:
10830 if (arg || nr_args)
10832 ret = io_unregister_iowq_aff(ctx);
10834 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10836 if (!arg || nr_args != 2)
10838 ret = io_register_iowq_max_workers(ctx, arg);
10845 if (io_register_op_must_quiesce(opcode)) {
10846 /* bring the ctx back to life */
10847 percpu_ref_reinit(&ctx->refs);
10848 reinit_completion(&ctx->ref_comp);
10853 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10854 void __user *, arg, unsigned int, nr_args)
10856 struct io_ring_ctx *ctx;
10865 if (f.file->f_op != &io_uring_fops)
10868 ctx = f.file->private_data;
10870 io_run_task_work();
10872 mutex_lock(&ctx->uring_lock);
10873 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10874 mutex_unlock(&ctx->uring_lock);
10875 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10876 ctx->cq_ev_fd != NULL, ret);
10882 static int __init io_uring_init(void)
10884 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10885 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10886 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10889 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10890 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10891 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10892 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10893 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10894 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10895 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10896 BUILD_BUG_SQE_ELEM(8, __u64, off);
10897 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10898 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10899 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10900 BUILD_BUG_SQE_ELEM(24, __u32, len);
10901 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10902 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10903 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10904 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10905 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10906 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10907 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10908 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10909 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10910 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10911 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10912 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10913 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10914 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10915 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10916 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10917 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10918 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10919 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10920 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10921 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
10923 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10924 sizeof(struct io_uring_rsrc_update));
10925 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10926 sizeof(struct io_uring_rsrc_update2));
10928 /* ->buf_index is u16 */
10929 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
10931 /* should fit into one byte */
10932 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10934 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10935 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
10937 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10941 __initcall(io_uring_init);