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_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81 #include <linux/freezer.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
101 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
102 IORING_REGISTER_LAST + IORING_OP_LAST)
104 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
105 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 u32 head ____cacheline_aligned_in_smp;
110 u32 tail ____cacheline_aligned_in_smp;
114 * This data is shared with the application through the mmap at offsets
115 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
117 * The offsets to the member fields are published through struct
118 * io_sqring_offsets when calling io_uring_setup.
122 * Head and tail offsets into the ring; the offsets need to be
123 * masked to get valid indices.
125 * The kernel controls head of the sq ring and the tail of the cq ring,
126 * and the application controls tail of the sq ring and the head of the
129 struct io_uring sq, cq;
131 * Bitmasks to apply to head and tail offsets (constant, equals
134 u32 sq_ring_mask, cq_ring_mask;
135 /* Ring sizes (constant, power of 2) */
136 u32 sq_ring_entries, cq_ring_entries;
138 * Number of invalid entries dropped by the kernel due to
139 * invalid index stored in array
141 * Written by the kernel, shouldn't be modified by the
142 * application (i.e. get number of "new events" by comparing to
145 * After a new SQ head value was read by the application this
146 * counter includes all submissions that were dropped reaching
147 * the new SQ head (and possibly more).
153 * Written by the kernel, shouldn't be modified by the
156 * The application needs a full memory barrier before checking
157 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
163 * Written by the application, shouldn't be modified by the
168 * Number of completion events lost because the queue was full;
169 * this should be avoided by the application by making sure
170 * there are not more requests pending than there is space in
171 * the completion queue.
173 * Written by the kernel, shouldn't be modified by the
174 * application (i.e. get number of "new events" by comparing to
177 * As completion events come in out of order this counter is not
178 * ordered with any other data.
182 * Ring buffer of completion events.
184 * The kernel writes completion events fresh every time they are
185 * produced, so the application is allowed to modify pending
188 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
191 enum io_uring_cmd_flags {
192 IO_URING_F_NONBLOCK = 1,
193 IO_URING_F_COMPLETE_DEFER = 2,
196 struct io_mapped_ubuf {
199 struct bio_vec *bvec;
200 unsigned int nr_bvecs;
201 unsigned long acct_pages;
207 struct list_head list;
214 struct fixed_rsrc_table {
218 struct fixed_rsrc_ref_node {
219 struct percpu_ref refs;
220 struct list_head node;
221 struct list_head rsrc_list;
222 struct fixed_rsrc_data *rsrc_data;
223 void (*rsrc_put)(struct io_ring_ctx *ctx,
224 struct io_rsrc_put *prsrc);
225 struct llist_node llist;
229 struct fixed_rsrc_data {
230 struct fixed_rsrc_table *table;
231 struct io_ring_ctx *ctx;
233 struct fixed_rsrc_ref_node *node;
234 struct percpu_ref refs;
235 struct completion done;
240 struct list_head list;
246 struct io_restriction {
247 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
248 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
249 u8 sqe_flags_allowed;
250 u8 sqe_flags_required;
255 IO_SQ_THREAD_SHOULD_STOP = 0,
256 IO_SQ_THREAD_SHOULD_PARK,
261 atomic_t park_pending;
264 /* ctx's that are using this sqd */
265 struct list_head ctx_list;
267 struct task_struct *thread;
268 struct wait_queue_head wait;
270 unsigned sq_thread_idle;
276 struct completion exited;
277 struct callback_head *park_task_work;
280 #define IO_IOPOLL_BATCH 8
281 #define IO_COMPL_BATCH 32
282 #define IO_REQ_CACHE_SIZE 32
283 #define IO_REQ_ALLOC_BATCH 8
285 struct io_comp_state {
286 struct io_kiocb *reqs[IO_COMPL_BATCH];
288 unsigned int locked_free_nr;
289 /* inline/task_work completion list, under ->uring_lock */
290 struct list_head free_list;
291 /* IRQ completion list, under ->completion_lock */
292 struct list_head locked_free_list;
295 struct io_submit_link {
296 struct io_kiocb *head;
297 struct io_kiocb *last;
300 struct io_submit_state {
301 struct blk_plug plug;
302 struct io_submit_link link;
305 * io_kiocb alloc cache
307 void *reqs[IO_REQ_CACHE_SIZE];
308 unsigned int free_reqs;
313 * Batch completion logic
315 struct io_comp_state comp;
318 * File reference cache
322 unsigned int file_refs;
323 unsigned int ios_left;
328 struct percpu_ref refs;
329 } ____cacheline_aligned_in_smp;
333 unsigned int compat: 1;
334 unsigned int cq_overflow_flushed: 1;
335 unsigned int drain_next: 1;
336 unsigned int eventfd_async: 1;
337 unsigned int restricted: 1;
340 * Ring buffer of indices into array of io_uring_sqe, which is
341 * mmapped by the application using the IORING_OFF_SQES offset.
343 * This indirection could e.g. be used to assign fixed
344 * io_uring_sqe entries to operations and only submit them to
345 * the queue when needed.
347 * The kernel modifies neither the indices array nor the entries
351 unsigned cached_sq_head;
354 unsigned sq_thread_idle;
355 unsigned cached_sq_dropped;
356 unsigned cached_cq_overflow;
357 unsigned long sq_check_overflow;
359 /* hashed buffered write serialization */
360 struct io_wq_hash *hash_map;
362 struct list_head defer_list;
363 struct list_head timeout_list;
364 struct list_head cq_overflow_list;
366 struct io_uring_sqe *sq_sqes;
367 } ____cacheline_aligned_in_smp;
370 struct mutex uring_lock;
371 wait_queue_head_t wait;
372 } ____cacheline_aligned_in_smp;
374 struct io_submit_state submit_state;
376 struct io_rings *rings;
378 /* Only used for accounting purposes */
379 struct mm_struct *mm_account;
381 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
382 struct io_sq_data *sq_data; /* if using sq thread polling */
384 struct wait_queue_head sqo_sq_wait;
385 struct list_head sqd_list;
388 * If used, fixed file set. Writers must ensure that ->refs is dead,
389 * readers must ensure that ->refs is alive as long as the file* is
390 * used. Only updated through io_uring_register(2).
392 struct fixed_rsrc_data *file_data;
393 unsigned nr_user_files;
395 /* if used, fixed mapped user buffers */
396 unsigned nr_user_bufs;
397 struct io_mapped_ubuf *user_bufs;
399 struct user_struct *user;
401 struct completion ref_comp;
403 #if defined(CONFIG_UNIX)
404 struct socket *ring_sock;
407 struct xarray io_buffers;
409 struct xarray personalities;
413 unsigned cached_cq_tail;
416 atomic_t cq_timeouts;
417 unsigned cq_last_tm_flush;
418 unsigned long cq_check_overflow;
419 struct wait_queue_head cq_wait;
420 struct fasync_struct *cq_fasync;
421 struct eventfd_ctx *cq_ev_fd;
422 } ____cacheline_aligned_in_smp;
425 spinlock_t completion_lock;
428 * ->iopoll_list is protected by the ctx->uring_lock for
429 * io_uring instances that don't use IORING_SETUP_SQPOLL.
430 * For SQPOLL, only the single threaded io_sq_thread() will
431 * manipulate the list, hence no extra locking is needed there.
433 struct list_head iopoll_list;
434 struct hlist_head *cancel_hash;
435 unsigned cancel_hash_bits;
436 bool poll_multi_file;
438 spinlock_t inflight_lock;
439 struct list_head inflight_list;
440 } ____cacheline_aligned_in_smp;
442 struct delayed_work rsrc_put_work;
443 struct llist_head rsrc_put_llist;
444 struct list_head rsrc_ref_list;
445 spinlock_t rsrc_ref_lock;
447 struct io_restriction restrictions;
450 struct callback_head *exit_task_work;
452 struct wait_queue_head hash_wait;
454 /* Keep this last, we don't need it for the fast path */
455 struct work_struct exit_work;
456 struct list_head tctx_list;
459 struct io_uring_task {
460 /* submission side */
462 struct wait_queue_head wait;
463 const struct io_ring_ctx *last;
465 struct percpu_counter inflight;
469 spinlock_t task_lock;
470 struct io_wq_work_list task_list;
471 unsigned long task_state;
472 struct callback_head task_work;
476 * First field must be the file pointer in all the
477 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
479 struct io_poll_iocb {
481 struct wait_queue_head *head;
485 struct wait_queue_entry wait;
488 struct io_poll_remove {
498 struct io_timeout_data {
499 struct io_kiocb *req;
500 struct hrtimer timer;
501 struct timespec64 ts;
502 enum hrtimer_mode mode;
507 struct sockaddr __user *addr;
508 int __user *addr_len;
510 unsigned long nofile;
530 struct list_head list;
531 /* head of the link, used by linked timeouts only */
532 struct io_kiocb *head;
535 struct io_timeout_rem {
540 struct timespec64 ts;
545 /* NOTE: kiocb has the file as the first member, so don't do it here */
553 struct sockaddr __user *addr;
560 struct user_msghdr __user *umsg;
566 struct io_buffer *kbuf;
572 struct filename *filename;
574 unsigned long nofile;
577 struct io_rsrc_update {
603 struct epoll_event event;
607 struct file *file_out;
608 struct file *file_in;
615 struct io_provide_buf {
629 const char __user *filename;
630 struct statx __user *buffer;
642 struct filename *oldpath;
643 struct filename *newpath;
651 struct filename *filename;
654 struct io_completion {
656 struct list_head list;
660 struct io_async_connect {
661 struct sockaddr_storage address;
664 struct io_async_msghdr {
665 struct iovec fast_iov[UIO_FASTIOV];
666 /* points to an allocated iov, if NULL we use fast_iov instead */
667 struct iovec *free_iov;
668 struct sockaddr __user *uaddr;
670 struct sockaddr_storage addr;
674 struct iovec fast_iov[UIO_FASTIOV];
675 const struct iovec *free_iovec;
676 struct iov_iter iter;
678 struct wait_page_queue wpq;
682 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
683 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
684 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
685 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
686 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
687 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
693 REQ_F_LINK_TIMEOUT_BIT,
695 REQ_F_NEED_CLEANUP_BIT,
697 REQ_F_BUFFER_SELECTED_BIT,
698 REQ_F_NO_FILE_TABLE_BIT,
699 REQ_F_LTIMEOUT_ACTIVE_BIT,
700 REQ_F_COMPLETE_INLINE_BIT,
702 /* not a real bit, just to check we're not overflowing the space */
708 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
709 /* drain existing IO first */
710 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
712 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
713 /* doesn't sever on completion < 0 */
714 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
716 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
717 /* IOSQE_BUFFER_SELECT */
718 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
720 /* fail rest of links */
721 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
722 /* on inflight list, should be cancelled and waited on exit reliably */
723 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
724 /* read/write uses file position */
725 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
726 /* must not punt to workers */
727 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
728 /* has or had linked timeout */
729 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
731 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
733 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
734 /* already went through poll handler */
735 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
736 /* buffer already selected */
737 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
738 /* doesn't need file table for this request */
739 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
740 /* linked timeout is active, i.e. prepared by link's head */
741 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
742 /* completion is deferred through io_comp_state */
743 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
747 struct io_poll_iocb poll;
748 struct io_poll_iocb *double_poll;
751 struct io_task_work {
752 struct io_wq_work_node node;
753 task_work_func_t func;
757 * NOTE! Each of the iocb union members has the file pointer
758 * as the first entry in their struct definition. So you can
759 * access the file pointer through any of the sub-structs,
760 * or directly as just 'ki_filp' in this struct.
766 struct io_poll_iocb poll;
767 struct io_poll_remove poll_remove;
768 struct io_accept accept;
770 struct io_cancel cancel;
771 struct io_timeout timeout;
772 struct io_timeout_rem timeout_rem;
773 struct io_connect connect;
774 struct io_sr_msg sr_msg;
776 struct io_close close;
777 struct io_rsrc_update rsrc_update;
778 struct io_fadvise fadvise;
779 struct io_madvise madvise;
780 struct io_epoll epoll;
781 struct io_splice splice;
782 struct io_provide_buf pbuf;
783 struct io_statx statx;
784 struct io_shutdown shutdown;
785 struct io_rename rename;
786 struct io_unlink unlink;
787 /* use only after cleaning per-op data, see io_clean_op() */
788 struct io_completion compl;
791 /* opcode allocated if it needs to store data for async defer */
794 /* polled IO has completed */
800 struct io_ring_ctx *ctx;
803 struct task_struct *task;
806 struct io_kiocb *link;
807 struct percpu_ref *fixed_rsrc_refs;
810 * 1. used with ctx->iopoll_list with reads/writes
811 * 2. to track reqs with ->files (see io_op_def::file_table)
813 struct list_head inflight_entry;
815 struct io_task_work io_task_work;
816 struct callback_head task_work;
818 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
819 struct hlist_node hash_node;
820 struct async_poll *apoll;
821 struct io_wq_work work;
824 struct io_tctx_node {
825 struct list_head ctx_node;
826 struct task_struct *task;
827 struct io_ring_ctx *ctx;
830 struct io_defer_entry {
831 struct list_head list;
832 struct io_kiocb *req;
837 /* needs req->file assigned */
838 unsigned needs_file : 1;
839 /* hash wq insertion if file is a regular file */
840 unsigned hash_reg_file : 1;
841 /* unbound wq insertion if file is a non-regular file */
842 unsigned unbound_nonreg_file : 1;
843 /* opcode is not supported by this kernel */
844 unsigned not_supported : 1;
845 /* set if opcode supports polled "wait" */
847 unsigned pollout : 1;
848 /* op supports buffer selection */
849 unsigned buffer_select : 1;
850 /* must always have async data allocated */
851 unsigned needs_async_data : 1;
852 /* should block plug */
854 /* size of async data needed, if any */
855 unsigned short async_size;
858 static const struct io_op_def io_op_defs[] = {
859 [IORING_OP_NOP] = {},
860 [IORING_OP_READV] = {
862 .unbound_nonreg_file = 1,
865 .needs_async_data = 1,
867 .async_size = sizeof(struct io_async_rw),
869 [IORING_OP_WRITEV] = {
872 .unbound_nonreg_file = 1,
874 .needs_async_data = 1,
876 .async_size = sizeof(struct io_async_rw),
878 [IORING_OP_FSYNC] = {
881 [IORING_OP_READ_FIXED] = {
883 .unbound_nonreg_file = 1,
886 .async_size = sizeof(struct io_async_rw),
888 [IORING_OP_WRITE_FIXED] = {
891 .unbound_nonreg_file = 1,
894 .async_size = sizeof(struct io_async_rw),
896 [IORING_OP_POLL_ADD] = {
898 .unbound_nonreg_file = 1,
900 [IORING_OP_POLL_REMOVE] = {},
901 [IORING_OP_SYNC_FILE_RANGE] = {
904 [IORING_OP_SENDMSG] = {
906 .unbound_nonreg_file = 1,
908 .needs_async_data = 1,
909 .async_size = sizeof(struct io_async_msghdr),
911 [IORING_OP_RECVMSG] = {
913 .unbound_nonreg_file = 1,
916 .needs_async_data = 1,
917 .async_size = sizeof(struct io_async_msghdr),
919 [IORING_OP_TIMEOUT] = {
920 .needs_async_data = 1,
921 .async_size = sizeof(struct io_timeout_data),
923 [IORING_OP_TIMEOUT_REMOVE] = {
924 /* used by timeout updates' prep() */
926 [IORING_OP_ACCEPT] = {
928 .unbound_nonreg_file = 1,
931 [IORING_OP_ASYNC_CANCEL] = {},
932 [IORING_OP_LINK_TIMEOUT] = {
933 .needs_async_data = 1,
934 .async_size = sizeof(struct io_timeout_data),
936 [IORING_OP_CONNECT] = {
938 .unbound_nonreg_file = 1,
940 .needs_async_data = 1,
941 .async_size = sizeof(struct io_async_connect),
943 [IORING_OP_FALLOCATE] = {
946 [IORING_OP_OPENAT] = {},
947 [IORING_OP_CLOSE] = {},
948 [IORING_OP_FILES_UPDATE] = {},
949 [IORING_OP_STATX] = {},
952 .unbound_nonreg_file = 1,
956 .async_size = sizeof(struct io_async_rw),
958 [IORING_OP_WRITE] = {
960 .unbound_nonreg_file = 1,
963 .async_size = sizeof(struct io_async_rw),
965 [IORING_OP_FADVISE] = {
968 [IORING_OP_MADVISE] = {},
971 .unbound_nonreg_file = 1,
976 .unbound_nonreg_file = 1,
980 [IORING_OP_OPENAT2] = {
982 [IORING_OP_EPOLL_CTL] = {
983 .unbound_nonreg_file = 1,
985 [IORING_OP_SPLICE] = {
988 .unbound_nonreg_file = 1,
990 [IORING_OP_PROVIDE_BUFFERS] = {},
991 [IORING_OP_REMOVE_BUFFERS] = {},
995 .unbound_nonreg_file = 1,
997 [IORING_OP_SHUTDOWN] = {
1000 [IORING_OP_RENAMEAT] = {},
1001 [IORING_OP_UNLINKAT] = {},
1004 static bool io_disarm_next(struct io_kiocb *req);
1005 static void io_uring_del_task_file(unsigned long index);
1006 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1007 struct task_struct *task,
1008 struct files_struct *files);
1009 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
1010 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
1011 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1012 struct io_ring_ctx *ctx);
1013 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
1015 static bool io_rw_reissue(struct io_kiocb *req);
1016 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1017 static void io_put_req(struct io_kiocb *req);
1018 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1019 static void io_double_put_req(struct io_kiocb *req);
1020 static void io_dismantle_req(struct io_kiocb *req);
1021 static void io_put_task(struct task_struct *task, int nr);
1022 static void io_queue_next(struct io_kiocb *req);
1023 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1024 static void __io_queue_linked_timeout(struct io_kiocb *req);
1025 static void io_queue_linked_timeout(struct io_kiocb *req);
1026 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1027 struct io_uring_rsrc_update *ip,
1029 static void __io_clean_op(struct io_kiocb *req);
1030 static struct file *io_file_get(struct io_submit_state *state,
1031 struct io_kiocb *req, int fd, bool fixed);
1032 static void __io_queue_sqe(struct io_kiocb *req);
1033 static void io_rsrc_put_work(struct work_struct *work);
1035 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1036 struct iov_iter *iter, bool needs_lock);
1037 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1038 const struct iovec *fast_iov,
1039 struct iov_iter *iter, bool force);
1040 static void io_req_task_queue(struct io_kiocb *req);
1041 static void io_submit_flush_completions(struct io_comp_state *cs,
1042 struct io_ring_ctx *ctx);
1044 static struct kmem_cache *req_cachep;
1046 static const struct file_operations io_uring_fops;
1048 struct sock *io_uring_get_socket(struct file *file)
1050 #if defined(CONFIG_UNIX)
1051 if (file->f_op == &io_uring_fops) {
1052 struct io_ring_ctx *ctx = file->private_data;
1054 return ctx->ring_sock->sk;
1059 EXPORT_SYMBOL(io_uring_get_socket);
1061 #define io_for_each_link(pos, head) \
1062 for (pos = (head); pos; pos = pos->link)
1064 static inline void io_clean_op(struct io_kiocb *req)
1066 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1070 static inline void io_set_resource_node(struct io_kiocb *req)
1072 struct io_ring_ctx *ctx = req->ctx;
1074 if (!req->fixed_rsrc_refs) {
1075 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1076 percpu_ref_get(req->fixed_rsrc_refs);
1080 static bool io_match_task(struct io_kiocb *head,
1081 struct task_struct *task,
1082 struct files_struct *files)
1084 struct io_kiocb *req;
1086 if (task && head->task != task) {
1087 /* in terms of cancelation, always match if req task is dead */
1088 if (head->task->flags & PF_EXITING)
1095 io_for_each_link(req, head) {
1096 if (req->flags & REQ_F_INFLIGHT)
1098 if (req->task->files == files)
1104 static inline void req_set_fail_links(struct io_kiocb *req)
1106 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1107 req->flags |= REQ_F_FAIL_LINK;
1110 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1112 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1114 complete(&ctx->ref_comp);
1117 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1119 return !req->timeout.off;
1122 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1124 struct io_ring_ctx *ctx;
1127 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1132 * Use 5 bits less than the max cq entries, that should give us around
1133 * 32 entries per hash list if totally full and uniformly spread.
1135 hash_bits = ilog2(p->cq_entries);
1139 ctx->cancel_hash_bits = hash_bits;
1140 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1142 if (!ctx->cancel_hash)
1144 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1146 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1147 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1150 ctx->flags = p->flags;
1151 init_waitqueue_head(&ctx->sqo_sq_wait);
1152 INIT_LIST_HEAD(&ctx->sqd_list);
1153 init_waitqueue_head(&ctx->cq_wait);
1154 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1155 init_completion(&ctx->ref_comp);
1156 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1157 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1158 mutex_init(&ctx->uring_lock);
1159 init_waitqueue_head(&ctx->wait);
1160 spin_lock_init(&ctx->completion_lock);
1161 INIT_LIST_HEAD(&ctx->iopoll_list);
1162 INIT_LIST_HEAD(&ctx->defer_list);
1163 INIT_LIST_HEAD(&ctx->timeout_list);
1164 spin_lock_init(&ctx->inflight_lock);
1165 INIT_LIST_HEAD(&ctx->inflight_list);
1166 spin_lock_init(&ctx->rsrc_ref_lock);
1167 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1168 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1169 init_llist_head(&ctx->rsrc_put_llist);
1170 INIT_LIST_HEAD(&ctx->tctx_list);
1171 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1172 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1175 kfree(ctx->cancel_hash);
1180 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1182 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1183 struct io_ring_ctx *ctx = req->ctx;
1185 return seq != ctx->cached_cq_tail
1186 + READ_ONCE(ctx->cached_cq_overflow);
1192 static void io_req_track_inflight(struct io_kiocb *req)
1194 struct io_ring_ctx *ctx = req->ctx;
1196 if (!(req->flags & REQ_F_INFLIGHT)) {
1197 req->flags |= REQ_F_INFLIGHT;
1199 spin_lock_irq(&ctx->inflight_lock);
1200 list_add(&req->inflight_entry, &ctx->inflight_list);
1201 spin_unlock_irq(&ctx->inflight_lock);
1205 static void io_prep_async_work(struct io_kiocb *req)
1207 const struct io_op_def *def = &io_op_defs[req->opcode];
1208 struct io_ring_ctx *ctx = req->ctx;
1210 if (!req->work.creds)
1211 req->work.creds = get_current_cred();
1213 if (req->flags & REQ_F_FORCE_ASYNC)
1214 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1216 if (req->flags & REQ_F_ISREG) {
1217 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1218 io_wq_hash_work(&req->work, file_inode(req->file));
1220 if (def->unbound_nonreg_file)
1221 req->work.flags |= IO_WQ_WORK_UNBOUND;
1225 static void io_prep_async_link(struct io_kiocb *req)
1227 struct io_kiocb *cur;
1229 io_for_each_link(cur, req)
1230 io_prep_async_work(cur);
1233 static void io_queue_async_work(struct io_kiocb *req)
1235 struct io_ring_ctx *ctx = req->ctx;
1236 struct io_kiocb *link = io_prep_linked_timeout(req);
1237 struct io_uring_task *tctx = req->task->io_uring;
1240 BUG_ON(!tctx->io_wq);
1242 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1243 &req->work, req->flags);
1244 /* init ->work of the whole link before punting */
1245 io_prep_async_link(req);
1246 io_wq_enqueue(tctx->io_wq, &req->work);
1248 io_queue_linked_timeout(link);
1251 static void io_kill_timeout(struct io_kiocb *req)
1253 struct io_timeout_data *io = req->async_data;
1256 ret = hrtimer_try_to_cancel(&io->timer);
1258 atomic_set(&req->ctx->cq_timeouts,
1259 atomic_read(&req->ctx->cq_timeouts) + 1);
1260 list_del_init(&req->timeout.list);
1261 io_cqring_fill_event(req, 0);
1262 io_put_req_deferred(req, 1);
1267 * Returns true if we found and killed one or more timeouts
1269 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1270 struct files_struct *files)
1272 struct io_kiocb *req, *tmp;
1275 spin_lock_irq(&ctx->completion_lock);
1276 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1277 if (io_match_task(req, tsk, files)) {
1278 io_kill_timeout(req);
1282 spin_unlock_irq(&ctx->completion_lock);
1283 return canceled != 0;
1286 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1289 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1290 struct io_defer_entry, list);
1292 if (req_need_defer(de->req, de->seq))
1294 list_del_init(&de->list);
1295 io_req_task_queue(de->req);
1297 } while (!list_empty(&ctx->defer_list));
1300 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1304 if (list_empty(&ctx->timeout_list))
1307 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1310 u32 events_needed, events_got;
1311 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1312 struct io_kiocb, timeout.list);
1314 if (io_is_timeout_noseq(req))
1318 * Since seq can easily wrap around over time, subtract
1319 * the last seq at which timeouts were flushed before comparing.
1320 * Assuming not more than 2^31-1 events have happened since,
1321 * these subtractions won't have wrapped, so we can check if
1322 * target is in [last_seq, current_seq] by comparing the two.
1324 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1325 events_got = seq - ctx->cq_last_tm_flush;
1326 if (events_got < events_needed)
1329 list_del_init(&req->timeout.list);
1330 io_kill_timeout(req);
1331 } while (!list_empty(&ctx->timeout_list));
1333 ctx->cq_last_tm_flush = seq;
1336 static void io_commit_cqring(struct io_ring_ctx *ctx)
1338 io_flush_timeouts(ctx);
1340 /* order cqe stores with ring update */
1341 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1343 if (unlikely(!list_empty(&ctx->defer_list)))
1344 __io_queue_deferred(ctx);
1347 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1349 struct io_rings *r = ctx->rings;
1351 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1354 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1356 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1359 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1361 struct io_rings *rings = ctx->rings;
1365 * writes to the cq entry need to come after reading head; the
1366 * control dependency is enough as we're using WRITE_ONCE to
1369 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1372 tail = ctx->cached_cq_tail++;
1373 return &rings->cqes[tail & ctx->cq_mask];
1376 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1380 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1382 if (!ctx->eventfd_async)
1384 return io_wq_current_is_worker();
1387 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1389 /* see waitqueue_active() comment */
1392 if (waitqueue_active(&ctx->wait))
1393 wake_up(&ctx->wait);
1394 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1395 wake_up(&ctx->sq_data->wait);
1396 if (io_should_trigger_evfd(ctx))
1397 eventfd_signal(ctx->cq_ev_fd, 1);
1398 if (waitqueue_active(&ctx->cq_wait)) {
1399 wake_up_interruptible(&ctx->cq_wait);
1400 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1404 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1406 /* see waitqueue_active() comment */
1409 if (ctx->flags & IORING_SETUP_SQPOLL) {
1410 if (waitqueue_active(&ctx->wait))
1411 wake_up(&ctx->wait);
1413 if (io_should_trigger_evfd(ctx))
1414 eventfd_signal(ctx->cq_ev_fd, 1);
1415 if (waitqueue_active(&ctx->cq_wait)) {
1416 wake_up_interruptible(&ctx->cq_wait);
1417 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1421 /* Returns true if there are no backlogged entries after the flush */
1422 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1423 struct task_struct *tsk,
1424 struct files_struct *files)
1426 struct io_rings *rings = ctx->rings;
1427 struct io_kiocb *req, *tmp;
1428 struct io_uring_cqe *cqe;
1429 unsigned long flags;
1430 bool all_flushed, posted;
1433 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1437 spin_lock_irqsave(&ctx->completion_lock, flags);
1438 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1439 if (!io_match_task(req, tsk, files))
1442 cqe = io_get_cqring(ctx);
1446 list_move(&req->compl.list, &list);
1448 WRITE_ONCE(cqe->user_data, req->user_data);
1449 WRITE_ONCE(cqe->res, req->result);
1450 WRITE_ONCE(cqe->flags, req->compl.cflags);
1452 ctx->cached_cq_overflow++;
1453 WRITE_ONCE(ctx->rings->cq_overflow,
1454 ctx->cached_cq_overflow);
1459 all_flushed = list_empty(&ctx->cq_overflow_list);
1461 clear_bit(0, &ctx->sq_check_overflow);
1462 clear_bit(0, &ctx->cq_check_overflow);
1463 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1467 io_commit_cqring(ctx);
1468 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1470 io_cqring_ev_posted(ctx);
1472 while (!list_empty(&list)) {
1473 req = list_first_entry(&list, struct io_kiocb, compl.list);
1474 list_del(&req->compl.list);
1481 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1482 struct task_struct *tsk,
1483 struct files_struct *files)
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, tsk, files);
1492 if (ctx->flags & IORING_SETUP_IOPOLL)
1493 mutex_unlock(&ctx->uring_lock);
1499 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1501 struct io_ring_ctx *ctx = req->ctx;
1502 struct io_uring_cqe *cqe;
1504 trace_io_uring_complete(ctx, req->user_data, res);
1507 * If we can't get a cq entry, userspace overflowed the
1508 * submission (by quite a lot). Increment the overflow count in
1511 cqe = io_get_cqring(ctx);
1513 WRITE_ONCE(cqe->user_data, req->user_data);
1514 WRITE_ONCE(cqe->res, res);
1515 WRITE_ONCE(cqe->flags, cflags);
1516 } else if (ctx->cq_overflow_flushed ||
1517 atomic_read(&req->task->io_uring->in_idle)) {
1519 * If we're in ring overflow flush mode, or in task cancel mode,
1520 * then we cannot store the request for later flushing, we need
1521 * to drop it on the floor.
1523 ctx->cached_cq_overflow++;
1524 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1526 if (list_empty(&ctx->cq_overflow_list)) {
1527 set_bit(0, &ctx->sq_check_overflow);
1528 set_bit(0, &ctx->cq_check_overflow);
1529 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1533 req->compl.cflags = cflags;
1534 refcount_inc(&req->refs);
1535 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1539 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1541 __io_cqring_fill_event(req, res, 0);
1544 static void io_req_complete_post(struct io_kiocb *req, long res,
1545 unsigned int cflags)
1547 struct io_ring_ctx *ctx = req->ctx;
1548 unsigned long flags;
1550 spin_lock_irqsave(&ctx->completion_lock, flags);
1551 __io_cqring_fill_event(req, res, cflags);
1553 * If we're the last reference to this request, add to our locked
1556 if (refcount_dec_and_test(&req->refs)) {
1557 struct io_comp_state *cs = &ctx->submit_state.comp;
1559 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1560 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1561 io_disarm_next(req);
1563 io_req_task_queue(req->link);
1567 io_dismantle_req(req);
1568 io_put_task(req->task, 1);
1569 list_add(&req->compl.list, &cs->locked_free_list);
1570 cs->locked_free_nr++;
1572 if (!percpu_ref_tryget(&ctx->refs))
1575 io_commit_cqring(ctx);
1576 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1579 io_cqring_ev_posted(ctx);
1580 percpu_ref_put(&ctx->refs);
1584 static void io_req_complete_state(struct io_kiocb *req, long res,
1585 unsigned int cflags)
1589 req->compl.cflags = cflags;
1590 req->flags |= REQ_F_COMPLETE_INLINE;
1593 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1594 long res, unsigned cflags)
1596 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1597 io_req_complete_state(req, res, cflags);
1599 io_req_complete_post(req, res, cflags);
1602 static inline void io_req_complete(struct io_kiocb *req, long res)
1604 __io_req_complete(req, 0, res, 0);
1607 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1609 struct io_submit_state *state = &ctx->submit_state;
1610 struct io_comp_state *cs = &state->comp;
1611 struct io_kiocb *req = NULL;
1614 * If we have more than a batch's worth of requests in our IRQ side
1615 * locked cache, grab the lock and move them over to our submission
1618 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1619 spin_lock_irq(&ctx->completion_lock);
1620 list_splice_init(&cs->locked_free_list, &cs->free_list);
1621 cs->locked_free_nr = 0;
1622 spin_unlock_irq(&ctx->completion_lock);
1625 while (!list_empty(&cs->free_list)) {
1626 req = list_first_entry(&cs->free_list, struct io_kiocb,
1628 list_del(&req->compl.list);
1629 state->reqs[state->free_reqs++] = req;
1630 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1637 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1639 struct io_submit_state *state = &ctx->submit_state;
1641 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1643 if (!state->free_reqs) {
1644 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1647 if (io_flush_cached_reqs(ctx))
1650 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1654 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1655 * retry single alloc to be on the safe side.
1657 if (unlikely(ret <= 0)) {
1658 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1659 if (!state->reqs[0])
1663 state->free_reqs = ret;
1667 return state->reqs[state->free_reqs];
1670 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1677 static void io_dismantle_req(struct io_kiocb *req)
1681 if (req->async_data)
1682 kfree(req->async_data);
1684 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1685 if (req->fixed_rsrc_refs)
1686 percpu_ref_put(req->fixed_rsrc_refs);
1687 if (req->work.creds) {
1688 put_cred(req->work.creds);
1689 req->work.creds = NULL;
1692 if (req->flags & REQ_F_INFLIGHT) {
1693 struct io_ring_ctx *ctx = req->ctx;
1694 unsigned long flags;
1696 spin_lock_irqsave(&ctx->inflight_lock, flags);
1697 list_del(&req->inflight_entry);
1698 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1699 req->flags &= ~REQ_F_INFLIGHT;
1703 /* must to be called somewhat shortly after putting a request */
1704 static inline void io_put_task(struct task_struct *task, int nr)
1706 struct io_uring_task *tctx = task->io_uring;
1708 percpu_counter_sub(&tctx->inflight, nr);
1709 if (unlikely(atomic_read(&tctx->in_idle)))
1710 wake_up(&tctx->wait);
1711 put_task_struct_many(task, nr);
1714 static void __io_free_req(struct io_kiocb *req)
1716 struct io_ring_ctx *ctx = req->ctx;
1718 io_dismantle_req(req);
1719 io_put_task(req->task, 1);
1721 kmem_cache_free(req_cachep, req);
1722 percpu_ref_put(&ctx->refs);
1725 static inline void io_remove_next_linked(struct io_kiocb *req)
1727 struct io_kiocb *nxt = req->link;
1729 req->link = nxt->link;
1733 static bool io_kill_linked_timeout(struct io_kiocb *req)
1734 __must_hold(&req->ctx->completion_lock)
1736 struct io_kiocb *link = req->link;
1737 bool cancelled = false;
1740 * Can happen if a linked timeout fired and link had been like
1741 * req -> link t-out -> link t-out [-> ...]
1743 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1744 struct io_timeout_data *io = link->async_data;
1747 io_remove_next_linked(req);
1748 link->timeout.head = NULL;
1749 ret = hrtimer_try_to_cancel(&io->timer);
1751 io_cqring_fill_event(link, -ECANCELED);
1752 io_put_req_deferred(link, 1);
1756 req->flags &= ~REQ_F_LINK_TIMEOUT;
1760 static void io_fail_links(struct io_kiocb *req)
1761 __must_hold(&req->ctx->completion_lock)
1763 struct io_kiocb *nxt, *link = req->link;
1770 trace_io_uring_fail_link(req, link);
1771 io_cqring_fill_event(link, -ECANCELED);
1772 io_put_req_deferred(link, 2);
1777 static bool io_disarm_next(struct io_kiocb *req)
1778 __must_hold(&req->ctx->completion_lock)
1780 bool posted = false;
1782 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1783 posted = io_kill_linked_timeout(req);
1784 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1785 posted |= (req->link != NULL);
1791 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1793 struct io_kiocb *nxt;
1796 * If LINK is set, we have dependent requests in this chain. If we
1797 * didn't fail this request, queue the first one up, moving any other
1798 * dependencies to the next request. In case of failure, fail the rest
1801 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1802 struct io_ring_ctx *ctx = req->ctx;
1803 unsigned long flags;
1806 spin_lock_irqsave(&ctx->completion_lock, flags);
1807 posted = io_disarm_next(req);
1809 io_commit_cqring(req->ctx);
1810 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1812 io_cqring_ev_posted(ctx);
1819 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1821 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1823 return __io_req_find_next(req);
1826 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1830 if (ctx->submit_state.comp.nr) {
1831 mutex_lock(&ctx->uring_lock);
1832 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1833 mutex_unlock(&ctx->uring_lock);
1835 percpu_ref_put(&ctx->refs);
1838 static bool __tctx_task_work(struct io_uring_task *tctx)
1840 struct io_ring_ctx *ctx = NULL;
1841 struct io_wq_work_list list;
1842 struct io_wq_work_node *node;
1844 if (wq_list_empty(&tctx->task_list))
1847 spin_lock_irq(&tctx->task_lock);
1848 list = tctx->task_list;
1849 INIT_WQ_LIST(&tctx->task_list);
1850 spin_unlock_irq(&tctx->task_lock);
1854 struct io_wq_work_node *next = node->next;
1855 struct io_kiocb *req;
1857 req = container_of(node, struct io_kiocb, io_task_work.node);
1858 if (req->ctx != ctx) {
1859 ctx_flush_and_put(ctx);
1861 percpu_ref_get(&ctx->refs);
1864 req->task_work.func(&req->task_work);
1868 ctx_flush_and_put(ctx);
1869 return list.first != NULL;
1872 static void tctx_task_work(struct callback_head *cb)
1874 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1876 clear_bit(0, &tctx->task_state);
1878 while (__tctx_task_work(tctx))
1882 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1883 enum task_work_notify_mode notify)
1885 struct io_uring_task *tctx = tsk->io_uring;
1886 struct io_wq_work_node *node, *prev;
1887 unsigned long flags;
1890 WARN_ON_ONCE(!tctx);
1892 spin_lock_irqsave(&tctx->task_lock, flags);
1893 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1894 spin_unlock_irqrestore(&tctx->task_lock, flags);
1896 /* task_work already pending, we're done */
1897 if (test_bit(0, &tctx->task_state) ||
1898 test_and_set_bit(0, &tctx->task_state))
1901 if (!task_work_add(tsk, &tctx->task_work, notify))
1905 * Slow path - we failed, find and delete work. if the work is not
1906 * in the list, it got run and we're fine.
1909 spin_lock_irqsave(&tctx->task_lock, flags);
1910 wq_list_for_each(node, prev, &tctx->task_list) {
1911 if (&req->io_task_work.node == node) {
1912 wq_list_del(&tctx->task_list, node, prev);
1917 spin_unlock_irqrestore(&tctx->task_lock, flags);
1918 clear_bit(0, &tctx->task_state);
1922 static int io_req_task_work_add(struct io_kiocb *req)
1924 struct task_struct *tsk = req->task;
1925 struct io_ring_ctx *ctx = req->ctx;
1926 enum task_work_notify_mode notify;
1929 if (tsk->flags & PF_EXITING)
1933 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1934 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1935 * processing task_work. There's no reliable way to tell if TWA_RESUME
1939 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1940 notify = TWA_SIGNAL;
1942 ret = io_task_work_add(tsk, req, notify);
1944 wake_up_process(tsk);
1949 static bool io_run_task_work_head(struct callback_head **work_head)
1951 struct callback_head *work, *next;
1952 bool executed = false;
1955 work = xchg(work_head, NULL);
1971 static void io_task_work_add_head(struct callback_head **work_head,
1972 struct callback_head *task_work)
1974 struct callback_head *head;
1977 head = READ_ONCE(*work_head);
1978 task_work->next = head;
1979 } while (cmpxchg(work_head, head, task_work) != head);
1982 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1983 task_work_func_t cb)
1985 init_task_work(&req->task_work, cb);
1986 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
1989 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1991 struct io_ring_ctx *ctx = req->ctx;
1993 spin_lock_irq(&ctx->completion_lock);
1994 io_cqring_fill_event(req, error);
1995 io_commit_cqring(ctx);
1996 spin_unlock_irq(&ctx->completion_lock);
1998 io_cqring_ev_posted(ctx);
1999 req_set_fail_links(req);
2000 io_double_put_req(req);
2003 static void io_req_task_cancel(struct callback_head *cb)
2005 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2006 struct io_ring_ctx *ctx = req->ctx;
2008 mutex_lock(&ctx->uring_lock);
2009 __io_req_task_cancel(req, req->result);
2010 mutex_unlock(&ctx->uring_lock);
2011 percpu_ref_put(&ctx->refs);
2014 static void __io_req_task_submit(struct io_kiocb *req)
2016 struct io_ring_ctx *ctx = req->ctx;
2018 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2019 mutex_lock(&ctx->uring_lock);
2020 if (!(current->flags & PF_EXITING) && !current->in_execve)
2021 __io_queue_sqe(req);
2023 __io_req_task_cancel(req, -EFAULT);
2024 mutex_unlock(&ctx->uring_lock);
2027 static void io_req_task_submit(struct callback_head *cb)
2029 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2031 __io_req_task_submit(req);
2034 static void io_req_task_queue(struct io_kiocb *req)
2038 req->task_work.func = io_req_task_submit;
2039 ret = io_req_task_work_add(req);
2040 if (unlikely(ret)) {
2041 req->result = -ECANCELED;
2042 percpu_ref_get(&req->ctx->refs);
2043 io_req_task_work_add_fallback(req, io_req_task_cancel);
2047 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2049 percpu_ref_get(&req->ctx->refs);
2051 req->task_work.func = io_req_task_cancel;
2053 if (unlikely(io_req_task_work_add(req)))
2054 io_req_task_work_add_fallback(req, io_req_task_cancel);
2057 static inline void io_queue_next(struct io_kiocb *req)
2059 struct io_kiocb *nxt = io_req_find_next(req);
2062 io_req_task_queue(nxt);
2065 static void io_free_req(struct io_kiocb *req)
2072 struct task_struct *task;
2077 static inline void io_init_req_batch(struct req_batch *rb)
2084 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2085 struct req_batch *rb)
2088 io_put_task(rb->task, rb->task_refs);
2090 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2093 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2094 struct io_submit_state *state)
2098 if (req->task != rb->task) {
2100 io_put_task(rb->task, rb->task_refs);
2101 rb->task = req->task;
2107 io_dismantle_req(req);
2108 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2109 state->reqs[state->free_reqs++] = req;
2111 list_add(&req->compl.list, &state->comp.free_list);
2114 static void io_submit_flush_completions(struct io_comp_state *cs,
2115 struct io_ring_ctx *ctx)
2118 struct io_kiocb *req;
2119 struct req_batch rb;
2121 io_init_req_batch(&rb);
2122 spin_lock_irq(&ctx->completion_lock);
2123 for (i = 0; i < nr; i++) {
2125 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2127 io_commit_cqring(ctx);
2128 spin_unlock_irq(&ctx->completion_lock);
2130 io_cqring_ev_posted(ctx);
2131 for (i = 0; i < nr; i++) {
2134 /* submission and completion refs */
2135 if (refcount_sub_and_test(2, &req->refs))
2136 io_req_free_batch(&rb, req, &ctx->submit_state);
2139 io_req_free_batch_finish(ctx, &rb);
2144 * Drop reference to request, return next in chain (if there is one) if this
2145 * was the last reference to this request.
2147 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2149 struct io_kiocb *nxt = NULL;
2151 if (refcount_dec_and_test(&req->refs)) {
2152 nxt = io_req_find_next(req);
2158 static void io_put_req(struct io_kiocb *req)
2160 if (refcount_dec_and_test(&req->refs))
2164 static void io_put_req_deferred_cb(struct callback_head *cb)
2166 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2171 static void io_free_req_deferred(struct io_kiocb *req)
2175 req->task_work.func = io_put_req_deferred_cb;
2176 ret = io_req_task_work_add(req);
2178 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2181 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2183 if (refcount_sub_and_test(refs, &req->refs))
2184 io_free_req_deferred(req);
2187 static void io_double_put_req(struct io_kiocb *req)
2189 /* drop both submit and complete references */
2190 if (refcount_sub_and_test(2, &req->refs))
2194 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2196 /* See comment at the top of this file */
2198 return __io_cqring_events(ctx);
2201 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2203 struct io_rings *rings = ctx->rings;
2205 /* make sure SQ entry isn't read before tail */
2206 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2209 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2211 unsigned int cflags;
2213 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2214 cflags |= IORING_CQE_F_BUFFER;
2215 req->flags &= ~REQ_F_BUFFER_SELECTED;
2220 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2222 struct io_buffer *kbuf;
2224 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2225 return io_put_kbuf(req, kbuf);
2228 static inline bool io_run_task_work(void)
2231 * Not safe to run on exiting task, and the task_work handling will
2232 * not add work to such a task.
2234 if (unlikely(current->flags & PF_EXITING))
2236 if (current->task_works) {
2237 __set_current_state(TASK_RUNNING);
2246 * Find and free completed poll iocbs
2248 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2249 struct list_head *done)
2251 struct req_batch rb;
2252 struct io_kiocb *req;
2254 /* order with ->result store in io_complete_rw_iopoll() */
2257 io_init_req_batch(&rb);
2258 while (!list_empty(done)) {
2261 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2262 list_del(&req->inflight_entry);
2264 if (READ_ONCE(req->result) == -EAGAIN) {
2265 req->iopoll_completed = 0;
2266 if (io_rw_reissue(req))
2270 if (req->flags & REQ_F_BUFFER_SELECTED)
2271 cflags = io_put_rw_kbuf(req);
2273 __io_cqring_fill_event(req, req->result, cflags);
2276 if (refcount_dec_and_test(&req->refs))
2277 io_req_free_batch(&rb, req, &ctx->submit_state);
2280 io_commit_cqring(ctx);
2281 io_cqring_ev_posted_iopoll(ctx);
2282 io_req_free_batch_finish(ctx, &rb);
2285 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2288 struct io_kiocb *req, *tmp;
2294 * Only spin for completions if we don't have multiple devices hanging
2295 * off our complete list, and we're under the requested amount.
2297 spin = !ctx->poll_multi_file && *nr_events < min;
2300 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2301 struct kiocb *kiocb = &req->rw.kiocb;
2304 * Move completed and retryable entries to our local lists.
2305 * If we find a request that requires polling, break out
2306 * and complete those lists first, if we have entries there.
2308 if (READ_ONCE(req->iopoll_completed)) {
2309 list_move_tail(&req->inflight_entry, &done);
2312 if (!list_empty(&done))
2315 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2319 /* iopoll may have completed current req */
2320 if (READ_ONCE(req->iopoll_completed))
2321 list_move_tail(&req->inflight_entry, &done);
2328 if (!list_empty(&done))
2329 io_iopoll_complete(ctx, nr_events, &done);
2335 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2336 * non-spinning poll check - we'll still enter the driver poll loop, but only
2337 * as a non-spinning completion check.
2339 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2342 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2345 ret = io_do_iopoll(ctx, nr_events, min);
2348 if (*nr_events >= min)
2356 * We can't just wait for polled events to come to us, we have to actively
2357 * find and complete them.
2359 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2361 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2364 mutex_lock(&ctx->uring_lock);
2365 while (!list_empty(&ctx->iopoll_list)) {
2366 unsigned int nr_events = 0;
2368 io_do_iopoll(ctx, &nr_events, 0);
2370 /* let it sleep and repeat later if can't complete a request */
2374 * Ensure we allow local-to-the-cpu processing to take place,
2375 * in this case we need to ensure that we reap all events.
2376 * Also let task_work, etc. to progress by releasing the mutex
2378 if (need_resched()) {
2379 mutex_unlock(&ctx->uring_lock);
2381 mutex_lock(&ctx->uring_lock);
2384 mutex_unlock(&ctx->uring_lock);
2387 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2389 unsigned int nr_events = 0;
2390 int iters = 0, ret = 0;
2393 * We disallow the app entering submit/complete with polling, but we
2394 * still need to lock the ring to prevent racing with polled issue
2395 * that got punted to a workqueue.
2397 mutex_lock(&ctx->uring_lock);
2400 * Don't enter poll loop if we already have events pending.
2401 * If we do, we can potentially be spinning for commands that
2402 * already triggered a CQE (eg in error).
2404 if (test_bit(0, &ctx->cq_check_overflow))
2405 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2406 if (io_cqring_events(ctx))
2410 * If a submit got punted to a workqueue, we can have the
2411 * application entering polling for a command before it gets
2412 * issued. That app will hold the uring_lock for the duration
2413 * of the poll right here, so we need to take a breather every
2414 * now and then to ensure that the issue has a chance to add
2415 * the poll to the issued list. Otherwise we can spin here
2416 * forever, while the workqueue is stuck trying to acquire the
2419 if (!(++iters & 7)) {
2420 mutex_unlock(&ctx->uring_lock);
2422 mutex_lock(&ctx->uring_lock);
2425 ret = io_iopoll_getevents(ctx, &nr_events, min);
2429 } while (min && !nr_events && !need_resched());
2431 mutex_unlock(&ctx->uring_lock);
2435 static void kiocb_end_write(struct io_kiocb *req)
2438 * Tell lockdep we inherited freeze protection from submission
2441 if (req->flags & REQ_F_ISREG) {
2442 struct inode *inode = file_inode(req->file);
2444 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2446 file_end_write(req->file);
2450 static bool io_resubmit_prep(struct io_kiocb *req)
2452 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2454 struct iov_iter iter;
2456 /* already prepared */
2457 if (req->async_data)
2460 switch (req->opcode) {
2461 case IORING_OP_READV:
2462 case IORING_OP_READ_FIXED:
2463 case IORING_OP_READ:
2466 case IORING_OP_WRITEV:
2467 case IORING_OP_WRITE_FIXED:
2468 case IORING_OP_WRITE:
2472 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2477 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2480 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2483 static bool io_rw_should_reissue(struct io_kiocb *req)
2485 umode_t mode = file_inode(req->file)->i_mode;
2486 struct io_ring_ctx *ctx = req->ctx;
2488 if (!S_ISBLK(mode) && !S_ISREG(mode))
2490 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2491 !(ctx->flags & IORING_SETUP_IOPOLL)))
2494 * If ref is dying, we might be running poll reap from the exit work.
2495 * Don't attempt to reissue from that path, just let it fail with
2498 if (percpu_ref_is_dying(&ctx->refs))
2504 static bool io_rw_reissue(struct io_kiocb *req)
2507 if (!io_rw_should_reissue(req))
2510 lockdep_assert_held(&req->ctx->uring_lock);
2512 if (io_resubmit_prep(req)) {
2513 refcount_inc(&req->refs);
2514 io_queue_async_work(req);
2517 req_set_fail_links(req);
2522 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2523 unsigned int issue_flags)
2527 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2529 if (res != req->result)
2530 req_set_fail_links(req);
2532 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2533 kiocb_end_write(req);
2534 if (req->flags & REQ_F_BUFFER_SELECTED)
2535 cflags = io_put_rw_kbuf(req);
2536 __io_req_complete(req, issue_flags, res, cflags);
2539 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2541 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2543 __io_complete_rw(req, res, res2, 0);
2546 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2548 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2551 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2552 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2553 struct io_async_rw *rw = req->async_data;
2556 iov_iter_revert(&rw->iter,
2557 req->result - iov_iter_count(&rw->iter));
2558 else if (!io_resubmit_prep(req))
2563 if (kiocb->ki_flags & IOCB_WRITE)
2564 kiocb_end_write(req);
2566 if (res != -EAGAIN && res != req->result)
2567 req_set_fail_links(req);
2569 WRITE_ONCE(req->result, res);
2570 /* order with io_poll_complete() checking ->result */
2572 WRITE_ONCE(req->iopoll_completed, 1);
2576 * After the iocb has been issued, it's safe to be found on the poll list.
2577 * Adding the kiocb to the list AFTER submission ensures that we don't
2578 * find it from a io_iopoll_getevents() thread before the issuer is done
2579 * accessing the kiocb cookie.
2581 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2583 struct io_ring_ctx *ctx = req->ctx;
2586 * Track whether we have multiple files in our lists. This will impact
2587 * how we do polling eventually, not spinning if we're on potentially
2588 * different devices.
2590 if (list_empty(&ctx->iopoll_list)) {
2591 ctx->poll_multi_file = false;
2592 } else if (!ctx->poll_multi_file) {
2593 struct io_kiocb *list_req;
2595 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2597 if (list_req->file != req->file)
2598 ctx->poll_multi_file = true;
2602 * For fast devices, IO may have already completed. If it has, add
2603 * it to the front so we find it first.
2605 if (READ_ONCE(req->iopoll_completed))
2606 list_add(&req->inflight_entry, &ctx->iopoll_list);
2608 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2611 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2612 * task context or in io worker task context. If current task context is
2613 * sq thread, we don't need to check whether should wake up sq thread.
2615 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2616 wq_has_sleeper(&ctx->sq_data->wait))
2617 wake_up(&ctx->sq_data->wait);
2620 static inline void io_state_file_put(struct io_submit_state *state)
2622 if (state->file_refs) {
2623 fput_many(state->file, state->file_refs);
2624 state->file_refs = 0;
2629 * Get as many references to a file as we have IOs left in this submission,
2630 * assuming most submissions are for one file, or at least that each file
2631 * has more than one submission.
2633 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2638 if (state->file_refs) {
2639 if (state->fd == fd) {
2643 io_state_file_put(state);
2645 state->file = fget_many(fd, state->ios_left);
2646 if (unlikely(!state->file))
2650 state->file_refs = state->ios_left - 1;
2654 static bool io_bdev_nowait(struct block_device *bdev)
2656 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2660 * If we tracked the file through the SCM inflight mechanism, we could support
2661 * any file. For now, just ensure that anything potentially problematic is done
2664 static bool io_file_supports_async(struct file *file, int rw)
2666 umode_t mode = file_inode(file)->i_mode;
2668 if (S_ISBLK(mode)) {
2669 if (IS_ENABLED(CONFIG_BLOCK) &&
2670 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2674 if (S_ISCHR(mode) || S_ISSOCK(mode))
2676 if (S_ISREG(mode)) {
2677 if (IS_ENABLED(CONFIG_BLOCK) &&
2678 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2679 file->f_op != &io_uring_fops)
2684 /* any ->read/write should understand O_NONBLOCK */
2685 if (file->f_flags & O_NONBLOCK)
2688 if (!(file->f_mode & FMODE_NOWAIT))
2692 return file->f_op->read_iter != NULL;
2694 return file->f_op->write_iter != NULL;
2697 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2699 struct io_ring_ctx *ctx = req->ctx;
2700 struct kiocb *kiocb = &req->rw.kiocb;
2701 struct file *file = req->file;
2705 if (S_ISREG(file_inode(file)->i_mode))
2706 req->flags |= REQ_F_ISREG;
2708 kiocb->ki_pos = READ_ONCE(sqe->off);
2709 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2710 req->flags |= REQ_F_CUR_POS;
2711 kiocb->ki_pos = file->f_pos;
2713 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2714 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2715 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2719 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2720 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2721 req->flags |= REQ_F_NOWAIT;
2723 ioprio = READ_ONCE(sqe->ioprio);
2725 ret = ioprio_check_cap(ioprio);
2729 kiocb->ki_ioprio = ioprio;
2731 kiocb->ki_ioprio = get_current_ioprio();
2733 if (ctx->flags & IORING_SETUP_IOPOLL) {
2734 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2735 !kiocb->ki_filp->f_op->iopoll)
2738 kiocb->ki_flags |= IOCB_HIPRI;
2739 kiocb->ki_complete = io_complete_rw_iopoll;
2740 req->iopoll_completed = 0;
2742 if (kiocb->ki_flags & IOCB_HIPRI)
2744 kiocb->ki_complete = io_complete_rw;
2747 req->rw.addr = READ_ONCE(sqe->addr);
2748 req->rw.len = READ_ONCE(sqe->len);
2749 req->buf_index = READ_ONCE(sqe->buf_index);
2753 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2759 case -ERESTARTNOINTR:
2760 case -ERESTARTNOHAND:
2761 case -ERESTART_RESTARTBLOCK:
2763 * We can't just restart the syscall, since previously
2764 * submitted sqes may already be in progress. Just fail this
2770 kiocb->ki_complete(kiocb, ret, 0);
2774 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2775 unsigned int issue_flags)
2777 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2778 struct io_async_rw *io = req->async_data;
2780 /* add previously done IO, if any */
2781 if (io && io->bytes_done > 0) {
2783 ret = io->bytes_done;
2785 ret += io->bytes_done;
2788 if (req->flags & REQ_F_CUR_POS)
2789 req->file->f_pos = kiocb->ki_pos;
2790 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2791 __io_complete_rw(req, ret, 0, issue_flags);
2793 io_rw_done(kiocb, ret);
2796 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2798 struct io_ring_ctx *ctx = req->ctx;
2799 size_t len = req->rw.len;
2800 struct io_mapped_ubuf *imu;
2801 u16 index, buf_index = req->buf_index;
2805 if (unlikely(buf_index >= ctx->nr_user_bufs))
2807 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2808 imu = &ctx->user_bufs[index];
2809 buf_addr = req->rw.addr;
2812 if (buf_addr + len < buf_addr)
2814 /* not inside the mapped region */
2815 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2819 * May not be a start of buffer, set size appropriately
2820 * and advance us to the beginning.
2822 offset = buf_addr - imu->ubuf;
2823 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2827 * Don't use iov_iter_advance() here, as it's really slow for
2828 * using the latter parts of a big fixed buffer - it iterates
2829 * over each segment manually. We can cheat a bit here, because
2832 * 1) it's a BVEC iter, we set it up
2833 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2834 * first and last bvec
2836 * So just find our index, and adjust the iterator afterwards.
2837 * If the offset is within the first bvec (or the whole first
2838 * bvec, just use iov_iter_advance(). This makes it easier
2839 * since we can just skip the first segment, which may not
2840 * be PAGE_SIZE aligned.
2842 const struct bio_vec *bvec = imu->bvec;
2844 if (offset <= bvec->bv_len) {
2845 iov_iter_advance(iter, offset);
2847 unsigned long seg_skip;
2849 /* skip first vec */
2850 offset -= bvec->bv_len;
2851 seg_skip = 1 + (offset >> PAGE_SHIFT);
2853 iter->bvec = bvec + seg_skip;
2854 iter->nr_segs -= seg_skip;
2855 iter->count -= bvec->bv_len + offset;
2856 iter->iov_offset = offset & ~PAGE_MASK;
2863 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2866 mutex_unlock(&ctx->uring_lock);
2869 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2872 * "Normal" inline submissions always hold the uring_lock, since we
2873 * grab it from the system call. Same is true for the SQPOLL offload.
2874 * The only exception is when we've detached the request and issue it
2875 * from an async worker thread, grab the lock for that case.
2878 mutex_lock(&ctx->uring_lock);
2881 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2882 int bgid, struct io_buffer *kbuf,
2885 struct io_buffer *head;
2887 if (req->flags & REQ_F_BUFFER_SELECTED)
2890 io_ring_submit_lock(req->ctx, needs_lock);
2892 lockdep_assert_held(&req->ctx->uring_lock);
2894 head = xa_load(&req->ctx->io_buffers, bgid);
2896 if (!list_empty(&head->list)) {
2897 kbuf = list_last_entry(&head->list, struct io_buffer,
2899 list_del(&kbuf->list);
2902 xa_erase(&req->ctx->io_buffers, bgid);
2904 if (*len > kbuf->len)
2907 kbuf = ERR_PTR(-ENOBUFS);
2910 io_ring_submit_unlock(req->ctx, needs_lock);
2915 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2918 struct io_buffer *kbuf;
2921 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2922 bgid = req->buf_index;
2923 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2926 req->rw.addr = (u64) (unsigned long) kbuf;
2927 req->flags |= REQ_F_BUFFER_SELECTED;
2928 return u64_to_user_ptr(kbuf->addr);
2931 #ifdef CONFIG_COMPAT
2932 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2935 struct compat_iovec __user *uiov;
2936 compat_ssize_t clen;
2940 uiov = u64_to_user_ptr(req->rw.addr);
2941 if (!access_ok(uiov, sizeof(*uiov)))
2943 if (__get_user(clen, &uiov->iov_len))
2949 buf = io_rw_buffer_select(req, &len, needs_lock);
2951 return PTR_ERR(buf);
2952 iov[0].iov_base = buf;
2953 iov[0].iov_len = (compat_size_t) len;
2958 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2961 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2965 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2968 len = iov[0].iov_len;
2971 buf = io_rw_buffer_select(req, &len, needs_lock);
2973 return PTR_ERR(buf);
2974 iov[0].iov_base = buf;
2975 iov[0].iov_len = len;
2979 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2982 if (req->flags & REQ_F_BUFFER_SELECTED) {
2983 struct io_buffer *kbuf;
2985 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2986 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2987 iov[0].iov_len = kbuf->len;
2990 if (req->rw.len != 1)
2993 #ifdef CONFIG_COMPAT
2994 if (req->ctx->compat)
2995 return io_compat_import(req, iov, needs_lock);
2998 return __io_iov_buffer_select(req, iov, needs_lock);
3001 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3002 struct iov_iter *iter, bool needs_lock)
3004 void __user *buf = u64_to_user_ptr(req->rw.addr);
3005 size_t sqe_len = req->rw.len;
3006 u8 opcode = req->opcode;
3009 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3011 return io_import_fixed(req, rw, iter);
3014 /* buffer index only valid with fixed read/write, or buffer select */
3015 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3018 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3019 if (req->flags & REQ_F_BUFFER_SELECT) {
3020 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3022 return PTR_ERR(buf);
3023 req->rw.len = sqe_len;
3026 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3031 if (req->flags & REQ_F_BUFFER_SELECT) {
3032 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3034 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3039 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3043 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3045 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3049 * For files that don't have ->read_iter() and ->write_iter(), handle them
3050 * by looping over ->read() or ->write() manually.
3052 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3054 struct kiocb *kiocb = &req->rw.kiocb;
3055 struct file *file = req->file;
3059 * Don't support polled IO through this interface, and we can't
3060 * support non-blocking either. For the latter, this just causes
3061 * the kiocb to be handled from an async context.
3063 if (kiocb->ki_flags & IOCB_HIPRI)
3065 if (kiocb->ki_flags & IOCB_NOWAIT)
3068 while (iov_iter_count(iter)) {
3072 if (!iov_iter_is_bvec(iter)) {
3073 iovec = iov_iter_iovec(iter);
3075 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3076 iovec.iov_len = req->rw.len;
3080 nr = file->f_op->read(file, iovec.iov_base,
3081 iovec.iov_len, io_kiocb_ppos(kiocb));
3083 nr = file->f_op->write(file, iovec.iov_base,
3084 iovec.iov_len, io_kiocb_ppos(kiocb));
3093 if (nr != iovec.iov_len)
3097 iov_iter_advance(iter, nr);
3103 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3104 const struct iovec *fast_iov, struct iov_iter *iter)
3106 struct io_async_rw *rw = req->async_data;
3108 memcpy(&rw->iter, iter, sizeof(*iter));
3109 rw->free_iovec = iovec;
3111 /* can only be fixed buffers, no need to do anything */
3112 if (iov_iter_is_bvec(iter))
3115 unsigned iov_off = 0;
3117 rw->iter.iov = rw->fast_iov;
3118 if (iter->iov != fast_iov) {
3119 iov_off = iter->iov - fast_iov;
3120 rw->iter.iov += iov_off;
3122 if (rw->fast_iov != fast_iov)
3123 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3124 sizeof(struct iovec) * iter->nr_segs);
3126 req->flags |= REQ_F_NEED_CLEANUP;
3130 static inline int __io_alloc_async_data(struct io_kiocb *req)
3132 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3133 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3134 return req->async_data == NULL;
3137 static int io_alloc_async_data(struct io_kiocb *req)
3139 if (!io_op_defs[req->opcode].needs_async_data)
3142 return __io_alloc_async_data(req);
3145 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3146 const struct iovec *fast_iov,
3147 struct iov_iter *iter, bool force)
3149 if (!force && !io_op_defs[req->opcode].needs_async_data)
3151 if (!req->async_data) {
3152 if (__io_alloc_async_data(req)) {
3157 io_req_map_rw(req, iovec, fast_iov, iter);
3162 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3164 struct io_async_rw *iorw = req->async_data;
3165 struct iovec *iov = iorw->fast_iov;
3168 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3169 if (unlikely(ret < 0))
3172 iorw->bytes_done = 0;
3173 iorw->free_iovec = iov;
3175 req->flags |= REQ_F_NEED_CLEANUP;
3179 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3181 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3183 return io_prep_rw(req, sqe);
3187 * This is our waitqueue callback handler, registered through lock_page_async()
3188 * when we initially tried to do the IO with the iocb armed our waitqueue.
3189 * This gets called when the page is unlocked, and we generally expect that to
3190 * happen when the page IO is completed and the page is now uptodate. This will
3191 * queue a task_work based retry of the operation, attempting to copy the data
3192 * again. If the latter fails because the page was NOT uptodate, then we will
3193 * do a thread based blocking retry of the operation. That's the unexpected
3196 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3197 int sync, void *arg)
3199 struct wait_page_queue *wpq;
3200 struct io_kiocb *req = wait->private;
3201 struct wait_page_key *key = arg;
3203 wpq = container_of(wait, struct wait_page_queue, wait);
3205 if (!wake_page_match(wpq, key))
3208 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3209 list_del_init(&wait->entry);
3211 /* submit ref gets dropped, acquire a new one */
3212 refcount_inc(&req->refs);
3213 io_req_task_queue(req);
3218 * This controls whether a given IO request should be armed for async page
3219 * based retry. If we return false here, the request is handed to the async
3220 * worker threads for retry. If we're doing buffered reads on a regular file,
3221 * we prepare a private wait_page_queue entry and retry the operation. This
3222 * will either succeed because the page is now uptodate and unlocked, or it
3223 * will register a callback when the page is unlocked at IO completion. Through
3224 * that callback, io_uring uses task_work to setup a retry of the operation.
3225 * That retry will attempt the buffered read again. The retry will generally
3226 * succeed, or in rare cases where it fails, we then fall back to using the
3227 * async worker threads for a blocking retry.
3229 static bool io_rw_should_retry(struct io_kiocb *req)
3231 struct io_async_rw *rw = req->async_data;
3232 struct wait_page_queue *wait = &rw->wpq;
3233 struct kiocb *kiocb = &req->rw.kiocb;
3235 /* never retry for NOWAIT, we just complete with -EAGAIN */
3236 if (req->flags & REQ_F_NOWAIT)
3239 /* Only for buffered IO */
3240 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3244 * just use poll if we can, and don't attempt if the fs doesn't
3245 * support callback based unlocks
3247 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3250 wait->wait.func = io_async_buf_func;
3251 wait->wait.private = req;
3252 wait->wait.flags = 0;
3253 INIT_LIST_HEAD(&wait->wait.entry);
3254 kiocb->ki_flags |= IOCB_WAITQ;
3255 kiocb->ki_flags &= ~IOCB_NOWAIT;
3256 kiocb->ki_waitq = wait;
3260 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3262 if (req->file->f_op->read_iter)
3263 return call_read_iter(req->file, &req->rw.kiocb, iter);
3264 else if (req->file->f_op->read)
3265 return loop_rw_iter(READ, req, iter);
3270 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3272 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3273 struct kiocb *kiocb = &req->rw.kiocb;
3274 struct iov_iter __iter, *iter = &__iter;
3275 struct io_async_rw *rw = req->async_data;
3276 ssize_t io_size, ret, ret2;
3277 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3283 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3287 io_size = iov_iter_count(iter);
3288 req->result = io_size;
3290 /* Ensure we clear previously set non-block flag */
3291 if (!force_nonblock)
3292 kiocb->ki_flags &= ~IOCB_NOWAIT;
3294 kiocb->ki_flags |= IOCB_NOWAIT;
3296 /* If the file doesn't support async, just async punt */
3297 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3298 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3299 return ret ?: -EAGAIN;
3302 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3303 if (unlikely(ret)) {
3308 ret = io_iter_do_read(req, iter);
3310 if (ret == -EIOCBQUEUED) {
3311 if (req->async_data)
3312 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3314 } else if (ret == -EAGAIN) {
3315 /* IOPOLL retry should happen for io-wq threads */
3316 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3318 /* no retry on NONBLOCK nor RWF_NOWAIT */
3319 if (req->flags & REQ_F_NOWAIT)
3321 /* some cases will consume bytes even on error returns */
3322 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3324 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3325 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3326 /* read all, failed, already did sync or don't want to retry */
3330 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3335 rw = req->async_data;
3336 /* now use our persistent iterator, if we aren't already */
3341 rw->bytes_done += ret;
3342 /* if we can retry, do so with the callbacks armed */
3343 if (!io_rw_should_retry(req)) {
3344 kiocb->ki_flags &= ~IOCB_WAITQ;
3349 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3350 * we get -EIOCBQUEUED, then we'll get a notification when the
3351 * desired page gets unlocked. We can also get a partial read
3352 * here, and if we do, then just retry at the new offset.
3354 ret = io_iter_do_read(req, iter);
3355 if (ret == -EIOCBQUEUED)
3357 /* we got some bytes, but not all. retry. */
3358 kiocb->ki_flags &= ~IOCB_WAITQ;
3359 } while (ret > 0 && ret < io_size);
3361 kiocb_done(kiocb, ret, issue_flags);
3363 /* it's faster to check here then delegate to kfree */
3369 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3371 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3373 return io_prep_rw(req, sqe);
3376 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3378 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3379 struct kiocb *kiocb = &req->rw.kiocb;
3380 struct iov_iter __iter, *iter = &__iter;
3381 struct io_async_rw *rw = req->async_data;
3382 ssize_t ret, ret2, io_size;
3383 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3389 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3393 io_size = iov_iter_count(iter);
3394 req->result = io_size;
3396 /* Ensure we clear previously set non-block flag */
3397 if (!force_nonblock)
3398 kiocb->ki_flags &= ~IOCB_NOWAIT;
3400 kiocb->ki_flags |= IOCB_NOWAIT;
3402 /* If the file doesn't support async, just async punt */
3403 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3406 /* file path doesn't support NOWAIT for non-direct_IO */
3407 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3408 (req->flags & REQ_F_ISREG))
3411 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3416 * Open-code file_start_write here to grab freeze protection,
3417 * which will be released by another thread in
3418 * io_complete_rw(). Fool lockdep by telling it the lock got
3419 * released so that it doesn't complain about the held lock when
3420 * we return to userspace.
3422 if (req->flags & REQ_F_ISREG) {
3423 sb_start_write(file_inode(req->file)->i_sb);
3424 __sb_writers_release(file_inode(req->file)->i_sb,
3427 kiocb->ki_flags |= IOCB_WRITE;
3429 if (req->file->f_op->write_iter)
3430 ret2 = call_write_iter(req->file, kiocb, iter);
3431 else if (req->file->f_op->write)
3432 ret2 = loop_rw_iter(WRITE, req, iter);
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 (ret2 == -EIOCBQUEUED && req->async_data)
3446 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3447 if (!force_nonblock || ret2 != -EAGAIN) {
3448 /* IOPOLL retry should happen for io-wq threads */
3449 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3452 kiocb_done(kiocb, ret2, issue_flags);
3455 /* some cases will consume bytes even on error returns */
3456 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3457 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3458 return ret ?: -EAGAIN;
3461 /* it's reportedly faster than delegating the null check to kfree() */
3467 static int io_renameat_prep(struct io_kiocb *req,
3468 const struct io_uring_sqe *sqe)
3470 struct io_rename *ren = &req->rename;
3471 const char __user *oldf, *newf;
3473 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3476 ren->old_dfd = READ_ONCE(sqe->fd);
3477 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3478 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3479 ren->new_dfd = READ_ONCE(sqe->len);
3480 ren->flags = READ_ONCE(sqe->rename_flags);
3482 ren->oldpath = getname(oldf);
3483 if (IS_ERR(ren->oldpath))
3484 return PTR_ERR(ren->oldpath);
3486 ren->newpath = getname(newf);
3487 if (IS_ERR(ren->newpath)) {
3488 putname(ren->oldpath);
3489 return PTR_ERR(ren->newpath);
3492 req->flags |= REQ_F_NEED_CLEANUP;
3496 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3498 struct io_rename *ren = &req->rename;
3501 if (issue_flags & IO_URING_F_NONBLOCK)
3504 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3505 ren->newpath, ren->flags);
3507 req->flags &= ~REQ_F_NEED_CLEANUP;
3509 req_set_fail_links(req);
3510 io_req_complete(req, ret);
3514 static int io_unlinkat_prep(struct io_kiocb *req,
3515 const struct io_uring_sqe *sqe)
3517 struct io_unlink *un = &req->unlink;
3518 const char __user *fname;
3520 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3523 un->dfd = READ_ONCE(sqe->fd);
3525 un->flags = READ_ONCE(sqe->unlink_flags);
3526 if (un->flags & ~AT_REMOVEDIR)
3529 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3530 un->filename = getname(fname);
3531 if (IS_ERR(un->filename))
3532 return PTR_ERR(un->filename);
3534 req->flags |= REQ_F_NEED_CLEANUP;
3538 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3540 struct io_unlink *un = &req->unlink;
3543 if (issue_flags & IO_URING_F_NONBLOCK)
3546 if (un->flags & AT_REMOVEDIR)
3547 ret = do_rmdir(un->dfd, un->filename);
3549 ret = do_unlinkat(un->dfd, un->filename);
3551 req->flags &= ~REQ_F_NEED_CLEANUP;
3553 req_set_fail_links(req);
3554 io_req_complete(req, ret);
3558 static int io_shutdown_prep(struct io_kiocb *req,
3559 const struct io_uring_sqe *sqe)
3561 #if defined(CONFIG_NET)
3562 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3564 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3568 req->shutdown.how = READ_ONCE(sqe->len);
3575 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3577 #if defined(CONFIG_NET)
3578 struct socket *sock;
3581 if (issue_flags & IO_URING_F_NONBLOCK)
3584 sock = sock_from_file(req->file);
3585 if (unlikely(!sock))
3588 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3590 req_set_fail_links(req);
3591 io_req_complete(req, ret);
3598 static int __io_splice_prep(struct io_kiocb *req,
3599 const struct io_uring_sqe *sqe)
3601 struct io_splice* sp = &req->splice;
3602 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3604 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3608 sp->len = READ_ONCE(sqe->len);
3609 sp->flags = READ_ONCE(sqe->splice_flags);
3611 if (unlikely(sp->flags & ~valid_flags))
3614 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3615 (sp->flags & SPLICE_F_FD_IN_FIXED));
3618 req->flags |= REQ_F_NEED_CLEANUP;
3620 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3622 * Splice operation will be punted aync, and here need to
3623 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3625 req->work.flags |= IO_WQ_WORK_UNBOUND;
3631 static int io_tee_prep(struct io_kiocb *req,
3632 const struct io_uring_sqe *sqe)
3634 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3636 return __io_splice_prep(req, sqe);
3639 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3641 struct io_splice *sp = &req->splice;
3642 struct file *in = sp->file_in;
3643 struct file *out = sp->file_out;
3644 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3647 if (issue_flags & IO_URING_F_NONBLOCK)
3650 ret = do_tee(in, out, sp->len, flags);
3652 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3653 req->flags &= ~REQ_F_NEED_CLEANUP;
3656 req_set_fail_links(req);
3657 io_req_complete(req, ret);
3661 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3663 struct io_splice* sp = &req->splice;
3665 sp->off_in = READ_ONCE(sqe->splice_off_in);
3666 sp->off_out = READ_ONCE(sqe->off);
3667 return __io_splice_prep(req, sqe);
3670 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3672 struct io_splice *sp = &req->splice;
3673 struct file *in = sp->file_in;
3674 struct file *out = sp->file_out;
3675 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3676 loff_t *poff_in, *poff_out;
3679 if (issue_flags & IO_URING_F_NONBLOCK)
3682 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3683 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3686 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3688 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3689 req->flags &= ~REQ_F_NEED_CLEANUP;
3692 req_set_fail_links(req);
3693 io_req_complete(req, ret);
3698 * IORING_OP_NOP just posts a completion event, nothing else.
3700 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3702 struct io_ring_ctx *ctx = req->ctx;
3704 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3707 __io_req_complete(req, issue_flags, 0, 0);
3711 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3713 struct io_ring_ctx *ctx = req->ctx;
3718 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3720 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3723 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3724 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3727 req->sync.off = READ_ONCE(sqe->off);
3728 req->sync.len = READ_ONCE(sqe->len);
3732 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3734 loff_t end = req->sync.off + req->sync.len;
3737 /* fsync always requires a blocking context */
3738 if (issue_flags & IO_URING_F_NONBLOCK)
3741 ret = vfs_fsync_range(req->file, req->sync.off,
3742 end > 0 ? end : LLONG_MAX,
3743 req->sync.flags & IORING_FSYNC_DATASYNC);
3745 req_set_fail_links(req);
3746 io_req_complete(req, ret);
3750 static int io_fallocate_prep(struct io_kiocb *req,
3751 const struct io_uring_sqe *sqe)
3753 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3755 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3758 req->sync.off = READ_ONCE(sqe->off);
3759 req->sync.len = READ_ONCE(sqe->addr);
3760 req->sync.mode = READ_ONCE(sqe->len);
3764 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3768 /* fallocate always requiring blocking context */
3769 if (issue_flags & IO_URING_F_NONBLOCK)
3771 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3774 req_set_fail_links(req);
3775 io_req_complete(req, ret);
3779 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3781 const char __user *fname;
3784 if (unlikely(sqe->ioprio || sqe->buf_index))
3786 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3789 /* open.how should be already initialised */
3790 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3791 req->open.how.flags |= O_LARGEFILE;
3793 req->open.dfd = READ_ONCE(sqe->fd);
3794 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3795 req->open.filename = getname(fname);
3796 if (IS_ERR(req->open.filename)) {
3797 ret = PTR_ERR(req->open.filename);
3798 req->open.filename = NULL;
3801 req->open.nofile = rlimit(RLIMIT_NOFILE);
3802 req->flags |= REQ_F_NEED_CLEANUP;
3806 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3810 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3812 mode = READ_ONCE(sqe->len);
3813 flags = READ_ONCE(sqe->open_flags);
3814 req->open.how = build_open_how(flags, mode);
3815 return __io_openat_prep(req, sqe);
3818 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3820 struct open_how __user *how;
3824 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3826 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3827 len = READ_ONCE(sqe->len);
3828 if (len < OPEN_HOW_SIZE_VER0)
3831 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3836 return __io_openat_prep(req, sqe);
3839 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3841 struct open_flags op;
3844 bool resolve_nonblock;
3847 ret = build_open_flags(&req->open.how, &op);
3850 nonblock_set = op.open_flag & O_NONBLOCK;
3851 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3852 if (issue_flags & IO_URING_F_NONBLOCK) {
3854 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3855 * it'll always -EAGAIN
3857 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3859 op.lookup_flags |= LOOKUP_CACHED;
3860 op.open_flag |= O_NONBLOCK;
3863 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3867 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3868 /* only retry if RESOLVE_CACHED wasn't already set by application */
3869 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3870 file == ERR_PTR(-EAGAIN)) {
3872 * We could hang on to this 'fd', but seems like marginal
3873 * gain for something that is now known to be a slower path.
3874 * So just put it, and we'll get a new one when we retry.
3882 ret = PTR_ERR(file);
3884 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3885 file->f_flags &= ~O_NONBLOCK;
3886 fsnotify_open(file);
3887 fd_install(ret, file);
3890 putname(req->open.filename);
3891 req->flags &= ~REQ_F_NEED_CLEANUP;
3893 req_set_fail_links(req);
3894 io_req_complete(req, ret);
3898 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3900 return io_openat2(req, issue_flags);
3903 static int io_remove_buffers_prep(struct io_kiocb *req,
3904 const struct io_uring_sqe *sqe)
3906 struct io_provide_buf *p = &req->pbuf;
3909 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3912 tmp = READ_ONCE(sqe->fd);
3913 if (!tmp || tmp > USHRT_MAX)
3916 memset(p, 0, sizeof(*p));
3918 p->bgid = READ_ONCE(sqe->buf_group);
3922 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3923 int bgid, unsigned nbufs)
3927 /* shouldn't happen */
3931 /* the head kbuf is the list itself */
3932 while (!list_empty(&buf->list)) {
3933 struct io_buffer *nxt;
3935 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3936 list_del(&nxt->list);
3943 xa_erase(&ctx->io_buffers, bgid);
3948 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3950 struct io_provide_buf *p = &req->pbuf;
3951 struct io_ring_ctx *ctx = req->ctx;
3952 struct io_buffer *head;
3954 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3956 io_ring_submit_lock(ctx, !force_nonblock);
3958 lockdep_assert_held(&ctx->uring_lock);
3961 head = xa_load(&ctx->io_buffers, p->bgid);
3963 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3965 req_set_fail_links(req);
3967 /* need to hold the lock to complete IOPOLL requests */
3968 if (ctx->flags & IORING_SETUP_IOPOLL) {
3969 __io_req_complete(req, issue_flags, ret, 0);
3970 io_ring_submit_unlock(ctx, !force_nonblock);
3972 io_ring_submit_unlock(ctx, !force_nonblock);
3973 __io_req_complete(req, issue_flags, ret, 0);
3978 static int io_provide_buffers_prep(struct io_kiocb *req,
3979 const struct io_uring_sqe *sqe)
3981 struct io_provide_buf *p = &req->pbuf;
3984 if (sqe->ioprio || sqe->rw_flags)
3987 tmp = READ_ONCE(sqe->fd);
3988 if (!tmp || tmp > USHRT_MAX)
3991 p->addr = READ_ONCE(sqe->addr);
3992 p->len = READ_ONCE(sqe->len);
3994 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3997 p->bgid = READ_ONCE(sqe->buf_group);
3998 tmp = READ_ONCE(sqe->off);
3999 if (tmp > USHRT_MAX)
4005 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4007 struct io_buffer *buf;
4008 u64 addr = pbuf->addr;
4009 int i, bid = pbuf->bid;
4011 for (i = 0; i < pbuf->nbufs; i++) {
4012 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4017 buf->len = pbuf->len;
4022 INIT_LIST_HEAD(&buf->list);
4025 list_add_tail(&buf->list, &(*head)->list);
4029 return i ? i : -ENOMEM;
4032 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4034 struct io_provide_buf *p = &req->pbuf;
4035 struct io_ring_ctx *ctx = req->ctx;
4036 struct io_buffer *head, *list;
4038 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4040 io_ring_submit_lock(ctx, !force_nonblock);
4042 lockdep_assert_held(&ctx->uring_lock);
4044 list = head = xa_load(&ctx->io_buffers, p->bgid);
4046 ret = io_add_buffers(p, &head);
4047 if (ret >= 0 && !list) {
4048 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4050 __io_remove_buffers(ctx, head, p->bgid, -1U);
4053 req_set_fail_links(req);
4055 /* need to hold the lock to complete IOPOLL requests */
4056 if (ctx->flags & IORING_SETUP_IOPOLL) {
4057 __io_req_complete(req, issue_flags, ret, 0);
4058 io_ring_submit_unlock(ctx, !force_nonblock);
4060 io_ring_submit_unlock(ctx, !force_nonblock);
4061 __io_req_complete(req, issue_flags, ret, 0);
4066 static int io_epoll_ctl_prep(struct io_kiocb *req,
4067 const struct io_uring_sqe *sqe)
4069 #if defined(CONFIG_EPOLL)
4070 if (sqe->ioprio || sqe->buf_index)
4072 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4075 req->epoll.epfd = READ_ONCE(sqe->fd);
4076 req->epoll.op = READ_ONCE(sqe->len);
4077 req->epoll.fd = READ_ONCE(sqe->off);
4079 if (ep_op_has_event(req->epoll.op)) {
4080 struct epoll_event __user *ev;
4082 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4083 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4093 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4095 #if defined(CONFIG_EPOLL)
4096 struct io_epoll *ie = &req->epoll;
4098 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4100 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4101 if (force_nonblock && ret == -EAGAIN)
4105 req_set_fail_links(req);
4106 __io_req_complete(req, issue_flags, ret, 0);
4113 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4115 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4116 if (sqe->ioprio || sqe->buf_index || sqe->off)
4118 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4121 req->madvise.addr = READ_ONCE(sqe->addr);
4122 req->madvise.len = READ_ONCE(sqe->len);
4123 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4130 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4132 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4133 struct io_madvise *ma = &req->madvise;
4136 if (issue_flags & IO_URING_F_NONBLOCK)
4139 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4141 req_set_fail_links(req);
4142 io_req_complete(req, ret);
4149 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4151 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4153 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4156 req->fadvise.offset = READ_ONCE(sqe->off);
4157 req->fadvise.len = READ_ONCE(sqe->len);
4158 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4162 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4164 struct io_fadvise *fa = &req->fadvise;
4167 if (issue_flags & IO_URING_F_NONBLOCK) {
4168 switch (fa->advice) {
4169 case POSIX_FADV_NORMAL:
4170 case POSIX_FADV_RANDOM:
4171 case POSIX_FADV_SEQUENTIAL:
4178 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4180 req_set_fail_links(req);
4181 io_req_complete(req, ret);
4185 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4187 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4189 if (sqe->ioprio || sqe->buf_index)
4191 if (req->flags & REQ_F_FIXED_FILE)
4194 req->statx.dfd = READ_ONCE(sqe->fd);
4195 req->statx.mask = READ_ONCE(sqe->len);
4196 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4197 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4198 req->statx.flags = READ_ONCE(sqe->statx_flags);
4203 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4205 struct io_statx *ctx = &req->statx;
4208 if (issue_flags & IO_URING_F_NONBLOCK) {
4209 /* only need file table for an actual valid fd */
4210 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4211 req->flags |= REQ_F_NO_FILE_TABLE;
4215 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4219 req_set_fail_links(req);
4220 io_req_complete(req, ret);
4224 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4226 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4228 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4229 sqe->rw_flags || sqe->buf_index)
4231 if (req->flags & REQ_F_FIXED_FILE)
4234 req->close.fd = READ_ONCE(sqe->fd);
4238 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4240 struct files_struct *files = current->files;
4241 struct io_close *close = &req->close;
4242 struct fdtable *fdt;
4248 spin_lock(&files->file_lock);
4249 fdt = files_fdtable(files);
4250 if (close->fd >= fdt->max_fds) {
4251 spin_unlock(&files->file_lock);
4254 file = fdt->fd[close->fd];
4256 spin_unlock(&files->file_lock);
4260 if (file->f_op == &io_uring_fops) {
4261 spin_unlock(&files->file_lock);
4266 /* if the file has a flush method, be safe and punt to async */
4267 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4268 spin_unlock(&files->file_lock);
4272 ret = __close_fd_get_file(close->fd, &file);
4273 spin_unlock(&files->file_lock);
4280 /* No ->flush() or already async, safely close from here */
4281 ret = filp_close(file, current->files);
4284 req_set_fail_links(req);
4287 __io_req_complete(req, issue_flags, ret, 0);
4291 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4293 struct io_ring_ctx *ctx = req->ctx;
4295 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4297 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4300 req->sync.off = READ_ONCE(sqe->off);
4301 req->sync.len = READ_ONCE(sqe->len);
4302 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4306 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4310 /* sync_file_range always requires a blocking context */
4311 if (issue_flags & IO_URING_F_NONBLOCK)
4314 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4317 req_set_fail_links(req);
4318 io_req_complete(req, ret);
4322 #if defined(CONFIG_NET)
4323 static int io_setup_async_msg(struct io_kiocb *req,
4324 struct io_async_msghdr *kmsg)
4326 struct io_async_msghdr *async_msg = req->async_data;
4330 if (io_alloc_async_data(req)) {
4331 kfree(kmsg->free_iov);
4334 async_msg = req->async_data;
4335 req->flags |= REQ_F_NEED_CLEANUP;
4336 memcpy(async_msg, kmsg, sizeof(*kmsg));
4337 async_msg->msg.msg_name = &async_msg->addr;
4338 /* if were using fast_iov, set it to the new one */
4339 if (!async_msg->free_iov)
4340 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4345 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4346 struct io_async_msghdr *iomsg)
4348 iomsg->msg.msg_name = &iomsg->addr;
4349 iomsg->free_iov = iomsg->fast_iov;
4350 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4351 req->sr_msg.msg_flags, &iomsg->free_iov);
4354 static int io_sendmsg_prep_async(struct io_kiocb *req)
4358 if (!io_op_defs[req->opcode].needs_async_data)
4360 ret = io_sendmsg_copy_hdr(req, req->async_data);
4362 req->flags |= REQ_F_NEED_CLEANUP;
4366 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4368 struct io_sr_msg *sr = &req->sr_msg;
4370 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4373 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4374 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4375 sr->len = READ_ONCE(sqe->len);
4377 #ifdef CONFIG_COMPAT
4378 if (req->ctx->compat)
4379 sr->msg_flags |= MSG_CMSG_COMPAT;
4384 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4386 struct io_async_msghdr iomsg, *kmsg;
4387 struct socket *sock;
4391 sock = sock_from_file(req->file);
4392 if (unlikely(!sock))
4395 kmsg = req->async_data;
4397 ret = io_sendmsg_copy_hdr(req, &iomsg);
4403 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4404 if (flags & MSG_DONTWAIT)
4405 req->flags |= REQ_F_NOWAIT;
4406 else if (issue_flags & IO_URING_F_NONBLOCK)
4407 flags |= MSG_DONTWAIT;
4409 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4410 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4411 return io_setup_async_msg(req, kmsg);
4412 if (ret == -ERESTARTSYS)
4415 /* fast path, check for non-NULL to avoid function call */
4417 kfree(kmsg->free_iov);
4418 req->flags &= ~REQ_F_NEED_CLEANUP;
4420 req_set_fail_links(req);
4421 __io_req_complete(req, issue_flags, ret, 0);
4425 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4427 struct io_sr_msg *sr = &req->sr_msg;
4430 struct socket *sock;
4434 sock = sock_from_file(req->file);
4435 if (unlikely(!sock))
4438 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4442 msg.msg_name = NULL;
4443 msg.msg_control = NULL;
4444 msg.msg_controllen = 0;
4445 msg.msg_namelen = 0;
4447 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4448 if (flags & MSG_DONTWAIT)
4449 req->flags |= REQ_F_NOWAIT;
4450 else if (issue_flags & IO_URING_F_NONBLOCK)
4451 flags |= MSG_DONTWAIT;
4453 msg.msg_flags = flags;
4454 ret = sock_sendmsg(sock, &msg);
4455 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4457 if (ret == -ERESTARTSYS)
4461 req_set_fail_links(req);
4462 __io_req_complete(req, issue_flags, ret, 0);
4466 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4467 struct io_async_msghdr *iomsg)
4469 struct io_sr_msg *sr = &req->sr_msg;
4470 struct iovec __user *uiov;
4474 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4475 &iomsg->uaddr, &uiov, &iov_len);
4479 if (req->flags & REQ_F_BUFFER_SELECT) {
4482 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4484 sr->len = iomsg->fast_iov[0].iov_len;
4485 iomsg->free_iov = NULL;
4487 iomsg->free_iov = iomsg->fast_iov;
4488 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4489 &iomsg->free_iov, &iomsg->msg.msg_iter,
4498 #ifdef CONFIG_COMPAT
4499 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4500 struct io_async_msghdr *iomsg)
4502 struct compat_msghdr __user *msg_compat;
4503 struct io_sr_msg *sr = &req->sr_msg;
4504 struct compat_iovec __user *uiov;
4509 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4510 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4515 uiov = compat_ptr(ptr);
4516 if (req->flags & REQ_F_BUFFER_SELECT) {
4517 compat_ssize_t clen;
4521 if (!access_ok(uiov, sizeof(*uiov)))
4523 if (__get_user(clen, &uiov->iov_len))
4528 iomsg->free_iov = NULL;
4530 iomsg->free_iov = iomsg->fast_iov;
4531 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4532 UIO_FASTIOV, &iomsg->free_iov,
4533 &iomsg->msg.msg_iter, true);
4542 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4543 struct io_async_msghdr *iomsg)
4545 iomsg->msg.msg_name = &iomsg->addr;
4547 #ifdef CONFIG_COMPAT
4548 if (req->ctx->compat)
4549 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4552 return __io_recvmsg_copy_hdr(req, iomsg);
4555 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4558 struct io_sr_msg *sr = &req->sr_msg;
4559 struct io_buffer *kbuf;
4561 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4566 req->flags |= REQ_F_BUFFER_SELECTED;
4570 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4572 return io_put_kbuf(req, req->sr_msg.kbuf);
4575 static int io_recvmsg_prep_async(struct io_kiocb *req)
4579 if (!io_op_defs[req->opcode].needs_async_data)
4581 ret = io_recvmsg_copy_hdr(req, req->async_data);
4583 req->flags |= REQ_F_NEED_CLEANUP;
4587 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4589 struct io_sr_msg *sr = &req->sr_msg;
4591 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4594 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4595 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4596 sr->len = READ_ONCE(sqe->len);
4597 sr->bgid = READ_ONCE(sqe->buf_group);
4599 #ifdef CONFIG_COMPAT
4600 if (req->ctx->compat)
4601 sr->msg_flags |= MSG_CMSG_COMPAT;
4606 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4608 struct io_async_msghdr iomsg, *kmsg;
4609 struct socket *sock;
4610 struct io_buffer *kbuf;
4612 int ret, cflags = 0;
4613 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4615 sock = sock_from_file(req->file);
4616 if (unlikely(!sock))
4619 kmsg = req->async_data;
4621 ret = io_recvmsg_copy_hdr(req, &iomsg);
4627 if (req->flags & REQ_F_BUFFER_SELECT) {
4628 kbuf = io_recv_buffer_select(req, !force_nonblock);
4630 return PTR_ERR(kbuf);
4631 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4632 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4633 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4634 1, req->sr_msg.len);
4637 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4638 if (flags & MSG_DONTWAIT)
4639 req->flags |= REQ_F_NOWAIT;
4640 else if (force_nonblock)
4641 flags |= MSG_DONTWAIT;
4643 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4644 kmsg->uaddr, flags);
4645 if (force_nonblock && ret == -EAGAIN)
4646 return io_setup_async_msg(req, kmsg);
4647 if (ret == -ERESTARTSYS)
4650 if (req->flags & REQ_F_BUFFER_SELECTED)
4651 cflags = io_put_recv_kbuf(req);
4652 /* fast path, check for non-NULL to avoid function call */
4654 kfree(kmsg->free_iov);
4655 req->flags &= ~REQ_F_NEED_CLEANUP;
4657 req_set_fail_links(req);
4658 __io_req_complete(req, issue_flags, ret, cflags);
4662 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4664 struct io_buffer *kbuf;
4665 struct io_sr_msg *sr = &req->sr_msg;
4667 void __user *buf = sr->buf;
4668 struct socket *sock;
4671 int ret, cflags = 0;
4672 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4674 sock = sock_from_file(req->file);
4675 if (unlikely(!sock))
4678 if (req->flags & REQ_F_BUFFER_SELECT) {
4679 kbuf = io_recv_buffer_select(req, !force_nonblock);
4681 return PTR_ERR(kbuf);
4682 buf = u64_to_user_ptr(kbuf->addr);
4685 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4689 msg.msg_name = NULL;
4690 msg.msg_control = NULL;
4691 msg.msg_controllen = 0;
4692 msg.msg_namelen = 0;
4693 msg.msg_iocb = NULL;
4696 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4697 if (flags & MSG_DONTWAIT)
4698 req->flags |= REQ_F_NOWAIT;
4699 else if (force_nonblock)
4700 flags |= MSG_DONTWAIT;
4702 ret = sock_recvmsg(sock, &msg, flags);
4703 if (force_nonblock && ret == -EAGAIN)
4705 if (ret == -ERESTARTSYS)
4708 if (req->flags & REQ_F_BUFFER_SELECTED)
4709 cflags = io_put_recv_kbuf(req);
4711 req_set_fail_links(req);
4712 __io_req_complete(req, issue_flags, ret, cflags);
4716 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4718 struct io_accept *accept = &req->accept;
4720 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4722 if (sqe->ioprio || sqe->len || sqe->buf_index)
4725 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4726 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4727 accept->flags = READ_ONCE(sqe->accept_flags);
4728 accept->nofile = rlimit(RLIMIT_NOFILE);
4732 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4734 struct io_accept *accept = &req->accept;
4735 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4736 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4739 if (req->file->f_flags & O_NONBLOCK)
4740 req->flags |= REQ_F_NOWAIT;
4742 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4743 accept->addr_len, accept->flags,
4745 if (ret == -EAGAIN && force_nonblock)
4748 if (ret == -ERESTARTSYS)
4750 req_set_fail_links(req);
4752 __io_req_complete(req, issue_flags, ret, 0);
4756 static int io_connect_prep_async(struct io_kiocb *req)
4758 struct io_async_connect *io = req->async_data;
4759 struct io_connect *conn = &req->connect;
4761 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4764 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4766 struct io_connect *conn = &req->connect;
4768 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4770 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4773 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4774 conn->addr_len = READ_ONCE(sqe->addr2);
4778 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4780 struct io_async_connect __io, *io;
4781 unsigned file_flags;
4783 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4785 if (req->async_data) {
4786 io = req->async_data;
4788 ret = move_addr_to_kernel(req->connect.addr,
4789 req->connect.addr_len,
4796 file_flags = force_nonblock ? O_NONBLOCK : 0;
4798 ret = __sys_connect_file(req->file, &io->address,
4799 req->connect.addr_len, file_flags);
4800 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4801 if (req->async_data)
4803 if (io_alloc_async_data(req)) {
4807 io = req->async_data;
4808 memcpy(req->async_data, &__io, sizeof(__io));
4811 if (ret == -ERESTARTSYS)
4815 req_set_fail_links(req);
4816 __io_req_complete(req, issue_flags, ret, 0);
4819 #else /* !CONFIG_NET */
4820 #define IO_NETOP_FN(op) \
4821 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4823 return -EOPNOTSUPP; \
4826 #define IO_NETOP_PREP(op) \
4828 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4830 return -EOPNOTSUPP; \
4833 #define IO_NETOP_PREP_ASYNC(op) \
4835 static int io_##op##_prep_async(struct io_kiocb *req) \
4837 return -EOPNOTSUPP; \
4840 IO_NETOP_PREP_ASYNC(sendmsg);
4841 IO_NETOP_PREP_ASYNC(recvmsg);
4842 IO_NETOP_PREP_ASYNC(connect);
4843 IO_NETOP_PREP(accept);
4846 #endif /* CONFIG_NET */
4848 struct io_poll_table {
4849 struct poll_table_struct pt;
4850 struct io_kiocb *req;
4854 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4855 __poll_t mask, task_work_func_t func)
4859 /* for instances that support it check for an event match first: */
4860 if (mask && !(mask & poll->events))
4863 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4865 list_del_init(&poll->wait.entry);
4868 req->task_work.func = func;
4869 percpu_ref_get(&req->ctx->refs);
4872 * If this fails, then the task is exiting. When a task exits, the
4873 * work gets canceled, so just cancel this request as well instead
4874 * of executing it. We can't safely execute it anyway, as we may not
4875 * have the needed state needed for it anyway.
4877 ret = io_req_task_work_add(req);
4878 if (unlikely(ret)) {
4879 WRITE_ONCE(poll->canceled, true);
4880 io_req_task_work_add_fallback(req, func);
4885 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4886 __acquires(&req->ctx->completion_lock)
4888 struct io_ring_ctx *ctx = req->ctx;
4890 if (!req->result && !READ_ONCE(poll->canceled)) {
4891 struct poll_table_struct pt = { ._key = poll->events };
4893 req->result = vfs_poll(req->file, &pt) & poll->events;
4896 spin_lock_irq(&ctx->completion_lock);
4897 if (!req->result && !READ_ONCE(poll->canceled)) {
4898 add_wait_queue(poll->head, &poll->wait);
4905 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4907 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4908 if (req->opcode == IORING_OP_POLL_ADD)
4909 return req->async_data;
4910 return req->apoll->double_poll;
4913 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4915 if (req->opcode == IORING_OP_POLL_ADD)
4917 return &req->apoll->poll;
4920 static void io_poll_remove_double(struct io_kiocb *req)
4922 struct io_poll_iocb *poll = io_poll_get_double(req);
4924 lockdep_assert_held(&req->ctx->completion_lock);
4926 if (poll && poll->head) {
4927 struct wait_queue_head *head = poll->head;
4929 spin_lock(&head->lock);
4930 list_del_init(&poll->wait.entry);
4931 if (poll->wait.private)
4932 refcount_dec(&req->refs);
4934 spin_unlock(&head->lock);
4938 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4940 struct io_ring_ctx *ctx = req->ctx;
4942 io_poll_remove_double(req);
4943 req->poll.done = true;
4944 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4945 io_commit_cqring(ctx);
4948 static void io_poll_task_func(struct callback_head *cb)
4950 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4951 struct io_ring_ctx *ctx = req->ctx;
4952 struct io_kiocb *nxt;
4954 if (io_poll_rewait(req, &req->poll)) {
4955 spin_unlock_irq(&ctx->completion_lock);
4957 hash_del(&req->hash_node);
4958 io_poll_complete(req, req->result, 0);
4959 spin_unlock_irq(&ctx->completion_lock);
4961 nxt = io_put_req_find_next(req);
4962 io_cqring_ev_posted(ctx);
4964 __io_req_task_submit(nxt);
4967 percpu_ref_put(&ctx->refs);
4970 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4971 int sync, void *key)
4973 struct io_kiocb *req = wait->private;
4974 struct io_poll_iocb *poll = io_poll_get_single(req);
4975 __poll_t mask = key_to_poll(key);
4977 /* for instances that support it check for an event match first: */
4978 if (mask && !(mask & poll->events))
4981 list_del_init(&wait->entry);
4983 if (poll && poll->head) {
4986 spin_lock(&poll->head->lock);
4987 done = list_empty(&poll->wait.entry);
4989 list_del_init(&poll->wait.entry);
4990 /* make sure double remove sees this as being gone */
4991 wait->private = NULL;
4992 spin_unlock(&poll->head->lock);
4994 /* use wait func handler, so it matches the rq type */
4995 poll->wait.func(&poll->wait, mode, sync, key);
4998 refcount_dec(&req->refs);
5002 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5003 wait_queue_func_t wake_func)
5007 poll->canceled = false;
5008 poll->events = events;
5009 INIT_LIST_HEAD(&poll->wait.entry);
5010 init_waitqueue_func_entry(&poll->wait, wake_func);
5013 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5014 struct wait_queue_head *head,
5015 struct io_poll_iocb **poll_ptr)
5017 struct io_kiocb *req = pt->req;
5020 * If poll->head is already set, it's because the file being polled
5021 * uses multiple waitqueues for poll handling (eg one for read, one
5022 * for write). Setup a separate io_poll_iocb if this happens.
5024 if (unlikely(poll->head)) {
5025 struct io_poll_iocb *poll_one = poll;
5027 /* already have a 2nd entry, fail a third attempt */
5029 pt->error = -EINVAL;
5032 /* double add on the same waitqueue head, ignore */
5033 if (poll->head == head)
5035 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5037 pt->error = -ENOMEM;
5040 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5041 refcount_inc(&req->refs);
5042 poll->wait.private = req;
5049 if (poll->events & EPOLLEXCLUSIVE)
5050 add_wait_queue_exclusive(head, &poll->wait);
5052 add_wait_queue(head, &poll->wait);
5055 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5056 struct poll_table_struct *p)
5058 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5059 struct async_poll *apoll = pt->req->apoll;
5061 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5064 static void io_async_task_func(struct callback_head *cb)
5066 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5067 struct async_poll *apoll = req->apoll;
5068 struct io_ring_ctx *ctx = req->ctx;
5070 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5072 if (io_poll_rewait(req, &apoll->poll)) {
5073 spin_unlock_irq(&ctx->completion_lock);
5074 percpu_ref_put(&ctx->refs);
5078 /* If req is still hashed, it cannot have been canceled. Don't check. */
5079 if (hash_hashed(&req->hash_node))
5080 hash_del(&req->hash_node);
5082 io_poll_remove_double(req);
5083 spin_unlock_irq(&ctx->completion_lock);
5085 if (!READ_ONCE(apoll->poll.canceled))
5086 __io_req_task_submit(req);
5088 __io_req_task_cancel(req, -ECANCELED);
5090 percpu_ref_put(&ctx->refs);
5091 kfree(apoll->double_poll);
5095 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5098 struct io_kiocb *req = wait->private;
5099 struct io_poll_iocb *poll = &req->apoll->poll;
5101 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5104 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5107 static void io_poll_req_insert(struct io_kiocb *req)
5109 struct io_ring_ctx *ctx = req->ctx;
5110 struct hlist_head *list;
5112 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5113 hlist_add_head(&req->hash_node, list);
5116 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5117 struct io_poll_iocb *poll,
5118 struct io_poll_table *ipt, __poll_t mask,
5119 wait_queue_func_t wake_func)
5120 __acquires(&ctx->completion_lock)
5122 struct io_ring_ctx *ctx = req->ctx;
5123 bool cancel = false;
5125 INIT_HLIST_NODE(&req->hash_node);
5126 io_init_poll_iocb(poll, mask, wake_func);
5127 poll->file = req->file;
5128 poll->wait.private = req;
5130 ipt->pt._key = mask;
5132 ipt->error = -EINVAL;
5134 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5136 spin_lock_irq(&ctx->completion_lock);
5137 if (likely(poll->head)) {
5138 spin_lock(&poll->head->lock);
5139 if (unlikely(list_empty(&poll->wait.entry))) {
5145 if (mask || ipt->error)
5146 list_del_init(&poll->wait.entry);
5148 WRITE_ONCE(poll->canceled, true);
5149 else if (!poll->done) /* actually waiting for an event */
5150 io_poll_req_insert(req);
5151 spin_unlock(&poll->head->lock);
5157 static bool io_arm_poll_handler(struct io_kiocb *req)
5159 const struct io_op_def *def = &io_op_defs[req->opcode];
5160 struct io_ring_ctx *ctx = req->ctx;
5161 struct async_poll *apoll;
5162 struct io_poll_table ipt;
5166 if (!req->file || !file_can_poll(req->file))
5168 if (req->flags & REQ_F_POLLED)
5172 else if (def->pollout)
5176 /* if we can't nonblock try, then no point in arming a poll handler */
5177 if (!io_file_supports_async(req->file, rw))
5180 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5181 if (unlikely(!apoll))
5183 apoll->double_poll = NULL;
5185 req->flags |= REQ_F_POLLED;
5190 mask |= POLLIN | POLLRDNORM;
5192 mask |= POLLOUT | POLLWRNORM;
5194 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5195 if ((req->opcode == IORING_OP_RECVMSG) &&
5196 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5199 mask |= POLLERR | POLLPRI;
5201 ipt.pt._qproc = io_async_queue_proc;
5203 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5205 if (ret || ipt.error) {
5206 io_poll_remove_double(req);
5207 spin_unlock_irq(&ctx->completion_lock);
5208 kfree(apoll->double_poll);
5212 spin_unlock_irq(&ctx->completion_lock);
5213 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5214 apoll->poll.events);
5218 static bool __io_poll_remove_one(struct io_kiocb *req,
5219 struct io_poll_iocb *poll)
5221 bool do_complete = false;
5223 spin_lock(&poll->head->lock);
5224 WRITE_ONCE(poll->canceled, true);
5225 if (!list_empty(&poll->wait.entry)) {
5226 list_del_init(&poll->wait.entry);
5229 spin_unlock(&poll->head->lock);
5230 hash_del(&req->hash_node);
5234 static bool io_poll_remove_one(struct io_kiocb *req)
5238 io_poll_remove_double(req);
5240 if (req->opcode == IORING_OP_POLL_ADD) {
5241 do_complete = __io_poll_remove_one(req, &req->poll);
5243 struct async_poll *apoll = req->apoll;
5245 /* non-poll requests have submit ref still */
5246 do_complete = __io_poll_remove_one(req, &apoll->poll);
5249 kfree(apoll->double_poll);
5255 io_cqring_fill_event(req, -ECANCELED);
5256 io_commit_cqring(req->ctx);
5257 req_set_fail_links(req);
5258 io_put_req_deferred(req, 1);
5265 * Returns true if we found and killed one or more poll requests
5267 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5268 struct files_struct *files)
5270 struct hlist_node *tmp;
5271 struct io_kiocb *req;
5274 spin_lock_irq(&ctx->completion_lock);
5275 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5276 struct hlist_head *list;
5278 list = &ctx->cancel_hash[i];
5279 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5280 if (io_match_task(req, tsk, files))
5281 posted += io_poll_remove_one(req);
5284 spin_unlock_irq(&ctx->completion_lock);
5287 io_cqring_ev_posted(ctx);
5292 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5294 struct hlist_head *list;
5295 struct io_kiocb *req;
5297 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5298 hlist_for_each_entry(req, list, hash_node) {
5299 if (sqe_addr != req->user_data)
5301 if (io_poll_remove_one(req))
5309 static int io_poll_remove_prep(struct io_kiocb *req,
5310 const struct io_uring_sqe *sqe)
5312 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5314 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5318 req->poll_remove.addr = READ_ONCE(sqe->addr);
5323 * Find a running poll command that matches one specified in sqe->addr,
5324 * and remove it if found.
5326 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5328 struct io_ring_ctx *ctx = req->ctx;
5331 spin_lock_irq(&ctx->completion_lock);
5332 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5333 spin_unlock_irq(&ctx->completion_lock);
5336 req_set_fail_links(req);
5337 io_req_complete(req, ret);
5341 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5344 struct io_kiocb *req = wait->private;
5345 struct io_poll_iocb *poll = &req->poll;
5347 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5350 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5351 struct poll_table_struct *p)
5353 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5355 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5358 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5360 struct io_poll_iocb *poll = &req->poll;
5363 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5365 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5368 events = READ_ONCE(sqe->poll32_events);
5370 events = swahw32(events);
5372 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5373 (events & EPOLLEXCLUSIVE);
5377 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5379 struct io_poll_iocb *poll = &req->poll;
5380 struct io_ring_ctx *ctx = req->ctx;
5381 struct io_poll_table ipt;
5384 ipt.pt._qproc = io_poll_queue_proc;
5386 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5389 if (mask) { /* no async, we'd stolen it */
5391 io_poll_complete(req, mask, 0);
5393 spin_unlock_irq(&ctx->completion_lock);
5396 io_cqring_ev_posted(ctx);
5402 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5404 struct io_timeout_data *data = container_of(timer,
5405 struct io_timeout_data, timer);
5406 struct io_kiocb *req = data->req;
5407 struct io_ring_ctx *ctx = req->ctx;
5408 unsigned long flags;
5410 spin_lock_irqsave(&ctx->completion_lock, flags);
5411 list_del_init(&req->timeout.list);
5412 atomic_set(&req->ctx->cq_timeouts,
5413 atomic_read(&req->ctx->cq_timeouts) + 1);
5415 io_cqring_fill_event(req, -ETIME);
5416 io_commit_cqring(ctx);
5417 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5419 io_cqring_ev_posted(ctx);
5420 req_set_fail_links(req);
5422 return HRTIMER_NORESTART;
5425 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5428 struct io_timeout_data *io;
5429 struct io_kiocb *req;
5432 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5433 if (user_data == req->user_data) {
5440 return ERR_PTR(ret);
5442 io = req->async_data;
5443 ret = hrtimer_try_to_cancel(&io->timer);
5445 return ERR_PTR(-EALREADY);
5446 list_del_init(&req->timeout.list);
5450 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5452 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5455 return PTR_ERR(req);
5457 req_set_fail_links(req);
5458 io_cqring_fill_event(req, -ECANCELED);
5459 io_put_req_deferred(req, 1);
5463 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5464 struct timespec64 *ts, enum hrtimer_mode mode)
5466 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5467 struct io_timeout_data *data;
5470 return PTR_ERR(req);
5472 req->timeout.off = 0; /* noseq */
5473 data = req->async_data;
5474 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5475 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5476 data->timer.function = io_timeout_fn;
5477 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5481 static int io_timeout_remove_prep(struct io_kiocb *req,
5482 const struct io_uring_sqe *sqe)
5484 struct io_timeout_rem *tr = &req->timeout_rem;
5486 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5488 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5490 if (sqe->ioprio || sqe->buf_index || sqe->len)
5493 tr->addr = READ_ONCE(sqe->addr);
5494 tr->flags = READ_ONCE(sqe->timeout_flags);
5495 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5496 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5498 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5500 } else if (tr->flags) {
5501 /* timeout removal doesn't support flags */
5508 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5510 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5515 * Remove or update an existing timeout command
5517 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5519 struct io_timeout_rem *tr = &req->timeout_rem;
5520 struct io_ring_ctx *ctx = req->ctx;
5523 spin_lock_irq(&ctx->completion_lock);
5524 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5525 ret = io_timeout_cancel(ctx, tr->addr);
5527 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5528 io_translate_timeout_mode(tr->flags));
5530 io_cqring_fill_event(req, ret);
5531 io_commit_cqring(ctx);
5532 spin_unlock_irq(&ctx->completion_lock);
5533 io_cqring_ev_posted(ctx);
5535 req_set_fail_links(req);
5540 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5541 bool is_timeout_link)
5543 struct io_timeout_data *data;
5545 u32 off = READ_ONCE(sqe->off);
5547 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5549 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5551 if (off && is_timeout_link)
5553 flags = READ_ONCE(sqe->timeout_flags);
5554 if (flags & ~IORING_TIMEOUT_ABS)
5557 req->timeout.off = off;
5559 if (!req->async_data && io_alloc_async_data(req))
5562 data = req->async_data;
5565 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5568 data->mode = io_translate_timeout_mode(flags);
5569 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5570 io_req_track_inflight(req);
5574 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5576 struct io_ring_ctx *ctx = req->ctx;
5577 struct io_timeout_data *data = req->async_data;
5578 struct list_head *entry;
5579 u32 tail, off = req->timeout.off;
5581 spin_lock_irq(&ctx->completion_lock);
5584 * sqe->off holds how many events that need to occur for this
5585 * timeout event to be satisfied. If it isn't set, then this is
5586 * a pure timeout request, sequence isn't used.
5588 if (io_is_timeout_noseq(req)) {
5589 entry = ctx->timeout_list.prev;
5593 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5594 req->timeout.target_seq = tail + off;
5596 /* Update the last seq here in case io_flush_timeouts() hasn't.
5597 * This is safe because ->completion_lock is held, and submissions
5598 * and completions are never mixed in the same ->completion_lock section.
5600 ctx->cq_last_tm_flush = tail;
5603 * Insertion sort, ensuring the first entry in the list is always
5604 * the one we need first.
5606 list_for_each_prev(entry, &ctx->timeout_list) {
5607 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5610 if (io_is_timeout_noseq(nxt))
5612 /* nxt.seq is behind @tail, otherwise would've been completed */
5613 if (off >= nxt->timeout.target_seq - tail)
5617 list_add(&req->timeout.list, entry);
5618 data->timer.function = io_timeout_fn;
5619 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5620 spin_unlock_irq(&ctx->completion_lock);
5624 struct io_cancel_data {
5625 struct io_ring_ctx *ctx;
5629 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5631 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5632 struct io_cancel_data *cd = data;
5634 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5637 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5638 struct io_ring_ctx *ctx)
5640 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5641 enum io_wq_cancel cancel_ret;
5644 if (!tctx || !tctx->io_wq)
5647 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5648 switch (cancel_ret) {
5649 case IO_WQ_CANCEL_OK:
5652 case IO_WQ_CANCEL_RUNNING:
5655 case IO_WQ_CANCEL_NOTFOUND:
5663 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5664 struct io_kiocb *req, __u64 sqe_addr,
5667 unsigned long flags;
5670 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5671 if (ret != -ENOENT) {
5672 spin_lock_irqsave(&ctx->completion_lock, flags);
5676 spin_lock_irqsave(&ctx->completion_lock, flags);
5677 ret = io_timeout_cancel(ctx, sqe_addr);
5680 ret = io_poll_cancel(ctx, sqe_addr);
5684 io_cqring_fill_event(req, ret);
5685 io_commit_cqring(ctx);
5686 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5687 io_cqring_ev_posted(ctx);
5690 req_set_fail_links(req);
5694 static int io_async_cancel_prep(struct io_kiocb *req,
5695 const struct io_uring_sqe *sqe)
5697 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5699 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5701 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5704 req->cancel.addr = READ_ONCE(sqe->addr);
5708 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5710 struct io_ring_ctx *ctx = req->ctx;
5711 u64 sqe_addr = req->cancel.addr;
5712 struct io_tctx_node *node;
5715 /* tasks should wait for their io-wq threads, so safe w/o sync */
5716 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5717 spin_lock_irq(&ctx->completion_lock);
5720 ret = io_timeout_cancel(ctx, sqe_addr);
5723 ret = io_poll_cancel(ctx, sqe_addr);
5726 spin_unlock_irq(&ctx->completion_lock);
5728 /* slow path, try all io-wq's */
5729 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5731 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5732 struct io_uring_task *tctx = node->task->io_uring;
5734 if (!tctx || !tctx->io_wq)
5736 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5740 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5742 spin_lock_irq(&ctx->completion_lock);
5744 io_cqring_fill_event(req, ret);
5745 io_commit_cqring(ctx);
5746 spin_unlock_irq(&ctx->completion_lock);
5747 io_cqring_ev_posted(ctx);
5750 req_set_fail_links(req);
5755 static int io_rsrc_update_prep(struct io_kiocb *req,
5756 const struct io_uring_sqe *sqe)
5758 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5760 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5762 if (sqe->ioprio || sqe->rw_flags)
5765 req->rsrc_update.offset = READ_ONCE(sqe->off);
5766 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5767 if (!req->rsrc_update.nr_args)
5769 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5773 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5775 struct io_ring_ctx *ctx = req->ctx;
5776 struct io_uring_rsrc_update up;
5779 if (issue_flags & IO_URING_F_NONBLOCK)
5782 up.offset = req->rsrc_update.offset;
5783 up.data = req->rsrc_update.arg;
5785 mutex_lock(&ctx->uring_lock);
5786 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5787 mutex_unlock(&ctx->uring_lock);
5790 req_set_fail_links(req);
5791 __io_req_complete(req, issue_flags, ret, 0);
5795 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5797 switch (req->opcode) {
5800 case IORING_OP_READV:
5801 case IORING_OP_READ_FIXED:
5802 case IORING_OP_READ:
5803 return io_read_prep(req, sqe);
5804 case IORING_OP_WRITEV:
5805 case IORING_OP_WRITE_FIXED:
5806 case IORING_OP_WRITE:
5807 return io_write_prep(req, sqe);
5808 case IORING_OP_POLL_ADD:
5809 return io_poll_add_prep(req, sqe);
5810 case IORING_OP_POLL_REMOVE:
5811 return io_poll_remove_prep(req, sqe);
5812 case IORING_OP_FSYNC:
5813 return io_fsync_prep(req, sqe);
5814 case IORING_OP_SYNC_FILE_RANGE:
5815 return io_sfr_prep(req, sqe);
5816 case IORING_OP_SENDMSG:
5817 case IORING_OP_SEND:
5818 return io_sendmsg_prep(req, sqe);
5819 case IORING_OP_RECVMSG:
5820 case IORING_OP_RECV:
5821 return io_recvmsg_prep(req, sqe);
5822 case IORING_OP_CONNECT:
5823 return io_connect_prep(req, sqe);
5824 case IORING_OP_TIMEOUT:
5825 return io_timeout_prep(req, sqe, false);
5826 case IORING_OP_TIMEOUT_REMOVE:
5827 return io_timeout_remove_prep(req, sqe);
5828 case IORING_OP_ASYNC_CANCEL:
5829 return io_async_cancel_prep(req, sqe);
5830 case IORING_OP_LINK_TIMEOUT:
5831 return io_timeout_prep(req, sqe, true);
5832 case IORING_OP_ACCEPT:
5833 return io_accept_prep(req, sqe);
5834 case IORING_OP_FALLOCATE:
5835 return io_fallocate_prep(req, sqe);
5836 case IORING_OP_OPENAT:
5837 return io_openat_prep(req, sqe);
5838 case IORING_OP_CLOSE:
5839 return io_close_prep(req, sqe);
5840 case IORING_OP_FILES_UPDATE:
5841 return io_rsrc_update_prep(req, sqe);
5842 case IORING_OP_STATX:
5843 return io_statx_prep(req, sqe);
5844 case IORING_OP_FADVISE:
5845 return io_fadvise_prep(req, sqe);
5846 case IORING_OP_MADVISE:
5847 return io_madvise_prep(req, sqe);
5848 case IORING_OP_OPENAT2:
5849 return io_openat2_prep(req, sqe);
5850 case IORING_OP_EPOLL_CTL:
5851 return io_epoll_ctl_prep(req, sqe);
5852 case IORING_OP_SPLICE:
5853 return io_splice_prep(req, sqe);
5854 case IORING_OP_PROVIDE_BUFFERS:
5855 return io_provide_buffers_prep(req, sqe);
5856 case IORING_OP_REMOVE_BUFFERS:
5857 return io_remove_buffers_prep(req, sqe);
5859 return io_tee_prep(req, sqe);
5860 case IORING_OP_SHUTDOWN:
5861 return io_shutdown_prep(req, sqe);
5862 case IORING_OP_RENAMEAT:
5863 return io_renameat_prep(req, sqe);
5864 case IORING_OP_UNLINKAT:
5865 return io_unlinkat_prep(req, sqe);
5868 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5873 static int io_req_prep_async(struct io_kiocb *req)
5875 switch (req->opcode) {
5876 case IORING_OP_READV:
5877 case IORING_OP_READ_FIXED:
5878 case IORING_OP_READ:
5879 return io_rw_prep_async(req, READ);
5880 case IORING_OP_WRITEV:
5881 case IORING_OP_WRITE_FIXED:
5882 case IORING_OP_WRITE:
5883 return io_rw_prep_async(req, WRITE);
5884 case IORING_OP_SENDMSG:
5885 case IORING_OP_SEND:
5886 return io_sendmsg_prep_async(req);
5887 case IORING_OP_RECVMSG:
5888 case IORING_OP_RECV:
5889 return io_recvmsg_prep_async(req);
5890 case IORING_OP_CONNECT:
5891 return io_connect_prep_async(req);
5896 static int io_req_defer_prep(struct io_kiocb *req)
5898 if (!io_op_defs[req->opcode].needs_async_data)
5900 /* some opcodes init it during the inital prep */
5901 if (req->async_data)
5903 if (__io_alloc_async_data(req))
5905 return io_req_prep_async(req);
5908 static u32 io_get_sequence(struct io_kiocb *req)
5910 struct io_kiocb *pos;
5911 struct io_ring_ctx *ctx = req->ctx;
5912 u32 total_submitted, nr_reqs = 0;
5914 io_for_each_link(pos, req)
5917 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5918 return total_submitted - nr_reqs;
5921 static int io_req_defer(struct io_kiocb *req)
5923 struct io_ring_ctx *ctx = req->ctx;
5924 struct io_defer_entry *de;
5928 /* Still need defer if there is pending req in defer list. */
5929 if (likely(list_empty_careful(&ctx->defer_list) &&
5930 !(req->flags & REQ_F_IO_DRAIN)))
5933 seq = io_get_sequence(req);
5934 /* Still a chance to pass the sequence check */
5935 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5938 ret = io_req_defer_prep(req);
5941 io_prep_async_link(req);
5942 de = kmalloc(sizeof(*de), GFP_KERNEL);
5946 spin_lock_irq(&ctx->completion_lock);
5947 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5948 spin_unlock_irq(&ctx->completion_lock);
5950 io_queue_async_work(req);
5951 return -EIOCBQUEUED;
5954 trace_io_uring_defer(ctx, req, req->user_data);
5957 list_add_tail(&de->list, &ctx->defer_list);
5958 spin_unlock_irq(&ctx->completion_lock);
5959 return -EIOCBQUEUED;
5962 static void __io_clean_op(struct io_kiocb *req)
5964 if (req->flags & REQ_F_BUFFER_SELECTED) {
5965 switch (req->opcode) {
5966 case IORING_OP_READV:
5967 case IORING_OP_READ_FIXED:
5968 case IORING_OP_READ:
5969 kfree((void *)(unsigned long)req->rw.addr);
5971 case IORING_OP_RECVMSG:
5972 case IORING_OP_RECV:
5973 kfree(req->sr_msg.kbuf);
5976 req->flags &= ~REQ_F_BUFFER_SELECTED;
5979 if (req->flags & REQ_F_NEED_CLEANUP) {
5980 switch (req->opcode) {
5981 case IORING_OP_READV:
5982 case IORING_OP_READ_FIXED:
5983 case IORING_OP_READ:
5984 case IORING_OP_WRITEV:
5985 case IORING_OP_WRITE_FIXED:
5986 case IORING_OP_WRITE: {
5987 struct io_async_rw *io = req->async_data;
5989 kfree(io->free_iovec);
5992 case IORING_OP_RECVMSG:
5993 case IORING_OP_SENDMSG: {
5994 struct io_async_msghdr *io = req->async_data;
5996 kfree(io->free_iov);
5999 case IORING_OP_SPLICE:
6001 io_put_file(req, req->splice.file_in,
6002 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6004 case IORING_OP_OPENAT:
6005 case IORING_OP_OPENAT2:
6006 if (req->open.filename)
6007 putname(req->open.filename);
6009 case IORING_OP_RENAMEAT:
6010 putname(req->rename.oldpath);
6011 putname(req->rename.newpath);
6013 case IORING_OP_UNLINKAT:
6014 putname(req->unlink.filename);
6017 req->flags &= ~REQ_F_NEED_CLEANUP;
6021 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6023 struct io_ring_ctx *ctx = req->ctx;
6024 const struct cred *creds = NULL;
6027 if (req->work.creds && req->work.creds != current_cred())
6028 creds = override_creds(req->work.creds);
6030 switch (req->opcode) {
6032 ret = io_nop(req, issue_flags);
6034 case IORING_OP_READV:
6035 case IORING_OP_READ_FIXED:
6036 case IORING_OP_READ:
6037 ret = io_read(req, issue_flags);
6039 case IORING_OP_WRITEV:
6040 case IORING_OP_WRITE_FIXED:
6041 case IORING_OP_WRITE:
6042 ret = io_write(req, issue_flags);
6044 case IORING_OP_FSYNC:
6045 ret = io_fsync(req, issue_flags);
6047 case IORING_OP_POLL_ADD:
6048 ret = io_poll_add(req, issue_flags);
6050 case IORING_OP_POLL_REMOVE:
6051 ret = io_poll_remove(req, issue_flags);
6053 case IORING_OP_SYNC_FILE_RANGE:
6054 ret = io_sync_file_range(req, issue_flags);
6056 case IORING_OP_SENDMSG:
6057 ret = io_sendmsg(req, issue_flags);
6059 case IORING_OP_SEND:
6060 ret = io_send(req, issue_flags);
6062 case IORING_OP_RECVMSG:
6063 ret = io_recvmsg(req, issue_flags);
6065 case IORING_OP_RECV:
6066 ret = io_recv(req, issue_flags);
6068 case IORING_OP_TIMEOUT:
6069 ret = io_timeout(req, issue_flags);
6071 case IORING_OP_TIMEOUT_REMOVE:
6072 ret = io_timeout_remove(req, issue_flags);
6074 case IORING_OP_ACCEPT:
6075 ret = io_accept(req, issue_flags);
6077 case IORING_OP_CONNECT:
6078 ret = io_connect(req, issue_flags);
6080 case IORING_OP_ASYNC_CANCEL:
6081 ret = io_async_cancel(req, issue_flags);
6083 case IORING_OP_FALLOCATE:
6084 ret = io_fallocate(req, issue_flags);
6086 case IORING_OP_OPENAT:
6087 ret = io_openat(req, issue_flags);
6089 case IORING_OP_CLOSE:
6090 ret = io_close(req, issue_flags);
6092 case IORING_OP_FILES_UPDATE:
6093 ret = io_files_update(req, issue_flags);
6095 case IORING_OP_STATX:
6096 ret = io_statx(req, issue_flags);
6098 case IORING_OP_FADVISE:
6099 ret = io_fadvise(req, issue_flags);
6101 case IORING_OP_MADVISE:
6102 ret = io_madvise(req, issue_flags);
6104 case IORING_OP_OPENAT2:
6105 ret = io_openat2(req, issue_flags);
6107 case IORING_OP_EPOLL_CTL:
6108 ret = io_epoll_ctl(req, issue_flags);
6110 case IORING_OP_SPLICE:
6111 ret = io_splice(req, issue_flags);
6113 case IORING_OP_PROVIDE_BUFFERS:
6114 ret = io_provide_buffers(req, issue_flags);
6116 case IORING_OP_REMOVE_BUFFERS:
6117 ret = io_remove_buffers(req, issue_flags);
6120 ret = io_tee(req, issue_flags);
6122 case IORING_OP_SHUTDOWN:
6123 ret = io_shutdown(req, issue_flags);
6125 case IORING_OP_RENAMEAT:
6126 ret = io_renameat(req, issue_flags);
6128 case IORING_OP_UNLINKAT:
6129 ret = io_unlinkat(req, issue_flags);
6137 revert_creds(creds);
6142 /* If the op doesn't have a file, we're not polling for it */
6143 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6144 const bool in_async = io_wq_current_is_worker();
6146 /* workqueue context doesn't hold uring_lock, grab it now */
6148 mutex_lock(&ctx->uring_lock);
6150 io_iopoll_req_issued(req, in_async);
6153 mutex_unlock(&ctx->uring_lock);
6159 static void io_wq_submit_work(struct io_wq_work *work)
6161 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6162 struct io_kiocb *timeout;
6165 timeout = io_prep_linked_timeout(req);
6167 io_queue_linked_timeout(timeout);
6169 if (work->flags & IO_WQ_WORK_CANCEL)
6174 ret = io_issue_sqe(req, 0);
6176 * We can get EAGAIN for polled IO even though we're
6177 * forcing a sync submission from here, since we can't
6178 * wait for request slots on the block side.
6186 /* avoid locking problems by failing it from a clean context */
6188 /* io-wq is going to take one down */
6189 refcount_inc(&req->refs);
6190 io_req_task_queue_fail(req, ret);
6194 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6197 struct fixed_rsrc_table *table;
6199 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6200 return table->files[index & IORING_FILE_TABLE_MASK];
6203 static struct file *io_file_get(struct io_submit_state *state,
6204 struct io_kiocb *req, int fd, bool fixed)
6206 struct io_ring_ctx *ctx = req->ctx;
6210 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6212 fd = array_index_nospec(fd, ctx->nr_user_files);
6213 file = io_file_from_index(ctx, fd);
6214 io_set_resource_node(req);
6216 trace_io_uring_file_get(ctx, fd);
6217 file = __io_file_get(state, fd);
6220 if (file && unlikely(file->f_op == &io_uring_fops))
6221 io_req_track_inflight(req);
6225 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6227 struct io_timeout_data *data = container_of(timer,
6228 struct io_timeout_data, timer);
6229 struct io_kiocb *prev, *req = data->req;
6230 struct io_ring_ctx *ctx = req->ctx;
6231 unsigned long flags;
6233 spin_lock_irqsave(&ctx->completion_lock, flags);
6234 prev = req->timeout.head;
6235 req->timeout.head = NULL;
6238 * We don't expect the list to be empty, that will only happen if we
6239 * race with the completion of the linked work.
6241 if (prev && refcount_inc_not_zero(&prev->refs))
6242 io_remove_next_linked(prev);
6245 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6248 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6249 io_put_req_deferred(prev, 1);
6251 io_req_complete_post(req, -ETIME, 0);
6252 io_put_req_deferred(req, 1);
6254 return HRTIMER_NORESTART;
6257 static void __io_queue_linked_timeout(struct io_kiocb *req)
6260 * If the back reference is NULL, then our linked request finished
6261 * before we got a chance to setup the timer
6263 if (req->timeout.head) {
6264 struct io_timeout_data *data = req->async_data;
6266 data->timer.function = io_link_timeout_fn;
6267 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6272 static void io_queue_linked_timeout(struct io_kiocb *req)
6274 struct io_ring_ctx *ctx = req->ctx;
6276 spin_lock_irq(&ctx->completion_lock);
6277 __io_queue_linked_timeout(req);
6278 spin_unlock_irq(&ctx->completion_lock);
6280 /* drop submission reference */
6284 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6286 struct io_kiocb *nxt = req->link;
6288 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6289 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6292 nxt->timeout.head = req;
6293 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6294 req->flags |= REQ_F_LINK_TIMEOUT;
6298 static void __io_queue_sqe(struct io_kiocb *req)
6300 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6303 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6306 * We async punt it if the file wasn't marked NOWAIT, or if the file
6307 * doesn't support non-blocking read/write attempts
6309 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6310 if (!io_arm_poll_handler(req)) {
6312 * Queued up for async execution, worker will release
6313 * submit reference when the iocb is actually submitted.
6315 io_queue_async_work(req);
6317 } else if (likely(!ret)) {
6318 /* drop submission reference */
6319 if (req->flags & REQ_F_COMPLETE_INLINE) {
6320 struct io_ring_ctx *ctx = req->ctx;
6321 struct io_comp_state *cs = &ctx->submit_state.comp;
6323 cs->reqs[cs->nr++] = req;
6324 if (cs->nr == ARRAY_SIZE(cs->reqs))
6325 io_submit_flush_completions(cs, ctx);
6330 req_set_fail_links(req);
6332 io_req_complete(req, ret);
6335 io_queue_linked_timeout(linked_timeout);
6338 static void io_queue_sqe(struct io_kiocb *req)
6342 ret = io_req_defer(req);
6344 if (ret != -EIOCBQUEUED) {
6346 req_set_fail_links(req);
6348 io_req_complete(req, ret);
6350 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6351 ret = io_req_defer_prep(req);
6354 io_queue_async_work(req);
6356 __io_queue_sqe(req);
6361 * Check SQE restrictions (opcode and flags).
6363 * Returns 'true' if SQE is allowed, 'false' otherwise.
6365 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6366 struct io_kiocb *req,
6367 unsigned int sqe_flags)
6369 if (!ctx->restricted)
6372 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6375 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6376 ctx->restrictions.sqe_flags_required)
6379 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6380 ctx->restrictions.sqe_flags_required))
6386 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6387 const struct io_uring_sqe *sqe)
6389 struct io_submit_state *state;
6390 unsigned int sqe_flags;
6391 int personality, ret = 0;
6393 req->opcode = READ_ONCE(sqe->opcode);
6394 /* same numerical values with corresponding REQ_F_*, safe to copy */
6395 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6396 req->user_data = READ_ONCE(sqe->user_data);
6397 req->async_data = NULL;
6401 req->fixed_rsrc_refs = NULL;
6402 /* one is dropped after submission, the other at completion */
6403 refcount_set(&req->refs, 2);
6404 req->task = current;
6406 req->work.list.next = NULL;
6407 req->work.creds = NULL;
6408 req->work.flags = 0;
6410 /* enforce forwards compatibility on users */
6411 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6416 if (unlikely(req->opcode >= IORING_OP_LAST))
6419 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6422 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6423 !io_op_defs[req->opcode].buffer_select)
6426 personality = READ_ONCE(sqe->personality);
6428 req->work.creds = xa_load(&ctx->personalities, personality);
6429 if (!req->work.creds)
6431 get_cred(req->work.creds);
6433 state = &ctx->submit_state;
6436 * Plug now if we have more than 1 IO left after this, and the target
6437 * is potentially a read/write to block based storage.
6439 if (!state->plug_started && state->ios_left > 1 &&
6440 io_op_defs[req->opcode].plug) {
6441 blk_start_plug(&state->plug);
6442 state->plug_started = true;
6445 if (io_op_defs[req->opcode].needs_file) {
6446 bool fixed = req->flags & REQ_F_FIXED_FILE;
6448 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6449 if (unlikely(!req->file))
6457 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6458 const struct io_uring_sqe *sqe)
6460 struct io_submit_link *link = &ctx->submit_state.link;
6463 ret = io_init_req(ctx, req, sqe);
6464 if (unlikely(ret)) {
6467 io_req_complete(req, ret);
6469 /* fail even hard links since we don't submit */
6470 link->head->flags |= REQ_F_FAIL_LINK;
6471 io_put_req(link->head);
6472 io_req_complete(link->head, -ECANCELED);
6477 ret = io_req_prep(req, sqe);
6481 /* don't need @sqe from now on */
6482 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6483 true, ctx->flags & IORING_SETUP_SQPOLL);
6486 * If we already have a head request, queue this one for async
6487 * submittal once the head completes. If we don't have a head but
6488 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6489 * submitted sync once the chain is complete. If none of those
6490 * conditions are true (normal request), then just queue it.
6493 struct io_kiocb *head = link->head;
6496 * Taking sequential execution of a link, draining both sides
6497 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6498 * requests in the link. So, it drains the head and the
6499 * next after the link request. The last one is done via
6500 * drain_next flag to persist the effect across calls.
6502 if (req->flags & REQ_F_IO_DRAIN) {
6503 head->flags |= REQ_F_IO_DRAIN;
6504 ctx->drain_next = 1;
6506 ret = io_req_defer_prep(req);
6509 trace_io_uring_link(ctx, req, head);
6510 link->last->link = req;
6513 /* last request of a link, enqueue the link */
6514 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6519 if (unlikely(ctx->drain_next)) {
6520 req->flags |= REQ_F_IO_DRAIN;
6521 ctx->drain_next = 0;
6523 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6535 * Batched submission is done, ensure local IO is flushed out.
6537 static void io_submit_state_end(struct io_submit_state *state,
6538 struct io_ring_ctx *ctx)
6540 if (state->link.head)
6541 io_queue_sqe(state->link.head);
6543 io_submit_flush_completions(&state->comp, ctx);
6544 if (state->plug_started)
6545 blk_finish_plug(&state->plug);
6546 io_state_file_put(state);
6550 * Start submission side cache.
6552 static void io_submit_state_start(struct io_submit_state *state,
6553 unsigned int max_ios)
6555 state->plug_started = false;
6556 state->ios_left = max_ios;
6557 /* set only head, no need to init link_last in advance */
6558 state->link.head = NULL;
6561 static void io_commit_sqring(struct io_ring_ctx *ctx)
6563 struct io_rings *rings = ctx->rings;
6566 * Ensure any loads from the SQEs are done at this point,
6567 * since once we write the new head, the application could
6568 * write new data to them.
6570 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6574 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6575 * that is mapped by userspace. This means that care needs to be taken to
6576 * ensure that reads are stable, as we cannot rely on userspace always
6577 * being a good citizen. If members of the sqe are validated and then later
6578 * used, it's important that those reads are done through READ_ONCE() to
6579 * prevent a re-load down the line.
6581 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6583 u32 *sq_array = ctx->sq_array;
6587 * The cached sq head (or cq tail) serves two purposes:
6589 * 1) allows us to batch the cost of updating the user visible
6591 * 2) allows the kernel side to track the head on its own, even
6592 * though the application is the one updating it.
6594 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6595 if (likely(head < ctx->sq_entries))
6596 return &ctx->sq_sqes[head];
6598 /* drop invalid entries */
6599 ctx->cached_sq_dropped++;
6600 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6604 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6608 /* if we have a backlog and couldn't flush it all, return BUSY */
6609 if (test_bit(0, &ctx->sq_check_overflow)) {
6610 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6614 /* make sure SQ entry isn't read before tail */
6615 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6617 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6620 percpu_counter_add(¤t->io_uring->inflight, nr);
6621 refcount_add(nr, ¤t->usage);
6622 io_submit_state_start(&ctx->submit_state, nr);
6624 while (submitted < nr) {
6625 const struct io_uring_sqe *sqe;
6626 struct io_kiocb *req;
6628 req = io_alloc_req(ctx);
6629 if (unlikely(!req)) {
6631 submitted = -EAGAIN;
6634 sqe = io_get_sqe(ctx);
6635 if (unlikely(!sqe)) {
6636 kmem_cache_free(req_cachep, req);
6639 /* will complete beyond this point, count as submitted */
6641 if (io_submit_sqe(ctx, req, sqe))
6645 if (unlikely(submitted != nr)) {
6646 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6647 struct io_uring_task *tctx = current->io_uring;
6648 int unused = nr - ref_used;
6650 percpu_ref_put_many(&ctx->refs, unused);
6651 percpu_counter_sub(&tctx->inflight, unused);
6652 put_task_struct_many(current, unused);
6655 io_submit_state_end(&ctx->submit_state, ctx);
6656 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6657 io_commit_sqring(ctx);
6662 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6664 /* Tell userspace we may need a wakeup call */
6665 spin_lock_irq(&ctx->completion_lock);
6666 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6667 spin_unlock_irq(&ctx->completion_lock);
6670 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6672 spin_lock_irq(&ctx->completion_lock);
6673 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6674 spin_unlock_irq(&ctx->completion_lock);
6677 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6679 unsigned int to_submit;
6682 to_submit = io_sqring_entries(ctx);
6683 /* if we're handling multiple rings, cap submit size for fairness */
6684 if (cap_entries && to_submit > 8)
6687 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6688 unsigned nr_events = 0;
6690 mutex_lock(&ctx->uring_lock);
6691 if (!list_empty(&ctx->iopoll_list))
6692 io_do_iopoll(ctx, &nr_events, 0);
6694 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6695 !(ctx->flags & IORING_SETUP_R_DISABLED))
6696 ret = io_submit_sqes(ctx, to_submit);
6697 mutex_unlock(&ctx->uring_lock);
6700 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6701 wake_up(&ctx->sqo_sq_wait);
6706 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6708 struct io_ring_ctx *ctx;
6709 unsigned sq_thread_idle = 0;
6711 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6712 if (sq_thread_idle < ctx->sq_thread_idle)
6713 sq_thread_idle = ctx->sq_thread_idle;
6716 sqd->sq_thread_idle = sq_thread_idle;
6719 static int io_sq_thread(void *data)
6721 struct io_sq_data *sqd = data;
6722 struct io_ring_ctx *ctx;
6723 unsigned long timeout = 0;
6724 char buf[TASK_COMM_LEN];
6727 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6728 set_task_comm(current, buf);
6729 current->pf_io_worker = NULL;
6731 if (sqd->sq_cpu != -1)
6732 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6734 set_cpus_allowed_ptr(current, cpu_online_mask);
6735 current->flags |= PF_NO_SETAFFINITY;
6737 mutex_lock(&sqd->lock);
6738 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6740 bool cap_entries, sqt_spin, needs_sched;
6742 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6743 mutex_unlock(&sqd->lock);
6745 mutex_lock(&sqd->lock);
6747 io_run_task_work_head(&sqd->park_task_work);
6748 timeout = jiffies + sqd->sq_thread_idle;
6751 if (fatal_signal_pending(current))
6754 cap_entries = !list_is_singular(&sqd->ctx_list);
6755 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6756 const struct cred *creds = NULL;
6758 if (ctx->sq_creds != current_cred())
6759 creds = override_creds(ctx->sq_creds);
6760 ret = __io_sq_thread(ctx, cap_entries);
6762 revert_creds(creds);
6763 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6767 if (sqt_spin || !time_after(jiffies, timeout)) {
6771 timeout = jiffies + sqd->sq_thread_idle;
6776 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6777 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6778 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6779 !list_empty_careful(&ctx->iopoll_list)) {
6780 needs_sched = false;
6783 if (io_sqring_entries(ctx)) {
6784 needs_sched = false;
6789 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6790 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6791 io_ring_set_wakeup_flag(ctx);
6793 mutex_unlock(&sqd->lock);
6796 mutex_lock(&sqd->lock);
6797 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6798 io_ring_clear_wakeup_flag(ctx);
6801 finish_wait(&sqd->wait, &wait);
6802 io_run_task_work_head(&sqd->park_task_work);
6803 timeout = jiffies + sqd->sq_thread_idle;
6806 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6807 io_uring_cancel_sqpoll(ctx);
6809 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6810 io_ring_set_wakeup_flag(ctx);
6811 mutex_unlock(&sqd->lock);
6814 io_run_task_work_head(&sqd->park_task_work);
6815 complete(&sqd->exited);
6819 struct io_wait_queue {
6820 struct wait_queue_entry wq;
6821 struct io_ring_ctx *ctx;
6823 unsigned nr_timeouts;
6826 static inline bool io_should_wake(struct io_wait_queue *iowq)
6828 struct io_ring_ctx *ctx = iowq->ctx;
6831 * Wake up if we have enough events, or if a timeout occurred since we
6832 * started waiting. For timeouts, we always want to return to userspace,
6833 * regardless of event count.
6835 return io_cqring_events(ctx) >= iowq->to_wait ||
6836 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6839 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6840 int wake_flags, void *key)
6842 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6846 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6847 * the task, and the next invocation will do it.
6849 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6850 return autoremove_wake_function(curr, mode, wake_flags, key);
6854 static int io_run_task_work_sig(void)
6856 if (io_run_task_work())
6858 if (!signal_pending(current))
6860 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6861 return -ERESTARTSYS;
6865 /* when returns >0, the caller should retry */
6866 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6867 struct io_wait_queue *iowq,
6868 signed long *timeout)
6872 /* make sure we run task_work before checking for signals */
6873 ret = io_run_task_work_sig();
6874 if (ret || io_should_wake(iowq))
6876 /* let the caller flush overflows, retry */
6877 if (test_bit(0, &ctx->cq_check_overflow))
6880 *timeout = schedule_timeout(*timeout);
6881 return !*timeout ? -ETIME : 1;
6885 * Wait until events become available, if we don't already have some. The
6886 * application must reap them itself, as they reside on the shared cq ring.
6888 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6889 const sigset_t __user *sig, size_t sigsz,
6890 struct __kernel_timespec __user *uts)
6892 struct io_wait_queue iowq = {
6895 .func = io_wake_function,
6896 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6899 .to_wait = min_events,
6901 struct io_rings *rings = ctx->rings;
6902 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6906 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6907 if (io_cqring_events(ctx) >= min_events)
6909 if (!io_run_task_work())
6914 #ifdef CONFIG_COMPAT
6915 if (in_compat_syscall())
6916 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6920 ret = set_user_sigmask(sig, sigsz);
6927 struct timespec64 ts;
6929 if (get_timespec64(&ts, uts))
6931 timeout = timespec64_to_jiffies(&ts);
6934 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6935 trace_io_uring_cqring_wait(ctx, min_events);
6937 /* if we can't even flush overflow, don't wait for more */
6938 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6942 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6943 TASK_INTERRUPTIBLE);
6944 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6945 finish_wait(&ctx->wait, &iowq.wq);
6949 restore_saved_sigmask_unless(ret == -EINTR);
6951 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6954 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6956 #if defined(CONFIG_UNIX)
6957 if (ctx->ring_sock) {
6958 struct sock *sock = ctx->ring_sock->sk;
6959 struct sk_buff *skb;
6961 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6967 for (i = 0; i < ctx->nr_user_files; i++) {
6970 file = io_file_from_index(ctx, i);
6977 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6979 struct fixed_rsrc_data *data;
6981 data = container_of(ref, struct fixed_rsrc_data, refs);
6982 complete(&data->done);
6985 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6987 spin_lock_bh(&ctx->rsrc_ref_lock);
6990 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6992 spin_unlock_bh(&ctx->rsrc_ref_lock);
6995 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6996 struct fixed_rsrc_data *rsrc_data,
6997 struct fixed_rsrc_ref_node *ref_node)
6999 io_rsrc_ref_lock(ctx);
7000 rsrc_data->node = ref_node;
7001 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
7002 io_rsrc_ref_unlock(ctx);
7003 percpu_ref_get(&rsrc_data->refs);
7006 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
7008 struct fixed_rsrc_ref_node *ref_node = NULL;
7010 io_rsrc_ref_lock(ctx);
7011 ref_node = data->node;
7013 io_rsrc_ref_unlock(ctx);
7015 percpu_ref_kill(&ref_node->refs);
7018 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7019 struct io_ring_ctx *ctx,
7020 void (*rsrc_put)(struct io_ring_ctx *ctx,
7021 struct io_rsrc_put *prsrc))
7023 struct fixed_rsrc_ref_node *backup_node;
7029 data->quiesce = true;
7032 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7035 backup_node->rsrc_data = data;
7036 backup_node->rsrc_put = rsrc_put;
7038 io_sqe_rsrc_kill_node(ctx, data);
7039 percpu_ref_kill(&data->refs);
7040 flush_delayed_work(&ctx->rsrc_put_work);
7042 ret = wait_for_completion_interruptible(&data->done);
7046 percpu_ref_resurrect(&data->refs);
7047 io_sqe_rsrc_set_node(ctx, data, backup_node);
7049 reinit_completion(&data->done);
7050 mutex_unlock(&ctx->uring_lock);
7051 ret = io_run_task_work_sig();
7052 mutex_lock(&ctx->uring_lock);
7054 data->quiesce = false;
7057 destroy_fixed_rsrc_ref_node(backup_node);
7061 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7063 struct fixed_rsrc_data *data;
7065 data = kzalloc(sizeof(*data), GFP_KERNEL);
7069 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7070 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7075 init_completion(&data->done);
7079 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7081 percpu_ref_exit(&data->refs);
7086 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7088 struct fixed_rsrc_data *data = ctx->file_data;
7089 unsigned nr_tables, i;
7093 * percpu_ref_is_dying() is to stop parallel files unregister
7094 * Since we possibly drop uring lock later in this function to
7097 if (!data || percpu_ref_is_dying(&data->refs))
7099 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7103 __io_sqe_files_unregister(ctx);
7104 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7105 for (i = 0; i < nr_tables; i++)
7106 kfree(data->table[i].files);
7107 free_fixed_rsrc_data(data);
7108 ctx->file_data = NULL;
7109 ctx->nr_user_files = 0;
7113 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7114 __releases(&sqd->lock)
7116 WARN_ON_ONCE(sqd->thread == current);
7119 * Do the dance but not conditional clear_bit() because it'd race with
7120 * other threads incrementing park_pending and setting the bit.
7122 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7123 if (atomic_dec_return(&sqd->park_pending))
7124 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7125 mutex_unlock(&sqd->lock);
7128 static void io_sq_thread_park(struct io_sq_data *sqd)
7129 __acquires(&sqd->lock)
7131 WARN_ON_ONCE(sqd->thread == current);
7133 atomic_inc(&sqd->park_pending);
7134 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7135 mutex_lock(&sqd->lock);
7137 wake_up_process(sqd->thread);
7140 static void io_sq_thread_stop(struct io_sq_data *sqd)
7142 WARN_ON_ONCE(sqd->thread == current);
7144 mutex_lock(&sqd->lock);
7145 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7147 wake_up_process(sqd->thread);
7148 mutex_unlock(&sqd->lock);
7149 wait_for_completion(&sqd->exited);
7152 static void io_put_sq_data(struct io_sq_data *sqd)
7154 if (refcount_dec_and_test(&sqd->refs)) {
7155 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7157 io_sq_thread_stop(sqd);
7162 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7164 struct io_sq_data *sqd = ctx->sq_data;
7167 io_sq_thread_park(sqd);
7168 list_del_init(&ctx->sqd_list);
7169 io_sqd_update_thread_idle(sqd);
7170 io_sq_thread_unpark(sqd);
7172 io_put_sq_data(sqd);
7173 ctx->sq_data = NULL;
7175 put_cred(ctx->sq_creds);
7179 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7181 struct io_ring_ctx *ctx_attach;
7182 struct io_sq_data *sqd;
7185 f = fdget(p->wq_fd);
7187 return ERR_PTR(-ENXIO);
7188 if (f.file->f_op != &io_uring_fops) {
7190 return ERR_PTR(-EINVAL);
7193 ctx_attach = f.file->private_data;
7194 sqd = ctx_attach->sq_data;
7197 return ERR_PTR(-EINVAL);
7199 if (sqd->task_tgid != current->tgid) {
7201 return ERR_PTR(-EPERM);
7204 refcount_inc(&sqd->refs);
7209 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7212 struct io_sq_data *sqd;
7215 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7216 sqd = io_attach_sq_data(p);
7221 /* fall through for EPERM case, setup new sqd/task */
7222 if (PTR_ERR(sqd) != -EPERM)
7226 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7228 return ERR_PTR(-ENOMEM);
7230 atomic_set(&sqd->park_pending, 0);
7231 refcount_set(&sqd->refs, 1);
7232 INIT_LIST_HEAD(&sqd->ctx_list);
7233 mutex_init(&sqd->lock);
7234 init_waitqueue_head(&sqd->wait);
7235 init_completion(&sqd->exited);
7239 #if defined(CONFIG_UNIX)
7241 * Ensure the UNIX gc is aware of our file set, so we are certain that
7242 * the io_uring can be safely unregistered on process exit, even if we have
7243 * loops in the file referencing.
7245 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7247 struct sock *sk = ctx->ring_sock->sk;
7248 struct scm_fp_list *fpl;
7249 struct sk_buff *skb;
7252 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7256 skb = alloc_skb(0, GFP_KERNEL);
7265 fpl->user = get_uid(current_user());
7266 for (i = 0; i < nr; i++) {
7267 struct file *file = io_file_from_index(ctx, i + offset);
7271 fpl->fp[nr_files] = get_file(file);
7272 unix_inflight(fpl->user, fpl->fp[nr_files]);
7277 fpl->max = SCM_MAX_FD;
7278 fpl->count = nr_files;
7279 UNIXCB(skb).fp = fpl;
7280 skb->destructor = unix_destruct_scm;
7281 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7282 skb_queue_head(&sk->sk_receive_queue, skb);
7284 for (i = 0; i < nr_files; i++)
7295 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7296 * causes regular reference counting to break down. We rely on the UNIX
7297 * garbage collection to take care of this problem for us.
7299 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7301 unsigned left, total;
7305 left = ctx->nr_user_files;
7307 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7309 ret = __io_sqe_files_scm(ctx, this_files, total);
7313 total += this_files;
7319 while (total < ctx->nr_user_files) {
7320 struct file *file = io_file_from_index(ctx, total);
7330 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7336 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7337 unsigned nr_tables, unsigned nr_files)
7341 for (i = 0; i < nr_tables; i++) {
7342 struct fixed_rsrc_table *table = &file_data->table[i];
7343 unsigned this_files;
7345 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7346 table->files = kcalloc(this_files, sizeof(struct file *),
7350 nr_files -= this_files;
7356 for (i = 0; i < nr_tables; i++) {
7357 struct fixed_rsrc_table *table = &file_data->table[i];
7358 kfree(table->files);
7363 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7365 struct file *file = prsrc->file;
7366 #if defined(CONFIG_UNIX)
7367 struct sock *sock = ctx->ring_sock->sk;
7368 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7369 struct sk_buff *skb;
7372 __skb_queue_head_init(&list);
7375 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7376 * remove this entry and rearrange the file array.
7378 skb = skb_dequeue(head);
7380 struct scm_fp_list *fp;
7382 fp = UNIXCB(skb).fp;
7383 for (i = 0; i < fp->count; i++) {
7386 if (fp->fp[i] != file)
7389 unix_notinflight(fp->user, fp->fp[i]);
7390 left = fp->count - 1 - i;
7392 memmove(&fp->fp[i], &fp->fp[i + 1],
7393 left * sizeof(struct file *));
7400 __skb_queue_tail(&list, skb);
7410 __skb_queue_tail(&list, skb);
7412 skb = skb_dequeue(head);
7415 if (skb_peek(&list)) {
7416 spin_lock_irq(&head->lock);
7417 while ((skb = __skb_dequeue(&list)) != NULL)
7418 __skb_queue_tail(head, skb);
7419 spin_unlock_irq(&head->lock);
7426 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7428 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7429 struct io_ring_ctx *ctx = rsrc_data->ctx;
7430 struct io_rsrc_put *prsrc, *tmp;
7432 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7433 list_del(&prsrc->list);
7434 ref_node->rsrc_put(ctx, prsrc);
7438 percpu_ref_exit(&ref_node->refs);
7440 percpu_ref_put(&rsrc_data->refs);
7443 static void io_rsrc_put_work(struct work_struct *work)
7445 struct io_ring_ctx *ctx;
7446 struct llist_node *node;
7448 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7449 node = llist_del_all(&ctx->rsrc_put_llist);
7452 struct fixed_rsrc_ref_node *ref_node;
7453 struct llist_node *next = node->next;
7455 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7456 __io_rsrc_put_work(ref_node);
7461 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7464 struct fixed_rsrc_table *table;
7466 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7467 return &table->files[i & IORING_FILE_TABLE_MASK];
7470 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7472 struct fixed_rsrc_ref_node *ref_node;
7473 struct fixed_rsrc_data *data;
7474 struct io_ring_ctx *ctx;
7475 bool first_add = false;
7478 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7479 data = ref_node->rsrc_data;
7482 io_rsrc_ref_lock(ctx);
7483 ref_node->done = true;
7485 while (!list_empty(&ctx->rsrc_ref_list)) {
7486 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7487 struct fixed_rsrc_ref_node, node);
7488 /* recycle ref nodes in order */
7489 if (!ref_node->done)
7491 list_del(&ref_node->node);
7492 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7494 io_rsrc_ref_unlock(ctx);
7496 if (percpu_ref_is_dying(&data->refs))
7500 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7502 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7505 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7506 struct io_ring_ctx *ctx)
7508 struct fixed_rsrc_ref_node *ref_node;
7510 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7514 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7519 INIT_LIST_HEAD(&ref_node->node);
7520 INIT_LIST_HEAD(&ref_node->rsrc_list);
7521 ref_node->done = false;
7525 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7526 struct fixed_rsrc_ref_node *ref_node)
7528 ref_node->rsrc_data = ctx->file_data;
7529 ref_node->rsrc_put = io_ring_file_put;
7532 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7534 percpu_ref_exit(&ref_node->refs);
7539 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7542 __s32 __user *fds = (__s32 __user *) arg;
7543 unsigned nr_tables, i;
7545 int fd, ret = -ENOMEM;
7546 struct fixed_rsrc_ref_node *ref_node;
7547 struct fixed_rsrc_data *file_data;
7553 if (nr_args > IORING_MAX_FIXED_FILES)
7556 file_data = alloc_fixed_rsrc_data(ctx);
7559 ctx->file_data = file_data;
7561 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7562 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7564 if (!file_data->table)
7567 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7570 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7571 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7575 /* allow sparse sets */
7585 * Don't allow io_uring instances to be registered. If UNIX
7586 * isn't enabled, then this causes a reference cycle and this
7587 * instance can never get freed. If UNIX is enabled we'll
7588 * handle it just fine, but there's still no point in allowing
7589 * a ring fd as it doesn't support regular read/write anyway.
7591 if (file->f_op == &io_uring_fops) {
7595 *io_fixed_file_slot(file_data, i) = file;
7598 ret = io_sqe_files_scm(ctx);
7600 io_sqe_files_unregister(ctx);
7604 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7606 io_sqe_files_unregister(ctx);
7609 init_fixed_file_ref_node(ctx, ref_node);
7611 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7614 for (i = 0; i < ctx->nr_user_files; i++) {
7615 file = io_file_from_index(ctx, i);
7619 for (i = 0; i < nr_tables; i++)
7620 kfree(file_data->table[i].files);
7621 ctx->nr_user_files = 0;
7623 free_fixed_rsrc_data(ctx->file_data);
7624 ctx->file_data = NULL;
7628 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7631 #if defined(CONFIG_UNIX)
7632 struct sock *sock = ctx->ring_sock->sk;
7633 struct sk_buff_head *head = &sock->sk_receive_queue;
7634 struct sk_buff *skb;
7637 * See if we can merge this file into an existing skb SCM_RIGHTS
7638 * file set. If there's no room, fall back to allocating a new skb
7639 * and filling it in.
7641 spin_lock_irq(&head->lock);
7642 skb = skb_peek(head);
7644 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7646 if (fpl->count < SCM_MAX_FD) {
7647 __skb_unlink(skb, head);
7648 spin_unlock_irq(&head->lock);
7649 fpl->fp[fpl->count] = get_file(file);
7650 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7652 spin_lock_irq(&head->lock);
7653 __skb_queue_head(head, skb);
7658 spin_unlock_irq(&head->lock);
7665 return __io_sqe_files_scm(ctx, 1, index);
7671 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7673 struct io_rsrc_put *prsrc;
7674 struct fixed_rsrc_ref_node *ref_node = data->node;
7676 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7681 list_add(&prsrc->list, &ref_node->rsrc_list);
7686 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7689 return io_queue_rsrc_removal(data, (void *)file);
7692 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7693 struct io_uring_rsrc_update *up,
7696 struct fixed_rsrc_data *data = ctx->file_data;
7697 struct fixed_rsrc_ref_node *ref_node;
7698 struct file *file, **file_slot;
7702 bool needs_switch = false;
7704 if (check_add_overflow(up->offset, nr_args, &done))
7706 if (done > ctx->nr_user_files)
7709 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7712 init_fixed_file_ref_node(ctx, ref_node);
7714 fds = u64_to_user_ptr(up->data);
7715 for (done = 0; done < nr_args; done++) {
7717 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7721 if (fd == IORING_REGISTER_FILES_SKIP)
7724 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7725 file_slot = io_fixed_file_slot(ctx->file_data, i);
7728 err = io_queue_file_removal(data, *file_slot);
7732 needs_switch = true;
7741 * Don't allow io_uring instances to be registered. If
7742 * UNIX isn't enabled, then this causes a reference
7743 * cycle and this instance can never get freed. If UNIX
7744 * is enabled we'll handle it just fine, but there's
7745 * still no point in allowing a ring fd as it doesn't
7746 * support regular read/write anyway.
7748 if (file->f_op == &io_uring_fops) {
7754 err = io_sqe_file_register(ctx, file, i);
7764 percpu_ref_kill(&data->node->refs);
7765 io_sqe_rsrc_set_node(ctx, data, ref_node);
7767 destroy_fixed_rsrc_ref_node(ref_node);
7769 return done ? done : err;
7772 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7775 struct io_uring_rsrc_update up;
7777 if (!ctx->file_data)
7781 if (copy_from_user(&up, arg, sizeof(up)))
7786 return __io_sqe_files_update(ctx, &up, nr_args);
7789 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7791 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7793 req = io_put_req_find_next(req);
7794 return req ? &req->work : NULL;
7797 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7799 struct io_wq_hash *hash;
7800 struct io_wq_data data;
7801 unsigned int concurrency;
7803 hash = ctx->hash_map;
7805 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7807 return ERR_PTR(-ENOMEM);
7808 refcount_set(&hash->refs, 1);
7809 init_waitqueue_head(&hash->wait);
7810 ctx->hash_map = hash;
7814 data.free_work = io_free_work;
7815 data.do_work = io_wq_submit_work;
7817 /* Do QD, or 4 * CPUS, whatever is smallest */
7818 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7820 return io_wq_create(concurrency, &data);
7823 static int io_uring_alloc_task_context(struct task_struct *task,
7824 struct io_ring_ctx *ctx)
7826 struct io_uring_task *tctx;
7829 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7830 if (unlikely(!tctx))
7833 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7834 if (unlikely(ret)) {
7839 tctx->io_wq = io_init_wq_offload(ctx);
7840 if (IS_ERR(tctx->io_wq)) {
7841 ret = PTR_ERR(tctx->io_wq);
7842 percpu_counter_destroy(&tctx->inflight);
7848 init_waitqueue_head(&tctx->wait);
7850 atomic_set(&tctx->in_idle, 0);
7851 task->io_uring = tctx;
7852 spin_lock_init(&tctx->task_lock);
7853 INIT_WQ_LIST(&tctx->task_list);
7854 tctx->task_state = 0;
7855 init_task_work(&tctx->task_work, tctx_task_work);
7859 void __io_uring_free(struct task_struct *tsk)
7861 struct io_uring_task *tctx = tsk->io_uring;
7863 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7864 WARN_ON_ONCE(tctx->io_wq);
7866 percpu_counter_destroy(&tctx->inflight);
7868 tsk->io_uring = NULL;
7871 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7872 struct io_uring_params *p)
7876 /* Retain compatibility with failing for an invalid attach attempt */
7877 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7878 IORING_SETUP_ATTACH_WQ) {
7881 f = fdget(p->wq_fd);
7884 if (f.file->f_op != &io_uring_fops) {
7890 if (ctx->flags & IORING_SETUP_SQPOLL) {
7891 struct task_struct *tsk;
7892 struct io_sq_data *sqd;
7896 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7899 sqd = io_get_sq_data(p, &attached);
7905 ctx->sq_creds = get_current_cred();
7907 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7908 if (!ctx->sq_thread_idle)
7909 ctx->sq_thread_idle = HZ;
7912 io_sq_thread_park(sqd);
7913 /* don't attach to a dying SQPOLL thread, would be racy */
7914 if (attached && !sqd->thread) {
7917 list_add(&ctx->sqd_list, &sqd->ctx_list);
7918 io_sqd_update_thread_idle(sqd);
7920 io_sq_thread_unpark(sqd);
7923 io_put_sq_data(sqd);
7924 ctx->sq_data = NULL;
7926 } else if (attached) {
7930 if (p->flags & IORING_SETUP_SQ_AFF) {
7931 int cpu = p->sq_thread_cpu;
7934 if (cpu >= nr_cpu_ids)
7936 if (!cpu_online(cpu))
7944 sqd->task_pid = current->pid;
7945 sqd->task_tgid = current->tgid;
7946 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7953 ret = io_uring_alloc_task_context(tsk, ctx);
7954 wake_up_new_task(tsk);
7957 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7958 /* Can't have SQ_AFF without SQPOLL */
7965 io_sq_thread_finish(ctx);
7968 complete(&ctx->sq_data->exited);
7972 static inline void __io_unaccount_mem(struct user_struct *user,
7973 unsigned long nr_pages)
7975 atomic_long_sub(nr_pages, &user->locked_vm);
7978 static inline int __io_account_mem(struct user_struct *user,
7979 unsigned long nr_pages)
7981 unsigned long page_limit, cur_pages, new_pages;
7983 /* Don't allow more pages than we can safely lock */
7984 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7987 cur_pages = atomic_long_read(&user->locked_vm);
7988 new_pages = cur_pages + nr_pages;
7989 if (new_pages > page_limit)
7991 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7992 new_pages) != cur_pages);
7997 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8000 __io_unaccount_mem(ctx->user, nr_pages);
8002 if (ctx->mm_account)
8003 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8006 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8011 ret = __io_account_mem(ctx->user, nr_pages);
8016 if (ctx->mm_account)
8017 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8022 static void io_mem_free(void *ptr)
8029 page = virt_to_head_page(ptr);
8030 if (put_page_testzero(page))
8031 free_compound_page(page);
8034 static void *io_mem_alloc(size_t size)
8036 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8037 __GFP_NORETRY | __GFP_ACCOUNT;
8039 return (void *) __get_free_pages(gfp_flags, get_order(size));
8042 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8045 struct io_rings *rings;
8046 size_t off, sq_array_size;
8048 off = struct_size(rings, cqes, cq_entries);
8049 if (off == SIZE_MAX)
8053 off = ALIGN(off, SMP_CACHE_BYTES);
8061 sq_array_size = array_size(sizeof(u32), sq_entries);
8062 if (sq_array_size == SIZE_MAX)
8065 if (check_add_overflow(off, sq_array_size, &off))
8071 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8075 if (!ctx->user_bufs)
8078 for (i = 0; i < ctx->nr_user_bufs; i++) {
8079 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8081 for (j = 0; j < imu->nr_bvecs; j++)
8082 unpin_user_page(imu->bvec[j].bv_page);
8084 if (imu->acct_pages)
8085 io_unaccount_mem(ctx, imu->acct_pages);
8090 kfree(ctx->user_bufs);
8091 ctx->user_bufs = NULL;
8092 ctx->nr_user_bufs = 0;
8096 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8097 void __user *arg, unsigned index)
8099 struct iovec __user *src;
8101 #ifdef CONFIG_COMPAT
8103 struct compat_iovec __user *ciovs;
8104 struct compat_iovec ciov;
8106 ciovs = (struct compat_iovec __user *) arg;
8107 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8110 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8111 dst->iov_len = ciov.iov_len;
8115 src = (struct iovec __user *) arg;
8116 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8122 * Not super efficient, but this is just a registration time. And we do cache
8123 * the last compound head, so generally we'll only do a full search if we don't
8126 * We check if the given compound head page has already been accounted, to
8127 * avoid double accounting it. This allows us to account the full size of the
8128 * page, not just the constituent pages of a huge page.
8130 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8131 int nr_pages, struct page *hpage)
8135 /* check current page array */
8136 for (i = 0; i < nr_pages; i++) {
8137 if (!PageCompound(pages[i]))
8139 if (compound_head(pages[i]) == hpage)
8143 /* check previously registered pages */
8144 for (i = 0; i < ctx->nr_user_bufs; i++) {
8145 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8147 for (j = 0; j < imu->nr_bvecs; j++) {
8148 if (!PageCompound(imu->bvec[j].bv_page))
8150 if (compound_head(imu->bvec[j].bv_page) == hpage)
8158 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8159 int nr_pages, struct io_mapped_ubuf *imu,
8160 struct page **last_hpage)
8164 for (i = 0; i < nr_pages; i++) {
8165 if (!PageCompound(pages[i])) {
8170 hpage = compound_head(pages[i]);
8171 if (hpage == *last_hpage)
8173 *last_hpage = hpage;
8174 if (headpage_already_acct(ctx, pages, i, hpage))
8176 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8180 if (!imu->acct_pages)
8183 ret = io_account_mem(ctx, imu->acct_pages);
8185 imu->acct_pages = 0;
8189 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8190 struct io_mapped_ubuf *imu,
8191 struct page **last_hpage)
8193 struct vm_area_struct **vmas = NULL;
8194 struct page **pages = NULL;
8195 unsigned long off, start, end, ubuf;
8197 int ret, pret, nr_pages, i;
8199 ubuf = (unsigned long) iov->iov_base;
8200 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8201 start = ubuf >> PAGE_SHIFT;
8202 nr_pages = end - start;
8206 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8210 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8215 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8221 mmap_read_lock(current->mm);
8222 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8224 if (pret == nr_pages) {
8225 /* don't support file backed memory */
8226 for (i = 0; i < nr_pages; i++) {
8227 struct vm_area_struct *vma = vmas[i];
8230 !is_file_hugepages(vma->vm_file)) {
8236 ret = pret < 0 ? pret : -EFAULT;
8238 mmap_read_unlock(current->mm);
8241 * if we did partial map, or found file backed vmas,
8242 * release any pages we did get
8245 unpin_user_pages(pages, pret);
8250 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8252 unpin_user_pages(pages, pret);
8257 off = ubuf & ~PAGE_MASK;
8258 size = iov->iov_len;
8259 for (i = 0; i < nr_pages; i++) {
8262 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8263 imu->bvec[i].bv_page = pages[i];
8264 imu->bvec[i].bv_len = vec_len;
8265 imu->bvec[i].bv_offset = off;
8269 /* store original address for later verification */
8271 imu->len = iov->iov_len;
8272 imu->nr_bvecs = nr_pages;
8280 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8284 if (!nr_args || nr_args > UIO_MAXIOV)
8287 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8289 if (!ctx->user_bufs)
8295 static int io_buffer_validate(struct iovec *iov)
8298 * Don't impose further limits on the size and buffer
8299 * constraints here, we'll -EINVAL later when IO is
8300 * submitted if they are wrong.
8302 if (!iov->iov_base || !iov->iov_len)
8305 /* arbitrary limit, but we need something */
8306 if (iov->iov_len > SZ_1G)
8312 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8313 unsigned int nr_args)
8317 struct page *last_hpage = NULL;
8319 ret = io_buffers_map_alloc(ctx, nr_args);
8323 for (i = 0; i < nr_args; i++) {
8324 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8326 ret = io_copy_iov(ctx, &iov, arg, i);
8330 ret = io_buffer_validate(&iov);
8334 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8338 ctx->nr_user_bufs++;
8342 io_sqe_buffers_unregister(ctx);
8347 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8349 __s32 __user *fds = arg;
8355 if (copy_from_user(&fd, fds, sizeof(*fds)))
8358 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8359 if (IS_ERR(ctx->cq_ev_fd)) {
8360 int ret = PTR_ERR(ctx->cq_ev_fd);
8361 ctx->cq_ev_fd = NULL;
8368 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8370 if (ctx->cq_ev_fd) {
8371 eventfd_ctx_put(ctx->cq_ev_fd);
8372 ctx->cq_ev_fd = NULL;
8379 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8381 struct io_buffer *buf;
8382 unsigned long index;
8384 xa_for_each(&ctx->io_buffers, index, buf)
8385 __io_remove_buffers(ctx, buf, index, -1U);
8388 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8390 struct io_kiocb *req, *nxt;
8392 list_for_each_entry_safe(req, nxt, list, compl.list) {
8393 if (tsk && req->task != tsk)
8395 list_del(&req->compl.list);
8396 kmem_cache_free(req_cachep, req);
8400 static void io_req_caches_free(struct io_ring_ctx *ctx)
8402 struct io_submit_state *submit_state = &ctx->submit_state;
8403 struct io_comp_state *cs = &ctx->submit_state.comp;
8405 mutex_lock(&ctx->uring_lock);
8407 if (submit_state->free_reqs) {
8408 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8409 submit_state->reqs);
8410 submit_state->free_reqs = 0;
8413 spin_lock_irq(&ctx->completion_lock);
8414 list_splice_init(&cs->locked_free_list, &cs->free_list);
8415 cs->locked_free_nr = 0;
8416 spin_unlock_irq(&ctx->completion_lock);
8418 io_req_cache_free(&cs->free_list, NULL);
8420 mutex_unlock(&ctx->uring_lock);
8423 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8426 * Some may use context even when all refs and requests have been put,
8427 * and they are free to do so while still holding uring_lock or
8428 * completion_lock, see __io_req_task_submit(). Wait for them to finish.
8430 mutex_lock(&ctx->uring_lock);
8431 mutex_unlock(&ctx->uring_lock);
8432 spin_lock_irq(&ctx->completion_lock);
8433 spin_unlock_irq(&ctx->completion_lock);
8435 io_sq_thread_finish(ctx);
8436 io_sqe_buffers_unregister(ctx);
8438 if (ctx->mm_account) {
8439 mmdrop(ctx->mm_account);
8440 ctx->mm_account = NULL;
8443 mutex_lock(&ctx->uring_lock);
8444 io_sqe_files_unregister(ctx);
8445 mutex_unlock(&ctx->uring_lock);
8446 io_eventfd_unregister(ctx);
8447 io_destroy_buffers(ctx);
8449 #if defined(CONFIG_UNIX)
8450 if (ctx->ring_sock) {
8451 ctx->ring_sock->file = NULL; /* so that iput() is called */
8452 sock_release(ctx->ring_sock);
8456 io_mem_free(ctx->rings);
8457 io_mem_free(ctx->sq_sqes);
8459 percpu_ref_exit(&ctx->refs);
8460 free_uid(ctx->user);
8461 io_req_caches_free(ctx);
8463 io_wq_put_hash(ctx->hash_map);
8464 kfree(ctx->cancel_hash);
8468 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8470 struct io_ring_ctx *ctx = file->private_data;
8473 poll_wait(file, &ctx->cq_wait, wait);
8475 * synchronizes with barrier from wq_has_sleeper call in
8479 if (!io_sqring_full(ctx))
8480 mask |= EPOLLOUT | EPOLLWRNORM;
8483 * Don't flush cqring overflow list here, just do a simple check.
8484 * Otherwise there could possible be ABBA deadlock:
8487 * lock(&ctx->uring_lock);
8489 * lock(&ctx->uring_lock);
8492 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8493 * pushs them to do the flush.
8495 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8496 mask |= EPOLLIN | EPOLLRDNORM;
8501 static int io_uring_fasync(int fd, struct file *file, int on)
8503 struct io_ring_ctx *ctx = file->private_data;
8505 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8508 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8510 const struct cred *creds;
8512 creds = xa_erase(&ctx->personalities, id);
8521 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8523 return io_run_task_work_head(&ctx->exit_task_work);
8526 struct io_tctx_exit {
8527 struct callback_head task_work;
8528 struct completion completion;
8529 struct io_ring_ctx *ctx;
8532 static void io_tctx_exit_cb(struct callback_head *cb)
8534 struct io_uring_task *tctx = current->io_uring;
8535 struct io_tctx_exit *work;
8537 work = container_of(cb, struct io_tctx_exit, task_work);
8539 * When @in_idle, we're in cancellation and it's racy to remove the
8540 * node. It'll be removed by the end of cancellation, just ignore it.
8542 if (!atomic_read(&tctx->in_idle))
8543 io_uring_del_task_file((unsigned long)work->ctx);
8544 complete(&work->completion);
8547 static void io_ring_exit_work(struct work_struct *work)
8549 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8550 unsigned long timeout = jiffies + HZ * 60 * 5;
8551 struct io_tctx_exit exit;
8552 struct io_tctx_node *node;
8556 * If we're doing polled IO and end up having requests being
8557 * submitted async (out-of-line), then completions can come in while
8558 * we're waiting for refs to drop. We need to reap these manually,
8559 * as nobody else will be looking for them.
8562 io_uring_try_cancel_requests(ctx, NULL, NULL);
8564 WARN_ON_ONCE(time_after(jiffies, timeout));
8565 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8567 mutex_lock(&ctx->uring_lock);
8568 while (!list_empty(&ctx->tctx_list)) {
8569 WARN_ON_ONCE(time_after(jiffies, timeout));
8571 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8574 init_completion(&exit.completion);
8575 init_task_work(&exit.task_work, io_tctx_exit_cb);
8576 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8577 if (WARN_ON_ONCE(ret))
8579 wake_up_process(node->task);
8581 mutex_unlock(&ctx->uring_lock);
8582 wait_for_completion(&exit.completion);
8584 mutex_lock(&ctx->uring_lock);
8586 mutex_unlock(&ctx->uring_lock);
8588 io_ring_ctx_free(ctx);
8591 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8593 unsigned long index;
8594 struct creds *creds;
8596 mutex_lock(&ctx->uring_lock);
8597 percpu_ref_kill(&ctx->refs);
8598 /* if force is set, the ring is going away. always drop after that */
8599 ctx->cq_overflow_flushed = 1;
8601 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8602 xa_for_each(&ctx->personalities, index, creds)
8603 io_unregister_personality(ctx, index);
8604 mutex_unlock(&ctx->uring_lock);
8606 /* prevent SQPOLL from submitting new requests */
8608 io_sq_thread_park(ctx->sq_data);
8609 list_del_init(&ctx->sqd_list);
8610 io_sqd_update_thread_idle(ctx->sq_data);
8611 io_sq_thread_unpark(ctx->sq_data);
8614 io_kill_timeouts(ctx, NULL, NULL);
8615 io_poll_remove_all(ctx, NULL, NULL);
8617 /* if we failed setting up the ctx, we might not have any rings */
8618 io_iopoll_try_reap_events(ctx);
8620 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8622 * Use system_unbound_wq to avoid spawning tons of event kworkers
8623 * if we're exiting a ton of rings at the same time. It just adds
8624 * noise and overhead, there's no discernable change in runtime
8625 * over using system_wq.
8627 queue_work(system_unbound_wq, &ctx->exit_work);
8630 static int io_uring_release(struct inode *inode, struct file *file)
8632 struct io_ring_ctx *ctx = file->private_data;
8634 file->private_data = NULL;
8635 io_ring_ctx_wait_and_kill(ctx);
8639 struct io_task_cancel {
8640 struct task_struct *task;
8641 struct files_struct *files;
8644 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8646 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8647 struct io_task_cancel *cancel = data;
8650 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8651 unsigned long flags;
8652 struct io_ring_ctx *ctx = req->ctx;
8654 /* protect against races with linked timeouts */
8655 spin_lock_irqsave(&ctx->completion_lock, flags);
8656 ret = io_match_task(req, cancel->task, cancel->files);
8657 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8659 ret = io_match_task(req, cancel->task, cancel->files);
8664 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8665 struct task_struct *task,
8666 struct files_struct *files)
8668 struct io_defer_entry *de;
8671 spin_lock_irq(&ctx->completion_lock);
8672 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8673 if (io_match_task(de->req, task, files)) {
8674 list_cut_position(&list, &ctx->defer_list, &de->list);
8678 spin_unlock_irq(&ctx->completion_lock);
8679 if (list_empty(&list))
8682 while (!list_empty(&list)) {
8683 de = list_first_entry(&list, struct io_defer_entry, list);
8684 list_del_init(&de->list);
8685 req_set_fail_links(de->req);
8686 io_put_req(de->req);
8687 io_req_complete(de->req, -ECANCELED);
8693 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8695 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8697 return req->ctx == data;
8700 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8702 struct io_tctx_node *node;
8703 enum io_wq_cancel cret;
8706 mutex_lock(&ctx->uring_lock);
8707 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8708 struct io_uring_task *tctx = node->task->io_uring;
8711 * io_wq will stay alive while we hold uring_lock, because it's
8712 * killed after ctx nodes, which requires to take the lock.
8714 if (!tctx || !tctx->io_wq)
8716 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8717 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8719 mutex_unlock(&ctx->uring_lock);
8724 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8725 struct task_struct *task,
8726 struct files_struct *files)
8728 struct io_task_cancel cancel = { .task = task, .files = files, };
8729 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8732 enum io_wq_cancel cret;
8736 ret |= io_uring_try_cancel_iowq(ctx);
8737 } else if (tctx && tctx->io_wq) {
8739 * Cancels requests of all rings, not only @ctx, but
8740 * it's fine as the task is in exit/exec.
8742 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8744 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8747 /* SQPOLL thread does its own polling */
8748 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8749 (ctx->sq_data && ctx->sq_data->thread == current)) {
8750 while (!list_empty_careful(&ctx->iopoll_list)) {
8751 io_iopoll_try_reap_events(ctx);
8756 ret |= io_cancel_defer_files(ctx, task, files);
8757 ret |= io_poll_remove_all(ctx, task, files);
8758 ret |= io_kill_timeouts(ctx, task, files);
8759 ret |= io_run_task_work();
8760 ret |= io_run_ctx_fallback(ctx);
8761 io_cqring_overflow_flush(ctx, true, task, files);
8768 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8769 struct task_struct *task,
8770 struct files_struct *files)
8772 struct io_kiocb *req;
8775 spin_lock_irq(&ctx->inflight_lock);
8776 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8777 cnt += io_match_task(req, task, files);
8778 spin_unlock_irq(&ctx->inflight_lock);
8782 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8783 struct task_struct *task,
8784 struct files_struct *files)
8786 while (!list_empty_careful(&ctx->inflight_list)) {
8790 inflight = io_uring_count_inflight(ctx, task, files);
8794 io_uring_try_cancel_requests(ctx, task, files);
8796 prepare_to_wait(&task->io_uring->wait, &wait,
8797 TASK_UNINTERRUPTIBLE);
8798 if (inflight == io_uring_count_inflight(ctx, task, files))
8800 finish_wait(&task->io_uring->wait, &wait);
8805 * Note that this task has used io_uring. We use it for cancelation purposes.
8807 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8809 struct io_uring_task *tctx = current->io_uring;
8810 struct io_tctx_node *node;
8813 if (unlikely(!tctx)) {
8814 ret = io_uring_alloc_task_context(current, ctx);
8817 tctx = current->io_uring;
8819 if (tctx->last != ctx) {
8820 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8823 node = kmalloc(sizeof(*node), GFP_KERNEL);
8827 node->task = current;
8829 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8836 mutex_lock(&ctx->uring_lock);
8837 list_add(&node->ctx_node, &ctx->tctx_list);
8838 mutex_unlock(&ctx->uring_lock);
8846 * Remove this io_uring_file -> task mapping.
8848 static void io_uring_del_task_file(unsigned long index)
8850 struct io_uring_task *tctx = current->io_uring;
8851 struct io_tctx_node *node;
8855 node = xa_erase(&tctx->xa, index);
8859 WARN_ON_ONCE(current != node->task);
8860 WARN_ON_ONCE(list_empty(&node->ctx_node));
8862 mutex_lock(&node->ctx->uring_lock);
8863 list_del(&node->ctx_node);
8864 mutex_unlock(&node->ctx->uring_lock);
8866 if (tctx->last == node->ctx)
8871 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8873 struct io_tctx_node *node;
8874 unsigned long index;
8876 xa_for_each(&tctx->xa, index, node)
8877 io_uring_del_task_file(index);
8879 io_wq_put_and_exit(tctx->io_wq);
8884 static s64 tctx_inflight(struct io_uring_task *tctx)
8886 return percpu_counter_sum(&tctx->inflight);
8889 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8891 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8892 struct io_ring_ctx *ctx = work->ctx;
8893 struct io_sq_data *sqd = ctx->sq_data;
8896 io_uring_cancel_sqpoll(ctx);
8897 complete(&work->completion);
8900 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8902 struct io_sq_data *sqd = ctx->sq_data;
8903 struct io_tctx_exit work = { .ctx = ctx, };
8904 struct task_struct *task;
8906 io_sq_thread_park(sqd);
8907 list_del_init(&ctx->sqd_list);
8908 io_sqd_update_thread_idle(sqd);
8911 init_completion(&work.completion);
8912 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8913 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
8914 wake_up_process(task);
8916 io_sq_thread_unpark(sqd);
8919 wait_for_completion(&work.completion);
8922 void __io_uring_files_cancel(struct files_struct *files)
8924 struct io_uring_task *tctx = current->io_uring;
8925 struct io_tctx_node *node;
8926 unsigned long index;
8928 /* make sure overflow events are dropped */
8929 atomic_inc(&tctx->in_idle);
8930 xa_for_each(&tctx->xa, index, node) {
8931 struct io_ring_ctx *ctx = node->ctx;
8934 io_sqpoll_cancel_sync(ctx);
8937 io_uring_cancel_files(ctx, current, files);
8939 io_uring_try_cancel_requests(ctx, current, NULL);
8941 atomic_dec(&tctx->in_idle);
8944 io_uring_clean_tctx(tctx);
8947 /* should only be called by SQPOLL task */
8948 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8950 struct io_sq_data *sqd = ctx->sq_data;
8951 struct io_uring_task *tctx = current->io_uring;
8955 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
8957 atomic_inc(&tctx->in_idle);
8959 /* read completions before cancelations */
8960 inflight = tctx_inflight(tctx);
8963 io_uring_try_cancel_requests(ctx, current, NULL);
8965 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8967 * If we've seen completions, retry without waiting. This
8968 * avoids a race where a completion comes in before we did
8969 * prepare_to_wait().
8971 if (inflight == tctx_inflight(tctx))
8973 finish_wait(&tctx->wait, &wait);
8975 atomic_dec(&tctx->in_idle);
8979 * Find any io_uring fd that this task has registered or done IO on, and cancel
8982 void __io_uring_task_cancel(void)
8984 struct io_uring_task *tctx = current->io_uring;
8988 /* make sure overflow events are dropped */
8989 atomic_inc(&tctx->in_idle);
8991 /* read completions before cancelations */
8992 inflight = tctx_inflight(tctx);
8995 __io_uring_files_cancel(NULL);
8997 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9000 * If we've seen completions, retry without waiting. This
9001 * avoids a race where a completion comes in before we did
9002 * prepare_to_wait().
9004 if (inflight == tctx_inflight(tctx))
9006 finish_wait(&tctx->wait, &wait);
9009 atomic_dec(&tctx->in_idle);
9011 io_uring_clean_tctx(tctx);
9012 /* all current's requests should be gone, we can kill tctx */
9013 __io_uring_free(current);
9016 static void *io_uring_validate_mmap_request(struct file *file,
9017 loff_t pgoff, size_t sz)
9019 struct io_ring_ctx *ctx = file->private_data;
9020 loff_t offset = pgoff << PAGE_SHIFT;
9025 case IORING_OFF_SQ_RING:
9026 case IORING_OFF_CQ_RING:
9029 case IORING_OFF_SQES:
9033 return ERR_PTR(-EINVAL);
9036 page = virt_to_head_page(ptr);
9037 if (sz > page_size(page))
9038 return ERR_PTR(-EINVAL);
9045 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9047 size_t sz = vma->vm_end - vma->vm_start;
9051 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9053 return PTR_ERR(ptr);
9055 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9056 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9059 #else /* !CONFIG_MMU */
9061 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9063 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9066 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9068 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9071 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9072 unsigned long addr, unsigned long len,
9073 unsigned long pgoff, unsigned long flags)
9077 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9079 return PTR_ERR(ptr);
9081 return (unsigned long) ptr;
9084 #endif /* !CONFIG_MMU */
9086 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9091 if (!io_sqring_full(ctx))
9093 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9095 if (!io_sqring_full(ctx))
9098 } while (!signal_pending(current));
9100 finish_wait(&ctx->sqo_sq_wait, &wait);
9104 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9105 struct __kernel_timespec __user **ts,
9106 const sigset_t __user **sig)
9108 struct io_uring_getevents_arg arg;
9111 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9112 * is just a pointer to the sigset_t.
9114 if (!(flags & IORING_ENTER_EXT_ARG)) {
9115 *sig = (const sigset_t __user *) argp;
9121 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9122 * timespec and sigset_t pointers if good.
9124 if (*argsz != sizeof(arg))
9126 if (copy_from_user(&arg, argp, sizeof(arg)))
9128 *sig = u64_to_user_ptr(arg.sigmask);
9129 *argsz = arg.sigmask_sz;
9130 *ts = u64_to_user_ptr(arg.ts);
9134 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9135 u32, min_complete, u32, flags, const void __user *, argp,
9138 struct io_ring_ctx *ctx;
9145 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9146 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9154 if (f.file->f_op != &io_uring_fops)
9158 ctx = f.file->private_data;
9159 if (!percpu_ref_tryget(&ctx->refs))
9163 if (ctx->flags & IORING_SETUP_R_DISABLED)
9167 * For SQ polling, the thread will do all submissions and completions.
9168 * Just return the requested submit count, and wake the thread if
9172 if (ctx->flags & IORING_SETUP_SQPOLL) {
9173 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9176 if (unlikely(ctx->sq_data->thread == NULL)) {
9179 if (flags & IORING_ENTER_SQ_WAKEUP)
9180 wake_up(&ctx->sq_data->wait);
9181 if (flags & IORING_ENTER_SQ_WAIT) {
9182 ret = io_sqpoll_wait_sq(ctx);
9186 submitted = to_submit;
9187 } else if (to_submit) {
9188 ret = io_uring_add_task_file(ctx);
9191 mutex_lock(&ctx->uring_lock);
9192 submitted = io_submit_sqes(ctx, to_submit);
9193 mutex_unlock(&ctx->uring_lock);
9195 if (submitted != to_submit)
9198 if (flags & IORING_ENTER_GETEVENTS) {
9199 const sigset_t __user *sig;
9200 struct __kernel_timespec __user *ts;
9202 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9206 min_complete = min(min_complete, ctx->cq_entries);
9209 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9210 * space applications don't need to do io completion events
9211 * polling again, they can rely on io_sq_thread to do polling
9212 * work, which can reduce cpu usage and uring_lock contention.
9214 if (ctx->flags & IORING_SETUP_IOPOLL &&
9215 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9216 ret = io_iopoll_check(ctx, min_complete);
9218 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9223 percpu_ref_put(&ctx->refs);
9226 return submitted ? submitted : ret;
9229 #ifdef CONFIG_PROC_FS
9230 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9231 const struct cred *cred)
9233 struct user_namespace *uns = seq_user_ns(m);
9234 struct group_info *gi;
9239 seq_printf(m, "%5d\n", id);
9240 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9241 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9242 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9243 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9244 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9245 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9246 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9247 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9248 seq_puts(m, "\n\tGroups:\t");
9249 gi = cred->group_info;
9250 for (g = 0; g < gi->ngroups; g++) {
9251 seq_put_decimal_ull(m, g ? " " : "",
9252 from_kgid_munged(uns, gi->gid[g]));
9254 seq_puts(m, "\n\tCapEff:\t");
9255 cap = cred->cap_effective;
9256 CAP_FOR_EACH_U32(__capi)
9257 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9262 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9264 struct io_sq_data *sq = NULL;
9269 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9270 * since fdinfo case grabs it in the opposite direction of normal use
9271 * cases. If we fail to get the lock, we just don't iterate any
9272 * structures that could be going away outside the io_uring mutex.
9274 has_lock = mutex_trylock(&ctx->uring_lock);
9276 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9282 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9283 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9284 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9285 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9286 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9289 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9291 seq_printf(m, "%5u: <none>\n", i);
9293 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9294 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9295 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9297 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9298 (unsigned int) buf->len);
9300 if (has_lock && !xa_empty(&ctx->personalities)) {
9301 unsigned long index;
9302 const struct cred *cred;
9304 seq_printf(m, "Personalities:\n");
9305 xa_for_each(&ctx->personalities, index, cred)
9306 io_uring_show_cred(m, index, cred);
9308 seq_printf(m, "PollList:\n");
9309 spin_lock_irq(&ctx->completion_lock);
9310 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9311 struct hlist_head *list = &ctx->cancel_hash[i];
9312 struct io_kiocb *req;
9314 hlist_for_each_entry(req, list, hash_node)
9315 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9316 req->task->task_works != NULL);
9318 spin_unlock_irq(&ctx->completion_lock);
9320 mutex_unlock(&ctx->uring_lock);
9323 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9325 struct io_ring_ctx *ctx = f->private_data;
9327 if (percpu_ref_tryget(&ctx->refs)) {
9328 __io_uring_show_fdinfo(ctx, m);
9329 percpu_ref_put(&ctx->refs);
9334 static const struct file_operations io_uring_fops = {
9335 .release = io_uring_release,
9336 .mmap = io_uring_mmap,
9338 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9339 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9341 .poll = io_uring_poll,
9342 .fasync = io_uring_fasync,
9343 #ifdef CONFIG_PROC_FS
9344 .show_fdinfo = io_uring_show_fdinfo,
9348 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9349 struct io_uring_params *p)
9351 struct io_rings *rings;
9352 size_t size, sq_array_offset;
9354 /* make sure these are sane, as we already accounted them */
9355 ctx->sq_entries = p->sq_entries;
9356 ctx->cq_entries = p->cq_entries;
9358 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9359 if (size == SIZE_MAX)
9362 rings = io_mem_alloc(size);
9367 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9368 rings->sq_ring_mask = p->sq_entries - 1;
9369 rings->cq_ring_mask = p->cq_entries - 1;
9370 rings->sq_ring_entries = p->sq_entries;
9371 rings->cq_ring_entries = p->cq_entries;
9372 ctx->sq_mask = rings->sq_ring_mask;
9373 ctx->cq_mask = rings->cq_ring_mask;
9375 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9376 if (size == SIZE_MAX) {
9377 io_mem_free(ctx->rings);
9382 ctx->sq_sqes = io_mem_alloc(size);
9383 if (!ctx->sq_sqes) {
9384 io_mem_free(ctx->rings);
9392 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9396 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9400 ret = io_uring_add_task_file(ctx);
9405 fd_install(fd, file);
9410 * Allocate an anonymous fd, this is what constitutes the application
9411 * visible backing of an io_uring instance. The application mmaps this
9412 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9413 * we have to tie this fd to a socket for file garbage collection purposes.
9415 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9418 #if defined(CONFIG_UNIX)
9421 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9424 return ERR_PTR(ret);
9427 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9428 O_RDWR | O_CLOEXEC);
9429 #if defined(CONFIG_UNIX)
9431 sock_release(ctx->ring_sock);
9432 ctx->ring_sock = NULL;
9434 ctx->ring_sock->file = file;
9440 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9441 struct io_uring_params __user *params)
9443 struct io_ring_ctx *ctx;
9449 if (entries > IORING_MAX_ENTRIES) {
9450 if (!(p->flags & IORING_SETUP_CLAMP))
9452 entries = IORING_MAX_ENTRIES;
9456 * Use twice as many entries for the CQ ring. It's possible for the
9457 * application to drive a higher depth than the size of the SQ ring,
9458 * since the sqes are only used at submission time. This allows for
9459 * some flexibility in overcommitting a bit. If the application has
9460 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9461 * of CQ ring entries manually.
9463 p->sq_entries = roundup_pow_of_two(entries);
9464 if (p->flags & IORING_SETUP_CQSIZE) {
9466 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9467 * to a power-of-two, if it isn't already. We do NOT impose
9468 * any cq vs sq ring sizing.
9472 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9473 if (!(p->flags & IORING_SETUP_CLAMP))
9475 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9477 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9478 if (p->cq_entries < p->sq_entries)
9481 p->cq_entries = 2 * p->sq_entries;
9484 ctx = io_ring_ctx_alloc(p);
9487 ctx->compat = in_compat_syscall();
9488 if (!capable(CAP_IPC_LOCK))
9489 ctx->user = get_uid(current_user());
9492 * This is just grabbed for accounting purposes. When a process exits,
9493 * the mm is exited and dropped before the files, hence we need to hang
9494 * on to this mm purely for the purposes of being able to unaccount
9495 * memory (locked/pinned vm). It's not used for anything else.
9497 mmgrab(current->mm);
9498 ctx->mm_account = current->mm;
9500 ret = io_allocate_scq_urings(ctx, p);
9504 ret = io_sq_offload_create(ctx, p);
9508 memset(&p->sq_off, 0, sizeof(p->sq_off));
9509 p->sq_off.head = offsetof(struct io_rings, sq.head);
9510 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9511 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9512 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9513 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9514 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9515 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9517 memset(&p->cq_off, 0, sizeof(p->cq_off));
9518 p->cq_off.head = offsetof(struct io_rings, cq.head);
9519 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9520 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9521 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9522 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9523 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9524 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9526 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9527 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9528 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9529 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9530 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9532 if (copy_to_user(params, p, sizeof(*p))) {
9537 file = io_uring_get_file(ctx);
9539 ret = PTR_ERR(file);
9544 * Install ring fd as the very last thing, so we don't risk someone
9545 * having closed it before we finish setup
9547 ret = io_uring_install_fd(ctx, file);
9549 /* fput will clean it up */
9554 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9557 io_ring_ctx_wait_and_kill(ctx);
9562 * Sets up an aio uring context, and returns the fd. Applications asks for a
9563 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9564 * params structure passed in.
9566 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9568 struct io_uring_params p;
9571 if (copy_from_user(&p, params, sizeof(p)))
9573 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9578 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9579 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9580 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9581 IORING_SETUP_R_DISABLED))
9584 return io_uring_create(entries, &p, params);
9587 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9588 struct io_uring_params __user *, params)
9590 return io_uring_setup(entries, params);
9593 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9595 struct io_uring_probe *p;
9599 size = struct_size(p, ops, nr_args);
9600 if (size == SIZE_MAX)
9602 p = kzalloc(size, GFP_KERNEL);
9607 if (copy_from_user(p, arg, size))
9610 if (memchr_inv(p, 0, size))
9613 p->last_op = IORING_OP_LAST - 1;
9614 if (nr_args > IORING_OP_LAST)
9615 nr_args = IORING_OP_LAST;
9617 for (i = 0; i < nr_args; i++) {
9619 if (!io_op_defs[i].not_supported)
9620 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9625 if (copy_to_user(arg, p, size))
9632 static int io_register_personality(struct io_ring_ctx *ctx)
9634 const struct cred *creds;
9638 creds = get_current_cred();
9640 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9641 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9648 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9649 unsigned int nr_args)
9651 struct io_uring_restriction *res;
9655 /* Restrictions allowed only if rings started disabled */
9656 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9659 /* We allow only a single restrictions registration */
9660 if (ctx->restrictions.registered)
9663 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9666 size = array_size(nr_args, sizeof(*res));
9667 if (size == SIZE_MAX)
9670 res = memdup_user(arg, size);
9672 return PTR_ERR(res);
9676 for (i = 0; i < nr_args; i++) {
9677 switch (res[i].opcode) {
9678 case IORING_RESTRICTION_REGISTER_OP:
9679 if (res[i].register_op >= IORING_REGISTER_LAST) {
9684 __set_bit(res[i].register_op,
9685 ctx->restrictions.register_op);
9687 case IORING_RESTRICTION_SQE_OP:
9688 if (res[i].sqe_op >= IORING_OP_LAST) {
9693 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9695 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9696 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9698 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9699 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9708 /* Reset all restrictions if an error happened */
9710 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9712 ctx->restrictions.registered = true;
9718 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9720 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9723 if (ctx->restrictions.registered)
9724 ctx->restricted = 1;
9726 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9727 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9728 wake_up(&ctx->sq_data->wait);
9732 static bool io_register_op_must_quiesce(int op)
9735 case IORING_UNREGISTER_FILES:
9736 case IORING_REGISTER_FILES_UPDATE:
9737 case IORING_REGISTER_PROBE:
9738 case IORING_REGISTER_PERSONALITY:
9739 case IORING_UNREGISTER_PERSONALITY:
9746 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9747 void __user *arg, unsigned nr_args)
9748 __releases(ctx->uring_lock)
9749 __acquires(ctx->uring_lock)
9754 * We're inside the ring mutex, if the ref is already dying, then
9755 * someone else killed the ctx or is already going through
9756 * io_uring_register().
9758 if (percpu_ref_is_dying(&ctx->refs))
9761 if (io_register_op_must_quiesce(opcode)) {
9762 percpu_ref_kill(&ctx->refs);
9765 * Drop uring mutex before waiting for references to exit. If
9766 * another thread is currently inside io_uring_enter() it might
9767 * need to grab the uring_lock to make progress. If we hold it
9768 * here across the drain wait, then we can deadlock. It's safe
9769 * to drop the mutex here, since no new references will come in
9770 * after we've killed the percpu ref.
9772 mutex_unlock(&ctx->uring_lock);
9774 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9777 ret = io_run_task_work_sig();
9782 mutex_lock(&ctx->uring_lock);
9785 percpu_ref_resurrect(&ctx->refs);
9790 if (ctx->restricted) {
9791 if (opcode >= IORING_REGISTER_LAST) {
9796 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9803 case IORING_REGISTER_BUFFERS:
9804 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9806 case IORING_UNREGISTER_BUFFERS:
9810 ret = io_sqe_buffers_unregister(ctx);
9812 case IORING_REGISTER_FILES:
9813 ret = io_sqe_files_register(ctx, arg, nr_args);
9815 case IORING_UNREGISTER_FILES:
9819 ret = io_sqe_files_unregister(ctx);
9821 case IORING_REGISTER_FILES_UPDATE:
9822 ret = io_sqe_files_update(ctx, arg, nr_args);
9824 case IORING_REGISTER_EVENTFD:
9825 case IORING_REGISTER_EVENTFD_ASYNC:
9829 ret = io_eventfd_register(ctx, arg);
9832 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9833 ctx->eventfd_async = 1;
9835 ctx->eventfd_async = 0;
9837 case IORING_UNREGISTER_EVENTFD:
9841 ret = io_eventfd_unregister(ctx);
9843 case IORING_REGISTER_PROBE:
9845 if (!arg || nr_args > 256)
9847 ret = io_probe(ctx, arg, nr_args);
9849 case IORING_REGISTER_PERSONALITY:
9853 ret = io_register_personality(ctx);
9855 case IORING_UNREGISTER_PERSONALITY:
9859 ret = io_unregister_personality(ctx, nr_args);
9861 case IORING_REGISTER_ENABLE_RINGS:
9865 ret = io_register_enable_rings(ctx);
9867 case IORING_REGISTER_RESTRICTIONS:
9868 ret = io_register_restrictions(ctx, arg, nr_args);
9876 if (io_register_op_must_quiesce(opcode)) {
9877 /* bring the ctx back to life */
9878 percpu_ref_reinit(&ctx->refs);
9880 reinit_completion(&ctx->ref_comp);
9885 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9886 void __user *, arg, unsigned int, nr_args)
9888 struct io_ring_ctx *ctx;
9897 if (f.file->f_op != &io_uring_fops)
9900 ctx = f.file->private_data;
9904 mutex_lock(&ctx->uring_lock);
9905 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9906 mutex_unlock(&ctx->uring_lock);
9907 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9908 ctx->cq_ev_fd != NULL, ret);
9914 static int __init io_uring_init(void)
9916 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9917 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9918 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9921 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9922 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9923 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9924 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9925 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9926 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9927 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9928 BUILD_BUG_SQE_ELEM(8, __u64, off);
9929 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9930 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9931 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9932 BUILD_BUG_SQE_ELEM(24, __u32, len);
9933 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9934 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9935 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9936 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9937 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9938 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9939 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9940 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9941 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9942 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9943 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9944 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9945 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9946 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9947 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9948 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9949 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9950 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9951 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9953 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9954 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9955 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9959 __initcall(io_uring_init);