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 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
98 #define IORING_FILE_TABLE_SHIFT 9
99 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
100 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
101 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
102 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
103 IORING_REGISTER_LAST + IORING_OP_LAST)
105 #define IO_RSRC_TAG_TABLE_SHIFT 9
106 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
107 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
109 #define IORING_MAX_REG_BUFFERS (1U << 14)
111 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
112 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
114 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
115 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
117 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
120 u32 head ____cacheline_aligned_in_smp;
121 u32 tail ____cacheline_aligned_in_smp;
125 * This data is shared with the application through the mmap at offsets
126 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
128 * The offsets to the member fields are published through struct
129 * io_sqring_offsets when calling io_uring_setup.
133 * Head and tail offsets into the ring; the offsets need to be
134 * masked to get valid indices.
136 * The kernel controls head of the sq ring and the tail of the cq ring,
137 * and the application controls tail of the sq ring and the head of the
140 struct io_uring sq, cq;
142 * Bitmasks to apply to head and tail offsets (constant, equals
145 u32 sq_ring_mask, cq_ring_mask;
146 /* Ring sizes (constant, power of 2) */
147 u32 sq_ring_entries, cq_ring_entries;
149 * Number of invalid entries dropped by the kernel due to
150 * invalid index stored in array
152 * Written by the kernel, shouldn't be modified by the
153 * application (i.e. get number of "new events" by comparing to
156 * After a new SQ head value was read by the application this
157 * counter includes all submissions that were dropped reaching
158 * the new SQ head (and possibly more).
164 * Written by the kernel, shouldn't be modified by the
167 * The application needs a full memory barrier before checking
168 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
174 * Written by the application, shouldn't be modified by the
179 * Number of completion events lost because the queue was full;
180 * this should be avoided by the application by making sure
181 * there are not more requests pending than there is space in
182 * the completion queue.
184 * Written by the kernel, shouldn't be modified by the
185 * application (i.e. get number of "new events" by comparing to
188 * As completion events come in out of order this counter is not
189 * ordered with any other data.
193 * Ring buffer of completion events.
195 * The kernel writes completion events fresh every time they are
196 * produced, so the application is allowed to modify pending
199 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
202 enum io_uring_cmd_flags {
203 IO_URING_F_NONBLOCK = 1,
204 IO_URING_F_COMPLETE_DEFER = 2,
207 struct io_mapped_ubuf {
210 unsigned int nr_bvecs;
211 unsigned long acct_pages;
212 struct bio_vec bvec[];
217 struct io_overflow_cqe {
218 struct io_uring_cqe cqe;
219 struct list_head list;
222 struct io_fixed_file {
223 /* file * with additional FFS_* flags */
224 unsigned long file_ptr;
228 struct list_head list;
233 struct io_mapped_ubuf *buf;
237 struct io_file_table {
238 /* two level table */
239 struct io_fixed_file **files;
242 struct io_rsrc_node {
243 struct percpu_ref refs;
244 struct list_head node;
245 struct list_head rsrc_list;
246 struct io_rsrc_data *rsrc_data;
247 struct llist_node llist;
251 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
253 struct io_rsrc_data {
254 struct io_ring_ctx *ctx;
260 struct completion done;
265 struct list_head list;
271 struct io_restriction {
272 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
273 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
274 u8 sqe_flags_allowed;
275 u8 sqe_flags_required;
280 IO_SQ_THREAD_SHOULD_STOP = 0,
281 IO_SQ_THREAD_SHOULD_PARK,
286 atomic_t park_pending;
289 /* ctx's that are using this sqd */
290 struct list_head ctx_list;
292 struct task_struct *thread;
293 struct wait_queue_head wait;
295 unsigned sq_thread_idle;
301 struct completion exited;
304 #define IO_IOPOLL_BATCH 8
305 #define IO_COMPL_BATCH 32
306 #define IO_REQ_CACHE_SIZE 32
307 #define IO_REQ_ALLOC_BATCH 8
309 struct io_comp_state {
310 struct io_kiocb *reqs[IO_COMPL_BATCH];
312 /* inline/task_work completion list, under ->uring_lock */
313 struct list_head free_list;
316 struct io_submit_link {
317 struct io_kiocb *head;
318 struct io_kiocb *last;
321 struct io_submit_state {
322 struct blk_plug plug;
323 struct io_submit_link link;
326 * io_kiocb alloc cache
328 void *reqs[IO_REQ_CACHE_SIZE];
329 unsigned int free_reqs;
334 * Batch completion logic
336 struct io_comp_state comp;
339 * File reference cache
343 unsigned int file_refs;
344 unsigned int ios_left;
348 /* const or read-mostly hot data */
350 struct percpu_ref refs;
352 struct io_rings *rings;
354 unsigned int compat: 1;
355 unsigned int drain_next: 1;
356 unsigned int eventfd_async: 1;
357 unsigned int restricted: 1;
358 unsigned int off_timeout_used: 1;
359 unsigned int drain_active: 1;
360 } ____cacheline_aligned_in_smp;
362 /* submission data */
364 struct mutex uring_lock;
367 * Ring buffer of indices into array of io_uring_sqe, which is
368 * mmapped by the application using the IORING_OFF_SQES offset.
370 * This indirection could e.g. be used to assign fixed
371 * io_uring_sqe entries to operations and only submit them to
372 * the queue when needed.
374 * The kernel modifies neither the indices array nor the entries
378 struct io_uring_sqe *sq_sqes;
379 unsigned cached_sq_head;
381 struct list_head defer_list;
384 * Fixed resources fast path, should be accessed only under
385 * uring_lock, and updated through io_uring_register(2)
387 struct io_rsrc_node *rsrc_node;
388 struct io_file_table file_table;
389 unsigned nr_user_files;
390 unsigned nr_user_bufs;
391 struct io_mapped_ubuf **user_bufs;
393 struct io_submit_state submit_state;
394 struct list_head timeout_list;
395 struct list_head cq_overflow_list;
396 struct xarray io_buffers;
397 struct xarray personalities;
399 unsigned sq_thread_idle;
400 } ____cacheline_aligned_in_smp;
402 /* IRQ completion list, under ->completion_lock */
403 struct list_head locked_free_list;
404 unsigned int locked_free_nr;
406 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
407 struct io_sq_data *sq_data; /* if using sq thread polling */
409 struct wait_queue_head sqo_sq_wait;
410 struct list_head sqd_list;
412 unsigned long check_cq_overflow;
415 unsigned cached_cq_tail;
417 struct eventfd_ctx *cq_ev_fd;
418 struct wait_queue_head poll_wait;
419 struct wait_queue_head cq_wait;
421 atomic_t cq_timeouts;
422 struct fasync_struct *cq_fasync;
423 unsigned cq_last_tm_flush;
424 } ____cacheline_aligned_in_smp;
427 spinlock_t completion_lock;
430 * ->iopoll_list is protected by the ctx->uring_lock for
431 * io_uring instances that don't use IORING_SETUP_SQPOLL.
432 * For SQPOLL, only the single threaded io_sq_thread() will
433 * manipulate the list, hence no extra locking is needed there.
435 struct list_head iopoll_list;
436 struct hlist_head *cancel_hash;
437 unsigned cancel_hash_bits;
438 bool poll_multi_queue;
439 } ____cacheline_aligned_in_smp;
441 struct io_restriction restrictions;
443 /* slow path rsrc auxilary data, used by update/register */
445 struct io_rsrc_node *rsrc_backup_node;
446 struct io_mapped_ubuf *dummy_ubuf;
447 struct io_rsrc_data *file_data;
448 struct io_rsrc_data *buf_data;
450 struct delayed_work rsrc_put_work;
451 struct llist_head rsrc_put_llist;
452 struct list_head rsrc_ref_list;
453 spinlock_t rsrc_ref_lock;
456 /* Keep this last, we don't need it for the fast path */
458 #if defined(CONFIG_UNIX)
459 struct socket *ring_sock;
461 /* hashed buffered write serialization */
462 struct io_wq_hash *hash_map;
464 /* Only used for accounting purposes */
465 struct user_struct *user;
466 struct mm_struct *mm_account;
468 /* ctx exit and cancelation */
469 struct llist_head fallback_llist;
470 struct delayed_work fallback_work;
471 struct work_struct exit_work;
472 struct list_head tctx_list;
473 struct completion ref_comp;
477 struct io_uring_task {
478 /* submission side */
481 struct wait_queue_head wait;
482 const struct io_ring_ctx *last;
484 struct percpu_counter inflight;
485 atomic_t inflight_tracked;
488 spinlock_t task_lock;
489 struct io_wq_work_list task_list;
490 unsigned long task_state;
491 struct callback_head task_work;
495 * First field must be the file pointer in all the
496 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
498 struct io_poll_iocb {
500 struct wait_queue_head *head;
504 struct wait_queue_entry wait;
507 struct io_poll_update {
513 bool update_user_data;
521 struct io_timeout_data {
522 struct io_kiocb *req;
523 struct hrtimer timer;
524 struct timespec64 ts;
525 enum hrtimer_mode mode;
530 struct sockaddr __user *addr;
531 int __user *addr_len;
533 unsigned long nofile;
553 struct list_head list;
554 /* head of the link, used by linked timeouts only */
555 struct io_kiocb *head;
558 struct io_timeout_rem {
563 struct timespec64 ts;
568 /* NOTE: kiocb has the file as the first member, so don't do it here */
576 struct sockaddr __user *addr;
583 struct compat_msghdr __user *umsg_compat;
584 struct user_msghdr __user *umsg;
590 struct io_buffer *kbuf;
596 struct filename *filename;
598 unsigned long nofile;
601 struct io_rsrc_update {
627 struct epoll_event event;
631 struct file *file_out;
632 struct file *file_in;
639 struct io_provide_buf {
653 const char __user *filename;
654 struct statx __user *buffer;
666 struct filename *oldpath;
667 struct filename *newpath;
675 struct filename *filename;
678 struct io_completion {
680 struct list_head list;
684 struct io_async_connect {
685 struct sockaddr_storage address;
688 struct io_async_msghdr {
689 struct iovec fast_iov[UIO_FASTIOV];
690 /* points to an allocated iov, if NULL we use fast_iov instead */
691 struct iovec *free_iov;
692 struct sockaddr __user *uaddr;
694 struct sockaddr_storage addr;
698 struct iovec fast_iov[UIO_FASTIOV];
699 const struct iovec *free_iovec;
700 struct iov_iter iter;
702 struct wait_page_queue wpq;
706 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
707 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
708 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
709 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
710 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
711 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
713 /* first byte is taken by user flags, shift it to not overlap */
718 REQ_F_LINK_TIMEOUT_BIT,
719 REQ_F_NEED_CLEANUP_BIT,
721 REQ_F_BUFFER_SELECTED_BIT,
722 REQ_F_LTIMEOUT_ACTIVE_BIT,
723 REQ_F_COMPLETE_INLINE_BIT,
725 REQ_F_DONT_REISSUE_BIT,
727 /* keep async read/write and isreg together and in order */
728 REQ_F_ASYNC_READ_BIT,
729 REQ_F_ASYNC_WRITE_BIT,
732 /* not a real bit, just to check we're not overflowing the space */
738 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
739 /* drain existing IO first */
740 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
742 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
743 /* doesn't sever on completion < 0 */
744 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
746 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
747 /* IOSQE_BUFFER_SELECT */
748 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
750 /* fail rest of links */
751 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
752 /* on inflight list, should be cancelled and waited on exit reliably */
753 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
754 /* read/write uses file position */
755 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
756 /* must not punt to workers */
757 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
758 /* has or had linked timeout */
759 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
761 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
762 /* already went through poll handler */
763 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
764 /* buffer already selected */
765 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
766 /* linked timeout is active, i.e. prepared by link's head */
767 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
768 /* completion is deferred through io_comp_state */
769 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
770 /* caller should reissue async */
771 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
772 /* don't attempt request reissue, see io_rw_reissue() */
773 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
774 /* supports async reads */
775 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
776 /* supports async writes */
777 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
779 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
780 /* has creds assigned */
781 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
785 struct io_poll_iocb poll;
786 struct io_poll_iocb *double_poll;
789 typedef void (*io_req_tw_func_t)(struct io_kiocb *req);
791 struct io_task_work {
793 struct io_wq_work_node node;
794 struct llist_node fallback_node;
796 io_req_tw_func_t func;
800 IORING_RSRC_FILE = 0,
801 IORING_RSRC_BUFFER = 1,
805 * NOTE! Each of the iocb union members has the file pointer
806 * as the first entry in their struct definition. So you can
807 * access the file pointer through any of the sub-structs,
808 * or directly as just 'ki_filp' in this struct.
814 struct io_poll_iocb poll;
815 struct io_poll_update poll_update;
816 struct io_accept accept;
818 struct io_cancel cancel;
819 struct io_timeout timeout;
820 struct io_timeout_rem timeout_rem;
821 struct io_connect connect;
822 struct io_sr_msg sr_msg;
824 struct io_close close;
825 struct io_rsrc_update rsrc_update;
826 struct io_fadvise fadvise;
827 struct io_madvise madvise;
828 struct io_epoll epoll;
829 struct io_splice splice;
830 struct io_provide_buf pbuf;
831 struct io_statx statx;
832 struct io_shutdown shutdown;
833 struct io_rename rename;
834 struct io_unlink unlink;
835 /* use only after cleaning per-op data, see io_clean_op() */
836 struct io_completion compl;
839 /* opcode allocated if it needs to store data for async defer */
842 /* polled IO has completed */
848 struct io_ring_ctx *ctx;
851 struct task_struct *task;
854 struct io_kiocb *link;
855 struct percpu_ref *fixed_rsrc_refs;
857 /* used with ctx->iopoll_list with reads/writes */
858 struct list_head inflight_entry;
859 struct io_task_work io_task_work;
860 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
861 struct hlist_node hash_node;
862 struct async_poll *apoll;
863 struct io_wq_work work;
864 const struct cred *creds;
866 /* store used ubuf, so we can prevent reloading */
867 struct io_mapped_ubuf *imu;
870 struct io_tctx_node {
871 struct list_head ctx_node;
872 struct task_struct *task;
873 struct io_ring_ctx *ctx;
876 struct io_defer_entry {
877 struct list_head list;
878 struct io_kiocb *req;
883 /* needs req->file assigned */
884 unsigned needs_file : 1;
885 /* hash wq insertion if file is a regular file */
886 unsigned hash_reg_file : 1;
887 /* unbound wq insertion if file is a non-regular file */
888 unsigned unbound_nonreg_file : 1;
889 /* opcode is not supported by this kernel */
890 unsigned not_supported : 1;
891 /* set if opcode supports polled "wait" */
893 unsigned pollout : 1;
894 /* op supports buffer selection */
895 unsigned buffer_select : 1;
896 /* do prep async if is going to be punted */
897 unsigned needs_async_setup : 1;
898 /* should block plug */
900 /* size of async data needed, if any */
901 unsigned short async_size;
904 static const struct io_op_def io_op_defs[] = {
905 [IORING_OP_NOP] = {},
906 [IORING_OP_READV] = {
908 .unbound_nonreg_file = 1,
911 .needs_async_setup = 1,
913 .async_size = sizeof(struct io_async_rw),
915 [IORING_OP_WRITEV] = {
918 .unbound_nonreg_file = 1,
920 .needs_async_setup = 1,
922 .async_size = sizeof(struct io_async_rw),
924 [IORING_OP_FSYNC] = {
927 [IORING_OP_READ_FIXED] = {
929 .unbound_nonreg_file = 1,
932 .async_size = sizeof(struct io_async_rw),
934 [IORING_OP_WRITE_FIXED] = {
937 .unbound_nonreg_file = 1,
940 .async_size = sizeof(struct io_async_rw),
942 [IORING_OP_POLL_ADD] = {
944 .unbound_nonreg_file = 1,
946 [IORING_OP_POLL_REMOVE] = {},
947 [IORING_OP_SYNC_FILE_RANGE] = {
950 [IORING_OP_SENDMSG] = {
952 .unbound_nonreg_file = 1,
954 .needs_async_setup = 1,
955 .async_size = sizeof(struct io_async_msghdr),
957 [IORING_OP_RECVMSG] = {
959 .unbound_nonreg_file = 1,
962 .needs_async_setup = 1,
963 .async_size = sizeof(struct io_async_msghdr),
965 [IORING_OP_TIMEOUT] = {
966 .async_size = sizeof(struct io_timeout_data),
968 [IORING_OP_TIMEOUT_REMOVE] = {
969 /* used by timeout updates' prep() */
971 [IORING_OP_ACCEPT] = {
973 .unbound_nonreg_file = 1,
976 [IORING_OP_ASYNC_CANCEL] = {},
977 [IORING_OP_LINK_TIMEOUT] = {
978 .async_size = sizeof(struct io_timeout_data),
980 [IORING_OP_CONNECT] = {
982 .unbound_nonreg_file = 1,
984 .needs_async_setup = 1,
985 .async_size = sizeof(struct io_async_connect),
987 [IORING_OP_FALLOCATE] = {
990 [IORING_OP_OPENAT] = {},
991 [IORING_OP_CLOSE] = {},
992 [IORING_OP_FILES_UPDATE] = {},
993 [IORING_OP_STATX] = {},
996 .unbound_nonreg_file = 1,
1000 .async_size = sizeof(struct io_async_rw),
1002 [IORING_OP_WRITE] = {
1004 .unbound_nonreg_file = 1,
1007 .async_size = sizeof(struct io_async_rw),
1009 [IORING_OP_FADVISE] = {
1012 [IORING_OP_MADVISE] = {},
1013 [IORING_OP_SEND] = {
1015 .unbound_nonreg_file = 1,
1018 [IORING_OP_RECV] = {
1020 .unbound_nonreg_file = 1,
1024 [IORING_OP_OPENAT2] = {
1026 [IORING_OP_EPOLL_CTL] = {
1027 .unbound_nonreg_file = 1,
1029 [IORING_OP_SPLICE] = {
1032 .unbound_nonreg_file = 1,
1034 [IORING_OP_PROVIDE_BUFFERS] = {},
1035 [IORING_OP_REMOVE_BUFFERS] = {},
1039 .unbound_nonreg_file = 1,
1041 [IORING_OP_SHUTDOWN] = {
1044 [IORING_OP_RENAMEAT] = {},
1045 [IORING_OP_UNLINKAT] = {},
1048 static bool io_disarm_next(struct io_kiocb *req);
1049 static void io_uring_del_tctx_node(unsigned long index);
1050 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1051 struct task_struct *task,
1053 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1054 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1056 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1057 long res, unsigned int cflags);
1058 static void io_put_req(struct io_kiocb *req);
1059 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1060 static void io_dismantle_req(struct io_kiocb *req);
1061 static void io_put_task(struct task_struct *task, int nr);
1062 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1063 static void io_queue_linked_timeout(struct io_kiocb *req);
1064 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1065 struct io_uring_rsrc_update2 *up,
1067 static void io_clean_op(struct io_kiocb *req);
1068 static struct file *io_file_get(struct io_submit_state *state,
1069 struct io_kiocb *req, int fd, bool fixed);
1070 static void __io_queue_sqe(struct io_kiocb *req);
1071 static void io_rsrc_put_work(struct work_struct *work);
1073 static void io_req_task_queue(struct io_kiocb *req);
1074 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1075 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1076 static int io_req_prep_async(struct io_kiocb *req);
1078 static void io_fallback_req_func(struct work_struct *unused);
1080 static struct kmem_cache *req_cachep;
1082 static const struct file_operations io_uring_fops;
1084 struct sock *io_uring_get_socket(struct file *file)
1086 #if defined(CONFIG_UNIX)
1087 if (file->f_op == &io_uring_fops) {
1088 struct io_ring_ctx *ctx = file->private_data;
1090 return ctx->ring_sock->sk;
1095 EXPORT_SYMBOL(io_uring_get_socket);
1097 #define io_for_each_link(pos, head) \
1098 for (pos = (head); pos; pos = pos->link)
1100 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1102 struct io_ring_ctx *ctx = req->ctx;
1104 if (!req->fixed_rsrc_refs) {
1105 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1106 percpu_ref_get(req->fixed_rsrc_refs);
1110 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1112 bool got = percpu_ref_tryget(ref);
1114 /* already at zero, wait for ->release() */
1116 wait_for_completion(compl);
1117 percpu_ref_resurrect(ref);
1119 percpu_ref_put(ref);
1122 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1125 struct io_kiocb *req;
1127 if (task && head->task != task)
1132 io_for_each_link(req, head) {
1133 if (req->flags & REQ_F_INFLIGHT)
1139 static inline void req_set_fail(struct io_kiocb *req)
1141 req->flags |= REQ_F_FAIL;
1144 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1146 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1148 complete(&ctx->ref_comp);
1151 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1153 return !req->timeout.off;
1156 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1158 struct io_ring_ctx *ctx;
1161 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1166 * Use 5 bits less than the max cq entries, that should give us around
1167 * 32 entries per hash list if totally full and uniformly spread.
1169 hash_bits = ilog2(p->cq_entries);
1173 ctx->cancel_hash_bits = hash_bits;
1174 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1176 if (!ctx->cancel_hash)
1178 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1180 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1181 if (!ctx->dummy_ubuf)
1183 /* set invalid range, so io_import_fixed() fails meeting it */
1184 ctx->dummy_ubuf->ubuf = -1UL;
1186 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1187 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1190 ctx->flags = p->flags;
1191 init_waitqueue_head(&ctx->sqo_sq_wait);
1192 INIT_LIST_HEAD(&ctx->sqd_list);
1193 init_waitqueue_head(&ctx->poll_wait);
1194 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1195 init_completion(&ctx->ref_comp);
1196 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1197 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1198 mutex_init(&ctx->uring_lock);
1199 init_waitqueue_head(&ctx->cq_wait);
1200 spin_lock_init(&ctx->completion_lock);
1201 INIT_LIST_HEAD(&ctx->iopoll_list);
1202 INIT_LIST_HEAD(&ctx->defer_list);
1203 INIT_LIST_HEAD(&ctx->timeout_list);
1204 spin_lock_init(&ctx->rsrc_ref_lock);
1205 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1206 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1207 init_llist_head(&ctx->rsrc_put_llist);
1208 INIT_LIST_HEAD(&ctx->tctx_list);
1209 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1210 INIT_LIST_HEAD(&ctx->locked_free_list);
1211 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1214 kfree(ctx->dummy_ubuf);
1215 kfree(ctx->cancel_hash);
1220 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1222 struct io_rings *r = ctx->rings;
1224 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1228 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1230 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1231 struct io_ring_ctx *ctx = req->ctx;
1233 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1239 static void io_req_track_inflight(struct io_kiocb *req)
1241 if (!(req->flags & REQ_F_INFLIGHT)) {
1242 req->flags |= REQ_F_INFLIGHT;
1243 atomic_inc(¤t->io_uring->inflight_tracked);
1247 static void io_prep_async_work(struct io_kiocb *req)
1249 const struct io_op_def *def = &io_op_defs[req->opcode];
1250 struct io_ring_ctx *ctx = req->ctx;
1252 if (!(req->flags & REQ_F_CREDS)) {
1253 req->flags |= REQ_F_CREDS;
1254 req->creds = get_current_cred();
1257 req->work.list.next = NULL;
1258 req->work.flags = 0;
1259 if (req->flags & REQ_F_FORCE_ASYNC)
1260 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1262 if (req->flags & REQ_F_ISREG) {
1263 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1264 io_wq_hash_work(&req->work, file_inode(req->file));
1265 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1266 if (def->unbound_nonreg_file)
1267 req->work.flags |= IO_WQ_WORK_UNBOUND;
1270 switch (req->opcode) {
1271 case IORING_OP_SPLICE:
1273 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1274 req->work.flags |= IO_WQ_WORK_UNBOUND;
1279 static void io_prep_async_link(struct io_kiocb *req)
1281 struct io_kiocb *cur;
1283 if (req->flags & REQ_F_LINK_TIMEOUT) {
1284 struct io_ring_ctx *ctx = req->ctx;
1286 spin_lock_irq(&ctx->completion_lock);
1287 io_for_each_link(cur, req)
1288 io_prep_async_work(cur);
1289 spin_unlock_irq(&ctx->completion_lock);
1291 io_for_each_link(cur, req)
1292 io_prep_async_work(cur);
1296 static void io_queue_async_work(struct io_kiocb *req)
1298 struct io_ring_ctx *ctx = req->ctx;
1299 struct io_kiocb *link = io_prep_linked_timeout(req);
1300 struct io_uring_task *tctx = req->task->io_uring;
1303 BUG_ON(!tctx->io_wq);
1305 /* init ->work of the whole link before punting */
1306 io_prep_async_link(req);
1309 * Not expected to happen, but if we do have a bug where this _can_
1310 * happen, catch it here and ensure the request is marked as
1311 * canceled. That will make io-wq go through the usual work cancel
1312 * procedure rather than attempt to run this request (or create a new
1315 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1316 req->work.flags |= IO_WQ_WORK_CANCEL;
1318 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1319 &req->work, req->flags);
1320 io_wq_enqueue(tctx->io_wq, &req->work);
1322 io_queue_linked_timeout(link);
1325 static void io_kill_timeout(struct io_kiocb *req, int status)
1326 __must_hold(&req->ctx->completion_lock)
1328 struct io_timeout_data *io = req->async_data;
1330 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1331 atomic_set(&req->ctx->cq_timeouts,
1332 atomic_read(&req->ctx->cq_timeouts) + 1);
1333 list_del_init(&req->timeout.list);
1334 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1335 io_put_req_deferred(req, 1);
1339 static void io_queue_deferred(struct io_ring_ctx *ctx)
1341 while (!list_empty(&ctx->defer_list)) {
1342 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1343 struct io_defer_entry, list);
1345 if (req_need_defer(de->req, de->seq))
1347 list_del_init(&de->list);
1348 io_req_task_queue(de->req);
1353 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1355 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1357 while (!list_empty(&ctx->timeout_list)) {
1358 u32 events_needed, events_got;
1359 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1360 struct io_kiocb, timeout.list);
1362 if (io_is_timeout_noseq(req))
1366 * Since seq can easily wrap around over time, subtract
1367 * the last seq at which timeouts were flushed before comparing.
1368 * Assuming not more than 2^31-1 events have happened since,
1369 * these subtractions won't have wrapped, so we can check if
1370 * target is in [last_seq, current_seq] by comparing the two.
1372 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1373 events_got = seq - ctx->cq_last_tm_flush;
1374 if (events_got < events_needed)
1377 list_del_init(&req->timeout.list);
1378 io_kill_timeout(req, 0);
1380 ctx->cq_last_tm_flush = seq;
1383 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1385 if (ctx->off_timeout_used)
1386 io_flush_timeouts(ctx);
1387 if (ctx->drain_active)
1388 io_queue_deferred(ctx);
1391 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1393 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1394 __io_commit_cqring_flush(ctx);
1395 /* order cqe stores with ring update */
1396 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1399 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1401 struct io_rings *r = ctx->rings;
1403 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1406 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1408 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1411 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1413 struct io_rings *rings = ctx->rings;
1414 unsigned tail, mask = ctx->cq_entries - 1;
1417 * writes to the cq entry need to come after reading head; the
1418 * control dependency is enough as we're using WRITE_ONCE to
1421 if (__io_cqring_events(ctx) == ctx->cq_entries)
1424 tail = ctx->cached_cq_tail++;
1425 return &rings->cqes[tail & mask];
1428 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1430 if (likely(!ctx->cq_ev_fd))
1432 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1434 return !ctx->eventfd_async || io_wq_current_is_worker();
1437 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1439 /* see waitqueue_active() comment */
1442 if (waitqueue_active(&ctx->cq_wait))
1443 wake_up(&ctx->cq_wait);
1444 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1445 wake_up(&ctx->sq_data->wait);
1446 if (io_should_trigger_evfd(ctx))
1447 eventfd_signal(ctx->cq_ev_fd, 1);
1448 if (waitqueue_active(&ctx->poll_wait)) {
1449 wake_up_interruptible(&ctx->poll_wait);
1450 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1454 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1456 /* see waitqueue_active() comment */
1459 if (ctx->flags & IORING_SETUP_SQPOLL) {
1460 if (waitqueue_active(&ctx->cq_wait))
1461 wake_up(&ctx->cq_wait);
1463 if (io_should_trigger_evfd(ctx))
1464 eventfd_signal(ctx->cq_ev_fd, 1);
1465 if (waitqueue_active(&ctx->poll_wait)) {
1466 wake_up_interruptible(&ctx->poll_wait);
1467 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1471 /* Returns true if there are no backlogged entries after the flush */
1472 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1474 unsigned long flags;
1475 bool all_flushed, posted;
1477 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1481 spin_lock_irqsave(&ctx->completion_lock, flags);
1482 while (!list_empty(&ctx->cq_overflow_list)) {
1483 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1484 struct io_overflow_cqe *ocqe;
1488 ocqe = list_first_entry(&ctx->cq_overflow_list,
1489 struct io_overflow_cqe, list);
1491 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1493 io_account_cq_overflow(ctx);
1496 list_del(&ocqe->list);
1500 all_flushed = list_empty(&ctx->cq_overflow_list);
1502 clear_bit(0, &ctx->check_cq_overflow);
1503 WRITE_ONCE(ctx->rings->sq_flags,
1504 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1508 io_commit_cqring(ctx);
1509 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1511 io_cqring_ev_posted(ctx);
1515 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1519 if (test_bit(0, &ctx->check_cq_overflow)) {
1520 /* iopoll syncs against uring_lock, not completion_lock */
1521 if (ctx->flags & IORING_SETUP_IOPOLL)
1522 mutex_lock(&ctx->uring_lock);
1523 ret = __io_cqring_overflow_flush(ctx, force);
1524 if (ctx->flags & IORING_SETUP_IOPOLL)
1525 mutex_unlock(&ctx->uring_lock);
1532 * Shamelessly stolen from the mm implementation of page reference checking,
1533 * see commit f958d7b528b1 for details.
1535 #define req_ref_zero_or_close_to_overflow(req) \
1536 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1538 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1540 return atomic_inc_not_zero(&req->refs);
1543 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1545 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1546 return atomic_sub_and_test(refs, &req->refs);
1549 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1551 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1552 return atomic_dec_and_test(&req->refs);
1555 static inline void req_ref_put(struct io_kiocb *req)
1557 WARN_ON_ONCE(req_ref_put_and_test(req));
1560 static inline void req_ref_get(struct io_kiocb *req)
1562 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1563 atomic_inc(&req->refs);
1566 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1567 long res, unsigned int cflags)
1569 struct io_overflow_cqe *ocqe;
1571 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1574 * If we're in ring overflow flush mode, or in task cancel mode,
1575 * or cannot allocate an overflow entry, then we need to drop it
1578 io_account_cq_overflow(ctx);
1581 if (list_empty(&ctx->cq_overflow_list)) {
1582 set_bit(0, &ctx->check_cq_overflow);
1583 WRITE_ONCE(ctx->rings->sq_flags,
1584 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1587 ocqe->cqe.user_data = user_data;
1588 ocqe->cqe.res = res;
1589 ocqe->cqe.flags = cflags;
1590 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1594 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1595 long res, unsigned int cflags)
1597 struct io_uring_cqe *cqe;
1599 trace_io_uring_complete(ctx, user_data, res, cflags);
1602 * If we can't get a cq entry, userspace overflowed the
1603 * submission (by quite a lot). Increment the overflow count in
1606 cqe = io_get_cqe(ctx);
1608 WRITE_ONCE(cqe->user_data, user_data);
1609 WRITE_ONCE(cqe->res, res);
1610 WRITE_ONCE(cqe->flags, cflags);
1613 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1616 /* not as hot to bloat with inlining */
1617 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1618 long res, unsigned int cflags)
1620 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1623 static void io_req_complete_post(struct io_kiocb *req, long res,
1624 unsigned int cflags)
1626 struct io_ring_ctx *ctx = req->ctx;
1627 unsigned long flags;
1629 spin_lock_irqsave(&ctx->completion_lock, flags);
1630 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1632 * If we're the last reference to this request, add to our locked
1635 if (req_ref_put_and_test(req)) {
1636 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1637 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1638 io_disarm_next(req);
1640 io_req_task_queue(req->link);
1644 io_dismantle_req(req);
1645 io_put_task(req->task, 1);
1646 list_add(&req->compl.list, &ctx->locked_free_list);
1647 ctx->locked_free_nr++;
1649 if (!percpu_ref_tryget(&ctx->refs))
1652 io_commit_cqring(ctx);
1653 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1656 io_cqring_ev_posted(ctx);
1657 percpu_ref_put(&ctx->refs);
1661 static inline bool io_req_needs_clean(struct io_kiocb *req)
1663 return req->flags & IO_REQ_CLEAN_FLAGS;
1666 static void io_req_complete_state(struct io_kiocb *req, long res,
1667 unsigned int cflags)
1669 if (io_req_needs_clean(req))
1672 req->compl.cflags = cflags;
1673 req->flags |= REQ_F_COMPLETE_INLINE;
1676 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1677 long res, unsigned cflags)
1679 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1680 io_req_complete_state(req, res, cflags);
1682 io_req_complete_post(req, res, cflags);
1685 static inline void io_req_complete(struct io_kiocb *req, long res)
1687 __io_req_complete(req, 0, res, 0);
1690 static void io_req_complete_failed(struct io_kiocb *req, long res)
1694 io_req_complete_post(req, res, 0);
1697 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1698 struct io_comp_state *cs)
1700 spin_lock_irq(&ctx->completion_lock);
1701 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1702 ctx->locked_free_nr = 0;
1703 spin_unlock_irq(&ctx->completion_lock);
1706 /* Returns true IFF there are requests in the cache */
1707 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1709 struct io_submit_state *state = &ctx->submit_state;
1710 struct io_comp_state *cs = &state->comp;
1714 * If we have more than a batch's worth of requests in our IRQ side
1715 * locked cache, grab the lock and move them over to our submission
1718 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1719 io_flush_cached_locked_reqs(ctx, cs);
1721 nr = state->free_reqs;
1722 while (!list_empty(&cs->free_list)) {
1723 struct io_kiocb *req = list_first_entry(&cs->free_list,
1724 struct io_kiocb, compl.list);
1726 list_del(&req->compl.list);
1727 state->reqs[nr++] = req;
1728 if (nr == ARRAY_SIZE(state->reqs))
1732 state->free_reqs = nr;
1736 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1738 struct io_submit_state *state = &ctx->submit_state;
1740 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1742 if (!state->free_reqs) {
1743 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1746 if (io_flush_cached_reqs(ctx))
1749 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1753 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1754 * retry single alloc to be on the safe side.
1756 if (unlikely(ret <= 0)) {
1757 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1758 if (!state->reqs[0])
1764 * Don't initialise the fields below on every allocation, but
1765 * do that in advance and keep valid on free.
1767 for (i = 0; i < ret; i++) {
1768 struct io_kiocb *req = state->reqs[i];
1772 req->async_data = NULL;
1773 /* not necessary, but safer to zero */
1776 state->free_reqs = ret;
1780 return state->reqs[state->free_reqs];
1783 static inline void io_put_file(struct file *file)
1789 static void io_dismantle_req(struct io_kiocb *req)
1791 unsigned int flags = req->flags;
1793 if (io_req_needs_clean(req))
1795 if (!(flags & REQ_F_FIXED_FILE))
1796 io_put_file(req->file);
1797 if (req->fixed_rsrc_refs)
1798 percpu_ref_put(req->fixed_rsrc_refs);
1799 if (req->async_data) {
1800 kfree(req->async_data);
1801 req->async_data = NULL;
1805 /* must to be called somewhat shortly after putting a request */
1806 static inline void io_put_task(struct task_struct *task, int nr)
1808 struct io_uring_task *tctx = task->io_uring;
1810 percpu_counter_sub(&tctx->inflight, nr);
1811 if (unlikely(atomic_read(&tctx->in_idle)))
1812 wake_up(&tctx->wait);
1813 put_task_struct_many(task, nr);
1816 static void __io_free_req(struct io_kiocb *req)
1818 struct io_ring_ctx *ctx = req->ctx;
1820 io_dismantle_req(req);
1821 io_put_task(req->task, 1);
1823 kmem_cache_free(req_cachep, req);
1824 percpu_ref_put(&ctx->refs);
1827 static inline void io_remove_next_linked(struct io_kiocb *req)
1829 struct io_kiocb *nxt = req->link;
1831 req->link = nxt->link;
1835 static bool io_kill_linked_timeout(struct io_kiocb *req)
1836 __must_hold(&req->ctx->completion_lock)
1838 struct io_kiocb *link = req->link;
1841 * Can happen if a linked timeout fired and link had been like
1842 * req -> link t-out -> link t-out [-> ...]
1844 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1845 struct io_timeout_data *io = link->async_data;
1847 io_remove_next_linked(req);
1848 link->timeout.head = NULL;
1849 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1850 io_cqring_fill_event(link->ctx, link->user_data,
1852 io_put_req_deferred(link, 1);
1859 static void io_fail_links(struct io_kiocb *req)
1860 __must_hold(&req->ctx->completion_lock)
1862 struct io_kiocb *nxt, *link = req->link;
1869 trace_io_uring_fail_link(req, link);
1870 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1871 io_put_req_deferred(link, 2);
1876 static bool io_disarm_next(struct io_kiocb *req)
1877 __must_hold(&req->ctx->completion_lock)
1879 bool posted = false;
1881 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1882 posted = io_kill_linked_timeout(req);
1883 if (unlikely((req->flags & REQ_F_FAIL) &&
1884 !(req->flags & REQ_F_HARDLINK))) {
1885 posted |= (req->link != NULL);
1891 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1893 struct io_kiocb *nxt;
1896 * If LINK is set, we have dependent requests in this chain. If we
1897 * didn't fail this request, queue the first one up, moving any other
1898 * dependencies to the next request. In case of failure, fail the rest
1901 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1902 struct io_ring_ctx *ctx = req->ctx;
1903 unsigned long flags;
1906 spin_lock_irqsave(&ctx->completion_lock, flags);
1907 posted = io_disarm_next(req);
1909 io_commit_cqring(req->ctx);
1910 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1912 io_cqring_ev_posted(ctx);
1919 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1921 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1923 return __io_req_find_next(req);
1926 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1930 if (ctx->submit_state.comp.nr) {
1931 mutex_lock(&ctx->uring_lock);
1932 io_submit_flush_completions(ctx);
1933 mutex_unlock(&ctx->uring_lock);
1935 percpu_ref_put(&ctx->refs);
1938 static void tctx_task_work(struct callback_head *cb)
1940 struct io_ring_ctx *ctx = NULL;
1941 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1945 struct io_wq_work_node *node;
1947 spin_lock_irq(&tctx->task_lock);
1948 node = tctx->task_list.first;
1949 INIT_WQ_LIST(&tctx->task_list);
1950 spin_unlock_irq(&tctx->task_lock);
1953 struct io_wq_work_node *next = node->next;
1954 struct io_kiocb *req = container_of(node, struct io_kiocb,
1957 if (req->ctx != ctx) {
1958 ctx_flush_and_put(ctx);
1960 percpu_ref_get(&ctx->refs);
1962 req->io_task_work.func(req);
1965 if (wq_list_empty(&tctx->task_list)) {
1966 spin_lock_irq(&tctx->task_lock);
1967 clear_bit(0, &tctx->task_state);
1968 if (wq_list_empty(&tctx->task_list)) {
1969 spin_unlock_irq(&tctx->task_lock);
1972 spin_unlock_irq(&tctx->task_lock);
1973 /* another tctx_task_work() is enqueued, yield */
1974 if (test_and_set_bit(0, &tctx->task_state))
1980 ctx_flush_and_put(ctx);
1983 static void io_req_task_work_add(struct io_kiocb *req)
1985 struct task_struct *tsk = req->task;
1986 struct io_uring_task *tctx = tsk->io_uring;
1987 enum task_work_notify_mode notify;
1988 struct io_wq_work_node *node;
1989 unsigned long flags;
1991 WARN_ON_ONCE(!tctx);
1993 spin_lock_irqsave(&tctx->task_lock, flags);
1994 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1995 spin_unlock_irqrestore(&tctx->task_lock, flags);
1997 /* task_work already pending, we're done */
1998 if (test_bit(0, &tctx->task_state) ||
1999 test_and_set_bit(0, &tctx->task_state))
2001 if (unlikely(tsk->flags & PF_EXITING))
2005 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2006 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2007 * processing task_work. There's no reliable way to tell if TWA_RESUME
2010 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2011 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2012 wake_up_process(tsk);
2016 clear_bit(0, &tctx->task_state);
2017 spin_lock_irqsave(&tctx->task_lock, flags);
2018 node = tctx->task_list.first;
2019 INIT_WQ_LIST(&tctx->task_list);
2020 spin_unlock_irqrestore(&tctx->task_lock, flags);
2023 req = container_of(node, struct io_kiocb, io_task_work.node);
2025 if (llist_add(&req->io_task_work.fallback_node,
2026 &req->ctx->fallback_llist))
2027 schedule_delayed_work(&req->ctx->fallback_work, 1);
2031 static void io_req_task_cancel(struct io_kiocb *req)
2033 struct io_ring_ctx *ctx = req->ctx;
2035 /* ctx is guaranteed to stay alive while we hold uring_lock */
2036 mutex_lock(&ctx->uring_lock);
2037 io_req_complete_failed(req, req->result);
2038 mutex_unlock(&ctx->uring_lock);
2041 static void io_req_task_submit(struct io_kiocb *req)
2043 struct io_ring_ctx *ctx = req->ctx;
2045 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2046 mutex_lock(&ctx->uring_lock);
2047 if (!(req->task->flags & PF_EXITING) && !req->task->in_execve)
2048 __io_queue_sqe(req);
2050 io_req_complete_failed(req, -EFAULT);
2051 mutex_unlock(&ctx->uring_lock);
2054 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2057 req->io_task_work.func = io_req_task_cancel;
2058 io_req_task_work_add(req);
2061 static void io_req_task_queue(struct io_kiocb *req)
2063 req->io_task_work.func = io_req_task_submit;
2064 io_req_task_work_add(req);
2067 static void io_req_task_queue_reissue(struct io_kiocb *req)
2069 req->io_task_work.func = io_queue_async_work;
2070 io_req_task_work_add(req);
2073 static inline void io_queue_next(struct io_kiocb *req)
2075 struct io_kiocb *nxt = io_req_find_next(req);
2078 io_req_task_queue(nxt);
2081 static void io_free_req(struct io_kiocb *req)
2088 struct task_struct *task;
2093 static inline void io_init_req_batch(struct req_batch *rb)
2100 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2101 struct req_batch *rb)
2104 io_put_task(rb->task, rb->task_refs);
2106 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2109 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2110 struct io_submit_state *state)
2113 io_dismantle_req(req);
2115 if (req->task != rb->task) {
2117 io_put_task(rb->task, rb->task_refs);
2118 rb->task = req->task;
2124 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2125 state->reqs[state->free_reqs++] = req;
2127 list_add(&req->compl.list, &state->comp.free_list);
2130 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2132 struct io_comp_state *cs = &ctx->submit_state.comp;
2134 struct req_batch rb;
2136 spin_lock_irq(&ctx->completion_lock);
2137 for (i = 0; i < nr; i++) {
2138 struct io_kiocb *req = cs->reqs[i];
2140 __io_cqring_fill_event(ctx, req->user_data, req->result,
2143 io_commit_cqring(ctx);
2144 spin_unlock_irq(&ctx->completion_lock);
2145 io_cqring_ev_posted(ctx);
2147 io_init_req_batch(&rb);
2148 for (i = 0; i < nr; i++) {
2149 struct io_kiocb *req = cs->reqs[i];
2151 /* submission and completion refs */
2152 if (req_ref_sub_and_test(req, 2))
2153 io_req_free_batch(&rb, req, &ctx->submit_state);
2156 io_req_free_batch_finish(ctx, &rb);
2161 * Drop reference to request, return next in chain (if there is one) if this
2162 * was the last reference to this request.
2164 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2166 struct io_kiocb *nxt = NULL;
2168 if (req_ref_put_and_test(req)) {
2169 nxt = io_req_find_next(req);
2175 static inline void io_put_req(struct io_kiocb *req)
2177 if (req_ref_put_and_test(req))
2181 static void io_free_req_deferred(struct io_kiocb *req)
2183 req->io_task_work.func = io_free_req;
2184 io_req_task_work_add(req);
2187 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2189 if (req_ref_sub_and_test(req, refs))
2190 io_free_req_deferred(req);
2193 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2195 /* See comment at the top of this file */
2197 return __io_cqring_events(ctx);
2200 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2202 struct io_rings *rings = ctx->rings;
2204 /* make sure SQ entry isn't read before tail */
2205 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2208 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2210 unsigned int cflags;
2212 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2213 cflags |= IORING_CQE_F_BUFFER;
2214 req->flags &= ~REQ_F_BUFFER_SELECTED;
2219 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2221 struct io_buffer *kbuf;
2223 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2224 return io_put_kbuf(req, kbuf);
2227 static inline bool io_run_task_work(void)
2229 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2230 __set_current_state(TASK_RUNNING);
2231 tracehook_notify_signal();
2239 * Find and free completed poll iocbs
2241 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2242 struct list_head *done, bool resubmit)
2244 struct req_batch rb;
2245 struct io_kiocb *req;
2247 /* order with ->result store in io_complete_rw_iopoll() */
2250 io_init_req_batch(&rb);
2251 while (!list_empty(done)) {
2254 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2255 list_del(&req->inflight_entry);
2257 if (READ_ONCE(req->result) == -EAGAIN && resubmit &&
2258 !(req->flags & REQ_F_DONT_REISSUE)) {
2259 req->iopoll_completed = 0;
2261 io_req_task_queue_reissue(req);
2265 if (req->flags & REQ_F_BUFFER_SELECTED)
2266 cflags = io_put_rw_kbuf(req);
2268 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2271 if (req_ref_put_and_test(req))
2272 io_req_free_batch(&rb, req, &ctx->submit_state);
2275 io_commit_cqring(ctx);
2276 io_cqring_ev_posted_iopoll(ctx);
2277 io_req_free_batch_finish(ctx, &rb);
2280 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2281 long min, bool resubmit)
2283 struct io_kiocb *req, *tmp;
2289 * Only spin for completions if we don't have multiple devices hanging
2290 * off our complete list, and we're under the requested amount.
2292 spin = !ctx->poll_multi_queue && *nr_events < min;
2295 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2296 struct kiocb *kiocb = &req->rw.kiocb;
2299 * Move completed and retryable entries to our local lists.
2300 * If we find a request that requires polling, break out
2301 * and complete those lists first, if we have entries there.
2303 if (READ_ONCE(req->iopoll_completed)) {
2304 list_move_tail(&req->inflight_entry, &done);
2307 if (!list_empty(&done))
2310 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2314 /* iopoll may have completed current req */
2315 if (READ_ONCE(req->iopoll_completed))
2316 list_move_tail(&req->inflight_entry, &done);
2323 if (!list_empty(&done))
2324 io_iopoll_complete(ctx, nr_events, &done, resubmit);
2330 * We can't just wait for polled events to come to us, we have to actively
2331 * find and complete them.
2333 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2335 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2338 mutex_lock(&ctx->uring_lock);
2339 while (!list_empty(&ctx->iopoll_list)) {
2340 unsigned int nr_events = 0;
2342 io_do_iopoll(ctx, &nr_events, 0, false);
2344 /* let it sleep and repeat later if can't complete a request */
2348 * Ensure we allow local-to-the-cpu processing to take place,
2349 * in this case we need to ensure that we reap all events.
2350 * Also let task_work, etc. to progress by releasing the mutex
2352 if (need_resched()) {
2353 mutex_unlock(&ctx->uring_lock);
2355 mutex_lock(&ctx->uring_lock);
2358 mutex_unlock(&ctx->uring_lock);
2361 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2363 unsigned int nr_events = 0;
2367 * We disallow the app entering submit/complete with polling, but we
2368 * still need to lock the ring to prevent racing with polled issue
2369 * that got punted to a workqueue.
2371 mutex_lock(&ctx->uring_lock);
2373 * Don't enter poll loop if we already have events pending.
2374 * If we do, we can potentially be spinning for commands that
2375 * already triggered a CQE (eg in error).
2377 if (test_bit(0, &ctx->check_cq_overflow))
2378 __io_cqring_overflow_flush(ctx, false);
2379 if (io_cqring_events(ctx))
2383 * If a submit got punted to a workqueue, we can have the
2384 * application entering polling for a command before it gets
2385 * issued. That app will hold the uring_lock for the duration
2386 * of the poll right here, so we need to take a breather every
2387 * now and then to ensure that the issue has a chance to add
2388 * the poll to the issued list. Otherwise we can spin here
2389 * forever, while the workqueue is stuck trying to acquire the
2392 if (list_empty(&ctx->iopoll_list)) {
2393 u32 tail = ctx->cached_cq_tail;
2395 mutex_unlock(&ctx->uring_lock);
2397 mutex_lock(&ctx->uring_lock);
2399 /* some requests don't go through iopoll_list */
2400 if (tail != ctx->cached_cq_tail ||
2401 list_empty(&ctx->iopoll_list))
2404 ret = io_do_iopoll(ctx, &nr_events, min, true);
2405 } while (!ret && nr_events < min && !need_resched());
2407 mutex_unlock(&ctx->uring_lock);
2411 static void kiocb_end_write(struct io_kiocb *req)
2414 * Tell lockdep we inherited freeze protection from submission
2417 if (req->flags & REQ_F_ISREG) {
2418 struct super_block *sb = file_inode(req->file)->i_sb;
2420 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2426 static bool io_resubmit_prep(struct io_kiocb *req)
2428 struct io_async_rw *rw = req->async_data;
2431 return !io_req_prep_async(req);
2432 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2433 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2437 static bool io_rw_should_reissue(struct io_kiocb *req)
2439 umode_t mode = file_inode(req->file)->i_mode;
2440 struct io_ring_ctx *ctx = req->ctx;
2442 if (!S_ISBLK(mode) && !S_ISREG(mode))
2444 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2445 !(ctx->flags & IORING_SETUP_IOPOLL)))
2448 * If ref is dying, we might be running poll reap from the exit work.
2449 * Don't attempt to reissue from that path, just let it fail with
2452 if (percpu_ref_is_dying(&ctx->refs))
2455 * Play it safe and assume not safe to re-import and reissue if we're
2456 * not in the original thread group (or in task context).
2458 if (!same_thread_group(req->task, current) || !in_task())
2463 static bool io_resubmit_prep(struct io_kiocb *req)
2467 static bool io_rw_should_reissue(struct io_kiocb *req)
2473 static void io_fallback_req_func(struct work_struct *work)
2475 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
2476 fallback_work.work);
2477 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
2478 struct io_kiocb *req, *tmp;
2480 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
2481 req->io_task_work.func(req);
2484 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2485 unsigned int issue_flags)
2489 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2490 kiocb_end_write(req);
2491 if (res != req->result) {
2492 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2493 io_rw_should_reissue(req)) {
2494 req->flags |= REQ_F_REISSUE;
2499 if (req->flags & REQ_F_BUFFER_SELECTED)
2500 cflags = io_put_rw_kbuf(req);
2501 __io_req_complete(req, issue_flags, res, cflags);
2504 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2506 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2508 __io_complete_rw(req, res, res2, 0);
2511 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2513 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2515 if (kiocb->ki_flags & IOCB_WRITE)
2516 kiocb_end_write(req);
2517 if (unlikely(res != req->result)) {
2518 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2519 io_resubmit_prep(req))) {
2521 req->flags |= REQ_F_DONT_REISSUE;
2525 WRITE_ONCE(req->result, res);
2526 /* order with io_iopoll_complete() checking ->result */
2528 WRITE_ONCE(req->iopoll_completed, 1);
2532 * After the iocb has been issued, it's safe to be found on the poll list.
2533 * Adding the kiocb to the list AFTER submission ensures that we don't
2534 * find it from a io_do_iopoll() thread before the issuer is done
2535 * accessing the kiocb cookie.
2537 static void io_iopoll_req_issued(struct io_kiocb *req)
2539 struct io_ring_ctx *ctx = req->ctx;
2540 const bool in_async = io_wq_current_is_worker();
2542 /* workqueue context doesn't hold uring_lock, grab it now */
2543 if (unlikely(in_async))
2544 mutex_lock(&ctx->uring_lock);
2547 * Track whether we have multiple files in our lists. This will impact
2548 * how we do polling eventually, not spinning if we're on potentially
2549 * different devices.
2551 if (list_empty(&ctx->iopoll_list)) {
2552 ctx->poll_multi_queue = false;
2553 } else if (!ctx->poll_multi_queue) {
2554 struct io_kiocb *list_req;
2555 unsigned int queue_num0, queue_num1;
2557 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2560 if (list_req->file != req->file) {
2561 ctx->poll_multi_queue = true;
2563 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2564 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2565 if (queue_num0 != queue_num1)
2566 ctx->poll_multi_queue = true;
2571 * For fast devices, IO may have already completed. If it has, add
2572 * it to the front so we find it first.
2574 if (READ_ONCE(req->iopoll_completed))
2575 list_add(&req->inflight_entry, &ctx->iopoll_list);
2577 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2579 if (unlikely(in_async)) {
2581 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2582 * in sq thread task context or in io worker task context. If
2583 * current task context is sq thread, we don't need to check
2584 * whether should wake up sq thread.
2586 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2587 wq_has_sleeper(&ctx->sq_data->wait))
2588 wake_up(&ctx->sq_data->wait);
2590 mutex_unlock(&ctx->uring_lock);
2594 static inline void io_state_file_put(struct io_submit_state *state)
2596 if (state->file_refs) {
2597 fput_many(state->file, state->file_refs);
2598 state->file_refs = 0;
2603 * Get as many references to a file as we have IOs left in this submission,
2604 * assuming most submissions are for one file, or at least that each file
2605 * has more than one submission.
2607 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2612 if (state->file_refs) {
2613 if (state->fd == fd) {
2617 io_state_file_put(state);
2619 state->file = fget_many(fd, state->ios_left);
2620 if (unlikely(!state->file))
2624 state->file_refs = state->ios_left - 1;
2628 static bool io_bdev_nowait(struct block_device *bdev)
2630 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2634 * If we tracked the file through the SCM inflight mechanism, we could support
2635 * any file. For now, just ensure that anything potentially problematic is done
2638 static bool __io_file_supports_async(struct file *file, int rw)
2640 umode_t mode = file_inode(file)->i_mode;
2642 if (S_ISBLK(mode)) {
2643 if (IS_ENABLED(CONFIG_BLOCK) &&
2644 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2650 if (S_ISREG(mode)) {
2651 if (IS_ENABLED(CONFIG_BLOCK) &&
2652 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2653 file->f_op != &io_uring_fops)
2658 /* any ->read/write should understand O_NONBLOCK */
2659 if (file->f_flags & O_NONBLOCK)
2662 if (!(file->f_mode & FMODE_NOWAIT))
2666 return file->f_op->read_iter != NULL;
2668 return file->f_op->write_iter != NULL;
2671 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2673 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2675 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2678 return __io_file_supports_async(req->file, rw);
2681 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2683 struct io_ring_ctx *ctx = req->ctx;
2684 struct kiocb *kiocb = &req->rw.kiocb;
2685 struct file *file = req->file;
2689 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2690 req->flags |= REQ_F_ISREG;
2692 kiocb->ki_pos = READ_ONCE(sqe->off);
2693 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2694 req->flags |= REQ_F_CUR_POS;
2695 kiocb->ki_pos = file->f_pos;
2697 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2698 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2699 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2703 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2704 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2705 req->flags |= REQ_F_NOWAIT;
2707 ioprio = READ_ONCE(sqe->ioprio);
2709 ret = ioprio_check_cap(ioprio);
2713 kiocb->ki_ioprio = ioprio;
2715 kiocb->ki_ioprio = get_current_ioprio();
2717 if (ctx->flags & IORING_SETUP_IOPOLL) {
2718 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2719 !kiocb->ki_filp->f_op->iopoll)
2722 kiocb->ki_flags |= IOCB_HIPRI;
2723 kiocb->ki_complete = io_complete_rw_iopoll;
2724 req->iopoll_completed = 0;
2726 if (kiocb->ki_flags & IOCB_HIPRI)
2728 kiocb->ki_complete = io_complete_rw;
2731 if (req->opcode == IORING_OP_READ_FIXED ||
2732 req->opcode == IORING_OP_WRITE_FIXED) {
2734 io_req_set_rsrc_node(req);
2737 req->rw.addr = READ_ONCE(sqe->addr);
2738 req->rw.len = READ_ONCE(sqe->len);
2739 req->buf_index = READ_ONCE(sqe->buf_index);
2743 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2749 case -ERESTARTNOINTR:
2750 case -ERESTARTNOHAND:
2751 case -ERESTART_RESTARTBLOCK:
2753 * We can't just restart the syscall, since previously
2754 * submitted sqes may already be in progress. Just fail this
2760 kiocb->ki_complete(kiocb, ret, 0);
2764 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2765 unsigned int issue_flags)
2767 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2768 struct io_async_rw *io = req->async_data;
2769 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2771 /* add previously done IO, if any */
2772 if (io && io->bytes_done > 0) {
2774 ret = io->bytes_done;
2776 ret += io->bytes_done;
2779 if (req->flags & REQ_F_CUR_POS)
2780 req->file->f_pos = kiocb->ki_pos;
2781 if (ret >= 0 && check_reissue)
2782 __io_complete_rw(req, ret, 0, issue_flags);
2784 io_rw_done(kiocb, ret);
2786 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2787 req->flags &= ~REQ_F_REISSUE;
2788 if (io_resubmit_prep(req)) {
2790 io_req_task_queue_reissue(req);
2795 if (req->flags & REQ_F_BUFFER_SELECTED)
2796 cflags = io_put_rw_kbuf(req);
2797 __io_req_complete(req, issue_flags, ret, cflags);
2802 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2803 struct io_mapped_ubuf *imu)
2805 size_t len = req->rw.len;
2806 u64 buf_end, buf_addr = req->rw.addr;
2809 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2811 /* not inside the mapped region */
2812 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2816 * May not be a start of buffer, set size appropriately
2817 * and advance us to the beginning.
2819 offset = buf_addr - imu->ubuf;
2820 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2824 * Don't use iov_iter_advance() here, as it's really slow for
2825 * using the latter parts of a big fixed buffer - it iterates
2826 * over each segment manually. We can cheat a bit here, because
2829 * 1) it's a BVEC iter, we set it up
2830 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2831 * first and last bvec
2833 * So just find our index, and adjust the iterator afterwards.
2834 * If the offset is within the first bvec (or the whole first
2835 * bvec, just use iov_iter_advance(). This makes it easier
2836 * since we can just skip the first segment, which may not
2837 * be PAGE_SIZE aligned.
2839 const struct bio_vec *bvec = imu->bvec;
2841 if (offset <= bvec->bv_len) {
2842 iov_iter_advance(iter, offset);
2844 unsigned long seg_skip;
2846 /* skip first vec */
2847 offset -= bvec->bv_len;
2848 seg_skip = 1 + (offset >> PAGE_SHIFT);
2850 iter->bvec = bvec + seg_skip;
2851 iter->nr_segs -= seg_skip;
2852 iter->count -= bvec->bv_len + offset;
2853 iter->iov_offset = offset & ~PAGE_MASK;
2860 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2862 struct io_ring_ctx *ctx = req->ctx;
2863 struct io_mapped_ubuf *imu = req->imu;
2864 u16 index, buf_index = req->buf_index;
2867 if (unlikely(buf_index >= ctx->nr_user_bufs))
2869 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2870 imu = READ_ONCE(ctx->user_bufs[index]);
2873 return __io_import_fixed(req, rw, iter, imu);
2876 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2879 mutex_unlock(&ctx->uring_lock);
2882 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2885 * "Normal" inline submissions always hold the uring_lock, since we
2886 * grab it from the system call. Same is true for the SQPOLL offload.
2887 * The only exception is when we've detached the request and issue it
2888 * from an async worker thread, grab the lock for that case.
2891 mutex_lock(&ctx->uring_lock);
2894 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2895 int bgid, struct io_buffer *kbuf,
2898 struct io_buffer *head;
2900 if (req->flags & REQ_F_BUFFER_SELECTED)
2903 io_ring_submit_lock(req->ctx, needs_lock);
2905 lockdep_assert_held(&req->ctx->uring_lock);
2907 head = xa_load(&req->ctx->io_buffers, bgid);
2909 if (!list_empty(&head->list)) {
2910 kbuf = list_last_entry(&head->list, struct io_buffer,
2912 list_del(&kbuf->list);
2915 xa_erase(&req->ctx->io_buffers, bgid);
2917 if (*len > kbuf->len)
2920 kbuf = ERR_PTR(-ENOBUFS);
2923 io_ring_submit_unlock(req->ctx, needs_lock);
2928 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2931 struct io_buffer *kbuf;
2934 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2935 bgid = req->buf_index;
2936 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2939 req->rw.addr = (u64) (unsigned long) kbuf;
2940 req->flags |= REQ_F_BUFFER_SELECTED;
2941 return u64_to_user_ptr(kbuf->addr);
2944 #ifdef CONFIG_COMPAT
2945 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2948 struct compat_iovec __user *uiov;
2949 compat_ssize_t clen;
2953 uiov = u64_to_user_ptr(req->rw.addr);
2954 if (!access_ok(uiov, sizeof(*uiov)))
2956 if (__get_user(clen, &uiov->iov_len))
2962 buf = io_rw_buffer_select(req, &len, needs_lock);
2964 return PTR_ERR(buf);
2965 iov[0].iov_base = buf;
2966 iov[0].iov_len = (compat_size_t) len;
2971 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2974 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2978 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2981 len = iov[0].iov_len;
2984 buf = io_rw_buffer_select(req, &len, needs_lock);
2986 return PTR_ERR(buf);
2987 iov[0].iov_base = buf;
2988 iov[0].iov_len = len;
2992 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2995 if (req->flags & REQ_F_BUFFER_SELECTED) {
2996 struct io_buffer *kbuf;
2998 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2999 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3000 iov[0].iov_len = kbuf->len;
3003 if (req->rw.len != 1)
3006 #ifdef CONFIG_COMPAT
3007 if (req->ctx->compat)
3008 return io_compat_import(req, iov, needs_lock);
3011 return __io_iov_buffer_select(req, iov, needs_lock);
3014 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3015 struct iov_iter *iter, bool needs_lock)
3017 void __user *buf = u64_to_user_ptr(req->rw.addr);
3018 size_t sqe_len = req->rw.len;
3019 u8 opcode = req->opcode;
3022 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3024 return io_import_fixed(req, rw, iter);
3027 /* buffer index only valid with fixed read/write, or buffer select */
3028 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3031 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3032 if (req->flags & REQ_F_BUFFER_SELECT) {
3033 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3035 return PTR_ERR(buf);
3036 req->rw.len = sqe_len;
3039 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3044 if (req->flags & REQ_F_BUFFER_SELECT) {
3045 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3047 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3052 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3056 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3058 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3062 * For files that don't have ->read_iter() and ->write_iter(), handle them
3063 * by looping over ->read() or ->write() manually.
3065 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3067 struct kiocb *kiocb = &req->rw.kiocb;
3068 struct file *file = req->file;
3072 * Don't support polled IO through this interface, and we can't
3073 * support non-blocking either. For the latter, this just causes
3074 * the kiocb to be handled from an async context.
3076 if (kiocb->ki_flags & IOCB_HIPRI)
3078 if (kiocb->ki_flags & IOCB_NOWAIT)
3081 while (iov_iter_count(iter)) {
3085 if (!iov_iter_is_bvec(iter)) {
3086 iovec = iov_iter_iovec(iter);
3088 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3089 iovec.iov_len = req->rw.len;
3093 nr = file->f_op->read(file, iovec.iov_base,
3094 iovec.iov_len, io_kiocb_ppos(kiocb));
3096 nr = file->f_op->write(file, iovec.iov_base,
3097 iovec.iov_len, io_kiocb_ppos(kiocb));
3106 if (nr != iovec.iov_len)
3110 iov_iter_advance(iter, nr);
3116 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3117 const struct iovec *fast_iov, struct iov_iter *iter)
3119 struct io_async_rw *rw = req->async_data;
3121 memcpy(&rw->iter, iter, sizeof(*iter));
3122 rw->free_iovec = iovec;
3124 /* can only be fixed buffers, no need to do anything */
3125 if (iov_iter_is_bvec(iter))
3128 unsigned iov_off = 0;
3130 rw->iter.iov = rw->fast_iov;
3131 if (iter->iov != fast_iov) {
3132 iov_off = iter->iov - fast_iov;
3133 rw->iter.iov += iov_off;
3135 if (rw->fast_iov != fast_iov)
3136 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3137 sizeof(struct iovec) * iter->nr_segs);
3139 req->flags |= REQ_F_NEED_CLEANUP;
3143 static inline int io_alloc_async_data(struct io_kiocb *req)
3145 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3146 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3147 return req->async_data == NULL;
3150 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3151 const struct iovec *fast_iov,
3152 struct iov_iter *iter, bool force)
3154 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3156 if (!req->async_data) {
3157 if (io_alloc_async_data(req)) {
3162 io_req_map_rw(req, iovec, fast_iov, iter);
3167 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3169 struct io_async_rw *iorw = req->async_data;
3170 struct iovec *iov = iorw->fast_iov;
3173 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3174 if (unlikely(ret < 0))
3177 iorw->bytes_done = 0;
3178 iorw->free_iovec = iov;
3180 req->flags |= REQ_F_NEED_CLEANUP;
3184 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3186 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3188 return io_prep_rw(req, sqe);
3192 * This is our waitqueue callback handler, registered through lock_page_async()
3193 * when we initially tried to do the IO with the iocb armed our waitqueue.
3194 * This gets called when the page is unlocked, and we generally expect that to
3195 * happen when the page IO is completed and the page is now uptodate. This will
3196 * queue a task_work based retry of the operation, attempting to copy the data
3197 * again. If the latter fails because the page was NOT uptodate, then we will
3198 * do a thread based blocking retry of the operation. That's the unexpected
3201 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3202 int sync, void *arg)
3204 struct wait_page_queue *wpq;
3205 struct io_kiocb *req = wait->private;
3206 struct wait_page_key *key = arg;
3208 wpq = container_of(wait, struct wait_page_queue, wait);
3210 if (!wake_page_match(wpq, key))
3213 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3214 list_del_init(&wait->entry);
3216 /* submit ref gets dropped, acquire a new one */
3218 io_req_task_queue(req);
3223 * This controls whether a given IO request should be armed for async page
3224 * based retry. If we return false here, the request is handed to the async
3225 * worker threads for retry. If we're doing buffered reads on a regular file,
3226 * we prepare a private wait_page_queue entry and retry the operation. This
3227 * will either succeed because the page is now uptodate and unlocked, or it
3228 * will register a callback when the page is unlocked at IO completion. Through
3229 * that callback, io_uring uses task_work to setup a retry of the operation.
3230 * That retry will attempt the buffered read again. The retry will generally
3231 * succeed, or in rare cases where it fails, we then fall back to using the
3232 * async worker threads for a blocking retry.
3234 static bool io_rw_should_retry(struct io_kiocb *req)
3236 struct io_async_rw *rw = req->async_data;
3237 struct wait_page_queue *wait = &rw->wpq;
3238 struct kiocb *kiocb = &req->rw.kiocb;
3240 /* never retry for NOWAIT, we just complete with -EAGAIN */
3241 if (req->flags & REQ_F_NOWAIT)
3244 /* Only for buffered IO */
3245 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3249 * just use poll if we can, and don't attempt if the fs doesn't
3250 * support callback based unlocks
3252 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3255 wait->wait.func = io_async_buf_func;
3256 wait->wait.private = req;
3257 wait->wait.flags = 0;
3258 INIT_LIST_HEAD(&wait->wait.entry);
3259 kiocb->ki_flags |= IOCB_WAITQ;
3260 kiocb->ki_flags &= ~IOCB_NOWAIT;
3261 kiocb->ki_waitq = wait;
3265 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3267 if (req->file->f_op->read_iter)
3268 return call_read_iter(req->file, &req->rw.kiocb, iter);
3269 else if (req->file->f_op->read)
3270 return loop_rw_iter(READ, req, iter);
3275 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3277 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3278 struct kiocb *kiocb = &req->rw.kiocb;
3279 struct iov_iter __iter, *iter = &__iter;
3280 struct io_async_rw *rw = req->async_data;
3281 ssize_t io_size, ret, ret2;
3282 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3288 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3292 io_size = iov_iter_count(iter);
3293 req->result = io_size;
3295 /* Ensure we clear previously set non-block flag */
3296 if (!force_nonblock)
3297 kiocb->ki_flags &= ~IOCB_NOWAIT;
3299 kiocb->ki_flags |= IOCB_NOWAIT;
3301 /* If the file doesn't support async, just async punt */
3302 if (force_nonblock && !io_file_supports_async(req, READ)) {
3303 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3304 return ret ?: -EAGAIN;
3307 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3308 if (unlikely(ret)) {
3313 ret = io_iter_do_read(req, iter);
3315 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3316 req->flags &= ~REQ_F_REISSUE;
3317 /* IOPOLL retry should happen for io-wq threads */
3318 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3320 /* no retry on NONBLOCK nor RWF_NOWAIT */
3321 if (req->flags & REQ_F_NOWAIT)
3323 /* some cases will consume bytes even on error returns */
3324 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3326 } else if (ret == -EIOCBQUEUED) {
3328 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3329 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3330 /* read all, failed, already did sync or don't want to retry */
3334 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3339 rw = req->async_data;
3340 /* now use our persistent iterator, if we aren't already */
3345 rw->bytes_done += ret;
3346 /* if we can retry, do so with the callbacks armed */
3347 if (!io_rw_should_retry(req)) {
3348 kiocb->ki_flags &= ~IOCB_WAITQ;
3353 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3354 * we get -EIOCBQUEUED, then we'll get a notification when the
3355 * desired page gets unlocked. We can also get a partial read
3356 * here, and if we do, then just retry at the new offset.
3358 ret = io_iter_do_read(req, iter);
3359 if (ret == -EIOCBQUEUED)
3361 /* we got some bytes, but not all. retry. */
3362 kiocb->ki_flags &= ~IOCB_WAITQ;
3363 } while (ret > 0 && ret < io_size);
3365 kiocb_done(kiocb, ret, issue_flags);
3367 /* it's faster to check here then delegate to kfree */
3373 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3375 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3377 return io_prep_rw(req, sqe);
3380 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3382 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3383 struct kiocb *kiocb = &req->rw.kiocb;
3384 struct iov_iter __iter, *iter = &__iter;
3385 struct io_async_rw *rw = req->async_data;
3386 ssize_t ret, ret2, io_size;
3387 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3393 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3397 io_size = iov_iter_count(iter);
3398 req->result = io_size;
3400 /* Ensure we clear previously set non-block flag */
3401 if (!force_nonblock)
3402 kiocb->ki_flags &= ~IOCB_NOWAIT;
3404 kiocb->ki_flags |= IOCB_NOWAIT;
3406 /* If the file doesn't support async, just async punt */
3407 if (force_nonblock && !io_file_supports_async(req, WRITE))
3410 /* file path doesn't support NOWAIT for non-direct_IO */
3411 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3412 (req->flags & REQ_F_ISREG))
3415 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3420 * Open-code file_start_write here to grab freeze protection,
3421 * which will be released by another thread in
3422 * io_complete_rw(). Fool lockdep by telling it the lock got
3423 * released so that it doesn't complain about the held lock when
3424 * we return to userspace.
3426 if (req->flags & REQ_F_ISREG) {
3427 sb_start_write(file_inode(req->file)->i_sb);
3428 __sb_writers_release(file_inode(req->file)->i_sb,
3431 kiocb->ki_flags |= IOCB_WRITE;
3433 if (req->file->f_op->write_iter)
3434 ret2 = call_write_iter(req->file, kiocb, iter);
3435 else if (req->file->f_op->write)
3436 ret2 = loop_rw_iter(WRITE, req, iter);
3440 if (req->flags & REQ_F_REISSUE) {
3441 req->flags &= ~REQ_F_REISSUE;
3446 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3447 * retry them without IOCB_NOWAIT.
3449 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3451 /* no retry on NONBLOCK nor RWF_NOWAIT */
3452 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3454 if (!force_nonblock || ret2 != -EAGAIN) {
3455 /* IOPOLL retry should happen for io-wq threads */
3456 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3459 kiocb_done(kiocb, ret2, issue_flags);
3462 /* some cases will consume bytes even on error returns */
3463 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3464 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3465 return ret ?: -EAGAIN;
3468 /* it's reportedly faster than delegating the null check to kfree() */
3474 static int io_renameat_prep(struct io_kiocb *req,
3475 const struct io_uring_sqe *sqe)
3477 struct io_rename *ren = &req->rename;
3478 const char __user *oldf, *newf;
3480 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3482 if (sqe->ioprio || sqe->buf_index)
3484 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3487 ren->old_dfd = READ_ONCE(sqe->fd);
3488 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3489 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3490 ren->new_dfd = READ_ONCE(sqe->len);
3491 ren->flags = READ_ONCE(sqe->rename_flags);
3493 ren->oldpath = getname(oldf);
3494 if (IS_ERR(ren->oldpath))
3495 return PTR_ERR(ren->oldpath);
3497 ren->newpath = getname(newf);
3498 if (IS_ERR(ren->newpath)) {
3499 putname(ren->oldpath);
3500 return PTR_ERR(ren->newpath);
3503 req->flags |= REQ_F_NEED_CLEANUP;
3507 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3509 struct io_rename *ren = &req->rename;
3512 if (issue_flags & IO_URING_F_NONBLOCK)
3515 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3516 ren->newpath, ren->flags);
3518 req->flags &= ~REQ_F_NEED_CLEANUP;
3521 io_req_complete(req, ret);
3525 static int io_unlinkat_prep(struct io_kiocb *req,
3526 const struct io_uring_sqe *sqe)
3528 struct io_unlink *un = &req->unlink;
3529 const char __user *fname;
3531 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3533 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3535 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3538 un->dfd = READ_ONCE(sqe->fd);
3540 un->flags = READ_ONCE(sqe->unlink_flags);
3541 if (un->flags & ~AT_REMOVEDIR)
3544 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3545 un->filename = getname(fname);
3546 if (IS_ERR(un->filename))
3547 return PTR_ERR(un->filename);
3549 req->flags |= REQ_F_NEED_CLEANUP;
3553 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3555 struct io_unlink *un = &req->unlink;
3558 if (issue_flags & IO_URING_F_NONBLOCK)
3561 if (un->flags & AT_REMOVEDIR)
3562 ret = do_rmdir(un->dfd, un->filename);
3564 ret = do_unlinkat(un->dfd, un->filename);
3566 req->flags &= ~REQ_F_NEED_CLEANUP;
3569 io_req_complete(req, ret);
3573 static int io_shutdown_prep(struct io_kiocb *req,
3574 const struct io_uring_sqe *sqe)
3576 #if defined(CONFIG_NET)
3577 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3579 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3583 req->shutdown.how = READ_ONCE(sqe->len);
3590 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3592 #if defined(CONFIG_NET)
3593 struct socket *sock;
3596 if (issue_flags & IO_URING_F_NONBLOCK)
3599 sock = sock_from_file(req->file);
3600 if (unlikely(!sock))
3603 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3606 io_req_complete(req, ret);
3613 static int __io_splice_prep(struct io_kiocb *req,
3614 const struct io_uring_sqe *sqe)
3616 struct io_splice *sp = &req->splice;
3617 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3619 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3623 sp->len = READ_ONCE(sqe->len);
3624 sp->flags = READ_ONCE(sqe->splice_flags);
3626 if (unlikely(sp->flags & ~valid_flags))
3629 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3630 (sp->flags & SPLICE_F_FD_IN_FIXED));
3633 req->flags |= REQ_F_NEED_CLEANUP;
3637 static int io_tee_prep(struct io_kiocb *req,
3638 const struct io_uring_sqe *sqe)
3640 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3642 return __io_splice_prep(req, sqe);
3645 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3647 struct io_splice *sp = &req->splice;
3648 struct file *in = sp->file_in;
3649 struct file *out = sp->file_out;
3650 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3653 if (issue_flags & IO_URING_F_NONBLOCK)
3656 ret = do_tee(in, out, sp->len, flags);
3658 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3660 req->flags &= ~REQ_F_NEED_CLEANUP;
3664 io_req_complete(req, ret);
3668 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3670 struct io_splice *sp = &req->splice;
3672 sp->off_in = READ_ONCE(sqe->splice_off_in);
3673 sp->off_out = READ_ONCE(sqe->off);
3674 return __io_splice_prep(req, sqe);
3677 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3679 struct io_splice *sp = &req->splice;
3680 struct file *in = sp->file_in;
3681 struct file *out = sp->file_out;
3682 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3683 loff_t *poff_in, *poff_out;
3686 if (issue_flags & IO_URING_F_NONBLOCK)
3689 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3690 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3693 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3695 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3697 req->flags &= ~REQ_F_NEED_CLEANUP;
3701 io_req_complete(req, ret);
3706 * IORING_OP_NOP just posts a completion event, nothing else.
3708 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3710 struct io_ring_ctx *ctx = req->ctx;
3712 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3715 __io_req_complete(req, issue_flags, 0, 0);
3719 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3721 struct io_ring_ctx *ctx = req->ctx;
3726 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3728 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3731 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3732 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3735 req->sync.off = READ_ONCE(sqe->off);
3736 req->sync.len = READ_ONCE(sqe->len);
3740 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3742 loff_t end = req->sync.off + req->sync.len;
3745 /* fsync always requires a blocking context */
3746 if (issue_flags & IO_URING_F_NONBLOCK)
3749 ret = vfs_fsync_range(req->file, req->sync.off,
3750 end > 0 ? end : LLONG_MAX,
3751 req->sync.flags & IORING_FSYNC_DATASYNC);
3754 io_req_complete(req, ret);
3758 static int io_fallocate_prep(struct io_kiocb *req,
3759 const struct io_uring_sqe *sqe)
3761 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3763 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3766 req->sync.off = READ_ONCE(sqe->off);
3767 req->sync.len = READ_ONCE(sqe->addr);
3768 req->sync.mode = READ_ONCE(sqe->len);
3772 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3776 /* fallocate always requiring blocking context */
3777 if (issue_flags & IO_URING_F_NONBLOCK)
3779 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3783 io_req_complete(req, ret);
3787 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3789 const char __user *fname;
3792 if (unlikely(sqe->ioprio || sqe->buf_index))
3794 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3797 /* open.how should be already initialised */
3798 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3799 req->open.how.flags |= O_LARGEFILE;
3801 req->open.dfd = READ_ONCE(sqe->fd);
3802 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3803 req->open.filename = getname(fname);
3804 if (IS_ERR(req->open.filename)) {
3805 ret = PTR_ERR(req->open.filename);
3806 req->open.filename = NULL;
3809 req->open.nofile = rlimit(RLIMIT_NOFILE);
3810 req->flags |= REQ_F_NEED_CLEANUP;
3814 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3818 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3820 mode = READ_ONCE(sqe->len);
3821 flags = READ_ONCE(sqe->open_flags);
3822 req->open.how = build_open_how(flags, mode);
3823 return __io_openat_prep(req, sqe);
3826 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3828 struct open_how __user *how;
3832 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3834 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3835 len = READ_ONCE(sqe->len);
3836 if (len < OPEN_HOW_SIZE_VER0)
3839 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3844 return __io_openat_prep(req, sqe);
3847 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3849 struct open_flags op;
3852 bool resolve_nonblock;
3855 ret = build_open_flags(&req->open.how, &op);
3858 nonblock_set = op.open_flag & O_NONBLOCK;
3859 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3860 if (issue_flags & IO_URING_F_NONBLOCK) {
3862 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3863 * it'll always -EAGAIN
3865 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3867 op.lookup_flags |= LOOKUP_CACHED;
3868 op.open_flag |= O_NONBLOCK;
3871 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3875 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3878 * We could hang on to this 'fd' on retrying, but seems like
3879 * marginal gain for something that is now known to be a slower
3880 * path. So just put it, and we'll get a new one when we retry.
3884 ret = PTR_ERR(file);
3885 /* only retry if RESOLVE_CACHED wasn't already set by application */
3886 if (ret == -EAGAIN &&
3887 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3892 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3893 file->f_flags &= ~O_NONBLOCK;
3894 fsnotify_open(file);
3895 fd_install(ret, file);
3897 putname(req->open.filename);
3898 req->flags &= ~REQ_F_NEED_CLEANUP;
3901 __io_req_complete(req, issue_flags, ret, 0);
3905 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3907 return io_openat2(req, issue_flags);
3910 static int io_remove_buffers_prep(struct io_kiocb *req,
3911 const struct io_uring_sqe *sqe)
3913 struct io_provide_buf *p = &req->pbuf;
3916 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3919 tmp = READ_ONCE(sqe->fd);
3920 if (!tmp || tmp > USHRT_MAX)
3923 memset(p, 0, sizeof(*p));
3925 p->bgid = READ_ONCE(sqe->buf_group);
3929 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3930 int bgid, unsigned nbufs)
3934 /* shouldn't happen */
3938 /* the head kbuf is the list itself */
3939 while (!list_empty(&buf->list)) {
3940 struct io_buffer *nxt;
3942 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3943 list_del(&nxt->list);
3950 xa_erase(&ctx->io_buffers, bgid);
3955 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3957 struct io_provide_buf *p = &req->pbuf;
3958 struct io_ring_ctx *ctx = req->ctx;
3959 struct io_buffer *head;
3961 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3963 io_ring_submit_lock(ctx, !force_nonblock);
3965 lockdep_assert_held(&ctx->uring_lock);
3968 head = xa_load(&ctx->io_buffers, p->bgid);
3970 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3974 /* complete before unlock, IOPOLL may need the lock */
3975 __io_req_complete(req, issue_flags, ret, 0);
3976 io_ring_submit_unlock(ctx, !force_nonblock);
3980 static int io_provide_buffers_prep(struct io_kiocb *req,
3981 const struct io_uring_sqe *sqe)
3983 unsigned long size, tmp_check;
3984 struct io_provide_buf *p = &req->pbuf;
3987 if (sqe->ioprio || sqe->rw_flags)
3990 tmp = READ_ONCE(sqe->fd);
3991 if (!tmp || tmp > USHRT_MAX)
3994 p->addr = READ_ONCE(sqe->addr);
3995 p->len = READ_ONCE(sqe->len);
3997 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4000 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4003 size = (unsigned long)p->len * p->nbufs;
4004 if (!access_ok(u64_to_user_ptr(p->addr), size))
4007 p->bgid = READ_ONCE(sqe->buf_group);
4008 tmp = READ_ONCE(sqe->off);
4009 if (tmp > USHRT_MAX)
4015 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4017 struct io_buffer *buf;
4018 u64 addr = pbuf->addr;
4019 int i, bid = pbuf->bid;
4021 for (i = 0; i < pbuf->nbufs; i++) {
4022 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4027 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4032 INIT_LIST_HEAD(&buf->list);
4035 list_add_tail(&buf->list, &(*head)->list);
4039 return i ? i : -ENOMEM;
4042 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4044 struct io_provide_buf *p = &req->pbuf;
4045 struct io_ring_ctx *ctx = req->ctx;
4046 struct io_buffer *head, *list;
4048 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4050 io_ring_submit_lock(ctx, !force_nonblock);
4052 lockdep_assert_held(&ctx->uring_lock);
4054 list = head = xa_load(&ctx->io_buffers, p->bgid);
4056 ret = io_add_buffers(p, &head);
4057 if (ret >= 0 && !list) {
4058 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4060 __io_remove_buffers(ctx, head, p->bgid, -1U);
4064 /* complete before unlock, IOPOLL may need the lock */
4065 __io_req_complete(req, issue_flags, ret, 0);
4066 io_ring_submit_unlock(ctx, !force_nonblock);
4070 static int io_epoll_ctl_prep(struct io_kiocb *req,
4071 const struct io_uring_sqe *sqe)
4073 #if defined(CONFIG_EPOLL)
4074 if (sqe->ioprio || sqe->buf_index)
4076 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4079 req->epoll.epfd = READ_ONCE(sqe->fd);
4080 req->epoll.op = READ_ONCE(sqe->len);
4081 req->epoll.fd = READ_ONCE(sqe->off);
4083 if (ep_op_has_event(req->epoll.op)) {
4084 struct epoll_event __user *ev;
4086 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4087 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4097 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4099 #if defined(CONFIG_EPOLL)
4100 struct io_epoll *ie = &req->epoll;
4102 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4104 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4105 if (force_nonblock && ret == -EAGAIN)
4110 __io_req_complete(req, issue_flags, ret, 0);
4117 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4119 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4120 if (sqe->ioprio || sqe->buf_index || sqe->off)
4122 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4125 req->madvise.addr = READ_ONCE(sqe->addr);
4126 req->madvise.len = READ_ONCE(sqe->len);
4127 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4134 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4136 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4137 struct io_madvise *ma = &req->madvise;
4140 if (issue_flags & IO_URING_F_NONBLOCK)
4143 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4146 io_req_complete(req, ret);
4153 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4155 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4157 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4160 req->fadvise.offset = READ_ONCE(sqe->off);
4161 req->fadvise.len = READ_ONCE(sqe->len);
4162 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4166 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4168 struct io_fadvise *fa = &req->fadvise;
4171 if (issue_flags & IO_URING_F_NONBLOCK) {
4172 switch (fa->advice) {
4173 case POSIX_FADV_NORMAL:
4174 case POSIX_FADV_RANDOM:
4175 case POSIX_FADV_SEQUENTIAL:
4182 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4185 __io_req_complete(req, issue_flags, ret, 0);
4189 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4191 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4193 if (sqe->ioprio || sqe->buf_index)
4195 if (req->flags & REQ_F_FIXED_FILE)
4198 req->statx.dfd = READ_ONCE(sqe->fd);
4199 req->statx.mask = READ_ONCE(sqe->len);
4200 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4201 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4202 req->statx.flags = READ_ONCE(sqe->statx_flags);
4207 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4209 struct io_statx *ctx = &req->statx;
4212 if (issue_flags & IO_URING_F_NONBLOCK)
4215 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4220 io_req_complete(req, ret);
4224 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4226 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4228 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4229 sqe->rw_flags || sqe->buf_index)
4231 if (req->flags & REQ_F_FIXED_FILE)
4234 req->close.fd = READ_ONCE(sqe->fd);
4238 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4240 struct files_struct *files = current->files;
4241 struct io_close *close = &req->close;
4242 struct fdtable *fdt;
4243 struct file *file = NULL;
4246 spin_lock(&files->file_lock);
4247 fdt = files_fdtable(files);
4248 if (close->fd >= fdt->max_fds) {
4249 spin_unlock(&files->file_lock);
4252 file = fdt->fd[close->fd];
4253 if (!file || file->f_op == &io_uring_fops) {
4254 spin_unlock(&files->file_lock);
4259 /* if the file has a flush method, be safe and punt to async */
4260 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4261 spin_unlock(&files->file_lock);
4265 ret = __close_fd_get_file(close->fd, &file);
4266 spin_unlock(&files->file_lock);
4273 /* No ->flush() or already async, safely close from here */
4274 ret = filp_close(file, current->files);
4280 __io_req_complete(req, issue_flags, ret, 0);
4284 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4286 struct io_ring_ctx *ctx = req->ctx;
4288 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4290 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4293 req->sync.off = READ_ONCE(sqe->off);
4294 req->sync.len = READ_ONCE(sqe->len);
4295 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4299 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4303 /* sync_file_range always requires a blocking context */
4304 if (issue_flags & IO_URING_F_NONBLOCK)
4307 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4311 io_req_complete(req, ret);
4315 #if defined(CONFIG_NET)
4316 static int io_setup_async_msg(struct io_kiocb *req,
4317 struct io_async_msghdr *kmsg)
4319 struct io_async_msghdr *async_msg = req->async_data;
4323 if (io_alloc_async_data(req)) {
4324 kfree(kmsg->free_iov);
4327 async_msg = req->async_data;
4328 req->flags |= REQ_F_NEED_CLEANUP;
4329 memcpy(async_msg, kmsg, sizeof(*kmsg));
4330 async_msg->msg.msg_name = &async_msg->addr;
4331 /* if were using fast_iov, set it to the new one */
4332 if (!async_msg->free_iov)
4333 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4338 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4339 struct io_async_msghdr *iomsg)
4341 iomsg->msg.msg_name = &iomsg->addr;
4342 iomsg->free_iov = iomsg->fast_iov;
4343 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4344 req->sr_msg.msg_flags, &iomsg->free_iov);
4347 static int io_sendmsg_prep_async(struct io_kiocb *req)
4351 ret = io_sendmsg_copy_hdr(req, req->async_data);
4353 req->flags |= REQ_F_NEED_CLEANUP;
4357 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4359 struct io_sr_msg *sr = &req->sr_msg;
4361 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4364 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4365 sr->len = READ_ONCE(sqe->len);
4366 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4367 if (sr->msg_flags & MSG_DONTWAIT)
4368 req->flags |= REQ_F_NOWAIT;
4370 #ifdef CONFIG_COMPAT
4371 if (req->ctx->compat)
4372 sr->msg_flags |= MSG_CMSG_COMPAT;
4377 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4379 struct io_async_msghdr iomsg, *kmsg;
4380 struct socket *sock;
4385 sock = sock_from_file(req->file);
4386 if (unlikely(!sock))
4389 kmsg = req->async_data;
4391 ret = io_sendmsg_copy_hdr(req, &iomsg);
4397 flags = req->sr_msg.msg_flags;
4398 if (issue_flags & IO_URING_F_NONBLOCK)
4399 flags |= MSG_DONTWAIT;
4400 if (flags & MSG_WAITALL)
4401 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4403 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4404 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4405 return io_setup_async_msg(req, kmsg);
4406 if (ret == -ERESTARTSYS)
4409 /* fast path, check for non-NULL to avoid function call */
4411 kfree(kmsg->free_iov);
4412 req->flags &= ~REQ_F_NEED_CLEANUP;
4415 __io_req_complete(req, issue_flags, ret, 0);
4419 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4421 struct io_sr_msg *sr = &req->sr_msg;
4424 struct socket *sock;
4429 sock = sock_from_file(req->file);
4430 if (unlikely(!sock))
4433 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4437 msg.msg_name = NULL;
4438 msg.msg_control = NULL;
4439 msg.msg_controllen = 0;
4440 msg.msg_namelen = 0;
4442 flags = req->sr_msg.msg_flags;
4443 if (issue_flags & IO_URING_F_NONBLOCK)
4444 flags |= MSG_DONTWAIT;
4445 if (flags & MSG_WAITALL)
4446 min_ret = iov_iter_count(&msg.msg_iter);
4448 msg.msg_flags = flags;
4449 ret = sock_sendmsg(sock, &msg);
4450 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4452 if (ret == -ERESTARTSYS)
4457 __io_req_complete(req, issue_flags, ret, 0);
4461 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4462 struct io_async_msghdr *iomsg)
4464 struct io_sr_msg *sr = &req->sr_msg;
4465 struct iovec __user *uiov;
4469 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4470 &iomsg->uaddr, &uiov, &iov_len);
4474 if (req->flags & REQ_F_BUFFER_SELECT) {
4477 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4479 sr->len = iomsg->fast_iov[0].iov_len;
4480 iomsg->free_iov = NULL;
4482 iomsg->free_iov = iomsg->fast_iov;
4483 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4484 &iomsg->free_iov, &iomsg->msg.msg_iter,
4493 #ifdef CONFIG_COMPAT
4494 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4495 struct io_async_msghdr *iomsg)
4497 struct io_sr_msg *sr = &req->sr_msg;
4498 struct compat_iovec __user *uiov;
4503 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4508 uiov = compat_ptr(ptr);
4509 if (req->flags & REQ_F_BUFFER_SELECT) {
4510 compat_ssize_t clen;
4514 if (!access_ok(uiov, sizeof(*uiov)))
4516 if (__get_user(clen, &uiov->iov_len))
4521 iomsg->free_iov = NULL;
4523 iomsg->free_iov = iomsg->fast_iov;
4524 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4525 UIO_FASTIOV, &iomsg->free_iov,
4526 &iomsg->msg.msg_iter, true);
4535 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4536 struct io_async_msghdr *iomsg)
4538 iomsg->msg.msg_name = &iomsg->addr;
4540 #ifdef CONFIG_COMPAT
4541 if (req->ctx->compat)
4542 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4545 return __io_recvmsg_copy_hdr(req, iomsg);
4548 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4551 struct io_sr_msg *sr = &req->sr_msg;
4552 struct io_buffer *kbuf;
4554 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4559 req->flags |= REQ_F_BUFFER_SELECTED;
4563 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4565 return io_put_kbuf(req, req->sr_msg.kbuf);
4568 static int io_recvmsg_prep_async(struct io_kiocb *req)
4572 ret = io_recvmsg_copy_hdr(req, req->async_data);
4574 req->flags |= REQ_F_NEED_CLEANUP;
4578 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4580 struct io_sr_msg *sr = &req->sr_msg;
4582 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4585 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4586 sr->len = READ_ONCE(sqe->len);
4587 sr->bgid = READ_ONCE(sqe->buf_group);
4588 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4589 if (sr->msg_flags & MSG_DONTWAIT)
4590 req->flags |= REQ_F_NOWAIT;
4592 #ifdef CONFIG_COMPAT
4593 if (req->ctx->compat)
4594 sr->msg_flags |= MSG_CMSG_COMPAT;
4599 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4601 struct io_async_msghdr iomsg, *kmsg;
4602 struct socket *sock;
4603 struct io_buffer *kbuf;
4606 int ret, cflags = 0;
4607 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4609 sock = sock_from_file(req->file);
4610 if (unlikely(!sock))
4613 kmsg = req->async_data;
4615 ret = io_recvmsg_copy_hdr(req, &iomsg);
4621 if (req->flags & REQ_F_BUFFER_SELECT) {
4622 kbuf = io_recv_buffer_select(req, !force_nonblock);
4624 return PTR_ERR(kbuf);
4625 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4626 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4627 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4628 1, req->sr_msg.len);
4631 flags = req->sr_msg.msg_flags;
4633 flags |= MSG_DONTWAIT;
4634 if (flags & MSG_WAITALL)
4635 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4637 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4638 kmsg->uaddr, flags);
4639 if (force_nonblock && ret == -EAGAIN)
4640 return io_setup_async_msg(req, kmsg);
4641 if (ret == -ERESTARTSYS)
4644 if (req->flags & REQ_F_BUFFER_SELECTED)
4645 cflags = io_put_recv_kbuf(req);
4646 /* fast path, check for non-NULL to avoid function call */
4648 kfree(kmsg->free_iov);
4649 req->flags &= ~REQ_F_NEED_CLEANUP;
4650 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4652 __io_req_complete(req, issue_flags, ret, cflags);
4656 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4658 struct io_buffer *kbuf;
4659 struct io_sr_msg *sr = &req->sr_msg;
4661 void __user *buf = sr->buf;
4662 struct socket *sock;
4666 int ret, cflags = 0;
4667 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4669 sock = sock_from_file(req->file);
4670 if (unlikely(!sock))
4673 if (req->flags & REQ_F_BUFFER_SELECT) {
4674 kbuf = io_recv_buffer_select(req, !force_nonblock);
4676 return PTR_ERR(kbuf);
4677 buf = u64_to_user_ptr(kbuf->addr);
4680 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4684 msg.msg_name = NULL;
4685 msg.msg_control = NULL;
4686 msg.msg_controllen = 0;
4687 msg.msg_namelen = 0;
4688 msg.msg_iocb = NULL;
4691 flags = req->sr_msg.msg_flags;
4693 flags |= MSG_DONTWAIT;
4694 if (flags & MSG_WAITALL)
4695 min_ret = iov_iter_count(&msg.msg_iter);
4697 ret = sock_recvmsg(sock, &msg, flags);
4698 if (force_nonblock && ret == -EAGAIN)
4700 if (ret == -ERESTARTSYS)
4703 if (req->flags & REQ_F_BUFFER_SELECTED)
4704 cflags = io_put_recv_kbuf(req);
4705 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4707 __io_req_complete(req, issue_flags, ret, cflags);
4711 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4713 struct io_accept *accept = &req->accept;
4715 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4717 if (sqe->ioprio || sqe->len || sqe->buf_index)
4720 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4721 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4722 accept->flags = READ_ONCE(sqe->accept_flags);
4723 accept->nofile = rlimit(RLIMIT_NOFILE);
4727 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4729 struct io_accept *accept = &req->accept;
4730 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4731 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4734 if (req->file->f_flags & O_NONBLOCK)
4735 req->flags |= REQ_F_NOWAIT;
4737 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4738 accept->addr_len, accept->flags,
4740 if (ret == -EAGAIN && force_nonblock)
4743 if (ret == -ERESTARTSYS)
4747 __io_req_complete(req, issue_flags, ret, 0);
4751 static int io_connect_prep_async(struct io_kiocb *req)
4753 struct io_async_connect *io = req->async_data;
4754 struct io_connect *conn = &req->connect;
4756 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4759 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4761 struct io_connect *conn = &req->connect;
4763 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4765 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4768 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4769 conn->addr_len = READ_ONCE(sqe->addr2);
4773 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4775 struct io_async_connect __io, *io;
4776 unsigned file_flags;
4778 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4780 if (req->async_data) {
4781 io = req->async_data;
4783 ret = move_addr_to_kernel(req->connect.addr,
4784 req->connect.addr_len,
4791 file_flags = force_nonblock ? O_NONBLOCK : 0;
4793 ret = __sys_connect_file(req->file, &io->address,
4794 req->connect.addr_len, file_flags);
4795 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4796 if (req->async_data)
4798 if (io_alloc_async_data(req)) {
4802 memcpy(req->async_data, &__io, sizeof(__io));
4805 if (ret == -ERESTARTSYS)
4810 __io_req_complete(req, issue_flags, ret, 0);
4813 #else /* !CONFIG_NET */
4814 #define IO_NETOP_FN(op) \
4815 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4817 return -EOPNOTSUPP; \
4820 #define IO_NETOP_PREP(op) \
4822 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4824 return -EOPNOTSUPP; \
4827 #define IO_NETOP_PREP_ASYNC(op) \
4829 static int io_##op##_prep_async(struct io_kiocb *req) \
4831 return -EOPNOTSUPP; \
4834 IO_NETOP_PREP_ASYNC(sendmsg);
4835 IO_NETOP_PREP_ASYNC(recvmsg);
4836 IO_NETOP_PREP_ASYNC(connect);
4837 IO_NETOP_PREP(accept);
4840 #endif /* CONFIG_NET */
4842 struct io_poll_table {
4843 struct poll_table_struct pt;
4844 struct io_kiocb *req;
4849 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4850 __poll_t mask, io_req_tw_func_t func)
4852 /* for instances that support it check for an event match first: */
4853 if (mask && !(mask & poll->events))
4856 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4858 list_del_init(&poll->wait.entry);
4861 req->io_task_work.func = func;
4864 * If this fails, then the task is exiting. When a task exits, the
4865 * work gets canceled, so just cancel this request as well instead
4866 * of executing it. We can't safely execute it anyway, as we may not
4867 * have the needed state needed for it anyway.
4869 io_req_task_work_add(req);
4873 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4874 __acquires(&req->ctx->completion_lock)
4876 struct io_ring_ctx *ctx = req->ctx;
4878 if (unlikely(req->task->flags & PF_EXITING))
4879 WRITE_ONCE(poll->canceled, true);
4881 if (!req->result && !READ_ONCE(poll->canceled)) {
4882 struct poll_table_struct pt = { ._key = poll->events };
4884 req->result = vfs_poll(req->file, &pt) & poll->events;
4887 spin_lock_irq(&ctx->completion_lock);
4888 if (!req->result && !READ_ONCE(poll->canceled)) {
4889 add_wait_queue(poll->head, &poll->wait);
4896 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4898 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4899 if (req->opcode == IORING_OP_POLL_ADD)
4900 return req->async_data;
4901 return req->apoll->double_poll;
4904 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4906 if (req->opcode == IORING_OP_POLL_ADD)
4908 return &req->apoll->poll;
4911 static void io_poll_remove_double(struct io_kiocb *req)
4912 __must_hold(&req->ctx->completion_lock)
4914 struct io_poll_iocb *poll = io_poll_get_double(req);
4916 lockdep_assert_held(&req->ctx->completion_lock);
4918 if (poll && poll->head) {
4919 struct wait_queue_head *head = poll->head;
4921 spin_lock(&head->lock);
4922 list_del_init(&poll->wait.entry);
4923 if (poll->wait.private)
4926 spin_unlock(&head->lock);
4930 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4931 __must_hold(&req->ctx->completion_lock)
4933 struct io_ring_ctx *ctx = req->ctx;
4934 unsigned flags = IORING_CQE_F_MORE;
4937 if (READ_ONCE(req->poll.canceled)) {
4939 req->poll.events |= EPOLLONESHOT;
4941 error = mangle_poll(mask);
4943 if (req->poll.events & EPOLLONESHOT)
4945 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4946 req->poll.done = true;
4949 if (flags & IORING_CQE_F_MORE)
4952 io_commit_cqring(ctx);
4953 return !(flags & IORING_CQE_F_MORE);
4956 static void io_poll_task_func(struct io_kiocb *req)
4958 struct io_ring_ctx *ctx = req->ctx;
4959 struct io_kiocb *nxt;
4961 if (io_poll_rewait(req, &req->poll)) {
4962 spin_unlock_irq(&ctx->completion_lock);
4966 done = io_poll_complete(req, req->result);
4968 io_poll_remove_double(req);
4969 hash_del(&req->hash_node);
4972 add_wait_queue(req->poll.head, &req->poll.wait);
4974 spin_unlock_irq(&ctx->completion_lock);
4975 io_cqring_ev_posted(ctx);
4978 nxt = io_put_req_find_next(req);
4980 io_req_task_submit(nxt);
4985 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4986 int sync, void *key)
4988 struct io_kiocb *req = wait->private;
4989 struct io_poll_iocb *poll = io_poll_get_single(req);
4990 __poll_t mask = key_to_poll(key);
4992 /* for instances that support it check for an event match first: */
4993 if (mask && !(mask & poll->events))
4995 if (!(poll->events & EPOLLONESHOT))
4996 return poll->wait.func(&poll->wait, mode, sync, key);
4998 list_del_init(&wait->entry);
5003 spin_lock(&poll->head->lock);
5004 done = list_empty(&poll->wait.entry);
5006 list_del_init(&poll->wait.entry);
5007 /* make sure double remove sees this as being gone */
5008 wait->private = NULL;
5009 spin_unlock(&poll->head->lock);
5011 /* use wait func handler, so it matches the rq type */
5012 poll->wait.func(&poll->wait, mode, sync, key);
5019 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5020 wait_queue_func_t wake_func)
5024 poll->canceled = false;
5025 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5026 /* mask in events that we always want/need */
5027 poll->events = events | IO_POLL_UNMASK;
5028 INIT_LIST_HEAD(&poll->wait.entry);
5029 init_waitqueue_func_entry(&poll->wait, wake_func);
5032 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5033 struct wait_queue_head *head,
5034 struct io_poll_iocb **poll_ptr)
5036 struct io_kiocb *req = pt->req;
5039 * The file being polled uses multiple waitqueues for poll handling
5040 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5043 if (unlikely(pt->nr_entries)) {
5044 struct io_poll_iocb *poll_one = poll;
5046 /* already have a 2nd entry, fail a third attempt */
5048 pt->error = -EINVAL;
5052 * Can't handle multishot for double wait for now, turn it
5053 * into one-shot mode.
5055 if (!(poll_one->events & EPOLLONESHOT))
5056 poll_one->events |= EPOLLONESHOT;
5057 /* double add on the same waitqueue head, ignore */
5058 if (poll_one->head == head)
5060 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5062 pt->error = -ENOMEM;
5065 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5067 poll->wait.private = req;
5074 if (poll->events & EPOLLEXCLUSIVE)
5075 add_wait_queue_exclusive(head, &poll->wait);
5077 add_wait_queue(head, &poll->wait);
5080 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5081 struct poll_table_struct *p)
5083 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5084 struct async_poll *apoll = pt->req->apoll;
5086 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5089 static void io_async_task_func(struct io_kiocb *req)
5091 struct async_poll *apoll = req->apoll;
5092 struct io_ring_ctx *ctx = req->ctx;
5094 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5096 if (io_poll_rewait(req, &apoll->poll)) {
5097 spin_unlock_irq(&ctx->completion_lock);
5101 hash_del(&req->hash_node);
5102 io_poll_remove_double(req);
5103 spin_unlock_irq(&ctx->completion_lock);
5105 if (!READ_ONCE(apoll->poll.canceled))
5106 io_req_task_submit(req);
5108 io_req_complete_failed(req, -ECANCELED);
5111 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5114 struct io_kiocb *req = wait->private;
5115 struct io_poll_iocb *poll = &req->apoll->poll;
5117 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5120 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5123 static void io_poll_req_insert(struct io_kiocb *req)
5125 struct io_ring_ctx *ctx = req->ctx;
5126 struct hlist_head *list;
5128 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5129 hlist_add_head(&req->hash_node, list);
5132 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5133 struct io_poll_iocb *poll,
5134 struct io_poll_table *ipt, __poll_t mask,
5135 wait_queue_func_t wake_func)
5136 __acquires(&ctx->completion_lock)
5138 struct io_ring_ctx *ctx = req->ctx;
5139 bool cancel = false;
5141 INIT_HLIST_NODE(&req->hash_node);
5142 io_init_poll_iocb(poll, mask, wake_func);
5143 poll->file = req->file;
5144 poll->wait.private = req;
5146 ipt->pt._key = mask;
5149 ipt->nr_entries = 0;
5151 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5152 if (unlikely(!ipt->nr_entries) && !ipt->error)
5153 ipt->error = -EINVAL;
5155 spin_lock_irq(&ctx->completion_lock);
5156 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5157 io_poll_remove_double(req);
5158 if (likely(poll->head)) {
5159 spin_lock(&poll->head->lock);
5160 if (unlikely(list_empty(&poll->wait.entry))) {
5166 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5167 list_del_init(&poll->wait.entry);
5169 WRITE_ONCE(poll->canceled, true);
5170 else if (!poll->done) /* actually waiting for an event */
5171 io_poll_req_insert(req);
5172 spin_unlock(&poll->head->lock);
5184 static int io_arm_poll_handler(struct io_kiocb *req)
5186 const struct io_op_def *def = &io_op_defs[req->opcode];
5187 struct io_ring_ctx *ctx = req->ctx;
5188 struct async_poll *apoll;
5189 struct io_poll_table ipt;
5190 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5193 if (!req->file || !file_can_poll(req->file))
5194 return IO_APOLL_ABORTED;
5195 if (req->flags & REQ_F_POLLED)
5196 return IO_APOLL_ABORTED;
5197 if (!def->pollin && !def->pollout)
5198 return IO_APOLL_ABORTED;
5202 mask |= POLLIN | POLLRDNORM;
5204 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5205 if ((req->opcode == IORING_OP_RECVMSG) &&
5206 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5210 mask |= POLLOUT | POLLWRNORM;
5213 /* if we can't nonblock try, then no point in arming a poll handler */
5214 if (!io_file_supports_async(req, rw))
5215 return IO_APOLL_ABORTED;
5217 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5218 if (unlikely(!apoll))
5219 return IO_APOLL_ABORTED;
5220 apoll->double_poll = NULL;
5222 req->flags |= REQ_F_POLLED;
5223 ipt.pt._qproc = io_async_queue_proc;
5225 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5227 if (ret || ipt.error) {
5228 spin_unlock_irq(&ctx->completion_lock);
5230 return IO_APOLL_READY;
5231 return IO_APOLL_ABORTED;
5233 spin_unlock_irq(&ctx->completion_lock);
5234 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5235 mask, apoll->poll.events);
5239 static bool __io_poll_remove_one(struct io_kiocb *req,
5240 struct io_poll_iocb *poll, bool do_cancel)
5241 __must_hold(&req->ctx->completion_lock)
5243 bool do_complete = false;
5247 spin_lock(&poll->head->lock);
5249 WRITE_ONCE(poll->canceled, true);
5250 if (!list_empty(&poll->wait.entry)) {
5251 list_del_init(&poll->wait.entry);
5254 spin_unlock(&poll->head->lock);
5255 hash_del(&req->hash_node);
5259 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5260 __must_hold(&req->ctx->completion_lock)
5264 io_poll_remove_double(req);
5265 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5267 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5268 /* non-poll requests have submit ref still */
5274 static bool io_poll_remove_one(struct io_kiocb *req)
5275 __must_hold(&req->ctx->completion_lock)
5279 do_complete = io_poll_remove_waitqs(req);
5281 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5282 io_commit_cqring(req->ctx);
5284 io_put_req_deferred(req, 1);
5291 * Returns true if we found and killed one or more poll requests
5293 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5296 struct hlist_node *tmp;
5297 struct io_kiocb *req;
5300 spin_lock_irq(&ctx->completion_lock);
5301 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5302 struct hlist_head *list;
5304 list = &ctx->cancel_hash[i];
5305 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5306 if (io_match_task(req, tsk, cancel_all))
5307 posted += io_poll_remove_one(req);
5310 spin_unlock_irq(&ctx->completion_lock);
5313 io_cqring_ev_posted(ctx);
5318 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5320 __must_hold(&ctx->completion_lock)
5322 struct hlist_head *list;
5323 struct io_kiocb *req;
5325 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5326 hlist_for_each_entry(req, list, hash_node) {
5327 if (sqe_addr != req->user_data)
5329 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5336 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5338 __must_hold(&ctx->completion_lock)
5340 struct io_kiocb *req;
5342 req = io_poll_find(ctx, sqe_addr, poll_only);
5345 if (io_poll_remove_one(req))
5351 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5356 events = READ_ONCE(sqe->poll32_events);
5358 events = swahw32(events);
5360 if (!(flags & IORING_POLL_ADD_MULTI))
5361 events |= EPOLLONESHOT;
5362 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5365 static int io_poll_update_prep(struct io_kiocb *req,
5366 const struct io_uring_sqe *sqe)
5368 struct io_poll_update *upd = &req->poll_update;
5371 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5373 if (sqe->ioprio || sqe->buf_index)
5375 flags = READ_ONCE(sqe->len);
5376 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5377 IORING_POLL_ADD_MULTI))
5379 /* meaningless without update */
5380 if (flags == IORING_POLL_ADD_MULTI)
5383 upd->old_user_data = READ_ONCE(sqe->addr);
5384 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5385 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5387 upd->new_user_data = READ_ONCE(sqe->off);
5388 if (!upd->update_user_data && upd->new_user_data)
5390 if (upd->update_events)
5391 upd->events = io_poll_parse_events(sqe, flags);
5392 else if (sqe->poll32_events)
5398 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5401 struct io_kiocb *req = wait->private;
5402 struct io_poll_iocb *poll = &req->poll;
5404 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5407 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5408 struct poll_table_struct *p)
5410 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5412 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5415 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5417 struct io_poll_iocb *poll = &req->poll;
5420 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5422 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5424 flags = READ_ONCE(sqe->len);
5425 if (flags & ~IORING_POLL_ADD_MULTI)
5428 poll->events = io_poll_parse_events(sqe, flags);
5432 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5434 struct io_poll_iocb *poll = &req->poll;
5435 struct io_ring_ctx *ctx = req->ctx;
5436 struct io_poll_table ipt;
5439 ipt.pt._qproc = io_poll_queue_proc;
5441 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5444 if (mask) { /* no async, we'd stolen it */
5446 io_poll_complete(req, mask);
5448 spin_unlock_irq(&ctx->completion_lock);
5451 io_cqring_ev_posted(ctx);
5452 if (poll->events & EPOLLONESHOT)
5458 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5460 struct io_ring_ctx *ctx = req->ctx;
5461 struct io_kiocb *preq;
5465 spin_lock_irq(&ctx->completion_lock);
5466 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5472 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5474 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5479 * Don't allow racy completion with singleshot, as we cannot safely
5480 * update those. For multishot, if we're racing with completion, just
5481 * let completion re-add it.
5483 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5484 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5488 /* we now have a detached poll request. reissue. */
5492 spin_unlock_irq(&ctx->completion_lock);
5494 io_req_complete(req, ret);
5497 /* only mask one event flags, keep behavior flags */
5498 if (req->poll_update.update_events) {
5499 preq->poll.events &= ~0xffff;
5500 preq->poll.events |= req->poll_update.events & 0xffff;
5501 preq->poll.events |= IO_POLL_UNMASK;
5503 if (req->poll_update.update_user_data)
5504 preq->user_data = req->poll_update.new_user_data;
5505 spin_unlock_irq(&ctx->completion_lock);
5507 /* complete update request, we're done with it */
5508 io_req_complete(req, ret);
5511 ret = io_poll_add(preq, issue_flags);
5514 io_req_complete(preq, ret);
5520 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5522 struct io_timeout_data *data = container_of(timer,
5523 struct io_timeout_data, timer);
5524 struct io_kiocb *req = data->req;
5525 struct io_ring_ctx *ctx = req->ctx;
5526 unsigned long flags;
5528 spin_lock_irqsave(&ctx->completion_lock, flags);
5529 list_del_init(&req->timeout.list);
5530 atomic_set(&req->ctx->cq_timeouts,
5531 atomic_read(&req->ctx->cq_timeouts) + 1);
5533 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5534 io_commit_cqring(ctx);
5535 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5537 io_cqring_ev_posted(ctx);
5540 return HRTIMER_NORESTART;
5543 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5545 __must_hold(&ctx->completion_lock)
5547 struct io_timeout_data *io;
5548 struct io_kiocb *req;
5551 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5552 found = user_data == req->user_data;
5557 return ERR_PTR(-ENOENT);
5559 io = req->async_data;
5560 if (hrtimer_try_to_cancel(&io->timer) == -1)
5561 return ERR_PTR(-EALREADY);
5562 list_del_init(&req->timeout.list);
5566 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5567 __must_hold(&ctx->completion_lock)
5569 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5572 return PTR_ERR(req);
5575 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5576 io_put_req_deferred(req, 1);
5580 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5581 struct timespec64 *ts, enum hrtimer_mode mode)
5582 __must_hold(&ctx->completion_lock)
5584 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5585 struct io_timeout_data *data;
5588 return PTR_ERR(req);
5590 req->timeout.off = 0; /* noseq */
5591 data = req->async_data;
5592 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5593 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5594 data->timer.function = io_timeout_fn;
5595 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5599 static int io_timeout_remove_prep(struct io_kiocb *req,
5600 const struct io_uring_sqe *sqe)
5602 struct io_timeout_rem *tr = &req->timeout_rem;
5604 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5606 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5608 if (sqe->ioprio || sqe->buf_index || sqe->len)
5611 tr->addr = READ_ONCE(sqe->addr);
5612 tr->flags = READ_ONCE(sqe->timeout_flags);
5613 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5614 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5616 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5618 } else if (tr->flags) {
5619 /* timeout removal doesn't support flags */
5626 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5628 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5633 * Remove or update an existing timeout command
5635 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5637 struct io_timeout_rem *tr = &req->timeout_rem;
5638 struct io_ring_ctx *ctx = req->ctx;
5641 spin_lock_irq(&ctx->completion_lock);
5642 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5643 ret = io_timeout_cancel(ctx, tr->addr);
5645 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5646 io_translate_timeout_mode(tr->flags));
5648 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5649 io_commit_cqring(ctx);
5650 spin_unlock_irq(&ctx->completion_lock);
5651 io_cqring_ev_posted(ctx);
5658 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5659 bool is_timeout_link)
5661 struct io_timeout_data *data;
5663 u32 off = READ_ONCE(sqe->off);
5665 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5667 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5669 if (off && is_timeout_link)
5671 flags = READ_ONCE(sqe->timeout_flags);
5672 if (flags & ~IORING_TIMEOUT_ABS)
5675 req->timeout.off = off;
5676 if (unlikely(off && !req->ctx->off_timeout_used))
5677 req->ctx->off_timeout_used = true;
5679 if (!req->async_data && io_alloc_async_data(req))
5682 data = req->async_data;
5685 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5688 data->mode = io_translate_timeout_mode(flags);
5689 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5690 if (is_timeout_link)
5691 io_req_track_inflight(req);
5695 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5697 struct io_ring_ctx *ctx = req->ctx;
5698 struct io_timeout_data *data = req->async_data;
5699 struct list_head *entry;
5700 u32 tail, off = req->timeout.off;
5702 spin_lock_irq(&ctx->completion_lock);
5705 * sqe->off holds how many events that need to occur for this
5706 * timeout event to be satisfied. If it isn't set, then this is
5707 * a pure timeout request, sequence isn't used.
5709 if (io_is_timeout_noseq(req)) {
5710 entry = ctx->timeout_list.prev;
5714 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5715 req->timeout.target_seq = tail + off;
5717 /* Update the last seq here in case io_flush_timeouts() hasn't.
5718 * This is safe because ->completion_lock is held, and submissions
5719 * and completions are never mixed in the same ->completion_lock section.
5721 ctx->cq_last_tm_flush = tail;
5724 * Insertion sort, ensuring the first entry in the list is always
5725 * the one we need first.
5727 list_for_each_prev(entry, &ctx->timeout_list) {
5728 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5731 if (io_is_timeout_noseq(nxt))
5733 /* nxt.seq is behind @tail, otherwise would've been completed */
5734 if (off >= nxt->timeout.target_seq - tail)
5738 list_add(&req->timeout.list, entry);
5739 data->timer.function = io_timeout_fn;
5740 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5741 spin_unlock_irq(&ctx->completion_lock);
5745 struct io_cancel_data {
5746 struct io_ring_ctx *ctx;
5750 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5752 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5753 struct io_cancel_data *cd = data;
5755 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5758 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5759 struct io_ring_ctx *ctx)
5761 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5762 enum io_wq_cancel cancel_ret;
5765 if (!tctx || !tctx->io_wq)
5768 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5769 switch (cancel_ret) {
5770 case IO_WQ_CANCEL_OK:
5773 case IO_WQ_CANCEL_RUNNING:
5776 case IO_WQ_CANCEL_NOTFOUND:
5784 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5785 struct io_kiocb *req, __u64 sqe_addr,
5788 unsigned long flags;
5791 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5792 spin_lock_irqsave(&ctx->completion_lock, flags);
5795 ret = io_timeout_cancel(ctx, sqe_addr);
5798 ret = io_poll_cancel(ctx, sqe_addr, false);
5802 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5803 io_commit_cqring(ctx);
5804 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5805 io_cqring_ev_posted(ctx);
5811 static int io_async_cancel_prep(struct io_kiocb *req,
5812 const struct io_uring_sqe *sqe)
5814 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5816 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5818 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5821 req->cancel.addr = READ_ONCE(sqe->addr);
5825 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5827 struct io_ring_ctx *ctx = req->ctx;
5828 u64 sqe_addr = req->cancel.addr;
5829 struct io_tctx_node *node;
5832 /* tasks should wait for their io-wq threads, so safe w/o sync */
5833 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5834 spin_lock_irq(&ctx->completion_lock);
5837 ret = io_timeout_cancel(ctx, sqe_addr);
5840 ret = io_poll_cancel(ctx, sqe_addr, false);
5843 spin_unlock_irq(&ctx->completion_lock);
5845 /* slow path, try all io-wq's */
5846 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5848 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5849 struct io_uring_task *tctx = node->task->io_uring;
5851 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5855 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5857 spin_lock_irq(&ctx->completion_lock);
5859 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5860 io_commit_cqring(ctx);
5861 spin_unlock_irq(&ctx->completion_lock);
5862 io_cqring_ev_posted(ctx);
5870 static int io_rsrc_update_prep(struct io_kiocb *req,
5871 const struct io_uring_sqe *sqe)
5873 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5875 if (sqe->ioprio || sqe->rw_flags)
5878 req->rsrc_update.offset = READ_ONCE(sqe->off);
5879 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5880 if (!req->rsrc_update.nr_args)
5882 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5886 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5888 struct io_ring_ctx *ctx = req->ctx;
5889 struct io_uring_rsrc_update2 up;
5892 if (issue_flags & IO_URING_F_NONBLOCK)
5895 up.offset = req->rsrc_update.offset;
5896 up.data = req->rsrc_update.arg;
5901 mutex_lock(&ctx->uring_lock);
5902 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5903 &up, req->rsrc_update.nr_args);
5904 mutex_unlock(&ctx->uring_lock);
5908 __io_req_complete(req, issue_flags, ret, 0);
5912 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5914 switch (req->opcode) {
5917 case IORING_OP_READV:
5918 case IORING_OP_READ_FIXED:
5919 case IORING_OP_READ:
5920 return io_read_prep(req, sqe);
5921 case IORING_OP_WRITEV:
5922 case IORING_OP_WRITE_FIXED:
5923 case IORING_OP_WRITE:
5924 return io_write_prep(req, sqe);
5925 case IORING_OP_POLL_ADD:
5926 return io_poll_add_prep(req, sqe);
5927 case IORING_OP_POLL_REMOVE:
5928 return io_poll_update_prep(req, sqe);
5929 case IORING_OP_FSYNC:
5930 return io_fsync_prep(req, sqe);
5931 case IORING_OP_SYNC_FILE_RANGE:
5932 return io_sfr_prep(req, sqe);
5933 case IORING_OP_SENDMSG:
5934 case IORING_OP_SEND:
5935 return io_sendmsg_prep(req, sqe);
5936 case IORING_OP_RECVMSG:
5937 case IORING_OP_RECV:
5938 return io_recvmsg_prep(req, sqe);
5939 case IORING_OP_CONNECT:
5940 return io_connect_prep(req, sqe);
5941 case IORING_OP_TIMEOUT:
5942 return io_timeout_prep(req, sqe, false);
5943 case IORING_OP_TIMEOUT_REMOVE:
5944 return io_timeout_remove_prep(req, sqe);
5945 case IORING_OP_ASYNC_CANCEL:
5946 return io_async_cancel_prep(req, sqe);
5947 case IORING_OP_LINK_TIMEOUT:
5948 return io_timeout_prep(req, sqe, true);
5949 case IORING_OP_ACCEPT:
5950 return io_accept_prep(req, sqe);
5951 case IORING_OP_FALLOCATE:
5952 return io_fallocate_prep(req, sqe);
5953 case IORING_OP_OPENAT:
5954 return io_openat_prep(req, sqe);
5955 case IORING_OP_CLOSE:
5956 return io_close_prep(req, sqe);
5957 case IORING_OP_FILES_UPDATE:
5958 return io_rsrc_update_prep(req, sqe);
5959 case IORING_OP_STATX:
5960 return io_statx_prep(req, sqe);
5961 case IORING_OP_FADVISE:
5962 return io_fadvise_prep(req, sqe);
5963 case IORING_OP_MADVISE:
5964 return io_madvise_prep(req, sqe);
5965 case IORING_OP_OPENAT2:
5966 return io_openat2_prep(req, sqe);
5967 case IORING_OP_EPOLL_CTL:
5968 return io_epoll_ctl_prep(req, sqe);
5969 case IORING_OP_SPLICE:
5970 return io_splice_prep(req, sqe);
5971 case IORING_OP_PROVIDE_BUFFERS:
5972 return io_provide_buffers_prep(req, sqe);
5973 case IORING_OP_REMOVE_BUFFERS:
5974 return io_remove_buffers_prep(req, sqe);
5976 return io_tee_prep(req, sqe);
5977 case IORING_OP_SHUTDOWN:
5978 return io_shutdown_prep(req, sqe);
5979 case IORING_OP_RENAMEAT:
5980 return io_renameat_prep(req, sqe);
5981 case IORING_OP_UNLINKAT:
5982 return io_unlinkat_prep(req, sqe);
5985 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5990 static int io_req_prep_async(struct io_kiocb *req)
5992 if (!io_op_defs[req->opcode].needs_async_setup)
5994 if (WARN_ON_ONCE(req->async_data))
5996 if (io_alloc_async_data(req))
5999 switch (req->opcode) {
6000 case IORING_OP_READV:
6001 return io_rw_prep_async(req, READ);
6002 case IORING_OP_WRITEV:
6003 return io_rw_prep_async(req, WRITE);
6004 case IORING_OP_SENDMSG:
6005 return io_sendmsg_prep_async(req);
6006 case IORING_OP_RECVMSG:
6007 return io_recvmsg_prep_async(req);
6008 case IORING_OP_CONNECT:
6009 return io_connect_prep_async(req);
6011 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6016 static u32 io_get_sequence(struct io_kiocb *req)
6018 u32 seq = req->ctx->cached_sq_head;
6020 /* need original cached_sq_head, but it was increased for each req */
6021 io_for_each_link(req, req)
6026 static bool io_drain_req(struct io_kiocb *req)
6028 struct io_kiocb *pos;
6029 struct io_ring_ctx *ctx = req->ctx;
6030 struct io_defer_entry *de;
6035 * If we need to drain a request in the middle of a link, drain the
6036 * head request and the next request/link after the current link.
6037 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6038 * maintained for every request of our link.
6040 if (ctx->drain_next) {
6041 req->flags |= REQ_F_IO_DRAIN;
6042 ctx->drain_next = false;
6044 /* not interested in head, start from the first linked */
6045 io_for_each_link(pos, req->link) {
6046 if (pos->flags & REQ_F_IO_DRAIN) {
6047 ctx->drain_next = true;
6048 req->flags |= REQ_F_IO_DRAIN;
6053 /* Still need defer if there is pending req in defer list. */
6054 if (likely(list_empty_careful(&ctx->defer_list) &&
6055 !(req->flags & REQ_F_IO_DRAIN))) {
6056 ctx->drain_active = false;
6060 seq = io_get_sequence(req);
6061 /* Still a chance to pass the sequence check */
6062 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6065 ret = io_req_prep_async(req);
6068 io_prep_async_link(req);
6069 de = kmalloc(sizeof(*de), GFP_KERNEL);
6073 io_req_complete_failed(req, ret);
6077 spin_lock_irq(&ctx->completion_lock);
6078 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6079 spin_unlock_irq(&ctx->completion_lock);
6081 io_queue_async_work(req);
6085 trace_io_uring_defer(ctx, req, req->user_data);
6088 list_add_tail(&de->list, &ctx->defer_list);
6089 spin_unlock_irq(&ctx->completion_lock);
6093 static void io_clean_op(struct io_kiocb *req)
6095 if (req->flags & REQ_F_BUFFER_SELECTED) {
6096 switch (req->opcode) {
6097 case IORING_OP_READV:
6098 case IORING_OP_READ_FIXED:
6099 case IORING_OP_READ:
6100 kfree((void *)(unsigned long)req->rw.addr);
6102 case IORING_OP_RECVMSG:
6103 case IORING_OP_RECV:
6104 kfree(req->sr_msg.kbuf);
6109 if (req->flags & REQ_F_NEED_CLEANUP) {
6110 switch (req->opcode) {
6111 case IORING_OP_READV:
6112 case IORING_OP_READ_FIXED:
6113 case IORING_OP_READ:
6114 case IORING_OP_WRITEV:
6115 case IORING_OP_WRITE_FIXED:
6116 case IORING_OP_WRITE: {
6117 struct io_async_rw *io = req->async_data;
6119 kfree(io->free_iovec);
6122 case IORING_OP_RECVMSG:
6123 case IORING_OP_SENDMSG: {
6124 struct io_async_msghdr *io = req->async_data;
6126 kfree(io->free_iov);
6129 case IORING_OP_SPLICE:
6131 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6132 io_put_file(req->splice.file_in);
6134 case IORING_OP_OPENAT:
6135 case IORING_OP_OPENAT2:
6136 if (req->open.filename)
6137 putname(req->open.filename);
6139 case IORING_OP_RENAMEAT:
6140 putname(req->rename.oldpath);
6141 putname(req->rename.newpath);
6143 case IORING_OP_UNLINKAT:
6144 putname(req->unlink.filename);
6148 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6149 kfree(req->apoll->double_poll);
6153 if (req->flags & REQ_F_INFLIGHT) {
6154 struct io_uring_task *tctx = req->task->io_uring;
6156 atomic_dec(&tctx->inflight_tracked);
6158 if (req->flags & REQ_F_CREDS)
6159 put_cred(req->creds);
6161 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6164 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6166 struct io_ring_ctx *ctx = req->ctx;
6167 const struct cred *creds = NULL;
6170 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6171 creds = override_creds(req->creds);
6173 switch (req->opcode) {
6175 ret = io_nop(req, issue_flags);
6177 case IORING_OP_READV:
6178 case IORING_OP_READ_FIXED:
6179 case IORING_OP_READ:
6180 ret = io_read(req, issue_flags);
6182 case IORING_OP_WRITEV:
6183 case IORING_OP_WRITE_FIXED:
6184 case IORING_OP_WRITE:
6185 ret = io_write(req, issue_flags);
6187 case IORING_OP_FSYNC:
6188 ret = io_fsync(req, issue_flags);
6190 case IORING_OP_POLL_ADD:
6191 ret = io_poll_add(req, issue_flags);
6193 case IORING_OP_POLL_REMOVE:
6194 ret = io_poll_update(req, issue_flags);
6196 case IORING_OP_SYNC_FILE_RANGE:
6197 ret = io_sync_file_range(req, issue_flags);
6199 case IORING_OP_SENDMSG:
6200 ret = io_sendmsg(req, issue_flags);
6202 case IORING_OP_SEND:
6203 ret = io_send(req, issue_flags);
6205 case IORING_OP_RECVMSG:
6206 ret = io_recvmsg(req, issue_flags);
6208 case IORING_OP_RECV:
6209 ret = io_recv(req, issue_flags);
6211 case IORING_OP_TIMEOUT:
6212 ret = io_timeout(req, issue_flags);
6214 case IORING_OP_TIMEOUT_REMOVE:
6215 ret = io_timeout_remove(req, issue_flags);
6217 case IORING_OP_ACCEPT:
6218 ret = io_accept(req, issue_flags);
6220 case IORING_OP_CONNECT:
6221 ret = io_connect(req, issue_flags);
6223 case IORING_OP_ASYNC_CANCEL:
6224 ret = io_async_cancel(req, issue_flags);
6226 case IORING_OP_FALLOCATE:
6227 ret = io_fallocate(req, issue_flags);
6229 case IORING_OP_OPENAT:
6230 ret = io_openat(req, issue_flags);
6232 case IORING_OP_CLOSE:
6233 ret = io_close(req, issue_flags);
6235 case IORING_OP_FILES_UPDATE:
6236 ret = io_files_update(req, issue_flags);
6238 case IORING_OP_STATX:
6239 ret = io_statx(req, issue_flags);
6241 case IORING_OP_FADVISE:
6242 ret = io_fadvise(req, issue_flags);
6244 case IORING_OP_MADVISE:
6245 ret = io_madvise(req, issue_flags);
6247 case IORING_OP_OPENAT2:
6248 ret = io_openat2(req, issue_flags);
6250 case IORING_OP_EPOLL_CTL:
6251 ret = io_epoll_ctl(req, issue_flags);
6253 case IORING_OP_SPLICE:
6254 ret = io_splice(req, issue_flags);
6256 case IORING_OP_PROVIDE_BUFFERS:
6257 ret = io_provide_buffers(req, issue_flags);
6259 case IORING_OP_REMOVE_BUFFERS:
6260 ret = io_remove_buffers(req, issue_flags);
6263 ret = io_tee(req, issue_flags);
6265 case IORING_OP_SHUTDOWN:
6266 ret = io_shutdown(req, issue_flags);
6268 case IORING_OP_RENAMEAT:
6269 ret = io_renameat(req, issue_flags);
6271 case IORING_OP_UNLINKAT:
6272 ret = io_unlinkat(req, issue_flags);
6280 revert_creds(creds);
6283 /* If the op doesn't have a file, we're not polling for it */
6284 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6285 io_iopoll_req_issued(req);
6290 static void io_wq_submit_work(struct io_wq_work *work)
6292 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6293 struct io_kiocb *timeout;
6296 timeout = io_prep_linked_timeout(req);
6298 io_queue_linked_timeout(timeout);
6300 if (work->flags & IO_WQ_WORK_CANCEL)
6305 ret = io_issue_sqe(req, 0);
6307 * We can get EAGAIN for polled IO even though we're
6308 * forcing a sync submission from here, since we can't
6309 * wait for request slots on the block side.
6317 /* avoid locking problems by failing it from a clean context */
6319 /* io-wq is going to take one down */
6321 io_req_task_queue_fail(req, ret);
6325 #define FFS_ASYNC_READ 0x1UL
6326 #define FFS_ASYNC_WRITE 0x2UL
6328 #define FFS_ISREG 0x4UL
6330 #define FFS_ISREG 0x0UL
6332 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6334 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6337 struct io_fixed_file *table_l2;
6339 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6340 return &table_l2[i & IORING_FILE_TABLE_MASK];
6343 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6346 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6348 return (struct file *) (slot->file_ptr & FFS_MASK);
6351 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6353 unsigned long file_ptr = (unsigned long) file;
6355 if (__io_file_supports_async(file, READ))
6356 file_ptr |= FFS_ASYNC_READ;
6357 if (__io_file_supports_async(file, WRITE))
6358 file_ptr |= FFS_ASYNC_WRITE;
6359 if (S_ISREG(file_inode(file)->i_mode))
6360 file_ptr |= FFS_ISREG;
6361 file_slot->file_ptr = file_ptr;
6364 static struct file *io_file_get(struct io_submit_state *state,
6365 struct io_kiocb *req, int fd, bool fixed)
6367 struct io_ring_ctx *ctx = req->ctx;
6371 unsigned long file_ptr;
6373 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6375 fd = array_index_nospec(fd, ctx->nr_user_files);
6376 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6377 file = (struct file *) (file_ptr & FFS_MASK);
6378 file_ptr &= ~FFS_MASK;
6379 /* mask in overlapping REQ_F and FFS bits */
6380 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6381 io_req_set_rsrc_node(req);
6383 trace_io_uring_file_get(ctx, fd);
6384 file = __io_file_get(state, fd);
6386 /* we don't allow fixed io_uring files */
6387 if (file && unlikely(file->f_op == &io_uring_fops))
6388 io_req_track_inflight(req);
6394 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6396 struct io_timeout_data *data = container_of(timer,
6397 struct io_timeout_data, timer);
6398 struct io_kiocb *prev, *req = data->req;
6399 struct io_ring_ctx *ctx = req->ctx;
6400 unsigned long flags;
6402 spin_lock_irqsave(&ctx->completion_lock, flags);
6403 prev = req->timeout.head;
6404 req->timeout.head = NULL;
6407 * We don't expect the list to be empty, that will only happen if we
6408 * race with the completion of the linked work.
6411 io_remove_next_linked(prev);
6412 if (!req_ref_inc_not_zero(prev))
6415 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6418 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6419 io_put_req_deferred(prev, 1);
6420 io_put_req_deferred(req, 1);
6422 io_req_complete_post(req, -ETIME, 0);
6424 return HRTIMER_NORESTART;
6427 static void io_queue_linked_timeout(struct io_kiocb *req)
6429 struct io_ring_ctx *ctx = req->ctx;
6431 spin_lock_irq(&ctx->completion_lock);
6433 * If the back reference is NULL, then our linked request finished
6434 * before we got a chance to setup the timer
6436 if (req->timeout.head) {
6437 struct io_timeout_data *data = req->async_data;
6439 data->timer.function = io_link_timeout_fn;
6440 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6443 spin_unlock_irq(&ctx->completion_lock);
6444 /* drop submission reference */
6448 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6450 struct io_kiocb *nxt = req->link;
6452 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6453 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6456 nxt->timeout.head = req;
6457 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6458 req->flags |= REQ_F_LINK_TIMEOUT;
6462 static void __io_queue_sqe(struct io_kiocb *req)
6464 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6468 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6471 * We async punt it if the file wasn't marked NOWAIT, or if the file
6472 * doesn't support non-blocking read/write attempts
6475 /* drop submission reference */
6476 if (req->flags & REQ_F_COMPLETE_INLINE) {
6477 struct io_ring_ctx *ctx = req->ctx;
6478 struct io_comp_state *cs = &ctx->submit_state.comp;
6480 cs->reqs[cs->nr++] = req;
6481 if (cs->nr == ARRAY_SIZE(cs->reqs))
6482 io_submit_flush_completions(ctx);
6486 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6487 switch (io_arm_poll_handler(req)) {
6488 case IO_APOLL_READY:
6490 case IO_APOLL_ABORTED:
6492 * Queued up for async execution, worker will release
6493 * submit reference when the iocb is actually submitted.
6495 io_queue_async_work(req);
6499 io_req_complete_failed(req, ret);
6502 io_queue_linked_timeout(linked_timeout);
6505 static inline void io_queue_sqe(struct io_kiocb *req)
6507 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6510 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6511 __io_queue_sqe(req);
6513 int ret = io_req_prep_async(req);
6516 io_req_complete_failed(req, ret);
6518 io_queue_async_work(req);
6523 * Check SQE restrictions (opcode and flags).
6525 * Returns 'true' if SQE is allowed, 'false' otherwise.
6527 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6528 struct io_kiocb *req,
6529 unsigned int sqe_flags)
6531 if (likely(!ctx->restricted))
6534 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6537 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6538 ctx->restrictions.sqe_flags_required)
6541 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6542 ctx->restrictions.sqe_flags_required))
6548 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6549 const struct io_uring_sqe *sqe)
6551 struct io_submit_state *state;
6552 unsigned int sqe_flags;
6553 int personality, ret = 0;
6555 req->opcode = READ_ONCE(sqe->opcode);
6556 /* same numerical values with corresponding REQ_F_*, safe to copy */
6557 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6558 req->user_data = READ_ONCE(sqe->user_data);
6560 req->fixed_rsrc_refs = NULL;
6561 /* one is dropped after submission, the other at completion */
6562 atomic_set(&req->refs, 2);
6563 req->task = current;
6565 /* enforce forwards compatibility on users */
6566 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6568 if (unlikely(req->opcode >= IORING_OP_LAST))
6570 if (!io_check_restriction(ctx, req, sqe_flags))
6573 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6574 !io_op_defs[req->opcode].buffer_select)
6576 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6577 ctx->drain_active = true;
6579 personality = READ_ONCE(sqe->personality);
6581 req->creds = xa_load(&ctx->personalities, personality);
6584 get_cred(req->creds);
6585 req->flags |= REQ_F_CREDS;
6587 state = &ctx->submit_state;
6590 * Plug now if we have more than 1 IO left after this, and the target
6591 * is potentially a read/write to block based storage.
6593 if (!state->plug_started && state->ios_left > 1 &&
6594 io_op_defs[req->opcode].plug) {
6595 blk_start_plug(&state->plug);
6596 state->plug_started = true;
6599 if (io_op_defs[req->opcode].needs_file) {
6600 bool fixed = req->flags & REQ_F_FIXED_FILE;
6602 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6603 if (unlikely(!req->file))
6611 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6612 const struct io_uring_sqe *sqe)
6614 struct io_submit_link *link = &ctx->submit_state.link;
6617 ret = io_init_req(ctx, req, sqe);
6618 if (unlikely(ret)) {
6621 /* fail even hard links since we don't submit */
6622 req_set_fail(link->head);
6623 io_req_complete_failed(link->head, -ECANCELED);
6626 io_req_complete_failed(req, ret);
6630 ret = io_req_prep(req, sqe);
6634 /* don't need @sqe from now on */
6635 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6637 ctx->flags & IORING_SETUP_SQPOLL);
6640 * If we already have a head request, queue this one for async
6641 * submittal once the head completes. If we don't have a head but
6642 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6643 * submitted sync once the chain is complete. If none of those
6644 * conditions are true (normal request), then just queue it.
6647 struct io_kiocb *head = link->head;
6649 ret = io_req_prep_async(req);
6652 trace_io_uring_link(ctx, req, head);
6653 link->last->link = req;
6656 /* last request of a link, enqueue the link */
6657 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6662 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6674 * Batched submission is done, ensure local IO is flushed out.
6676 static void io_submit_state_end(struct io_submit_state *state,
6677 struct io_ring_ctx *ctx)
6679 if (state->link.head)
6680 io_queue_sqe(state->link.head);
6682 io_submit_flush_completions(ctx);
6683 if (state->plug_started)
6684 blk_finish_plug(&state->plug);
6685 io_state_file_put(state);
6689 * Start submission side cache.
6691 static void io_submit_state_start(struct io_submit_state *state,
6692 unsigned int max_ios)
6694 state->plug_started = false;
6695 state->ios_left = max_ios;
6696 /* set only head, no need to init link_last in advance */
6697 state->link.head = NULL;
6700 static void io_commit_sqring(struct io_ring_ctx *ctx)
6702 struct io_rings *rings = ctx->rings;
6705 * Ensure any loads from the SQEs are done at this point,
6706 * since once we write the new head, the application could
6707 * write new data to them.
6709 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6713 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6714 * that is mapped by userspace. This means that care needs to be taken to
6715 * ensure that reads are stable, as we cannot rely on userspace always
6716 * being a good citizen. If members of the sqe are validated and then later
6717 * used, it's important that those reads are done through READ_ONCE() to
6718 * prevent a re-load down the line.
6720 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6722 unsigned head, mask = ctx->sq_entries - 1;
6723 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6726 * The cached sq head (or cq tail) serves two purposes:
6728 * 1) allows us to batch the cost of updating the user visible
6730 * 2) allows the kernel side to track the head on its own, even
6731 * though the application is the one updating it.
6733 head = READ_ONCE(ctx->sq_array[sq_idx]);
6734 if (likely(head < ctx->sq_entries))
6735 return &ctx->sq_sqes[head];
6737 /* drop invalid entries */
6739 WRITE_ONCE(ctx->rings->sq_dropped,
6740 READ_ONCE(ctx->rings->sq_dropped) + 1);
6744 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6746 struct io_uring_task *tctx;
6749 /* make sure SQ entry isn't read before tail */
6750 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6751 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6754 tctx = current->io_uring;
6755 tctx->cached_refs -= nr;
6756 if (unlikely(tctx->cached_refs < 0)) {
6757 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6759 percpu_counter_add(&tctx->inflight, refill);
6760 refcount_add(refill, ¤t->usage);
6761 tctx->cached_refs += refill;
6763 io_submit_state_start(&ctx->submit_state, nr);
6765 while (submitted < nr) {
6766 const struct io_uring_sqe *sqe;
6767 struct io_kiocb *req;
6769 req = io_alloc_req(ctx);
6770 if (unlikely(!req)) {
6772 submitted = -EAGAIN;
6775 sqe = io_get_sqe(ctx);
6776 if (unlikely(!sqe)) {
6777 kmem_cache_free(req_cachep, req);
6780 /* will complete beyond this point, count as submitted */
6782 if (io_submit_sqe(ctx, req, sqe))
6786 if (unlikely(submitted != nr)) {
6787 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6788 int unused = nr - ref_used;
6790 current->io_uring->cached_refs += unused;
6791 percpu_ref_put_many(&ctx->refs, unused);
6794 io_submit_state_end(&ctx->submit_state, ctx);
6795 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6796 io_commit_sqring(ctx);
6801 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6803 return READ_ONCE(sqd->state);
6806 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6808 /* Tell userspace we may need a wakeup call */
6809 spin_lock_irq(&ctx->completion_lock);
6810 WRITE_ONCE(ctx->rings->sq_flags,
6811 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6812 spin_unlock_irq(&ctx->completion_lock);
6815 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6817 spin_lock_irq(&ctx->completion_lock);
6818 WRITE_ONCE(ctx->rings->sq_flags,
6819 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6820 spin_unlock_irq(&ctx->completion_lock);
6823 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6825 unsigned int to_submit;
6828 to_submit = io_sqring_entries(ctx);
6829 /* if we're handling multiple rings, cap submit size for fairness */
6830 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6831 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6833 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6834 unsigned nr_events = 0;
6835 const struct cred *creds = NULL;
6837 if (ctx->sq_creds != current_cred())
6838 creds = override_creds(ctx->sq_creds);
6840 mutex_lock(&ctx->uring_lock);
6841 if (!list_empty(&ctx->iopoll_list))
6842 io_do_iopoll(ctx, &nr_events, 0, true);
6845 * Don't submit if refs are dying, good for io_uring_register(),
6846 * but also it is relied upon by io_ring_exit_work()
6848 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6849 !(ctx->flags & IORING_SETUP_R_DISABLED))
6850 ret = io_submit_sqes(ctx, to_submit);
6851 mutex_unlock(&ctx->uring_lock);
6853 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6854 wake_up(&ctx->sqo_sq_wait);
6856 revert_creds(creds);
6862 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6864 struct io_ring_ctx *ctx;
6865 unsigned sq_thread_idle = 0;
6867 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6868 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6869 sqd->sq_thread_idle = sq_thread_idle;
6872 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6874 bool did_sig = false;
6875 struct ksignal ksig;
6877 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6878 signal_pending(current)) {
6879 mutex_unlock(&sqd->lock);
6880 if (signal_pending(current))
6881 did_sig = get_signal(&ksig);
6883 mutex_lock(&sqd->lock);
6885 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6888 static int io_sq_thread(void *data)
6890 struct io_sq_data *sqd = data;
6891 struct io_ring_ctx *ctx;
6892 unsigned long timeout = 0;
6893 char buf[TASK_COMM_LEN];
6896 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6897 set_task_comm(current, buf);
6899 if (sqd->sq_cpu != -1)
6900 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6902 set_cpus_allowed_ptr(current, cpu_online_mask);
6903 current->flags |= PF_NO_SETAFFINITY;
6905 mutex_lock(&sqd->lock);
6907 bool cap_entries, sqt_spin = false;
6909 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6910 if (io_sqd_handle_event(sqd))
6912 timeout = jiffies + sqd->sq_thread_idle;
6915 cap_entries = !list_is_singular(&sqd->ctx_list);
6916 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6917 int ret = __io_sq_thread(ctx, cap_entries);
6919 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6922 if (io_run_task_work())
6925 if (sqt_spin || !time_after(jiffies, timeout)) {
6928 timeout = jiffies + sqd->sq_thread_idle;
6932 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6933 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6934 bool needs_sched = true;
6936 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6937 io_ring_set_wakeup_flag(ctx);
6939 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6940 !list_empty_careful(&ctx->iopoll_list)) {
6941 needs_sched = false;
6944 if (io_sqring_entries(ctx)) {
6945 needs_sched = false;
6951 mutex_unlock(&sqd->lock);
6953 mutex_lock(&sqd->lock);
6955 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6956 io_ring_clear_wakeup_flag(ctx);
6959 finish_wait(&sqd->wait, &wait);
6960 timeout = jiffies + sqd->sq_thread_idle;
6963 io_uring_cancel_generic(true, sqd);
6965 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6966 io_ring_set_wakeup_flag(ctx);
6968 mutex_unlock(&sqd->lock);
6970 complete(&sqd->exited);
6974 struct io_wait_queue {
6975 struct wait_queue_entry wq;
6976 struct io_ring_ctx *ctx;
6978 unsigned nr_timeouts;
6981 static inline bool io_should_wake(struct io_wait_queue *iowq)
6983 struct io_ring_ctx *ctx = iowq->ctx;
6986 * Wake up if we have enough events, or if a timeout occurred since we
6987 * started waiting. For timeouts, we always want to return to userspace,
6988 * regardless of event count.
6990 return io_cqring_events(ctx) >= iowq->to_wait ||
6991 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6994 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6995 int wake_flags, void *key)
6997 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7001 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7002 * the task, and the next invocation will do it.
7004 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7005 return autoremove_wake_function(curr, mode, wake_flags, key);
7009 static int io_run_task_work_sig(void)
7011 if (io_run_task_work())
7013 if (!signal_pending(current))
7015 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7016 return -ERESTARTSYS;
7020 /* when returns >0, the caller should retry */
7021 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7022 struct io_wait_queue *iowq,
7023 signed long *timeout)
7027 /* make sure we run task_work before checking for signals */
7028 ret = io_run_task_work_sig();
7029 if (ret || io_should_wake(iowq))
7031 /* let the caller flush overflows, retry */
7032 if (test_bit(0, &ctx->check_cq_overflow))
7035 *timeout = schedule_timeout(*timeout);
7036 return !*timeout ? -ETIME : 1;
7040 * Wait until events become available, if we don't already have some. The
7041 * application must reap them itself, as they reside on the shared cq ring.
7043 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7044 const sigset_t __user *sig, size_t sigsz,
7045 struct __kernel_timespec __user *uts)
7047 struct io_wait_queue iowq = {
7050 .func = io_wake_function,
7051 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7054 .to_wait = min_events,
7056 struct io_rings *rings = ctx->rings;
7057 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7061 io_cqring_overflow_flush(ctx, false);
7062 if (io_cqring_events(ctx) >= min_events)
7064 if (!io_run_task_work())
7069 #ifdef CONFIG_COMPAT
7070 if (in_compat_syscall())
7071 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7075 ret = set_user_sigmask(sig, sigsz);
7082 struct timespec64 ts;
7084 if (get_timespec64(&ts, uts))
7086 timeout = timespec64_to_jiffies(&ts);
7089 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7090 trace_io_uring_cqring_wait(ctx, min_events);
7092 /* if we can't even flush overflow, don't wait for more */
7093 if (!io_cqring_overflow_flush(ctx, false)) {
7097 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7098 TASK_INTERRUPTIBLE);
7099 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7100 finish_wait(&ctx->cq_wait, &iowq.wq);
7104 restore_saved_sigmask_unless(ret == -EINTR);
7106 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7109 static void io_free_page_table(void **table, size_t size)
7111 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7113 for (i = 0; i < nr_tables; i++)
7118 static void **io_alloc_page_table(size_t size)
7120 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7121 size_t init_size = size;
7124 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7128 for (i = 0; i < nr_tables; i++) {
7129 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7131 table[i] = kzalloc(this_size, GFP_KERNEL);
7133 io_free_page_table(table, init_size);
7141 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7143 percpu_ref_exit(&ref_node->refs);
7147 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7148 struct io_rsrc_data *data_to_kill)
7150 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7151 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7154 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7156 rsrc_node->rsrc_data = data_to_kill;
7157 spin_lock_irq(&ctx->rsrc_ref_lock);
7158 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7159 spin_unlock_irq(&ctx->rsrc_ref_lock);
7161 atomic_inc(&data_to_kill->refs);
7162 percpu_ref_kill(&rsrc_node->refs);
7163 ctx->rsrc_node = NULL;
7166 if (!ctx->rsrc_node) {
7167 ctx->rsrc_node = ctx->rsrc_backup_node;
7168 ctx->rsrc_backup_node = NULL;
7172 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7174 if (ctx->rsrc_backup_node)
7176 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7177 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7180 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7184 /* As we may drop ->uring_lock, other task may have started quiesce */
7188 data->quiesce = true;
7190 ret = io_rsrc_node_switch_start(ctx);
7193 io_rsrc_node_switch(ctx, data);
7195 /* kill initial ref, already quiesced if zero */
7196 if (atomic_dec_and_test(&data->refs))
7198 mutex_unlock(&ctx->uring_lock);
7199 flush_delayed_work(&ctx->rsrc_put_work);
7200 ret = wait_for_completion_interruptible(&data->done);
7202 mutex_lock(&ctx->uring_lock);
7206 atomic_inc(&data->refs);
7207 /* wait for all works potentially completing data->done */
7208 flush_delayed_work(&ctx->rsrc_put_work);
7209 reinit_completion(&data->done);
7211 ret = io_run_task_work_sig();
7212 mutex_lock(&ctx->uring_lock);
7214 data->quiesce = false;
7219 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7221 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7222 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7224 return &data->tags[table_idx][off];
7227 static void io_rsrc_data_free(struct io_rsrc_data *data)
7229 size_t size = data->nr * sizeof(data->tags[0][0]);
7232 io_free_page_table((void **)data->tags, size);
7236 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7237 u64 __user *utags, unsigned nr,
7238 struct io_rsrc_data **pdata)
7240 struct io_rsrc_data *data;
7244 data = kzalloc(sizeof(*data), GFP_KERNEL);
7247 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7255 data->do_put = do_put;
7258 for (i = 0; i < nr; i++) {
7259 u64 *tag_slot = io_get_tag_slot(data, i);
7261 if (copy_from_user(tag_slot, &utags[i],
7267 atomic_set(&data->refs, 1);
7268 init_completion(&data->done);
7272 io_rsrc_data_free(data);
7276 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7278 size_t size = nr_files * sizeof(struct io_fixed_file);
7280 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7281 return !!table->files;
7284 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7286 size_t size = nr_files * sizeof(struct io_fixed_file);
7288 io_free_page_table((void **)table->files, size);
7289 table->files = NULL;
7292 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7294 #if defined(CONFIG_UNIX)
7295 if (ctx->ring_sock) {
7296 struct sock *sock = ctx->ring_sock->sk;
7297 struct sk_buff *skb;
7299 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7305 for (i = 0; i < ctx->nr_user_files; i++) {
7308 file = io_file_from_index(ctx, i);
7313 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7314 io_rsrc_data_free(ctx->file_data);
7315 ctx->file_data = NULL;
7316 ctx->nr_user_files = 0;
7319 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7323 if (!ctx->file_data)
7325 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7327 __io_sqe_files_unregister(ctx);
7331 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7332 __releases(&sqd->lock)
7334 WARN_ON_ONCE(sqd->thread == current);
7337 * Do the dance but not conditional clear_bit() because it'd race with
7338 * other threads incrementing park_pending and setting the bit.
7340 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7341 if (atomic_dec_return(&sqd->park_pending))
7342 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7343 mutex_unlock(&sqd->lock);
7346 static void io_sq_thread_park(struct io_sq_data *sqd)
7347 __acquires(&sqd->lock)
7349 WARN_ON_ONCE(sqd->thread == current);
7351 atomic_inc(&sqd->park_pending);
7352 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7353 mutex_lock(&sqd->lock);
7355 wake_up_process(sqd->thread);
7358 static void io_sq_thread_stop(struct io_sq_data *sqd)
7360 WARN_ON_ONCE(sqd->thread == current);
7361 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7363 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7364 mutex_lock(&sqd->lock);
7366 wake_up_process(sqd->thread);
7367 mutex_unlock(&sqd->lock);
7368 wait_for_completion(&sqd->exited);
7371 static void io_put_sq_data(struct io_sq_data *sqd)
7373 if (refcount_dec_and_test(&sqd->refs)) {
7374 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7376 io_sq_thread_stop(sqd);
7381 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7383 struct io_sq_data *sqd = ctx->sq_data;
7386 io_sq_thread_park(sqd);
7387 list_del_init(&ctx->sqd_list);
7388 io_sqd_update_thread_idle(sqd);
7389 io_sq_thread_unpark(sqd);
7391 io_put_sq_data(sqd);
7392 ctx->sq_data = NULL;
7396 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7398 struct io_ring_ctx *ctx_attach;
7399 struct io_sq_data *sqd;
7402 f = fdget(p->wq_fd);
7404 return ERR_PTR(-ENXIO);
7405 if (f.file->f_op != &io_uring_fops) {
7407 return ERR_PTR(-EINVAL);
7410 ctx_attach = f.file->private_data;
7411 sqd = ctx_attach->sq_data;
7414 return ERR_PTR(-EINVAL);
7416 if (sqd->task_tgid != current->tgid) {
7418 return ERR_PTR(-EPERM);
7421 refcount_inc(&sqd->refs);
7426 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7429 struct io_sq_data *sqd;
7432 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7433 sqd = io_attach_sq_data(p);
7438 /* fall through for EPERM case, setup new sqd/task */
7439 if (PTR_ERR(sqd) != -EPERM)
7443 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7445 return ERR_PTR(-ENOMEM);
7447 atomic_set(&sqd->park_pending, 0);
7448 refcount_set(&sqd->refs, 1);
7449 INIT_LIST_HEAD(&sqd->ctx_list);
7450 mutex_init(&sqd->lock);
7451 init_waitqueue_head(&sqd->wait);
7452 init_completion(&sqd->exited);
7456 #if defined(CONFIG_UNIX)
7458 * Ensure the UNIX gc is aware of our file set, so we are certain that
7459 * the io_uring can be safely unregistered on process exit, even if we have
7460 * loops in the file referencing.
7462 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7464 struct sock *sk = ctx->ring_sock->sk;
7465 struct scm_fp_list *fpl;
7466 struct sk_buff *skb;
7469 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7473 skb = alloc_skb(0, GFP_KERNEL);
7482 fpl->user = get_uid(current_user());
7483 for (i = 0; i < nr; i++) {
7484 struct file *file = io_file_from_index(ctx, i + offset);
7488 fpl->fp[nr_files] = get_file(file);
7489 unix_inflight(fpl->user, fpl->fp[nr_files]);
7494 fpl->max = SCM_MAX_FD;
7495 fpl->count = nr_files;
7496 UNIXCB(skb).fp = fpl;
7497 skb->destructor = unix_destruct_scm;
7498 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7499 skb_queue_head(&sk->sk_receive_queue, skb);
7501 for (i = 0; i < nr_files; i++)
7512 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7513 * causes regular reference counting to break down. We rely on the UNIX
7514 * garbage collection to take care of this problem for us.
7516 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7518 unsigned left, total;
7522 left = ctx->nr_user_files;
7524 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7526 ret = __io_sqe_files_scm(ctx, this_files, total);
7530 total += this_files;
7536 while (total < ctx->nr_user_files) {
7537 struct file *file = io_file_from_index(ctx, total);
7547 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7553 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7555 struct file *file = prsrc->file;
7556 #if defined(CONFIG_UNIX)
7557 struct sock *sock = ctx->ring_sock->sk;
7558 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7559 struct sk_buff *skb;
7562 __skb_queue_head_init(&list);
7565 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7566 * remove this entry and rearrange the file array.
7568 skb = skb_dequeue(head);
7570 struct scm_fp_list *fp;
7572 fp = UNIXCB(skb).fp;
7573 for (i = 0; i < fp->count; i++) {
7576 if (fp->fp[i] != file)
7579 unix_notinflight(fp->user, fp->fp[i]);
7580 left = fp->count - 1 - i;
7582 memmove(&fp->fp[i], &fp->fp[i + 1],
7583 left * sizeof(struct file *));
7590 __skb_queue_tail(&list, skb);
7600 __skb_queue_tail(&list, skb);
7602 skb = skb_dequeue(head);
7605 if (skb_peek(&list)) {
7606 spin_lock_irq(&head->lock);
7607 while ((skb = __skb_dequeue(&list)) != NULL)
7608 __skb_queue_tail(head, skb);
7609 spin_unlock_irq(&head->lock);
7616 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7618 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7619 struct io_ring_ctx *ctx = rsrc_data->ctx;
7620 struct io_rsrc_put *prsrc, *tmp;
7622 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7623 list_del(&prsrc->list);
7626 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7628 io_ring_submit_lock(ctx, lock_ring);
7629 spin_lock_irq(&ctx->completion_lock);
7630 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7632 io_commit_cqring(ctx);
7633 spin_unlock_irq(&ctx->completion_lock);
7634 io_cqring_ev_posted(ctx);
7635 io_ring_submit_unlock(ctx, lock_ring);
7638 rsrc_data->do_put(ctx, prsrc);
7642 io_rsrc_node_destroy(ref_node);
7643 if (atomic_dec_and_test(&rsrc_data->refs))
7644 complete(&rsrc_data->done);
7647 static void io_rsrc_put_work(struct work_struct *work)
7649 struct io_ring_ctx *ctx;
7650 struct llist_node *node;
7652 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7653 node = llist_del_all(&ctx->rsrc_put_llist);
7656 struct io_rsrc_node *ref_node;
7657 struct llist_node *next = node->next;
7659 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7660 __io_rsrc_put_work(ref_node);
7665 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7667 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7668 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7669 unsigned long flags;
7670 bool first_add = false;
7672 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7675 while (!list_empty(&ctx->rsrc_ref_list)) {
7676 node = list_first_entry(&ctx->rsrc_ref_list,
7677 struct io_rsrc_node, node);
7678 /* recycle ref nodes in order */
7681 list_del(&node->node);
7682 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7684 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7687 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7690 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7692 struct io_rsrc_node *ref_node;
7694 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7698 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7703 INIT_LIST_HEAD(&ref_node->node);
7704 INIT_LIST_HEAD(&ref_node->rsrc_list);
7705 ref_node->done = false;
7709 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7710 unsigned nr_args, u64 __user *tags)
7712 __s32 __user *fds = (__s32 __user *) arg;
7721 if (nr_args > IORING_MAX_FIXED_FILES)
7723 ret = io_rsrc_node_switch_start(ctx);
7726 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7732 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7735 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7736 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7740 /* allow sparse sets */
7743 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7750 if (unlikely(!file))
7754 * Don't allow io_uring instances to be registered. If UNIX
7755 * isn't enabled, then this causes a reference cycle and this
7756 * instance can never get freed. If UNIX is enabled we'll
7757 * handle it just fine, but there's still no point in allowing
7758 * a ring fd as it doesn't support regular read/write anyway.
7760 if (file->f_op == &io_uring_fops) {
7764 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7767 ret = io_sqe_files_scm(ctx);
7769 __io_sqe_files_unregister(ctx);
7773 io_rsrc_node_switch(ctx, NULL);
7776 for (i = 0; i < ctx->nr_user_files; i++) {
7777 file = io_file_from_index(ctx, i);
7781 io_free_file_tables(&ctx->file_table, nr_args);
7782 ctx->nr_user_files = 0;
7784 io_rsrc_data_free(ctx->file_data);
7785 ctx->file_data = NULL;
7789 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7792 #if defined(CONFIG_UNIX)
7793 struct sock *sock = ctx->ring_sock->sk;
7794 struct sk_buff_head *head = &sock->sk_receive_queue;
7795 struct sk_buff *skb;
7798 * See if we can merge this file into an existing skb SCM_RIGHTS
7799 * file set. If there's no room, fall back to allocating a new skb
7800 * and filling it in.
7802 spin_lock_irq(&head->lock);
7803 skb = skb_peek(head);
7805 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7807 if (fpl->count < SCM_MAX_FD) {
7808 __skb_unlink(skb, head);
7809 spin_unlock_irq(&head->lock);
7810 fpl->fp[fpl->count] = get_file(file);
7811 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7813 spin_lock_irq(&head->lock);
7814 __skb_queue_head(head, skb);
7819 spin_unlock_irq(&head->lock);
7826 return __io_sqe_files_scm(ctx, 1, index);
7832 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7833 struct io_rsrc_node *node, void *rsrc)
7835 struct io_rsrc_put *prsrc;
7837 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7841 prsrc->tag = *io_get_tag_slot(data, idx);
7843 list_add(&prsrc->list, &node->rsrc_list);
7847 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7848 struct io_uring_rsrc_update2 *up,
7851 u64 __user *tags = u64_to_user_ptr(up->tags);
7852 __s32 __user *fds = u64_to_user_ptr(up->data);
7853 struct io_rsrc_data *data = ctx->file_data;
7854 struct io_fixed_file *file_slot;
7858 bool needs_switch = false;
7860 if (!ctx->file_data)
7862 if (up->offset + nr_args > ctx->nr_user_files)
7865 for (done = 0; done < nr_args; done++) {
7868 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7869 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7873 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7877 if (fd == IORING_REGISTER_FILES_SKIP)
7880 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7881 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7883 if (file_slot->file_ptr) {
7884 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7885 err = io_queue_rsrc_removal(data, up->offset + done,
7886 ctx->rsrc_node, file);
7889 file_slot->file_ptr = 0;
7890 needs_switch = true;
7899 * Don't allow io_uring instances to be registered. If
7900 * UNIX isn't enabled, then this causes a reference
7901 * cycle and this instance can never get freed. If UNIX
7902 * is enabled we'll handle it just fine, but there's
7903 * still no point in allowing a ring fd as it doesn't
7904 * support regular read/write anyway.
7906 if (file->f_op == &io_uring_fops) {
7911 *io_get_tag_slot(data, up->offset + done) = tag;
7912 io_fixed_file_set(file_slot, file);
7913 err = io_sqe_file_register(ctx, file, i);
7915 file_slot->file_ptr = 0;
7923 io_rsrc_node_switch(ctx, data);
7924 return done ? done : err;
7927 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7929 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7931 req = io_put_req_find_next(req);
7932 return req ? &req->work : NULL;
7935 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7936 struct task_struct *task)
7938 struct io_wq_hash *hash;
7939 struct io_wq_data data;
7940 unsigned int concurrency;
7942 mutex_lock(&ctx->uring_lock);
7943 hash = ctx->hash_map;
7945 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7947 mutex_unlock(&ctx->uring_lock);
7948 return ERR_PTR(-ENOMEM);
7950 refcount_set(&hash->refs, 1);
7951 init_waitqueue_head(&hash->wait);
7952 ctx->hash_map = hash;
7954 mutex_unlock(&ctx->uring_lock);
7958 data.free_work = io_free_work;
7959 data.do_work = io_wq_submit_work;
7961 /* Do QD, or 4 * CPUS, whatever is smallest */
7962 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7964 return io_wq_create(concurrency, &data);
7967 static int io_uring_alloc_task_context(struct task_struct *task,
7968 struct io_ring_ctx *ctx)
7970 struct io_uring_task *tctx;
7973 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7974 if (unlikely(!tctx))
7977 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7978 if (unlikely(ret)) {
7983 tctx->io_wq = io_init_wq_offload(ctx, task);
7984 if (IS_ERR(tctx->io_wq)) {
7985 ret = PTR_ERR(tctx->io_wq);
7986 percpu_counter_destroy(&tctx->inflight);
7992 init_waitqueue_head(&tctx->wait);
7993 atomic_set(&tctx->in_idle, 0);
7994 atomic_set(&tctx->inflight_tracked, 0);
7995 task->io_uring = tctx;
7996 spin_lock_init(&tctx->task_lock);
7997 INIT_WQ_LIST(&tctx->task_list);
7998 init_task_work(&tctx->task_work, tctx_task_work);
8002 void __io_uring_free(struct task_struct *tsk)
8004 struct io_uring_task *tctx = tsk->io_uring;
8006 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8007 WARN_ON_ONCE(tctx->io_wq);
8008 WARN_ON_ONCE(tctx->cached_refs);
8010 percpu_counter_destroy(&tctx->inflight);
8012 tsk->io_uring = NULL;
8015 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8016 struct io_uring_params *p)
8020 /* Retain compatibility with failing for an invalid attach attempt */
8021 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8022 IORING_SETUP_ATTACH_WQ) {
8025 f = fdget(p->wq_fd);
8028 if (f.file->f_op != &io_uring_fops) {
8034 if (ctx->flags & IORING_SETUP_SQPOLL) {
8035 struct task_struct *tsk;
8036 struct io_sq_data *sqd;
8039 sqd = io_get_sq_data(p, &attached);
8045 ctx->sq_creds = get_current_cred();
8047 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8048 if (!ctx->sq_thread_idle)
8049 ctx->sq_thread_idle = HZ;
8051 io_sq_thread_park(sqd);
8052 list_add(&ctx->sqd_list, &sqd->ctx_list);
8053 io_sqd_update_thread_idle(sqd);
8054 /* don't attach to a dying SQPOLL thread, would be racy */
8055 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8056 io_sq_thread_unpark(sqd);
8063 if (p->flags & IORING_SETUP_SQ_AFF) {
8064 int cpu = p->sq_thread_cpu;
8067 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8074 sqd->task_pid = current->pid;
8075 sqd->task_tgid = current->tgid;
8076 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8083 ret = io_uring_alloc_task_context(tsk, ctx);
8084 wake_up_new_task(tsk);
8087 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8088 /* Can't have SQ_AFF without SQPOLL */
8095 complete(&ctx->sq_data->exited);
8097 io_sq_thread_finish(ctx);
8101 static inline void __io_unaccount_mem(struct user_struct *user,
8102 unsigned long nr_pages)
8104 atomic_long_sub(nr_pages, &user->locked_vm);
8107 static inline int __io_account_mem(struct user_struct *user,
8108 unsigned long nr_pages)
8110 unsigned long page_limit, cur_pages, new_pages;
8112 /* Don't allow more pages than we can safely lock */
8113 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8116 cur_pages = atomic_long_read(&user->locked_vm);
8117 new_pages = cur_pages + nr_pages;
8118 if (new_pages > page_limit)
8120 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8121 new_pages) != cur_pages);
8126 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8129 __io_unaccount_mem(ctx->user, nr_pages);
8131 if (ctx->mm_account)
8132 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8135 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8140 ret = __io_account_mem(ctx->user, nr_pages);
8145 if (ctx->mm_account)
8146 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8151 static void io_mem_free(void *ptr)
8158 page = virt_to_head_page(ptr);
8159 if (put_page_testzero(page))
8160 free_compound_page(page);
8163 static void *io_mem_alloc(size_t size)
8165 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8166 __GFP_NORETRY | __GFP_ACCOUNT;
8168 return (void *) __get_free_pages(gfp_flags, get_order(size));
8171 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8174 struct io_rings *rings;
8175 size_t off, sq_array_size;
8177 off = struct_size(rings, cqes, cq_entries);
8178 if (off == SIZE_MAX)
8182 off = ALIGN(off, SMP_CACHE_BYTES);
8190 sq_array_size = array_size(sizeof(u32), sq_entries);
8191 if (sq_array_size == SIZE_MAX)
8194 if (check_add_overflow(off, sq_array_size, &off))
8200 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8202 struct io_mapped_ubuf *imu = *slot;
8205 if (imu != ctx->dummy_ubuf) {
8206 for (i = 0; i < imu->nr_bvecs; i++)
8207 unpin_user_page(imu->bvec[i].bv_page);
8208 if (imu->acct_pages)
8209 io_unaccount_mem(ctx, imu->acct_pages);
8215 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8217 io_buffer_unmap(ctx, &prsrc->buf);
8221 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8225 for (i = 0; i < ctx->nr_user_bufs; i++)
8226 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8227 kfree(ctx->user_bufs);
8228 io_rsrc_data_free(ctx->buf_data);
8229 ctx->user_bufs = NULL;
8230 ctx->buf_data = NULL;
8231 ctx->nr_user_bufs = 0;
8234 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8241 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8243 __io_sqe_buffers_unregister(ctx);
8247 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8248 void __user *arg, unsigned index)
8250 struct iovec __user *src;
8252 #ifdef CONFIG_COMPAT
8254 struct compat_iovec __user *ciovs;
8255 struct compat_iovec ciov;
8257 ciovs = (struct compat_iovec __user *) arg;
8258 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8261 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8262 dst->iov_len = ciov.iov_len;
8266 src = (struct iovec __user *) arg;
8267 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8273 * Not super efficient, but this is just a registration time. And we do cache
8274 * the last compound head, so generally we'll only do a full search if we don't
8277 * We check if the given compound head page has already been accounted, to
8278 * avoid double accounting it. This allows us to account the full size of the
8279 * page, not just the constituent pages of a huge page.
8281 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8282 int nr_pages, struct page *hpage)
8286 /* check current page array */
8287 for (i = 0; i < nr_pages; i++) {
8288 if (!PageCompound(pages[i]))
8290 if (compound_head(pages[i]) == hpage)
8294 /* check previously registered pages */
8295 for (i = 0; i < ctx->nr_user_bufs; i++) {
8296 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8298 for (j = 0; j < imu->nr_bvecs; j++) {
8299 if (!PageCompound(imu->bvec[j].bv_page))
8301 if (compound_head(imu->bvec[j].bv_page) == hpage)
8309 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8310 int nr_pages, struct io_mapped_ubuf *imu,
8311 struct page **last_hpage)
8315 imu->acct_pages = 0;
8316 for (i = 0; i < nr_pages; i++) {
8317 if (!PageCompound(pages[i])) {
8322 hpage = compound_head(pages[i]);
8323 if (hpage == *last_hpage)
8325 *last_hpage = hpage;
8326 if (headpage_already_acct(ctx, pages, i, hpage))
8328 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8332 if (!imu->acct_pages)
8335 ret = io_account_mem(ctx, imu->acct_pages);
8337 imu->acct_pages = 0;
8341 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8342 struct io_mapped_ubuf **pimu,
8343 struct page **last_hpage)
8345 struct io_mapped_ubuf *imu = NULL;
8346 struct vm_area_struct **vmas = NULL;
8347 struct page **pages = NULL;
8348 unsigned long off, start, end, ubuf;
8350 int ret, pret, nr_pages, i;
8352 if (!iov->iov_base) {
8353 *pimu = ctx->dummy_ubuf;
8357 ubuf = (unsigned long) iov->iov_base;
8358 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8359 start = ubuf >> PAGE_SHIFT;
8360 nr_pages = end - start;
8365 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8369 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8374 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8379 mmap_read_lock(current->mm);
8380 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8382 if (pret == nr_pages) {
8383 /* don't support file backed memory */
8384 for (i = 0; i < nr_pages; i++) {
8385 struct vm_area_struct *vma = vmas[i];
8387 if (vma_is_shmem(vma))
8390 !is_file_hugepages(vma->vm_file)) {
8396 ret = pret < 0 ? pret : -EFAULT;
8398 mmap_read_unlock(current->mm);
8401 * if we did partial map, or found file backed vmas,
8402 * release any pages we did get
8405 unpin_user_pages(pages, pret);
8409 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8411 unpin_user_pages(pages, pret);
8415 off = ubuf & ~PAGE_MASK;
8416 size = iov->iov_len;
8417 for (i = 0; i < nr_pages; i++) {
8420 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8421 imu->bvec[i].bv_page = pages[i];
8422 imu->bvec[i].bv_len = vec_len;
8423 imu->bvec[i].bv_offset = off;
8427 /* store original address for later verification */
8429 imu->ubuf_end = ubuf + iov->iov_len;
8430 imu->nr_bvecs = nr_pages;
8441 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8443 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8444 return ctx->user_bufs ? 0 : -ENOMEM;
8447 static int io_buffer_validate(struct iovec *iov)
8449 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8452 * Don't impose further limits on the size and buffer
8453 * constraints here, we'll -EINVAL later when IO is
8454 * submitted if they are wrong.
8457 return iov->iov_len ? -EFAULT : 0;
8461 /* arbitrary limit, but we need something */
8462 if (iov->iov_len > SZ_1G)
8465 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8471 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8472 unsigned int nr_args, u64 __user *tags)
8474 struct page *last_hpage = NULL;
8475 struct io_rsrc_data *data;
8481 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8483 ret = io_rsrc_node_switch_start(ctx);
8486 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8489 ret = io_buffers_map_alloc(ctx, nr_args);
8491 io_rsrc_data_free(data);
8495 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8496 ret = io_copy_iov(ctx, &iov, arg, i);
8499 ret = io_buffer_validate(&iov);
8502 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8507 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8513 WARN_ON_ONCE(ctx->buf_data);
8515 ctx->buf_data = data;
8517 __io_sqe_buffers_unregister(ctx);
8519 io_rsrc_node_switch(ctx, NULL);
8523 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8524 struct io_uring_rsrc_update2 *up,
8525 unsigned int nr_args)
8527 u64 __user *tags = u64_to_user_ptr(up->tags);
8528 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8529 struct page *last_hpage = NULL;
8530 bool needs_switch = false;
8536 if (up->offset + nr_args > ctx->nr_user_bufs)
8539 for (done = 0; done < nr_args; done++) {
8540 struct io_mapped_ubuf *imu;
8541 int offset = up->offset + done;
8544 err = io_copy_iov(ctx, &iov, iovs, done);
8547 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8551 err = io_buffer_validate(&iov);
8554 if (!iov.iov_base && tag) {
8558 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8562 i = array_index_nospec(offset, ctx->nr_user_bufs);
8563 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8564 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8565 ctx->rsrc_node, ctx->user_bufs[i]);
8566 if (unlikely(err)) {
8567 io_buffer_unmap(ctx, &imu);
8570 ctx->user_bufs[i] = NULL;
8571 needs_switch = true;
8574 ctx->user_bufs[i] = imu;
8575 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8579 io_rsrc_node_switch(ctx, ctx->buf_data);
8580 return done ? done : err;
8583 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8585 __s32 __user *fds = arg;
8591 if (copy_from_user(&fd, fds, sizeof(*fds)))
8594 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8595 if (IS_ERR(ctx->cq_ev_fd)) {
8596 int ret = PTR_ERR(ctx->cq_ev_fd);
8598 ctx->cq_ev_fd = NULL;
8605 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8607 if (ctx->cq_ev_fd) {
8608 eventfd_ctx_put(ctx->cq_ev_fd);
8609 ctx->cq_ev_fd = NULL;
8616 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8618 struct io_buffer *buf;
8619 unsigned long index;
8621 xa_for_each(&ctx->io_buffers, index, buf)
8622 __io_remove_buffers(ctx, buf, index, -1U);
8625 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8627 struct io_kiocb *req, *nxt;
8629 list_for_each_entry_safe(req, nxt, list, compl.list) {
8630 if (tsk && req->task != tsk)
8632 list_del(&req->compl.list);
8633 kmem_cache_free(req_cachep, req);
8637 static void io_req_caches_free(struct io_ring_ctx *ctx)
8639 struct io_submit_state *submit_state = &ctx->submit_state;
8640 struct io_comp_state *cs = &ctx->submit_state.comp;
8642 mutex_lock(&ctx->uring_lock);
8644 if (submit_state->free_reqs) {
8645 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8646 submit_state->reqs);
8647 submit_state->free_reqs = 0;
8650 io_flush_cached_locked_reqs(ctx, cs);
8651 io_req_cache_free(&cs->free_list, NULL);
8652 mutex_unlock(&ctx->uring_lock);
8655 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8657 if (data && !atomic_dec_and_test(&data->refs))
8658 wait_for_completion(&data->done);
8661 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8663 io_sq_thread_finish(ctx);
8665 if (ctx->mm_account) {
8666 mmdrop(ctx->mm_account);
8667 ctx->mm_account = NULL;
8670 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8671 io_wait_rsrc_data(ctx->buf_data);
8672 io_wait_rsrc_data(ctx->file_data);
8674 mutex_lock(&ctx->uring_lock);
8676 __io_sqe_buffers_unregister(ctx);
8678 __io_sqe_files_unregister(ctx);
8680 __io_cqring_overflow_flush(ctx, true);
8681 mutex_unlock(&ctx->uring_lock);
8682 io_eventfd_unregister(ctx);
8683 io_destroy_buffers(ctx);
8685 put_cred(ctx->sq_creds);
8687 /* there are no registered resources left, nobody uses it */
8689 io_rsrc_node_destroy(ctx->rsrc_node);
8690 if (ctx->rsrc_backup_node)
8691 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8692 flush_delayed_work(&ctx->rsrc_put_work);
8694 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8695 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8697 #if defined(CONFIG_UNIX)
8698 if (ctx->ring_sock) {
8699 ctx->ring_sock->file = NULL; /* so that iput() is called */
8700 sock_release(ctx->ring_sock);
8704 io_mem_free(ctx->rings);
8705 io_mem_free(ctx->sq_sqes);
8707 percpu_ref_exit(&ctx->refs);
8708 free_uid(ctx->user);
8709 io_req_caches_free(ctx);
8711 io_wq_put_hash(ctx->hash_map);
8712 kfree(ctx->cancel_hash);
8713 kfree(ctx->dummy_ubuf);
8717 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8719 struct io_ring_ctx *ctx = file->private_data;
8722 poll_wait(file, &ctx->poll_wait, wait);
8724 * synchronizes with barrier from wq_has_sleeper call in
8728 if (!io_sqring_full(ctx))
8729 mask |= EPOLLOUT | EPOLLWRNORM;
8732 * Don't flush cqring overflow list here, just do a simple check.
8733 * Otherwise there could possible be ABBA deadlock:
8736 * lock(&ctx->uring_lock);
8738 * lock(&ctx->uring_lock);
8741 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8742 * pushs them to do the flush.
8744 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8745 mask |= EPOLLIN | EPOLLRDNORM;
8750 static int io_uring_fasync(int fd, struct file *file, int on)
8752 struct io_ring_ctx *ctx = file->private_data;
8754 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8757 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8759 const struct cred *creds;
8761 creds = xa_erase(&ctx->personalities, id);
8770 struct io_tctx_exit {
8771 struct callback_head task_work;
8772 struct completion completion;
8773 struct io_ring_ctx *ctx;
8776 static void io_tctx_exit_cb(struct callback_head *cb)
8778 struct io_uring_task *tctx = current->io_uring;
8779 struct io_tctx_exit *work;
8781 work = container_of(cb, struct io_tctx_exit, task_work);
8783 * When @in_idle, we're in cancellation and it's racy to remove the
8784 * node. It'll be removed by the end of cancellation, just ignore it.
8786 if (!atomic_read(&tctx->in_idle))
8787 io_uring_del_tctx_node((unsigned long)work->ctx);
8788 complete(&work->completion);
8791 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8793 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8795 return req->ctx == data;
8798 static void io_ring_exit_work(struct work_struct *work)
8800 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8801 unsigned long timeout = jiffies + HZ * 60 * 5;
8802 struct io_tctx_exit exit;
8803 struct io_tctx_node *node;
8807 * If we're doing polled IO and end up having requests being
8808 * submitted async (out-of-line), then completions can come in while
8809 * we're waiting for refs to drop. We need to reap these manually,
8810 * as nobody else will be looking for them.
8813 io_uring_try_cancel_requests(ctx, NULL, true);
8815 struct io_sq_data *sqd = ctx->sq_data;
8816 struct task_struct *tsk;
8818 io_sq_thread_park(sqd);
8820 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8821 io_wq_cancel_cb(tsk->io_uring->io_wq,
8822 io_cancel_ctx_cb, ctx, true);
8823 io_sq_thread_unpark(sqd);
8826 WARN_ON_ONCE(time_after(jiffies, timeout));
8827 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8829 init_completion(&exit.completion);
8830 init_task_work(&exit.task_work, io_tctx_exit_cb);
8833 * Some may use context even when all refs and requests have been put,
8834 * and they are free to do so while still holding uring_lock or
8835 * completion_lock, see io_req_task_submit(). Apart from other work,
8836 * this lock/unlock section also waits them to finish.
8838 mutex_lock(&ctx->uring_lock);
8839 while (!list_empty(&ctx->tctx_list)) {
8840 WARN_ON_ONCE(time_after(jiffies, timeout));
8842 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8844 /* don't spin on a single task if cancellation failed */
8845 list_rotate_left(&ctx->tctx_list);
8846 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8847 if (WARN_ON_ONCE(ret))
8849 wake_up_process(node->task);
8851 mutex_unlock(&ctx->uring_lock);
8852 wait_for_completion(&exit.completion);
8853 mutex_lock(&ctx->uring_lock);
8855 mutex_unlock(&ctx->uring_lock);
8856 spin_lock_irq(&ctx->completion_lock);
8857 spin_unlock_irq(&ctx->completion_lock);
8859 io_ring_ctx_free(ctx);
8862 /* Returns true if we found and killed one or more timeouts */
8863 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8866 struct io_kiocb *req, *tmp;
8869 spin_lock_irq(&ctx->completion_lock);
8870 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8871 if (io_match_task(req, tsk, cancel_all)) {
8872 io_kill_timeout(req, -ECANCELED);
8877 io_commit_cqring(ctx);
8878 spin_unlock_irq(&ctx->completion_lock);
8880 io_cqring_ev_posted(ctx);
8881 return canceled != 0;
8884 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8886 unsigned long index;
8887 struct creds *creds;
8889 mutex_lock(&ctx->uring_lock);
8890 percpu_ref_kill(&ctx->refs);
8892 __io_cqring_overflow_flush(ctx, true);
8893 xa_for_each(&ctx->personalities, index, creds)
8894 io_unregister_personality(ctx, index);
8895 mutex_unlock(&ctx->uring_lock);
8897 io_kill_timeouts(ctx, NULL, true);
8898 io_poll_remove_all(ctx, NULL, true);
8900 /* if we failed setting up the ctx, we might not have any rings */
8901 io_iopoll_try_reap_events(ctx);
8903 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8905 * Use system_unbound_wq to avoid spawning tons of event kworkers
8906 * if we're exiting a ton of rings at the same time. It just adds
8907 * noise and overhead, there's no discernable change in runtime
8908 * over using system_wq.
8910 queue_work(system_unbound_wq, &ctx->exit_work);
8913 static int io_uring_release(struct inode *inode, struct file *file)
8915 struct io_ring_ctx *ctx = file->private_data;
8917 file->private_data = NULL;
8918 io_ring_ctx_wait_and_kill(ctx);
8922 struct io_task_cancel {
8923 struct task_struct *task;
8927 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8929 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8930 struct io_task_cancel *cancel = data;
8933 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8934 unsigned long flags;
8935 struct io_ring_ctx *ctx = req->ctx;
8937 /* protect against races with linked timeouts */
8938 spin_lock_irqsave(&ctx->completion_lock, flags);
8939 ret = io_match_task(req, cancel->task, cancel->all);
8940 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8942 ret = io_match_task(req, cancel->task, cancel->all);
8947 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8948 struct task_struct *task, bool cancel_all)
8950 struct io_defer_entry *de;
8953 spin_lock_irq(&ctx->completion_lock);
8954 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8955 if (io_match_task(de->req, task, cancel_all)) {
8956 list_cut_position(&list, &ctx->defer_list, &de->list);
8960 spin_unlock_irq(&ctx->completion_lock);
8961 if (list_empty(&list))
8964 while (!list_empty(&list)) {
8965 de = list_first_entry(&list, struct io_defer_entry, list);
8966 list_del_init(&de->list);
8967 io_req_complete_failed(de->req, -ECANCELED);
8973 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8975 struct io_tctx_node *node;
8976 enum io_wq_cancel cret;
8979 mutex_lock(&ctx->uring_lock);
8980 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8981 struct io_uring_task *tctx = node->task->io_uring;
8984 * io_wq will stay alive while we hold uring_lock, because it's
8985 * killed after ctx nodes, which requires to take the lock.
8987 if (!tctx || !tctx->io_wq)
8989 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8990 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8992 mutex_unlock(&ctx->uring_lock);
8997 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8998 struct task_struct *task,
9001 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9002 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9005 enum io_wq_cancel cret;
9009 ret |= io_uring_try_cancel_iowq(ctx);
9010 } else if (tctx && tctx->io_wq) {
9012 * Cancels requests of all rings, not only @ctx, but
9013 * it's fine as the task is in exit/exec.
9015 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9017 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9020 /* SQPOLL thread does its own polling */
9021 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9022 (ctx->sq_data && ctx->sq_data->thread == current)) {
9023 while (!list_empty_careful(&ctx->iopoll_list)) {
9024 io_iopoll_try_reap_events(ctx);
9029 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9030 ret |= io_poll_remove_all(ctx, task, cancel_all);
9031 ret |= io_kill_timeouts(ctx, task, cancel_all);
9033 ret |= io_run_task_work();
9040 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9042 struct io_uring_task *tctx = current->io_uring;
9043 struct io_tctx_node *node;
9046 if (unlikely(!tctx)) {
9047 ret = io_uring_alloc_task_context(current, ctx);
9050 tctx = current->io_uring;
9052 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9053 node = kmalloc(sizeof(*node), GFP_KERNEL);
9057 node->task = current;
9059 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9066 mutex_lock(&ctx->uring_lock);
9067 list_add(&node->ctx_node, &ctx->tctx_list);
9068 mutex_unlock(&ctx->uring_lock);
9075 * Note that this task has used io_uring. We use it for cancelation purposes.
9077 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9079 struct io_uring_task *tctx = current->io_uring;
9081 if (likely(tctx && tctx->last == ctx))
9083 return __io_uring_add_tctx_node(ctx);
9087 * Remove this io_uring_file -> task mapping.
9089 static void io_uring_del_tctx_node(unsigned long index)
9091 struct io_uring_task *tctx = current->io_uring;
9092 struct io_tctx_node *node;
9096 node = xa_erase(&tctx->xa, index);
9100 WARN_ON_ONCE(current != node->task);
9101 WARN_ON_ONCE(list_empty(&node->ctx_node));
9103 mutex_lock(&node->ctx->uring_lock);
9104 list_del(&node->ctx_node);
9105 mutex_unlock(&node->ctx->uring_lock);
9107 if (tctx->last == node->ctx)
9112 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9114 struct io_wq *wq = tctx->io_wq;
9115 struct io_tctx_node *node;
9116 unsigned long index;
9118 xa_for_each(&tctx->xa, index, node)
9119 io_uring_del_tctx_node(index);
9122 * Must be after io_uring_del_task_file() (removes nodes under
9123 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9126 io_wq_put_and_exit(wq);
9130 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9133 return atomic_read(&tctx->inflight_tracked);
9134 return percpu_counter_sum(&tctx->inflight);
9137 static void io_uring_drop_tctx_refs(struct task_struct *task)
9139 struct io_uring_task *tctx = task->io_uring;
9140 unsigned int refs = tctx->cached_refs;
9142 tctx->cached_refs = 0;
9143 percpu_counter_sub(&tctx->inflight, refs);
9144 put_task_struct_many(task, refs);
9148 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9149 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9151 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9153 struct io_uring_task *tctx = current->io_uring;
9154 struct io_ring_ctx *ctx;
9158 WARN_ON_ONCE(sqd && sqd->thread != current);
9160 if (!current->io_uring)
9163 io_wq_exit_start(tctx->io_wq);
9165 io_uring_drop_tctx_refs(current);
9166 atomic_inc(&tctx->in_idle);
9168 /* read completions before cancelations */
9169 inflight = tctx_inflight(tctx, !cancel_all);
9174 struct io_tctx_node *node;
9175 unsigned long index;
9177 xa_for_each(&tctx->xa, index, node) {
9178 /* sqpoll task will cancel all its requests */
9179 if (node->ctx->sq_data)
9181 io_uring_try_cancel_requests(node->ctx, current,
9185 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9186 io_uring_try_cancel_requests(ctx, current,
9190 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9192 * If we've seen completions, retry without waiting. This
9193 * avoids a race where a completion comes in before we did
9194 * prepare_to_wait().
9196 if (inflight == tctx_inflight(tctx, !cancel_all))
9198 finish_wait(&tctx->wait, &wait);
9200 atomic_dec(&tctx->in_idle);
9202 io_uring_clean_tctx(tctx);
9204 /* for exec all current's requests should be gone, kill tctx */
9205 __io_uring_free(current);
9209 void __io_uring_cancel(struct files_struct *files)
9211 io_uring_cancel_generic(!files, NULL);
9214 static void *io_uring_validate_mmap_request(struct file *file,
9215 loff_t pgoff, size_t sz)
9217 struct io_ring_ctx *ctx = file->private_data;
9218 loff_t offset = pgoff << PAGE_SHIFT;
9223 case IORING_OFF_SQ_RING:
9224 case IORING_OFF_CQ_RING:
9227 case IORING_OFF_SQES:
9231 return ERR_PTR(-EINVAL);
9234 page = virt_to_head_page(ptr);
9235 if (sz > page_size(page))
9236 return ERR_PTR(-EINVAL);
9243 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9245 size_t sz = vma->vm_end - vma->vm_start;
9249 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9251 return PTR_ERR(ptr);
9253 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9254 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9257 #else /* !CONFIG_MMU */
9259 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9261 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9264 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9266 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9269 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9270 unsigned long addr, unsigned long len,
9271 unsigned long pgoff, unsigned long flags)
9275 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9277 return PTR_ERR(ptr);
9279 return (unsigned long) ptr;
9282 #endif /* !CONFIG_MMU */
9284 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9289 if (!io_sqring_full(ctx))
9291 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9293 if (!io_sqring_full(ctx))
9296 } while (!signal_pending(current));
9298 finish_wait(&ctx->sqo_sq_wait, &wait);
9302 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9303 struct __kernel_timespec __user **ts,
9304 const sigset_t __user **sig)
9306 struct io_uring_getevents_arg arg;
9309 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9310 * is just a pointer to the sigset_t.
9312 if (!(flags & IORING_ENTER_EXT_ARG)) {
9313 *sig = (const sigset_t __user *) argp;
9319 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9320 * timespec and sigset_t pointers if good.
9322 if (*argsz != sizeof(arg))
9324 if (copy_from_user(&arg, argp, sizeof(arg)))
9326 *sig = u64_to_user_ptr(arg.sigmask);
9327 *argsz = arg.sigmask_sz;
9328 *ts = u64_to_user_ptr(arg.ts);
9332 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9333 u32, min_complete, u32, flags, const void __user *, argp,
9336 struct io_ring_ctx *ctx;
9343 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9344 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9348 if (unlikely(!f.file))
9352 if (unlikely(f.file->f_op != &io_uring_fops))
9356 ctx = f.file->private_data;
9357 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9361 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9365 * For SQ polling, the thread will do all submissions and completions.
9366 * Just return the requested submit count, and wake the thread if
9370 if (ctx->flags & IORING_SETUP_SQPOLL) {
9371 io_cqring_overflow_flush(ctx, false);
9374 if (unlikely(ctx->sq_data->thread == NULL))
9376 if (flags & IORING_ENTER_SQ_WAKEUP)
9377 wake_up(&ctx->sq_data->wait);
9378 if (flags & IORING_ENTER_SQ_WAIT) {
9379 ret = io_sqpoll_wait_sq(ctx);
9383 submitted = to_submit;
9384 } else if (to_submit) {
9385 ret = io_uring_add_tctx_node(ctx);
9388 mutex_lock(&ctx->uring_lock);
9389 submitted = io_submit_sqes(ctx, to_submit);
9390 mutex_unlock(&ctx->uring_lock);
9392 if (submitted != to_submit)
9395 if (flags & IORING_ENTER_GETEVENTS) {
9396 const sigset_t __user *sig;
9397 struct __kernel_timespec __user *ts;
9399 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9403 min_complete = min(min_complete, ctx->cq_entries);
9406 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9407 * space applications don't need to do io completion events
9408 * polling again, they can rely on io_sq_thread to do polling
9409 * work, which can reduce cpu usage and uring_lock contention.
9411 if (ctx->flags & IORING_SETUP_IOPOLL &&
9412 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9413 ret = io_iopoll_check(ctx, min_complete);
9415 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9420 percpu_ref_put(&ctx->refs);
9423 return submitted ? submitted : ret;
9426 #ifdef CONFIG_PROC_FS
9427 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9428 const struct cred *cred)
9430 struct user_namespace *uns = seq_user_ns(m);
9431 struct group_info *gi;
9436 seq_printf(m, "%5d\n", id);
9437 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9438 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9439 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9440 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9441 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9442 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9443 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9444 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9445 seq_puts(m, "\n\tGroups:\t");
9446 gi = cred->group_info;
9447 for (g = 0; g < gi->ngroups; g++) {
9448 seq_put_decimal_ull(m, g ? " " : "",
9449 from_kgid_munged(uns, gi->gid[g]));
9451 seq_puts(m, "\n\tCapEff:\t");
9452 cap = cred->cap_effective;
9453 CAP_FOR_EACH_U32(__capi)
9454 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9459 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9461 struct io_sq_data *sq = NULL;
9466 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9467 * since fdinfo case grabs it in the opposite direction of normal use
9468 * cases. If we fail to get the lock, we just don't iterate any
9469 * structures that could be going away outside the io_uring mutex.
9471 has_lock = mutex_trylock(&ctx->uring_lock);
9473 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9479 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9480 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9481 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9482 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9483 struct file *f = io_file_from_index(ctx, i);
9486 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9488 seq_printf(m, "%5u: <none>\n", i);
9490 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9491 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9492 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9493 unsigned int len = buf->ubuf_end - buf->ubuf;
9495 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9497 if (has_lock && !xa_empty(&ctx->personalities)) {
9498 unsigned long index;
9499 const struct cred *cred;
9501 seq_printf(m, "Personalities:\n");
9502 xa_for_each(&ctx->personalities, index, cred)
9503 io_uring_show_cred(m, index, cred);
9505 seq_printf(m, "PollList:\n");
9506 spin_lock_irq(&ctx->completion_lock);
9507 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9508 struct hlist_head *list = &ctx->cancel_hash[i];
9509 struct io_kiocb *req;
9511 hlist_for_each_entry(req, list, hash_node)
9512 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9513 req->task->task_works != NULL);
9515 spin_unlock_irq(&ctx->completion_lock);
9517 mutex_unlock(&ctx->uring_lock);
9520 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9522 struct io_ring_ctx *ctx = f->private_data;
9524 if (percpu_ref_tryget(&ctx->refs)) {
9525 __io_uring_show_fdinfo(ctx, m);
9526 percpu_ref_put(&ctx->refs);
9531 static const struct file_operations io_uring_fops = {
9532 .release = io_uring_release,
9533 .mmap = io_uring_mmap,
9535 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9536 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9538 .poll = io_uring_poll,
9539 .fasync = io_uring_fasync,
9540 #ifdef CONFIG_PROC_FS
9541 .show_fdinfo = io_uring_show_fdinfo,
9545 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9546 struct io_uring_params *p)
9548 struct io_rings *rings;
9549 size_t size, sq_array_offset;
9551 /* make sure these are sane, as we already accounted them */
9552 ctx->sq_entries = p->sq_entries;
9553 ctx->cq_entries = p->cq_entries;
9555 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9556 if (size == SIZE_MAX)
9559 rings = io_mem_alloc(size);
9564 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9565 rings->sq_ring_mask = p->sq_entries - 1;
9566 rings->cq_ring_mask = p->cq_entries - 1;
9567 rings->sq_ring_entries = p->sq_entries;
9568 rings->cq_ring_entries = p->cq_entries;
9570 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9571 if (size == SIZE_MAX) {
9572 io_mem_free(ctx->rings);
9577 ctx->sq_sqes = io_mem_alloc(size);
9578 if (!ctx->sq_sqes) {
9579 io_mem_free(ctx->rings);
9587 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9591 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9595 ret = io_uring_add_tctx_node(ctx);
9600 fd_install(fd, file);
9605 * Allocate an anonymous fd, this is what constitutes the application
9606 * visible backing of an io_uring instance. The application mmaps this
9607 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9608 * we have to tie this fd to a socket for file garbage collection purposes.
9610 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9613 #if defined(CONFIG_UNIX)
9616 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9619 return ERR_PTR(ret);
9622 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9623 O_RDWR | O_CLOEXEC);
9624 #if defined(CONFIG_UNIX)
9626 sock_release(ctx->ring_sock);
9627 ctx->ring_sock = NULL;
9629 ctx->ring_sock->file = file;
9635 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9636 struct io_uring_params __user *params)
9638 struct io_ring_ctx *ctx;
9644 if (entries > IORING_MAX_ENTRIES) {
9645 if (!(p->flags & IORING_SETUP_CLAMP))
9647 entries = IORING_MAX_ENTRIES;
9651 * Use twice as many entries for the CQ ring. It's possible for the
9652 * application to drive a higher depth than the size of the SQ ring,
9653 * since the sqes are only used at submission time. This allows for
9654 * some flexibility in overcommitting a bit. If the application has
9655 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9656 * of CQ ring entries manually.
9658 p->sq_entries = roundup_pow_of_two(entries);
9659 if (p->flags & IORING_SETUP_CQSIZE) {
9661 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9662 * to a power-of-two, if it isn't already. We do NOT impose
9663 * any cq vs sq ring sizing.
9667 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9668 if (!(p->flags & IORING_SETUP_CLAMP))
9670 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9672 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9673 if (p->cq_entries < p->sq_entries)
9676 p->cq_entries = 2 * p->sq_entries;
9679 ctx = io_ring_ctx_alloc(p);
9682 ctx->compat = in_compat_syscall();
9683 if (!capable(CAP_IPC_LOCK))
9684 ctx->user = get_uid(current_user());
9687 * This is just grabbed for accounting purposes. When a process exits,
9688 * the mm is exited and dropped before the files, hence we need to hang
9689 * on to this mm purely for the purposes of being able to unaccount
9690 * memory (locked/pinned vm). It's not used for anything else.
9692 mmgrab(current->mm);
9693 ctx->mm_account = current->mm;
9695 ret = io_allocate_scq_urings(ctx, p);
9699 ret = io_sq_offload_create(ctx, p);
9702 /* always set a rsrc node */
9703 ret = io_rsrc_node_switch_start(ctx);
9706 io_rsrc_node_switch(ctx, NULL);
9708 memset(&p->sq_off, 0, sizeof(p->sq_off));
9709 p->sq_off.head = offsetof(struct io_rings, sq.head);
9710 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9711 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9712 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9713 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9714 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9715 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9717 memset(&p->cq_off, 0, sizeof(p->cq_off));
9718 p->cq_off.head = offsetof(struct io_rings, cq.head);
9719 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9720 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9721 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9722 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9723 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9724 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9726 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9727 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9728 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9729 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9730 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9731 IORING_FEAT_RSRC_TAGS;
9733 if (copy_to_user(params, p, sizeof(*p))) {
9738 file = io_uring_get_file(ctx);
9740 ret = PTR_ERR(file);
9745 * Install ring fd as the very last thing, so we don't risk someone
9746 * having closed it before we finish setup
9748 ret = io_uring_install_fd(ctx, file);
9750 /* fput will clean it up */
9755 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9758 io_ring_ctx_wait_and_kill(ctx);
9763 * Sets up an aio uring context, and returns the fd. Applications asks for a
9764 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9765 * params structure passed in.
9767 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9769 struct io_uring_params p;
9772 if (copy_from_user(&p, params, sizeof(p)))
9774 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9779 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9780 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9781 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9782 IORING_SETUP_R_DISABLED))
9785 return io_uring_create(entries, &p, params);
9788 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9789 struct io_uring_params __user *, params)
9791 return io_uring_setup(entries, params);
9794 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9796 struct io_uring_probe *p;
9800 size = struct_size(p, ops, nr_args);
9801 if (size == SIZE_MAX)
9803 p = kzalloc(size, GFP_KERNEL);
9808 if (copy_from_user(p, arg, size))
9811 if (memchr_inv(p, 0, size))
9814 p->last_op = IORING_OP_LAST - 1;
9815 if (nr_args > IORING_OP_LAST)
9816 nr_args = IORING_OP_LAST;
9818 for (i = 0; i < nr_args; i++) {
9820 if (!io_op_defs[i].not_supported)
9821 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9826 if (copy_to_user(arg, p, size))
9833 static int io_register_personality(struct io_ring_ctx *ctx)
9835 const struct cred *creds;
9839 creds = get_current_cred();
9841 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9842 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9849 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9850 unsigned int nr_args)
9852 struct io_uring_restriction *res;
9856 /* Restrictions allowed only if rings started disabled */
9857 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9860 /* We allow only a single restrictions registration */
9861 if (ctx->restrictions.registered)
9864 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9867 size = array_size(nr_args, sizeof(*res));
9868 if (size == SIZE_MAX)
9871 res = memdup_user(arg, size);
9873 return PTR_ERR(res);
9877 for (i = 0; i < nr_args; i++) {
9878 switch (res[i].opcode) {
9879 case IORING_RESTRICTION_REGISTER_OP:
9880 if (res[i].register_op >= IORING_REGISTER_LAST) {
9885 __set_bit(res[i].register_op,
9886 ctx->restrictions.register_op);
9888 case IORING_RESTRICTION_SQE_OP:
9889 if (res[i].sqe_op >= IORING_OP_LAST) {
9894 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9896 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9897 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9899 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9900 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9909 /* Reset all restrictions if an error happened */
9911 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9913 ctx->restrictions.registered = true;
9919 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9921 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9924 if (ctx->restrictions.registered)
9925 ctx->restricted = 1;
9927 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9928 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9929 wake_up(&ctx->sq_data->wait);
9933 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9934 struct io_uring_rsrc_update2 *up,
9942 if (check_add_overflow(up->offset, nr_args, &tmp))
9944 err = io_rsrc_node_switch_start(ctx);
9949 case IORING_RSRC_FILE:
9950 return __io_sqe_files_update(ctx, up, nr_args);
9951 case IORING_RSRC_BUFFER:
9952 return __io_sqe_buffers_update(ctx, up, nr_args);
9957 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9960 struct io_uring_rsrc_update2 up;
9964 memset(&up, 0, sizeof(up));
9965 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9967 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9970 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9971 unsigned size, unsigned type)
9973 struct io_uring_rsrc_update2 up;
9975 if (size != sizeof(up))
9977 if (copy_from_user(&up, arg, sizeof(up)))
9979 if (!up.nr || up.resv)
9981 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9984 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9985 unsigned int size, unsigned int type)
9987 struct io_uring_rsrc_register rr;
9989 /* keep it extendible */
9990 if (size != sizeof(rr))
9993 memset(&rr, 0, sizeof(rr));
9994 if (copy_from_user(&rr, arg, size))
9996 if (!rr.nr || rr.resv || rr.resv2)
10000 case IORING_RSRC_FILE:
10001 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10002 rr.nr, u64_to_user_ptr(rr.tags));
10003 case IORING_RSRC_BUFFER:
10004 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10005 rr.nr, u64_to_user_ptr(rr.tags));
10010 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10013 struct io_uring_task *tctx = current->io_uring;
10014 cpumask_var_t new_mask;
10017 if (!tctx || !tctx->io_wq)
10020 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10023 cpumask_clear(new_mask);
10024 if (len > cpumask_size())
10025 len = cpumask_size();
10027 if (copy_from_user(new_mask, arg, len)) {
10028 free_cpumask_var(new_mask);
10032 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10033 free_cpumask_var(new_mask);
10037 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10039 struct io_uring_task *tctx = current->io_uring;
10041 if (!tctx || !tctx->io_wq)
10044 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10047 static bool io_register_op_must_quiesce(int op)
10050 case IORING_REGISTER_BUFFERS:
10051 case IORING_UNREGISTER_BUFFERS:
10052 case IORING_REGISTER_FILES:
10053 case IORING_UNREGISTER_FILES:
10054 case IORING_REGISTER_FILES_UPDATE:
10055 case IORING_REGISTER_PROBE:
10056 case IORING_REGISTER_PERSONALITY:
10057 case IORING_UNREGISTER_PERSONALITY:
10058 case IORING_REGISTER_FILES2:
10059 case IORING_REGISTER_FILES_UPDATE2:
10060 case IORING_REGISTER_BUFFERS2:
10061 case IORING_REGISTER_BUFFERS_UPDATE:
10062 case IORING_REGISTER_IOWQ_AFF:
10063 case IORING_UNREGISTER_IOWQ_AFF:
10070 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10071 void __user *arg, unsigned nr_args)
10072 __releases(ctx->uring_lock)
10073 __acquires(ctx->uring_lock)
10078 * We're inside the ring mutex, if the ref is already dying, then
10079 * someone else killed the ctx or is already going through
10080 * io_uring_register().
10082 if (percpu_ref_is_dying(&ctx->refs))
10085 if (ctx->restricted) {
10086 if (opcode >= IORING_REGISTER_LAST)
10088 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10089 if (!test_bit(opcode, ctx->restrictions.register_op))
10093 if (io_register_op_must_quiesce(opcode)) {
10094 percpu_ref_kill(&ctx->refs);
10097 * Drop uring mutex before waiting for references to exit. If
10098 * another thread is currently inside io_uring_enter() it might
10099 * need to grab the uring_lock to make progress. If we hold it
10100 * here across the drain wait, then we can deadlock. It's safe
10101 * to drop the mutex here, since no new references will come in
10102 * after we've killed the percpu ref.
10104 mutex_unlock(&ctx->uring_lock);
10106 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10109 ret = io_run_task_work_sig();
10113 mutex_lock(&ctx->uring_lock);
10116 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10122 case IORING_REGISTER_BUFFERS:
10123 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10125 case IORING_UNREGISTER_BUFFERS:
10127 if (arg || nr_args)
10129 ret = io_sqe_buffers_unregister(ctx);
10131 case IORING_REGISTER_FILES:
10132 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10134 case IORING_UNREGISTER_FILES:
10136 if (arg || nr_args)
10138 ret = io_sqe_files_unregister(ctx);
10140 case IORING_REGISTER_FILES_UPDATE:
10141 ret = io_register_files_update(ctx, arg, nr_args);
10143 case IORING_REGISTER_EVENTFD:
10144 case IORING_REGISTER_EVENTFD_ASYNC:
10148 ret = io_eventfd_register(ctx, arg);
10151 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10152 ctx->eventfd_async = 1;
10154 ctx->eventfd_async = 0;
10156 case IORING_UNREGISTER_EVENTFD:
10158 if (arg || nr_args)
10160 ret = io_eventfd_unregister(ctx);
10162 case IORING_REGISTER_PROBE:
10164 if (!arg || nr_args > 256)
10166 ret = io_probe(ctx, arg, nr_args);
10168 case IORING_REGISTER_PERSONALITY:
10170 if (arg || nr_args)
10172 ret = io_register_personality(ctx);
10174 case IORING_UNREGISTER_PERSONALITY:
10178 ret = io_unregister_personality(ctx, nr_args);
10180 case IORING_REGISTER_ENABLE_RINGS:
10182 if (arg || nr_args)
10184 ret = io_register_enable_rings(ctx);
10186 case IORING_REGISTER_RESTRICTIONS:
10187 ret = io_register_restrictions(ctx, arg, nr_args);
10189 case IORING_REGISTER_FILES2:
10190 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10192 case IORING_REGISTER_FILES_UPDATE2:
10193 ret = io_register_rsrc_update(ctx, arg, nr_args,
10196 case IORING_REGISTER_BUFFERS2:
10197 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10199 case IORING_REGISTER_BUFFERS_UPDATE:
10200 ret = io_register_rsrc_update(ctx, arg, nr_args,
10201 IORING_RSRC_BUFFER);
10203 case IORING_REGISTER_IOWQ_AFF:
10205 if (!arg || !nr_args)
10207 ret = io_register_iowq_aff(ctx, arg, nr_args);
10209 case IORING_UNREGISTER_IOWQ_AFF:
10211 if (arg || nr_args)
10213 ret = io_unregister_iowq_aff(ctx);
10220 if (io_register_op_must_quiesce(opcode)) {
10221 /* bring the ctx back to life */
10222 percpu_ref_reinit(&ctx->refs);
10223 reinit_completion(&ctx->ref_comp);
10228 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10229 void __user *, arg, unsigned int, nr_args)
10231 struct io_ring_ctx *ctx;
10240 if (f.file->f_op != &io_uring_fops)
10243 ctx = f.file->private_data;
10245 io_run_task_work();
10247 mutex_lock(&ctx->uring_lock);
10248 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10249 mutex_unlock(&ctx->uring_lock);
10250 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10251 ctx->cq_ev_fd != NULL, ret);
10257 static int __init io_uring_init(void)
10259 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10260 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10261 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10264 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10265 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10266 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10267 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10268 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10269 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10270 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10271 BUILD_BUG_SQE_ELEM(8, __u64, off);
10272 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10273 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10274 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10275 BUILD_BUG_SQE_ELEM(24, __u32, len);
10276 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10277 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10278 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10279 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10280 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10281 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10282 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10283 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10284 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10285 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10286 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10287 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10288 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10289 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10290 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10291 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10292 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10293 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10294 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10295 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10297 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10298 sizeof(struct io_uring_rsrc_update));
10299 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10300 sizeof(struct io_uring_rsrc_update2));
10301 /* should fit into one byte */
10302 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10304 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10305 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10307 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10311 __initcall(io_uring_init);