1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81 #include <linux/tracehook.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp;
116 u32 tail ____cacheline_aligned_in_smp;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq, cq;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask, cq_ring_mask;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries, cq_ring_entries;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
197 enum io_uring_cmd_flags {
198 IO_URING_F_NONBLOCK = 1,
199 IO_URING_F_COMPLETE_DEFER = 2,
202 struct io_mapped_ubuf {
205 unsigned int nr_bvecs;
206 unsigned long acct_pages;
207 struct bio_vec bvec[];
212 struct io_overflow_cqe {
213 struct io_uring_cqe cqe;
214 struct list_head list;
217 struct io_fixed_file {
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr;
223 struct list_head list;
228 struct io_mapped_ubuf *buf;
232 struct io_file_table {
233 struct io_fixed_file *files;
236 struct io_rsrc_node {
237 struct percpu_ref refs;
238 struct list_head node;
239 struct list_head rsrc_list;
240 struct io_rsrc_data *rsrc_data;
241 struct llist_node llist;
245 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
247 struct io_rsrc_data {
248 struct io_ring_ctx *ctx;
254 struct completion done;
259 struct list_head list;
265 struct io_restriction {
266 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268 u8 sqe_flags_allowed;
269 u8 sqe_flags_required;
274 IO_SQ_THREAD_SHOULD_STOP = 0,
275 IO_SQ_THREAD_SHOULD_PARK,
280 atomic_t park_pending;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list;
286 struct task_struct *thread;
287 struct wait_queue_head wait;
289 unsigned sq_thread_idle;
295 struct completion exited;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_submit_link {
303 struct io_kiocb *head;
304 struct io_kiocb *last;
307 struct io_submit_state {
308 struct blk_plug plug;
309 struct io_submit_link link;
312 * io_kiocb alloc cache
314 void *reqs[IO_REQ_CACHE_SIZE];
315 unsigned int free_reqs;
320 * Batch completion logic
322 struct io_kiocb *compl_reqs[IO_COMPL_BATCH];
323 unsigned int compl_nr;
324 /* inline/task_work completion list, under ->uring_lock */
325 struct list_head free_list;
327 unsigned int ios_left;
331 /* const or read-mostly hot data */
333 struct percpu_ref refs;
335 struct io_rings *rings;
337 unsigned int compat: 1;
338 unsigned int drain_next: 1;
339 unsigned int eventfd_async: 1;
340 unsigned int restricted: 1;
341 unsigned int off_timeout_used: 1;
342 unsigned int drain_active: 1;
343 } ____cacheline_aligned_in_smp;
345 /* submission data */
347 struct mutex uring_lock;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 struct io_uring_sqe *sq_sqes;
362 unsigned cached_sq_head;
364 struct list_head defer_list;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node *rsrc_node;
371 struct io_file_table file_table;
372 unsigned nr_user_files;
373 unsigned nr_user_bufs;
374 struct io_mapped_ubuf **user_bufs;
376 struct io_submit_state submit_state;
377 struct list_head timeout_list;
378 struct list_head ltimeout_list;
379 struct list_head cq_overflow_list;
380 struct xarray io_buffers;
381 struct xarray personalities;
383 unsigned sq_thread_idle;
384 } ____cacheline_aligned_in_smp;
386 /* IRQ completion list, under ->completion_lock */
387 struct list_head locked_free_list;
388 unsigned int locked_free_nr;
390 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
391 struct io_sq_data *sq_data; /* if using sq thread polling */
393 struct wait_queue_head sqo_sq_wait;
394 struct list_head sqd_list;
396 unsigned long check_cq_overflow;
399 unsigned cached_cq_tail;
401 struct eventfd_ctx *cq_ev_fd;
402 struct wait_queue_head poll_wait;
403 struct wait_queue_head cq_wait;
405 atomic_t cq_timeouts;
406 struct fasync_struct *cq_fasync;
407 unsigned cq_last_tm_flush;
408 } ____cacheline_aligned_in_smp;
411 spinlock_t completion_lock;
413 spinlock_t timeout_lock;
416 * ->iopoll_list is protected by the ctx->uring_lock for
417 * io_uring instances that don't use IORING_SETUP_SQPOLL.
418 * For SQPOLL, only the single threaded io_sq_thread() will
419 * manipulate the list, hence no extra locking is needed there.
421 struct list_head iopoll_list;
422 struct hlist_head *cancel_hash;
423 unsigned cancel_hash_bits;
424 bool poll_multi_queue;
425 } ____cacheline_aligned_in_smp;
427 struct io_restriction restrictions;
429 /* slow path rsrc auxilary data, used by update/register */
431 struct io_rsrc_node *rsrc_backup_node;
432 struct io_mapped_ubuf *dummy_ubuf;
433 struct io_rsrc_data *file_data;
434 struct io_rsrc_data *buf_data;
436 struct delayed_work rsrc_put_work;
437 struct llist_head rsrc_put_llist;
438 struct list_head rsrc_ref_list;
439 spinlock_t rsrc_ref_lock;
442 /* Keep this last, we don't need it for the fast path */
444 #if defined(CONFIG_UNIX)
445 struct socket *ring_sock;
447 /* hashed buffered write serialization */
448 struct io_wq_hash *hash_map;
450 /* Only used for accounting purposes */
451 struct user_struct *user;
452 struct mm_struct *mm_account;
454 /* ctx exit and cancelation */
455 struct llist_head fallback_llist;
456 struct delayed_work fallback_work;
457 struct work_struct exit_work;
458 struct list_head tctx_list;
459 struct completion ref_comp;
463 struct io_uring_task {
464 /* submission side */
467 struct wait_queue_head wait;
468 const struct io_ring_ctx *last;
470 struct percpu_counter inflight;
471 atomic_t inflight_tracked;
474 spinlock_t task_lock;
475 struct io_wq_work_list task_list;
476 struct callback_head task_work;
481 * First field must be the file pointer in all the
482 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
484 struct io_poll_iocb {
486 struct wait_queue_head *head;
490 struct wait_queue_entry wait;
493 struct io_poll_update {
499 bool update_user_data;
507 struct io_timeout_data {
508 struct io_kiocb *req;
509 struct hrtimer timer;
510 struct timespec64 ts;
511 enum hrtimer_mode mode;
517 struct sockaddr __user *addr;
518 int __user *addr_len;
521 unsigned long nofile;
541 struct list_head list;
542 /* head of the link, used by linked timeouts only */
543 struct io_kiocb *head;
544 /* for linked completions */
545 struct io_kiocb *prev;
548 struct io_timeout_rem {
553 struct timespec64 ts;
558 /* NOTE: kiocb has the file as the first member, so don't do it here */
566 struct sockaddr __user *addr;
573 struct compat_msghdr __user *umsg_compat;
574 struct user_msghdr __user *umsg;
580 struct io_buffer *kbuf;
587 struct filename *filename;
589 unsigned long nofile;
592 struct io_rsrc_update {
618 struct epoll_event event;
622 struct file *file_out;
623 struct file *file_in;
630 struct io_provide_buf {
644 const char __user *filename;
645 struct statx __user *buffer;
657 struct filename *oldpath;
658 struct filename *newpath;
666 struct filename *filename;
669 struct io_completion {
674 struct io_async_connect {
675 struct sockaddr_storage address;
678 struct io_async_msghdr {
679 struct iovec fast_iov[UIO_FASTIOV];
680 /* points to an allocated iov, if NULL we use fast_iov instead */
681 struct iovec *free_iov;
682 struct sockaddr __user *uaddr;
684 struct sockaddr_storage addr;
688 struct iovec fast_iov[UIO_FASTIOV];
689 const struct iovec *free_iovec;
690 struct iov_iter iter;
692 struct wait_page_queue wpq;
696 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
697 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
698 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
699 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
700 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
701 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
703 /* first byte is taken by user flags, shift it to not overlap */
708 REQ_F_LINK_TIMEOUT_BIT,
709 REQ_F_NEED_CLEANUP_BIT,
711 REQ_F_BUFFER_SELECTED_BIT,
712 REQ_F_COMPLETE_INLINE_BIT,
714 REQ_F_DONT_REISSUE_BIT,
717 REQ_F_ARM_LTIMEOUT_BIT,
718 /* keep async read/write and isreg together and in order */
719 REQ_F_NOWAIT_READ_BIT,
720 REQ_F_NOWAIT_WRITE_BIT,
723 /* not a real bit, just to check we're not overflowing the space */
729 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
730 /* drain existing IO first */
731 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
733 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
734 /* doesn't sever on completion < 0 */
735 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
737 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
738 /* IOSQE_BUFFER_SELECT */
739 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
741 /* fail rest of links */
742 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
743 /* on inflight list, should be cancelled and waited on exit reliably */
744 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
745 /* read/write uses file position */
746 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
747 /* must not punt to workers */
748 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
749 /* has or had linked timeout */
750 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
752 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
753 /* already went through poll handler */
754 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
755 /* buffer already selected */
756 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
757 /* completion is deferred through io_comp_state */
758 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
759 /* caller should reissue async */
760 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
761 /* don't attempt request reissue, see io_rw_reissue() */
762 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
763 /* supports async reads */
764 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
765 /* supports async writes */
766 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
768 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
769 /* has creds assigned */
770 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
771 /* skip refcounting if not set */
772 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
773 /* there is a linked timeout that has to be armed */
774 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
778 struct io_poll_iocb poll;
779 struct io_poll_iocb *double_poll;
782 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
784 struct io_task_work {
786 struct io_wq_work_node node;
787 struct llist_node fallback_node;
789 io_req_tw_func_t func;
793 IORING_RSRC_FILE = 0,
794 IORING_RSRC_BUFFER = 1,
798 * NOTE! Each of the iocb union members has the file pointer
799 * as the first entry in their struct definition. So you can
800 * access the file pointer through any of the sub-structs,
801 * or directly as just 'ki_filp' in this struct.
807 struct io_poll_iocb poll;
808 struct io_poll_update poll_update;
809 struct io_accept accept;
811 struct io_cancel cancel;
812 struct io_timeout timeout;
813 struct io_timeout_rem timeout_rem;
814 struct io_connect connect;
815 struct io_sr_msg sr_msg;
817 struct io_close close;
818 struct io_rsrc_update rsrc_update;
819 struct io_fadvise fadvise;
820 struct io_madvise madvise;
821 struct io_epoll epoll;
822 struct io_splice splice;
823 struct io_provide_buf pbuf;
824 struct io_statx statx;
825 struct io_shutdown shutdown;
826 struct io_rename rename;
827 struct io_unlink unlink;
828 /* use only after cleaning per-op data, see io_clean_op() */
829 struct io_completion compl;
832 /* opcode allocated if it needs to store data for async defer */
835 /* polled IO has completed */
841 struct io_ring_ctx *ctx;
844 struct task_struct *task;
847 struct io_kiocb *link;
848 struct percpu_ref *fixed_rsrc_refs;
850 /* used with ctx->iopoll_list with reads/writes */
851 struct list_head inflight_entry;
852 struct io_task_work io_task_work;
853 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
854 struct hlist_node hash_node;
855 struct async_poll *apoll;
856 struct io_wq_work work;
857 const struct cred *creds;
859 /* store used ubuf, so we can prevent reloading */
860 struct io_mapped_ubuf *imu;
863 struct io_tctx_node {
864 struct list_head ctx_node;
865 struct task_struct *task;
866 struct io_ring_ctx *ctx;
869 struct io_defer_entry {
870 struct list_head list;
871 struct io_kiocb *req;
876 /* needs req->file assigned */
877 unsigned needs_file : 1;
878 /* hash wq insertion if file is a regular file */
879 unsigned hash_reg_file : 1;
880 /* unbound wq insertion if file is a non-regular file */
881 unsigned unbound_nonreg_file : 1;
882 /* opcode is not supported by this kernel */
883 unsigned not_supported : 1;
884 /* set if opcode supports polled "wait" */
886 unsigned pollout : 1;
887 /* op supports buffer selection */
888 unsigned buffer_select : 1;
889 /* do prep async if is going to be punted */
890 unsigned needs_async_setup : 1;
891 /* should block plug */
893 /* size of async data needed, if any */
894 unsigned short async_size;
897 static const struct io_op_def io_op_defs[] = {
898 [IORING_OP_NOP] = {},
899 [IORING_OP_READV] = {
901 .unbound_nonreg_file = 1,
904 .needs_async_setup = 1,
906 .async_size = sizeof(struct io_async_rw),
908 [IORING_OP_WRITEV] = {
911 .unbound_nonreg_file = 1,
913 .needs_async_setup = 1,
915 .async_size = sizeof(struct io_async_rw),
917 [IORING_OP_FSYNC] = {
920 [IORING_OP_READ_FIXED] = {
922 .unbound_nonreg_file = 1,
925 .async_size = sizeof(struct io_async_rw),
927 [IORING_OP_WRITE_FIXED] = {
930 .unbound_nonreg_file = 1,
933 .async_size = sizeof(struct io_async_rw),
935 [IORING_OP_POLL_ADD] = {
937 .unbound_nonreg_file = 1,
939 [IORING_OP_POLL_REMOVE] = {},
940 [IORING_OP_SYNC_FILE_RANGE] = {
943 [IORING_OP_SENDMSG] = {
945 .unbound_nonreg_file = 1,
947 .needs_async_setup = 1,
948 .async_size = sizeof(struct io_async_msghdr),
950 [IORING_OP_RECVMSG] = {
952 .unbound_nonreg_file = 1,
955 .needs_async_setup = 1,
956 .async_size = sizeof(struct io_async_msghdr),
958 [IORING_OP_TIMEOUT] = {
959 .async_size = sizeof(struct io_timeout_data),
961 [IORING_OP_TIMEOUT_REMOVE] = {
962 /* used by timeout updates' prep() */
964 [IORING_OP_ACCEPT] = {
966 .unbound_nonreg_file = 1,
969 [IORING_OP_ASYNC_CANCEL] = {},
970 [IORING_OP_LINK_TIMEOUT] = {
971 .async_size = sizeof(struct io_timeout_data),
973 [IORING_OP_CONNECT] = {
975 .unbound_nonreg_file = 1,
977 .needs_async_setup = 1,
978 .async_size = sizeof(struct io_async_connect),
980 [IORING_OP_FALLOCATE] = {
983 [IORING_OP_OPENAT] = {},
984 [IORING_OP_CLOSE] = {},
985 [IORING_OP_FILES_UPDATE] = {},
986 [IORING_OP_STATX] = {},
989 .unbound_nonreg_file = 1,
993 .async_size = sizeof(struct io_async_rw),
995 [IORING_OP_WRITE] = {
997 .unbound_nonreg_file = 1,
1000 .async_size = sizeof(struct io_async_rw),
1002 [IORING_OP_FADVISE] = {
1005 [IORING_OP_MADVISE] = {},
1006 [IORING_OP_SEND] = {
1008 .unbound_nonreg_file = 1,
1011 [IORING_OP_RECV] = {
1013 .unbound_nonreg_file = 1,
1017 [IORING_OP_OPENAT2] = {
1019 [IORING_OP_EPOLL_CTL] = {
1020 .unbound_nonreg_file = 1,
1022 [IORING_OP_SPLICE] = {
1025 .unbound_nonreg_file = 1,
1027 [IORING_OP_PROVIDE_BUFFERS] = {},
1028 [IORING_OP_REMOVE_BUFFERS] = {},
1032 .unbound_nonreg_file = 1,
1034 [IORING_OP_SHUTDOWN] = {
1037 [IORING_OP_RENAMEAT] = {},
1038 [IORING_OP_UNLINKAT] = {},
1041 /* requests with any of those set should undergo io_disarm_next() */
1042 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1044 static bool io_disarm_next(struct io_kiocb *req);
1045 static void io_uring_del_tctx_node(unsigned long index);
1046 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1047 struct task_struct *task,
1049 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1051 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1052 long res, unsigned int cflags);
1053 static void io_put_req(struct io_kiocb *req);
1054 static void io_put_req_deferred(struct io_kiocb *req);
1055 static void io_dismantle_req(struct io_kiocb *req);
1056 static void io_queue_linked_timeout(struct io_kiocb *req);
1057 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1058 struct io_uring_rsrc_update2 *up,
1060 static void io_clean_op(struct io_kiocb *req);
1061 static struct file *io_file_get(struct io_ring_ctx *ctx,
1062 struct io_kiocb *req, int fd, bool fixed);
1063 static void __io_queue_sqe(struct io_kiocb *req);
1064 static void io_rsrc_put_work(struct work_struct *work);
1066 static void io_req_task_queue(struct io_kiocb *req);
1067 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1068 static int io_req_prep_async(struct io_kiocb *req);
1070 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1071 unsigned int issue_flags, u32 slot_index);
1073 static struct kmem_cache *req_cachep;
1075 static const struct file_operations io_uring_fops;
1077 struct sock *io_uring_get_socket(struct file *file)
1079 #if defined(CONFIG_UNIX)
1080 if (file->f_op == &io_uring_fops) {
1081 struct io_ring_ctx *ctx = file->private_data;
1083 return ctx->ring_sock->sk;
1088 EXPORT_SYMBOL(io_uring_get_socket);
1090 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1093 mutex_lock(&ctx->uring_lock);
1098 #define io_for_each_link(pos, head) \
1099 for (pos = (head); pos; pos = pos->link)
1102 * Shamelessly stolen from the mm implementation of page reference checking,
1103 * see commit f958d7b528b1 for details.
1105 #define req_ref_zero_or_close_to_overflow(req) \
1106 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1108 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1110 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1111 return atomic_inc_not_zero(&req->refs);
1114 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1116 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1119 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1120 return atomic_dec_and_test(&req->refs);
1123 static inline void req_ref_put(struct io_kiocb *req)
1125 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1126 WARN_ON_ONCE(req_ref_put_and_test(req));
1129 static inline void req_ref_get(struct io_kiocb *req)
1131 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1132 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1133 atomic_inc(&req->refs);
1136 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1138 if (!(req->flags & REQ_F_REFCOUNT)) {
1139 req->flags |= REQ_F_REFCOUNT;
1140 atomic_set(&req->refs, nr);
1144 static inline void io_req_set_refcount(struct io_kiocb *req)
1146 __io_req_set_refcount(req, 1);
1149 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1151 struct io_ring_ctx *ctx = req->ctx;
1153 if (!req->fixed_rsrc_refs) {
1154 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1155 percpu_ref_get(req->fixed_rsrc_refs);
1159 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1161 bool got = percpu_ref_tryget(ref);
1163 /* already at zero, wait for ->release() */
1165 wait_for_completion(compl);
1166 percpu_ref_resurrect(ref);
1168 percpu_ref_put(ref);
1171 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1174 struct io_kiocb *req;
1176 if (task && head->task != task)
1181 io_for_each_link(req, head) {
1182 if (req->flags & REQ_F_INFLIGHT)
1188 static inline void req_set_fail(struct io_kiocb *req)
1190 req->flags |= REQ_F_FAIL;
1193 static inline void req_fail_link_node(struct io_kiocb *req, int res)
1199 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1201 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1203 complete(&ctx->ref_comp);
1206 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1208 return !req->timeout.off;
1211 static void io_fallback_req_func(struct work_struct *work)
1213 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1214 fallback_work.work);
1215 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1216 struct io_kiocb *req, *tmp;
1217 bool locked = false;
1219 percpu_ref_get(&ctx->refs);
1220 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1221 req->io_task_work.func(req, &locked);
1224 if (ctx->submit_state.compl_nr)
1225 io_submit_flush_completions(ctx);
1226 mutex_unlock(&ctx->uring_lock);
1228 percpu_ref_put(&ctx->refs);
1232 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1234 struct io_ring_ctx *ctx;
1237 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1242 * Use 5 bits less than the max cq entries, that should give us around
1243 * 32 entries per hash list if totally full and uniformly spread.
1245 hash_bits = ilog2(p->cq_entries);
1249 ctx->cancel_hash_bits = hash_bits;
1250 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1252 if (!ctx->cancel_hash)
1254 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1256 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1257 if (!ctx->dummy_ubuf)
1259 /* set invalid range, so io_import_fixed() fails meeting it */
1260 ctx->dummy_ubuf->ubuf = -1UL;
1262 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1263 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1266 ctx->flags = p->flags;
1267 init_waitqueue_head(&ctx->sqo_sq_wait);
1268 INIT_LIST_HEAD(&ctx->sqd_list);
1269 init_waitqueue_head(&ctx->poll_wait);
1270 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1271 init_completion(&ctx->ref_comp);
1272 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1273 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1274 mutex_init(&ctx->uring_lock);
1275 init_waitqueue_head(&ctx->cq_wait);
1276 spin_lock_init(&ctx->completion_lock);
1277 spin_lock_init(&ctx->timeout_lock);
1278 INIT_LIST_HEAD(&ctx->iopoll_list);
1279 INIT_LIST_HEAD(&ctx->defer_list);
1280 INIT_LIST_HEAD(&ctx->timeout_list);
1281 INIT_LIST_HEAD(&ctx->ltimeout_list);
1282 spin_lock_init(&ctx->rsrc_ref_lock);
1283 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1284 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1285 init_llist_head(&ctx->rsrc_put_llist);
1286 INIT_LIST_HEAD(&ctx->tctx_list);
1287 INIT_LIST_HEAD(&ctx->submit_state.free_list);
1288 INIT_LIST_HEAD(&ctx->locked_free_list);
1289 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1292 kfree(ctx->dummy_ubuf);
1293 kfree(ctx->cancel_hash);
1298 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1300 struct io_rings *r = ctx->rings;
1302 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1306 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1308 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1309 struct io_ring_ctx *ctx = req->ctx;
1311 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1317 #define FFS_ASYNC_READ 0x1UL
1318 #define FFS_ASYNC_WRITE 0x2UL
1320 #define FFS_ISREG 0x4UL
1322 #define FFS_ISREG 0x0UL
1324 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1326 static inline bool io_req_ffs_set(struct io_kiocb *req)
1328 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1331 static void io_req_track_inflight(struct io_kiocb *req)
1333 if (!(req->flags & REQ_F_INFLIGHT)) {
1334 req->flags |= REQ_F_INFLIGHT;
1335 atomic_inc(¤t->io_uring->inflight_tracked);
1339 static inline void io_unprep_linked_timeout(struct io_kiocb *req)
1341 req->flags &= ~REQ_F_LINK_TIMEOUT;
1344 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1346 if (WARN_ON_ONCE(!req->link))
1349 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1350 req->flags |= REQ_F_LINK_TIMEOUT;
1352 /* linked timeouts should have two refs once prep'ed */
1353 io_req_set_refcount(req);
1354 __io_req_set_refcount(req->link, 2);
1358 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1360 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1362 return __io_prep_linked_timeout(req);
1365 static void io_prep_async_work(struct io_kiocb *req)
1367 const struct io_op_def *def = &io_op_defs[req->opcode];
1368 struct io_ring_ctx *ctx = req->ctx;
1370 if (!(req->flags & REQ_F_CREDS)) {
1371 req->flags |= REQ_F_CREDS;
1372 req->creds = get_current_cred();
1375 req->work.list.next = NULL;
1376 req->work.flags = 0;
1377 if (req->flags & REQ_F_FORCE_ASYNC)
1378 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1380 if (req->flags & REQ_F_ISREG) {
1381 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1382 io_wq_hash_work(&req->work, file_inode(req->file));
1383 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1384 if (def->unbound_nonreg_file)
1385 req->work.flags |= IO_WQ_WORK_UNBOUND;
1388 switch (req->opcode) {
1389 case IORING_OP_SPLICE:
1391 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1392 req->work.flags |= IO_WQ_WORK_UNBOUND;
1397 static void io_prep_async_link(struct io_kiocb *req)
1399 struct io_kiocb *cur;
1401 if (req->flags & REQ_F_LINK_TIMEOUT) {
1402 struct io_ring_ctx *ctx = req->ctx;
1404 spin_lock(&ctx->completion_lock);
1405 io_for_each_link(cur, req)
1406 io_prep_async_work(cur);
1407 spin_unlock(&ctx->completion_lock);
1409 io_for_each_link(cur, req)
1410 io_prep_async_work(cur);
1414 static void io_queue_async_work(struct io_kiocb *req, bool *locked)
1416 struct io_ring_ctx *ctx = req->ctx;
1417 struct io_kiocb *link = io_prep_linked_timeout(req);
1418 struct io_uring_task *tctx = req->task->io_uring;
1420 /* must not take the lock, NULL it as a precaution */
1424 BUG_ON(!tctx->io_wq);
1426 /* init ->work of the whole link before punting */
1427 io_prep_async_link(req);
1430 * Not expected to happen, but if we do have a bug where this _can_
1431 * happen, catch it here and ensure the request is marked as
1432 * canceled. That will make io-wq go through the usual work cancel
1433 * procedure rather than attempt to run this request (or create a new
1436 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1437 req->work.flags |= IO_WQ_WORK_CANCEL;
1439 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1440 &req->work, req->flags);
1441 io_wq_enqueue(tctx->io_wq, &req->work);
1443 io_queue_linked_timeout(link);
1446 static void io_kill_timeout(struct io_kiocb *req, int status)
1447 __must_hold(&req->ctx->completion_lock)
1448 __must_hold(&req->ctx->timeout_lock)
1450 struct io_timeout_data *io = req->async_data;
1452 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1453 atomic_set(&req->ctx->cq_timeouts,
1454 atomic_read(&req->ctx->cq_timeouts) + 1);
1455 list_del_init(&req->timeout.list);
1456 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1457 io_put_req_deferred(req);
1461 static void io_queue_deferred(struct io_ring_ctx *ctx)
1463 while (!list_empty(&ctx->defer_list)) {
1464 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1465 struct io_defer_entry, list);
1467 if (req_need_defer(de->req, de->seq))
1469 list_del_init(&de->list);
1470 io_req_task_queue(de->req);
1475 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1476 __must_hold(&ctx->completion_lock)
1478 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1480 spin_lock_irq(&ctx->timeout_lock);
1481 while (!list_empty(&ctx->timeout_list)) {
1482 u32 events_needed, events_got;
1483 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1484 struct io_kiocb, timeout.list);
1486 if (io_is_timeout_noseq(req))
1490 * Since seq can easily wrap around over time, subtract
1491 * the last seq at which timeouts were flushed before comparing.
1492 * Assuming not more than 2^31-1 events have happened since,
1493 * these subtractions won't have wrapped, so we can check if
1494 * target is in [last_seq, current_seq] by comparing the two.
1496 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1497 events_got = seq - ctx->cq_last_tm_flush;
1498 if (events_got < events_needed)
1501 list_del_init(&req->timeout.list);
1502 io_kill_timeout(req, 0);
1504 ctx->cq_last_tm_flush = seq;
1505 spin_unlock_irq(&ctx->timeout_lock);
1508 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1510 if (ctx->off_timeout_used)
1511 io_flush_timeouts(ctx);
1512 if (ctx->drain_active)
1513 io_queue_deferred(ctx);
1516 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1518 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1519 __io_commit_cqring_flush(ctx);
1520 /* order cqe stores with ring update */
1521 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1524 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1526 struct io_rings *r = ctx->rings;
1528 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1531 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1533 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1536 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1538 struct io_rings *rings = ctx->rings;
1539 unsigned tail, mask = ctx->cq_entries - 1;
1542 * writes to the cq entry need to come after reading head; the
1543 * control dependency is enough as we're using WRITE_ONCE to
1546 if (__io_cqring_events(ctx) == ctx->cq_entries)
1549 tail = ctx->cached_cq_tail++;
1550 return &rings->cqes[tail & mask];
1553 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1555 if (likely(!ctx->cq_ev_fd))
1557 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1559 return !ctx->eventfd_async || io_wq_current_is_worker();
1563 * This should only get called when at least one event has been posted.
1564 * Some applications rely on the eventfd notification count only changing
1565 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1566 * 1:1 relationship between how many times this function is called (and
1567 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1569 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1572 * wake_up_all() may seem excessive, but io_wake_function() and
1573 * io_should_wake() handle the termination of the loop and only
1574 * wake as many waiters as we need to.
1576 if (wq_has_sleeper(&ctx->cq_wait))
1577 wake_up_all(&ctx->cq_wait);
1578 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1579 wake_up(&ctx->sq_data->wait);
1580 if (io_should_trigger_evfd(ctx))
1581 eventfd_signal(ctx->cq_ev_fd, 1);
1582 if (waitqueue_active(&ctx->poll_wait)) {
1583 wake_up_interruptible(&ctx->poll_wait);
1584 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1588 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1590 if (ctx->flags & IORING_SETUP_SQPOLL) {
1591 if (wq_has_sleeper(&ctx->cq_wait))
1592 wake_up_all(&ctx->cq_wait);
1594 if (io_should_trigger_evfd(ctx))
1595 eventfd_signal(ctx->cq_ev_fd, 1);
1596 if (waitqueue_active(&ctx->poll_wait)) {
1597 wake_up_interruptible(&ctx->poll_wait);
1598 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1602 /* Returns true if there are no backlogged entries after the flush */
1603 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1605 bool all_flushed, posted;
1607 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1611 spin_lock(&ctx->completion_lock);
1612 while (!list_empty(&ctx->cq_overflow_list)) {
1613 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1614 struct io_overflow_cqe *ocqe;
1618 ocqe = list_first_entry(&ctx->cq_overflow_list,
1619 struct io_overflow_cqe, list);
1621 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1623 io_account_cq_overflow(ctx);
1626 list_del(&ocqe->list);
1630 all_flushed = list_empty(&ctx->cq_overflow_list);
1632 clear_bit(0, &ctx->check_cq_overflow);
1633 WRITE_ONCE(ctx->rings->sq_flags,
1634 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1638 io_commit_cqring(ctx);
1639 spin_unlock(&ctx->completion_lock);
1641 io_cqring_ev_posted(ctx);
1645 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1649 if (test_bit(0, &ctx->check_cq_overflow)) {
1650 /* iopoll syncs against uring_lock, not completion_lock */
1651 if (ctx->flags & IORING_SETUP_IOPOLL)
1652 mutex_lock(&ctx->uring_lock);
1653 ret = __io_cqring_overflow_flush(ctx, false);
1654 if (ctx->flags & IORING_SETUP_IOPOLL)
1655 mutex_unlock(&ctx->uring_lock);
1661 /* must to be called somewhat shortly after putting a request */
1662 static inline void io_put_task(struct task_struct *task, int nr)
1664 struct io_uring_task *tctx = task->io_uring;
1666 if (likely(task == current)) {
1667 tctx->cached_refs += nr;
1669 percpu_counter_sub(&tctx->inflight, nr);
1670 if (unlikely(atomic_read(&tctx->in_idle)))
1671 wake_up(&tctx->wait);
1672 put_task_struct_many(task, nr);
1676 static void io_task_refs_refill(struct io_uring_task *tctx)
1678 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1680 percpu_counter_add(&tctx->inflight, refill);
1681 refcount_add(refill, ¤t->usage);
1682 tctx->cached_refs += refill;
1685 static inline void io_get_task_refs(int nr)
1687 struct io_uring_task *tctx = current->io_uring;
1689 tctx->cached_refs -= nr;
1690 if (unlikely(tctx->cached_refs < 0))
1691 io_task_refs_refill(tctx);
1694 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1695 long res, unsigned int cflags)
1697 struct io_overflow_cqe *ocqe;
1699 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1702 * If we're in ring overflow flush mode, or in task cancel mode,
1703 * or cannot allocate an overflow entry, then we need to drop it
1706 io_account_cq_overflow(ctx);
1709 if (list_empty(&ctx->cq_overflow_list)) {
1710 set_bit(0, &ctx->check_cq_overflow);
1711 WRITE_ONCE(ctx->rings->sq_flags,
1712 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1715 ocqe->cqe.user_data = user_data;
1716 ocqe->cqe.res = res;
1717 ocqe->cqe.flags = cflags;
1718 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1722 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1723 long res, unsigned int cflags)
1725 struct io_uring_cqe *cqe;
1727 trace_io_uring_complete(ctx, user_data, res, cflags);
1730 * If we can't get a cq entry, userspace overflowed the
1731 * submission (by quite a lot). Increment the overflow count in
1734 cqe = io_get_cqe(ctx);
1736 WRITE_ONCE(cqe->user_data, user_data);
1737 WRITE_ONCE(cqe->res, res);
1738 WRITE_ONCE(cqe->flags, cflags);
1741 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1744 /* not as hot to bloat with inlining */
1745 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1746 long res, unsigned int cflags)
1748 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1751 static void io_req_complete_post(struct io_kiocb *req, long res,
1752 unsigned int cflags)
1754 struct io_ring_ctx *ctx = req->ctx;
1756 spin_lock(&ctx->completion_lock);
1757 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1759 * If we're the last reference to this request, add to our locked
1762 if (req_ref_put_and_test(req)) {
1763 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1764 if (req->flags & IO_DISARM_MASK)
1765 io_disarm_next(req);
1767 io_req_task_queue(req->link);
1771 io_dismantle_req(req);
1772 io_put_task(req->task, 1);
1773 list_add(&req->inflight_entry, &ctx->locked_free_list);
1774 ctx->locked_free_nr++;
1776 if (!percpu_ref_tryget(&ctx->refs))
1779 io_commit_cqring(ctx);
1780 spin_unlock(&ctx->completion_lock);
1783 io_cqring_ev_posted(ctx);
1784 percpu_ref_put(&ctx->refs);
1788 static inline bool io_req_needs_clean(struct io_kiocb *req)
1790 return req->flags & IO_REQ_CLEAN_FLAGS;
1793 static void io_req_complete_state(struct io_kiocb *req, long res,
1794 unsigned int cflags)
1796 if (io_req_needs_clean(req))
1799 req->compl.cflags = cflags;
1800 req->flags |= REQ_F_COMPLETE_INLINE;
1803 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1804 long res, unsigned cflags)
1806 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1807 io_req_complete_state(req, res, cflags);
1809 io_req_complete_post(req, res, cflags);
1812 static inline void io_req_complete(struct io_kiocb *req, long res)
1814 __io_req_complete(req, 0, res, 0);
1817 static void io_req_complete_failed(struct io_kiocb *req, long res)
1820 io_req_complete_post(req, res, 0);
1824 * Don't initialise the fields below on every allocation, but do that in
1825 * advance and keep them valid across allocations.
1827 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1831 req->async_data = NULL;
1832 /* not necessary, but safer to zero */
1836 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1837 struct io_submit_state *state)
1839 spin_lock(&ctx->completion_lock);
1840 list_splice_init(&ctx->locked_free_list, &state->free_list);
1841 ctx->locked_free_nr = 0;
1842 spin_unlock(&ctx->completion_lock);
1845 /* Returns true IFF there are requests in the cache */
1846 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1848 struct io_submit_state *state = &ctx->submit_state;
1852 * If we have more than a batch's worth of requests in our IRQ side
1853 * locked cache, grab the lock and move them over to our submission
1856 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1857 io_flush_cached_locked_reqs(ctx, state);
1859 nr = state->free_reqs;
1860 while (!list_empty(&state->free_list)) {
1861 struct io_kiocb *req = list_first_entry(&state->free_list,
1862 struct io_kiocb, inflight_entry);
1864 list_del(&req->inflight_entry);
1865 state->reqs[nr++] = req;
1866 if (nr == ARRAY_SIZE(state->reqs))
1870 state->free_reqs = nr;
1875 * A request might get retired back into the request caches even before opcode
1876 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1877 * Because of that, io_alloc_req() should be called only under ->uring_lock
1878 * and with extra caution to not get a request that is still worked on.
1880 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1881 __must_hold(&ctx->uring_lock)
1883 struct io_submit_state *state = &ctx->submit_state;
1884 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1887 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1889 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1892 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1896 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1897 * retry single alloc to be on the safe side.
1899 if (unlikely(ret <= 0)) {
1900 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1901 if (!state->reqs[0])
1906 for (i = 0; i < ret; i++)
1907 io_preinit_req(state->reqs[i], ctx);
1908 state->free_reqs = ret;
1911 return state->reqs[state->free_reqs];
1914 static inline void io_put_file(struct file *file)
1920 static void io_dismantle_req(struct io_kiocb *req)
1922 unsigned int flags = req->flags;
1924 if (io_req_needs_clean(req))
1926 if (!(flags & REQ_F_FIXED_FILE))
1927 io_put_file(req->file);
1928 if (req->fixed_rsrc_refs)
1929 percpu_ref_put(req->fixed_rsrc_refs);
1930 if (req->async_data) {
1931 kfree(req->async_data);
1932 req->async_data = NULL;
1936 static void __io_free_req(struct io_kiocb *req)
1938 struct io_ring_ctx *ctx = req->ctx;
1940 io_dismantle_req(req);
1941 io_put_task(req->task, 1);
1943 spin_lock(&ctx->completion_lock);
1944 list_add(&req->inflight_entry, &ctx->locked_free_list);
1945 ctx->locked_free_nr++;
1946 spin_unlock(&ctx->completion_lock);
1948 percpu_ref_put(&ctx->refs);
1951 static inline void io_remove_next_linked(struct io_kiocb *req)
1953 struct io_kiocb *nxt = req->link;
1955 req->link = nxt->link;
1959 static bool io_kill_linked_timeout(struct io_kiocb *req)
1960 __must_hold(&req->ctx->completion_lock)
1961 __must_hold(&req->ctx->timeout_lock)
1963 struct io_kiocb *link = req->link;
1965 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1966 struct io_timeout_data *io = link->async_data;
1968 io_remove_next_linked(req);
1969 link->timeout.head = NULL;
1970 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1971 list_del(&link->timeout.list);
1972 io_cqring_fill_event(link->ctx, link->user_data,
1974 io_put_req_deferred(link);
1981 static void io_fail_links(struct io_kiocb *req)
1982 __must_hold(&req->ctx->completion_lock)
1984 struct io_kiocb *nxt, *link = req->link;
1988 long res = -ECANCELED;
1990 if (link->flags & REQ_F_FAIL)
1996 trace_io_uring_fail_link(req, link);
1997 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
1998 io_put_req_deferred(link);
2003 static bool io_disarm_next(struct io_kiocb *req)
2004 __must_hold(&req->ctx->completion_lock)
2006 bool posted = false;
2008 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2009 struct io_kiocb *link = req->link;
2011 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2012 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2013 io_remove_next_linked(req);
2014 io_cqring_fill_event(link->ctx, link->user_data,
2016 io_put_req_deferred(link);
2019 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2020 struct io_ring_ctx *ctx = req->ctx;
2022 spin_lock_irq(&ctx->timeout_lock);
2023 posted = io_kill_linked_timeout(req);
2024 spin_unlock_irq(&ctx->timeout_lock);
2026 if (unlikely((req->flags & REQ_F_FAIL) &&
2027 !(req->flags & REQ_F_HARDLINK))) {
2028 posted |= (req->link != NULL);
2034 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2036 struct io_kiocb *nxt;
2039 * If LINK is set, we have dependent requests in this chain. If we
2040 * didn't fail this request, queue the first one up, moving any other
2041 * dependencies to the next request. In case of failure, fail the rest
2044 if (req->flags & IO_DISARM_MASK) {
2045 struct io_ring_ctx *ctx = req->ctx;
2048 spin_lock(&ctx->completion_lock);
2049 posted = io_disarm_next(req);
2051 io_commit_cqring(req->ctx);
2052 spin_unlock(&ctx->completion_lock);
2054 io_cqring_ev_posted(ctx);
2061 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2063 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2065 return __io_req_find_next(req);
2068 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2073 if (ctx->submit_state.compl_nr)
2074 io_submit_flush_completions(ctx);
2075 mutex_unlock(&ctx->uring_lock);
2078 percpu_ref_put(&ctx->refs);
2081 static void tctx_task_work(struct callback_head *cb)
2083 bool locked = false;
2084 struct io_ring_ctx *ctx = NULL;
2085 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2089 struct io_wq_work_node *node;
2091 spin_lock_irq(&tctx->task_lock);
2092 node = tctx->task_list.first;
2093 INIT_WQ_LIST(&tctx->task_list);
2095 tctx->task_running = false;
2096 spin_unlock_irq(&tctx->task_lock);
2101 struct io_wq_work_node *next = node->next;
2102 struct io_kiocb *req = container_of(node, struct io_kiocb,
2105 if (req->ctx != ctx) {
2106 ctx_flush_and_put(ctx, &locked);
2108 /* if not contended, grab and improve batching */
2109 locked = mutex_trylock(&ctx->uring_lock);
2110 percpu_ref_get(&ctx->refs);
2112 req->io_task_work.func(req, &locked);
2119 ctx_flush_and_put(ctx, &locked);
2122 static void io_req_task_work_add(struct io_kiocb *req)
2124 struct task_struct *tsk = req->task;
2125 struct io_uring_task *tctx = tsk->io_uring;
2126 enum task_work_notify_mode notify;
2127 struct io_wq_work_node *node;
2128 unsigned long flags;
2131 WARN_ON_ONCE(!tctx);
2133 spin_lock_irqsave(&tctx->task_lock, flags);
2134 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2135 running = tctx->task_running;
2137 tctx->task_running = true;
2138 spin_unlock_irqrestore(&tctx->task_lock, flags);
2140 /* task_work already pending, we're done */
2145 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2146 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2147 * processing task_work. There's no reliable way to tell if TWA_RESUME
2150 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2151 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2152 wake_up_process(tsk);
2156 spin_lock_irqsave(&tctx->task_lock, flags);
2157 tctx->task_running = false;
2158 node = tctx->task_list.first;
2159 INIT_WQ_LIST(&tctx->task_list);
2160 spin_unlock_irqrestore(&tctx->task_lock, flags);
2163 req = container_of(node, struct io_kiocb, io_task_work.node);
2165 if (llist_add(&req->io_task_work.fallback_node,
2166 &req->ctx->fallback_llist))
2167 schedule_delayed_work(&req->ctx->fallback_work, 1);
2171 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2173 struct io_ring_ctx *ctx = req->ctx;
2175 /* not needed for normal modes, but SQPOLL depends on it */
2176 io_tw_lock(ctx, locked);
2177 io_req_complete_failed(req, req->result);
2180 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2182 struct io_ring_ctx *ctx = req->ctx;
2184 io_tw_lock(ctx, locked);
2185 /* req->task == current here, checking PF_EXITING is safe */
2186 if (likely(!(req->task->flags & PF_EXITING)))
2187 __io_queue_sqe(req);
2189 io_req_complete_failed(req, -EFAULT);
2192 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2195 req->io_task_work.func = io_req_task_cancel;
2196 io_req_task_work_add(req);
2199 static void io_req_task_queue(struct io_kiocb *req)
2201 req->io_task_work.func = io_req_task_submit;
2202 io_req_task_work_add(req);
2205 static void io_req_task_queue_reissue(struct io_kiocb *req)
2207 req->io_task_work.func = io_queue_async_work;
2208 io_req_task_work_add(req);
2211 static inline void io_queue_next(struct io_kiocb *req)
2213 struct io_kiocb *nxt = io_req_find_next(req);
2216 io_req_task_queue(nxt);
2219 static void io_free_req(struct io_kiocb *req)
2225 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2231 struct task_struct *task;
2236 static inline void io_init_req_batch(struct req_batch *rb)
2243 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2244 struct req_batch *rb)
2247 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2249 io_put_task(rb->task, rb->task_refs);
2252 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2253 struct io_submit_state *state)
2256 io_dismantle_req(req);
2258 if (req->task != rb->task) {
2260 io_put_task(rb->task, rb->task_refs);
2261 rb->task = req->task;
2267 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2268 state->reqs[state->free_reqs++] = req;
2270 list_add(&req->inflight_entry, &state->free_list);
2273 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2274 __must_hold(&ctx->uring_lock)
2276 struct io_submit_state *state = &ctx->submit_state;
2277 int i, nr = state->compl_nr;
2278 struct req_batch rb;
2280 spin_lock(&ctx->completion_lock);
2281 for (i = 0; i < nr; i++) {
2282 struct io_kiocb *req = state->compl_reqs[i];
2284 __io_cqring_fill_event(ctx, req->user_data, req->result,
2287 io_commit_cqring(ctx);
2288 spin_unlock(&ctx->completion_lock);
2289 io_cqring_ev_posted(ctx);
2291 io_init_req_batch(&rb);
2292 for (i = 0; i < nr; i++) {
2293 struct io_kiocb *req = state->compl_reqs[i];
2295 if (req_ref_put_and_test(req))
2296 io_req_free_batch(&rb, req, &ctx->submit_state);
2299 io_req_free_batch_finish(ctx, &rb);
2300 state->compl_nr = 0;
2304 * Drop reference to request, return next in chain (if there is one) if this
2305 * was the last reference to this request.
2307 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2309 struct io_kiocb *nxt = NULL;
2311 if (req_ref_put_and_test(req)) {
2312 nxt = io_req_find_next(req);
2318 static inline void io_put_req(struct io_kiocb *req)
2320 if (req_ref_put_and_test(req))
2324 static inline void io_put_req_deferred(struct io_kiocb *req)
2326 if (req_ref_put_and_test(req)) {
2327 req->io_task_work.func = io_free_req_work;
2328 io_req_task_work_add(req);
2332 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2334 /* See comment at the top of this file */
2336 return __io_cqring_events(ctx);
2339 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2341 struct io_rings *rings = ctx->rings;
2343 /* make sure SQ entry isn't read before tail */
2344 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2347 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2349 unsigned int cflags;
2351 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2352 cflags |= IORING_CQE_F_BUFFER;
2353 req->flags &= ~REQ_F_BUFFER_SELECTED;
2358 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2360 struct io_buffer *kbuf;
2362 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2364 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2365 return io_put_kbuf(req, kbuf);
2368 static inline bool io_run_task_work(void)
2370 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2371 __set_current_state(TASK_RUNNING);
2372 tracehook_notify_signal();
2380 * Find and free completed poll iocbs
2382 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2383 struct list_head *done)
2385 struct req_batch rb;
2386 struct io_kiocb *req;
2388 /* order with ->result store in io_complete_rw_iopoll() */
2391 io_init_req_batch(&rb);
2392 while (!list_empty(done)) {
2393 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2394 list_del(&req->inflight_entry);
2396 if (READ_ONCE(req->result) == -EAGAIN &&
2397 !(req->flags & REQ_F_DONT_REISSUE)) {
2398 req->iopoll_completed = 0;
2399 io_req_task_queue_reissue(req);
2403 __io_cqring_fill_event(ctx, req->user_data, req->result,
2404 io_put_rw_kbuf(req));
2407 if (req_ref_put_and_test(req))
2408 io_req_free_batch(&rb, req, &ctx->submit_state);
2411 io_commit_cqring(ctx);
2412 io_cqring_ev_posted_iopoll(ctx);
2413 io_req_free_batch_finish(ctx, &rb);
2416 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2419 struct io_kiocb *req, *tmp;
2424 * Only spin for completions if we don't have multiple devices hanging
2425 * off our complete list, and we're under the requested amount.
2427 spin = !ctx->poll_multi_queue && *nr_events < min;
2429 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2430 struct kiocb *kiocb = &req->rw.kiocb;
2434 * Move completed and retryable entries to our local lists.
2435 * If we find a request that requires polling, break out
2436 * and complete those lists first, if we have entries there.
2438 if (READ_ONCE(req->iopoll_completed)) {
2439 list_move_tail(&req->inflight_entry, &done);
2442 if (!list_empty(&done))
2445 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2446 if (unlikely(ret < 0))
2451 /* iopoll may have completed current req */
2452 if (READ_ONCE(req->iopoll_completed))
2453 list_move_tail(&req->inflight_entry, &done);
2456 if (!list_empty(&done))
2457 io_iopoll_complete(ctx, nr_events, &done);
2463 * We can't just wait for polled events to come to us, we have to actively
2464 * find and complete them.
2466 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2468 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2471 mutex_lock(&ctx->uring_lock);
2472 while (!list_empty(&ctx->iopoll_list)) {
2473 unsigned int nr_events = 0;
2475 io_do_iopoll(ctx, &nr_events, 0);
2477 /* let it sleep and repeat later if can't complete a request */
2481 * Ensure we allow local-to-the-cpu processing to take place,
2482 * in this case we need to ensure that we reap all events.
2483 * Also let task_work, etc. to progress by releasing the mutex
2485 if (need_resched()) {
2486 mutex_unlock(&ctx->uring_lock);
2488 mutex_lock(&ctx->uring_lock);
2491 mutex_unlock(&ctx->uring_lock);
2494 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2496 unsigned int nr_events = 0;
2500 * We disallow the app entering submit/complete with polling, but we
2501 * still need to lock the ring to prevent racing with polled issue
2502 * that got punted to a workqueue.
2504 mutex_lock(&ctx->uring_lock);
2506 * Don't enter poll loop if we already have events pending.
2507 * If we do, we can potentially be spinning for commands that
2508 * already triggered a CQE (eg in error).
2510 if (test_bit(0, &ctx->check_cq_overflow))
2511 __io_cqring_overflow_flush(ctx, false);
2512 if (io_cqring_events(ctx))
2516 * If a submit got punted to a workqueue, we can have the
2517 * application entering polling for a command before it gets
2518 * issued. That app will hold the uring_lock for the duration
2519 * of the poll right here, so we need to take a breather every
2520 * now and then to ensure that the issue has a chance to add
2521 * the poll to the issued list. Otherwise we can spin here
2522 * forever, while the workqueue is stuck trying to acquire the
2525 if (list_empty(&ctx->iopoll_list)) {
2526 u32 tail = ctx->cached_cq_tail;
2528 mutex_unlock(&ctx->uring_lock);
2530 mutex_lock(&ctx->uring_lock);
2532 /* some requests don't go through iopoll_list */
2533 if (tail != ctx->cached_cq_tail ||
2534 list_empty(&ctx->iopoll_list))
2537 ret = io_do_iopoll(ctx, &nr_events, min);
2538 } while (!ret && nr_events < min && !need_resched());
2540 mutex_unlock(&ctx->uring_lock);
2544 static void kiocb_end_write(struct io_kiocb *req)
2547 * Tell lockdep we inherited freeze protection from submission
2550 if (req->flags & REQ_F_ISREG) {
2551 struct super_block *sb = file_inode(req->file)->i_sb;
2553 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2559 static bool io_resubmit_prep(struct io_kiocb *req)
2561 struct io_async_rw *rw = req->async_data;
2564 return !io_req_prep_async(req);
2565 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2566 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2570 static bool io_rw_should_reissue(struct io_kiocb *req)
2572 umode_t mode = file_inode(req->file)->i_mode;
2573 struct io_ring_ctx *ctx = req->ctx;
2575 if (!S_ISBLK(mode) && !S_ISREG(mode))
2577 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2578 !(ctx->flags & IORING_SETUP_IOPOLL)))
2581 * If ref is dying, we might be running poll reap from the exit work.
2582 * Don't attempt to reissue from that path, just let it fail with
2585 if (percpu_ref_is_dying(&ctx->refs))
2588 * Play it safe and assume not safe to re-import and reissue if we're
2589 * not in the original thread group (or in task context).
2591 if (!same_thread_group(req->task, current) || !in_task())
2596 static bool io_resubmit_prep(struct io_kiocb *req)
2600 static bool io_rw_should_reissue(struct io_kiocb *req)
2606 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2608 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2609 kiocb_end_write(req);
2610 if (res != req->result) {
2611 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2612 io_rw_should_reissue(req)) {
2613 req->flags |= REQ_F_REISSUE;
2622 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2624 unsigned int cflags = io_put_rw_kbuf(req);
2625 long res = req->result;
2628 struct io_ring_ctx *ctx = req->ctx;
2629 struct io_submit_state *state = &ctx->submit_state;
2631 io_req_complete_state(req, res, cflags);
2632 state->compl_reqs[state->compl_nr++] = req;
2633 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2634 io_submit_flush_completions(ctx);
2636 io_req_complete_post(req, res, cflags);
2640 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2641 unsigned int issue_flags)
2643 if (__io_complete_rw_common(req, res))
2645 __io_req_complete(req, 0, req->result, io_put_rw_kbuf(req));
2648 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2650 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2652 if (__io_complete_rw_common(req, res))
2655 req->io_task_work.func = io_req_task_complete;
2656 io_req_task_work_add(req);
2659 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2661 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2663 if (kiocb->ki_flags & IOCB_WRITE)
2664 kiocb_end_write(req);
2665 if (unlikely(res != req->result)) {
2666 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2667 io_resubmit_prep(req))) {
2669 req->flags |= REQ_F_DONT_REISSUE;
2673 WRITE_ONCE(req->result, res);
2674 /* order with io_iopoll_complete() checking ->result */
2676 WRITE_ONCE(req->iopoll_completed, 1);
2680 * After the iocb has been issued, it's safe to be found on the poll list.
2681 * Adding the kiocb to the list AFTER submission ensures that we don't
2682 * find it from a io_do_iopoll() thread before the issuer is done
2683 * accessing the kiocb cookie.
2685 static void io_iopoll_req_issued(struct io_kiocb *req)
2687 struct io_ring_ctx *ctx = req->ctx;
2688 const bool in_async = io_wq_current_is_worker();
2690 /* workqueue context doesn't hold uring_lock, grab it now */
2691 if (unlikely(in_async))
2692 mutex_lock(&ctx->uring_lock);
2695 * Track whether we have multiple files in our lists. This will impact
2696 * how we do polling eventually, not spinning if we're on potentially
2697 * different devices.
2699 if (list_empty(&ctx->iopoll_list)) {
2700 ctx->poll_multi_queue = false;
2701 } else if (!ctx->poll_multi_queue) {
2702 struct io_kiocb *list_req;
2703 unsigned int queue_num0, queue_num1;
2705 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2708 if (list_req->file != req->file) {
2709 ctx->poll_multi_queue = true;
2711 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2712 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2713 if (queue_num0 != queue_num1)
2714 ctx->poll_multi_queue = true;
2719 * For fast devices, IO may have already completed. If it has, add
2720 * it to the front so we find it first.
2722 if (READ_ONCE(req->iopoll_completed))
2723 list_add(&req->inflight_entry, &ctx->iopoll_list);
2725 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2727 if (unlikely(in_async)) {
2729 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2730 * in sq thread task context or in io worker task context. If
2731 * current task context is sq thread, we don't need to check
2732 * whether should wake up sq thread.
2734 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2735 wq_has_sleeper(&ctx->sq_data->wait))
2736 wake_up(&ctx->sq_data->wait);
2738 mutex_unlock(&ctx->uring_lock);
2742 static bool io_bdev_nowait(struct block_device *bdev)
2744 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2748 * If we tracked the file through the SCM inflight mechanism, we could support
2749 * any file. For now, just ensure that anything potentially problematic is done
2752 static bool __io_file_supports_nowait(struct file *file, int rw)
2754 umode_t mode = file_inode(file)->i_mode;
2756 if (S_ISBLK(mode)) {
2757 if (IS_ENABLED(CONFIG_BLOCK) &&
2758 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2764 if (S_ISREG(mode)) {
2765 if (IS_ENABLED(CONFIG_BLOCK) &&
2766 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2767 file->f_op != &io_uring_fops)
2772 /* any ->read/write should understand O_NONBLOCK */
2773 if (file->f_flags & O_NONBLOCK)
2776 if (!(file->f_mode & FMODE_NOWAIT))
2780 return file->f_op->read_iter != NULL;
2782 return file->f_op->write_iter != NULL;
2785 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2787 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2789 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2792 return __io_file_supports_nowait(req->file, rw);
2795 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2797 struct io_ring_ctx *ctx = req->ctx;
2798 struct kiocb *kiocb = &req->rw.kiocb;
2799 struct file *file = req->file;
2803 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2804 req->flags |= REQ_F_ISREG;
2806 kiocb->ki_pos = READ_ONCE(sqe->off);
2807 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2808 req->flags |= REQ_F_CUR_POS;
2809 kiocb->ki_pos = file->f_pos;
2811 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2812 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2813 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2817 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2818 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2819 req->flags |= REQ_F_NOWAIT;
2821 ioprio = READ_ONCE(sqe->ioprio);
2823 ret = ioprio_check_cap(ioprio);
2827 kiocb->ki_ioprio = ioprio;
2829 kiocb->ki_ioprio = get_current_ioprio();
2831 if (ctx->flags & IORING_SETUP_IOPOLL) {
2832 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2833 !kiocb->ki_filp->f_op->iopoll)
2836 kiocb->ki_flags |= IOCB_HIPRI;
2837 kiocb->ki_complete = io_complete_rw_iopoll;
2838 req->iopoll_completed = 0;
2840 if (kiocb->ki_flags & IOCB_HIPRI)
2842 kiocb->ki_complete = io_complete_rw;
2845 if (req->opcode == IORING_OP_READ_FIXED ||
2846 req->opcode == IORING_OP_WRITE_FIXED) {
2848 io_req_set_rsrc_node(req);
2851 req->rw.addr = READ_ONCE(sqe->addr);
2852 req->rw.len = READ_ONCE(sqe->len);
2853 req->buf_index = READ_ONCE(sqe->buf_index);
2857 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2863 case -ERESTARTNOINTR:
2864 case -ERESTARTNOHAND:
2865 case -ERESTART_RESTARTBLOCK:
2867 * We can't just restart the syscall, since previously
2868 * submitted sqes may already be in progress. Just fail this
2874 kiocb->ki_complete(kiocb, ret, 0);
2878 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2879 unsigned int issue_flags)
2881 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2882 struct io_async_rw *io = req->async_data;
2883 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2885 /* add previously done IO, if any */
2886 if (io && io->bytes_done > 0) {
2888 ret = io->bytes_done;
2890 ret += io->bytes_done;
2893 if (req->flags & REQ_F_CUR_POS)
2894 req->file->f_pos = kiocb->ki_pos;
2895 if (ret >= 0 && check_reissue)
2896 __io_complete_rw(req, ret, 0, issue_flags);
2898 io_rw_done(kiocb, ret);
2900 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2901 req->flags &= ~REQ_F_REISSUE;
2902 if (io_resubmit_prep(req)) {
2903 io_req_task_queue_reissue(req);
2906 __io_req_complete(req, issue_flags, ret,
2907 io_put_rw_kbuf(req));
2912 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2913 struct io_mapped_ubuf *imu)
2915 size_t len = req->rw.len;
2916 u64 buf_end, buf_addr = req->rw.addr;
2919 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2921 /* not inside the mapped region */
2922 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2926 * May not be a start of buffer, set size appropriately
2927 * and advance us to the beginning.
2929 offset = buf_addr - imu->ubuf;
2930 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2934 * Don't use iov_iter_advance() here, as it's really slow for
2935 * using the latter parts of a big fixed buffer - it iterates
2936 * over each segment manually. We can cheat a bit here, because
2939 * 1) it's a BVEC iter, we set it up
2940 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2941 * first and last bvec
2943 * So just find our index, and adjust the iterator afterwards.
2944 * If the offset is within the first bvec (or the whole first
2945 * bvec, just use iov_iter_advance(). This makes it easier
2946 * since we can just skip the first segment, which may not
2947 * be PAGE_SIZE aligned.
2949 const struct bio_vec *bvec = imu->bvec;
2951 if (offset <= bvec->bv_len) {
2952 iov_iter_advance(iter, offset);
2954 unsigned long seg_skip;
2956 /* skip first vec */
2957 offset -= bvec->bv_len;
2958 seg_skip = 1 + (offset >> PAGE_SHIFT);
2960 iter->bvec = bvec + seg_skip;
2961 iter->nr_segs -= seg_skip;
2962 iter->count -= bvec->bv_len + offset;
2963 iter->iov_offset = offset & ~PAGE_MASK;
2970 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2972 struct io_ring_ctx *ctx = req->ctx;
2973 struct io_mapped_ubuf *imu = req->imu;
2974 u16 index, buf_index = req->buf_index;
2977 if (unlikely(buf_index >= ctx->nr_user_bufs))
2979 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2980 imu = READ_ONCE(ctx->user_bufs[index]);
2983 return __io_import_fixed(req, rw, iter, imu);
2986 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2989 mutex_unlock(&ctx->uring_lock);
2992 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2995 * "Normal" inline submissions always hold the uring_lock, since we
2996 * grab it from the system call. Same is true for the SQPOLL offload.
2997 * The only exception is when we've detached the request and issue it
2998 * from an async worker thread, grab the lock for that case.
3001 mutex_lock(&ctx->uring_lock);
3004 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3005 int bgid, struct io_buffer *kbuf,
3008 struct io_buffer *head;
3010 if (req->flags & REQ_F_BUFFER_SELECTED)
3013 io_ring_submit_lock(req->ctx, needs_lock);
3015 lockdep_assert_held(&req->ctx->uring_lock);
3017 head = xa_load(&req->ctx->io_buffers, bgid);
3019 if (!list_empty(&head->list)) {
3020 kbuf = list_last_entry(&head->list, struct io_buffer,
3022 list_del(&kbuf->list);
3025 xa_erase(&req->ctx->io_buffers, bgid);
3027 if (*len > kbuf->len)
3030 kbuf = ERR_PTR(-ENOBUFS);
3033 io_ring_submit_unlock(req->ctx, needs_lock);
3038 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3041 struct io_buffer *kbuf;
3044 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3045 bgid = req->buf_index;
3046 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3049 req->rw.addr = (u64) (unsigned long) kbuf;
3050 req->flags |= REQ_F_BUFFER_SELECTED;
3051 return u64_to_user_ptr(kbuf->addr);
3054 #ifdef CONFIG_COMPAT
3055 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3058 struct compat_iovec __user *uiov;
3059 compat_ssize_t clen;
3063 uiov = u64_to_user_ptr(req->rw.addr);
3064 if (!access_ok(uiov, sizeof(*uiov)))
3066 if (__get_user(clen, &uiov->iov_len))
3072 buf = io_rw_buffer_select(req, &len, needs_lock);
3074 return PTR_ERR(buf);
3075 iov[0].iov_base = buf;
3076 iov[0].iov_len = (compat_size_t) len;
3081 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3084 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3088 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3091 len = iov[0].iov_len;
3094 buf = io_rw_buffer_select(req, &len, needs_lock);
3096 return PTR_ERR(buf);
3097 iov[0].iov_base = buf;
3098 iov[0].iov_len = len;
3102 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3105 if (req->flags & REQ_F_BUFFER_SELECTED) {
3106 struct io_buffer *kbuf;
3108 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3109 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3110 iov[0].iov_len = kbuf->len;
3113 if (req->rw.len != 1)
3116 #ifdef CONFIG_COMPAT
3117 if (req->ctx->compat)
3118 return io_compat_import(req, iov, needs_lock);
3121 return __io_iov_buffer_select(req, iov, needs_lock);
3124 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3125 struct iov_iter *iter, bool needs_lock)
3127 void __user *buf = u64_to_user_ptr(req->rw.addr);
3128 size_t sqe_len = req->rw.len;
3129 u8 opcode = req->opcode;
3132 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3134 return io_import_fixed(req, rw, iter);
3137 /* buffer index only valid with fixed read/write, or buffer select */
3138 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3141 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3142 if (req->flags & REQ_F_BUFFER_SELECT) {
3143 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3145 return PTR_ERR(buf);
3146 req->rw.len = sqe_len;
3149 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3154 if (req->flags & REQ_F_BUFFER_SELECT) {
3155 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3157 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3162 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3166 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3168 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3172 * For files that don't have ->read_iter() and ->write_iter(), handle them
3173 * by looping over ->read() or ->write() manually.
3175 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3177 struct kiocb *kiocb = &req->rw.kiocb;
3178 struct file *file = req->file;
3182 * Don't support polled IO through this interface, and we can't
3183 * support non-blocking either. For the latter, this just causes
3184 * the kiocb to be handled from an async context.
3186 if (kiocb->ki_flags & IOCB_HIPRI)
3188 if (kiocb->ki_flags & IOCB_NOWAIT)
3191 while (iov_iter_count(iter)) {
3195 if (!iov_iter_is_bvec(iter)) {
3196 iovec = iov_iter_iovec(iter);
3198 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3199 iovec.iov_len = req->rw.len;
3203 nr = file->f_op->read(file, iovec.iov_base,
3204 iovec.iov_len, io_kiocb_ppos(kiocb));
3206 nr = file->f_op->write(file, iovec.iov_base,
3207 iovec.iov_len, io_kiocb_ppos(kiocb));
3216 if (nr != iovec.iov_len)
3220 iov_iter_advance(iter, nr);
3226 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3227 const struct iovec *fast_iov, struct iov_iter *iter)
3229 struct io_async_rw *rw = req->async_data;
3231 memcpy(&rw->iter, iter, sizeof(*iter));
3232 rw->free_iovec = iovec;
3234 /* can only be fixed buffers, no need to do anything */
3235 if (iov_iter_is_bvec(iter))
3238 unsigned iov_off = 0;
3240 rw->iter.iov = rw->fast_iov;
3241 if (iter->iov != fast_iov) {
3242 iov_off = iter->iov - fast_iov;
3243 rw->iter.iov += iov_off;
3245 if (rw->fast_iov != fast_iov)
3246 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3247 sizeof(struct iovec) * iter->nr_segs);
3249 req->flags |= REQ_F_NEED_CLEANUP;
3253 static inline int io_alloc_async_data(struct io_kiocb *req)
3255 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3256 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3257 return req->async_data == NULL;
3260 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3261 const struct iovec *fast_iov,
3262 struct iov_iter *iter, bool force)
3264 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3266 if (!req->async_data) {
3267 if (io_alloc_async_data(req)) {
3272 io_req_map_rw(req, iovec, fast_iov, iter);
3277 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3279 struct io_async_rw *iorw = req->async_data;
3280 struct iovec *iov = iorw->fast_iov;
3283 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3284 if (unlikely(ret < 0))
3287 iorw->bytes_done = 0;
3288 iorw->free_iovec = iov;
3290 req->flags |= REQ_F_NEED_CLEANUP;
3294 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3296 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3298 return io_prep_rw(req, sqe);
3302 * This is our waitqueue callback handler, registered through lock_page_async()
3303 * when we initially tried to do the IO with the iocb armed our waitqueue.
3304 * This gets called when the page is unlocked, and we generally expect that to
3305 * happen when the page IO is completed and the page is now uptodate. This will
3306 * queue a task_work based retry of the operation, attempting to copy the data
3307 * again. If the latter fails because the page was NOT uptodate, then we will
3308 * do a thread based blocking retry of the operation. That's the unexpected
3311 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3312 int sync, void *arg)
3314 struct wait_page_queue *wpq;
3315 struct io_kiocb *req = wait->private;
3316 struct wait_page_key *key = arg;
3318 wpq = container_of(wait, struct wait_page_queue, wait);
3320 if (!wake_page_match(wpq, key))
3323 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3324 list_del_init(&wait->entry);
3325 io_req_task_queue(req);
3330 * This controls whether a given IO request should be armed for async page
3331 * based retry. If we return false here, the request is handed to the async
3332 * worker threads for retry. If we're doing buffered reads on a regular file,
3333 * we prepare a private wait_page_queue entry and retry the operation. This
3334 * will either succeed because the page is now uptodate and unlocked, or it
3335 * will register a callback when the page is unlocked at IO completion. Through
3336 * that callback, io_uring uses task_work to setup a retry of the operation.
3337 * That retry will attempt the buffered read again. The retry will generally
3338 * succeed, or in rare cases where it fails, we then fall back to using the
3339 * async worker threads for a blocking retry.
3341 static bool io_rw_should_retry(struct io_kiocb *req)
3343 struct io_async_rw *rw = req->async_data;
3344 struct wait_page_queue *wait = &rw->wpq;
3345 struct kiocb *kiocb = &req->rw.kiocb;
3347 /* never retry for NOWAIT, we just complete with -EAGAIN */
3348 if (req->flags & REQ_F_NOWAIT)
3351 /* Only for buffered IO */
3352 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3356 * just use poll if we can, and don't attempt if the fs doesn't
3357 * support callback based unlocks
3359 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3362 wait->wait.func = io_async_buf_func;
3363 wait->wait.private = req;
3364 wait->wait.flags = 0;
3365 INIT_LIST_HEAD(&wait->wait.entry);
3366 kiocb->ki_flags |= IOCB_WAITQ;
3367 kiocb->ki_flags &= ~IOCB_NOWAIT;
3368 kiocb->ki_waitq = wait;
3372 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3374 if (req->file->f_op->read_iter)
3375 return call_read_iter(req->file, &req->rw.kiocb, iter);
3376 else if (req->file->f_op->read)
3377 return loop_rw_iter(READ, req, iter);
3382 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3384 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3385 struct kiocb *kiocb = &req->rw.kiocb;
3386 struct iov_iter __iter, *iter = &__iter;
3387 struct io_async_rw *rw = req->async_data;
3388 ssize_t io_size, ret, ret2;
3389 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3395 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3399 io_size = iov_iter_count(iter);
3400 req->result = io_size;
3402 /* Ensure we clear previously set non-block flag */
3403 if (!force_nonblock)
3404 kiocb->ki_flags &= ~IOCB_NOWAIT;
3406 kiocb->ki_flags |= IOCB_NOWAIT;
3408 /* If the file doesn't support async, just async punt */
3409 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3410 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3411 return ret ?: -EAGAIN;
3414 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3415 if (unlikely(ret)) {
3420 ret = io_iter_do_read(req, iter);
3422 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3423 req->flags &= ~REQ_F_REISSUE;
3424 /* IOPOLL retry should happen for io-wq threads */
3425 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3427 /* no retry on NONBLOCK nor RWF_NOWAIT */
3428 if (req->flags & REQ_F_NOWAIT)
3430 /* some cases will consume bytes even on error returns */
3431 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3433 } else if (ret == -EIOCBQUEUED) {
3435 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3436 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3437 /* read all, failed, already did sync or don't want to retry */
3441 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3446 rw = req->async_data;
3447 /* now use our persistent iterator, if we aren't already */
3452 rw->bytes_done += ret;
3453 /* if we can retry, do so with the callbacks armed */
3454 if (!io_rw_should_retry(req)) {
3455 kiocb->ki_flags &= ~IOCB_WAITQ;
3460 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3461 * we get -EIOCBQUEUED, then we'll get a notification when the
3462 * desired page gets unlocked. We can also get a partial read
3463 * here, and if we do, then just retry at the new offset.
3465 ret = io_iter_do_read(req, iter);
3466 if (ret == -EIOCBQUEUED)
3468 /* we got some bytes, but not all. retry. */
3469 kiocb->ki_flags &= ~IOCB_WAITQ;
3470 } while (ret > 0 && ret < io_size);
3472 kiocb_done(kiocb, ret, issue_flags);
3474 /* it's faster to check here then delegate to kfree */
3480 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3482 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3484 return io_prep_rw(req, sqe);
3487 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3489 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3490 struct kiocb *kiocb = &req->rw.kiocb;
3491 struct iov_iter __iter, *iter = &__iter;
3492 struct io_async_rw *rw = req->async_data;
3493 ssize_t ret, ret2, io_size;
3494 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3500 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3504 io_size = iov_iter_count(iter);
3505 req->result = io_size;
3507 /* Ensure we clear previously set non-block flag */
3508 if (!force_nonblock)
3509 kiocb->ki_flags &= ~IOCB_NOWAIT;
3511 kiocb->ki_flags |= IOCB_NOWAIT;
3513 /* If the file doesn't support async, just async punt */
3514 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3517 /* file path doesn't support NOWAIT for non-direct_IO */
3518 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3519 (req->flags & REQ_F_ISREG))
3522 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3527 * Open-code file_start_write here to grab freeze protection,
3528 * which will be released by another thread in
3529 * io_complete_rw(). Fool lockdep by telling it the lock got
3530 * released so that it doesn't complain about the held lock when
3531 * we return to userspace.
3533 if (req->flags & REQ_F_ISREG) {
3534 sb_start_write(file_inode(req->file)->i_sb);
3535 __sb_writers_release(file_inode(req->file)->i_sb,
3538 kiocb->ki_flags |= IOCB_WRITE;
3540 if (req->file->f_op->write_iter)
3541 ret2 = call_write_iter(req->file, kiocb, iter);
3542 else if (req->file->f_op->write)
3543 ret2 = loop_rw_iter(WRITE, req, iter);
3547 if (req->flags & REQ_F_REISSUE) {
3548 req->flags &= ~REQ_F_REISSUE;
3553 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3554 * retry them without IOCB_NOWAIT.
3556 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3558 /* no retry on NONBLOCK nor RWF_NOWAIT */
3559 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3561 if (!force_nonblock || ret2 != -EAGAIN) {
3562 /* IOPOLL retry should happen for io-wq threads */
3563 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3566 kiocb_done(kiocb, ret2, issue_flags);
3569 /* some cases will consume bytes even on error returns */
3570 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3571 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3572 return ret ?: -EAGAIN;
3575 /* it's reportedly faster than delegating the null check to kfree() */
3581 static int io_renameat_prep(struct io_kiocb *req,
3582 const struct io_uring_sqe *sqe)
3584 struct io_rename *ren = &req->rename;
3585 const char __user *oldf, *newf;
3587 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3589 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3591 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3594 ren->old_dfd = READ_ONCE(sqe->fd);
3595 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3596 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3597 ren->new_dfd = READ_ONCE(sqe->len);
3598 ren->flags = READ_ONCE(sqe->rename_flags);
3600 ren->oldpath = getname(oldf);
3601 if (IS_ERR(ren->oldpath))
3602 return PTR_ERR(ren->oldpath);
3604 ren->newpath = getname(newf);
3605 if (IS_ERR(ren->newpath)) {
3606 putname(ren->oldpath);
3607 return PTR_ERR(ren->newpath);
3610 req->flags |= REQ_F_NEED_CLEANUP;
3614 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3616 struct io_rename *ren = &req->rename;
3619 if (issue_flags & IO_URING_F_NONBLOCK)
3622 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3623 ren->newpath, ren->flags);
3625 req->flags &= ~REQ_F_NEED_CLEANUP;
3628 io_req_complete(req, ret);
3632 static int io_unlinkat_prep(struct io_kiocb *req,
3633 const struct io_uring_sqe *sqe)
3635 struct io_unlink *un = &req->unlink;
3636 const char __user *fname;
3638 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3640 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3643 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3646 un->dfd = READ_ONCE(sqe->fd);
3648 un->flags = READ_ONCE(sqe->unlink_flags);
3649 if (un->flags & ~AT_REMOVEDIR)
3652 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3653 un->filename = getname(fname);
3654 if (IS_ERR(un->filename))
3655 return PTR_ERR(un->filename);
3657 req->flags |= REQ_F_NEED_CLEANUP;
3661 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3663 struct io_unlink *un = &req->unlink;
3666 if (issue_flags & IO_URING_F_NONBLOCK)
3669 if (un->flags & AT_REMOVEDIR)
3670 ret = do_rmdir(un->dfd, un->filename);
3672 ret = do_unlinkat(un->dfd, un->filename);
3674 req->flags &= ~REQ_F_NEED_CLEANUP;
3677 io_req_complete(req, ret);
3681 static int io_shutdown_prep(struct io_kiocb *req,
3682 const struct io_uring_sqe *sqe)
3684 #if defined(CONFIG_NET)
3685 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3687 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3688 sqe->buf_index || sqe->splice_fd_in))
3691 req->shutdown.how = READ_ONCE(sqe->len);
3698 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3700 #if defined(CONFIG_NET)
3701 struct socket *sock;
3704 if (issue_flags & IO_URING_F_NONBLOCK)
3707 sock = sock_from_file(req->file);
3708 if (unlikely(!sock))
3711 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3714 io_req_complete(req, ret);
3721 static int __io_splice_prep(struct io_kiocb *req,
3722 const struct io_uring_sqe *sqe)
3724 struct io_splice *sp = &req->splice;
3725 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3727 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3731 sp->len = READ_ONCE(sqe->len);
3732 sp->flags = READ_ONCE(sqe->splice_flags);
3734 if (unlikely(sp->flags & ~valid_flags))
3737 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3738 (sp->flags & SPLICE_F_FD_IN_FIXED));
3741 req->flags |= REQ_F_NEED_CLEANUP;
3745 static int io_tee_prep(struct io_kiocb *req,
3746 const struct io_uring_sqe *sqe)
3748 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3750 return __io_splice_prep(req, sqe);
3753 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3755 struct io_splice *sp = &req->splice;
3756 struct file *in = sp->file_in;
3757 struct file *out = sp->file_out;
3758 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3761 if (issue_flags & IO_URING_F_NONBLOCK)
3764 ret = do_tee(in, out, sp->len, flags);
3766 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3768 req->flags &= ~REQ_F_NEED_CLEANUP;
3772 io_req_complete(req, ret);
3776 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3778 struct io_splice *sp = &req->splice;
3780 sp->off_in = READ_ONCE(sqe->splice_off_in);
3781 sp->off_out = READ_ONCE(sqe->off);
3782 return __io_splice_prep(req, sqe);
3785 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3787 struct io_splice *sp = &req->splice;
3788 struct file *in = sp->file_in;
3789 struct file *out = sp->file_out;
3790 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3791 loff_t *poff_in, *poff_out;
3794 if (issue_flags & IO_URING_F_NONBLOCK)
3797 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3798 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3801 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3803 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3805 req->flags &= ~REQ_F_NEED_CLEANUP;
3809 io_req_complete(req, ret);
3814 * IORING_OP_NOP just posts a completion event, nothing else.
3816 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3818 struct io_ring_ctx *ctx = req->ctx;
3820 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3823 __io_req_complete(req, issue_flags, 0, 0);
3827 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3829 struct io_ring_ctx *ctx = req->ctx;
3834 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3836 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
3840 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3841 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3844 req->sync.off = READ_ONCE(sqe->off);
3845 req->sync.len = READ_ONCE(sqe->len);
3849 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3851 loff_t end = req->sync.off + req->sync.len;
3854 /* fsync always requires a blocking context */
3855 if (issue_flags & IO_URING_F_NONBLOCK)
3858 ret = vfs_fsync_range(req->file, req->sync.off,
3859 end > 0 ? end : LLONG_MAX,
3860 req->sync.flags & IORING_FSYNC_DATASYNC);
3863 io_req_complete(req, ret);
3867 static int io_fallocate_prep(struct io_kiocb *req,
3868 const struct io_uring_sqe *sqe)
3870 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
3873 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3876 req->sync.off = READ_ONCE(sqe->off);
3877 req->sync.len = READ_ONCE(sqe->addr);
3878 req->sync.mode = READ_ONCE(sqe->len);
3882 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3886 /* fallocate always requiring blocking context */
3887 if (issue_flags & IO_URING_F_NONBLOCK)
3889 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3893 io_req_complete(req, ret);
3897 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3899 const char __user *fname;
3902 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3904 if (unlikely(sqe->ioprio || sqe->buf_index))
3906 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3909 /* open.how should be already initialised */
3910 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3911 req->open.how.flags |= O_LARGEFILE;
3913 req->open.dfd = READ_ONCE(sqe->fd);
3914 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3915 req->open.filename = getname(fname);
3916 if (IS_ERR(req->open.filename)) {
3917 ret = PTR_ERR(req->open.filename);
3918 req->open.filename = NULL;
3922 req->open.file_slot = READ_ONCE(sqe->file_index);
3923 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
3926 req->open.nofile = rlimit(RLIMIT_NOFILE);
3927 req->flags |= REQ_F_NEED_CLEANUP;
3931 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3933 u64 mode = READ_ONCE(sqe->len);
3934 u64 flags = READ_ONCE(sqe->open_flags);
3936 req->open.how = build_open_how(flags, mode);
3937 return __io_openat_prep(req, sqe);
3940 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3942 struct open_how __user *how;
3946 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3947 len = READ_ONCE(sqe->len);
3948 if (len < OPEN_HOW_SIZE_VER0)
3951 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3956 return __io_openat_prep(req, sqe);
3959 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3961 struct open_flags op;
3963 bool resolve_nonblock, nonblock_set;
3964 bool fixed = !!req->open.file_slot;
3967 ret = build_open_flags(&req->open.how, &op);
3970 nonblock_set = op.open_flag & O_NONBLOCK;
3971 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3972 if (issue_flags & IO_URING_F_NONBLOCK) {
3974 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3975 * it'll always -EAGAIN
3977 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3979 op.lookup_flags |= LOOKUP_CACHED;
3980 op.open_flag |= O_NONBLOCK;
3984 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3989 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3992 * We could hang on to this 'fd' on retrying, but seems like
3993 * marginal gain for something that is now known to be a slower
3994 * path. So just put it, and we'll get a new one when we retry.
3999 ret = PTR_ERR(file);
4000 /* only retry if RESOLVE_CACHED wasn't already set by application */
4001 if (ret == -EAGAIN &&
4002 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4007 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4008 file->f_flags &= ~O_NONBLOCK;
4009 fsnotify_open(file);
4012 fd_install(ret, file);
4014 ret = io_install_fixed_file(req, file, issue_flags,
4015 req->open.file_slot - 1);
4017 putname(req->open.filename);
4018 req->flags &= ~REQ_F_NEED_CLEANUP;
4021 __io_req_complete(req, issue_flags, ret, 0);
4025 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4027 return io_openat2(req, issue_flags);
4030 static int io_remove_buffers_prep(struct io_kiocb *req,
4031 const struct io_uring_sqe *sqe)
4033 struct io_provide_buf *p = &req->pbuf;
4036 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4040 tmp = READ_ONCE(sqe->fd);
4041 if (!tmp || tmp > USHRT_MAX)
4044 memset(p, 0, sizeof(*p));
4046 p->bgid = READ_ONCE(sqe->buf_group);
4050 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4051 int bgid, unsigned nbufs)
4055 /* shouldn't happen */
4059 /* the head kbuf is the list itself */
4060 while (!list_empty(&buf->list)) {
4061 struct io_buffer *nxt;
4063 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4064 list_del(&nxt->list);
4071 xa_erase(&ctx->io_buffers, bgid);
4076 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4078 struct io_provide_buf *p = &req->pbuf;
4079 struct io_ring_ctx *ctx = req->ctx;
4080 struct io_buffer *head;
4082 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4084 io_ring_submit_lock(ctx, !force_nonblock);
4086 lockdep_assert_held(&ctx->uring_lock);
4089 head = xa_load(&ctx->io_buffers, p->bgid);
4091 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4095 /* complete before unlock, IOPOLL may need the lock */
4096 __io_req_complete(req, issue_flags, ret, 0);
4097 io_ring_submit_unlock(ctx, !force_nonblock);
4101 static int io_provide_buffers_prep(struct io_kiocb *req,
4102 const struct io_uring_sqe *sqe)
4104 unsigned long size, tmp_check;
4105 struct io_provide_buf *p = &req->pbuf;
4108 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4111 tmp = READ_ONCE(sqe->fd);
4112 if (!tmp || tmp > USHRT_MAX)
4115 p->addr = READ_ONCE(sqe->addr);
4116 p->len = READ_ONCE(sqe->len);
4118 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4121 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4124 size = (unsigned long)p->len * p->nbufs;
4125 if (!access_ok(u64_to_user_ptr(p->addr), size))
4128 p->bgid = READ_ONCE(sqe->buf_group);
4129 tmp = READ_ONCE(sqe->off);
4130 if (tmp > USHRT_MAX)
4136 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4138 struct io_buffer *buf;
4139 u64 addr = pbuf->addr;
4140 int i, bid = pbuf->bid;
4142 for (i = 0; i < pbuf->nbufs; i++) {
4143 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4148 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4153 INIT_LIST_HEAD(&buf->list);
4156 list_add_tail(&buf->list, &(*head)->list);
4160 return i ? i : -ENOMEM;
4163 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4165 struct io_provide_buf *p = &req->pbuf;
4166 struct io_ring_ctx *ctx = req->ctx;
4167 struct io_buffer *head, *list;
4169 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4171 io_ring_submit_lock(ctx, !force_nonblock);
4173 lockdep_assert_held(&ctx->uring_lock);
4175 list = head = xa_load(&ctx->io_buffers, p->bgid);
4177 ret = io_add_buffers(p, &head);
4178 if (ret >= 0 && !list) {
4179 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4181 __io_remove_buffers(ctx, head, p->bgid, -1U);
4185 /* complete before unlock, IOPOLL may need the lock */
4186 __io_req_complete(req, issue_flags, ret, 0);
4187 io_ring_submit_unlock(ctx, !force_nonblock);
4191 static int io_epoll_ctl_prep(struct io_kiocb *req,
4192 const struct io_uring_sqe *sqe)
4194 #if defined(CONFIG_EPOLL)
4195 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4197 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4200 req->epoll.epfd = READ_ONCE(sqe->fd);
4201 req->epoll.op = READ_ONCE(sqe->len);
4202 req->epoll.fd = READ_ONCE(sqe->off);
4204 if (ep_op_has_event(req->epoll.op)) {
4205 struct epoll_event __user *ev;
4207 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4208 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4218 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4220 #if defined(CONFIG_EPOLL)
4221 struct io_epoll *ie = &req->epoll;
4223 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4225 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4226 if (force_nonblock && ret == -EAGAIN)
4231 __io_req_complete(req, issue_flags, ret, 0);
4238 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4240 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4241 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4243 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4246 req->madvise.addr = READ_ONCE(sqe->addr);
4247 req->madvise.len = READ_ONCE(sqe->len);
4248 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4255 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4257 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4258 struct io_madvise *ma = &req->madvise;
4261 if (issue_flags & IO_URING_F_NONBLOCK)
4264 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4267 io_req_complete(req, ret);
4274 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4276 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4278 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4281 req->fadvise.offset = READ_ONCE(sqe->off);
4282 req->fadvise.len = READ_ONCE(sqe->len);
4283 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4287 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4289 struct io_fadvise *fa = &req->fadvise;
4292 if (issue_flags & IO_URING_F_NONBLOCK) {
4293 switch (fa->advice) {
4294 case POSIX_FADV_NORMAL:
4295 case POSIX_FADV_RANDOM:
4296 case POSIX_FADV_SEQUENTIAL:
4303 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4306 __io_req_complete(req, issue_flags, ret, 0);
4310 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4312 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4314 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4316 if (req->flags & REQ_F_FIXED_FILE)
4319 req->statx.dfd = READ_ONCE(sqe->fd);
4320 req->statx.mask = READ_ONCE(sqe->len);
4321 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4322 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4323 req->statx.flags = READ_ONCE(sqe->statx_flags);
4328 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4330 struct io_statx *ctx = &req->statx;
4333 if (issue_flags & IO_URING_F_NONBLOCK)
4336 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4341 io_req_complete(req, ret);
4345 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4347 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4349 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4350 sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
4352 if (req->flags & REQ_F_FIXED_FILE)
4355 req->close.fd = READ_ONCE(sqe->fd);
4359 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4361 struct files_struct *files = current->files;
4362 struct io_close *close = &req->close;
4363 struct fdtable *fdt;
4364 struct file *file = NULL;
4367 spin_lock(&files->file_lock);
4368 fdt = files_fdtable(files);
4369 if (close->fd >= fdt->max_fds) {
4370 spin_unlock(&files->file_lock);
4373 file = fdt->fd[close->fd];
4374 if (!file || file->f_op == &io_uring_fops) {
4375 spin_unlock(&files->file_lock);
4380 /* if the file has a flush method, be safe and punt to async */
4381 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4382 spin_unlock(&files->file_lock);
4386 ret = __close_fd_get_file(close->fd, &file);
4387 spin_unlock(&files->file_lock);
4394 /* No ->flush() or already async, safely close from here */
4395 ret = filp_close(file, current->files);
4401 __io_req_complete(req, issue_flags, ret, 0);
4405 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4407 struct io_ring_ctx *ctx = req->ctx;
4409 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4411 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4415 req->sync.off = READ_ONCE(sqe->off);
4416 req->sync.len = READ_ONCE(sqe->len);
4417 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4421 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4425 /* sync_file_range always requires a blocking context */
4426 if (issue_flags & IO_URING_F_NONBLOCK)
4429 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4433 io_req_complete(req, ret);
4437 #if defined(CONFIG_NET)
4438 static int io_setup_async_msg(struct io_kiocb *req,
4439 struct io_async_msghdr *kmsg)
4441 struct io_async_msghdr *async_msg = req->async_data;
4445 if (io_alloc_async_data(req)) {
4446 kfree(kmsg->free_iov);
4449 async_msg = req->async_data;
4450 req->flags |= REQ_F_NEED_CLEANUP;
4451 memcpy(async_msg, kmsg, sizeof(*kmsg));
4452 async_msg->msg.msg_name = &async_msg->addr;
4453 /* if were using fast_iov, set it to the new one */
4454 if (!async_msg->free_iov)
4455 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4460 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4461 struct io_async_msghdr *iomsg)
4463 iomsg->msg.msg_name = &iomsg->addr;
4464 iomsg->free_iov = iomsg->fast_iov;
4465 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4466 req->sr_msg.msg_flags, &iomsg->free_iov);
4469 static int io_sendmsg_prep_async(struct io_kiocb *req)
4473 ret = io_sendmsg_copy_hdr(req, req->async_data);
4475 req->flags |= REQ_F_NEED_CLEANUP;
4479 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4481 struct io_sr_msg *sr = &req->sr_msg;
4483 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4486 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4487 sr->len = READ_ONCE(sqe->len);
4488 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4489 if (sr->msg_flags & MSG_DONTWAIT)
4490 req->flags |= REQ_F_NOWAIT;
4492 #ifdef CONFIG_COMPAT
4493 if (req->ctx->compat)
4494 sr->msg_flags |= MSG_CMSG_COMPAT;
4499 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4501 struct io_async_msghdr iomsg, *kmsg;
4502 struct socket *sock;
4507 sock = sock_from_file(req->file);
4508 if (unlikely(!sock))
4511 kmsg = req->async_data;
4513 ret = io_sendmsg_copy_hdr(req, &iomsg);
4519 flags = req->sr_msg.msg_flags;
4520 if (issue_flags & IO_URING_F_NONBLOCK)
4521 flags |= MSG_DONTWAIT;
4522 if (flags & MSG_WAITALL)
4523 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4525 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4526 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4527 return io_setup_async_msg(req, kmsg);
4528 if (ret == -ERESTARTSYS)
4531 /* fast path, check for non-NULL to avoid function call */
4533 kfree(kmsg->free_iov);
4534 req->flags &= ~REQ_F_NEED_CLEANUP;
4537 __io_req_complete(req, issue_flags, ret, 0);
4541 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4543 struct io_sr_msg *sr = &req->sr_msg;
4546 struct socket *sock;
4551 sock = sock_from_file(req->file);
4552 if (unlikely(!sock))
4555 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4559 msg.msg_name = NULL;
4560 msg.msg_control = NULL;
4561 msg.msg_controllen = 0;
4562 msg.msg_namelen = 0;
4564 flags = req->sr_msg.msg_flags;
4565 if (issue_flags & IO_URING_F_NONBLOCK)
4566 flags |= MSG_DONTWAIT;
4567 if (flags & MSG_WAITALL)
4568 min_ret = iov_iter_count(&msg.msg_iter);
4570 msg.msg_flags = flags;
4571 ret = sock_sendmsg(sock, &msg);
4572 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4574 if (ret == -ERESTARTSYS)
4579 __io_req_complete(req, issue_flags, ret, 0);
4583 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4584 struct io_async_msghdr *iomsg)
4586 struct io_sr_msg *sr = &req->sr_msg;
4587 struct iovec __user *uiov;
4591 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4592 &iomsg->uaddr, &uiov, &iov_len);
4596 if (req->flags & REQ_F_BUFFER_SELECT) {
4599 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4601 sr->len = iomsg->fast_iov[0].iov_len;
4602 iomsg->free_iov = NULL;
4604 iomsg->free_iov = iomsg->fast_iov;
4605 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4606 &iomsg->free_iov, &iomsg->msg.msg_iter,
4615 #ifdef CONFIG_COMPAT
4616 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4617 struct io_async_msghdr *iomsg)
4619 struct io_sr_msg *sr = &req->sr_msg;
4620 struct compat_iovec __user *uiov;
4625 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4630 uiov = compat_ptr(ptr);
4631 if (req->flags & REQ_F_BUFFER_SELECT) {
4632 compat_ssize_t clen;
4636 if (!access_ok(uiov, sizeof(*uiov)))
4638 if (__get_user(clen, &uiov->iov_len))
4643 iomsg->free_iov = NULL;
4645 iomsg->free_iov = iomsg->fast_iov;
4646 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4647 UIO_FASTIOV, &iomsg->free_iov,
4648 &iomsg->msg.msg_iter, true);
4657 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4658 struct io_async_msghdr *iomsg)
4660 iomsg->msg.msg_name = &iomsg->addr;
4662 #ifdef CONFIG_COMPAT
4663 if (req->ctx->compat)
4664 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4667 return __io_recvmsg_copy_hdr(req, iomsg);
4670 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4673 struct io_sr_msg *sr = &req->sr_msg;
4674 struct io_buffer *kbuf;
4676 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4681 req->flags |= REQ_F_BUFFER_SELECTED;
4685 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4687 return io_put_kbuf(req, req->sr_msg.kbuf);
4690 static int io_recvmsg_prep_async(struct io_kiocb *req)
4694 ret = io_recvmsg_copy_hdr(req, req->async_data);
4696 req->flags |= REQ_F_NEED_CLEANUP;
4700 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4702 struct io_sr_msg *sr = &req->sr_msg;
4704 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4707 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4708 sr->len = READ_ONCE(sqe->len);
4709 sr->bgid = READ_ONCE(sqe->buf_group);
4710 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4711 if (sr->msg_flags & MSG_DONTWAIT)
4712 req->flags |= REQ_F_NOWAIT;
4714 #ifdef CONFIG_COMPAT
4715 if (req->ctx->compat)
4716 sr->msg_flags |= MSG_CMSG_COMPAT;
4721 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4723 struct io_async_msghdr iomsg, *kmsg;
4724 struct socket *sock;
4725 struct io_buffer *kbuf;
4728 int ret, cflags = 0;
4729 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4731 sock = sock_from_file(req->file);
4732 if (unlikely(!sock))
4735 kmsg = req->async_data;
4737 ret = io_recvmsg_copy_hdr(req, &iomsg);
4743 if (req->flags & REQ_F_BUFFER_SELECT) {
4744 kbuf = io_recv_buffer_select(req, !force_nonblock);
4746 return PTR_ERR(kbuf);
4747 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4748 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4749 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4750 1, req->sr_msg.len);
4753 flags = req->sr_msg.msg_flags;
4755 flags |= MSG_DONTWAIT;
4756 if (flags & MSG_WAITALL)
4757 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4759 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4760 kmsg->uaddr, flags);
4761 if (force_nonblock && ret == -EAGAIN)
4762 return io_setup_async_msg(req, kmsg);
4763 if (ret == -ERESTARTSYS)
4766 if (req->flags & REQ_F_BUFFER_SELECTED)
4767 cflags = io_put_recv_kbuf(req);
4768 /* fast path, check for non-NULL to avoid function call */
4770 kfree(kmsg->free_iov);
4771 req->flags &= ~REQ_F_NEED_CLEANUP;
4772 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4774 __io_req_complete(req, issue_flags, ret, cflags);
4778 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4780 struct io_buffer *kbuf;
4781 struct io_sr_msg *sr = &req->sr_msg;
4783 void __user *buf = sr->buf;
4784 struct socket *sock;
4788 int ret, cflags = 0;
4789 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4791 sock = sock_from_file(req->file);
4792 if (unlikely(!sock))
4795 if (req->flags & REQ_F_BUFFER_SELECT) {
4796 kbuf = io_recv_buffer_select(req, !force_nonblock);
4798 return PTR_ERR(kbuf);
4799 buf = u64_to_user_ptr(kbuf->addr);
4802 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4806 msg.msg_name = NULL;
4807 msg.msg_control = NULL;
4808 msg.msg_controllen = 0;
4809 msg.msg_namelen = 0;
4810 msg.msg_iocb = NULL;
4813 flags = req->sr_msg.msg_flags;
4815 flags |= MSG_DONTWAIT;
4816 if (flags & MSG_WAITALL)
4817 min_ret = iov_iter_count(&msg.msg_iter);
4819 ret = sock_recvmsg(sock, &msg, flags);
4820 if (force_nonblock && ret == -EAGAIN)
4822 if (ret == -ERESTARTSYS)
4825 if (req->flags & REQ_F_BUFFER_SELECTED)
4826 cflags = io_put_recv_kbuf(req);
4827 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4829 __io_req_complete(req, issue_flags, ret, cflags);
4833 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4835 struct io_accept *accept = &req->accept;
4837 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4839 if (sqe->ioprio || sqe->len || sqe->buf_index)
4842 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4843 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4844 accept->flags = READ_ONCE(sqe->accept_flags);
4845 accept->nofile = rlimit(RLIMIT_NOFILE);
4847 accept->file_slot = READ_ONCE(sqe->file_index);
4848 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
4849 (accept->flags & SOCK_CLOEXEC)))
4851 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
4853 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
4854 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
4858 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4860 struct io_accept *accept = &req->accept;
4861 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4862 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4863 bool fixed = !!accept->file_slot;
4867 if (req->file->f_flags & O_NONBLOCK)
4868 req->flags |= REQ_F_NOWAIT;
4871 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
4872 if (unlikely(fd < 0))
4875 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
4880 ret = PTR_ERR(file);
4881 if (ret == -EAGAIN && force_nonblock)
4883 if (ret == -ERESTARTSYS)
4886 } else if (!fixed) {
4887 fd_install(fd, file);
4890 ret = io_install_fixed_file(req, file, issue_flags,
4891 accept->file_slot - 1);
4893 __io_req_complete(req, issue_flags, ret, 0);
4897 static int io_connect_prep_async(struct io_kiocb *req)
4899 struct io_async_connect *io = req->async_data;
4900 struct io_connect *conn = &req->connect;
4902 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4905 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4907 struct io_connect *conn = &req->connect;
4909 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4911 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
4915 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4916 conn->addr_len = READ_ONCE(sqe->addr2);
4920 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4922 struct io_async_connect __io, *io;
4923 unsigned file_flags;
4925 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4927 if (req->async_data) {
4928 io = req->async_data;
4930 ret = move_addr_to_kernel(req->connect.addr,
4931 req->connect.addr_len,
4938 file_flags = force_nonblock ? O_NONBLOCK : 0;
4940 ret = __sys_connect_file(req->file, &io->address,
4941 req->connect.addr_len, file_flags);
4942 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4943 if (req->async_data)
4945 if (io_alloc_async_data(req)) {
4949 memcpy(req->async_data, &__io, sizeof(__io));
4952 if (ret == -ERESTARTSYS)
4957 __io_req_complete(req, issue_flags, ret, 0);
4960 #else /* !CONFIG_NET */
4961 #define IO_NETOP_FN(op) \
4962 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4964 return -EOPNOTSUPP; \
4967 #define IO_NETOP_PREP(op) \
4969 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4971 return -EOPNOTSUPP; \
4974 #define IO_NETOP_PREP_ASYNC(op) \
4976 static int io_##op##_prep_async(struct io_kiocb *req) \
4978 return -EOPNOTSUPP; \
4981 IO_NETOP_PREP_ASYNC(sendmsg);
4982 IO_NETOP_PREP_ASYNC(recvmsg);
4983 IO_NETOP_PREP_ASYNC(connect);
4984 IO_NETOP_PREP(accept);
4987 #endif /* CONFIG_NET */
4989 struct io_poll_table {
4990 struct poll_table_struct pt;
4991 struct io_kiocb *req;
4996 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4997 __poll_t mask, io_req_tw_func_t func)
4999 /* for instances that support it check for an event match first: */
5000 if (mask && !(mask & poll->events))
5003 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5005 list_del_init(&poll->wait.entry);
5008 req->io_task_work.func = func;
5011 * If this fails, then the task is exiting. When a task exits, the
5012 * work gets canceled, so just cancel this request as well instead
5013 * of executing it. We can't safely execute it anyway, as we may not
5014 * have the needed state needed for it anyway.
5016 io_req_task_work_add(req);
5020 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5021 __acquires(&req->ctx->completion_lock)
5023 struct io_ring_ctx *ctx = req->ctx;
5025 /* req->task == current here, checking PF_EXITING is safe */
5026 if (unlikely(req->task->flags & PF_EXITING))
5027 WRITE_ONCE(poll->canceled, true);
5029 if (!req->result && !READ_ONCE(poll->canceled)) {
5030 struct poll_table_struct pt = { ._key = poll->events };
5032 req->result = vfs_poll(req->file, &pt) & poll->events;
5035 spin_lock(&ctx->completion_lock);
5036 if (!req->result && !READ_ONCE(poll->canceled)) {
5037 add_wait_queue(poll->head, &poll->wait);
5044 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5046 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5047 if (req->opcode == IORING_OP_POLL_ADD)
5048 return req->async_data;
5049 return req->apoll->double_poll;
5052 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5054 if (req->opcode == IORING_OP_POLL_ADD)
5056 return &req->apoll->poll;
5059 static void io_poll_remove_double(struct io_kiocb *req)
5060 __must_hold(&req->ctx->completion_lock)
5062 struct io_poll_iocb *poll = io_poll_get_double(req);
5064 lockdep_assert_held(&req->ctx->completion_lock);
5066 if (poll && poll->head) {
5067 struct wait_queue_head *head = poll->head;
5069 spin_lock_irq(&head->lock);
5070 list_del_init(&poll->wait.entry);
5071 if (poll->wait.private)
5074 spin_unlock_irq(&head->lock);
5078 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5079 __must_hold(&req->ctx->completion_lock)
5081 struct io_ring_ctx *ctx = req->ctx;
5082 unsigned flags = IORING_CQE_F_MORE;
5085 if (READ_ONCE(req->poll.canceled)) {
5087 req->poll.events |= EPOLLONESHOT;
5089 error = mangle_poll(mask);
5091 if (req->poll.events & EPOLLONESHOT)
5093 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5094 req->poll.done = true;
5097 if (flags & IORING_CQE_F_MORE)
5100 io_commit_cqring(ctx);
5101 return !(flags & IORING_CQE_F_MORE);
5104 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5106 struct io_ring_ctx *ctx = req->ctx;
5107 struct io_kiocb *nxt;
5109 if (io_poll_rewait(req, &req->poll)) {
5110 spin_unlock(&ctx->completion_lock);
5114 done = io_poll_complete(req, req->result);
5116 io_poll_remove_double(req);
5117 hash_del(&req->hash_node);
5120 add_wait_queue(req->poll.head, &req->poll.wait);
5122 spin_unlock(&ctx->completion_lock);
5123 io_cqring_ev_posted(ctx);
5126 nxt = io_put_req_find_next(req);
5128 io_req_task_submit(nxt, locked);
5133 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5134 int sync, void *key)
5136 struct io_kiocb *req = wait->private;
5137 struct io_poll_iocb *poll = io_poll_get_single(req);
5138 __poll_t mask = key_to_poll(key);
5139 unsigned long flags;
5141 /* for instances that support it check for an event match first: */
5142 if (mask && !(mask & poll->events))
5144 if (!(poll->events & EPOLLONESHOT))
5145 return poll->wait.func(&poll->wait, mode, sync, key);
5147 list_del_init(&wait->entry);
5152 spin_lock_irqsave(&poll->head->lock, flags);
5153 done = list_empty(&poll->wait.entry);
5155 list_del_init(&poll->wait.entry);
5156 /* make sure double remove sees this as being gone */
5157 wait->private = NULL;
5158 spin_unlock_irqrestore(&poll->head->lock, flags);
5160 /* use wait func handler, so it matches the rq type */
5161 poll->wait.func(&poll->wait, mode, sync, key);
5168 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5169 wait_queue_func_t wake_func)
5173 poll->canceled = false;
5174 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5175 /* mask in events that we always want/need */
5176 poll->events = events | IO_POLL_UNMASK;
5177 INIT_LIST_HEAD(&poll->wait.entry);
5178 init_waitqueue_func_entry(&poll->wait, wake_func);
5181 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5182 struct wait_queue_head *head,
5183 struct io_poll_iocb **poll_ptr)
5185 struct io_kiocb *req = pt->req;
5188 * The file being polled uses multiple waitqueues for poll handling
5189 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5192 if (unlikely(pt->nr_entries)) {
5193 struct io_poll_iocb *poll_one = poll;
5195 /* double add on the same waitqueue head, ignore */
5196 if (poll_one->head == head)
5198 /* already have a 2nd entry, fail a third attempt */
5200 if ((*poll_ptr)->head == head)
5202 pt->error = -EINVAL;
5206 * Can't handle multishot for double wait for now, turn it
5207 * into one-shot mode.
5209 if (!(poll_one->events & EPOLLONESHOT))
5210 poll_one->events |= EPOLLONESHOT;
5211 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5213 pt->error = -ENOMEM;
5216 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5218 poll->wait.private = req;
5225 if (poll->events & EPOLLEXCLUSIVE)
5226 add_wait_queue_exclusive(head, &poll->wait);
5228 add_wait_queue(head, &poll->wait);
5231 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5232 struct poll_table_struct *p)
5234 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5235 struct async_poll *apoll = pt->req->apoll;
5237 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5240 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5242 struct async_poll *apoll = req->apoll;
5243 struct io_ring_ctx *ctx = req->ctx;
5245 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5247 if (io_poll_rewait(req, &apoll->poll)) {
5248 spin_unlock(&ctx->completion_lock);
5252 hash_del(&req->hash_node);
5253 io_poll_remove_double(req);
5254 spin_unlock(&ctx->completion_lock);
5256 if (!READ_ONCE(apoll->poll.canceled))
5257 io_req_task_submit(req, locked);
5259 io_req_complete_failed(req, -ECANCELED);
5262 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5265 struct io_kiocb *req = wait->private;
5266 struct io_poll_iocb *poll = &req->apoll->poll;
5268 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5271 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5274 static void io_poll_req_insert(struct io_kiocb *req)
5276 struct io_ring_ctx *ctx = req->ctx;
5277 struct hlist_head *list;
5279 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5280 hlist_add_head(&req->hash_node, list);
5283 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5284 struct io_poll_iocb *poll,
5285 struct io_poll_table *ipt, __poll_t mask,
5286 wait_queue_func_t wake_func)
5287 __acquires(&ctx->completion_lock)
5289 struct io_ring_ctx *ctx = req->ctx;
5290 bool cancel = false;
5292 INIT_HLIST_NODE(&req->hash_node);
5293 io_init_poll_iocb(poll, mask, wake_func);
5294 poll->file = req->file;
5295 poll->wait.private = req;
5297 ipt->pt._key = mask;
5300 ipt->nr_entries = 0;
5302 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5303 if (unlikely(!ipt->nr_entries) && !ipt->error)
5304 ipt->error = -EINVAL;
5306 spin_lock(&ctx->completion_lock);
5307 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5308 io_poll_remove_double(req);
5309 if (likely(poll->head)) {
5310 spin_lock_irq(&poll->head->lock);
5311 if (unlikely(list_empty(&poll->wait.entry))) {
5317 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5318 list_del_init(&poll->wait.entry);
5320 WRITE_ONCE(poll->canceled, true);
5321 else if (!poll->done) /* actually waiting for an event */
5322 io_poll_req_insert(req);
5323 spin_unlock_irq(&poll->head->lock);
5335 static int io_arm_poll_handler(struct io_kiocb *req)
5337 const struct io_op_def *def = &io_op_defs[req->opcode];
5338 struct io_ring_ctx *ctx = req->ctx;
5339 struct async_poll *apoll;
5340 struct io_poll_table ipt;
5341 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5344 if (!req->file || !file_can_poll(req->file))
5345 return IO_APOLL_ABORTED;
5346 if (req->flags & REQ_F_POLLED)
5347 return IO_APOLL_ABORTED;
5348 if (!def->pollin && !def->pollout)
5349 return IO_APOLL_ABORTED;
5353 mask |= POLLIN | POLLRDNORM;
5355 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5356 if ((req->opcode == IORING_OP_RECVMSG) &&
5357 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5361 mask |= POLLOUT | POLLWRNORM;
5364 /* if we can't nonblock try, then no point in arming a poll handler */
5365 if (!io_file_supports_nowait(req, rw))
5366 return IO_APOLL_ABORTED;
5368 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5369 if (unlikely(!apoll))
5370 return IO_APOLL_ABORTED;
5371 apoll->double_poll = NULL;
5373 req->flags |= REQ_F_POLLED;
5374 ipt.pt._qproc = io_async_queue_proc;
5375 io_req_set_refcount(req);
5377 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5379 spin_unlock(&ctx->completion_lock);
5380 if (ret || ipt.error)
5381 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5383 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5384 mask, apoll->poll.events);
5388 static bool __io_poll_remove_one(struct io_kiocb *req,
5389 struct io_poll_iocb *poll, bool do_cancel)
5390 __must_hold(&req->ctx->completion_lock)
5392 bool do_complete = false;
5396 spin_lock_irq(&poll->head->lock);
5398 WRITE_ONCE(poll->canceled, true);
5399 if (!list_empty(&poll->wait.entry)) {
5400 list_del_init(&poll->wait.entry);
5403 spin_unlock_irq(&poll->head->lock);
5404 hash_del(&req->hash_node);
5408 static bool io_poll_remove_one(struct io_kiocb *req)
5409 __must_hold(&req->ctx->completion_lock)
5413 io_poll_remove_double(req);
5414 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5417 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5418 io_commit_cqring(req->ctx);
5420 io_put_req_deferred(req);
5426 * Returns true if we found and killed one or more poll requests
5428 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5431 struct hlist_node *tmp;
5432 struct io_kiocb *req;
5435 spin_lock(&ctx->completion_lock);
5436 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5437 struct hlist_head *list;
5439 list = &ctx->cancel_hash[i];
5440 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5441 if (io_match_task(req, tsk, cancel_all))
5442 posted += io_poll_remove_one(req);
5445 spin_unlock(&ctx->completion_lock);
5448 io_cqring_ev_posted(ctx);
5453 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5455 __must_hold(&ctx->completion_lock)
5457 struct hlist_head *list;
5458 struct io_kiocb *req;
5460 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5461 hlist_for_each_entry(req, list, hash_node) {
5462 if (sqe_addr != req->user_data)
5464 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5471 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5473 __must_hold(&ctx->completion_lock)
5475 struct io_kiocb *req;
5477 req = io_poll_find(ctx, sqe_addr, poll_only);
5480 if (io_poll_remove_one(req))
5486 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5491 events = READ_ONCE(sqe->poll32_events);
5493 events = swahw32(events);
5495 if (!(flags & IORING_POLL_ADD_MULTI))
5496 events |= EPOLLONESHOT;
5497 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5500 static int io_poll_update_prep(struct io_kiocb *req,
5501 const struct io_uring_sqe *sqe)
5503 struct io_poll_update *upd = &req->poll_update;
5506 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5508 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5510 flags = READ_ONCE(sqe->len);
5511 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5512 IORING_POLL_ADD_MULTI))
5514 /* meaningless without update */
5515 if (flags == IORING_POLL_ADD_MULTI)
5518 upd->old_user_data = READ_ONCE(sqe->addr);
5519 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5520 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5522 upd->new_user_data = READ_ONCE(sqe->off);
5523 if (!upd->update_user_data && upd->new_user_data)
5525 if (upd->update_events)
5526 upd->events = io_poll_parse_events(sqe, flags);
5527 else if (sqe->poll32_events)
5533 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5536 struct io_kiocb *req = wait->private;
5537 struct io_poll_iocb *poll = &req->poll;
5539 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5542 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5543 struct poll_table_struct *p)
5545 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5547 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5550 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5552 struct io_poll_iocb *poll = &req->poll;
5555 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5557 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5559 flags = READ_ONCE(sqe->len);
5560 if (flags & ~IORING_POLL_ADD_MULTI)
5563 io_req_set_refcount(req);
5564 poll->events = io_poll_parse_events(sqe, flags);
5568 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5570 struct io_poll_iocb *poll = &req->poll;
5571 struct io_ring_ctx *ctx = req->ctx;
5572 struct io_poll_table ipt;
5575 ipt.pt._qproc = io_poll_queue_proc;
5577 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5580 if (mask) { /* no async, we'd stolen it */
5582 io_poll_complete(req, mask);
5584 spin_unlock(&ctx->completion_lock);
5587 io_cqring_ev_posted(ctx);
5588 if (poll->events & EPOLLONESHOT)
5594 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5596 struct io_ring_ctx *ctx = req->ctx;
5597 struct io_kiocb *preq;
5601 spin_lock(&ctx->completion_lock);
5602 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5608 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5610 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5615 * Don't allow racy completion with singleshot, as we cannot safely
5616 * update those. For multishot, if we're racing with completion, just
5617 * let completion re-add it.
5619 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5620 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5624 /* we now have a detached poll request. reissue. */
5628 spin_unlock(&ctx->completion_lock);
5630 io_req_complete(req, ret);
5633 /* only mask one event flags, keep behavior flags */
5634 if (req->poll_update.update_events) {
5635 preq->poll.events &= ~0xffff;
5636 preq->poll.events |= req->poll_update.events & 0xffff;
5637 preq->poll.events |= IO_POLL_UNMASK;
5639 if (req->poll_update.update_user_data)
5640 preq->user_data = req->poll_update.new_user_data;
5641 spin_unlock(&ctx->completion_lock);
5643 /* complete update request, we're done with it */
5644 io_req_complete(req, ret);
5647 ret = io_poll_add(preq, issue_flags);
5650 io_req_complete(preq, ret);
5656 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5659 io_req_complete_post(req, -ETIME, 0);
5662 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5664 struct io_timeout_data *data = container_of(timer,
5665 struct io_timeout_data, timer);
5666 struct io_kiocb *req = data->req;
5667 struct io_ring_ctx *ctx = req->ctx;
5668 unsigned long flags;
5670 spin_lock_irqsave(&ctx->timeout_lock, flags);
5671 list_del_init(&req->timeout.list);
5672 atomic_set(&req->ctx->cq_timeouts,
5673 atomic_read(&req->ctx->cq_timeouts) + 1);
5674 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5676 req->io_task_work.func = io_req_task_timeout;
5677 io_req_task_work_add(req);
5678 return HRTIMER_NORESTART;
5681 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5683 __must_hold(&ctx->timeout_lock)
5685 struct io_timeout_data *io;
5686 struct io_kiocb *req;
5689 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5690 found = user_data == req->user_data;
5695 return ERR_PTR(-ENOENT);
5697 io = req->async_data;
5698 if (hrtimer_try_to_cancel(&io->timer) == -1)
5699 return ERR_PTR(-EALREADY);
5700 list_del_init(&req->timeout.list);
5704 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5705 __must_hold(&ctx->completion_lock)
5706 __must_hold(&ctx->timeout_lock)
5708 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5711 return PTR_ERR(req);
5714 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5715 io_put_req_deferred(req);
5719 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
5721 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
5722 case IORING_TIMEOUT_BOOTTIME:
5723 return CLOCK_BOOTTIME;
5724 case IORING_TIMEOUT_REALTIME:
5725 return CLOCK_REALTIME;
5727 /* can't happen, vetted at prep time */
5731 return CLOCK_MONOTONIC;
5735 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5736 struct timespec64 *ts, enum hrtimer_mode mode)
5737 __must_hold(&ctx->timeout_lock)
5739 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5740 struct io_timeout_data *data;
5743 return PTR_ERR(req);
5745 req->timeout.off = 0; /* noseq */
5746 data = req->async_data;
5747 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5748 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
5749 data->timer.function = io_timeout_fn;
5750 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5754 static int io_timeout_remove_prep(struct io_kiocb *req,
5755 const struct io_uring_sqe *sqe)
5757 struct io_timeout_rem *tr = &req->timeout_rem;
5759 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5761 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5763 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
5766 tr->addr = READ_ONCE(sqe->addr);
5767 tr->flags = READ_ONCE(sqe->timeout_flags);
5768 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5769 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5771 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5773 } else if (tr->flags) {
5774 /* timeout removal doesn't support flags */
5781 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5783 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5788 * Remove or update an existing timeout command
5790 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5792 struct io_timeout_rem *tr = &req->timeout_rem;
5793 struct io_ring_ctx *ctx = req->ctx;
5796 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
5797 spin_lock(&ctx->completion_lock);
5798 spin_lock_irq(&ctx->timeout_lock);
5799 ret = io_timeout_cancel(ctx, tr->addr);
5800 spin_unlock_irq(&ctx->timeout_lock);
5801 spin_unlock(&ctx->completion_lock);
5803 spin_lock_irq(&ctx->timeout_lock);
5804 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5805 io_translate_timeout_mode(tr->flags));
5806 spin_unlock_irq(&ctx->timeout_lock);
5811 io_req_complete_post(req, ret, 0);
5815 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5816 bool is_timeout_link)
5818 struct io_timeout_data *data;
5820 u32 off = READ_ONCE(sqe->off);
5822 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5824 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
5827 if (off && is_timeout_link)
5829 flags = READ_ONCE(sqe->timeout_flags);
5830 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
5832 /* more than one clock specified is invalid, obviously */
5833 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
5836 INIT_LIST_HEAD(&req->timeout.list);
5837 req->timeout.off = off;
5838 if (unlikely(off && !req->ctx->off_timeout_used))
5839 req->ctx->off_timeout_used = true;
5841 if (!req->async_data && io_alloc_async_data(req))
5844 data = req->async_data;
5846 data->flags = flags;
5848 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5851 data->mode = io_translate_timeout_mode(flags);
5852 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
5854 if (is_timeout_link) {
5855 struct io_submit_link *link = &req->ctx->submit_state.link;
5859 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5861 req->timeout.head = link->last;
5862 link->last->flags |= REQ_F_ARM_LTIMEOUT;
5867 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5869 struct io_ring_ctx *ctx = req->ctx;
5870 struct io_timeout_data *data = req->async_data;
5871 struct list_head *entry;
5872 u32 tail, off = req->timeout.off;
5874 spin_lock_irq(&ctx->timeout_lock);
5877 * sqe->off holds how many events that need to occur for this
5878 * timeout event to be satisfied. If it isn't set, then this is
5879 * a pure timeout request, sequence isn't used.
5881 if (io_is_timeout_noseq(req)) {
5882 entry = ctx->timeout_list.prev;
5886 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5887 req->timeout.target_seq = tail + off;
5889 /* Update the last seq here in case io_flush_timeouts() hasn't.
5890 * This is safe because ->completion_lock is held, and submissions
5891 * and completions are never mixed in the same ->completion_lock section.
5893 ctx->cq_last_tm_flush = tail;
5896 * Insertion sort, ensuring the first entry in the list is always
5897 * the one we need first.
5899 list_for_each_prev(entry, &ctx->timeout_list) {
5900 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5903 if (io_is_timeout_noseq(nxt))
5905 /* nxt.seq is behind @tail, otherwise would've been completed */
5906 if (off >= nxt->timeout.target_seq - tail)
5910 list_add(&req->timeout.list, entry);
5911 data->timer.function = io_timeout_fn;
5912 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5913 spin_unlock_irq(&ctx->timeout_lock);
5917 struct io_cancel_data {
5918 struct io_ring_ctx *ctx;
5922 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5924 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5925 struct io_cancel_data *cd = data;
5927 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5930 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5931 struct io_ring_ctx *ctx)
5933 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5934 enum io_wq_cancel cancel_ret;
5937 if (!tctx || !tctx->io_wq)
5940 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5941 switch (cancel_ret) {
5942 case IO_WQ_CANCEL_OK:
5945 case IO_WQ_CANCEL_RUNNING:
5948 case IO_WQ_CANCEL_NOTFOUND:
5956 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
5958 struct io_ring_ctx *ctx = req->ctx;
5961 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
5963 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5967 spin_lock(&ctx->completion_lock);
5968 spin_lock_irq(&ctx->timeout_lock);
5969 ret = io_timeout_cancel(ctx, sqe_addr);
5970 spin_unlock_irq(&ctx->timeout_lock);
5973 ret = io_poll_cancel(ctx, sqe_addr, false);
5975 spin_unlock(&ctx->completion_lock);
5979 static int io_async_cancel_prep(struct io_kiocb *req,
5980 const struct io_uring_sqe *sqe)
5982 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5984 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5986 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
5990 req->cancel.addr = READ_ONCE(sqe->addr);
5994 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5996 struct io_ring_ctx *ctx = req->ctx;
5997 u64 sqe_addr = req->cancel.addr;
5998 struct io_tctx_node *node;
6001 ret = io_try_cancel_userdata(req, sqe_addr);
6005 /* slow path, try all io-wq's */
6006 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6008 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6009 struct io_uring_task *tctx = node->task->io_uring;
6011 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6015 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6019 io_req_complete_post(req, ret, 0);
6023 static int io_rsrc_update_prep(struct io_kiocb *req,
6024 const struct io_uring_sqe *sqe)
6026 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6028 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6031 req->rsrc_update.offset = READ_ONCE(sqe->off);
6032 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6033 if (!req->rsrc_update.nr_args)
6035 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6039 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6041 struct io_ring_ctx *ctx = req->ctx;
6042 struct io_uring_rsrc_update2 up;
6045 if (issue_flags & IO_URING_F_NONBLOCK)
6048 up.offset = req->rsrc_update.offset;
6049 up.data = req->rsrc_update.arg;
6054 mutex_lock(&ctx->uring_lock);
6055 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6056 &up, req->rsrc_update.nr_args);
6057 mutex_unlock(&ctx->uring_lock);
6061 __io_req_complete(req, issue_flags, ret, 0);
6065 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6067 switch (req->opcode) {
6070 case IORING_OP_READV:
6071 case IORING_OP_READ_FIXED:
6072 case IORING_OP_READ:
6073 return io_read_prep(req, sqe);
6074 case IORING_OP_WRITEV:
6075 case IORING_OP_WRITE_FIXED:
6076 case IORING_OP_WRITE:
6077 return io_write_prep(req, sqe);
6078 case IORING_OP_POLL_ADD:
6079 return io_poll_add_prep(req, sqe);
6080 case IORING_OP_POLL_REMOVE:
6081 return io_poll_update_prep(req, sqe);
6082 case IORING_OP_FSYNC:
6083 return io_fsync_prep(req, sqe);
6084 case IORING_OP_SYNC_FILE_RANGE:
6085 return io_sfr_prep(req, sqe);
6086 case IORING_OP_SENDMSG:
6087 case IORING_OP_SEND:
6088 return io_sendmsg_prep(req, sqe);
6089 case IORING_OP_RECVMSG:
6090 case IORING_OP_RECV:
6091 return io_recvmsg_prep(req, sqe);
6092 case IORING_OP_CONNECT:
6093 return io_connect_prep(req, sqe);
6094 case IORING_OP_TIMEOUT:
6095 return io_timeout_prep(req, sqe, false);
6096 case IORING_OP_TIMEOUT_REMOVE:
6097 return io_timeout_remove_prep(req, sqe);
6098 case IORING_OP_ASYNC_CANCEL:
6099 return io_async_cancel_prep(req, sqe);
6100 case IORING_OP_LINK_TIMEOUT:
6101 return io_timeout_prep(req, sqe, true);
6102 case IORING_OP_ACCEPT:
6103 return io_accept_prep(req, sqe);
6104 case IORING_OP_FALLOCATE:
6105 return io_fallocate_prep(req, sqe);
6106 case IORING_OP_OPENAT:
6107 return io_openat_prep(req, sqe);
6108 case IORING_OP_CLOSE:
6109 return io_close_prep(req, sqe);
6110 case IORING_OP_FILES_UPDATE:
6111 return io_rsrc_update_prep(req, sqe);
6112 case IORING_OP_STATX:
6113 return io_statx_prep(req, sqe);
6114 case IORING_OP_FADVISE:
6115 return io_fadvise_prep(req, sqe);
6116 case IORING_OP_MADVISE:
6117 return io_madvise_prep(req, sqe);
6118 case IORING_OP_OPENAT2:
6119 return io_openat2_prep(req, sqe);
6120 case IORING_OP_EPOLL_CTL:
6121 return io_epoll_ctl_prep(req, sqe);
6122 case IORING_OP_SPLICE:
6123 return io_splice_prep(req, sqe);
6124 case IORING_OP_PROVIDE_BUFFERS:
6125 return io_provide_buffers_prep(req, sqe);
6126 case IORING_OP_REMOVE_BUFFERS:
6127 return io_remove_buffers_prep(req, sqe);
6129 return io_tee_prep(req, sqe);
6130 case IORING_OP_SHUTDOWN:
6131 return io_shutdown_prep(req, sqe);
6132 case IORING_OP_RENAMEAT:
6133 return io_renameat_prep(req, sqe);
6134 case IORING_OP_UNLINKAT:
6135 return io_unlinkat_prep(req, sqe);
6138 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6143 static int io_req_prep_async(struct io_kiocb *req)
6145 if (!io_op_defs[req->opcode].needs_async_setup)
6147 if (WARN_ON_ONCE(req->async_data))
6149 if (io_alloc_async_data(req))
6152 switch (req->opcode) {
6153 case IORING_OP_READV:
6154 return io_rw_prep_async(req, READ);
6155 case IORING_OP_WRITEV:
6156 return io_rw_prep_async(req, WRITE);
6157 case IORING_OP_SENDMSG:
6158 return io_sendmsg_prep_async(req);
6159 case IORING_OP_RECVMSG:
6160 return io_recvmsg_prep_async(req);
6161 case IORING_OP_CONNECT:
6162 return io_connect_prep_async(req);
6164 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6169 static u32 io_get_sequence(struct io_kiocb *req)
6171 u32 seq = req->ctx->cached_sq_head;
6173 /* need original cached_sq_head, but it was increased for each req */
6174 io_for_each_link(req, req)
6179 static bool io_drain_req(struct io_kiocb *req)
6181 struct io_kiocb *pos;
6182 struct io_ring_ctx *ctx = req->ctx;
6183 struct io_defer_entry *de;
6188 * If we need to drain a request in the middle of a link, drain the
6189 * head request and the next request/link after the current link.
6190 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6191 * maintained for every request of our link.
6193 if (ctx->drain_next) {
6194 req->flags |= REQ_F_IO_DRAIN;
6195 ctx->drain_next = false;
6197 /* not interested in head, start from the first linked */
6198 io_for_each_link(pos, req->link) {
6199 if (pos->flags & REQ_F_IO_DRAIN) {
6200 ctx->drain_next = true;
6201 req->flags |= REQ_F_IO_DRAIN;
6206 /* Still need defer if there is pending req in defer list. */
6207 if (likely(list_empty_careful(&ctx->defer_list) &&
6208 !(req->flags & REQ_F_IO_DRAIN))) {
6209 ctx->drain_active = false;
6213 seq = io_get_sequence(req);
6214 /* Still a chance to pass the sequence check */
6215 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6218 ret = io_req_prep_async(req);
6221 io_prep_async_link(req);
6222 de = kmalloc(sizeof(*de), GFP_KERNEL);
6226 io_req_complete_failed(req, ret);
6230 spin_lock(&ctx->completion_lock);
6231 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6232 spin_unlock(&ctx->completion_lock);
6234 io_queue_async_work(req, NULL);
6238 trace_io_uring_defer(ctx, req, req->user_data);
6241 list_add_tail(&de->list, &ctx->defer_list);
6242 spin_unlock(&ctx->completion_lock);
6246 static void io_clean_op(struct io_kiocb *req)
6248 if (req->flags & REQ_F_BUFFER_SELECTED) {
6249 switch (req->opcode) {
6250 case IORING_OP_READV:
6251 case IORING_OP_READ_FIXED:
6252 case IORING_OP_READ:
6253 kfree((void *)(unsigned long)req->rw.addr);
6255 case IORING_OP_RECVMSG:
6256 case IORING_OP_RECV:
6257 kfree(req->sr_msg.kbuf);
6262 if (req->flags & REQ_F_NEED_CLEANUP) {
6263 switch (req->opcode) {
6264 case IORING_OP_READV:
6265 case IORING_OP_READ_FIXED:
6266 case IORING_OP_READ:
6267 case IORING_OP_WRITEV:
6268 case IORING_OP_WRITE_FIXED:
6269 case IORING_OP_WRITE: {
6270 struct io_async_rw *io = req->async_data;
6272 kfree(io->free_iovec);
6275 case IORING_OP_RECVMSG:
6276 case IORING_OP_SENDMSG: {
6277 struct io_async_msghdr *io = req->async_data;
6279 kfree(io->free_iov);
6282 case IORING_OP_SPLICE:
6284 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6285 io_put_file(req->splice.file_in);
6287 case IORING_OP_OPENAT:
6288 case IORING_OP_OPENAT2:
6289 if (req->open.filename)
6290 putname(req->open.filename);
6292 case IORING_OP_RENAMEAT:
6293 putname(req->rename.oldpath);
6294 putname(req->rename.newpath);
6296 case IORING_OP_UNLINKAT:
6297 putname(req->unlink.filename);
6301 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6302 kfree(req->apoll->double_poll);
6306 if (req->flags & REQ_F_INFLIGHT) {
6307 struct io_uring_task *tctx = req->task->io_uring;
6309 atomic_dec(&tctx->inflight_tracked);
6311 if (req->flags & REQ_F_CREDS)
6312 put_cred(req->creds);
6314 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6317 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6319 struct io_ring_ctx *ctx = req->ctx;
6320 const struct cred *creds = NULL;
6323 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6324 creds = override_creds(req->creds);
6326 switch (req->opcode) {
6328 ret = io_nop(req, issue_flags);
6330 case IORING_OP_READV:
6331 case IORING_OP_READ_FIXED:
6332 case IORING_OP_READ:
6333 ret = io_read(req, issue_flags);
6335 case IORING_OP_WRITEV:
6336 case IORING_OP_WRITE_FIXED:
6337 case IORING_OP_WRITE:
6338 ret = io_write(req, issue_flags);
6340 case IORING_OP_FSYNC:
6341 ret = io_fsync(req, issue_flags);
6343 case IORING_OP_POLL_ADD:
6344 ret = io_poll_add(req, issue_flags);
6346 case IORING_OP_POLL_REMOVE:
6347 ret = io_poll_update(req, issue_flags);
6349 case IORING_OP_SYNC_FILE_RANGE:
6350 ret = io_sync_file_range(req, issue_flags);
6352 case IORING_OP_SENDMSG:
6353 ret = io_sendmsg(req, issue_flags);
6355 case IORING_OP_SEND:
6356 ret = io_send(req, issue_flags);
6358 case IORING_OP_RECVMSG:
6359 ret = io_recvmsg(req, issue_flags);
6361 case IORING_OP_RECV:
6362 ret = io_recv(req, issue_flags);
6364 case IORING_OP_TIMEOUT:
6365 ret = io_timeout(req, issue_flags);
6367 case IORING_OP_TIMEOUT_REMOVE:
6368 ret = io_timeout_remove(req, issue_flags);
6370 case IORING_OP_ACCEPT:
6371 ret = io_accept(req, issue_flags);
6373 case IORING_OP_CONNECT:
6374 ret = io_connect(req, issue_flags);
6376 case IORING_OP_ASYNC_CANCEL:
6377 ret = io_async_cancel(req, issue_flags);
6379 case IORING_OP_FALLOCATE:
6380 ret = io_fallocate(req, issue_flags);
6382 case IORING_OP_OPENAT:
6383 ret = io_openat(req, issue_flags);
6385 case IORING_OP_CLOSE:
6386 ret = io_close(req, issue_flags);
6388 case IORING_OP_FILES_UPDATE:
6389 ret = io_files_update(req, issue_flags);
6391 case IORING_OP_STATX:
6392 ret = io_statx(req, issue_flags);
6394 case IORING_OP_FADVISE:
6395 ret = io_fadvise(req, issue_flags);
6397 case IORING_OP_MADVISE:
6398 ret = io_madvise(req, issue_flags);
6400 case IORING_OP_OPENAT2:
6401 ret = io_openat2(req, issue_flags);
6403 case IORING_OP_EPOLL_CTL:
6404 ret = io_epoll_ctl(req, issue_flags);
6406 case IORING_OP_SPLICE:
6407 ret = io_splice(req, issue_flags);
6409 case IORING_OP_PROVIDE_BUFFERS:
6410 ret = io_provide_buffers(req, issue_flags);
6412 case IORING_OP_REMOVE_BUFFERS:
6413 ret = io_remove_buffers(req, issue_flags);
6416 ret = io_tee(req, issue_flags);
6418 case IORING_OP_SHUTDOWN:
6419 ret = io_shutdown(req, issue_flags);
6421 case IORING_OP_RENAMEAT:
6422 ret = io_renameat(req, issue_flags);
6424 case IORING_OP_UNLINKAT:
6425 ret = io_unlinkat(req, issue_flags);
6433 revert_creds(creds);
6436 /* If the op doesn't have a file, we're not polling for it */
6437 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6438 io_iopoll_req_issued(req);
6443 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6445 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6447 req = io_put_req_find_next(req);
6448 return req ? &req->work : NULL;
6451 static void io_wq_submit_work(struct io_wq_work *work)
6453 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6454 struct io_kiocb *timeout;
6457 /* one will be dropped by ->io_free_work() after returning to io-wq */
6458 if (!(req->flags & REQ_F_REFCOUNT))
6459 __io_req_set_refcount(req, 2);
6463 timeout = io_prep_linked_timeout(req);
6465 io_queue_linked_timeout(timeout);
6467 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6468 if (work->flags & IO_WQ_WORK_CANCEL)
6473 ret = io_issue_sqe(req, 0);
6475 * We can get EAGAIN for polled IO even though we're
6476 * forcing a sync submission from here, since we can't
6477 * wait for request slots on the block side.
6485 /* avoid locking problems by failing it from a clean context */
6487 io_req_task_queue_fail(req, ret);
6490 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6493 return &table->files[i];
6496 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6499 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6501 return (struct file *) (slot->file_ptr & FFS_MASK);
6504 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6506 unsigned long file_ptr = (unsigned long) file;
6508 if (__io_file_supports_nowait(file, READ))
6509 file_ptr |= FFS_ASYNC_READ;
6510 if (__io_file_supports_nowait(file, WRITE))
6511 file_ptr |= FFS_ASYNC_WRITE;
6512 if (S_ISREG(file_inode(file)->i_mode))
6513 file_ptr |= FFS_ISREG;
6514 file_slot->file_ptr = file_ptr;
6517 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6518 struct io_kiocb *req, int fd)
6521 unsigned long file_ptr;
6523 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6525 fd = array_index_nospec(fd, ctx->nr_user_files);
6526 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6527 file = (struct file *) (file_ptr & FFS_MASK);
6528 file_ptr &= ~FFS_MASK;
6529 /* mask in overlapping REQ_F and FFS bits */
6530 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6531 io_req_set_rsrc_node(req);
6535 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6536 struct io_kiocb *req, int fd)
6538 struct file *file = fget(fd);
6540 trace_io_uring_file_get(ctx, fd);
6542 /* we don't allow fixed io_uring files */
6543 if (file && unlikely(file->f_op == &io_uring_fops))
6544 io_req_track_inflight(req);
6548 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6549 struct io_kiocb *req, int fd, bool fixed)
6552 return io_file_get_fixed(ctx, req, fd);
6554 return io_file_get_normal(ctx, req, fd);
6557 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6559 struct io_kiocb *prev = req->timeout.prev;
6563 ret = io_try_cancel_userdata(req, prev->user_data);
6564 io_req_complete_post(req, ret ?: -ETIME, 0);
6567 io_req_complete_post(req, -ETIME, 0);
6571 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6573 struct io_timeout_data *data = container_of(timer,
6574 struct io_timeout_data, timer);
6575 struct io_kiocb *prev, *req = data->req;
6576 struct io_ring_ctx *ctx = req->ctx;
6577 unsigned long flags;
6579 spin_lock_irqsave(&ctx->timeout_lock, flags);
6580 prev = req->timeout.head;
6581 req->timeout.head = NULL;
6584 * We don't expect the list to be empty, that will only happen if we
6585 * race with the completion of the linked work.
6588 io_remove_next_linked(prev);
6589 if (!req_ref_inc_not_zero(prev))
6592 list_del(&req->timeout.list);
6593 req->timeout.prev = prev;
6594 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6596 req->io_task_work.func = io_req_task_link_timeout;
6597 io_req_task_work_add(req);
6598 return HRTIMER_NORESTART;
6601 static void io_queue_linked_timeout(struct io_kiocb *req)
6603 struct io_ring_ctx *ctx = req->ctx;
6605 spin_lock_irq(&ctx->timeout_lock);
6607 * If the back reference is NULL, then our linked request finished
6608 * before we got a chance to setup the timer
6610 if (req->timeout.head) {
6611 struct io_timeout_data *data = req->async_data;
6613 data->timer.function = io_link_timeout_fn;
6614 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6616 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6618 spin_unlock_irq(&ctx->timeout_lock);
6619 /* drop submission reference */
6623 static void __io_queue_sqe(struct io_kiocb *req)
6624 __must_hold(&req->ctx->uring_lock)
6626 struct io_kiocb *linked_timeout;
6630 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6633 * We async punt it if the file wasn't marked NOWAIT, or if the file
6634 * doesn't support non-blocking read/write attempts
6637 if (req->flags & REQ_F_COMPLETE_INLINE) {
6638 struct io_ring_ctx *ctx = req->ctx;
6639 struct io_submit_state *state = &ctx->submit_state;
6641 state->compl_reqs[state->compl_nr++] = req;
6642 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6643 io_submit_flush_completions(ctx);
6647 linked_timeout = io_prep_linked_timeout(req);
6649 io_queue_linked_timeout(linked_timeout);
6650 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6651 linked_timeout = io_prep_linked_timeout(req);
6653 switch (io_arm_poll_handler(req)) {
6654 case IO_APOLL_READY:
6656 io_unprep_linked_timeout(req);
6658 case IO_APOLL_ABORTED:
6660 * Queued up for async execution, worker will release
6661 * submit reference when the iocb is actually submitted.
6663 io_queue_async_work(req, NULL);
6668 io_queue_linked_timeout(linked_timeout);
6670 io_req_complete_failed(req, ret);
6674 static inline void io_queue_sqe(struct io_kiocb *req)
6675 __must_hold(&req->ctx->uring_lock)
6677 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6680 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
6681 __io_queue_sqe(req);
6682 } else if (req->flags & REQ_F_FAIL) {
6683 io_req_complete_failed(req, req->result);
6685 int ret = io_req_prep_async(req);
6688 io_req_complete_failed(req, ret);
6690 io_queue_async_work(req, NULL);
6695 * Check SQE restrictions (opcode and flags).
6697 * Returns 'true' if SQE is allowed, 'false' otherwise.
6699 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6700 struct io_kiocb *req,
6701 unsigned int sqe_flags)
6703 if (likely(!ctx->restricted))
6706 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6709 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6710 ctx->restrictions.sqe_flags_required)
6713 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6714 ctx->restrictions.sqe_flags_required))
6720 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6721 const struct io_uring_sqe *sqe)
6722 __must_hold(&ctx->uring_lock)
6724 struct io_submit_state *state;
6725 unsigned int sqe_flags;
6726 int personality, ret = 0;
6728 /* req is partially pre-initialised, see io_preinit_req() */
6729 req->opcode = READ_ONCE(sqe->opcode);
6730 /* same numerical values with corresponding REQ_F_*, safe to copy */
6731 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6732 req->user_data = READ_ONCE(sqe->user_data);
6734 req->fixed_rsrc_refs = NULL;
6735 req->task = current;
6737 /* enforce forwards compatibility on users */
6738 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6740 if (unlikely(req->opcode >= IORING_OP_LAST))
6742 if (!io_check_restriction(ctx, req, sqe_flags))
6745 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6746 !io_op_defs[req->opcode].buffer_select)
6748 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6749 ctx->drain_active = true;
6751 personality = READ_ONCE(sqe->personality);
6753 req->creds = xa_load(&ctx->personalities, personality);
6756 get_cred(req->creds);
6757 req->flags |= REQ_F_CREDS;
6759 state = &ctx->submit_state;
6762 * Plug now if we have more than 1 IO left after this, and the target
6763 * is potentially a read/write to block based storage.
6765 if (!state->plug_started && state->ios_left > 1 &&
6766 io_op_defs[req->opcode].plug) {
6767 blk_start_plug(&state->plug);
6768 state->plug_started = true;
6771 if (io_op_defs[req->opcode].needs_file) {
6772 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
6773 (sqe_flags & IOSQE_FIXED_FILE));
6774 if (unlikely(!req->file))
6782 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6783 const struct io_uring_sqe *sqe)
6784 __must_hold(&ctx->uring_lock)
6786 struct io_submit_link *link = &ctx->submit_state.link;
6789 ret = io_init_req(ctx, req, sqe);
6790 if (unlikely(ret)) {
6792 /* fail even hard links since we don't submit */
6795 * we can judge a link req is failed or cancelled by if
6796 * REQ_F_FAIL is set, but the head is an exception since
6797 * it may be set REQ_F_FAIL because of other req's failure
6798 * so let's leverage req->result to distinguish if a head
6799 * is set REQ_F_FAIL because of its failure or other req's
6800 * failure so that we can set the correct ret code for it.
6801 * init result here to avoid affecting the normal path.
6803 if (!(link->head->flags & REQ_F_FAIL))
6804 req_fail_link_node(link->head, -ECANCELED);
6805 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6807 * the current req is a normal req, we should return
6808 * error and thus break the submittion loop.
6810 io_req_complete_failed(req, ret);
6813 req_fail_link_node(req, ret);
6815 ret = io_req_prep(req, sqe);
6820 /* don't need @sqe from now on */
6821 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6823 ctx->flags & IORING_SETUP_SQPOLL);
6826 * If we already have a head request, queue this one for async
6827 * submittal once the head completes. If we don't have a head but
6828 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6829 * submitted sync once the chain is complete. If none of those
6830 * conditions are true (normal request), then just queue it.
6833 struct io_kiocb *head = link->head;
6835 if (!(req->flags & REQ_F_FAIL)) {
6836 ret = io_req_prep_async(req);
6837 if (unlikely(ret)) {
6838 req_fail_link_node(req, ret);
6839 if (!(head->flags & REQ_F_FAIL))
6840 req_fail_link_node(head, -ECANCELED);
6843 trace_io_uring_link(ctx, req, head);
6844 link->last->link = req;
6847 /* last request of a link, enqueue the link */
6848 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6853 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6865 * Batched submission is done, ensure local IO is flushed out.
6867 static void io_submit_state_end(struct io_submit_state *state,
6868 struct io_ring_ctx *ctx)
6870 if (state->link.head)
6871 io_queue_sqe(state->link.head);
6872 if (state->compl_nr)
6873 io_submit_flush_completions(ctx);
6874 if (state->plug_started)
6875 blk_finish_plug(&state->plug);
6879 * Start submission side cache.
6881 static void io_submit_state_start(struct io_submit_state *state,
6882 unsigned int max_ios)
6884 state->plug_started = false;
6885 state->ios_left = max_ios;
6886 /* set only head, no need to init link_last in advance */
6887 state->link.head = NULL;
6890 static void io_commit_sqring(struct io_ring_ctx *ctx)
6892 struct io_rings *rings = ctx->rings;
6895 * Ensure any loads from the SQEs are done at this point,
6896 * since once we write the new head, the application could
6897 * write new data to them.
6899 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6903 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6904 * that is mapped by userspace. This means that care needs to be taken to
6905 * ensure that reads are stable, as we cannot rely on userspace always
6906 * being a good citizen. If members of the sqe are validated and then later
6907 * used, it's important that those reads are done through READ_ONCE() to
6908 * prevent a re-load down the line.
6910 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6912 unsigned head, mask = ctx->sq_entries - 1;
6913 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6916 * The cached sq head (or cq tail) serves two purposes:
6918 * 1) allows us to batch the cost of updating the user visible
6920 * 2) allows the kernel side to track the head on its own, even
6921 * though the application is the one updating it.
6923 head = READ_ONCE(ctx->sq_array[sq_idx]);
6924 if (likely(head < ctx->sq_entries))
6925 return &ctx->sq_sqes[head];
6927 /* drop invalid entries */
6929 WRITE_ONCE(ctx->rings->sq_dropped,
6930 READ_ONCE(ctx->rings->sq_dropped) + 1);
6934 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6935 __must_hold(&ctx->uring_lock)
6939 /* make sure SQ entry isn't read before tail */
6940 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6941 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6943 io_get_task_refs(nr);
6945 io_submit_state_start(&ctx->submit_state, nr);
6946 while (submitted < nr) {
6947 const struct io_uring_sqe *sqe;
6948 struct io_kiocb *req;
6950 req = io_alloc_req(ctx);
6951 if (unlikely(!req)) {
6953 submitted = -EAGAIN;
6956 sqe = io_get_sqe(ctx);
6957 if (unlikely(!sqe)) {
6958 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
6961 /* will complete beyond this point, count as submitted */
6963 if (io_submit_sqe(ctx, req, sqe))
6967 if (unlikely(submitted != nr)) {
6968 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6969 int unused = nr - ref_used;
6971 current->io_uring->cached_refs += unused;
6972 percpu_ref_put_many(&ctx->refs, unused);
6975 io_submit_state_end(&ctx->submit_state, ctx);
6976 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6977 io_commit_sqring(ctx);
6982 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6984 return READ_ONCE(sqd->state);
6987 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6989 /* Tell userspace we may need a wakeup call */
6990 spin_lock(&ctx->completion_lock);
6991 WRITE_ONCE(ctx->rings->sq_flags,
6992 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6993 spin_unlock(&ctx->completion_lock);
6996 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6998 spin_lock(&ctx->completion_lock);
6999 WRITE_ONCE(ctx->rings->sq_flags,
7000 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7001 spin_unlock(&ctx->completion_lock);
7004 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7006 unsigned int to_submit;
7009 to_submit = io_sqring_entries(ctx);
7010 /* if we're handling multiple rings, cap submit size for fairness */
7011 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7012 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7014 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7015 unsigned nr_events = 0;
7016 const struct cred *creds = NULL;
7018 if (ctx->sq_creds != current_cred())
7019 creds = override_creds(ctx->sq_creds);
7021 mutex_lock(&ctx->uring_lock);
7022 if (!list_empty(&ctx->iopoll_list))
7023 io_do_iopoll(ctx, &nr_events, 0);
7026 * Don't submit if refs are dying, good for io_uring_register(),
7027 * but also it is relied upon by io_ring_exit_work()
7029 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7030 !(ctx->flags & IORING_SETUP_R_DISABLED))
7031 ret = io_submit_sqes(ctx, to_submit);
7032 mutex_unlock(&ctx->uring_lock);
7034 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7035 wake_up(&ctx->sqo_sq_wait);
7037 revert_creds(creds);
7043 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7045 struct io_ring_ctx *ctx;
7046 unsigned sq_thread_idle = 0;
7048 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7049 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7050 sqd->sq_thread_idle = sq_thread_idle;
7053 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7055 bool did_sig = false;
7056 struct ksignal ksig;
7058 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7059 signal_pending(current)) {
7060 mutex_unlock(&sqd->lock);
7061 if (signal_pending(current))
7062 did_sig = get_signal(&ksig);
7064 mutex_lock(&sqd->lock);
7066 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7069 static int io_sq_thread(void *data)
7071 struct io_sq_data *sqd = data;
7072 struct io_ring_ctx *ctx;
7073 unsigned long timeout = 0;
7074 char buf[TASK_COMM_LEN];
7077 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7078 set_task_comm(current, buf);
7080 if (sqd->sq_cpu != -1)
7081 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7083 set_cpus_allowed_ptr(current, cpu_online_mask);
7084 current->flags |= PF_NO_SETAFFINITY;
7086 mutex_lock(&sqd->lock);
7088 bool cap_entries, sqt_spin = false;
7090 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7091 if (io_sqd_handle_event(sqd))
7093 timeout = jiffies + sqd->sq_thread_idle;
7096 cap_entries = !list_is_singular(&sqd->ctx_list);
7097 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7098 int ret = __io_sq_thread(ctx, cap_entries);
7100 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7103 if (io_run_task_work())
7106 if (sqt_spin || !time_after(jiffies, timeout)) {
7109 timeout = jiffies + sqd->sq_thread_idle;
7113 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7114 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7115 bool needs_sched = true;
7117 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7118 io_ring_set_wakeup_flag(ctx);
7120 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7121 !list_empty_careful(&ctx->iopoll_list)) {
7122 needs_sched = false;
7125 if (io_sqring_entries(ctx)) {
7126 needs_sched = false;
7132 mutex_unlock(&sqd->lock);
7134 mutex_lock(&sqd->lock);
7136 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7137 io_ring_clear_wakeup_flag(ctx);
7140 finish_wait(&sqd->wait, &wait);
7141 timeout = jiffies + sqd->sq_thread_idle;
7144 io_uring_cancel_generic(true, sqd);
7146 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7147 io_ring_set_wakeup_flag(ctx);
7149 mutex_unlock(&sqd->lock);
7151 complete(&sqd->exited);
7155 struct io_wait_queue {
7156 struct wait_queue_entry wq;
7157 struct io_ring_ctx *ctx;
7159 unsigned nr_timeouts;
7162 static inline bool io_should_wake(struct io_wait_queue *iowq)
7164 struct io_ring_ctx *ctx = iowq->ctx;
7165 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7168 * Wake up if we have enough events, or if a timeout occurred since we
7169 * started waiting. For timeouts, we always want to return to userspace,
7170 * regardless of event count.
7172 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7175 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7176 int wake_flags, void *key)
7178 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7182 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7183 * the task, and the next invocation will do it.
7185 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7186 return autoremove_wake_function(curr, mode, wake_flags, key);
7190 static int io_run_task_work_sig(void)
7192 if (io_run_task_work())
7194 if (!signal_pending(current))
7196 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7197 return -ERESTARTSYS;
7201 /* when returns >0, the caller should retry */
7202 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7203 struct io_wait_queue *iowq,
7204 signed long *timeout)
7208 /* make sure we run task_work before checking for signals */
7209 ret = io_run_task_work_sig();
7210 if (ret || io_should_wake(iowq))
7212 /* let the caller flush overflows, retry */
7213 if (test_bit(0, &ctx->check_cq_overflow))
7216 *timeout = schedule_timeout(*timeout);
7217 return !*timeout ? -ETIME : 1;
7221 * Wait until events become available, if we don't already have some. The
7222 * application must reap them itself, as they reside on the shared cq ring.
7224 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7225 const sigset_t __user *sig, size_t sigsz,
7226 struct __kernel_timespec __user *uts)
7228 struct io_wait_queue iowq;
7229 struct io_rings *rings = ctx->rings;
7230 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7234 io_cqring_overflow_flush(ctx);
7235 if (io_cqring_events(ctx) >= min_events)
7237 if (!io_run_task_work())
7242 #ifdef CONFIG_COMPAT
7243 if (in_compat_syscall())
7244 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7248 ret = set_user_sigmask(sig, sigsz);
7255 struct timespec64 ts;
7257 if (get_timespec64(&ts, uts))
7259 timeout = timespec64_to_jiffies(&ts);
7262 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7263 iowq.wq.private = current;
7264 INIT_LIST_HEAD(&iowq.wq.entry);
7266 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7267 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7269 trace_io_uring_cqring_wait(ctx, min_events);
7271 /* if we can't even flush overflow, don't wait for more */
7272 if (!io_cqring_overflow_flush(ctx)) {
7276 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7277 TASK_INTERRUPTIBLE);
7278 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7279 finish_wait(&ctx->cq_wait, &iowq.wq);
7283 restore_saved_sigmask_unless(ret == -EINTR);
7285 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7288 static void io_free_page_table(void **table, size_t size)
7290 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7292 for (i = 0; i < nr_tables; i++)
7297 static void **io_alloc_page_table(size_t size)
7299 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7300 size_t init_size = size;
7303 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7307 for (i = 0; i < nr_tables; i++) {
7308 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7310 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7312 io_free_page_table(table, init_size);
7320 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7322 percpu_ref_exit(&ref_node->refs);
7326 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7328 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7329 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7330 unsigned long flags;
7331 bool first_add = false;
7333 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7336 while (!list_empty(&ctx->rsrc_ref_list)) {
7337 node = list_first_entry(&ctx->rsrc_ref_list,
7338 struct io_rsrc_node, node);
7339 /* recycle ref nodes in order */
7342 list_del(&node->node);
7343 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7345 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7348 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7351 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7353 struct io_rsrc_node *ref_node;
7355 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7359 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7364 INIT_LIST_HEAD(&ref_node->node);
7365 INIT_LIST_HEAD(&ref_node->rsrc_list);
7366 ref_node->done = false;
7370 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7371 struct io_rsrc_data *data_to_kill)
7373 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7374 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7377 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7379 rsrc_node->rsrc_data = data_to_kill;
7380 spin_lock_irq(&ctx->rsrc_ref_lock);
7381 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7382 spin_unlock_irq(&ctx->rsrc_ref_lock);
7384 atomic_inc(&data_to_kill->refs);
7385 percpu_ref_kill(&rsrc_node->refs);
7386 ctx->rsrc_node = NULL;
7389 if (!ctx->rsrc_node) {
7390 ctx->rsrc_node = ctx->rsrc_backup_node;
7391 ctx->rsrc_backup_node = NULL;
7395 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7397 if (ctx->rsrc_backup_node)
7399 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7400 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7403 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7407 /* As we may drop ->uring_lock, other task may have started quiesce */
7411 data->quiesce = true;
7413 ret = io_rsrc_node_switch_start(ctx);
7416 io_rsrc_node_switch(ctx, data);
7418 /* kill initial ref, already quiesced if zero */
7419 if (atomic_dec_and_test(&data->refs))
7421 mutex_unlock(&ctx->uring_lock);
7422 flush_delayed_work(&ctx->rsrc_put_work);
7423 ret = wait_for_completion_interruptible(&data->done);
7425 mutex_lock(&ctx->uring_lock);
7429 atomic_inc(&data->refs);
7430 /* wait for all works potentially completing data->done */
7431 flush_delayed_work(&ctx->rsrc_put_work);
7432 reinit_completion(&data->done);
7434 ret = io_run_task_work_sig();
7435 mutex_lock(&ctx->uring_lock);
7437 data->quiesce = false;
7442 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7444 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7445 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7447 return &data->tags[table_idx][off];
7450 static void io_rsrc_data_free(struct io_rsrc_data *data)
7452 size_t size = data->nr * sizeof(data->tags[0][0]);
7455 io_free_page_table((void **)data->tags, size);
7459 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7460 u64 __user *utags, unsigned nr,
7461 struct io_rsrc_data **pdata)
7463 struct io_rsrc_data *data;
7467 data = kzalloc(sizeof(*data), GFP_KERNEL);
7470 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7478 data->do_put = do_put;
7481 for (i = 0; i < nr; i++) {
7482 u64 *tag_slot = io_get_tag_slot(data, i);
7484 if (copy_from_user(tag_slot, &utags[i],
7490 atomic_set(&data->refs, 1);
7491 init_completion(&data->done);
7495 io_rsrc_data_free(data);
7499 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7501 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7502 GFP_KERNEL_ACCOUNT);
7503 return !!table->files;
7506 static void io_free_file_tables(struct io_file_table *table)
7508 kvfree(table->files);
7509 table->files = NULL;
7512 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7514 #if defined(CONFIG_UNIX)
7515 if (ctx->ring_sock) {
7516 struct sock *sock = ctx->ring_sock->sk;
7517 struct sk_buff *skb;
7519 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7525 for (i = 0; i < ctx->nr_user_files; i++) {
7528 file = io_file_from_index(ctx, i);
7533 io_free_file_tables(&ctx->file_table);
7534 io_rsrc_data_free(ctx->file_data);
7535 ctx->file_data = NULL;
7536 ctx->nr_user_files = 0;
7539 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7543 if (!ctx->file_data)
7545 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7547 __io_sqe_files_unregister(ctx);
7551 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7552 __releases(&sqd->lock)
7554 WARN_ON_ONCE(sqd->thread == current);
7557 * Do the dance but not conditional clear_bit() because it'd race with
7558 * other threads incrementing park_pending and setting the bit.
7560 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7561 if (atomic_dec_return(&sqd->park_pending))
7562 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7563 mutex_unlock(&sqd->lock);
7566 static void io_sq_thread_park(struct io_sq_data *sqd)
7567 __acquires(&sqd->lock)
7569 WARN_ON_ONCE(sqd->thread == current);
7571 atomic_inc(&sqd->park_pending);
7572 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7573 mutex_lock(&sqd->lock);
7575 wake_up_process(sqd->thread);
7578 static void io_sq_thread_stop(struct io_sq_data *sqd)
7580 WARN_ON_ONCE(sqd->thread == current);
7581 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7583 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7584 mutex_lock(&sqd->lock);
7586 wake_up_process(sqd->thread);
7587 mutex_unlock(&sqd->lock);
7588 wait_for_completion(&sqd->exited);
7591 static void io_put_sq_data(struct io_sq_data *sqd)
7593 if (refcount_dec_and_test(&sqd->refs)) {
7594 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7596 io_sq_thread_stop(sqd);
7601 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7603 struct io_sq_data *sqd = ctx->sq_data;
7606 io_sq_thread_park(sqd);
7607 list_del_init(&ctx->sqd_list);
7608 io_sqd_update_thread_idle(sqd);
7609 io_sq_thread_unpark(sqd);
7611 io_put_sq_data(sqd);
7612 ctx->sq_data = NULL;
7616 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7618 struct io_ring_ctx *ctx_attach;
7619 struct io_sq_data *sqd;
7622 f = fdget(p->wq_fd);
7624 return ERR_PTR(-ENXIO);
7625 if (f.file->f_op != &io_uring_fops) {
7627 return ERR_PTR(-EINVAL);
7630 ctx_attach = f.file->private_data;
7631 sqd = ctx_attach->sq_data;
7634 return ERR_PTR(-EINVAL);
7636 if (sqd->task_tgid != current->tgid) {
7638 return ERR_PTR(-EPERM);
7641 refcount_inc(&sqd->refs);
7646 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7649 struct io_sq_data *sqd;
7652 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7653 sqd = io_attach_sq_data(p);
7658 /* fall through for EPERM case, setup new sqd/task */
7659 if (PTR_ERR(sqd) != -EPERM)
7663 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7665 return ERR_PTR(-ENOMEM);
7667 atomic_set(&sqd->park_pending, 0);
7668 refcount_set(&sqd->refs, 1);
7669 INIT_LIST_HEAD(&sqd->ctx_list);
7670 mutex_init(&sqd->lock);
7671 init_waitqueue_head(&sqd->wait);
7672 init_completion(&sqd->exited);
7676 #if defined(CONFIG_UNIX)
7678 * Ensure the UNIX gc is aware of our file set, so we are certain that
7679 * the io_uring can be safely unregistered on process exit, even if we have
7680 * loops in the file referencing.
7682 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7684 struct sock *sk = ctx->ring_sock->sk;
7685 struct scm_fp_list *fpl;
7686 struct sk_buff *skb;
7689 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7693 skb = alloc_skb(0, GFP_KERNEL);
7702 fpl->user = get_uid(current_user());
7703 for (i = 0; i < nr; i++) {
7704 struct file *file = io_file_from_index(ctx, i + offset);
7708 fpl->fp[nr_files] = get_file(file);
7709 unix_inflight(fpl->user, fpl->fp[nr_files]);
7714 fpl->max = SCM_MAX_FD;
7715 fpl->count = nr_files;
7716 UNIXCB(skb).fp = fpl;
7717 skb->destructor = unix_destruct_scm;
7718 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7719 skb_queue_head(&sk->sk_receive_queue, skb);
7721 for (i = 0; i < nr_files; i++)
7732 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7733 * causes regular reference counting to break down. We rely on the UNIX
7734 * garbage collection to take care of this problem for us.
7736 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7738 unsigned left, total;
7742 left = ctx->nr_user_files;
7744 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7746 ret = __io_sqe_files_scm(ctx, this_files, total);
7750 total += this_files;
7756 while (total < ctx->nr_user_files) {
7757 struct file *file = io_file_from_index(ctx, total);
7767 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7773 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7775 struct file *file = prsrc->file;
7776 #if defined(CONFIG_UNIX)
7777 struct sock *sock = ctx->ring_sock->sk;
7778 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7779 struct sk_buff *skb;
7782 __skb_queue_head_init(&list);
7785 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7786 * remove this entry and rearrange the file array.
7788 skb = skb_dequeue(head);
7790 struct scm_fp_list *fp;
7792 fp = UNIXCB(skb).fp;
7793 for (i = 0; i < fp->count; i++) {
7796 if (fp->fp[i] != file)
7799 unix_notinflight(fp->user, fp->fp[i]);
7800 left = fp->count - 1 - i;
7802 memmove(&fp->fp[i], &fp->fp[i + 1],
7803 left * sizeof(struct file *));
7810 __skb_queue_tail(&list, skb);
7820 __skb_queue_tail(&list, skb);
7822 skb = skb_dequeue(head);
7825 if (skb_peek(&list)) {
7826 spin_lock_irq(&head->lock);
7827 while ((skb = __skb_dequeue(&list)) != NULL)
7828 __skb_queue_tail(head, skb);
7829 spin_unlock_irq(&head->lock);
7836 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7838 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7839 struct io_ring_ctx *ctx = rsrc_data->ctx;
7840 struct io_rsrc_put *prsrc, *tmp;
7842 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7843 list_del(&prsrc->list);
7846 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7848 io_ring_submit_lock(ctx, lock_ring);
7849 spin_lock(&ctx->completion_lock);
7850 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7852 io_commit_cqring(ctx);
7853 spin_unlock(&ctx->completion_lock);
7854 io_cqring_ev_posted(ctx);
7855 io_ring_submit_unlock(ctx, lock_ring);
7858 rsrc_data->do_put(ctx, prsrc);
7862 io_rsrc_node_destroy(ref_node);
7863 if (atomic_dec_and_test(&rsrc_data->refs))
7864 complete(&rsrc_data->done);
7867 static void io_rsrc_put_work(struct work_struct *work)
7869 struct io_ring_ctx *ctx;
7870 struct llist_node *node;
7872 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7873 node = llist_del_all(&ctx->rsrc_put_llist);
7876 struct io_rsrc_node *ref_node;
7877 struct llist_node *next = node->next;
7879 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7880 __io_rsrc_put_work(ref_node);
7885 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7886 unsigned nr_args, u64 __user *tags)
7888 __s32 __user *fds = (__s32 __user *) arg;
7897 if (nr_args > IORING_MAX_FIXED_FILES)
7899 if (nr_args > rlimit(RLIMIT_NOFILE))
7901 ret = io_rsrc_node_switch_start(ctx);
7904 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7910 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7913 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7914 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7918 /* allow sparse sets */
7921 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7928 if (unlikely(!file))
7932 * Don't allow io_uring instances to be registered. If UNIX
7933 * isn't enabled, then this causes a reference cycle and this
7934 * instance can never get freed. If UNIX is enabled we'll
7935 * handle it just fine, but there's still no point in allowing
7936 * a ring fd as it doesn't support regular read/write anyway.
7938 if (file->f_op == &io_uring_fops) {
7942 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7945 ret = io_sqe_files_scm(ctx);
7947 __io_sqe_files_unregister(ctx);
7951 io_rsrc_node_switch(ctx, NULL);
7954 for (i = 0; i < ctx->nr_user_files; i++) {
7955 file = io_file_from_index(ctx, i);
7959 io_free_file_tables(&ctx->file_table);
7960 ctx->nr_user_files = 0;
7962 io_rsrc_data_free(ctx->file_data);
7963 ctx->file_data = NULL;
7967 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7970 #if defined(CONFIG_UNIX)
7971 struct sock *sock = ctx->ring_sock->sk;
7972 struct sk_buff_head *head = &sock->sk_receive_queue;
7973 struct sk_buff *skb;
7976 * See if we can merge this file into an existing skb SCM_RIGHTS
7977 * file set. If there's no room, fall back to allocating a new skb
7978 * and filling it in.
7980 spin_lock_irq(&head->lock);
7981 skb = skb_peek(head);
7983 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7985 if (fpl->count < SCM_MAX_FD) {
7986 __skb_unlink(skb, head);
7987 spin_unlock_irq(&head->lock);
7988 fpl->fp[fpl->count] = get_file(file);
7989 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7991 spin_lock_irq(&head->lock);
7992 __skb_queue_head(head, skb);
7997 spin_unlock_irq(&head->lock);
8004 return __io_sqe_files_scm(ctx, 1, index);
8010 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8011 unsigned int issue_flags, u32 slot_index)
8013 struct io_ring_ctx *ctx = req->ctx;
8014 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8015 struct io_fixed_file *file_slot;
8018 io_ring_submit_lock(ctx, !force_nonblock);
8019 if (file->f_op == &io_uring_fops)
8022 if (!ctx->file_data)
8025 if (slot_index >= ctx->nr_user_files)
8028 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8029 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8031 if (file_slot->file_ptr)
8034 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8035 io_fixed_file_set(file_slot, file);
8036 ret = io_sqe_file_register(ctx, file, slot_index);
8038 file_slot->file_ptr = 0;
8044 io_ring_submit_unlock(ctx, !force_nonblock);
8050 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8051 struct io_rsrc_node *node, void *rsrc)
8053 struct io_rsrc_put *prsrc;
8055 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8059 prsrc->tag = *io_get_tag_slot(data, idx);
8061 list_add(&prsrc->list, &node->rsrc_list);
8065 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8066 struct io_uring_rsrc_update2 *up,
8069 u64 __user *tags = u64_to_user_ptr(up->tags);
8070 __s32 __user *fds = u64_to_user_ptr(up->data);
8071 struct io_rsrc_data *data = ctx->file_data;
8072 struct io_fixed_file *file_slot;
8076 bool needs_switch = false;
8078 if (!ctx->file_data)
8080 if (up->offset + nr_args > ctx->nr_user_files)
8083 for (done = 0; done < nr_args; done++) {
8086 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8087 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8091 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8095 if (fd == IORING_REGISTER_FILES_SKIP)
8098 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8099 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8101 if (file_slot->file_ptr) {
8102 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8103 err = io_queue_rsrc_removal(data, up->offset + done,
8104 ctx->rsrc_node, file);
8107 file_slot->file_ptr = 0;
8108 needs_switch = true;
8117 * Don't allow io_uring instances to be registered. If
8118 * UNIX isn't enabled, then this causes a reference
8119 * cycle and this instance can never get freed. If UNIX
8120 * is enabled we'll handle it just fine, but there's
8121 * still no point in allowing a ring fd as it doesn't
8122 * support regular read/write anyway.
8124 if (file->f_op == &io_uring_fops) {
8129 *io_get_tag_slot(data, up->offset + done) = tag;
8130 io_fixed_file_set(file_slot, file);
8131 err = io_sqe_file_register(ctx, file, i);
8133 file_slot->file_ptr = 0;
8141 io_rsrc_node_switch(ctx, data);
8142 return done ? done : err;
8145 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8146 struct task_struct *task)
8148 struct io_wq_hash *hash;
8149 struct io_wq_data data;
8150 unsigned int concurrency;
8152 mutex_lock(&ctx->uring_lock);
8153 hash = ctx->hash_map;
8155 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8157 mutex_unlock(&ctx->uring_lock);
8158 return ERR_PTR(-ENOMEM);
8160 refcount_set(&hash->refs, 1);
8161 init_waitqueue_head(&hash->wait);
8162 ctx->hash_map = hash;
8164 mutex_unlock(&ctx->uring_lock);
8168 data.free_work = io_wq_free_work;
8169 data.do_work = io_wq_submit_work;
8171 /* Do QD, or 4 * CPUS, whatever is smallest */
8172 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8174 return io_wq_create(concurrency, &data);
8177 static int io_uring_alloc_task_context(struct task_struct *task,
8178 struct io_ring_ctx *ctx)
8180 struct io_uring_task *tctx;
8183 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8184 if (unlikely(!tctx))
8187 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8188 if (unlikely(ret)) {
8193 tctx->io_wq = io_init_wq_offload(ctx, task);
8194 if (IS_ERR(tctx->io_wq)) {
8195 ret = PTR_ERR(tctx->io_wq);
8196 percpu_counter_destroy(&tctx->inflight);
8202 init_waitqueue_head(&tctx->wait);
8203 atomic_set(&tctx->in_idle, 0);
8204 atomic_set(&tctx->inflight_tracked, 0);
8205 task->io_uring = tctx;
8206 spin_lock_init(&tctx->task_lock);
8207 INIT_WQ_LIST(&tctx->task_list);
8208 init_task_work(&tctx->task_work, tctx_task_work);
8212 void __io_uring_free(struct task_struct *tsk)
8214 struct io_uring_task *tctx = tsk->io_uring;
8216 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8217 WARN_ON_ONCE(tctx->io_wq);
8218 WARN_ON_ONCE(tctx->cached_refs);
8220 percpu_counter_destroy(&tctx->inflight);
8222 tsk->io_uring = NULL;
8225 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8226 struct io_uring_params *p)
8230 /* Retain compatibility with failing for an invalid attach attempt */
8231 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8232 IORING_SETUP_ATTACH_WQ) {
8235 f = fdget(p->wq_fd);
8238 if (f.file->f_op != &io_uring_fops) {
8244 if (ctx->flags & IORING_SETUP_SQPOLL) {
8245 struct task_struct *tsk;
8246 struct io_sq_data *sqd;
8249 sqd = io_get_sq_data(p, &attached);
8255 ctx->sq_creds = get_current_cred();
8257 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8258 if (!ctx->sq_thread_idle)
8259 ctx->sq_thread_idle = HZ;
8261 io_sq_thread_park(sqd);
8262 list_add(&ctx->sqd_list, &sqd->ctx_list);
8263 io_sqd_update_thread_idle(sqd);
8264 /* don't attach to a dying SQPOLL thread, would be racy */
8265 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8266 io_sq_thread_unpark(sqd);
8273 if (p->flags & IORING_SETUP_SQ_AFF) {
8274 int cpu = p->sq_thread_cpu;
8277 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8284 sqd->task_pid = current->pid;
8285 sqd->task_tgid = current->tgid;
8286 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8293 ret = io_uring_alloc_task_context(tsk, ctx);
8294 wake_up_new_task(tsk);
8297 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8298 /* Can't have SQ_AFF without SQPOLL */
8305 complete(&ctx->sq_data->exited);
8307 io_sq_thread_finish(ctx);
8311 static inline void __io_unaccount_mem(struct user_struct *user,
8312 unsigned long nr_pages)
8314 atomic_long_sub(nr_pages, &user->locked_vm);
8317 static inline int __io_account_mem(struct user_struct *user,
8318 unsigned long nr_pages)
8320 unsigned long page_limit, cur_pages, new_pages;
8322 /* Don't allow more pages than we can safely lock */
8323 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8326 cur_pages = atomic_long_read(&user->locked_vm);
8327 new_pages = cur_pages + nr_pages;
8328 if (new_pages > page_limit)
8330 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8331 new_pages) != cur_pages);
8336 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8339 __io_unaccount_mem(ctx->user, nr_pages);
8341 if (ctx->mm_account)
8342 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8345 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8350 ret = __io_account_mem(ctx->user, nr_pages);
8355 if (ctx->mm_account)
8356 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8361 static void io_mem_free(void *ptr)
8368 page = virt_to_head_page(ptr);
8369 if (put_page_testzero(page))
8370 free_compound_page(page);
8373 static void *io_mem_alloc(size_t size)
8375 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8376 __GFP_NORETRY | __GFP_ACCOUNT;
8378 return (void *) __get_free_pages(gfp_flags, get_order(size));
8381 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8384 struct io_rings *rings;
8385 size_t off, sq_array_size;
8387 off = struct_size(rings, cqes, cq_entries);
8388 if (off == SIZE_MAX)
8392 off = ALIGN(off, SMP_CACHE_BYTES);
8400 sq_array_size = array_size(sizeof(u32), sq_entries);
8401 if (sq_array_size == SIZE_MAX)
8404 if (check_add_overflow(off, sq_array_size, &off))
8410 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8412 struct io_mapped_ubuf *imu = *slot;
8415 if (imu != ctx->dummy_ubuf) {
8416 for (i = 0; i < imu->nr_bvecs; i++)
8417 unpin_user_page(imu->bvec[i].bv_page);
8418 if (imu->acct_pages)
8419 io_unaccount_mem(ctx, imu->acct_pages);
8425 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8427 io_buffer_unmap(ctx, &prsrc->buf);
8431 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8435 for (i = 0; i < ctx->nr_user_bufs; i++)
8436 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8437 kfree(ctx->user_bufs);
8438 io_rsrc_data_free(ctx->buf_data);
8439 ctx->user_bufs = NULL;
8440 ctx->buf_data = NULL;
8441 ctx->nr_user_bufs = 0;
8444 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8451 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8453 __io_sqe_buffers_unregister(ctx);
8457 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8458 void __user *arg, unsigned index)
8460 struct iovec __user *src;
8462 #ifdef CONFIG_COMPAT
8464 struct compat_iovec __user *ciovs;
8465 struct compat_iovec ciov;
8467 ciovs = (struct compat_iovec __user *) arg;
8468 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8471 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8472 dst->iov_len = ciov.iov_len;
8476 src = (struct iovec __user *) arg;
8477 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8483 * Not super efficient, but this is just a registration time. And we do cache
8484 * the last compound head, so generally we'll only do a full search if we don't
8487 * We check if the given compound head page has already been accounted, to
8488 * avoid double accounting it. This allows us to account the full size of the
8489 * page, not just the constituent pages of a huge page.
8491 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8492 int nr_pages, struct page *hpage)
8496 /* check current page array */
8497 for (i = 0; i < nr_pages; i++) {
8498 if (!PageCompound(pages[i]))
8500 if (compound_head(pages[i]) == hpage)
8504 /* check previously registered pages */
8505 for (i = 0; i < ctx->nr_user_bufs; i++) {
8506 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8508 for (j = 0; j < imu->nr_bvecs; j++) {
8509 if (!PageCompound(imu->bvec[j].bv_page))
8511 if (compound_head(imu->bvec[j].bv_page) == hpage)
8519 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8520 int nr_pages, struct io_mapped_ubuf *imu,
8521 struct page **last_hpage)
8525 imu->acct_pages = 0;
8526 for (i = 0; i < nr_pages; i++) {
8527 if (!PageCompound(pages[i])) {
8532 hpage = compound_head(pages[i]);
8533 if (hpage == *last_hpage)
8535 *last_hpage = hpage;
8536 if (headpage_already_acct(ctx, pages, i, hpage))
8538 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8542 if (!imu->acct_pages)
8545 ret = io_account_mem(ctx, imu->acct_pages);
8547 imu->acct_pages = 0;
8551 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8552 struct io_mapped_ubuf **pimu,
8553 struct page **last_hpage)
8555 struct io_mapped_ubuf *imu = NULL;
8556 struct vm_area_struct **vmas = NULL;
8557 struct page **pages = NULL;
8558 unsigned long off, start, end, ubuf;
8560 int ret, pret, nr_pages, i;
8562 if (!iov->iov_base) {
8563 *pimu = ctx->dummy_ubuf;
8567 ubuf = (unsigned long) iov->iov_base;
8568 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8569 start = ubuf >> PAGE_SHIFT;
8570 nr_pages = end - start;
8575 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8579 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8584 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8589 mmap_read_lock(current->mm);
8590 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8592 if (pret == nr_pages) {
8593 /* don't support file backed memory */
8594 for (i = 0; i < nr_pages; i++) {
8595 struct vm_area_struct *vma = vmas[i];
8597 if (vma_is_shmem(vma))
8600 !is_file_hugepages(vma->vm_file)) {
8606 ret = pret < 0 ? pret : -EFAULT;
8608 mmap_read_unlock(current->mm);
8611 * if we did partial map, or found file backed vmas,
8612 * release any pages we did get
8615 unpin_user_pages(pages, pret);
8619 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8621 unpin_user_pages(pages, pret);
8625 off = ubuf & ~PAGE_MASK;
8626 size = iov->iov_len;
8627 for (i = 0; i < nr_pages; i++) {
8630 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8631 imu->bvec[i].bv_page = pages[i];
8632 imu->bvec[i].bv_len = vec_len;
8633 imu->bvec[i].bv_offset = off;
8637 /* store original address for later verification */
8639 imu->ubuf_end = ubuf + iov->iov_len;
8640 imu->nr_bvecs = nr_pages;
8651 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8653 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8654 return ctx->user_bufs ? 0 : -ENOMEM;
8657 static int io_buffer_validate(struct iovec *iov)
8659 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8662 * Don't impose further limits on the size and buffer
8663 * constraints here, we'll -EINVAL later when IO is
8664 * submitted if they are wrong.
8667 return iov->iov_len ? -EFAULT : 0;
8671 /* arbitrary limit, but we need something */
8672 if (iov->iov_len > SZ_1G)
8675 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8681 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8682 unsigned int nr_args, u64 __user *tags)
8684 struct page *last_hpage = NULL;
8685 struct io_rsrc_data *data;
8691 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8693 ret = io_rsrc_node_switch_start(ctx);
8696 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8699 ret = io_buffers_map_alloc(ctx, nr_args);
8701 io_rsrc_data_free(data);
8705 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8706 ret = io_copy_iov(ctx, &iov, arg, i);
8709 ret = io_buffer_validate(&iov);
8712 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8717 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8723 WARN_ON_ONCE(ctx->buf_data);
8725 ctx->buf_data = data;
8727 __io_sqe_buffers_unregister(ctx);
8729 io_rsrc_node_switch(ctx, NULL);
8733 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8734 struct io_uring_rsrc_update2 *up,
8735 unsigned int nr_args)
8737 u64 __user *tags = u64_to_user_ptr(up->tags);
8738 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8739 struct page *last_hpage = NULL;
8740 bool needs_switch = false;
8746 if (up->offset + nr_args > ctx->nr_user_bufs)
8749 for (done = 0; done < nr_args; done++) {
8750 struct io_mapped_ubuf *imu;
8751 int offset = up->offset + done;
8754 err = io_copy_iov(ctx, &iov, iovs, done);
8757 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8761 err = io_buffer_validate(&iov);
8764 if (!iov.iov_base && tag) {
8768 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8772 i = array_index_nospec(offset, ctx->nr_user_bufs);
8773 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8774 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8775 ctx->rsrc_node, ctx->user_bufs[i]);
8776 if (unlikely(err)) {
8777 io_buffer_unmap(ctx, &imu);
8780 ctx->user_bufs[i] = NULL;
8781 needs_switch = true;
8784 ctx->user_bufs[i] = imu;
8785 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8789 io_rsrc_node_switch(ctx, ctx->buf_data);
8790 return done ? done : err;
8793 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8795 __s32 __user *fds = arg;
8801 if (copy_from_user(&fd, fds, sizeof(*fds)))
8804 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8805 if (IS_ERR(ctx->cq_ev_fd)) {
8806 int ret = PTR_ERR(ctx->cq_ev_fd);
8808 ctx->cq_ev_fd = NULL;
8815 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8817 if (ctx->cq_ev_fd) {
8818 eventfd_ctx_put(ctx->cq_ev_fd);
8819 ctx->cq_ev_fd = NULL;
8826 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8828 struct io_buffer *buf;
8829 unsigned long index;
8831 xa_for_each(&ctx->io_buffers, index, buf)
8832 __io_remove_buffers(ctx, buf, index, -1U);
8835 static void io_req_cache_free(struct list_head *list)
8837 struct io_kiocb *req, *nxt;
8839 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
8840 list_del(&req->inflight_entry);
8841 kmem_cache_free(req_cachep, req);
8845 static void io_req_caches_free(struct io_ring_ctx *ctx)
8847 struct io_submit_state *state = &ctx->submit_state;
8849 mutex_lock(&ctx->uring_lock);
8851 if (state->free_reqs) {
8852 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
8853 state->free_reqs = 0;
8856 io_flush_cached_locked_reqs(ctx, state);
8857 io_req_cache_free(&state->free_list);
8858 mutex_unlock(&ctx->uring_lock);
8861 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8863 if (data && !atomic_dec_and_test(&data->refs))
8864 wait_for_completion(&data->done);
8867 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8869 io_sq_thread_finish(ctx);
8871 if (ctx->mm_account) {
8872 mmdrop(ctx->mm_account);
8873 ctx->mm_account = NULL;
8876 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8877 io_wait_rsrc_data(ctx->buf_data);
8878 io_wait_rsrc_data(ctx->file_data);
8880 mutex_lock(&ctx->uring_lock);
8882 __io_sqe_buffers_unregister(ctx);
8884 __io_sqe_files_unregister(ctx);
8886 __io_cqring_overflow_flush(ctx, true);
8887 mutex_unlock(&ctx->uring_lock);
8888 io_eventfd_unregister(ctx);
8889 io_destroy_buffers(ctx);
8891 put_cred(ctx->sq_creds);
8893 /* there are no registered resources left, nobody uses it */
8895 io_rsrc_node_destroy(ctx->rsrc_node);
8896 if (ctx->rsrc_backup_node)
8897 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8898 flush_delayed_work(&ctx->rsrc_put_work);
8900 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8901 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8903 #if defined(CONFIG_UNIX)
8904 if (ctx->ring_sock) {
8905 ctx->ring_sock->file = NULL; /* so that iput() is called */
8906 sock_release(ctx->ring_sock);
8909 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
8911 io_mem_free(ctx->rings);
8912 io_mem_free(ctx->sq_sqes);
8914 percpu_ref_exit(&ctx->refs);
8915 free_uid(ctx->user);
8916 io_req_caches_free(ctx);
8918 io_wq_put_hash(ctx->hash_map);
8919 kfree(ctx->cancel_hash);
8920 kfree(ctx->dummy_ubuf);
8924 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8926 struct io_ring_ctx *ctx = file->private_data;
8929 poll_wait(file, &ctx->poll_wait, wait);
8931 * synchronizes with barrier from wq_has_sleeper call in
8935 if (!io_sqring_full(ctx))
8936 mask |= EPOLLOUT | EPOLLWRNORM;
8939 * Don't flush cqring overflow list here, just do a simple check.
8940 * Otherwise there could possible be ABBA deadlock:
8943 * lock(&ctx->uring_lock);
8945 * lock(&ctx->uring_lock);
8948 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8949 * pushs them to do the flush.
8951 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8952 mask |= EPOLLIN | EPOLLRDNORM;
8957 static int io_uring_fasync(int fd, struct file *file, int on)
8959 struct io_ring_ctx *ctx = file->private_data;
8961 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8964 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8966 const struct cred *creds;
8968 creds = xa_erase(&ctx->personalities, id);
8977 struct io_tctx_exit {
8978 struct callback_head task_work;
8979 struct completion completion;
8980 struct io_ring_ctx *ctx;
8983 static void io_tctx_exit_cb(struct callback_head *cb)
8985 struct io_uring_task *tctx = current->io_uring;
8986 struct io_tctx_exit *work;
8988 work = container_of(cb, struct io_tctx_exit, task_work);
8990 * When @in_idle, we're in cancellation and it's racy to remove the
8991 * node. It'll be removed by the end of cancellation, just ignore it.
8993 if (!atomic_read(&tctx->in_idle))
8994 io_uring_del_tctx_node((unsigned long)work->ctx);
8995 complete(&work->completion);
8998 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9000 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9002 return req->ctx == data;
9005 static void io_ring_exit_work(struct work_struct *work)
9007 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9008 unsigned long timeout = jiffies + HZ * 60 * 5;
9009 unsigned long interval = HZ / 20;
9010 struct io_tctx_exit exit;
9011 struct io_tctx_node *node;
9015 * If we're doing polled IO and end up having requests being
9016 * submitted async (out-of-line), then completions can come in while
9017 * we're waiting for refs to drop. We need to reap these manually,
9018 * as nobody else will be looking for them.
9021 io_uring_try_cancel_requests(ctx, NULL, true);
9023 struct io_sq_data *sqd = ctx->sq_data;
9024 struct task_struct *tsk;
9026 io_sq_thread_park(sqd);
9028 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9029 io_wq_cancel_cb(tsk->io_uring->io_wq,
9030 io_cancel_ctx_cb, ctx, true);
9031 io_sq_thread_unpark(sqd);
9034 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9035 /* there is little hope left, don't run it too often */
9038 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9040 init_completion(&exit.completion);
9041 init_task_work(&exit.task_work, io_tctx_exit_cb);
9044 * Some may use context even when all refs and requests have been put,
9045 * and they are free to do so while still holding uring_lock or
9046 * completion_lock, see io_req_task_submit(). Apart from other work,
9047 * this lock/unlock section also waits them to finish.
9049 mutex_lock(&ctx->uring_lock);
9050 while (!list_empty(&ctx->tctx_list)) {
9051 WARN_ON_ONCE(time_after(jiffies, timeout));
9053 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9055 /* don't spin on a single task if cancellation failed */
9056 list_rotate_left(&ctx->tctx_list);
9057 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9058 if (WARN_ON_ONCE(ret))
9060 wake_up_process(node->task);
9062 mutex_unlock(&ctx->uring_lock);
9063 wait_for_completion(&exit.completion);
9064 mutex_lock(&ctx->uring_lock);
9066 mutex_unlock(&ctx->uring_lock);
9067 spin_lock(&ctx->completion_lock);
9068 spin_unlock(&ctx->completion_lock);
9070 io_ring_ctx_free(ctx);
9073 /* Returns true if we found and killed one or more timeouts */
9074 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9077 struct io_kiocb *req, *tmp;
9080 spin_lock(&ctx->completion_lock);
9081 spin_lock_irq(&ctx->timeout_lock);
9082 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9083 if (io_match_task(req, tsk, cancel_all)) {
9084 io_kill_timeout(req, -ECANCELED);
9088 spin_unlock_irq(&ctx->timeout_lock);
9090 io_commit_cqring(ctx);
9091 spin_unlock(&ctx->completion_lock);
9093 io_cqring_ev_posted(ctx);
9094 return canceled != 0;
9097 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9099 unsigned long index;
9100 struct creds *creds;
9102 mutex_lock(&ctx->uring_lock);
9103 percpu_ref_kill(&ctx->refs);
9105 __io_cqring_overflow_flush(ctx, true);
9106 xa_for_each(&ctx->personalities, index, creds)
9107 io_unregister_personality(ctx, index);
9108 mutex_unlock(&ctx->uring_lock);
9110 io_kill_timeouts(ctx, NULL, true);
9111 io_poll_remove_all(ctx, NULL, true);
9113 /* if we failed setting up the ctx, we might not have any rings */
9114 io_iopoll_try_reap_events(ctx);
9116 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9118 * Use system_unbound_wq to avoid spawning tons of event kworkers
9119 * if we're exiting a ton of rings at the same time. It just adds
9120 * noise and overhead, there's no discernable change in runtime
9121 * over using system_wq.
9123 queue_work(system_unbound_wq, &ctx->exit_work);
9126 static int io_uring_release(struct inode *inode, struct file *file)
9128 struct io_ring_ctx *ctx = file->private_data;
9130 file->private_data = NULL;
9131 io_ring_ctx_wait_and_kill(ctx);
9135 struct io_task_cancel {
9136 struct task_struct *task;
9140 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9142 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9143 struct io_task_cancel *cancel = data;
9146 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9147 struct io_ring_ctx *ctx = req->ctx;
9149 /* protect against races with linked timeouts */
9150 spin_lock(&ctx->completion_lock);
9151 ret = io_match_task(req, cancel->task, cancel->all);
9152 spin_unlock(&ctx->completion_lock);
9154 ret = io_match_task(req, cancel->task, cancel->all);
9159 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9160 struct task_struct *task, bool cancel_all)
9162 struct io_defer_entry *de;
9165 spin_lock(&ctx->completion_lock);
9166 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9167 if (io_match_task(de->req, task, cancel_all)) {
9168 list_cut_position(&list, &ctx->defer_list, &de->list);
9172 spin_unlock(&ctx->completion_lock);
9173 if (list_empty(&list))
9176 while (!list_empty(&list)) {
9177 de = list_first_entry(&list, struct io_defer_entry, list);
9178 list_del_init(&de->list);
9179 io_req_complete_failed(de->req, -ECANCELED);
9185 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9187 struct io_tctx_node *node;
9188 enum io_wq_cancel cret;
9191 mutex_lock(&ctx->uring_lock);
9192 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9193 struct io_uring_task *tctx = node->task->io_uring;
9196 * io_wq will stay alive while we hold uring_lock, because it's
9197 * killed after ctx nodes, which requires to take the lock.
9199 if (!tctx || !tctx->io_wq)
9201 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9202 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9204 mutex_unlock(&ctx->uring_lock);
9209 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9210 struct task_struct *task,
9213 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9214 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9217 enum io_wq_cancel cret;
9221 ret |= io_uring_try_cancel_iowq(ctx);
9222 } else if (tctx && tctx->io_wq) {
9224 * Cancels requests of all rings, not only @ctx, but
9225 * it's fine as the task is in exit/exec.
9227 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9229 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9232 /* SQPOLL thread does its own polling */
9233 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9234 (ctx->sq_data && ctx->sq_data->thread == current)) {
9235 while (!list_empty_careful(&ctx->iopoll_list)) {
9236 io_iopoll_try_reap_events(ctx);
9241 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9242 ret |= io_poll_remove_all(ctx, task, cancel_all);
9243 ret |= io_kill_timeouts(ctx, task, cancel_all);
9245 ret |= io_run_task_work();
9252 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9254 struct io_uring_task *tctx = current->io_uring;
9255 struct io_tctx_node *node;
9258 if (unlikely(!tctx)) {
9259 ret = io_uring_alloc_task_context(current, ctx);
9262 tctx = current->io_uring;
9264 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9265 node = kmalloc(sizeof(*node), GFP_KERNEL);
9269 node->task = current;
9271 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9278 mutex_lock(&ctx->uring_lock);
9279 list_add(&node->ctx_node, &ctx->tctx_list);
9280 mutex_unlock(&ctx->uring_lock);
9287 * Note that this task has used io_uring. We use it for cancelation purposes.
9289 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9291 struct io_uring_task *tctx = current->io_uring;
9293 if (likely(tctx && tctx->last == ctx))
9295 return __io_uring_add_tctx_node(ctx);
9299 * Remove this io_uring_file -> task mapping.
9301 static void io_uring_del_tctx_node(unsigned long index)
9303 struct io_uring_task *tctx = current->io_uring;
9304 struct io_tctx_node *node;
9308 node = xa_erase(&tctx->xa, index);
9312 WARN_ON_ONCE(current != node->task);
9313 WARN_ON_ONCE(list_empty(&node->ctx_node));
9315 mutex_lock(&node->ctx->uring_lock);
9316 list_del(&node->ctx_node);
9317 mutex_unlock(&node->ctx->uring_lock);
9319 if (tctx->last == node->ctx)
9324 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9326 struct io_wq *wq = tctx->io_wq;
9327 struct io_tctx_node *node;
9328 unsigned long index;
9330 xa_for_each(&tctx->xa, index, node)
9331 io_uring_del_tctx_node(index);
9334 * Must be after io_uring_del_task_file() (removes nodes under
9335 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9337 io_wq_put_and_exit(wq);
9342 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9345 return atomic_read(&tctx->inflight_tracked);
9346 return percpu_counter_sum(&tctx->inflight);
9349 static void io_uring_drop_tctx_refs(struct task_struct *task)
9351 struct io_uring_task *tctx = task->io_uring;
9352 unsigned int refs = tctx->cached_refs;
9355 tctx->cached_refs = 0;
9356 percpu_counter_sub(&tctx->inflight, refs);
9357 put_task_struct_many(task, refs);
9362 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9363 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9365 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9367 struct io_uring_task *tctx = current->io_uring;
9368 struct io_ring_ctx *ctx;
9372 WARN_ON_ONCE(sqd && sqd->thread != current);
9374 if (!current->io_uring)
9377 io_wq_exit_start(tctx->io_wq);
9379 atomic_inc(&tctx->in_idle);
9381 io_uring_drop_tctx_refs(current);
9382 /* read completions before cancelations */
9383 inflight = tctx_inflight(tctx, !cancel_all);
9388 struct io_tctx_node *node;
9389 unsigned long index;
9391 xa_for_each(&tctx->xa, index, node) {
9392 /* sqpoll task will cancel all its requests */
9393 if (node->ctx->sq_data)
9395 io_uring_try_cancel_requests(node->ctx, current,
9399 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9400 io_uring_try_cancel_requests(ctx, current,
9404 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9405 io_uring_drop_tctx_refs(current);
9407 * If we've seen completions, retry without waiting. This
9408 * avoids a race where a completion comes in before we did
9409 * prepare_to_wait().
9411 if (inflight == tctx_inflight(tctx, !cancel_all))
9413 finish_wait(&tctx->wait, &wait);
9415 atomic_dec(&tctx->in_idle);
9417 io_uring_clean_tctx(tctx);
9419 /* for exec all current's requests should be gone, kill tctx */
9420 __io_uring_free(current);
9424 void __io_uring_cancel(bool cancel_all)
9426 io_uring_cancel_generic(cancel_all, NULL);
9429 static void *io_uring_validate_mmap_request(struct file *file,
9430 loff_t pgoff, size_t sz)
9432 struct io_ring_ctx *ctx = file->private_data;
9433 loff_t offset = pgoff << PAGE_SHIFT;
9438 case IORING_OFF_SQ_RING:
9439 case IORING_OFF_CQ_RING:
9442 case IORING_OFF_SQES:
9446 return ERR_PTR(-EINVAL);
9449 page = virt_to_head_page(ptr);
9450 if (sz > page_size(page))
9451 return ERR_PTR(-EINVAL);
9458 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9460 size_t sz = vma->vm_end - vma->vm_start;
9464 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9466 return PTR_ERR(ptr);
9468 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9469 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9472 #else /* !CONFIG_MMU */
9474 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9476 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9479 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9481 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9484 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9485 unsigned long addr, unsigned long len,
9486 unsigned long pgoff, unsigned long flags)
9490 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9492 return PTR_ERR(ptr);
9494 return (unsigned long) ptr;
9497 #endif /* !CONFIG_MMU */
9499 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9504 if (!io_sqring_full(ctx))
9506 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9508 if (!io_sqring_full(ctx))
9511 } while (!signal_pending(current));
9513 finish_wait(&ctx->sqo_sq_wait, &wait);
9517 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9518 struct __kernel_timespec __user **ts,
9519 const sigset_t __user **sig)
9521 struct io_uring_getevents_arg arg;
9524 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9525 * is just a pointer to the sigset_t.
9527 if (!(flags & IORING_ENTER_EXT_ARG)) {
9528 *sig = (const sigset_t __user *) argp;
9534 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9535 * timespec and sigset_t pointers if good.
9537 if (*argsz != sizeof(arg))
9539 if (copy_from_user(&arg, argp, sizeof(arg)))
9541 *sig = u64_to_user_ptr(arg.sigmask);
9542 *argsz = arg.sigmask_sz;
9543 *ts = u64_to_user_ptr(arg.ts);
9547 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9548 u32, min_complete, u32, flags, const void __user *, argp,
9551 struct io_ring_ctx *ctx;
9558 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9559 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9563 if (unlikely(!f.file))
9567 if (unlikely(f.file->f_op != &io_uring_fops))
9571 ctx = f.file->private_data;
9572 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9576 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9580 * For SQ polling, the thread will do all submissions and completions.
9581 * Just return the requested submit count, and wake the thread if
9585 if (ctx->flags & IORING_SETUP_SQPOLL) {
9586 io_cqring_overflow_flush(ctx);
9588 if (unlikely(ctx->sq_data->thread == NULL)) {
9592 if (flags & IORING_ENTER_SQ_WAKEUP)
9593 wake_up(&ctx->sq_data->wait);
9594 if (flags & IORING_ENTER_SQ_WAIT) {
9595 ret = io_sqpoll_wait_sq(ctx);
9599 submitted = to_submit;
9600 } else if (to_submit) {
9601 ret = io_uring_add_tctx_node(ctx);
9604 mutex_lock(&ctx->uring_lock);
9605 submitted = io_submit_sqes(ctx, to_submit);
9606 mutex_unlock(&ctx->uring_lock);
9608 if (submitted != to_submit)
9611 if (flags & IORING_ENTER_GETEVENTS) {
9612 const sigset_t __user *sig;
9613 struct __kernel_timespec __user *ts;
9615 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9619 min_complete = min(min_complete, ctx->cq_entries);
9622 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9623 * space applications don't need to do io completion events
9624 * polling again, they can rely on io_sq_thread to do polling
9625 * work, which can reduce cpu usage and uring_lock contention.
9627 if (ctx->flags & IORING_SETUP_IOPOLL &&
9628 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9629 ret = io_iopoll_check(ctx, min_complete);
9631 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9636 percpu_ref_put(&ctx->refs);
9639 return submitted ? submitted : ret;
9642 #ifdef CONFIG_PROC_FS
9643 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9644 const struct cred *cred)
9646 struct user_namespace *uns = seq_user_ns(m);
9647 struct group_info *gi;
9652 seq_printf(m, "%5d\n", id);
9653 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9654 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9655 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9656 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9657 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9658 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9659 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9660 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9661 seq_puts(m, "\n\tGroups:\t");
9662 gi = cred->group_info;
9663 for (g = 0; g < gi->ngroups; g++) {
9664 seq_put_decimal_ull(m, g ? " " : "",
9665 from_kgid_munged(uns, gi->gid[g]));
9667 seq_puts(m, "\n\tCapEff:\t");
9668 cap = cred->cap_effective;
9669 CAP_FOR_EACH_U32(__capi)
9670 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9675 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9677 struct io_sq_data *sq = NULL;
9682 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9683 * since fdinfo case grabs it in the opposite direction of normal use
9684 * cases. If we fail to get the lock, we just don't iterate any
9685 * structures that could be going away outside the io_uring mutex.
9687 has_lock = mutex_trylock(&ctx->uring_lock);
9689 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9695 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9696 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9697 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9698 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9699 struct file *f = io_file_from_index(ctx, i);
9702 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9704 seq_printf(m, "%5u: <none>\n", i);
9706 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9707 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9708 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9709 unsigned int len = buf->ubuf_end - buf->ubuf;
9711 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9713 if (has_lock && !xa_empty(&ctx->personalities)) {
9714 unsigned long index;
9715 const struct cred *cred;
9717 seq_printf(m, "Personalities:\n");
9718 xa_for_each(&ctx->personalities, index, cred)
9719 io_uring_show_cred(m, index, cred);
9721 seq_printf(m, "PollList:\n");
9722 spin_lock(&ctx->completion_lock);
9723 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9724 struct hlist_head *list = &ctx->cancel_hash[i];
9725 struct io_kiocb *req;
9727 hlist_for_each_entry(req, list, hash_node)
9728 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9729 req->task->task_works != NULL);
9731 spin_unlock(&ctx->completion_lock);
9733 mutex_unlock(&ctx->uring_lock);
9736 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9738 struct io_ring_ctx *ctx = f->private_data;
9740 if (percpu_ref_tryget(&ctx->refs)) {
9741 __io_uring_show_fdinfo(ctx, m);
9742 percpu_ref_put(&ctx->refs);
9747 static const struct file_operations io_uring_fops = {
9748 .release = io_uring_release,
9749 .mmap = io_uring_mmap,
9751 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9752 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9754 .poll = io_uring_poll,
9755 .fasync = io_uring_fasync,
9756 #ifdef CONFIG_PROC_FS
9757 .show_fdinfo = io_uring_show_fdinfo,
9761 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9762 struct io_uring_params *p)
9764 struct io_rings *rings;
9765 size_t size, sq_array_offset;
9767 /* make sure these are sane, as we already accounted them */
9768 ctx->sq_entries = p->sq_entries;
9769 ctx->cq_entries = p->cq_entries;
9771 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9772 if (size == SIZE_MAX)
9775 rings = io_mem_alloc(size);
9780 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9781 rings->sq_ring_mask = p->sq_entries - 1;
9782 rings->cq_ring_mask = p->cq_entries - 1;
9783 rings->sq_ring_entries = p->sq_entries;
9784 rings->cq_ring_entries = p->cq_entries;
9786 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9787 if (size == SIZE_MAX) {
9788 io_mem_free(ctx->rings);
9793 ctx->sq_sqes = io_mem_alloc(size);
9794 if (!ctx->sq_sqes) {
9795 io_mem_free(ctx->rings);
9803 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9807 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9811 ret = io_uring_add_tctx_node(ctx);
9816 fd_install(fd, file);
9821 * Allocate an anonymous fd, this is what constitutes the application
9822 * visible backing of an io_uring instance. The application mmaps this
9823 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9824 * we have to tie this fd to a socket for file garbage collection purposes.
9826 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9829 #if defined(CONFIG_UNIX)
9832 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9835 return ERR_PTR(ret);
9838 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9839 O_RDWR | O_CLOEXEC);
9840 #if defined(CONFIG_UNIX)
9842 sock_release(ctx->ring_sock);
9843 ctx->ring_sock = NULL;
9845 ctx->ring_sock->file = file;
9851 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9852 struct io_uring_params __user *params)
9854 struct io_ring_ctx *ctx;
9860 if (entries > IORING_MAX_ENTRIES) {
9861 if (!(p->flags & IORING_SETUP_CLAMP))
9863 entries = IORING_MAX_ENTRIES;
9867 * Use twice as many entries for the CQ ring. It's possible for the
9868 * application to drive a higher depth than the size of the SQ ring,
9869 * since the sqes are only used at submission time. This allows for
9870 * some flexibility in overcommitting a bit. If the application has
9871 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9872 * of CQ ring entries manually.
9874 p->sq_entries = roundup_pow_of_two(entries);
9875 if (p->flags & IORING_SETUP_CQSIZE) {
9877 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9878 * to a power-of-two, if it isn't already. We do NOT impose
9879 * any cq vs sq ring sizing.
9883 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9884 if (!(p->flags & IORING_SETUP_CLAMP))
9886 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9888 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9889 if (p->cq_entries < p->sq_entries)
9892 p->cq_entries = 2 * p->sq_entries;
9895 ctx = io_ring_ctx_alloc(p);
9898 ctx->compat = in_compat_syscall();
9899 if (!capable(CAP_IPC_LOCK))
9900 ctx->user = get_uid(current_user());
9903 * This is just grabbed for accounting purposes. When a process exits,
9904 * the mm is exited and dropped before the files, hence we need to hang
9905 * on to this mm purely for the purposes of being able to unaccount
9906 * memory (locked/pinned vm). It's not used for anything else.
9908 mmgrab(current->mm);
9909 ctx->mm_account = current->mm;
9911 ret = io_allocate_scq_urings(ctx, p);
9915 ret = io_sq_offload_create(ctx, p);
9918 /* always set a rsrc node */
9919 ret = io_rsrc_node_switch_start(ctx);
9922 io_rsrc_node_switch(ctx, NULL);
9924 memset(&p->sq_off, 0, sizeof(p->sq_off));
9925 p->sq_off.head = offsetof(struct io_rings, sq.head);
9926 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9927 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9928 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9929 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9930 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9931 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9933 memset(&p->cq_off, 0, sizeof(p->cq_off));
9934 p->cq_off.head = offsetof(struct io_rings, cq.head);
9935 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9936 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9937 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9938 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9939 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9940 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9942 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9943 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9944 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9945 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9946 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9947 IORING_FEAT_RSRC_TAGS;
9949 if (copy_to_user(params, p, sizeof(*p))) {
9954 file = io_uring_get_file(ctx);
9956 ret = PTR_ERR(file);
9961 * Install ring fd as the very last thing, so we don't risk someone
9962 * having closed it before we finish setup
9964 ret = io_uring_install_fd(ctx, file);
9966 /* fput will clean it up */
9971 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9974 io_ring_ctx_wait_and_kill(ctx);
9979 * Sets up an aio uring context, and returns the fd. Applications asks for a
9980 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9981 * params structure passed in.
9983 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9985 struct io_uring_params p;
9988 if (copy_from_user(&p, params, sizeof(p)))
9990 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9995 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9996 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9997 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9998 IORING_SETUP_R_DISABLED))
10001 return io_uring_create(entries, &p, params);
10004 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10005 struct io_uring_params __user *, params)
10007 return io_uring_setup(entries, params);
10010 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10012 struct io_uring_probe *p;
10016 size = struct_size(p, ops, nr_args);
10017 if (size == SIZE_MAX)
10019 p = kzalloc(size, GFP_KERNEL);
10024 if (copy_from_user(p, arg, size))
10027 if (memchr_inv(p, 0, size))
10030 p->last_op = IORING_OP_LAST - 1;
10031 if (nr_args > IORING_OP_LAST)
10032 nr_args = IORING_OP_LAST;
10034 for (i = 0; i < nr_args; i++) {
10036 if (!io_op_defs[i].not_supported)
10037 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10042 if (copy_to_user(arg, p, size))
10049 static int io_register_personality(struct io_ring_ctx *ctx)
10051 const struct cred *creds;
10055 creds = get_current_cred();
10057 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10058 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10066 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10067 unsigned int nr_args)
10069 struct io_uring_restriction *res;
10073 /* Restrictions allowed only if rings started disabled */
10074 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10077 /* We allow only a single restrictions registration */
10078 if (ctx->restrictions.registered)
10081 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10084 size = array_size(nr_args, sizeof(*res));
10085 if (size == SIZE_MAX)
10088 res = memdup_user(arg, size);
10090 return PTR_ERR(res);
10094 for (i = 0; i < nr_args; i++) {
10095 switch (res[i].opcode) {
10096 case IORING_RESTRICTION_REGISTER_OP:
10097 if (res[i].register_op >= IORING_REGISTER_LAST) {
10102 __set_bit(res[i].register_op,
10103 ctx->restrictions.register_op);
10105 case IORING_RESTRICTION_SQE_OP:
10106 if (res[i].sqe_op >= IORING_OP_LAST) {
10111 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10113 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10114 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10116 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10117 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10126 /* Reset all restrictions if an error happened */
10128 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10130 ctx->restrictions.registered = true;
10136 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10138 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10141 if (ctx->restrictions.registered)
10142 ctx->restricted = 1;
10144 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10145 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10146 wake_up(&ctx->sq_data->wait);
10150 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10151 struct io_uring_rsrc_update2 *up,
10159 if (check_add_overflow(up->offset, nr_args, &tmp))
10161 err = io_rsrc_node_switch_start(ctx);
10166 case IORING_RSRC_FILE:
10167 return __io_sqe_files_update(ctx, up, nr_args);
10168 case IORING_RSRC_BUFFER:
10169 return __io_sqe_buffers_update(ctx, up, nr_args);
10174 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10177 struct io_uring_rsrc_update2 up;
10181 memset(&up, 0, sizeof(up));
10182 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10184 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10187 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10188 unsigned size, unsigned type)
10190 struct io_uring_rsrc_update2 up;
10192 if (size != sizeof(up))
10194 if (copy_from_user(&up, arg, sizeof(up)))
10196 if (!up.nr || up.resv)
10198 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10201 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10202 unsigned int size, unsigned int type)
10204 struct io_uring_rsrc_register rr;
10206 /* keep it extendible */
10207 if (size != sizeof(rr))
10210 memset(&rr, 0, sizeof(rr));
10211 if (copy_from_user(&rr, arg, size))
10213 if (!rr.nr || rr.resv || rr.resv2)
10217 case IORING_RSRC_FILE:
10218 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10219 rr.nr, u64_to_user_ptr(rr.tags));
10220 case IORING_RSRC_BUFFER:
10221 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10222 rr.nr, u64_to_user_ptr(rr.tags));
10227 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10230 struct io_uring_task *tctx = current->io_uring;
10231 cpumask_var_t new_mask;
10234 if (!tctx || !tctx->io_wq)
10237 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10240 cpumask_clear(new_mask);
10241 if (len > cpumask_size())
10242 len = cpumask_size();
10244 if (copy_from_user(new_mask, arg, len)) {
10245 free_cpumask_var(new_mask);
10249 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10250 free_cpumask_var(new_mask);
10254 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10256 struct io_uring_task *tctx = current->io_uring;
10258 if (!tctx || !tctx->io_wq)
10261 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10264 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10267 struct io_uring_task *tctx = current->io_uring;
10268 __u32 new_count[2];
10271 if (!tctx || !tctx->io_wq)
10273 if (copy_from_user(new_count, arg, sizeof(new_count)))
10275 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10276 if (new_count[i] > INT_MAX)
10279 ret = io_wq_max_workers(tctx->io_wq, new_count);
10283 if (copy_to_user(arg, new_count, sizeof(new_count)))
10289 static bool io_register_op_must_quiesce(int op)
10292 case IORING_REGISTER_BUFFERS:
10293 case IORING_UNREGISTER_BUFFERS:
10294 case IORING_REGISTER_FILES:
10295 case IORING_UNREGISTER_FILES:
10296 case IORING_REGISTER_FILES_UPDATE:
10297 case IORING_REGISTER_PROBE:
10298 case IORING_REGISTER_PERSONALITY:
10299 case IORING_UNREGISTER_PERSONALITY:
10300 case IORING_REGISTER_FILES2:
10301 case IORING_REGISTER_FILES_UPDATE2:
10302 case IORING_REGISTER_BUFFERS2:
10303 case IORING_REGISTER_BUFFERS_UPDATE:
10304 case IORING_REGISTER_IOWQ_AFF:
10305 case IORING_UNREGISTER_IOWQ_AFF:
10306 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10313 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10317 percpu_ref_kill(&ctx->refs);
10320 * Drop uring mutex before waiting for references to exit. If another
10321 * thread is currently inside io_uring_enter() it might need to grab the
10322 * uring_lock to make progress. If we hold it here across the drain
10323 * wait, then we can deadlock. It's safe to drop the mutex here, since
10324 * no new references will come in after we've killed the percpu ref.
10326 mutex_unlock(&ctx->uring_lock);
10328 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10331 ret = io_run_task_work_sig();
10332 } while (ret >= 0);
10333 mutex_lock(&ctx->uring_lock);
10336 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10340 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10341 void __user *arg, unsigned nr_args)
10342 __releases(ctx->uring_lock)
10343 __acquires(ctx->uring_lock)
10348 * We're inside the ring mutex, if the ref is already dying, then
10349 * someone else killed the ctx or is already going through
10350 * io_uring_register().
10352 if (percpu_ref_is_dying(&ctx->refs))
10355 if (ctx->restricted) {
10356 if (opcode >= IORING_REGISTER_LAST)
10358 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10359 if (!test_bit(opcode, ctx->restrictions.register_op))
10363 if (io_register_op_must_quiesce(opcode)) {
10364 ret = io_ctx_quiesce(ctx);
10370 case IORING_REGISTER_BUFFERS:
10371 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10373 case IORING_UNREGISTER_BUFFERS:
10375 if (arg || nr_args)
10377 ret = io_sqe_buffers_unregister(ctx);
10379 case IORING_REGISTER_FILES:
10380 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10382 case IORING_UNREGISTER_FILES:
10384 if (arg || nr_args)
10386 ret = io_sqe_files_unregister(ctx);
10388 case IORING_REGISTER_FILES_UPDATE:
10389 ret = io_register_files_update(ctx, arg, nr_args);
10391 case IORING_REGISTER_EVENTFD:
10392 case IORING_REGISTER_EVENTFD_ASYNC:
10396 ret = io_eventfd_register(ctx, arg);
10399 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10400 ctx->eventfd_async = 1;
10402 ctx->eventfd_async = 0;
10404 case IORING_UNREGISTER_EVENTFD:
10406 if (arg || nr_args)
10408 ret = io_eventfd_unregister(ctx);
10410 case IORING_REGISTER_PROBE:
10412 if (!arg || nr_args > 256)
10414 ret = io_probe(ctx, arg, nr_args);
10416 case IORING_REGISTER_PERSONALITY:
10418 if (arg || nr_args)
10420 ret = io_register_personality(ctx);
10422 case IORING_UNREGISTER_PERSONALITY:
10426 ret = io_unregister_personality(ctx, nr_args);
10428 case IORING_REGISTER_ENABLE_RINGS:
10430 if (arg || nr_args)
10432 ret = io_register_enable_rings(ctx);
10434 case IORING_REGISTER_RESTRICTIONS:
10435 ret = io_register_restrictions(ctx, arg, nr_args);
10437 case IORING_REGISTER_FILES2:
10438 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10440 case IORING_REGISTER_FILES_UPDATE2:
10441 ret = io_register_rsrc_update(ctx, arg, nr_args,
10444 case IORING_REGISTER_BUFFERS2:
10445 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10447 case IORING_REGISTER_BUFFERS_UPDATE:
10448 ret = io_register_rsrc_update(ctx, arg, nr_args,
10449 IORING_RSRC_BUFFER);
10451 case IORING_REGISTER_IOWQ_AFF:
10453 if (!arg || !nr_args)
10455 ret = io_register_iowq_aff(ctx, arg, nr_args);
10457 case IORING_UNREGISTER_IOWQ_AFF:
10459 if (arg || nr_args)
10461 ret = io_unregister_iowq_aff(ctx);
10463 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10465 if (!arg || nr_args != 2)
10467 ret = io_register_iowq_max_workers(ctx, arg);
10474 if (io_register_op_must_quiesce(opcode)) {
10475 /* bring the ctx back to life */
10476 percpu_ref_reinit(&ctx->refs);
10477 reinit_completion(&ctx->ref_comp);
10482 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10483 void __user *, arg, unsigned int, nr_args)
10485 struct io_ring_ctx *ctx;
10494 if (f.file->f_op != &io_uring_fops)
10497 ctx = f.file->private_data;
10499 io_run_task_work();
10501 mutex_lock(&ctx->uring_lock);
10502 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10503 mutex_unlock(&ctx->uring_lock);
10504 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10505 ctx->cq_ev_fd != NULL, ret);
10511 static int __init io_uring_init(void)
10513 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10514 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10515 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10518 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10519 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10520 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10521 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10522 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10523 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10524 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10525 BUILD_BUG_SQE_ELEM(8, __u64, off);
10526 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10527 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10528 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10529 BUILD_BUG_SQE_ELEM(24, __u32, len);
10530 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10531 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10532 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10533 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10534 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10535 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10536 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10537 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10538 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10539 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10540 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10541 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10542 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10543 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10544 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10545 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10546 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10547 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10548 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10549 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10550 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
10552 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10553 sizeof(struct io_uring_rsrc_update));
10554 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10555 sizeof(struct io_uring_rsrc_update2));
10557 /* ->buf_index is u16 */
10558 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
10560 /* should fit into one byte */
10561 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10563 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10564 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10566 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10570 __initcall(io_uring_init);