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>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define IO_RSRC_TAG_TABLE_SHIFT 9
104 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
105 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
107 #define IORING_MAX_REG_BUFFERS (1U << 14)
109 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
110 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
113 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
116 u32 head ____cacheline_aligned_in_smp;
117 u32 tail ____cacheline_aligned_in_smp;
121 * This data is shared with the application through the mmap at offsets
122 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
124 * The offsets to the member fields are published through struct
125 * io_sqring_offsets when calling io_uring_setup.
129 * Head and tail offsets into the ring; the offsets need to be
130 * masked to get valid indices.
132 * The kernel controls head of the sq ring and the tail of the cq ring,
133 * and the application controls tail of the sq ring and the head of the
136 struct io_uring sq, cq;
138 * Bitmasks to apply to head and tail offsets (constant, equals
141 u32 sq_ring_mask, cq_ring_mask;
142 /* Ring sizes (constant, power of 2) */
143 u32 sq_ring_entries, cq_ring_entries;
145 * Number of invalid entries dropped by the kernel due to
146 * invalid index stored in array
148 * Written by the kernel, shouldn't be modified by the
149 * application (i.e. get number of "new events" by comparing to
152 * After a new SQ head value was read by the application this
153 * counter includes all submissions that were dropped reaching
154 * the new SQ head (and possibly more).
160 * Written by the kernel, shouldn't be modified by the
163 * The application needs a full memory barrier before checking
164 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
170 * Written by the application, shouldn't be modified by the
175 * Number of completion events lost because the queue was full;
176 * this should be avoided by the application by making sure
177 * there are not more requests pending than there is space in
178 * the completion queue.
180 * Written by the kernel, shouldn't be modified by the
181 * application (i.e. get number of "new events" by comparing to
184 * As completion events come in out of order this counter is not
185 * ordered with any other data.
189 * Ring buffer of completion events.
191 * The kernel writes completion events fresh every time they are
192 * produced, so the application is allowed to modify pending
195 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
198 enum io_uring_cmd_flags {
199 IO_URING_F_NONBLOCK = 1,
200 IO_URING_F_COMPLETE_DEFER = 2,
203 struct io_mapped_ubuf {
206 unsigned int nr_bvecs;
207 unsigned long acct_pages;
208 struct bio_vec bvec[];
213 struct io_overflow_cqe {
214 struct io_uring_cqe cqe;
215 struct list_head list;
218 struct io_fixed_file {
219 /* file * with additional FFS_* flags */
220 unsigned long file_ptr;
224 struct list_head list;
229 struct io_mapped_ubuf *buf;
233 struct io_file_table {
234 /* two level table */
235 struct io_fixed_file **files;
238 struct io_rsrc_node {
239 struct percpu_ref refs;
240 struct list_head node;
241 struct list_head rsrc_list;
242 struct io_rsrc_data *rsrc_data;
243 struct llist_node llist;
247 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
249 struct io_rsrc_data {
250 struct io_ring_ctx *ctx;
256 struct completion done;
261 struct list_head list;
267 struct io_restriction {
268 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
269 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
270 u8 sqe_flags_allowed;
271 u8 sqe_flags_required;
276 IO_SQ_THREAD_SHOULD_STOP = 0,
277 IO_SQ_THREAD_SHOULD_PARK,
282 atomic_t park_pending;
285 /* ctx's that are using this sqd */
286 struct list_head ctx_list;
288 struct task_struct *thread;
289 struct wait_queue_head wait;
291 unsigned sq_thread_idle;
297 struct completion exited;
300 #define IO_IOPOLL_BATCH 8
301 #define IO_COMPL_BATCH 32
302 #define IO_REQ_CACHE_SIZE 32
303 #define IO_REQ_ALLOC_BATCH 8
305 struct io_comp_state {
306 struct io_kiocb *reqs[IO_COMPL_BATCH];
308 /* inline/task_work completion list, under ->uring_lock */
309 struct list_head free_list;
312 struct io_submit_link {
313 struct io_kiocb *head;
314 struct io_kiocb *last;
317 struct io_submit_state {
318 struct blk_plug plug;
319 struct io_submit_link link;
322 * io_kiocb alloc cache
324 void *reqs[IO_REQ_CACHE_SIZE];
325 unsigned int free_reqs;
330 * Batch completion logic
332 struct io_comp_state comp;
335 * File reference cache
339 unsigned int file_refs;
340 unsigned int ios_left;
345 struct percpu_ref refs;
346 } ____cacheline_aligned_in_smp;
350 unsigned int compat: 1;
351 unsigned int drain_next: 1;
352 unsigned int eventfd_async: 1;
353 unsigned int restricted: 1;
356 * Ring buffer of indices into array of io_uring_sqe, which is
357 * mmapped by the application using the IORING_OFF_SQES offset.
359 * This indirection could e.g. be used to assign fixed
360 * io_uring_sqe entries to operations and only submit them to
361 * the queue when needed.
363 * The kernel modifies neither the indices array nor the entries
367 unsigned cached_sq_head;
369 unsigned sq_thread_idle;
370 unsigned cached_sq_dropped;
371 unsigned long sq_check_overflow;
373 struct list_head defer_list;
374 struct list_head timeout_list;
375 struct list_head cq_overflow_list;
377 struct io_uring_sqe *sq_sqes;
378 } ____cacheline_aligned_in_smp;
381 struct mutex uring_lock;
382 wait_queue_head_t wait;
383 } ____cacheline_aligned_in_smp;
385 struct io_submit_state submit_state;
386 /* IRQ completion list, under ->completion_lock */
387 struct list_head locked_free_list;
388 unsigned int locked_free_nr;
390 struct io_rings *rings;
392 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
393 struct io_sq_data *sq_data; /* if using sq thread polling */
395 struct wait_queue_head sqo_sq_wait;
396 struct list_head sqd_list;
399 * Fixed resources fast path, should be accessed only under uring_lock,
400 * and updated through io_uring_register(2)
402 struct io_rsrc_node *rsrc_node;
404 struct io_file_table file_table;
405 unsigned nr_user_files;
406 unsigned nr_user_bufs;
407 struct io_mapped_ubuf **user_bufs;
409 struct xarray io_buffers;
410 struct xarray personalities;
414 unsigned cached_cq_tail;
416 atomic_t cq_timeouts;
417 unsigned cq_last_tm_flush;
419 unsigned long cq_check_overflow;
420 struct wait_queue_head cq_wait;
421 struct fasync_struct *cq_fasync;
422 struct eventfd_ctx *cq_ev_fd;
423 } ____cacheline_aligned_in_smp;
426 spinlock_t completion_lock;
429 * ->iopoll_list is protected by the ctx->uring_lock for
430 * io_uring instances that don't use IORING_SETUP_SQPOLL.
431 * For SQPOLL, only the single threaded io_sq_thread() will
432 * manipulate the list, hence no extra locking is needed there.
434 struct list_head iopoll_list;
435 struct hlist_head *cancel_hash;
436 unsigned cancel_hash_bits;
437 bool poll_multi_file;
438 } ____cacheline_aligned_in_smp;
440 struct io_restriction restrictions;
442 /* slow path rsrc auxilary data, used by update/register */
444 struct io_rsrc_node *rsrc_backup_node;
445 struct io_mapped_ubuf *dummy_ubuf;
446 struct io_rsrc_data *file_data;
447 struct io_rsrc_data *buf_data;
449 struct delayed_work rsrc_put_work;
450 struct llist_head rsrc_put_llist;
451 struct list_head rsrc_ref_list;
452 spinlock_t rsrc_ref_lock;
455 /* Keep this last, we don't need it for the fast path */
457 #if defined(CONFIG_UNIX)
458 struct socket *ring_sock;
460 /* hashed buffered write serialization */
461 struct io_wq_hash *hash_map;
463 /* Only used for accounting purposes */
464 struct user_struct *user;
465 struct mm_struct *mm_account;
467 /* ctx exit and cancelation */
468 struct callback_head *exit_task_work;
469 struct work_struct exit_work;
470 struct list_head tctx_list;
471 struct completion ref_comp;
475 struct io_uring_task {
476 /* submission side */
479 struct wait_queue_head wait;
480 const struct io_ring_ctx *last;
482 struct percpu_counter inflight;
483 atomic_t inflight_tracked;
486 spinlock_t task_lock;
487 struct io_wq_work_list task_list;
488 unsigned long task_state;
489 struct callback_head task_work;
493 * First field must be the file pointer in all the
494 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
496 struct io_poll_iocb {
498 struct wait_queue_head *head;
502 struct wait_queue_entry wait;
505 struct io_poll_update {
511 bool update_user_data;
519 struct io_timeout_data {
520 struct io_kiocb *req;
521 struct hrtimer timer;
522 struct timespec64 ts;
523 enum hrtimer_mode mode;
528 struct sockaddr __user *addr;
529 int __user *addr_len;
531 unsigned long nofile;
551 struct list_head list;
552 /* head of the link, used by linked timeouts only */
553 struct io_kiocb *head;
556 struct io_timeout_rem {
561 struct timespec64 ts;
566 /* NOTE: kiocb has the file as the first member, so don't do it here */
574 struct sockaddr __user *addr;
581 struct compat_msghdr __user *umsg_compat;
582 struct user_msghdr __user *umsg;
588 struct io_buffer *kbuf;
594 struct filename *filename;
596 unsigned long nofile;
599 struct io_rsrc_update {
625 struct epoll_event event;
629 struct file *file_out;
630 struct file *file_in;
637 struct io_provide_buf {
651 const char __user *filename;
652 struct statx __user *buffer;
664 struct filename *oldpath;
665 struct filename *newpath;
673 struct filename *filename;
676 struct io_completion {
678 struct list_head list;
682 struct io_async_connect {
683 struct sockaddr_storage address;
686 struct io_async_msghdr {
687 struct iovec fast_iov[UIO_FASTIOV];
688 /* points to an allocated iov, if NULL we use fast_iov instead */
689 struct iovec *free_iov;
690 struct sockaddr __user *uaddr;
692 struct sockaddr_storage addr;
696 struct iovec fast_iov[UIO_FASTIOV];
697 const struct iovec *free_iovec;
698 struct iov_iter iter;
700 struct wait_page_queue wpq;
704 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
705 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
706 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
707 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
708 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
709 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
711 /* first byte is taken by user flags, shift it to not overlap */
716 REQ_F_LINK_TIMEOUT_BIT,
717 REQ_F_NEED_CLEANUP_BIT,
719 REQ_F_BUFFER_SELECTED_BIT,
720 REQ_F_LTIMEOUT_ACTIVE_BIT,
721 REQ_F_COMPLETE_INLINE_BIT,
723 REQ_F_DONT_REISSUE_BIT,
724 /* keep async read/write and isreg together and in order */
725 REQ_F_ASYNC_READ_BIT,
726 REQ_F_ASYNC_WRITE_BIT,
729 /* not a real bit, just to check we're not overflowing the space */
735 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
736 /* drain existing IO first */
737 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
739 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
740 /* doesn't sever on completion < 0 */
741 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
743 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
744 /* IOSQE_BUFFER_SELECT */
745 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
747 /* fail rest of links */
748 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
749 /* on inflight list, should be cancelled and waited on exit reliably */
750 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
751 /* read/write uses file position */
752 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
753 /* must not punt to workers */
754 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
755 /* has or had linked timeout */
756 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
758 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
759 /* already went through poll handler */
760 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
761 /* buffer already selected */
762 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
763 /* linked timeout is active, i.e. prepared by link's head */
764 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
765 /* completion is deferred through io_comp_state */
766 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
767 /* caller should reissue async */
768 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
769 /* don't attempt request reissue, see io_rw_reissue() */
770 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
771 /* supports async reads */
772 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
773 /* supports async writes */
774 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
776 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
780 struct io_poll_iocb poll;
781 struct io_poll_iocb *double_poll;
784 struct io_task_work {
785 struct io_wq_work_node node;
786 task_work_func_t func;
790 IORING_RSRC_FILE = 0,
791 IORING_RSRC_BUFFER = 1,
795 * NOTE! Each of the iocb union members has the file pointer
796 * as the first entry in their struct definition. So you can
797 * access the file pointer through any of the sub-structs,
798 * or directly as just 'ki_filp' in this struct.
804 struct io_poll_iocb poll;
805 struct io_poll_update poll_update;
806 struct io_accept accept;
808 struct io_cancel cancel;
809 struct io_timeout timeout;
810 struct io_timeout_rem timeout_rem;
811 struct io_connect connect;
812 struct io_sr_msg sr_msg;
814 struct io_close close;
815 struct io_rsrc_update rsrc_update;
816 struct io_fadvise fadvise;
817 struct io_madvise madvise;
818 struct io_epoll epoll;
819 struct io_splice splice;
820 struct io_provide_buf pbuf;
821 struct io_statx statx;
822 struct io_shutdown shutdown;
823 struct io_rename rename;
824 struct io_unlink unlink;
825 /* use only after cleaning per-op data, see io_clean_op() */
826 struct io_completion compl;
829 /* opcode allocated if it needs to store data for async defer */
832 /* polled IO has completed */
838 struct io_ring_ctx *ctx;
841 struct task_struct *task;
844 struct io_kiocb *link;
845 struct percpu_ref *fixed_rsrc_refs;
847 /* used with ctx->iopoll_list with reads/writes */
848 struct list_head inflight_entry;
850 struct io_task_work io_task_work;
851 struct callback_head task_work;
853 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
854 struct hlist_node hash_node;
855 struct async_poll *apoll;
856 struct io_wq_work work;
857 /* store used ubuf, so we can prevent reloading */
858 struct io_mapped_ubuf *imu;
861 struct io_tctx_node {
862 struct list_head ctx_node;
863 struct task_struct *task;
864 struct io_ring_ctx *ctx;
867 struct io_defer_entry {
868 struct list_head list;
869 struct io_kiocb *req;
874 /* needs req->file assigned */
875 unsigned needs_file : 1;
876 /* hash wq insertion if file is a regular file */
877 unsigned hash_reg_file : 1;
878 /* unbound wq insertion if file is a non-regular file */
879 unsigned unbound_nonreg_file : 1;
880 /* opcode is not supported by this kernel */
881 unsigned not_supported : 1;
882 /* set if opcode supports polled "wait" */
884 unsigned pollout : 1;
885 /* op supports buffer selection */
886 unsigned buffer_select : 1;
887 /* do prep async if is going to be punted */
888 unsigned needs_async_setup : 1;
889 /* should block plug */
891 /* size of async data needed, if any */
892 unsigned short async_size;
895 static const struct io_op_def io_op_defs[] = {
896 [IORING_OP_NOP] = {},
897 [IORING_OP_READV] = {
899 .unbound_nonreg_file = 1,
902 .needs_async_setup = 1,
904 .async_size = sizeof(struct io_async_rw),
906 [IORING_OP_WRITEV] = {
909 .unbound_nonreg_file = 1,
911 .needs_async_setup = 1,
913 .async_size = sizeof(struct io_async_rw),
915 [IORING_OP_FSYNC] = {
918 [IORING_OP_READ_FIXED] = {
920 .unbound_nonreg_file = 1,
923 .async_size = sizeof(struct io_async_rw),
925 [IORING_OP_WRITE_FIXED] = {
928 .unbound_nonreg_file = 1,
931 .async_size = sizeof(struct io_async_rw),
933 [IORING_OP_POLL_ADD] = {
935 .unbound_nonreg_file = 1,
937 [IORING_OP_POLL_REMOVE] = {},
938 [IORING_OP_SYNC_FILE_RANGE] = {
941 [IORING_OP_SENDMSG] = {
943 .unbound_nonreg_file = 1,
945 .needs_async_setup = 1,
946 .async_size = sizeof(struct io_async_msghdr),
948 [IORING_OP_RECVMSG] = {
950 .unbound_nonreg_file = 1,
953 .needs_async_setup = 1,
954 .async_size = sizeof(struct io_async_msghdr),
956 [IORING_OP_TIMEOUT] = {
957 .async_size = sizeof(struct io_timeout_data),
959 [IORING_OP_TIMEOUT_REMOVE] = {
960 /* used by timeout updates' prep() */
962 [IORING_OP_ACCEPT] = {
964 .unbound_nonreg_file = 1,
967 [IORING_OP_ASYNC_CANCEL] = {},
968 [IORING_OP_LINK_TIMEOUT] = {
969 .async_size = sizeof(struct io_timeout_data),
971 [IORING_OP_CONNECT] = {
973 .unbound_nonreg_file = 1,
975 .needs_async_setup = 1,
976 .async_size = sizeof(struct io_async_connect),
978 [IORING_OP_FALLOCATE] = {
981 [IORING_OP_OPENAT] = {},
982 [IORING_OP_CLOSE] = {},
983 [IORING_OP_FILES_UPDATE] = {},
984 [IORING_OP_STATX] = {},
987 .unbound_nonreg_file = 1,
991 .async_size = sizeof(struct io_async_rw),
993 [IORING_OP_WRITE] = {
995 .unbound_nonreg_file = 1,
998 .async_size = sizeof(struct io_async_rw),
1000 [IORING_OP_FADVISE] = {
1003 [IORING_OP_MADVISE] = {},
1004 [IORING_OP_SEND] = {
1006 .unbound_nonreg_file = 1,
1009 [IORING_OP_RECV] = {
1011 .unbound_nonreg_file = 1,
1015 [IORING_OP_OPENAT2] = {
1017 [IORING_OP_EPOLL_CTL] = {
1018 .unbound_nonreg_file = 1,
1020 [IORING_OP_SPLICE] = {
1023 .unbound_nonreg_file = 1,
1025 [IORING_OP_PROVIDE_BUFFERS] = {},
1026 [IORING_OP_REMOVE_BUFFERS] = {},
1030 .unbound_nonreg_file = 1,
1032 [IORING_OP_SHUTDOWN] = {
1035 [IORING_OP_RENAMEAT] = {},
1036 [IORING_OP_UNLINKAT] = {},
1039 static bool io_disarm_next(struct io_kiocb *req);
1040 static void io_uring_del_tctx_node(unsigned long index);
1041 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1042 struct task_struct *task,
1044 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1045 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1047 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1048 long res, unsigned int cflags);
1049 static void io_put_req(struct io_kiocb *req);
1050 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1051 static void io_dismantle_req(struct io_kiocb *req);
1052 static void io_put_task(struct task_struct *task, int nr);
1053 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1054 static void io_queue_linked_timeout(struct io_kiocb *req);
1055 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1056 struct io_uring_rsrc_update2 *up,
1058 static void io_clean_op(struct io_kiocb *req);
1059 static struct file *io_file_get(struct io_submit_state *state,
1060 struct io_kiocb *req, int fd, bool fixed);
1061 static void __io_queue_sqe(struct io_kiocb *req);
1062 static void io_rsrc_put_work(struct work_struct *work);
1064 static void io_req_task_queue(struct io_kiocb *req);
1065 static void io_submit_flush_completions(struct io_comp_state *cs,
1066 struct io_ring_ctx *ctx);
1067 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1068 static int io_req_prep_async(struct io_kiocb *req);
1070 static struct kmem_cache *req_cachep;
1072 static const struct file_operations io_uring_fops;
1074 struct sock *io_uring_get_socket(struct file *file)
1076 #if defined(CONFIG_UNIX)
1077 if (file->f_op == &io_uring_fops) {
1078 struct io_ring_ctx *ctx = file->private_data;
1080 return ctx->ring_sock->sk;
1085 EXPORT_SYMBOL(io_uring_get_socket);
1087 #define io_for_each_link(pos, head) \
1088 for (pos = (head); pos; pos = pos->link)
1090 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1092 struct io_ring_ctx *ctx = req->ctx;
1094 if (!req->fixed_rsrc_refs) {
1095 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1096 percpu_ref_get(req->fixed_rsrc_refs);
1100 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1102 bool got = percpu_ref_tryget(ref);
1104 /* already at zero, wait for ->release() */
1106 wait_for_completion(compl);
1107 percpu_ref_resurrect(ref);
1109 percpu_ref_put(ref);
1112 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1115 struct io_kiocb *req;
1117 if (task && head->task != task)
1122 io_for_each_link(req, head) {
1123 if (req->flags & REQ_F_INFLIGHT)
1129 static inline void req_set_fail(struct io_kiocb *req)
1131 req->flags |= REQ_F_FAIL;
1134 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1136 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1138 complete(&ctx->ref_comp);
1141 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1143 return !req->timeout.off;
1146 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1148 struct io_ring_ctx *ctx;
1151 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1156 * Use 5 bits less than the max cq entries, that should give us around
1157 * 32 entries per hash list if totally full and uniformly spread.
1159 hash_bits = ilog2(p->cq_entries);
1163 ctx->cancel_hash_bits = hash_bits;
1164 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1166 if (!ctx->cancel_hash)
1168 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1170 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1171 if (!ctx->dummy_ubuf)
1173 /* set invalid range, so io_import_fixed() fails meeting it */
1174 ctx->dummy_ubuf->ubuf = -1UL;
1176 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1177 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1180 ctx->flags = p->flags;
1181 init_waitqueue_head(&ctx->sqo_sq_wait);
1182 INIT_LIST_HEAD(&ctx->sqd_list);
1183 init_waitqueue_head(&ctx->cq_wait);
1184 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1185 init_completion(&ctx->ref_comp);
1186 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1187 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1188 mutex_init(&ctx->uring_lock);
1189 init_waitqueue_head(&ctx->wait);
1190 spin_lock_init(&ctx->completion_lock);
1191 INIT_LIST_HEAD(&ctx->iopoll_list);
1192 INIT_LIST_HEAD(&ctx->defer_list);
1193 INIT_LIST_HEAD(&ctx->timeout_list);
1194 spin_lock_init(&ctx->rsrc_ref_lock);
1195 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1196 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1197 init_llist_head(&ctx->rsrc_put_llist);
1198 INIT_LIST_HEAD(&ctx->tctx_list);
1199 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1200 INIT_LIST_HEAD(&ctx->locked_free_list);
1203 kfree(ctx->dummy_ubuf);
1204 kfree(ctx->cancel_hash);
1209 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1211 struct io_rings *r = ctx->rings;
1213 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1217 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1219 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1220 struct io_ring_ctx *ctx = req->ctx;
1222 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1228 static void io_req_track_inflight(struct io_kiocb *req)
1230 if (!(req->flags & REQ_F_INFLIGHT)) {
1231 req->flags |= REQ_F_INFLIGHT;
1232 atomic_inc(¤t->io_uring->inflight_tracked);
1236 static void io_prep_async_work(struct io_kiocb *req)
1238 const struct io_op_def *def = &io_op_defs[req->opcode];
1239 struct io_ring_ctx *ctx = req->ctx;
1241 if (!req->work.creds)
1242 req->work.creds = get_current_cred();
1244 req->work.list.next = NULL;
1245 req->work.flags = 0;
1246 if (req->flags & REQ_F_FORCE_ASYNC)
1247 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1249 if (req->flags & REQ_F_ISREG) {
1250 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1251 io_wq_hash_work(&req->work, file_inode(req->file));
1252 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1253 if (def->unbound_nonreg_file)
1254 req->work.flags |= IO_WQ_WORK_UNBOUND;
1257 switch (req->opcode) {
1258 case IORING_OP_SPLICE:
1260 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1261 req->work.flags |= IO_WQ_WORK_UNBOUND;
1266 static void io_prep_async_link(struct io_kiocb *req)
1268 struct io_kiocb *cur;
1270 io_for_each_link(cur, req)
1271 io_prep_async_work(cur);
1274 static void io_queue_async_work(struct io_kiocb *req)
1276 struct io_ring_ctx *ctx = req->ctx;
1277 struct io_kiocb *link = io_prep_linked_timeout(req);
1278 struct io_uring_task *tctx = req->task->io_uring;
1281 BUG_ON(!tctx->io_wq);
1283 /* init ->work of the whole link before punting */
1284 io_prep_async_link(req);
1285 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1286 &req->work, req->flags);
1287 io_wq_enqueue(tctx->io_wq, &req->work);
1289 io_queue_linked_timeout(link);
1292 static void io_kill_timeout(struct io_kiocb *req, int status)
1293 __must_hold(&req->ctx->completion_lock)
1295 struct io_timeout_data *io = req->async_data;
1297 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1298 atomic_set(&req->ctx->cq_timeouts,
1299 atomic_read(&req->ctx->cq_timeouts) + 1);
1300 list_del_init(&req->timeout.list);
1301 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1302 io_put_req_deferred(req, 1);
1306 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1309 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1310 struct io_defer_entry, list);
1312 if (req_need_defer(de->req, de->seq))
1314 list_del_init(&de->list);
1315 io_req_task_queue(de->req);
1317 } while (!list_empty(&ctx->defer_list));
1320 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1324 if (list_empty(&ctx->timeout_list))
1327 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1330 u32 events_needed, events_got;
1331 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1332 struct io_kiocb, timeout.list);
1334 if (io_is_timeout_noseq(req))
1338 * Since seq can easily wrap around over time, subtract
1339 * the last seq at which timeouts were flushed before comparing.
1340 * Assuming not more than 2^31-1 events have happened since,
1341 * these subtractions won't have wrapped, so we can check if
1342 * target is in [last_seq, current_seq] by comparing the two.
1344 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1345 events_got = seq - ctx->cq_last_tm_flush;
1346 if (events_got < events_needed)
1349 list_del_init(&req->timeout.list);
1350 io_kill_timeout(req, 0);
1351 } while (!list_empty(&ctx->timeout_list));
1353 ctx->cq_last_tm_flush = seq;
1356 static void io_commit_cqring(struct io_ring_ctx *ctx)
1358 io_flush_timeouts(ctx);
1360 /* order cqe stores with ring update */
1361 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1363 if (unlikely(!list_empty(&ctx->defer_list)))
1364 __io_queue_deferred(ctx);
1367 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1369 struct io_rings *r = ctx->rings;
1371 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1374 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1376 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1379 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1381 struct io_rings *rings = ctx->rings;
1382 unsigned tail, mask = ctx->cq_entries - 1;
1385 * writes to the cq entry need to come after reading head; the
1386 * control dependency is enough as we're using WRITE_ONCE to
1389 if (__io_cqring_events(ctx) == ctx->cq_entries)
1392 tail = ctx->cached_cq_tail++;
1393 return &rings->cqes[tail & mask];
1396 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1398 if (likely(!ctx->cq_ev_fd))
1400 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1402 return !ctx->eventfd_async || io_wq_current_is_worker();
1405 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1407 /* see waitqueue_active() comment */
1410 if (waitqueue_active(&ctx->wait))
1411 wake_up(&ctx->wait);
1412 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1413 wake_up(&ctx->sq_data->wait);
1414 if (io_should_trigger_evfd(ctx))
1415 eventfd_signal(ctx->cq_ev_fd, 1);
1416 if (waitqueue_active(&ctx->cq_wait)) {
1417 wake_up_interruptible(&ctx->cq_wait);
1418 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1422 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1424 /* see waitqueue_active() comment */
1427 if (ctx->flags & IORING_SETUP_SQPOLL) {
1428 if (waitqueue_active(&ctx->wait))
1429 wake_up(&ctx->wait);
1431 if (io_should_trigger_evfd(ctx))
1432 eventfd_signal(ctx->cq_ev_fd, 1);
1433 if (waitqueue_active(&ctx->cq_wait)) {
1434 wake_up_interruptible(&ctx->cq_wait);
1435 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1439 /* Returns true if there are no backlogged entries after the flush */
1440 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1442 unsigned long flags;
1443 bool all_flushed, posted;
1445 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1449 spin_lock_irqsave(&ctx->completion_lock, flags);
1450 while (!list_empty(&ctx->cq_overflow_list)) {
1451 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1452 struct io_overflow_cqe *ocqe;
1456 ocqe = list_first_entry(&ctx->cq_overflow_list,
1457 struct io_overflow_cqe, list);
1459 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1461 io_account_cq_overflow(ctx);
1464 list_del(&ocqe->list);
1468 all_flushed = list_empty(&ctx->cq_overflow_list);
1470 clear_bit(0, &ctx->sq_check_overflow);
1471 clear_bit(0, &ctx->cq_check_overflow);
1472 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1476 io_commit_cqring(ctx);
1477 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1479 io_cqring_ev_posted(ctx);
1483 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1487 if (test_bit(0, &ctx->cq_check_overflow)) {
1488 /* iopoll syncs against uring_lock, not completion_lock */
1489 if (ctx->flags & IORING_SETUP_IOPOLL)
1490 mutex_lock(&ctx->uring_lock);
1491 ret = __io_cqring_overflow_flush(ctx, force);
1492 if (ctx->flags & IORING_SETUP_IOPOLL)
1493 mutex_unlock(&ctx->uring_lock);
1500 * Shamelessly stolen from the mm implementation of page reference checking,
1501 * see commit f958d7b528b1 for details.
1503 #define req_ref_zero_or_close_to_overflow(req) \
1504 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1506 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1508 return atomic_inc_not_zero(&req->refs);
1511 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1513 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1514 return atomic_sub_and_test(refs, &req->refs);
1517 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1519 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1520 return atomic_dec_and_test(&req->refs);
1523 static inline void req_ref_put(struct io_kiocb *req)
1525 WARN_ON_ONCE(req_ref_put_and_test(req));
1528 static inline void req_ref_get(struct io_kiocb *req)
1530 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1531 atomic_inc(&req->refs);
1534 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1535 long res, unsigned int cflags)
1537 struct io_overflow_cqe *ocqe;
1539 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1542 * If we're in ring overflow flush mode, or in task cancel mode,
1543 * or cannot allocate an overflow entry, then we need to drop it
1546 io_account_cq_overflow(ctx);
1549 if (list_empty(&ctx->cq_overflow_list)) {
1550 set_bit(0, &ctx->sq_check_overflow);
1551 set_bit(0, &ctx->cq_check_overflow);
1552 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1554 ocqe->cqe.user_data = user_data;
1555 ocqe->cqe.res = res;
1556 ocqe->cqe.flags = cflags;
1557 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1561 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1562 long res, unsigned int cflags)
1564 struct io_uring_cqe *cqe;
1566 trace_io_uring_complete(ctx, user_data, res, cflags);
1569 * If we can't get a cq entry, userspace overflowed the
1570 * submission (by quite a lot). Increment the overflow count in
1573 cqe = io_get_cqe(ctx);
1575 WRITE_ONCE(cqe->user_data, user_data);
1576 WRITE_ONCE(cqe->res, res);
1577 WRITE_ONCE(cqe->flags, cflags);
1580 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1583 /* not as hot to bloat with inlining */
1584 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1585 long res, unsigned int cflags)
1587 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1590 static void io_req_complete_post(struct io_kiocb *req, long res,
1591 unsigned int cflags)
1593 struct io_ring_ctx *ctx = req->ctx;
1594 unsigned long flags;
1596 spin_lock_irqsave(&ctx->completion_lock, flags);
1597 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1599 * If we're the last reference to this request, add to our locked
1602 if (req_ref_put_and_test(req)) {
1603 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1604 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1605 io_disarm_next(req);
1607 io_req_task_queue(req->link);
1611 io_dismantle_req(req);
1612 io_put_task(req->task, 1);
1613 list_add(&req->compl.list, &ctx->locked_free_list);
1614 ctx->locked_free_nr++;
1616 if (!percpu_ref_tryget(&ctx->refs))
1619 io_commit_cqring(ctx);
1620 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1623 io_cqring_ev_posted(ctx);
1624 percpu_ref_put(&ctx->refs);
1628 static inline bool io_req_needs_clean(struct io_kiocb *req)
1630 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1631 REQ_F_POLLED | REQ_F_INFLIGHT);
1634 static void io_req_complete_state(struct io_kiocb *req, long res,
1635 unsigned int cflags)
1637 if (io_req_needs_clean(req))
1640 req->compl.cflags = cflags;
1641 req->flags |= REQ_F_COMPLETE_INLINE;
1644 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1645 long res, unsigned cflags)
1647 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1648 io_req_complete_state(req, res, cflags);
1650 io_req_complete_post(req, res, cflags);
1653 static inline void io_req_complete(struct io_kiocb *req, long res)
1655 __io_req_complete(req, 0, res, 0);
1658 static void io_req_complete_failed(struct io_kiocb *req, long res)
1662 io_req_complete_post(req, res, 0);
1665 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1666 struct io_comp_state *cs)
1668 spin_lock_irq(&ctx->completion_lock);
1669 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1670 ctx->locked_free_nr = 0;
1671 spin_unlock_irq(&ctx->completion_lock);
1674 /* Returns true IFF there are requests in the cache */
1675 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1677 struct io_submit_state *state = &ctx->submit_state;
1678 struct io_comp_state *cs = &state->comp;
1682 * If we have more than a batch's worth of requests in our IRQ side
1683 * locked cache, grab the lock and move them over to our submission
1686 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1687 io_flush_cached_locked_reqs(ctx, cs);
1689 nr = state->free_reqs;
1690 while (!list_empty(&cs->free_list)) {
1691 struct io_kiocb *req = list_first_entry(&cs->free_list,
1692 struct io_kiocb, compl.list);
1694 list_del(&req->compl.list);
1695 state->reqs[nr++] = req;
1696 if (nr == ARRAY_SIZE(state->reqs))
1700 state->free_reqs = nr;
1704 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1706 struct io_submit_state *state = &ctx->submit_state;
1708 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1710 if (!state->free_reqs) {
1711 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1714 if (io_flush_cached_reqs(ctx))
1717 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1721 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1722 * retry single alloc to be on the safe side.
1724 if (unlikely(ret <= 0)) {
1725 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1726 if (!state->reqs[0])
1730 state->free_reqs = ret;
1734 return state->reqs[state->free_reqs];
1737 static inline void io_put_file(struct file *file)
1743 static void io_dismantle_req(struct io_kiocb *req)
1745 unsigned int flags = req->flags;
1747 if (io_req_needs_clean(req))
1749 if (!(flags & REQ_F_FIXED_FILE))
1750 io_put_file(req->file);
1751 if (req->fixed_rsrc_refs)
1752 percpu_ref_put(req->fixed_rsrc_refs);
1753 if (req->async_data)
1754 kfree(req->async_data);
1755 if (req->work.creds) {
1756 put_cred(req->work.creds);
1757 req->work.creds = NULL;
1761 /* must to be called somewhat shortly after putting a request */
1762 static inline void io_put_task(struct task_struct *task, int nr)
1764 struct io_uring_task *tctx = task->io_uring;
1766 percpu_counter_sub(&tctx->inflight, nr);
1767 if (unlikely(atomic_read(&tctx->in_idle)))
1768 wake_up(&tctx->wait);
1769 put_task_struct_many(task, nr);
1772 static void __io_free_req(struct io_kiocb *req)
1774 struct io_ring_ctx *ctx = req->ctx;
1776 io_dismantle_req(req);
1777 io_put_task(req->task, 1);
1779 kmem_cache_free(req_cachep, req);
1780 percpu_ref_put(&ctx->refs);
1783 static inline void io_remove_next_linked(struct io_kiocb *req)
1785 struct io_kiocb *nxt = req->link;
1787 req->link = nxt->link;
1791 static bool io_kill_linked_timeout(struct io_kiocb *req)
1792 __must_hold(&req->ctx->completion_lock)
1794 struct io_kiocb *link = req->link;
1797 * Can happen if a linked timeout fired and link had been like
1798 * req -> link t-out -> link t-out [-> ...]
1800 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1801 struct io_timeout_data *io = link->async_data;
1803 io_remove_next_linked(req);
1804 link->timeout.head = NULL;
1805 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1806 io_cqring_fill_event(link->ctx, link->user_data,
1808 io_put_req_deferred(link, 1);
1815 static void io_fail_links(struct io_kiocb *req)
1816 __must_hold(&req->ctx->completion_lock)
1818 struct io_kiocb *nxt, *link = req->link;
1825 trace_io_uring_fail_link(req, link);
1826 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1827 io_put_req_deferred(link, 2);
1832 static bool io_disarm_next(struct io_kiocb *req)
1833 __must_hold(&req->ctx->completion_lock)
1835 bool posted = false;
1837 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1838 posted = io_kill_linked_timeout(req);
1839 if (unlikely((req->flags & REQ_F_FAIL) &&
1840 !(req->flags & REQ_F_HARDLINK))) {
1841 posted |= (req->link != NULL);
1847 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1849 struct io_kiocb *nxt;
1852 * If LINK is set, we have dependent requests in this chain. If we
1853 * didn't fail this request, queue the first one up, moving any other
1854 * dependencies to the next request. In case of failure, fail the rest
1857 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1858 struct io_ring_ctx *ctx = req->ctx;
1859 unsigned long flags;
1862 spin_lock_irqsave(&ctx->completion_lock, flags);
1863 posted = io_disarm_next(req);
1865 io_commit_cqring(req->ctx);
1866 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1868 io_cqring_ev_posted(ctx);
1875 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1877 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1879 return __io_req_find_next(req);
1882 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1886 if (ctx->submit_state.comp.nr) {
1887 mutex_lock(&ctx->uring_lock);
1888 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1889 mutex_unlock(&ctx->uring_lock);
1891 percpu_ref_put(&ctx->refs);
1894 static bool __tctx_task_work(struct io_uring_task *tctx)
1896 struct io_ring_ctx *ctx = NULL;
1897 struct io_wq_work_list list;
1898 struct io_wq_work_node *node;
1900 if (wq_list_empty(&tctx->task_list))
1903 spin_lock_irq(&tctx->task_lock);
1904 list = tctx->task_list;
1905 INIT_WQ_LIST(&tctx->task_list);
1906 spin_unlock_irq(&tctx->task_lock);
1910 struct io_wq_work_node *next = node->next;
1911 struct io_kiocb *req;
1913 req = container_of(node, struct io_kiocb, io_task_work.node);
1914 if (req->ctx != ctx) {
1915 ctx_flush_and_put(ctx);
1917 percpu_ref_get(&ctx->refs);
1920 req->task_work.func(&req->task_work);
1924 ctx_flush_and_put(ctx);
1925 return list.first != NULL;
1928 static void tctx_task_work(struct callback_head *cb)
1930 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1932 clear_bit(0, &tctx->task_state);
1934 while (__tctx_task_work(tctx))
1938 static int io_req_task_work_add(struct io_kiocb *req)
1940 struct task_struct *tsk = req->task;
1941 struct io_uring_task *tctx = tsk->io_uring;
1942 enum task_work_notify_mode notify;
1943 struct io_wq_work_node *node, *prev;
1944 unsigned long flags;
1947 if (unlikely(tsk->flags & PF_EXITING))
1950 WARN_ON_ONCE(!tctx);
1952 spin_lock_irqsave(&tctx->task_lock, flags);
1953 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1954 spin_unlock_irqrestore(&tctx->task_lock, flags);
1956 /* task_work already pending, we're done */
1957 if (test_bit(0, &tctx->task_state) ||
1958 test_and_set_bit(0, &tctx->task_state))
1962 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1963 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1964 * processing task_work. There's no reliable way to tell if TWA_RESUME
1967 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1969 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1970 wake_up_process(tsk);
1975 * Slow path - we failed, find and delete work. if the work is not
1976 * in the list, it got run and we're fine.
1978 spin_lock_irqsave(&tctx->task_lock, flags);
1979 wq_list_for_each(node, prev, &tctx->task_list) {
1980 if (&req->io_task_work.node == node) {
1981 wq_list_del(&tctx->task_list, node, prev);
1986 spin_unlock_irqrestore(&tctx->task_lock, flags);
1987 clear_bit(0, &tctx->task_state);
1991 static bool io_run_task_work_head(struct callback_head **work_head)
1993 struct callback_head *work, *next;
1994 bool executed = false;
1997 work = xchg(work_head, NULL);
2013 static void io_task_work_add_head(struct callback_head **work_head,
2014 struct callback_head *task_work)
2016 struct callback_head *head;
2019 head = READ_ONCE(*work_head);
2020 task_work->next = head;
2021 } while (cmpxchg(work_head, head, task_work) != head);
2024 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2025 task_work_func_t cb)
2027 init_task_work(&req->task_work, cb);
2028 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2031 static void io_req_task_cancel(struct callback_head *cb)
2033 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2034 struct io_ring_ctx *ctx = req->ctx;
2036 /* ctx is guaranteed to stay alive while we hold uring_lock */
2037 mutex_lock(&ctx->uring_lock);
2038 io_req_complete_failed(req, req->result);
2039 mutex_unlock(&ctx->uring_lock);
2042 static void __io_req_task_submit(struct io_kiocb *req)
2044 struct io_ring_ctx *ctx = req->ctx;
2046 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2047 mutex_lock(&ctx->uring_lock);
2048 if (!(current->flags & PF_EXITING) && !current->in_execve)
2049 __io_queue_sqe(req);
2051 io_req_complete_failed(req, -EFAULT);
2052 mutex_unlock(&ctx->uring_lock);
2055 static void io_req_task_submit(struct callback_head *cb)
2057 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2059 __io_req_task_submit(req);
2062 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2065 req->task_work.func = io_req_task_cancel;
2067 if (unlikely(io_req_task_work_add(req)))
2068 io_req_task_work_add_fallback(req, io_req_task_cancel);
2071 static void io_req_task_queue(struct io_kiocb *req)
2073 req->task_work.func = io_req_task_submit;
2075 if (unlikely(io_req_task_work_add(req)))
2076 io_req_task_queue_fail(req, -ECANCELED);
2079 static inline void io_queue_next(struct io_kiocb *req)
2081 struct io_kiocb *nxt = io_req_find_next(req);
2084 io_req_task_queue(nxt);
2087 static void io_free_req(struct io_kiocb *req)
2094 struct task_struct *task;
2099 static inline void io_init_req_batch(struct req_batch *rb)
2106 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2107 struct req_batch *rb)
2110 io_put_task(rb->task, rb->task_refs);
2112 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2115 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2116 struct io_submit_state *state)
2119 io_dismantle_req(req);
2121 if (req->task != rb->task) {
2123 io_put_task(rb->task, rb->task_refs);
2124 rb->task = req->task;
2130 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2131 state->reqs[state->free_reqs++] = req;
2133 list_add(&req->compl.list, &state->comp.free_list);
2136 static void io_submit_flush_completions(struct io_comp_state *cs,
2137 struct io_ring_ctx *ctx)
2140 struct io_kiocb *req;
2141 struct req_batch rb;
2143 io_init_req_batch(&rb);
2144 spin_lock_irq(&ctx->completion_lock);
2145 for (i = 0; i < nr; i++) {
2147 __io_cqring_fill_event(ctx, req->user_data, req->result,
2150 io_commit_cqring(ctx);
2151 spin_unlock_irq(&ctx->completion_lock);
2153 io_cqring_ev_posted(ctx);
2154 for (i = 0; i < nr; i++) {
2157 /* submission and completion refs */
2158 if (req_ref_sub_and_test(req, 2))
2159 io_req_free_batch(&rb, req, &ctx->submit_state);
2162 io_req_free_batch_finish(ctx, &rb);
2167 * Drop reference to request, return next in chain (if there is one) if this
2168 * was the last reference to this request.
2170 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2172 struct io_kiocb *nxt = NULL;
2174 if (req_ref_put_and_test(req)) {
2175 nxt = io_req_find_next(req);
2181 static inline void io_put_req(struct io_kiocb *req)
2183 if (req_ref_put_and_test(req))
2187 static void io_put_req_deferred_cb(struct callback_head *cb)
2189 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2194 static void io_free_req_deferred(struct io_kiocb *req)
2196 req->task_work.func = io_put_req_deferred_cb;
2197 if (unlikely(io_req_task_work_add(req)))
2198 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2201 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2203 if (req_ref_sub_and_test(req, refs))
2204 io_free_req_deferred(req);
2207 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2209 /* See comment at the top of this file */
2211 return __io_cqring_events(ctx);
2214 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2216 struct io_rings *rings = ctx->rings;
2218 /* make sure SQ entry isn't read before tail */
2219 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2222 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2224 unsigned int cflags;
2226 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2227 cflags |= IORING_CQE_F_BUFFER;
2228 req->flags &= ~REQ_F_BUFFER_SELECTED;
2233 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2235 struct io_buffer *kbuf;
2237 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2238 return io_put_kbuf(req, kbuf);
2241 static inline bool io_run_task_work(void)
2244 * Not safe to run on exiting task, and the task_work handling will
2245 * not add work to such a task.
2247 if (unlikely(current->flags & PF_EXITING))
2249 if (current->task_works) {
2250 __set_current_state(TASK_RUNNING);
2259 * Find and free completed poll iocbs
2261 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2262 struct list_head *done)
2264 struct req_batch rb;
2265 struct io_kiocb *req;
2267 /* order with ->result store in io_complete_rw_iopoll() */
2270 io_init_req_batch(&rb);
2271 while (!list_empty(done)) {
2274 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2275 list_del(&req->inflight_entry);
2277 if (READ_ONCE(req->result) == -EAGAIN &&
2278 !(req->flags & REQ_F_DONT_REISSUE)) {
2279 req->iopoll_completed = 0;
2281 io_queue_async_work(req);
2285 if (req->flags & REQ_F_BUFFER_SELECTED)
2286 cflags = io_put_rw_kbuf(req);
2288 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2291 if (req_ref_put_and_test(req))
2292 io_req_free_batch(&rb, req, &ctx->submit_state);
2295 io_commit_cqring(ctx);
2296 io_cqring_ev_posted_iopoll(ctx);
2297 io_req_free_batch_finish(ctx, &rb);
2300 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2303 struct io_kiocb *req, *tmp;
2309 * Only spin for completions if we don't have multiple devices hanging
2310 * off our complete list, and we're under the requested amount.
2312 spin = !ctx->poll_multi_file && *nr_events < min;
2315 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2316 struct kiocb *kiocb = &req->rw.kiocb;
2319 * Move completed and retryable entries to our local lists.
2320 * If we find a request that requires polling, break out
2321 * and complete those lists first, if we have entries there.
2323 if (READ_ONCE(req->iopoll_completed)) {
2324 list_move_tail(&req->inflight_entry, &done);
2327 if (!list_empty(&done))
2330 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2334 /* iopoll may have completed current req */
2335 if (READ_ONCE(req->iopoll_completed))
2336 list_move_tail(&req->inflight_entry, &done);
2343 if (!list_empty(&done))
2344 io_iopoll_complete(ctx, nr_events, &done);
2350 * We can't just wait for polled events to come to us, we have to actively
2351 * find and complete them.
2353 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2355 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2358 mutex_lock(&ctx->uring_lock);
2359 while (!list_empty(&ctx->iopoll_list)) {
2360 unsigned int nr_events = 0;
2362 io_do_iopoll(ctx, &nr_events, 0);
2364 /* let it sleep and repeat later if can't complete a request */
2368 * Ensure we allow local-to-the-cpu processing to take place,
2369 * in this case we need to ensure that we reap all events.
2370 * Also let task_work, etc. to progress by releasing the mutex
2372 if (need_resched()) {
2373 mutex_unlock(&ctx->uring_lock);
2375 mutex_lock(&ctx->uring_lock);
2378 mutex_unlock(&ctx->uring_lock);
2381 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2383 unsigned int nr_events = 0;
2387 * We disallow the app entering submit/complete with polling, but we
2388 * still need to lock the ring to prevent racing with polled issue
2389 * that got punted to a workqueue.
2391 mutex_lock(&ctx->uring_lock);
2393 * Don't enter poll loop if we already have events pending.
2394 * If we do, we can potentially be spinning for commands that
2395 * already triggered a CQE (eg in error).
2397 if (test_bit(0, &ctx->cq_check_overflow))
2398 __io_cqring_overflow_flush(ctx, false);
2399 if (io_cqring_events(ctx))
2403 * If a submit got punted to a workqueue, we can have the
2404 * application entering polling for a command before it gets
2405 * issued. That app will hold the uring_lock for the duration
2406 * of the poll right here, so we need to take a breather every
2407 * now and then to ensure that the issue has a chance to add
2408 * the poll to the issued list. Otherwise we can spin here
2409 * forever, while the workqueue is stuck trying to acquire the
2412 if (list_empty(&ctx->iopoll_list)) {
2413 mutex_unlock(&ctx->uring_lock);
2415 mutex_lock(&ctx->uring_lock);
2417 if (list_empty(&ctx->iopoll_list))
2420 ret = io_do_iopoll(ctx, &nr_events, min);
2421 } while (!ret && nr_events < min && !need_resched());
2423 mutex_unlock(&ctx->uring_lock);
2427 static void kiocb_end_write(struct io_kiocb *req)
2430 * Tell lockdep we inherited freeze protection from submission
2433 if (req->flags & REQ_F_ISREG) {
2434 struct super_block *sb = file_inode(req->file)->i_sb;
2436 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2442 static bool io_resubmit_prep(struct io_kiocb *req)
2444 struct io_async_rw *rw = req->async_data;
2447 return !io_req_prep_async(req);
2448 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2449 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2453 static bool io_rw_should_reissue(struct io_kiocb *req)
2455 umode_t mode = file_inode(req->file)->i_mode;
2456 struct io_ring_ctx *ctx = req->ctx;
2458 if (!S_ISBLK(mode) && !S_ISREG(mode))
2460 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2461 !(ctx->flags & IORING_SETUP_IOPOLL)))
2464 * If ref is dying, we might be running poll reap from the exit work.
2465 * Don't attempt to reissue from that path, just let it fail with
2468 if (percpu_ref_is_dying(&ctx->refs))
2473 static bool io_resubmit_prep(struct io_kiocb *req)
2477 static bool io_rw_should_reissue(struct io_kiocb *req)
2483 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2484 unsigned int issue_flags)
2488 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2489 kiocb_end_write(req);
2490 if (res != req->result) {
2491 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2492 io_rw_should_reissue(req)) {
2493 req->flags |= REQ_F_REISSUE;
2498 if (req->flags & REQ_F_BUFFER_SELECTED)
2499 cflags = io_put_rw_kbuf(req);
2500 __io_req_complete(req, issue_flags, res, cflags);
2503 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2505 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2507 __io_complete_rw(req, res, res2, 0);
2510 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2512 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2514 if (kiocb->ki_flags & IOCB_WRITE)
2515 kiocb_end_write(req);
2516 if (unlikely(res != req->result)) {
2517 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2518 io_resubmit_prep(req))) {
2520 req->flags |= REQ_F_DONT_REISSUE;
2524 WRITE_ONCE(req->result, res);
2525 /* order with io_iopoll_complete() checking ->result */
2527 WRITE_ONCE(req->iopoll_completed, 1);
2531 * After the iocb has been issued, it's safe to be found on the poll list.
2532 * Adding the kiocb to the list AFTER submission ensures that we don't
2533 * find it from a io_do_iopoll() thread before the issuer is done
2534 * accessing the kiocb cookie.
2536 static void io_iopoll_req_issued(struct io_kiocb *req)
2538 struct io_ring_ctx *ctx = req->ctx;
2539 const bool in_async = io_wq_current_is_worker();
2541 /* workqueue context doesn't hold uring_lock, grab it now */
2542 if (unlikely(in_async))
2543 mutex_lock(&ctx->uring_lock);
2546 * Track whether we have multiple files in our lists. This will impact
2547 * how we do polling eventually, not spinning if we're on potentially
2548 * different devices.
2550 if (list_empty(&ctx->iopoll_list)) {
2551 ctx->poll_multi_file = false;
2552 } else if (!ctx->poll_multi_file) {
2553 struct io_kiocb *list_req;
2555 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2557 if (list_req->file != req->file)
2558 ctx->poll_multi_file = true;
2562 * For fast devices, IO may have already completed. If it has, add
2563 * it to the front so we find it first.
2565 if (READ_ONCE(req->iopoll_completed))
2566 list_add(&req->inflight_entry, &ctx->iopoll_list);
2568 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2570 if (unlikely(in_async)) {
2572 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2573 * in sq thread task context or in io worker task context. If
2574 * current task context is sq thread, we don't need to check
2575 * whether should wake up sq thread.
2577 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2578 wq_has_sleeper(&ctx->sq_data->wait))
2579 wake_up(&ctx->sq_data->wait);
2581 mutex_unlock(&ctx->uring_lock);
2585 static inline void io_state_file_put(struct io_submit_state *state)
2587 if (state->file_refs) {
2588 fput_many(state->file, state->file_refs);
2589 state->file_refs = 0;
2594 * Get as many references to a file as we have IOs left in this submission,
2595 * assuming most submissions are for one file, or at least that each file
2596 * has more than one submission.
2598 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2603 if (state->file_refs) {
2604 if (state->fd == fd) {
2608 io_state_file_put(state);
2610 state->file = fget_many(fd, state->ios_left);
2611 if (unlikely(!state->file))
2615 state->file_refs = state->ios_left - 1;
2619 static bool io_bdev_nowait(struct block_device *bdev)
2621 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2625 * If we tracked the file through the SCM inflight mechanism, we could support
2626 * any file. For now, just ensure that anything potentially problematic is done
2629 static bool __io_file_supports_async(struct file *file, int rw)
2631 umode_t mode = file_inode(file)->i_mode;
2633 if (S_ISBLK(mode)) {
2634 if (IS_ENABLED(CONFIG_BLOCK) &&
2635 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2641 if (S_ISREG(mode)) {
2642 if (IS_ENABLED(CONFIG_BLOCK) &&
2643 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2644 file->f_op != &io_uring_fops)
2649 /* any ->read/write should understand O_NONBLOCK */
2650 if (file->f_flags & O_NONBLOCK)
2653 if (!(file->f_mode & FMODE_NOWAIT))
2657 return file->f_op->read_iter != NULL;
2659 return file->f_op->write_iter != NULL;
2662 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2664 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2666 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2669 return __io_file_supports_async(req->file, rw);
2672 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2674 struct io_ring_ctx *ctx = req->ctx;
2675 struct kiocb *kiocb = &req->rw.kiocb;
2676 struct file *file = req->file;
2680 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2681 req->flags |= REQ_F_ISREG;
2683 kiocb->ki_pos = READ_ONCE(sqe->off);
2684 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2685 req->flags |= REQ_F_CUR_POS;
2686 kiocb->ki_pos = file->f_pos;
2688 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2689 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2690 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2694 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2695 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2696 req->flags |= REQ_F_NOWAIT;
2698 ioprio = READ_ONCE(sqe->ioprio);
2700 ret = ioprio_check_cap(ioprio);
2704 kiocb->ki_ioprio = ioprio;
2706 kiocb->ki_ioprio = get_current_ioprio();
2708 if (ctx->flags & IORING_SETUP_IOPOLL) {
2709 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2710 !kiocb->ki_filp->f_op->iopoll)
2713 kiocb->ki_flags |= IOCB_HIPRI;
2714 kiocb->ki_complete = io_complete_rw_iopoll;
2715 req->iopoll_completed = 0;
2717 if (kiocb->ki_flags & IOCB_HIPRI)
2719 kiocb->ki_complete = io_complete_rw;
2722 if (req->opcode == IORING_OP_READ_FIXED ||
2723 req->opcode == IORING_OP_WRITE_FIXED) {
2725 io_req_set_rsrc_node(req);
2728 req->rw.addr = READ_ONCE(sqe->addr);
2729 req->rw.len = READ_ONCE(sqe->len);
2730 req->buf_index = READ_ONCE(sqe->buf_index);
2734 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2740 case -ERESTARTNOINTR:
2741 case -ERESTARTNOHAND:
2742 case -ERESTART_RESTARTBLOCK:
2744 * We can't just restart the syscall, since previously
2745 * submitted sqes may already be in progress. Just fail this
2751 kiocb->ki_complete(kiocb, ret, 0);
2755 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2756 unsigned int issue_flags)
2758 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2759 struct io_async_rw *io = req->async_data;
2760 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2762 /* add previously done IO, if any */
2763 if (io && io->bytes_done > 0) {
2765 ret = io->bytes_done;
2767 ret += io->bytes_done;
2770 if (req->flags & REQ_F_CUR_POS)
2771 req->file->f_pos = kiocb->ki_pos;
2772 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2773 __io_complete_rw(req, ret, 0, issue_flags);
2775 io_rw_done(kiocb, ret);
2777 if (check_reissue && req->flags & REQ_F_REISSUE) {
2778 req->flags &= ~REQ_F_REISSUE;
2779 if (io_resubmit_prep(req)) {
2781 io_queue_async_work(req);
2786 if (req->flags & REQ_F_BUFFER_SELECTED)
2787 cflags = io_put_rw_kbuf(req);
2788 __io_req_complete(req, issue_flags, ret, cflags);
2793 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2794 struct io_mapped_ubuf *imu)
2796 size_t len = req->rw.len;
2797 u64 buf_end, buf_addr = req->rw.addr;
2800 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2802 /* not inside the mapped region */
2803 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2807 * May not be a start of buffer, set size appropriately
2808 * and advance us to the beginning.
2810 offset = buf_addr - imu->ubuf;
2811 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2815 * Don't use iov_iter_advance() here, as it's really slow for
2816 * using the latter parts of a big fixed buffer - it iterates
2817 * over each segment manually. We can cheat a bit here, because
2820 * 1) it's a BVEC iter, we set it up
2821 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2822 * first and last bvec
2824 * So just find our index, and adjust the iterator afterwards.
2825 * If the offset is within the first bvec (or the whole first
2826 * bvec, just use iov_iter_advance(). This makes it easier
2827 * since we can just skip the first segment, which may not
2828 * be PAGE_SIZE aligned.
2830 const struct bio_vec *bvec = imu->bvec;
2832 if (offset <= bvec->bv_len) {
2833 iov_iter_advance(iter, offset);
2835 unsigned long seg_skip;
2837 /* skip first vec */
2838 offset -= bvec->bv_len;
2839 seg_skip = 1 + (offset >> PAGE_SHIFT);
2841 iter->bvec = bvec + seg_skip;
2842 iter->nr_segs -= seg_skip;
2843 iter->count -= bvec->bv_len + offset;
2844 iter->iov_offset = offset & ~PAGE_MASK;
2851 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2853 struct io_ring_ctx *ctx = req->ctx;
2854 struct io_mapped_ubuf *imu = req->imu;
2855 u16 index, buf_index = req->buf_index;
2858 if (unlikely(buf_index >= ctx->nr_user_bufs))
2860 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2861 imu = READ_ONCE(ctx->user_bufs[index]);
2864 return __io_import_fixed(req, rw, iter, imu);
2867 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2870 mutex_unlock(&ctx->uring_lock);
2873 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2876 * "Normal" inline submissions always hold the uring_lock, since we
2877 * grab it from the system call. Same is true for the SQPOLL offload.
2878 * The only exception is when we've detached the request and issue it
2879 * from an async worker thread, grab the lock for that case.
2882 mutex_lock(&ctx->uring_lock);
2885 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2886 int bgid, struct io_buffer *kbuf,
2889 struct io_buffer *head;
2891 if (req->flags & REQ_F_BUFFER_SELECTED)
2894 io_ring_submit_lock(req->ctx, needs_lock);
2896 lockdep_assert_held(&req->ctx->uring_lock);
2898 head = xa_load(&req->ctx->io_buffers, bgid);
2900 if (!list_empty(&head->list)) {
2901 kbuf = list_last_entry(&head->list, struct io_buffer,
2903 list_del(&kbuf->list);
2906 xa_erase(&req->ctx->io_buffers, bgid);
2908 if (*len > kbuf->len)
2911 kbuf = ERR_PTR(-ENOBUFS);
2914 io_ring_submit_unlock(req->ctx, needs_lock);
2919 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2922 struct io_buffer *kbuf;
2925 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2926 bgid = req->buf_index;
2927 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2930 req->rw.addr = (u64) (unsigned long) kbuf;
2931 req->flags |= REQ_F_BUFFER_SELECTED;
2932 return u64_to_user_ptr(kbuf->addr);
2935 #ifdef CONFIG_COMPAT
2936 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2939 struct compat_iovec __user *uiov;
2940 compat_ssize_t clen;
2944 uiov = u64_to_user_ptr(req->rw.addr);
2945 if (!access_ok(uiov, sizeof(*uiov)))
2947 if (__get_user(clen, &uiov->iov_len))
2953 buf = io_rw_buffer_select(req, &len, needs_lock);
2955 return PTR_ERR(buf);
2956 iov[0].iov_base = buf;
2957 iov[0].iov_len = (compat_size_t) len;
2962 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2965 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2969 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2972 len = iov[0].iov_len;
2975 buf = io_rw_buffer_select(req, &len, needs_lock);
2977 return PTR_ERR(buf);
2978 iov[0].iov_base = buf;
2979 iov[0].iov_len = len;
2983 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2986 if (req->flags & REQ_F_BUFFER_SELECTED) {
2987 struct io_buffer *kbuf;
2989 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2990 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2991 iov[0].iov_len = kbuf->len;
2994 if (req->rw.len != 1)
2997 #ifdef CONFIG_COMPAT
2998 if (req->ctx->compat)
2999 return io_compat_import(req, iov, needs_lock);
3002 return __io_iov_buffer_select(req, iov, needs_lock);
3005 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3006 struct iov_iter *iter, bool needs_lock)
3008 void __user *buf = u64_to_user_ptr(req->rw.addr);
3009 size_t sqe_len = req->rw.len;
3010 u8 opcode = req->opcode;
3013 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3015 return io_import_fixed(req, rw, iter);
3018 /* buffer index only valid with fixed read/write, or buffer select */
3019 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3022 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3023 if (req->flags & REQ_F_BUFFER_SELECT) {
3024 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3026 return PTR_ERR(buf);
3027 req->rw.len = sqe_len;
3030 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3035 if (req->flags & REQ_F_BUFFER_SELECT) {
3036 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3038 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3043 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3047 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3049 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3053 * For files that don't have ->read_iter() and ->write_iter(), handle them
3054 * by looping over ->read() or ->write() manually.
3056 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3058 struct kiocb *kiocb = &req->rw.kiocb;
3059 struct file *file = req->file;
3063 * Don't support polled IO through this interface, and we can't
3064 * support non-blocking either. For the latter, this just causes
3065 * the kiocb to be handled from an async context.
3067 if (kiocb->ki_flags & IOCB_HIPRI)
3069 if (kiocb->ki_flags & IOCB_NOWAIT)
3072 while (iov_iter_count(iter)) {
3076 if (!iov_iter_is_bvec(iter)) {
3077 iovec = iov_iter_iovec(iter);
3079 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3080 iovec.iov_len = req->rw.len;
3084 nr = file->f_op->read(file, iovec.iov_base,
3085 iovec.iov_len, io_kiocb_ppos(kiocb));
3087 nr = file->f_op->write(file, iovec.iov_base,
3088 iovec.iov_len, io_kiocb_ppos(kiocb));
3097 if (nr != iovec.iov_len)
3101 iov_iter_advance(iter, nr);
3107 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3108 const struct iovec *fast_iov, struct iov_iter *iter)
3110 struct io_async_rw *rw = req->async_data;
3112 memcpy(&rw->iter, iter, sizeof(*iter));
3113 rw->free_iovec = iovec;
3115 /* can only be fixed buffers, no need to do anything */
3116 if (iov_iter_is_bvec(iter))
3119 unsigned iov_off = 0;
3121 rw->iter.iov = rw->fast_iov;
3122 if (iter->iov != fast_iov) {
3123 iov_off = iter->iov - fast_iov;
3124 rw->iter.iov += iov_off;
3126 if (rw->fast_iov != fast_iov)
3127 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3128 sizeof(struct iovec) * iter->nr_segs);
3130 req->flags |= REQ_F_NEED_CLEANUP;
3134 static inline int io_alloc_async_data(struct io_kiocb *req)
3136 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3137 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3138 return req->async_data == NULL;
3141 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3142 const struct iovec *fast_iov,
3143 struct iov_iter *iter, bool force)
3145 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3147 if (!req->async_data) {
3148 if (io_alloc_async_data(req)) {
3153 io_req_map_rw(req, iovec, fast_iov, iter);
3158 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3160 struct io_async_rw *iorw = req->async_data;
3161 struct iovec *iov = iorw->fast_iov;
3164 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3165 if (unlikely(ret < 0))
3168 iorw->bytes_done = 0;
3169 iorw->free_iovec = iov;
3171 req->flags |= REQ_F_NEED_CLEANUP;
3175 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3177 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3179 return io_prep_rw(req, sqe);
3183 * This is our waitqueue callback handler, registered through lock_page_async()
3184 * when we initially tried to do the IO with the iocb armed our waitqueue.
3185 * This gets called when the page is unlocked, and we generally expect that to
3186 * happen when the page IO is completed and the page is now uptodate. This will
3187 * queue a task_work based retry of the operation, attempting to copy the data
3188 * again. If the latter fails because the page was NOT uptodate, then we will
3189 * do a thread based blocking retry of the operation. That's the unexpected
3192 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3193 int sync, void *arg)
3195 struct wait_page_queue *wpq;
3196 struct io_kiocb *req = wait->private;
3197 struct wait_page_key *key = arg;
3199 wpq = container_of(wait, struct wait_page_queue, wait);
3201 if (!wake_page_match(wpq, key))
3204 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3205 list_del_init(&wait->entry);
3207 /* submit ref gets dropped, acquire a new one */
3209 io_req_task_queue(req);
3214 * This controls whether a given IO request should be armed for async page
3215 * based retry. If we return false here, the request is handed to the async
3216 * worker threads for retry. If we're doing buffered reads on a regular file,
3217 * we prepare a private wait_page_queue entry and retry the operation. This
3218 * will either succeed because the page is now uptodate and unlocked, or it
3219 * will register a callback when the page is unlocked at IO completion. Through
3220 * that callback, io_uring uses task_work to setup a retry of the operation.
3221 * That retry will attempt the buffered read again. The retry will generally
3222 * succeed, or in rare cases where it fails, we then fall back to using the
3223 * async worker threads for a blocking retry.
3225 static bool io_rw_should_retry(struct io_kiocb *req)
3227 struct io_async_rw *rw = req->async_data;
3228 struct wait_page_queue *wait = &rw->wpq;
3229 struct kiocb *kiocb = &req->rw.kiocb;
3231 /* never retry for NOWAIT, we just complete with -EAGAIN */
3232 if (req->flags & REQ_F_NOWAIT)
3235 /* Only for buffered IO */
3236 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3240 * just use poll if we can, and don't attempt if the fs doesn't
3241 * support callback based unlocks
3243 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3246 wait->wait.func = io_async_buf_func;
3247 wait->wait.private = req;
3248 wait->wait.flags = 0;
3249 INIT_LIST_HEAD(&wait->wait.entry);
3250 kiocb->ki_flags |= IOCB_WAITQ;
3251 kiocb->ki_flags &= ~IOCB_NOWAIT;
3252 kiocb->ki_waitq = wait;
3256 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3258 if (req->file->f_op->read_iter)
3259 return call_read_iter(req->file, &req->rw.kiocb, iter);
3260 else if (req->file->f_op->read)
3261 return loop_rw_iter(READ, req, iter);
3266 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3268 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3269 struct kiocb *kiocb = &req->rw.kiocb;
3270 struct iov_iter __iter, *iter = &__iter;
3271 struct io_async_rw *rw = req->async_data;
3272 ssize_t io_size, ret, ret2;
3273 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3279 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3283 io_size = iov_iter_count(iter);
3284 req->result = io_size;
3286 /* Ensure we clear previously set non-block flag */
3287 if (!force_nonblock)
3288 kiocb->ki_flags &= ~IOCB_NOWAIT;
3290 kiocb->ki_flags |= IOCB_NOWAIT;
3292 /* If the file doesn't support async, just async punt */
3293 if (force_nonblock && !io_file_supports_async(req, READ)) {
3294 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3295 return ret ?: -EAGAIN;
3298 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3299 if (unlikely(ret)) {
3304 ret = io_iter_do_read(req, iter);
3306 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3307 req->flags &= ~REQ_F_REISSUE;
3308 /* IOPOLL retry should happen for io-wq threads */
3309 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3311 /* no retry on NONBLOCK nor RWF_NOWAIT */
3312 if (req->flags & REQ_F_NOWAIT)
3314 /* some cases will consume bytes even on error returns */
3315 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3317 } else if (ret == -EIOCBQUEUED) {
3319 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3320 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3321 /* read all, failed, already did sync or don't want to retry */
3325 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3330 rw = req->async_data;
3331 /* now use our persistent iterator, if we aren't already */
3336 rw->bytes_done += ret;
3337 /* if we can retry, do so with the callbacks armed */
3338 if (!io_rw_should_retry(req)) {
3339 kiocb->ki_flags &= ~IOCB_WAITQ;
3344 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3345 * we get -EIOCBQUEUED, then we'll get a notification when the
3346 * desired page gets unlocked. We can also get a partial read
3347 * here, and if we do, then just retry at the new offset.
3349 ret = io_iter_do_read(req, iter);
3350 if (ret == -EIOCBQUEUED)
3352 /* we got some bytes, but not all. retry. */
3353 kiocb->ki_flags &= ~IOCB_WAITQ;
3354 } while (ret > 0 && ret < io_size);
3356 kiocb_done(kiocb, ret, issue_flags);
3358 /* it's faster to check here then delegate to kfree */
3364 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3366 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3368 return io_prep_rw(req, sqe);
3371 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3373 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3374 struct kiocb *kiocb = &req->rw.kiocb;
3375 struct iov_iter __iter, *iter = &__iter;
3376 struct io_async_rw *rw = req->async_data;
3377 ssize_t ret, ret2, io_size;
3378 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3384 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3388 io_size = iov_iter_count(iter);
3389 req->result = io_size;
3391 /* Ensure we clear previously set non-block flag */
3392 if (!force_nonblock)
3393 kiocb->ki_flags &= ~IOCB_NOWAIT;
3395 kiocb->ki_flags |= IOCB_NOWAIT;
3397 /* If the file doesn't support async, just async punt */
3398 if (force_nonblock && !io_file_supports_async(req, WRITE))
3401 /* file path doesn't support NOWAIT for non-direct_IO */
3402 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3403 (req->flags & REQ_F_ISREG))
3406 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3411 * Open-code file_start_write here to grab freeze protection,
3412 * which will be released by another thread in
3413 * io_complete_rw(). Fool lockdep by telling it the lock got
3414 * released so that it doesn't complain about the held lock when
3415 * we return to userspace.
3417 if (req->flags & REQ_F_ISREG) {
3418 sb_start_write(file_inode(req->file)->i_sb);
3419 __sb_writers_release(file_inode(req->file)->i_sb,
3422 kiocb->ki_flags |= IOCB_WRITE;
3424 if (req->file->f_op->write_iter)
3425 ret2 = call_write_iter(req->file, kiocb, iter);
3426 else if (req->file->f_op->write)
3427 ret2 = loop_rw_iter(WRITE, req, iter);
3431 if (req->flags & REQ_F_REISSUE) {
3432 req->flags &= ~REQ_F_REISSUE;
3437 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3438 * retry them without IOCB_NOWAIT.
3440 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3442 /* no retry on NONBLOCK nor RWF_NOWAIT */
3443 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3445 if (!force_nonblock || ret2 != -EAGAIN) {
3446 /* IOPOLL retry should happen for io-wq threads */
3447 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3450 kiocb_done(kiocb, ret2, issue_flags);
3453 /* some cases will consume bytes even on error returns */
3454 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3455 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3456 return ret ?: -EAGAIN;
3459 /* it's reportedly faster than delegating the null check to kfree() */
3465 static int io_renameat_prep(struct io_kiocb *req,
3466 const struct io_uring_sqe *sqe)
3468 struct io_rename *ren = &req->rename;
3469 const char __user *oldf, *newf;
3471 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3474 ren->old_dfd = READ_ONCE(sqe->fd);
3475 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3476 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3477 ren->new_dfd = READ_ONCE(sqe->len);
3478 ren->flags = READ_ONCE(sqe->rename_flags);
3480 ren->oldpath = getname(oldf);
3481 if (IS_ERR(ren->oldpath))
3482 return PTR_ERR(ren->oldpath);
3484 ren->newpath = getname(newf);
3485 if (IS_ERR(ren->newpath)) {
3486 putname(ren->oldpath);
3487 return PTR_ERR(ren->newpath);
3490 req->flags |= REQ_F_NEED_CLEANUP;
3494 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3496 struct io_rename *ren = &req->rename;
3499 if (issue_flags & IO_URING_F_NONBLOCK)
3502 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3503 ren->newpath, ren->flags);
3505 req->flags &= ~REQ_F_NEED_CLEANUP;
3508 io_req_complete(req, ret);
3512 static int io_unlinkat_prep(struct io_kiocb *req,
3513 const struct io_uring_sqe *sqe)
3515 struct io_unlink *un = &req->unlink;
3516 const char __user *fname;
3518 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3521 un->dfd = READ_ONCE(sqe->fd);
3523 un->flags = READ_ONCE(sqe->unlink_flags);
3524 if (un->flags & ~AT_REMOVEDIR)
3527 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3528 un->filename = getname(fname);
3529 if (IS_ERR(un->filename))
3530 return PTR_ERR(un->filename);
3532 req->flags |= REQ_F_NEED_CLEANUP;
3536 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3538 struct io_unlink *un = &req->unlink;
3541 if (issue_flags & IO_URING_F_NONBLOCK)
3544 if (un->flags & AT_REMOVEDIR)
3545 ret = do_rmdir(un->dfd, un->filename);
3547 ret = do_unlinkat(un->dfd, un->filename);
3549 req->flags &= ~REQ_F_NEED_CLEANUP;
3552 io_req_complete(req, ret);
3556 static int io_shutdown_prep(struct io_kiocb *req,
3557 const struct io_uring_sqe *sqe)
3559 #if defined(CONFIG_NET)
3560 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3562 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3566 req->shutdown.how = READ_ONCE(sqe->len);
3573 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3575 #if defined(CONFIG_NET)
3576 struct socket *sock;
3579 if (issue_flags & IO_URING_F_NONBLOCK)
3582 sock = sock_from_file(req->file);
3583 if (unlikely(!sock))
3586 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3589 io_req_complete(req, ret);
3596 static int __io_splice_prep(struct io_kiocb *req,
3597 const struct io_uring_sqe *sqe)
3599 struct io_splice* sp = &req->splice;
3600 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3602 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3606 sp->len = READ_ONCE(sqe->len);
3607 sp->flags = READ_ONCE(sqe->splice_flags);
3609 if (unlikely(sp->flags & ~valid_flags))
3612 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3613 (sp->flags & SPLICE_F_FD_IN_FIXED));
3616 req->flags |= REQ_F_NEED_CLEANUP;
3620 static int io_tee_prep(struct io_kiocb *req,
3621 const struct io_uring_sqe *sqe)
3623 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3625 return __io_splice_prep(req, sqe);
3628 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3630 struct io_splice *sp = &req->splice;
3631 struct file *in = sp->file_in;
3632 struct file *out = sp->file_out;
3633 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3636 if (issue_flags & IO_URING_F_NONBLOCK)
3639 ret = do_tee(in, out, sp->len, flags);
3641 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3643 req->flags &= ~REQ_F_NEED_CLEANUP;
3647 io_req_complete(req, ret);
3651 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3653 struct io_splice* sp = &req->splice;
3655 sp->off_in = READ_ONCE(sqe->splice_off_in);
3656 sp->off_out = READ_ONCE(sqe->off);
3657 return __io_splice_prep(req, sqe);
3660 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3662 struct io_splice *sp = &req->splice;
3663 struct file *in = sp->file_in;
3664 struct file *out = sp->file_out;
3665 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3666 loff_t *poff_in, *poff_out;
3669 if (issue_flags & IO_URING_F_NONBLOCK)
3672 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3673 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3676 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3678 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3680 req->flags &= ~REQ_F_NEED_CLEANUP;
3684 io_req_complete(req, ret);
3689 * IORING_OP_NOP just posts a completion event, nothing else.
3691 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3693 struct io_ring_ctx *ctx = req->ctx;
3695 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3698 __io_req_complete(req, issue_flags, 0, 0);
3702 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3704 struct io_ring_ctx *ctx = req->ctx;
3709 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3711 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3714 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3715 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3718 req->sync.off = READ_ONCE(sqe->off);
3719 req->sync.len = READ_ONCE(sqe->len);
3723 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3725 loff_t end = req->sync.off + req->sync.len;
3728 /* fsync always requires a blocking context */
3729 if (issue_flags & IO_URING_F_NONBLOCK)
3732 ret = vfs_fsync_range(req->file, req->sync.off,
3733 end > 0 ? end : LLONG_MAX,
3734 req->sync.flags & IORING_FSYNC_DATASYNC);
3737 io_req_complete(req, ret);
3741 static int io_fallocate_prep(struct io_kiocb *req,
3742 const struct io_uring_sqe *sqe)
3744 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3746 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3749 req->sync.off = READ_ONCE(sqe->off);
3750 req->sync.len = READ_ONCE(sqe->addr);
3751 req->sync.mode = READ_ONCE(sqe->len);
3755 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3759 /* fallocate always requiring blocking context */
3760 if (issue_flags & IO_URING_F_NONBLOCK)
3762 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3766 io_req_complete(req, ret);
3770 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3772 const char __user *fname;
3775 if (unlikely(sqe->ioprio || sqe->buf_index))
3777 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3780 /* open.how should be already initialised */
3781 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3782 req->open.how.flags |= O_LARGEFILE;
3784 req->open.dfd = READ_ONCE(sqe->fd);
3785 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3786 req->open.filename = getname(fname);
3787 if (IS_ERR(req->open.filename)) {
3788 ret = PTR_ERR(req->open.filename);
3789 req->open.filename = NULL;
3792 req->open.nofile = rlimit(RLIMIT_NOFILE);
3793 req->flags |= REQ_F_NEED_CLEANUP;
3797 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3801 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3803 mode = READ_ONCE(sqe->len);
3804 flags = READ_ONCE(sqe->open_flags);
3805 req->open.how = build_open_how(flags, mode);
3806 return __io_openat_prep(req, sqe);
3809 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3811 struct open_how __user *how;
3815 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3817 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3818 len = READ_ONCE(sqe->len);
3819 if (len < OPEN_HOW_SIZE_VER0)
3822 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3827 return __io_openat_prep(req, sqe);
3830 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3832 struct open_flags op;
3835 bool resolve_nonblock;
3838 ret = build_open_flags(&req->open.how, &op);
3841 nonblock_set = op.open_flag & O_NONBLOCK;
3842 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3843 if (issue_flags & IO_URING_F_NONBLOCK) {
3845 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3846 * it'll always -EAGAIN
3848 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3850 op.lookup_flags |= LOOKUP_CACHED;
3851 op.open_flag |= O_NONBLOCK;
3854 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3858 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3859 /* only retry if RESOLVE_CACHED wasn't already set by application */
3860 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3861 file == ERR_PTR(-EAGAIN)) {
3863 * We could hang on to this 'fd', but seems like marginal
3864 * gain for something that is now known to be a slower path.
3865 * So just put it, and we'll get a new one when we retry.
3873 ret = PTR_ERR(file);
3875 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3876 file->f_flags &= ~O_NONBLOCK;
3877 fsnotify_open(file);
3878 fd_install(ret, file);
3881 putname(req->open.filename);
3882 req->flags &= ~REQ_F_NEED_CLEANUP;
3885 __io_req_complete(req, issue_flags, ret, 0);
3889 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3891 return io_openat2(req, issue_flags);
3894 static int io_remove_buffers_prep(struct io_kiocb *req,
3895 const struct io_uring_sqe *sqe)
3897 struct io_provide_buf *p = &req->pbuf;
3900 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3903 tmp = READ_ONCE(sqe->fd);
3904 if (!tmp || tmp > USHRT_MAX)
3907 memset(p, 0, sizeof(*p));
3909 p->bgid = READ_ONCE(sqe->buf_group);
3913 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3914 int bgid, unsigned nbufs)
3918 /* shouldn't happen */
3922 /* the head kbuf is the list itself */
3923 while (!list_empty(&buf->list)) {
3924 struct io_buffer *nxt;
3926 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3927 list_del(&nxt->list);
3934 xa_erase(&ctx->io_buffers, bgid);
3939 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3941 struct io_provide_buf *p = &req->pbuf;
3942 struct io_ring_ctx *ctx = req->ctx;
3943 struct io_buffer *head;
3945 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3947 io_ring_submit_lock(ctx, !force_nonblock);
3949 lockdep_assert_held(&ctx->uring_lock);
3952 head = xa_load(&ctx->io_buffers, p->bgid);
3954 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3958 /* complete before unlock, IOPOLL may need the lock */
3959 __io_req_complete(req, issue_flags, ret, 0);
3960 io_ring_submit_unlock(ctx, !force_nonblock);
3964 static int io_provide_buffers_prep(struct io_kiocb *req,
3965 const struct io_uring_sqe *sqe)
3967 unsigned long size, tmp_check;
3968 struct io_provide_buf *p = &req->pbuf;
3971 if (sqe->ioprio || sqe->rw_flags)
3974 tmp = READ_ONCE(sqe->fd);
3975 if (!tmp || tmp > USHRT_MAX)
3978 p->addr = READ_ONCE(sqe->addr);
3979 p->len = READ_ONCE(sqe->len);
3981 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3984 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3987 size = (unsigned long)p->len * p->nbufs;
3988 if (!access_ok(u64_to_user_ptr(p->addr), size))
3991 p->bgid = READ_ONCE(sqe->buf_group);
3992 tmp = READ_ONCE(sqe->off);
3993 if (tmp > USHRT_MAX)
3999 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4001 struct io_buffer *buf;
4002 u64 addr = pbuf->addr;
4003 int i, bid = pbuf->bid;
4005 for (i = 0; i < pbuf->nbufs; i++) {
4006 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4011 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4016 INIT_LIST_HEAD(&buf->list);
4019 list_add_tail(&buf->list, &(*head)->list);
4023 return i ? i : -ENOMEM;
4026 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4028 struct io_provide_buf *p = &req->pbuf;
4029 struct io_ring_ctx *ctx = req->ctx;
4030 struct io_buffer *head, *list;
4032 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4034 io_ring_submit_lock(ctx, !force_nonblock);
4036 lockdep_assert_held(&ctx->uring_lock);
4038 list = head = xa_load(&ctx->io_buffers, p->bgid);
4040 ret = io_add_buffers(p, &head);
4041 if (ret >= 0 && !list) {
4042 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4044 __io_remove_buffers(ctx, head, p->bgid, -1U);
4048 /* complete before unlock, IOPOLL may need the lock */
4049 __io_req_complete(req, issue_flags, ret, 0);
4050 io_ring_submit_unlock(ctx, !force_nonblock);
4054 static int io_epoll_ctl_prep(struct io_kiocb *req,
4055 const struct io_uring_sqe *sqe)
4057 #if defined(CONFIG_EPOLL)
4058 if (sqe->ioprio || sqe->buf_index)
4060 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4063 req->epoll.epfd = READ_ONCE(sqe->fd);
4064 req->epoll.op = READ_ONCE(sqe->len);
4065 req->epoll.fd = READ_ONCE(sqe->off);
4067 if (ep_op_has_event(req->epoll.op)) {
4068 struct epoll_event __user *ev;
4070 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4071 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4081 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4083 #if defined(CONFIG_EPOLL)
4084 struct io_epoll *ie = &req->epoll;
4086 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4088 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4089 if (force_nonblock && ret == -EAGAIN)
4094 __io_req_complete(req, issue_flags, ret, 0);
4101 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4103 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4104 if (sqe->ioprio || sqe->buf_index || sqe->off)
4106 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4109 req->madvise.addr = READ_ONCE(sqe->addr);
4110 req->madvise.len = READ_ONCE(sqe->len);
4111 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4118 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4120 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4121 struct io_madvise *ma = &req->madvise;
4124 if (issue_flags & IO_URING_F_NONBLOCK)
4127 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4130 io_req_complete(req, ret);
4137 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4139 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4141 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4144 req->fadvise.offset = READ_ONCE(sqe->off);
4145 req->fadvise.len = READ_ONCE(sqe->len);
4146 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4150 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4152 struct io_fadvise *fa = &req->fadvise;
4155 if (issue_flags & IO_URING_F_NONBLOCK) {
4156 switch (fa->advice) {
4157 case POSIX_FADV_NORMAL:
4158 case POSIX_FADV_RANDOM:
4159 case POSIX_FADV_SEQUENTIAL:
4166 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4169 __io_req_complete(req, issue_flags, ret, 0);
4173 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4175 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4177 if (sqe->ioprio || sqe->buf_index)
4179 if (req->flags & REQ_F_FIXED_FILE)
4182 req->statx.dfd = READ_ONCE(sqe->fd);
4183 req->statx.mask = READ_ONCE(sqe->len);
4184 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4185 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4186 req->statx.flags = READ_ONCE(sqe->statx_flags);
4191 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4193 struct io_statx *ctx = &req->statx;
4196 if (issue_flags & IO_URING_F_NONBLOCK)
4199 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4204 io_req_complete(req, ret);
4208 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4210 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4212 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4213 sqe->rw_flags || sqe->buf_index)
4215 if (req->flags & REQ_F_FIXED_FILE)
4218 req->close.fd = READ_ONCE(sqe->fd);
4222 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4224 struct files_struct *files = current->files;
4225 struct io_close *close = &req->close;
4226 struct fdtable *fdt;
4227 struct file *file = NULL;
4230 spin_lock(&files->file_lock);
4231 fdt = files_fdtable(files);
4232 if (close->fd >= fdt->max_fds) {
4233 spin_unlock(&files->file_lock);
4236 file = fdt->fd[close->fd];
4237 if (!file || file->f_op == &io_uring_fops) {
4238 spin_unlock(&files->file_lock);
4243 /* if the file has a flush method, be safe and punt to async */
4244 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4245 spin_unlock(&files->file_lock);
4249 ret = __close_fd_get_file(close->fd, &file);
4250 spin_unlock(&files->file_lock);
4257 /* No ->flush() or already async, safely close from here */
4258 ret = filp_close(file, current->files);
4264 __io_req_complete(req, issue_flags, ret, 0);
4268 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4270 struct io_ring_ctx *ctx = req->ctx;
4272 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4274 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4277 req->sync.off = READ_ONCE(sqe->off);
4278 req->sync.len = READ_ONCE(sqe->len);
4279 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4283 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4287 /* sync_file_range always requires a blocking context */
4288 if (issue_flags & IO_URING_F_NONBLOCK)
4291 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4295 io_req_complete(req, ret);
4299 #if defined(CONFIG_NET)
4300 static int io_setup_async_msg(struct io_kiocb *req,
4301 struct io_async_msghdr *kmsg)
4303 struct io_async_msghdr *async_msg = req->async_data;
4307 if (io_alloc_async_data(req)) {
4308 kfree(kmsg->free_iov);
4311 async_msg = req->async_data;
4312 req->flags |= REQ_F_NEED_CLEANUP;
4313 memcpy(async_msg, kmsg, sizeof(*kmsg));
4314 async_msg->msg.msg_name = &async_msg->addr;
4315 /* if were using fast_iov, set it to the new one */
4316 if (!async_msg->free_iov)
4317 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4322 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4323 struct io_async_msghdr *iomsg)
4325 iomsg->msg.msg_name = &iomsg->addr;
4326 iomsg->free_iov = iomsg->fast_iov;
4327 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4328 req->sr_msg.msg_flags, &iomsg->free_iov);
4331 static int io_sendmsg_prep_async(struct io_kiocb *req)
4335 ret = io_sendmsg_copy_hdr(req, req->async_data);
4337 req->flags |= REQ_F_NEED_CLEANUP;
4341 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4343 struct io_sr_msg *sr = &req->sr_msg;
4345 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4348 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4349 sr->len = READ_ONCE(sqe->len);
4350 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4351 if (sr->msg_flags & MSG_DONTWAIT)
4352 req->flags |= REQ_F_NOWAIT;
4354 #ifdef CONFIG_COMPAT
4355 if (req->ctx->compat)
4356 sr->msg_flags |= MSG_CMSG_COMPAT;
4361 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4363 struct io_async_msghdr iomsg, *kmsg;
4364 struct socket *sock;
4369 sock = sock_from_file(req->file);
4370 if (unlikely(!sock))
4373 kmsg = req->async_data;
4375 ret = io_sendmsg_copy_hdr(req, &iomsg);
4381 flags = req->sr_msg.msg_flags;
4382 if (issue_flags & IO_URING_F_NONBLOCK)
4383 flags |= MSG_DONTWAIT;
4384 if (flags & MSG_WAITALL)
4385 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4387 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4388 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4389 return io_setup_async_msg(req, kmsg);
4390 if (ret == -ERESTARTSYS)
4393 /* fast path, check for non-NULL to avoid function call */
4395 kfree(kmsg->free_iov);
4396 req->flags &= ~REQ_F_NEED_CLEANUP;
4399 __io_req_complete(req, issue_flags, ret, 0);
4403 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4405 struct io_sr_msg *sr = &req->sr_msg;
4408 struct socket *sock;
4413 sock = sock_from_file(req->file);
4414 if (unlikely(!sock))
4417 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4421 msg.msg_name = NULL;
4422 msg.msg_control = NULL;
4423 msg.msg_controllen = 0;
4424 msg.msg_namelen = 0;
4426 flags = req->sr_msg.msg_flags;
4427 if (issue_flags & IO_URING_F_NONBLOCK)
4428 flags |= MSG_DONTWAIT;
4429 if (flags & MSG_WAITALL)
4430 min_ret = iov_iter_count(&msg.msg_iter);
4432 msg.msg_flags = flags;
4433 ret = sock_sendmsg(sock, &msg);
4434 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4436 if (ret == -ERESTARTSYS)
4441 __io_req_complete(req, issue_flags, ret, 0);
4445 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4446 struct io_async_msghdr *iomsg)
4448 struct io_sr_msg *sr = &req->sr_msg;
4449 struct iovec __user *uiov;
4453 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4454 &iomsg->uaddr, &uiov, &iov_len);
4458 if (req->flags & REQ_F_BUFFER_SELECT) {
4461 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4463 sr->len = iomsg->fast_iov[0].iov_len;
4464 iomsg->free_iov = NULL;
4466 iomsg->free_iov = iomsg->fast_iov;
4467 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4468 &iomsg->free_iov, &iomsg->msg.msg_iter,
4477 #ifdef CONFIG_COMPAT
4478 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4479 struct io_async_msghdr *iomsg)
4481 struct io_sr_msg *sr = &req->sr_msg;
4482 struct compat_iovec __user *uiov;
4487 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4492 uiov = compat_ptr(ptr);
4493 if (req->flags & REQ_F_BUFFER_SELECT) {
4494 compat_ssize_t clen;
4498 if (!access_ok(uiov, sizeof(*uiov)))
4500 if (__get_user(clen, &uiov->iov_len))
4505 iomsg->free_iov = NULL;
4507 iomsg->free_iov = iomsg->fast_iov;
4508 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4509 UIO_FASTIOV, &iomsg->free_iov,
4510 &iomsg->msg.msg_iter, true);
4519 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4520 struct io_async_msghdr *iomsg)
4522 iomsg->msg.msg_name = &iomsg->addr;
4524 #ifdef CONFIG_COMPAT
4525 if (req->ctx->compat)
4526 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4529 return __io_recvmsg_copy_hdr(req, iomsg);
4532 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4535 struct io_sr_msg *sr = &req->sr_msg;
4536 struct io_buffer *kbuf;
4538 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4543 req->flags |= REQ_F_BUFFER_SELECTED;
4547 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4549 return io_put_kbuf(req, req->sr_msg.kbuf);
4552 static int io_recvmsg_prep_async(struct io_kiocb *req)
4556 ret = io_recvmsg_copy_hdr(req, req->async_data);
4558 req->flags |= REQ_F_NEED_CLEANUP;
4562 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4564 struct io_sr_msg *sr = &req->sr_msg;
4566 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4569 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4570 sr->len = READ_ONCE(sqe->len);
4571 sr->bgid = READ_ONCE(sqe->buf_group);
4572 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4573 if (sr->msg_flags & MSG_DONTWAIT)
4574 req->flags |= REQ_F_NOWAIT;
4576 #ifdef CONFIG_COMPAT
4577 if (req->ctx->compat)
4578 sr->msg_flags |= MSG_CMSG_COMPAT;
4583 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4585 struct io_async_msghdr iomsg, *kmsg;
4586 struct socket *sock;
4587 struct io_buffer *kbuf;
4590 int ret, cflags = 0;
4591 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4593 sock = sock_from_file(req->file);
4594 if (unlikely(!sock))
4597 kmsg = req->async_data;
4599 ret = io_recvmsg_copy_hdr(req, &iomsg);
4605 if (req->flags & REQ_F_BUFFER_SELECT) {
4606 kbuf = io_recv_buffer_select(req, !force_nonblock);
4608 return PTR_ERR(kbuf);
4609 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4610 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4611 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4612 1, req->sr_msg.len);
4615 flags = req->sr_msg.msg_flags;
4617 flags |= MSG_DONTWAIT;
4618 if (flags & MSG_WAITALL)
4619 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4621 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4622 kmsg->uaddr, flags);
4623 if (force_nonblock && ret == -EAGAIN)
4624 return io_setup_async_msg(req, kmsg);
4625 if (ret == -ERESTARTSYS)
4628 if (req->flags & REQ_F_BUFFER_SELECTED)
4629 cflags = io_put_recv_kbuf(req);
4630 /* fast path, check for non-NULL to avoid function call */
4632 kfree(kmsg->free_iov);
4633 req->flags &= ~REQ_F_NEED_CLEANUP;
4634 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4636 __io_req_complete(req, issue_flags, ret, cflags);
4640 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4642 struct io_buffer *kbuf;
4643 struct io_sr_msg *sr = &req->sr_msg;
4645 void __user *buf = sr->buf;
4646 struct socket *sock;
4650 int ret, cflags = 0;
4651 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4653 sock = sock_from_file(req->file);
4654 if (unlikely(!sock))
4657 if (req->flags & REQ_F_BUFFER_SELECT) {
4658 kbuf = io_recv_buffer_select(req, !force_nonblock);
4660 return PTR_ERR(kbuf);
4661 buf = u64_to_user_ptr(kbuf->addr);
4664 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4668 msg.msg_name = NULL;
4669 msg.msg_control = NULL;
4670 msg.msg_controllen = 0;
4671 msg.msg_namelen = 0;
4672 msg.msg_iocb = NULL;
4675 flags = req->sr_msg.msg_flags;
4677 flags |= MSG_DONTWAIT;
4678 if (flags & MSG_WAITALL)
4679 min_ret = iov_iter_count(&msg.msg_iter);
4681 ret = sock_recvmsg(sock, &msg, flags);
4682 if (force_nonblock && ret == -EAGAIN)
4684 if (ret == -ERESTARTSYS)
4687 if (req->flags & REQ_F_BUFFER_SELECTED)
4688 cflags = io_put_recv_kbuf(req);
4689 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4691 __io_req_complete(req, issue_flags, ret, cflags);
4695 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4697 struct io_accept *accept = &req->accept;
4699 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4701 if (sqe->ioprio || sqe->len || sqe->buf_index)
4704 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4705 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4706 accept->flags = READ_ONCE(sqe->accept_flags);
4707 accept->nofile = rlimit(RLIMIT_NOFILE);
4711 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4713 struct io_accept *accept = &req->accept;
4714 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4715 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4718 if (req->file->f_flags & O_NONBLOCK)
4719 req->flags |= REQ_F_NOWAIT;
4721 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4722 accept->addr_len, accept->flags,
4724 if (ret == -EAGAIN && force_nonblock)
4727 if (ret == -ERESTARTSYS)
4731 __io_req_complete(req, issue_flags, ret, 0);
4735 static int io_connect_prep_async(struct io_kiocb *req)
4737 struct io_async_connect *io = req->async_data;
4738 struct io_connect *conn = &req->connect;
4740 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4743 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4745 struct io_connect *conn = &req->connect;
4747 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4749 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4752 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4753 conn->addr_len = READ_ONCE(sqe->addr2);
4757 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4759 struct io_async_connect __io, *io;
4760 unsigned file_flags;
4762 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4764 if (req->async_data) {
4765 io = req->async_data;
4767 ret = move_addr_to_kernel(req->connect.addr,
4768 req->connect.addr_len,
4775 file_flags = force_nonblock ? O_NONBLOCK : 0;
4777 ret = __sys_connect_file(req->file, &io->address,
4778 req->connect.addr_len, file_flags);
4779 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4780 if (req->async_data)
4782 if (io_alloc_async_data(req)) {
4786 memcpy(req->async_data, &__io, sizeof(__io));
4789 if (ret == -ERESTARTSYS)
4794 __io_req_complete(req, issue_flags, ret, 0);
4797 #else /* !CONFIG_NET */
4798 #define IO_NETOP_FN(op) \
4799 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4801 return -EOPNOTSUPP; \
4804 #define IO_NETOP_PREP(op) \
4806 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4808 return -EOPNOTSUPP; \
4811 #define IO_NETOP_PREP_ASYNC(op) \
4813 static int io_##op##_prep_async(struct io_kiocb *req) \
4815 return -EOPNOTSUPP; \
4818 IO_NETOP_PREP_ASYNC(sendmsg);
4819 IO_NETOP_PREP_ASYNC(recvmsg);
4820 IO_NETOP_PREP_ASYNC(connect);
4821 IO_NETOP_PREP(accept);
4824 #endif /* CONFIG_NET */
4826 struct io_poll_table {
4827 struct poll_table_struct pt;
4828 struct io_kiocb *req;
4832 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4833 __poll_t mask, task_work_func_t func)
4837 /* for instances that support it check for an event match first: */
4838 if (mask && !(mask & poll->events))
4841 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4843 list_del_init(&poll->wait.entry);
4846 req->task_work.func = func;
4849 * If this fails, then the task is exiting. When a task exits, the
4850 * work gets canceled, so just cancel this request as well instead
4851 * of executing it. We can't safely execute it anyway, as we may not
4852 * have the needed state needed for it anyway.
4854 ret = io_req_task_work_add(req);
4855 if (unlikely(ret)) {
4856 WRITE_ONCE(poll->canceled, true);
4857 io_req_task_work_add_fallback(req, func);
4862 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4863 __acquires(&req->ctx->completion_lock)
4865 struct io_ring_ctx *ctx = req->ctx;
4867 if (!req->result && !READ_ONCE(poll->canceled)) {
4868 struct poll_table_struct pt = { ._key = poll->events };
4870 req->result = vfs_poll(req->file, &pt) & poll->events;
4873 spin_lock_irq(&ctx->completion_lock);
4874 if (!req->result && !READ_ONCE(poll->canceled)) {
4875 add_wait_queue(poll->head, &poll->wait);
4882 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4884 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4885 if (req->opcode == IORING_OP_POLL_ADD)
4886 return req->async_data;
4887 return req->apoll->double_poll;
4890 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4892 if (req->opcode == IORING_OP_POLL_ADD)
4894 return &req->apoll->poll;
4897 static void io_poll_remove_double(struct io_kiocb *req)
4898 __must_hold(&req->ctx->completion_lock)
4900 struct io_poll_iocb *poll = io_poll_get_double(req);
4902 lockdep_assert_held(&req->ctx->completion_lock);
4904 if (poll && poll->head) {
4905 struct wait_queue_head *head = poll->head;
4907 spin_lock(&head->lock);
4908 list_del_init(&poll->wait.entry);
4909 if (poll->wait.private)
4912 spin_unlock(&head->lock);
4916 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4917 __must_hold(&req->ctx->completion_lock)
4919 struct io_ring_ctx *ctx = req->ctx;
4920 unsigned flags = IORING_CQE_F_MORE;
4923 if (READ_ONCE(req->poll.canceled)) {
4925 req->poll.events |= EPOLLONESHOT;
4927 error = mangle_poll(mask);
4929 if (req->poll.events & EPOLLONESHOT)
4931 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4932 io_poll_remove_waitqs(req);
4933 req->poll.done = true;
4936 if (flags & IORING_CQE_F_MORE)
4939 io_commit_cqring(ctx);
4940 return !(flags & IORING_CQE_F_MORE);
4943 static void io_poll_task_func(struct callback_head *cb)
4945 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4946 struct io_ring_ctx *ctx = req->ctx;
4947 struct io_kiocb *nxt;
4949 if (io_poll_rewait(req, &req->poll)) {
4950 spin_unlock_irq(&ctx->completion_lock);
4954 done = io_poll_complete(req, req->result);
4956 hash_del(&req->hash_node);
4959 add_wait_queue(req->poll.head, &req->poll.wait);
4961 spin_unlock_irq(&ctx->completion_lock);
4962 io_cqring_ev_posted(ctx);
4965 nxt = io_put_req_find_next(req);
4967 __io_req_task_submit(nxt);
4972 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4973 int sync, void *key)
4975 struct io_kiocb *req = wait->private;
4976 struct io_poll_iocb *poll = io_poll_get_single(req);
4977 __poll_t mask = key_to_poll(key);
4979 /* for instances that support it check for an event match first: */
4980 if (mask && !(mask & poll->events))
4982 if (!(poll->events & EPOLLONESHOT))
4983 return poll->wait.func(&poll->wait, mode, sync, key);
4985 list_del_init(&wait->entry);
4987 if (poll && poll->head) {
4990 spin_lock(&poll->head->lock);
4991 done = list_empty(&poll->wait.entry);
4993 list_del_init(&poll->wait.entry);
4994 /* make sure double remove sees this as being gone */
4995 wait->private = NULL;
4996 spin_unlock(&poll->head->lock);
4998 /* use wait func handler, so it matches the rq type */
4999 poll->wait.func(&poll->wait, mode, sync, key);
5006 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5007 wait_queue_func_t wake_func)
5011 poll->canceled = false;
5012 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5013 /* mask in events that we always want/need */
5014 poll->events = events | IO_POLL_UNMASK;
5015 INIT_LIST_HEAD(&poll->wait.entry);
5016 init_waitqueue_func_entry(&poll->wait, wake_func);
5019 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5020 struct wait_queue_head *head,
5021 struct io_poll_iocb **poll_ptr)
5023 struct io_kiocb *req = pt->req;
5026 * If poll->head is already set, it's because the file being polled
5027 * uses multiple waitqueues for poll handling (eg one for read, one
5028 * for write). Setup a separate io_poll_iocb if this happens.
5030 if (unlikely(poll->head)) {
5031 struct io_poll_iocb *poll_one = poll;
5033 /* already have a 2nd entry, fail a third attempt */
5035 pt->error = -EINVAL;
5039 * Can't handle multishot for double wait for now, turn it
5040 * into one-shot mode.
5042 if (!(poll_one->events & EPOLLONESHOT))
5043 poll_one->events |= EPOLLONESHOT;
5044 /* double add on the same waitqueue head, ignore */
5045 if (poll_one->head == head)
5047 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5049 pt->error = -ENOMEM;
5052 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5054 poll->wait.private = req;
5061 if (poll->events & EPOLLEXCLUSIVE)
5062 add_wait_queue_exclusive(head, &poll->wait);
5064 add_wait_queue(head, &poll->wait);
5067 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5068 struct poll_table_struct *p)
5070 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5071 struct async_poll *apoll = pt->req->apoll;
5073 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5076 static void io_async_task_func(struct callback_head *cb)
5078 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5079 struct async_poll *apoll = req->apoll;
5080 struct io_ring_ctx *ctx = req->ctx;
5082 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5084 if (io_poll_rewait(req, &apoll->poll)) {
5085 spin_unlock_irq(&ctx->completion_lock);
5089 hash_del(&req->hash_node);
5090 io_poll_remove_double(req);
5091 spin_unlock_irq(&ctx->completion_lock);
5093 if (!READ_ONCE(apoll->poll.canceled))
5094 __io_req_task_submit(req);
5096 io_req_complete_failed(req, -ECANCELED);
5099 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5102 struct io_kiocb *req = wait->private;
5103 struct io_poll_iocb *poll = &req->apoll->poll;
5105 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5108 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5111 static void io_poll_req_insert(struct io_kiocb *req)
5113 struct io_ring_ctx *ctx = req->ctx;
5114 struct hlist_head *list;
5116 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5117 hlist_add_head(&req->hash_node, list);
5120 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5121 struct io_poll_iocb *poll,
5122 struct io_poll_table *ipt, __poll_t mask,
5123 wait_queue_func_t wake_func)
5124 __acquires(&ctx->completion_lock)
5126 struct io_ring_ctx *ctx = req->ctx;
5127 bool cancel = false;
5129 INIT_HLIST_NODE(&req->hash_node);
5130 io_init_poll_iocb(poll, mask, wake_func);
5131 poll->file = req->file;
5132 poll->wait.private = req;
5134 ipt->pt._key = mask;
5136 ipt->error = -EINVAL;
5138 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5140 spin_lock_irq(&ctx->completion_lock);
5141 if (likely(poll->head)) {
5142 spin_lock(&poll->head->lock);
5143 if (unlikely(list_empty(&poll->wait.entry))) {
5149 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5150 list_del_init(&poll->wait.entry);
5152 WRITE_ONCE(poll->canceled, true);
5153 else if (!poll->done) /* actually waiting for an event */
5154 io_poll_req_insert(req);
5155 spin_unlock(&poll->head->lock);
5161 static bool io_arm_poll_handler(struct io_kiocb *req)
5163 const struct io_op_def *def = &io_op_defs[req->opcode];
5164 struct io_ring_ctx *ctx = req->ctx;
5165 struct async_poll *apoll;
5166 struct io_poll_table ipt;
5170 if (!req->file || !file_can_poll(req->file))
5172 if (req->flags & REQ_F_POLLED)
5176 else if (def->pollout)
5180 /* if we can't nonblock try, then no point in arming a poll handler */
5181 if (!io_file_supports_async(req, rw))
5184 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5185 if (unlikely(!apoll))
5187 apoll->double_poll = NULL;
5189 req->flags |= REQ_F_POLLED;
5192 mask = EPOLLONESHOT;
5194 mask |= POLLIN | POLLRDNORM;
5196 mask |= POLLOUT | POLLWRNORM;
5198 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5199 if ((req->opcode == IORING_OP_RECVMSG) &&
5200 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5203 mask |= POLLERR | POLLPRI;
5205 ipt.pt._qproc = io_async_queue_proc;
5207 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5209 if (ret || ipt.error) {
5210 io_poll_remove_double(req);
5211 spin_unlock_irq(&ctx->completion_lock);
5214 spin_unlock_irq(&ctx->completion_lock);
5215 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5216 apoll->poll.events);
5220 static bool __io_poll_remove_one(struct io_kiocb *req,
5221 struct io_poll_iocb *poll, bool do_cancel)
5222 __must_hold(&req->ctx->completion_lock)
5224 bool do_complete = false;
5228 spin_lock(&poll->head->lock);
5230 WRITE_ONCE(poll->canceled, true);
5231 if (!list_empty(&poll->wait.entry)) {
5232 list_del_init(&poll->wait.entry);
5235 spin_unlock(&poll->head->lock);
5236 hash_del(&req->hash_node);
5240 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5241 __must_hold(&req->ctx->completion_lock)
5245 io_poll_remove_double(req);
5246 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5248 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5249 /* non-poll requests have submit ref still */
5255 static bool io_poll_remove_one(struct io_kiocb *req)
5256 __must_hold(&req->ctx->completion_lock)
5260 do_complete = io_poll_remove_waitqs(req);
5262 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5263 io_commit_cqring(req->ctx);
5265 io_put_req_deferred(req, 1);
5272 * Returns true if we found and killed one or more poll requests
5274 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5277 struct hlist_node *tmp;
5278 struct io_kiocb *req;
5281 spin_lock_irq(&ctx->completion_lock);
5282 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5283 struct hlist_head *list;
5285 list = &ctx->cancel_hash[i];
5286 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5287 if (io_match_task(req, tsk, cancel_all))
5288 posted += io_poll_remove_one(req);
5291 spin_unlock_irq(&ctx->completion_lock);
5294 io_cqring_ev_posted(ctx);
5299 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5301 __must_hold(&ctx->completion_lock)
5303 struct hlist_head *list;
5304 struct io_kiocb *req;
5306 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5307 hlist_for_each_entry(req, list, hash_node) {
5308 if (sqe_addr != req->user_data)
5310 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5317 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5319 __must_hold(&ctx->completion_lock)
5321 struct io_kiocb *req;
5323 req = io_poll_find(ctx, sqe_addr, poll_only);
5326 if (io_poll_remove_one(req))
5332 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5337 events = READ_ONCE(sqe->poll32_events);
5339 events = swahw32(events);
5341 if (!(flags & IORING_POLL_ADD_MULTI))
5342 events |= EPOLLONESHOT;
5343 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5346 static int io_poll_update_prep(struct io_kiocb *req,
5347 const struct io_uring_sqe *sqe)
5349 struct io_poll_update *upd = &req->poll_update;
5352 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5354 if (sqe->ioprio || sqe->buf_index)
5356 flags = READ_ONCE(sqe->len);
5357 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5358 IORING_POLL_ADD_MULTI))
5360 /* meaningless without update */
5361 if (flags == IORING_POLL_ADD_MULTI)
5364 upd->old_user_data = READ_ONCE(sqe->addr);
5365 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5366 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5368 upd->new_user_data = READ_ONCE(sqe->off);
5369 if (!upd->update_user_data && upd->new_user_data)
5371 if (upd->update_events)
5372 upd->events = io_poll_parse_events(sqe, flags);
5373 else if (sqe->poll32_events)
5379 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5382 struct io_kiocb *req = wait->private;
5383 struct io_poll_iocb *poll = &req->poll;
5385 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5388 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5389 struct poll_table_struct *p)
5391 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5393 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5396 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5398 struct io_poll_iocb *poll = &req->poll;
5401 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5403 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5405 flags = READ_ONCE(sqe->len);
5406 if (flags & ~IORING_POLL_ADD_MULTI)
5409 poll->events = io_poll_parse_events(sqe, flags);
5413 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5415 struct io_poll_iocb *poll = &req->poll;
5416 struct io_ring_ctx *ctx = req->ctx;
5417 struct io_poll_table ipt;
5420 ipt.pt._qproc = io_poll_queue_proc;
5422 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5425 if (mask) { /* no async, we'd stolen it */
5427 io_poll_complete(req, mask);
5429 spin_unlock_irq(&ctx->completion_lock);
5432 io_cqring_ev_posted(ctx);
5433 if (poll->events & EPOLLONESHOT)
5439 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5441 struct io_ring_ctx *ctx = req->ctx;
5442 struct io_kiocb *preq;
5446 spin_lock_irq(&ctx->completion_lock);
5447 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5453 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5455 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5460 * Don't allow racy completion with singleshot, as we cannot safely
5461 * update those. For multishot, if we're racing with completion, just
5462 * let completion re-add it.
5464 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5465 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5469 /* we now have a detached poll request. reissue. */
5473 spin_unlock_irq(&ctx->completion_lock);
5475 io_req_complete(req, ret);
5478 /* only mask one event flags, keep behavior flags */
5479 if (req->poll_update.update_events) {
5480 preq->poll.events &= ~0xffff;
5481 preq->poll.events |= req->poll_update.events & 0xffff;
5482 preq->poll.events |= IO_POLL_UNMASK;
5484 if (req->poll_update.update_user_data)
5485 preq->user_data = req->poll_update.new_user_data;
5486 spin_unlock_irq(&ctx->completion_lock);
5488 /* complete update request, we're done with it */
5489 io_req_complete(req, ret);
5492 ret = io_poll_add(preq, issue_flags);
5495 io_req_complete(preq, ret);
5501 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5503 struct io_timeout_data *data = container_of(timer,
5504 struct io_timeout_data, timer);
5505 struct io_kiocb *req = data->req;
5506 struct io_ring_ctx *ctx = req->ctx;
5507 unsigned long flags;
5509 spin_lock_irqsave(&ctx->completion_lock, flags);
5510 list_del_init(&req->timeout.list);
5511 atomic_set(&req->ctx->cq_timeouts,
5512 atomic_read(&req->ctx->cq_timeouts) + 1);
5514 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5515 io_commit_cqring(ctx);
5516 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5518 io_cqring_ev_posted(ctx);
5521 return HRTIMER_NORESTART;
5524 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5526 __must_hold(&ctx->completion_lock)
5528 struct io_timeout_data *io;
5529 struct io_kiocb *req;
5532 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5533 found = user_data == req->user_data;
5538 return ERR_PTR(-ENOENT);
5540 io = req->async_data;
5541 if (hrtimer_try_to_cancel(&io->timer) == -1)
5542 return ERR_PTR(-EALREADY);
5543 list_del_init(&req->timeout.list);
5547 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5548 __must_hold(&ctx->completion_lock)
5550 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5553 return PTR_ERR(req);
5556 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5557 io_put_req_deferred(req, 1);
5561 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5562 struct timespec64 *ts, enum hrtimer_mode mode)
5563 __must_hold(&ctx->completion_lock)
5565 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5566 struct io_timeout_data *data;
5569 return PTR_ERR(req);
5571 req->timeout.off = 0; /* noseq */
5572 data = req->async_data;
5573 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5574 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5575 data->timer.function = io_timeout_fn;
5576 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5580 static int io_timeout_remove_prep(struct io_kiocb *req,
5581 const struct io_uring_sqe *sqe)
5583 struct io_timeout_rem *tr = &req->timeout_rem;
5585 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5587 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5589 if (sqe->ioprio || sqe->buf_index || sqe->len)
5592 tr->addr = READ_ONCE(sqe->addr);
5593 tr->flags = READ_ONCE(sqe->timeout_flags);
5594 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5595 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5597 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5599 } else if (tr->flags) {
5600 /* timeout removal doesn't support flags */
5607 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5609 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5614 * Remove or update an existing timeout command
5616 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5618 struct io_timeout_rem *tr = &req->timeout_rem;
5619 struct io_ring_ctx *ctx = req->ctx;
5622 spin_lock_irq(&ctx->completion_lock);
5623 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5624 ret = io_timeout_cancel(ctx, tr->addr);
5626 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5627 io_translate_timeout_mode(tr->flags));
5629 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5630 io_commit_cqring(ctx);
5631 spin_unlock_irq(&ctx->completion_lock);
5632 io_cqring_ev_posted(ctx);
5639 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5640 bool is_timeout_link)
5642 struct io_timeout_data *data;
5644 u32 off = READ_ONCE(sqe->off);
5646 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5648 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5650 if (off && is_timeout_link)
5652 flags = READ_ONCE(sqe->timeout_flags);
5653 if (flags & ~IORING_TIMEOUT_ABS)
5656 req->timeout.off = off;
5658 if (!req->async_data && io_alloc_async_data(req))
5661 data = req->async_data;
5664 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5667 data->mode = io_translate_timeout_mode(flags);
5668 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5669 if (is_timeout_link)
5670 io_req_track_inflight(req);
5674 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5676 struct io_ring_ctx *ctx = req->ctx;
5677 struct io_timeout_data *data = req->async_data;
5678 struct list_head *entry;
5679 u32 tail, off = req->timeout.off;
5681 spin_lock_irq(&ctx->completion_lock);
5684 * sqe->off holds how many events that need to occur for this
5685 * timeout event to be satisfied. If it isn't set, then this is
5686 * a pure timeout request, sequence isn't used.
5688 if (io_is_timeout_noseq(req)) {
5689 entry = ctx->timeout_list.prev;
5693 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5694 req->timeout.target_seq = tail + off;
5696 /* Update the last seq here in case io_flush_timeouts() hasn't.
5697 * This is safe because ->completion_lock is held, and submissions
5698 * and completions are never mixed in the same ->completion_lock section.
5700 ctx->cq_last_tm_flush = tail;
5703 * Insertion sort, ensuring the first entry in the list is always
5704 * the one we need first.
5706 list_for_each_prev(entry, &ctx->timeout_list) {
5707 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5710 if (io_is_timeout_noseq(nxt))
5712 /* nxt.seq is behind @tail, otherwise would've been completed */
5713 if (off >= nxt->timeout.target_seq - tail)
5717 list_add(&req->timeout.list, entry);
5718 data->timer.function = io_timeout_fn;
5719 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5720 spin_unlock_irq(&ctx->completion_lock);
5724 struct io_cancel_data {
5725 struct io_ring_ctx *ctx;
5729 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5731 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5732 struct io_cancel_data *cd = data;
5734 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5737 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5738 struct io_ring_ctx *ctx)
5740 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5741 enum io_wq_cancel cancel_ret;
5744 if (!tctx || !tctx->io_wq)
5747 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5748 switch (cancel_ret) {
5749 case IO_WQ_CANCEL_OK:
5752 case IO_WQ_CANCEL_RUNNING:
5755 case IO_WQ_CANCEL_NOTFOUND:
5763 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5764 struct io_kiocb *req, __u64 sqe_addr,
5767 unsigned long flags;
5770 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5771 spin_lock_irqsave(&ctx->completion_lock, flags);
5774 ret = io_timeout_cancel(ctx, sqe_addr);
5777 ret = io_poll_cancel(ctx, sqe_addr, false);
5781 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5782 io_commit_cqring(ctx);
5783 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5784 io_cqring_ev_posted(ctx);
5790 static int io_async_cancel_prep(struct io_kiocb *req,
5791 const struct io_uring_sqe *sqe)
5793 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5795 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5797 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5800 req->cancel.addr = READ_ONCE(sqe->addr);
5804 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5806 struct io_ring_ctx *ctx = req->ctx;
5807 u64 sqe_addr = req->cancel.addr;
5808 struct io_tctx_node *node;
5811 /* tasks should wait for their io-wq threads, so safe w/o sync */
5812 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5813 spin_lock_irq(&ctx->completion_lock);
5816 ret = io_timeout_cancel(ctx, sqe_addr);
5819 ret = io_poll_cancel(ctx, sqe_addr, false);
5822 spin_unlock_irq(&ctx->completion_lock);
5824 /* slow path, try all io-wq's */
5825 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5827 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5828 struct io_uring_task *tctx = node->task->io_uring;
5830 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5834 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5836 spin_lock_irq(&ctx->completion_lock);
5838 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5839 io_commit_cqring(ctx);
5840 spin_unlock_irq(&ctx->completion_lock);
5841 io_cqring_ev_posted(ctx);
5849 static int io_rsrc_update_prep(struct io_kiocb *req,
5850 const struct io_uring_sqe *sqe)
5852 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5854 if (sqe->ioprio || sqe->rw_flags)
5857 req->rsrc_update.offset = READ_ONCE(sqe->off);
5858 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5859 if (!req->rsrc_update.nr_args)
5861 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5865 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5867 struct io_ring_ctx *ctx = req->ctx;
5868 struct io_uring_rsrc_update2 up;
5871 if (issue_flags & IO_URING_F_NONBLOCK)
5874 up.offset = req->rsrc_update.offset;
5875 up.data = req->rsrc_update.arg;
5880 mutex_lock(&ctx->uring_lock);
5881 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5882 &up, req->rsrc_update.nr_args);
5883 mutex_unlock(&ctx->uring_lock);
5887 __io_req_complete(req, issue_flags, ret, 0);
5891 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5893 switch (req->opcode) {
5896 case IORING_OP_READV:
5897 case IORING_OP_READ_FIXED:
5898 case IORING_OP_READ:
5899 return io_read_prep(req, sqe);
5900 case IORING_OP_WRITEV:
5901 case IORING_OP_WRITE_FIXED:
5902 case IORING_OP_WRITE:
5903 return io_write_prep(req, sqe);
5904 case IORING_OP_POLL_ADD:
5905 return io_poll_add_prep(req, sqe);
5906 case IORING_OP_POLL_REMOVE:
5907 return io_poll_update_prep(req, sqe);
5908 case IORING_OP_FSYNC:
5909 return io_fsync_prep(req, sqe);
5910 case IORING_OP_SYNC_FILE_RANGE:
5911 return io_sfr_prep(req, sqe);
5912 case IORING_OP_SENDMSG:
5913 case IORING_OP_SEND:
5914 return io_sendmsg_prep(req, sqe);
5915 case IORING_OP_RECVMSG:
5916 case IORING_OP_RECV:
5917 return io_recvmsg_prep(req, sqe);
5918 case IORING_OP_CONNECT:
5919 return io_connect_prep(req, sqe);
5920 case IORING_OP_TIMEOUT:
5921 return io_timeout_prep(req, sqe, false);
5922 case IORING_OP_TIMEOUT_REMOVE:
5923 return io_timeout_remove_prep(req, sqe);
5924 case IORING_OP_ASYNC_CANCEL:
5925 return io_async_cancel_prep(req, sqe);
5926 case IORING_OP_LINK_TIMEOUT:
5927 return io_timeout_prep(req, sqe, true);
5928 case IORING_OP_ACCEPT:
5929 return io_accept_prep(req, sqe);
5930 case IORING_OP_FALLOCATE:
5931 return io_fallocate_prep(req, sqe);
5932 case IORING_OP_OPENAT:
5933 return io_openat_prep(req, sqe);
5934 case IORING_OP_CLOSE:
5935 return io_close_prep(req, sqe);
5936 case IORING_OP_FILES_UPDATE:
5937 return io_rsrc_update_prep(req, sqe);
5938 case IORING_OP_STATX:
5939 return io_statx_prep(req, sqe);
5940 case IORING_OP_FADVISE:
5941 return io_fadvise_prep(req, sqe);
5942 case IORING_OP_MADVISE:
5943 return io_madvise_prep(req, sqe);
5944 case IORING_OP_OPENAT2:
5945 return io_openat2_prep(req, sqe);
5946 case IORING_OP_EPOLL_CTL:
5947 return io_epoll_ctl_prep(req, sqe);
5948 case IORING_OP_SPLICE:
5949 return io_splice_prep(req, sqe);
5950 case IORING_OP_PROVIDE_BUFFERS:
5951 return io_provide_buffers_prep(req, sqe);
5952 case IORING_OP_REMOVE_BUFFERS:
5953 return io_remove_buffers_prep(req, sqe);
5955 return io_tee_prep(req, sqe);
5956 case IORING_OP_SHUTDOWN:
5957 return io_shutdown_prep(req, sqe);
5958 case IORING_OP_RENAMEAT:
5959 return io_renameat_prep(req, sqe);
5960 case IORING_OP_UNLINKAT:
5961 return io_unlinkat_prep(req, sqe);
5964 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5969 static int io_req_prep_async(struct io_kiocb *req)
5971 if (!io_op_defs[req->opcode].needs_async_setup)
5973 if (WARN_ON_ONCE(req->async_data))
5975 if (io_alloc_async_data(req))
5978 switch (req->opcode) {
5979 case IORING_OP_READV:
5980 return io_rw_prep_async(req, READ);
5981 case IORING_OP_WRITEV:
5982 return io_rw_prep_async(req, WRITE);
5983 case IORING_OP_SENDMSG:
5984 return io_sendmsg_prep_async(req);
5985 case IORING_OP_RECVMSG:
5986 return io_recvmsg_prep_async(req);
5987 case IORING_OP_CONNECT:
5988 return io_connect_prep_async(req);
5990 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5995 static u32 io_get_sequence(struct io_kiocb *req)
5997 struct io_kiocb *pos;
5998 struct io_ring_ctx *ctx = req->ctx;
5999 u32 total_submitted, nr_reqs = 0;
6001 io_for_each_link(pos, req)
6004 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6005 return total_submitted - nr_reqs;
6008 static int io_req_defer(struct io_kiocb *req)
6010 struct io_ring_ctx *ctx = req->ctx;
6011 struct io_defer_entry *de;
6015 /* Still need defer if there is pending req in defer list. */
6016 if (likely(list_empty_careful(&ctx->defer_list) &&
6017 !(req->flags & REQ_F_IO_DRAIN)))
6020 seq = io_get_sequence(req);
6021 /* Still a chance to pass the sequence check */
6022 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6025 ret = io_req_prep_async(req);
6028 io_prep_async_link(req);
6029 de = kmalloc(sizeof(*de), GFP_KERNEL);
6033 spin_lock_irq(&ctx->completion_lock);
6034 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6035 spin_unlock_irq(&ctx->completion_lock);
6037 io_queue_async_work(req);
6038 return -EIOCBQUEUED;
6041 trace_io_uring_defer(ctx, req, req->user_data);
6044 list_add_tail(&de->list, &ctx->defer_list);
6045 spin_unlock_irq(&ctx->completion_lock);
6046 return -EIOCBQUEUED;
6049 static void io_clean_op(struct io_kiocb *req)
6051 if (req->flags & REQ_F_BUFFER_SELECTED) {
6052 switch (req->opcode) {
6053 case IORING_OP_READV:
6054 case IORING_OP_READ_FIXED:
6055 case IORING_OP_READ:
6056 kfree((void *)(unsigned long)req->rw.addr);
6058 case IORING_OP_RECVMSG:
6059 case IORING_OP_RECV:
6060 kfree(req->sr_msg.kbuf);
6063 req->flags &= ~REQ_F_BUFFER_SELECTED;
6066 if (req->flags & REQ_F_NEED_CLEANUP) {
6067 switch (req->opcode) {
6068 case IORING_OP_READV:
6069 case IORING_OP_READ_FIXED:
6070 case IORING_OP_READ:
6071 case IORING_OP_WRITEV:
6072 case IORING_OP_WRITE_FIXED:
6073 case IORING_OP_WRITE: {
6074 struct io_async_rw *io = req->async_data;
6076 kfree(io->free_iovec);
6079 case IORING_OP_RECVMSG:
6080 case IORING_OP_SENDMSG: {
6081 struct io_async_msghdr *io = req->async_data;
6083 kfree(io->free_iov);
6086 case IORING_OP_SPLICE:
6088 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6089 io_put_file(req->splice.file_in);
6091 case IORING_OP_OPENAT:
6092 case IORING_OP_OPENAT2:
6093 if (req->open.filename)
6094 putname(req->open.filename);
6096 case IORING_OP_RENAMEAT:
6097 putname(req->rename.oldpath);
6098 putname(req->rename.newpath);
6100 case IORING_OP_UNLINKAT:
6101 putname(req->unlink.filename);
6104 req->flags &= ~REQ_F_NEED_CLEANUP;
6106 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6107 kfree(req->apoll->double_poll);
6111 if (req->flags & REQ_F_INFLIGHT) {
6112 struct io_uring_task *tctx = req->task->io_uring;
6114 atomic_dec(&tctx->inflight_tracked);
6115 req->flags &= ~REQ_F_INFLIGHT;
6119 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6121 struct io_ring_ctx *ctx = req->ctx;
6122 const struct cred *creds = NULL;
6125 if (req->work.creds && req->work.creds != current_cred())
6126 creds = override_creds(req->work.creds);
6128 switch (req->opcode) {
6130 ret = io_nop(req, issue_flags);
6132 case IORING_OP_READV:
6133 case IORING_OP_READ_FIXED:
6134 case IORING_OP_READ:
6135 ret = io_read(req, issue_flags);
6137 case IORING_OP_WRITEV:
6138 case IORING_OP_WRITE_FIXED:
6139 case IORING_OP_WRITE:
6140 ret = io_write(req, issue_flags);
6142 case IORING_OP_FSYNC:
6143 ret = io_fsync(req, issue_flags);
6145 case IORING_OP_POLL_ADD:
6146 ret = io_poll_add(req, issue_flags);
6148 case IORING_OP_POLL_REMOVE:
6149 ret = io_poll_update(req, issue_flags);
6151 case IORING_OP_SYNC_FILE_RANGE:
6152 ret = io_sync_file_range(req, issue_flags);
6154 case IORING_OP_SENDMSG:
6155 ret = io_sendmsg(req, issue_flags);
6157 case IORING_OP_SEND:
6158 ret = io_send(req, issue_flags);
6160 case IORING_OP_RECVMSG:
6161 ret = io_recvmsg(req, issue_flags);
6163 case IORING_OP_RECV:
6164 ret = io_recv(req, issue_flags);
6166 case IORING_OP_TIMEOUT:
6167 ret = io_timeout(req, issue_flags);
6169 case IORING_OP_TIMEOUT_REMOVE:
6170 ret = io_timeout_remove(req, issue_flags);
6172 case IORING_OP_ACCEPT:
6173 ret = io_accept(req, issue_flags);
6175 case IORING_OP_CONNECT:
6176 ret = io_connect(req, issue_flags);
6178 case IORING_OP_ASYNC_CANCEL:
6179 ret = io_async_cancel(req, issue_flags);
6181 case IORING_OP_FALLOCATE:
6182 ret = io_fallocate(req, issue_flags);
6184 case IORING_OP_OPENAT:
6185 ret = io_openat(req, issue_flags);
6187 case IORING_OP_CLOSE:
6188 ret = io_close(req, issue_flags);
6190 case IORING_OP_FILES_UPDATE:
6191 ret = io_files_update(req, issue_flags);
6193 case IORING_OP_STATX:
6194 ret = io_statx(req, issue_flags);
6196 case IORING_OP_FADVISE:
6197 ret = io_fadvise(req, issue_flags);
6199 case IORING_OP_MADVISE:
6200 ret = io_madvise(req, issue_flags);
6202 case IORING_OP_OPENAT2:
6203 ret = io_openat2(req, issue_flags);
6205 case IORING_OP_EPOLL_CTL:
6206 ret = io_epoll_ctl(req, issue_flags);
6208 case IORING_OP_SPLICE:
6209 ret = io_splice(req, issue_flags);
6211 case IORING_OP_PROVIDE_BUFFERS:
6212 ret = io_provide_buffers(req, issue_flags);
6214 case IORING_OP_REMOVE_BUFFERS:
6215 ret = io_remove_buffers(req, issue_flags);
6218 ret = io_tee(req, issue_flags);
6220 case IORING_OP_SHUTDOWN:
6221 ret = io_shutdown(req, issue_flags);
6223 case IORING_OP_RENAMEAT:
6224 ret = io_renameat(req, issue_flags);
6226 case IORING_OP_UNLINKAT:
6227 ret = io_unlinkat(req, issue_flags);
6235 revert_creds(creds);
6238 /* If the op doesn't have a file, we're not polling for it */
6239 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6240 io_iopoll_req_issued(req);
6245 static void io_wq_submit_work(struct io_wq_work *work)
6247 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6248 struct io_kiocb *timeout;
6251 timeout = io_prep_linked_timeout(req);
6253 io_queue_linked_timeout(timeout);
6255 if (work->flags & IO_WQ_WORK_CANCEL)
6260 ret = io_issue_sqe(req, 0);
6262 * We can get EAGAIN for polled IO even though we're
6263 * forcing a sync submission from here, since we can't
6264 * wait for request slots on the block side.
6272 /* avoid locking problems by failing it from a clean context */
6274 /* io-wq is going to take one down */
6276 io_req_task_queue_fail(req, ret);
6280 #define FFS_ASYNC_READ 0x1UL
6281 #define FFS_ASYNC_WRITE 0x2UL
6283 #define FFS_ISREG 0x4UL
6285 #define FFS_ISREG 0x0UL
6287 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6289 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6292 struct io_fixed_file *table_l2;
6294 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6295 return &table_l2[i & IORING_FILE_TABLE_MASK];
6298 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6301 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6303 return (struct file *) (slot->file_ptr & FFS_MASK);
6306 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6308 unsigned long file_ptr = (unsigned long) file;
6310 if (__io_file_supports_async(file, READ))
6311 file_ptr |= FFS_ASYNC_READ;
6312 if (__io_file_supports_async(file, WRITE))
6313 file_ptr |= FFS_ASYNC_WRITE;
6314 if (S_ISREG(file_inode(file)->i_mode))
6315 file_ptr |= FFS_ISREG;
6316 file_slot->file_ptr = file_ptr;
6319 static struct file *io_file_get(struct io_submit_state *state,
6320 struct io_kiocb *req, int fd, bool fixed)
6322 struct io_ring_ctx *ctx = req->ctx;
6326 unsigned long file_ptr;
6328 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6330 fd = array_index_nospec(fd, ctx->nr_user_files);
6331 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6332 file = (struct file *) (file_ptr & FFS_MASK);
6333 file_ptr &= ~FFS_MASK;
6334 /* mask in overlapping REQ_F and FFS bits */
6335 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6336 io_req_set_rsrc_node(req);
6338 trace_io_uring_file_get(ctx, fd);
6339 file = __io_file_get(state, fd);
6341 /* we don't allow fixed io_uring files */
6342 if (file && unlikely(file->f_op == &io_uring_fops))
6343 io_req_track_inflight(req);
6349 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6351 struct io_timeout_data *data = container_of(timer,
6352 struct io_timeout_data, timer);
6353 struct io_kiocb *prev, *req = data->req;
6354 struct io_ring_ctx *ctx = req->ctx;
6355 unsigned long flags;
6357 spin_lock_irqsave(&ctx->completion_lock, flags);
6358 prev = req->timeout.head;
6359 req->timeout.head = NULL;
6362 * We don't expect the list to be empty, that will only happen if we
6363 * race with the completion of the linked work.
6366 io_remove_next_linked(prev);
6367 if (!req_ref_inc_not_zero(prev))
6370 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6373 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6374 io_put_req_deferred(prev, 1);
6375 io_put_req_deferred(req, 1);
6377 io_req_complete_post(req, -ETIME, 0);
6379 return HRTIMER_NORESTART;
6382 static void io_queue_linked_timeout(struct io_kiocb *req)
6384 struct io_ring_ctx *ctx = req->ctx;
6386 spin_lock_irq(&ctx->completion_lock);
6388 * If the back reference is NULL, then our linked request finished
6389 * before we got a chance to setup the timer
6391 if (req->timeout.head) {
6392 struct io_timeout_data *data = req->async_data;
6394 data->timer.function = io_link_timeout_fn;
6395 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6398 spin_unlock_irq(&ctx->completion_lock);
6399 /* drop submission reference */
6403 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6405 struct io_kiocb *nxt = req->link;
6407 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6408 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6411 nxt->timeout.head = req;
6412 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6413 req->flags |= REQ_F_LINK_TIMEOUT;
6417 static void __io_queue_sqe(struct io_kiocb *req)
6419 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6422 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6425 * We async punt it if the file wasn't marked NOWAIT, or if the file
6426 * doesn't support non-blocking read/write attempts
6429 /* drop submission reference */
6430 if (req->flags & REQ_F_COMPLETE_INLINE) {
6431 struct io_ring_ctx *ctx = req->ctx;
6432 struct io_comp_state *cs = &ctx->submit_state.comp;
6434 cs->reqs[cs->nr++] = req;
6435 if (cs->nr == ARRAY_SIZE(cs->reqs))
6436 io_submit_flush_completions(cs, ctx);
6440 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6441 if (!io_arm_poll_handler(req)) {
6443 * Queued up for async execution, worker will release
6444 * submit reference when the iocb is actually submitted.
6446 io_queue_async_work(req);
6449 io_req_complete_failed(req, ret);
6452 io_queue_linked_timeout(linked_timeout);
6455 static void io_queue_sqe(struct io_kiocb *req)
6459 ret = io_req_defer(req);
6461 if (ret != -EIOCBQUEUED) {
6463 io_req_complete_failed(req, ret);
6465 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6466 ret = io_req_prep_async(req);
6469 io_queue_async_work(req);
6471 __io_queue_sqe(req);
6476 * Check SQE restrictions (opcode and flags).
6478 * Returns 'true' if SQE is allowed, 'false' otherwise.
6480 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6481 struct io_kiocb *req,
6482 unsigned int sqe_flags)
6484 if (!ctx->restricted)
6487 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6490 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6491 ctx->restrictions.sqe_flags_required)
6494 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6495 ctx->restrictions.sqe_flags_required))
6501 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6502 const struct io_uring_sqe *sqe)
6504 struct io_submit_state *state;
6505 unsigned int sqe_flags;
6506 int personality, ret = 0;
6508 req->opcode = READ_ONCE(sqe->opcode);
6509 /* same numerical values with corresponding REQ_F_*, safe to copy */
6510 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6511 req->user_data = READ_ONCE(sqe->user_data);
6512 req->async_data = NULL;
6516 req->fixed_rsrc_refs = NULL;
6517 /* one is dropped after submission, the other at completion */
6518 atomic_set(&req->refs, 2);
6519 req->task = current;
6521 req->work.creds = NULL;
6523 /* enforce forwards compatibility on users */
6524 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6526 if (unlikely(req->opcode >= IORING_OP_LAST))
6528 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6531 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6532 !io_op_defs[req->opcode].buffer_select)
6535 personality = READ_ONCE(sqe->personality);
6537 req->work.creds = xa_load(&ctx->personalities, personality);
6538 if (!req->work.creds)
6540 get_cred(req->work.creds);
6542 state = &ctx->submit_state;
6545 * Plug now if we have more than 1 IO left after this, and the target
6546 * is potentially a read/write to block based storage.
6548 if (!state->plug_started && state->ios_left > 1 &&
6549 io_op_defs[req->opcode].plug) {
6550 blk_start_plug(&state->plug);
6551 state->plug_started = true;
6554 if (io_op_defs[req->opcode].needs_file) {
6555 bool fixed = req->flags & REQ_F_FIXED_FILE;
6557 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6558 if (unlikely(!req->file))
6566 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6567 const struct io_uring_sqe *sqe)
6569 struct io_submit_link *link = &ctx->submit_state.link;
6572 ret = io_init_req(ctx, req, sqe);
6573 if (unlikely(ret)) {
6576 /* fail even hard links since we don't submit */
6577 req_set_fail(link->head);
6578 io_req_complete_failed(link->head, -ECANCELED);
6581 io_req_complete_failed(req, ret);
6584 ret = io_req_prep(req, sqe);
6588 /* don't need @sqe from now on */
6589 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6590 true, ctx->flags & IORING_SETUP_SQPOLL);
6593 * If we already have a head request, queue this one for async
6594 * submittal once the head completes. If we don't have a head but
6595 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6596 * submitted sync once the chain is complete. If none of those
6597 * conditions are true (normal request), then just queue it.
6600 struct io_kiocb *head = link->head;
6603 * Taking sequential execution of a link, draining both sides
6604 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6605 * requests in the link. So, it drains the head and the
6606 * next after the link request. The last one is done via
6607 * drain_next flag to persist the effect across calls.
6609 if (req->flags & REQ_F_IO_DRAIN) {
6610 head->flags |= REQ_F_IO_DRAIN;
6611 ctx->drain_next = 1;
6613 ret = io_req_prep_async(req);
6616 trace_io_uring_link(ctx, req, head);
6617 link->last->link = req;
6620 /* last request of a link, enqueue the link */
6621 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6626 if (unlikely(ctx->drain_next)) {
6627 req->flags |= REQ_F_IO_DRAIN;
6628 ctx->drain_next = 0;
6630 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6642 * Batched submission is done, ensure local IO is flushed out.
6644 static void io_submit_state_end(struct io_submit_state *state,
6645 struct io_ring_ctx *ctx)
6647 if (state->link.head)
6648 io_queue_sqe(state->link.head);
6650 io_submit_flush_completions(&state->comp, ctx);
6651 if (state->plug_started)
6652 blk_finish_plug(&state->plug);
6653 io_state_file_put(state);
6657 * Start submission side cache.
6659 static void io_submit_state_start(struct io_submit_state *state,
6660 unsigned int max_ios)
6662 state->plug_started = false;
6663 state->ios_left = max_ios;
6664 /* set only head, no need to init link_last in advance */
6665 state->link.head = NULL;
6668 static void io_commit_sqring(struct io_ring_ctx *ctx)
6670 struct io_rings *rings = ctx->rings;
6673 * Ensure any loads from the SQEs are done at this point,
6674 * since once we write the new head, the application could
6675 * write new data to them.
6677 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6681 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6682 * that is mapped by userspace. This means that care needs to be taken to
6683 * ensure that reads are stable, as we cannot rely on userspace always
6684 * being a good citizen. If members of the sqe are validated and then later
6685 * used, it's important that those reads are done through READ_ONCE() to
6686 * prevent a re-load down the line.
6688 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6690 u32 *sq_array = ctx->sq_array;
6691 unsigned head, mask = ctx->sq_entries - 1;
6694 * The cached sq head (or cq tail) serves two purposes:
6696 * 1) allows us to batch the cost of updating the user visible
6698 * 2) allows the kernel side to track the head on its own, even
6699 * though the application is the one updating it.
6701 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & mask]);
6702 if (likely(head < ctx->sq_entries))
6703 return &ctx->sq_sqes[head];
6705 /* drop invalid entries */
6706 ctx->cached_sq_dropped++;
6707 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6711 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6713 struct io_uring_task *tctx;
6716 /* make sure SQ entry isn't read before tail */
6717 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6718 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6721 tctx = current->io_uring;
6722 tctx->cached_refs -= nr;
6723 if (unlikely(tctx->cached_refs < 0)) {
6724 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6726 percpu_counter_add(&tctx->inflight, refill);
6727 refcount_add(refill, ¤t->usage);
6728 tctx->cached_refs += refill;
6730 io_submit_state_start(&ctx->submit_state, nr);
6732 while (submitted < nr) {
6733 const struct io_uring_sqe *sqe;
6734 struct io_kiocb *req;
6736 req = io_alloc_req(ctx);
6737 if (unlikely(!req)) {
6739 submitted = -EAGAIN;
6742 sqe = io_get_sqe(ctx);
6743 if (unlikely(!sqe)) {
6744 kmem_cache_free(req_cachep, req);
6747 /* will complete beyond this point, count as submitted */
6749 if (io_submit_sqe(ctx, req, sqe))
6753 if (unlikely(submitted != nr)) {
6754 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6755 int unused = nr - ref_used;
6757 current->io_uring->cached_refs += unused;
6758 percpu_ref_put_many(&ctx->refs, unused);
6761 io_submit_state_end(&ctx->submit_state, ctx);
6762 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6763 io_commit_sqring(ctx);
6768 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6770 return READ_ONCE(sqd->state);
6773 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6775 /* Tell userspace we may need a wakeup call */
6776 spin_lock_irq(&ctx->completion_lock);
6777 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6778 spin_unlock_irq(&ctx->completion_lock);
6781 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6783 spin_lock_irq(&ctx->completion_lock);
6784 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6785 spin_unlock_irq(&ctx->completion_lock);
6788 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6790 unsigned int to_submit;
6793 to_submit = io_sqring_entries(ctx);
6794 /* if we're handling multiple rings, cap submit size for fairness */
6795 if (cap_entries && to_submit > 8)
6798 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6799 unsigned nr_events = 0;
6801 mutex_lock(&ctx->uring_lock);
6802 if (!list_empty(&ctx->iopoll_list))
6803 io_do_iopoll(ctx, &nr_events, 0);
6806 * Don't submit if refs are dying, good for io_uring_register(),
6807 * but also it is relied upon by io_ring_exit_work()
6809 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6810 !(ctx->flags & IORING_SETUP_R_DISABLED))
6811 ret = io_submit_sqes(ctx, to_submit);
6812 mutex_unlock(&ctx->uring_lock);
6814 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6815 wake_up(&ctx->sqo_sq_wait);
6821 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6823 struct io_ring_ctx *ctx;
6824 unsigned sq_thread_idle = 0;
6826 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6827 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6828 sqd->sq_thread_idle = sq_thread_idle;
6831 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6833 bool did_sig = false;
6834 struct ksignal ksig;
6836 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6837 signal_pending(current)) {
6838 mutex_unlock(&sqd->lock);
6839 if (signal_pending(current))
6840 did_sig = get_signal(&ksig);
6842 mutex_lock(&sqd->lock);
6845 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6848 static int io_sq_thread(void *data)
6850 struct io_sq_data *sqd = data;
6851 struct io_ring_ctx *ctx;
6852 unsigned long timeout = 0;
6853 char buf[TASK_COMM_LEN];
6856 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6857 set_task_comm(current, buf);
6859 if (sqd->sq_cpu != -1)
6860 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6862 set_cpus_allowed_ptr(current, cpu_online_mask);
6863 current->flags |= PF_NO_SETAFFINITY;
6865 mutex_lock(&sqd->lock);
6868 bool cap_entries, sqt_spin, needs_sched;
6870 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6871 if (io_sqd_handle_event(sqd))
6873 timeout = jiffies + sqd->sq_thread_idle;
6878 cap_entries = !list_is_singular(&sqd->ctx_list);
6879 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6880 const struct cred *creds = NULL;
6882 if (ctx->sq_creds != current_cred())
6883 creds = override_creds(ctx->sq_creds);
6884 ret = __io_sq_thread(ctx, cap_entries);
6886 revert_creds(creds);
6887 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6891 if (sqt_spin || !time_after(jiffies, timeout)) {
6895 timeout = jiffies + sqd->sq_thread_idle;
6899 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6900 if (!io_sqd_events_pending(sqd)) {
6902 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6903 io_ring_set_wakeup_flag(ctx);
6905 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6906 !list_empty_careful(&ctx->iopoll_list)) {
6907 needs_sched = false;
6910 if (io_sqring_entries(ctx)) {
6911 needs_sched = false;
6917 mutex_unlock(&sqd->lock);
6919 mutex_lock(&sqd->lock);
6921 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6922 io_ring_clear_wakeup_flag(ctx);
6925 finish_wait(&sqd->wait, &wait);
6926 timeout = jiffies + sqd->sq_thread_idle;
6929 io_uring_cancel_generic(true, sqd);
6931 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6932 io_ring_set_wakeup_flag(ctx);
6934 mutex_unlock(&sqd->lock);
6936 complete(&sqd->exited);
6940 struct io_wait_queue {
6941 struct wait_queue_entry wq;
6942 struct io_ring_ctx *ctx;
6944 unsigned nr_timeouts;
6947 static inline bool io_should_wake(struct io_wait_queue *iowq)
6949 struct io_ring_ctx *ctx = iowq->ctx;
6952 * Wake up if we have enough events, or if a timeout occurred since we
6953 * started waiting. For timeouts, we always want to return to userspace,
6954 * regardless of event count.
6956 return io_cqring_events(ctx) >= iowq->to_wait ||
6957 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6960 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6961 int wake_flags, void *key)
6963 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6967 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6968 * the task, and the next invocation will do it.
6970 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6971 return autoremove_wake_function(curr, mode, wake_flags, key);
6975 static int io_run_task_work_sig(void)
6977 if (io_run_task_work())
6979 if (!signal_pending(current))
6981 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6982 return -ERESTARTSYS;
6986 /* when returns >0, the caller should retry */
6987 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6988 struct io_wait_queue *iowq,
6989 signed long *timeout)
6993 /* make sure we run task_work before checking for signals */
6994 ret = io_run_task_work_sig();
6995 if (ret || io_should_wake(iowq))
6997 /* let the caller flush overflows, retry */
6998 if (test_bit(0, &ctx->cq_check_overflow))
7001 *timeout = schedule_timeout(*timeout);
7002 return !*timeout ? -ETIME : 1;
7006 * Wait until events become available, if we don't already have some. The
7007 * application must reap them itself, as they reside on the shared cq ring.
7009 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7010 const sigset_t __user *sig, size_t sigsz,
7011 struct __kernel_timespec __user *uts)
7013 struct io_wait_queue iowq = {
7016 .func = io_wake_function,
7017 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7020 .to_wait = min_events,
7022 struct io_rings *rings = ctx->rings;
7023 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7027 io_cqring_overflow_flush(ctx, false);
7028 if (io_cqring_events(ctx) >= min_events)
7030 if (!io_run_task_work())
7035 #ifdef CONFIG_COMPAT
7036 if (in_compat_syscall())
7037 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7041 ret = set_user_sigmask(sig, sigsz);
7048 struct timespec64 ts;
7050 if (get_timespec64(&ts, uts))
7052 timeout = timespec64_to_jiffies(&ts);
7055 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7056 trace_io_uring_cqring_wait(ctx, min_events);
7058 /* if we can't even flush overflow, don't wait for more */
7059 if (!io_cqring_overflow_flush(ctx, false)) {
7063 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7064 TASK_INTERRUPTIBLE);
7065 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7066 finish_wait(&ctx->wait, &iowq.wq);
7070 restore_saved_sigmask_unless(ret == -EINTR);
7072 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7075 static void io_free_page_table(void **table, size_t size)
7077 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7079 for (i = 0; i < nr_tables; i++)
7084 static void **io_alloc_page_table(size_t size)
7086 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7087 size_t init_size = size;
7090 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7094 for (i = 0; i < nr_tables; i++) {
7095 unsigned int this_size = min(size, PAGE_SIZE);
7097 table[i] = kzalloc(this_size, GFP_KERNEL);
7099 io_free_page_table(table, init_size);
7107 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7109 spin_lock_bh(&ctx->rsrc_ref_lock);
7112 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7114 spin_unlock_bh(&ctx->rsrc_ref_lock);
7117 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7119 percpu_ref_exit(&ref_node->refs);
7123 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7124 struct io_rsrc_data *data_to_kill)
7126 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7127 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7130 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7132 rsrc_node->rsrc_data = data_to_kill;
7133 io_rsrc_ref_lock(ctx);
7134 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7135 io_rsrc_ref_unlock(ctx);
7137 atomic_inc(&data_to_kill->refs);
7138 percpu_ref_kill(&rsrc_node->refs);
7139 ctx->rsrc_node = NULL;
7142 if (!ctx->rsrc_node) {
7143 ctx->rsrc_node = ctx->rsrc_backup_node;
7144 ctx->rsrc_backup_node = NULL;
7148 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7150 if (ctx->rsrc_backup_node)
7152 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7153 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7156 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7160 /* As we may drop ->uring_lock, other task may have started quiesce */
7164 data->quiesce = true;
7166 ret = io_rsrc_node_switch_start(ctx);
7169 io_rsrc_node_switch(ctx, data);
7171 /* kill initial ref, already quiesced if zero */
7172 if (atomic_dec_and_test(&data->refs))
7174 flush_delayed_work(&ctx->rsrc_put_work);
7175 ret = wait_for_completion_interruptible(&data->done);
7179 atomic_inc(&data->refs);
7180 /* wait for all works potentially completing data->done */
7181 flush_delayed_work(&ctx->rsrc_put_work);
7182 reinit_completion(&data->done);
7184 mutex_unlock(&ctx->uring_lock);
7185 ret = io_run_task_work_sig();
7186 mutex_lock(&ctx->uring_lock);
7188 data->quiesce = false;
7193 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7195 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7196 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7198 return &data->tags[table_idx][off];
7201 static void io_rsrc_data_free(struct io_rsrc_data *data)
7203 size_t size = data->nr * sizeof(data->tags[0][0]);
7206 io_free_page_table((void **)data->tags, size);
7210 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7211 u64 __user *utags, unsigned nr,
7212 struct io_rsrc_data **pdata)
7214 struct io_rsrc_data *data;
7218 data = kzalloc(sizeof(*data), GFP_KERNEL);
7221 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7229 data->do_put = do_put;
7232 for (i = 0; i < nr; i++) {
7233 if (copy_from_user(io_get_tag_slot(data, i), &utags[i],
7234 sizeof(data->tags[i])))
7239 atomic_set(&data->refs, 1);
7240 init_completion(&data->done);
7244 io_rsrc_data_free(data);
7248 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7250 size_t size = nr_files * sizeof(struct io_fixed_file);
7252 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7253 return !!table->files;
7256 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7258 size_t size = nr_files * sizeof(struct io_fixed_file);
7260 io_free_page_table((void **)table->files, size);
7261 table->files = NULL;
7264 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7266 #if defined(CONFIG_UNIX)
7267 if (ctx->ring_sock) {
7268 struct sock *sock = ctx->ring_sock->sk;
7269 struct sk_buff *skb;
7271 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7277 for (i = 0; i < ctx->nr_user_files; i++) {
7280 file = io_file_from_index(ctx, i);
7285 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7286 io_rsrc_data_free(ctx->file_data);
7287 ctx->file_data = NULL;
7288 ctx->nr_user_files = 0;
7291 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7295 if (!ctx->file_data)
7297 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7299 __io_sqe_files_unregister(ctx);
7303 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7304 __releases(&sqd->lock)
7306 WARN_ON_ONCE(sqd->thread == current);
7309 * Do the dance but not conditional clear_bit() because it'd race with
7310 * other threads incrementing park_pending and setting the bit.
7312 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7313 if (atomic_dec_return(&sqd->park_pending))
7314 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7315 mutex_unlock(&sqd->lock);
7318 static void io_sq_thread_park(struct io_sq_data *sqd)
7319 __acquires(&sqd->lock)
7321 WARN_ON_ONCE(sqd->thread == current);
7323 atomic_inc(&sqd->park_pending);
7324 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7325 mutex_lock(&sqd->lock);
7327 wake_up_process(sqd->thread);
7330 static void io_sq_thread_stop(struct io_sq_data *sqd)
7332 WARN_ON_ONCE(sqd->thread == current);
7333 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7335 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7336 mutex_lock(&sqd->lock);
7338 wake_up_process(sqd->thread);
7339 mutex_unlock(&sqd->lock);
7340 wait_for_completion(&sqd->exited);
7343 static void io_put_sq_data(struct io_sq_data *sqd)
7345 if (refcount_dec_and_test(&sqd->refs)) {
7346 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7348 io_sq_thread_stop(sqd);
7353 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7355 struct io_sq_data *sqd = ctx->sq_data;
7358 io_sq_thread_park(sqd);
7359 list_del_init(&ctx->sqd_list);
7360 io_sqd_update_thread_idle(sqd);
7361 io_sq_thread_unpark(sqd);
7363 io_put_sq_data(sqd);
7364 ctx->sq_data = NULL;
7368 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7370 struct io_ring_ctx *ctx_attach;
7371 struct io_sq_data *sqd;
7374 f = fdget(p->wq_fd);
7376 return ERR_PTR(-ENXIO);
7377 if (f.file->f_op != &io_uring_fops) {
7379 return ERR_PTR(-EINVAL);
7382 ctx_attach = f.file->private_data;
7383 sqd = ctx_attach->sq_data;
7386 return ERR_PTR(-EINVAL);
7388 if (sqd->task_tgid != current->tgid) {
7390 return ERR_PTR(-EPERM);
7393 refcount_inc(&sqd->refs);
7398 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7401 struct io_sq_data *sqd;
7404 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7405 sqd = io_attach_sq_data(p);
7410 /* fall through for EPERM case, setup new sqd/task */
7411 if (PTR_ERR(sqd) != -EPERM)
7415 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7417 return ERR_PTR(-ENOMEM);
7419 atomic_set(&sqd->park_pending, 0);
7420 refcount_set(&sqd->refs, 1);
7421 INIT_LIST_HEAD(&sqd->ctx_list);
7422 mutex_init(&sqd->lock);
7423 init_waitqueue_head(&sqd->wait);
7424 init_completion(&sqd->exited);
7428 #if defined(CONFIG_UNIX)
7430 * Ensure the UNIX gc is aware of our file set, so we are certain that
7431 * the io_uring can be safely unregistered on process exit, even if we have
7432 * loops in the file referencing.
7434 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7436 struct sock *sk = ctx->ring_sock->sk;
7437 struct scm_fp_list *fpl;
7438 struct sk_buff *skb;
7441 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7445 skb = alloc_skb(0, GFP_KERNEL);
7454 fpl->user = get_uid(current_user());
7455 for (i = 0; i < nr; i++) {
7456 struct file *file = io_file_from_index(ctx, i + offset);
7460 fpl->fp[nr_files] = get_file(file);
7461 unix_inflight(fpl->user, fpl->fp[nr_files]);
7466 fpl->max = SCM_MAX_FD;
7467 fpl->count = nr_files;
7468 UNIXCB(skb).fp = fpl;
7469 skb->destructor = unix_destruct_scm;
7470 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7471 skb_queue_head(&sk->sk_receive_queue, skb);
7473 for (i = 0; i < nr_files; i++)
7484 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7485 * causes regular reference counting to break down. We rely on the UNIX
7486 * garbage collection to take care of this problem for us.
7488 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7490 unsigned left, total;
7494 left = ctx->nr_user_files;
7496 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7498 ret = __io_sqe_files_scm(ctx, this_files, total);
7502 total += this_files;
7508 while (total < ctx->nr_user_files) {
7509 struct file *file = io_file_from_index(ctx, total);
7519 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7525 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7527 struct file *file = prsrc->file;
7528 #if defined(CONFIG_UNIX)
7529 struct sock *sock = ctx->ring_sock->sk;
7530 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7531 struct sk_buff *skb;
7534 __skb_queue_head_init(&list);
7537 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7538 * remove this entry and rearrange the file array.
7540 skb = skb_dequeue(head);
7542 struct scm_fp_list *fp;
7544 fp = UNIXCB(skb).fp;
7545 for (i = 0; i < fp->count; i++) {
7548 if (fp->fp[i] != file)
7551 unix_notinflight(fp->user, fp->fp[i]);
7552 left = fp->count - 1 - i;
7554 memmove(&fp->fp[i], &fp->fp[i + 1],
7555 left * sizeof(struct file *));
7562 __skb_queue_tail(&list, skb);
7572 __skb_queue_tail(&list, skb);
7574 skb = skb_dequeue(head);
7577 if (skb_peek(&list)) {
7578 spin_lock_irq(&head->lock);
7579 while ((skb = __skb_dequeue(&list)) != NULL)
7580 __skb_queue_tail(head, skb);
7581 spin_unlock_irq(&head->lock);
7588 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7590 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7591 struct io_ring_ctx *ctx = rsrc_data->ctx;
7592 struct io_rsrc_put *prsrc, *tmp;
7594 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7595 list_del(&prsrc->list);
7598 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7600 io_ring_submit_lock(ctx, lock_ring);
7601 spin_lock_irq(&ctx->completion_lock);
7602 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7604 io_commit_cqring(ctx);
7605 spin_unlock_irq(&ctx->completion_lock);
7606 io_cqring_ev_posted(ctx);
7607 io_ring_submit_unlock(ctx, lock_ring);
7610 rsrc_data->do_put(ctx, prsrc);
7614 io_rsrc_node_destroy(ref_node);
7615 if (atomic_dec_and_test(&rsrc_data->refs))
7616 complete(&rsrc_data->done);
7619 static void io_rsrc_put_work(struct work_struct *work)
7621 struct io_ring_ctx *ctx;
7622 struct llist_node *node;
7624 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7625 node = llist_del_all(&ctx->rsrc_put_llist);
7628 struct io_rsrc_node *ref_node;
7629 struct llist_node *next = node->next;
7631 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7632 __io_rsrc_put_work(ref_node);
7637 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7639 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7640 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7641 bool first_add = false;
7643 io_rsrc_ref_lock(ctx);
7646 while (!list_empty(&ctx->rsrc_ref_list)) {
7647 node = list_first_entry(&ctx->rsrc_ref_list,
7648 struct io_rsrc_node, node);
7649 /* recycle ref nodes in order */
7652 list_del(&node->node);
7653 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7655 io_rsrc_ref_unlock(ctx);
7658 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7661 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7663 struct io_rsrc_node *ref_node;
7665 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7669 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7674 INIT_LIST_HEAD(&ref_node->node);
7675 INIT_LIST_HEAD(&ref_node->rsrc_list);
7676 ref_node->done = false;
7680 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7681 unsigned nr_args, u64 __user *tags)
7683 __s32 __user *fds = (__s32 __user *) arg;
7692 if (nr_args > IORING_MAX_FIXED_FILES)
7694 ret = io_rsrc_node_switch_start(ctx);
7697 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7703 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7706 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7707 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7711 /* allow sparse sets */
7714 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7721 if (unlikely(!file))
7725 * Don't allow io_uring instances to be registered. If UNIX
7726 * isn't enabled, then this causes a reference cycle and this
7727 * instance can never get freed. If UNIX is enabled we'll
7728 * handle it just fine, but there's still no point in allowing
7729 * a ring fd as it doesn't support regular read/write anyway.
7731 if (file->f_op == &io_uring_fops) {
7735 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7738 ret = io_sqe_files_scm(ctx);
7740 __io_sqe_files_unregister(ctx);
7744 io_rsrc_node_switch(ctx, NULL);
7747 for (i = 0; i < ctx->nr_user_files; i++) {
7748 file = io_file_from_index(ctx, i);
7752 io_free_file_tables(&ctx->file_table, nr_args);
7753 ctx->nr_user_files = 0;
7755 io_rsrc_data_free(ctx->file_data);
7756 ctx->file_data = NULL;
7760 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7763 #if defined(CONFIG_UNIX)
7764 struct sock *sock = ctx->ring_sock->sk;
7765 struct sk_buff_head *head = &sock->sk_receive_queue;
7766 struct sk_buff *skb;
7769 * See if we can merge this file into an existing skb SCM_RIGHTS
7770 * file set. If there's no room, fall back to allocating a new skb
7771 * and filling it in.
7773 spin_lock_irq(&head->lock);
7774 skb = skb_peek(head);
7776 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7778 if (fpl->count < SCM_MAX_FD) {
7779 __skb_unlink(skb, head);
7780 spin_unlock_irq(&head->lock);
7781 fpl->fp[fpl->count] = get_file(file);
7782 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7784 spin_lock_irq(&head->lock);
7785 __skb_queue_head(head, skb);
7790 spin_unlock_irq(&head->lock);
7797 return __io_sqe_files_scm(ctx, 1, index);
7803 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7804 struct io_rsrc_node *node, void *rsrc)
7806 struct io_rsrc_put *prsrc;
7808 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7812 prsrc->tag = *io_get_tag_slot(data, idx);
7814 list_add(&prsrc->list, &node->rsrc_list);
7818 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7819 struct io_uring_rsrc_update2 *up,
7822 u64 __user *tags = u64_to_user_ptr(up->tags);
7823 __s32 __user *fds = u64_to_user_ptr(up->data);
7824 struct io_rsrc_data *data = ctx->file_data;
7825 struct io_fixed_file *file_slot;
7829 bool needs_switch = false;
7831 if (!ctx->file_data)
7833 if (up->offset + nr_args > ctx->nr_user_files)
7836 for (done = 0; done < nr_args; done++) {
7839 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7840 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7844 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7848 if (fd == IORING_REGISTER_FILES_SKIP)
7851 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7852 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7854 if (file_slot->file_ptr) {
7855 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7856 err = io_queue_rsrc_removal(data, up->offset + done,
7857 ctx->rsrc_node, file);
7860 file_slot->file_ptr = 0;
7861 needs_switch = true;
7870 * Don't allow io_uring instances to be registered. If
7871 * UNIX isn't enabled, then this causes a reference
7872 * cycle and this instance can never get freed. If UNIX
7873 * is enabled we'll handle it just fine, but there's
7874 * still no point in allowing a ring fd as it doesn't
7875 * support regular read/write anyway.
7877 if (file->f_op == &io_uring_fops) {
7882 *io_get_tag_slot(data, up->offset + done) = tag;
7883 io_fixed_file_set(file_slot, file);
7884 err = io_sqe_file_register(ctx, file, i);
7886 file_slot->file_ptr = 0;
7894 io_rsrc_node_switch(ctx, data);
7895 return done ? done : err;
7898 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7900 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7902 req = io_put_req_find_next(req);
7903 return req ? &req->work : NULL;
7906 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7907 struct task_struct *task)
7909 struct io_wq_hash *hash;
7910 struct io_wq_data data;
7911 unsigned int concurrency;
7913 hash = ctx->hash_map;
7915 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7917 return ERR_PTR(-ENOMEM);
7918 refcount_set(&hash->refs, 1);
7919 init_waitqueue_head(&hash->wait);
7920 ctx->hash_map = hash;
7925 data.free_work = io_free_work;
7926 data.do_work = io_wq_submit_work;
7928 /* Do QD, or 4 * CPUS, whatever is smallest */
7929 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7931 return io_wq_create(concurrency, &data);
7934 static int io_uring_alloc_task_context(struct task_struct *task,
7935 struct io_ring_ctx *ctx)
7937 struct io_uring_task *tctx;
7940 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7941 if (unlikely(!tctx))
7944 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7945 if (unlikely(ret)) {
7950 tctx->io_wq = io_init_wq_offload(ctx, task);
7951 if (IS_ERR(tctx->io_wq)) {
7952 ret = PTR_ERR(tctx->io_wq);
7953 percpu_counter_destroy(&tctx->inflight);
7959 init_waitqueue_head(&tctx->wait);
7960 atomic_set(&tctx->in_idle, 0);
7961 atomic_set(&tctx->inflight_tracked, 0);
7962 task->io_uring = tctx;
7963 spin_lock_init(&tctx->task_lock);
7964 INIT_WQ_LIST(&tctx->task_list);
7965 init_task_work(&tctx->task_work, tctx_task_work);
7969 void __io_uring_free(struct task_struct *tsk)
7971 struct io_uring_task *tctx = tsk->io_uring;
7973 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7974 WARN_ON_ONCE(tctx->io_wq);
7975 WARN_ON_ONCE(tctx->cached_refs);
7977 percpu_counter_destroy(&tctx->inflight);
7979 tsk->io_uring = NULL;
7982 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7983 struct io_uring_params *p)
7987 /* Retain compatibility with failing for an invalid attach attempt */
7988 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7989 IORING_SETUP_ATTACH_WQ) {
7992 f = fdget(p->wq_fd);
7996 if (f.file->f_op != &io_uring_fops)
7999 if (ctx->flags & IORING_SETUP_SQPOLL) {
8000 struct task_struct *tsk;
8001 struct io_sq_data *sqd;
8004 sqd = io_get_sq_data(p, &attached);
8010 ctx->sq_creds = get_current_cred();
8012 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8013 if (!ctx->sq_thread_idle)
8014 ctx->sq_thread_idle = HZ;
8016 io_sq_thread_park(sqd);
8017 list_add(&ctx->sqd_list, &sqd->ctx_list);
8018 io_sqd_update_thread_idle(sqd);
8019 /* don't attach to a dying SQPOLL thread, would be racy */
8020 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8021 io_sq_thread_unpark(sqd);
8028 if (p->flags & IORING_SETUP_SQ_AFF) {
8029 int cpu = p->sq_thread_cpu;
8032 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8039 sqd->task_pid = current->pid;
8040 sqd->task_tgid = current->tgid;
8041 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8048 ret = io_uring_alloc_task_context(tsk, ctx);
8049 wake_up_new_task(tsk);
8052 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8053 /* Can't have SQ_AFF without SQPOLL */
8060 complete(&ctx->sq_data->exited);
8062 io_sq_thread_finish(ctx);
8066 static inline void __io_unaccount_mem(struct user_struct *user,
8067 unsigned long nr_pages)
8069 atomic_long_sub(nr_pages, &user->locked_vm);
8072 static inline int __io_account_mem(struct user_struct *user,
8073 unsigned long nr_pages)
8075 unsigned long page_limit, cur_pages, new_pages;
8077 /* Don't allow more pages than we can safely lock */
8078 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8081 cur_pages = atomic_long_read(&user->locked_vm);
8082 new_pages = cur_pages + nr_pages;
8083 if (new_pages > page_limit)
8085 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8086 new_pages) != cur_pages);
8091 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8094 __io_unaccount_mem(ctx->user, nr_pages);
8096 if (ctx->mm_account)
8097 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8100 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8105 ret = __io_account_mem(ctx->user, nr_pages);
8110 if (ctx->mm_account)
8111 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8116 static void io_mem_free(void *ptr)
8123 page = virt_to_head_page(ptr);
8124 if (put_page_testzero(page))
8125 free_compound_page(page);
8128 static void *io_mem_alloc(size_t size)
8130 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8131 __GFP_NORETRY | __GFP_ACCOUNT;
8133 return (void *) __get_free_pages(gfp_flags, get_order(size));
8136 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8139 struct io_rings *rings;
8140 size_t off, sq_array_size;
8142 off = struct_size(rings, cqes, cq_entries);
8143 if (off == SIZE_MAX)
8147 off = ALIGN(off, SMP_CACHE_BYTES);
8155 sq_array_size = array_size(sizeof(u32), sq_entries);
8156 if (sq_array_size == SIZE_MAX)
8159 if (check_add_overflow(off, sq_array_size, &off))
8165 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8167 struct io_mapped_ubuf *imu = *slot;
8170 if (imu != ctx->dummy_ubuf) {
8171 for (i = 0; i < imu->nr_bvecs; i++)
8172 unpin_user_page(imu->bvec[i].bv_page);
8173 if (imu->acct_pages)
8174 io_unaccount_mem(ctx, imu->acct_pages);
8180 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8182 io_buffer_unmap(ctx, &prsrc->buf);
8186 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8190 for (i = 0; i < ctx->nr_user_bufs; i++)
8191 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8192 kfree(ctx->user_bufs);
8193 io_rsrc_data_free(ctx->buf_data);
8194 ctx->user_bufs = NULL;
8195 ctx->buf_data = NULL;
8196 ctx->nr_user_bufs = 0;
8199 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8206 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8208 __io_sqe_buffers_unregister(ctx);
8212 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8213 void __user *arg, unsigned index)
8215 struct iovec __user *src;
8217 #ifdef CONFIG_COMPAT
8219 struct compat_iovec __user *ciovs;
8220 struct compat_iovec ciov;
8222 ciovs = (struct compat_iovec __user *) arg;
8223 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8226 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8227 dst->iov_len = ciov.iov_len;
8231 src = (struct iovec __user *) arg;
8232 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8238 * Not super efficient, but this is just a registration time. And we do cache
8239 * the last compound head, so generally we'll only do a full search if we don't
8242 * We check if the given compound head page has already been accounted, to
8243 * avoid double accounting it. This allows us to account the full size of the
8244 * page, not just the constituent pages of a huge page.
8246 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8247 int nr_pages, struct page *hpage)
8251 /* check current page array */
8252 for (i = 0; i < nr_pages; i++) {
8253 if (!PageCompound(pages[i]))
8255 if (compound_head(pages[i]) == hpage)
8259 /* check previously registered pages */
8260 for (i = 0; i < ctx->nr_user_bufs; i++) {
8261 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8263 for (j = 0; j < imu->nr_bvecs; j++) {
8264 if (!PageCompound(imu->bvec[j].bv_page))
8266 if (compound_head(imu->bvec[j].bv_page) == hpage)
8274 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8275 int nr_pages, struct io_mapped_ubuf *imu,
8276 struct page **last_hpage)
8280 imu->acct_pages = 0;
8281 for (i = 0; i < nr_pages; i++) {
8282 if (!PageCompound(pages[i])) {
8287 hpage = compound_head(pages[i]);
8288 if (hpage == *last_hpage)
8290 *last_hpage = hpage;
8291 if (headpage_already_acct(ctx, pages, i, hpage))
8293 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8297 if (!imu->acct_pages)
8300 ret = io_account_mem(ctx, imu->acct_pages);
8302 imu->acct_pages = 0;
8306 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8307 struct io_mapped_ubuf **pimu,
8308 struct page **last_hpage)
8310 struct io_mapped_ubuf *imu = NULL;
8311 struct vm_area_struct **vmas = NULL;
8312 struct page **pages = NULL;
8313 unsigned long off, start, end, ubuf;
8315 int ret, pret, nr_pages, i;
8317 if (!iov->iov_base) {
8318 *pimu = ctx->dummy_ubuf;
8322 ubuf = (unsigned long) iov->iov_base;
8323 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8324 start = ubuf >> PAGE_SHIFT;
8325 nr_pages = end - start;
8330 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8334 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8339 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8344 mmap_read_lock(current->mm);
8345 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8347 if (pret == nr_pages) {
8348 /* don't support file backed memory */
8349 for (i = 0; i < nr_pages; i++) {
8350 struct vm_area_struct *vma = vmas[i];
8352 if (vma_is_shmem(vma))
8355 !is_file_hugepages(vma->vm_file)) {
8361 ret = pret < 0 ? pret : -EFAULT;
8363 mmap_read_unlock(current->mm);
8366 * if we did partial map, or found file backed vmas,
8367 * release any pages we did get
8370 unpin_user_pages(pages, pret);
8374 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8376 unpin_user_pages(pages, pret);
8380 off = ubuf & ~PAGE_MASK;
8381 size = iov->iov_len;
8382 for (i = 0; i < nr_pages; i++) {
8385 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8386 imu->bvec[i].bv_page = pages[i];
8387 imu->bvec[i].bv_len = vec_len;
8388 imu->bvec[i].bv_offset = off;
8392 /* store original address for later verification */
8394 imu->ubuf_end = ubuf + iov->iov_len;
8395 imu->nr_bvecs = nr_pages;
8406 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8408 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8409 return ctx->user_bufs ? 0 : -ENOMEM;
8412 static int io_buffer_validate(struct iovec *iov)
8414 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8417 * Don't impose further limits on the size and buffer
8418 * constraints here, we'll -EINVAL later when IO is
8419 * submitted if they are wrong.
8422 return iov->iov_len ? -EFAULT : 0;
8426 /* arbitrary limit, but we need something */
8427 if (iov->iov_len > SZ_1G)
8430 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8436 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8437 unsigned int nr_args, u64 __user *tags)
8439 struct page *last_hpage = NULL;
8440 struct io_rsrc_data *data;
8446 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8448 ret = io_rsrc_node_switch_start(ctx);
8451 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8454 ret = io_buffers_map_alloc(ctx, nr_args);
8456 io_rsrc_data_free(data);
8460 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8461 ret = io_copy_iov(ctx, &iov, arg, i);
8464 ret = io_buffer_validate(&iov);
8467 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8472 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8478 WARN_ON_ONCE(ctx->buf_data);
8480 ctx->buf_data = data;
8482 __io_sqe_buffers_unregister(ctx);
8484 io_rsrc_node_switch(ctx, NULL);
8488 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8489 struct io_uring_rsrc_update2 *up,
8490 unsigned int nr_args)
8492 u64 __user *tags = u64_to_user_ptr(up->tags);
8493 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8494 struct page *last_hpage = NULL;
8495 bool needs_switch = false;
8501 if (up->offset + nr_args > ctx->nr_user_bufs)
8504 for (done = 0; done < nr_args; done++) {
8505 struct io_mapped_ubuf *imu;
8506 int offset = up->offset + done;
8509 err = io_copy_iov(ctx, &iov, iovs, done);
8512 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8516 err = io_buffer_validate(&iov);
8519 if (!iov.iov_base && tag) {
8523 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8527 i = array_index_nospec(offset, ctx->nr_user_bufs);
8528 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8529 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8530 ctx->rsrc_node, ctx->user_bufs[i]);
8531 if (unlikely(err)) {
8532 io_buffer_unmap(ctx, &imu);
8535 ctx->user_bufs[i] = NULL;
8536 needs_switch = true;
8539 ctx->user_bufs[i] = imu;
8540 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8544 io_rsrc_node_switch(ctx, ctx->buf_data);
8545 return done ? done : err;
8548 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8550 __s32 __user *fds = arg;
8556 if (copy_from_user(&fd, fds, sizeof(*fds)))
8559 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8560 if (IS_ERR(ctx->cq_ev_fd)) {
8561 int ret = PTR_ERR(ctx->cq_ev_fd);
8562 ctx->cq_ev_fd = NULL;
8569 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8571 if (ctx->cq_ev_fd) {
8572 eventfd_ctx_put(ctx->cq_ev_fd);
8573 ctx->cq_ev_fd = NULL;
8580 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8582 struct io_buffer *buf;
8583 unsigned long index;
8585 xa_for_each(&ctx->io_buffers, index, buf)
8586 __io_remove_buffers(ctx, buf, index, -1U);
8589 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8591 struct io_kiocb *req, *nxt;
8593 list_for_each_entry_safe(req, nxt, list, compl.list) {
8594 if (tsk && req->task != tsk)
8596 list_del(&req->compl.list);
8597 kmem_cache_free(req_cachep, req);
8601 static void io_req_caches_free(struct io_ring_ctx *ctx)
8603 struct io_submit_state *submit_state = &ctx->submit_state;
8604 struct io_comp_state *cs = &ctx->submit_state.comp;
8606 mutex_lock(&ctx->uring_lock);
8608 if (submit_state->free_reqs) {
8609 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8610 submit_state->reqs);
8611 submit_state->free_reqs = 0;
8614 io_flush_cached_locked_reqs(ctx, cs);
8615 io_req_cache_free(&cs->free_list, NULL);
8616 mutex_unlock(&ctx->uring_lock);
8619 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8623 if (!atomic_dec_and_test(&data->refs))
8624 wait_for_completion(&data->done);
8628 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8630 io_sq_thread_finish(ctx);
8632 if (ctx->mm_account) {
8633 mmdrop(ctx->mm_account);
8634 ctx->mm_account = NULL;
8637 mutex_lock(&ctx->uring_lock);
8638 if (io_wait_rsrc_data(ctx->buf_data))
8639 __io_sqe_buffers_unregister(ctx);
8640 if (io_wait_rsrc_data(ctx->file_data))
8641 __io_sqe_files_unregister(ctx);
8643 __io_cqring_overflow_flush(ctx, true);
8644 mutex_unlock(&ctx->uring_lock);
8645 io_eventfd_unregister(ctx);
8646 io_destroy_buffers(ctx);
8648 put_cred(ctx->sq_creds);
8650 /* there are no registered resources left, nobody uses it */
8652 io_rsrc_node_destroy(ctx->rsrc_node);
8653 if (ctx->rsrc_backup_node)
8654 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8655 flush_delayed_work(&ctx->rsrc_put_work);
8657 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8658 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8660 #if defined(CONFIG_UNIX)
8661 if (ctx->ring_sock) {
8662 ctx->ring_sock->file = NULL; /* so that iput() is called */
8663 sock_release(ctx->ring_sock);
8667 io_mem_free(ctx->rings);
8668 io_mem_free(ctx->sq_sqes);
8670 percpu_ref_exit(&ctx->refs);
8671 free_uid(ctx->user);
8672 io_req_caches_free(ctx);
8674 io_wq_put_hash(ctx->hash_map);
8675 kfree(ctx->cancel_hash);
8676 kfree(ctx->dummy_ubuf);
8680 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8682 struct io_ring_ctx *ctx = file->private_data;
8685 poll_wait(file, &ctx->cq_wait, wait);
8687 * synchronizes with barrier from wq_has_sleeper call in
8691 if (!io_sqring_full(ctx))
8692 mask |= EPOLLOUT | EPOLLWRNORM;
8695 * Don't flush cqring overflow list here, just do a simple check.
8696 * Otherwise there could possible be ABBA deadlock:
8699 * lock(&ctx->uring_lock);
8701 * lock(&ctx->uring_lock);
8704 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8705 * pushs them to do the flush.
8707 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8708 mask |= EPOLLIN | EPOLLRDNORM;
8713 static int io_uring_fasync(int fd, struct file *file, int on)
8715 struct io_ring_ctx *ctx = file->private_data;
8717 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8720 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8722 const struct cred *creds;
8724 creds = xa_erase(&ctx->personalities, id);
8733 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8735 return io_run_task_work_head(&ctx->exit_task_work);
8738 struct io_tctx_exit {
8739 struct callback_head task_work;
8740 struct completion completion;
8741 struct io_ring_ctx *ctx;
8744 static void io_tctx_exit_cb(struct callback_head *cb)
8746 struct io_uring_task *tctx = current->io_uring;
8747 struct io_tctx_exit *work;
8749 work = container_of(cb, struct io_tctx_exit, task_work);
8751 * When @in_idle, we're in cancellation and it's racy to remove the
8752 * node. It'll be removed by the end of cancellation, just ignore it.
8754 if (!atomic_read(&tctx->in_idle))
8755 io_uring_del_tctx_node((unsigned long)work->ctx);
8756 complete(&work->completion);
8759 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8761 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8763 return req->ctx == data;
8766 static void io_ring_exit_work(struct work_struct *work)
8768 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8769 unsigned long timeout = jiffies + HZ * 60 * 5;
8770 struct io_tctx_exit exit;
8771 struct io_tctx_node *node;
8775 * If we're doing polled IO and end up having requests being
8776 * submitted async (out-of-line), then completions can come in while
8777 * we're waiting for refs to drop. We need to reap these manually,
8778 * as nobody else will be looking for them.
8781 io_uring_try_cancel_requests(ctx, NULL, true);
8783 struct io_sq_data *sqd = ctx->sq_data;
8784 struct task_struct *tsk;
8786 io_sq_thread_park(sqd);
8788 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8789 io_wq_cancel_cb(tsk->io_uring->io_wq,
8790 io_cancel_ctx_cb, ctx, true);
8791 io_sq_thread_unpark(sqd);
8794 WARN_ON_ONCE(time_after(jiffies, timeout));
8795 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8797 init_completion(&exit.completion);
8798 init_task_work(&exit.task_work, io_tctx_exit_cb);
8801 * Some may use context even when all refs and requests have been put,
8802 * and they are free to do so while still holding uring_lock or
8803 * completion_lock, see __io_req_task_submit(). Apart from other work,
8804 * this lock/unlock section also waits them to finish.
8806 mutex_lock(&ctx->uring_lock);
8807 while (!list_empty(&ctx->tctx_list)) {
8808 WARN_ON_ONCE(time_after(jiffies, timeout));
8810 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8812 /* don't spin on a single task if cancellation failed */
8813 list_rotate_left(&ctx->tctx_list);
8814 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8815 if (WARN_ON_ONCE(ret))
8817 wake_up_process(node->task);
8819 mutex_unlock(&ctx->uring_lock);
8820 wait_for_completion(&exit.completion);
8821 mutex_lock(&ctx->uring_lock);
8823 mutex_unlock(&ctx->uring_lock);
8824 spin_lock_irq(&ctx->completion_lock);
8825 spin_unlock_irq(&ctx->completion_lock);
8827 io_ring_ctx_free(ctx);
8830 /* Returns true if we found and killed one or more timeouts */
8831 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8834 struct io_kiocb *req, *tmp;
8837 spin_lock_irq(&ctx->completion_lock);
8838 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8839 if (io_match_task(req, tsk, cancel_all)) {
8840 io_kill_timeout(req, -ECANCELED);
8845 io_commit_cqring(ctx);
8846 spin_unlock_irq(&ctx->completion_lock);
8848 io_cqring_ev_posted(ctx);
8849 return canceled != 0;
8852 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8854 unsigned long index;
8855 struct creds *creds;
8857 mutex_lock(&ctx->uring_lock);
8858 percpu_ref_kill(&ctx->refs);
8860 __io_cqring_overflow_flush(ctx, true);
8861 xa_for_each(&ctx->personalities, index, creds)
8862 io_unregister_personality(ctx, index);
8863 mutex_unlock(&ctx->uring_lock);
8865 io_kill_timeouts(ctx, NULL, true);
8866 io_poll_remove_all(ctx, NULL, true);
8868 /* if we failed setting up the ctx, we might not have any rings */
8869 io_iopoll_try_reap_events(ctx);
8871 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8873 * Use system_unbound_wq to avoid spawning tons of event kworkers
8874 * if we're exiting a ton of rings at the same time. It just adds
8875 * noise and overhead, there's no discernable change in runtime
8876 * over using system_wq.
8878 queue_work(system_unbound_wq, &ctx->exit_work);
8881 static int io_uring_release(struct inode *inode, struct file *file)
8883 struct io_ring_ctx *ctx = file->private_data;
8885 file->private_data = NULL;
8886 io_ring_ctx_wait_and_kill(ctx);
8890 struct io_task_cancel {
8891 struct task_struct *task;
8895 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8897 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8898 struct io_task_cancel *cancel = data;
8901 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8902 unsigned long flags;
8903 struct io_ring_ctx *ctx = req->ctx;
8905 /* protect against races with linked timeouts */
8906 spin_lock_irqsave(&ctx->completion_lock, flags);
8907 ret = io_match_task(req, cancel->task, cancel->all);
8908 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8910 ret = io_match_task(req, cancel->task, cancel->all);
8915 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8916 struct task_struct *task, bool cancel_all)
8918 struct io_defer_entry *de;
8921 spin_lock_irq(&ctx->completion_lock);
8922 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8923 if (io_match_task(de->req, task, cancel_all)) {
8924 list_cut_position(&list, &ctx->defer_list, &de->list);
8928 spin_unlock_irq(&ctx->completion_lock);
8929 if (list_empty(&list))
8932 while (!list_empty(&list)) {
8933 de = list_first_entry(&list, struct io_defer_entry, list);
8934 list_del_init(&de->list);
8935 io_req_complete_failed(de->req, -ECANCELED);
8941 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8943 struct io_tctx_node *node;
8944 enum io_wq_cancel cret;
8947 mutex_lock(&ctx->uring_lock);
8948 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8949 struct io_uring_task *tctx = node->task->io_uring;
8952 * io_wq will stay alive while we hold uring_lock, because it's
8953 * killed after ctx nodes, which requires to take the lock.
8955 if (!tctx || !tctx->io_wq)
8957 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8958 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8960 mutex_unlock(&ctx->uring_lock);
8965 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8966 struct task_struct *task,
8969 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8970 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8973 enum io_wq_cancel cret;
8977 ret |= io_uring_try_cancel_iowq(ctx);
8978 } else if (tctx && tctx->io_wq) {
8980 * Cancels requests of all rings, not only @ctx, but
8981 * it's fine as the task is in exit/exec.
8983 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8985 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8988 /* SQPOLL thread does its own polling */
8989 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
8990 (ctx->sq_data && ctx->sq_data->thread == current)) {
8991 while (!list_empty_careful(&ctx->iopoll_list)) {
8992 io_iopoll_try_reap_events(ctx);
8997 ret |= io_cancel_defer_files(ctx, task, cancel_all);
8998 ret |= io_poll_remove_all(ctx, task, cancel_all);
8999 ret |= io_kill_timeouts(ctx, task, cancel_all);
9000 ret |= io_run_task_work();
9001 ret |= io_run_ctx_fallback(ctx);
9008 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9010 struct io_uring_task *tctx = current->io_uring;
9011 struct io_tctx_node *node;
9014 if (unlikely(!tctx)) {
9015 ret = io_uring_alloc_task_context(current, ctx);
9018 tctx = current->io_uring;
9020 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9021 node = kmalloc(sizeof(*node), GFP_KERNEL);
9025 node->task = current;
9027 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9034 mutex_lock(&ctx->uring_lock);
9035 list_add(&node->ctx_node, &ctx->tctx_list);
9036 mutex_unlock(&ctx->uring_lock);
9043 * Note that this task has used io_uring. We use it for cancelation purposes.
9045 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9047 struct io_uring_task *tctx = current->io_uring;
9049 if (likely(tctx && tctx->last == ctx))
9051 return __io_uring_add_tctx_node(ctx);
9055 * Remove this io_uring_file -> task mapping.
9057 static void io_uring_del_tctx_node(unsigned long index)
9059 struct io_uring_task *tctx = current->io_uring;
9060 struct io_tctx_node *node;
9064 node = xa_erase(&tctx->xa, index);
9068 WARN_ON_ONCE(current != node->task);
9069 WARN_ON_ONCE(list_empty(&node->ctx_node));
9071 mutex_lock(&node->ctx->uring_lock);
9072 list_del(&node->ctx_node);
9073 mutex_unlock(&node->ctx->uring_lock);
9075 if (tctx->last == node->ctx)
9080 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9082 struct io_wq *wq = tctx->io_wq;
9083 struct io_tctx_node *node;
9084 unsigned long index;
9086 xa_for_each(&tctx->xa, index, node)
9087 io_uring_del_tctx_node(index);
9090 * Must be after io_uring_del_task_file() (removes nodes under
9091 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9094 io_wq_put_and_exit(wq);
9098 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9101 return atomic_read(&tctx->inflight_tracked);
9102 return percpu_counter_sum(&tctx->inflight);
9105 static void io_uring_drop_tctx_refs(struct task_struct *task)
9107 struct io_uring_task *tctx = task->io_uring;
9108 unsigned int refs = tctx->cached_refs;
9110 tctx->cached_refs = 0;
9111 percpu_counter_sub(&tctx->inflight, refs);
9112 put_task_struct_many(task, refs);
9116 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9117 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9119 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9121 struct io_uring_task *tctx = current->io_uring;
9122 struct io_ring_ctx *ctx;
9126 WARN_ON_ONCE(sqd && sqd->thread != current);
9128 if (!current->io_uring)
9131 io_wq_exit_start(tctx->io_wq);
9133 io_uring_drop_tctx_refs(current);
9134 atomic_inc(&tctx->in_idle);
9136 /* read completions before cancelations */
9137 inflight = tctx_inflight(tctx, !cancel_all);
9142 struct io_tctx_node *node;
9143 unsigned long index;
9145 xa_for_each(&tctx->xa, index, node) {
9146 /* sqpoll task will cancel all its requests */
9147 if (node->ctx->sq_data)
9149 io_uring_try_cancel_requests(node->ctx, current,
9153 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9154 io_uring_try_cancel_requests(ctx, current,
9158 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9160 * If we've seen completions, retry without waiting. This
9161 * avoids a race where a completion comes in before we did
9162 * prepare_to_wait().
9164 if (inflight == tctx_inflight(tctx, !cancel_all))
9166 finish_wait(&tctx->wait, &wait);
9168 atomic_dec(&tctx->in_idle);
9170 io_uring_clean_tctx(tctx);
9172 /* for exec all current's requests should be gone, kill tctx */
9173 __io_uring_free(current);
9177 void __io_uring_cancel(struct files_struct *files)
9179 io_uring_cancel_generic(!files, NULL);
9182 static void *io_uring_validate_mmap_request(struct file *file,
9183 loff_t pgoff, size_t sz)
9185 struct io_ring_ctx *ctx = file->private_data;
9186 loff_t offset = pgoff << PAGE_SHIFT;
9191 case IORING_OFF_SQ_RING:
9192 case IORING_OFF_CQ_RING:
9195 case IORING_OFF_SQES:
9199 return ERR_PTR(-EINVAL);
9202 page = virt_to_head_page(ptr);
9203 if (sz > page_size(page))
9204 return ERR_PTR(-EINVAL);
9211 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9213 size_t sz = vma->vm_end - vma->vm_start;
9217 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9219 return PTR_ERR(ptr);
9221 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9222 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9225 #else /* !CONFIG_MMU */
9227 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9229 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9232 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9234 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9237 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9238 unsigned long addr, unsigned long len,
9239 unsigned long pgoff, unsigned long flags)
9243 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9245 return PTR_ERR(ptr);
9247 return (unsigned long) ptr;
9250 #endif /* !CONFIG_MMU */
9252 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9257 if (!io_sqring_full(ctx))
9259 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9261 if (!io_sqring_full(ctx))
9264 } while (!signal_pending(current));
9266 finish_wait(&ctx->sqo_sq_wait, &wait);
9270 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9271 struct __kernel_timespec __user **ts,
9272 const sigset_t __user **sig)
9274 struct io_uring_getevents_arg arg;
9277 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9278 * is just a pointer to the sigset_t.
9280 if (!(flags & IORING_ENTER_EXT_ARG)) {
9281 *sig = (const sigset_t __user *) argp;
9287 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9288 * timespec and sigset_t pointers if good.
9290 if (*argsz != sizeof(arg))
9292 if (copy_from_user(&arg, argp, sizeof(arg)))
9294 *sig = u64_to_user_ptr(arg.sigmask);
9295 *argsz = arg.sigmask_sz;
9296 *ts = u64_to_user_ptr(arg.ts);
9300 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9301 u32, min_complete, u32, flags, const void __user *, argp,
9304 struct io_ring_ctx *ctx;
9311 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9312 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9316 if (unlikely(!f.file))
9320 if (unlikely(f.file->f_op != &io_uring_fops))
9324 ctx = f.file->private_data;
9325 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9329 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9333 * For SQ polling, the thread will do all submissions and completions.
9334 * Just return the requested submit count, and wake the thread if
9338 if (ctx->flags & IORING_SETUP_SQPOLL) {
9339 io_cqring_overflow_flush(ctx, false);
9342 if (unlikely(ctx->sq_data->thread == NULL)) {
9345 if (flags & IORING_ENTER_SQ_WAKEUP)
9346 wake_up(&ctx->sq_data->wait);
9347 if (flags & IORING_ENTER_SQ_WAIT) {
9348 ret = io_sqpoll_wait_sq(ctx);
9352 submitted = to_submit;
9353 } else if (to_submit) {
9354 ret = io_uring_add_tctx_node(ctx);
9357 mutex_lock(&ctx->uring_lock);
9358 submitted = io_submit_sqes(ctx, to_submit);
9359 mutex_unlock(&ctx->uring_lock);
9361 if (submitted != to_submit)
9364 if (flags & IORING_ENTER_GETEVENTS) {
9365 const sigset_t __user *sig;
9366 struct __kernel_timespec __user *ts;
9368 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9372 min_complete = min(min_complete, ctx->cq_entries);
9375 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9376 * space applications don't need to do io completion events
9377 * polling again, they can rely on io_sq_thread to do polling
9378 * work, which can reduce cpu usage and uring_lock contention.
9380 if (ctx->flags & IORING_SETUP_IOPOLL &&
9381 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9382 ret = io_iopoll_check(ctx, min_complete);
9384 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9389 percpu_ref_put(&ctx->refs);
9392 return submitted ? submitted : ret;
9395 #ifdef CONFIG_PROC_FS
9396 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9397 const struct cred *cred)
9399 struct user_namespace *uns = seq_user_ns(m);
9400 struct group_info *gi;
9405 seq_printf(m, "%5d\n", id);
9406 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9407 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9408 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9409 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9410 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9411 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9412 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9413 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9414 seq_puts(m, "\n\tGroups:\t");
9415 gi = cred->group_info;
9416 for (g = 0; g < gi->ngroups; g++) {
9417 seq_put_decimal_ull(m, g ? " " : "",
9418 from_kgid_munged(uns, gi->gid[g]));
9420 seq_puts(m, "\n\tCapEff:\t");
9421 cap = cred->cap_effective;
9422 CAP_FOR_EACH_U32(__capi)
9423 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9428 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9430 struct io_sq_data *sq = NULL;
9435 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9436 * since fdinfo case grabs it in the opposite direction of normal use
9437 * cases. If we fail to get the lock, we just don't iterate any
9438 * structures that could be going away outside the io_uring mutex.
9440 has_lock = mutex_trylock(&ctx->uring_lock);
9442 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9448 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9449 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9450 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9451 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9452 struct file *f = io_file_from_index(ctx, i);
9455 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9457 seq_printf(m, "%5u: <none>\n", i);
9459 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9460 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9461 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9462 unsigned int len = buf->ubuf_end - buf->ubuf;
9464 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9466 if (has_lock && !xa_empty(&ctx->personalities)) {
9467 unsigned long index;
9468 const struct cred *cred;
9470 seq_printf(m, "Personalities:\n");
9471 xa_for_each(&ctx->personalities, index, cred)
9472 io_uring_show_cred(m, index, cred);
9474 seq_printf(m, "PollList:\n");
9475 spin_lock_irq(&ctx->completion_lock);
9476 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9477 struct hlist_head *list = &ctx->cancel_hash[i];
9478 struct io_kiocb *req;
9480 hlist_for_each_entry(req, list, hash_node)
9481 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9482 req->task->task_works != NULL);
9484 spin_unlock_irq(&ctx->completion_lock);
9486 mutex_unlock(&ctx->uring_lock);
9489 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9491 struct io_ring_ctx *ctx = f->private_data;
9493 if (percpu_ref_tryget(&ctx->refs)) {
9494 __io_uring_show_fdinfo(ctx, m);
9495 percpu_ref_put(&ctx->refs);
9500 static const struct file_operations io_uring_fops = {
9501 .release = io_uring_release,
9502 .mmap = io_uring_mmap,
9504 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9505 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9507 .poll = io_uring_poll,
9508 .fasync = io_uring_fasync,
9509 #ifdef CONFIG_PROC_FS
9510 .show_fdinfo = io_uring_show_fdinfo,
9514 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9515 struct io_uring_params *p)
9517 struct io_rings *rings;
9518 size_t size, sq_array_offset;
9520 /* make sure these are sane, as we already accounted them */
9521 ctx->sq_entries = p->sq_entries;
9522 ctx->cq_entries = p->cq_entries;
9524 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9525 if (size == SIZE_MAX)
9528 rings = io_mem_alloc(size);
9533 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9534 rings->sq_ring_mask = p->sq_entries - 1;
9535 rings->cq_ring_mask = p->cq_entries - 1;
9536 rings->sq_ring_entries = p->sq_entries;
9537 rings->cq_ring_entries = p->cq_entries;
9539 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9540 if (size == SIZE_MAX) {
9541 io_mem_free(ctx->rings);
9546 ctx->sq_sqes = io_mem_alloc(size);
9547 if (!ctx->sq_sqes) {
9548 io_mem_free(ctx->rings);
9556 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9560 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9564 ret = io_uring_add_tctx_node(ctx);
9569 fd_install(fd, file);
9574 * Allocate an anonymous fd, this is what constitutes the application
9575 * visible backing of an io_uring instance. The application mmaps this
9576 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9577 * we have to tie this fd to a socket for file garbage collection purposes.
9579 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9582 #if defined(CONFIG_UNIX)
9585 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9588 return ERR_PTR(ret);
9591 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9592 O_RDWR | O_CLOEXEC);
9593 #if defined(CONFIG_UNIX)
9595 sock_release(ctx->ring_sock);
9596 ctx->ring_sock = NULL;
9598 ctx->ring_sock->file = file;
9604 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9605 struct io_uring_params __user *params)
9607 struct io_ring_ctx *ctx;
9613 if (entries > IORING_MAX_ENTRIES) {
9614 if (!(p->flags & IORING_SETUP_CLAMP))
9616 entries = IORING_MAX_ENTRIES;
9620 * Use twice as many entries for the CQ ring. It's possible for the
9621 * application to drive a higher depth than the size of the SQ ring,
9622 * since the sqes are only used at submission time. This allows for
9623 * some flexibility in overcommitting a bit. If the application has
9624 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9625 * of CQ ring entries manually.
9627 p->sq_entries = roundup_pow_of_two(entries);
9628 if (p->flags & IORING_SETUP_CQSIZE) {
9630 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9631 * to a power-of-two, if it isn't already. We do NOT impose
9632 * any cq vs sq ring sizing.
9636 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9637 if (!(p->flags & IORING_SETUP_CLAMP))
9639 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9641 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9642 if (p->cq_entries < p->sq_entries)
9645 p->cq_entries = 2 * p->sq_entries;
9648 ctx = io_ring_ctx_alloc(p);
9651 ctx->compat = in_compat_syscall();
9652 if (!capable(CAP_IPC_LOCK))
9653 ctx->user = get_uid(current_user());
9656 * This is just grabbed for accounting purposes. When a process exits,
9657 * the mm is exited and dropped before the files, hence we need to hang
9658 * on to this mm purely for the purposes of being able to unaccount
9659 * memory (locked/pinned vm). It's not used for anything else.
9661 mmgrab(current->mm);
9662 ctx->mm_account = current->mm;
9664 ret = io_allocate_scq_urings(ctx, p);
9668 ret = io_sq_offload_create(ctx, p);
9671 /* always set a rsrc node */
9672 ret = io_rsrc_node_switch_start(ctx);
9675 io_rsrc_node_switch(ctx, NULL);
9677 memset(&p->sq_off, 0, sizeof(p->sq_off));
9678 p->sq_off.head = offsetof(struct io_rings, sq.head);
9679 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9680 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9681 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9682 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9683 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9684 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9686 memset(&p->cq_off, 0, sizeof(p->cq_off));
9687 p->cq_off.head = offsetof(struct io_rings, cq.head);
9688 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9689 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9690 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9691 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9692 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9693 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9695 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9696 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9697 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9698 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9699 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9700 IORING_FEAT_RSRC_TAGS;
9702 if (copy_to_user(params, p, sizeof(*p))) {
9707 file = io_uring_get_file(ctx);
9709 ret = PTR_ERR(file);
9714 * Install ring fd as the very last thing, so we don't risk someone
9715 * having closed it before we finish setup
9717 ret = io_uring_install_fd(ctx, file);
9719 /* fput will clean it up */
9724 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9727 io_ring_ctx_wait_and_kill(ctx);
9732 * Sets up an aio uring context, and returns the fd. Applications asks for a
9733 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9734 * params structure passed in.
9736 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9738 struct io_uring_params p;
9741 if (copy_from_user(&p, params, sizeof(p)))
9743 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9748 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9749 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9750 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9751 IORING_SETUP_R_DISABLED))
9754 return io_uring_create(entries, &p, params);
9757 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9758 struct io_uring_params __user *, params)
9760 return io_uring_setup(entries, params);
9763 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9765 struct io_uring_probe *p;
9769 size = struct_size(p, ops, nr_args);
9770 if (size == SIZE_MAX)
9772 p = kzalloc(size, GFP_KERNEL);
9777 if (copy_from_user(p, arg, size))
9780 if (memchr_inv(p, 0, size))
9783 p->last_op = IORING_OP_LAST - 1;
9784 if (nr_args > IORING_OP_LAST)
9785 nr_args = IORING_OP_LAST;
9787 for (i = 0; i < nr_args; i++) {
9789 if (!io_op_defs[i].not_supported)
9790 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9795 if (copy_to_user(arg, p, size))
9802 static int io_register_personality(struct io_ring_ctx *ctx)
9804 const struct cred *creds;
9808 creds = get_current_cred();
9810 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9811 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9818 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9819 unsigned int nr_args)
9821 struct io_uring_restriction *res;
9825 /* Restrictions allowed only if rings started disabled */
9826 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9829 /* We allow only a single restrictions registration */
9830 if (ctx->restrictions.registered)
9833 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9836 size = array_size(nr_args, sizeof(*res));
9837 if (size == SIZE_MAX)
9840 res = memdup_user(arg, size);
9842 return PTR_ERR(res);
9846 for (i = 0; i < nr_args; i++) {
9847 switch (res[i].opcode) {
9848 case IORING_RESTRICTION_REGISTER_OP:
9849 if (res[i].register_op >= IORING_REGISTER_LAST) {
9854 __set_bit(res[i].register_op,
9855 ctx->restrictions.register_op);
9857 case IORING_RESTRICTION_SQE_OP:
9858 if (res[i].sqe_op >= IORING_OP_LAST) {
9863 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9865 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9866 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9868 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9869 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9878 /* Reset all restrictions if an error happened */
9880 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9882 ctx->restrictions.registered = true;
9888 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9890 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9893 if (ctx->restrictions.registered)
9894 ctx->restricted = 1;
9896 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9897 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9898 wake_up(&ctx->sq_data->wait);
9902 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9903 struct io_uring_rsrc_update2 *up,
9911 if (check_add_overflow(up->offset, nr_args, &tmp))
9913 err = io_rsrc_node_switch_start(ctx);
9918 case IORING_RSRC_FILE:
9919 return __io_sqe_files_update(ctx, up, nr_args);
9920 case IORING_RSRC_BUFFER:
9921 return __io_sqe_buffers_update(ctx, up, nr_args);
9926 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9929 struct io_uring_rsrc_update2 up;
9933 memset(&up, 0, sizeof(up));
9934 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9936 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9939 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9940 unsigned size, unsigned type)
9942 struct io_uring_rsrc_update2 up;
9944 if (size != sizeof(up))
9946 if (copy_from_user(&up, arg, sizeof(up)))
9948 if (!up.nr || up.resv)
9950 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9953 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9954 unsigned int size, unsigned int type)
9956 struct io_uring_rsrc_register rr;
9958 /* keep it extendible */
9959 if (size != sizeof(rr))
9962 memset(&rr, 0, sizeof(rr));
9963 if (copy_from_user(&rr, arg, size))
9965 if (!rr.nr || rr.resv || rr.resv2)
9969 case IORING_RSRC_FILE:
9970 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9971 rr.nr, u64_to_user_ptr(rr.tags));
9972 case IORING_RSRC_BUFFER:
9973 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9974 rr.nr, u64_to_user_ptr(rr.tags));
9979 static bool io_register_op_must_quiesce(int op)
9982 case IORING_REGISTER_BUFFERS:
9983 case IORING_UNREGISTER_BUFFERS:
9984 case IORING_REGISTER_FILES:
9985 case IORING_UNREGISTER_FILES:
9986 case IORING_REGISTER_FILES_UPDATE:
9987 case IORING_REGISTER_PROBE:
9988 case IORING_REGISTER_PERSONALITY:
9989 case IORING_UNREGISTER_PERSONALITY:
9990 case IORING_REGISTER_FILES2:
9991 case IORING_REGISTER_FILES_UPDATE2:
9992 case IORING_REGISTER_BUFFERS2:
9993 case IORING_REGISTER_BUFFERS_UPDATE:
10000 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10001 void __user *arg, unsigned nr_args)
10002 __releases(ctx->uring_lock)
10003 __acquires(ctx->uring_lock)
10008 * We're inside the ring mutex, if the ref is already dying, then
10009 * someone else killed the ctx or is already going through
10010 * io_uring_register().
10012 if (percpu_ref_is_dying(&ctx->refs))
10015 if (ctx->restricted) {
10016 if (opcode >= IORING_REGISTER_LAST)
10018 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10019 if (!test_bit(opcode, ctx->restrictions.register_op))
10023 if (io_register_op_must_quiesce(opcode)) {
10024 percpu_ref_kill(&ctx->refs);
10027 * Drop uring mutex before waiting for references to exit. If
10028 * another thread is currently inside io_uring_enter() it might
10029 * need to grab the uring_lock to make progress. If we hold it
10030 * here across the drain wait, then we can deadlock. It's safe
10031 * to drop the mutex here, since no new references will come in
10032 * after we've killed the percpu ref.
10034 mutex_unlock(&ctx->uring_lock);
10036 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10039 ret = io_run_task_work_sig();
10043 mutex_lock(&ctx->uring_lock);
10046 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10052 case IORING_REGISTER_BUFFERS:
10053 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10055 case IORING_UNREGISTER_BUFFERS:
10057 if (arg || nr_args)
10059 ret = io_sqe_buffers_unregister(ctx);
10061 case IORING_REGISTER_FILES:
10062 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10064 case IORING_UNREGISTER_FILES:
10066 if (arg || nr_args)
10068 ret = io_sqe_files_unregister(ctx);
10070 case IORING_REGISTER_FILES_UPDATE:
10071 ret = io_register_files_update(ctx, arg, nr_args);
10073 case IORING_REGISTER_EVENTFD:
10074 case IORING_REGISTER_EVENTFD_ASYNC:
10078 ret = io_eventfd_register(ctx, arg);
10081 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10082 ctx->eventfd_async = 1;
10084 ctx->eventfd_async = 0;
10086 case IORING_UNREGISTER_EVENTFD:
10088 if (arg || nr_args)
10090 ret = io_eventfd_unregister(ctx);
10092 case IORING_REGISTER_PROBE:
10094 if (!arg || nr_args > 256)
10096 ret = io_probe(ctx, arg, nr_args);
10098 case IORING_REGISTER_PERSONALITY:
10100 if (arg || nr_args)
10102 ret = io_register_personality(ctx);
10104 case IORING_UNREGISTER_PERSONALITY:
10108 ret = io_unregister_personality(ctx, nr_args);
10110 case IORING_REGISTER_ENABLE_RINGS:
10112 if (arg || nr_args)
10114 ret = io_register_enable_rings(ctx);
10116 case IORING_REGISTER_RESTRICTIONS:
10117 ret = io_register_restrictions(ctx, arg, nr_args);
10119 case IORING_REGISTER_FILES2:
10120 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10122 case IORING_REGISTER_FILES_UPDATE2:
10123 ret = io_register_rsrc_update(ctx, arg, nr_args,
10126 case IORING_REGISTER_BUFFERS2:
10127 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10129 case IORING_REGISTER_BUFFERS_UPDATE:
10130 ret = io_register_rsrc_update(ctx, arg, nr_args,
10131 IORING_RSRC_BUFFER);
10138 if (io_register_op_must_quiesce(opcode)) {
10139 /* bring the ctx back to life */
10140 percpu_ref_reinit(&ctx->refs);
10141 reinit_completion(&ctx->ref_comp);
10146 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10147 void __user *, arg, unsigned int, nr_args)
10149 struct io_ring_ctx *ctx;
10158 if (f.file->f_op != &io_uring_fops)
10161 ctx = f.file->private_data;
10163 io_run_task_work();
10165 mutex_lock(&ctx->uring_lock);
10166 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10167 mutex_unlock(&ctx->uring_lock);
10168 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10169 ctx->cq_ev_fd != NULL, ret);
10175 static int __init io_uring_init(void)
10177 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10178 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10179 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10182 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10183 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10184 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10185 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10186 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10187 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10188 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10189 BUILD_BUG_SQE_ELEM(8, __u64, off);
10190 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10191 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10192 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10193 BUILD_BUG_SQE_ELEM(24, __u32, len);
10194 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10195 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10196 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10197 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10198 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10199 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10200 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10201 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10202 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10203 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10204 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10205 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10206 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10207 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10208 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10209 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10210 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10211 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10212 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10214 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10215 sizeof(struct io_uring_rsrc_update));
10216 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10217 sizeof(struct io_uring_rsrc_update2));
10218 /* should fit into one byte */
10219 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10221 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10222 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10223 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10227 __initcall(io_uring_init);