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 unsigned cq_last_tm_flush;
407 } ____cacheline_aligned_in_smp;
410 spinlock_t completion_lock;
412 spinlock_t timeout_lock;
415 * ->iopoll_list is protected by the ctx->uring_lock for
416 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417 * For SQPOLL, only the single threaded io_sq_thread() will
418 * manipulate the list, hence no extra locking is needed there.
420 struct list_head iopoll_list;
421 struct hlist_head *cancel_hash;
422 unsigned cancel_hash_bits;
423 bool poll_multi_queue;
424 } ____cacheline_aligned_in_smp;
426 struct io_restriction restrictions;
428 /* slow path rsrc auxilary data, used by update/register */
430 struct io_rsrc_node *rsrc_backup_node;
431 struct io_mapped_ubuf *dummy_ubuf;
432 struct io_rsrc_data *file_data;
433 struct io_rsrc_data *buf_data;
435 struct delayed_work rsrc_put_work;
436 struct llist_head rsrc_put_llist;
437 struct list_head rsrc_ref_list;
438 spinlock_t rsrc_ref_lock;
441 /* Keep this last, we don't need it for the fast path */
443 #if defined(CONFIG_UNIX)
444 struct socket *ring_sock;
446 /* hashed buffered write serialization */
447 struct io_wq_hash *hash_map;
449 /* Only used for accounting purposes */
450 struct user_struct *user;
451 struct mm_struct *mm_account;
453 /* ctx exit and cancelation */
454 struct llist_head fallback_llist;
455 struct delayed_work fallback_work;
456 struct work_struct exit_work;
457 struct list_head tctx_list;
458 struct completion ref_comp;
460 bool iowq_limits_set;
464 struct io_uring_task {
465 /* submission side */
468 struct wait_queue_head wait;
469 const struct io_ring_ctx *last;
471 struct percpu_counter inflight;
472 atomic_t inflight_tracked;
475 spinlock_t task_lock;
476 struct io_wq_work_list task_list;
477 struct callback_head task_work;
482 * First field must be the file pointer in all the
483 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
485 struct io_poll_iocb {
487 struct wait_queue_head *head;
489 struct wait_queue_entry wait;
492 struct io_poll_update {
498 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;
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;
886 /* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
887 struct io_buffer *kbuf;
891 struct io_tctx_node {
892 struct list_head ctx_node;
893 struct task_struct *task;
894 struct io_ring_ctx *ctx;
897 struct io_defer_entry {
898 struct list_head list;
899 struct io_kiocb *req;
904 /* needs req->file assigned */
905 unsigned needs_file : 1;
906 /* hash wq insertion if file is a regular file */
907 unsigned hash_reg_file : 1;
908 /* unbound wq insertion if file is a non-regular file */
909 unsigned unbound_nonreg_file : 1;
910 /* opcode is not supported by this kernel */
911 unsigned not_supported : 1;
912 /* set if opcode supports polled "wait" */
914 unsigned pollout : 1;
915 /* op supports buffer selection */
916 unsigned buffer_select : 1;
917 /* do prep async if is going to be punted */
918 unsigned needs_async_setup : 1;
919 /* should block plug */
921 /* size of async data needed, if any */
922 unsigned short async_size;
925 static const struct io_op_def io_op_defs[] = {
926 [IORING_OP_NOP] = {},
927 [IORING_OP_READV] = {
929 .unbound_nonreg_file = 1,
932 .needs_async_setup = 1,
934 .async_size = sizeof(struct io_async_rw),
936 [IORING_OP_WRITEV] = {
939 .unbound_nonreg_file = 1,
941 .needs_async_setup = 1,
943 .async_size = sizeof(struct io_async_rw),
945 [IORING_OP_FSYNC] = {
948 [IORING_OP_READ_FIXED] = {
950 .unbound_nonreg_file = 1,
953 .async_size = sizeof(struct io_async_rw),
955 [IORING_OP_WRITE_FIXED] = {
958 .unbound_nonreg_file = 1,
961 .async_size = sizeof(struct io_async_rw),
963 [IORING_OP_POLL_ADD] = {
965 .unbound_nonreg_file = 1,
967 [IORING_OP_POLL_REMOVE] = {},
968 [IORING_OP_SYNC_FILE_RANGE] = {
971 [IORING_OP_SENDMSG] = {
973 .unbound_nonreg_file = 1,
975 .needs_async_setup = 1,
976 .async_size = sizeof(struct io_async_msghdr),
978 [IORING_OP_RECVMSG] = {
980 .unbound_nonreg_file = 1,
983 .needs_async_setup = 1,
984 .async_size = sizeof(struct io_async_msghdr),
986 [IORING_OP_TIMEOUT] = {
987 .async_size = sizeof(struct io_timeout_data),
989 [IORING_OP_TIMEOUT_REMOVE] = {
990 /* used by timeout updates' prep() */
992 [IORING_OP_ACCEPT] = {
994 .unbound_nonreg_file = 1,
997 [IORING_OP_ASYNC_CANCEL] = {},
998 [IORING_OP_LINK_TIMEOUT] = {
999 .async_size = sizeof(struct io_timeout_data),
1001 [IORING_OP_CONNECT] = {
1003 .unbound_nonreg_file = 1,
1005 .needs_async_setup = 1,
1006 .async_size = sizeof(struct io_async_connect),
1008 [IORING_OP_FALLOCATE] = {
1011 [IORING_OP_OPENAT] = {},
1012 [IORING_OP_CLOSE] = {},
1013 [IORING_OP_FILES_UPDATE] = {},
1014 [IORING_OP_STATX] = {},
1015 [IORING_OP_READ] = {
1017 .unbound_nonreg_file = 1,
1021 .async_size = sizeof(struct io_async_rw),
1023 [IORING_OP_WRITE] = {
1026 .unbound_nonreg_file = 1,
1029 .async_size = sizeof(struct io_async_rw),
1031 [IORING_OP_FADVISE] = {
1034 [IORING_OP_MADVISE] = {},
1035 [IORING_OP_SEND] = {
1037 .unbound_nonreg_file = 1,
1040 [IORING_OP_RECV] = {
1042 .unbound_nonreg_file = 1,
1046 [IORING_OP_OPENAT2] = {
1048 [IORING_OP_EPOLL_CTL] = {
1049 .unbound_nonreg_file = 1,
1051 [IORING_OP_SPLICE] = {
1054 .unbound_nonreg_file = 1,
1056 [IORING_OP_PROVIDE_BUFFERS] = {},
1057 [IORING_OP_REMOVE_BUFFERS] = {},
1061 .unbound_nonreg_file = 1,
1063 [IORING_OP_SHUTDOWN] = {
1066 [IORING_OP_RENAMEAT] = {},
1067 [IORING_OP_UNLINKAT] = {},
1068 [IORING_OP_MKDIRAT] = {},
1069 [IORING_OP_SYMLINKAT] = {},
1070 [IORING_OP_LINKAT] = {},
1073 /* requests with any of those set should undergo io_disarm_next() */
1074 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1076 static bool io_disarm_next(struct io_kiocb *req);
1077 static void io_uring_del_tctx_node(unsigned long index);
1078 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1079 struct task_struct *task,
1081 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1083 static void io_fill_cqe_req(struct io_kiocb *req, s32 res, u32 cflags);
1085 static void io_put_req(struct io_kiocb *req);
1086 static void io_put_req_deferred(struct io_kiocb *req);
1087 static void io_dismantle_req(struct io_kiocb *req);
1088 static void io_queue_linked_timeout(struct io_kiocb *req);
1089 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1090 struct io_uring_rsrc_update2 *up,
1092 static void io_clean_op(struct io_kiocb *req);
1093 static struct file *io_file_get(struct io_ring_ctx *ctx,
1094 struct io_kiocb *req, int fd, bool fixed);
1095 static void __io_queue_sqe(struct io_kiocb *req);
1096 static void io_rsrc_put_work(struct work_struct *work);
1098 static void io_req_task_queue(struct io_kiocb *req);
1099 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1100 static int io_req_prep_async(struct io_kiocb *req);
1102 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1103 unsigned int issue_flags, u32 slot_index);
1104 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
1106 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
1108 static struct kmem_cache *req_cachep;
1110 static const struct file_operations io_uring_fops;
1112 struct sock *io_uring_get_socket(struct file *file)
1114 #if defined(CONFIG_UNIX)
1115 if (file->f_op == &io_uring_fops) {
1116 struct io_ring_ctx *ctx = file->private_data;
1118 return ctx->ring_sock->sk;
1123 EXPORT_SYMBOL(io_uring_get_socket);
1125 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1128 mutex_lock(&ctx->uring_lock);
1133 #define io_for_each_link(pos, head) \
1134 for (pos = (head); pos; pos = pos->link)
1137 * Shamelessly stolen from the mm implementation of page reference checking,
1138 * see commit f958d7b528b1 for details.
1140 #define req_ref_zero_or_close_to_overflow(req) \
1141 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1143 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1145 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1146 return atomic_inc_not_zero(&req->refs);
1149 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1151 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1154 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1155 return atomic_dec_and_test(&req->refs);
1158 static inline void req_ref_get(struct io_kiocb *req)
1160 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1161 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1162 atomic_inc(&req->refs);
1165 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1167 if (!(req->flags & REQ_F_REFCOUNT)) {
1168 req->flags |= REQ_F_REFCOUNT;
1169 atomic_set(&req->refs, nr);
1173 static inline void io_req_set_refcount(struct io_kiocb *req)
1175 __io_req_set_refcount(req, 1);
1178 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1180 struct io_ring_ctx *ctx = req->ctx;
1182 if (!req->fixed_rsrc_refs) {
1183 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1184 percpu_ref_get(req->fixed_rsrc_refs);
1188 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1190 bool got = percpu_ref_tryget(ref);
1192 /* already at zero, wait for ->release() */
1194 wait_for_completion(compl);
1195 percpu_ref_resurrect(ref);
1197 percpu_ref_put(ref);
1200 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1202 __must_hold(&req->ctx->timeout_lock)
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 bool io_match_linked(struct io_kiocb *head)
1220 struct io_kiocb *req;
1222 io_for_each_link(req, head) {
1223 if (req->flags & REQ_F_INFLIGHT)
1230 * As io_match_task() but protected against racing with linked timeouts.
1231 * User must not hold timeout_lock.
1233 static bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
1238 if (task && head->task != task)
1243 if (head->flags & REQ_F_LINK_TIMEOUT) {
1244 struct io_ring_ctx *ctx = head->ctx;
1246 /* protect against races with linked timeouts */
1247 spin_lock_irq(&ctx->timeout_lock);
1248 matched = io_match_linked(head);
1249 spin_unlock_irq(&ctx->timeout_lock);
1251 matched = io_match_linked(head);
1256 static inline void req_set_fail(struct io_kiocb *req)
1258 req->flags |= REQ_F_FAIL;
1261 static inline void req_fail_link_node(struct io_kiocb *req, int res)
1267 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1269 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1271 complete(&ctx->ref_comp);
1274 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1276 return !req->timeout.off;
1279 static void io_fallback_req_func(struct work_struct *work)
1281 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1282 fallback_work.work);
1283 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1284 struct io_kiocb *req, *tmp;
1285 bool locked = false;
1287 percpu_ref_get(&ctx->refs);
1288 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1289 req->io_task_work.func(req, &locked);
1292 if (ctx->submit_state.compl_nr)
1293 io_submit_flush_completions(ctx);
1294 mutex_unlock(&ctx->uring_lock);
1296 percpu_ref_put(&ctx->refs);
1300 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1302 struct io_ring_ctx *ctx;
1305 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1310 * Use 5 bits less than the max cq entries, that should give us around
1311 * 32 entries per hash list if totally full and uniformly spread.
1313 hash_bits = ilog2(p->cq_entries);
1317 ctx->cancel_hash_bits = hash_bits;
1318 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1320 if (!ctx->cancel_hash)
1322 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1324 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1325 if (!ctx->dummy_ubuf)
1327 /* set invalid range, so io_import_fixed() fails meeting it */
1328 ctx->dummy_ubuf->ubuf = -1UL;
1330 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1331 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1334 ctx->flags = p->flags;
1335 init_waitqueue_head(&ctx->sqo_sq_wait);
1336 INIT_LIST_HEAD(&ctx->sqd_list);
1337 init_waitqueue_head(&ctx->poll_wait);
1338 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1339 init_completion(&ctx->ref_comp);
1340 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1341 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1342 mutex_init(&ctx->uring_lock);
1343 init_waitqueue_head(&ctx->cq_wait);
1344 spin_lock_init(&ctx->completion_lock);
1345 spin_lock_init(&ctx->timeout_lock);
1346 INIT_LIST_HEAD(&ctx->iopoll_list);
1347 INIT_LIST_HEAD(&ctx->defer_list);
1348 INIT_LIST_HEAD(&ctx->timeout_list);
1349 INIT_LIST_HEAD(&ctx->ltimeout_list);
1350 spin_lock_init(&ctx->rsrc_ref_lock);
1351 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1352 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1353 init_llist_head(&ctx->rsrc_put_llist);
1354 INIT_LIST_HEAD(&ctx->tctx_list);
1355 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1356 INIT_LIST_HEAD(&ctx->locked_free_list);
1357 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1360 kfree(ctx->dummy_ubuf);
1361 kfree(ctx->cancel_hash);
1366 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1368 struct io_rings *r = ctx->rings;
1370 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1374 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1376 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1377 struct io_ring_ctx *ctx = req->ctx;
1379 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1385 #define FFS_ASYNC_READ 0x1UL
1386 #define FFS_ASYNC_WRITE 0x2UL
1388 #define FFS_ISREG 0x4UL
1390 #define FFS_ISREG 0x0UL
1392 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1394 static inline bool io_req_ffs_set(struct io_kiocb *req)
1396 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1399 static void io_req_track_inflight(struct io_kiocb *req)
1401 if (!(req->flags & REQ_F_INFLIGHT)) {
1402 req->flags |= REQ_F_INFLIGHT;
1403 atomic_inc(&req->task->io_uring->inflight_tracked);
1407 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1409 if (WARN_ON_ONCE(!req->link))
1412 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1413 req->flags |= REQ_F_LINK_TIMEOUT;
1415 /* linked timeouts should have two refs once prep'ed */
1416 io_req_set_refcount(req);
1417 __io_req_set_refcount(req->link, 2);
1421 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1423 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1425 return __io_prep_linked_timeout(req);
1428 static void io_prep_async_work(struct io_kiocb *req)
1430 const struct io_op_def *def = &io_op_defs[req->opcode];
1431 struct io_ring_ctx *ctx = req->ctx;
1433 if (!(req->flags & REQ_F_CREDS)) {
1434 req->flags |= REQ_F_CREDS;
1435 req->creds = get_current_cred();
1438 req->work.list.next = NULL;
1439 req->work.flags = 0;
1440 if (req->flags & REQ_F_FORCE_ASYNC)
1441 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1443 if (req->flags & REQ_F_ISREG) {
1444 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1445 io_wq_hash_work(&req->work, file_inode(req->file));
1446 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1447 if (def->unbound_nonreg_file)
1448 req->work.flags |= IO_WQ_WORK_UNBOUND;
1452 static void io_prep_async_link(struct io_kiocb *req)
1454 struct io_kiocb *cur;
1456 if (req->flags & REQ_F_LINK_TIMEOUT) {
1457 struct io_ring_ctx *ctx = req->ctx;
1459 spin_lock_irq(&ctx->timeout_lock);
1460 io_for_each_link(cur, req)
1461 io_prep_async_work(cur);
1462 spin_unlock_irq(&ctx->timeout_lock);
1464 io_for_each_link(cur, req)
1465 io_prep_async_work(cur);
1469 static void io_queue_async_work(struct io_kiocb *req, bool *locked)
1471 struct io_ring_ctx *ctx = req->ctx;
1472 struct io_kiocb *link = io_prep_linked_timeout(req);
1473 struct io_uring_task *tctx = req->task->io_uring;
1475 /* must not take the lock, NULL it as a precaution */
1479 BUG_ON(!tctx->io_wq);
1481 /* init ->work of the whole link before punting */
1482 io_prep_async_link(req);
1485 * Not expected to happen, but if we do have a bug where this _can_
1486 * happen, catch it here and ensure the request is marked as
1487 * canceled. That will make io-wq go through the usual work cancel
1488 * procedure rather than attempt to run this request (or create a new
1491 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1492 req->work.flags |= IO_WQ_WORK_CANCEL;
1494 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1495 &req->work, req->flags);
1496 io_wq_enqueue(tctx->io_wq, &req->work);
1498 io_queue_linked_timeout(link);
1501 static void io_kill_timeout(struct io_kiocb *req, int status)
1502 __must_hold(&req->ctx->completion_lock)
1503 __must_hold(&req->ctx->timeout_lock)
1505 struct io_timeout_data *io = req->async_data;
1507 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1510 atomic_set(&req->ctx->cq_timeouts,
1511 atomic_read(&req->ctx->cq_timeouts) + 1);
1512 list_del_init(&req->timeout.list);
1513 io_fill_cqe_req(req, status, 0);
1514 io_put_req_deferred(req);
1518 static void io_queue_deferred(struct io_ring_ctx *ctx)
1520 while (!list_empty(&ctx->defer_list)) {
1521 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1522 struct io_defer_entry, list);
1524 if (req_need_defer(de->req, de->seq))
1526 list_del_init(&de->list);
1527 io_req_task_queue(de->req);
1532 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1533 __must_hold(&ctx->completion_lock)
1535 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1536 struct io_kiocb *req, *tmp;
1538 spin_lock_irq(&ctx->timeout_lock);
1539 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1540 u32 events_needed, events_got;
1542 if (io_is_timeout_noseq(req))
1546 * Since seq can easily wrap around over time, subtract
1547 * the last seq at which timeouts were flushed before comparing.
1548 * Assuming not more than 2^31-1 events have happened since,
1549 * these subtractions won't have wrapped, so we can check if
1550 * target is in [last_seq, current_seq] by comparing the two.
1552 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1553 events_got = seq - ctx->cq_last_tm_flush;
1554 if (events_got < events_needed)
1557 io_kill_timeout(req, 0);
1559 ctx->cq_last_tm_flush = seq;
1560 spin_unlock_irq(&ctx->timeout_lock);
1563 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1565 if (ctx->off_timeout_used)
1566 io_flush_timeouts(ctx);
1567 if (ctx->drain_active)
1568 io_queue_deferred(ctx);
1571 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1573 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1574 __io_commit_cqring_flush(ctx);
1575 /* order cqe stores with ring update */
1576 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1579 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1581 struct io_rings *r = ctx->rings;
1583 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1586 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1588 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1591 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1593 struct io_rings *rings = ctx->rings;
1594 unsigned tail, mask = ctx->cq_entries - 1;
1597 * writes to the cq entry need to come after reading head; the
1598 * control dependency is enough as we're using WRITE_ONCE to
1601 if (__io_cqring_events(ctx) == ctx->cq_entries)
1604 tail = ctx->cached_cq_tail++;
1605 return &rings->cqes[tail & mask];
1608 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1610 if (likely(!ctx->cq_ev_fd))
1612 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1614 return !ctx->eventfd_async || io_wq_current_is_worker();
1618 * This should only get called when at least one event has been posted.
1619 * Some applications rely on the eventfd notification count only changing
1620 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1621 * 1:1 relationship between how many times this function is called (and
1622 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1624 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1627 * wake_up_all() may seem excessive, but io_wake_function() and
1628 * io_should_wake() handle the termination of the loop and only
1629 * wake as many waiters as we need to.
1631 if (wq_has_sleeper(&ctx->cq_wait))
1632 wake_up_all(&ctx->cq_wait);
1633 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1634 wake_up(&ctx->sq_data->wait);
1635 if (io_should_trigger_evfd(ctx))
1636 eventfd_signal(ctx->cq_ev_fd, 1);
1637 if (waitqueue_active(&ctx->poll_wait))
1638 wake_up_interruptible(&ctx->poll_wait);
1641 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1643 /* see waitqueue_active() comment */
1646 if (ctx->flags & IORING_SETUP_SQPOLL) {
1647 if (waitqueue_active(&ctx->cq_wait))
1648 wake_up_all(&ctx->cq_wait);
1650 if (io_should_trigger_evfd(ctx))
1651 eventfd_signal(ctx->cq_ev_fd, 1);
1652 if (waitqueue_active(&ctx->poll_wait))
1653 wake_up_interruptible(&ctx->poll_wait);
1656 /* Returns true if there are no backlogged entries after the flush */
1657 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1659 bool all_flushed, posted;
1661 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1665 spin_lock(&ctx->completion_lock);
1666 while (!list_empty(&ctx->cq_overflow_list)) {
1667 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1668 struct io_overflow_cqe *ocqe;
1672 ocqe = list_first_entry(&ctx->cq_overflow_list,
1673 struct io_overflow_cqe, list);
1675 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1677 io_account_cq_overflow(ctx);
1680 list_del(&ocqe->list);
1684 all_flushed = list_empty(&ctx->cq_overflow_list);
1686 clear_bit(0, &ctx->check_cq_overflow);
1687 WRITE_ONCE(ctx->rings->sq_flags,
1688 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1692 io_commit_cqring(ctx);
1693 spin_unlock(&ctx->completion_lock);
1695 io_cqring_ev_posted(ctx);
1699 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1703 if (test_bit(0, &ctx->check_cq_overflow)) {
1704 /* iopoll syncs against uring_lock, not completion_lock */
1705 if (ctx->flags & IORING_SETUP_IOPOLL)
1706 mutex_lock(&ctx->uring_lock);
1707 ret = __io_cqring_overflow_flush(ctx, false);
1708 if (ctx->flags & IORING_SETUP_IOPOLL)
1709 mutex_unlock(&ctx->uring_lock);
1715 /* must to be called somewhat shortly after putting a request */
1716 static inline void io_put_task(struct task_struct *task, int nr)
1718 struct io_uring_task *tctx = task->io_uring;
1720 if (likely(task == current)) {
1721 tctx->cached_refs += nr;
1723 percpu_counter_sub(&tctx->inflight, nr);
1724 if (unlikely(atomic_read(&tctx->in_idle)))
1725 wake_up(&tctx->wait);
1726 put_task_struct_many(task, nr);
1730 static void io_task_refs_refill(struct io_uring_task *tctx)
1732 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1734 percpu_counter_add(&tctx->inflight, refill);
1735 refcount_add(refill, ¤t->usage);
1736 tctx->cached_refs += refill;
1739 static inline void io_get_task_refs(int nr)
1741 struct io_uring_task *tctx = current->io_uring;
1743 tctx->cached_refs -= nr;
1744 if (unlikely(tctx->cached_refs < 0))
1745 io_task_refs_refill(tctx);
1748 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
1750 struct io_uring_task *tctx = task->io_uring;
1751 unsigned int refs = tctx->cached_refs;
1754 tctx->cached_refs = 0;
1755 percpu_counter_sub(&tctx->inflight, refs);
1756 put_task_struct_many(task, refs);
1760 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1761 s32 res, u32 cflags)
1763 struct io_overflow_cqe *ocqe;
1765 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1768 * If we're in ring overflow flush mode, or in task cancel mode,
1769 * or cannot allocate an overflow entry, then we need to drop it
1772 io_account_cq_overflow(ctx);
1775 if (list_empty(&ctx->cq_overflow_list)) {
1776 set_bit(0, &ctx->check_cq_overflow);
1777 WRITE_ONCE(ctx->rings->sq_flags,
1778 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1781 ocqe->cqe.user_data = user_data;
1782 ocqe->cqe.res = res;
1783 ocqe->cqe.flags = cflags;
1784 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1788 static inline bool __io_fill_cqe(struct io_ring_ctx *ctx, u64 user_data,
1789 s32 res, u32 cflags)
1791 struct io_uring_cqe *cqe;
1793 trace_io_uring_complete(ctx, user_data, res, cflags);
1796 * If we can't get a cq entry, userspace overflowed the
1797 * submission (by quite a lot). Increment the overflow count in
1800 cqe = io_get_cqe(ctx);
1802 WRITE_ONCE(cqe->user_data, user_data);
1803 WRITE_ONCE(cqe->res, res);
1804 WRITE_ONCE(cqe->flags, cflags);
1807 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1810 static noinline void io_fill_cqe_req(struct io_kiocb *req, s32 res, u32 cflags)
1812 __io_fill_cqe(req->ctx, req->user_data, res, cflags);
1815 static noinline bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data,
1816 s32 res, u32 cflags)
1819 return __io_fill_cqe(ctx, user_data, res, cflags);
1822 static void io_req_complete_post(struct io_kiocb *req, s32 res,
1825 struct io_ring_ctx *ctx = req->ctx;
1827 spin_lock(&ctx->completion_lock);
1828 __io_fill_cqe(ctx, req->user_data, res, cflags);
1830 * If we're the last reference to this request, add to our locked
1833 if (req_ref_put_and_test(req)) {
1834 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1835 if (req->flags & IO_DISARM_MASK)
1836 io_disarm_next(req);
1838 io_req_task_queue(req->link);
1842 io_dismantle_req(req);
1843 io_put_task(req->task, 1);
1844 list_add(&req->inflight_entry, &ctx->locked_free_list);
1845 ctx->locked_free_nr++;
1847 if (!percpu_ref_tryget(&ctx->refs))
1850 io_commit_cqring(ctx);
1851 spin_unlock(&ctx->completion_lock);
1854 io_cqring_ev_posted(ctx);
1855 percpu_ref_put(&ctx->refs);
1859 static inline bool io_req_needs_clean(struct io_kiocb *req)
1861 return req->flags & IO_REQ_CLEAN_FLAGS;
1864 static inline void io_req_complete_state(struct io_kiocb *req, s32 res,
1867 if (io_req_needs_clean(req))
1870 req->compl.cflags = cflags;
1871 req->flags |= REQ_F_COMPLETE_INLINE;
1874 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1875 s32 res, u32 cflags)
1877 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1878 io_req_complete_state(req, res, cflags);
1880 io_req_complete_post(req, res, cflags);
1883 static inline void io_req_complete(struct io_kiocb *req, s32 res)
1885 __io_req_complete(req, 0, res, 0);
1888 static void io_req_complete_failed(struct io_kiocb *req, s32 res)
1891 io_req_complete_post(req, res, 0);
1894 static void io_req_complete_fail_submit(struct io_kiocb *req)
1897 * We don't submit, fail them all, for that replace hardlinks with
1898 * normal links. Extra REQ_F_LINK is tolerated.
1900 req->flags &= ~REQ_F_HARDLINK;
1901 req->flags |= REQ_F_LINK;
1902 io_req_complete_failed(req, req->result);
1906 * Don't initialise the fields below on every allocation, but do that in
1907 * advance and keep them valid across allocations.
1909 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1913 req->async_data = NULL;
1914 /* not necessary, but safer to zero */
1918 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1919 struct io_submit_state *state)
1921 spin_lock(&ctx->completion_lock);
1922 list_splice_init(&ctx->locked_free_list, &state->free_list);
1923 ctx->locked_free_nr = 0;
1924 spin_unlock(&ctx->completion_lock);
1927 /* Returns true IFF there are requests in the cache */
1928 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1930 struct io_submit_state *state = &ctx->submit_state;
1934 * If we have more than a batch's worth of requests in our IRQ side
1935 * locked cache, grab the lock and move them over to our submission
1938 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1939 io_flush_cached_locked_reqs(ctx, state);
1941 nr = state->free_reqs;
1942 while (!list_empty(&state->free_list)) {
1943 struct io_kiocb *req = list_first_entry(&state->free_list,
1944 struct io_kiocb, inflight_entry);
1946 list_del(&req->inflight_entry);
1947 state->reqs[nr++] = req;
1948 if (nr == ARRAY_SIZE(state->reqs))
1952 state->free_reqs = nr;
1957 * A request might get retired back into the request caches even before opcode
1958 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1959 * Because of that, io_alloc_req() should be called only under ->uring_lock
1960 * and with extra caution to not get a request that is still worked on.
1962 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1963 __must_hold(&ctx->uring_lock)
1965 struct io_submit_state *state = &ctx->submit_state;
1966 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1969 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1971 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1974 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1978 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1979 * retry single alloc to be on the safe side.
1981 if (unlikely(ret <= 0)) {
1982 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1983 if (!state->reqs[0])
1988 for (i = 0; i < ret; i++)
1989 io_preinit_req(state->reqs[i], ctx);
1990 state->free_reqs = ret;
1993 return state->reqs[state->free_reqs];
1996 static inline void io_put_file(struct file *file)
2002 static void io_dismantle_req(struct io_kiocb *req)
2004 unsigned int flags = req->flags;
2006 if (io_req_needs_clean(req))
2008 if (!(flags & REQ_F_FIXED_FILE))
2009 io_put_file(req->file);
2010 if (req->fixed_rsrc_refs)
2011 percpu_ref_put(req->fixed_rsrc_refs);
2012 if (req->async_data) {
2013 kfree(req->async_data);
2014 req->async_data = NULL;
2018 static void __io_free_req(struct io_kiocb *req)
2020 struct io_ring_ctx *ctx = req->ctx;
2022 io_dismantle_req(req);
2023 io_put_task(req->task, 1);
2025 spin_lock(&ctx->completion_lock);
2026 list_add(&req->inflight_entry, &ctx->locked_free_list);
2027 ctx->locked_free_nr++;
2028 spin_unlock(&ctx->completion_lock);
2030 percpu_ref_put(&ctx->refs);
2033 static inline void io_remove_next_linked(struct io_kiocb *req)
2035 struct io_kiocb *nxt = req->link;
2037 req->link = nxt->link;
2041 static bool io_kill_linked_timeout(struct io_kiocb *req)
2042 __must_hold(&req->ctx->completion_lock)
2043 __must_hold(&req->ctx->timeout_lock)
2045 struct io_kiocb *link = req->link;
2047 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2048 struct io_timeout_data *io = link->async_data;
2050 io_remove_next_linked(req);
2051 link->timeout.head = NULL;
2052 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2053 list_del(&link->timeout.list);
2054 io_fill_cqe_req(link, -ECANCELED, 0);
2055 io_put_req_deferred(link);
2062 static void io_fail_links(struct io_kiocb *req)
2063 __must_hold(&req->ctx->completion_lock)
2065 struct io_kiocb *nxt, *link = req->link;
2069 long res = -ECANCELED;
2071 if (link->flags & REQ_F_FAIL)
2077 trace_io_uring_fail_link(req, link);
2078 io_fill_cqe_req(link, res, 0);
2079 io_put_req_deferred(link);
2084 static bool io_disarm_next(struct io_kiocb *req)
2085 __must_hold(&req->ctx->completion_lock)
2087 bool posted = false;
2089 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2090 struct io_kiocb *link = req->link;
2092 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2093 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2094 io_remove_next_linked(req);
2095 io_fill_cqe_req(link, -ECANCELED, 0);
2096 io_put_req_deferred(link);
2099 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2100 struct io_ring_ctx *ctx = req->ctx;
2102 spin_lock_irq(&ctx->timeout_lock);
2103 posted = io_kill_linked_timeout(req);
2104 spin_unlock_irq(&ctx->timeout_lock);
2106 if (unlikely((req->flags & REQ_F_FAIL) &&
2107 !(req->flags & REQ_F_HARDLINK))) {
2108 posted |= (req->link != NULL);
2114 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2116 struct io_kiocb *nxt;
2119 * If LINK is set, we have dependent requests in this chain. If we
2120 * didn't fail this request, queue the first one up, moving any other
2121 * dependencies to the next request. In case of failure, fail the rest
2124 if (req->flags & IO_DISARM_MASK) {
2125 struct io_ring_ctx *ctx = req->ctx;
2128 spin_lock(&ctx->completion_lock);
2129 posted = io_disarm_next(req);
2131 io_commit_cqring(req->ctx);
2132 spin_unlock(&ctx->completion_lock);
2134 io_cqring_ev_posted(ctx);
2141 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2143 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2145 return __io_req_find_next(req);
2148 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2153 if (ctx->submit_state.compl_nr)
2154 io_submit_flush_completions(ctx);
2155 mutex_unlock(&ctx->uring_lock);
2158 percpu_ref_put(&ctx->refs);
2161 static void tctx_task_work(struct callback_head *cb)
2163 bool locked = false;
2164 struct io_ring_ctx *ctx = NULL;
2165 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2169 struct io_wq_work_node *node;
2171 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2172 io_submit_flush_completions(ctx);
2174 spin_lock_irq(&tctx->task_lock);
2175 node = tctx->task_list.first;
2176 INIT_WQ_LIST(&tctx->task_list);
2178 tctx->task_running = false;
2179 spin_unlock_irq(&tctx->task_lock);
2184 struct io_wq_work_node *next = node->next;
2185 struct io_kiocb *req = container_of(node, struct io_kiocb,
2188 if (req->ctx != ctx) {
2189 ctx_flush_and_put(ctx, &locked);
2191 /* if not contended, grab and improve batching */
2192 locked = mutex_trylock(&ctx->uring_lock);
2193 percpu_ref_get(&ctx->refs);
2195 req->io_task_work.func(req, &locked);
2202 ctx_flush_and_put(ctx, &locked);
2204 /* relaxed read is enough as only the task itself sets ->in_idle */
2205 if (unlikely(atomic_read(&tctx->in_idle)))
2206 io_uring_drop_tctx_refs(current);
2209 static void io_req_task_work_add(struct io_kiocb *req)
2211 struct task_struct *tsk = req->task;
2212 struct io_uring_task *tctx = tsk->io_uring;
2213 enum task_work_notify_mode notify;
2214 struct io_wq_work_node *node;
2215 unsigned long flags;
2218 WARN_ON_ONCE(!tctx);
2220 spin_lock_irqsave(&tctx->task_lock, flags);
2221 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2222 running = tctx->task_running;
2224 tctx->task_running = true;
2225 spin_unlock_irqrestore(&tctx->task_lock, flags);
2227 /* task_work already pending, we're done */
2232 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2233 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2234 * processing task_work. There's no reliable way to tell if TWA_RESUME
2237 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2238 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2239 wake_up_process(tsk);
2243 spin_lock_irqsave(&tctx->task_lock, flags);
2244 tctx->task_running = false;
2245 node = tctx->task_list.first;
2246 INIT_WQ_LIST(&tctx->task_list);
2247 spin_unlock_irqrestore(&tctx->task_lock, flags);
2250 req = container_of(node, struct io_kiocb, io_task_work.node);
2252 if (llist_add(&req->io_task_work.fallback_node,
2253 &req->ctx->fallback_llist))
2254 schedule_delayed_work(&req->ctx->fallback_work, 1);
2258 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2260 struct io_ring_ctx *ctx = req->ctx;
2262 /* not needed for normal modes, but SQPOLL depends on it */
2263 io_tw_lock(ctx, locked);
2264 io_req_complete_failed(req, req->result);
2267 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2269 struct io_ring_ctx *ctx = req->ctx;
2271 io_tw_lock(ctx, locked);
2272 /* req->task == current here, checking PF_EXITING is safe */
2273 if (likely(!(req->task->flags & PF_EXITING)))
2274 __io_queue_sqe(req);
2276 io_req_complete_failed(req, -EFAULT);
2279 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2282 req->io_task_work.func = io_req_task_cancel;
2283 io_req_task_work_add(req);
2286 static void io_req_task_queue(struct io_kiocb *req)
2288 req->io_task_work.func = io_req_task_submit;
2289 io_req_task_work_add(req);
2292 static void io_req_task_queue_reissue(struct io_kiocb *req)
2294 req->io_task_work.func = io_queue_async_work;
2295 io_req_task_work_add(req);
2298 static inline void io_queue_next(struct io_kiocb *req)
2300 struct io_kiocb *nxt = io_req_find_next(req);
2303 io_req_task_queue(nxt);
2306 static void io_free_req(struct io_kiocb *req)
2312 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2318 struct task_struct *task;
2323 static inline void io_init_req_batch(struct req_batch *rb)
2330 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2331 struct req_batch *rb)
2334 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2336 io_put_task(rb->task, rb->task_refs);
2339 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2340 struct io_submit_state *state)
2343 io_dismantle_req(req);
2345 if (req->task != rb->task) {
2347 io_put_task(rb->task, rb->task_refs);
2348 rb->task = req->task;
2354 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2355 state->reqs[state->free_reqs++] = req;
2357 list_add(&req->inflight_entry, &state->free_list);
2360 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2361 __must_hold(&ctx->uring_lock)
2363 struct io_submit_state *state = &ctx->submit_state;
2364 int i, nr = state->compl_nr;
2365 struct req_batch rb;
2367 spin_lock(&ctx->completion_lock);
2368 for (i = 0; i < nr; i++) {
2369 struct io_kiocb *req = state->compl_reqs[i];
2371 __io_fill_cqe(ctx, req->user_data, req->result,
2374 io_commit_cqring(ctx);
2375 spin_unlock(&ctx->completion_lock);
2376 io_cqring_ev_posted(ctx);
2378 io_init_req_batch(&rb);
2379 for (i = 0; i < nr; i++) {
2380 struct io_kiocb *req = state->compl_reqs[i];
2382 if (req_ref_put_and_test(req))
2383 io_req_free_batch(&rb, req, &ctx->submit_state);
2386 io_req_free_batch_finish(ctx, &rb);
2387 state->compl_nr = 0;
2391 * Drop reference to request, return next in chain (if there is one) if this
2392 * was the last reference to this request.
2394 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2396 struct io_kiocb *nxt = NULL;
2398 if (req_ref_put_and_test(req)) {
2399 nxt = io_req_find_next(req);
2405 static inline void io_put_req(struct io_kiocb *req)
2407 if (req_ref_put_and_test(req))
2411 static inline void io_put_req_deferred(struct io_kiocb *req)
2413 if (req_ref_put_and_test(req)) {
2414 req->io_task_work.func = io_free_req_work;
2415 io_req_task_work_add(req);
2419 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2421 /* See comment at the top of this file */
2423 return __io_cqring_events(ctx);
2426 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2428 struct io_rings *rings = ctx->rings;
2430 /* make sure SQ entry isn't read before tail */
2431 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2434 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2436 unsigned int cflags;
2438 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2439 cflags |= IORING_CQE_F_BUFFER;
2440 req->flags &= ~REQ_F_BUFFER_SELECTED;
2445 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2447 struct io_buffer *kbuf;
2449 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2451 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2452 return io_put_kbuf(req, kbuf);
2455 static inline bool io_run_task_work(void)
2457 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2458 __set_current_state(TASK_RUNNING);
2459 tracehook_notify_signal();
2467 * Find and free completed poll iocbs
2469 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2470 struct list_head *done)
2472 struct req_batch rb;
2473 struct io_kiocb *req;
2475 /* order with ->result store in io_complete_rw_iopoll() */
2478 io_init_req_batch(&rb);
2479 while (!list_empty(done)) {
2480 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2481 list_del(&req->inflight_entry);
2483 io_fill_cqe_req(req, req->result, io_put_rw_kbuf(req));
2486 if (req_ref_put_and_test(req))
2487 io_req_free_batch(&rb, req, &ctx->submit_state);
2490 io_commit_cqring(ctx);
2491 io_cqring_ev_posted_iopoll(ctx);
2492 io_req_free_batch_finish(ctx, &rb);
2495 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2498 struct io_kiocb *req, *tmp;
2503 * Only spin for completions if we don't have multiple devices hanging
2504 * off our complete list, and we're under the requested amount.
2506 spin = !ctx->poll_multi_queue && *nr_events < min;
2508 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2509 struct kiocb *kiocb = &req->rw.kiocb;
2513 * Move completed and retryable entries to our local lists.
2514 * If we find a request that requires polling, break out
2515 * and complete those lists first, if we have entries there.
2517 if (READ_ONCE(req->iopoll_completed)) {
2518 list_move_tail(&req->inflight_entry, &done);
2521 if (!list_empty(&done))
2524 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2525 if (unlikely(ret < 0))
2530 /* iopoll may have completed current req */
2531 if (READ_ONCE(req->iopoll_completed))
2532 list_move_tail(&req->inflight_entry, &done);
2535 if (!list_empty(&done))
2536 io_iopoll_complete(ctx, nr_events, &done);
2542 * We can't just wait for polled events to come to us, we have to actively
2543 * find and complete them.
2545 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2547 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2550 mutex_lock(&ctx->uring_lock);
2551 while (!list_empty(&ctx->iopoll_list)) {
2552 unsigned int nr_events = 0;
2554 io_do_iopoll(ctx, &nr_events, 0);
2556 /* let it sleep and repeat later if can't complete a request */
2560 * Ensure we allow local-to-the-cpu processing to take place,
2561 * in this case we need to ensure that we reap all events.
2562 * Also let task_work, etc. to progress by releasing the mutex
2564 if (need_resched()) {
2565 mutex_unlock(&ctx->uring_lock);
2567 mutex_lock(&ctx->uring_lock);
2570 mutex_unlock(&ctx->uring_lock);
2573 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2575 unsigned int nr_events = 0;
2579 * We disallow the app entering submit/complete with polling, but we
2580 * still need to lock the ring to prevent racing with polled issue
2581 * that got punted to a workqueue.
2583 mutex_lock(&ctx->uring_lock);
2585 * Don't enter poll loop if we already have events pending.
2586 * If we do, we can potentially be spinning for commands that
2587 * already triggered a CQE (eg in error).
2589 if (test_bit(0, &ctx->check_cq_overflow))
2590 __io_cqring_overflow_flush(ctx, false);
2591 if (io_cqring_events(ctx))
2595 * If a submit got punted to a workqueue, we can have the
2596 * application entering polling for a command before it gets
2597 * issued. That app will hold the uring_lock for the duration
2598 * of the poll right here, so we need to take a breather every
2599 * now and then to ensure that the issue has a chance to add
2600 * the poll to the issued list. Otherwise we can spin here
2601 * forever, while the workqueue is stuck trying to acquire the
2604 if (list_empty(&ctx->iopoll_list)) {
2605 u32 tail = ctx->cached_cq_tail;
2607 mutex_unlock(&ctx->uring_lock);
2609 mutex_lock(&ctx->uring_lock);
2611 /* some requests don't go through iopoll_list */
2612 if (tail != ctx->cached_cq_tail ||
2613 list_empty(&ctx->iopoll_list))
2616 ret = io_do_iopoll(ctx, &nr_events, min);
2617 } while (!ret && nr_events < min && !need_resched());
2619 mutex_unlock(&ctx->uring_lock);
2623 static void kiocb_end_write(struct io_kiocb *req)
2626 * Tell lockdep we inherited freeze protection from submission
2629 if (req->flags & REQ_F_ISREG) {
2630 struct super_block *sb = file_inode(req->file)->i_sb;
2632 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2638 static bool io_resubmit_prep(struct io_kiocb *req)
2640 struct io_async_rw *rw = req->async_data;
2643 return !io_req_prep_async(req);
2644 iov_iter_restore(&rw->iter, &rw->iter_state);
2648 static bool io_rw_should_reissue(struct io_kiocb *req)
2650 umode_t mode = file_inode(req->file)->i_mode;
2651 struct io_ring_ctx *ctx = req->ctx;
2653 if (!S_ISBLK(mode) && !S_ISREG(mode))
2655 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2656 !(ctx->flags & IORING_SETUP_IOPOLL)))
2659 * If ref is dying, we might be running poll reap from the exit work.
2660 * Don't attempt to reissue from that path, just let it fail with
2663 if (percpu_ref_is_dying(&ctx->refs))
2666 * Play it safe and assume not safe to re-import and reissue if we're
2667 * not in the original thread group (or in task context).
2669 if (!same_thread_group(req->task, current) || !in_task())
2674 static bool io_resubmit_prep(struct io_kiocb *req)
2678 static bool io_rw_should_reissue(struct io_kiocb *req)
2684 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2686 if (req->rw.kiocb.ki_flags & IOCB_WRITE) {
2687 kiocb_end_write(req);
2688 fsnotify_modify(req->file);
2690 fsnotify_access(req->file);
2692 if (res != req->result) {
2693 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2694 io_rw_should_reissue(req)) {
2695 req->flags |= REQ_F_REISSUE;
2704 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2706 unsigned int cflags = io_put_rw_kbuf(req);
2707 int res = req->result;
2710 struct io_ring_ctx *ctx = req->ctx;
2711 struct io_submit_state *state = &ctx->submit_state;
2713 io_req_complete_state(req, res, cflags);
2714 state->compl_reqs[state->compl_nr++] = req;
2715 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2716 io_submit_flush_completions(ctx);
2718 io_req_complete_post(req, res, cflags);
2722 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2723 unsigned int issue_flags)
2725 if (__io_complete_rw_common(req, res))
2727 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2730 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2732 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2734 if (__io_complete_rw_common(req, res))
2737 req->io_task_work.func = io_req_task_complete;
2738 io_req_task_work_add(req);
2741 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2743 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2745 if (kiocb->ki_flags & IOCB_WRITE)
2746 kiocb_end_write(req);
2747 if (unlikely(res != req->result)) {
2748 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2749 req->flags |= REQ_F_REISSUE;
2754 WRITE_ONCE(req->result, res);
2755 /* order with io_iopoll_complete() checking ->result */
2757 WRITE_ONCE(req->iopoll_completed, 1);
2761 * After the iocb has been issued, it's safe to be found on the poll list.
2762 * Adding the kiocb to the list AFTER submission ensures that we don't
2763 * find it from a io_do_iopoll() thread before the issuer is done
2764 * accessing the kiocb cookie.
2766 static void io_iopoll_req_issued(struct io_kiocb *req)
2768 struct io_ring_ctx *ctx = req->ctx;
2769 const bool in_async = io_wq_current_is_worker();
2771 /* workqueue context doesn't hold uring_lock, grab it now */
2772 if (unlikely(in_async))
2773 mutex_lock(&ctx->uring_lock);
2776 * Track whether we have multiple files in our lists. This will impact
2777 * how we do polling eventually, not spinning if we're on potentially
2778 * different devices.
2780 if (list_empty(&ctx->iopoll_list)) {
2781 ctx->poll_multi_queue = false;
2782 } else if (!ctx->poll_multi_queue) {
2783 struct io_kiocb *list_req;
2784 unsigned int queue_num0, queue_num1;
2786 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2789 if (list_req->file != req->file) {
2790 ctx->poll_multi_queue = true;
2792 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2793 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2794 if (queue_num0 != queue_num1)
2795 ctx->poll_multi_queue = true;
2800 * For fast devices, IO may have already completed. If it has, add
2801 * it to the front so we find it first.
2803 if (READ_ONCE(req->iopoll_completed))
2804 list_add(&req->inflight_entry, &ctx->iopoll_list);
2806 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2808 if (unlikely(in_async)) {
2810 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2811 * in sq thread task context or in io worker task context. If
2812 * current task context is sq thread, we don't need to check
2813 * whether should wake up sq thread.
2815 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2816 wq_has_sleeper(&ctx->sq_data->wait))
2817 wake_up(&ctx->sq_data->wait);
2819 mutex_unlock(&ctx->uring_lock);
2823 static bool io_bdev_nowait(struct block_device *bdev)
2825 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2829 * If we tracked the file through the SCM inflight mechanism, we could support
2830 * any file. For now, just ensure that anything potentially problematic is done
2833 static bool __io_file_supports_nowait(struct file *file, int rw)
2835 umode_t mode = file_inode(file)->i_mode;
2837 if (S_ISBLK(mode)) {
2838 if (IS_ENABLED(CONFIG_BLOCK) &&
2839 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2845 if (S_ISREG(mode)) {
2846 if (IS_ENABLED(CONFIG_BLOCK) &&
2847 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2848 file->f_op != &io_uring_fops)
2853 /* any ->read/write should understand O_NONBLOCK */
2854 if (file->f_flags & O_NONBLOCK)
2857 if (!(file->f_mode & FMODE_NOWAIT))
2861 return file->f_op->read_iter != NULL;
2863 return file->f_op->write_iter != NULL;
2866 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2868 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2870 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2873 return __io_file_supports_nowait(req->file, rw);
2876 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2879 struct io_ring_ctx *ctx = req->ctx;
2880 struct kiocb *kiocb = &req->rw.kiocb;
2881 struct file *file = req->file;
2885 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2886 req->flags |= REQ_F_ISREG;
2888 kiocb->ki_pos = READ_ONCE(sqe->off);
2889 if (kiocb->ki_pos == -1) {
2890 if (!(file->f_mode & FMODE_STREAM)) {
2891 req->flags |= REQ_F_CUR_POS;
2892 kiocb->ki_pos = file->f_pos;
2897 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2898 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2899 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2904 * If the file is marked O_NONBLOCK, still allow retry for it if it
2905 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2906 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2908 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2909 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2910 req->flags |= REQ_F_NOWAIT;
2912 ioprio = READ_ONCE(sqe->ioprio);
2914 ret = ioprio_check_cap(ioprio);
2918 kiocb->ki_ioprio = ioprio;
2920 kiocb->ki_ioprio = get_current_ioprio();
2922 if (ctx->flags & IORING_SETUP_IOPOLL) {
2923 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2924 !kiocb->ki_filp->f_op->iopoll)
2927 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2928 kiocb->ki_complete = io_complete_rw_iopoll;
2929 req->iopoll_completed = 0;
2931 if (kiocb->ki_flags & IOCB_HIPRI)
2933 kiocb->ki_complete = io_complete_rw;
2936 /* used for fixed read/write too - just read unconditionally */
2937 req->buf_index = READ_ONCE(sqe->buf_index);
2940 if (req->opcode == IORING_OP_READ_FIXED ||
2941 req->opcode == IORING_OP_WRITE_FIXED) {
2942 struct io_ring_ctx *ctx = req->ctx;
2945 if (unlikely(req->buf_index >= ctx->nr_user_bufs))
2947 index = array_index_nospec(req->buf_index, ctx->nr_user_bufs);
2948 req->imu = ctx->user_bufs[index];
2949 io_req_set_rsrc_node(req);
2952 req->rw.addr = READ_ONCE(sqe->addr);
2953 req->rw.len = READ_ONCE(sqe->len);
2957 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2963 case -ERESTARTNOINTR:
2964 case -ERESTARTNOHAND:
2965 case -ERESTART_RESTARTBLOCK:
2967 * We can't just restart the syscall, since previously
2968 * submitted sqes may already be in progress. Just fail this
2974 kiocb->ki_complete(kiocb, ret, 0);
2978 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2979 unsigned int issue_flags)
2981 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2982 struct io_async_rw *io = req->async_data;
2984 /* add previously done IO, if any */
2985 if (io && io->bytes_done > 0) {
2987 ret = io->bytes_done;
2989 ret += io->bytes_done;
2992 if (req->flags & REQ_F_CUR_POS)
2993 req->file->f_pos = kiocb->ki_pos;
2994 if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
2995 __io_complete_rw(req, ret, 0, issue_flags);
2997 io_rw_done(kiocb, ret);
2999 if (req->flags & REQ_F_REISSUE) {
3000 req->flags &= ~REQ_F_REISSUE;
3001 if (io_resubmit_prep(req)) {
3002 io_req_task_queue_reissue(req);
3004 unsigned int cflags = io_put_rw_kbuf(req);
3005 struct io_ring_ctx *ctx = req->ctx;
3008 if (!(issue_flags & IO_URING_F_NONBLOCK)) {
3009 mutex_lock(&ctx->uring_lock);
3010 __io_req_complete(req, issue_flags, ret, cflags);
3011 mutex_unlock(&ctx->uring_lock);
3013 __io_req_complete(req, issue_flags, ret, cflags);
3019 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
3020 struct io_mapped_ubuf *imu)
3022 size_t len = req->rw.len;
3023 u64 buf_end, buf_addr = req->rw.addr;
3026 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
3028 /* not inside the mapped region */
3029 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
3033 * May not be a start of buffer, set size appropriately
3034 * and advance us to the beginning.
3036 offset = buf_addr - imu->ubuf;
3037 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3041 * Don't use iov_iter_advance() here, as it's really slow for
3042 * using the latter parts of a big fixed buffer - it iterates
3043 * over each segment manually. We can cheat a bit here, because
3046 * 1) it's a BVEC iter, we set it up
3047 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3048 * first and last bvec
3050 * So just find our index, and adjust the iterator afterwards.
3051 * If the offset is within the first bvec (or the whole first
3052 * bvec, just use iov_iter_advance(). This makes it easier
3053 * since we can just skip the first segment, which may not
3054 * be PAGE_SIZE aligned.
3056 const struct bio_vec *bvec = imu->bvec;
3058 if (offset <= bvec->bv_len) {
3059 iov_iter_advance(iter, offset);
3061 unsigned long seg_skip;
3063 /* skip first vec */
3064 offset -= bvec->bv_len;
3065 seg_skip = 1 + (offset >> PAGE_SHIFT);
3067 iter->bvec = bvec + seg_skip;
3068 iter->nr_segs -= seg_skip;
3069 iter->count -= bvec->bv_len + offset;
3070 iter->iov_offset = offset & ~PAGE_MASK;
3077 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3079 if (WARN_ON_ONCE(!req->imu))
3081 return __io_import_fixed(req, rw, iter, req->imu);
3084 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3087 mutex_unlock(&ctx->uring_lock);
3090 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3093 * "Normal" inline submissions always hold the uring_lock, since we
3094 * grab it from the system call. Same is true for the SQPOLL offload.
3095 * The only exception is when we've detached the request and issue it
3096 * from an async worker thread, grab the lock for that case.
3099 mutex_lock(&ctx->uring_lock);
3102 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3103 int bgid, struct io_buffer *kbuf,
3106 struct io_buffer *head;
3108 if (req->flags & REQ_F_BUFFER_SELECTED)
3111 io_ring_submit_lock(req->ctx, needs_lock);
3113 lockdep_assert_held(&req->ctx->uring_lock);
3115 head = xa_load(&req->ctx->io_buffers, bgid);
3117 if (!list_empty(&head->list)) {
3118 kbuf = list_last_entry(&head->list, struct io_buffer,
3120 list_del(&kbuf->list);
3123 xa_erase(&req->ctx->io_buffers, bgid);
3125 if (*len > kbuf->len)
3128 kbuf = ERR_PTR(-ENOBUFS);
3131 io_ring_submit_unlock(req->ctx, needs_lock);
3136 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3139 struct io_buffer *kbuf;
3142 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3143 bgid = req->buf_index;
3144 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3147 req->rw.addr = (u64) (unsigned long) kbuf;
3148 req->flags |= REQ_F_BUFFER_SELECTED;
3149 return u64_to_user_ptr(kbuf->addr);
3152 #ifdef CONFIG_COMPAT
3153 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3156 struct compat_iovec __user *uiov;
3157 compat_ssize_t clen;
3161 uiov = u64_to_user_ptr(req->rw.addr);
3162 if (!access_ok(uiov, sizeof(*uiov)))
3164 if (__get_user(clen, &uiov->iov_len))
3170 buf = io_rw_buffer_select(req, &len, needs_lock);
3172 return PTR_ERR(buf);
3173 iov[0].iov_base = buf;
3174 iov[0].iov_len = (compat_size_t) len;
3179 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3182 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3186 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3189 len = iov[0].iov_len;
3192 buf = io_rw_buffer_select(req, &len, needs_lock);
3194 return PTR_ERR(buf);
3195 iov[0].iov_base = buf;
3196 iov[0].iov_len = len;
3200 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3203 if (req->flags & REQ_F_BUFFER_SELECTED) {
3204 struct io_buffer *kbuf;
3206 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3207 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3208 iov[0].iov_len = kbuf->len;
3211 if (req->rw.len != 1)
3214 #ifdef CONFIG_COMPAT
3215 if (req->ctx->compat)
3216 return io_compat_import(req, iov, needs_lock);
3219 return __io_iov_buffer_select(req, iov, needs_lock);
3222 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3223 struct iov_iter *iter, bool needs_lock)
3225 void __user *buf = u64_to_user_ptr(req->rw.addr);
3226 size_t sqe_len = req->rw.len;
3227 u8 opcode = req->opcode;
3230 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3232 return io_import_fixed(req, rw, iter);
3235 /* buffer index only valid with fixed read/write, or buffer select */
3236 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3239 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3240 if (req->flags & REQ_F_BUFFER_SELECT) {
3241 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3243 return PTR_ERR(buf);
3244 req->rw.len = sqe_len;
3247 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3252 if (req->flags & REQ_F_BUFFER_SELECT) {
3253 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3255 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3260 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3264 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3266 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3270 * For files that don't have ->read_iter() and ->write_iter(), handle them
3271 * by looping over ->read() or ->write() manually.
3273 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3275 struct kiocb *kiocb = &req->rw.kiocb;
3276 struct file *file = req->file;
3280 * Don't support polled IO through this interface, and we can't
3281 * support non-blocking either. For the latter, this just causes
3282 * the kiocb to be handled from an async context.
3284 if (kiocb->ki_flags & IOCB_HIPRI)
3286 if (kiocb->ki_flags & IOCB_NOWAIT)
3289 while (iov_iter_count(iter)) {
3293 if (!iov_iter_is_bvec(iter)) {
3294 iovec = iov_iter_iovec(iter);
3296 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3297 iovec.iov_len = req->rw.len;
3301 nr = file->f_op->read(file, iovec.iov_base,
3302 iovec.iov_len, io_kiocb_ppos(kiocb));
3304 nr = file->f_op->write(file, iovec.iov_base,
3305 iovec.iov_len, io_kiocb_ppos(kiocb));
3314 if (!iov_iter_is_bvec(iter)) {
3315 iov_iter_advance(iter, nr);
3322 if (nr != iovec.iov_len)
3329 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3330 const struct iovec *fast_iov, struct iov_iter *iter)
3332 struct io_async_rw *rw = req->async_data;
3334 memcpy(&rw->iter, iter, sizeof(*iter));
3335 rw->free_iovec = iovec;
3337 /* can only be fixed buffers, no need to do anything */
3338 if (iov_iter_is_bvec(iter))
3341 unsigned iov_off = 0;
3343 rw->iter.iov = rw->fast_iov;
3344 if (iter->iov != fast_iov) {
3345 iov_off = iter->iov - fast_iov;
3346 rw->iter.iov += iov_off;
3348 if (rw->fast_iov != fast_iov)
3349 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3350 sizeof(struct iovec) * iter->nr_segs);
3352 req->flags |= REQ_F_NEED_CLEANUP;
3356 static inline int io_alloc_async_data(struct io_kiocb *req)
3358 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3359 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3360 return req->async_data == NULL;
3363 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3364 const struct iovec *fast_iov,
3365 struct iov_iter *iter, bool force)
3367 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3369 if (!req->async_data) {
3370 struct io_async_rw *iorw;
3372 if (io_alloc_async_data(req)) {
3377 io_req_map_rw(req, iovec, fast_iov, iter);
3378 iorw = req->async_data;
3379 /* we've copied and mapped the iter, ensure state is saved */
3380 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3385 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3387 struct io_async_rw *iorw = req->async_data;
3388 struct iovec *iov = iorw->fast_iov;
3391 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3392 if (unlikely(ret < 0))
3395 iorw->bytes_done = 0;
3396 iorw->free_iovec = iov;
3398 req->flags |= REQ_F_NEED_CLEANUP;
3399 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3403 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3405 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3407 return io_prep_rw(req, sqe, READ);
3411 * This is our waitqueue callback handler, registered through lock_page_async()
3412 * when we initially tried to do the IO with the iocb armed our waitqueue.
3413 * This gets called when the page is unlocked, and we generally expect that to
3414 * happen when the page IO is completed and the page is now uptodate. This will
3415 * queue a task_work based retry of the operation, attempting to copy the data
3416 * again. If the latter fails because the page was NOT uptodate, then we will
3417 * do a thread based blocking retry of the operation. That's the unexpected
3420 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3421 int sync, void *arg)
3423 struct wait_page_queue *wpq;
3424 struct io_kiocb *req = wait->private;
3425 struct wait_page_key *key = arg;
3427 wpq = container_of(wait, struct wait_page_queue, wait);
3429 if (!wake_page_match(wpq, key))
3432 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3433 list_del_init(&wait->entry);
3434 io_req_task_queue(req);
3439 * This controls whether a given IO request should be armed for async page
3440 * based retry. If we return false here, the request is handed to the async
3441 * worker threads for retry. If we're doing buffered reads on a regular file,
3442 * we prepare a private wait_page_queue entry and retry the operation. This
3443 * will either succeed because the page is now uptodate and unlocked, or it
3444 * will register a callback when the page is unlocked at IO completion. Through
3445 * that callback, io_uring uses task_work to setup a retry of the operation.
3446 * That retry will attempt the buffered read again. The retry will generally
3447 * succeed, or in rare cases where it fails, we then fall back to using the
3448 * async worker threads for a blocking retry.
3450 static bool io_rw_should_retry(struct io_kiocb *req)
3452 struct io_async_rw *rw = req->async_data;
3453 struct wait_page_queue *wait = &rw->wpq;
3454 struct kiocb *kiocb = &req->rw.kiocb;
3456 /* never retry for NOWAIT, we just complete with -EAGAIN */
3457 if (req->flags & REQ_F_NOWAIT)
3460 /* Only for buffered IO */
3461 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3465 * just use poll if we can, and don't attempt if the fs doesn't
3466 * support callback based unlocks
3468 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3471 wait->wait.func = io_async_buf_func;
3472 wait->wait.private = req;
3473 wait->wait.flags = 0;
3474 INIT_LIST_HEAD(&wait->wait.entry);
3475 kiocb->ki_flags |= IOCB_WAITQ;
3476 kiocb->ki_flags &= ~IOCB_NOWAIT;
3477 kiocb->ki_waitq = wait;
3481 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3483 if (req->file->f_op->read_iter)
3484 return call_read_iter(req->file, &req->rw.kiocb, iter);
3485 else if (req->file->f_op->read)
3486 return loop_rw_iter(READ, req, iter);
3491 static bool need_read_all(struct io_kiocb *req)
3493 return req->flags & REQ_F_ISREG ||
3494 S_ISBLK(file_inode(req->file)->i_mode);
3497 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3499 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3500 struct kiocb *kiocb = &req->rw.kiocb;
3501 struct iov_iter __iter, *iter = &__iter;
3502 struct io_async_rw *rw = req->async_data;
3503 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3504 struct iov_iter_state __state, *state;
3509 state = &rw->iter_state;
3511 * We come here from an earlier attempt, restore our state to
3512 * match in case it doesn't. It's cheap enough that we don't
3513 * need to make this conditional.
3515 iov_iter_restore(iter, state);
3518 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3522 iov_iter_save_state(iter, state);
3524 req->result = iov_iter_count(iter);
3526 /* Ensure we clear previously set non-block flag */
3527 if (!force_nonblock)
3528 kiocb->ki_flags &= ~IOCB_NOWAIT;
3530 kiocb->ki_flags |= IOCB_NOWAIT;
3532 /* If the file doesn't support async, just async punt */
3533 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3534 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3535 return ret ?: -EAGAIN;
3538 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3539 if (unlikely(ret)) {
3544 ret = io_iter_do_read(req, iter);
3546 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3547 req->flags &= ~REQ_F_REISSUE;
3548 /* IOPOLL retry should happen for io-wq threads */
3549 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3551 /* no retry on NONBLOCK nor RWF_NOWAIT */
3552 if (req->flags & REQ_F_NOWAIT)
3555 } else if (ret == -EIOCBQUEUED) {
3557 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3558 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3559 /* read all, failed, already did sync or don't want to retry */
3564 * Don't depend on the iter state matching what was consumed, or being
3565 * untouched in case of error. Restore it and we'll advance it
3566 * manually if we need to.
3568 iov_iter_restore(iter, state);
3570 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3575 rw = req->async_data;
3577 * Now use our persistent iterator and state, if we aren't already.
3578 * We've restored and mapped the iter to match.
3580 if (iter != &rw->iter) {
3582 state = &rw->iter_state;
3587 * We end up here because of a partial read, either from
3588 * above or inside this loop. Advance the iter by the bytes
3589 * that were consumed.
3591 iov_iter_advance(iter, ret);
3592 if (!iov_iter_count(iter))
3594 rw->bytes_done += ret;
3595 iov_iter_save_state(iter, state);
3597 /* if we can retry, do so with the callbacks armed */
3598 if (!io_rw_should_retry(req)) {
3599 kiocb->ki_flags &= ~IOCB_WAITQ;
3604 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3605 * we get -EIOCBQUEUED, then we'll get a notification when the
3606 * desired page gets unlocked. We can also get a partial read
3607 * here, and if we do, then just retry at the new offset.
3609 ret = io_iter_do_read(req, iter);
3610 if (ret == -EIOCBQUEUED)
3612 /* we got some bytes, but not all. retry. */
3613 kiocb->ki_flags &= ~IOCB_WAITQ;
3614 iov_iter_restore(iter, state);
3617 kiocb_done(kiocb, ret, issue_flags);
3619 /* it's faster to check here then delegate to kfree */
3625 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3627 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3629 return io_prep_rw(req, sqe, WRITE);
3632 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3634 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3635 struct kiocb *kiocb = &req->rw.kiocb;
3636 struct iov_iter __iter, *iter = &__iter;
3637 struct io_async_rw *rw = req->async_data;
3638 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3639 struct iov_iter_state __state, *state;
3644 state = &rw->iter_state;
3645 iov_iter_restore(iter, state);
3648 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3652 iov_iter_save_state(iter, state);
3654 req->result = iov_iter_count(iter);
3656 /* Ensure we clear previously set non-block flag */
3657 if (!force_nonblock)
3658 kiocb->ki_flags &= ~IOCB_NOWAIT;
3660 kiocb->ki_flags |= IOCB_NOWAIT;
3662 /* If the file doesn't support async, just async punt */
3663 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3666 /* file path doesn't support NOWAIT for non-direct_IO */
3667 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3668 (req->flags & REQ_F_ISREG))
3671 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3676 * Open-code file_start_write here to grab freeze protection,
3677 * which will be released by another thread in
3678 * io_complete_rw(). Fool lockdep by telling it the lock got
3679 * released so that it doesn't complain about the held lock when
3680 * we return to userspace.
3682 if (req->flags & REQ_F_ISREG) {
3683 sb_start_write(file_inode(req->file)->i_sb);
3684 __sb_writers_release(file_inode(req->file)->i_sb,
3687 kiocb->ki_flags |= IOCB_WRITE;
3689 if (req->file->f_op->write_iter)
3690 ret2 = call_write_iter(req->file, kiocb, iter);
3691 else if (req->file->f_op->write)
3692 ret2 = loop_rw_iter(WRITE, req, iter);
3696 if (req->flags & REQ_F_REISSUE) {
3697 req->flags &= ~REQ_F_REISSUE;
3702 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3703 * retry them without IOCB_NOWAIT.
3705 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3707 /* no retry on NONBLOCK nor RWF_NOWAIT */
3708 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3710 if (!force_nonblock || ret2 != -EAGAIN) {
3711 /* IOPOLL retry should happen for io-wq threads */
3712 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3715 kiocb_done(kiocb, ret2, issue_flags);
3718 iov_iter_restore(iter, state);
3719 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3721 if (kiocb->ki_flags & IOCB_WRITE)
3722 kiocb_end_write(req);
3728 /* it's reportedly faster than delegating the null check to kfree() */
3734 static int io_renameat_prep(struct io_kiocb *req,
3735 const struct io_uring_sqe *sqe)
3737 struct io_rename *ren = &req->rename;
3738 const char __user *oldf, *newf;
3740 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3742 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3744 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3747 ren->old_dfd = READ_ONCE(sqe->fd);
3748 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3749 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3750 ren->new_dfd = READ_ONCE(sqe->len);
3751 ren->flags = READ_ONCE(sqe->rename_flags);
3753 ren->oldpath = getname(oldf);
3754 if (IS_ERR(ren->oldpath))
3755 return PTR_ERR(ren->oldpath);
3757 ren->newpath = getname(newf);
3758 if (IS_ERR(ren->newpath)) {
3759 putname(ren->oldpath);
3760 return PTR_ERR(ren->newpath);
3763 req->flags |= REQ_F_NEED_CLEANUP;
3767 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3769 struct io_rename *ren = &req->rename;
3772 if (issue_flags & IO_URING_F_NONBLOCK)
3775 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3776 ren->newpath, ren->flags);
3778 req->flags &= ~REQ_F_NEED_CLEANUP;
3781 io_req_complete(req, ret);
3785 static int io_unlinkat_prep(struct io_kiocb *req,
3786 const struct io_uring_sqe *sqe)
3788 struct io_unlink *un = &req->unlink;
3789 const char __user *fname;
3791 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3793 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3796 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3799 un->dfd = READ_ONCE(sqe->fd);
3801 un->flags = READ_ONCE(sqe->unlink_flags);
3802 if (un->flags & ~AT_REMOVEDIR)
3805 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3806 un->filename = getname(fname);
3807 if (IS_ERR(un->filename))
3808 return PTR_ERR(un->filename);
3810 req->flags |= REQ_F_NEED_CLEANUP;
3814 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3816 struct io_unlink *un = &req->unlink;
3819 if (issue_flags & IO_URING_F_NONBLOCK)
3822 if (un->flags & AT_REMOVEDIR)
3823 ret = do_rmdir(un->dfd, un->filename);
3825 ret = do_unlinkat(un->dfd, un->filename);
3827 req->flags &= ~REQ_F_NEED_CLEANUP;
3830 io_req_complete(req, ret);
3834 static int io_mkdirat_prep(struct io_kiocb *req,
3835 const struct io_uring_sqe *sqe)
3837 struct io_mkdir *mkd = &req->mkdir;
3838 const char __user *fname;
3840 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3842 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3845 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3848 mkd->dfd = READ_ONCE(sqe->fd);
3849 mkd->mode = READ_ONCE(sqe->len);
3851 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3852 mkd->filename = getname(fname);
3853 if (IS_ERR(mkd->filename))
3854 return PTR_ERR(mkd->filename);
3856 req->flags |= REQ_F_NEED_CLEANUP;
3860 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3862 struct io_mkdir *mkd = &req->mkdir;
3865 if (issue_flags & IO_URING_F_NONBLOCK)
3868 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3870 req->flags &= ~REQ_F_NEED_CLEANUP;
3873 io_req_complete(req, ret);
3877 static int io_symlinkat_prep(struct io_kiocb *req,
3878 const struct io_uring_sqe *sqe)
3880 struct io_symlink *sl = &req->symlink;
3881 const char __user *oldpath, *newpath;
3883 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3885 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3888 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3891 sl->new_dfd = READ_ONCE(sqe->fd);
3892 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3893 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3895 sl->oldpath = getname(oldpath);
3896 if (IS_ERR(sl->oldpath))
3897 return PTR_ERR(sl->oldpath);
3899 sl->newpath = getname(newpath);
3900 if (IS_ERR(sl->newpath)) {
3901 putname(sl->oldpath);
3902 return PTR_ERR(sl->newpath);
3905 req->flags |= REQ_F_NEED_CLEANUP;
3909 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3911 struct io_symlink *sl = &req->symlink;
3914 if (issue_flags & IO_URING_F_NONBLOCK)
3917 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3919 req->flags &= ~REQ_F_NEED_CLEANUP;
3922 io_req_complete(req, ret);
3926 static int io_linkat_prep(struct io_kiocb *req,
3927 const struct io_uring_sqe *sqe)
3929 struct io_hardlink *lnk = &req->hardlink;
3930 const char __user *oldf, *newf;
3932 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3934 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3936 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3939 lnk->old_dfd = READ_ONCE(sqe->fd);
3940 lnk->new_dfd = READ_ONCE(sqe->len);
3941 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3942 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3943 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3945 lnk->oldpath = getname(oldf);
3946 if (IS_ERR(lnk->oldpath))
3947 return PTR_ERR(lnk->oldpath);
3949 lnk->newpath = getname(newf);
3950 if (IS_ERR(lnk->newpath)) {
3951 putname(lnk->oldpath);
3952 return PTR_ERR(lnk->newpath);
3955 req->flags |= REQ_F_NEED_CLEANUP;
3959 static int io_linkat(struct io_kiocb *req, int issue_flags)
3961 struct io_hardlink *lnk = &req->hardlink;
3964 if (issue_flags & IO_URING_F_NONBLOCK)
3967 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3968 lnk->newpath, lnk->flags);
3970 req->flags &= ~REQ_F_NEED_CLEANUP;
3973 io_req_complete(req, ret);
3977 static int io_shutdown_prep(struct io_kiocb *req,
3978 const struct io_uring_sqe *sqe)
3980 #if defined(CONFIG_NET)
3981 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3983 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3984 sqe->buf_index || sqe->splice_fd_in))
3987 req->shutdown.how = READ_ONCE(sqe->len);
3994 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3996 #if defined(CONFIG_NET)
3997 struct socket *sock;
4000 if (issue_flags & IO_URING_F_NONBLOCK)
4003 sock = sock_from_file(req->file);
4004 if (unlikely(!sock))
4007 ret = __sys_shutdown_sock(sock, req->shutdown.how);
4010 io_req_complete(req, ret);
4017 static int __io_splice_prep(struct io_kiocb *req,
4018 const struct io_uring_sqe *sqe)
4020 struct io_splice *sp = &req->splice;
4021 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
4023 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4026 sp->len = READ_ONCE(sqe->len);
4027 sp->flags = READ_ONCE(sqe->splice_flags);
4028 if (unlikely(sp->flags & ~valid_flags))
4030 sp->splice_fd_in = READ_ONCE(sqe->splice_fd_in);
4034 static int io_tee_prep(struct io_kiocb *req,
4035 const struct io_uring_sqe *sqe)
4037 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
4039 return __io_splice_prep(req, sqe);
4042 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
4044 struct io_splice *sp = &req->splice;
4045 struct file *out = sp->file_out;
4046 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4050 if (issue_flags & IO_URING_F_NONBLOCK)
4053 in = io_file_get(req->ctx, req, sp->splice_fd_in,
4054 (sp->flags & SPLICE_F_FD_IN_FIXED));
4061 ret = do_tee(in, out, sp->len, flags);
4063 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4068 io_req_complete(req, ret);
4072 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4074 struct io_splice *sp = &req->splice;
4076 sp->off_in = READ_ONCE(sqe->splice_off_in);
4077 sp->off_out = READ_ONCE(sqe->off);
4078 return __io_splice_prep(req, sqe);
4081 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4083 struct io_splice *sp = &req->splice;
4084 struct file *out = sp->file_out;
4085 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4086 loff_t *poff_in, *poff_out;
4090 if (issue_flags & IO_URING_F_NONBLOCK)
4093 in = io_file_get(req->ctx, req, sp->splice_fd_in,
4094 (sp->flags & SPLICE_F_FD_IN_FIXED));
4100 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4101 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4104 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4106 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4111 io_req_complete(req, ret);
4116 * IORING_OP_NOP just posts a completion event, nothing else.
4118 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4120 struct io_ring_ctx *ctx = req->ctx;
4122 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4125 __io_req_complete(req, issue_flags, 0, 0);
4129 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4131 struct io_ring_ctx *ctx = req->ctx;
4133 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4135 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4139 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4140 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4143 req->sync.off = READ_ONCE(sqe->off);
4144 req->sync.len = READ_ONCE(sqe->len);
4148 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4150 loff_t end = req->sync.off + req->sync.len;
4153 /* fsync always requires a blocking context */
4154 if (issue_flags & IO_URING_F_NONBLOCK)
4157 ret = vfs_fsync_range(req->file, req->sync.off,
4158 end > 0 ? end : LLONG_MAX,
4159 req->sync.flags & IORING_FSYNC_DATASYNC);
4162 io_req_complete(req, ret);
4166 static int io_fallocate_prep(struct io_kiocb *req,
4167 const struct io_uring_sqe *sqe)
4169 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4172 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4175 req->sync.off = READ_ONCE(sqe->off);
4176 req->sync.len = READ_ONCE(sqe->addr);
4177 req->sync.mode = READ_ONCE(sqe->len);
4181 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4185 /* fallocate always requiring blocking context */
4186 if (issue_flags & IO_URING_F_NONBLOCK)
4188 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4193 fsnotify_modify(req->file);
4194 io_req_complete(req, ret);
4198 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4200 const char __user *fname;
4203 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4205 if (unlikely(sqe->ioprio || sqe->buf_index))
4207 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4210 /* open.how should be already initialised */
4211 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4212 req->open.how.flags |= O_LARGEFILE;
4214 req->open.dfd = READ_ONCE(sqe->fd);
4215 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4216 req->open.filename = getname(fname);
4217 if (IS_ERR(req->open.filename)) {
4218 ret = PTR_ERR(req->open.filename);
4219 req->open.filename = NULL;
4223 req->open.file_slot = READ_ONCE(sqe->file_index);
4224 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4227 req->open.nofile = rlimit(RLIMIT_NOFILE);
4228 req->flags |= REQ_F_NEED_CLEANUP;
4232 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4234 u64 mode = READ_ONCE(sqe->len);
4235 u64 flags = READ_ONCE(sqe->open_flags);
4237 req->open.how = build_open_how(flags, mode);
4238 return __io_openat_prep(req, sqe);
4241 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4243 struct open_how __user *how;
4247 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4248 len = READ_ONCE(sqe->len);
4249 if (len < OPEN_HOW_SIZE_VER0)
4252 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4257 return __io_openat_prep(req, sqe);
4260 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4262 struct open_flags op;
4264 bool resolve_nonblock, nonblock_set;
4265 bool fixed = !!req->open.file_slot;
4268 ret = build_open_flags(&req->open.how, &op);
4271 nonblock_set = op.open_flag & O_NONBLOCK;
4272 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4273 if (issue_flags & IO_URING_F_NONBLOCK) {
4275 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4276 * it'll always -EAGAIN
4278 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4280 op.lookup_flags |= LOOKUP_CACHED;
4281 op.open_flag |= O_NONBLOCK;
4285 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4290 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4293 * We could hang on to this 'fd' on retrying, but seems like
4294 * marginal gain for something that is now known to be a slower
4295 * path. So just put it, and we'll get a new one when we retry.
4300 ret = PTR_ERR(file);
4301 /* only retry if RESOLVE_CACHED wasn't already set by application */
4302 if (ret == -EAGAIN &&
4303 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4308 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4309 file->f_flags &= ~O_NONBLOCK;
4310 fsnotify_open(file);
4313 fd_install(ret, file);
4315 ret = io_install_fixed_file(req, file, issue_flags,
4316 req->open.file_slot - 1);
4318 putname(req->open.filename);
4319 req->flags &= ~REQ_F_NEED_CLEANUP;
4322 __io_req_complete(req, issue_flags, ret, 0);
4326 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4328 return io_openat2(req, issue_flags);
4331 static int io_remove_buffers_prep(struct io_kiocb *req,
4332 const struct io_uring_sqe *sqe)
4334 struct io_provide_buf *p = &req->pbuf;
4337 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4341 tmp = READ_ONCE(sqe->fd);
4342 if (!tmp || tmp > USHRT_MAX)
4345 memset(p, 0, sizeof(*p));
4347 p->bgid = READ_ONCE(sqe->buf_group);
4351 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4352 int bgid, unsigned nbufs)
4356 /* shouldn't happen */
4360 /* the head kbuf is the list itself */
4361 while (!list_empty(&buf->list)) {
4362 struct io_buffer *nxt;
4364 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4365 list_del(&nxt->list);
4373 xa_erase(&ctx->io_buffers, bgid);
4378 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4380 struct io_provide_buf *p = &req->pbuf;
4381 struct io_ring_ctx *ctx = req->ctx;
4382 struct io_buffer *head;
4384 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4386 io_ring_submit_lock(ctx, !force_nonblock);
4388 lockdep_assert_held(&ctx->uring_lock);
4391 head = xa_load(&ctx->io_buffers, p->bgid);
4393 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4397 /* complete before unlock, IOPOLL may need the lock */
4398 __io_req_complete(req, issue_flags, ret, 0);
4399 io_ring_submit_unlock(ctx, !force_nonblock);
4403 static int io_provide_buffers_prep(struct io_kiocb *req,
4404 const struct io_uring_sqe *sqe)
4406 unsigned long size, tmp_check;
4407 struct io_provide_buf *p = &req->pbuf;
4410 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4413 tmp = READ_ONCE(sqe->fd);
4414 if (!tmp || tmp > USHRT_MAX)
4417 p->addr = READ_ONCE(sqe->addr);
4418 p->len = READ_ONCE(sqe->len);
4420 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4423 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4426 size = (unsigned long)p->len * p->nbufs;
4427 if (!access_ok(u64_to_user_ptr(p->addr), size))
4430 p->bgid = READ_ONCE(sqe->buf_group);
4431 tmp = READ_ONCE(sqe->off);
4432 if (tmp > USHRT_MAX)
4438 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4440 struct io_buffer *buf;
4441 u64 addr = pbuf->addr;
4442 int i, bid = pbuf->bid;
4444 for (i = 0; i < pbuf->nbufs; i++) {
4445 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4450 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4455 INIT_LIST_HEAD(&buf->list);
4458 list_add_tail(&buf->list, &(*head)->list);
4463 return i ? i : -ENOMEM;
4466 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4468 struct io_provide_buf *p = &req->pbuf;
4469 struct io_ring_ctx *ctx = req->ctx;
4470 struct io_buffer *head, *list;
4472 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4474 io_ring_submit_lock(ctx, !force_nonblock);
4476 lockdep_assert_held(&ctx->uring_lock);
4478 list = head = xa_load(&ctx->io_buffers, p->bgid);
4480 ret = io_add_buffers(p, &head);
4481 if (ret >= 0 && !list) {
4482 ret = xa_insert(&ctx->io_buffers, p->bgid, head,
4483 GFP_KERNEL_ACCOUNT);
4485 __io_remove_buffers(ctx, head, p->bgid, -1U);
4489 /* complete before unlock, IOPOLL may need the lock */
4490 __io_req_complete(req, issue_flags, ret, 0);
4491 io_ring_submit_unlock(ctx, !force_nonblock);
4495 static int io_epoll_ctl_prep(struct io_kiocb *req,
4496 const struct io_uring_sqe *sqe)
4498 #if defined(CONFIG_EPOLL)
4499 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4501 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4504 req->epoll.epfd = READ_ONCE(sqe->fd);
4505 req->epoll.op = READ_ONCE(sqe->len);
4506 req->epoll.fd = READ_ONCE(sqe->off);
4508 if (ep_op_has_event(req->epoll.op)) {
4509 struct epoll_event __user *ev;
4511 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4512 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4522 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4524 #if defined(CONFIG_EPOLL)
4525 struct io_epoll *ie = &req->epoll;
4527 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4529 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4530 if (force_nonblock && ret == -EAGAIN)
4535 __io_req_complete(req, issue_flags, ret, 0);
4542 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4544 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4545 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4547 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4550 req->madvise.addr = READ_ONCE(sqe->addr);
4551 req->madvise.len = READ_ONCE(sqe->len);
4552 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4559 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4561 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4562 struct io_madvise *ma = &req->madvise;
4565 if (issue_flags & IO_URING_F_NONBLOCK)
4568 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4571 io_req_complete(req, ret);
4578 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4580 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4582 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4585 req->fadvise.offset = READ_ONCE(sqe->off);
4586 req->fadvise.len = READ_ONCE(sqe->len);
4587 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4591 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4593 struct io_fadvise *fa = &req->fadvise;
4596 if (issue_flags & IO_URING_F_NONBLOCK) {
4597 switch (fa->advice) {
4598 case POSIX_FADV_NORMAL:
4599 case POSIX_FADV_RANDOM:
4600 case POSIX_FADV_SEQUENTIAL:
4607 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4610 __io_req_complete(req, issue_flags, ret, 0);
4614 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4616 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4618 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4620 if (req->flags & REQ_F_FIXED_FILE)
4623 req->statx.dfd = READ_ONCE(sqe->fd);
4624 req->statx.mask = READ_ONCE(sqe->len);
4625 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4626 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4627 req->statx.flags = READ_ONCE(sqe->statx_flags);
4632 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4634 struct io_statx *ctx = &req->statx;
4637 if (issue_flags & IO_URING_F_NONBLOCK)
4640 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4645 io_req_complete(req, ret);
4649 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4651 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4653 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4654 sqe->rw_flags || sqe->buf_index)
4656 if (req->flags & REQ_F_FIXED_FILE)
4659 req->close.fd = READ_ONCE(sqe->fd);
4660 req->close.file_slot = READ_ONCE(sqe->file_index);
4661 if (req->close.file_slot && req->close.fd)
4667 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4669 struct files_struct *files = current->files;
4670 struct io_close *close = &req->close;
4671 struct fdtable *fdt;
4672 struct file *file = NULL;
4675 if (req->close.file_slot) {
4676 ret = io_close_fixed(req, issue_flags);
4680 spin_lock(&files->file_lock);
4681 fdt = files_fdtable(files);
4682 if (close->fd >= fdt->max_fds) {
4683 spin_unlock(&files->file_lock);
4686 file = fdt->fd[close->fd];
4687 if (!file || file->f_op == &io_uring_fops) {
4688 spin_unlock(&files->file_lock);
4693 /* if the file has a flush method, be safe and punt to async */
4694 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4695 spin_unlock(&files->file_lock);
4699 ret = __close_fd_get_file(close->fd, &file);
4700 spin_unlock(&files->file_lock);
4707 /* No ->flush() or already async, safely close from here */
4708 ret = filp_close(file, current->files);
4714 __io_req_complete(req, issue_flags, ret, 0);
4718 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4720 struct io_ring_ctx *ctx = req->ctx;
4722 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4724 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4728 req->sync.off = READ_ONCE(sqe->off);
4729 req->sync.len = READ_ONCE(sqe->len);
4730 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4734 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4738 /* sync_file_range always requires a blocking context */
4739 if (issue_flags & IO_URING_F_NONBLOCK)
4742 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4746 io_req_complete(req, ret);
4750 #if defined(CONFIG_NET)
4751 static int io_setup_async_msg(struct io_kiocb *req,
4752 struct io_async_msghdr *kmsg)
4754 struct io_async_msghdr *async_msg = req->async_data;
4758 if (io_alloc_async_data(req)) {
4759 kfree(kmsg->free_iov);
4762 async_msg = req->async_data;
4763 req->flags |= REQ_F_NEED_CLEANUP;
4764 memcpy(async_msg, kmsg, sizeof(*kmsg));
4765 async_msg->msg.msg_name = &async_msg->addr;
4766 /* if were using fast_iov, set it to the new one */
4767 if (!async_msg->free_iov)
4768 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4773 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4774 struct io_async_msghdr *iomsg)
4776 iomsg->msg.msg_name = &iomsg->addr;
4777 iomsg->free_iov = iomsg->fast_iov;
4778 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4779 req->sr_msg.msg_flags, &iomsg->free_iov);
4782 static int io_sendmsg_prep_async(struct io_kiocb *req)
4786 ret = io_sendmsg_copy_hdr(req, req->async_data);
4788 req->flags |= REQ_F_NEED_CLEANUP;
4792 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4794 struct io_sr_msg *sr = &req->sr_msg;
4796 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4798 if (unlikely(sqe->addr2 || sqe->file_index))
4800 if (unlikely(sqe->addr2 || sqe->file_index || sqe->ioprio))
4803 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4804 sr->len = READ_ONCE(sqe->len);
4805 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4806 if (sr->msg_flags & MSG_DONTWAIT)
4807 req->flags |= REQ_F_NOWAIT;
4809 #ifdef CONFIG_COMPAT
4810 if (req->ctx->compat)
4811 sr->msg_flags |= MSG_CMSG_COMPAT;
4816 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4818 struct io_async_msghdr iomsg, *kmsg;
4819 struct socket *sock;
4824 sock = sock_from_file(req->file);
4825 if (unlikely(!sock))
4828 kmsg = req->async_data;
4830 ret = io_sendmsg_copy_hdr(req, &iomsg);
4836 flags = req->sr_msg.msg_flags;
4837 if (issue_flags & IO_URING_F_NONBLOCK)
4838 flags |= MSG_DONTWAIT;
4839 if (flags & MSG_WAITALL)
4840 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4842 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4843 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4844 return io_setup_async_msg(req, kmsg);
4845 if (ret == -ERESTARTSYS)
4848 /* fast path, check for non-NULL to avoid function call */
4850 kfree(kmsg->free_iov);
4851 req->flags &= ~REQ_F_NEED_CLEANUP;
4854 __io_req_complete(req, issue_flags, ret, 0);
4858 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4860 struct io_sr_msg *sr = &req->sr_msg;
4863 struct socket *sock;
4868 sock = sock_from_file(req->file);
4869 if (unlikely(!sock))
4872 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4876 msg.msg_name = NULL;
4877 msg.msg_control = NULL;
4878 msg.msg_controllen = 0;
4879 msg.msg_namelen = 0;
4881 flags = req->sr_msg.msg_flags;
4882 if (issue_flags & IO_URING_F_NONBLOCK)
4883 flags |= MSG_DONTWAIT;
4884 if (flags & MSG_WAITALL)
4885 min_ret = iov_iter_count(&msg.msg_iter);
4887 msg.msg_flags = flags;
4888 ret = sock_sendmsg(sock, &msg);
4889 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4891 if (ret == -ERESTARTSYS)
4896 __io_req_complete(req, issue_flags, ret, 0);
4900 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4901 struct io_async_msghdr *iomsg)
4903 struct io_sr_msg *sr = &req->sr_msg;
4904 struct iovec __user *uiov;
4908 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4909 &iomsg->uaddr, &uiov, &iov_len);
4913 if (req->flags & REQ_F_BUFFER_SELECT) {
4916 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4918 sr->len = iomsg->fast_iov[0].iov_len;
4919 iomsg->free_iov = NULL;
4921 iomsg->free_iov = iomsg->fast_iov;
4922 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4923 &iomsg->free_iov, &iomsg->msg.msg_iter,
4932 #ifdef CONFIG_COMPAT
4933 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4934 struct io_async_msghdr *iomsg)
4936 struct io_sr_msg *sr = &req->sr_msg;
4937 struct compat_iovec __user *uiov;
4942 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4947 uiov = compat_ptr(ptr);
4948 if (req->flags & REQ_F_BUFFER_SELECT) {
4949 compat_ssize_t clen;
4953 if (!access_ok(uiov, sizeof(*uiov)))
4955 if (__get_user(clen, &uiov->iov_len))
4960 iomsg->free_iov = NULL;
4962 iomsg->free_iov = iomsg->fast_iov;
4963 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4964 UIO_FASTIOV, &iomsg->free_iov,
4965 &iomsg->msg.msg_iter, true);
4974 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4975 struct io_async_msghdr *iomsg)
4977 iomsg->msg.msg_name = &iomsg->addr;
4979 #ifdef CONFIG_COMPAT
4980 if (req->ctx->compat)
4981 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4984 return __io_recvmsg_copy_hdr(req, iomsg);
4987 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4990 struct io_sr_msg *sr = &req->sr_msg;
4991 struct io_buffer *kbuf;
4993 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4998 req->flags |= REQ_F_BUFFER_SELECTED;
5002 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
5004 return io_put_kbuf(req, req->sr_msg.kbuf);
5007 static int io_recvmsg_prep_async(struct io_kiocb *req)
5011 ret = io_recvmsg_copy_hdr(req, req->async_data);
5013 req->flags |= REQ_F_NEED_CLEANUP;
5017 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5019 struct io_sr_msg *sr = &req->sr_msg;
5021 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5023 if (unlikely(sqe->addr2 || sqe->file_index))
5025 if (unlikely(sqe->addr2 || sqe->file_index || sqe->ioprio))
5028 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
5029 sr->len = READ_ONCE(sqe->len);
5030 sr->bgid = READ_ONCE(sqe->buf_group);
5031 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
5032 if (sr->msg_flags & MSG_DONTWAIT)
5033 req->flags |= REQ_F_NOWAIT;
5035 #ifdef CONFIG_COMPAT
5036 if (req->ctx->compat)
5037 sr->msg_flags |= MSG_CMSG_COMPAT;
5042 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
5044 struct io_async_msghdr iomsg, *kmsg;
5045 struct socket *sock;
5046 struct io_buffer *kbuf;
5049 int ret, cflags = 0;
5050 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5052 sock = sock_from_file(req->file);
5053 if (unlikely(!sock))
5056 kmsg = req->async_data;
5058 ret = io_recvmsg_copy_hdr(req, &iomsg);
5064 if (req->flags & REQ_F_BUFFER_SELECT) {
5065 kbuf = io_recv_buffer_select(req, !force_nonblock);
5067 return PTR_ERR(kbuf);
5068 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5069 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5070 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5071 1, req->sr_msg.len);
5074 flags = req->sr_msg.msg_flags;
5076 flags |= MSG_DONTWAIT;
5077 if (flags & MSG_WAITALL)
5078 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5080 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5081 kmsg->uaddr, flags);
5082 if (force_nonblock && ret == -EAGAIN)
5083 return io_setup_async_msg(req, kmsg);
5084 if (ret == -ERESTARTSYS)
5087 if (req->flags & REQ_F_BUFFER_SELECTED)
5088 cflags = io_put_recv_kbuf(req);
5089 /* fast path, check for non-NULL to avoid function call */
5091 kfree(kmsg->free_iov);
5092 req->flags &= ~REQ_F_NEED_CLEANUP;
5093 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5095 __io_req_complete(req, issue_flags, ret, cflags);
5099 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5101 struct io_buffer *kbuf;
5102 struct io_sr_msg *sr = &req->sr_msg;
5104 void __user *buf = sr->buf;
5105 struct socket *sock;
5109 int ret, cflags = 0;
5110 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5112 sock = sock_from_file(req->file);
5113 if (unlikely(!sock))
5116 if (req->flags & REQ_F_BUFFER_SELECT) {
5117 kbuf = io_recv_buffer_select(req, !force_nonblock);
5119 return PTR_ERR(kbuf);
5120 buf = u64_to_user_ptr(kbuf->addr);
5123 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5127 msg.msg_name = NULL;
5128 msg.msg_control = NULL;
5129 msg.msg_controllen = 0;
5130 msg.msg_namelen = 0;
5131 msg.msg_iocb = NULL;
5134 flags = req->sr_msg.msg_flags;
5136 flags |= MSG_DONTWAIT;
5137 if (flags & MSG_WAITALL)
5138 min_ret = iov_iter_count(&msg.msg_iter);
5140 ret = sock_recvmsg(sock, &msg, flags);
5141 if (force_nonblock && ret == -EAGAIN)
5143 if (ret == -ERESTARTSYS)
5146 if (req->flags & REQ_F_BUFFER_SELECTED)
5147 cflags = io_put_recv_kbuf(req);
5148 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5150 __io_req_complete(req, issue_flags, ret, cflags);
5154 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5156 struct io_accept *accept = &req->accept;
5158 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5160 if (sqe->ioprio || sqe->len || sqe->buf_index)
5163 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5164 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5165 accept->flags = READ_ONCE(sqe->accept_flags);
5166 accept->nofile = rlimit(RLIMIT_NOFILE);
5168 accept->file_slot = READ_ONCE(sqe->file_index);
5169 if (accept->file_slot && (accept->flags & SOCK_CLOEXEC))
5171 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5173 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5174 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5178 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5180 struct io_accept *accept = &req->accept;
5181 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5182 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5183 bool fixed = !!accept->file_slot;
5187 if (req->file->f_flags & O_NONBLOCK)
5188 req->flags |= REQ_F_NOWAIT;
5191 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5192 if (unlikely(fd < 0))
5195 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5200 ret = PTR_ERR(file);
5201 if (ret == -EAGAIN && force_nonblock)
5203 if (ret == -ERESTARTSYS)
5206 } else if (!fixed) {
5207 fd_install(fd, file);
5210 ret = io_install_fixed_file(req, file, issue_flags,
5211 accept->file_slot - 1);
5213 __io_req_complete(req, issue_flags, ret, 0);
5217 static int io_connect_prep_async(struct io_kiocb *req)
5219 struct io_async_connect *io = req->async_data;
5220 struct io_connect *conn = &req->connect;
5222 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5225 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5227 struct io_connect *conn = &req->connect;
5229 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5231 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5235 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5236 conn->addr_len = READ_ONCE(sqe->addr2);
5240 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5242 struct io_async_connect __io, *io;
5243 unsigned file_flags;
5245 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5247 if (req->async_data) {
5248 io = req->async_data;
5250 ret = move_addr_to_kernel(req->connect.addr,
5251 req->connect.addr_len,
5258 file_flags = force_nonblock ? O_NONBLOCK : 0;
5260 ret = __sys_connect_file(req->file, &io->address,
5261 req->connect.addr_len, file_flags);
5262 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5263 if (req->async_data)
5265 if (io_alloc_async_data(req)) {
5269 memcpy(req->async_data, &__io, sizeof(__io));
5272 if (ret == -ERESTARTSYS)
5277 __io_req_complete(req, issue_flags, ret, 0);
5280 #else /* !CONFIG_NET */
5281 #define IO_NETOP_FN(op) \
5282 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5284 return -EOPNOTSUPP; \
5287 #define IO_NETOP_PREP(op) \
5289 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5291 return -EOPNOTSUPP; \
5294 #define IO_NETOP_PREP_ASYNC(op) \
5296 static int io_##op##_prep_async(struct io_kiocb *req) \
5298 return -EOPNOTSUPP; \
5301 IO_NETOP_PREP_ASYNC(sendmsg);
5302 IO_NETOP_PREP_ASYNC(recvmsg);
5303 IO_NETOP_PREP_ASYNC(connect);
5304 IO_NETOP_PREP(accept);
5307 #endif /* CONFIG_NET */
5309 struct io_poll_table {
5310 struct poll_table_struct pt;
5311 struct io_kiocb *req;
5316 #define IO_POLL_CANCEL_FLAG BIT(31)
5317 #define IO_POLL_REF_MASK GENMASK(30, 0)
5320 * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can
5321 * bump it and acquire ownership. It's disallowed to modify requests while not
5322 * owning it, that prevents from races for enqueueing task_work's and b/w
5323 * arming poll and wakeups.
5325 static inline bool io_poll_get_ownership(struct io_kiocb *req)
5327 return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK);
5330 static void io_poll_mark_cancelled(struct io_kiocb *req)
5332 atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs);
5335 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5337 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5338 if (req->opcode == IORING_OP_POLL_ADD)
5339 return req->async_data;
5340 return req->apoll->double_poll;
5343 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5345 if (req->opcode == IORING_OP_POLL_ADD)
5347 return &req->apoll->poll;
5350 static void io_poll_req_insert(struct io_kiocb *req)
5352 struct io_ring_ctx *ctx = req->ctx;
5353 struct hlist_head *list;
5355 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5356 hlist_add_head(&req->hash_node, list);
5359 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5360 wait_queue_func_t wake_func)
5363 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5364 /* mask in events that we always want/need */
5365 poll->events = events | IO_POLL_UNMASK;
5366 INIT_LIST_HEAD(&poll->wait.entry);
5367 init_waitqueue_func_entry(&poll->wait, wake_func);
5370 static inline void io_poll_remove_entry(struct io_poll_iocb *poll)
5372 struct wait_queue_head *head = smp_load_acquire(&poll->head);
5375 spin_lock_irq(&head->lock);
5376 list_del_init(&poll->wait.entry);
5378 spin_unlock_irq(&head->lock);
5382 static void io_poll_remove_entries(struct io_kiocb *req)
5384 struct io_poll_iocb *poll = io_poll_get_single(req);
5385 struct io_poll_iocb *poll_double = io_poll_get_double(req);
5388 * While we hold the waitqueue lock and the waitqueue is nonempty,
5389 * wake_up_pollfree() will wait for us. However, taking the waitqueue
5390 * lock in the first place can race with the waitqueue being freed.
5392 * We solve this as eventpoll does: by taking advantage of the fact that
5393 * all users of wake_up_pollfree() will RCU-delay the actual free. If
5394 * we enter rcu_read_lock() and see that the pointer to the queue is
5395 * non-NULL, we can then lock it without the memory being freed out from
5398 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
5399 * case the caller deletes the entry from the queue, leaving it empty.
5400 * In that case, only RCU prevents the queue memory from being freed.
5403 io_poll_remove_entry(poll);
5405 io_poll_remove_entry(poll_double);
5410 * All poll tw should go through this. Checks for poll events, manages
5411 * references, does rewait, etc.
5413 * Returns a negative error on failure. >0 when no action require, which is
5414 * either spurious wakeup or multishot CQE is served. 0 when it's done with
5415 * the request, then the mask is stored in req->result.
5417 static int io_poll_check_events(struct io_kiocb *req)
5419 struct io_ring_ctx *ctx = req->ctx;
5420 struct io_poll_iocb *poll = io_poll_get_single(req);
5423 /* req->task == current here, checking PF_EXITING is safe */
5424 if (unlikely(req->task->flags & PF_EXITING))
5425 io_poll_mark_cancelled(req);
5428 v = atomic_read(&req->poll_refs);
5430 /* tw handler should be the owner, and so have some references */
5431 if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK)))
5433 if (v & IO_POLL_CANCEL_FLAG)
5437 struct poll_table_struct pt = { ._key = poll->events };
5439 req->result = vfs_poll(req->file, &pt) & poll->events;
5442 /* multishot, just fill an CQE and proceed */
5443 if (req->result && !(poll->events & EPOLLONESHOT)) {
5444 __poll_t mask = mangle_poll(req->result & poll->events);
5447 spin_lock(&ctx->completion_lock);
5448 filled = io_fill_cqe_aux(ctx, req->user_data, mask,
5450 io_commit_cqring(ctx);
5451 spin_unlock(&ctx->completion_lock);
5452 if (unlikely(!filled))
5454 io_cqring_ev_posted(ctx);
5455 } else if (req->result) {
5460 * Release all references, retry if someone tried to restart
5461 * task_work while we were executing it.
5463 } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs));
5468 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5470 struct io_ring_ctx *ctx = req->ctx;
5473 ret = io_poll_check_events(req);
5478 req->result = mangle_poll(req->result & req->poll.events);
5484 io_poll_remove_entries(req);
5485 spin_lock(&ctx->completion_lock);
5486 hash_del(&req->hash_node);
5487 spin_unlock(&ctx->completion_lock);
5488 io_req_complete_post(req, req->result, 0);
5491 static void io_apoll_task_func(struct io_kiocb *req, bool *locked)
5493 struct io_ring_ctx *ctx = req->ctx;
5496 ret = io_poll_check_events(req);
5500 io_poll_remove_entries(req);
5501 spin_lock(&ctx->completion_lock);
5502 hash_del(&req->hash_node);
5503 spin_unlock(&ctx->completion_lock);
5506 io_req_task_submit(req, locked);
5508 io_req_complete_failed(req, ret);
5511 static void __io_poll_execute(struct io_kiocb *req, int mask)
5514 if (req->opcode == IORING_OP_POLL_ADD)
5515 req->io_task_work.func = io_poll_task_func;
5517 req->io_task_work.func = io_apoll_task_func;
5519 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5520 io_req_task_work_add(req);
5523 static inline void io_poll_execute(struct io_kiocb *req, int res)
5525 if (io_poll_get_ownership(req))
5526 __io_poll_execute(req, res);
5529 static void io_poll_cancel_req(struct io_kiocb *req)
5531 io_poll_mark_cancelled(req);
5532 /* kick tw, which should complete the request */
5533 io_poll_execute(req, 0);
5536 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5539 struct io_kiocb *req = wait->private;
5540 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
5542 __poll_t mask = key_to_poll(key);
5544 if (unlikely(mask & POLLFREE)) {
5545 io_poll_mark_cancelled(req);
5546 /* we have to kick tw in case it's not already */
5547 io_poll_execute(req, 0);
5550 * If the waitqueue is being freed early but someone is already
5551 * holds ownership over it, we have to tear down the request as
5552 * best we can. That means immediately removing the request from
5553 * its waitqueue and preventing all further accesses to the
5554 * waitqueue via the request.
5556 list_del_init(&poll->wait.entry);
5559 * Careful: this *must* be the last step, since as soon
5560 * as req->head is NULL'ed out, the request can be
5561 * completed and freed, since aio_poll_complete_work()
5562 * will no longer need to take the waitqueue lock.
5564 smp_store_release(&poll->head, NULL);
5568 /* for instances that support it check for an event match first */
5569 if (mask && !(mask & poll->events))
5572 if (io_poll_get_ownership(req))
5573 __io_poll_execute(req, mask);
5577 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5578 struct wait_queue_head *head,
5579 struct io_poll_iocb **poll_ptr)
5581 struct io_kiocb *req = pt->req;
5584 * The file being polled uses multiple waitqueues for poll handling
5585 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5588 if (unlikely(pt->nr_entries)) {
5589 struct io_poll_iocb *first = poll;
5591 /* double add on the same waitqueue head, ignore */
5592 if (first->head == head)
5594 /* already have a 2nd entry, fail a third attempt */
5596 if ((*poll_ptr)->head == head)
5598 pt->error = -EINVAL;
5602 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5604 pt->error = -ENOMEM;
5607 io_init_poll_iocb(poll, first->events, first->wait.func);
5613 poll->wait.private = req;
5615 if (poll->events & EPOLLEXCLUSIVE)
5616 add_wait_queue_exclusive(head, &poll->wait);
5618 add_wait_queue(head, &poll->wait);
5621 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5622 struct poll_table_struct *p)
5624 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5626 __io_queue_proc(&pt->req->poll, pt, head,
5627 (struct io_poll_iocb **) &pt->req->async_data);
5630 static int __io_arm_poll_handler(struct io_kiocb *req,
5631 struct io_poll_iocb *poll,
5632 struct io_poll_table *ipt, __poll_t mask)
5634 struct io_ring_ctx *ctx = req->ctx;
5637 INIT_HLIST_NODE(&req->hash_node);
5638 io_init_poll_iocb(poll, mask, io_poll_wake);
5639 poll->file = req->file;
5640 poll->wait.private = req;
5642 ipt->pt._key = mask;
5645 ipt->nr_entries = 0;
5648 * Take the ownership to delay any tw execution up until we're done
5649 * with poll arming. see io_poll_get_ownership().
5651 atomic_set(&req->poll_refs, 1);
5652 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5654 if (mask && (poll->events & EPOLLONESHOT)) {
5655 io_poll_remove_entries(req);
5656 /* no one else has access to the req, forget about the ref */
5659 if (!mask && unlikely(ipt->error || !ipt->nr_entries)) {
5660 io_poll_remove_entries(req);
5662 ipt->error = -EINVAL;
5666 spin_lock(&ctx->completion_lock);
5667 io_poll_req_insert(req);
5668 spin_unlock(&ctx->completion_lock);
5671 /* can't multishot if failed, just queue the event we've got */
5672 if (unlikely(ipt->error || !ipt->nr_entries)) {
5673 poll->events |= EPOLLONESHOT;
5676 __io_poll_execute(req, mask);
5681 * Release ownership. If someone tried to queue a tw while it was
5682 * locked, kick it off for them.
5684 v = atomic_dec_return(&req->poll_refs);
5685 if (unlikely(v & IO_POLL_REF_MASK))
5686 __io_poll_execute(req, 0);
5690 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5691 struct poll_table_struct *p)
5693 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5694 struct async_poll *apoll = pt->req->apoll;
5696 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5705 static int io_arm_poll_handler(struct io_kiocb *req)
5707 const struct io_op_def *def = &io_op_defs[req->opcode];
5708 struct io_ring_ctx *ctx = req->ctx;
5709 struct async_poll *apoll;
5710 struct io_poll_table ipt;
5711 __poll_t mask = EPOLLONESHOT | POLLERR | POLLPRI;
5714 if (!req->file || !file_can_poll(req->file))
5715 return IO_APOLL_ABORTED;
5716 if (req->flags & REQ_F_POLLED)
5717 return IO_APOLL_ABORTED;
5718 if (!def->pollin && !def->pollout)
5719 return IO_APOLL_ABORTED;
5722 mask |= POLLIN | POLLRDNORM;
5724 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5725 if ((req->opcode == IORING_OP_RECVMSG) &&
5726 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5729 mask |= POLLOUT | POLLWRNORM;
5732 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5733 if (unlikely(!apoll))
5734 return IO_APOLL_ABORTED;
5735 apoll->double_poll = NULL;
5737 req->flags |= REQ_F_POLLED;
5738 ipt.pt._qproc = io_async_queue_proc;
5740 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask);
5741 if (ret || ipt.error)
5742 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5744 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5745 mask, apoll->poll.events);
5750 * Returns true if we found and killed one or more poll requests
5752 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5755 struct hlist_node *tmp;
5756 struct io_kiocb *req;
5760 spin_lock(&ctx->completion_lock);
5761 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5762 struct hlist_head *list;
5764 list = &ctx->cancel_hash[i];
5765 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5766 if (io_match_task_safe(req, tsk, cancel_all)) {
5767 hlist_del_init(&req->hash_node);
5768 io_poll_cancel_req(req);
5773 spin_unlock(&ctx->completion_lock);
5777 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5779 __must_hold(&ctx->completion_lock)
5781 struct hlist_head *list;
5782 struct io_kiocb *req;
5784 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5785 hlist_for_each_entry(req, list, hash_node) {
5786 if (sqe_addr != req->user_data)
5788 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5795 static bool io_poll_disarm(struct io_kiocb *req)
5796 __must_hold(&ctx->completion_lock)
5798 if (!io_poll_get_ownership(req))
5800 io_poll_remove_entries(req);
5801 hash_del(&req->hash_node);
5805 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5807 __must_hold(&ctx->completion_lock)
5809 struct io_kiocb *req = io_poll_find(ctx, sqe_addr, poll_only);
5813 io_poll_cancel_req(req);
5817 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5822 events = READ_ONCE(sqe->poll32_events);
5824 events = swahw32(events);
5826 if (!(flags & IORING_POLL_ADD_MULTI))
5827 events |= EPOLLONESHOT;
5828 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5831 static int io_poll_update_prep(struct io_kiocb *req,
5832 const struct io_uring_sqe *sqe)
5834 struct io_poll_update *upd = &req->poll_update;
5837 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5839 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5841 flags = READ_ONCE(sqe->len);
5842 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5843 IORING_POLL_ADD_MULTI))
5845 /* meaningless without update */
5846 if (flags == IORING_POLL_ADD_MULTI)
5849 upd->old_user_data = READ_ONCE(sqe->addr);
5850 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5851 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5853 upd->new_user_data = READ_ONCE(sqe->off);
5854 if (!upd->update_user_data && upd->new_user_data)
5856 if (upd->update_events)
5857 upd->events = io_poll_parse_events(sqe, flags);
5858 else if (sqe->poll32_events)
5864 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5866 struct io_poll_iocb *poll = &req->poll;
5869 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5871 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5873 flags = READ_ONCE(sqe->len);
5874 if (flags & ~IORING_POLL_ADD_MULTI)
5877 io_req_set_refcount(req);
5878 poll->events = io_poll_parse_events(sqe, flags);
5882 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5884 struct io_poll_iocb *poll = &req->poll;
5885 struct io_poll_table ipt;
5888 ipt.pt._qproc = io_poll_queue_proc;
5890 ret = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events);
5891 if (!ret && ipt.error)
5893 ret = ret ?: ipt.error;
5895 __io_req_complete(req, issue_flags, ret, 0);
5899 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5901 struct io_ring_ctx *ctx = req->ctx;
5902 struct io_kiocb *preq;
5905 spin_lock(&ctx->completion_lock);
5906 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5907 if (!preq || !io_poll_disarm(preq)) {
5908 spin_unlock(&ctx->completion_lock);
5909 ret = preq ? -EALREADY : -ENOENT;
5912 spin_unlock(&ctx->completion_lock);
5914 if (req->poll_update.update_events || req->poll_update.update_user_data) {
5915 /* only mask one event flags, keep behavior flags */
5916 if (req->poll_update.update_events) {
5917 preq->poll.events &= ~0xffff;
5918 preq->poll.events |= req->poll_update.events & 0xffff;
5919 preq->poll.events |= IO_POLL_UNMASK;
5921 if (req->poll_update.update_user_data)
5922 preq->user_data = req->poll_update.new_user_data;
5924 ret2 = io_poll_add(preq, issue_flags);
5925 /* successfully updated, don't complete poll request */
5930 io_req_complete(preq, -ECANCELED);
5934 /* complete update request, we're done with it */
5935 io_req_complete(req, ret);
5939 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5942 io_req_complete_post(req, -ETIME, 0);
5945 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5947 struct io_timeout_data *data = container_of(timer,
5948 struct io_timeout_data, timer);
5949 struct io_kiocb *req = data->req;
5950 struct io_ring_ctx *ctx = req->ctx;
5951 unsigned long flags;
5953 spin_lock_irqsave(&ctx->timeout_lock, flags);
5954 list_del_init(&req->timeout.list);
5955 atomic_set(&req->ctx->cq_timeouts,
5956 atomic_read(&req->ctx->cq_timeouts) + 1);
5957 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5959 req->io_task_work.func = io_req_task_timeout;
5960 io_req_task_work_add(req);
5961 return HRTIMER_NORESTART;
5964 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5966 __must_hold(&ctx->timeout_lock)
5968 struct io_timeout_data *io;
5969 struct io_kiocb *req;
5972 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5973 found = user_data == req->user_data;
5978 return ERR_PTR(-ENOENT);
5980 io = req->async_data;
5981 if (hrtimer_try_to_cancel(&io->timer) == -1)
5982 return ERR_PTR(-EALREADY);
5983 list_del_init(&req->timeout.list);
5987 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5988 __must_hold(&ctx->completion_lock)
5989 __must_hold(&ctx->timeout_lock)
5991 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5994 return PTR_ERR(req);
5997 io_fill_cqe_req(req, -ECANCELED, 0);
5998 io_put_req_deferred(req);
6002 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
6004 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
6005 case IORING_TIMEOUT_BOOTTIME:
6006 return CLOCK_BOOTTIME;
6007 case IORING_TIMEOUT_REALTIME:
6008 return CLOCK_REALTIME;
6010 /* can't happen, vetted at prep time */
6014 return CLOCK_MONOTONIC;
6018 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6019 struct timespec64 *ts, enum hrtimer_mode mode)
6020 __must_hold(&ctx->timeout_lock)
6022 struct io_timeout_data *io;
6023 struct io_kiocb *req;
6026 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6027 found = user_data == req->user_data;
6034 io = req->async_data;
6035 if (hrtimer_try_to_cancel(&io->timer) == -1)
6037 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6038 io->timer.function = io_link_timeout_fn;
6039 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6043 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6044 struct timespec64 *ts, enum hrtimer_mode mode)
6045 __must_hold(&ctx->timeout_lock)
6047 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6048 struct io_timeout_data *data;
6051 return PTR_ERR(req);
6053 req->timeout.off = 0; /* noseq */
6054 data = req->async_data;
6055 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6056 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6057 data->timer.function = io_timeout_fn;
6058 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6062 static int io_timeout_remove_prep(struct io_kiocb *req,
6063 const struct io_uring_sqe *sqe)
6065 struct io_timeout_rem *tr = &req->timeout_rem;
6067 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6069 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6071 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6074 tr->ltimeout = false;
6075 tr->addr = READ_ONCE(sqe->addr);
6076 tr->flags = READ_ONCE(sqe->timeout_flags);
6077 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6078 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6080 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6081 tr->ltimeout = true;
6082 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6084 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6086 } else if (tr->flags) {
6087 /* timeout removal doesn't support flags */
6094 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6096 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6101 * Remove or update an existing timeout command
6103 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6105 struct io_timeout_rem *tr = &req->timeout_rem;
6106 struct io_ring_ctx *ctx = req->ctx;
6109 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6110 spin_lock(&ctx->completion_lock);
6111 spin_lock_irq(&ctx->timeout_lock);
6112 ret = io_timeout_cancel(ctx, tr->addr);
6113 spin_unlock_irq(&ctx->timeout_lock);
6114 spin_unlock(&ctx->completion_lock);
6116 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6118 spin_lock_irq(&ctx->timeout_lock);
6120 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6122 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6123 spin_unlock_irq(&ctx->timeout_lock);
6128 io_req_complete_post(req, ret, 0);
6132 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6133 bool is_timeout_link)
6135 struct io_timeout_data *data;
6137 u32 off = READ_ONCE(sqe->off);
6139 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6141 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6144 if (off && is_timeout_link)
6146 flags = READ_ONCE(sqe->timeout_flags);
6147 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6149 /* more than one clock specified is invalid, obviously */
6150 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6153 INIT_LIST_HEAD(&req->timeout.list);
6154 req->timeout.off = off;
6155 if (unlikely(off && !req->ctx->off_timeout_used))
6156 req->ctx->off_timeout_used = true;
6158 if (!req->async_data && io_alloc_async_data(req))
6161 data = req->async_data;
6163 data->flags = flags;
6165 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6168 INIT_LIST_HEAD(&req->timeout.list);
6169 data->mode = io_translate_timeout_mode(flags);
6170 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6172 if (is_timeout_link) {
6173 struct io_submit_link *link = &req->ctx->submit_state.link;
6177 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6179 req->timeout.head = link->last;
6180 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6185 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6187 struct io_ring_ctx *ctx = req->ctx;
6188 struct io_timeout_data *data = req->async_data;
6189 struct list_head *entry;
6190 u32 tail, off = req->timeout.off;
6192 spin_lock_irq(&ctx->timeout_lock);
6195 * sqe->off holds how many events that need to occur for this
6196 * timeout event to be satisfied. If it isn't set, then this is
6197 * a pure timeout request, sequence isn't used.
6199 if (io_is_timeout_noseq(req)) {
6200 entry = ctx->timeout_list.prev;
6204 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6205 req->timeout.target_seq = tail + off;
6207 /* Update the last seq here in case io_flush_timeouts() hasn't.
6208 * This is safe because ->completion_lock is held, and submissions
6209 * and completions are never mixed in the same ->completion_lock section.
6211 ctx->cq_last_tm_flush = tail;
6214 * Insertion sort, ensuring the first entry in the list is always
6215 * the one we need first.
6217 list_for_each_prev(entry, &ctx->timeout_list) {
6218 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6221 if (io_is_timeout_noseq(nxt))
6223 /* nxt.seq is behind @tail, otherwise would've been completed */
6224 if (off >= nxt->timeout.target_seq - tail)
6228 list_add(&req->timeout.list, entry);
6229 data->timer.function = io_timeout_fn;
6230 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6231 spin_unlock_irq(&ctx->timeout_lock);
6235 struct io_cancel_data {
6236 struct io_ring_ctx *ctx;
6240 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6242 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6243 struct io_cancel_data *cd = data;
6245 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6248 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6249 struct io_ring_ctx *ctx)
6251 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6252 enum io_wq_cancel cancel_ret;
6255 if (!tctx || !tctx->io_wq)
6258 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6259 switch (cancel_ret) {
6260 case IO_WQ_CANCEL_OK:
6263 case IO_WQ_CANCEL_RUNNING:
6266 case IO_WQ_CANCEL_NOTFOUND:
6274 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6276 struct io_ring_ctx *ctx = req->ctx;
6279 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6281 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6285 spin_lock(&ctx->completion_lock);
6286 spin_lock_irq(&ctx->timeout_lock);
6287 ret = io_timeout_cancel(ctx, sqe_addr);
6288 spin_unlock_irq(&ctx->timeout_lock);
6291 ret = io_poll_cancel(ctx, sqe_addr, false);
6293 spin_unlock(&ctx->completion_lock);
6297 static int io_async_cancel_prep(struct io_kiocb *req,
6298 const struct io_uring_sqe *sqe)
6300 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6302 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6304 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6308 req->cancel.addr = READ_ONCE(sqe->addr);
6312 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6314 struct io_ring_ctx *ctx = req->ctx;
6315 u64 sqe_addr = req->cancel.addr;
6316 struct io_tctx_node *node;
6319 ret = io_try_cancel_userdata(req, sqe_addr);
6323 /* slow path, try all io-wq's */
6324 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6326 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6327 struct io_uring_task *tctx = node->task->io_uring;
6329 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6333 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6337 io_req_complete_post(req, ret, 0);
6341 static int io_rsrc_update_prep(struct io_kiocb *req,
6342 const struct io_uring_sqe *sqe)
6344 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6346 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6349 req->rsrc_update.offset = READ_ONCE(sqe->off);
6350 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6351 if (!req->rsrc_update.nr_args)
6353 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6357 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6359 struct io_ring_ctx *ctx = req->ctx;
6360 struct io_uring_rsrc_update2 up;
6363 up.offset = req->rsrc_update.offset;
6364 up.data = req->rsrc_update.arg;
6370 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6371 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6372 &up, req->rsrc_update.nr_args);
6373 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6377 __io_req_complete(req, issue_flags, ret, 0);
6381 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6383 switch (req->opcode) {
6386 case IORING_OP_READV:
6387 case IORING_OP_READ_FIXED:
6388 case IORING_OP_READ:
6389 return io_read_prep(req, sqe);
6390 case IORING_OP_WRITEV:
6391 case IORING_OP_WRITE_FIXED:
6392 case IORING_OP_WRITE:
6393 return io_write_prep(req, sqe);
6394 case IORING_OP_POLL_ADD:
6395 return io_poll_add_prep(req, sqe);
6396 case IORING_OP_POLL_REMOVE:
6397 return io_poll_update_prep(req, sqe);
6398 case IORING_OP_FSYNC:
6399 return io_fsync_prep(req, sqe);
6400 case IORING_OP_SYNC_FILE_RANGE:
6401 return io_sfr_prep(req, sqe);
6402 case IORING_OP_SENDMSG:
6403 case IORING_OP_SEND:
6404 return io_sendmsg_prep(req, sqe);
6405 case IORING_OP_RECVMSG:
6406 case IORING_OP_RECV:
6407 return io_recvmsg_prep(req, sqe);
6408 case IORING_OP_CONNECT:
6409 return io_connect_prep(req, sqe);
6410 case IORING_OP_TIMEOUT:
6411 return io_timeout_prep(req, sqe, false);
6412 case IORING_OP_TIMEOUT_REMOVE:
6413 return io_timeout_remove_prep(req, sqe);
6414 case IORING_OP_ASYNC_CANCEL:
6415 return io_async_cancel_prep(req, sqe);
6416 case IORING_OP_LINK_TIMEOUT:
6417 return io_timeout_prep(req, sqe, true);
6418 case IORING_OP_ACCEPT:
6419 return io_accept_prep(req, sqe);
6420 case IORING_OP_FALLOCATE:
6421 return io_fallocate_prep(req, sqe);
6422 case IORING_OP_OPENAT:
6423 return io_openat_prep(req, sqe);
6424 case IORING_OP_CLOSE:
6425 return io_close_prep(req, sqe);
6426 case IORING_OP_FILES_UPDATE:
6427 return io_rsrc_update_prep(req, sqe);
6428 case IORING_OP_STATX:
6429 return io_statx_prep(req, sqe);
6430 case IORING_OP_FADVISE:
6431 return io_fadvise_prep(req, sqe);
6432 case IORING_OP_MADVISE:
6433 return io_madvise_prep(req, sqe);
6434 case IORING_OP_OPENAT2:
6435 return io_openat2_prep(req, sqe);
6436 case IORING_OP_EPOLL_CTL:
6437 return io_epoll_ctl_prep(req, sqe);
6438 case IORING_OP_SPLICE:
6439 return io_splice_prep(req, sqe);
6440 case IORING_OP_PROVIDE_BUFFERS:
6441 return io_provide_buffers_prep(req, sqe);
6442 case IORING_OP_REMOVE_BUFFERS:
6443 return io_remove_buffers_prep(req, sqe);
6445 return io_tee_prep(req, sqe);
6446 case IORING_OP_SHUTDOWN:
6447 return io_shutdown_prep(req, sqe);
6448 case IORING_OP_RENAMEAT:
6449 return io_renameat_prep(req, sqe);
6450 case IORING_OP_UNLINKAT:
6451 return io_unlinkat_prep(req, sqe);
6452 case IORING_OP_MKDIRAT:
6453 return io_mkdirat_prep(req, sqe);
6454 case IORING_OP_SYMLINKAT:
6455 return io_symlinkat_prep(req, sqe);
6456 case IORING_OP_LINKAT:
6457 return io_linkat_prep(req, sqe);
6460 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6465 static int io_req_prep_async(struct io_kiocb *req)
6467 if (!io_op_defs[req->opcode].needs_async_setup)
6469 if (WARN_ON_ONCE(req->async_data))
6471 if (io_alloc_async_data(req))
6474 switch (req->opcode) {
6475 case IORING_OP_READV:
6476 return io_rw_prep_async(req, READ);
6477 case IORING_OP_WRITEV:
6478 return io_rw_prep_async(req, WRITE);
6479 case IORING_OP_SENDMSG:
6480 return io_sendmsg_prep_async(req);
6481 case IORING_OP_RECVMSG:
6482 return io_recvmsg_prep_async(req);
6483 case IORING_OP_CONNECT:
6484 return io_connect_prep_async(req);
6486 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6491 static u32 io_get_sequence(struct io_kiocb *req)
6493 u32 seq = req->ctx->cached_sq_head;
6495 /* need original cached_sq_head, but it was increased for each req */
6496 io_for_each_link(req, req)
6501 static bool io_drain_req(struct io_kiocb *req)
6503 struct io_kiocb *pos;
6504 struct io_ring_ctx *ctx = req->ctx;
6505 struct io_defer_entry *de;
6509 if (req->flags & REQ_F_FAIL) {
6510 io_req_complete_fail_submit(req);
6515 * If we need to drain a request in the middle of a link, drain the
6516 * head request and the next request/link after the current link.
6517 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6518 * maintained for every request of our link.
6520 if (ctx->drain_next) {
6521 req->flags |= REQ_F_IO_DRAIN;
6522 ctx->drain_next = false;
6524 /* not interested in head, start from the first linked */
6525 io_for_each_link(pos, req->link) {
6526 if (pos->flags & REQ_F_IO_DRAIN) {
6527 ctx->drain_next = true;
6528 req->flags |= REQ_F_IO_DRAIN;
6533 /* Still need defer if there is pending req in defer list. */
6534 spin_lock(&ctx->completion_lock);
6535 if (likely(list_empty_careful(&ctx->defer_list) &&
6536 !(req->flags & REQ_F_IO_DRAIN))) {
6537 spin_unlock(&ctx->completion_lock);
6538 ctx->drain_active = false;
6541 spin_unlock(&ctx->completion_lock);
6543 seq = io_get_sequence(req);
6544 /* Still a chance to pass the sequence check */
6545 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6548 ret = io_req_prep_async(req);
6551 io_prep_async_link(req);
6552 de = kmalloc(sizeof(*de), GFP_KERNEL);
6556 io_req_complete_failed(req, ret);
6560 spin_lock(&ctx->completion_lock);
6561 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6562 spin_unlock(&ctx->completion_lock);
6564 io_queue_async_work(req, NULL);
6568 trace_io_uring_defer(ctx, req, req->user_data);
6571 list_add_tail(&de->list, &ctx->defer_list);
6572 spin_unlock(&ctx->completion_lock);
6576 static void io_clean_op(struct io_kiocb *req)
6578 if (req->flags & REQ_F_BUFFER_SELECTED) {
6579 switch (req->opcode) {
6580 case IORING_OP_READV:
6581 case IORING_OP_READ_FIXED:
6582 case IORING_OP_READ:
6583 kfree((void *)(unsigned long)req->rw.addr);
6585 case IORING_OP_RECVMSG:
6586 case IORING_OP_RECV:
6587 kfree(req->sr_msg.kbuf);
6592 if (req->flags & REQ_F_NEED_CLEANUP) {
6593 switch (req->opcode) {
6594 case IORING_OP_READV:
6595 case IORING_OP_READ_FIXED:
6596 case IORING_OP_READ:
6597 case IORING_OP_WRITEV:
6598 case IORING_OP_WRITE_FIXED:
6599 case IORING_OP_WRITE: {
6600 struct io_async_rw *io = req->async_data;
6602 kfree(io->free_iovec);
6605 case IORING_OP_RECVMSG:
6606 case IORING_OP_SENDMSG: {
6607 struct io_async_msghdr *io = req->async_data;
6609 kfree(io->free_iov);
6612 case IORING_OP_OPENAT:
6613 case IORING_OP_OPENAT2:
6614 if (req->open.filename)
6615 putname(req->open.filename);
6617 case IORING_OP_RENAMEAT:
6618 putname(req->rename.oldpath);
6619 putname(req->rename.newpath);
6621 case IORING_OP_UNLINKAT:
6622 putname(req->unlink.filename);
6624 case IORING_OP_MKDIRAT:
6625 putname(req->mkdir.filename);
6627 case IORING_OP_SYMLINKAT:
6628 putname(req->symlink.oldpath);
6629 putname(req->symlink.newpath);
6631 case IORING_OP_LINKAT:
6632 putname(req->hardlink.oldpath);
6633 putname(req->hardlink.newpath);
6637 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6638 kfree(req->apoll->double_poll);
6642 if (req->flags & REQ_F_INFLIGHT) {
6643 struct io_uring_task *tctx = req->task->io_uring;
6645 atomic_dec(&tctx->inflight_tracked);
6647 if (req->flags & REQ_F_CREDS)
6648 put_cred(req->creds);
6650 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6653 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6655 struct io_ring_ctx *ctx = req->ctx;
6656 const struct cred *creds = NULL;
6659 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6660 creds = override_creds(req->creds);
6662 switch (req->opcode) {
6664 ret = io_nop(req, issue_flags);
6666 case IORING_OP_READV:
6667 case IORING_OP_READ_FIXED:
6668 case IORING_OP_READ:
6669 ret = io_read(req, issue_flags);
6671 case IORING_OP_WRITEV:
6672 case IORING_OP_WRITE_FIXED:
6673 case IORING_OP_WRITE:
6674 ret = io_write(req, issue_flags);
6676 case IORING_OP_FSYNC:
6677 ret = io_fsync(req, issue_flags);
6679 case IORING_OP_POLL_ADD:
6680 ret = io_poll_add(req, issue_flags);
6682 case IORING_OP_POLL_REMOVE:
6683 ret = io_poll_update(req, issue_flags);
6685 case IORING_OP_SYNC_FILE_RANGE:
6686 ret = io_sync_file_range(req, issue_flags);
6688 case IORING_OP_SENDMSG:
6689 ret = io_sendmsg(req, issue_flags);
6691 case IORING_OP_SEND:
6692 ret = io_send(req, issue_flags);
6694 case IORING_OP_RECVMSG:
6695 ret = io_recvmsg(req, issue_flags);
6697 case IORING_OP_RECV:
6698 ret = io_recv(req, issue_flags);
6700 case IORING_OP_TIMEOUT:
6701 ret = io_timeout(req, issue_flags);
6703 case IORING_OP_TIMEOUT_REMOVE:
6704 ret = io_timeout_remove(req, issue_flags);
6706 case IORING_OP_ACCEPT:
6707 ret = io_accept(req, issue_flags);
6709 case IORING_OP_CONNECT:
6710 ret = io_connect(req, issue_flags);
6712 case IORING_OP_ASYNC_CANCEL:
6713 ret = io_async_cancel(req, issue_flags);
6715 case IORING_OP_FALLOCATE:
6716 ret = io_fallocate(req, issue_flags);
6718 case IORING_OP_OPENAT:
6719 ret = io_openat(req, issue_flags);
6721 case IORING_OP_CLOSE:
6722 ret = io_close(req, issue_flags);
6724 case IORING_OP_FILES_UPDATE:
6725 ret = io_files_update(req, issue_flags);
6727 case IORING_OP_STATX:
6728 ret = io_statx(req, issue_flags);
6730 case IORING_OP_FADVISE:
6731 ret = io_fadvise(req, issue_flags);
6733 case IORING_OP_MADVISE:
6734 ret = io_madvise(req, issue_flags);
6736 case IORING_OP_OPENAT2:
6737 ret = io_openat2(req, issue_flags);
6739 case IORING_OP_EPOLL_CTL:
6740 ret = io_epoll_ctl(req, issue_flags);
6742 case IORING_OP_SPLICE:
6743 ret = io_splice(req, issue_flags);
6745 case IORING_OP_PROVIDE_BUFFERS:
6746 ret = io_provide_buffers(req, issue_flags);
6748 case IORING_OP_REMOVE_BUFFERS:
6749 ret = io_remove_buffers(req, issue_flags);
6752 ret = io_tee(req, issue_flags);
6754 case IORING_OP_SHUTDOWN:
6755 ret = io_shutdown(req, issue_flags);
6757 case IORING_OP_RENAMEAT:
6758 ret = io_renameat(req, issue_flags);
6760 case IORING_OP_UNLINKAT:
6761 ret = io_unlinkat(req, issue_flags);
6763 case IORING_OP_MKDIRAT:
6764 ret = io_mkdirat(req, issue_flags);
6766 case IORING_OP_SYMLINKAT:
6767 ret = io_symlinkat(req, issue_flags);
6769 case IORING_OP_LINKAT:
6770 ret = io_linkat(req, issue_flags);
6778 revert_creds(creds);
6781 /* If the op doesn't have a file, we're not polling for it */
6782 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6783 io_iopoll_req_issued(req);
6788 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6790 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6792 req = io_put_req_find_next(req);
6793 return req ? &req->work : NULL;
6796 static void io_wq_submit_work(struct io_wq_work *work)
6798 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6799 struct io_kiocb *timeout;
6802 /* one will be dropped by ->io_free_work() after returning to io-wq */
6803 if (!(req->flags & REQ_F_REFCOUNT))
6804 __io_req_set_refcount(req, 2);
6808 timeout = io_prep_linked_timeout(req);
6810 io_queue_linked_timeout(timeout);
6812 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6813 if (work->flags & IO_WQ_WORK_CANCEL)
6818 ret = io_issue_sqe(req, 0);
6820 * We can get EAGAIN for polled IO even though we're
6821 * forcing a sync submission from here, since we can't
6822 * wait for request slots on the block side.
6824 if (ret != -EAGAIN || !(req->ctx->flags & IORING_SETUP_IOPOLL))
6830 /* avoid locking problems by failing it from a clean context */
6832 io_req_task_queue_fail(req, ret);
6835 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6838 return &table->files[i];
6841 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6844 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6846 return (struct file *) (slot->file_ptr & FFS_MASK);
6849 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6851 unsigned long file_ptr = (unsigned long) file;
6853 if (__io_file_supports_nowait(file, READ))
6854 file_ptr |= FFS_ASYNC_READ;
6855 if (__io_file_supports_nowait(file, WRITE))
6856 file_ptr |= FFS_ASYNC_WRITE;
6857 if (S_ISREG(file_inode(file)->i_mode))
6858 file_ptr |= FFS_ISREG;
6859 file_slot->file_ptr = file_ptr;
6862 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6863 struct io_kiocb *req, int fd)
6866 unsigned long file_ptr;
6868 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6870 fd = array_index_nospec(fd, ctx->nr_user_files);
6871 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6872 file = (struct file *) (file_ptr & FFS_MASK);
6873 file_ptr &= ~FFS_MASK;
6874 /* mask in overlapping REQ_F and FFS bits */
6875 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6876 io_req_set_rsrc_node(req);
6880 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6881 struct io_kiocb *req, int fd)
6883 struct file *file = fget(fd);
6885 trace_io_uring_file_get(ctx, fd);
6887 /* we don't allow fixed io_uring files */
6888 if (file && unlikely(file->f_op == &io_uring_fops))
6889 io_req_track_inflight(req);
6893 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6894 struct io_kiocb *req, int fd, bool fixed)
6897 return io_file_get_fixed(ctx, req, fd);
6899 return io_file_get_normal(ctx, req, fd);
6902 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6904 struct io_kiocb *prev = req->timeout.prev;
6908 if (!(req->task->flags & PF_EXITING))
6909 ret = io_try_cancel_userdata(req, prev->user_data);
6910 io_req_complete_post(req, ret ?: -ETIME, 0);
6913 io_req_complete_post(req, -ETIME, 0);
6917 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6919 struct io_timeout_data *data = container_of(timer,
6920 struct io_timeout_data, timer);
6921 struct io_kiocb *prev, *req = data->req;
6922 struct io_ring_ctx *ctx = req->ctx;
6923 unsigned long flags;
6925 spin_lock_irqsave(&ctx->timeout_lock, flags);
6926 prev = req->timeout.head;
6927 req->timeout.head = NULL;
6930 * We don't expect the list to be empty, that will only happen if we
6931 * race with the completion of the linked work.
6934 io_remove_next_linked(prev);
6935 if (!req_ref_inc_not_zero(prev))
6938 list_del(&req->timeout.list);
6939 req->timeout.prev = prev;
6940 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6942 req->io_task_work.func = io_req_task_link_timeout;
6943 io_req_task_work_add(req);
6944 return HRTIMER_NORESTART;
6947 static void io_queue_linked_timeout(struct io_kiocb *req)
6949 struct io_ring_ctx *ctx = req->ctx;
6951 spin_lock_irq(&ctx->timeout_lock);
6953 * If the back reference is NULL, then our linked request finished
6954 * before we got a chance to setup the timer
6956 if (req->timeout.head) {
6957 struct io_timeout_data *data = req->async_data;
6959 data->timer.function = io_link_timeout_fn;
6960 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6962 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6964 spin_unlock_irq(&ctx->timeout_lock);
6965 /* drop submission reference */
6969 static void __io_queue_sqe(struct io_kiocb *req)
6970 __must_hold(&req->ctx->uring_lock)
6972 struct io_kiocb *linked_timeout;
6976 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6979 * We async punt it if the file wasn't marked NOWAIT, or if the file
6980 * doesn't support non-blocking read/write attempts
6983 if (req->flags & REQ_F_COMPLETE_INLINE) {
6984 struct io_ring_ctx *ctx = req->ctx;
6985 struct io_submit_state *state = &ctx->submit_state;
6987 state->compl_reqs[state->compl_nr++] = req;
6988 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6989 io_submit_flush_completions(ctx);
6993 linked_timeout = io_prep_linked_timeout(req);
6995 io_queue_linked_timeout(linked_timeout);
6996 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6997 linked_timeout = io_prep_linked_timeout(req);
6999 switch (io_arm_poll_handler(req)) {
7000 case IO_APOLL_READY:
7002 io_queue_linked_timeout(linked_timeout);
7004 case IO_APOLL_ABORTED:
7006 * Queued up for async execution, worker will release
7007 * submit reference when the iocb is actually submitted.
7009 io_queue_async_work(req, NULL);
7014 io_queue_linked_timeout(linked_timeout);
7016 io_req_complete_failed(req, ret);
7020 static inline void io_queue_sqe(struct io_kiocb *req)
7021 __must_hold(&req->ctx->uring_lock)
7023 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7026 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7027 __io_queue_sqe(req);
7028 } else if (req->flags & REQ_F_FAIL) {
7029 io_req_complete_fail_submit(req);
7031 int ret = io_req_prep_async(req);
7034 io_req_complete_failed(req, ret);
7036 io_queue_async_work(req, NULL);
7041 * Check SQE restrictions (opcode and flags).
7043 * Returns 'true' if SQE is allowed, 'false' otherwise.
7045 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7046 struct io_kiocb *req,
7047 unsigned int sqe_flags)
7049 if (likely(!ctx->restricted))
7052 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7055 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7056 ctx->restrictions.sqe_flags_required)
7059 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7060 ctx->restrictions.sqe_flags_required))
7066 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7067 const struct io_uring_sqe *sqe)
7068 __must_hold(&ctx->uring_lock)
7070 struct io_submit_state *state;
7071 unsigned int sqe_flags;
7072 int personality, ret = 0;
7074 /* req is partially pre-initialised, see io_preinit_req() */
7075 req->opcode = READ_ONCE(sqe->opcode);
7076 /* same numerical values with corresponding REQ_F_*, safe to copy */
7077 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7078 req->user_data = READ_ONCE(sqe->user_data);
7080 req->fixed_rsrc_refs = NULL;
7081 req->task = current;
7083 /* enforce forwards compatibility on users */
7084 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7086 if (unlikely(req->opcode >= IORING_OP_LAST))
7088 if (!io_check_restriction(ctx, req, sqe_flags))
7091 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7092 !io_op_defs[req->opcode].buffer_select)
7094 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7095 ctx->drain_active = true;
7097 personality = READ_ONCE(sqe->personality);
7099 req->creds = xa_load(&ctx->personalities, personality);
7102 get_cred(req->creds);
7103 req->flags |= REQ_F_CREDS;
7105 state = &ctx->submit_state;
7108 * Plug now if we have more than 1 IO left after this, and the target
7109 * is potentially a read/write to block based storage.
7111 if (!state->plug_started && state->ios_left > 1 &&
7112 io_op_defs[req->opcode].plug) {
7113 blk_start_plug(&state->plug);
7114 state->plug_started = true;
7117 if (io_op_defs[req->opcode].needs_file) {
7118 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7119 (sqe_flags & IOSQE_FIXED_FILE));
7120 if (unlikely(!req->file))
7128 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7129 const struct io_uring_sqe *sqe)
7130 __must_hold(&ctx->uring_lock)
7132 struct io_submit_link *link = &ctx->submit_state.link;
7135 ret = io_init_req(ctx, req, sqe);
7136 if (unlikely(ret)) {
7138 /* fail even hard links since we don't submit */
7141 * we can judge a link req is failed or cancelled by if
7142 * REQ_F_FAIL is set, but the head is an exception since
7143 * it may be set REQ_F_FAIL because of other req's failure
7144 * so let's leverage req->result to distinguish if a head
7145 * is set REQ_F_FAIL because of its failure or other req's
7146 * failure so that we can set the correct ret code for it.
7147 * init result here to avoid affecting the normal path.
7149 if (!(link->head->flags & REQ_F_FAIL))
7150 req_fail_link_node(link->head, -ECANCELED);
7151 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7153 * the current req is a normal req, we should return
7154 * error and thus break the submittion loop.
7156 io_req_complete_failed(req, ret);
7159 req_fail_link_node(req, ret);
7161 ret = io_req_prep(req, sqe);
7166 /* don't need @sqe from now on */
7167 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7169 ctx->flags & IORING_SETUP_SQPOLL);
7172 * If we already have a head request, queue this one for async
7173 * submittal once the head completes. If we don't have a head but
7174 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7175 * submitted sync once the chain is complete. If none of those
7176 * conditions are true (normal request), then just queue it.
7179 struct io_kiocb *head = link->head;
7181 if (!(req->flags & REQ_F_FAIL)) {
7182 ret = io_req_prep_async(req);
7183 if (unlikely(ret)) {
7184 req_fail_link_node(req, ret);
7185 if (!(head->flags & REQ_F_FAIL))
7186 req_fail_link_node(head, -ECANCELED);
7189 trace_io_uring_link(ctx, req, head);
7190 link->last->link = req;
7193 /* last request of a link, enqueue the link */
7194 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7199 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7211 * Batched submission is done, ensure local IO is flushed out.
7213 static void io_submit_state_end(struct io_submit_state *state,
7214 struct io_ring_ctx *ctx)
7216 if (state->link.head)
7217 io_queue_sqe(state->link.head);
7218 if (state->compl_nr)
7219 io_submit_flush_completions(ctx);
7220 if (state->plug_started)
7221 blk_finish_plug(&state->plug);
7225 * Start submission side cache.
7227 static void io_submit_state_start(struct io_submit_state *state,
7228 unsigned int max_ios)
7230 state->plug_started = false;
7231 state->ios_left = max_ios;
7232 /* set only head, no need to init link_last in advance */
7233 state->link.head = NULL;
7236 static void io_commit_sqring(struct io_ring_ctx *ctx)
7238 struct io_rings *rings = ctx->rings;
7241 * Ensure any loads from the SQEs are done at this point,
7242 * since once we write the new head, the application could
7243 * write new data to them.
7245 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7249 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7250 * that is mapped by userspace. This means that care needs to be taken to
7251 * ensure that reads are stable, as we cannot rely on userspace always
7252 * being a good citizen. If members of the sqe are validated and then later
7253 * used, it's important that those reads are done through READ_ONCE() to
7254 * prevent a re-load down the line.
7256 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7258 unsigned head, mask = ctx->sq_entries - 1;
7259 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7262 * The cached sq head (or cq tail) serves two purposes:
7264 * 1) allows us to batch the cost of updating the user visible
7266 * 2) allows the kernel side to track the head on its own, even
7267 * though the application is the one updating it.
7269 head = READ_ONCE(ctx->sq_array[sq_idx]);
7270 if (likely(head < ctx->sq_entries))
7271 return &ctx->sq_sqes[head];
7273 /* drop invalid entries */
7275 WRITE_ONCE(ctx->rings->sq_dropped,
7276 READ_ONCE(ctx->rings->sq_dropped) + 1);
7280 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7281 __must_hold(&ctx->uring_lock)
7285 /* make sure SQ entry isn't read before tail */
7286 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7287 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7289 io_get_task_refs(nr);
7291 io_submit_state_start(&ctx->submit_state, nr);
7292 while (submitted < nr) {
7293 const struct io_uring_sqe *sqe;
7294 struct io_kiocb *req;
7296 req = io_alloc_req(ctx);
7297 if (unlikely(!req)) {
7299 submitted = -EAGAIN;
7302 sqe = io_get_sqe(ctx);
7303 if (unlikely(!sqe)) {
7304 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7307 /* will complete beyond this point, count as submitted */
7309 if (io_submit_sqe(ctx, req, sqe))
7313 if (unlikely(submitted != nr)) {
7314 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7315 int unused = nr - ref_used;
7317 current->io_uring->cached_refs += unused;
7318 percpu_ref_put_many(&ctx->refs, unused);
7321 io_submit_state_end(&ctx->submit_state, ctx);
7322 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7323 io_commit_sqring(ctx);
7328 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7330 return READ_ONCE(sqd->state);
7333 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7335 /* Tell userspace we may need a wakeup call */
7336 spin_lock(&ctx->completion_lock);
7337 WRITE_ONCE(ctx->rings->sq_flags,
7338 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7339 spin_unlock(&ctx->completion_lock);
7342 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7344 spin_lock(&ctx->completion_lock);
7345 WRITE_ONCE(ctx->rings->sq_flags,
7346 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7347 spin_unlock(&ctx->completion_lock);
7350 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7352 unsigned int to_submit;
7355 to_submit = io_sqring_entries(ctx);
7356 /* if we're handling multiple rings, cap submit size for fairness */
7357 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7358 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7360 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7361 unsigned nr_events = 0;
7362 const struct cred *creds = NULL;
7364 if (ctx->sq_creds != current_cred())
7365 creds = override_creds(ctx->sq_creds);
7367 mutex_lock(&ctx->uring_lock);
7368 if (!list_empty(&ctx->iopoll_list))
7369 io_do_iopoll(ctx, &nr_events, 0);
7372 * Don't submit if refs are dying, good for io_uring_register(),
7373 * but also it is relied upon by io_ring_exit_work()
7375 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7376 !(ctx->flags & IORING_SETUP_R_DISABLED))
7377 ret = io_submit_sqes(ctx, to_submit);
7378 mutex_unlock(&ctx->uring_lock);
7380 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7381 wake_up(&ctx->sqo_sq_wait);
7383 revert_creds(creds);
7389 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7391 struct io_ring_ctx *ctx;
7392 unsigned sq_thread_idle = 0;
7394 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7395 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7396 sqd->sq_thread_idle = sq_thread_idle;
7399 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7401 bool did_sig = false;
7402 struct ksignal ksig;
7404 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7405 signal_pending(current)) {
7406 mutex_unlock(&sqd->lock);
7407 if (signal_pending(current))
7408 did_sig = get_signal(&ksig);
7410 mutex_lock(&sqd->lock);
7412 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7415 static int io_sq_thread(void *data)
7417 struct io_sq_data *sqd = data;
7418 struct io_ring_ctx *ctx;
7419 unsigned long timeout = 0;
7420 char buf[TASK_COMM_LEN];
7423 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7424 set_task_comm(current, buf);
7426 if (sqd->sq_cpu != -1)
7427 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7429 set_cpus_allowed_ptr(current, cpu_online_mask);
7430 current->flags |= PF_NO_SETAFFINITY;
7432 mutex_lock(&sqd->lock);
7434 bool cap_entries, sqt_spin = false;
7436 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7437 if (io_sqd_handle_event(sqd))
7439 timeout = jiffies + sqd->sq_thread_idle;
7442 cap_entries = !list_is_singular(&sqd->ctx_list);
7443 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7444 int ret = __io_sq_thread(ctx, cap_entries);
7446 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7449 if (io_run_task_work())
7452 if (sqt_spin || !time_after(jiffies, timeout)) {
7455 timeout = jiffies + sqd->sq_thread_idle;
7459 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7460 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7461 bool needs_sched = true;
7463 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7464 io_ring_set_wakeup_flag(ctx);
7466 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7467 !list_empty_careful(&ctx->iopoll_list)) {
7468 needs_sched = false;
7471 if (io_sqring_entries(ctx)) {
7472 needs_sched = false;
7478 mutex_unlock(&sqd->lock);
7480 mutex_lock(&sqd->lock);
7482 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7483 io_ring_clear_wakeup_flag(ctx);
7486 finish_wait(&sqd->wait, &wait);
7487 timeout = jiffies + sqd->sq_thread_idle;
7490 io_uring_cancel_generic(true, sqd);
7492 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7493 io_ring_set_wakeup_flag(ctx);
7495 mutex_unlock(&sqd->lock);
7497 complete(&sqd->exited);
7501 struct io_wait_queue {
7502 struct wait_queue_entry wq;
7503 struct io_ring_ctx *ctx;
7505 unsigned nr_timeouts;
7508 static inline bool io_should_wake(struct io_wait_queue *iowq)
7510 struct io_ring_ctx *ctx = iowq->ctx;
7511 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7514 * Wake up if we have enough events, or if a timeout occurred since we
7515 * started waiting. For timeouts, we always want to return to userspace,
7516 * regardless of event count.
7518 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7521 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7522 int wake_flags, void *key)
7524 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7528 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7529 * the task, and the next invocation will do it.
7531 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7532 return autoremove_wake_function(curr, mode, wake_flags, key);
7536 static int io_run_task_work_sig(void)
7538 if (io_run_task_work())
7540 if (!signal_pending(current))
7542 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7543 return -ERESTARTSYS;
7547 /* when returns >0, the caller should retry */
7548 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7549 struct io_wait_queue *iowq,
7554 /* make sure we run task_work before checking for signals */
7555 ret = io_run_task_work_sig();
7556 if (ret || io_should_wake(iowq))
7558 /* let the caller flush overflows, retry */
7559 if (test_bit(0, &ctx->check_cq_overflow))
7562 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
7568 * Wait until events become available, if we don't already have some. The
7569 * application must reap them itself, as they reside on the shared cq ring.
7571 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7572 const sigset_t __user *sig, size_t sigsz,
7573 struct __kernel_timespec __user *uts)
7575 struct io_wait_queue iowq;
7576 struct io_rings *rings = ctx->rings;
7577 ktime_t timeout = KTIME_MAX;
7581 io_cqring_overflow_flush(ctx);
7582 if (io_cqring_events(ctx) >= min_events)
7584 if (!io_run_task_work())
7589 struct timespec64 ts;
7591 if (get_timespec64(&ts, uts))
7593 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
7597 #ifdef CONFIG_COMPAT
7598 if (in_compat_syscall())
7599 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7603 ret = set_user_sigmask(sig, sigsz);
7609 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7610 iowq.wq.private = current;
7611 INIT_LIST_HEAD(&iowq.wq.entry);
7613 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7614 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7616 trace_io_uring_cqring_wait(ctx, min_events);
7618 /* if we can't even flush overflow, don't wait for more */
7619 if (!io_cqring_overflow_flush(ctx)) {
7623 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7624 TASK_INTERRUPTIBLE);
7625 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
7626 finish_wait(&ctx->cq_wait, &iowq.wq);
7630 restore_saved_sigmask_unless(ret == -EINTR);
7632 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7635 static void io_free_page_table(void **table, size_t size)
7637 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7639 for (i = 0; i < nr_tables; i++)
7644 static void **io_alloc_page_table(size_t size)
7646 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7647 size_t init_size = size;
7650 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7654 for (i = 0; i < nr_tables; i++) {
7655 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7657 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7659 io_free_page_table(table, init_size);
7667 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7669 percpu_ref_exit(&ref_node->refs);
7673 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7675 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7676 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7677 unsigned long flags;
7678 bool first_add = false;
7679 unsigned long delay = HZ;
7681 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7684 /* if we are mid-quiesce then do not delay */
7685 if (node->rsrc_data->quiesce)
7688 while (!list_empty(&ctx->rsrc_ref_list)) {
7689 node = list_first_entry(&ctx->rsrc_ref_list,
7690 struct io_rsrc_node, node);
7691 /* recycle ref nodes in order */
7694 list_del(&node->node);
7695 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7697 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7700 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7703 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7705 struct io_rsrc_node *ref_node;
7707 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7711 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7716 INIT_LIST_HEAD(&ref_node->node);
7717 INIT_LIST_HEAD(&ref_node->rsrc_list);
7718 ref_node->done = false;
7722 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7723 struct io_rsrc_data *data_to_kill)
7725 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7726 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7729 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7731 rsrc_node->rsrc_data = data_to_kill;
7732 spin_lock_irq(&ctx->rsrc_ref_lock);
7733 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7734 spin_unlock_irq(&ctx->rsrc_ref_lock);
7736 atomic_inc(&data_to_kill->refs);
7737 percpu_ref_kill(&rsrc_node->refs);
7738 ctx->rsrc_node = NULL;
7741 if (!ctx->rsrc_node) {
7742 ctx->rsrc_node = ctx->rsrc_backup_node;
7743 ctx->rsrc_backup_node = NULL;
7747 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7749 if (ctx->rsrc_backup_node)
7751 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7752 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7755 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7759 /* As we may drop ->uring_lock, other task may have started quiesce */
7763 data->quiesce = true;
7765 ret = io_rsrc_node_switch_start(ctx);
7768 io_rsrc_node_switch(ctx, data);
7770 /* kill initial ref, already quiesced if zero */
7771 if (atomic_dec_and_test(&data->refs))
7773 mutex_unlock(&ctx->uring_lock);
7774 flush_delayed_work(&ctx->rsrc_put_work);
7775 ret = wait_for_completion_interruptible(&data->done);
7777 mutex_lock(&ctx->uring_lock);
7778 if (atomic_read(&data->refs) > 0) {
7780 * it has been revived by another thread while
7783 mutex_unlock(&ctx->uring_lock);
7789 atomic_inc(&data->refs);
7790 /* wait for all works potentially completing data->done */
7791 flush_delayed_work(&ctx->rsrc_put_work);
7792 reinit_completion(&data->done);
7794 ret = io_run_task_work_sig();
7795 mutex_lock(&ctx->uring_lock);
7797 data->quiesce = false;
7802 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7804 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7805 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7807 return &data->tags[table_idx][off];
7810 static void io_rsrc_data_free(struct io_rsrc_data *data)
7812 size_t size = data->nr * sizeof(data->tags[0][0]);
7815 io_free_page_table((void **)data->tags, size);
7819 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7820 u64 __user *utags, unsigned nr,
7821 struct io_rsrc_data **pdata)
7823 struct io_rsrc_data *data;
7827 data = kzalloc(sizeof(*data), GFP_KERNEL);
7830 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7838 data->do_put = do_put;
7841 for (i = 0; i < nr; i++) {
7842 u64 *tag_slot = io_get_tag_slot(data, i);
7844 if (copy_from_user(tag_slot, &utags[i],
7850 atomic_set(&data->refs, 1);
7851 init_completion(&data->done);
7855 io_rsrc_data_free(data);
7859 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7861 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7862 GFP_KERNEL_ACCOUNT);
7863 return !!table->files;
7866 static void io_free_file_tables(struct io_file_table *table)
7868 kvfree(table->files);
7869 table->files = NULL;
7872 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7874 #if defined(CONFIG_UNIX)
7875 if (ctx->ring_sock) {
7876 struct sock *sock = ctx->ring_sock->sk;
7877 struct sk_buff *skb;
7879 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7885 for (i = 0; i < ctx->nr_user_files; i++) {
7888 file = io_file_from_index(ctx, i);
7893 io_free_file_tables(&ctx->file_table);
7894 io_rsrc_data_free(ctx->file_data);
7895 ctx->file_data = NULL;
7896 ctx->nr_user_files = 0;
7899 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7901 unsigned nr = ctx->nr_user_files;
7904 if (!ctx->file_data)
7908 * Quiesce may unlock ->uring_lock, and while it's not held
7909 * prevent new requests using the table.
7911 ctx->nr_user_files = 0;
7912 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7913 ctx->nr_user_files = nr;
7915 __io_sqe_files_unregister(ctx);
7919 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7920 __releases(&sqd->lock)
7922 WARN_ON_ONCE(sqd->thread == current);
7925 * Do the dance but not conditional clear_bit() because it'd race with
7926 * other threads incrementing park_pending and setting the bit.
7928 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7929 if (atomic_dec_return(&sqd->park_pending))
7930 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7931 mutex_unlock(&sqd->lock);
7934 static void io_sq_thread_park(struct io_sq_data *sqd)
7935 __acquires(&sqd->lock)
7937 WARN_ON_ONCE(sqd->thread == current);
7939 atomic_inc(&sqd->park_pending);
7940 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7941 mutex_lock(&sqd->lock);
7943 wake_up_process(sqd->thread);
7946 static void io_sq_thread_stop(struct io_sq_data *sqd)
7948 WARN_ON_ONCE(sqd->thread == current);
7949 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7951 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7952 mutex_lock(&sqd->lock);
7954 wake_up_process(sqd->thread);
7955 mutex_unlock(&sqd->lock);
7956 wait_for_completion(&sqd->exited);
7959 static void io_put_sq_data(struct io_sq_data *sqd)
7961 if (refcount_dec_and_test(&sqd->refs)) {
7962 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7964 io_sq_thread_stop(sqd);
7969 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7971 struct io_sq_data *sqd = ctx->sq_data;
7974 io_sq_thread_park(sqd);
7975 list_del_init(&ctx->sqd_list);
7976 io_sqd_update_thread_idle(sqd);
7977 io_sq_thread_unpark(sqd);
7979 io_put_sq_data(sqd);
7980 ctx->sq_data = NULL;
7984 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7986 struct io_ring_ctx *ctx_attach;
7987 struct io_sq_data *sqd;
7990 f = fdget(p->wq_fd);
7992 return ERR_PTR(-ENXIO);
7993 if (f.file->f_op != &io_uring_fops) {
7995 return ERR_PTR(-EINVAL);
7998 ctx_attach = f.file->private_data;
7999 sqd = ctx_attach->sq_data;
8002 return ERR_PTR(-EINVAL);
8004 if (sqd->task_tgid != current->tgid) {
8006 return ERR_PTR(-EPERM);
8009 refcount_inc(&sqd->refs);
8014 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
8017 struct io_sq_data *sqd;
8020 if (p->flags & IORING_SETUP_ATTACH_WQ) {
8021 sqd = io_attach_sq_data(p);
8026 /* fall through for EPERM case, setup new sqd/task */
8027 if (PTR_ERR(sqd) != -EPERM)
8031 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
8033 return ERR_PTR(-ENOMEM);
8035 atomic_set(&sqd->park_pending, 0);
8036 refcount_set(&sqd->refs, 1);
8037 INIT_LIST_HEAD(&sqd->ctx_list);
8038 mutex_init(&sqd->lock);
8039 init_waitqueue_head(&sqd->wait);
8040 init_completion(&sqd->exited);
8044 #if defined(CONFIG_UNIX)
8046 * Ensure the UNIX gc is aware of our file set, so we are certain that
8047 * the io_uring can be safely unregistered on process exit, even if we have
8048 * loops in the file referencing.
8050 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8052 struct sock *sk = ctx->ring_sock->sk;
8053 struct scm_fp_list *fpl;
8054 struct sk_buff *skb;
8057 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8061 skb = alloc_skb(0, GFP_KERNEL);
8070 fpl->user = get_uid(current_user());
8071 for (i = 0; i < nr; i++) {
8072 struct file *file = io_file_from_index(ctx, i + offset);
8076 fpl->fp[nr_files] = get_file(file);
8077 unix_inflight(fpl->user, fpl->fp[nr_files]);
8082 fpl->max = SCM_MAX_FD;
8083 fpl->count = nr_files;
8084 UNIXCB(skb).fp = fpl;
8085 skb->destructor = unix_destruct_scm;
8086 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8087 skb_queue_head(&sk->sk_receive_queue, skb);
8089 for (i = 0; i < nr; i++) {
8090 struct file *file = io_file_from_index(ctx, i + offset);
8097 free_uid(fpl->user);
8105 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8106 * causes regular reference counting to break down. We rely on the UNIX
8107 * garbage collection to take care of this problem for us.
8109 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8111 unsigned left, total;
8115 left = ctx->nr_user_files;
8117 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8119 ret = __io_sqe_files_scm(ctx, this_files, total);
8123 total += this_files;
8129 while (total < ctx->nr_user_files) {
8130 struct file *file = io_file_from_index(ctx, total);
8140 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8146 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8148 struct file *file = prsrc->file;
8149 #if defined(CONFIG_UNIX)
8150 struct sock *sock = ctx->ring_sock->sk;
8151 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8152 struct sk_buff *skb;
8155 __skb_queue_head_init(&list);
8158 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8159 * remove this entry and rearrange the file array.
8161 skb = skb_dequeue(head);
8163 struct scm_fp_list *fp;
8165 fp = UNIXCB(skb).fp;
8166 for (i = 0; i < fp->count; i++) {
8169 if (fp->fp[i] != file)
8172 unix_notinflight(fp->user, fp->fp[i]);
8173 left = fp->count - 1 - i;
8175 memmove(&fp->fp[i], &fp->fp[i + 1],
8176 left * sizeof(struct file *));
8183 __skb_queue_tail(&list, skb);
8193 __skb_queue_tail(&list, skb);
8195 skb = skb_dequeue(head);
8198 if (skb_peek(&list)) {
8199 spin_lock_irq(&head->lock);
8200 while ((skb = __skb_dequeue(&list)) != NULL)
8201 __skb_queue_tail(head, skb);
8202 spin_unlock_irq(&head->lock);
8209 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8211 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8212 struct io_ring_ctx *ctx = rsrc_data->ctx;
8213 struct io_rsrc_put *prsrc, *tmp;
8215 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8216 list_del(&prsrc->list);
8219 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8221 io_ring_submit_lock(ctx, lock_ring);
8222 spin_lock(&ctx->completion_lock);
8223 io_fill_cqe_aux(ctx, prsrc->tag, 0, 0);
8224 io_commit_cqring(ctx);
8225 spin_unlock(&ctx->completion_lock);
8226 io_cqring_ev_posted(ctx);
8227 io_ring_submit_unlock(ctx, lock_ring);
8230 rsrc_data->do_put(ctx, prsrc);
8234 io_rsrc_node_destroy(ref_node);
8235 if (atomic_dec_and_test(&rsrc_data->refs))
8236 complete(&rsrc_data->done);
8239 static void io_rsrc_put_work(struct work_struct *work)
8241 struct io_ring_ctx *ctx;
8242 struct llist_node *node;
8244 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8245 node = llist_del_all(&ctx->rsrc_put_llist);
8248 struct io_rsrc_node *ref_node;
8249 struct llist_node *next = node->next;
8251 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8252 __io_rsrc_put_work(ref_node);
8257 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8258 unsigned nr_args, u64 __user *tags)
8260 __s32 __user *fds = (__s32 __user *) arg;
8269 if (nr_args > IORING_MAX_FIXED_FILES)
8271 if (nr_args > rlimit(RLIMIT_NOFILE))
8273 ret = io_rsrc_node_switch_start(ctx);
8276 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8282 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8285 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8286 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8290 /* allow sparse sets */
8293 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8300 if (unlikely(!file))
8304 * Don't allow io_uring instances to be registered. If UNIX
8305 * isn't enabled, then this causes a reference cycle and this
8306 * instance can never get freed. If UNIX is enabled we'll
8307 * handle it just fine, but there's still no point in allowing
8308 * a ring fd as it doesn't support regular read/write anyway.
8310 if (file->f_op == &io_uring_fops) {
8314 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8317 ret = io_sqe_files_scm(ctx);
8319 __io_sqe_files_unregister(ctx);
8323 io_rsrc_node_switch(ctx, NULL);
8326 for (i = 0; i < ctx->nr_user_files; i++) {
8327 file = io_file_from_index(ctx, i);
8331 io_free_file_tables(&ctx->file_table);
8332 ctx->nr_user_files = 0;
8334 io_rsrc_data_free(ctx->file_data);
8335 ctx->file_data = NULL;
8339 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8342 #if defined(CONFIG_UNIX)
8343 struct sock *sock = ctx->ring_sock->sk;
8344 struct sk_buff_head *head = &sock->sk_receive_queue;
8345 struct sk_buff *skb;
8348 * See if we can merge this file into an existing skb SCM_RIGHTS
8349 * file set. If there's no room, fall back to allocating a new skb
8350 * and filling it in.
8352 spin_lock_irq(&head->lock);
8353 skb = skb_peek(head);
8355 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8357 if (fpl->count < SCM_MAX_FD) {
8358 __skb_unlink(skb, head);
8359 spin_unlock_irq(&head->lock);
8360 fpl->fp[fpl->count] = get_file(file);
8361 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8363 spin_lock_irq(&head->lock);
8364 __skb_queue_head(head, skb);
8369 spin_unlock_irq(&head->lock);
8376 return __io_sqe_files_scm(ctx, 1, index);
8382 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8383 struct io_rsrc_node *node, void *rsrc)
8385 u64 *tag_slot = io_get_tag_slot(data, idx);
8386 struct io_rsrc_put *prsrc;
8388 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8392 prsrc->tag = *tag_slot;
8395 list_add(&prsrc->list, &node->rsrc_list);
8399 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8400 unsigned int issue_flags, u32 slot_index)
8402 struct io_ring_ctx *ctx = req->ctx;
8403 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8404 bool needs_switch = false;
8405 struct io_fixed_file *file_slot;
8408 io_ring_submit_lock(ctx, !force_nonblock);
8409 if (file->f_op == &io_uring_fops)
8412 if (!ctx->file_data)
8415 if (slot_index >= ctx->nr_user_files)
8418 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8419 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8421 if (file_slot->file_ptr) {
8422 struct file *old_file;
8424 ret = io_rsrc_node_switch_start(ctx);
8428 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8429 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8430 ctx->rsrc_node, old_file);
8433 file_slot->file_ptr = 0;
8434 needs_switch = true;
8437 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8438 io_fixed_file_set(file_slot, file);
8439 ret = io_sqe_file_register(ctx, file, slot_index);
8441 file_slot->file_ptr = 0;
8448 io_rsrc_node_switch(ctx, ctx->file_data);
8449 io_ring_submit_unlock(ctx, !force_nonblock);
8455 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8457 unsigned int offset = req->close.file_slot - 1;
8458 struct io_ring_ctx *ctx = req->ctx;
8459 struct io_fixed_file *file_slot;
8463 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8465 if (unlikely(!ctx->file_data))
8468 if (offset >= ctx->nr_user_files)
8470 ret = io_rsrc_node_switch_start(ctx);
8474 offset = array_index_nospec(offset, ctx->nr_user_files);
8475 file_slot = io_fixed_file_slot(&ctx->file_table, offset);
8477 if (!file_slot->file_ptr)
8480 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8481 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8485 file_slot->file_ptr = 0;
8486 io_rsrc_node_switch(ctx, ctx->file_data);
8489 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8493 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8494 struct io_uring_rsrc_update2 *up,
8497 u64 __user *tags = u64_to_user_ptr(up->tags);
8498 __s32 __user *fds = u64_to_user_ptr(up->data);
8499 struct io_rsrc_data *data = ctx->file_data;
8500 struct io_fixed_file *file_slot;
8504 bool needs_switch = false;
8506 if (!ctx->file_data)
8508 if (up->offset + nr_args > ctx->nr_user_files)
8511 for (done = 0; done < nr_args; done++) {
8514 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8515 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8519 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8523 if (fd == IORING_REGISTER_FILES_SKIP)
8526 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8527 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8529 if (file_slot->file_ptr) {
8530 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8531 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
8534 file_slot->file_ptr = 0;
8535 needs_switch = true;
8544 * Don't allow io_uring instances to be registered. If
8545 * UNIX isn't enabled, then this causes a reference
8546 * cycle and this instance can never get freed. If UNIX
8547 * is enabled we'll handle it just fine, but there's
8548 * still no point in allowing a ring fd as it doesn't
8549 * support regular read/write anyway.
8551 if (file->f_op == &io_uring_fops) {
8556 *io_get_tag_slot(data, i) = tag;
8557 io_fixed_file_set(file_slot, file);
8558 err = io_sqe_file_register(ctx, file, i);
8560 file_slot->file_ptr = 0;
8568 io_rsrc_node_switch(ctx, data);
8569 return done ? done : err;
8572 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8573 struct task_struct *task)
8575 struct io_wq_hash *hash;
8576 struct io_wq_data data;
8577 unsigned int concurrency;
8579 mutex_lock(&ctx->uring_lock);
8580 hash = ctx->hash_map;
8582 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8584 mutex_unlock(&ctx->uring_lock);
8585 return ERR_PTR(-ENOMEM);
8587 refcount_set(&hash->refs, 1);
8588 init_waitqueue_head(&hash->wait);
8589 ctx->hash_map = hash;
8591 mutex_unlock(&ctx->uring_lock);
8595 data.free_work = io_wq_free_work;
8596 data.do_work = io_wq_submit_work;
8598 /* Do QD, or 4 * CPUS, whatever is smallest */
8599 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8601 return io_wq_create(concurrency, &data);
8604 static int io_uring_alloc_task_context(struct task_struct *task,
8605 struct io_ring_ctx *ctx)
8607 struct io_uring_task *tctx;
8610 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8611 if (unlikely(!tctx))
8614 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8615 if (unlikely(ret)) {
8620 tctx->io_wq = io_init_wq_offload(ctx, task);
8621 if (IS_ERR(tctx->io_wq)) {
8622 ret = PTR_ERR(tctx->io_wq);
8623 percpu_counter_destroy(&tctx->inflight);
8629 init_waitqueue_head(&tctx->wait);
8630 atomic_set(&tctx->in_idle, 0);
8631 atomic_set(&tctx->inflight_tracked, 0);
8632 task->io_uring = tctx;
8633 spin_lock_init(&tctx->task_lock);
8634 INIT_WQ_LIST(&tctx->task_list);
8635 init_task_work(&tctx->task_work, tctx_task_work);
8639 void __io_uring_free(struct task_struct *tsk)
8641 struct io_uring_task *tctx = tsk->io_uring;
8643 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8644 WARN_ON_ONCE(tctx->io_wq);
8645 WARN_ON_ONCE(tctx->cached_refs);
8647 percpu_counter_destroy(&tctx->inflight);
8649 tsk->io_uring = NULL;
8652 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8653 struct io_uring_params *p)
8657 /* Retain compatibility with failing for an invalid attach attempt */
8658 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8659 IORING_SETUP_ATTACH_WQ) {
8662 f = fdget(p->wq_fd);
8665 if (f.file->f_op != &io_uring_fops) {
8671 if (ctx->flags & IORING_SETUP_SQPOLL) {
8672 struct task_struct *tsk;
8673 struct io_sq_data *sqd;
8676 sqd = io_get_sq_data(p, &attached);
8682 ctx->sq_creds = get_current_cred();
8684 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8685 if (!ctx->sq_thread_idle)
8686 ctx->sq_thread_idle = HZ;
8688 io_sq_thread_park(sqd);
8689 list_add(&ctx->sqd_list, &sqd->ctx_list);
8690 io_sqd_update_thread_idle(sqd);
8691 /* don't attach to a dying SQPOLL thread, would be racy */
8692 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8693 io_sq_thread_unpark(sqd);
8700 if (p->flags & IORING_SETUP_SQ_AFF) {
8701 int cpu = p->sq_thread_cpu;
8704 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8711 sqd->task_pid = current->pid;
8712 sqd->task_tgid = current->tgid;
8713 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8720 ret = io_uring_alloc_task_context(tsk, ctx);
8721 wake_up_new_task(tsk);
8724 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8725 /* Can't have SQ_AFF without SQPOLL */
8732 complete(&ctx->sq_data->exited);
8734 io_sq_thread_finish(ctx);
8738 static inline void __io_unaccount_mem(struct user_struct *user,
8739 unsigned long nr_pages)
8741 atomic_long_sub(nr_pages, &user->locked_vm);
8744 static inline int __io_account_mem(struct user_struct *user,
8745 unsigned long nr_pages)
8747 unsigned long page_limit, cur_pages, new_pages;
8749 /* Don't allow more pages than we can safely lock */
8750 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8753 cur_pages = atomic_long_read(&user->locked_vm);
8754 new_pages = cur_pages + nr_pages;
8755 if (new_pages > page_limit)
8757 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8758 new_pages) != cur_pages);
8763 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8766 __io_unaccount_mem(ctx->user, nr_pages);
8768 if (ctx->mm_account)
8769 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8772 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8777 ret = __io_account_mem(ctx->user, nr_pages);
8782 if (ctx->mm_account)
8783 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8788 static void io_mem_free(void *ptr)
8795 page = virt_to_head_page(ptr);
8796 if (put_page_testzero(page))
8797 free_compound_page(page);
8800 static void *io_mem_alloc(size_t size)
8802 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
8804 return (void *) __get_free_pages(gfp, get_order(size));
8807 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8810 struct io_rings *rings;
8811 size_t off, sq_array_size;
8813 off = struct_size(rings, cqes, cq_entries);
8814 if (off == SIZE_MAX)
8818 off = ALIGN(off, SMP_CACHE_BYTES);
8826 sq_array_size = array_size(sizeof(u32), sq_entries);
8827 if (sq_array_size == SIZE_MAX)
8830 if (check_add_overflow(off, sq_array_size, &off))
8836 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8838 struct io_mapped_ubuf *imu = *slot;
8841 if (imu != ctx->dummy_ubuf) {
8842 for (i = 0; i < imu->nr_bvecs; i++)
8843 unpin_user_page(imu->bvec[i].bv_page);
8844 if (imu->acct_pages)
8845 io_unaccount_mem(ctx, imu->acct_pages);
8851 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8853 io_buffer_unmap(ctx, &prsrc->buf);
8857 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8861 for (i = 0; i < ctx->nr_user_bufs; i++)
8862 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8863 kfree(ctx->user_bufs);
8864 io_rsrc_data_free(ctx->buf_data);
8865 ctx->user_bufs = NULL;
8866 ctx->buf_data = NULL;
8867 ctx->nr_user_bufs = 0;
8870 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8872 unsigned nr = ctx->nr_user_bufs;
8879 * Quiesce may unlock ->uring_lock, and while it's not held
8880 * prevent new requests using the table.
8882 ctx->nr_user_bufs = 0;
8883 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8884 ctx->nr_user_bufs = nr;
8886 __io_sqe_buffers_unregister(ctx);
8890 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8891 void __user *arg, unsigned index)
8893 struct iovec __user *src;
8895 #ifdef CONFIG_COMPAT
8897 struct compat_iovec __user *ciovs;
8898 struct compat_iovec ciov;
8900 ciovs = (struct compat_iovec __user *) arg;
8901 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8904 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8905 dst->iov_len = ciov.iov_len;
8909 src = (struct iovec __user *) arg;
8910 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8916 * Not super efficient, but this is just a registration time. And we do cache
8917 * the last compound head, so generally we'll only do a full search if we don't
8920 * We check if the given compound head page has already been accounted, to
8921 * avoid double accounting it. This allows us to account the full size of the
8922 * page, not just the constituent pages of a huge page.
8924 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8925 int nr_pages, struct page *hpage)
8929 /* check current page array */
8930 for (i = 0; i < nr_pages; i++) {
8931 if (!PageCompound(pages[i]))
8933 if (compound_head(pages[i]) == hpage)
8937 /* check previously registered pages */
8938 for (i = 0; i < ctx->nr_user_bufs; i++) {
8939 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8941 for (j = 0; j < imu->nr_bvecs; j++) {
8942 if (!PageCompound(imu->bvec[j].bv_page))
8944 if (compound_head(imu->bvec[j].bv_page) == hpage)
8952 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8953 int nr_pages, struct io_mapped_ubuf *imu,
8954 struct page **last_hpage)
8958 imu->acct_pages = 0;
8959 for (i = 0; i < nr_pages; i++) {
8960 if (!PageCompound(pages[i])) {
8965 hpage = compound_head(pages[i]);
8966 if (hpage == *last_hpage)
8968 *last_hpage = hpage;
8969 if (headpage_already_acct(ctx, pages, i, hpage))
8971 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8975 if (!imu->acct_pages)
8978 ret = io_account_mem(ctx, imu->acct_pages);
8980 imu->acct_pages = 0;
8984 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8985 struct io_mapped_ubuf **pimu,
8986 struct page **last_hpage)
8988 struct io_mapped_ubuf *imu = NULL;
8989 struct vm_area_struct **vmas = NULL;
8990 struct page **pages = NULL;
8991 unsigned long off, start, end, ubuf;
8993 int ret, pret, nr_pages, i;
8995 if (!iov->iov_base) {
8996 *pimu = ctx->dummy_ubuf;
9000 ubuf = (unsigned long) iov->iov_base;
9001 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
9002 start = ubuf >> PAGE_SHIFT;
9003 nr_pages = end - start;
9008 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
9012 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
9017 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
9022 mmap_read_lock(current->mm);
9023 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
9025 if (pret == nr_pages) {
9026 /* don't support file backed memory */
9027 for (i = 0; i < nr_pages; i++) {
9028 struct vm_area_struct *vma = vmas[i];
9030 if (vma_is_shmem(vma))
9033 !is_file_hugepages(vma->vm_file)) {
9039 ret = pret < 0 ? pret : -EFAULT;
9041 mmap_read_unlock(current->mm);
9044 * if we did partial map, or found file backed vmas,
9045 * release any pages we did get
9048 unpin_user_pages(pages, pret);
9052 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9054 unpin_user_pages(pages, pret);
9058 off = ubuf & ~PAGE_MASK;
9059 size = iov->iov_len;
9060 for (i = 0; i < nr_pages; i++) {
9063 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9064 imu->bvec[i].bv_page = pages[i];
9065 imu->bvec[i].bv_len = vec_len;
9066 imu->bvec[i].bv_offset = off;
9070 /* store original address for later verification */
9072 imu->ubuf_end = ubuf + iov->iov_len;
9073 imu->nr_bvecs = nr_pages;
9084 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9086 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9087 return ctx->user_bufs ? 0 : -ENOMEM;
9090 static int io_buffer_validate(struct iovec *iov)
9092 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9095 * Don't impose further limits on the size and buffer
9096 * constraints here, we'll -EINVAL later when IO is
9097 * submitted if they are wrong.
9100 return iov->iov_len ? -EFAULT : 0;
9104 /* arbitrary limit, but we need something */
9105 if (iov->iov_len > SZ_1G)
9108 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9114 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9115 unsigned int nr_args, u64 __user *tags)
9117 struct page *last_hpage = NULL;
9118 struct io_rsrc_data *data;
9124 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9126 ret = io_rsrc_node_switch_start(ctx);
9129 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9132 ret = io_buffers_map_alloc(ctx, nr_args);
9134 io_rsrc_data_free(data);
9138 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9139 ret = io_copy_iov(ctx, &iov, arg, i);
9142 ret = io_buffer_validate(&iov);
9145 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9150 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9156 WARN_ON_ONCE(ctx->buf_data);
9158 ctx->buf_data = data;
9160 __io_sqe_buffers_unregister(ctx);
9162 io_rsrc_node_switch(ctx, NULL);
9166 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9167 struct io_uring_rsrc_update2 *up,
9168 unsigned int nr_args)
9170 u64 __user *tags = u64_to_user_ptr(up->tags);
9171 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9172 struct page *last_hpage = NULL;
9173 bool needs_switch = false;
9179 if (up->offset + nr_args > ctx->nr_user_bufs)
9182 for (done = 0; done < nr_args; done++) {
9183 struct io_mapped_ubuf *imu;
9184 int offset = up->offset + done;
9187 err = io_copy_iov(ctx, &iov, iovs, done);
9190 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9194 err = io_buffer_validate(&iov);
9197 if (!iov.iov_base && tag) {
9201 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9205 i = array_index_nospec(offset, ctx->nr_user_bufs);
9206 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9207 err = io_queue_rsrc_removal(ctx->buf_data, i,
9208 ctx->rsrc_node, ctx->user_bufs[i]);
9209 if (unlikely(err)) {
9210 io_buffer_unmap(ctx, &imu);
9213 ctx->user_bufs[i] = NULL;
9214 needs_switch = true;
9217 ctx->user_bufs[i] = imu;
9218 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9222 io_rsrc_node_switch(ctx, ctx->buf_data);
9223 return done ? done : err;
9226 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9228 __s32 __user *fds = arg;
9234 if (copy_from_user(&fd, fds, sizeof(*fds)))
9237 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9238 if (IS_ERR(ctx->cq_ev_fd)) {
9239 int ret = PTR_ERR(ctx->cq_ev_fd);
9241 ctx->cq_ev_fd = NULL;
9248 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9250 if (ctx->cq_ev_fd) {
9251 eventfd_ctx_put(ctx->cq_ev_fd);
9252 ctx->cq_ev_fd = NULL;
9259 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9261 struct io_buffer *buf;
9262 unsigned long index;
9264 xa_for_each(&ctx->io_buffers, index, buf)
9265 __io_remove_buffers(ctx, buf, index, -1U);
9268 static void io_req_cache_free(struct list_head *list)
9270 struct io_kiocb *req, *nxt;
9272 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9273 list_del(&req->inflight_entry);
9274 kmem_cache_free(req_cachep, req);
9278 static void io_req_caches_free(struct io_ring_ctx *ctx)
9280 struct io_submit_state *state = &ctx->submit_state;
9282 mutex_lock(&ctx->uring_lock);
9284 if (state->free_reqs) {
9285 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9286 state->free_reqs = 0;
9289 io_flush_cached_locked_reqs(ctx, state);
9290 io_req_cache_free(&state->free_list);
9291 mutex_unlock(&ctx->uring_lock);
9294 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9296 if (data && !atomic_dec_and_test(&data->refs))
9297 wait_for_completion(&data->done);
9300 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9302 io_sq_thread_finish(ctx);
9304 if (ctx->mm_account) {
9305 mmdrop(ctx->mm_account);
9306 ctx->mm_account = NULL;
9309 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9310 io_wait_rsrc_data(ctx->buf_data);
9311 io_wait_rsrc_data(ctx->file_data);
9313 mutex_lock(&ctx->uring_lock);
9315 __io_sqe_buffers_unregister(ctx);
9317 __io_sqe_files_unregister(ctx);
9319 __io_cqring_overflow_flush(ctx, true);
9320 mutex_unlock(&ctx->uring_lock);
9321 io_eventfd_unregister(ctx);
9322 io_destroy_buffers(ctx);
9324 put_cred(ctx->sq_creds);
9326 /* there are no registered resources left, nobody uses it */
9328 io_rsrc_node_destroy(ctx->rsrc_node);
9329 if (ctx->rsrc_backup_node)
9330 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9331 flush_delayed_work(&ctx->rsrc_put_work);
9333 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9334 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9336 #if defined(CONFIG_UNIX)
9337 if (ctx->ring_sock) {
9338 ctx->ring_sock->file = NULL; /* so that iput() is called */
9339 sock_release(ctx->ring_sock);
9342 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9344 io_mem_free(ctx->rings);
9345 io_mem_free(ctx->sq_sqes);
9347 percpu_ref_exit(&ctx->refs);
9348 free_uid(ctx->user);
9349 io_req_caches_free(ctx);
9351 io_wq_put_hash(ctx->hash_map);
9352 kfree(ctx->cancel_hash);
9353 kfree(ctx->dummy_ubuf);
9357 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9359 struct io_ring_ctx *ctx = file->private_data;
9362 poll_wait(file, &ctx->poll_wait, wait);
9364 * synchronizes with barrier from wq_has_sleeper call in
9368 if (!io_sqring_full(ctx))
9369 mask |= EPOLLOUT | EPOLLWRNORM;
9372 * Don't flush cqring overflow list here, just do a simple check.
9373 * Otherwise there could possible be ABBA deadlock:
9376 * lock(&ctx->uring_lock);
9378 * lock(&ctx->uring_lock);
9381 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9382 * pushs them to do the flush.
9384 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9385 mask |= EPOLLIN | EPOLLRDNORM;
9390 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9392 const struct cred *creds;
9394 creds = xa_erase(&ctx->personalities, id);
9403 struct io_tctx_exit {
9404 struct callback_head task_work;
9405 struct completion completion;
9406 struct io_ring_ctx *ctx;
9409 static void io_tctx_exit_cb(struct callback_head *cb)
9411 struct io_uring_task *tctx = current->io_uring;
9412 struct io_tctx_exit *work;
9414 work = container_of(cb, struct io_tctx_exit, task_work);
9416 * When @in_idle, we're in cancellation and it's racy to remove the
9417 * node. It'll be removed by the end of cancellation, just ignore it.
9419 if (!atomic_read(&tctx->in_idle))
9420 io_uring_del_tctx_node((unsigned long)work->ctx);
9421 complete(&work->completion);
9424 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9426 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9428 return req->ctx == data;
9431 static void io_ring_exit_work(struct work_struct *work)
9433 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9434 unsigned long timeout = jiffies + HZ * 60 * 5;
9435 unsigned long interval = HZ / 20;
9436 struct io_tctx_exit exit;
9437 struct io_tctx_node *node;
9441 * If we're doing polled IO and end up having requests being
9442 * submitted async (out-of-line), then completions can come in while
9443 * we're waiting for refs to drop. We need to reap these manually,
9444 * as nobody else will be looking for them.
9447 io_uring_try_cancel_requests(ctx, NULL, true);
9449 struct io_sq_data *sqd = ctx->sq_data;
9450 struct task_struct *tsk;
9452 io_sq_thread_park(sqd);
9454 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9455 io_wq_cancel_cb(tsk->io_uring->io_wq,
9456 io_cancel_ctx_cb, ctx, true);
9457 io_sq_thread_unpark(sqd);
9460 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9461 /* there is little hope left, don't run it too often */
9464 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9466 init_completion(&exit.completion);
9467 init_task_work(&exit.task_work, io_tctx_exit_cb);
9470 * Some may use context even when all refs and requests have been put,
9471 * and they are free to do so while still holding uring_lock or
9472 * completion_lock, see io_req_task_submit(). Apart from other work,
9473 * this lock/unlock section also waits them to finish.
9475 mutex_lock(&ctx->uring_lock);
9476 while (!list_empty(&ctx->tctx_list)) {
9477 WARN_ON_ONCE(time_after(jiffies, timeout));
9479 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9481 /* don't spin on a single task if cancellation failed */
9482 list_rotate_left(&ctx->tctx_list);
9483 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9484 if (WARN_ON_ONCE(ret))
9486 wake_up_process(node->task);
9488 mutex_unlock(&ctx->uring_lock);
9489 wait_for_completion(&exit.completion);
9490 mutex_lock(&ctx->uring_lock);
9492 mutex_unlock(&ctx->uring_lock);
9493 spin_lock(&ctx->completion_lock);
9494 spin_unlock(&ctx->completion_lock);
9496 io_ring_ctx_free(ctx);
9499 /* Returns true if we found and killed one or more timeouts */
9500 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9503 struct io_kiocb *req, *tmp;
9506 spin_lock(&ctx->completion_lock);
9507 spin_lock_irq(&ctx->timeout_lock);
9508 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9509 if (io_match_task(req, tsk, cancel_all)) {
9510 io_kill_timeout(req, -ECANCELED);
9514 spin_unlock_irq(&ctx->timeout_lock);
9516 io_commit_cqring(ctx);
9517 spin_unlock(&ctx->completion_lock);
9519 io_cqring_ev_posted(ctx);
9520 return canceled != 0;
9523 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9525 unsigned long index;
9526 struct creds *creds;
9528 mutex_lock(&ctx->uring_lock);
9529 percpu_ref_kill(&ctx->refs);
9531 __io_cqring_overflow_flush(ctx, true);
9532 xa_for_each(&ctx->personalities, index, creds)
9533 io_unregister_personality(ctx, index);
9534 mutex_unlock(&ctx->uring_lock);
9536 io_kill_timeouts(ctx, NULL, true);
9537 io_poll_remove_all(ctx, NULL, true);
9539 /* if we failed setting up the ctx, we might not have any rings */
9540 io_iopoll_try_reap_events(ctx);
9542 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9544 * Use system_unbound_wq to avoid spawning tons of event kworkers
9545 * if we're exiting a ton of rings at the same time. It just adds
9546 * noise and overhead, there's no discernable change in runtime
9547 * over using system_wq.
9549 queue_work(system_unbound_wq, &ctx->exit_work);
9552 static int io_uring_release(struct inode *inode, struct file *file)
9554 struct io_ring_ctx *ctx = file->private_data;
9556 file->private_data = NULL;
9557 io_ring_ctx_wait_and_kill(ctx);
9561 struct io_task_cancel {
9562 struct task_struct *task;
9566 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9568 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9569 struct io_task_cancel *cancel = data;
9571 return io_match_task_safe(req, cancel->task, cancel->all);
9574 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9575 struct task_struct *task, bool cancel_all)
9577 struct io_defer_entry *de;
9580 spin_lock(&ctx->completion_lock);
9581 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9582 if (io_match_task_safe(de->req, task, cancel_all)) {
9583 list_cut_position(&list, &ctx->defer_list, &de->list);
9587 spin_unlock(&ctx->completion_lock);
9588 if (list_empty(&list))
9591 while (!list_empty(&list)) {
9592 de = list_first_entry(&list, struct io_defer_entry, list);
9593 list_del_init(&de->list);
9594 io_req_complete_failed(de->req, -ECANCELED);
9600 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9602 struct io_tctx_node *node;
9603 enum io_wq_cancel cret;
9606 mutex_lock(&ctx->uring_lock);
9607 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9608 struct io_uring_task *tctx = node->task->io_uring;
9611 * io_wq will stay alive while we hold uring_lock, because it's
9612 * killed after ctx nodes, which requires to take the lock.
9614 if (!tctx || !tctx->io_wq)
9616 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9617 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9619 mutex_unlock(&ctx->uring_lock);
9624 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9625 struct task_struct *task,
9628 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9629 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9632 enum io_wq_cancel cret;
9636 ret |= io_uring_try_cancel_iowq(ctx);
9637 } else if (tctx && tctx->io_wq) {
9639 * Cancels requests of all rings, not only @ctx, but
9640 * it's fine as the task is in exit/exec.
9642 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9644 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9647 /* SQPOLL thread does its own polling */
9648 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9649 (ctx->sq_data && ctx->sq_data->thread == current)) {
9650 while (!list_empty_careful(&ctx->iopoll_list)) {
9651 io_iopoll_try_reap_events(ctx);
9656 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9657 ret |= io_poll_remove_all(ctx, task, cancel_all);
9658 ret |= io_kill_timeouts(ctx, task, cancel_all);
9660 ret |= io_run_task_work();
9667 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9669 struct io_uring_task *tctx = current->io_uring;
9670 struct io_tctx_node *node;
9673 if (unlikely(!tctx)) {
9674 ret = io_uring_alloc_task_context(current, ctx);
9678 tctx = current->io_uring;
9679 if (ctx->iowq_limits_set) {
9680 unsigned int limits[2] = { ctx->iowq_limits[0],
9681 ctx->iowq_limits[1], };
9683 ret = io_wq_max_workers(tctx->io_wq, limits);
9688 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9689 node = kmalloc(sizeof(*node), GFP_KERNEL);
9693 node->task = current;
9695 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9702 mutex_lock(&ctx->uring_lock);
9703 list_add(&node->ctx_node, &ctx->tctx_list);
9704 mutex_unlock(&ctx->uring_lock);
9711 * Note that this task has used io_uring. We use it for cancelation purposes.
9713 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9715 struct io_uring_task *tctx = current->io_uring;
9717 if (likely(tctx && tctx->last == ctx))
9719 return __io_uring_add_tctx_node(ctx);
9723 * Remove this io_uring_file -> task mapping.
9725 static void io_uring_del_tctx_node(unsigned long index)
9727 struct io_uring_task *tctx = current->io_uring;
9728 struct io_tctx_node *node;
9732 node = xa_erase(&tctx->xa, index);
9736 WARN_ON_ONCE(current != node->task);
9737 WARN_ON_ONCE(list_empty(&node->ctx_node));
9739 mutex_lock(&node->ctx->uring_lock);
9740 list_del(&node->ctx_node);
9741 mutex_unlock(&node->ctx->uring_lock);
9743 if (tctx->last == node->ctx)
9748 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9750 struct io_wq *wq = tctx->io_wq;
9751 struct io_tctx_node *node;
9752 unsigned long index;
9754 xa_for_each(&tctx->xa, index, node) {
9755 io_uring_del_tctx_node(index);
9760 * Must be after io_uring_del_task_file() (removes nodes under
9761 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9763 io_wq_put_and_exit(wq);
9768 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9771 return atomic_read(&tctx->inflight_tracked);
9772 return percpu_counter_sum(&tctx->inflight);
9776 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9777 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9779 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9781 struct io_uring_task *tctx = current->io_uring;
9782 struct io_ring_ctx *ctx;
9786 WARN_ON_ONCE(sqd && sqd->thread != current);
9788 if (!current->io_uring)
9791 io_wq_exit_start(tctx->io_wq);
9793 atomic_inc(&tctx->in_idle);
9795 io_uring_drop_tctx_refs(current);
9796 /* read completions before cancelations */
9797 inflight = tctx_inflight(tctx, !cancel_all);
9802 struct io_tctx_node *node;
9803 unsigned long index;
9805 xa_for_each(&tctx->xa, index, node) {
9806 /* sqpoll task will cancel all its requests */
9807 if (node->ctx->sq_data)
9809 io_uring_try_cancel_requests(node->ctx, current,
9813 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9814 io_uring_try_cancel_requests(ctx, current,
9818 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
9820 io_uring_drop_tctx_refs(current);
9823 * If we've seen completions, retry without waiting. This
9824 * avoids a race where a completion comes in before we did
9825 * prepare_to_wait().
9827 if (inflight == tctx_inflight(tctx, !cancel_all))
9829 finish_wait(&tctx->wait, &wait);
9832 io_uring_clean_tctx(tctx);
9835 * We shouldn't run task_works after cancel, so just leave
9836 * ->in_idle set for normal exit.
9838 atomic_dec(&tctx->in_idle);
9839 /* for exec all current's requests should be gone, kill tctx */
9840 __io_uring_free(current);
9844 void __io_uring_cancel(bool cancel_all)
9846 io_uring_cancel_generic(cancel_all, NULL);
9849 static void *io_uring_validate_mmap_request(struct file *file,
9850 loff_t pgoff, size_t sz)
9852 struct io_ring_ctx *ctx = file->private_data;
9853 loff_t offset = pgoff << PAGE_SHIFT;
9858 case IORING_OFF_SQ_RING:
9859 case IORING_OFF_CQ_RING:
9862 case IORING_OFF_SQES:
9866 return ERR_PTR(-EINVAL);
9869 page = virt_to_head_page(ptr);
9870 if (sz > page_size(page))
9871 return ERR_PTR(-EINVAL);
9878 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9880 size_t sz = vma->vm_end - vma->vm_start;
9884 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9886 return PTR_ERR(ptr);
9888 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9889 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9892 #else /* !CONFIG_MMU */
9894 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9896 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9899 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9901 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9904 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9905 unsigned long addr, unsigned long len,
9906 unsigned long pgoff, unsigned long flags)
9910 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9912 return PTR_ERR(ptr);
9914 return (unsigned long) ptr;
9917 #endif /* !CONFIG_MMU */
9919 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9924 if (!io_sqring_full(ctx))
9926 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9928 if (!io_sqring_full(ctx))
9931 } while (!signal_pending(current));
9933 finish_wait(&ctx->sqo_sq_wait, &wait);
9937 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9938 struct __kernel_timespec __user **ts,
9939 const sigset_t __user **sig)
9941 struct io_uring_getevents_arg arg;
9944 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9945 * is just a pointer to the sigset_t.
9947 if (!(flags & IORING_ENTER_EXT_ARG)) {
9948 *sig = (const sigset_t __user *) argp;
9954 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9955 * timespec and sigset_t pointers if good.
9957 if (*argsz != sizeof(arg))
9959 if (copy_from_user(&arg, argp, sizeof(arg)))
9963 *sig = u64_to_user_ptr(arg.sigmask);
9964 *argsz = arg.sigmask_sz;
9965 *ts = u64_to_user_ptr(arg.ts);
9969 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9970 u32, min_complete, u32, flags, const void __user *, argp,
9973 struct io_ring_ctx *ctx;
9980 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9981 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9985 if (unlikely(!f.file))
9989 if (unlikely(f.file->f_op != &io_uring_fops))
9993 ctx = f.file->private_data;
9994 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9998 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
10002 * For SQ polling, the thread will do all submissions and completions.
10003 * Just return the requested submit count, and wake the thread if
10004 * we were asked to.
10007 if (ctx->flags & IORING_SETUP_SQPOLL) {
10008 io_cqring_overflow_flush(ctx);
10010 if (unlikely(ctx->sq_data->thread == NULL)) {
10014 if (flags & IORING_ENTER_SQ_WAKEUP)
10015 wake_up(&ctx->sq_data->wait);
10016 if (flags & IORING_ENTER_SQ_WAIT) {
10017 ret = io_sqpoll_wait_sq(ctx);
10021 submitted = to_submit;
10022 } else if (to_submit) {
10023 ret = io_uring_add_tctx_node(ctx);
10026 mutex_lock(&ctx->uring_lock);
10027 submitted = io_submit_sqes(ctx, to_submit);
10028 mutex_unlock(&ctx->uring_lock);
10030 if (submitted != to_submit)
10033 if (flags & IORING_ENTER_GETEVENTS) {
10034 const sigset_t __user *sig;
10035 struct __kernel_timespec __user *ts;
10037 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10041 min_complete = min(min_complete, ctx->cq_entries);
10044 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10045 * space applications don't need to do io completion events
10046 * polling again, they can rely on io_sq_thread to do polling
10047 * work, which can reduce cpu usage and uring_lock contention.
10049 if (ctx->flags & IORING_SETUP_IOPOLL &&
10050 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10051 ret = io_iopoll_check(ctx, min_complete);
10053 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10058 percpu_ref_put(&ctx->refs);
10061 return submitted ? submitted : ret;
10064 #ifdef CONFIG_PROC_FS
10065 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10066 const struct cred *cred)
10068 struct user_namespace *uns = seq_user_ns(m);
10069 struct group_info *gi;
10074 seq_printf(m, "%5d\n", id);
10075 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10076 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10077 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10078 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10079 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10080 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10081 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10082 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10083 seq_puts(m, "\n\tGroups:\t");
10084 gi = cred->group_info;
10085 for (g = 0; g < gi->ngroups; g++) {
10086 seq_put_decimal_ull(m, g ? " " : "",
10087 from_kgid_munged(uns, gi->gid[g]));
10089 seq_puts(m, "\n\tCapEff:\t");
10090 cap = cred->cap_effective;
10091 CAP_FOR_EACH_U32(__capi)
10092 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10097 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10099 struct io_sq_data *sq = NULL;
10104 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10105 * since fdinfo case grabs it in the opposite direction of normal use
10106 * cases. If we fail to get the lock, we just don't iterate any
10107 * structures that could be going away outside the io_uring mutex.
10109 has_lock = mutex_trylock(&ctx->uring_lock);
10111 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10117 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10118 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10119 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10120 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10121 struct file *f = io_file_from_index(ctx, i);
10124 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10126 seq_printf(m, "%5u: <none>\n", i);
10128 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10129 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10130 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10131 unsigned int len = buf->ubuf_end - buf->ubuf;
10133 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10135 if (has_lock && !xa_empty(&ctx->personalities)) {
10136 unsigned long index;
10137 const struct cred *cred;
10139 seq_printf(m, "Personalities:\n");
10140 xa_for_each(&ctx->personalities, index, cred)
10141 io_uring_show_cred(m, index, cred);
10143 seq_printf(m, "PollList:\n");
10144 spin_lock(&ctx->completion_lock);
10145 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10146 struct hlist_head *list = &ctx->cancel_hash[i];
10147 struct io_kiocb *req;
10149 hlist_for_each_entry(req, list, hash_node)
10150 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10151 req->task->task_works != NULL);
10153 spin_unlock(&ctx->completion_lock);
10155 mutex_unlock(&ctx->uring_lock);
10158 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10160 struct io_ring_ctx *ctx = f->private_data;
10162 if (percpu_ref_tryget(&ctx->refs)) {
10163 __io_uring_show_fdinfo(ctx, m);
10164 percpu_ref_put(&ctx->refs);
10169 static const struct file_operations io_uring_fops = {
10170 .release = io_uring_release,
10171 .mmap = io_uring_mmap,
10173 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10174 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10176 .poll = io_uring_poll,
10177 #ifdef CONFIG_PROC_FS
10178 .show_fdinfo = io_uring_show_fdinfo,
10182 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10183 struct io_uring_params *p)
10185 struct io_rings *rings;
10186 size_t size, sq_array_offset;
10188 /* make sure these are sane, as we already accounted them */
10189 ctx->sq_entries = p->sq_entries;
10190 ctx->cq_entries = p->cq_entries;
10192 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10193 if (size == SIZE_MAX)
10196 rings = io_mem_alloc(size);
10200 ctx->rings = rings;
10201 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10202 rings->sq_ring_mask = p->sq_entries - 1;
10203 rings->cq_ring_mask = p->cq_entries - 1;
10204 rings->sq_ring_entries = p->sq_entries;
10205 rings->cq_ring_entries = p->cq_entries;
10207 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10208 if (size == SIZE_MAX) {
10209 io_mem_free(ctx->rings);
10214 ctx->sq_sqes = io_mem_alloc(size);
10215 if (!ctx->sq_sqes) {
10216 io_mem_free(ctx->rings);
10224 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10228 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10232 ret = io_uring_add_tctx_node(ctx);
10237 fd_install(fd, file);
10242 * Allocate an anonymous fd, this is what constitutes the application
10243 * visible backing of an io_uring instance. The application mmaps this
10244 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10245 * we have to tie this fd to a socket for file garbage collection purposes.
10247 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10250 #if defined(CONFIG_UNIX)
10253 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10256 return ERR_PTR(ret);
10259 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10260 O_RDWR | O_CLOEXEC);
10261 #if defined(CONFIG_UNIX)
10262 if (IS_ERR(file)) {
10263 sock_release(ctx->ring_sock);
10264 ctx->ring_sock = NULL;
10266 ctx->ring_sock->file = file;
10272 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10273 struct io_uring_params __user *params)
10275 struct io_ring_ctx *ctx;
10281 if (entries > IORING_MAX_ENTRIES) {
10282 if (!(p->flags & IORING_SETUP_CLAMP))
10284 entries = IORING_MAX_ENTRIES;
10288 * Use twice as many entries for the CQ ring. It's possible for the
10289 * application to drive a higher depth than the size of the SQ ring,
10290 * since the sqes are only used at submission time. This allows for
10291 * some flexibility in overcommitting a bit. If the application has
10292 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10293 * of CQ ring entries manually.
10295 p->sq_entries = roundup_pow_of_two(entries);
10296 if (p->flags & IORING_SETUP_CQSIZE) {
10298 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10299 * to a power-of-two, if it isn't already. We do NOT impose
10300 * any cq vs sq ring sizing.
10302 if (!p->cq_entries)
10304 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10305 if (!(p->flags & IORING_SETUP_CLAMP))
10307 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10309 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10310 if (p->cq_entries < p->sq_entries)
10313 p->cq_entries = 2 * p->sq_entries;
10316 ctx = io_ring_ctx_alloc(p);
10319 ctx->compat = in_compat_syscall();
10320 if (!capable(CAP_IPC_LOCK))
10321 ctx->user = get_uid(current_user());
10324 * This is just grabbed for accounting purposes. When a process exits,
10325 * the mm is exited and dropped before the files, hence we need to hang
10326 * on to this mm purely for the purposes of being able to unaccount
10327 * memory (locked/pinned vm). It's not used for anything else.
10329 mmgrab(current->mm);
10330 ctx->mm_account = current->mm;
10332 ret = io_allocate_scq_urings(ctx, p);
10336 ret = io_sq_offload_create(ctx, p);
10339 /* always set a rsrc node */
10340 ret = io_rsrc_node_switch_start(ctx);
10343 io_rsrc_node_switch(ctx, NULL);
10345 memset(&p->sq_off, 0, sizeof(p->sq_off));
10346 p->sq_off.head = offsetof(struct io_rings, sq.head);
10347 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10348 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10349 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10350 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10351 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10352 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10354 memset(&p->cq_off, 0, sizeof(p->cq_off));
10355 p->cq_off.head = offsetof(struct io_rings, cq.head);
10356 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10357 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10358 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10359 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10360 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10361 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10363 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10364 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10365 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10366 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10367 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10368 IORING_FEAT_RSRC_TAGS;
10370 if (copy_to_user(params, p, sizeof(*p))) {
10375 file = io_uring_get_file(ctx);
10376 if (IS_ERR(file)) {
10377 ret = PTR_ERR(file);
10382 * Install ring fd as the very last thing, so we don't risk someone
10383 * having closed it before we finish setup
10385 ret = io_uring_install_fd(ctx, file);
10387 /* fput will clean it up */
10392 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10395 io_ring_ctx_wait_and_kill(ctx);
10400 * Sets up an aio uring context, and returns the fd. Applications asks for a
10401 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10402 * params structure passed in.
10404 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10406 struct io_uring_params p;
10409 if (copy_from_user(&p, params, sizeof(p)))
10411 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10416 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10417 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10418 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10419 IORING_SETUP_R_DISABLED))
10422 return io_uring_create(entries, &p, params);
10425 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10426 struct io_uring_params __user *, params)
10428 return io_uring_setup(entries, params);
10431 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10433 struct io_uring_probe *p;
10437 size = struct_size(p, ops, nr_args);
10438 if (size == SIZE_MAX)
10440 p = kzalloc(size, GFP_KERNEL);
10445 if (copy_from_user(p, arg, size))
10448 if (memchr_inv(p, 0, size))
10451 p->last_op = IORING_OP_LAST - 1;
10452 if (nr_args > IORING_OP_LAST)
10453 nr_args = IORING_OP_LAST;
10455 for (i = 0; i < nr_args; i++) {
10457 if (!io_op_defs[i].not_supported)
10458 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10463 if (copy_to_user(arg, p, size))
10470 static int io_register_personality(struct io_ring_ctx *ctx)
10472 const struct cred *creds;
10476 creds = get_current_cred();
10478 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10479 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10487 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10488 unsigned int nr_args)
10490 struct io_uring_restriction *res;
10494 /* Restrictions allowed only if rings started disabled */
10495 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10498 /* We allow only a single restrictions registration */
10499 if (ctx->restrictions.registered)
10502 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10505 size = array_size(nr_args, sizeof(*res));
10506 if (size == SIZE_MAX)
10509 res = memdup_user(arg, size);
10511 return PTR_ERR(res);
10515 for (i = 0; i < nr_args; i++) {
10516 switch (res[i].opcode) {
10517 case IORING_RESTRICTION_REGISTER_OP:
10518 if (res[i].register_op >= IORING_REGISTER_LAST) {
10523 __set_bit(res[i].register_op,
10524 ctx->restrictions.register_op);
10526 case IORING_RESTRICTION_SQE_OP:
10527 if (res[i].sqe_op >= IORING_OP_LAST) {
10532 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10534 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10535 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10537 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10538 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10547 /* Reset all restrictions if an error happened */
10549 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10551 ctx->restrictions.registered = true;
10557 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10559 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10562 if (ctx->restrictions.registered)
10563 ctx->restricted = 1;
10565 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10566 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10567 wake_up(&ctx->sq_data->wait);
10571 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10572 struct io_uring_rsrc_update2 *up,
10578 if (check_add_overflow(up->offset, nr_args, &tmp))
10580 err = io_rsrc_node_switch_start(ctx);
10585 case IORING_RSRC_FILE:
10586 return __io_sqe_files_update(ctx, up, nr_args);
10587 case IORING_RSRC_BUFFER:
10588 return __io_sqe_buffers_update(ctx, up, nr_args);
10593 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10596 struct io_uring_rsrc_update2 up;
10600 memset(&up, 0, sizeof(up));
10601 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10603 if (up.resv || up.resv2)
10605 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10608 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10609 unsigned size, unsigned type)
10611 struct io_uring_rsrc_update2 up;
10613 if (size != sizeof(up))
10615 if (copy_from_user(&up, arg, sizeof(up)))
10617 if (!up.nr || up.resv || up.resv2)
10619 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10622 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10623 unsigned int size, unsigned int type)
10625 struct io_uring_rsrc_register rr;
10627 /* keep it extendible */
10628 if (size != sizeof(rr))
10631 memset(&rr, 0, sizeof(rr));
10632 if (copy_from_user(&rr, arg, size))
10634 if (!rr.nr || rr.resv || rr.resv2)
10638 case IORING_RSRC_FILE:
10639 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10640 rr.nr, u64_to_user_ptr(rr.tags));
10641 case IORING_RSRC_BUFFER:
10642 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10643 rr.nr, u64_to_user_ptr(rr.tags));
10648 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10651 struct io_uring_task *tctx = current->io_uring;
10652 cpumask_var_t new_mask;
10655 if (!tctx || !tctx->io_wq)
10658 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10661 cpumask_clear(new_mask);
10662 if (len > cpumask_size())
10663 len = cpumask_size();
10665 if (in_compat_syscall()) {
10666 ret = compat_get_bitmap(cpumask_bits(new_mask),
10667 (const compat_ulong_t __user *)arg,
10668 len * 8 /* CHAR_BIT */);
10670 ret = copy_from_user(new_mask, arg, len);
10674 free_cpumask_var(new_mask);
10678 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10679 free_cpumask_var(new_mask);
10683 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10685 struct io_uring_task *tctx = current->io_uring;
10687 if (!tctx || !tctx->io_wq)
10690 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10693 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10695 __must_hold(&ctx->uring_lock)
10697 struct io_tctx_node *node;
10698 struct io_uring_task *tctx = NULL;
10699 struct io_sq_data *sqd = NULL;
10700 __u32 new_count[2];
10703 if (copy_from_user(new_count, arg, sizeof(new_count)))
10705 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10706 if (new_count[i] > INT_MAX)
10709 if (ctx->flags & IORING_SETUP_SQPOLL) {
10710 sqd = ctx->sq_data;
10713 * Observe the correct sqd->lock -> ctx->uring_lock
10714 * ordering. Fine to drop uring_lock here, we hold
10715 * a ref to the ctx.
10717 refcount_inc(&sqd->refs);
10718 mutex_unlock(&ctx->uring_lock);
10719 mutex_lock(&sqd->lock);
10720 mutex_lock(&ctx->uring_lock);
10722 tctx = sqd->thread->io_uring;
10725 tctx = current->io_uring;
10728 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
10730 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10732 ctx->iowq_limits[i] = new_count[i];
10733 ctx->iowq_limits_set = true;
10736 if (tctx && tctx->io_wq) {
10737 ret = io_wq_max_workers(tctx->io_wq, new_count);
10741 memset(new_count, 0, sizeof(new_count));
10745 mutex_unlock(&sqd->lock);
10746 io_put_sq_data(sqd);
10749 if (copy_to_user(arg, new_count, sizeof(new_count)))
10752 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10756 /* now propagate the restriction to all registered users */
10757 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
10758 struct io_uring_task *tctx = node->task->io_uring;
10760 if (WARN_ON_ONCE(!tctx->io_wq))
10763 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10764 new_count[i] = ctx->iowq_limits[i];
10765 /* ignore errors, it always returns zero anyway */
10766 (void)io_wq_max_workers(tctx->io_wq, new_count);
10771 mutex_unlock(&sqd->lock);
10772 io_put_sq_data(sqd);
10777 static bool io_register_op_must_quiesce(int op)
10780 case IORING_REGISTER_BUFFERS:
10781 case IORING_UNREGISTER_BUFFERS:
10782 case IORING_REGISTER_FILES:
10783 case IORING_UNREGISTER_FILES:
10784 case IORING_REGISTER_FILES_UPDATE:
10785 case IORING_REGISTER_PROBE:
10786 case IORING_REGISTER_PERSONALITY:
10787 case IORING_UNREGISTER_PERSONALITY:
10788 case IORING_REGISTER_FILES2:
10789 case IORING_REGISTER_FILES_UPDATE2:
10790 case IORING_REGISTER_BUFFERS2:
10791 case IORING_REGISTER_BUFFERS_UPDATE:
10792 case IORING_REGISTER_IOWQ_AFF:
10793 case IORING_UNREGISTER_IOWQ_AFF:
10794 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10801 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10805 percpu_ref_kill(&ctx->refs);
10808 * Drop uring mutex before waiting for references to exit. If another
10809 * thread is currently inside io_uring_enter() it might need to grab the
10810 * uring_lock to make progress. If we hold it here across the drain
10811 * wait, then we can deadlock. It's safe to drop the mutex here, since
10812 * no new references will come in after we've killed the percpu ref.
10814 mutex_unlock(&ctx->uring_lock);
10816 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10819 ret = io_run_task_work_sig();
10820 } while (ret >= 0);
10821 mutex_lock(&ctx->uring_lock);
10824 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10828 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10829 void __user *arg, unsigned nr_args)
10830 __releases(ctx->uring_lock)
10831 __acquires(ctx->uring_lock)
10836 * We're inside the ring mutex, if the ref is already dying, then
10837 * someone else killed the ctx or is already going through
10838 * io_uring_register().
10840 if (percpu_ref_is_dying(&ctx->refs))
10843 if (ctx->restricted) {
10844 if (opcode >= IORING_REGISTER_LAST)
10846 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10847 if (!test_bit(opcode, ctx->restrictions.register_op))
10851 if (io_register_op_must_quiesce(opcode)) {
10852 ret = io_ctx_quiesce(ctx);
10858 case IORING_REGISTER_BUFFERS:
10859 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10861 case IORING_UNREGISTER_BUFFERS:
10863 if (arg || nr_args)
10865 ret = io_sqe_buffers_unregister(ctx);
10867 case IORING_REGISTER_FILES:
10868 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10870 case IORING_UNREGISTER_FILES:
10872 if (arg || nr_args)
10874 ret = io_sqe_files_unregister(ctx);
10876 case IORING_REGISTER_FILES_UPDATE:
10877 ret = io_register_files_update(ctx, arg, nr_args);
10879 case IORING_REGISTER_EVENTFD:
10880 case IORING_REGISTER_EVENTFD_ASYNC:
10884 ret = io_eventfd_register(ctx, arg);
10887 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10888 ctx->eventfd_async = 1;
10890 ctx->eventfd_async = 0;
10892 case IORING_UNREGISTER_EVENTFD:
10894 if (arg || nr_args)
10896 ret = io_eventfd_unregister(ctx);
10898 case IORING_REGISTER_PROBE:
10900 if (!arg || nr_args > 256)
10902 ret = io_probe(ctx, arg, nr_args);
10904 case IORING_REGISTER_PERSONALITY:
10906 if (arg || nr_args)
10908 ret = io_register_personality(ctx);
10910 case IORING_UNREGISTER_PERSONALITY:
10914 ret = io_unregister_personality(ctx, nr_args);
10916 case IORING_REGISTER_ENABLE_RINGS:
10918 if (arg || nr_args)
10920 ret = io_register_enable_rings(ctx);
10922 case IORING_REGISTER_RESTRICTIONS:
10923 ret = io_register_restrictions(ctx, arg, nr_args);
10925 case IORING_REGISTER_FILES2:
10926 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10928 case IORING_REGISTER_FILES_UPDATE2:
10929 ret = io_register_rsrc_update(ctx, arg, nr_args,
10932 case IORING_REGISTER_BUFFERS2:
10933 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10935 case IORING_REGISTER_BUFFERS_UPDATE:
10936 ret = io_register_rsrc_update(ctx, arg, nr_args,
10937 IORING_RSRC_BUFFER);
10939 case IORING_REGISTER_IOWQ_AFF:
10941 if (!arg || !nr_args)
10943 ret = io_register_iowq_aff(ctx, arg, nr_args);
10945 case IORING_UNREGISTER_IOWQ_AFF:
10947 if (arg || nr_args)
10949 ret = io_unregister_iowq_aff(ctx);
10951 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10953 if (!arg || nr_args != 2)
10955 ret = io_register_iowq_max_workers(ctx, arg);
10962 if (io_register_op_must_quiesce(opcode)) {
10963 /* bring the ctx back to life */
10964 percpu_ref_reinit(&ctx->refs);
10965 reinit_completion(&ctx->ref_comp);
10970 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10971 void __user *, arg, unsigned int, nr_args)
10973 struct io_ring_ctx *ctx;
10982 if (f.file->f_op != &io_uring_fops)
10985 ctx = f.file->private_data;
10987 io_run_task_work();
10989 mutex_lock(&ctx->uring_lock);
10990 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10991 mutex_unlock(&ctx->uring_lock);
10992 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10993 ctx->cq_ev_fd != NULL, ret);
10999 static int __init io_uring_init(void)
11001 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
11002 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
11003 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
11006 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
11007 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
11008 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
11009 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
11010 BUILD_BUG_SQE_ELEM(1, __u8, flags);
11011 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
11012 BUILD_BUG_SQE_ELEM(4, __s32, fd);
11013 BUILD_BUG_SQE_ELEM(8, __u64, off);
11014 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
11015 BUILD_BUG_SQE_ELEM(16, __u64, addr);
11016 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
11017 BUILD_BUG_SQE_ELEM(24, __u32, len);
11018 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
11019 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
11020 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
11021 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
11022 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
11023 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
11024 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
11025 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
11026 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
11027 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
11028 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
11029 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
11030 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
11031 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
11032 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
11033 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
11034 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
11035 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
11036 BUILD_BUG_SQE_ELEM(42, __u16, personality);
11037 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
11038 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
11040 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
11041 sizeof(struct io_uring_rsrc_update));
11042 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
11043 sizeof(struct io_uring_rsrc_update2));
11045 /* ->buf_index is u16 */
11046 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
11048 /* should fit into one byte */
11049 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
11051 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
11052 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
11054 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11058 __initcall(io_uring_init);