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>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 enum io_uring_cmd_flags {
191 IO_URING_F_NONBLOCK = 1,
192 IO_URING_F_COMPLETE_DEFER = 2,
195 struct io_mapped_ubuf {
198 unsigned int nr_bvecs;
199 unsigned long acct_pages;
200 struct bio_vec bvec[];
205 struct io_overflow_cqe {
206 struct io_uring_cqe cqe;
207 struct list_head list;
210 struct io_fixed_file {
211 /* file * with additional FFS_* flags */
212 unsigned long file_ptr;
216 struct list_head list;
221 struct io_mapped_ubuf *buf;
225 struct io_file_table {
226 /* two level table */
227 struct io_fixed_file **files;
230 struct io_rsrc_node {
231 struct percpu_ref refs;
232 struct list_head node;
233 struct list_head rsrc_list;
234 struct io_rsrc_data *rsrc_data;
235 struct llist_node llist;
239 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
241 struct io_rsrc_data {
242 struct io_ring_ctx *ctx;
247 struct completion done;
252 struct list_head list;
258 struct io_restriction {
259 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
260 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
261 u8 sqe_flags_allowed;
262 u8 sqe_flags_required;
267 IO_SQ_THREAD_SHOULD_STOP = 0,
268 IO_SQ_THREAD_SHOULD_PARK,
273 atomic_t park_pending;
276 /* ctx's that are using this sqd */
277 struct list_head ctx_list;
279 struct task_struct *thread;
280 struct wait_queue_head wait;
282 unsigned sq_thread_idle;
288 struct completion exited;
289 struct callback_head *park_task_work;
292 #define IO_IOPOLL_BATCH 8
293 #define IO_COMPL_BATCH 32
294 #define IO_REQ_CACHE_SIZE 32
295 #define IO_REQ_ALLOC_BATCH 8
297 struct io_comp_state {
298 struct io_kiocb *reqs[IO_COMPL_BATCH];
300 unsigned int locked_free_nr;
301 /* inline/task_work completion list, under ->uring_lock */
302 struct list_head free_list;
303 /* IRQ completion list, under ->completion_lock */
304 struct list_head locked_free_list;
307 struct io_submit_link {
308 struct io_kiocb *head;
309 struct io_kiocb *last;
312 struct io_submit_state {
313 struct blk_plug plug;
314 struct io_submit_link link;
317 * io_kiocb alloc cache
319 void *reqs[IO_REQ_CACHE_SIZE];
320 unsigned int free_reqs;
325 * Batch completion logic
327 struct io_comp_state comp;
330 * File reference cache
334 unsigned int file_refs;
335 unsigned int ios_left;
340 struct percpu_ref refs;
341 } ____cacheline_aligned_in_smp;
345 unsigned int compat: 1;
346 unsigned int drain_next: 1;
347 unsigned int eventfd_async: 1;
348 unsigned int restricted: 1;
351 * Ring buffer of indices into array of io_uring_sqe, which is
352 * mmapped by the application using the IORING_OFF_SQES offset.
354 * This indirection could e.g. be used to assign fixed
355 * io_uring_sqe entries to operations and only submit them to
356 * the queue when needed.
358 * The kernel modifies neither the indices array nor the entries
362 unsigned cached_sq_head;
365 unsigned sq_thread_idle;
366 unsigned cached_sq_dropped;
367 unsigned cached_cq_overflow;
368 unsigned long sq_check_overflow;
370 /* hashed buffered write serialization */
371 struct io_wq_hash *hash_map;
373 struct list_head defer_list;
374 struct list_head timeout_list;
375 struct list_head cq_overflow_list;
377 struct io_uring_sqe *sq_sqes;
378 } ____cacheline_aligned_in_smp;
381 struct mutex uring_lock;
382 wait_queue_head_t wait;
383 } ____cacheline_aligned_in_smp;
385 struct io_submit_state submit_state;
387 struct io_rings *rings;
389 /* Only used for accounting purposes */
390 struct mm_struct *mm_account;
392 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
393 struct io_sq_data *sq_data; /* if using sq thread polling */
395 struct wait_queue_head sqo_sq_wait;
396 struct list_head sqd_list;
399 * If used, fixed file set. Writers must ensure that ->refs is dead,
400 * readers must ensure that ->refs is alive as long as the file* is
401 * used. Only updated through io_uring_register(2).
403 struct io_rsrc_data *file_data;
404 struct io_file_table file_table;
405 unsigned nr_user_files;
407 /* if used, fixed mapped user buffers */
408 struct io_rsrc_data *buf_data;
409 unsigned nr_user_bufs;
410 struct io_mapped_ubuf **user_bufs;
412 struct user_struct *user;
414 struct completion ref_comp;
416 #if defined(CONFIG_UNIX)
417 struct socket *ring_sock;
420 struct xarray io_buffers;
422 struct xarray personalities;
426 unsigned cached_cq_tail;
429 atomic_t cq_timeouts;
430 unsigned cq_last_tm_flush;
432 unsigned long cq_check_overflow;
433 struct wait_queue_head cq_wait;
434 struct fasync_struct *cq_fasync;
435 struct eventfd_ctx *cq_ev_fd;
436 } ____cacheline_aligned_in_smp;
439 spinlock_t completion_lock;
442 * ->iopoll_list is protected by the ctx->uring_lock for
443 * io_uring instances that don't use IORING_SETUP_SQPOLL.
444 * For SQPOLL, only the single threaded io_sq_thread() will
445 * manipulate the list, hence no extra locking is needed there.
447 struct list_head iopoll_list;
448 struct hlist_head *cancel_hash;
449 unsigned cancel_hash_bits;
450 bool poll_multi_file;
451 } ____cacheline_aligned_in_smp;
453 struct delayed_work rsrc_put_work;
454 struct llist_head rsrc_put_llist;
455 struct list_head rsrc_ref_list;
456 spinlock_t rsrc_ref_lock;
457 struct io_rsrc_node *rsrc_node;
458 struct io_rsrc_node *rsrc_backup_node;
459 struct io_mapped_ubuf *dummy_ubuf;
461 struct io_restriction restrictions;
464 struct callback_head *exit_task_work;
466 /* Keep this last, we don't need it for the fast path */
467 struct work_struct exit_work;
468 struct list_head tctx_list;
471 struct io_uring_task {
472 /* submission side */
474 struct wait_queue_head wait;
475 const struct io_ring_ctx *last;
477 struct percpu_counter inflight;
478 atomic_t inflight_tracked;
481 spinlock_t task_lock;
482 struct io_wq_work_list task_list;
483 unsigned long task_state;
484 struct callback_head task_work;
488 * First field must be the file pointer in all the
489 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
491 struct io_poll_iocb {
493 struct wait_queue_head *head;
497 struct wait_queue_entry wait;
500 struct io_poll_update {
506 bool update_user_data;
514 struct io_timeout_data {
515 struct io_kiocb *req;
516 struct hrtimer timer;
517 struct timespec64 ts;
518 enum hrtimer_mode mode;
523 struct sockaddr __user *addr;
524 int __user *addr_len;
526 unsigned long nofile;
546 struct list_head list;
547 /* head of the link, used by linked timeouts only */
548 struct io_kiocb *head;
551 struct io_timeout_rem {
556 struct timespec64 ts;
561 /* NOTE: kiocb has the file as the first member, so don't do it here */
569 struct sockaddr __user *addr;
576 struct compat_msghdr __user *umsg_compat;
577 struct user_msghdr __user *umsg;
583 struct io_buffer *kbuf;
589 struct filename *filename;
591 unsigned long nofile;
594 struct io_rsrc_update {
620 struct epoll_event event;
624 struct file *file_out;
625 struct file *file_in;
632 struct io_provide_buf {
646 const char __user *filename;
647 struct statx __user *buffer;
659 struct filename *oldpath;
660 struct filename *newpath;
668 struct filename *filename;
671 struct io_completion {
673 struct list_head list;
677 struct io_async_connect {
678 struct sockaddr_storage address;
681 struct io_async_msghdr {
682 struct iovec fast_iov[UIO_FASTIOV];
683 /* points to an allocated iov, if NULL we use fast_iov instead */
684 struct iovec *free_iov;
685 struct sockaddr __user *uaddr;
687 struct sockaddr_storage addr;
691 struct iovec fast_iov[UIO_FASTIOV];
692 const struct iovec *free_iovec;
693 struct iov_iter iter;
695 struct wait_page_queue wpq;
699 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
700 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
701 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
702 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
703 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
704 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
706 /* first byte is taken by user flags, shift it to not overlap */
707 REQ_F_FAIL_LINK_BIT = 8,
711 REQ_F_LINK_TIMEOUT_BIT,
712 REQ_F_NEED_CLEANUP_BIT,
714 REQ_F_BUFFER_SELECTED_BIT,
715 REQ_F_LTIMEOUT_ACTIVE_BIT,
716 REQ_F_COMPLETE_INLINE_BIT,
718 REQ_F_DONT_REISSUE_BIT,
719 /* keep async read/write and isreg together and in order */
720 REQ_F_ASYNC_READ_BIT,
721 REQ_F_ASYNC_WRITE_BIT,
724 /* not a real bit, just to check we're not overflowing the space */
730 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
731 /* drain existing IO first */
732 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
734 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
735 /* doesn't sever on completion < 0 */
736 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
738 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
739 /* IOSQE_BUFFER_SELECT */
740 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
742 /* fail rest of links */
743 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
744 /* on inflight list, should be cancelled and waited on exit reliably */
745 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
746 /* read/write uses file position */
747 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
748 /* must not punt to workers */
749 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
750 /* has or had linked timeout */
751 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
753 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
754 /* already went through poll handler */
755 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
756 /* buffer already selected */
757 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
758 /* linked timeout is active, i.e. prepared by link's head */
759 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
760 /* completion is deferred through io_comp_state */
761 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
762 /* caller should reissue async */
763 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
764 /* don't attempt request reissue, see io_rw_reissue() */
765 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
766 /* supports async reads */
767 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
768 /* supports async writes */
769 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
771 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
775 struct io_poll_iocb poll;
776 struct io_poll_iocb *double_poll;
779 struct io_task_work {
780 struct io_wq_work_node node;
781 task_work_func_t func;
785 * NOTE! Each of the iocb union members has the file pointer
786 * as the first entry in their struct definition. So you can
787 * access the file pointer through any of the sub-structs,
788 * or directly as just 'ki_filp' in this struct.
794 struct io_poll_iocb poll;
795 struct io_poll_update poll_update;
796 struct io_accept accept;
798 struct io_cancel cancel;
799 struct io_timeout timeout;
800 struct io_timeout_rem timeout_rem;
801 struct io_connect connect;
802 struct io_sr_msg sr_msg;
804 struct io_close close;
805 struct io_rsrc_update rsrc_update;
806 struct io_fadvise fadvise;
807 struct io_madvise madvise;
808 struct io_epoll epoll;
809 struct io_splice splice;
810 struct io_provide_buf pbuf;
811 struct io_statx statx;
812 struct io_shutdown shutdown;
813 struct io_rename rename;
814 struct io_unlink unlink;
815 /* use only after cleaning per-op data, see io_clean_op() */
816 struct io_completion compl;
819 /* opcode allocated if it needs to store data for async defer */
822 /* polled IO has completed */
828 struct io_ring_ctx *ctx;
831 struct task_struct *task;
834 struct io_kiocb *link;
835 struct percpu_ref *fixed_rsrc_refs;
837 /* used with ctx->iopoll_list with reads/writes */
838 struct list_head inflight_entry;
840 struct io_task_work io_task_work;
841 struct callback_head task_work;
843 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
844 struct hlist_node hash_node;
845 struct async_poll *apoll;
846 struct io_wq_work work;
847 /* store used ubuf, so we can prevent reloading */
848 struct io_mapped_ubuf *imu;
851 struct io_tctx_node {
852 struct list_head ctx_node;
853 struct task_struct *task;
854 struct io_ring_ctx *ctx;
857 struct io_defer_entry {
858 struct list_head list;
859 struct io_kiocb *req;
864 /* needs req->file assigned */
865 unsigned needs_file : 1;
866 /* hash wq insertion if file is a regular file */
867 unsigned hash_reg_file : 1;
868 /* unbound wq insertion if file is a non-regular file */
869 unsigned unbound_nonreg_file : 1;
870 /* opcode is not supported by this kernel */
871 unsigned not_supported : 1;
872 /* set if opcode supports polled "wait" */
874 unsigned pollout : 1;
875 /* op supports buffer selection */
876 unsigned buffer_select : 1;
877 /* do prep async if is going to be punted */
878 unsigned needs_async_setup : 1;
879 /* should block plug */
881 /* size of async data needed, if any */
882 unsigned short async_size;
885 static const struct io_op_def io_op_defs[] = {
886 [IORING_OP_NOP] = {},
887 [IORING_OP_READV] = {
889 .unbound_nonreg_file = 1,
892 .needs_async_setup = 1,
894 .async_size = sizeof(struct io_async_rw),
896 [IORING_OP_WRITEV] = {
899 .unbound_nonreg_file = 1,
901 .needs_async_setup = 1,
903 .async_size = sizeof(struct io_async_rw),
905 [IORING_OP_FSYNC] = {
908 [IORING_OP_READ_FIXED] = {
910 .unbound_nonreg_file = 1,
913 .async_size = sizeof(struct io_async_rw),
915 [IORING_OP_WRITE_FIXED] = {
918 .unbound_nonreg_file = 1,
921 .async_size = sizeof(struct io_async_rw),
923 [IORING_OP_POLL_ADD] = {
925 .unbound_nonreg_file = 1,
927 [IORING_OP_POLL_REMOVE] = {},
928 [IORING_OP_SYNC_FILE_RANGE] = {
931 [IORING_OP_SENDMSG] = {
933 .unbound_nonreg_file = 1,
935 .needs_async_setup = 1,
936 .async_size = sizeof(struct io_async_msghdr),
938 [IORING_OP_RECVMSG] = {
940 .unbound_nonreg_file = 1,
943 .needs_async_setup = 1,
944 .async_size = sizeof(struct io_async_msghdr),
946 [IORING_OP_TIMEOUT] = {
947 .async_size = sizeof(struct io_timeout_data),
949 [IORING_OP_TIMEOUT_REMOVE] = {
950 /* used by timeout updates' prep() */
952 [IORING_OP_ACCEPT] = {
954 .unbound_nonreg_file = 1,
957 [IORING_OP_ASYNC_CANCEL] = {},
958 [IORING_OP_LINK_TIMEOUT] = {
959 .async_size = sizeof(struct io_timeout_data),
961 [IORING_OP_CONNECT] = {
963 .unbound_nonreg_file = 1,
965 .needs_async_setup = 1,
966 .async_size = sizeof(struct io_async_connect),
968 [IORING_OP_FALLOCATE] = {
971 [IORING_OP_OPENAT] = {},
972 [IORING_OP_CLOSE] = {},
973 [IORING_OP_FILES_UPDATE] = {},
974 [IORING_OP_STATX] = {},
977 .unbound_nonreg_file = 1,
981 .async_size = sizeof(struct io_async_rw),
983 [IORING_OP_WRITE] = {
985 .unbound_nonreg_file = 1,
988 .async_size = sizeof(struct io_async_rw),
990 [IORING_OP_FADVISE] = {
993 [IORING_OP_MADVISE] = {},
996 .unbound_nonreg_file = 1,
1001 .unbound_nonreg_file = 1,
1005 [IORING_OP_OPENAT2] = {
1007 [IORING_OP_EPOLL_CTL] = {
1008 .unbound_nonreg_file = 1,
1010 [IORING_OP_SPLICE] = {
1013 .unbound_nonreg_file = 1,
1015 [IORING_OP_PROVIDE_BUFFERS] = {},
1016 [IORING_OP_REMOVE_BUFFERS] = {},
1020 .unbound_nonreg_file = 1,
1022 [IORING_OP_SHUTDOWN] = {
1025 [IORING_OP_RENAMEAT] = {},
1026 [IORING_OP_UNLINKAT] = {},
1029 static bool io_disarm_next(struct io_kiocb *req);
1030 static void io_uring_del_task_file(unsigned long index);
1031 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1032 struct task_struct *task,
1033 struct files_struct *files);
1034 static void io_uring_cancel_sqpoll(struct io_sq_data *sqd);
1035 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1037 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1038 long res, unsigned int cflags);
1039 static void io_put_req(struct io_kiocb *req);
1040 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1041 static void io_dismantle_req(struct io_kiocb *req);
1042 static void io_put_task(struct task_struct *task, int nr);
1043 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1044 static void io_queue_linked_timeout(struct io_kiocb *req);
1045 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1046 struct io_uring_rsrc_update2 *up,
1048 static void io_clean_op(struct io_kiocb *req);
1049 static struct file *io_file_get(struct io_submit_state *state,
1050 struct io_kiocb *req, int fd, bool fixed);
1051 static void __io_queue_sqe(struct io_kiocb *req);
1052 static void io_rsrc_put_work(struct work_struct *work);
1054 static void io_req_task_queue(struct io_kiocb *req);
1055 static void io_submit_flush_completions(struct io_comp_state *cs,
1056 struct io_ring_ctx *ctx);
1057 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1058 static int io_req_prep_async(struct io_kiocb *req);
1060 static struct kmem_cache *req_cachep;
1062 static const struct file_operations io_uring_fops;
1064 struct sock *io_uring_get_socket(struct file *file)
1066 #if defined(CONFIG_UNIX)
1067 if (file->f_op == &io_uring_fops) {
1068 struct io_ring_ctx *ctx = file->private_data;
1070 return ctx->ring_sock->sk;
1075 EXPORT_SYMBOL(io_uring_get_socket);
1077 #define io_for_each_link(pos, head) \
1078 for (pos = (head); pos; pos = pos->link)
1080 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1082 struct io_ring_ctx *ctx = req->ctx;
1084 if (!req->fixed_rsrc_refs) {
1085 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1086 percpu_ref_get(req->fixed_rsrc_refs);
1090 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1092 bool got = percpu_ref_tryget(ref);
1094 /* already at zero, wait for ->release() */
1096 wait_for_completion(compl);
1097 percpu_ref_resurrect(ref);
1099 percpu_ref_put(ref);
1102 static bool io_match_task(struct io_kiocb *head,
1103 struct task_struct *task,
1104 struct files_struct *files)
1106 struct io_kiocb *req;
1108 if (task && head->task != task)
1113 io_for_each_link(req, head) {
1114 if (req->flags & REQ_F_INFLIGHT)
1120 static inline void req_set_fail_links(struct io_kiocb *req)
1122 if (req->flags & REQ_F_LINK)
1123 req->flags |= REQ_F_FAIL_LINK;
1126 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1128 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1130 complete(&ctx->ref_comp);
1133 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1135 return !req->timeout.off;
1138 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1140 struct io_ring_ctx *ctx;
1143 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1148 * Use 5 bits less than the max cq entries, that should give us around
1149 * 32 entries per hash list if totally full and uniformly spread.
1151 hash_bits = ilog2(p->cq_entries);
1155 ctx->cancel_hash_bits = hash_bits;
1156 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1158 if (!ctx->cancel_hash)
1160 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1162 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1163 if (!ctx->dummy_ubuf)
1165 /* set invalid range, so io_import_fixed() fails meeting it */
1166 ctx->dummy_ubuf->ubuf = -1UL;
1168 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1169 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1172 ctx->flags = p->flags;
1173 init_waitqueue_head(&ctx->sqo_sq_wait);
1174 INIT_LIST_HEAD(&ctx->sqd_list);
1175 init_waitqueue_head(&ctx->cq_wait);
1176 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1177 init_completion(&ctx->ref_comp);
1178 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1179 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1180 mutex_init(&ctx->uring_lock);
1181 init_waitqueue_head(&ctx->wait);
1182 spin_lock_init(&ctx->completion_lock);
1183 INIT_LIST_HEAD(&ctx->iopoll_list);
1184 INIT_LIST_HEAD(&ctx->defer_list);
1185 INIT_LIST_HEAD(&ctx->timeout_list);
1186 spin_lock_init(&ctx->rsrc_ref_lock);
1187 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1188 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1189 init_llist_head(&ctx->rsrc_put_llist);
1190 INIT_LIST_HEAD(&ctx->tctx_list);
1191 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1192 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1195 kfree(ctx->dummy_ubuf);
1196 kfree(ctx->cancel_hash);
1201 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1203 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1204 struct io_ring_ctx *ctx = req->ctx;
1206 return seq + ctx->cq_extra != ctx->cached_cq_tail
1207 + READ_ONCE(ctx->cached_cq_overflow);
1213 static void io_req_track_inflight(struct io_kiocb *req)
1215 if (!(req->flags & REQ_F_INFLIGHT)) {
1216 req->flags |= REQ_F_INFLIGHT;
1217 atomic_inc(¤t->io_uring->inflight_tracked);
1221 static void io_prep_async_work(struct io_kiocb *req)
1223 const struct io_op_def *def = &io_op_defs[req->opcode];
1224 struct io_ring_ctx *ctx = req->ctx;
1226 if (!req->work.creds)
1227 req->work.creds = get_current_cred();
1229 req->work.list.next = NULL;
1230 req->work.flags = 0;
1231 if (req->flags & REQ_F_FORCE_ASYNC)
1232 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1234 if (req->flags & REQ_F_ISREG) {
1235 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1236 io_wq_hash_work(&req->work, file_inode(req->file));
1237 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1238 if (def->unbound_nonreg_file)
1239 req->work.flags |= IO_WQ_WORK_UNBOUND;
1242 switch (req->opcode) {
1243 case IORING_OP_SPLICE:
1245 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1246 req->work.flags |= IO_WQ_WORK_UNBOUND;
1251 static void io_prep_async_link(struct io_kiocb *req)
1253 struct io_kiocb *cur;
1255 io_for_each_link(cur, req)
1256 io_prep_async_work(cur);
1259 static void io_queue_async_work(struct io_kiocb *req)
1261 struct io_ring_ctx *ctx = req->ctx;
1262 struct io_kiocb *link = io_prep_linked_timeout(req);
1263 struct io_uring_task *tctx = req->task->io_uring;
1266 BUG_ON(!tctx->io_wq);
1268 /* init ->work of the whole link before punting */
1269 io_prep_async_link(req);
1270 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1271 &req->work, req->flags);
1272 io_wq_enqueue(tctx->io_wq, &req->work);
1274 io_queue_linked_timeout(link);
1277 static void io_kill_timeout(struct io_kiocb *req, int status)
1278 __must_hold(&req->ctx->completion_lock)
1280 struct io_timeout_data *io = req->async_data;
1282 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1283 atomic_set(&req->ctx->cq_timeouts,
1284 atomic_read(&req->ctx->cq_timeouts) + 1);
1285 list_del_init(&req->timeout.list);
1286 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1287 io_put_req_deferred(req, 1);
1291 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1294 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1295 struct io_defer_entry, list);
1297 if (req_need_defer(de->req, de->seq))
1299 list_del_init(&de->list);
1300 io_req_task_queue(de->req);
1302 } while (!list_empty(&ctx->defer_list));
1305 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1309 if (list_empty(&ctx->timeout_list))
1312 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1315 u32 events_needed, events_got;
1316 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1317 struct io_kiocb, timeout.list);
1319 if (io_is_timeout_noseq(req))
1323 * Since seq can easily wrap around over time, subtract
1324 * the last seq at which timeouts were flushed before comparing.
1325 * Assuming not more than 2^31-1 events have happened since,
1326 * these subtractions won't have wrapped, so we can check if
1327 * target is in [last_seq, current_seq] by comparing the two.
1329 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1330 events_got = seq - ctx->cq_last_tm_flush;
1331 if (events_got < events_needed)
1334 list_del_init(&req->timeout.list);
1335 io_kill_timeout(req, 0);
1336 } while (!list_empty(&ctx->timeout_list));
1338 ctx->cq_last_tm_flush = seq;
1341 static void io_commit_cqring(struct io_ring_ctx *ctx)
1343 io_flush_timeouts(ctx);
1345 /* order cqe stores with ring update */
1346 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1348 if (unlikely(!list_empty(&ctx->defer_list)))
1349 __io_queue_deferred(ctx);
1352 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1354 struct io_rings *r = ctx->rings;
1356 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1359 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1361 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1364 static inline struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1366 struct io_rings *rings = ctx->rings;
1370 * writes to the cq entry need to come after reading head; the
1371 * control dependency is enough as we're using WRITE_ONCE to
1374 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1377 tail = ctx->cached_cq_tail++;
1378 return &rings->cqes[tail & ctx->cq_mask];
1381 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1383 if (likely(!ctx->cq_ev_fd))
1385 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1387 return !ctx->eventfd_async || io_wq_current_is_worker();
1390 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1392 /* see waitqueue_active() comment */
1395 if (waitqueue_active(&ctx->wait))
1396 wake_up(&ctx->wait);
1397 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1398 wake_up(&ctx->sq_data->wait);
1399 if (io_should_trigger_evfd(ctx))
1400 eventfd_signal(ctx->cq_ev_fd, 1);
1401 if (waitqueue_active(&ctx->cq_wait)) {
1402 wake_up_interruptible(&ctx->cq_wait);
1403 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1407 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1409 /* see waitqueue_active() comment */
1412 if (ctx->flags & IORING_SETUP_SQPOLL) {
1413 if (waitqueue_active(&ctx->wait))
1414 wake_up(&ctx->wait);
1416 if (io_should_trigger_evfd(ctx))
1417 eventfd_signal(ctx->cq_ev_fd, 1);
1418 if (waitqueue_active(&ctx->cq_wait)) {
1419 wake_up_interruptible(&ctx->cq_wait);
1420 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1424 /* Returns true if there are no backlogged entries after the flush */
1425 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1427 struct io_rings *rings = ctx->rings;
1428 unsigned long flags;
1429 bool all_flushed, posted;
1431 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1435 spin_lock_irqsave(&ctx->completion_lock, flags);
1436 while (!list_empty(&ctx->cq_overflow_list)) {
1437 struct io_uring_cqe *cqe = io_get_cqring(ctx);
1438 struct io_overflow_cqe *ocqe;
1442 ocqe = list_first_entry(&ctx->cq_overflow_list,
1443 struct io_overflow_cqe, list);
1445 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1447 WRITE_ONCE(ctx->rings->cq_overflow,
1448 ++ctx->cached_cq_overflow);
1450 list_del(&ocqe->list);
1454 all_flushed = list_empty(&ctx->cq_overflow_list);
1456 clear_bit(0, &ctx->sq_check_overflow);
1457 clear_bit(0, &ctx->cq_check_overflow);
1458 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1462 io_commit_cqring(ctx);
1463 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1465 io_cqring_ev_posted(ctx);
1469 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1473 if (test_bit(0, &ctx->cq_check_overflow)) {
1474 /* iopoll syncs against uring_lock, not completion_lock */
1475 if (ctx->flags & IORING_SETUP_IOPOLL)
1476 mutex_lock(&ctx->uring_lock);
1477 ret = __io_cqring_overflow_flush(ctx, force);
1478 if (ctx->flags & IORING_SETUP_IOPOLL)
1479 mutex_unlock(&ctx->uring_lock);
1486 * Shamelessly stolen from the mm implementation of page reference checking,
1487 * see commit f958d7b528b1 for details.
1489 #define req_ref_zero_or_close_to_overflow(req) \
1490 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1492 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1494 return atomic_inc_not_zero(&req->refs);
1497 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1499 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1500 return atomic_sub_and_test(refs, &req->refs);
1503 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1505 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1506 return atomic_dec_and_test(&req->refs);
1509 static inline void req_ref_put(struct io_kiocb *req)
1511 WARN_ON_ONCE(req_ref_put_and_test(req));
1514 static inline void req_ref_get(struct io_kiocb *req)
1516 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1517 atomic_inc(&req->refs);
1520 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1521 long res, unsigned int cflags)
1523 struct io_overflow_cqe *ocqe;
1525 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1528 * If we're in ring overflow flush mode, or in task cancel mode,
1529 * or cannot allocate an overflow entry, then we need to drop it
1532 WRITE_ONCE(ctx->rings->cq_overflow, ++ctx->cached_cq_overflow);
1535 if (list_empty(&ctx->cq_overflow_list)) {
1536 set_bit(0, &ctx->sq_check_overflow);
1537 set_bit(0, &ctx->cq_check_overflow);
1538 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1540 ocqe->cqe.user_data = user_data;
1541 ocqe->cqe.res = res;
1542 ocqe->cqe.flags = cflags;
1543 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1547 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1548 long res, unsigned int cflags)
1550 struct io_uring_cqe *cqe;
1552 trace_io_uring_complete(ctx, user_data, res, cflags);
1555 * If we can't get a cq entry, userspace overflowed the
1556 * submission (by quite a lot). Increment the overflow count in
1559 cqe = io_get_cqring(ctx);
1561 WRITE_ONCE(cqe->user_data, user_data);
1562 WRITE_ONCE(cqe->res, res);
1563 WRITE_ONCE(cqe->flags, cflags);
1566 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1569 /* not as hot to bloat with inlining */
1570 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1571 long res, unsigned int cflags)
1573 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1576 static void io_req_complete_post(struct io_kiocb *req, long res,
1577 unsigned int cflags)
1579 struct io_ring_ctx *ctx = req->ctx;
1580 unsigned long flags;
1582 spin_lock_irqsave(&ctx->completion_lock, flags);
1583 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1585 * If we're the last reference to this request, add to our locked
1588 if (req_ref_put_and_test(req)) {
1589 struct io_comp_state *cs = &ctx->submit_state.comp;
1591 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1592 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1593 io_disarm_next(req);
1595 io_req_task_queue(req->link);
1599 io_dismantle_req(req);
1600 io_put_task(req->task, 1);
1601 list_add(&req->compl.list, &cs->locked_free_list);
1602 cs->locked_free_nr++;
1604 if (!percpu_ref_tryget(&ctx->refs))
1607 io_commit_cqring(ctx);
1608 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1611 io_cqring_ev_posted(ctx);
1612 percpu_ref_put(&ctx->refs);
1616 static inline bool io_req_needs_clean(struct io_kiocb *req)
1618 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1619 REQ_F_POLLED | REQ_F_INFLIGHT);
1622 static void io_req_complete_state(struct io_kiocb *req, long res,
1623 unsigned int cflags)
1625 if (io_req_needs_clean(req))
1628 req->compl.cflags = cflags;
1629 req->flags |= REQ_F_COMPLETE_INLINE;
1632 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1633 long res, unsigned cflags)
1635 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1636 io_req_complete_state(req, res, cflags);
1638 io_req_complete_post(req, res, cflags);
1641 static inline void io_req_complete(struct io_kiocb *req, long res)
1643 __io_req_complete(req, 0, res, 0);
1646 static void io_req_complete_failed(struct io_kiocb *req, long res)
1648 req_set_fail_links(req);
1650 io_req_complete_post(req, res, 0);
1653 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1654 struct io_comp_state *cs)
1656 spin_lock_irq(&ctx->completion_lock);
1657 list_splice_init(&cs->locked_free_list, &cs->free_list);
1658 cs->locked_free_nr = 0;
1659 spin_unlock_irq(&ctx->completion_lock);
1662 /* Returns true IFF there are requests in the cache */
1663 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1665 struct io_submit_state *state = &ctx->submit_state;
1666 struct io_comp_state *cs = &state->comp;
1670 * If we have more than a batch's worth of requests in our IRQ side
1671 * locked cache, grab the lock and move them over to our submission
1674 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH)
1675 io_flush_cached_locked_reqs(ctx, cs);
1677 nr = state->free_reqs;
1678 while (!list_empty(&cs->free_list)) {
1679 struct io_kiocb *req = list_first_entry(&cs->free_list,
1680 struct io_kiocb, compl.list);
1682 list_del(&req->compl.list);
1683 state->reqs[nr++] = req;
1684 if (nr == ARRAY_SIZE(state->reqs))
1688 state->free_reqs = nr;
1692 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1694 struct io_submit_state *state = &ctx->submit_state;
1696 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1698 if (!state->free_reqs) {
1699 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1702 if (io_flush_cached_reqs(ctx))
1705 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1709 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1710 * retry single alloc to be on the safe side.
1712 if (unlikely(ret <= 0)) {
1713 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1714 if (!state->reqs[0])
1718 state->free_reqs = ret;
1722 return state->reqs[state->free_reqs];
1725 static inline void io_put_file(struct file *file)
1731 static void io_dismantle_req(struct io_kiocb *req)
1733 unsigned int flags = req->flags;
1735 if (io_req_needs_clean(req))
1737 if (!(flags & REQ_F_FIXED_FILE))
1738 io_put_file(req->file);
1739 if (req->fixed_rsrc_refs)
1740 percpu_ref_put(req->fixed_rsrc_refs);
1741 if (req->async_data)
1742 kfree(req->async_data);
1743 if (req->work.creds) {
1744 put_cred(req->work.creds);
1745 req->work.creds = NULL;
1749 /* must to be called somewhat shortly after putting a request */
1750 static inline void io_put_task(struct task_struct *task, int nr)
1752 struct io_uring_task *tctx = task->io_uring;
1754 percpu_counter_sub(&tctx->inflight, nr);
1755 if (unlikely(atomic_read(&tctx->in_idle)))
1756 wake_up(&tctx->wait);
1757 put_task_struct_many(task, nr);
1760 static void __io_free_req(struct io_kiocb *req)
1762 struct io_ring_ctx *ctx = req->ctx;
1764 io_dismantle_req(req);
1765 io_put_task(req->task, 1);
1767 kmem_cache_free(req_cachep, req);
1768 percpu_ref_put(&ctx->refs);
1771 static inline void io_remove_next_linked(struct io_kiocb *req)
1773 struct io_kiocb *nxt = req->link;
1775 req->link = nxt->link;
1779 static bool io_kill_linked_timeout(struct io_kiocb *req)
1780 __must_hold(&req->ctx->completion_lock)
1782 struct io_kiocb *link = req->link;
1785 * Can happen if a linked timeout fired and link had been like
1786 * req -> link t-out -> link t-out [-> ...]
1788 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1789 struct io_timeout_data *io = link->async_data;
1791 io_remove_next_linked(req);
1792 link->timeout.head = NULL;
1793 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1794 io_cqring_fill_event(link->ctx, link->user_data,
1796 io_put_req_deferred(link, 1);
1803 static void io_fail_links(struct io_kiocb *req)
1804 __must_hold(&req->ctx->completion_lock)
1806 struct io_kiocb *nxt, *link = req->link;
1813 trace_io_uring_fail_link(req, link);
1814 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1815 io_put_req_deferred(link, 2);
1820 static bool io_disarm_next(struct io_kiocb *req)
1821 __must_hold(&req->ctx->completion_lock)
1823 bool posted = false;
1825 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1826 posted = io_kill_linked_timeout(req);
1827 if (unlikely((req->flags & REQ_F_FAIL_LINK) &&
1828 !(req->flags & REQ_F_HARDLINK))) {
1829 posted |= (req->link != NULL);
1835 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1837 struct io_kiocb *nxt;
1840 * If LINK is set, we have dependent requests in this chain. If we
1841 * didn't fail this request, queue the first one up, moving any other
1842 * dependencies to the next request. In case of failure, fail the rest
1845 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1846 struct io_ring_ctx *ctx = req->ctx;
1847 unsigned long flags;
1850 spin_lock_irqsave(&ctx->completion_lock, flags);
1851 posted = io_disarm_next(req);
1853 io_commit_cqring(req->ctx);
1854 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1856 io_cqring_ev_posted(ctx);
1863 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1865 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1867 return __io_req_find_next(req);
1870 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1874 if (ctx->submit_state.comp.nr) {
1875 mutex_lock(&ctx->uring_lock);
1876 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1877 mutex_unlock(&ctx->uring_lock);
1879 percpu_ref_put(&ctx->refs);
1882 static bool __tctx_task_work(struct io_uring_task *tctx)
1884 struct io_ring_ctx *ctx = NULL;
1885 struct io_wq_work_list list;
1886 struct io_wq_work_node *node;
1888 if (wq_list_empty(&tctx->task_list))
1891 spin_lock_irq(&tctx->task_lock);
1892 list = tctx->task_list;
1893 INIT_WQ_LIST(&tctx->task_list);
1894 spin_unlock_irq(&tctx->task_lock);
1898 struct io_wq_work_node *next = node->next;
1899 struct io_kiocb *req;
1901 req = container_of(node, struct io_kiocb, io_task_work.node);
1902 if (req->ctx != ctx) {
1903 ctx_flush_and_put(ctx);
1905 percpu_ref_get(&ctx->refs);
1908 req->task_work.func(&req->task_work);
1912 ctx_flush_and_put(ctx);
1913 return list.first != NULL;
1916 static void tctx_task_work(struct callback_head *cb)
1918 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1920 clear_bit(0, &tctx->task_state);
1922 while (__tctx_task_work(tctx))
1926 static int io_req_task_work_add(struct io_kiocb *req)
1928 struct task_struct *tsk = req->task;
1929 struct io_uring_task *tctx = tsk->io_uring;
1930 enum task_work_notify_mode notify;
1931 struct io_wq_work_node *node, *prev;
1932 unsigned long flags;
1935 if (unlikely(tsk->flags & PF_EXITING))
1938 WARN_ON_ONCE(!tctx);
1940 spin_lock_irqsave(&tctx->task_lock, flags);
1941 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1942 spin_unlock_irqrestore(&tctx->task_lock, flags);
1944 /* task_work already pending, we're done */
1945 if (test_bit(0, &tctx->task_state) ||
1946 test_and_set_bit(0, &tctx->task_state))
1950 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1951 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1952 * processing task_work. There's no reliable way to tell if TWA_RESUME
1955 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1957 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1958 wake_up_process(tsk);
1963 * Slow path - we failed, find and delete work. if the work is not
1964 * in the list, it got run and we're fine.
1966 spin_lock_irqsave(&tctx->task_lock, flags);
1967 wq_list_for_each(node, prev, &tctx->task_list) {
1968 if (&req->io_task_work.node == node) {
1969 wq_list_del(&tctx->task_list, node, prev);
1974 spin_unlock_irqrestore(&tctx->task_lock, flags);
1975 clear_bit(0, &tctx->task_state);
1979 static bool io_run_task_work_head(struct callback_head **work_head)
1981 struct callback_head *work, *next;
1982 bool executed = false;
1985 work = xchg(work_head, NULL);
2001 static void io_task_work_add_head(struct callback_head **work_head,
2002 struct callback_head *task_work)
2004 struct callback_head *head;
2007 head = READ_ONCE(*work_head);
2008 task_work->next = head;
2009 } while (cmpxchg(work_head, head, task_work) != head);
2012 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2013 task_work_func_t cb)
2015 init_task_work(&req->task_work, cb);
2016 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2019 static void io_req_task_cancel(struct callback_head *cb)
2021 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2022 struct io_ring_ctx *ctx = req->ctx;
2024 /* ctx is guaranteed to stay alive while we hold uring_lock */
2025 mutex_lock(&ctx->uring_lock);
2026 io_req_complete_failed(req, req->result);
2027 mutex_unlock(&ctx->uring_lock);
2030 static void __io_req_task_submit(struct io_kiocb *req)
2032 struct io_ring_ctx *ctx = req->ctx;
2034 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2035 mutex_lock(&ctx->uring_lock);
2036 if (!(current->flags & PF_EXITING) && !current->in_execve)
2037 __io_queue_sqe(req);
2039 io_req_complete_failed(req, -EFAULT);
2040 mutex_unlock(&ctx->uring_lock);
2043 static void io_req_task_submit(struct callback_head *cb)
2045 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2047 __io_req_task_submit(req);
2050 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2053 req->task_work.func = io_req_task_cancel;
2055 if (unlikely(io_req_task_work_add(req)))
2056 io_req_task_work_add_fallback(req, io_req_task_cancel);
2059 static void io_req_task_queue(struct io_kiocb *req)
2061 req->task_work.func = io_req_task_submit;
2063 if (unlikely(io_req_task_work_add(req)))
2064 io_req_task_queue_fail(req, -ECANCELED);
2067 static inline void io_queue_next(struct io_kiocb *req)
2069 struct io_kiocb *nxt = io_req_find_next(req);
2072 io_req_task_queue(nxt);
2075 static void io_free_req(struct io_kiocb *req)
2082 struct task_struct *task;
2087 static inline void io_init_req_batch(struct req_batch *rb)
2094 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2095 struct req_batch *rb)
2098 io_put_task(rb->task, rb->task_refs);
2100 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2103 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2104 struct io_submit_state *state)
2107 io_dismantle_req(req);
2109 if (req->task != rb->task) {
2111 io_put_task(rb->task, rb->task_refs);
2112 rb->task = req->task;
2118 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2119 state->reqs[state->free_reqs++] = req;
2121 list_add(&req->compl.list, &state->comp.free_list);
2124 static void io_submit_flush_completions(struct io_comp_state *cs,
2125 struct io_ring_ctx *ctx)
2128 struct io_kiocb *req;
2129 struct req_batch rb;
2131 io_init_req_batch(&rb);
2132 spin_lock_irq(&ctx->completion_lock);
2133 for (i = 0; i < nr; i++) {
2135 __io_cqring_fill_event(ctx, req->user_data, req->result,
2138 io_commit_cqring(ctx);
2139 spin_unlock_irq(&ctx->completion_lock);
2141 io_cqring_ev_posted(ctx);
2142 for (i = 0; i < nr; i++) {
2145 /* submission and completion refs */
2146 if (req_ref_sub_and_test(req, 2))
2147 io_req_free_batch(&rb, req, &ctx->submit_state);
2150 io_req_free_batch_finish(ctx, &rb);
2155 * Drop reference to request, return next in chain (if there is one) if this
2156 * was the last reference to this request.
2158 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2160 struct io_kiocb *nxt = NULL;
2162 if (req_ref_put_and_test(req)) {
2163 nxt = io_req_find_next(req);
2169 static inline void io_put_req(struct io_kiocb *req)
2171 if (req_ref_put_and_test(req))
2175 static void io_put_req_deferred_cb(struct callback_head *cb)
2177 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2182 static void io_free_req_deferred(struct io_kiocb *req)
2184 req->task_work.func = io_put_req_deferred_cb;
2185 if (unlikely(io_req_task_work_add(req)))
2186 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2189 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2191 if (req_ref_sub_and_test(req, refs))
2192 io_free_req_deferred(req);
2195 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2197 /* See comment at the top of this file */
2199 return __io_cqring_events(ctx);
2202 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2204 struct io_rings *rings = ctx->rings;
2206 /* make sure SQ entry isn't read before tail */
2207 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2210 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2212 unsigned int cflags;
2214 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2215 cflags |= IORING_CQE_F_BUFFER;
2216 req->flags &= ~REQ_F_BUFFER_SELECTED;
2221 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2223 struct io_buffer *kbuf;
2225 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2226 return io_put_kbuf(req, kbuf);
2229 static inline bool io_run_task_work(void)
2232 * Not safe to run on exiting task, and the task_work handling will
2233 * not add work to such a task.
2235 if (unlikely(current->flags & PF_EXITING))
2237 if (current->task_works) {
2238 __set_current_state(TASK_RUNNING);
2247 * Find and free completed poll iocbs
2249 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2250 struct list_head *done)
2252 struct req_batch rb;
2253 struct io_kiocb *req;
2255 /* order with ->result store in io_complete_rw_iopoll() */
2258 io_init_req_batch(&rb);
2259 while (!list_empty(done)) {
2262 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2263 list_del(&req->inflight_entry);
2265 if (READ_ONCE(req->result) == -EAGAIN &&
2266 !(req->flags & REQ_F_DONT_REISSUE)) {
2267 req->iopoll_completed = 0;
2269 io_queue_async_work(req);
2273 if (req->flags & REQ_F_BUFFER_SELECTED)
2274 cflags = io_put_rw_kbuf(req);
2276 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2279 if (req_ref_put_and_test(req))
2280 io_req_free_batch(&rb, req, &ctx->submit_state);
2283 io_commit_cqring(ctx);
2284 io_cqring_ev_posted_iopoll(ctx);
2285 io_req_free_batch_finish(ctx, &rb);
2288 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2291 struct io_kiocb *req, *tmp;
2297 * Only spin for completions if we don't have multiple devices hanging
2298 * off our complete list, and we're under the requested amount.
2300 spin = !ctx->poll_multi_file && *nr_events < min;
2303 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2304 struct kiocb *kiocb = &req->rw.kiocb;
2307 * Move completed and retryable entries to our local lists.
2308 * If we find a request that requires polling, break out
2309 * and complete those lists first, if we have entries there.
2311 if (READ_ONCE(req->iopoll_completed)) {
2312 list_move_tail(&req->inflight_entry, &done);
2315 if (!list_empty(&done))
2318 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2322 /* iopoll may have completed current req */
2323 if (READ_ONCE(req->iopoll_completed))
2324 list_move_tail(&req->inflight_entry, &done);
2331 if (!list_empty(&done))
2332 io_iopoll_complete(ctx, nr_events, &done);
2338 * We can't just wait for polled events to come to us, we have to actively
2339 * find and complete them.
2341 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2343 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2346 mutex_lock(&ctx->uring_lock);
2347 while (!list_empty(&ctx->iopoll_list)) {
2348 unsigned int nr_events = 0;
2350 io_do_iopoll(ctx, &nr_events, 0);
2352 /* let it sleep and repeat later if can't complete a request */
2356 * Ensure we allow local-to-the-cpu processing to take place,
2357 * in this case we need to ensure that we reap all events.
2358 * Also let task_work, etc. to progress by releasing the mutex
2360 if (need_resched()) {
2361 mutex_unlock(&ctx->uring_lock);
2363 mutex_lock(&ctx->uring_lock);
2366 mutex_unlock(&ctx->uring_lock);
2369 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2371 unsigned int nr_events = 0;
2375 * We disallow the app entering submit/complete with polling, but we
2376 * still need to lock the ring to prevent racing with polled issue
2377 * that got punted to a workqueue.
2379 mutex_lock(&ctx->uring_lock);
2381 * Don't enter poll loop if we already have events pending.
2382 * If we do, we can potentially be spinning for commands that
2383 * already triggered a CQE (eg in error).
2385 if (test_bit(0, &ctx->cq_check_overflow))
2386 __io_cqring_overflow_flush(ctx, false);
2387 if (io_cqring_events(ctx))
2391 * If a submit got punted to a workqueue, we can have the
2392 * application entering polling for a command before it gets
2393 * issued. That app will hold the uring_lock for the duration
2394 * of the poll right here, so we need to take a breather every
2395 * now and then to ensure that the issue has a chance to add
2396 * the poll to the issued list. Otherwise we can spin here
2397 * forever, while the workqueue is stuck trying to acquire the
2400 if (list_empty(&ctx->iopoll_list)) {
2401 mutex_unlock(&ctx->uring_lock);
2403 mutex_lock(&ctx->uring_lock);
2405 if (list_empty(&ctx->iopoll_list))
2408 ret = io_do_iopoll(ctx, &nr_events, min);
2409 } while (!ret && nr_events < min && !need_resched());
2411 mutex_unlock(&ctx->uring_lock);
2415 static void kiocb_end_write(struct io_kiocb *req)
2418 * Tell lockdep we inherited freeze protection from submission
2421 if (req->flags & REQ_F_ISREG) {
2422 struct super_block *sb = file_inode(req->file)->i_sb;
2424 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2430 static bool io_resubmit_prep(struct io_kiocb *req)
2432 struct io_async_rw *rw = req->async_data;
2435 return !io_req_prep_async(req);
2436 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2437 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2441 static bool io_rw_should_reissue(struct io_kiocb *req)
2443 umode_t mode = file_inode(req->file)->i_mode;
2444 struct io_ring_ctx *ctx = req->ctx;
2446 if (!S_ISBLK(mode) && !S_ISREG(mode))
2448 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2449 !(ctx->flags & IORING_SETUP_IOPOLL)))
2452 * If ref is dying, we might be running poll reap from the exit work.
2453 * Don't attempt to reissue from that path, just let it fail with
2456 if (percpu_ref_is_dying(&ctx->refs))
2461 static bool io_resubmit_prep(struct io_kiocb *req)
2465 static bool io_rw_should_reissue(struct io_kiocb *req)
2471 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2472 unsigned int issue_flags)
2476 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2477 kiocb_end_write(req);
2478 if (res != req->result) {
2479 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2480 io_rw_should_reissue(req)) {
2481 req->flags |= REQ_F_REISSUE;
2484 req_set_fail_links(req);
2486 if (req->flags & REQ_F_BUFFER_SELECTED)
2487 cflags = io_put_rw_kbuf(req);
2488 __io_req_complete(req, issue_flags, res, cflags);
2491 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2493 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2495 __io_complete_rw(req, res, res2, 0);
2498 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2500 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2502 if (kiocb->ki_flags & IOCB_WRITE)
2503 kiocb_end_write(req);
2504 if (unlikely(res != req->result)) {
2505 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2506 io_resubmit_prep(req))) {
2507 req_set_fail_links(req);
2508 req->flags |= REQ_F_DONT_REISSUE;
2512 WRITE_ONCE(req->result, res);
2513 /* order with io_iopoll_complete() checking ->result */
2515 WRITE_ONCE(req->iopoll_completed, 1);
2519 * After the iocb has been issued, it's safe to be found on the poll list.
2520 * Adding the kiocb to the list AFTER submission ensures that we don't
2521 * find it from a io_do_iopoll() thread before the issuer is done
2522 * accessing the kiocb cookie.
2524 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2526 struct io_ring_ctx *ctx = req->ctx;
2529 * Track whether we have multiple files in our lists. This will impact
2530 * how we do polling eventually, not spinning if we're on potentially
2531 * different devices.
2533 if (list_empty(&ctx->iopoll_list)) {
2534 ctx->poll_multi_file = false;
2535 } else if (!ctx->poll_multi_file) {
2536 struct io_kiocb *list_req;
2538 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2540 if (list_req->file != req->file)
2541 ctx->poll_multi_file = true;
2545 * For fast devices, IO may have already completed. If it has, add
2546 * it to the front so we find it first.
2548 if (READ_ONCE(req->iopoll_completed))
2549 list_add(&req->inflight_entry, &ctx->iopoll_list);
2551 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2554 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2555 * task context or in io worker task context. If current task context is
2556 * sq thread, we don't need to check whether should wake up sq thread.
2558 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2559 wq_has_sleeper(&ctx->sq_data->wait))
2560 wake_up(&ctx->sq_data->wait);
2563 static inline void io_state_file_put(struct io_submit_state *state)
2565 if (state->file_refs) {
2566 fput_many(state->file, state->file_refs);
2567 state->file_refs = 0;
2572 * Get as many references to a file as we have IOs left in this submission,
2573 * assuming most submissions are for one file, or at least that each file
2574 * has more than one submission.
2576 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2581 if (state->file_refs) {
2582 if (state->fd == fd) {
2586 io_state_file_put(state);
2588 state->file = fget_many(fd, state->ios_left);
2589 if (unlikely(!state->file))
2593 state->file_refs = state->ios_left - 1;
2597 static bool io_bdev_nowait(struct block_device *bdev)
2599 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2603 * If we tracked the file through the SCM inflight mechanism, we could support
2604 * any file. For now, just ensure that anything potentially problematic is done
2607 static bool __io_file_supports_async(struct file *file, int rw)
2609 umode_t mode = file_inode(file)->i_mode;
2611 if (S_ISBLK(mode)) {
2612 if (IS_ENABLED(CONFIG_BLOCK) &&
2613 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2617 if (S_ISCHR(mode) || S_ISSOCK(mode))
2619 if (S_ISREG(mode)) {
2620 if (IS_ENABLED(CONFIG_BLOCK) &&
2621 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2622 file->f_op != &io_uring_fops)
2627 /* any ->read/write should understand O_NONBLOCK */
2628 if (file->f_flags & O_NONBLOCK)
2631 if (!(file->f_mode & FMODE_NOWAIT))
2635 return file->f_op->read_iter != NULL;
2637 return file->f_op->write_iter != NULL;
2640 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2642 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2644 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2647 return __io_file_supports_async(req->file, rw);
2650 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2652 struct io_ring_ctx *ctx = req->ctx;
2653 struct kiocb *kiocb = &req->rw.kiocb;
2654 struct file *file = req->file;
2658 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2659 req->flags |= REQ_F_ISREG;
2661 kiocb->ki_pos = READ_ONCE(sqe->off);
2662 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2663 req->flags |= REQ_F_CUR_POS;
2664 kiocb->ki_pos = file->f_pos;
2666 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2667 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2668 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2672 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2673 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2674 req->flags |= REQ_F_NOWAIT;
2676 ioprio = READ_ONCE(sqe->ioprio);
2678 ret = ioprio_check_cap(ioprio);
2682 kiocb->ki_ioprio = ioprio;
2684 kiocb->ki_ioprio = get_current_ioprio();
2686 if (ctx->flags & IORING_SETUP_IOPOLL) {
2687 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2688 !kiocb->ki_filp->f_op->iopoll)
2691 kiocb->ki_flags |= IOCB_HIPRI;
2692 kiocb->ki_complete = io_complete_rw_iopoll;
2693 req->iopoll_completed = 0;
2695 if (kiocb->ki_flags & IOCB_HIPRI)
2697 kiocb->ki_complete = io_complete_rw;
2700 if (req->opcode == IORING_OP_READ_FIXED ||
2701 req->opcode == IORING_OP_WRITE_FIXED) {
2703 io_req_set_rsrc_node(req);
2706 req->rw.addr = READ_ONCE(sqe->addr);
2707 req->rw.len = READ_ONCE(sqe->len);
2708 req->buf_index = READ_ONCE(sqe->buf_index);
2712 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2718 case -ERESTARTNOINTR:
2719 case -ERESTARTNOHAND:
2720 case -ERESTART_RESTARTBLOCK:
2722 * We can't just restart the syscall, since previously
2723 * submitted sqes may already be in progress. Just fail this
2729 kiocb->ki_complete(kiocb, ret, 0);
2733 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2734 unsigned int issue_flags)
2736 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2737 struct io_async_rw *io = req->async_data;
2738 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2740 /* add previously done IO, if any */
2741 if (io && io->bytes_done > 0) {
2743 ret = io->bytes_done;
2745 ret += io->bytes_done;
2748 if (req->flags & REQ_F_CUR_POS)
2749 req->file->f_pos = kiocb->ki_pos;
2750 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2751 __io_complete_rw(req, ret, 0, issue_flags);
2753 io_rw_done(kiocb, ret);
2755 if (check_reissue && req->flags & REQ_F_REISSUE) {
2756 req->flags &= ~REQ_F_REISSUE;
2757 if (io_resubmit_prep(req)) {
2759 io_queue_async_work(req);
2763 req_set_fail_links(req);
2764 if (req->flags & REQ_F_BUFFER_SELECTED)
2765 cflags = io_put_rw_kbuf(req);
2766 __io_req_complete(req, issue_flags, ret, cflags);
2771 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2772 struct io_mapped_ubuf *imu)
2774 size_t len = req->rw.len;
2775 u64 buf_end, buf_addr = req->rw.addr;
2778 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2780 /* not inside the mapped region */
2781 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2785 * May not be a start of buffer, set size appropriately
2786 * and advance us to the beginning.
2788 offset = buf_addr - imu->ubuf;
2789 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2793 * Don't use iov_iter_advance() here, as it's really slow for
2794 * using the latter parts of a big fixed buffer - it iterates
2795 * over each segment manually. We can cheat a bit here, because
2798 * 1) it's a BVEC iter, we set it up
2799 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2800 * first and last bvec
2802 * So just find our index, and adjust the iterator afterwards.
2803 * If the offset is within the first bvec (or the whole first
2804 * bvec, just use iov_iter_advance(). This makes it easier
2805 * since we can just skip the first segment, which may not
2806 * be PAGE_SIZE aligned.
2808 const struct bio_vec *bvec = imu->bvec;
2810 if (offset <= bvec->bv_len) {
2811 iov_iter_advance(iter, offset);
2813 unsigned long seg_skip;
2815 /* skip first vec */
2816 offset -= bvec->bv_len;
2817 seg_skip = 1 + (offset >> PAGE_SHIFT);
2819 iter->bvec = bvec + seg_skip;
2820 iter->nr_segs -= seg_skip;
2821 iter->count -= bvec->bv_len + offset;
2822 iter->iov_offset = offset & ~PAGE_MASK;
2829 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2831 struct io_ring_ctx *ctx = req->ctx;
2832 struct io_mapped_ubuf *imu = req->imu;
2833 u16 index, buf_index = req->buf_index;
2836 if (unlikely(buf_index >= ctx->nr_user_bufs))
2838 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2839 imu = READ_ONCE(ctx->user_bufs[index]);
2842 return __io_import_fixed(req, rw, iter, imu);
2845 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2848 mutex_unlock(&ctx->uring_lock);
2851 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2854 * "Normal" inline submissions always hold the uring_lock, since we
2855 * grab it from the system call. Same is true for the SQPOLL offload.
2856 * The only exception is when we've detached the request and issue it
2857 * from an async worker thread, grab the lock for that case.
2860 mutex_lock(&ctx->uring_lock);
2863 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2864 int bgid, struct io_buffer *kbuf,
2867 struct io_buffer *head;
2869 if (req->flags & REQ_F_BUFFER_SELECTED)
2872 io_ring_submit_lock(req->ctx, needs_lock);
2874 lockdep_assert_held(&req->ctx->uring_lock);
2876 head = xa_load(&req->ctx->io_buffers, bgid);
2878 if (!list_empty(&head->list)) {
2879 kbuf = list_last_entry(&head->list, struct io_buffer,
2881 list_del(&kbuf->list);
2884 xa_erase(&req->ctx->io_buffers, bgid);
2886 if (*len > kbuf->len)
2889 kbuf = ERR_PTR(-ENOBUFS);
2892 io_ring_submit_unlock(req->ctx, needs_lock);
2897 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2900 struct io_buffer *kbuf;
2903 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2904 bgid = req->buf_index;
2905 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2908 req->rw.addr = (u64) (unsigned long) kbuf;
2909 req->flags |= REQ_F_BUFFER_SELECTED;
2910 return u64_to_user_ptr(kbuf->addr);
2913 #ifdef CONFIG_COMPAT
2914 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2917 struct compat_iovec __user *uiov;
2918 compat_ssize_t clen;
2922 uiov = u64_to_user_ptr(req->rw.addr);
2923 if (!access_ok(uiov, sizeof(*uiov)))
2925 if (__get_user(clen, &uiov->iov_len))
2931 buf = io_rw_buffer_select(req, &len, needs_lock);
2933 return PTR_ERR(buf);
2934 iov[0].iov_base = buf;
2935 iov[0].iov_len = (compat_size_t) len;
2940 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2943 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2947 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2950 len = iov[0].iov_len;
2953 buf = io_rw_buffer_select(req, &len, needs_lock);
2955 return PTR_ERR(buf);
2956 iov[0].iov_base = buf;
2957 iov[0].iov_len = len;
2961 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2964 if (req->flags & REQ_F_BUFFER_SELECTED) {
2965 struct io_buffer *kbuf;
2967 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2968 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2969 iov[0].iov_len = kbuf->len;
2972 if (req->rw.len != 1)
2975 #ifdef CONFIG_COMPAT
2976 if (req->ctx->compat)
2977 return io_compat_import(req, iov, needs_lock);
2980 return __io_iov_buffer_select(req, iov, needs_lock);
2983 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2984 struct iov_iter *iter, bool needs_lock)
2986 void __user *buf = u64_to_user_ptr(req->rw.addr);
2987 size_t sqe_len = req->rw.len;
2988 u8 opcode = req->opcode;
2991 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2993 return io_import_fixed(req, rw, iter);
2996 /* buffer index only valid with fixed read/write, or buffer select */
2997 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3000 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3001 if (req->flags & REQ_F_BUFFER_SELECT) {
3002 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3004 return PTR_ERR(buf);
3005 req->rw.len = sqe_len;
3008 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3013 if (req->flags & REQ_F_BUFFER_SELECT) {
3014 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3016 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3021 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3025 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3027 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3031 * For files that don't have ->read_iter() and ->write_iter(), handle them
3032 * by looping over ->read() or ->write() manually.
3034 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3036 struct kiocb *kiocb = &req->rw.kiocb;
3037 struct file *file = req->file;
3041 * Don't support polled IO through this interface, and we can't
3042 * support non-blocking either. For the latter, this just causes
3043 * the kiocb to be handled from an async context.
3045 if (kiocb->ki_flags & IOCB_HIPRI)
3047 if (kiocb->ki_flags & IOCB_NOWAIT)
3050 while (iov_iter_count(iter)) {
3054 if (!iov_iter_is_bvec(iter)) {
3055 iovec = iov_iter_iovec(iter);
3057 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3058 iovec.iov_len = req->rw.len;
3062 nr = file->f_op->read(file, iovec.iov_base,
3063 iovec.iov_len, io_kiocb_ppos(kiocb));
3065 nr = file->f_op->write(file, iovec.iov_base,
3066 iovec.iov_len, io_kiocb_ppos(kiocb));
3075 if (nr != iovec.iov_len)
3079 iov_iter_advance(iter, nr);
3085 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3086 const struct iovec *fast_iov, struct iov_iter *iter)
3088 struct io_async_rw *rw = req->async_data;
3090 memcpy(&rw->iter, iter, sizeof(*iter));
3091 rw->free_iovec = iovec;
3093 /* can only be fixed buffers, no need to do anything */
3094 if (iov_iter_is_bvec(iter))
3097 unsigned iov_off = 0;
3099 rw->iter.iov = rw->fast_iov;
3100 if (iter->iov != fast_iov) {
3101 iov_off = iter->iov - fast_iov;
3102 rw->iter.iov += iov_off;
3104 if (rw->fast_iov != fast_iov)
3105 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3106 sizeof(struct iovec) * iter->nr_segs);
3108 req->flags |= REQ_F_NEED_CLEANUP;
3112 static inline int io_alloc_async_data(struct io_kiocb *req)
3114 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3115 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3116 return req->async_data == NULL;
3119 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3120 const struct iovec *fast_iov,
3121 struct iov_iter *iter, bool force)
3123 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3125 if (!req->async_data) {
3126 if (io_alloc_async_data(req)) {
3131 io_req_map_rw(req, iovec, fast_iov, iter);
3136 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3138 struct io_async_rw *iorw = req->async_data;
3139 struct iovec *iov = iorw->fast_iov;
3142 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3143 if (unlikely(ret < 0))
3146 iorw->bytes_done = 0;
3147 iorw->free_iovec = iov;
3149 req->flags |= REQ_F_NEED_CLEANUP;
3153 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3155 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3157 return io_prep_rw(req, sqe);
3161 * This is our waitqueue callback handler, registered through lock_page_async()
3162 * when we initially tried to do the IO with the iocb armed our waitqueue.
3163 * This gets called when the page is unlocked, and we generally expect that to
3164 * happen when the page IO is completed and the page is now uptodate. This will
3165 * queue a task_work based retry of the operation, attempting to copy the data
3166 * again. If the latter fails because the page was NOT uptodate, then we will
3167 * do a thread based blocking retry of the operation. That's the unexpected
3170 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3171 int sync, void *arg)
3173 struct wait_page_queue *wpq;
3174 struct io_kiocb *req = wait->private;
3175 struct wait_page_key *key = arg;
3177 wpq = container_of(wait, struct wait_page_queue, wait);
3179 if (!wake_page_match(wpq, key))
3182 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3183 list_del_init(&wait->entry);
3185 /* submit ref gets dropped, acquire a new one */
3187 io_req_task_queue(req);
3192 * This controls whether a given IO request should be armed for async page
3193 * based retry. If we return false here, the request is handed to the async
3194 * worker threads for retry. If we're doing buffered reads on a regular file,
3195 * we prepare a private wait_page_queue entry and retry the operation. This
3196 * will either succeed because the page is now uptodate and unlocked, or it
3197 * will register a callback when the page is unlocked at IO completion. Through
3198 * that callback, io_uring uses task_work to setup a retry of the operation.
3199 * That retry will attempt the buffered read again. The retry will generally
3200 * succeed, or in rare cases where it fails, we then fall back to using the
3201 * async worker threads for a blocking retry.
3203 static bool io_rw_should_retry(struct io_kiocb *req)
3205 struct io_async_rw *rw = req->async_data;
3206 struct wait_page_queue *wait = &rw->wpq;
3207 struct kiocb *kiocb = &req->rw.kiocb;
3209 /* never retry for NOWAIT, we just complete with -EAGAIN */
3210 if (req->flags & REQ_F_NOWAIT)
3213 /* Only for buffered IO */
3214 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3218 * just use poll if we can, and don't attempt if the fs doesn't
3219 * support callback based unlocks
3221 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3224 wait->wait.func = io_async_buf_func;
3225 wait->wait.private = req;
3226 wait->wait.flags = 0;
3227 INIT_LIST_HEAD(&wait->wait.entry);
3228 kiocb->ki_flags |= IOCB_WAITQ;
3229 kiocb->ki_flags &= ~IOCB_NOWAIT;
3230 kiocb->ki_waitq = wait;
3234 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3236 if (req->file->f_op->read_iter)
3237 return call_read_iter(req->file, &req->rw.kiocb, iter);
3238 else if (req->file->f_op->read)
3239 return loop_rw_iter(READ, req, iter);
3244 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3246 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3247 struct kiocb *kiocb = &req->rw.kiocb;
3248 struct iov_iter __iter, *iter = &__iter;
3249 struct io_async_rw *rw = req->async_data;
3250 ssize_t io_size, ret, ret2;
3251 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3257 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3261 io_size = iov_iter_count(iter);
3262 req->result = io_size;
3264 /* Ensure we clear previously set non-block flag */
3265 if (!force_nonblock)
3266 kiocb->ki_flags &= ~IOCB_NOWAIT;
3268 kiocb->ki_flags |= IOCB_NOWAIT;
3270 /* If the file doesn't support async, just async punt */
3271 if (force_nonblock && !io_file_supports_async(req, READ)) {
3272 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3273 return ret ?: -EAGAIN;
3276 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3277 if (unlikely(ret)) {
3282 ret = io_iter_do_read(req, iter);
3284 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3285 req->flags &= ~REQ_F_REISSUE;
3286 /* IOPOLL retry should happen for io-wq threads */
3287 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3289 /* no retry on NONBLOCK nor RWF_NOWAIT */
3290 if (req->flags & REQ_F_NOWAIT)
3292 /* some cases will consume bytes even on error returns */
3293 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3295 } else if (ret == -EIOCBQUEUED) {
3297 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3298 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3299 /* read all, failed, already did sync or don't want to retry */
3303 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3308 rw = req->async_data;
3309 /* now use our persistent iterator, if we aren't already */
3314 rw->bytes_done += ret;
3315 /* if we can retry, do so with the callbacks armed */
3316 if (!io_rw_should_retry(req)) {
3317 kiocb->ki_flags &= ~IOCB_WAITQ;
3322 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3323 * we get -EIOCBQUEUED, then we'll get a notification when the
3324 * desired page gets unlocked. We can also get a partial read
3325 * here, and if we do, then just retry at the new offset.
3327 ret = io_iter_do_read(req, iter);
3328 if (ret == -EIOCBQUEUED)
3330 /* we got some bytes, but not all. retry. */
3331 kiocb->ki_flags &= ~IOCB_WAITQ;
3332 } while (ret > 0 && ret < io_size);
3334 kiocb_done(kiocb, ret, issue_flags);
3336 /* it's faster to check here then delegate to kfree */
3342 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3344 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3346 return io_prep_rw(req, sqe);
3349 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3351 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3352 struct kiocb *kiocb = &req->rw.kiocb;
3353 struct iov_iter __iter, *iter = &__iter;
3354 struct io_async_rw *rw = req->async_data;
3355 ssize_t ret, ret2, io_size;
3356 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3362 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3366 io_size = iov_iter_count(iter);
3367 req->result = io_size;
3369 /* Ensure we clear previously set non-block flag */
3370 if (!force_nonblock)
3371 kiocb->ki_flags &= ~IOCB_NOWAIT;
3373 kiocb->ki_flags |= IOCB_NOWAIT;
3375 /* If the file doesn't support async, just async punt */
3376 if (force_nonblock && !io_file_supports_async(req, WRITE))
3379 /* file path doesn't support NOWAIT for non-direct_IO */
3380 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3381 (req->flags & REQ_F_ISREG))
3384 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3389 * Open-code file_start_write here to grab freeze protection,
3390 * which will be released by another thread in
3391 * io_complete_rw(). Fool lockdep by telling it the lock got
3392 * released so that it doesn't complain about the held lock when
3393 * we return to userspace.
3395 if (req->flags & REQ_F_ISREG) {
3396 sb_start_write(file_inode(req->file)->i_sb);
3397 __sb_writers_release(file_inode(req->file)->i_sb,
3400 kiocb->ki_flags |= IOCB_WRITE;
3402 if (req->file->f_op->write_iter)
3403 ret2 = call_write_iter(req->file, kiocb, iter);
3404 else if (req->file->f_op->write)
3405 ret2 = loop_rw_iter(WRITE, req, iter);
3409 if (req->flags & REQ_F_REISSUE) {
3410 req->flags &= ~REQ_F_REISSUE;
3415 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3416 * retry them without IOCB_NOWAIT.
3418 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3420 /* no retry on NONBLOCK nor RWF_NOWAIT */
3421 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3423 if (!force_nonblock || ret2 != -EAGAIN) {
3424 /* IOPOLL retry should happen for io-wq threads */
3425 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3428 kiocb_done(kiocb, ret2, issue_flags);
3431 /* some cases will consume bytes even on error returns */
3432 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3433 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3434 return ret ?: -EAGAIN;
3437 /* it's reportedly faster than delegating the null check to kfree() */
3443 static int io_renameat_prep(struct io_kiocb *req,
3444 const struct io_uring_sqe *sqe)
3446 struct io_rename *ren = &req->rename;
3447 const char __user *oldf, *newf;
3449 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3452 ren->old_dfd = READ_ONCE(sqe->fd);
3453 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3454 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3455 ren->new_dfd = READ_ONCE(sqe->len);
3456 ren->flags = READ_ONCE(sqe->rename_flags);
3458 ren->oldpath = getname(oldf);
3459 if (IS_ERR(ren->oldpath))
3460 return PTR_ERR(ren->oldpath);
3462 ren->newpath = getname(newf);
3463 if (IS_ERR(ren->newpath)) {
3464 putname(ren->oldpath);
3465 return PTR_ERR(ren->newpath);
3468 req->flags |= REQ_F_NEED_CLEANUP;
3472 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3474 struct io_rename *ren = &req->rename;
3477 if (issue_flags & IO_URING_F_NONBLOCK)
3480 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3481 ren->newpath, ren->flags);
3483 req->flags &= ~REQ_F_NEED_CLEANUP;
3485 req_set_fail_links(req);
3486 io_req_complete(req, ret);
3490 static int io_unlinkat_prep(struct io_kiocb *req,
3491 const struct io_uring_sqe *sqe)
3493 struct io_unlink *un = &req->unlink;
3494 const char __user *fname;
3496 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3499 un->dfd = READ_ONCE(sqe->fd);
3501 un->flags = READ_ONCE(sqe->unlink_flags);
3502 if (un->flags & ~AT_REMOVEDIR)
3505 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3506 un->filename = getname(fname);
3507 if (IS_ERR(un->filename))
3508 return PTR_ERR(un->filename);
3510 req->flags |= REQ_F_NEED_CLEANUP;
3514 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3516 struct io_unlink *un = &req->unlink;
3519 if (issue_flags & IO_URING_F_NONBLOCK)
3522 if (un->flags & AT_REMOVEDIR)
3523 ret = do_rmdir(un->dfd, un->filename);
3525 ret = do_unlinkat(un->dfd, un->filename);
3527 req->flags &= ~REQ_F_NEED_CLEANUP;
3529 req_set_fail_links(req);
3530 io_req_complete(req, ret);
3534 static int io_shutdown_prep(struct io_kiocb *req,
3535 const struct io_uring_sqe *sqe)
3537 #if defined(CONFIG_NET)
3538 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3540 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3544 req->shutdown.how = READ_ONCE(sqe->len);
3551 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3553 #if defined(CONFIG_NET)
3554 struct socket *sock;
3557 if (issue_flags & IO_URING_F_NONBLOCK)
3560 sock = sock_from_file(req->file);
3561 if (unlikely(!sock))
3564 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3566 req_set_fail_links(req);
3567 io_req_complete(req, ret);
3574 static int __io_splice_prep(struct io_kiocb *req,
3575 const struct io_uring_sqe *sqe)
3577 struct io_splice* sp = &req->splice;
3578 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3580 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3584 sp->len = READ_ONCE(sqe->len);
3585 sp->flags = READ_ONCE(sqe->splice_flags);
3587 if (unlikely(sp->flags & ~valid_flags))
3590 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3591 (sp->flags & SPLICE_F_FD_IN_FIXED));
3594 req->flags |= REQ_F_NEED_CLEANUP;
3598 static int io_tee_prep(struct io_kiocb *req,
3599 const struct io_uring_sqe *sqe)
3601 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3603 return __io_splice_prep(req, sqe);
3606 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3608 struct io_splice *sp = &req->splice;
3609 struct file *in = sp->file_in;
3610 struct file *out = sp->file_out;
3611 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3614 if (issue_flags & IO_URING_F_NONBLOCK)
3617 ret = do_tee(in, out, sp->len, flags);
3619 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3621 req->flags &= ~REQ_F_NEED_CLEANUP;
3624 req_set_fail_links(req);
3625 io_req_complete(req, ret);
3629 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3631 struct io_splice* sp = &req->splice;
3633 sp->off_in = READ_ONCE(sqe->splice_off_in);
3634 sp->off_out = READ_ONCE(sqe->off);
3635 return __io_splice_prep(req, sqe);
3638 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3640 struct io_splice *sp = &req->splice;
3641 struct file *in = sp->file_in;
3642 struct file *out = sp->file_out;
3643 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3644 loff_t *poff_in, *poff_out;
3647 if (issue_flags & IO_URING_F_NONBLOCK)
3650 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3651 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3654 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3656 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3658 req->flags &= ~REQ_F_NEED_CLEANUP;
3661 req_set_fail_links(req);
3662 io_req_complete(req, ret);
3667 * IORING_OP_NOP just posts a completion event, nothing else.
3669 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3671 struct io_ring_ctx *ctx = req->ctx;
3673 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3676 __io_req_complete(req, issue_flags, 0, 0);
3680 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3682 struct io_ring_ctx *ctx = req->ctx;
3687 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3689 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3692 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3693 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3696 req->sync.off = READ_ONCE(sqe->off);
3697 req->sync.len = READ_ONCE(sqe->len);
3701 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3703 loff_t end = req->sync.off + req->sync.len;
3706 /* fsync always requires a blocking context */
3707 if (issue_flags & IO_URING_F_NONBLOCK)
3710 ret = vfs_fsync_range(req->file, req->sync.off,
3711 end > 0 ? end : LLONG_MAX,
3712 req->sync.flags & IORING_FSYNC_DATASYNC);
3714 req_set_fail_links(req);
3715 io_req_complete(req, ret);
3719 static int io_fallocate_prep(struct io_kiocb *req,
3720 const struct io_uring_sqe *sqe)
3722 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3724 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3727 req->sync.off = READ_ONCE(sqe->off);
3728 req->sync.len = READ_ONCE(sqe->addr);
3729 req->sync.mode = READ_ONCE(sqe->len);
3733 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3737 /* fallocate always requiring blocking context */
3738 if (issue_flags & IO_URING_F_NONBLOCK)
3740 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3743 req_set_fail_links(req);
3744 io_req_complete(req, ret);
3748 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3750 const char __user *fname;
3753 if (unlikely(sqe->ioprio || sqe->buf_index))
3755 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3758 /* open.how should be already initialised */
3759 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3760 req->open.how.flags |= O_LARGEFILE;
3762 req->open.dfd = READ_ONCE(sqe->fd);
3763 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3764 req->open.filename = getname(fname);
3765 if (IS_ERR(req->open.filename)) {
3766 ret = PTR_ERR(req->open.filename);
3767 req->open.filename = NULL;
3770 req->open.nofile = rlimit(RLIMIT_NOFILE);
3771 req->flags |= REQ_F_NEED_CLEANUP;
3775 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3779 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3781 mode = READ_ONCE(sqe->len);
3782 flags = READ_ONCE(sqe->open_flags);
3783 req->open.how = build_open_how(flags, mode);
3784 return __io_openat_prep(req, sqe);
3787 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3789 struct open_how __user *how;
3793 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3795 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3796 len = READ_ONCE(sqe->len);
3797 if (len < OPEN_HOW_SIZE_VER0)
3800 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3805 return __io_openat_prep(req, sqe);
3808 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3810 struct open_flags op;
3813 bool resolve_nonblock;
3816 ret = build_open_flags(&req->open.how, &op);
3819 nonblock_set = op.open_flag & O_NONBLOCK;
3820 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3821 if (issue_flags & IO_URING_F_NONBLOCK) {
3823 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3824 * it'll always -EAGAIN
3826 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3828 op.lookup_flags |= LOOKUP_CACHED;
3829 op.open_flag |= O_NONBLOCK;
3832 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3836 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3837 /* only retry if RESOLVE_CACHED wasn't already set by application */
3838 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3839 file == ERR_PTR(-EAGAIN)) {
3841 * We could hang on to this 'fd', but seems like marginal
3842 * gain for something that is now known to be a slower path.
3843 * So just put it, and we'll get a new one when we retry.
3851 ret = PTR_ERR(file);
3853 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3854 file->f_flags &= ~O_NONBLOCK;
3855 fsnotify_open(file);
3856 fd_install(ret, file);
3859 putname(req->open.filename);
3860 req->flags &= ~REQ_F_NEED_CLEANUP;
3862 req_set_fail_links(req);
3863 __io_req_complete(req, issue_flags, ret, 0);
3867 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3869 return io_openat2(req, issue_flags);
3872 static int io_remove_buffers_prep(struct io_kiocb *req,
3873 const struct io_uring_sqe *sqe)
3875 struct io_provide_buf *p = &req->pbuf;
3878 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3881 tmp = READ_ONCE(sqe->fd);
3882 if (!tmp || tmp > USHRT_MAX)
3885 memset(p, 0, sizeof(*p));
3887 p->bgid = READ_ONCE(sqe->buf_group);
3891 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3892 int bgid, unsigned nbufs)
3896 /* shouldn't happen */
3900 /* the head kbuf is the list itself */
3901 while (!list_empty(&buf->list)) {
3902 struct io_buffer *nxt;
3904 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3905 list_del(&nxt->list);
3912 xa_erase(&ctx->io_buffers, bgid);
3917 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3919 struct io_provide_buf *p = &req->pbuf;
3920 struct io_ring_ctx *ctx = req->ctx;
3921 struct io_buffer *head;
3923 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3925 io_ring_submit_lock(ctx, !force_nonblock);
3927 lockdep_assert_held(&ctx->uring_lock);
3930 head = xa_load(&ctx->io_buffers, p->bgid);
3932 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3934 req_set_fail_links(req);
3936 /* complete before unlock, IOPOLL may need the lock */
3937 __io_req_complete(req, issue_flags, ret, 0);
3938 io_ring_submit_unlock(ctx, !force_nonblock);
3942 static int io_provide_buffers_prep(struct io_kiocb *req,
3943 const struct io_uring_sqe *sqe)
3945 unsigned long size, tmp_check;
3946 struct io_provide_buf *p = &req->pbuf;
3949 if (sqe->ioprio || sqe->rw_flags)
3952 tmp = READ_ONCE(sqe->fd);
3953 if (!tmp || tmp > USHRT_MAX)
3956 p->addr = READ_ONCE(sqe->addr);
3957 p->len = READ_ONCE(sqe->len);
3959 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3962 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3965 size = (unsigned long)p->len * p->nbufs;
3966 if (!access_ok(u64_to_user_ptr(p->addr), size))
3969 p->bgid = READ_ONCE(sqe->buf_group);
3970 tmp = READ_ONCE(sqe->off);
3971 if (tmp > USHRT_MAX)
3977 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3979 struct io_buffer *buf;
3980 u64 addr = pbuf->addr;
3981 int i, bid = pbuf->bid;
3983 for (i = 0; i < pbuf->nbufs; i++) {
3984 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3989 buf->len = pbuf->len;
3994 INIT_LIST_HEAD(&buf->list);
3997 list_add_tail(&buf->list, &(*head)->list);
4001 return i ? i : -ENOMEM;
4004 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4006 struct io_provide_buf *p = &req->pbuf;
4007 struct io_ring_ctx *ctx = req->ctx;
4008 struct io_buffer *head, *list;
4010 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4012 io_ring_submit_lock(ctx, !force_nonblock);
4014 lockdep_assert_held(&ctx->uring_lock);
4016 list = head = xa_load(&ctx->io_buffers, p->bgid);
4018 ret = io_add_buffers(p, &head);
4019 if (ret >= 0 && !list) {
4020 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4022 __io_remove_buffers(ctx, head, p->bgid, -1U);
4025 req_set_fail_links(req);
4026 /* complete before unlock, IOPOLL may need the lock */
4027 __io_req_complete(req, issue_flags, ret, 0);
4028 io_ring_submit_unlock(ctx, !force_nonblock);
4032 static int io_epoll_ctl_prep(struct io_kiocb *req,
4033 const struct io_uring_sqe *sqe)
4035 #if defined(CONFIG_EPOLL)
4036 if (sqe->ioprio || sqe->buf_index)
4038 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4041 req->epoll.epfd = READ_ONCE(sqe->fd);
4042 req->epoll.op = READ_ONCE(sqe->len);
4043 req->epoll.fd = READ_ONCE(sqe->off);
4045 if (ep_op_has_event(req->epoll.op)) {
4046 struct epoll_event __user *ev;
4048 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4049 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4059 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4061 #if defined(CONFIG_EPOLL)
4062 struct io_epoll *ie = &req->epoll;
4064 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4066 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4067 if (force_nonblock && ret == -EAGAIN)
4071 req_set_fail_links(req);
4072 __io_req_complete(req, issue_flags, ret, 0);
4079 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4081 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4082 if (sqe->ioprio || sqe->buf_index || sqe->off)
4084 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4087 req->madvise.addr = READ_ONCE(sqe->addr);
4088 req->madvise.len = READ_ONCE(sqe->len);
4089 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4096 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4098 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4099 struct io_madvise *ma = &req->madvise;
4102 if (issue_flags & IO_URING_F_NONBLOCK)
4105 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4107 req_set_fail_links(req);
4108 io_req_complete(req, ret);
4115 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4117 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4119 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4122 req->fadvise.offset = READ_ONCE(sqe->off);
4123 req->fadvise.len = READ_ONCE(sqe->len);
4124 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4128 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4130 struct io_fadvise *fa = &req->fadvise;
4133 if (issue_flags & IO_URING_F_NONBLOCK) {
4134 switch (fa->advice) {
4135 case POSIX_FADV_NORMAL:
4136 case POSIX_FADV_RANDOM:
4137 case POSIX_FADV_SEQUENTIAL:
4144 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4146 req_set_fail_links(req);
4147 __io_req_complete(req, issue_flags, ret, 0);
4151 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4153 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4155 if (sqe->ioprio || sqe->buf_index)
4157 if (req->flags & REQ_F_FIXED_FILE)
4160 req->statx.dfd = READ_ONCE(sqe->fd);
4161 req->statx.mask = READ_ONCE(sqe->len);
4162 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4163 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4164 req->statx.flags = READ_ONCE(sqe->statx_flags);
4169 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4171 struct io_statx *ctx = &req->statx;
4174 if (issue_flags & IO_URING_F_NONBLOCK)
4177 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4181 req_set_fail_links(req);
4182 io_req_complete(req, ret);
4186 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4188 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4190 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4191 sqe->rw_flags || sqe->buf_index)
4193 if (req->flags & REQ_F_FIXED_FILE)
4196 req->close.fd = READ_ONCE(sqe->fd);
4200 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4202 struct files_struct *files = current->files;
4203 struct io_close *close = &req->close;
4204 struct fdtable *fdt;
4205 struct file *file = NULL;
4208 spin_lock(&files->file_lock);
4209 fdt = files_fdtable(files);
4210 if (close->fd >= fdt->max_fds) {
4211 spin_unlock(&files->file_lock);
4214 file = fdt->fd[close->fd];
4215 if (!file || file->f_op == &io_uring_fops) {
4216 spin_unlock(&files->file_lock);
4221 /* if the file has a flush method, be safe and punt to async */
4222 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4223 spin_unlock(&files->file_lock);
4227 ret = __close_fd_get_file(close->fd, &file);
4228 spin_unlock(&files->file_lock);
4235 /* No ->flush() or already async, safely close from here */
4236 ret = filp_close(file, current->files);
4239 req_set_fail_links(req);
4242 __io_req_complete(req, issue_flags, ret, 0);
4246 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4248 struct io_ring_ctx *ctx = req->ctx;
4250 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4252 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4255 req->sync.off = READ_ONCE(sqe->off);
4256 req->sync.len = READ_ONCE(sqe->len);
4257 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4261 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4265 /* sync_file_range always requires a blocking context */
4266 if (issue_flags & IO_URING_F_NONBLOCK)
4269 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4272 req_set_fail_links(req);
4273 io_req_complete(req, ret);
4277 #if defined(CONFIG_NET)
4278 static int io_setup_async_msg(struct io_kiocb *req,
4279 struct io_async_msghdr *kmsg)
4281 struct io_async_msghdr *async_msg = req->async_data;
4285 if (io_alloc_async_data(req)) {
4286 kfree(kmsg->free_iov);
4289 async_msg = req->async_data;
4290 req->flags |= REQ_F_NEED_CLEANUP;
4291 memcpy(async_msg, kmsg, sizeof(*kmsg));
4292 async_msg->msg.msg_name = &async_msg->addr;
4293 /* if were using fast_iov, set it to the new one */
4294 if (!async_msg->free_iov)
4295 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4300 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4301 struct io_async_msghdr *iomsg)
4303 iomsg->msg.msg_name = &iomsg->addr;
4304 iomsg->free_iov = iomsg->fast_iov;
4305 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4306 req->sr_msg.msg_flags, &iomsg->free_iov);
4309 static int io_sendmsg_prep_async(struct io_kiocb *req)
4313 ret = io_sendmsg_copy_hdr(req, req->async_data);
4315 req->flags |= REQ_F_NEED_CLEANUP;
4319 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4321 struct io_sr_msg *sr = &req->sr_msg;
4323 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4326 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4327 sr->len = READ_ONCE(sqe->len);
4328 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4329 if (sr->msg_flags & MSG_DONTWAIT)
4330 req->flags |= REQ_F_NOWAIT;
4332 #ifdef CONFIG_COMPAT
4333 if (req->ctx->compat)
4334 sr->msg_flags |= MSG_CMSG_COMPAT;
4339 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4341 struct io_async_msghdr iomsg, *kmsg;
4342 struct socket *sock;
4347 sock = sock_from_file(req->file);
4348 if (unlikely(!sock))
4351 kmsg = req->async_data;
4353 ret = io_sendmsg_copy_hdr(req, &iomsg);
4359 flags = req->sr_msg.msg_flags;
4360 if (issue_flags & IO_URING_F_NONBLOCK)
4361 flags |= MSG_DONTWAIT;
4362 if (flags & MSG_WAITALL)
4363 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4365 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4366 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4367 return io_setup_async_msg(req, kmsg);
4368 if (ret == -ERESTARTSYS)
4371 /* fast path, check for non-NULL to avoid function call */
4373 kfree(kmsg->free_iov);
4374 req->flags &= ~REQ_F_NEED_CLEANUP;
4376 req_set_fail_links(req);
4377 __io_req_complete(req, issue_flags, ret, 0);
4381 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4383 struct io_sr_msg *sr = &req->sr_msg;
4386 struct socket *sock;
4391 sock = sock_from_file(req->file);
4392 if (unlikely(!sock))
4395 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4399 msg.msg_name = NULL;
4400 msg.msg_control = NULL;
4401 msg.msg_controllen = 0;
4402 msg.msg_namelen = 0;
4404 flags = req->sr_msg.msg_flags;
4405 if (issue_flags & IO_URING_F_NONBLOCK)
4406 flags |= MSG_DONTWAIT;
4407 if (flags & MSG_WAITALL)
4408 min_ret = iov_iter_count(&msg.msg_iter);
4410 msg.msg_flags = flags;
4411 ret = sock_sendmsg(sock, &msg);
4412 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4414 if (ret == -ERESTARTSYS)
4418 req_set_fail_links(req);
4419 __io_req_complete(req, issue_flags, ret, 0);
4423 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4424 struct io_async_msghdr *iomsg)
4426 struct io_sr_msg *sr = &req->sr_msg;
4427 struct iovec __user *uiov;
4431 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4432 &iomsg->uaddr, &uiov, &iov_len);
4436 if (req->flags & REQ_F_BUFFER_SELECT) {
4439 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4441 sr->len = iomsg->fast_iov[0].iov_len;
4442 iomsg->free_iov = NULL;
4444 iomsg->free_iov = iomsg->fast_iov;
4445 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4446 &iomsg->free_iov, &iomsg->msg.msg_iter,
4455 #ifdef CONFIG_COMPAT
4456 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4457 struct io_async_msghdr *iomsg)
4459 struct io_sr_msg *sr = &req->sr_msg;
4460 struct compat_iovec __user *uiov;
4465 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4470 uiov = compat_ptr(ptr);
4471 if (req->flags & REQ_F_BUFFER_SELECT) {
4472 compat_ssize_t clen;
4476 if (!access_ok(uiov, sizeof(*uiov)))
4478 if (__get_user(clen, &uiov->iov_len))
4483 iomsg->free_iov = NULL;
4485 iomsg->free_iov = iomsg->fast_iov;
4486 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4487 UIO_FASTIOV, &iomsg->free_iov,
4488 &iomsg->msg.msg_iter, true);
4497 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4498 struct io_async_msghdr *iomsg)
4500 iomsg->msg.msg_name = &iomsg->addr;
4502 #ifdef CONFIG_COMPAT
4503 if (req->ctx->compat)
4504 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4507 return __io_recvmsg_copy_hdr(req, iomsg);
4510 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4513 struct io_sr_msg *sr = &req->sr_msg;
4514 struct io_buffer *kbuf;
4516 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4521 req->flags |= REQ_F_BUFFER_SELECTED;
4525 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4527 return io_put_kbuf(req, req->sr_msg.kbuf);
4530 static int io_recvmsg_prep_async(struct io_kiocb *req)
4534 ret = io_recvmsg_copy_hdr(req, req->async_data);
4536 req->flags |= REQ_F_NEED_CLEANUP;
4540 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4542 struct io_sr_msg *sr = &req->sr_msg;
4544 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4547 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4548 sr->len = READ_ONCE(sqe->len);
4549 sr->bgid = READ_ONCE(sqe->buf_group);
4550 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4551 if (sr->msg_flags & MSG_DONTWAIT)
4552 req->flags |= REQ_F_NOWAIT;
4554 #ifdef CONFIG_COMPAT
4555 if (req->ctx->compat)
4556 sr->msg_flags |= MSG_CMSG_COMPAT;
4561 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4563 struct io_async_msghdr iomsg, *kmsg;
4564 struct socket *sock;
4565 struct io_buffer *kbuf;
4568 int ret, cflags = 0;
4569 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4571 sock = sock_from_file(req->file);
4572 if (unlikely(!sock))
4575 kmsg = req->async_data;
4577 ret = io_recvmsg_copy_hdr(req, &iomsg);
4583 if (req->flags & REQ_F_BUFFER_SELECT) {
4584 kbuf = io_recv_buffer_select(req, !force_nonblock);
4586 return PTR_ERR(kbuf);
4587 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4588 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4589 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4590 1, req->sr_msg.len);
4593 flags = req->sr_msg.msg_flags;
4595 flags |= MSG_DONTWAIT;
4596 if (flags & MSG_WAITALL)
4597 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4599 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4600 kmsg->uaddr, flags);
4601 if (force_nonblock && ret == -EAGAIN)
4602 return io_setup_async_msg(req, kmsg);
4603 if (ret == -ERESTARTSYS)
4606 if (req->flags & REQ_F_BUFFER_SELECTED)
4607 cflags = io_put_recv_kbuf(req);
4608 /* fast path, check for non-NULL to avoid function call */
4610 kfree(kmsg->free_iov);
4611 req->flags &= ~REQ_F_NEED_CLEANUP;
4612 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4613 req_set_fail_links(req);
4614 __io_req_complete(req, issue_flags, ret, cflags);
4618 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4620 struct io_buffer *kbuf;
4621 struct io_sr_msg *sr = &req->sr_msg;
4623 void __user *buf = sr->buf;
4624 struct socket *sock;
4628 int ret, cflags = 0;
4629 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4631 sock = sock_from_file(req->file);
4632 if (unlikely(!sock))
4635 if (req->flags & REQ_F_BUFFER_SELECT) {
4636 kbuf = io_recv_buffer_select(req, !force_nonblock);
4638 return PTR_ERR(kbuf);
4639 buf = u64_to_user_ptr(kbuf->addr);
4642 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4646 msg.msg_name = NULL;
4647 msg.msg_control = NULL;
4648 msg.msg_controllen = 0;
4649 msg.msg_namelen = 0;
4650 msg.msg_iocb = NULL;
4653 flags = req->sr_msg.msg_flags;
4655 flags |= MSG_DONTWAIT;
4656 if (flags & MSG_WAITALL)
4657 min_ret = iov_iter_count(&msg.msg_iter);
4659 ret = sock_recvmsg(sock, &msg, flags);
4660 if (force_nonblock && ret == -EAGAIN)
4662 if (ret == -ERESTARTSYS)
4665 if (req->flags & REQ_F_BUFFER_SELECTED)
4666 cflags = io_put_recv_kbuf(req);
4667 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4668 req_set_fail_links(req);
4669 __io_req_complete(req, issue_flags, ret, cflags);
4673 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4675 struct io_accept *accept = &req->accept;
4677 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4679 if (sqe->ioprio || sqe->len || sqe->buf_index)
4682 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4683 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4684 accept->flags = READ_ONCE(sqe->accept_flags);
4685 accept->nofile = rlimit(RLIMIT_NOFILE);
4689 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4691 struct io_accept *accept = &req->accept;
4692 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4693 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4696 if (req->file->f_flags & O_NONBLOCK)
4697 req->flags |= REQ_F_NOWAIT;
4699 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4700 accept->addr_len, accept->flags,
4702 if (ret == -EAGAIN && force_nonblock)
4705 if (ret == -ERESTARTSYS)
4707 req_set_fail_links(req);
4709 __io_req_complete(req, issue_flags, ret, 0);
4713 static int io_connect_prep_async(struct io_kiocb *req)
4715 struct io_async_connect *io = req->async_data;
4716 struct io_connect *conn = &req->connect;
4718 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4721 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4723 struct io_connect *conn = &req->connect;
4725 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4727 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4730 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4731 conn->addr_len = READ_ONCE(sqe->addr2);
4735 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4737 struct io_async_connect __io, *io;
4738 unsigned file_flags;
4740 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4742 if (req->async_data) {
4743 io = req->async_data;
4745 ret = move_addr_to_kernel(req->connect.addr,
4746 req->connect.addr_len,
4753 file_flags = force_nonblock ? O_NONBLOCK : 0;
4755 ret = __sys_connect_file(req->file, &io->address,
4756 req->connect.addr_len, file_flags);
4757 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4758 if (req->async_data)
4760 if (io_alloc_async_data(req)) {
4764 memcpy(req->async_data, &__io, sizeof(__io));
4767 if (ret == -ERESTARTSYS)
4771 req_set_fail_links(req);
4772 __io_req_complete(req, issue_flags, ret, 0);
4775 #else /* !CONFIG_NET */
4776 #define IO_NETOP_FN(op) \
4777 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4779 return -EOPNOTSUPP; \
4782 #define IO_NETOP_PREP(op) \
4784 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4786 return -EOPNOTSUPP; \
4789 #define IO_NETOP_PREP_ASYNC(op) \
4791 static int io_##op##_prep_async(struct io_kiocb *req) \
4793 return -EOPNOTSUPP; \
4796 IO_NETOP_PREP_ASYNC(sendmsg);
4797 IO_NETOP_PREP_ASYNC(recvmsg);
4798 IO_NETOP_PREP_ASYNC(connect);
4799 IO_NETOP_PREP(accept);
4802 #endif /* CONFIG_NET */
4804 struct io_poll_table {
4805 struct poll_table_struct pt;
4806 struct io_kiocb *req;
4810 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4811 __poll_t mask, task_work_func_t func)
4815 /* for instances that support it check for an event match first: */
4816 if (mask && !(mask & poll->events))
4819 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4821 list_del_init(&poll->wait.entry);
4824 req->task_work.func = func;
4827 * If this fails, then the task is exiting. When a task exits, the
4828 * work gets canceled, so just cancel this request as well instead
4829 * of executing it. We can't safely execute it anyway, as we may not
4830 * have the needed state needed for it anyway.
4832 ret = io_req_task_work_add(req);
4833 if (unlikely(ret)) {
4834 WRITE_ONCE(poll->canceled, true);
4835 io_req_task_work_add_fallback(req, func);
4840 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4841 __acquires(&req->ctx->completion_lock)
4843 struct io_ring_ctx *ctx = req->ctx;
4845 if (!req->result && !READ_ONCE(poll->canceled)) {
4846 struct poll_table_struct pt = { ._key = poll->events };
4848 req->result = vfs_poll(req->file, &pt) & poll->events;
4851 spin_lock_irq(&ctx->completion_lock);
4852 if (!req->result && !READ_ONCE(poll->canceled)) {
4853 add_wait_queue(poll->head, &poll->wait);
4860 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4862 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4863 if (req->opcode == IORING_OP_POLL_ADD)
4864 return req->async_data;
4865 return req->apoll->double_poll;
4868 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4870 if (req->opcode == IORING_OP_POLL_ADD)
4872 return &req->apoll->poll;
4875 static void io_poll_remove_double(struct io_kiocb *req)
4876 __must_hold(&req->ctx->completion_lock)
4878 struct io_poll_iocb *poll = io_poll_get_double(req);
4880 lockdep_assert_held(&req->ctx->completion_lock);
4882 if (poll && poll->head) {
4883 struct wait_queue_head *head = poll->head;
4885 spin_lock(&head->lock);
4886 list_del_init(&poll->wait.entry);
4887 if (poll->wait.private)
4890 spin_unlock(&head->lock);
4894 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4895 __must_hold(&req->ctx->completion_lock)
4897 struct io_ring_ctx *ctx = req->ctx;
4898 unsigned flags = IORING_CQE_F_MORE;
4901 if (READ_ONCE(req->poll.canceled)) {
4903 req->poll.events |= EPOLLONESHOT;
4905 error = mangle_poll(mask);
4907 if (req->poll.events & EPOLLONESHOT)
4909 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4910 io_poll_remove_waitqs(req);
4911 req->poll.done = true;
4914 if (flags & IORING_CQE_F_MORE)
4917 io_commit_cqring(ctx);
4918 return !(flags & IORING_CQE_F_MORE);
4921 static void io_poll_task_func(struct callback_head *cb)
4923 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4924 struct io_ring_ctx *ctx = req->ctx;
4925 struct io_kiocb *nxt;
4927 if (io_poll_rewait(req, &req->poll)) {
4928 spin_unlock_irq(&ctx->completion_lock);
4932 done = io_poll_complete(req, req->result);
4934 hash_del(&req->hash_node);
4937 add_wait_queue(req->poll.head, &req->poll.wait);
4939 spin_unlock_irq(&ctx->completion_lock);
4940 io_cqring_ev_posted(ctx);
4943 nxt = io_put_req_find_next(req);
4945 __io_req_task_submit(nxt);
4950 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4951 int sync, void *key)
4953 struct io_kiocb *req = wait->private;
4954 struct io_poll_iocb *poll = io_poll_get_single(req);
4955 __poll_t mask = key_to_poll(key);
4957 /* for instances that support it check for an event match first: */
4958 if (mask && !(mask & poll->events))
4960 if (!(poll->events & EPOLLONESHOT))
4961 return poll->wait.func(&poll->wait, mode, sync, key);
4963 list_del_init(&wait->entry);
4965 if (poll && poll->head) {
4968 spin_lock(&poll->head->lock);
4969 done = list_empty(&poll->wait.entry);
4971 list_del_init(&poll->wait.entry);
4972 /* make sure double remove sees this as being gone */
4973 wait->private = NULL;
4974 spin_unlock(&poll->head->lock);
4976 /* use wait func handler, so it matches the rq type */
4977 poll->wait.func(&poll->wait, mode, sync, key);
4984 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4985 wait_queue_func_t wake_func)
4989 poll->canceled = false;
4990 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
4991 /* mask in events that we always want/need */
4992 poll->events = events | IO_POLL_UNMASK;
4993 INIT_LIST_HEAD(&poll->wait.entry);
4994 init_waitqueue_func_entry(&poll->wait, wake_func);
4997 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4998 struct wait_queue_head *head,
4999 struct io_poll_iocb **poll_ptr)
5001 struct io_kiocb *req = pt->req;
5004 * If poll->head is already set, it's because the file being polled
5005 * uses multiple waitqueues for poll handling (eg one for read, one
5006 * for write). Setup a separate io_poll_iocb if this happens.
5008 if (unlikely(poll->head)) {
5009 struct io_poll_iocb *poll_one = poll;
5011 /* already have a 2nd entry, fail a third attempt */
5013 pt->error = -EINVAL;
5017 * Can't handle multishot for double wait for now, turn it
5018 * into one-shot mode.
5020 if (!(req->poll.events & EPOLLONESHOT))
5021 req->poll.events |= EPOLLONESHOT;
5022 /* double add on the same waitqueue head, ignore */
5023 if (poll->head == head)
5025 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5027 pt->error = -ENOMEM;
5030 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5032 poll->wait.private = req;
5039 if (poll->events & EPOLLEXCLUSIVE)
5040 add_wait_queue_exclusive(head, &poll->wait);
5042 add_wait_queue(head, &poll->wait);
5045 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5046 struct poll_table_struct *p)
5048 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5049 struct async_poll *apoll = pt->req->apoll;
5051 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5054 static void io_async_task_func(struct callback_head *cb)
5056 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5057 struct async_poll *apoll = req->apoll;
5058 struct io_ring_ctx *ctx = req->ctx;
5060 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5062 if (io_poll_rewait(req, &apoll->poll)) {
5063 spin_unlock_irq(&ctx->completion_lock);
5067 hash_del(&req->hash_node);
5068 io_poll_remove_double(req);
5069 spin_unlock_irq(&ctx->completion_lock);
5071 if (!READ_ONCE(apoll->poll.canceled))
5072 __io_req_task_submit(req);
5074 io_req_complete_failed(req, -ECANCELED);
5077 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5080 struct io_kiocb *req = wait->private;
5081 struct io_poll_iocb *poll = &req->apoll->poll;
5083 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5086 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5089 static void io_poll_req_insert(struct io_kiocb *req)
5091 struct io_ring_ctx *ctx = req->ctx;
5092 struct hlist_head *list;
5094 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5095 hlist_add_head(&req->hash_node, list);
5098 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5099 struct io_poll_iocb *poll,
5100 struct io_poll_table *ipt, __poll_t mask,
5101 wait_queue_func_t wake_func)
5102 __acquires(&ctx->completion_lock)
5104 struct io_ring_ctx *ctx = req->ctx;
5105 bool cancel = false;
5107 INIT_HLIST_NODE(&req->hash_node);
5108 io_init_poll_iocb(poll, mask, wake_func);
5109 poll->file = req->file;
5110 poll->wait.private = req;
5112 ipt->pt._key = mask;
5114 ipt->error = -EINVAL;
5116 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5118 spin_lock_irq(&ctx->completion_lock);
5119 if (likely(poll->head)) {
5120 spin_lock(&poll->head->lock);
5121 if (unlikely(list_empty(&poll->wait.entry))) {
5127 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5128 list_del_init(&poll->wait.entry);
5130 WRITE_ONCE(poll->canceled, true);
5131 else if (!poll->done) /* actually waiting for an event */
5132 io_poll_req_insert(req);
5133 spin_unlock(&poll->head->lock);
5139 static bool io_arm_poll_handler(struct io_kiocb *req)
5141 const struct io_op_def *def = &io_op_defs[req->opcode];
5142 struct io_ring_ctx *ctx = req->ctx;
5143 struct async_poll *apoll;
5144 struct io_poll_table ipt;
5148 if (!req->file || !file_can_poll(req->file))
5150 if (req->flags & REQ_F_POLLED)
5154 else if (def->pollout)
5158 /* if we can't nonblock try, then no point in arming a poll handler */
5159 if (!io_file_supports_async(req, rw))
5162 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5163 if (unlikely(!apoll))
5165 apoll->double_poll = NULL;
5167 req->flags |= REQ_F_POLLED;
5170 mask = EPOLLONESHOT;
5172 mask |= POLLIN | POLLRDNORM;
5174 mask |= POLLOUT | POLLWRNORM;
5176 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5177 if ((req->opcode == IORING_OP_RECVMSG) &&
5178 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5181 mask |= POLLERR | POLLPRI;
5183 ipt.pt._qproc = io_async_queue_proc;
5185 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5187 if (ret || ipt.error) {
5188 io_poll_remove_double(req);
5189 spin_unlock_irq(&ctx->completion_lock);
5192 spin_unlock_irq(&ctx->completion_lock);
5193 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5194 apoll->poll.events);
5198 static bool __io_poll_remove_one(struct io_kiocb *req,
5199 struct io_poll_iocb *poll, bool do_cancel)
5200 __must_hold(&req->ctx->completion_lock)
5202 bool do_complete = false;
5206 spin_lock(&poll->head->lock);
5208 WRITE_ONCE(poll->canceled, true);
5209 if (!list_empty(&poll->wait.entry)) {
5210 list_del_init(&poll->wait.entry);
5213 spin_unlock(&poll->head->lock);
5214 hash_del(&req->hash_node);
5218 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5219 __must_hold(&req->ctx->completion_lock)
5223 io_poll_remove_double(req);
5224 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5226 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5227 /* non-poll requests have submit ref still */
5233 static bool io_poll_remove_one(struct io_kiocb *req)
5234 __must_hold(&req->ctx->completion_lock)
5238 do_complete = io_poll_remove_waitqs(req);
5240 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5241 io_commit_cqring(req->ctx);
5242 req_set_fail_links(req);
5243 io_put_req_deferred(req, 1);
5250 * Returns true if we found and killed one or more poll requests
5252 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5253 struct files_struct *files)
5255 struct hlist_node *tmp;
5256 struct io_kiocb *req;
5259 spin_lock_irq(&ctx->completion_lock);
5260 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5261 struct hlist_head *list;
5263 list = &ctx->cancel_hash[i];
5264 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5265 if (io_match_task(req, tsk, files))
5266 posted += io_poll_remove_one(req);
5269 spin_unlock_irq(&ctx->completion_lock);
5272 io_cqring_ev_posted(ctx);
5277 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5279 __must_hold(&ctx->completion_lock)
5281 struct hlist_head *list;
5282 struct io_kiocb *req;
5284 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5285 hlist_for_each_entry(req, list, hash_node) {
5286 if (sqe_addr != req->user_data)
5288 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5295 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5297 __must_hold(&ctx->completion_lock)
5299 struct io_kiocb *req;
5301 req = io_poll_find(ctx, sqe_addr, poll_only);
5304 if (io_poll_remove_one(req))
5310 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5315 events = READ_ONCE(sqe->poll32_events);
5317 events = swahw32(events);
5319 if (!(flags & IORING_POLL_ADD_MULTI))
5320 events |= EPOLLONESHOT;
5321 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5324 static int io_poll_update_prep(struct io_kiocb *req,
5325 const struct io_uring_sqe *sqe)
5327 struct io_poll_update *upd = &req->poll_update;
5330 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5332 if (sqe->ioprio || sqe->buf_index)
5334 flags = READ_ONCE(sqe->len);
5335 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5336 IORING_POLL_ADD_MULTI))
5338 /* meaningless without update */
5339 if (flags == IORING_POLL_ADD_MULTI)
5342 upd->old_user_data = READ_ONCE(sqe->addr);
5343 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5344 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5346 upd->new_user_data = READ_ONCE(sqe->off);
5347 if (!upd->update_user_data && upd->new_user_data)
5349 if (upd->update_events)
5350 upd->events = io_poll_parse_events(sqe, flags);
5351 else if (sqe->poll32_events)
5357 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5360 struct io_kiocb *req = wait->private;
5361 struct io_poll_iocb *poll = &req->poll;
5363 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5366 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5367 struct poll_table_struct *p)
5369 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5371 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5374 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5376 struct io_poll_iocb *poll = &req->poll;
5379 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5381 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5383 flags = READ_ONCE(sqe->len);
5384 if (flags & ~IORING_POLL_ADD_MULTI)
5387 poll->events = io_poll_parse_events(sqe, flags);
5391 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5393 struct io_poll_iocb *poll = &req->poll;
5394 struct io_ring_ctx *ctx = req->ctx;
5395 struct io_poll_table ipt;
5398 ipt.pt._qproc = io_poll_queue_proc;
5400 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5403 if (mask) { /* no async, we'd stolen it */
5405 io_poll_complete(req, mask);
5407 spin_unlock_irq(&ctx->completion_lock);
5410 io_cqring_ev_posted(ctx);
5411 if (poll->events & EPOLLONESHOT)
5417 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5419 struct io_ring_ctx *ctx = req->ctx;
5420 struct io_kiocb *preq;
5424 spin_lock_irq(&ctx->completion_lock);
5425 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5431 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5433 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5438 * Don't allow racy completion with singleshot, as we cannot safely
5439 * update those. For multishot, if we're racing with completion, just
5440 * let completion re-add it.
5442 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5443 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5447 /* we now have a detached poll request. reissue. */
5451 spin_unlock_irq(&ctx->completion_lock);
5452 req_set_fail_links(req);
5453 io_req_complete(req, ret);
5456 /* only mask one event flags, keep behavior flags */
5457 if (req->poll_update.update_events) {
5458 preq->poll.events &= ~0xffff;
5459 preq->poll.events |= req->poll_update.events & 0xffff;
5460 preq->poll.events |= IO_POLL_UNMASK;
5462 if (req->poll_update.update_user_data)
5463 preq->user_data = req->poll_update.new_user_data;
5464 spin_unlock_irq(&ctx->completion_lock);
5466 /* complete update request, we're done with it */
5467 io_req_complete(req, ret);
5470 ret = io_poll_add(preq, issue_flags);
5472 req_set_fail_links(preq);
5473 io_req_complete(preq, ret);
5479 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5481 struct io_timeout_data *data = container_of(timer,
5482 struct io_timeout_data, timer);
5483 struct io_kiocb *req = data->req;
5484 struct io_ring_ctx *ctx = req->ctx;
5485 unsigned long flags;
5487 spin_lock_irqsave(&ctx->completion_lock, flags);
5488 list_del_init(&req->timeout.list);
5489 atomic_set(&req->ctx->cq_timeouts,
5490 atomic_read(&req->ctx->cq_timeouts) + 1);
5492 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5493 io_commit_cqring(ctx);
5494 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5496 io_cqring_ev_posted(ctx);
5497 req_set_fail_links(req);
5499 return HRTIMER_NORESTART;
5502 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5504 __must_hold(&ctx->completion_lock)
5506 struct io_timeout_data *io;
5507 struct io_kiocb *req;
5510 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5511 found = user_data == req->user_data;
5516 return ERR_PTR(-ENOENT);
5518 io = req->async_data;
5519 if (hrtimer_try_to_cancel(&io->timer) == -1)
5520 return ERR_PTR(-EALREADY);
5521 list_del_init(&req->timeout.list);
5525 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5526 __must_hold(&ctx->completion_lock)
5528 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5531 return PTR_ERR(req);
5533 req_set_fail_links(req);
5534 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5535 io_put_req_deferred(req, 1);
5539 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5540 struct timespec64 *ts, enum hrtimer_mode mode)
5541 __must_hold(&ctx->completion_lock)
5543 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5544 struct io_timeout_data *data;
5547 return PTR_ERR(req);
5549 req->timeout.off = 0; /* noseq */
5550 data = req->async_data;
5551 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5552 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5553 data->timer.function = io_timeout_fn;
5554 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5558 static int io_timeout_remove_prep(struct io_kiocb *req,
5559 const struct io_uring_sqe *sqe)
5561 struct io_timeout_rem *tr = &req->timeout_rem;
5563 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5565 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5567 if (sqe->ioprio || sqe->buf_index || sqe->len)
5570 tr->addr = READ_ONCE(sqe->addr);
5571 tr->flags = READ_ONCE(sqe->timeout_flags);
5572 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5573 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5575 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5577 } else if (tr->flags) {
5578 /* timeout removal doesn't support flags */
5585 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5587 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5592 * Remove or update an existing timeout command
5594 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5596 struct io_timeout_rem *tr = &req->timeout_rem;
5597 struct io_ring_ctx *ctx = req->ctx;
5600 spin_lock_irq(&ctx->completion_lock);
5601 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5602 ret = io_timeout_cancel(ctx, tr->addr);
5604 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5605 io_translate_timeout_mode(tr->flags));
5607 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5608 io_commit_cqring(ctx);
5609 spin_unlock_irq(&ctx->completion_lock);
5610 io_cqring_ev_posted(ctx);
5612 req_set_fail_links(req);
5617 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5618 bool is_timeout_link)
5620 struct io_timeout_data *data;
5622 u32 off = READ_ONCE(sqe->off);
5624 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5626 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5628 if (off && is_timeout_link)
5630 flags = READ_ONCE(sqe->timeout_flags);
5631 if (flags & ~IORING_TIMEOUT_ABS)
5634 req->timeout.off = off;
5636 if (!req->async_data && io_alloc_async_data(req))
5639 data = req->async_data;
5642 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5645 data->mode = io_translate_timeout_mode(flags);
5646 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5647 if (is_timeout_link)
5648 io_req_track_inflight(req);
5652 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5654 struct io_ring_ctx *ctx = req->ctx;
5655 struct io_timeout_data *data = req->async_data;
5656 struct list_head *entry;
5657 u32 tail, off = req->timeout.off;
5659 spin_lock_irq(&ctx->completion_lock);
5662 * sqe->off holds how many events that need to occur for this
5663 * timeout event to be satisfied. If it isn't set, then this is
5664 * a pure timeout request, sequence isn't used.
5666 if (io_is_timeout_noseq(req)) {
5667 entry = ctx->timeout_list.prev;
5671 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5672 req->timeout.target_seq = tail + off;
5674 /* Update the last seq here in case io_flush_timeouts() hasn't.
5675 * This is safe because ->completion_lock is held, and submissions
5676 * and completions are never mixed in the same ->completion_lock section.
5678 ctx->cq_last_tm_flush = tail;
5681 * Insertion sort, ensuring the first entry in the list is always
5682 * the one we need first.
5684 list_for_each_prev(entry, &ctx->timeout_list) {
5685 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5688 if (io_is_timeout_noseq(nxt))
5690 /* nxt.seq is behind @tail, otherwise would've been completed */
5691 if (off >= nxt->timeout.target_seq - tail)
5695 list_add(&req->timeout.list, entry);
5696 data->timer.function = io_timeout_fn;
5697 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5698 spin_unlock_irq(&ctx->completion_lock);
5702 struct io_cancel_data {
5703 struct io_ring_ctx *ctx;
5707 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5709 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5710 struct io_cancel_data *cd = data;
5712 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5715 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5716 struct io_ring_ctx *ctx)
5718 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5719 enum io_wq_cancel cancel_ret;
5722 if (!tctx || !tctx->io_wq)
5725 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5726 switch (cancel_ret) {
5727 case IO_WQ_CANCEL_OK:
5730 case IO_WQ_CANCEL_RUNNING:
5733 case IO_WQ_CANCEL_NOTFOUND:
5741 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5742 struct io_kiocb *req, __u64 sqe_addr,
5745 unsigned long flags;
5748 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5749 spin_lock_irqsave(&ctx->completion_lock, flags);
5752 ret = io_timeout_cancel(ctx, sqe_addr);
5755 ret = io_poll_cancel(ctx, sqe_addr, false);
5759 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5760 io_commit_cqring(ctx);
5761 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5762 io_cqring_ev_posted(ctx);
5765 req_set_fail_links(req);
5768 static int io_async_cancel_prep(struct io_kiocb *req,
5769 const struct io_uring_sqe *sqe)
5771 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5773 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5775 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5778 req->cancel.addr = READ_ONCE(sqe->addr);
5782 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5784 struct io_ring_ctx *ctx = req->ctx;
5785 u64 sqe_addr = req->cancel.addr;
5786 struct io_tctx_node *node;
5789 /* tasks should wait for their io-wq threads, so safe w/o sync */
5790 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5791 spin_lock_irq(&ctx->completion_lock);
5794 ret = io_timeout_cancel(ctx, sqe_addr);
5797 ret = io_poll_cancel(ctx, sqe_addr, false);
5800 spin_unlock_irq(&ctx->completion_lock);
5802 /* slow path, try all io-wq's */
5803 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5805 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5806 struct io_uring_task *tctx = node->task->io_uring;
5808 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5812 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5814 spin_lock_irq(&ctx->completion_lock);
5816 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5817 io_commit_cqring(ctx);
5818 spin_unlock_irq(&ctx->completion_lock);
5819 io_cqring_ev_posted(ctx);
5822 req_set_fail_links(req);
5827 static int io_rsrc_update_prep(struct io_kiocb *req,
5828 const struct io_uring_sqe *sqe)
5830 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5832 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5834 if (sqe->ioprio || sqe->rw_flags)
5837 req->rsrc_update.offset = READ_ONCE(sqe->off);
5838 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5839 if (!req->rsrc_update.nr_args)
5841 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5845 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5847 struct io_ring_ctx *ctx = req->ctx;
5848 struct io_uring_rsrc_update2 up;
5851 if (issue_flags & IO_URING_F_NONBLOCK)
5854 up.offset = req->rsrc_update.offset;
5855 up.data = req->rsrc_update.arg;
5860 mutex_lock(&ctx->uring_lock);
5861 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5862 &up, req->rsrc_update.nr_args);
5863 mutex_unlock(&ctx->uring_lock);
5866 req_set_fail_links(req);
5867 __io_req_complete(req, issue_flags, ret, 0);
5871 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5873 switch (req->opcode) {
5876 case IORING_OP_READV:
5877 case IORING_OP_READ_FIXED:
5878 case IORING_OP_READ:
5879 return io_read_prep(req, sqe);
5880 case IORING_OP_WRITEV:
5881 case IORING_OP_WRITE_FIXED:
5882 case IORING_OP_WRITE:
5883 return io_write_prep(req, sqe);
5884 case IORING_OP_POLL_ADD:
5885 return io_poll_add_prep(req, sqe);
5886 case IORING_OP_POLL_REMOVE:
5887 return io_poll_update_prep(req, sqe);
5888 case IORING_OP_FSYNC:
5889 return io_fsync_prep(req, sqe);
5890 case IORING_OP_SYNC_FILE_RANGE:
5891 return io_sfr_prep(req, sqe);
5892 case IORING_OP_SENDMSG:
5893 case IORING_OP_SEND:
5894 return io_sendmsg_prep(req, sqe);
5895 case IORING_OP_RECVMSG:
5896 case IORING_OP_RECV:
5897 return io_recvmsg_prep(req, sqe);
5898 case IORING_OP_CONNECT:
5899 return io_connect_prep(req, sqe);
5900 case IORING_OP_TIMEOUT:
5901 return io_timeout_prep(req, sqe, false);
5902 case IORING_OP_TIMEOUT_REMOVE:
5903 return io_timeout_remove_prep(req, sqe);
5904 case IORING_OP_ASYNC_CANCEL:
5905 return io_async_cancel_prep(req, sqe);
5906 case IORING_OP_LINK_TIMEOUT:
5907 return io_timeout_prep(req, sqe, true);
5908 case IORING_OP_ACCEPT:
5909 return io_accept_prep(req, sqe);
5910 case IORING_OP_FALLOCATE:
5911 return io_fallocate_prep(req, sqe);
5912 case IORING_OP_OPENAT:
5913 return io_openat_prep(req, sqe);
5914 case IORING_OP_CLOSE:
5915 return io_close_prep(req, sqe);
5916 case IORING_OP_FILES_UPDATE:
5917 return io_rsrc_update_prep(req, sqe);
5918 case IORING_OP_STATX:
5919 return io_statx_prep(req, sqe);
5920 case IORING_OP_FADVISE:
5921 return io_fadvise_prep(req, sqe);
5922 case IORING_OP_MADVISE:
5923 return io_madvise_prep(req, sqe);
5924 case IORING_OP_OPENAT2:
5925 return io_openat2_prep(req, sqe);
5926 case IORING_OP_EPOLL_CTL:
5927 return io_epoll_ctl_prep(req, sqe);
5928 case IORING_OP_SPLICE:
5929 return io_splice_prep(req, sqe);
5930 case IORING_OP_PROVIDE_BUFFERS:
5931 return io_provide_buffers_prep(req, sqe);
5932 case IORING_OP_REMOVE_BUFFERS:
5933 return io_remove_buffers_prep(req, sqe);
5935 return io_tee_prep(req, sqe);
5936 case IORING_OP_SHUTDOWN:
5937 return io_shutdown_prep(req, sqe);
5938 case IORING_OP_RENAMEAT:
5939 return io_renameat_prep(req, sqe);
5940 case IORING_OP_UNLINKAT:
5941 return io_unlinkat_prep(req, sqe);
5944 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5949 static int io_req_prep_async(struct io_kiocb *req)
5951 if (!io_op_defs[req->opcode].needs_async_setup)
5953 if (WARN_ON_ONCE(req->async_data))
5955 if (io_alloc_async_data(req))
5958 switch (req->opcode) {
5959 case IORING_OP_READV:
5960 return io_rw_prep_async(req, READ);
5961 case IORING_OP_WRITEV:
5962 return io_rw_prep_async(req, WRITE);
5963 case IORING_OP_SENDMSG:
5964 return io_sendmsg_prep_async(req);
5965 case IORING_OP_RECVMSG:
5966 return io_recvmsg_prep_async(req);
5967 case IORING_OP_CONNECT:
5968 return io_connect_prep_async(req);
5970 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5975 static u32 io_get_sequence(struct io_kiocb *req)
5977 struct io_kiocb *pos;
5978 struct io_ring_ctx *ctx = req->ctx;
5979 u32 total_submitted, nr_reqs = 0;
5981 io_for_each_link(pos, req)
5984 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5985 return total_submitted - nr_reqs;
5988 static int io_req_defer(struct io_kiocb *req)
5990 struct io_ring_ctx *ctx = req->ctx;
5991 struct io_defer_entry *de;
5995 /* Still need defer if there is pending req in defer list. */
5996 if (likely(list_empty_careful(&ctx->defer_list) &&
5997 !(req->flags & REQ_F_IO_DRAIN)))
6000 seq = io_get_sequence(req);
6001 /* Still a chance to pass the sequence check */
6002 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6005 ret = io_req_prep_async(req);
6008 io_prep_async_link(req);
6009 de = kmalloc(sizeof(*de), GFP_KERNEL);
6013 spin_lock_irq(&ctx->completion_lock);
6014 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6015 spin_unlock_irq(&ctx->completion_lock);
6017 io_queue_async_work(req);
6018 return -EIOCBQUEUED;
6021 trace_io_uring_defer(ctx, req, req->user_data);
6024 list_add_tail(&de->list, &ctx->defer_list);
6025 spin_unlock_irq(&ctx->completion_lock);
6026 return -EIOCBQUEUED;
6029 static void io_clean_op(struct io_kiocb *req)
6031 if (req->flags & REQ_F_BUFFER_SELECTED) {
6032 switch (req->opcode) {
6033 case IORING_OP_READV:
6034 case IORING_OP_READ_FIXED:
6035 case IORING_OP_READ:
6036 kfree((void *)(unsigned long)req->rw.addr);
6038 case IORING_OP_RECVMSG:
6039 case IORING_OP_RECV:
6040 kfree(req->sr_msg.kbuf);
6043 req->flags &= ~REQ_F_BUFFER_SELECTED;
6046 if (req->flags & REQ_F_NEED_CLEANUP) {
6047 switch (req->opcode) {
6048 case IORING_OP_READV:
6049 case IORING_OP_READ_FIXED:
6050 case IORING_OP_READ:
6051 case IORING_OP_WRITEV:
6052 case IORING_OP_WRITE_FIXED:
6053 case IORING_OP_WRITE: {
6054 struct io_async_rw *io = req->async_data;
6056 kfree(io->free_iovec);
6059 case IORING_OP_RECVMSG:
6060 case IORING_OP_SENDMSG: {
6061 struct io_async_msghdr *io = req->async_data;
6063 kfree(io->free_iov);
6066 case IORING_OP_SPLICE:
6068 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6069 io_put_file(req->splice.file_in);
6071 case IORING_OP_OPENAT:
6072 case IORING_OP_OPENAT2:
6073 if (req->open.filename)
6074 putname(req->open.filename);
6076 case IORING_OP_RENAMEAT:
6077 putname(req->rename.oldpath);
6078 putname(req->rename.newpath);
6080 case IORING_OP_UNLINKAT:
6081 putname(req->unlink.filename);
6084 req->flags &= ~REQ_F_NEED_CLEANUP;
6086 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6087 kfree(req->apoll->double_poll);
6091 if (req->flags & REQ_F_INFLIGHT) {
6092 struct io_uring_task *tctx = req->task->io_uring;
6094 atomic_dec(&tctx->inflight_tracked);
6095 req->flags &= ~REQ_F_INFLIGHT;
6099 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6101 struct io_ring_ctx *ctx = req->ctx;
6102 const struct cred *creds = NULL;
6105 if (req->work.creds && req->work.creds != current_cred())
6106 creds = override_creds(req->work.creds);
6108 switch (req->opcode) {
6110 ret = io_nop(req, issue_flags);
6112 case IORING_OP_READV:
6113 case IORING_OP_READ_FIXED:
6114 case IORING_OP_READ:
6115 ret = io_read(req, issue_flags);
6117 case IORING_OP_WRITEV:
6118 case IORING_OP_WRITE_FIXED:
6119 case IORING_OP_WRITE:
6120 ret = io_write(req, issue_flags);
6122 case IORING_OP_FSYNC:
6123 ret = io_fsync(req, issue_flags);
6125 case IORING_OP_POLL_ADD:
6126 ret = io_poll_add(req, issue_flags);
6128 case IORING_OP_POLL_REMOVE:
6129 ret = io_poll_update(req, issue_flags);
6131 case IORING_OP_SYNC_FILE_RANGE:
6132 ret = io_sync_file_range(req, issue_flags);
6134 case IORING_OP_SENDMSG:
6135 ret = io_sendmsg(req, issue_flags);
6137 case IORING_OP_SEND:
6138 ret = io_send(req, issue_flags);
6140 case IORING_OP_RECVMSG:
6141 ret = io_recvmsg(req, issue_flags);
6143 case IORING_OP_RECV:
6144 ret = io_recv(req, issue_flags);
6146 case IORING_OP_TIMEOUT:
6147 ret = io_timeout(req, issue_flags);
6149 case IORING_OP_TIMEOUT_REMOVE:
6150 ret = io_timeout_remove(req, issue_flags);
6152 case IORING_OP_ACCEPT:
6153 ret = io_accept(req, issue_flags);
6155 case IORING_OP_CONNECT:
6156 ret = io_connect(req, issue_flags);
6158 case IORING_OP_ASYNC_CANCEL:
6159 ret = io_async_cancel(req, issue_flags);
6161 case IORING_OP_FALLOCATE:
6162 ret = io_fallocate(req, issue_flags);
6164 case IORING_OP_OPENAT:
6165 ret = io_openat(req, issue_flags);
6167 case IORING_OP_CLOSE:
6168 ret = io_close(req, issue_flags);
6170 case IORING_OP_FILES_UPDATE:
6171 ret = io_files_update(req, issue_flags);
6173 case IORING_OP_STATX:
6174 ret = io_statx(req, issue_flags);
6176 case IORING_OP_FADVISE:
6177 ret = io_fadvise(req, issue_flags);
6179 case IORING_OP_MADVISE:
6180 ret = io_madvise(req, issue_flags);
6182 case IORING_OP_OPENAT2:
6183 ret = io_openat2(req, issue_flags);
6185 case IORING_OP_EPOLL_CTL:
6186 ret = io_epoll_ctl(req, issue_flags);
6188 case IORING_OP_SPLICE:
6189 ret = io_splice(req, issue_flags);
6191 case IORING_OP_PROVIDE_BUFFERS:
6192 ret = io_provide_buffers(req, issue_flags);
6194 case IORING_OP_REMOVE_BUFFERS:
6195 ret = io_remove_buffers(req, issue_flags);
6198 ret = io_tee(req, issue_flags);
6200 case IORING_OP_SHUTDOWN:
6201 ret = io_shutdown(req, issue_flags);
6203 case IORING_OP_RENAMEAT:
6204 ret = io_renameat(req, issue_flags);
6206 case IORING_OP_UNLINKAT:
6207 ret = io_unlinkat(req, issue_flags);
6215 revert_creds(creds);
6220 /* If the op doesn't have a file, we're not polling for it */
6221 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6222 const bool in_async = io_wq_current_is_worker();
6224 /* workqueue context doesn't hold uring_lock, grab it now */
6226 mutex_lock(&ctx->uring_lock);
6228 io_iopoll_req_issued(req, in_async);
6231 mutex_unlock(&ctx->uring_lock);
6237 static void io_wq_submit_work(struct io_wq_work *work)
6239 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6240 struct io_kiocb *timeout;
6243 timeout = io_prep_linked_timeout(req);
6245 io_queue_linked_timeout(timeout);
6247 if (work->flags & IO_WQ_WORK_CANCEL)
6252 ret = io_issue_sqe(req, 0);
6254 * We can get EAGAIN for polled IO even though we're
6255 * forcing a sync submission from here, since we can't
6256 * wait for request slots on the block side.
6264 /* avoid locking problems by failing it from a clean context */
6266 /* io-wq is going to take one down */
6268 io_req_task_queue_fail(req, ret);
6272 #define FFS_ASYNC_READ 0x1UL
6273 #define FFS_ASYNC_WRITE 0x2UL
6275 #define FFS_ISREG 0x4UL
6277 #define FFS_ISREG 0x0UL
6279 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6281 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6284 struct io_fixed_file *table_l2;
6286 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6287 return &table_l2[i & IORING_FILE_TABLE_MASK];
6290 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6293 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6295 return (struct file *) (slot->file_ptr & FFS_MASK);
6298 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6300 unsigned long file_ptr = (unsigned long) file;
6302 if (__io_file_supports_async(file, READ))
6303 file_ptr |= FFS_ASYNC_READ;
6304 if (__io_file_supports_async(file, WRITE))
6305 file_ptr |= FFS_ASYNC_WRITE;
6306 if (S_ISREG(file_inode(file)->i_mode))
6307 file_ptr |= FFS_ISREG;
6308 file_slot->file_ptr = file_ptr;
6311 static struct file *io_file_get(struct io_submit_state *state,
6312 struct io_kiocb *req, int fd, bool fixed)
6314 struct io_ring_ctx *ctx = req->ctx;
6318 unsigned long file_ptr;
6320 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6322 fd = array_index_nospec(fd, ctx->nr_user_files);
6323 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6324 file = (struct file *) (file_ptr & FFS_MASK);
6325 file_ptr &= ~FFS_MASK;
6326 /* mask in overlapping REQ_F and FFS bits */
6327 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6328 io_req_set_rsrc_node(req);
6330 trace_io_uring_file_get(ctx, fd);
6331 file = __io_file_get(state, fd);
6333 /* we don't allow fixed io_uring files */
6334 if (file && unlikely(file->f_op == &io_uring_fops))
6335 io_req_track_inflight(req);
6341 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6343 struct io_timeout_data *data = container_of(timer,
6344 struct io_timeout_data, timer);
6345 struct io_kiocb *prev, *req = data->req;
6346 struct io_ring_ctx *ctx = req->ctx;
6347 unsigned long flags;
6349 spin_lock_irqsave(&ctx->completion_lock, flags);
6350 prev = req->timeout.head;
6351 req->timeout.head = NULL;
6354 * We don't expect the list to be empty, that will only happen if we
6355 * race with the completion of the linked work.
6357 if (prev && req_ref_inc_not_zero(prev))
6358 io_remove_next_linked(prev);
6361 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6364 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6365 io_put_req_deferred(prev, 1);
6367 io_req_complete_post(req, -ETIME, 0);
6369 io_put_req_deferred(req, 1);
6370 return HRTIMER_NORESTART;
6373 static void io_queue_linked_timeout(struct io_kiocb *req)
6375 struct io_ring_ctx *ctx = req->ctx;
6377 spin_lock_irq(&ctx->completion_lock);
6379 * If the back reference is NULL, then our linked request finished
6380 * before we got a chance to setup the timer
6382 if (req->timeout.head) {
6383 struct io_timeout_data *data = req->async_data;
6385 data->timer.function = io_link_timeout_fn;
6386 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6389 spin_unlock_irq(&ctx->completion_lock);
6390 /* drop submission reference */
6394 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6396 struct io_kiocb *nxt = req->link;
6398 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6399 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6402 nxt->timeout.head = req;
6403 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6404 req->flags |= REQ_F_LINK_TIMEOUT;
6408 static void __io_queue_sqe(struct io_kiocb *req)
6410 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6413 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6416 * We async punt it if the file wasn't marked NOWAIT, or if the file
6417 * doesn't support non-blocking read/write attempts
6420 /* drop submission reference */
6421 if (req->flags & REQ_F_COMPLETE_INLINE) {
6422 struct io_ring_ctx *ctx = req->ctx;
6423 struct io_comp_state *cs = &ctx->submit_state.comp;
6425 cs->reqs[cs->nr++] = req;
6426 if (cs->nr == ARRAY_SIZE(cs->reqs))
6427 io_submit_flush_completions(cs, ctx);
6431 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6432 if (!io_arm_poll_handler(req)) {
6434 * Queued up for async execution, worker will release
6435 * submit reference when the iocb is actually submitted.
6437 io_queue_async_work(req);
6440 io_req_complete_failed(req, ret);
6443 io_queue_linked_timeout(linked_timeout);
6446 static void io_queue_sqe(struct io_kiocb *req)
6450 ret = io_req_defer(req);
6452 if (ret != -EIOCBQUEUED) {
6454 io_req_complete_failed(req, ret);
6456 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6457 ret = io_req_prep_async(req);
6460 io_queue_async_work(req);
6462 __io_queue_sqe(req);
6467 * Check SQE restrictions (opcode and flags).
6469 * Returns 'true' if SQE is allowed, 'false' otherwise.
6471 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6472 struct io_kiocb *req,
6473 unsigned int sqe_flags)
6475 if (!ctx->restricted)
6478 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6481 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6482 ctx->restrictions.sqe_flags_required)
6485 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6486 ctx->restrictions.sqe_flags_required))
6492 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6493 const struct io_uring_sqe *sqe)
6495 struct io_submit_state *state;
6496 unsigned int sqe_flags;
6497 int personality, ret = 0;
6499 req->opcode = READ_ONCE(sqe->opcode);
6500 /* same numerical values with corresponding REQ_F_*, safe to copy */
6501 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6502 req->user_data = READ_ONCE(sqe->user_data);
6503 req->async_data = NULL;
6507 req->fixed_rsrc_refs = NULL;
6508 /* one is dropped after submission, the other at completion */
6509 atomic_set(&req->refs, 2);
6510 req->task = current;
6512 req->work.creds = NULL;
6514 /* enforce forwards compatibility on users */
6515 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6517 if (unlikely(req->opcode >= IORING_OP_LAST))
6519 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6522 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6523 !io_op_defs[req->opcode].buffer_select)
6526 personality = READ_ONCE(sqe->personality);
6528 req->work.creds = xa_load(&ctx->personalities, personality);
6529 if (!req->work.creds)
6531 get_cred(req->work.creds);
6533 state = &ctx->submit_state;
6536 * Plug now if we have more than 1 IO left after this, and the target
6537 * is potentially a read/write to block based storage.
6539 if (!state->plug_started && state->ios_left > 1 &&
6540 io_op_defs[req->opcode].plug) {
6541 blk_start_plug(&state->plug);
6542 state->plug_started = true;
6545 if (io_op_defs[req->opcode].needs_file) {
6546 bool fixed = req->flags & REQ_F_FIXED_FILE;
6548 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6549 if (unlikely(!req->file))
6557 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6558 const struct io_uring_sqe *sqe)
6560 struct io_submit_link *link = &ctx->submit_state.link;
6563 ret = io_init_req(ctx, req, sqe);
6564 if (unlikely(ret)) {
6567 /* fail even hard links since we don't submit */
6568 link->head->flags |= REQ_F_FAIL_LINK;
6569 io_req_complete_failed(link->head, -ECANCELED);
6572 io_req_complete_failed(req, ret);
6575 ret = io_req_prep(req, sqe);
6579 /* don't need @sqe from now on */
6580 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6581 true, ctx->flags & IORING_SETUP_SQPOLL);
6584 * If we already have a head request, queue this one for async
6585 * submittal once the head completes. If we don't have a head but
6586 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6587 * submitted sync once the chain is complete. If none of those
6588 * conditions are true (normal request), then just queue it.
6591 struct io_kiocb *head = link->head;
6594 * Taking sequential execution of a link, draining both sides
6595 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6596 * requests in the link. So, it drains the head and the
6597 * next after the link request. The last one is done via
6598 * drain_next flag to persist the effect across calls.
6600 if (req->flags & REQ_F_IO_DRAIN) {
6601 head->flags |= REQ_F_IO_DRAIN;
6602 ctx->drain_next = 1;
6604 ret = io_req_prep_async(req);
6607 trace_io_uring_link(ctx, req, head);
6608 link->last->link = req;
6611 /* last request of a link, enqueue the link */
6612 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6617 if (unlikely(ctx->drain_next)) {
6618 req->flags |= REQ_F_IO_DRAIN;
6619 ctx->drain_next = 0;
6621 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6633 * Batched submission is done, ensure local IO is flushed out.
6635 static void io_submit_state_end(struct io_submit_state *state,
6636 struct io_ring_ctx *ctx)
6638 if (state->link.head)
6639 io_queue_sqe(state->link.head);
6641 io_submit_flush_completions(&state->comp, ctx);
6642 if (state->plug_started)
6643 blk_finish_plug(&state->plug);
6644 io_state_file_put(state);
6648 * Start submission side cache.
6650 static void io_submit_state_start(struct io_submit_state *state,
6651 unsigned int max_ios)
6653 state->plug_started = false;
6654 state->ios_left = max_ios;
6655 /* set only head, no need to init link_last in advance */
6656 state->link.head = NULL;
6659 static void io_commit_sqring(struct io_ring_ctx *ctx)
6661 struct io_rings *rings = ctx->rings;
6664 * Ensure any loads from the SQEs are done at this point,
6665 * since once we write the new head, the application could
6666 * write new data to them.
6668 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6672 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6673 * that is mapped by userspace. This means that care needs to be taken to
6674 * ensure that reads are stable, as we cannot rely on userspace always
6675 * being a good citizen. If members of the sqe are validated and then later
6676 * used, it's important that those reads are done through READ_ONCE() to
6677 * prevent a re-load down the line.
6679 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6681 u32 *sq_array = ctx->sq_array;
6685 * The cached sq head (or cq tail) serves two purposes:
6687 * 1) allows us to batch the cost of updating the user visible
6689 * 2) allows the kernel side to track the head on its own, even
6690 * though the application is the one updating it.
6692 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6693 if (likely(head < ctx->sq_entries))
6694 return &ctx->sq_sqes[head];
6696 /* drop invalid entries */
6697 ctx->cached_sq_dropped++;
6698 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6702 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6706 /* make sure SQ entry isn't read before tail */
6707 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6709 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6712 percpu_counter_add(¤t->io_uring->inflight, nr);
6713 refcount_add(nr, ¤t->usage);
6714 io_submit_state_start(&ctx->submit_state, nr);
6716 while (submitted < nr) {
6717 const struct io_uring_sqe *sqe;
6718 struct io_kiocb *req;
6720 req = io_alloc_req(ctx);
6721 if (unlikely(!req)) {
6723 submitted = -EAGAIN;
6726 sqe = io_get_sqe(ctx);
6727 if (unlikely(!sqe)) {
6728 kmem_cache_free(req_cachep, req);
6731 /* will complete beyond this point, count as submitted */
6733 if (io_submit_sqe(ctx, req, sqe))
6737 if (unlikely(submitted != nr)) {
6738 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6739 struct io_uring_task *tctx = current->io_uring;
6740 int unused = nr - ref_used;
6742 percpu_ref_put_many(&ctx->refs, unused);
6743 percpu_counter_sub(&tctx->inflight, unused);
6744 put_task_struct_many(current, unused);
6747 io_submit_state_end(&ctx->submit_state, ctx);
6748 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6749 io_commit_sqring(ctx);
6754 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6756 /* Tell userspace we may need a wakeup call */
6757 spin_lock_irq(&ctx->completion_lock);
6758 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6759 spin_unlock_irq(&ctx->completion_lock);
6762 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6764 spin_lock_irq(&ctx->completion_lock);
6765 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6766 spin_unlock_irq(&ctx->completion_lock);
6769 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6771 unsigned int to_submit;
6774 to_submit = io_sqring_entries(ctx);
6775 /* if we're handling multiple rings, cap submit size for fairness */
6776 if (cap_entries && to_submit > 8)
6779 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6780 unsigned nr_events = 0;
6782 mutex_lock(&ctx->uring_lock);
6783 if (!list_empty(&ctx->iopoll_list))
6784 io_do_iopoll(ctx, &nr_events, 0);
6787 * Don't submit if refs are dying, good for io_uring_register(),
6788 * but also it is relied upon by io_ring_exit_work()
6790 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6791 !(ctx->flags & IORING_SETUP_R_DISABLED))
6792 ret = io_submit_sqes(ctx, to_submit);
6793 mutex_unlock(&ctx->uring_lock);
6796 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6797 wake_up(&ctx->sqo_sq_wait);
6802 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6804 struct io_ring_ctx *ctx;
6805 unsigned sq_thread_idle = 0;
6807 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6808 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6809 sqd->sq_thread_idle = sq_thread_idle;
6812 static int io_sq_thread(void *data)
6814 struct io_sq_data *sqd = data;
6815 struct io_ring_ctx *ctx;
6816 unsigned long timeout = 0;
6817 char buf[TASK_COMM_LEN];
6820 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6821 set_task_comm(current, buf);
6823 if (sqd->sq_cpu != -1)
6824 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6826 set_cpus_allowed_ptr(current, cpu_online_mask);
6827 current->flags |= PF_NO_SETAFFINITY;
6829 mutex_lock(&sqd->lock);
6830 /* a user may had exited before the thread started */
6831 io_run_task_work_head(&sqd->park_task_work);
6833 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6835 bool cap_entries, sqt_spin, needs_sched;
6837 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6838 signal_pending(current)) {
6839 bool did_sig = false;
6841 mutex_unlock(&sqd->lock);
6842 if (signal_pending(current)) {
6843 struct ksignal ksig;
6845 did_sig = get_signal(&ksig);
6848 mutex_lock(&sqd->lock);
6850 io_run_task_work_head(&sqd->park_task_work);
6853 timeout = jiffies + sqd->sq_thread_idle;
6857 cap_entries = !list_is_singular(&sqd->ctx_list);
6858 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6859 const struct cred *creds = NULL;
6861 if (ctx->sq_creds != current_cred())
6862 creds = override_creds(ctx->sq_creds);
6863 ret = __io_sq_thread(ctx, cap_entries);
6865 revert_creds(creds);
6866 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6870 if (sqt_spin || !time_after(jiffies, timeout)) {
6874 timeout = jiffies + sqd->sq_thread_idle;
6878 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6879 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6880 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6881 io_ring_set_wakeup_flag(ctx);
6884 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6885 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6886 !list_empty_careful(&ctx->iopoll_list)) {
6887 needs_sched = false;
6890 if (io_sqring_entries(ctx)) {
6891 needs_sched = false;
6897 mutex_unlock(&sqd->lock);
6899 mutex_lock(&sqd->lock);
6901 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6902 io_ring_clear_wakeup_flag(ctx);
6905 finish_wait(&sqd->wait, &wait);
6906 io_run_task_work_head(&sqd->park_task_work);
6907 timeout = jiffies + sqd->sq_thread_idle;
6910 io_uring_cancel_sqpoll(sqd);
6912 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6913 io_ring_set_wakeup_flag(ctx);
6915 io_run_task_work_head(&sqd->park_task_work);
6916 mutex_unlock(&sqd->lock);
6918 complete(&sqd->exited);
6922 struct io_wait_queue {
6923 struct wait_queue_entry wq;
6924 struct io_ring_ctx *ctx;
6926 unsigned nr_timeouts;
6929 static inline bool io_should_wake(struct io_wait_queue *iowq)
6931 struct io_ring_ctx *ctx = iowq->ctx;
6934 * Wake up if we have enough events, or if a timeout occurred since we
6935 * started waiting. For timeouts, we always want to return to userspace,
6936 * regardless of event count.
6938 return io_cqring_events(ctx) >= iowq->to_wait ||
6939 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6942 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6943 int wake_flags, void *key)
6945 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6949 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6950 * the task, and the next invocation will do it.
6952 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6953 return autoremove_wake_function(curr, mode, wake_flags, key);
6957 static int io_run_task_work_sig(void)
6959 if (io_run_task_work())
6961 if (!signal_pending(current))
6963 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6964 return -ERESTARTSYS;
6968 /* when returns >0, the caller should retry */
6969 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6970 struct io_wait_queue *iowq,
6971 signed long *timeout)
6975 /* make sure we run task_work before checking for signals */
6976 ret = io_run_task_work_sig();
6977 if (ret || io_should_wake(iowq))
6979 /* let the caller flush overflows, retry */
6980 if (test_bit(0, &ctx->cq_check_overflow))
6983 *timeout = schedule_timeout(*timeout);
6984 return !*timeout ? -ETIME : 1;
6988 * Wait until events become available, if we don't already have some. The
6989 * application must reap them itself, as they reside on the shared cq ring.
6991 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6992 const sigset_t __user *sig, size_t sigsz,
6993 struct __kernel_timespec __user *uts)
6995 struct io_wait_queue iowq = {
6998 .func = io_wake_function,
6999 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7002 .to_wait = min_events,
7004 struct io_rings *rings = ctx->rings;
7005 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7009 io_cqring_overflow_flush(ctx, false);
7010 if (io_cqring_events(ctx) >= min_events)
7012 if (!io_run_task_work())
7017 #ifdef CONFIG_COMPAT
7018 if (in_compat_syscall())
7019 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7023 ret = set_user_sigmask(sig, sigsz);
7030 struct timespec64 ts;
7032 if (get_timespec64(&ts, uts))
7034 timeout = timespec64_to_jiffies(&ts);
7037 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7038 trace_io_uring_cqring_wait(ctx, min_events);
7040 /* if we can't even flush overflow, don't wait for more */
7041 if (!io_cqring_overflow_flush(ctx, false)) {
7045 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7046 TASK_INTERRUPTIBLE);
7047 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7048 finish_wait(&ctx->wait, &iowq.wq);
7052 restore_saved_sigmask_unless(ret == -EINTR);
7054 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7057 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7059 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7061 for (i = 0; i < nr_tables; i++)
7062 kfree(table->files[i]);
7063 kfree(table->files);
7064 table->files = NULL;
7067 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7069 spin_lock_bh(&ctx->rsrc_ref_lock);
7072 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7074 spin_unlock_bh(&ctx->rsrc_ref_lock);
7077 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7079 percpu_ref_exit(&ref_node->refs);
7083 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7084 struct io_rsrc_data *data_to_kill)
7086 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7087 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7090 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7092 rsrc_node->rsrc_data = data_to_kill;
7093 io_rsrc_ref_lock(ctx);
7094 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7095 io_rsrc_ref_unlock(ctx);
7097 atomic_inc(&data_to_kill->refs);
7098 percpu_ref_kill(&rsrc_node->refs);
7099 ctx->rsrc_node = NULL;
7102 if (!ctx->rsrc_node) {
7103 ctx->rsrc_node = ctx->rsrc_backup_node;
7104 ctx->rsrc_backup_node = NULL;
7108 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7110 if (ctx->rsrc_backup_node)
7112 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7113 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7116 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7120 /* As we may drop ->uring_lock, other task may have started quiesce */
7124 data->quiesce = true;
7126 ret = io_rsrc_node_switch_start(ctx);
7129 io_rsrc_node_switch(ctx, data);
7131 /* kill initial ref, already quiesced if zero */
7132 if (atomic_dec_and_test(&data->refs))
7134 flush_delayed_work(&ctx->rsrc_put_work);
7135 ret = wait_for_completion_interruptible(&data->done);
7139 atomic_inc(&data->refs);
7140 /* wait for all works potentially completing data->done */
7141 flush_delayed_work(&ctx->rsrc_put_work);
7142 reinit_completion(&data->done);
7144 mutex_unlock(&ctx->uring_lock);
7145 ret = io_run_task_work_sig();
7146 mutex_lock(&ctx->uring_lock);
7148 data->quiesce = false;
7153 static void io_rsrc_data_free(struct io_rsrc_data *data)
7159 static struct io_rsrc_data *io_rsrc_data_alloc(struct io_ring_ctx *ctx,
7160 rsrc_put_fn *do_put,
7163 struct io_rsrc_data *data;
7165 data = kzalloc(sizeof(*data), GFP_KERNEL);
7169 data->tags = kvcalloc(nr, sizeof(*data->tags), GFP_KERNEL);
7175 atomic_set(&data->refs, 1);
7177 data->do_put = do_put;
7178 init_completion(&data->done);
7182 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7184 #if defined(CONFIG_UNIX)
7185 if (ctx->ring_sock) {
7186 struct sock *sock = ctx->ring_sock->sk;
7187 struct sk_buff *skb;
7189 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7195 for (i = 0; i < ctx->nr_user_files; i++) {
7198 file = io_file_from_index(ctx, i);
7203 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7204 io_rsrc_data_free(ctx->file_data);
7205 ctx->file_data = NULL;
7206 ctx->nr_user_files = 0;
7209 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7213 if (!ctx->file_data)
7215 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7217 __io_sqe_files_unregister(ctx);
7221 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7222 __releases(&sqd->lock)
7224 WARN_ON_ONCE(sqd->thread == current);
7227 * Do the dance but not conditional clear_bit() because it'd race with
7228 * other threads incrementing park_pending and setting the bit.
7230 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7231 if (atomic_dec_return(&sqd->park_pending))
7232 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7233 mutex_unlock(&sqd->lock);
7236 static void io_sq_thread_park(struct io_sq_data *sqd)
7237 __acquires(&sqd->lock)
7239 WARN_ON_ONCE(sqd->thread == current);
7241 atomic_inc(&sqd->park_pending);
7242 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7243 mutex_lock(&sqd->lock);
7245 wake_up_process(sqd->thread);
7248 static void io_sq_thread_stop(struct io_sq_data *sqd)
7250 WARN_ON_ONCE(sqd->thread == current);
7251 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7253 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7254 mutex_lock(&sqd->lock);
7256 wake_up_process(sqd->thread);
7257 mutex_unlock(&sqd->lock);
7258 wait_for_completion(&sqd->exited);
7261 static void io_put_sq_data(struct io_sq_data *sqd)
7263 if (refcount_dec_and_test(&sqd->refs)) {
7264 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7266 io_sq_thread_stop(sqd);
7271 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7273 struct io_sq_data *sqd = ctx->sq_data;
7276 io_sq_thread_park(sqd);
7277 list_del_init(&ctx->sqd_list);
7278 io_sqd_update_thread_idle(sqd);
7279 io_sq_thread_unpark(sqd);
7281 io_put_sq_data(sqd);
7282 ctx->sq_data = NULL;
7286 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7288 struct io_ring_ctx *ctx_attach;
7289 struct io_sq_data *sqd;
7292 f = fdget(p->wq_fd);
7294 return ERR_PTR(-ENXIO);
7295 if (f.file->f_op != &io_uring_fops) {
7297 return ERR_PTR(-EINVAL);
7300 ctx_attach = f.file->private_data;
7301 sqd = ctx_attach->sq_data;
7304 return ERR_PTR(-EINVAL);
7306 if (sqd->task_tgid != current->tgid) {
7308 return ERR_PTR(-EPERM);
7311 refcount_inc(&sqd->refs);
7316 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7319 struct io_sq_data *sqd;
7322 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7323 sqd = io_attach_sq_data(p);
7328 /* fall through for EPERM case, setup new sqd/task */
7329 if (PTR_ERR(sqd) != -EPERM)
7333 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7335 return ERR_PTR(-ENOMEM);
7337 atomic_set(&sqd->park_pending, 0);
7338 refcount_set(&sqd->refs, 1);
7339 INIT_LIST_HEAD(&sqd->ctx_list);
7340 mutex_init(&sqd->lock);
7341 init_waitqueue_head(&sqd->wait);
7342 init_completion(&sqd->exited);
7346 #if defined(CONFIG_UNIX)
7348 * Ensure the UNIX gc is aware of our file set, so we are certain that
7349 * the io_uring can be safely unregistered on process exit, even if we have
7350 * loops in the file referencing.
7352 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7354 struct sock *sk = ctx->ring_sock->sk;
7355 struct scm_fp_list *fpl;
7356 struct sk_buff *skb;
7359 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7363 skb = alloc_skb(0, GFP_KERNEL);
7372 fpl->user = get_uid(current_user());
7373 for (i = 0; i < nr; i++) {
7374 struct file *file = io_file_from_index(ctx, i + offset);
7378 fpl->fp[nr_files] = get_file(file);
7379 unix_inflight(fpl->user, fpl->fp[nr_files]);
7384 fpl->max = SCM_MAX_FD;
7385 fpl->count = nr_files;
7386 UNIXCB(skb).fp = fpl;
7387 skb->destructor = unix_destruct_scm;
7388 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7389 skb_queue_head(&sk->sk_receive_queue, skb);
7391 for (i = 0; i < nr_files; i++)
7402 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7403 * causes regular reference counting to break down. We rely on the UNIX
7404 * garbage collection to take care of this problem for us.
7406 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7408 unsigned left, total;
7412 left = ctx->nr_user_files;
7414 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7416 ret = __io_sqe_files_scm(ctx, this_files, total);
7420 total += this_files;
7426 while (total < ctx->nr_user_files) {
7427 struct file *file = io_file_from_index(ctx, total);
7437 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7443 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7445 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7447 table->files = kcalloc(nr_tables, sizeof(*table->files), GFP_KERNEL);
7451 for (i = 0; i < nr_tables; i++) {
7452 unsigned int this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7454 table->files[i] = kcalloc(this_files, sizeof(*table->files[i]),
7456 if (!table->files[i])
7458 nr_files -= this_files;
7464 io_free_file_tables(table, nr_tables * IORING_MAX_FILES_TABLE);
7468 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7470 struct file *file = prsrc->file;
7471 #if defined(CONFIG_UNIX)
7472 struct sock *sock = ctx->ring_sock->sk;
7473 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7474 struct sk_buff *skb;
7477 __skb_queue_head_init(&list);
7480 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7481 * remove this entry and rearrange the file array.
7483 skb = skb_dequeue(head);
7485 struct scm_fp_list *fp;
7487 fp = UNIXCB(skb).fp;
7488 for (i = 0; i < fp->count; i++) {
7491 if (fp->fp[i] != file)
7494 unix_notinflight(fp->user, fp->fp[i]);
7495 left = fp->count - 1 - i;
7497 memmove(&fp->fp[i], &fp->fp[i + 1],
7498 left * sizeof(struct file *));
7505 __skb_queue_tail(&list, skb);
7515 __skb_queue_tail(&list, skb);
7517 skb = skb_dequeue(head);
7520 if (skb_peek(&list)) {
7521 spin_lock_irq(&head->lock);
7522 while ((skb = __skb_dequeue(&list)) != NULL)
7523 __skb_queue_tail(head, skb);
7524 spin_unlock_irq(&head->lock);
7531 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7533 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7534 struct io_ring_ctx *ctx = rsrc_data->ctx;
7535 struct io_rsrc_put *prsrc, *tmp;
7537 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7538 list_del(&prsrc->list);
7541 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7542 unsigned long flags;
7544 io_ring_submit_lock(ctx, lock_ring);
7545 spin_lock_irqsave(&ctx->completion_lock, flags);
7546 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7548 io_commit_cqring(ctx);
7549 spin_unlock_irqrestore(&ctx->completion_lock, flags);
7550 io_cqring_ev_posted(ctx);
7551 io_ring_submit_unlock(ctx, lock_ring);
7554 rsrc_data->do_put(ctx, prsrc);
7558 io_rsrc_node_destroy(ref_node);
7559 if (atomic_dec_and_test(&rsrc_data->refs))
7560 complete(&rsrc_data->done);
7563 static void io_rsrc_put_work(struct work_struct *work)
7565 struct io_ring_ctx *ctx;
7566 struct llist_node *node;
7568 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7569 node = llist_del_all(&ctx->rsrc_put_llist);
7572 struct io_rsrc_node *ref_node;
7573 struct llist_node *next = node->next;
7575 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7576 __io_rsrc_put_work(ref_node);
7581 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7583 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7584 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7585 bool first_add = false;
7587 io_rsrc_ref_lock(ctx);
7590 while (!list_empty(&ctx->rsrc_ref_list)) {
7591 node = list_first_entry(&ctx->rsrc_ref_list,
7592 struct io_rsrc_node, node);
7593 /* recycle ref nodes in order */
7596 list_del(&node->node);
7597 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7599 io_rsrc_ref_unlock(ctx);
7602 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7605 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7607 struct io_rsrc_node *ref_node;
7609 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7613 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7618 INIT_LIST_HEAD(&ref_node->node);
7619 INIT_LIST_HEAD(&ref_node->rsrc_list);
7620 ref_node->done = false;
7624 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7625 unsigned nr_args, u64 __user *tags)
7627 __s32 __user *fds = (__s32 __user *) arg;
7631 struct io_rsrc_data *file_data;
7637 if (nr_args > IORING_MAX_FIXED_FILES)
7639 ret = io_rsrc_node_switch_start(ctx);
7643 file_data = io_rsrc_data_alloc(ctx, io_rsrc_file_put, nr_args);
7646 ctx->file_data = file_data;
7648 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7651 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7654 if ((tags && copy_from_user(&tag, &tags[i], sizeof(tag))) ||
7655 copy_from_user(&fd, &fds[i], sizeof(fd))) {
7659 /* allow sparse sets */
7669 if (unlikely(!file))
7673 * Don't allow io_uring instances to be registered. If UNIX
7674 * isn't enabled, then this causes a reference cycle and this
7675 * instance can never get freed. If UNIX is enabled we'll
7676 * handle it just fine, but there's still no point in allowing
7677 * a ring fd as it doesn't support regular read/write anyway.
7679 if (file->f_op == &io_uring_fops) {
7683 ctx->file_data->tags[i] = tag;
7684 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7687 ret = io_sqe_files_scm(ctx);
7689 __io_sqe_files_unregister(ctx);
7693 io_rsrc_node_switch(ctx, NULL);
7696 for (i = 0; i < ctx->nr_user_files; i++) {
7697 file = io_file_from_index(ctx, i);
7701 io_free_file_tables(&ctx->file_table, nr_args);
7702 ctx->nr_user_files = 0;
7704 io_rsrc_data_free(ctx->file_data);
7705 ctx->file_data = NULL;
7709 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7712 #if defined(CONFIG_UNIX)
7713 struct sock *sock = ctx->ring_sock->sk;
7714 struct sk_buff_head *head = &sock->sk_receive_queue;
7715 struct sk_buff *skb;
7718 * See if we can merge this file into an existing skb SCM_RIGHTS
7719 * file set. If there's no room, fall back to allocating a new skb
7720 * and filling it in.
7722 spin_lock_irq(&head->lock);
7723 skb = skb_peek(head);
7725 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7727 if (fpl->count < SCM_MAX_FD) {
7728 __skb_unlink(skb, head);
7729 spin_unlock_irq(&head->lock);
7730 fpl->fp[fpl->count] = get_file(file);
7731 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7733 spin_lock_irq(&head->lock);
7734 __skb_queue_head(head, skb);
7739 spin_unlock_irq(&head->lock);
7746 return __io_sqe_files_scm(ctx, 1, index);
7752 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7753 struct io_rsrc_node *node, void *rsrc)
7755 struct io_rsrc_put *prsrc;
7757 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7761 prsrc->tag = data->tags[idx];
7763 list_add(&prsrc->list, &node->rsrc_list);
7767 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7768 struct io_uring_rsrc_update2 *up,
7771 u64 __user *tags = u64_to_user_ptr(up->tags);
7772 __s32 __user *fds = u64_to_user_ptr(up->data);
7773 struct io_rsrc_data *data = ctx->file_data;
7774 struct io_fixed_file *file_slot;
7778 bool needs_switch = false;
7780 if (!ctx->file_data)
7782 if (up->offset + nr_args > ctx->nr_user_files)
7785 for (done = 0; done < nr_args; done++) {
7788 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7789 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7793 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7797 if (fd == IORING_REGISTER_FILES_SKIP)
7800 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7801 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7803 if (file_slot->file_ptr) {
7804 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7805 err = io_queue_rsrc_removal(data, up->offset + done,
7806 ctx->rsrc_node, file);
7809 file_slot->file_ptr = 0;
7810 needs_switch = true;
7819 * Don't allow io_uring instances to be registered. If
7820 * UNIX isn't enabled, then this causes a reference
7821 * cycle and this instance can never get freed. If UNIX
7822 * is enabled we'll handle it just fine, but there's
7823 * still no point in allowing a ring fd as it doesn't
7824 * support regular read/write anyway.
7826 if (file->f_op == &io_uring_fops) {
7831 data->tags[up->offset + done] = tag;
7832 io_fixed_file_set(file_slot, file);
7833 err = io_sqe_file_register(ctx, file, i);
7835 file_slot->file_ptr = 0;
7843 io_rsrc_node_switch(ctx, data);
7844 return done ? done : err;
7847 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7849 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7851 req = io_put_req_find_next(req);
7852 return req ? &req->work : NULL;
7855 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7856 struct task_struct *task)
7858 struct io_wq_hash *hash;
7859 struct io_wq_data data;
7860 unsigned int concurrency;
7862 hash = ctx->hash_map;
7864 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7866 return ERR_PTR(-ENOMEM);
7867 refcount_set(&hash->refs, 1);
7868 init_waitqueue_head(&hash->wait);
7869 ctx->hash_map = hash;
7874 data.free_work = io_free_work;
7875 data.do_work = io_wq_submit_work;
7877 /* Do QD, or 4 * CPUS, whatever is smallest */
7878 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7880 return io_wq_create(concurrency, &data);
7883 static int io_uring_alloc_task_context(struct task_struct *task,
7884 struct io_ring_ctx *ctx)
7886 struct io_uring_task *tctx;
7889 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7890 if (unlikely(!tctx))
7893 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7894 if (unlikely(ret)) {
7899 tctx->io_wq = io_init_wq_offload(ctx, task);
7900 if (IS_ERR(tctx->io_wq)) {
7901 ret = PTR_ERR(tctx->io_wq);
7902 percpu_counter_destroy(&tctx->inflight);
7908 init_waitqueue_head(&tctx->wait);
7910 atomic_set(&tctx->in_idle, 0);
7911 atomic_set(&tctx->inflight_tracked, 0);
7912 task->io_uring = tctx;
7913 spin_lock_init(&tctx->task_lock);
7914 INIT_WQ_LIST(&tctx->task_list);
7915 tctx->task_state = 0;
7916 init_task_work(&tctx->task_work, tctx_task_work);
7920 void __io_uring_free(struct task_struct *tsk)
7922 struct io_uring_task *tctx = tsk->io_uring;
7924 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7925 WARN_ON_ONCE(tctx->io_wq);
7927 percpu_counter_destroy(&tctx->inflight);
7929 tsk->io_uring = NULL;
7932 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7933 struct io_uring_params *p)
7937 /* Retain compatibility with failing for an invalid attach attempt */
7938 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7939 IORING_SETUP_ATTACH_WQ) {
7942 f = fdget(p->wq_fd);
7946 if (f.file->f_op != &io_uring_fops)
7949 if (ctx->flags & IORING_SETUP_SQPOLL) {
7950 struct task_struct *tsk;
7951 struct io_sq_data *sqd;
7954 sqd = io_get_sq_data(p, &attached);
7960 ctx->sq_creds = get_current_cred();
7962 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7963 if (!ctx->sq_thread_idle)
7964 ctx->sq_thread_idle = HZ;
7966 io_sq_thread_park(sqd);
7967 list_add(&ctx->sqd_list, &sqd->ctx_list);
7968 io_sqd_update_thread_idle(sqd);
7969 /* don't attach to a dying SQPOLL thread, would be racy */
7970 ret = (attached && !sqd->thread) ? -ENXIO : 0;
7971 io_sq_thread_unpark(sqd);
7978 if (p->flags & IORING_SETUP_SQ_AFF) {
7979 int cpu = p->sq_thread_cpu;
7982 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
7989 sqd->task_pid = current->pid;
7990 sqd->task_tgid = current->tgid;
7991 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7998 ret = io_uring_alloc_task_context(tsk, ctx);
7999 wake_up_new_task(tsk);
8002 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8003 /* Can't have SQ_AFF without SQPOLL */
8010 complete(&ctx->sq_data->exited);
8012 io_sq_thread_finish(ctx);
8016 static inline void __io_unaccount_mem(struct user_struct *user,
8017 unsigned long nr_pages)
8019 atomic_long_sub(nr_pages, &user->locked_vm);
8022 static inline int __io_account_mem(struct user_struct *user,
8023 unsigned long nr_pages)
8025 unsigned long page_limit, cur_pages, new_pages;
8027 /* Don't allow more pages than we can safely lock */
8028 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8031 cur_pages = atomic_long_read(&user->locked_vm);
8032 new_pages = cur_pages + nr_pages;
8033 if (new_pages > page_limit)
8035 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8036 new_pages) != cur_pages);
8041 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8044 __io_unaccount_mem(ctx->user, nr_pages);
8046 if (ctx->mm_account)
8047 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8050 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8055 ret = __io_account_mem(ctx->user, nr_pages);
8060 if (ctx->mm_account)
8061 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8066 static void io_mem_free(void *ptr)
8073 page = virt_to_head_page(ptr);
8074 if (put_page_testzero(page))
8075 free_compound_page(page);
8078 static void *io_mem_alloc(size_t size)
8080 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8081 __GFP_NORETRY | __GFP_ACCOUNT;
8083 return (void *) __get_free_pages(gfp_flags, get_order(size));
8086 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8089 struct io_rings *rings;
8090 size_t off, sq_array_size;
8092 off = struct_size(rings, cqes, cq_entries);
8093 if (off == SIZE_MAX)
8097 off = ALIGN(off, SMP_CACHE_BYTES);
8105 sq_array_size = array_size(sizeof(u32), sq_entries);
8106 if (sq_array_size == SIZE_MAX)
8109 if (check_add_overflow(off, sq_array_size, &off))
8115 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8117 struct io_mapped_ubuf *imu = *slot;
8120 if (imu != ctx->dummy_ubuf) {
8121 for (i = 0; i < imu->nr_bvecs; i++)
8122 unpin_user_page(imu->bvec[i].bv_page);
8123 if (imu->acct_pages)
8124 io_unaccount_mem(ctx, imu->acct_pages);
8130 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8132 io_buffer_unmap(ctx, &prsrc->buf);
8136 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8140 for (i = 0; i < ctx->nr_user_bufs; i++)
8141 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8142 kfree(ctx->user_bufs);
8143 kfree(ctx->buf_data);
8144 ctx->user_bufs = NULL;
8145 ctx->buf_data = NULL;
8146 ctx->nr_user_bufs = 0;
8149 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8156 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8158 __io_sqe_buffers_unregister(ctx);
8162 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8163 void __user *arg, unsigned index)
8165 struct iovec __user *src;
8167 #ifdef CONFIG_COMPAT
8169 struct compat_iovec __user *ciovs;
8170 struct compat_iovec ciov;
8172 ciovs = (struct compat_iovec __user *) arg;
8173 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8176 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8177 dst->iov_len = ciov.iov_len;
8181 src = (struct iovec __user *) arg;
8182 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8188 * Not super efficient, but this is just a registration time. And we do cache
8189 * the last compound head, so generally we'll only do a full search if we don't
8192 * We check if the given compound head page has already been accounted, to
8193 * avoid double accounting it. This allows us to account the full size of the
8194 * page, not just the constituent pages of a huge page.
8196 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8197 int nr_pages, struct page *hpage)
8201 /* check current page array */
8202 for (i = 0; i < nr_pages; i++) {
8203 if (!PageCompound(pages[i]))
8205 if (compound_head(pages[i]) == hpage)
8209 /* check previously registered pages */
8210 for (i = 0; i < ctx->nr_user_bufs; i++) {
8211 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8213 for (j = 0; j < imu->nr_bvecs; j++) {
8214 if (!PageCompound(imu->bvec[j].bv_page))
8216 if (compound_head(imu->bvec[j].bv_page) == hpage)
8224 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8225 int nr_pages, struct io_mapped_ubuf *imu,
8226 struct page **last_hpage)
8230 for (i = 0; i < nr_pages; i++) {
8231 if (!PageCompound(pages[i])) {
8236 hpage = compound_head(pages[i]);
8237 if (hpage == *last_hpage)
8239 *last_hpage = hpage;
8240 if (headpage_already_acct(ctx, pages, i, hpage))
8242 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8246 if (!imu->acct_pages)
8249 ret = io_account_mem(ctx, imu->acct_pages);
8251 imu->acct_pages = 0;
8255 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8256 struct io_mapped_ubuf **pimu,
8257 struct page **last_hpage)
8259 struct io_mapped_ubuf *imu = NULL;
8260 struct vm_area_struct **vmas = NULL;
8261 struct page **pages = NULL;
8262 unsigned long off, start, end, ubuf;
8264 int ret, pret, nr_pages, i;
8266 if (!iov->iov_base) {
8267 *pimu = ctx->dummy_ubuf;
8271 ubuf = (unsigned long) iov->iov_base;
8272 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8273 start = ubuf >> PAGE_SHIFT;
8274 nr_pages = end - start;
8279 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8283 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8288 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8293 mmap_read_lock(current->mm);
8294 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8296 if (pret == nr_pages) {
8297 /* don't support file backed memory */
8298 for (i = 0; i < nr_pages; i++) {
8299 struct vm_area_struct *vma = vmas[i];
8302 !is_file_hugepages(vma->vm_file)) {
8308 ret = pret < 0 ? pret : -EFAULT;
8310 mmap_read_unlock(current->mm);
8313 * if we did partial map, or found file backed vmas,
8314 * release any pages we did get
8317 unpin_user_pages(pages, pret);
8321 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8323 unpin_user_pages(pages, pret);
8327 off = ubuf & ~PAGE_MASK;
8328 size = iov->iov_len;
8329 for (i = 0; i < nr_pages; i++) {
8332 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8333 imu->bvec[i].bv_page = pages[i];
8334 imu->bvec[i].bv_len = vec_len;
8335 imu->bvec[i].bv_offset = off;
8339 /* store original address for later verification */
8341 imu->ubuf_end = ubuf + iov->iov_len;
8342 imu->nr_bvecs = nr_pages;
8353 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8355 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8356 return ctx->user_bufs ? 0 : -ENOMEM;
8359 static int io_buffer_validate(struct iovec *iov)
8361 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8364 * Don't impose further limits on the size and buffer
8365 * constraints here, we'll -EINVAL later when IO is
8366 * submitted if they are wrong.
8369 return iov->iov_len ? -EFAULT : 0;
8373 /* arbitrary limit, but we need something */
8374 if (iov->iov_len > SZ_1G)
8377 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8383 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8384 unsigned int nr_args, u64 __user *tags)
8386 struct page *last_hpage = NULL;
8387 struct io_rsrc_data *data;
8393 if (!nr_args || nr_args > UIO_MAXIOV)
8395 ret = io_rsrc_node_switch_start(ctx);
8398 data = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, nr_args);
8401 ret = io_buffers_map_alloc(ctx, nr_args);
8407 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8410 if (tags && copy_from_user(&tag, &tags[i], sizeof(tag))) {
8414 ret = io_copy_iov(ctx, &iov, arg, i);
8417 ret = io_buffer_validate(&iov);
8420 if (!iov.iov_base && tag)
8423 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8427 data->tags[i] = tag;
8430 WARN_ON_ONCE(ctx->buf_data);
8432 ctx->buf_data = data;
8434 __io_sqe_buffers_unregister(ctx);
8436 io_rsrc_node_switch(ctx, NULL);
8440 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8441 struct io_uring_rsrc_update2 *up,
8442 unsigned int nr_args)
8444 u64 __user *tags = u64_to_user_ptr(up->tags);
8445 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8446 struct page *last_hpage = NULL;
8447 bool needs_switch = false;
8453 if (up->offset + nr_args > ctx->nr_user_bufs)
8456 for (done = 0; done < nr_args; done++) {
8457 struct io_mapped_ubuf *imu;
8458 int offset = up->offset + done;
8461 err = io_copy_iov(ctx, &iov, iovs, done);
8464 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8468 err = io_buffer_validate(&iov);
8471 if (!iov.iov_base && tag)
8473 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8477 i = array_index_nospec(offset, ctx->nr_user_bufs);
8478 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8479 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8480 ctx->rsrc_node, ctx->user_bufs[i]);
8481 if (unlikely(err)) {
8482 io_buffer_unmap(ctx, &imu);
8485 ctx->user_bufs[i] = NULL;
8486 needs_switch = true;
8489 ctx->user_bufs[i] = imu;
8490 ctx->buf_data->tags[offset] = tag;
8494 io_rsrc_node_switch(ctx, ctx->buf_data);
8495 return done ? done : err;
8498 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8500 __s32 __user *fds = arg;
8506 if (copy_from_user(&fd, fds, sizeof(*fds)))
8509 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8510 if (IS_ERR(ctx->cq_ev_fd)) {
8511 int ret = PTR_ERR(ctx->cq_ev_fd);
8512 ctx->cq_ev_fd = NULL;
8519 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8521 if (ctx->cq_ev_fd) {
8522 eventfd_ctx_put(ctx->cq_ev_fd);
8523 ctx->cq_ev_fd = NULL;
8530 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8532 struct io_buffer *buf;
8533 unsigned long index;
8535 xa_for_each(&ctx->io_buffers, index, buf)
8536 __io_remove_buffers(ctx, buf, index, -1U);
8539 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8541 struct io_kiocb *req, *nxt;
8543 list_for_each_entry_safe(req, nxt, list, compl.list) {
8544 if (tsk && req->task != tsk)
8546 list_del(&req->compl.list);
8547 kmem_cache_free(req_cachep, req);
8551 static void io_req_caches_free(struct io_ring_ctx *ctx)
8553 struct io_submit_state *submit_state = &ctx->submit_state;
8554 struct io_comp_state *cs = &ctx->submit_state.comp;
8556 mutex_lock(&ctx->uring_lock);
8558 if (submit_state->free_reqs) {
8559 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8560 submit_state->reqs);
8561 submit_state->free_reqs = 0;
8564 io_flush_cached_locked_reqs(ctx, cs);
8565 io_req_cache_free(&cs->free_list, NULL);
8566 mutex_unlock(&ctx->uring_lock);
8569 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8573 if (!atomic_dec_and_test(&data->refs))
8574 wait_for_completion(&data->done);
8578 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8580 io_sq_thread_finish(ctx);
8582 if (ctx->mm_account) {
8583 mmdrop(ctx->mm_account);
8584 ctx->mm_account = NULL;
8587 mutex_lock(&ctx->uring_lock);
8588 if (io_wait_rsrc_data(ctx->buf_data))
8589 __io_sqe_buffers_unregister(ctx);
8590 if (io_wait_rsrc_data(ctx->file_data))
8591 __io_sqe_files_unregister(ctx);
8593 __io_cqring_overflow_flush(ctx, true);
8594 mutex_unlock(&ctx->uring_lock);
8595 io_eventfd_unregister(ctx);
8596 io_destroy_buffers(ctx);
8598 put_cred(ctx->sq_creds);
8600 /* there are no registered resources left, nobody uses it */
8602 io_rsrc_node_destroy(ctx->rsrc_node);
8603 if (ctx->rsrc_backup_node)
8604 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8605 flush_delayed_work(&ctx->rsrc_put_work);
8607 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8608 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8610 #if defined(CONFIG_UNIX)
8611 if (ctx->ring_sock) {
8612 ctx->ring_sock->file = NULL; /* so that iput() is called */
8613 sock_release(ctx->ring_sock);
8617 io_mem_free(ctx->rings);
8618 io_mem_free(ctx->sq_sqes);
8620 percpu_ref_exit(&ctx->refs);
8621 free_uid(ctx->user);
8622 io_req_caches_free(ctx);
8624 io_wq_put_hash(ctx->hash_map);
8625 kfree(ctx->cancel_hash);
8626 kfree(ctx->dummy_ubuf);
8630 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8632 struct io_ring_ctx *ctx = file->private_data;
8635 poll_wait(file, &ctx->cq_wait, wait);
8637 * synchronizes with barrier from wq_has_sleeper call in
8641 if (!io_sqring_full(ctx))
8642 mask |= EPOLLOUT | EPOLLWRNORM;
8645 * Don't flush cqring overflow list here, just do a simple check.
8646 * Otherwise there could possible be ABBA deadlock:
8649 * lock(&ctx->uring_lock);
8651 * lock(&ctx->uring_lock);
8654 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8655 * pushs them to do the flush.
8657 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8658 mask |= EPOLLIN | EPOLLRDNORM;
8663 static int io_uring_fasync(int fd, struct file *file, int on)
8665 struct io_ring_ctx *ctx = file->private_data;
8667 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8670 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8672 const struct cred *creds;
8674 creds = xa_erase(&ctx->personalities, id);
8683 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8685 return io_run_task_work_head(&ctx->exit_task_work);
8688 struct io_tctx_exit {
8689 struct callback_head task_work;
8690 struct completion completion;
8691 struct io_ring_ctx *ctx;
8694 static void io_tctx_exit_cb(struct callback_head *cb)
8696 struct io_uring_task *tctx = current->io_uring;
8697 struct io_tctx_exit *work;
8699 work = container_of(cb, struct io_tctx_exit, task_work);
8701 * When @in_idle, we're in cancellation and it's racy to remove the
8702 * node. It'll be removed by the end of cancellation, just ignore it.
8704 if (!atomic_read(&tctx->in_idle))
8705 io_uring_del_task_file((unsigned long)work->ctx);
8706 complete(&work->completion);
8709 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8711 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8713 return req->ctx == data;
8716 static void io_ring_exit_work(struct work_struct *work)
8718 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8719 unsigned long timeout = jiffies + HZ * 60 * 5;
8720 struct io_tctx_exit exit;
8721 struct io_tctx_node *node;
8725 * If we're doing polled IO and end up having requests being
8726 * submitted async (out-of-line), then completions can come in while
8727 * we're waiting for refs to drop. We need to reap these manually,
8728 * as nobody else will be looking for them.
8731 io_uring_try_cancel_requests(ctx, NULL, NULL);
8733 struct io_sq_data *sqd = ctx->sq_data;
8734 struct task_struct *tsk;
8736 io_sq_thread_park(sqd);
8738 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8739 io_wq_cancel_cb(tsk->io_uring->io_wq,
8740 io_cancel_ctx_cb, ctx, true);
8741 io_sq_thread_unpark(sqd);
8744 WARN_ON_ONCE(time_after(jiffies, timeout));
8745 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8747 init_completion(&exit.completion);
8748 init_task_work(&exit.task_work, io_tctx_exit_cb);
8751 * Some may use context even when all refs and requests have been put,
8752 * and they are free to do so while still holding uring_lock or
8753 * completion_lock, see __io_req_task_submit(). Apart from other work,
8754 * this lock/unlock section also waits them to finish.
8756 mutex_lock(&ctx->uring_lock);
8757 while (!list_empty(&ctx->tctx_list)) {
8758 WARN_ON_ONCE(time_after(jiffies, timeout));
8760 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8762 /* don't spin on a single task if cancellation failed */
8763 list_rotate_left(&ctx->tctx_list);
8764 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8765 if (WARN_ON_ONCE(ret))
8767 wake_up_process(node->task);
8769 mutex_unlock(&ctx->uring_lock);
8770 wait_for_completion(&exit.completion);
8771 mutex_lock(&ctx->uring_lock);
8773 mutex_unlock(&ctx->uring_lock);
8774 spin_lock_irq(&ctx->completion_lock);
8775 spin_unlock_irq(&ctx->completion_lock);
8777 io_ring_ctx_free(ctx);
8780 /* Returns true if we found and killed one or more timeouts */
8781 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8782 struct files_struct *files)
8784 struct io_kiocb *req, *tmp;
8787 spin_lock_irq(&ctx->completion_lock);
8788 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8789 if (io_match_task(req, tsk, files)) {
8790 io_kill_timeout(req, -ECANCELED);
8795 io_commit_cqring(ctx);
8796 spin_unlock_irq(&ctx->completion_lock);
8798 io_cqring_ev_posted(ctx);
8799 return canceled != 0;
8802 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8804 unsigned long index;
8805 struct creds *creds;
8807 mutex_lock(&ctx->uring_lock);
8808 percpu_ref_kill(&ctx->refs);
8810 __io_cqring_overflow_flush(ctx, true);
8811 xa_for_each(&ctx->personalities, index, creds)
8812 io_unregister_personality(ctx, index);
8813 mutex_unlock(&ctx->uring_lock);
8815 io_kill_timeouts(ctx, NULL, NULL);
8816 io_poll_remove_all(ctx, NULL, NULL);
8818 /* if we failed setting up the ctx, we might not have any rings */
8819 io_iopoll_try_reap_events(ctx);
8821 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8823 * Use system_unbound_wq to avoid spawning tons of event kworkers
8824 * if we're exiting a ton of rings at the same time. It just adds
8825 * noise and overhead, there's no discernable change in runtime
8826 * over using system_wq.
8828 queue_work(system_unbound_wq, &ctx->exit_work);
8831 static int io_uring_release(struct inode *inode, struct file *file)
8833 struct io_ring_ctx *ctx = file->private_data;
8835 file->private_data = NULL;
8836 io_ring_ctx_wait_and_kill(ctx);
8840 struct io_task_cancel {
8841 struct task_struct *task;
8842 struct files_struct *files;
8845 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8847 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8848 struct io_task_cancel *cancel = data;
8851 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8852 unsigned long flags;
8853 struct io_ring_ctx *ctx = req->ctx;
8855 /* protect against races with linked timeouts */
8856 spin_lock_irqsave(&ctx->completion_lock, flags);
8857 ret = io_match_task(req, cancel->task, cancel->files);
8858 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8860 ret = io_match_task(req, cancel->task, cancel->files);
8865 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8866 struct task_struct *task,
8867 struct files_struct *files)
8869 struct io_defer_entry *de;
8872 spin_lock_irq(&ctx->completion_lock);
8873 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8874 if (io_match_task(de->req, task, files)) {
8875 list_cut_position(&list, &ctx->defer_list, &de->list);
8879 spin_unlock_irq(&ctx->completion_lock);
8880 if (list_empty(&list))
8883 while (!list_empty(&list)) {
8884 de = list_first_entry(&list, struct io_defer_entry, list);
8885 list_del_init(&de->list);
8886 io_req_complete_failed(de->req, -ECANCELED);
8892 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8894 struct io_tctx_node *node;
8895 enum io_wq_cancel cret;
8898 mutex_lock(&ctx->uring_lock);
8899 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8900 struct io_uring_task *tctx = node->task->io_uring;
8903 * io_wq will stay alive while we hold uring_lock, because it's
8904 * killed after ctx nodes, which requires to take the lock.
8906 if (!tctx || !tctx->io_wq)
8908 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8909 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8911 mutex_unlock(&ctx->uring_lock);
8916 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8917 struct task_struct *task,
8918 struct files_struct *files)
8920 struct io_task_cancel cancel = { .task = task, .files = files, };
8921 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8924 enum io_wq_cancel cret;
8928 ret |= io_uring_try_cancel_iowq(ctx);
8929 } else if (tctx && tctx->io_wq) {
8931 * Cancels requests of all rings, not only @ctx, but
8932 * it's fine as the task is in exit/exec.
8934 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8936 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8939 /* SQPOLL thread does its own polling */
8940 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8941 (ctx->sq_data && ctx->sq_data->thread == current)) {
8942 while (!list_empty_careful(&ctx->iopoll_list)) {
8943 io_iopoll_try_reap_events(ctx);
8948 ret |= io_cancel_defer_files(ctx, task, files);
8949 ret |= io_poll_remove_all(ctx, task, files);
8950 ret |= io_kill_timeouts(ctx, task, files);
8951 ret |= io_run_task_work();
8952 ret |= io_run_ctx_fallback(ctx);
8959 static int __io_uring_add_task_file(struct io_ring_ctx *ctx)
8961 struct io_uring_task *tctx = current->io_uring;
8962 struct io_tctx_node *node;
8965 if (unlikely(!tctx)) {
8966 ret = io_uring_alloc_task_context(current, ctx);
8969 tctx = current->io_uring;
8971 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
8972 node = kmalloc(sizeof(*node), GFP_KERNEL);
8976 node->task = current;
8978 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8985 mutex_lock(&ctx->uring_lock);
8986 list_add(&node->ctx_node, &ctx->tctx_list);
8987 mutex_unlock(&ctx->uring_lock);
8994 * Note that this task has used io_uring. We use it for cancelation purposes.
8996 static inline int io_uring_add_task_file(struct io_ring_ctx *ctx)
8998 struct io_uring_task *tctx = current->io_uring;
9000 if (likely(tctx && tctx->last == ctx))
9002 return __io_uring_add_task_file(ctx);
9006 * Remove this io_uring_file -> task mapping.
9008 static void io_uring_del_task_file(unsigned long index)
9010 struct io_uring_task *tctx = current->io_uring;
9011 struct io_tctx_node *node;
9015 node = xa_erase(&tctx->xa, index);
9019 WARN_ON_ONCE(current != node->task);
9020 WARN_ON_ONCE(list_empty(&node->ctx_node));
9022 mutex_lock(&node->ctx->uring_lock);
9023 list_del(&node->ctx_node);
9024 mutex_unlock(&node->ctx->uring_lock);
9026 if (tctx->last == node->ctx)
9031 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9033 struct io_tctx_node *node;
9034 unsigned long index;
9036 xa_for_each(&tctx->xa, index, node)
9037 io_uring_del_task_file(index);
9039 io_wq_put_and_exit(tctx->io_wq);
9044 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9047 return atomic_read(&tctx->inflight_tracked);
9048 return percpu_counter_sum(&tctx->inflight);
9051 static void io_uring_try_cancel(struct files_struct *files)
9053 struct io_uring_task *tctx = current->io_uring;
9054 struct io_tctx_node *node;
9055 unsigned long index;
9057 xa_for_each(&tctx->xa, index, node) {
9058 struct io_ring_ctx *ctx = node->ctx;
9060 /* sqpoll task will cancel all its requests */
9062 io_uring_try_cancel_requests(ctx, current, files);
9066 /* should only be called by SQPOLL task */
9067 static void io_uring_cancel_sqpoll(struct io_sq_data *sqd)
9069 struct io_uring_task *tctx = current->io_uring;
9070 struct io_ring_ctx *ctx;
9074 if (!current->io_uring)
9076 WARN_ON_ONCE(!sqd || sqd->thread != current);
9078 atomic_inc(&tctx->in_idle);
9080 /* read completions before cancelations */
9081 inflight = tctx_inflight(tctx, false);
9084 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9085 io_uring_try_cancel_requests(ctx, current, NULL);
9087 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9089 * If we've seen completions, retry without waiting. This
9090 * avoids a race where a completion comes in before we did
9091 * prepare_to_wait().
9093 if (inflight == tctx_inflight(tctx, false))
9095 finish_wait(&tctx->wait, &wait);
9097 atomic_dec(&tctx->in_idle);
9101 * Find any io_uring fd that this task has registered or done IO on, and cancel
9104 void __io_uring_cancel(struct files_struct *files)
9106 struct io_uring_task *tctx = current->io_uring;
9110 /* make sure overflow events are dropped */
9111 atomic_inc(&tctx->in_idle);
9113 /* read completions before cancelations */
9114 inflight = tctx_inflight(tctx, !!files);
9117 io_uring_try_cancel(files);
9118 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9121 * If we've seen completions, retry without waiting. This
9122 * avoids a race where a completion comes in before we did
9123 * prepare_to_wait().
9125 if (inflight == tctx_inflight(tctx, !!files))
9127 finish_wait(&tctx->wait, &wait);
9129 atomic_dec(&tctx->in_idle);
9131 io_uring_clean_tctx(tctx);
9133 /* for exec all current's requests should be gone, kill tctx */
9134 __io_uring_free(current);
9138 static void *io_uring_validate_mmap_request(struct file *file,
9139 loff_t pgoff, size_t sz)
9141 struct io_ring_ctx *ctx = file->private_data;
9142 loff_t offset = pgoff << PAGE_SHIFT;
9147 case IORING_OFF_SQ_RING:
9148 case IORING_OFF_CQ_RING:
9151 case IORING_OFF_SQES:
9155 return ERR_PTR(-EINVAL);
9158 page = virt_to_head_page(ptr);
9159 if (sz > page_size(page))
9160 return ERR_PTR(-EINVAL);
9167 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9169 size_t sz = vma->vm_end - vma->vm_start;
9173 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9175 return PTR_ERR(ptr);
9177 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9178 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9181 #else /* !CONFIG_MMU */
9183 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9185 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9188 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9190 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9193 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9194 unsigned long addr, unsigned long len,
9195 unsigned long pgoff, unsigned long flags)
9199 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9201 return PTR_ERR(ptr);
9203 return (unsigned long) ptr;
9206 #endif /* !CONFIG_MMU */
9208 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9213 if (!io_sqring_full(ctx))
9215 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9217 if (!io_sqring_full(ctx))
9220 } while (!signal_pending(current));
9222 finish_wait(&ctx->sqo_sq_wait, &wait);
9226 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9227 struct __kernel_timespec __user **ts,
9228 const sigset_t __user **sig)
9230 struct io_uring_getevents_arg arg;
9233 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9234 * is just a pointer to the sigset_t.
9236 if (!(flags & IORING_ENTER_EXT_ARG)) {
9237 *sig = (const sigset_t __user *) argp;
9243 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9244 * timespec and sigset_t pointers if good.
9246 if (*argsz != sizeof(arg))
9248 if (copy_from_user(&arg, argp, sizeof(arg)))
9250 *sig = u64_to_user_ptr(arg.sigmask);
9251 *argsz = arg.sigmask_sz;
9252 *ts = u64_to_user_ptr(arg.ts);
9256 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9257 u32, min_complete, u32, flags, const void __user *, argp,
9260 struct io_ring_ctx *ctx;
9267 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9268 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9272 if (unlikely(!f.file))
9276 if (unlikely(f.file->f_op != &io_uring_fops))
9280 ctx = f.file->private_data;
9281 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9285 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9289 * For SQ polling, the thread will do all submissions and completions.
9290 * Just return the requested submit count, and wake the thread if
9294 if (ctx->flags & IORING_SETUP_SQPOLL) {
9295 io_cqring_overflow_flush(ctx, false);
9298 if (unlikely(ctx->sq_data->thread == NULL)) {
9301 if (flags & IORING_ENTER_SQ_WAKEUP)
9302 wake_up(&ctx->sq_data->wait);
9303 if (flags & IORING_ENTER_SQ_WAIT) {
9304 ret = io_sqpoll_wait_sq(ctx);
9308 submitted = to_submit;
9309 } else if (to_submit) {
9310 ret = io_uring_add_task_file(ctx);
9313 mutex_lock(&ctx->uring_lock);
9314 submitted = io_submit_sqes(ctx, to_submit);
9315 mutex_unlock(&ctx->uring_lock);
9317 if (submitted != to_submit)
9320 if (flags & IORING_ENTER_GETEVENTS) {
9321 const sigset_t __user *sig;
9322 struct __kernel_timespec __user *ts;
9324 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9328 min_complete = min(min_complete, ctx->cq_entries);
9331 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9332 * space applications don't need to do io completion events
9333 * polling again, they can rely on io_sq_thread to do polling
9334 * work, which can reduce cpu usage and uring_lock contention.
9336 if (ctx->flags & IORING_SETUP_IOPOLL &&
9337 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9338 ret = io_iopoll_check(ctx, min_complete);
9340 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9345 percpu_ref_put(&ctx->refs);
9348 return submitted ? submitted : ret;
9351 #ifdef CONFIG_PROC_FS
9352 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9353 const struct cred *cred)
9355 struct user_namespace *uns = seq_user_ns(m);
9356 struct group_info *gi;
9361 seq_printf(m, "%5d\n", id);
9362 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9363 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9364 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9365 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9366 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9367 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9368 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9369 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9370 seq_puts(m, "\n\tGroups:\t");
9371 gi = cred->group_info;
9372 for (g = 0; g < gi->ngroups; g++) {
9373 seq_put_decimal_ull(m, g ? " " : "",
9374 from_kgid_munged(uns, gi->gid[g]));
9376 seq_puts(m, "\n\tCapEff:\t");
9377 cap = cred->cap_effective;
9378 CAP_FOR_EACH_U32(__capi)
9379 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9384 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9386 struct io_sq_data *sq = NULL;
9391 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9392 * since fdinfo case grabs it in the opposite direction of normal use
9393 * cases. If we fail to get the lock, we just don't iterate any
9394 * structures that could be going away outside the io_uring mutex.
9396 has_lock = mutex_trylock(&ctx->uring_lock);
9398 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9404 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9405 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9406 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9407 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9408 struct file *f = io_file_from_index(ctx, i);
9411 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9413 seq_printf(m, "%5u: <none>\n", i);
9415 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9416 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9417 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9418 unsigned int len = buf->ubuf_end - buf->ubuf;
9420 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9422 if (has_lock && !xa_empty(&ctx->personalities)) {
9423 unsigned long index;
9424 const struct cred *cred;
9426 seq_printf(m, "Personalities:\n");
9427 xa_for_each(&ctx->personalities, index, cred)
9428 io_uring_show_cred(m, index, cred);
9430 seq_printf(m, "PollList:\n");
9431 spin_lock_irq(&ctx->completion_lock);
9432 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9433 struct hlist_head *list = &ctx->cancel_hash[i];
9434 struct io_kiocb *req;
9436 hlist_for_each_entry(req, list, hash_node)
9437 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9438 req->task->task_works != NULL);
9440 spin_unlock_irq(&ctx->completion_lock);
9442 mutex_unlock(&ctx->uring_lock);
9445 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9447 struct io_ring_ctx *ctx = f->private_data;
9449 if (percpu_ref_tryget(&ctx->refs)) {
9450 __io_uring_show_fdinfo(ctx, m);
9451 percpu_ref_put(&ctx->refs);
9456 static const struct file_operations io_uring_fops = {
9457 .release = io_uring_release,
9458 .mmap = io_uring_mmap,
9460 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9461 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9463 .poll = io_uring_poll,
9464 .fasync = io_uring_fasync,
9465 #ifdef CONFIG_PROC_FS
9466 .show_fdinfo = io_uring_show_fdinfo,
9470 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9471 struct io_uring_params *p)
9473 struct io_rings *rings;
9474 size_t size, sq_array_offset;
9476 /* make sure these are sane, as we already accounted them */
9477 ctx->sq_entries = p->sq_entries;
9478 ctx->cq_entries = p->cq_entries;
9480 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9481 if (size == SIZE_MAX)
9484 rings = io_mem_alloc(size);
9489 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9490 rings->sq_ring_mask = p->sq_entries - 1;
9491 rings->cq_ring_mask = p->cq_entries - 1;
9492 rings->sq_ring_entries = p->sq_entries;
9493 rings->cq_ring_entries = p->cq_entries;
9494 ctx->sq_mask = rings->sq_ring_mask;
9495 ctx->cq_mask = rings->cq_ring_mask;
9497 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9498 if (size == SIZE_MAX) {
9499 io_mem_free(ctx->rings);
9504 ctx->sq_sqes = io_mem_alloc(size);
9505 if (!ctx->sq_sqes) {
9506 io_mem_free(ctx->rings);
9514 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9518 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9522 ret = io_uring_add_task_file(ctx);
9527 fd_install(fd, file);
9532 * Allocate an anonymous fd, this is what constitutes the application
9533 * visible backing of an io_uring instance. The application mmaps this
9534 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9535 * we have to tie this fd to a socket for file garbage collection purposes.
9537 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9540 #if defined(CONFIG_UNIX)
9543 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9546 return ERR_PTR(ret);
9549 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9550 O_RDWR | O_CLOEXEC);
9551 #if defined(CONFIG_UNIX)
9553 sock_release(ctx->ring_sock);
9554 ctx->ring_sock = NULL;
9556 ctx->ring_sock->file = file;
9562 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9563 struct io_uring_params __user *params)
9565 struct io_ring_ctx *ctx;
9571 if (entries > IORING_MAX_ENTRIES) {
9572 if (!(p->flags & IORING_SETUP_CLAMP))
9574 entries = IORING_MAX_ENTRIES;
9578 * Use twice as many entries for the CQ ring. It's possible for the
9579 * application to drive a higher depth than the size of the SQ ring,
9580 * since the sqes are only used at submission time. This allows for
9581 * some flexibility in overcommitting a bit. If the application has
9582 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9583 * of CQ ring entries manually.
9585 p->sq_entries = roundup_pow_of_two(entries);
9586 if (p->flags & IORING_SETUP_CQSIZE) {
9588 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9589 * to a power-of-two, if it isn't already. We do NOT impose
9590 * any cq vs sq ring sizing.
9594 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9595 if (!(p->flags & IORING_SETUP_CLAMP))
9597 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9599 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9600 if (p->cq_entries < p->sq_entries)
9603 p->cq_entries = 2 * p->sq_entries;
9606 ctx = io_ring_ctx_alloc(p);
9609 ctx->compat = in_compat_syscall();
9610 if (!capable(CAP_IPC_LOCK))
9611 ctx->user = get_uid(current_user());
9614 * This is just grabbed for accounting purposes. When a process exits,
9615 * the mm is exited and dropped before the files, hence we need to hang
9616 * on to this mm purely for the purposes of being able to unaccount
9617 * memory (locked/pinned vm). It's not used for anything else.
9619 mmgrab(current->mm);
9620 ctx->mm_account = current->mm;
9622 ret = io_allocate_scq_urings(ctx, p);
9626 ret = io_sq_offload_create(ctx, p);
9629 /* always set a rsrc node */
9630 ret = io_rsrc_node_switch_start(ctx);
9633 io_rsrc_node_switch(ctx, NULL);
9635 memset(&p->sq_off, 0, sizeof(p->sq_off));
9636 p->sq_off.head = offsetof(struct io_rings, sq.head);
9637 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9638 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9639 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9640 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9641 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9642 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9644 memset(&p->cq_off, 0, sizeof(p->cq_off));
9645 p->cq_off.head = offsetof(struct io_rings, cq.head);
9646 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9647 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9648 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9649 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9650 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9651 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9653 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9654 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9655 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9656 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9657 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9659 if (copy_to_user(params, p, sizeof(*p))) {
9664 file = io_uring_get_file(ctx);
9666 ret = PTR_ERR(file);
9671 * Install ring fd as the very last thing, so we don't risk someone
9672 * having closed it before we finish setup
9674 ret = io_uring_install_fd(ctx, file);
9676 /* fput will clean it up */
9681 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9684 io_ring_ctx_wait_and_kill(ctx);
9689 * Sets up an aio uring context, and returns the fd. Applications asks for a
9690 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9691 * params structure passed in.
9693 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9695 struct io_uring_params p;
9698 if (copy_from_user(&p, params, sizeof(p)))
9700 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9705 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9706 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9707 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9708 IORING_SETUP_R_DISABLED))
9711 return io_uring_create(entries, &p, params);
9714 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9715 struct io_uring_params __user *, params)
9717 return io_uring_setup(entries, params);
9720 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9722 struct io_uring_probe *p;
9726 size = struct_size(p, ops, nr_args);
9727 if (size == SIZE_MAX)
9729 p = kzalloc(size, GFP_KERNEL);
9734 if (copy_from_user(p, arg, size))
9737 if (memchr_inv(p, 0, size))
9740 p->last_op = IORING_OP_LAST - 1;
9741 if (nr_args > IORING_OP_LAST)
9742 nr_args = IORING_OP_LAST;
9744 for (i = 0; i < nr_args; i++) {
9746 if (!io_op_defs[i].not_supported)
9747 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9752 if (copy_to_user(arg, p, size))
9759 static int io_register_personality(struct io_ring_ctx *ctx)
9761 const struct cred *creds;
9765 creds = get_current_cred();
9767 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9768 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9775 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9776 unsigned int nr_args)
9778 struct io_uring_restriction *res;
9782 /* Restrictions allowed only if rings started disabled */
9783 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9786 /* We allow only a single restrictions registration */
9787 if (ctx->restrictions.registered)
9790 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9793 size = array_size(nr_args, sizeof(*res));
9794 if (size == SIZE_MAX)
9797 res = memdup_user(arg, size);
9799 return PTR_ERR(res);
9803 for (i = 0; i < nr_args; i++) {
9804 switch (res[i].opcode) {
9805 case IORING_RESTRICTION_REGISTER_OP:
9806 if (res[i].register_op >= IORING_REGISTER_LAST) {
9811 __set_bit(res[i].register_op,
9812 ctx->restrictions.register_op);
9814 case IORING_RESTRICTION_SQE_OP:
9815 if (res[i].sqe_op >= IORING_OP_LAST) {
9820 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9822 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9823 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9825 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9826 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9835 /* Reset all restrictions if an error happened */
9837 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9839 ctx->restrictions.registered = true;
9845 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9847 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9850 if (ctx->restrictions.registered)
9851 ctx->restricted = 1;
9853 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9854 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9855 wake_up(&ctx->sq_data->wait);
9859 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9860 struct io_uring_rsrc_update2 *up,
9868 if (check_add_overflow(up->offset, nr_args, &tmp))
9870 err = io_rsrc_node_switch_start(ctx);
9875 case IORING_RSRC_FILE:
9876 return __io_sqe_files_update(ctx, up, nr_args);
9877 case IORING_RSRC_BUFFER:
9878 return __io_sqe_buffers_update(ctx, up, nr_args);
9883 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9886 struct io_uring_rsrc_update2 up;
9890 memset(&up, 0, sizeof(up));
9891 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9893 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9896 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9899 struct io_uring_rsrc_update2 up;
9901 if (size != sizeof(up))
9903 if (copy_from_user(&up, arg, sizeof(up)))
9907 return __io_register_rsrc_update(ctx, up.type, &up, up.nr);
9910 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9913 struct io_uring_rsrc_register rr;
9915 /* keep it extendible */
9916 if (size != sizeof(rr))
9919 memset(&rr, 0, sizeof(rr));
9920 if (copy_from_user(&rr, arg, size))
9926 case IORING_RSRC_FILE:
9927 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9928 rr.nr, u64_to_user_ptr(rr.tags));
9929 case IORING_RSRC_BUFFER:
9930 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9931 rr.nr, u64_to_user_ptr(rr.tags));
9936 static bool io_register_op_must_quiesce(int op)
9939 case IORING_REGISTER_BUFFERS:
9940 case IORING_UNREGISTER_BUFFERS:
9941 case IORING_REGISTER_FILES:
9942 case IORING_UNREGISTER_FILES:
9943 case IORING_REGISTER_FILES_UPDATE:
9944 case IORING_REGISTER_PROBE:
9945 case IORING_REGISTER_PERSONALITY:
9946 case IORING_UNREGISTER_PERSONALITY:
9947 case IORING_REGISTER_RSRC:
9948 case IORING_REGISTER_RSRC_UPDATE:
9955 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9956 void __user *arg, unsigned nr_args)
9957 __releases(ctx->uring_lock)
9958 __acquires(ctx->uring_lock)
9963 * We're inside the ring mutex, if the ref is already dying, then
9964 * someone else killed the ctx or is already going through
9965 * io_uring_register().
9967 if (percpu_ref_is_dying(&ctx->refs))
9970 if (ctx->restricted) {
9971 if (opcode >= IORING_REGISTER_LAST)
9973 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
9974 if (!test_bit(opcode, ctx->restrictions.register_op))
9978 if (io_register_op_must_quiesce(opcode)) {
9979 percpu_ref_kill(&ctx->refs);
9982 * Drop uring mutex before waiting for references to exit. If
9983 * another thread is currently inside io_uring_enter() it might
9984 * need to grab the uring_lock to make progress. If we hold it
9985 * here across the drain wait, then we can deadlock. It's safe
9986 * to drop the mutex here, since no new references will come in
9987 * after we've killed the percpu ref.
9989 mutex_unlock(&ctx->uring_lock);
9991 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9994 ret = io_run_task_work_sig();
9998 mutex_lock(&ctx->uring_lock);
10001 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10007 case IORING_REGISTER_BUFFERS:
10008 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10010 case IORING_UNREGISTER_BUFFERS:
10012 if (arg || nr_args)
10014 ret = io_sqe_buffers_unregister(ctx);
10016 case IORING_REGISTER_FILES:
10017 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10019 case IORING_UNREGISTER_FILES:
10021 if (arg || nr_args)
10023 ret = io_sqe_files_unregister(ctx);
10025 case IORING_REGISTER_FILES_UPDATE:
10026 ret = io_register_files_update(ctx, arg, nr_args);
10028 case IORING_REGISTER_EVENTFD:
10029 case IORING_REGISTER_EVENTFD_ASYNC:
10033 ret = io_eventfd_register(ctx, arg);
10036 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10037 ctx->eventfd_async = 1;
10039 ctx->eventfd_async = 0;
10041 case IORING_UNREGISTER_EVENTFD:
10043 if (arg || nr_args)
10045 ret = io_eventfd_unregister(ctx);
10047 case IORING_REGISTER_PROBE:
10049 if (!arg || nr_args > 256)
10051 ret = io_probe(ctx, arg, nr_args);
10053 case IORING_REGISTER_PERSONALITY:
10055 if (arg || nr_args)
10057 ret = io_register_personality(ctx);
10059 case IORING_UNREGISTER_PERSONALITY:
10063 ret = io_unregister_personality(ctx, nr_args);
10065 case IORING_REGISTER_ENABLE_RINGS:
10067 if (arg || nr_args)
10069 ret = io_register_enable_rings(ctx);
10071 case IORING_REGISTER_RESTRICTIONS:
10072 ret = io_register_restrictions(ctx, arg, nr_args);
10074 case IORING_REGISTER_RSRC:
10075 ret = io_register_rsrc(ctx, arg, nr_args);
10077 case IORING_REGISTER_RSRC_UPDATE:
10078 ret = io_register_rsrc_update(ctx, arg, nr_args);
10085 if (io_register_op_must_quiesce(opcode)) {
10086 /* bring the ctx back to life */
10087 percpu_ref_reinit(&ctx->refs);
10088 reinit_completion(&ctx->ref_comp);
10093 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10094 void __user *, arg, unsigned int, nr_args)
10096 struct io_ring_ctx *ctx;
10105 if (f.file->f_op != &io_uring_fops)
10108 ctx = f.file->private_data;
10110 io_run_task_work();
10112 mutex_lock(&ctx->uring_lock);
10113 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10114 mutex_unlock(&ctx->uring_lock);
10115 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10116 ctx->cq_ev_fd != NULL, ret);
10122 static int __init io_uring_init(void)
10124 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10125 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10126 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10129 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10130 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10131 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10132 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10133 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10134 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10135 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10136 BUILD_BUG_SQE_ELEM(8, __u64, off);
10137 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10138 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10139 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10140 BUILD_BUG_SQE_ELEM(24, __u32, len);
10141 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10142 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10143 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10144 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10145 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10146 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10147 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10148 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10149 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10150 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10151 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10152 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10153 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10154 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10155 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10156 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10157 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10158 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10159 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10161 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10162 sizeof(struct io_uring_rsrc_update));
10163 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10164 sizeof(struct io_uring_rsrc_update2));
10165 /* should fit into one byte */
10166 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10168 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10169 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10170 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10174 __initcall(io_uring_init);