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/blk-mq.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/audit.h>
82 #include <linux/security.h>
84 #define CREATE_TRACE_POINTS
85 #include <trace/events/io_uring.h>
87 #include <uapi/linux/io_uring.h>
92 #define IORING_MAX_ENTRIES 32768
93 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
97 #define IORING_MAX_FIXED_FILES (1U << 15)
98 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
99 IORING_REGISTER_LAST + IORING_OP_LAST)
101 #define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
102 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
103 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
105 #define IORING_MAX_REG_BUFFERS (1U << 14)
107 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
108 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
110 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
111 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
113 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
114 REQ_F_POLLED | REQ_F_CREDS | REQ_F_ASYNC_DATA)
116 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
119 u32 head ____cacheline_aligned_in_smp;
120 u32 tail ____cacheline_aligned_in_smp;
124 * This data is shared with the application through the mmap at offsets
125 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
127 * The offsets to the member fields are published through struct
128 * io_sqring_offsets when calling io_uring_setup.
132 * Head and tail offsets into the ring; the offsets need to be
133 * masked to get valid indices.
135 * The kernel controls head of the sq ring and the tail of the cq ring,
136 * and the application controls tail of the sq ring and the head of the
139 struct io_uring sq, cq;
141 * Bitmasks to apply to head and tail offsets (constant, equals
144 u32 sq_ring_mask, cq_ring_mask;
145 /* Ring sizes (constant, power of 2) */
146 u32 sq_ring_entries, cq_ring_entries;
148 * Number of invalid entries dropped by the kernel due to
149 * invalid index stored in array
151 * Written by the kernel, shouldn't be modified by the
152 * application (i.e. get number of "new events" by comparing to
155 * After a new SQ head value was read by the application this
156 * counter includes all submissions that were dropped reaching
157 * the new SQ head (and possibly more).
163 * Written by the kernel, shouldn't be modified by the
166 * The application needs a full memory barrier before checking
167 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
173 * Written by the application, shouldn't be modified by the
178 * Number of completion events lost because the queue was full;
179 * this should be avoided by the application by making sure
180 * there are not more requests pending than there is space in
181 * the completion queue.
183 * Written by the kernel, shouldn't be modified by the
184 * application (i.e. get number of "new events" by comparing to
187 * As completion events come in out of order this counter is not
188 * ordered with any other data.
192 * Ring buffer of completion events.
194 * The kernel writes completion events fresh every time they are
195 * produced, so the application is allowed to modify pending
198 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
201 enum io_uring_cmd_flags {
202 IO_URING_F_COMPLETE_DEFER = 1,
203 IO_URING_F_UNLOCKED = 2,
204 /* int's last bit, sign checks are usually faster than a bit test */
205 IO_URING_F_NONBLOCK = INT_MIN,
208 struct io_mapped_ubuf {
211 unsigned int nr_bvecs;
212 unsigned long acct_pages;
213 struct bio_vec bvec[];
218 struct io_overflow_cqe {
219 struct io_uring_cqe cqe;
220 struct list_head list;
223 struct io_fixed_file {
224 /* file * with additional FFS_* flags */
225 unsigned long file_ptr;
229 struct list_head list;
234 struct io_mapped_ubuf *buf;
238 struct io_file_table {
239 struct io_fixed_file *files;
242 struct io_rsrc_node {
243 struct percpu_ref refs;
244 struct list_head node;
245 struct list_head rsrc_list;
246 struct io_rsrc_data *rsrc_data;
247 struct llist_node llist;
251 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
253 struct io_rsrc_data {
254 struct io_ring_ctx *ctx;
260 struct completion done;
264 struct io_buffer_list {
265 struct list_head list;
266 struct list_head buf_list;
271 struct list_head list;
278 struct io_restriction {
279 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
280 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
281 u8 sqe_flags_allowed;
282 u8 sqe_flags_required;
287 IO_SQ_THREAD_SHOULD_STOP = 0,
288 IO_SQ_THREAD_SHOULD_PARK,
293 atomic_t park_pending;
296 /* ctx's that are using this sqd */
297 struct list_head ctx_list;
299 struct task_struct *thread;
300 struct wait_queue_head wait;
302 unsigned sq_thread_idle;
308 struct completion exited;
311 #define IO_COMPL_BATCH 32
312 #define IO_REQ_CACHE_SIZE 32
313 #define IO_REQ_ALLOC_BATCH 8
315 struct io_submit_link {
316 struct io_kiocb *head;
317 struct io_kiocb *last;
320 struct io_submit_state {
321 /* inline/task_work completion list, under ->uring_lock */
322 struct io_wq_work_node free_list;
323 /* batch completion logic */
324 struct io_wq_work_list compl_reqs;
325 struct io_submit_link link;
330 unsigned short submit_nr;
331 struct blk_plug plug;
335 struct eventfd_ctx *cq_ev_fd;
336 unsigned int eventfd_async: 1;
340 #define IO_BUFFERS_HASH_BITS 5
343 /* const or read-mostly hot data */
345 struct percpu_ref refs;
347 struct io_rings *rings;
349 unsigned int compat: 1;
350 unsigned int drain_next: 1;
351 unsigned int restricted: 1;
352 unsigned int off_timeout_used: 1;
353 unsigned int drain_active: 1;
354 unsigned int drain_disabled: 1;
355 unsigned int has_evfd: 1;
356 } ____cacheline_aligned_in_smp;
358 /* submission data */
360 struct mutex uring_lock;
363 * Ring buffer of indices into array of io_uring_sqe, which is
364 * mmapped by the application using the IORING_OFF_SQES offset.
366 * This indirection could e.g. be used to assign fixed
367 * io_uring_sqe entries to operations and only submit them to
368 * the queue when needed.
370 * The kernel modifies neither the indices array nor the entries
374 struct io_uring_sqe *sq_sqes;
375 unsigned cached_sq_head;
377 struct list_head defer_list;
380 * Fixed resources fast path, should be accessed only under
381 * uring_lock, and updated through io_uring_register(2)
383 struct io_rsrc_node *rsrc_node;
384 int rsrc_cached_refs;
385 struct io_file_table file_table;
386 unsigned nr_user_files;
387 unsigned nr_user_bufs;
388 struct io_mapped_ubuf **user_bufs;
390 struct io_submit_state submit_state;
391 struct list_head timeout_list;
392 struct list_head ltimeout_list;
393 struct list_head cq_overflow_list;
394 struct list_head *io_buffers;
395 struct list_head io_buffers_cache;
396 struct list_head apoll_cache;
397 struct xarray personalities;
399 unsigned sq_thread_idle;
400 } ____cacheline_aligned_in_smp;
402 /* IRQ completion list, under ->completion_lock */
403 struct io_wq_work_list locked_free_list;
404 unsigned int locked_free_nr;
406 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
407 struct io_sq_data *sq_data; /* if using sq thread polling */
409 struct wait_queue_head sqo_sq_wait;
410 struct list_head sqd_list;
412 unsigned long check_cq_overflow;
415 unsigned cached_cq_tail;
417 struct io_ev_fd __rcu *io_ev_fd;
418 struct wait_queue_head cq_wait;
420 atomic_t cq_timeouts;
421 unsigned cq_last_tm_flush;
422 } ____cacheline_aligned_in_smp;
425 spinlock_t completion_lock;
427 spinlock_t timeout_lock;
430 * ->iopoll_list is protected by the ctx->uring_lock for
431 * io_uring instances that don't use IORING_SETUP_SQPOLL.
432 * For SQPOLL, only the single threaded io_sq_thread() will
433 * manipulate the list, hence no extra locking is needed there.
435 struct io_wq_work_list iopoll_list;
436 struct hlist_head *cancel_hash;
437 unsigned cancel_hash_bits;
438 bool poll_multi_queue;
440 struct list_head io_buffers_comp;
441 } ____cacheline_aligned_in_smp;
443 struct io_restriction restrictions;
445 /* slow path rsrc auxilary data, used by update/register */
447 struct io_rsrc_node *rsrc_backup_node;
448 struct io_mapped_ubuf *dummy_ubuf;
449 struct io_rsrc_data *file_data;
450 struct io_rsrc_data *buf_data;
452 struct delayed_work rsrc_put_work;
453 struct llist_head rsrc_put_llist;
454 struct list_head rsrc_ref_list;
455 spinlock_t rsrc_ref_lock;
457 struct list_head io_buffers_pages;
460 /* Keep this last, we don't need it for the fast path */
462 #if defined(CONFIG_UNIX)
463 struct socket *ring_sock;
465 /* hashed buffered write serialization */
466 struct io_wq_hash *hash_map;
468 /* Only used for accounting purposes */
469 struct user_struct *user;
470 struct mm_struct *mm_account;
472 /* ctx exit and cancelation */
473 struct llist_head fallback_llist;
474 struct delayed_work fallback_work;
475 struct work_struct exit_work;
476 struct list_head tctx_list;
477 struct completion ref_comp;
479 bool iowq_limits_set;
484 * Arbitrary limit, can be raised if need be
486 #define IO_RINGFD_REG_MAX 16
488 struct io_uring_task {
489 /* submission side */
492 struct wait_queue_head wait;
493 const struct io_ring_ctx *last;
495 struct percpu_counter inflight;
498 spinlock_t task_lock;
499 struct io_wq_work_list task_list;
500 struct io_wq_work_list prior_task_list;
501 struct callback_head task_work;
502 struct file **registered_rings;
507 * First field must be the file pointer in all the
508 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
510 struct io_poll_iocb {
512 struct wait_queue_head *head;
514 struct wait_queue_entry wait;
517 struct io_poll_update {
523 bool update_user_data;
532 struct io_timeout_data {
533 struct io_kiocb *req;
534 struct hrtimer timer;
535 struct timespec64 ts;
536 enum hrtimer_mode mode;
542 struct sockaddr __user *addr;
543 int __user *addr_len;
546 unsigned long nofile;
566 struct list_head list;
567 /* head of the link, used by linked timeouts only */
568 struct io_kiocb *head;
569 /* for linked completions */
570 struct io_kiocb *prev;
573 struct io_timeout_rem {
578 struct timespec64 ts;
584 /* NOTE: kiocb has the file as the first member, so don't do it here */
593 struct sockaddr __user *addr;
600 struct compat_msghdr __user *umsg_compat;
601 struct user_msghdr __user *umsg;
614 struct filename *filename;
616 unsigned long nofile;
619 struct io_rsrc_update {
645 struct epoll_event event;
649 struct file *file_out;
657 struct io_provide_buf {
671 struct filename *filename;
672 struct statx __user *buffer;
684 struct filename *oldpath;
685 struct filename *newpath;
693 struct filename *filename;
700 struct filename *filename;
706 struct filename *oldpath;
707 struct filename *newpath;
714 struct filename *oldpath;
715 struct filename *newpath;
725 struct io_async_connect {
726 struct sockaddr_storage address;
729 struct io_async_msghdr {
730 struct iovec fast_iov[UIO_FASTIOV];
731 /* points to an allocated iov, if NULL we use fast_iov instead */
732 struct iovec *free_iov;
733 struct sockaddr __user *uaddr;
735 struct sockaddr_storage addr;
739 struct iov_iter iter;
740 struct iov_iter_state iter_state;
741 struct iovec fast_iov[UIO_FASTIOV];
745 struct io_rw_state s;
746 const struct iovec *free_iovec;
748 struct wait_page_queue wpq;
752 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
753 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
754 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
755 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
756 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
757 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
758 REQ_F_CQE_SKIP_BIT = IOSQE_CQE_SKIP_SUCCESS_BIT,
760 /* first byte is taken by user flags, shift it to not overlap */
765 REQ_F_LINK_TIMEOUT_BIT,
766 REQ_F_NEED_CLEANUP_BIT,
768 REQ_F_BUFFER_SELECTED_BIT,
769 REQ_F_COMPLETE_INLINE_BIT,
773 REQ_F_ARM_LTIMEOUT_BIT,
774 REQ_F_ASYNC_DATA_BIT,
775 REQ_F_SKIP_LINK_CQES_BIT,
776 REQ_F_SINGLE_POLL_BIT,
777 REQ_F_DOUBLE_POLL_BIT,
778 REQ_F_PARTIAL_IO_BIT,
779 /* keep async read/write and isreg together and in order */
780 REQ_F_SUPPORT_NOWAIT_BIT,
783 /* not a real bit, just to check we're not overflowing the space */
789 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
790 /* drain existing IO first */
791 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
793 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
794 /* doesn't sever on completion < 0 */
795 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
797 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
798 /* IOSQE_BUFFER_SELECT */
799 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
800 /* IOSQE_CQE_SKIP_SUCCESS */
801 REQ_F_CQE_SKIP = BIT(REQ_F_CQE_SKIP_BIT),
803 /* fail rest of links */
804 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
805 /* on inflight list, should be cancelled and waited on exit reliably */
806 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
807 /* read/write uses file position */
808 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
809 /* must not punt to workers */
810 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
811 /* has or had linked timeout */
812 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
814 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
815 /* already went through poll handler */
816 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
817 /* buffer already selected */
818 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
819 /* completion is deferred through io_comp_state */
820 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
821 /* caller should reissue async */
822 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
823 /* supports async reads/writes */
824 REQ_F_SUPPORT_NOWAIT = BIT(REQ_F_SUPPORT_NOWAIT_BIT),
826 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
827 /* has creds assigned */
828 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
829 /* skip refcounting if not set */
830 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
831 /* there is a linked timeout that has to be armed */
832 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
833 /* ->async_data allocated */
834 REQ_F_ASYNC_DATA = BIT(REQ_F_ASYNC_DATA_BIT),
835 /* don't post CQEs while failing linked requests */
836 REQ_F_SKIP_LINK_CQES = BIT(REQ_F_SKIP_LINK_CQES_BIT),
837 /* single poll may be active */
838 REQ_F_SINGLE_POLL = BIT(REQ_F_SINGLE_POLL_BIT),
839 /* double poll may active */
840 REQ_F_DOUBLE_POLL = BIT(REQ_F_DOUBLE_POLL_BIT),
841 /* request has already done partial IO */
842 REQ_F_PARTIAL_IO = BIT(REQ_F_PARTIAL_IO_BIT),
846 struct io_poll_iocb poll;
847 struct io_poll_iocb *double_poll;
850 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
852 struct io_task_work {
854 struct io_wq_work_node node;
855 struct llist_node fallback_node;
857 io_req_tw_func_t func;
861 IORING_RSRC_FILE = 0,
862 IORING_RSRC_BUFFER = 1,
866 * NOTE! Each of the iocb union members has the file pointer
867 * as the first entry in their struct definition. So you can
868 * access the file pointer through any of the sub-structs,
869 * or directly as just 'file' in this struct.
875 struct io_poll_iocb poll;
876 struct io_poll_update poll_update;
877 struct io_accept accept;
879 struct io_cancel cancel;
880 struct io_timeout timeout;
881 struct io_timeout_rem timeout_rem;
882 struct io_connect connect;
883 struct io_sr_msg sr_msg;
885 struct io_close close;
886 struct io_rsrc_update rsrc_update;
887 struct io_fadvise fadvise;
888 struct io_madvise madvise;
889 struct io_epoll epoll;
890 struct io_splice splice;
891 struct io_provide_buf pbuf;
892 struct io_statx statx;
893 struct io_shutdown shutdown;
894 struct io_rename rename;
895 struct io_unlink unlink;
896 struct io_mkdir mkdir;
897 struct io_symlink symlink;
898 struct io_hardlink hardlink;
903 /* polled IO has completed */
910 /* fd initially, then cflags for completion */
916 struct io_ring_ctx *ctx;
917 struct task_struct *task;
919 struct percpu_ref *fixed_rsrc_refs;
920 /* store used ubuf, so we can prevent reloading */
921 struct io_mapped_ubuf *imu;
924 /* used by request caches, completion batching and iopoll */
925 struct io_wq_work_node comp_list;
926 /* cache ->apoll->events */
931 struct io_task_work io_task_work;
932 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
933 struct hlist_node hash_node;
934 /* internal polling, see IORING_FEAT_FAST_POLL */
935 struct async_poll *apoll;
936 /* opcode allocated if it needs to store data for async defer */
938 /* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
939 struct io_buffer *kbuf;
940 /* linked requests, IFF REQ_F_HARDLINK or REQ_F_LINK are set */
941 struct io_kiocb *link;
942 /* custom credentials, valid IFF REQ_F_CREDS is set */
943 const struct cred *creds;
944 struct io_wq_work work;
947 struct io_tctx_node {
948 struct list_head ctx_node;
949 struct task_struct *task;
950 struct io_ring_ctx *ctx;
953 struct io_defer_entry {
954 struct list_head list;
955 struct io_kiocb *req;
960 /* needs req->file assigned */
961 unsigned needs_file : 1;
962 /* should block plug */
964 /* hash wq insertion if file is a regular file */
965 unsigned hash_reg_file : 1;
966 /* unbound wq insertion if file is a non-regular file */
967 unsigned unbound_nonreg_file : 1;
968 /* set if opcode supports polled "wait" */
970 unsigned pollout : 1;
971 unsigned poll_exclusive : 1;
972 /* op supports buffer selection */
973 unsigned buffer_select : 1;
974 /* do prep async if is going to be punted */
975 unsigned needs_async_setup : 1;
976 /* opcode is not supported by this kernel */
977 unsigned not_supported : 1;
979 unsigned audit_skip : 1;
980 /* size of async data needed, if any */
981 unsigned short async_size;
984 static const struct io_op_def io_op_defs[] = {
985 [IORING_OP_NOP] = {},
986 [IORING_OP_READV] = {
988 .unbound_nonreg_file = 1,
991 .needs_async_setup = 1,
994 .async_size = sizeof(struct io_async_rw),
996 [IORING_OP_WRITEV] = {
999 .unbound_nonreg_file = 1,
1001 .needs_async_setup = 1,
1004 .async_size = sizeof(struct io_async_rw),
1006 [IORING_OP_FSYNC] = {
1010 [IORING_OP_READ_FIXED] = {
1012 .unbound_nonreg_file = 1,
1016 .async_size = sizeof(struct io_async_rw),
1018 [IORING_OP_WRITE_FIXED] = {
1021 .unbound_nonreg_file = 1,
1025 .async_size = sizeof(struct io_async_rw),
1027 [IORING_OP_POLL_ADD] = {
1029 .unbound_nonreg_file = 1,
1032 [IORING_OP_POLL_REMOVE] = {
1035 [IORING_OP_SYNC_FILE_RANGE] = {
1039 [IORING_OP_SENDMSG] = {
1041 .unbound_nonreg_file = 1,
1043 .needs_async_setup = 1,
1044 .async_size = sizeof(struct io_async_msghdr),
1046 [IORING_OP_RECVMSG] = {
1048 .unbound_nonreg_file = 1,
1051 .needs_async_setup = 1,
1052 .async_size = sizeof(struct io_async_msghdr),
1054 [IORING_OP_TIMEOUT] = {
1056 .async_size = sizeof(struct io_timeout_data),
1058 [IORING_OP_TIMEOUT_REMOVE] = {
1059 /* used by timeout updates' prep() */
1062 [IORING_OP_ACCEPT] = {
1064 .unbound_nonreg_file = 1,
1066 .poll_exclusive = 1,
1068 [IORING_OP_ASYNC_CANCEL] = {
1071 [IORING_OP_LINK_TIMEOUT] = {
1073 .async_size = sizeof(struct io_timeout_data),
1075 [IORING_OP_CONNECT] = {
1077 .unbound_nonreg_file = 1,
1079 .needs_async_setup = 1,
1080 .async_size = sizeof(struct io_async_connect),
1082 [IORING_OP_FALLOCATE] = {
1085 [IORING_OP_OPENAT] = {},
1086 [IORING_OP_CLOSE] = {},
1087 [IORING_OP_FILES_UPDATE] = {
1090 [IORING_OP_STATX] = {
1093 [IORING_OP_READ] = {
1095 .unbound_nonreg_file = 1,
1100 .async_size = sizeof(struct io_async_rw),
1102 [IORING_OP_WRITE] = {
1105 .unbound_nonreg_file = 1,
1109 .async_size = sizeof(struct io_async_rw),
1111 [IORING_OP_FADVISE] = {
1115 [IORING_OP_MADVISE] = {},
1116 [IORING_OP_SEND] = {
1118 .unbound_nonreg_file = 1,
1122 [IORING_OP_RECV] = {
1124 .unbound_nonreg_file = 1,
1129 [IORING_OP_OPENAT2] = {
1131 [IORING_OP_EPOLL_CTL] = {
1132 .unbound_nonreg_file = 1,
1135 [IORING_OP_SPLICE] = {
1138 .unbound_nonreg_file = 1,
1141 [IORING_OP_PROVIDE_BUFFERS] = {
1144 [IORING_OP_REMOVE_BUFFERS] = {
1150 .unbound_nonreg_file = 1,
1153 [IORING_OP_SHUTDOWN] = {
1156 [IORING_OP_RENAMEAT] = {},
1157 [IORING_OP_UNLINKAT] = {},
1158 [IORING_OP_MKDIRAT] = {},
1159 [IORING_OP_SYMLINKAT] = {},
1160 [IORING_OP_LINKAT] = {},
1161 [IORING_OP_MSG_RING] = {
1166 /* requests with any of those set should undergo io_disarm_next() */
1167 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1169 static bool io_disarm_next(struct io_kiocb *req);
1170 static void io_uring_del_tctx_node(unsigned long index);
1171 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1172 struct task_struct *task,
1174 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1176 static void io_fill_cqe_req(struct io_kiocb *req, s32 res, u32 cflags);
1178 static void io_put_req(struct io_kiocb *req);
1179 static void io_put_req_deferred(struct io_kiocb *req);
1180 static void io_dismantle_req(struct io_kiocb *req);
1181 static void io_queue_linked_timeout(struct io_kiocb *req);
1182 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1183 struct io_uring_rsrc_update2 *up,
1185 static void io_clean_op(struct io_kiocb *req);
1186 static inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1187 unsigned issue_flags);
1188 static inline struct file *io_file_get_normal(struct io_kiocb *req, int fd);
1189 static void io_drop_inflight_file(struct io_kiocb *req);
1190 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags);
1191 static void __io_queue_sqe(struct io_kiocb *req);
1192 static void io_rsrc_put_work(struct work_struct *work);
1194 static void io_req_task_queue(struct io_kiocb *req);
1195 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
1196 static int io_req_prep_async(struct io_kiocb *req);
1198 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1199 unsigned int issue_flags, u32 slot_index);
1200 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
1202 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
1203 static void io_eventfd_signal(struct io_ring_ctx *ctx);
1205 static struct kmem_cache *req_cachep;
1207 static const struct file_operations io_uring_fops;
1209 struct sock *io_uring_get_socket(struct file *file)
1211 #if defined(CONFIG_UNIX)
1212 if (file->f_op == &io_uring_fops) {
1213 struct io_ring_ctx *ctx = file->private_data;
1215 return ctx->ring_sock->sk;
1220 EXPORT_SYMBOL(io_uring_get_socket);
1222 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1225 mutex_lock(&ctx->uring_lock);
1230 #define io_for_each_link(pos, head) \
1231 for (pos = (head); pos; pos = pos->link)
1234 * Shamelessly stolen from the mm implementation of page reference checking,
1235 * see commit f958d7b528b1 for details.
1237 #define req_ref_zero_or_close_to_overflow(req) \
1238 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1240 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1242 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1243 return atomic_inc_not_zero(&req->refs);
1246 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1248 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1251 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1252 return atomic_dec_and_test(&req->refs);
1255 static inline void req_ref_get(struct io_kiocb *req)
1257 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1258 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1259 atomic_inc(&req->refs);
1262 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
1264 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
1265 __io_submit_flush_completions(ctx);
1268 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1270 if (!(req->flags & REQ_F_REFCOUNT)) {
1271 req->flags |= REQ_F_REFCOUNT;
1272 atomic_set(&req->refs, nr);
1276 static inline void io_req_set_refcount(struct io_kiocb *req)
1278 __io_req_set_refcount(req, 1);
1281 #define IO_RSRC_REF_BATCH 100
1283 static inline void io_req_put_rsrc_locked(struct io_kiocb *req,
1284 struct io_ring_ctx *ctx)
1285 __must_hold(&ctx->uring_lock)
1287 struct percpu_ref *ref = req->fixed_rsrc_refs;
1290 if (ref == &ctx->rsrc_node->refs)
1291 ctx->rsrc_cached_refs++;
1293 percpu_ref_put(ref);
1297 static inline void io_req_put_rsrc(struct io_kiocb *req, struct io_ring_ctx *ctx)
1299 if (req->fixed_rsrc_refs)
1300 percpu_ref_put(req->fixed_rsrc_refs);
1303 static __cold void io_rsrc_refs_drop(struct io_ring_ctx *ctx)
1304 __must_hold(&ctx->uring_lock)
1306 if (ctx->rsrc_cached_refs) {
1307 percpu_ref_put_many(&ctx->rsrc_node->refs, ctx->rsrc_cached_refs);
1308 ctx->rsrc_cached_refs = 0;
1312 static void io_rsrc_refs_refill(struct io_ring_ctx *ctx)
1313 __must_hold(&ctx->uring_lock)
1315 ctx->rsrc_cached_refs += IO_RSRC_REF_BATCH;
1316 percpu_ref_get_many(&ctx->rsrc_node->refs, IO_RSRC_REF_BATCH);
1319 static inline void io_req_set_rsrc_node(struct io_kiocb *req,
1320 struct io_ring_ctx *ctx,
1321 unsigned int issue_flags)
1323 if (!req->fixed_rsrc_refs) {
1324 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1326 if (!(issue_flags & IO_URING_F_UNLOCKED)) {
1327 lockdep_assert_held(&ctx->uring_lock);
1328 ctx->rsrc_cached_refs--;
1329 if (unlikely(ctx->rsrc_cached_refs < 0))
1330 io_rsrc_refs_refill(ctx);
1332 percpu_ref_get(req->fixed_rsrc_refs);
1337 static unsigned int __io_put_kbuf(struct io_kiocb *req, struct list_head *list)
1339 struct io_buffer *kbuf = req->kbuf;
1340 unsigned int cflags;
1342 cflags = IORING_CQE_F_BUFFER | (kbuf->bid << IORING_CQE_BUFFER_SHIFT);
1343 req->flags &= ~REQ_F_BUFFER_SELECTED;
1344 list_add(&kbuf->list, list);
1349 static inline unsigned int io_put_kbuf_comp(struct io_kiocb *req)
1351 lockdep_assert_held(&req->ctx->completion_lock);
1353 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
1355 return __io_put_kbuf(req, &req->ctx->io_buffers_comp);
1358 static inline unsigned int io_put_kbuf(struct io_kiocb *req,
1359 unsigned issue_flags)
1361 unsigned int cflags;
1363 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
1367 * We can add this buffer back to two lists:
1369 * 1) The io_buffers_cache list. This one is protected by the
1370 * ctx->uring_lock. If we already hold this lock, add back to this
1371 * list as we can grab it from issue as well.
1372 * 2) The io_buffers_comp list. This one is protected by the
1373 * ctx->completion_lock.
1375 * We migrate buffers from the comp_list to the issue cache list
1378 if (issue_flags & IO_URING_F_UNLOCKED) {
1379 struct io_ring_ctx *ctx = req->ctx;
1381 spin_lock(&ctx->completion_lock);
1382 cflags = __io_put_kbuf(req, &ctx->io_buffers_comp);
1383 spin_unlock(&ctx->completion_lock);
1385 lockdep_assert_held(&req->ctx->uring_lock);
1387 cflags = __io_put_kbuf(req, &req->ctx->io_buffers_cache);
1393 static struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
1396 struct list_head *hash_list;
1397 struct io_buffer_list *bl;
1399 hash_list = &ctx->io_buffers[hash_32(bgid, IO_BUFFERS_HASH_BITS)];
1400 list_for_each_entry(bl, hash_list, list)
1401 if (bl->bgid == bgid || bgid == -1U)
1407 static void io_kbuf_recycle(struct io_kiocb *req, unsigned issue_flags)
1409 struct io_ring_ctx *ctx = req->ctx;
1410 struct io_buffer_list *bl;
1411 struct io_buffer *buf;
1413 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
1415 /* don't recycle if we already did IO to this buffer */
1416 if (req->flags & REQ_F_PARTIAL_IO)
1419 if (issue_flags & IO_URING_F_UNLOCKED)
1420 mutex_lock(&ctx->uring_lock);
1422 lockdep_assert_held(&ctx->uring_lock);
1425 bl = io_buffer_get_list(ctx, buf->bgid);
1426 list_add(&buf->list, &bl->buf_list);
1427 req->flags &= ~REQ_F_BUFFER_SELECTED;
1430 if (issue_flags & IO_URING_F_UNLOCKED)
1431 mutex_unlock(&ctx->uring_lock);
1434 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1436 __must_hold(&req->ctx->timeout_lock)
1438 if (task && head->task != task)
1444 * As io_match_task() but protected against racing with linked timeouts.
1445 * User must not hold timeout_lock.
1447 static bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
1450 if (task && head->task != task)
1455 static inline bool req_has_async_data(struct io_kiocb *req)
1457 return req->flags & REQ_F_ASYNC_DATA;
1460 static inline void req_set_fail(struct io_kiocb *req)
1462 req->flags |= REQ_F_FAIL;
1463 if (req->flags & REQ_F_CQE_SKIP) {
1464 req->flags &= ~REQ_F_CQE_SKIP;
1465 req->flags |= REQ_F_SKIP_LINK_CQES;
1469 static inline void req_fail_link_node(struct io_kiocb *req, int res)
1475 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
1477 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1479 complete(&ctx->ref_comp);
1482 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1484 return !req->timeout.off;
1487 static __cold void io_fallback_req_func(struct work_struct *work)
1489 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1490 fallback_work.work);
1491 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1492 struct io_kiocb *req, *tmp;
1493 bool locked = false;
1495 percpu_ref_get(&ctx->refs);
1496 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1497 req->io_task_work.func(req, &locked);
1500 io_submit_flush_completions(ctx);
1501 mutex_unlock(&ctx->uring_lock);
1503 percpu_ref_put(&ctx->refs);
1506 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1508 struct io_ring_ctx *ctx;
1511 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1516 * Use 5 bits less than the max cq entries, that should give us around
1517 * 32 entries per hash list if totally full and uniformly spread.
1519 hash_bits = ilog2(p->cq_entries);
1523 ctx->cancel_hash_bits = hash_bits;
1524 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1526 if (!ctx->cancel_hash)
1528 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1530 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1531 if (!ctx->dummy_ubuf)
1533 /* set invalid range, so io_import_fixed() fails meeting it */
1534 ctx->dummy_ubuf->ubuf = -1UL;
1536 ctx->io_buffers = kcalloc(1U << IO_BUFFERS_HASH_BITS,
1537 sizeof(struct list_head), GFP_KERNEL);
1538 if (!ctx->io_buffers)
1540 for (i = 0; i < (1U << IO_BUFFERS_HASH_BITS); i++)
1541 INIT_LIST_HEAD(&ctx->io_buffers[i]);
1543 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1544 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1547 ctx->flags = p->flags;
1548 init_waitqueue_head(&ctx->sqo_sq_wait);
1549 INIT_LIST_HEAD(&ctx->sqd_list);
1550 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1551 INIT_LIST_HEAD(&ctx->io_buffers_cache);
1552 INIT_LIST_HEAD(&ctx->apoll_cache);
1553 init_completion(&ctx->ref_comp);
1554 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1555 mutex_init(&ctx->uring_lock);
1556 init_waitqueue_head(&ctx->cq_wait);
1557 spin_lock_init(&ctx->completion_lock);
1558 spin_lock_init(&ctx->timeout_lock);
1559 INIT_WQ_LIST(&ctx->iopoll_list);
1560 INIT_LIST_HEAD(&ctx->io_buffers_pages);
1561 INIT_LIST_HEAD(&ctx->io_buffers_comp);
1562 INIT_LIST_HEAD(&ctx->defer_list);
1563 INIT_LIST_HEAD(&ctx->timeout_list);
1564 INIT_LIST_HEAD(&ctx->ltimeout_list);
1565 spin_lock_init(&ctx->rsrc_ref_lock);
1566 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1567 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1568 init_llist_head(&ctx->rsrc_put_llist);
1569 INIT_LIST_HEAD(&ctx->tctx_list);
1570 ctx->submit_state.free_list.next = NULL;
1571 INIT_WQ_LIST(&ctx->locked_free_list);
1572 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1573 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
1576 kfree(ctx->dummy_ubuf);
1577 kfree(ctx->cancel_hash);
1578 kfree(ctx->io_buffers);
1583 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1585 struct io_rings *r = ctx->rings;
1587 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1591 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1593 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1594 struct io_ring_ctx *ctx = req->ctx;
1596 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1602 #define FFS_NOWAIT 0x1UL
1603 #define FFS_ISREG 0x2UL
1604 #define FFS_MASK ~(FFS_NOWAIT|FFS_ISREG)
1606 static inline bool io_req_ffs_set(struct io_kiocb *req)
1608 return req->flags & REQ_F_FIXED_FILE;
1611 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1613 if (WARN_ON_ONCE(!req->link))
1616 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1617 req->flags |= REQ_F_LINK_TIMEOUT;
1619 /* linked timeouts should have two refs once prep'ed */
1620 io_req_set_refcount(req);
1621 __io_req_set_refcount(req->link, 2);
1625 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1627 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1629 return __io_prep_linked_timeout(req);
1632 static void io_prep_async_work(struct io_kiocb *req)
1634 const struct io_op_def *def = &io_op_defs[req->opcode];
1635 struct io_ring_ctx *ctx = req->ctx;
1637 if (!(req->flags & REQ_F_CREDS)) {
1638 req->flags |= REQ_F_CREDS;
1639 req->creds = get_current_cred();
1642 req->work.list.next = NULL;
1643 req->work.flags = 0;
1644 if (req->flags & REQ_F_FORCE_ASYNC)
1645 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1647 if (req->flags & REQ_F_ISREG) {
1648 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1649 io_wq_hash_work(&req->work, file_inode(req->file));
1650 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1651 if (def->unbound_nonreg_file)
1652 req->work.flags |= IO_WQ_WORK_UNBOUND;
1656 static void io_prep_async_link(struct io_kiocb *req)
1658 struct io_kiocb *cur;
1660 if (req->flags & REQ_F_LINK_TIMEOUT) {
1661 struct io_ring_ctx *ctx = req->ctx;
1663 spin_lock_irq(&ctx->timeout_lock);
1664 io_for_each_link(cur, req)
1665 io_prep_async_work(cur);
1666 spin_unlock_irq(&ctx->timeout_lock);
1668 io_for_each_link(cur, req)
1669 io_prep_async_work(cur);
1673 static inline void io_req_add_compl_list(struct io_kiocb *req)
1675 struct io_ring_ctx *ctx = req->ctx;
1676 struct io_submit_state *state = &ctx->submit_state;
1678 if (!(req->flags & REQ_F_CQE_SKIP))
1679 ctx->submit_state.flush_cqes = true;
1680 wq_list_add_tail(&req->comp_list, &state->compl_reqs);
1683 static void io_queue_async_work(struct io_kiocb *req, bool *dont_use)
1685 struct io_ring_ctx *ctx = req->ctx;
1686 struct io_kiocb *link = io_prep_linked_timeout(req);
1687 struct io_uring_task *tctx = req->task->io_uring;
1690 BUG_ON(!tctx->io_wq);
1692 /* init ->work of the whole link before punting */
1693 io_prep_async_link(req);
1696 * Not expected to happen, but if we do have a bug where this _can_
1697 * happen, catch it here and ensure the request is marked as
1698 * canceled. That will make io-wq go through the usual work cancel
1699 * procedure rather than attempt to run this request (or create a new
1702 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1703 req->work.flags |= IO_WQ_WORK_CANCEL;
1705 trace_io_uring_queue_async_work(ctx, req, req->user_data, req->opcode, req->flags,
1706 &req->work, io_wq_is_hashed(&req->work));
1707 io_wq_enqueue(tctx->io_wq, &req->work);
1709 io_queue_linked_timeout(link);
1712 static void io_kill_timeout(struct io_kiocb *req, int status)
1713 __must_hold(&req->ctx->completion_lock)
1714 __must_hold(&req->ctx->timeout_lock)
1716 struct io_timeout_data *io = req->async_data;
1718 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1721 atomic_set(&req->ctx->cq_timeouts,
1722 atomic_read(&req->ctx->cq_timeouts) + 1);
1723 list_del_init(&req->timeout.list);
1724 io_fill_cqe_req(req, status, 0);
1725 io_put_req_deferred(req);
1729 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
1731 while (!list_empty(&ctx->defer_list)) {
1732 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1733 struct io_defer_entry, list);
1735 if (req_need_defer(de->req, de->seq))
1737 list_del_init(&de->list);
1738 io_req_task_queue(de->req);
1743 static __cold void io_flush_timeouts(struct io_ring_ctx *ctx)
1744 __must_hold(&ctx->completion_lock)
1746 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1747 struct io_kiocb *req, *tmp;
1749 spin_lock_irq(&ctx->timeout_lock);
1750 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1751 u32 events_needed, events_got;
1753 if (io_is_timeout_noseq(req))
1757 * Since seq can easily wrap around over time, subtract
1758 * the last seq at which timeouts were flushed before comparing.
1759 * Assuming not more than 2^31-1 events have happened since,
1760 * these subtractions won't have wrapped, so we can check if
1761 * target is in [last_seq, current_seq] by comparing the two.
1763 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1764 events_got = seq - ctx->cq_last_tm_flush;
1765 if (events_got < events_needed)
1768 io_kill_timeout(req, 0);
1770 ctx->cq_last_tm_flush = seq;
1771 spin_unlock_irq(&ctx->timeout_lock);
1774 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1776 /* order cqe stores with ring update */
1777 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1780 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1782 if (ctx->off_timeout_used || ctx->drain_active) {
1783 spin_lock(&ctx->completion_lock);
1784 if (ctx->off_timeout_used)
1785 io_flush_timeouts(ctx);
1786 if (ctx->drain_active)
1787 io_queue_deferred(ctx);
1788 io_commit_cqring(ctx);
1789 spin_unlock(&ctx->completion_lock);
1792 io_eventfd_signal(ctx);
1795 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1797 struct io_rings *r = ctx->rings;
1799 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1802 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1804 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1807 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1809 struct io_rings *rings = ctx->rings;
1810 unsigned tail, mask = ctx->cq_entries - 1;
1813 * writes to the cq entry need to come after reading head; the
1814 * control dependency is enough as we're using WRITE_ONCE to
1817 if (__io_cqring_events(ctx) == ctx->cq_entries)
1820 tail = ctx->cached_cq_tail++;
1821 return &rings->cqes[tail & mask];
1824 static void io_eventfd_signal(struct io_ring_ctx *ctx)
1826 struct io_ev_fd *ev_fd;
1830 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
1831 * and eventfd_signal
1833 ev_fd = rcu_dereference(ctx->io_ev_fd);
1836 * Check again if ev_fd exists incase an io_eventfd_unregister call
1837 * completed between the NULL check of ctx->io_ev_fd at the start of
1838 * the function and rcu_read_lock.
1840 if (unlikely(!ev_fd))
1842 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1845 if (!ev_fd->eventfd_async || io_wq_current_is_worker())
1846 eventfd_signal(ev_fd->cq_ev_fd, 1);
1851 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
1854 * wake_up_all() may seem excessive, but io_wake_function() and
1855 * io_should_wake() handle the termination of the loop and only
1856 * wake as many waiters as we need to.
1858 if (wq_has_sleeper(&ctx->cq_wait))
1859 wake_up_all(&ctx->cq_wait);
1863 * This should only get called when at least one event has been posted.
1864 * Some applications rely on the eventfd notification count only changing
1865 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1866 * 1:1 relationship between how many times this function is called (and
1867 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1869 static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1871 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
1873 __io_commit_cqring_flush(ctx);
1875 io_cqring_wake(ctx);
1878 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1880 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
1882 __io_commit_cqring_flush(ctx);
1884 if (ctx->flags & IORING_SETUP_SQPOLL)
1885 io_cqring_wake(ctx);
1888 /* Returns true if there are no backlogged entries after the flush */
1889 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1891 bool all_flushed, posted;
1893 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1897 spin_lock(&ctx->completion_lock);
1898 while (!list_empty(&ctx->cq_overflow_list)) {
1899 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1900 struct io_overflow_cqe *ocqe;
1904 ocqe = list_first_entry(&ctx->cq_overflow_list,
1905 struct io_overflow_cqe, list);
1907 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1909 io_account_cq_overflow(ctx);
1912 list_del(&ocqe->list);
1916 all_flushed = list_empty(&ctx->cq_overflow_list);
1918 clear_bit(0, &ctx->check_cq_overflow);
1919 WRITE_ONCE(ctx->rings->sq_flags,
1920 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1924 io_commit_cqring(ctx);
1925 spin_unlock(&ctx->completion_lock);
1927 io_cqring_ev_posted(ctx);
1931 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1935 if (test_bit(0, &ctx->check_cq_overflow)) {
1936 /* iopoll syncs against uring_lock, not completion_lock */
1937 if (ctx->flags & IORING_SETUP_IOPOLL)
1938 mutex_lock(&ctx->uring_lock);
1939 ret = __io_cqring_overflow_flush(ctx, false);
1940 if (ctx->flags & IORING_SETUP_IOPOLL)
1941 mutex_unlock(&ctx->uring_lock);
1947 /* must to be called somewhat shortly after putting a request */
1948 static inline void io_put_task(struct task_struct *task, int nr)
1950 struct io_uring_task *tctx = task->io_uring;
1952 if (likely(task == current)) {
1953 tctx->cached_refs += nr;
1955 percpu_counter_sub(&tctx->inflight, nr);
1956 if (unlikely(atomic_read(&tctx->in_idle)))
1957 wake_up(&tctx->wait);
1958 put_task_struct_many(task, nr);
1962 static void io_task_refs_refill(struct io_uring_task *tctx)
1964 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1966 percpu_counter_add(&tctx->inflight, refill);
1967 refcount_add(refill, ¤t->usage);
1968 tctx->cached_refs += refill;
1971 static inline void io_get_task_refs(int nr)
1973 struct io_uring_task *tctx = current->io_uring;
1975 tctx->cached_refs -= nr;
1976 if (unlikely(tctx->cached_refs < 0))
1977 io_task_refs_refill(tctx);
1980 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
1982 struct io_uring_task *tctx = task->io_uring;
1983 unsigned int refs = tctx->cached_refs;
1986 tctx->cached_refs = 0;
1987 percpu_counter_sub(&tctx->inflight, refs);
1988 put_task_struct_many(task, refs);
1992 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1993 s32 res, u32 cflags)
1995 struct io_overflow_cqe *ocqe;
1997 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
2000 * If we're in ring overflow flush mode, or in task cancel mode,
2001 * or cannot allocate an overflow entry, then we need to drop it
2004 io_account_cq_overflow(ctx);
2007 if (list_empty(&ctx->cq_overflow_list)) {
2008 set_bit(0, &ctx->check_cq_overflow);
2009 WRITE_ONCE(ctx->rings->sq_flags,
2010 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
2013 ocqe->cqe.user_data = user_data;
2014 ocqe->cqe.res = res;
2015 ocqe->cqe.flags = cflags;
2016 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
2020 static inline bool __io_fill_cqe(struct io_ring_ctx *ctx, u64 user_data,
2021 s32 res, u32 cflags)
2023 struct io_uring_cqe *cqe;
2026 * If we can't get a cq entry, userspace overflowed the
2027 * submission (by quite a lot). Increment the overflow count in
2030 cqe = io_get_cqe(ctx);
2032 WRITE_ONCE(cqe->user_data, user_data);
2033 WRITE_ONCE(cqe->res, res);
2034 WRITE_ONCE(cqe->flags, cflags);
2037 return io_cqring_event_overflow(ctx, user_data, res, cflags);
2040 static inline bool __io_fill_cqe_req(struct io_kiocb *req, s32 res, u32 cflags)
2042 trace_io_uring_complete(req->ctx, req, req->user_data, res, cflags);
2043 return __io_fill_cqe(req->ctx, req->user_data, res, cflags);
2046 static noinline void io_fill_cqe_req(struct io_kiocb *req, s32 res, u32 cflags)
2048 if (!(req->flags & REQ_F_CQE_SKIP))
2049 __io_fill_cqe_req(req, res, cflags);
2052 static noinline bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data,
2053 s32 res, u32 cflags)
2056 trace_io_uring_complete(ctx, NULL, user_data, res, cflags);
2057 return __io_fill_cqe(ctx, user_data, res, cflags);
2060 static void __io_req_complete_post(struct io_kiocb *req, s32 res,
2063 struct io_ring_ctx *ctx = req->ctx;
2065 if (!(req->flags & REQ_F_CQE_SKIP))
2066 __io_fill_cqe_req(req, res, cflags);
2068 * If we're the last reference to this request, add to our locked
2071 if (req_ref_put_and_test(req)) {
2072 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
2073 if (req->flags & IO_DISARM_MASK)
2074 io_disarm_next(req);
2076 io_req_task_queue(req->link);
2080 io_req_put_rsrc(req, ctx);
2082 * Selected buffer deallocation in io_clean_op() assumes that
2083 * we don't hold ->completion_lock. Clean them here to avoid
2086 io_put_kbuf_comp(req);
2087 io_dismantle_req(req);
2088 io_put_task(req->task, 1);
2089 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
2090 ctx->locked_free_nr++;
2094 static void io_req_complete_post(struct io_kiocb *req, s32 res,
2097 struct io_ring_ctx *ctx = req->ctx;
2099 spin_lock(&ctx->completion_lock);
2100 __io_req_complete_post(req, res, cflags);
2101 io_commit_cqring(ctx);
2102 spin_unlock(&ctx->completion_lock);
2103 io_cqring_ev_posted(ctx);
2106 static inline void io_req_complete_state(struct io_kiocb *req, s32 res,
2110 req->cflags = cflags;
2111 req->flags |= REQ_F_COMPLETE_INLINE;
2114 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
2115 s32 res, u32 cflags)
2117 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
2118 io_req_complete_state(req, res, cflags);
2120 io_req_complete_post(req, res, cflags);
2123 static inline void io_req_complete(struct io_kiocb *req, s32 res)
2125 __io_req_complete(req, 0, res, 0);
2128 static void io_req_complete_failed(struct io_kiocb *req, s32 res)
2131 io_req_complete_post(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
2134 static void io_req_complete_fail_submit(struct io_kiocb *req)
2137 * We don't submit, fail them all, for that replace hardlinks with
2138 * normal links. Extra REQ_F_LINK is tolerated.
2140 req->flags &= ~REQ_F_HARDLINK;
2141 req->flags |= REQ_F_LINK;
2142 io_req_complete_failed(req, req->result);
2146 * Don't initialise the fields below on every allocation, but do that in
2147 * advance and keep them valid across allocations.
2149 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
2153 req->async_data = NULL;
2154 /* not necessary, but safer to zero */
2158 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
2159 struct io_submit_state *state)
2161 spin_lock(&ctx->completion_lock);
2162 wq_list_splice(&ctx->locked_free_list, &state->free_list);
2163 ctx->locked_free_nr = 0;
2164 spin_unlock(&ctx->completion_lock);
2167 /* Returns true IFF there are requests in the cache */
2168 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
2170 struct io_submit_state *state = &ctx->submit_state;
2173 * If we have more than a batch's worth of requests in our IRQ side
2174 * locked cache, grab the lock and move them over to our submission
2177 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
2178 io_flush_cached_locked_reqs(ctx, state);
2179 return !!state->free_list.next;
2183 * A request might get retired back into the request caches even before opcode
2184 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
2185 * Because of that, io_alloc_req() should be called only under ->uring_lock
2186 * and with extra caution to not get a request that is still worked on.
2188 static __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
2189 __must_hold(&ctx->uring_lock)
2191 struct io_submit_state *state = &ctx->submit_state;
2192 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
2193 void *reqs[IO_REQ_ALLOC_BATCH];
2194 struct io_kiocb *req;
2197 if (likely(state->free_list.next || io_flush_cached_reqs(ctx)))
2200 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
2203 * Bulk alloc is all-or-nothing. If we fail to get a batch,
2204 * retry single alloc to be on the safe side.
2206 if (unlikely(ret <= 0)) {
2207 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
2213 percpu_ref_get_many(&ctx->refs, ret);
2214 for (i = 0; i < ret; i++) {
2217 io_preinit_req(req, ctx);
2218 wq_stack_add_head(&req->comp_list, &state->free_list);
2223 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
2225 if (unlikely(!ctx->submit_state.free_list.next))
2226 return __io_alloc_req_refill(ctx);
2230 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
2232 struct io_wq_work_node *node;
2234 node = wq_stack_extract(&ctx->submit_state.free_list);
2235 return container_of(node, struct io_kiocb, comp_list);
2238 static inline void io_put_file(struct file *file)
2244 static inline void io_dismantle_req(struct io_kiocb *req)
2246 unsigned int flags = req->flags;
2248 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
2250 if (!(flags & REQ_F_FIXED_FILE))
2251 io_put_file(req->file);
2254 static __cold void __io_free_req(struct io_kiocb *req)
2256 struct io_ring_ctx *ctx = req->ctx;
2258 io_req_put_rsrc(req, ctx);
2259 io_dismantle_req(req);
2260 io_put_task(req->task, 1);
2262 spin_lock(&ctx->completion_lock);
2263 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
2264 ctx->locked_free_nr++;
2265 spin_unlock(&ctx->completion_lock);
2268 static inline void io_remove_next_linked(struct io_kiocb *req)
2270 struct io_kiocb *nxt = req->link;
2272 req->link = nxt->link;
2276 static bool io_kill_linked_timeout(struct io_kiocb *req)
2277 __must_hold(&req->ctx->completion_lock)
2278 __must_hold(&req->ctx->timeout_lock)
2280 struct io_kiocb *link = req->link;
2282 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2283 struct io_timeout_data *io = link->async_data;
2285 io_remove_next_linked(req);
2286 link->timeout.head = NULL;
2287 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2288 list_del(&link->timeout.list);
2289 /* leave REQ_F_CQE_SKIP to io_fill_cqe_req */
2290 io_fill_cqe_req(link, -ECANCELED, 0);
2291 io_put_req_deferred(link);
2298 static void io_fail_links(struct io_kiocb *req)
2299 __must_hold(&req->ctx->completion_lock)
2301 struct io_kiocb *nxt, *link = req->link;
2302 bool ignore_cqes = req->flags & REQ_F_SKIP_LINK_CQES;
2306 long res = -ECANCELED;
2308 if (link->flags & REQ_F_FAIL)
2314 trace_io_uring_fail_link(req->ctx, req, req->user_data,
2318 link->flags &= ~REQ_F_CQE_SKIP;
2319 io_fill_cqe_req(link, res, 0);
2321 io_put_req_deferred(link);
2326 static bool io_disarm_next(struct io_kiocb *req)
2327 __must_hold(&req->ctx->completion_lock)
2329 bool posted = false;
2331 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2332 struct io_kiocb *link = req->link;
2334 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2335 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2336 io_remove_next_linked(req);
2337 /* leave REQ_F_CQE_SKIP to io_fill_cqe_req */
2338 io_fill_cqe_req(link, -ECANCELED, 0);
2339 io_put_req_deferred(link);
2342 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2343 struct io_ring_ctx *ctx = req->ctx;
2345 spin_lock_irq(&ctx->timeout_lock);
2346 posted = io_kill_linked_timeout(req);
2347 spin_unlock_irq(&ctx->timeout_lock);
2349 if (unlikely((req->flags & REQ_F_FAIL) &&
2350 !(req->flags & REQ_F_HARDLINK))) {
2351 posted |= (req->link != NULL);
2357 static void __io_req_find_next_prep(struct io_kiocb *req)
2359 struct io_ring_ctx *ctx = req->ctx;
2362 spin_lock(&ctx->completion_lock);
2363 posted = io_disarm_next(req);
2365 io_commit_cqring(ctx);
2366 spin_unlock(&ctx->completion_lock);
2368 io_cqring_ev_posted(ctx);
2371 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2373 struct io_kiocb *nxt;
2375 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2378 * If LINK is set, we have dependent requests in this chain. If we
2379 * didn't fail this request, queue the first one up, moving any other
2380 * dependencies to the next request. In case of failure, fail the rest
2383 if (unlikely(req->flags & IO_DISARM_MASK))
2384 __io_req_find_next_prep(req);
2390 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2395 io_submit_flush_completions(ctx);
2396 mutex_unlock(&ctx->uring_lock);
2399 percpu_ref_put(&ctx->refs);
2402 static inline void ctx_commit_and_unlock(struct io_ring_ctx *ctx)
2404 io_commit_cqring(ctx);
2405 spin_unlock(&ctx->completion_lock);
2406 io_cqring_ev_posted(ctx);
2409 static void handle_prev_tw_list(struct io_wq_work_node *node,
2410 struct io_ring_ctx **ctx, bool *uring_locked)
2412 if (*ctx && !*uring_locked)
2413 spin_lock(&(*ctx)->completion_lock);
2416 struct io_wq_work_node *next = node->next;
2417 struct io_kiocb *req = container_of(node, struct io_kiocb,
2420 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
2422 if (req->ctx != *ctx) {
2423 if (unlikely(!*uring_locked && *ctx))
2424 ctx_commit_and_unlock(*ctx);
2426 ctx_flush_and_put(*ctx, uring_locked);
2428 /* if not contended, grab and improve batching */
2429 *uring_locked = mutex_trylock(&(*ctx)->uring_lock);
2430 percpu_ref_get(&(*ctx)->refs);
2431 if (unlikely(!*uring_locked))
2432 spin_lock(&(*ctx)->completion_lock);
2434 if (likely(*uring_locked))
2435 req->io_task_work.func(req, uring_locked);
2437 __io_req_complete_post(req, req->result,
2438 io_put_kbuf_comp(req));
2442 if (unlikely(!*uring_locked))
2443 ctx_commit_and_unlock(*ctx);
2446 static void handle_tw_list(struct io_wq_work_node *node,
2447 struct io_ring_ctx **ctx, bool *locked)
2450 struct io_wq_work_node *next = node->next;
2451 struct io_kiocb *req = container_of(node, struct io_kiocb,
2454 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
2456 if (req->ctx != *ctx) {
2457 ctx_flush_and_put(*ctx, locked);
2459 /* if not contended, grab and improve batching */
2460 *locked = mutex_trylock(&(*ctx)->uring_lock);
2461 percpu_ref_get(&(*ctx)->refs);
2463 req->io_task_work.func(req, locked);
2468 static void tctx_task_work(struct callback_head *cb)
2470 bool uring_locked = false;
2471 struct io_ring_ctx *ctx = NULL;
2472 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2476 struct io_wq_work_node *node1, *node2;
2478 if (!tctx->task_list.first &&
2479 !tctx->prior_task_list.first && uring_locked)
2480 io_submit_flush_completions(ctx);
2482 spin_lock_irq(&tctx->task_lock);
2483 node1 = tctx->prior_task_list.first;
2484 node2 = tctx->task_list.first;
2485 INIT_WQ_LIST(&tctx->task_list);
2486 INIT_WQ_LIST(&tctx->prior_task_list);
2487 if (!node2 && !node1)
2488 tctx->task_running = false;
2489 spin_unlock_irq(&tctx->task_lock);
2490 if (!node2 && !node1)
2494 handle_prev_tw_list(node1, &ctx, &uring_locked);
2497 handle_tw_list(node2, &ctx, &uring_locked);
2501 ctx_flush_and_put(ctx, &uring_locked);
2503 /* relaxed read is enough as only the task itself sets ->in_idle */
2504 if (unlikely(atomic_read(&tctx->in_idle)))
2505 io_uring_drop_tctx_refs(current);
2508 static void io_req_task_work_add(struct io_kiocb *req, bool priority)
2510 struct task_struct *tsk = req->task;
2511 struct io_uring_task *tctx = tsk->io_uring;
2512 enum task_work_notify_mode notify;
2513 struct io_wq_work_node *node;
2514 unsigned long flags;
2517 WARN_ON_ONCE(!tctx);
2519 io_drop_inflight_file(req);
2521 spin_lock_irqsave(&tctx->task_lock, flags);
2523 wq_list_add_tail(&req->io_task_work.node, &tctx->prior_task_list);
2525 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2526 running = tctx->task_running;
2528 tctx->task_running = true;
2529 spin_unlock_irqrestore(&tctx->task_lock, flags);
2531 /* task_work already pending, we're done */
2536 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2537 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2538 * processing task_work. There's no reliable way to tell if TWA_RESUME
2541 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2542 if (likely(!task_work_add(tsk, &tctx->task_work, notify))) {
2543 if (notify == TWA_NONE)
2544 wake_up_process(tsk);
2548 spin_lock_irqsave(&tctx->task_lock, flags);
2549 tctx->task_running = false;
2550 node = wq_list_merge(&tctx->prior_task_list, &tctx->task_list);
2551 spin_unlock_irqrestore(&tctx->task_lock, flags);
2554 req = container_of(node, struct io_kiocb, io_task_work.node);
2556 if (llist_add(&req->io_task_work.fallback_node,
2557 &req->ctx->fallback_llist))
2558 schedule_delayed_work(&req->ctx->fallback_work, 1);
2562 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2564 struct io_ring_ctx *ctx = req->ctx;
2566 /* not needed for normal modes, but SQPOLL depends on it */
2567 io_tw_lock(ctx, locked);
2568 io_req_complete_failed(req, req->result);
2571 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2573 struct io_ring_ctx *ctx = req->ctx;
2575 io_tw_lock(ctx, locked);
2576 /* req->task == current here, checking PF_EXITING is safe */
2577 if (likely(!(req->task->flags & PF_EXITING)))
2578 __io_queue_sqe(req);
2580 io_req_complete_failed(req, -EFAULT);
2583 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2586 req->io_task_work.func = io_req_task_cancel;
2587 io_req_task_work_add(req, false);
2590 static void io_req_task_queue(struct io_kiocb *req)
2592 req->io_task_work.func = io_req_task_submit;
2593 io_req_task_work_add(req, false);
2596 static void io_req_task_queue_reissue(struct io_kiocb *req)
2598 req->io_task_work.func = io_queue_async_work;
2599 io_req_task_work_add(req, false);
2602 static inline void io_queue_next(struct io_kiocb *req)
2604 struct io_kiocb *nxt = io_req_find_next(req);
2607 io_req_task_queue(nxt);
2610 static void io_free_req(struct io_kiocb *req)
2616 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2621 static void io_free_batch_list(struct io_ring_ctx *ctx,
2622 struct io_wq_work_node *node)
2623 __must_hold(&ctx->uring_lock)
2625 struct task_struct *task = NULL;
2629 struct io_kiocb *req = container_of(node, struct io_kiocb,
2632 if (unlikely(req->flags & REQ_F_REFCOUNT)) {
2633 node = req->comp_list.next;
2634 if (!req_ref_put_and_test(req))
2638 io_req_put_rsrc_locked(req, ctx);
2640 io_dismantle_req(req);
2642 if (req->task != task) {
2644 io_put_task(task, task_refs);
2649 node = req->comp_list.next;
2650 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
2654 io_put_task(task, task_refs);
2657 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
2658 __must_hold(&ctx->uring_lock)
2660 struct io_wq_work_node *node, *prev;
2661 struct io_submit_state *state = &ctx->submit_state;
2663 if (state->flush_cqes) {
2664 spin_lock(&ctx->completion_lock);
2665 wq_list_for_each(node, prev, &state->compl_reqs) {
2666 struct io_kiocb *req = container_of(node, struct io_kiocb,
2669 if (!(req->flags & REQ_F_CQE_SKIP))
2670 __io_fill_cqe_req(req, req->result, req->cflags);
2671 if ((req->flags & REQ_F_POLLED) && req->apoll) {
2672 struct async_poll *apoll = req->apoll;
2674 if (apoll->double_poll)
2675 kfree(apoll->double_poll);
2676 list_add(&apoll->poll.wait.entry,
2678 req->flags &= ~REQ_F_POLLED;
2682 io_commit_cqring(ctx);
2683 spin_unlock(&ctx->completion_lock);
2684 io_cqring_ev_posted(ctx);
2685 state->flush_cqes = false;
2688 io_free_batch_list(ctx, state->compl_reqs.first);
2689 INIT_WQ_LIST(&state->compl_reqs);
2693 * Drop reference to request, return next in chain (if there is one) if this
2694 * was the last reference to this request.
2696 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2698 struct io_kiocb *nxt = NULL;
2700 if (req_ref_put_and_test(req)) {
2701 nxt = io_req_find_next(req);
2707 static inline void io_put_req(struct io_kiocb *req)
2709 if (req_ref_put_and_test(req))
2713 static inline void io_put_req_deferred(struct io_kiocb *req)
2715 if (req_ref_put_and_test(req)) {
2716 req->io_task_work.func = io_free_req_work;
2717 io_req_task_work_add(req, false);
2721 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2723 /* See comment at the top of this file */
2725 return __io_cqring_events(ctx);
2728 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2730 struct io_rings *rings = ctx->rings;
2732 /* make sure SQ entry isn't read before tail */
2733 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2736 static inline bool io_run_task_work(void)
2738 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || task_work_pending(current)) {
2739 __set_current_state(TASK_RUNNING);
2740 clear_notify_signal();
2741 if (task_work_pending(current))
2749 static int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
2751 struct io_wq_work_node *pos, *start, *prev;
2752 unsigned int poll_flags = BLK_POLL_NOSLEEP;
2753 DEFINE_IO_COMP_BATCH(iob);
2757 * Only spin for completions if we don't have multiple devices hanging
2758 * off our complete list.
2760 if (ctx->poll_multi_queue || force_nonspin)
2761 poll_flags |= BLK_POLL_ONESHOT;
2763 wq_list_for_each(pos, start, &ctx->iopoll_list) {
2764 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
2765 struct kiocb *kiocb = &req->rw.kiocb;
2769 * Move completed and retryable entries to our local lists.
2770 * If we find a request that requires polling, break out
2771 * and complete those lists first, if we have entries there.
2773 if (READ_ONCE(req->iopoll_completed))
2776 ret = kiocb->ki_filp->f_op->iopoll(kiocb, &iob, poll_flags);
2777 if (unlikely(ret < 0))
2780 poll_flags |= BLK_POLL_ONESHOT;
2782 /* iopoll may have completed current req */
2783 if (!rq_list_empty(iob.req_list) ||
2784 READ_ONCE(req->iopoll_completed))
2788 if (!rq_list_empty(iob.req_list))
2794 wq_list_for_each_resume(pos, prev) {
2795 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
2797 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
2798 if (!smp_load_acquire(&req->iopoll_completed))
2800 if (unlikely(req->flags & REQ_F_CQE_SKIP))
2803 __io_fill_cqe_req(req, req->result, io_put_kbuf(req, 0));
2807 if (unlikely(!nr_events))
2810 io_commit_cqring(ctx);
2811 io_cqring_ev_posted_iopoll(ctx);
2812 pos = start ? start->next : ctx->iopoll_list.first;
2813 wq_list_cut(&ctx->iopoll_list, prev, start);
2814 io_free_batch_list(ctx, pos);
2819 * We can't just wait for polled events to come to us, we have to actively
2820 * find and complete them.
2822 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2824 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2827 mutex_lock(&ctx->uring_lock);
2828 while (!wq_list_empty(&ctx->iopoll_list)) {
2829 /* let it sleep and repeat later if can't complete a request */
2830 if (io_do_iopoll(ctx, true) == 0)
2833 * Ensure we allow local-to-the-cpu processing to take place,
2834 * in this case we need to ensure that we reap all events.
2835 * Also let task_work, etc. to progress by releasing the mutex
2837 if (need_resched()) {
2838 mutex_unlock(&ctx->uring_lock);
2840 mutex_lock(&ctx->uring_lock);
2843 mutex_unlock(&ctx->uring_lock);
2846 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2848 unsigned int nr_events = 0;
2852 * We disallow the app entering submit/complete with polling, but we
2853 * still need to lock the ring to prevent racing with polled issue
2854 * that got punted to a workqueue.
2856 mutex_lock(&ctx->uring_lock);
2858 * Don't enter poll loop if we already have events pending.
2859 * If we do, we can potentially be spinning for commands that
2860 * already triggered a CQE (eg in error).
2862 if (test_bit(0, &ctx->check_cq_overflow))
2863 __io_cqring_overflow_flush(ctx, false);
2864 if (io_cqring_events(ctx))
2868 * If a submit got punted to a workqueue, we can have the
2869 * application entering polling for a command before it gets
2870 * issued. That app will hold the uring_lock for the duration
2871 * of the poll right here, so we need to take a breather every
2872 * now and then to ensure that the issue has a chance to add
2873 * the poll to the issued list. Otherwise we can spin here
2874 * forever, while the workqueue is stuck trying to acquire the
2877 if (wq_list_empty(&ctx->iopoll_list)) {
2878 u32 tail = ctx->cached_cq_tail;
2880 mutex_unlock(&ctx->uring_lock);
2882 mutex_lock(&ctx->uring_lock);
2884 /* some requests don't go through iopoll_list */
2885 if (tail != ctx->cached_cq_tail ||
2886 wq_list_empty(&ctx->iopoll_list))
2889 ret = io_do_iopoll(ctx, !min);
2894 } while (nr_events < min && !need_resched());
2896 mutex_unlock(&ctx->uring_lock);
2900 static void kiocb_end_write(struct io_kiocb *req)
2903 * Tell lockdep we inherited freeze protection from submission
2906 if (req->flags & REQ_F_ISREG) {
2907 struct super_block *sb = file_inode(req->file)->i_sb;
2909 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2915 static bool io_resubmit_prep(struct io_kiocb *req)
2917 struct io_async_rw *rw = req->async_data;
2919 if (!req_has_async_data(req))
2920 return !io_req_prep_async(req);
2921 iov_iter_restore(&rw->s.iter, &rw->s.iter_state);
2925 static bool io_rw_should_reissue(struct io_kiocb *req)
2927 umode_t mode = file_inode(req->file)->i_mode;
2928 struct io_ring_ctx *ctx = req->ctx;
2930 if (!S_ISBLK(mode) && !S_ISREG(mode))
2932 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2933 !(ctx->flags & IORING_SETUP_IOPOLL)))
2936 * If ref is dying, we might be running poll reap from the exit work.
2937 * Don't attempt to reissue from that path, just let it fail with
2940 if (percpu_ref_is_dying(&ctx->refs))
2943 * Play it safe and assume not safe to re-import and reissue if we're
2944 * not in the original thread group (or in task context).
2946 if (!same_thread_group(req->task, current) || !in_task())
2951 static bool io_resubmit_prep(struct io_kiocb *req)
2955 static bool io_rw_should_reissue(struct io_kiocb *req)
2961 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2963 if (req->rw.kiocb.ki_flags & IOCB_WRITE) {
2964 kiocb_end_write(req);
2965 fsnotify_modify(req->file);
2967 fsnotify_access(req->file);
2969 if (unlikely(res != req->result)) {
2970 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2971 io_rw_should_reissue(req)) {
2972 req->flags |= REQ_F_REISSUE;
2981 static inline void io_req_task_complete(struct io_kiocb *req, bool *locked)
2983 int res = req->result;
2986 io_req_complete_state(req, res, io_put_kbuf(req, 0));
2987 io_req_add_compl_list(req);
2989 io_req_complete_post(req, res,
2990 io_put_kbuf(req, IO_URING_F_UNLOCKED));
2994 static void __io_complete_rw(struct io_kiocb *req, long res,
2995 unsigned int issue_flags)
2997 if (__io_complete_rw_common(req, res))
2999 __io_req_complete(req, issue_flags, req->result,
3000 io_put_kbuf(req, issue_flags));
3003 static void io_complete_rw(struct kiocb *kiocb, long res)
3005 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
3007 if (__io_complete_rw_common(req, res))
3010 req->io_task_work.func = io_req_task_complete;
3011 io_req_task_work_add(req, !!(req->ctx->flags & IORING_SETUP_SQPOLL));
3014 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
3016 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
3018 if (kiocb->ki_flags & IOCB_WRITE)
3019 kiocb_end_write(req);
3020 if (unlikely(res != req->result)) {
3021 if (res == -EAGAIN && io_rw_should_reissue(req)) {
3022 req->flags |= REQ_F_REISSUE;
3028 /* order with io_iopoll_complete() checking ->iopoll_completed */
3029 smp_store_release(&req->iopoll_completed, 1);
3033 * After the iocb has been issued, it's safe to be found on the poll list.
3034 * Adding the kiocb to the list AFTER submission ensures that we don't
3035 * find it from a io_do_iopoll() thread before the issuer is done
3036 * accessing the kiocb cookie.
3038 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
3040 struct io_ring_ctx *ctx = req->ctx;
3041 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
3043 /* workqueue context doesn't hold uring_lock, grab it now */
3044 if (unlikely(needs_lock))
3045 mutex_lock(&ctx->uring_lock);
3048 * Track whether we have multiple files in our lists. This will impact
3049 * how we do polling eventually, not spinning if we're on potentially
3050 * different devices.
3052 if (wq_list_empty(&ctx->iopoll_list)) {
3053 ctx->poll_multi_queue = false;
3054 } else if (!ctx->poll_multi_queue) {
3055 struct io_kiocb *list_req;
3057 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
3059 if (list_req->file != req->file)
3060 ctx->poll_multi_queue = true;
3064 * For fast devices, IO may have already completed. If it has, add
3065 * it to the front so we find it first.
3067 if (READ_ONCE(req->iopoll_completed))
3068 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
3070 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
3072 if (unlikely(needs_lock)) {
3074 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
3075 * in sq thread task context or in io worker task context. If
3076 * current task context is sq thread, we don't need to check
3077 * whether should wake up sq thread.
3079 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
3080 wq_has_sleeper(&ctx->sq_data->wait))
3081 wake_up(&ctx->sq_data->wait);
3083 mutex_unlock(&ctx->uring_lock);
3087 static bool io_bdev_nowait(struct block_device *bdev)
3089 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
3093 * If we tracked the file through the SCM inflight mechanism, we could support
3094 * any file. For now, just ensure that anything potentially problematic is done
3097 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
3099 if (S_ISBLK(mode)) {
3100 if (IS_ENABLED(CONFIG_BLOCK) &&
3101 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
3107 if (S_ISREG(mode)) {
3108 if (IS_ENABLED(CONFIG_BLOCK) &&
3109 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
3110 file->f_op != &io_uring_fops)
3115 /* any ->read/write should understand O_NONBLOCK */
3116 if (file->f_flags & O_NONBLOCK)
3118 return file->f_mode & FMODE_NOWAIT;
3122 * If we tracked the file through the SCM inflight mechanism, we could support
3123 * any file. For now, just ensure that anything potentially problematic is done
3126 static unsigned int io_file_get_flags(struct file *file)
3128 umode_t mode = file_inode(file)->i_mode;
3129 unsigned int res = 0;
3133 if (__io_file_supports_nowait(file, mode))
3138 static inline bool io_file_supports_nowait(struct io_kiocb *req)
3140 return req->flags & REQ_F_SUPPORT_NOWAIT;
3143 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3145 struct kiocb *kiocb = &req->rw.kiocb;
3149 kiocb->ki_pos = READ_ONCE(sqe->off);
3151 ioprio = READ_ONCE(sqe->ioprio);
3153 ret = ioprio_check_cap(ioprio);
3157 kiocb->ki_ioprio = ioprio;
3159 kiocb->ki_ioprio = get_current_ioprio();
3163 req->rw.addr = READ_ONCE(sqe->addr);
3164 req->rw.len = READ_ONCE(sqe->len);
3165 req->rw.flags = READ_ONCE(sqe->rw_flags);
3166 req->buf_index = READ_ONCE(sqe->buf_index);
3170 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
3176 case -ERESTARTNOINTR:
3177 case -ERESTARTNOHAND:
3178 case -ERESTART_RESTARTBLOCK:
3180 * We can't just restart the syscall, since previously
3181 * submitted sqes may already be in progress. Just fail this
3187 kiocb->ki_complete(kiocb, ret);
3191 static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req)
3193 struct kiocb *kiocb = &req->rw.kiocb;
3195 if (kiocb->ki_pos != -1)
3196 return &kiocb->ki_pos;
3198 if (!(req->file->f_mode & FMODE_STREAM)) {
3199 req->flags |= REQ_F_CUR_POS;
3200 kiocb->ki_pos = req->file->f_pos;
3201 return &kiocb->ki_pos;
3208 static void kiocb_done(struct io_kiocb *req, ssize_t ret,
3209 unsigned int issue_flags)
3211 struct io_async_rw *io = req->async_data;
3213 /* add previously done IO, if any */
3214 if (req_has_async_data(req) && io->bytes_done > 0) {
3216 ret = io->bytes_done;
3218 ret += io->bytes_done;
3221 if (req->flags & REQ_F_CUR_POS)
3222 req->file->f_pos = req->rw.kiocb.ki_pos;
3223 if (ret >= 0 && (req->rw.kiocb.ki_complete == io_complete_rw))
3224 __io_complete_rw(req, ret, issue_flags);
3226 io_rw_done(&req->rw.kiocb, ret);
3228 if (req->flags & REQ_F_REISSUE) {
3229 req->flags &= ~REQ_F_REISSUE;
3230 if (io_resubmit_prep(req))
3231 io_req_task_queue_reissue(req);
3233 io_req_task_queue_fail(req, ret);
3237 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
3238 struct io_mapped_ubuf *imu)
3240 size_t len = req->rw.len;
3241 u64 buf_end, buf_addr = req->rw.addr;
3244 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
3246 /* not inside the mapped region */
3247 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
3251 * May not be a start of buffer, set size appropriately
3252 * and advance us to the beginning.
3254 offset = buf_addr - imu->ubuf;
3255 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3259 * Don't use iov_iter_advance() here, as it's really slow for
3260 * using the latter parts of a big fixed buffer - it iterates
3261 * over each segment manually. We can cheat a bit here, because
3264 * 1) it's a BVEC iter, we set it up
3265 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3266 * first and last bvec
3268 * So just find our index, and adjust the iterator afterwards.
3269 * If the offset is within the first bvec (or the whole first
3270 * bvec, just use iov_iter_advance(). This makes it easier
3271 * since we can just skip the first segment, which may not
3272 * be PAGE_SIZE aligned.
3274 const struct bio_vec *bvec = imu->bvec;
3276 if (offset <= bvec->bv_len) {
3277 iov_iter_advance(iter, offset);
3279 unsigned long seg_skip;
3281 /* skip first vec */
3282 offset -= bvec->bv_len;
3283 seg_skip = 1 + (offset >> PAGE_SHIFT);
3285 iter->bvec = bvec + seg_skip;
3286 iter->nr_segs -= seg_skip;
3287 iter->count -= bvec->bv_len + offset;
3288 iter->iov_offset = offset & ~PAGE_MASK;
3295 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
3296 unsigned int issue_flags)
3298 struct io_mapped_ubuf *imu = req->imu;
3299 u16 index, buf_index = req->buf_index;
3302 struct io_ring_ctx *ctx = req->ctx;
3304 if (unlikely(buf_index >= ctx->nr_user_bufs))
3306 io_req_set_rsrc_node(req, ctx, issue_flags);
3307 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3308 imu = READ_ONCE(ctx->user_bufs[index]);
3311 return __io_import_fixed(req, rw, iter, imu);
3314 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3317 mutex_unlock(&ctx->uring_lock);
3320 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3323 * "Normal" inline submissions always hold the uring_lock, since we
3324 * grab it from the system call. Same is true for the SQPOLL offload.
3325 * The only exception is when we've detached the request and issue it
3326 * from an async worker thread, grab the lock for that case.
3329 mutex_lock(&ctx->uring_lock);
3332 static void io_buffer_add_list(struct io_ring_ctx *ctx,
3333 struct io_buffer_list *bl, unsigned int bgid)
3335 struct list_head *list;
3337 list = &ctx->io_buffers[hash_32(bgid, IO_BUFFERS_HASH_BITS)];
3338 INIT_LIST_HEAD(&bl->buf_list);
3340 list_add(&bl->list, list);
3343 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3344 int bgid, unsigned int issue_flags)
3346 struct io_buffer *kbuf = req->kbuf;
3347 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
3348 struct io_ring_ctx *ctx = req->ctx;
3349 struct io_buffer_list *bl;
3351 if (req->flags & REQ_F_BUFFER_SELECTED)
3354 io_ring_submit_lock(ctx, needs_lock);
3356 lockdep_assert_held(&ctx->uring_lock);
3358 bl = io_buffer_get_list(ctx, bgid);
3359 if (bl && !list_empty(&bl->buf_list)) {
3360 kbuf = list_first_entry(&bl->buf_list, struct io_buffer, list);
3361 list_del(&kbuf->list);
3362 if (*len > kbuf->len)
3364 req->flags |= REQ_F_BUFFER_SELECTED;
3367 kbuf = ERR_PTR(-ENOBUFS);
3370 io_ring_submit_unlock(req->ctx, needs_lock);
3374 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3375 unsigned int issue_flags)
3377 struct io_buffer *kbuf;
3380 bgid = req->buf_index;
3381 kbuf = io_buffer_select(req, len, bgid, issue_flags);
3384 return u64_to_user_ptr(kbuf->addr);
3387 #ifdef CONFIG_COMPAT
3388 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3389 unsigned int issue_flags)
3391 struct compat_iovec __user *uiov;
3392 compat_ssize_t clen;
3396 uiov = u64_to_user_ptr(req->rw.addr);
3397 if (!access_ok(uiov, sizeof(*uiov)))
3399 if (__get_user(clen, &uiov->iov_len))
3405 buf = io_rw_buffer_select(req, &len, issue_flags);
3407 return PTR_ERR(buf);
3408 iov[0].iov_base = buf;
3409 iov[0].iov_len = (compat_size_t) len;
3414 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3415 unsigned int issue_flags)
3417 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3421 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3424 len = iov[0].iov_len;
3427 buf = io_rw_buffer_select(req, &len, issue_flags);
3429 return PTR_ERR(buf);
3430 iov[0].iov_base = buf;
3431 iov[0].iov_len = len;
3435 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3436 unsigned int issue_flags)
3438 if (req->flags & REQ_F_BUFFER_SELECTED) {
3439 struct io_buffer *kbuf = req->kbuf;
3441 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3442 iov[0].iov_len = kbuf->len;
3445 if (req->rw.len != 1)
3448 #ifdef CONFIG_COMPAT
3449 if (req->ctx->compat)
3450 return io_compat_import(req, iov, issue_flags);
3453 return __io_iov_buffer_select(req, iov, issue_flags);
3456 static struct iovec *__io_import_iovec(int rw, struct io_kiocb *req,
3457 struct io_rw_state *s,
3458 unsigned int issue_flags)
3460 struct iov_iter *iter = &s->iter;
3461 u8 opcode = req->opcode;
3462 struct iovec *iovec;
3467 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3468 ret = io_import_fixed(req, rw, iter, issue_flags);
3470 return ERR_PTR(ret);
3474 /* buffer index only valid with fixed read/write, or buffer select */
3475 if (unlikely(req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT)))
3476 return ERR_PTR(-EINVAL);
3478 buf = u64_to_user_ptr(req->rw.addr);
3479 sqe_len = req->rw.len;
3481 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3482 if (req->flags & REQ_F_BUFFER_SELECT) {
3483 buf = io_rw_buffer_select(req, &sqe_len, issue_flags);
3485 return ERR_CAST(buf);
3486 req->rw.len = sqe_len;
3489 ret = import_single_range(rw, buf, sqe_len, s->fast_iov, iter);
3491 return ERR_PTR(ret);
3495 iovec = s->fast_iov;
3496 if (req->flags & REQ_F_BUFFER_SELECT) {
3497 ret = io_iov_buffer_select(req, iovec, issue_flags);
3499 return ERR_PTR(ret);
3500 iov_iter_init(iter, rw, iovec, 1, iovec->iov_len);
3504 ret = __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, &iovec, iter,
3506 if (unlikely(ret < 0))
3507 return ERR_PTR(ret);
3511 static inline int io_import_iovec(int rw, struct io_kiocb *req,
3512 struct iovec **iovec, struct io_rw_state *s,
3513 unsigned int issue_flags)
3515 *iovec = __io_import_iovec(rw, req, s, issue_flags);
3516 if (unlikely(IS_ERR(*iovec)))
3517 return PTR_ERR(*iovec);
3519 iov_iter_save_state(&s->iter, &s->iter_state);
3523 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3525 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3529 * For files that don't have ->read_iter() and ->write_iter(), handle them
3530 * by looping over ->read() or ->write() manually.
3532 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3534 struct kiocb *kiocb = &req->rw.kiocb;
3535 struct file *file = req->file;
3540 * Don't support polled IO through this interface, and we can't
3541 * support non-blocking either. For the latter, this just causes
3542 * the kiocb to be handled from an async context.
3544 if (kiocb->ki_flags & IOCB_HIPRI)
3546 if ((kiocb->ki_flags & IOCB_NOWAIT) &&
3547 !(kiocb->ki_filp->f_flags & O_NONBLOCK))
3550 ppos = io_kiocb_ppos(kiocb);
3552 while (iov_iter_count(iter)) {
3556 if (!iov_iter_is_bvec(iter)) {
3557 iovec = iov_iter_iovec(iter);
3559 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3560 iovec.iov_len = req->rw.len;
3564 nr = file->f_op->read(file, iovec.iov_base,
3565 iovec.iov_len, ppos);
3567 nr = file->f_op->write(file, iovec.iov_base,
3568 iovec.iov_len, ppos);
3577 if (!iov_iter_is_bvec(iter)) {
3578 iov_iter_advance(iter, nr);
3585 if (nr != iovec.iov_len)
3592 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3593 const struct iovec *fast_iov, struct iov_iter *iter)
3595 struct io_async_rw *rw = req->async_data;
3597 memcpy(&rw->s.iter, iter, sizeof(*iter));
3598 rw->free_iovec = iovec;
3600 /* can only be fixed buffers, no need to do anything */
3601 if (iov_iter_is_bvec(iter))
3604 unsigned iov_off = 0;
3606 rw->s.iter.iov = rw->s.fast_iov;
3607 if (iter->iov != fast_iov) {
3608 iov_off = iter->iov - fast_iov;
3609 rw->s.iter.iov += iov_off;
3611 if (rw->s.fast_iov != fast_iov)
3612 memcpy(rw->s.fast_iov + iov_off, fast_iov + iov_off,
3613 sizeof(struct iovec) * iter->nr_segs);
3615 req->flags |= REQ_F_NEED_CLEANUP;
3619 static inline bool io_alloc_async_data(struct io_kiocb *req)
3621 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3622 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3623 if (req->async_data) {
3624 req->flags |= REQ_F_ASYNC_DATA;
3630 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3631 struct io_rw_state *s, bool force)
3633 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3635 if (!req_has_async_data(req)) {
3636 struct io_async_rw *iorw;
3638 if (io_alloc_async_data(req)) {
3643 io_req_map_rw(req, iovec, s->fast_iov, &s->iter);
3644 iorw = req->async_data;
3645 /* we've copied and mapped the iter, ensure state is saved */
3646 iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state);
3651 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3653 struct io_async_rw *iorw = req->async_data;
3657 /* submission path, ->uring_lock should already be taken */
3658 ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
3659 if (unlikely(ret < 0))
3662 iorw->bytes_done = 0;
3663 iorw->free_iovec = iov;
3665 req->flags |= REQ_F_NEED_CLEANUP;
3670 * This is our waitqueue callback handler, registered through __folio_lock_async()
3671 * when we initially tried to do the IO with the iocb armed our waitqueue.
3672 * This gets called when the page is unlocked, and we generally expect that to
3673 * happen when the page IO is completed and the page is now uptodate. This will
3674 * queue a task_work based retry of the operation, attempting to copy the data
3675 * again. If the latter fails because the page was NOT uptodate, then we will
3676 * do a thread based blocking retry of the operation. That's the unexpected
3679 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3680 int sync, void *arg)
3682 struct wait_page_queue *wpq;
3683 struct io_kiocb *req = wait->private;
3684 struct wait_page_key *key = arg;
3686 wpq = container_of(wait, struct wait_page_queue, wait);
3688 if (!wake_page_match(wpq, key))
3691 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3692 list_del_init(&wait->entry);
3693 io_req_task_queue(req);
3698 * This controls whether a given IO request should be armed for async page
3699 * based retry. If we return false here, the request is handed to the async
3700 * worker threads for retry. If we're doing buffered reads on a regular file,
3701 * we prepare a private wait_page_queue entry and retry the operation. This
3702 * will either succeed because the page is now uptodate and unlocked, or it
3703 * will register a callback when the page is unlocked at IO completion. Through
3704 * that callback, io_uring uses task_work to setup a retry of the operation.
3705 * That retry will attempt the buffered read again. The retry will generally
3706 * succeed, or in rare cases where it fails, we then fall back to using the
3707 * async worker threads for a blocking retry.
3709 static bool io_rw_should_retry(struct io_kiocb *req)
3711 struct io_async_rw *rw = req->async_data;
3712 struct wait_page_queue *wait = &rw->wpq;
3713 struct kiocb *kiocb = &req->rw.kiocb;
3715 /* never retry for NOWAIT, we just complete with -EAGAIN */
3716 if (req->flags & REQ_F_NOWAIT)
3719 /* Only for buffered IO */
3720 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3724 * just use poll if we can, and don't attempt if the fs doesn't
3725 * support callback based unlocks
3727 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3730 wait->wait.func = io_async_buf_func;
3731 wait->wait.private = req;
3732 wait->wait.flags = 0;
3733 INIT_LIST_HEAD(&wait->wait.entry);
3734 kiocb->ki_flags |= IOCB_WAITQ;
3735 kiocb->ki_flags &= ~IOCB_NOWAIT;
3736 kiocb->ki_waitq = wait;
3740 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3742 if (likely(req->file->f_op->read_iter))
3743 return call_read_iter(req->file, &req->rw.kiocb, iter);
3744 else if (req->file->f_op->read)
3745 return loop_rw_iter(READ, req, iter);
3750 static bool need_read_all(struct io_kiocb *req)
3752 return req->flags & REQ_F_ISREG ||
3753 S_ISBLK(file_inode(req->file)->i_mode);
3756 static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
3758 struct kiocb *kiocb = &req->rw.kiocb;
3759 struct io_ring_ctx *ctx = req->ctx;
3760 struct file *file = req->file;
3763 if (unlikely(!file || !(file->f_mode & mode)))
3766 if (!io_req_ffs_set(req))
3767 req->flags |= io_file_get_flags(file) << REQ_F_SUPPORT_NOWAIT_BIT;
3769 kiocb->ki_flags = iocb_flags(file);
3770 ret = kiocb_set_rw_flags(kiocb, req->rw.flags);
3775 * If the file is marked O_NONBLOCK, still allow retry for it if it
3776 * supports async. Otherwise it's impossible to use O_NONBLOCK files
3777 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
3779 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
3780 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
3781 req->flags |= REQ_F_NOWAIT;
3783 if (ctx->flags & IORING_SETUP_IOPOLL) {
3784 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
3787 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
3788 kiocb->ki_complete = io_complete_rw_iopoll;
3789 req->iopoll_completed = 0;
3791 if (kiocb->ki_flags & IOCB_HIPRI)
3793 kiocb->ki_complete = io_complete_rw;
3799 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3801 struct io_rw_state __s, *s = &__s;
3802 struct iovec *iovec;
3803 struct kiocb *kiocb = &req->rw.kiocb;
3804 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3805 struct io_async_rw *rw;
3809 if (!req_has_async_data(req)) {
3810 ret = io_import_iovec(READ, req, &iovec, s, issue_flags);
3811 if (unlikely(ret < 0))
3815 * Safe and required to re-import if we're using provided
3816 * buffers, as we dropped the selected one before retry.
3818 if (req->flags & REQ_F_BUFFER_SELECT) {
3819 ret = io_import_iovec(READ, req, &iovec, s, issue_flags);
3820 if (unlikely(ret < 0))
3824 rw = req->async_data;
3827 * We come here from an earlier attempt, restore our state to
3828 * match in case it doesn't. It's cheap enough that we don't
3829 * need to make this conditional.
3831 iov_iter_restore(&s->iter, &s->iter_state);
3834 ret = io_rw_init_file(req, FMODE_READ);
3837 req->result = iov_iter_count(&s->iter);
3839 if (force_nonblock) {
3840 /* If the file doesn't support async, just async punt */
3841 if (unlikely(!io_file_supports_nowait(req))) {
3842 ret = io_setup_async_rw(req, iovec, s, true);
3843 return ret ?: -EAGAIN;
3845 kiocb->ki_flags |= IOCB_NOWAIT;
3847 /* Ensure we clear previously set non-block flag */
3848 kiocb->ki_flags &= ~IOCB_NOWAIT;
3851 ppos = io_kiocb_update_pos(req);
3853 ret = rw_verify_area(READ, req->file, ppos, req->result);
3854 if (unlikely(ret)) {
3859 ret = io_iter_do_read(req, &s->iter);
3861 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3862 req->flags &= ~REQ_F_REISSUE;
3863 /* if we can poll, just do that */
3864 if (req->opcode == IORING_OP_READ && file_can_poll(req->file))
3866 /* IOPOLL retry should happen for io-wq threads */
3867 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3869 /* no retry on NONBLOCK nor RWF_NOWAIT */
3870 if (req->flags & REQ_F_NOWAIT)
3873 } else if (ret == -EIOCBQUEUED) {
3875 } else if (ret == req->result || ret <= 0 || !force_nonblock ||
3876 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3877 /* read all, failed, already did sync or don't want to retry */
3882 * Don't depend on the iter state matching what was consumed, or being
3883 * untouched in case of error. Restore it and we'll advance it
3884 * manually if we need to.
3886 iov_iter_restore(&s->iter, &s->iter_state);
3888 ret2 = io_setup_async_rw(req, iovec, s, true);
3893 rw = req->async_data;
3896 * Now use our persistent iterator and state, if we aren't already.
3897 * We've restored and mapped the iter to match.
3902 * We end up here because of a partial read, either from
3903 * above or inside this loop. Advance the iter by the bytes
3904 * that were consumed.
3906 iov_iter_advance(&s->iter, ret);
3907 if (!iov_iter_count(&s->iter))
3909 rw->bytes_done += ret;
3910 iov_iter_save_state(&s->iter, &s->iter_state);
3912 /* if we can retry, do so with the callbacks armed */
3913 if (!io_rw_should_retry(req)) {
3914 kiocb->ki_flags &= ~IOCB_WAITQ;
3919 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3920 * we get -EIOCBQUEUED, then we'll get a notification when the
3921 * desired page gets unlocked. We can also get a partial read
3922 * here, and if we do, then just retry at the new offset.
3924 ret = io_iter_do_read(req, &s->iter);
3925 if (ret == -EIOCBQUEUED)
3927 /* we got some bytes, but not all. retry. */
3928 kiocb->ki_flags &= ~IOCB_WAITQ;
3929 iov_iter_restore(&s->iter, &s->iter_state);
3932 kiocb_done(req, ret, issue_flags);
3934 /* it's faster to check here then delegate to kfree */
3940 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3942 struct io_rw_state __s, *s = &__s;
3943 struct iovec *iovec;
3944 struct kiocb *kiocb = &req->rw.kiocb;
3945 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3949 if (!req_has_async_data(req)) {
3950 ret = io_import_iovec(WRITE, req, &iovec, s, issue_flags);
3951 if (unlikely(ret < 0))
3954 struct io_async_rw *rw = req->async_data;
3957 iov_iter_restore(&s->iter, &s->iter_state);
3960 ret = io_rw_init_file(req, FMODE_WRITE);
3963 req->result = iov_iter_count(&s->iter);
3965 if (force_nonblock) {
3966 /* If the file doesn't support async, just async punt */
3967 if (unlikely(!io_file_supports_nowait(req)))
3970 /* file path doesn't support NOWAIT for non-direct_IO */
3971 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3972 (req->flags & REQ_F_ISREG))
3975 kiocb->ki_flags |= IOCB_NOWAIT;
3977 /* Ensure we clear previously set non-block flag */
3978 kiocb->ki_flags &= ~IOCB_NOWAIT;
3981 ppos = io_kiocb_update_pos(req);
3983 ret = rw_verify_area(WRITE, req->file, ppos, req->result);
3988 * Open-code file_start_write here to grab freeze protection,
3989 * which will be released by another thread in
3990 * io_complete_rw(). Fool lockdep by telling it the lock got
3991 * released so that it doesn't complain about the held lock when
3992 * we return to userspace.
3994 if (req->flags & REQ_F_ISREG) {
3995 sb_start_write(file_inode(req->file)->i_sb);
3996 __sb_writers_release(file_inode(req->file)->i_sb,
3999 kiocb->ki_flags |= IOCB_WRITE;
4001 if (likely(req->file->f_op->write_iter))
4002 ret2 = call_write_iter(req->file, kiocb, &s->iter);
4003 else if (req->file->f_op->write)
4004 ret2 = loop_rw_iter(WRITE, req, &s->iter);
4008 if (req->flags & REQ_F_REISSUE) {
4009 req->flags &= ~REQ_F_REISSUE;
4014 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
4015 * retry them without IOCB_NOWAIT.
4017 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
4019 /* no retry on NONBLOCK nor RWF_NOWAIT */
4020 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
4022 if (!force_nonblock || ret2 != -EAGAIN) {
4023 /* IOPOLL retry should happen for io-wq threads */
4024 if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
4027 kiocb_done(req, ret2, issue_flags);
4030 iov_iter_restore(&s->iter, &s->iter_state);
4031 ret = io_setup_async_rw(req, iovec, s, false);
4032 return ret ?: -EAGAIN;
4035 /* it's reportedly faster than delegating the null check to kfree() */
4041 static int io_renameat_prep(struct io_kiocb *req,
4042 const struct io_uring_sqe *sqe)
4044 struct io_rename *ren = &req->rename;
4045 const char __user *oldf, *newf;
4047 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4049 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4051 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4054 ren->old_dfd = READ_ONCE(sqe->fd);
4055 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
4056 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4057 ren->new_dfd = READ_ONCE(sqe->len);
4058 ren->flags = READ_ONCE(sqe->rename_flags);
4060 ren->oldpath = getname(oldf);
4061 if (IS_ERR(ren->oldpath))
4062 return PTR_ERR(ren->oldpath);
4064 ren->newpath = getname(newf);
4065 if (IS_ERR(ren->newpath)) {
4066 putname(ren->oldpath);
4067 return PTR_ERR(ren->newpath);
4070 req->flags |= REQ_F_NEED_CLEANUP;
4074 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
4076 struct io_rename *ren = &req->rename;
4079 if (issue_flags & IO_URING_F_NONBLOCK)
4082 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
4083 ren->newpath, ren->flags);
4085 req->flags &= ~REQ_F_NEED_CLEANUP;
4088 io_req_complete(req, ret);
4092 static int io_unlinkat_prep(struct io_kiocb *req,
4093 const struct io_uring_sqe *sqe)
4095 struct io_unlink *un = &req->unlink;
4096 const char __user *fname;
4098 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4100 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4103 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4106 un->dfd = READ_ONCE(sqe->fd);
4108 un->flags = READ_ONCE(sqe->unlink_flags);
4109 if (un->flags & ~AT_REMOVEDIR)
4112 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4113 un->filename = getname(fname);
4114 if (IS_ERR(un->filename))
4115 return PTR_ERR(un->filename);
4117 req->flags |= REQ_F_NEED_CLEANUP;
4121 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
4123 struct io_unlink *un = &req->unlink;
4126 if (issue_flags & IO_URING_F_NONBLOCK)
4129 if (un->flags & AT_REMOVEDIR)
4130 ret = do_rmdir(un->dfd, un->filename);
4132 ret = do_unlinkat(un->dfd, un->filename);
4134 req->flags &= ~REQ_F_NEED_CLEANUP;
4137 io_req_complete(req, ret);
4141 static int io_mkdirat_prep(struct io_kiocb *req,
4142 const struct io_uring_sqe *sqe)
4144 struct io_mkdir *mkd = &req->mkdir;
4145 const char __user *fname;
4147 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4149 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
4152 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4155 mkd->dfd = READ_ONCE(sqe->fd);
4156 mkd->mode = READ_ONCE(sqe->len);
4158 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4159 mkd->filename = getname(fname);
4160 if (IS_ERR(mkd->filename))
4161 return PTR_ERR(mkd->filename);
4163 req->flags |= REQ_F_NEED_CLEANUP;
4167 static int io_mkdirat(struct io_kiocb *req, unsigned int issue_flags)
4169 struct io_mkdir *mkd = &req->mkdir;
4172 if (issue_flags & IO_URING_F_NONBLOCK)
4175 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
4177 req->flags &= ~REQ_F_NEED_CLEANUP;
4180 io_req_complete(req, ret);
4184 static int io_symlinkat_prep(struct io_kiocb *req,
4185 const struct io_uring_sqe *sqe)
4187 struct io_symlink *sl = &req->symlink;
4188 const char __user *oldpath, *newpath;
4190 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4192 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
4195 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4198 sl->new_dfd = READ_ONCE(sqe->fd);
4199 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
4200 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4202 sl->oldpath = getname(oldpath);
4203 if (IS_ERR(sl->oldpath))
4204 return PTR_ERR(sl->oldpath);
4206 sl->newpath = getname(newpath);
4207 if (IS_ERR(sl->newpath)) {
4208 putname(sl->oldpath);
4209 return PTR_ERR(sl->newpath);
4212 req->flags |= REQ_F_NEED_CLEANUP;
4216 static int io_symlinkat(struct io_kiocb *req, unsigned int issue_flags)
4218 struct io_symlink *sl = &req->symlink;
4221 if (issue_flags & IO_URING_F_NONBLOCK)
4224 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
4226 req->flags &= ~REQ_F_NEED_CLEANUP;
4229 io_req_complete(req, ret);
4233 static int io_linkat_prep(struct io_kiocb *req,
4234 const struct io_uring_sqe *sqe)
4236 struct io_hardlink *lnk = &req->hardlink;
4237 const char __user *oldf, *newf;
4239 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4241 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
4243 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4246 lnk->old_dfd = READ_ONCE(sqe->fd);
4247 lnk->new_dfd = READ_ONCE(sqe->len);
4248 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
4249 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4250 lnk->flags = READ_ONCE(sqe->hardlink_flags);
4252 lnk->oldpath = getname(oldf);
4253 if (IS_ERR(lnk->oldpath))
4254 return PTR_ERR(lnk->oldpath);
4256 lnk->newpath = getname(newf);
4257 if (IS_ERR(lnk->newpath)) {
4258 putname(lnk->oldpath);
4259 return PTR_ERR(lnk->newpath);
4262 req->flags |= REQ_F_NEED_CLEANUP;
4266 static int io_linkat(struct io_kiocb *req, unsigned int issue_flags)
4268 struct io_hardlink *lnk = &req->hardlink;
4271 if (issue_flags & IO_URING_F_NONBLOCK)
4274 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
4275 lnk->newpath, lnk->flags);
4277 req->flags &= ~REQ_F_NEED_CLEANUP;
4280 io_req_complete(req, ret);
4284 static int io_shutdown_prep(struct io_kiocb *req,
4285 const struct io_uring_sqe *sqe)
4287 #if defined(CONFIG_NET)
4288 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4290 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
4291 sqe->buf_index || sqe->splice_fd_in))
4294 req->shutdown.how = READ_ONCE(sqe->len);
4301 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
4303 #if defined(CONFIG_NET)
4304 struct socket *sock;
4307 if (issue_flags & IO_URING_F_NONBLOCK)
4310 sock = sock_from_file(req->file);
4311 if (unlikely(!sock))
4314 ret = __sys_shutdown_sock(sock, req->shutdown.how);
4317 io_req_complete(req, ret);
4324 static int __io_splice_prep(struct io_kiocb *req,
4325 const struct io_uring_sqe *sqe)
4327 struct io_splice *sp = &req->splice;
4328 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
4330 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4333 sp->len = READ_ONCE(sqe->len);
4334 sp->flags = READ_ONCE(sqe->splice_flags);
4335 if (unlikely(sp->flags & ~valid_flags))
4337 sp->splice_fd_in = READ_ONCE(sqe->splice_fd_in);
4341 static int io_tee_prep(struct io_kiocb *req,
4342 const struct io_uring_sqe *sqe)
4344 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
4346 return __io_splice_prep(req, sqe);
4349 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
4351 struct io_splice *sp = &req->splice;
4352 struct file *out = sp->file_out;
4353 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4357 if (issue_flags & IO_URING_F_NONBLOCK)
4360 if (sp->flags & SPLICE_F_FD_IN_FIXED)
4361 in = io_file_get_fixed(req, sp->splice_fd_in, issue_flags);
4363 in = io_file_get_normal(req, sp->splice_fd_in);
4370 ret = do_tee(in, out, sp->len, flags);
4372 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4377 io_req_complete(req, ret);
4381 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4383 struct io_splice *sp = &req->splice;
4385 sp->off_in = READ_ONCE(sqe->splice_off_in);
4386 sp->off_out = READ_ONCE(sqe->off);
4387 return __io_splice_prep(req, sqe);
4390 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4392 struct io_splice *sp = &req->splice;
4393 struct file *out = sp->file_out;
4394 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4395 loff_t *poff_in, *poff_out;
4399 if (issue_flags & IO_URING_F_NONBLOCK)
4402 if (sp->flags & SPLICE_F_FD_IN_FIXED)
4403 in = io_file_get_fixed(req, sp->splice_fd_in, issue_flags);
4405 in = io_file_get_normal(req, sp->splice_fd_in);
4411 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4412 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4415 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4417 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4422 io_req_complete(req, ret);
4427 * IORING_OP_NOP just posts a completion event, nothing else.
4429 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4431 struct io_ring_ctx *ctx = req->ctx;
4433 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4436 __io_req_complete(req, issue_flags, 0, 0);
4440 static int io_msg_ring_prep(struct io_kiocb *req,
4441 const struct io_uring_sqe *sqe)
4443 if (unlikely(sqe->addr || sqe->ioprio || sqe->rw_flags ||
4444 sqe->splice_fd_in || sqe->buf_index || sqe->personality))
4447 req->msg.user_data = READ_ONCE(sqe->off);
4448 req->msg.len = READ_ONCE(sqe->len);
4452 static int io_msg_ring(struct io_kiocb *req, unsigned int issue_flags)
4454 struct io_ring_ctx *target_ctx;
4455 struct io_msg *msg = &req->msg;
4460 if (req->file->f_op != &io_uring_fops)
4464 target_ctx = req->file->private_data;
4466 spin_lock(&target_ctx->completion_lock);
4467 filled = io_fill_cqe_aux(target_ctx, msg->user_data, msg->len, 0);
4468 io_commit_cqring(target_ctx);
4469 spin_unlock(&target_ctx->completion_lock);
4472 io_cqring_ev_posted(target_ctx);
4479 __io_req_complete(req, issue_flags, ret, 0);
4483 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4485 struct io_ring_ctx *ctx = req->ctx;
4487 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4489 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4493 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4494 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4497 req->sync.off = READ_ONCE(sqe->off);
4498 req->sync.len = READ_ONCE(sqe->len);
4502 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4504 loff_t end = req->sync.off + req->sync.len;
4507 /* fsync always requires a blocking context */
4508 if (issue_flags & IO_URING_F_NONBLOCK)
4511 ret = vfs_fsync_range(req->file, req->sync.off,
4512 end > 0 ? end : LLONG_MAX,
4513 req->sync.flags & IORING_FSYNC_DATASYNC);
4516 io_req_complete(req, ret);
4520 static int io_fallocate_prep(struct io_kiocb *req,
4521 const struct io_uring_sqe *sqe)
4523 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4526 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4529 req->sync.off = READ_ONCE(sqe->off);
4530 req->sync.len = READ_ONCE(sqe->addr);
4531 req->sync.mode = READ_ONCE(sqe->len);
4535 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4539 /* fallocate always requiring blocking context */
4540 if (issue_flags & IO_URING_F_NONBLOCK)
4542 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4547 fsnotify_modify(req->file);
4548 io_req_complete(req, ret);
4552 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4554 const char __user *fname;
4557 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4559 if (unlikely(sqe->ioprio || sqe->buf_index))
4561 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4564 /* open.how should be already initialised */
4565 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4566 req->open.how.flags |= O_LARGEFILE;
4568 req->open.dfd = READ_ONCE(sqe->fd);
4569 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4570 req->open.filename = getname(fname);
4571 if (IS_ERR(req->open.filename)) {
4572 ret = PTR_ERR(req->open.filename);
4573 req->open.filename = NULL;
4577 req->open.file_slot = READ_ONCE(sqe->file_index);
4578 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4581 req->open.nofile = rlimit(RLIMIT_NOFILE);
4582 req->flags |= REQ_F_NEED_CLEANUP;
4586 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4588 u64 mode = READ_ONCE(sqe->len);
4589 u64 flags = READ_ONCE(sqe->open_flags);
4591 req->open.how = build_open_how(flags, mode);
4592 return __io_openat_prep(req, sqe);
4595 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4597 struct open_how __user *how;
4601 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4602 len = READ_ONCE(sqe->len);
4603 if (len < OPEN_HOW_SIZE_VER0)
4606 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4611 return __io_openat_prep(req, sqe);
4614 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4616 struct open_flags op;
4618 bool resolve_nonblock, nonblock_set;
4619 bool fixed = !!req->open.file_slot;
4622 ret = build_open_flags(&req->open.how, &op);
4625 nonblock_set = op.open_flag & O_NONBLOCK;
4626 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4627 if (issue_flags & IO_URING_F_NONBLOCK) {
4629 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4630 * it'll always -EAGAIN
4632 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4634 op.lookup_flags |= LOOKUP_CACHED;
4635 op.open_flag |= O_NONBLOCK;
4639 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4644 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4647 * We could hang on to this 'fd' on retrying, but seems like
4648 * marginal gain for something that is now known to be a slower
4649 * path. So just put it, and we'll get a new one when we retry.
4654 ret = PTR_ERR(file);
4655 /* only retry if RESOLVE_CACHED wasn't already set by application */
4656 if (ret == -EAGAIN &&
4657 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4662 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4663 file->f_flags &= ~O_NONBLOCK;
4664 fsnotify_open(file);
4667 fd_install(ret, file);
4669 ret = io_install_fixed_file(req, file, issue_flags,
4670 req->open.file_slot - 1);
4672 putname(req->open.filename);
4673 req->flags &= ~REQ_F_NEED_CLEANUP;
4676 __io_req_complete(req, issue_flags, ret, 0);
4680 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4682 return io_openat2(req, issue_flags);
4685 static int io_remove_buffers_prep(struct io_kiocb *req,
4686 const struct io_uring_sqe *sqe)
4688 struct io_provide_buf *p = &req->pbuf;
4691 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4695 tmp = READ_ONCE(sqe->fd);
4696 if (!tmp || tmp > USHRT_MAX)
4699 memset(p, 0, sizeof(*p));
4701 p->bgid = READ_ONCE(sqe->buf_group);
4705 static int __io_remove_buffers(struct io_ring_ctx *ctx,
4706 struct io_buffer_list *bl, unsigned nbufs)
4710 /* shouldn't happen */
4714 /* the head kbuf is the list itself */
4715 while (!list_empty(&bl->buf_list)) {
4716 struct io_buffer *nxt;
4718 nxt = list_first_entry(&bl->buf_list, struct io_buffer, list);
4719 list_del(&nxt->list);
4729 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4731 struct io_provide_buf *p = &req->pbuf;
4732 struct io_ring_ctx *ctx = req->ctx;
4733 struct io_buffer_list *bl;
4735 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
4737 io_ring_submit_lock(ctx, needs_lock);
4739 lockdep_assert_held(&ctx->uring_lock);
4742 bl = io_buffer_get_list(ctx, p->bgid);
4744 ret = __io_remove_buffers(ctx, bl, p->nbufs);
4748 /* complete before unlock, IOPOLL may need the lock */
4749 __io_req_complete(req, issue_flags, ret, 0);
4750 io_ring_submit_unlock(ctx, needs_lock);
4754 static int io_provide_buffers_prep(struct io_kiocb *req,
4755 const struct io_uring_sqe *sqe)
4757 unsigned long size, tmp_check;
4758 struct io_provide_buf *p = &req->pbuf;
4761 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4764 tmp = READ_ONCE(sqe->fd);
4765 if (!tmp || tmp > USHRT_MAX)
4768 p->addr = READ_ONCE(sqe->addr);
4769 p->len = READ_ONCE(sqe->len);
4771 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4774 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4777 size = (unsigned long)p->len * p->nbufs;
4778 if (!access_ok(u64_to_user_ptr(p->addr), size))
4781 p->bgid = READ_ONCE(sqe->buf_group);
4782 tmp = READ_ONCE(sqe->off);
4783 if (tmp > USHRT_MAX)
4789 static int io_refill_buffer_cache(struct io_ring_ctx *ctx)
4791 struct io_buffer *buf;
4796 * Completions that don't happen inline (eg not under uring_lock) will
4797 * add to ->io_buffers_comp. If we don't have any free buffers, check
4798 * the completion list and splice those entries first.
4800 if (!list_empty_careful(&ctx->io_buffers_comp)) {
4801 spin_lock(&ctx->completion_lock);
4802 if (!list_empty(&ctx->io_buffers_comp)) {
4803 list_splice_init(&ctx->io_buffers_comp,
4804 &ctx->io_buffers_cache);
4805 spin_unlock(&ctx->completion_lock);
4808 spin_unlock(&ctx->completion_lock);
4812 * No free buffers and no completion entries either. Allocate a new
4813 * page worth of buffer entries and add those to our freelist.
4815 page = alloc_page(GFP_KERNEL_ACCOUNT);
4819 list_add(&page->lru, &ctx->io_buffers_pages);
4821 buf = page_address(page);
4822 bufs_in_page = PAGE_SIZE / sizeof(*buf);
4823 while (bufs_in_page) {
4824 list_add_tail(&buf->list, &ctx->io_buffers_cache);
4832 static int io_add_buffers(struct io_ring_ctx *ctx, struct io_provide_buf *pbuf,
4833 struct io_buffer_list *bl)
4835 struct io_buffer *buf;
4836 u64 addr = pbuf->addr;
4837 int i, bid = pbuf->bid;
4839 for (i = 0; i < pbuf->nbufs; i++) {
4840 if (list_empty(&ctx->io_buffers_cache) &&
4841 io_refill_buffer_cache(ctx))
4843 buf = list_first_entry(&ctx->io_buffers_cache, struct io_buffer,
4845 list_move_tail(&buf->list, &bl->buf_list);
4847 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4849 buf->bgid = pbuf->bgid;
4855 return i ? 0 : -ENOMEM;
4858 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4860 struct io_provide_buf *p = &req->pbuf;
4861 struct io_ring_ctx *ctx = req->ctx;
4862 struct io_buffer_list *bl;
4864 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
4866 io_ring_submit_lock(ctx, needs_lock);
4868 lockdep_assert_held(&ctx->uring_lock);
4870 bl = io_buffer_get_list(ctx, p->bgid);
4871 if (unlikely(!bl)) {
4872 bl = kmalloc(sizeof(*bl), GFP_KERNEL);
4877 io_buffer_add_list(ctx, bl, p->bgid);
4880 ret = io_add_buffers(ctx, p, bl);
4884 /* complete before unlock, IOPOLL may need the lock */
4885 __io_req_complete(req, issue_flags, ret, 0);
4886 io_ring_submit_unlock(ctx, needs_lock);
4890 static int io_epoll_ctl_prep(struct io_kiocb *req,
4891 const struct io_uring_sqe *sqe)
4893 #if defined(CONFIG_EPOLL)
4894 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4896 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4899 req->epoll.epfd = READ_ONCE(sqe->fd);
4900 req->epoll.op = READ_ONCE(sqe->len);
4901 req->epoll.fd = READ_ONCE(sqe->off);
4903 if (ep_op_has_event(req->epoll.op)) {
4904 struct epoll_event __user *ev;
4906 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4907 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4917 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4919 #if defined(CONFIG_EPOLL)
4920 struct io_epoll *ie = &req->epoll;
4922 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4924 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4925 if (force_nonblock && ret == -EAGAIN)
4930 __io_req_complete(req, issue_flags, ret, 0);
4937 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4939 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4940 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4942 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4945 req->madvise.addr = READ_ONCE(sqe->addr);
4946 req->madvise.len = READ_ONCE(sqe->len);
4947 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4954 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4956 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4957 struct io_madvise *ma = &req->madvise;
4960 if (issue_flags & IO_URING_F_NONBLOCK)
4963 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4966 io_req_complete(req, ret);
4973 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4975 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4977 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4980 req->fadvise.offset = READ_ONCE(sqe->off);
4981 req->fadvise.len = READ_ONCE(sqe->len);
4982 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4986 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4988 struct io_fadvise *fa = &req->fadvise;
4991 if (issue_flags & IO_URING_F_NONBLOCK) {
4992 switch (fa->advice) {
4993 case POSIX_FADV_NORMAL:
4994 case POSIX_FADV_RANDOM:
4995 case POSIX_FADV_SEQUENTIAL:
5002 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
5005 __io_req_complete(req, issue_flags, ret, 0);
5009 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5011 const char __user *path;
5013 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5015 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5017 if (req->flags & REQ_F_FIXED_FILE)
5020 req->statx.dfd = READ_ONCE(sqe->fd);
5021 req->statx.mask = READ_ONCE(sqe->len);
5022 path = u64_to_user_ptr(READ_ONCE(sqe->addr));
5023 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5024 req->statx.flags = READ_ONCE(sqe->statx_flags);
5026 req->statx.filename = getname_flags(path,
5027 getname_statx_lookup_flags(req->statx.flags),
5030 if (IS_ERR(req->statx.filename)) {
5031 int ret = PTR_ERR(req->statx.filename);
5033 req->statx.filename = NULL;
5037 req->flags |= REQ_F_NEED_CLEANUP;
5041 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
5043 struct io_statx *ctx = &req->statx;
5046 if (issue_flags & IO_URING_F_NONBLOCK)
5049 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
5054 io_req_complete(req, ret);
5058 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5060 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5062 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
5063 sqe->rw_flags || sqe->buf_index)
5065 if (req->flags & REQ_F_FIXED_FILE)
5068 req->close.fd = READ_ONCE(sqe->fd);
5069 req->close.file_slot = READ_ONCE(sqe->file_index);
5070 if (req->close.file_slot && req->close.fd)
5076 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
5078 struct files_struct *files = current->files;
5079 struct io_close *close = &req->close;
5080 struct fdtable *fdt;
5081 struct file *file = NULL;
5084 if (req->close.file_slot) {
5085 ret = io_close_fixed(req, issue_flags);
5089 spin_lock(&files->file_lock);
5090 fdt = files_fdtable(files);
5091 if (close->fd >= fdt->max_fds) {
5092 spin_unlock(&files->file_lock);
5095 file = fdt->fd[close->fd];
5096 if (!file || file->f_op == &io_uring_fops) {
5097 spin_unlock(&files->file_lock);
5102 /* if the file has a flush method, be safe and punt to async */
5103 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
5104 spin_unlock(&files->file_lock);
5108 ret = __close_fd_get_file(close->fd, &file);
5109 spin_unlock(&files->file_lock);
5116 /* No ->flush() or already async, safely close from here */
5117 ret = filp_close(file, current->files);
5123 __io_req_complete(req, issue_flags, ret, 0);
5127 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5129 struct io_ring_ctx *ctx = req->ctx;
5131 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
5133 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
5137 req->sync.off = READ_ONCE(sqe->off);
5138 req->sync.len = READ_ONCE(sqe->len);
5139 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
5143 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
5147 /* sync_file_range always requires a blocking context */
5148 if (issue_flags & IO_URING_F_NONBLOCK)
5151 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
5155 io_req_complete(req, ret);
5159 #if defined(CONFIG_NET)
5160 static int io_setup_async_msg(struct io_kiocb *req,
5161 struct io_async_msghdr *kmsg)
5163 struct io_async_msghdr *async_msg = req->async_data;
5167 if (io_alloc_async_data(req)) {
5168 kfree(kmsg->free_iov);
5171 async_msg = req->async_data;
5172 req->flags |= REQ_F_NEED_CLEANUP;
5173 memcpy(async_msg, kmsg, sizeof(*kmsg));
5174 async_msg->msg.msg_name = &async_msg->addr;
5175 /* if were using fast_iov, set it to the new one */
5176 if (!async_msg->free_iov)
5177 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
5182 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
5183 struct io_async_msghdr *iomsg)
5185 iomsg->msg.msg_name = &iomsg->addr;
5186 iomsg->free_iov = iomsg->fast_iov;
5187 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
5188 req->sr_msg.msg_flags, &iomsg->free_iov);
5191 static int io_sendmsg_prep_async(struct io_kiocb *req)
5195 ret = io_sendmsg_copy_hdr(req, req->async_data);
5197 req->flags |= REQ_F_NEED_CLEANUP;
5201 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5203 struct io_sr_msg *sr = &req->sr_msg;
5205 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5208 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
5209 sr->len = READ_ONCE(sqe->len);
5210 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
5211 if (sr->msg_flags & MSG_DONTWAIT)
5212 req->flags |= REQ_F_NOWAIT;
5214 #ifdef CONFIG_COMPAT
5215 if (req->ctx->compat)
5216 sr->msg_flags |= MSG_CMSG_COMPAT;
5221 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
5223 struct io_async_msghdr iomsg, *kmsg;
5224 struct socket *sock;
5229 sock = sock_from_file(req->file);
5230 if (unlikely(!sock))
5233 if (req_has_async_data(req)) {
5234 kmsg = req->async_data;
5236 ret = io_sendmsg_copy_hdr(req, &iomsg);
5242 flags = req->sr_msg.msg_flags;
5243 if (issue_flags & IO_URING_F_NONBLOCK)
5244 flags |= MSG_DONTWAIT;
5245 if (flags & MSG_WAITALL)
5246 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5248 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
5250 if (ret < min_ret) {
5251 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
5252 return io_setup_async_msg(req, kmsg);
5253 if (ret == -ERESTARTSYS)
5257 /* fast path, check for non-NULL to avoid function call */
5259 kfree(kmsg->free_iov);
5260 req->flags &= ~REQ_F_NEED_CLEANUP;
5261 __io_req_complete(req, issue_flags, ret, 0);
5265 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
5267 struct io_sr_msg *sr = &req->sr_msg;
5270 struct socket *sock;
5275 sock = sock_from_file(req->file);
5276 if (unlikely(!sock))
5279 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
5283 msg.msg_name = NULL;
5284 msg.msg_control = NULL;
5285 msg.msg_controllen = 0;
5286 msg.msg_namelen = 0;
5288 flags = req->sr_msg.msg_flags;
5289 if (issue_flags & IO_URING_F_NONBLOCK)
5290 flags |= MSG_DONTWAIT;
5291 if (flags & MSG_WAITALL)
5292 min_ret = iov_iter_count(&msg.msg_iter);
5294 msg.msg_flags = flags;
5295 ret = sock_sendmsg(sock, &msg);
5296 if (ret < min_ret) {
5297 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
5299 if (ret == -ERESTARTSYS)
5303 __io_req_complete(req, issue_flags, ret, 0);
5307 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
5308 struct io_async_msghdr *iomsg)
5310 struct io_sr_msg *sr = &req->sr_msg;
5311 struct iovec __user *uiov;
5315 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
5316 &iomsg->uaddr, &uiov, &iov_len);
5320 if (req->flags & REQ_F_BUFFER_SELECT) {
5323 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
5325 sr->len = iomsg->fast_iov[0].iov_len;
5326 iomsg->free_iov = NULL;
5328 iomsg->free_iov = iomsg->fast_iov;
5329 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
5330 &iomsg->free_iov, &iomsg->msg.msg_iter,
5339 #ifdef CONFIG_COMPAT
5340 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
5341 struct io_async_msghdr *iomsg)
5343 struct io_sr_msg *sr = &req->sr_msg;
5344 struct compat_iovec __user *uiov;
5349 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
5354 uiov = compat_ptr(ptr);
5355 if (req->flags & REQ_F_BUFFER_SELECT) {
5356 compat_ssize_t clen;
5360 if (!access_ok(uiov, sizeof(*uiov)))
5362 if (__get_user(clen, &uiov->iov_len))
5367 iomsg->free_iov = NULL;
5369 iomsg->free_iov = iomsg->fast_iov;
5370 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
5371 UIO_FASTIOV, &iomsg->free_iov,
5372 &iomsg->msg.msg_iter, true);
5381 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
5382 struct io_async_msghdr *iomsg)
5384 iomsg->msg.msg_name = &iomsg->addr;
5386 #ifdef CONFIG_COMPAT
5387 if (req->ctx->compat)
5388 return __io_compat_recvmsg_copy_hdr(req, iomsg);
5391 return __io_recvmsg_copy_hdr(req, iomsg);
5394 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
5395 unsigned int issue_flags)
5397 struct io_sr_msg *sr = &req->sr_msg;
5399 return io_buffer_select(req, &sr->len, sr->bgid, issue_flags);
5402 static int io_recvmsg_prep_async(struct io_kiocb *req)
5406 ret = io_recvmsg_copy_hdr(req, req->async_data);
5408 req->flags |= REQ_F_NEED_CLEANUP;
5412 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5414 struct io_sr_msg *sr = &req->sr_msg;
5416 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5419 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
5420 sr->len = READ_ONCE(sqe->len);
5421 sr->bgid = READ_ONCE(sqe->buf_group);
5422 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
5423 if (sr->msg_flags & MSG_DONTWAIT)
5424 req->flags |= REQ_F_NOWAIT;
5426 #ifdef CONFIG_COMPAT
5427 if (req->ctx->compat)
5428 sr->msg_flags |= MSG_CMSG_COMPAT;
5434 static bool io_net_retry(struct socket *sock, int flags)
5436 if (!(flags & MSG_WAITALL))
5438 return sock->type == SOCK_STREAM || sock->type == SOCK_SEQPACKET;
5441 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
5443 struct io_async_msghdr iomsg, *kmsg;
5444 struct io_sr_msg *sr = &req->sr_msg;
5445 struct socket *sock;
5446 struct io_buffer *kbuf;
5448 int ret, min_ret = 0;
5449 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5451 sock = sock_from_file(req->file);
5452 if (unlikely(!sock))
5455 if (req_has_async_data(req)) {
5456 kmsg = req->async_data;
5458 ret = io_recvmsg_copy_hdr(req, &iomsg);
5464 if (req->flags & REQ_F_BUFFER_SELECT) {
5465 kbuf = io_recv_buffer_select(req, issue_flags);
5467 return PTR_ERR(kbuf);
5468 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5469 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5470 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5471 1, req->sr_msg.len);
5474 flags = req->sr_msg.msg_flags;
5476 flags |= MSG_DONTWAIT;
5477 if (flags & MSG_WAITALL)
5478 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5480 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5481 kmsg->uaddr, flags);
5482 if (ret < min_ret) {
5483 if (ret == -EAGAIN && force_nonblock)
5484 return io_setup_async_msg(req, kmsg);
5485 if (ret == -ERESTARTSYS)
5487 if (ret > 0 && io_net_retry(sock, flags)) {
5489 req->flags |= REQ_F_PARTIAL_IO;
5490 return io_setup_async_msg(req, kmsg);
5493 } else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
5497 /* fast path, check for non-NULL to avoid function call */
5499 kfree(kmsg->free_iov);
5500 req->flags &= ~REQ_F_NEED_CLEANUP;
5503 else if (sr->done_io)
5505 __io_req_complete(req, issue_flags, ret, io_put_kbuf(req, issue_flags));
5509 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5511 struct io_buffer *kbuf;
5512 struct io_sr_msg *sr = &req->sr_msg;
5514 void __user *buf = sr->buf;
5515 struct socket *sock;
5518 int ret, min_ret = 0;
5519 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5521 sock = sock_from_file(req->file);
5522 if (unlikely(!sock))
5525 if (req->flags & REQ_F_BUFFER_SELECT) {
5526 kbuf = io_recv_buffer_select(req, issue_flags);
5528 return PTR_ERR(kbuf);
5529 buf = u64_to_user_ptr(kbuf->addr);
5532 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5536 msg.msg_name = NULL;
5537 msg.msg_control = NULL;
5538 msg.msg_controllen = 0;
5539 msg.msg_namelen = 0;
5540 msg.msg_iocb = NULL;
5543 flags = req->sr_msg.msg_flags;
5545 flags |= MSG_DONTWAIT;
5546 if (flags & MSG_WAITALL)
5547 min_ret = iov_iter_count(&msg.msg_iter);
5549 ret = sock_recvmsg(sock, &msg, flags);
5550 if (ret < min_ret) {
5551 if (ret == -EAGAIN && force_nonblock)
5553 if (ret == -ERESTARTSYS)
5555 if (ret > 0 && io_net_retry(sock, flags)) {
5559 req->flags |= REQ_F_PARTIAL_IO;
5563 } else if ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
5570 else if (sr->done_io)
5572 __io_req_complete(req, issue_flags, ret, io_put_kbuf(req, issue_flags));
5576 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5578 struct io_accept *accept = &req->accept;
5580 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5582 if (sqe->ioprio || sqe->len || sqe->buf_index)
5585 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5586 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5587 accept->flags = READ_ONCE(sqe->accept_flags);
5588 accept->nofile = rlimit(RLIMIT_NOFILE);
5590 accept->file_slot = READ_ONCE(sqe->file_index);
5591 if (accept->file_slot && (accept->flags & SOCK_CLOEXEC))
5593 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5595 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5596 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5600 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5602 struct io_accept *accept = &req->accept;
5603 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5604 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5605 bool fixed = !!accept->file_slot;
5610 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5611 if (unlikely(fd < 0))
5614 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5619 ret = PTR_ERR(file);
5620 if (ret == -EAGAIN && force_nonblock)
5622 if (ret == -ERESTARTSYS)
5625 } else if (!fixed) {
5626 fd_install(fd, file);
5629 ret = io_install_fixed_file(req, file, issue_flags,
5630 accept->file_slot - 1);
5632 __io_req_complete(req, issue_flags, ret, 0);
5636 static int io_connect_prep_async(struct io_kiocb *req)
5638 struct io_async_connect *io = req->async_data;
5639 struct io_connect *conn = &req->connect;
5641 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5644 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5646 struct io_connect *conn = &req->connect;
5648 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5650 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5654 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5655 conn->addr_len = READ_ONCE(sqe->addr2);
5659 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5661 struct io_async_connect __io, *io;
5662 unsigned file_flags;
5664 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5666 if (req_has_async_data(req)) {
5667 io = req->async_data;
5669 ret = move_addr_to_kernel(req->connect.addr,
5670 req->connect.addr_len,
5677 file_flags = force_nonblock ? O_NONBLOCK : 0;
5679 ret = __sys_connect_file(req->file, &io->address,
5680 req->connect.addr_len, file_flags);
5681 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5682 if (req_has_async_data(req))
5684 if (io_alloc_async_data(req)) {
5688 memcpy(req->async_data, &__io, sizeof(__io));
5691 if (ret == -ERESTARTSYS)
5696 __io_req_complete(req, issue_flags, ret, 0);
5699 #else /* !CONFIG_NET */
5700 #define IO_NETOP_FN(op) \
5701 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5703 return -EOPNOTSUPP; \
5706 #define IO_NETOP_PREP(op) \
5708 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5710 return -EOPNOTSUPP; \
5713 #define IO_NETOP_PREP_ASYNC(op) \
5715 static int io_##op##_prep_async(struct io_kiocb *req) \
5717 return -EOPNOTSUPP; \
5720 IO_NETOP_PREP_ASYNC(sendmsg);
5721 IO_NETOP_PREP_ASYNC(recvmsg);
5722 IO_NETOP_PREP_ASYNC(connect);
5723 IO_NETOP_PREP(accept);
5726 #endif /* CONFIG_NET */
5728 struct io_poll_table {
5729 struct poll_table_struct pt;
5730 struct io_kiocb *req;
5735 #define IO_POLL_CANCEL_FLAG BIT(31)
5736 #define IO_POLL_REF_MASK GENMASK(30, 0)
5739 * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can
5740 * bump it and acquire ownership. It's disallowed to modify requests while not
5741 * owning it, that prevents from races for enqueueing task_work's and b/w
5742 * arming poll and wakeups.
5744 static inline bool io_poll_get_ownership(struct io_kiocb *req)
5746 return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK);
5749 static void io_poll_mark_cancelled(struct io_kiocb *req)
5751 atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs);
5754 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5756 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5757 if (req->opcode == IORING_OP_POLL_ADD)
5758 return req->async_data;
5759 return req->apoll->double_poll;
5762 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5764 if (req->opcode == IORING_OP_POLL_ADD)
5766 return &req->apoll->poll;
5769 static void io_poll_req_insert(struct io_kiocb *req)
5771 struct io_ring_ctx *ctx = req->ctx;
5772 struct hlist_head *list;
5774 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5775 hlist_add_head(&req->hash_node, list);
5778 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5779 wait_queue_func_t wake_func)
5782 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5783 /* mask in events that we always want/need */
5784 poll->events = events | IO_POLL_UNMASK;
5785 INIT_LIST_HEAD(&poll->wait.entry);
5786 init_waitqueue_func_entry(&poll->wait, wake_func);
5789 static inline void io_poll_remove_entry(struct io_poll_iocb *poll)
5791 struct wait_queue_head *head = smp_load_acquire(&poll->head);
5794 spin_lock_irq(&head->lock);
5795 list_del_init(&poll->wait.entry);
5797 spin_unlock_irq(&head->lock);
5801 static void io_poll_remove_entries(struct io_kiocb *req)
5804 * Nothing to do if neither of those flags are set. Avoid dipping
5805 * into the poll/apoll/double cachelines if we can.
5807 if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL)))
5811 * While we hold the waitqueue lock and the waitqueue is nonempty,
5812 * wake_up_pollfree() will wait for us. However, taking the waitqueue
5813 * lock in the first place can race with the waitqueue being freed.
5815 * We solve this as eventpoll does: by taking advantage of the fact that
5816 * all users of wake_up_pollfree() will RCU-delay the actual free. If
5817 * we enter rcu_read_lock() and see that the pointer to the queue is
5818 * non-NULL, we can then lock it without the memory being freed out from
5821 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
5822 * case the caller deletes the entry from the queue, leaving it empty.
5823 * In that case, only RCU prevents the queue memory from being freed.
5826 if (req->flags & REQ_F_SINGLE_POLL)
5827 io_poll_remove_entry(io_poll_get_single(req));
5828 if (req->flags & REQ_F_DOUBLE_POLL)
5829 io_poll_remove_entry(io_poll_get_double(req));
5834 * All poll tw should go through this. Checks for poll events, manages
5835 * references, does rewait, etc.
5837 * Returns a negative error on failure. >0 when no action require, which is
5838 * either spurious wakeup or multishot CQE is served. 0 when it's done with
5839 * the request, then the mask is stored in req->result.
5841 static int io_poll_check_events(struct io_kiocb *req, bool locked)
5843 struct io_ring_ctx *ctx = req->ctx;
5846 /* req->task == current here, checking PF_EXITING is safe */
5847 if (unlikely(req->task->flags & PF_EXITING))
5848 io_poll_mark_cancelled(req);
5851 v = atomic_read(&req->poll_refs);
5853 /* tw handler should be the owner, and so have some references */
5854 if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK)))
5856 if (v & IO_POLL_CANCEL_FLAG)
5860 struct poll_table_struct pt = { ._key = req->apoll_events };
5861 unsigned flags = locked ? 0 : IO_URING_F_UNLOCKED;
5863 if (unlikely(!io_assign_file(req, flags)))
5865 req->result = vfs_poll(req->file, &pt) & req->apoll_events;
5868 /* multishot, just fill an CQE and proceed */
5869 if (req->result && !(req->apoll_events & EPOLLONESHOT)) {
5870 __poll_t mask = mangle_poll(req->result & req->apoll_events);
5873 spin_lock(&ctx->completion_lock);
5874 filled = io_fill_cqe_aux(ctx, req->user_data, mask,
5876 io_commit_cqring(ctx);
5877 spin_unlock(&ctx->completion_lock);
5878 if (unlikely(!filled))
5880 io_cqring_ev_posted(ctx);
5881 } else if (req->result) {
5886 * Release all references, retry if someone tried to restart
5887 * task_work while we were executing it.
5889 } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs));
5894 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5896 struct io_ring_ctx *ctx = req->ctx;
5899 ret = io_poll_check_events(req, *locked);
5904 req->result = mangle_poll(req->result & req->poll.events);
5910 io_poll_remove_entries(req);
5911 spin_lock(&ctx->completion_lock);
5912 hash_del(&req->hash_node);
5913 __io_req_complete_post(req, req->result, 0);
5914 io_commit_cqring(ctx);
5915 spin_unlock(&ctx->completion_lock);
5916 io_cqring_ev_posted(ctx);
5919 static void io_apoll_task_func(struct io_kiocb *req, bool *locked)
5921 struct io_ring_ctx *ctx = req->ctx;
5924 ret = io_poll_check_events(req, *locked);
5928 io_poll_remove_entries(req);
5929 spin_lock(&ctx->completion_lock);
5930 hash_del(&req->hash_node);
5931 spin_unlock(&ctx->completion_lock);
5934 io_req_task_submit(req, locked);
5936 io_req_complete_failed(req, ret);
5939 static void __io_poll_execute(struct io_kiocb *req, int mask, int events)
5943 * This is useful for poll that is armed on behalf of another
5944 * request, and where the wakeup path could be on a different
5945 * CPU. We want to avoid pulling in req->apoll->events for that
5948 req->apoll_events = events;
5949 if (req->opcode == IORING_OP_POLL_ADD)
5950 req->io_task_work.func = io_poll_task_func;
5952 req->io_task_work.func = io_apoll_task_func;
5954 trace_io_uring_task_add(req->ctx, req, req->user_data, req->opcode, mask);
5955 io_req_task_work_add(req, false);
5958 static inline void io_poll_execute(struct io_kiocb *req, int res, int events)
5960 if (io_poll_get_ownership(req))
5961 __io_poll_execute(req, res, events);
5964 static void io_poll_cancel_req(struct io_kiocb *req)
5966 io_poll_mark_cancelled(req);
5967 /* kick tw, which should complete the request */
5968 io_poll_execute(req, 0, 0);
5971 #define wqe_to_req(wait) ((void *)((unsigned long) (wait)->private & ~1))
5972 #define wqe_is_double(wait) ((unsigned long) (wait)->private & 1)
5974 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5977 struct io_kiocb *req = wqe_to_req(wait);
5978 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
5980 __poll_t mask = key_to_poll(key);
5982 if (unlikely(mask & POLLFREE)) {
5983 io_poll_mark_cancelled(req);
5984 /* we have to kick tw in case it's not already */
5985 io_poll_execute(req, 0, poll->events);
5988 * If the waitqueue is being freed early but someone is already
5989 * holds ownership over it, we have to tear down the request as
5990 * best we can. That means immediately removing the request from
5991 * its waitqueue and preventing all further accesses to the
5992 * waitqueue via the request.
5994 list_del_init(&poll->wait.entry);
5997 * Careful: this *must* be the last step, since as soon
5998 * as req->head is NULL'ed out, the request can be
5999 * completed and freed, since aio_poll_complete_work()
6000 * will no longer need to take the waitqueue lock.
6002 smp_store_release(&poll->head, NULL);
6006 /* for instances that support it check for an event match first */
6007 if (mask && !(mask & poll->events))
6010 if (io_poll_get_ownership(req)) {
6011 /* optional, saves extra locking for removal in tw handler */
6012 if (mask && poll->events & EPOLLONESHOT) {
6013 list_del_init(&poll->wait.entry);
6015 if (wqe_is_double(wait))
6016 req->flags &= ~REQ_F_DOUBLE_POLL;
6018 req->flags &= ~REQ_F_SINGLE_POLL;
6020 __io_poll_execute(req, mask, poll->events);
6025 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
6026 struct wait_queue_head *head,
6027 struct io_poll_iocb **poll_ptr)
6029 struct io_kiocb *req = pt->req;
6030 unsigned long wqe_private = (unsigned long) req;
6033 * The file being polled uses multiple waitqueues for poll handling
6034 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
6037 if (unlikely(pt->nr_entries)) {
6038 struct io_poll_iocb *first = poll;
6040 /* double add on the same waitqueue head, ignore */
6041 if (first->head == head)
6043 /* already have a 2nd entry, fail a third attempt */
6045 if ((*poll_ptr)->head == head)
6047 pt->error = -EINVAL;
6051 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
6053 pt->error = -ENOMEM;
6056 /* mark as double wq entry */
6058 req->flags |= REQ_F_DOUBLE_POLL;
6059 io_init_poll_iocb(poll, first->events, first->wait.func);
6061 if (req->opcode == IORING_OP_POLL_ADD)
6062 req->flags |= REQ_F_ASYNC_DATA;
6065 req->flags |= REQ_F_SINGLE_POLL;
6068 poll->wait.private = (void *) wqe_private;
6070 if (poll->events & EPOLLEXCLUSIVE)
6071 add_wait_queue_exclusive(head, &poll->wait);
6073 add_wait_queue(head, &poll->wait);
6076 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
6077 struct poll_table_struct *p)
6079 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
6081 __io_queue_proc(&pt->req->poll, pt, head,
6082 (struct io_poll_iocb **) &pt->req->async_data);
6085 static int __io_arm_poll_handler(struct io_kiocb *req,
6086 struct io_poll_iocb *poll,
6087 struct io_poll_table *ipt, __poll_t mask)
6089 struct io_ring_ctx *ctx = req->ctx;
6092 INIT_HLIST_NODE(&req->hash_node);
6093 io_init_poll_iocb(poll, mask, io_poll_wake);
6094 poll->file = req->file;
6096 ipt->pt._key = mask;
6099 ipt->nr_entries = 0;
6102 * Take the ownership to delay any tw execution up until we're done
6103 * with poll arming. see io_poll_get_ownership().
6105 atomic_set(&req->poll_refs, 1);
6106 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
6108 if (mask && (poll->events & EPOLLONESHOT)) {
6109 io_poll_remove_entries(req);
6110 /* no one else has access to the req, forget about the ref */
6113 if (!mask && unlikely(ipt->error || !ipt->nr_entries)) {
6114 io_poll_remove_entries(req);
6116 ipt->error = -EINVAL;
6120 spin_lock(&ctx->completion_lock);
6121 io_poll_req_insert(req);
6122 spin_unlock(&ctx->completion_lock);
6125 /* can't multishot if failed, just queue the event we've got */
6126 if (unlikely(ipt->error || !ipt->nr_entries))
6127 poll->events |= EPOLLONESHOT;
6128 __io_poll_execute(req, mask, poll->events);
6133 * Release ownership. If someone tried to queue a tw while it was
6134 * locked, kick it off for them.
6136 v = atomic_dec_return(&req->poll_refs);
6137 if (unlikely(v & IO_POLL_REF_MASK))
6138 __io_poll_execute(req, 0, poll->events);
6142 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
6143 struct poll_table_struct *p)
6145 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
6146 struct async_poll *apoll = pt->req->apoll;
6148 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
6157 static int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
6159 const struct io_op_def *def = &io_op_defs[req->opcode];
6160 struct io_ring_ctx *ctx = req->ctx;
6161 struct async_poll *apoll;
6162 struct io_poll_table ipt;
6163 __poll_t mask = EPOLLONESHOT | POLLERR | POLLPRI;
6166 if (!def->pollin && !def->pollout)
6167 return IO_APOLL_ABORTED;
6168 if (!file_can_poll(req->file) || (req->flags & REQ_F_POLLED))
6169 return IO_APOLL_ABORTED;
6172 mask |= POLLIN | POLLRDNORM;
6174 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
6175 if ((req->opcode == IORING_OP_RECVMSG) &&
6176 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
6179 mask |= POLLOUT | POLLWRNORM;
6181 if (def->poll_exclusive)
6182 mask |= EPOLLEXCLUSIVE;
6183 if (!(issue_flags & IO_URING_F_UNLOCKED) &&
6184 !list_empty(&ctx->apoll_cache)) {
6185 apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
6187 list_del_init(&apoll->poll.wait.entry);
6189 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
6190 if (unlikely(!apoll))
6191 return IO_APOLL_ABORTED;
6193 apoll->double_poll = NULL;
6195 req->flags |= REQ_F_POLLED;
6196 ipt.pt._qproc = io_async_queue_proc;
6198 io_kbuf_recycle(req, issue_flags);
6200 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask);
6201 if (ret || ipt.error)
6202 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
6204 trace_io_uring_poll_arm(ctx, req, req->user_data, req->opcode,
6205 mask, apoll->poll.events);
6210 * Returns true if we found and killed one or more poll requests
6212 static __cold bool io_poll_remove_all(struct io_ring_ctx *ctx,
6213 struct task_struct *tsk, bool cancel_all)
6215 struct hlist_node *tmp;
6216 struct io_kiocb *req;
6220 spin_lock(&ctx->completion_lock);
6221 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
6222 struct hlist_head *list;
6224 list = &ctx->cancel_hash[i];
6225 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
6226 if (io_match_task_safe(req, tsk, cancel_all)) {
6227 hlist_del_init(&req->hash_node);
6228 io_poll_cancel_req(req);
6233 spin_unlock(&ctx->completion_lock);
6237 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
6239 __must_hold(&ctx->completion_lock)
6241 struct hlist_head *list;
6242 struct io_kiocb *req;
6244 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
6245 hlist_for_each_entry(req, list, hash_node) {
6246 if (sqe_addr != req->user_data)
6248 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
6255 static bool io_poll_disarm(struct io_kiocb *req)
6256 __must_hold(&ctx->completion_lock)
6258 if (!io_poll_get_ownership(req))
6260 io_poll_remove_entries(req);
6261 hash_del(&req->hash_node);
6265 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
6267 __must_hold(&ctx->completion_lock)
6269 struct io_kiocb *req = io_poll_find(ctx, sqe_addr, poll_only);
6273 io_poll_cancel_req(req);
6277 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
6282 events = READ_ONCE(sqe->poll32_events);
6284 events = swahw32(events);
6286 if (!(flags & IORING_POLL_ADD_MULTI))
6287 events |= EPOLLONESHOT;
6288 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
6291 static int io_poll_update_prep(struct io_kiocb *req,
6292 const struct io_uring_sqe *sqe)
6294 struct io_poll_update *upd = &req->poll_update;
6297 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6299 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
6301 flags = READ_ONCE(sqe->len);
6302 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
6303 IORING_POLL_ADD_MULTI))
6305 /* meaningless without update */
6306 if (flags == IORING_POLL_ADD_MULTI)
6309 upd->old_user_data = READ_ONCE(sqe->addr);
6310 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
6311 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
6313 upd->new_user_data = READ_ONCE(sqe->off);
6314 if (!upd->update_user_data && upd->new_user_data)
6316 if (upd->update_events)
6317 upd->events = io_poll_parse_events(sqe, flags);
6318 else if (sqe->poll32_events)
6324 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6326 struct io_poll_iocb *poll = &req->poll;
6329 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6331 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
6333 flags = READ_ONCE(sqe->len);
6334 if (flags & ~IORING_POLL_ADD_MULTI)
6336 if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP))
6339 io_req_set_refcount(req);
6340 req->apoll_events = poll->events = io_poll_parse_events(sqe, flags);
6344 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
6346 struct io_poll_iocb *poll = &req->poll;
6347 struct io_poll_table ipt;
6350 ipt.pt._qproc = io_poll_queue_proc;
6352 ret = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events);
6353 ret = ret ?: ipt.error;
6355 __io_req_complete(req, issue_flags, ret, 0);
6359 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
6361 struct io_ring_ctx *ctx = req->ctx;
6362 struct io_kiocb *preq;
6366 spin_lock(&ctx->completion_lock);
6367 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
6368 if (!preq || !io_poll_disarm(preq)) {
6369 spin_unlock(&ctx->completion_lock);
6370 ret = preq ? -EALREADY : -ENOENT;
6373 spin_unlock(&ctx->completion_lock);
6375 if (req->poll_update.update_events || req->poll_update.update_user_data) {
6376 /* only mask one event flags, keep behavior flags */
6377 if (req->poll_update.update_events) {
6378 preq->poll.events &= ~0xffff;
6379 preq->poll.events |= req->poll_update.events & 0xffff;
6380 preq->poll.events |= IO_POLL_UNMASK;
6382 if (req->poll_update.update_user_data)
6383 preq->user_data = req->poll_update.new_user_data;
6385 ret2 = io_poll_add(preq, issue_flags);
6386 /* successfully updated, don't complete poll request */
6392 preq->result = -ECANCELED;
6393 locked = !(issue_flags & IO_URING_F_UNLOCKED);
6394 io_req_task_complete(preq, &locked);
6398 /* complete update request, we're done with it */
6399 __io_req_complete(req, issue_flags, ret, 0);
6403 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
6405 struct io_timeout_data *data = container_of(timer,
6406 struct io_timeout_data, timer);
6407 struct io_kiocb *req = data->req;
6408 struct io_ring_ctx *ctx = req->ctx;
6409 unsigned long flags;
6411 spin_lock_irqsave(&ctx->timeout_lock, flags);
6412 list_del_init(&req->timeout.list);
6413 atomic_set(&req->ctx->cq_timeouts,
6414 atomic_read(&req->ctx->cq_timeouts) + 1);
6415 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6417 if (!(data->flags & IORING_TIMEOUT_ETIME_SUCCESS))
6420 req->result = -ETIME;
6421 req->io_task_work.func = io_req_task_complete;
6422 io_req_task_work_add(req, false);
6423 return HRTIMER_NORESTART;
6426 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
6428 __must_hold(&ctx->timeout_lock)
6430 struct io_timeout_data *io;
6431 struct io_kiocb *req;
6434 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
6435 found = user_data == req->user_data;
6440 return ERR_PTR(-ENOENT);
6442 io = req->async_data;
6443 if (hrtimer_try_to_cancel(&io->timer) == -1)
6444 return ERR_PTR(-EALREADY);
6445 list_del_init(&req->timeout.list);
6449 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
6450 __must_hold(&ctx->completion_lock)
6451 __must_hold(&ctx->timeout_lock)
6453 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6456 return PTR_ERR(req);
6457 io_req_task_queue_fail(req, -ECANCELED);
6461 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
6463 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
6464 case IORING_TIMEOUT_BOOTTIME:
6465 return CLOCK_BOOTTIME;
6466 case IORING_TIMEOUT_REALTIME:
6467 return CLOCK_REALTIME;
6469 /* can't happen, vetted at prep time */
6473 return CLOCK_MONOTONIC;
6477 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6478 struct timespec64 *ts, enum hrtimer_mode mode)
6479 __must_hold(&ctx->timeout_lock)
6481 struct io_timeout_data *io;
6482 struct io_kiocb *req;
6485 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6486 found = user_data == req->user_data;
6493 io = req->async_data;
6494 if (hrtimer_try_to_cancel(&io->timer) == -1)
6496 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6497 io->timer.function = io_link_timeout_fn;
6498 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6502 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6503 struct timespec64 *ts, enum hrtimer_mode mode)
6504 __must_hold(&ctx->timeout_lock)
6506 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6507 struct io_timeout_data *data;
6510 return PTR_ERR(req);
6512 req->timeout.off = 0; /* noseq */
6513 data = req->async_data;
6514 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6515 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6516 data->timer.function = io_timeout_fn;
6517 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6521 static int io_timeout_remove_prep(struct io_kiocb *req,
6522 const struct io_uring_sqe *sqe)
6524 struct io_timeout_rem *tr = &req->timeout_rem;
6526 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6528 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6530 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6533 tr->ltimeout = false;
6534 tr->addr = READ_ONCE(sqe->addr);
6535 tr->flags = READ_ONCE(sqe->timeout_flags);
6536 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6537 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6539 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6540 tr->ltimeout = true;
6541 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6543 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6545 if (tr->ts.tv_sec < 0 || tr->ts.tv_nsec < 0)
6547 } else if (tr->flags) {
6548 /* timeout removal doesn't support flags */
6555 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6557 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6562 * Remove or update an existing timeout command
6564 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6566 struct io_timeout_rem *tr = &req->timeout_rem;
6567 struct io_ring_ctx *ctx = req->ctx;
6570 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6571 spin_lock(&ctx->completion_lock);
6572 spin_lock_irq(&ctx->timeout_lock);
6573 ret = io_timeout_cancel(ctx, tr->addr);
6574 spin_unlock_irq(&ctx->timeout_lock);
6575 spin_unlock(&ctx->completion_lock);
6577 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6579 spin_lock_irq(&ctx->timeout_lock);
6581 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6583 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6584 spin_unlock_irq(&ctx->timeout_lock);
6589 io_req_complete_post(req, ret, 0);
6593 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6594 bool is_timeout_link)
6596 struct io_timeout_data *data;
6598 u32 off = READ_ONCE(sqe->off);
6600 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6602 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6605 if (off && is_timeout_link)
6607 flags = READ_ONCE(sqe->timeout_flags);
6608 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK |
6609 IORING_TIMEOUT_ETIME_SUCCESS))
6611 /* more than one clock specified is invalid, obviously */
6612 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6615 INIT_LIST_HEAD(&req->timeout.list);
6616 req->timeout.off = off;
6617 if (unlikely(off && !req->ctx->off_timeout_used))
6618 req->ctx->off_timeout_used = true;
6620 if (WARN_ON_ONCE(req_has_async_data(req)))
6622 if (io_alloc_async_data(req))
6625 data = req->async_data;
6627 data->flags = flags;
6629 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6632 if (data->ts.tv_sec < 0 || data->ts.tv_nsec < 0)
6635 INIT_LIST_HEAD(&req->timeout.list);
6636 data->mode = io_translate_timeout_mode(flags);
6637 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6639 if (is_timeout_link) {
6640 struct io_submit_link *link = &req->ctx->submit_state.link;
6644 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6646 req->timeout.head = link->last;
6647 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6652 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6654 struct io_ring_ctx *ctx = req->ctx;
6655 struct io_timeout_data *data = req->async_data;
6656 struct list_head *entry;
6657 u32 tail, off = req->timeout.off;
6659 spin_lock_irq(&ctx->timeout_lock);
6662 * sqe->off holds how many events that need to occur for this
6663 * timeout event to be satisfied. If it isn't set, then this is
6664 * a pure timeout request, sequence isn't used.
6666 if (io_is_timeout_noseq(req)) {
6667 entry = ctx->timeout_list.prev;
6671 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6672 req->timeout.target_seq = tail + off;
6674 /* Update the last seq here in case io_flush_timeouts() hasn't.
6675 * This is safe because ->completion_lock is held, and submissions
6676 * and completions are never mixed in the same ->completion_lock section.
6678 ctx->cq_last_tm_flush = tail;
6681 * Insertion sort, ensuring the first entry in the list is always
6682 * the one we need first.
6684 list_for_each_prev(entry, &ctx->timeout_list) {
6685 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6688 if (io_is_timeout_noseq(nxt))
6690 /* nxt.seq is behind @tail, otherwise would've been completed */
6691 if (off >= nxt->timeout.target_seq - tail)
6695 list_add(&req->timeout.list, entry);
6696 data->timer.function = io_timeout_fn;
6697 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6698 spin_unlock_irq(&ctx->timeout_lock);
6702 struct io_cancel_data {
6703 struct io_ring_ctx *ctx;
6707 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6709 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6710 struct io_cancel_data *cd = data;
6712 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6715 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6716 struct io_ring_ctx *ctx)
6718 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6719 enum io_wq_cancel cancel_ret;
6722 if (!tctx || !tctx->io_wq)
6725 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6726 switch (cancel_ret) {
6727 case IO_WQ_CANCEL_OK:
6730 case IO_WQ_CANCEL_RUNNING:
6733 case IO_WQ_CANCEL_NOTFOUND:
6741 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6743 struct io_ring_ctx *ctx = req->ctx;
6746 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6748 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6750 * Fall-through even for -EALREADY, as we may have poll armed
6751 * that need unarming.
6756 spin_lock(&ctx->completion_lock);
6757 ret = io_poll_cancel(ctx, sqe_addr, false);
6761 spin_lock_irq(&ctx->timeout_lock);
6762 ret = io_timeout_cancel(ctx, sqe_addr);
6763 spin_unlock_irq(&ctx->timeout_lock);
6765 spin_unlock(&ctx->completion_lock);
6769 static int io_async_cancel_prep(struct io_kiocb *req,
6770 const struct io_uring_sqe *sqe)
6772 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6774 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6776 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6780 req->cancel.addr = READ_ONCE(sqe->addr);
6784 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6786 struct io_ring_ctx *ctx = req->ctx;
6787 u64 sqe_addr = req->cancel.addr;
6788 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
6789 struct io_tctx_node *node;
6792 ret = io_try_cancel_userdata(req, sqe_addr);
6796 /* slow path, try all io-wq's */
6797 io_ring_submit_lock(ctx, needs_lock);
6799 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6800 struct io_uring_task *tctx = node->task->io_uring;
6802 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6806 io_ring_submit_unlock(ctx, needs_lock);
6810 io_req_complete_post(req, ret, 0);
6814 static int io_rsrc_update_prep(struct io_kiocb *req,
6815 const struct io_uring_sqe *sqe)
6817 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6819 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6822 req->rsrc_update.offset = READ_ONCE(sqe->off);
6823 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6824 if (!req->rsrc_update.nr_args)
6826 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6830 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6832 struct io_ring_ctx *ctx = req->ctx;
6833 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
6834 struct io_uring_rsrc_update2 up;
6837 up.offset = req->rsrc_update.offset;
6838 up.data = req->rsrc_update.arg;
6844 io_ring_submit_lock(ctx, needs_lock);
6845 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6846 &up, req->rsrc_update.nr_args);
6847 io_ring_submit_unlock(ctx, needs_lock);
6851 __io_req_complete(req, issue_flags, ret, 0);
6855 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6857 switch (req->opcode) {
6860 case IORING_OP_READV:
6861 case IORING_OP_READ_FIXED:
6862 case IORING_OP_READ:
6863 case IORING_OP_WRITEV:
6864 case IORING_OP_WRITE_FIXED:
6865 case IORING_OP_WRITE:
6866 return io_prep_rw(req, sqe);
6867 case IORING_OP_POLL_ADD:
6868 return io_poll_add_prep(req, sqe);
6869 case IORING_OP_POLL_REMOVE:
6870 return io_poll_update_prep(req, sqe);
6871 case IORING_OP_FSYNC:
6872 return io_fsync_prep(req, sqe);
6873 case IORING_OP_SYNC_FILE_RANGE:
6874 return io_sfr_prep(req, sqe);
6875 case IORING_OP_SENDMSG:
6876 case IORING_OP_SEND:
6877 return io_sendmsg_prep(req, sqe);
6878 case IORING_OP_RECVMSG:
6879 case IORING_OP_RECV:
6880 return io_recvmsg_prep(req, sqe);
6881 case IORING_OP_CONNECT:
6882 return io_connect_prep(req, sqe);
6883 case IORING_OP_TIMEOUT:
6884 return io_timeout_prep(req, sqe, false);
6885 case IORING_OP_TIMEOUT_REMOVE:
6886 return io_timeout_remove_prep(req, sqe);
6887 case IORING_OP_ASYNC_CANCEL:
6888 return io_async_cancel_prep(req, sqe);
6889 case IORING_OP_LINK_TIMEOUT:
6890 return io_timeout_prep(req, sqe, true);
6891 case IORING_OP_ACCEPT:
6892 return io_accept_prep(req, sqe);
6893 case IORING_OP_FALLOCATE:
6894 return io_fallocate_prep(req, sqe);
6895 case IORING_OP_OPENAT:
6896 return io_openat_prep(req, sqe);
6897 case IORING_OP_CLOSE:
6898 return io_close_prep(req, sqe);
6899 case IORING_OP_FILES_UPDATE:
6900 return io_rsrc_update_prep(req, sqe);
6901 case IORING_OP_STATX:
6902 return io_statx_prep(req, sqe);
6903 case IORING_OP_FADVISE:
6904 return io_fadvise_prep(req, sqe);
6905 case IORING_OP_MADVISE:
6906 return io_madvise_prep(req, sqe);
6907 case IORING_OP_OPENAT2:
6908 return io_openat2_prep(req, sqe);
6909 case IORING_OP_EPOLL_CTL:
6910 return io_epoll_ctl_prep(req, sqe);
6911 case IORING_OP_SPLICE:
6912 return io_splice_prep(req, sqe);
6913 case IORING_OP_PROVIDE_BUFFERS:
6914 return io_provide_buffers_prep(req, sqe);
6915 case IORING_OP_REMOVE_BUFFERS:
6916 return io_remove_buffers_prep(req, sqe);
6918 return io_tee_prep(req, sqe);
6919 case IORING_OP_SHUTDOWN:
6920 return io_shutdown_prep(req, sqe);
6921 case IORING_OP_RENAMEAT:
6922 return io_renameat_prep(req, sqe);
6923 case IORING_OP_UNLINKAT:
6924 return io_unlinkat_prep(req, sqe);
6925 case IORING_OP_MKDIRAT:
6926 return io_mkdirat_prep(req, sqe);
6927 case IORING_OP_SYMLINKAT:
6928 return io_symlinkat_prep(req, sqe);
6929 case IORING_OP_LINKAT:
6930 return io_linkat_prep(req, sqe);
6931 case IORING_OP_MSG_RING:
6932 return io_msg_ring_prep(req, sqe);
6935 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6940 static int io_req_prep_async(struct io_kiocb *req)
6942 if (!io_op_defs[req->opcode].needs_async_setup)
6944 if (WARN_ON_ONCE(req_has_async_data(req)))
6946 if (io_alloc_async_data(req))
6949 switch (req->opcode) {
6950 case IORING_OP_READV:
6951 return io_rw_prep_async(req, READ);
6952 case IORING_OP_WRITEV:
6953 return io_rw_prep_async(req, WRITE);
6954 case IORING_OP_SENDMSG:
6955 return io_sendmsg_prep_async(req);
6956 case IORING_OP_RECVMSG:
6957 return io_recvmsg_prep_async(req);
6958 case IORING_OP_CONNECT:
6959 return io_connect_prep_async(req);
6961 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6966 static u32 io_get_sequence(struct io_kiocb *req)
6968 u32 seq = req->ctx->cached_sq_head;
6970 /* need original cached_sq_head, but it was increased for each req */
6971 io_for_each_link(req, req)
6976 static __cold void io_drain_req(struct io_kiocb *req)
6978 struct io_ring_ctx *ctx = req->ctx;
6979 struct io_defer_entry *de;
6981 u32 seq = io_get_sequence(req);
6983 /* Still need defer if there is pending req in defer list. */
6984 spin_lock(&ctx->completion_lock);
6985 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
6986 spin_unlock(&ctx->completion_lock);
6988 ctx->drain_active = false;
6989 io_req_task_queue(req);
6992 spin_unlock(&ctx->completion_lock);
6994 ret = io_req_prep_async(req);
6997 io_req_complete_failed(req, ret);
7000 io_prep_async_link(req);
7001 de = kmalloc(sizeof(*de), GFP_KERNEL);
7007 spin_lock(&ctx->completion_lock);
7008 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
7009 spin_unlock(&ctx->completion_lock);
7014 trace_io_uring_defer(ctx, req, req->user_data, req->opcode);
7017 list_add_tail(&de->list, &ctx->defer_list);
7018 spin_unlock(&ctx->completion_lock);
7021 static void io_clean_op(struct io_kiocb *req)
7023 if (req->flags & REQ_F_BUFFER_SELECTED) {
7024 spin_lock(&req->ctx->completion_lock);
7025 io_put_kbuf_comp(req);
7026 spin_unlock(&req->ctx->completion_lock);
7029 if (req->flags & REQ_F_NEED_CLEANUP) {
7030 switch (req->opcode) {
7031 case IORING_OP_READV:
7032 case IORING_OP_READ_FIXED:
7033 case IORING_OP_READ:
7034 case IORING_OP_WRITEV:
7035 case IORING_OP_WRITE_FIXED:
7036 case IORING_OP_WRITE: {
7037 struct io_async_rw *io = req->async_data;
7039 kfree(io->free_iovec);
7042 case IORING_OP_RECVMSG:
7043 case IORING_OP_SENDMSG: {
7044 struct io_async_msghdr *io = req->async_data;
7046 kfree(io->free_iov);
7049 case IORING_OP_OPENAT:
7050 case IORING_OP_OPENAT2:
7051 if (req->open.filename)
7052 putname(req->open.filename);
7054 case IORING_OP_RENAMEAT:
7055 putname(req->rename.oldpath);
7056 putname(req->rename.newpath);
7058 case IORING_OP_UNLINKAT:
7059 putname(req->unlink.filename);
7061 case IORING_OP_MKDIRAT:
7062 putname(req->mkdir.filename);
7064 case IORING_OP_SYMLINKAT:
7065 putname(req->symlink.oldpath);
7066 putname(req->symlink.newpath);
7068 case IORING_OP_LINKAT:
7069 putname(req->hardlink.oldpath);
7070 putname(req->hardlink.newpath);
7072 case IORING_OP_STATX:
7073 if (req->statx.filename)
7074 putname(req->statx.filename);
7078 if ((req->flags & REQ_F_POLLED) && req->apoll) {
7079 kfree(req->apoll->double_poll);
7083 if (req->flags & REQ_F_CREDS)
7084 put_cred(req->creds);
7085 if (req->flags & REQ_F_ASYNC_DATA) {
7086 kfree(req->async_data);
7087 req->async_data = NULL;
7089 req->flags &= ~IO_REQ_CLEAN_FLAGS;
7092 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
7094 if (req->file || !io_op_defs[req->opcode].needs_file)
7097 if (req->flags & REQ_F_FIXED_FILE)
7098 req->file = io_file_get_fixed(req, req->fd, issue_flags);
7100 req->file = io_file_get_normal(req, req->fd);
7105 req->result = -EBADF;
7109 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
7111 const struct cred *creds = NULL;
7114 if (unlikely(!io_assign_file(req, issue_flags)))
7117 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
7118 creds = override_creds(req->creds);
7120 if (!io_op_defs[req->opcode].audit_skip)
7121 audit_uring_entry(req->opcode);
7123 switch (req->opcode) {
7125 ret = io_nop(req, issue_flags);
7127 case IORING_OP_READV:
7128 case IORING_OP_READ_FIXED:
7129 case IORING_OP_READ:
7130 ret = io_read(req, issue_flags);
7132 case IORING_OP_WRITEV:
7133 case IORING_OP_WRITE_FIXED:
7134 case IORING_OP_WRITE:
7135 ret = io_write(req, issue_flags);
7137 case IORING_OP_FSYNC:
7138 ret = io_fsync(req, issue_flags);
7140 case IORING_OP_POLL_ADD:
7141 ret = io_poll_add(req, issue_flags);
7143 case IORING_OP_POLL_REMOVE:
7144 ret = io_poll_update(req, issue_flags);
7146 case IORING_OP_SYNC_FILE_RANGE:
7147 ret = io_sync_file_range(req, issue_flags);
7149 case IORING_OP_SENDMSG:
7150 ret = io_sendmsg(req, issue_flags);
7152 case IORING_OP_SEND:
7153 ret = io_send(req, issue_flags);
7155 case IORING_OP_RECVMSG:
7156 ret = io_recvmsg(req, issue_flags);
7158 case IORING_OP_RECV:
7159 ret = io_recv(req, issue_flags);
7161 case IORING_OP_TIMEOUT:
7162 ret = io_timeout(req, issue_flags);
7164 case IORING_OP_TIMEOUT_REMOVE:
7165 ret = io_timeout_remove(req, issue_flags);
7167 case IORING_OP_ACCEPT:
7168 ret = io_accept(req, issue_flags);
7170 case IORING_OP_CONNECT:
7171 ret = io_connect(req, issue_flags);
7173 case IORING_OP_ASYNC_CANCEL:
7174 ret = io_async_cancel(req, issue_flags);
7176 case IORING_OP_FALLOCATE:
7177 ret = io_fallocate(req, issue_flags);
7179 case IORING_OP_OPENAT:
7180 ret = io_openat(req, issue_flags);
7182 case IORING_OP_CLOSE:
7183 ret = io_close(req, issue_flags);
7185 case IORING_OP_FILES_UPDATE:
7186 ret = io_files_update(req, issue_flags);
7188 case IORING_OP_STATX:
7189 ret = io_statx(req, issue_flags);
7191 case IORING_OP_FADVISE:
7192 ret = io_fadvise(req, issue_flags);
7194 case IORING_OP_MADVISE:
7195 ret = io_madvise(req, issue_flags);
7197 case IORING_OP_OPENAT2:
7198 ret = io_openat2(req, issue_flags);
7200 case IORING_OP_EPOLL_CTL:
7201 ret = io_epoll_ctl(req, issue_flags);
7203 case IORING_OP_SPLICE:
7204 ret = io_splice(req, issue_flags);
7206 case IORING_OP_PROVIDE_BUFFERS:
7207 ret = io_provide_buffers(req, issue_flags);
7209 case IORING_OP_REMOVE_BUFFERS:
7210 ret = io_remove_buffers(req, issue_flags);
7213 ret = io_tee(req, issue_flags);
7215 case IORING_OP_SHUTDOWN:
7216 ret = io_shutdown(req, issue_flags);
7218 case IORING_OP_RENAMEAT:
7219 ret = io_renameat(req, issue_flags);
7221 case IORING_OP_UNLINKAT:
7222 ret = io_unlinkat(req, issue_flags);
7224 case IORING_OP_MKDIRAT:
7225 ret = io_mkdirat(req, issue_flags);
7227 case IORING_OP_SYMLINKAT:
7228 ret = io_symlinkat(req, issue_flags);
7230 case IORING_OP_LINKAT:
7231 ret = io_linkat(req, issue_flags);
7233 case IORING_OP_MSG_RING:
7234 ret = io_msg_ring(req, issue_flags);
7241 if (!io_op_defs[req->opcode].audit_skip)
7242 audit_uring_exit(!ret, ret);
7245 revert_creds(creds);
7248 /* If the op doesn't have a file, we're not polling for it */
7249 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
7250 io_iopoll_req_issued(req, issue_flags);
7255 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
7257 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7259 req = io_put_req_find_next(req);
7260 return req ? &req->work : NULL;
7263 static void io_wq_submit_work(struct io_wq_work *work)
7265 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7266 const struct io_op_def *def = &io_op_defs[req->opcode];
7267 unsigned int issue_flags = IO_URING_F_UNLOCKED;
7268 bool needs_poll = false;
7269 struct io_kiocb *timeout;
7270 int ret = 0, err = -ECANCELED;
7272 /* one will be dropped by ->io_free_work() after returning to io-wq */
7273 if (!(req->flags & REQ_F_REFCOUNT))
7274 __io_req_set_refcount(req, 2);
7278 timeout = io_prep_linked_timeout(req);
7280 io_queue_linked_timeout(timeout);
7283 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
7284 if (work->flags & IO_WQ_WORK_CANCEL) {
7286 io_req_task_queue_fail(req, err);
7289 if (!io_assign_file(req, issue_flags)) {
7291 work->flags |= IO_WQ_WORK_CANCEL;
7295 if (req->flags & REQ_F_FORCE_ASYNC) {
7296 bool opcode_poll = def->pollin || def->pollout;
7298 if (opcode_poll && file_can_poll(req->file)) {
7300 issue_flags |= IO_URING_F_NONBLOCK;
7305 ret = io_issue_sqe(req, issue_flags);
7309 * We can get EAGAIN for iopolled IO even though we're
7310 * forcing a sync submission from here, since we can't
7311 * wait for request slots on the block side.
7318 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
7320 /* aborted or ready, in either case retry blocking */
7322 issue_flags &= ~IO_URING_F_NONBLOCK;
7325 /* avoid locking problems by failing it from a clean context */
7327 io_req_task_queue_fail(req, ret);
7330 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
7333 return &table->files[i];
7336 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
7339 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
7341 return (struct file *) (slot->file_ptr & FFS_MASK);
7344 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
7346 unsigned long file_ptr = (unsigned long) file;
7348 file_ptr |= io_file_get_flags(file);
7349 file_slot->file_ptr = file_ptr;
7352 static inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
7353 unsigned int issue_flags)
7355 struct io_ring_ctx *ctx = req->ctx;
7356 struct file *file = NULL;
7357 unsigned long file_ptr;
7359 if (issue_flags & IO_URING_F_UNLOCKED)
7360 mutex_lock(&ctx->uring_lock);
7362 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
7364 fd = array_index_nospec(fd, ctx->nr_user_files);
7365 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
7366 file = (struct file *) (file_ptr & FFS_MASK);
7367 file_ptr &= ~FFS_MASK;
7368 /* mask in overlapping REQ_F and FFS bits */
7369 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
7370 io_req_set_rsrc_node(req, ctx, 0);
7372 if (issue_flags & IO_URING_F_UNLOCKED)
7373 mutex_unlock(&ctx->uring_lock);
7378 * Drop the file for requeue operations. Only used of req->file is the
7379 * io_uring descriptor itself.
7381 static void io_drop_inflight_file(struct io_kiocb *req)
7383 if (unlikely(req->flags & REQ_F_INFLIGHT)) {
7386 req->flags &= ~REQ_F_INFLIGHT;
7390 static struct file *io_file_get_normal(struct io_kiocb *req, int fd)
7392 struct file *file = fget(fd);
7394 trace_io_uring_file_get(req->ctx, req, req->user_data, fd);
7396 /* we don't allow fixed io_uring files */
7397 if (file && file->f_op == &io_uring_fops)
7398 req->flags |= REQ_F_INFLIGHT;
7402 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
7404 struct io_kiocb *prev = req->timeout.prev;
7408 if (!(req->task->flags & PF_EXITING))
7409 ret = io_try_cancel_userdata(req, prev->user_data);
7410 io_req_complete_post(req, ret ?: -ETIME, 0);
7413 io_req_complete_post(req, -ETIME, 0);
7417 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
7419 struct io_timeout_data *data = container_of(timer,
7420 struct io_timeout_data, timer);
7421 struct io_kiocb *prev, *req = data->req;
7422 struct io_ring_ctx *ctx = req->ctx;
7423 unsigned long flags;
7425 spin_lock_irqsave(&ctx->timeout_lock, flags);
7426 prev = req->timeout.head;
7427 req->timeout.head = NULL;
7430 * We don't expect the list to be empty, that will only happen if we
7431 * race with the completion of the linked work.
7434 io_remove_next_linked(prev);
7435 if (!req_ref_inc_not_zero(prev))
7438 list_del(&req->timeout.list);
7439 req->timeout.prev = prev;
7440 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
7442 req->io_task_work.func = io_req_task_link_timeout;
7443 io_req_task_work_add(req, false);
7444 return HRTIMER_NORESTART;
7447 static void io_queue_linked_timeout(struct io_kiocb *req)
7449 struct io_ring_ctx *ctx = req->ctx;
7451 spin_lock_irq(&ctx->timeout_lock);
7453 * If the back reference is NULL, then our linked request finished
7454 * before we got a chance to setup the timer
7456 if (req->timeout.head) {
7457 struct io_timeout_data *data = req->async_data;
7459 data->timer.function = io_link_timeout_fn;
7460 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
7462 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
7464 spin_unlock_irq(&ctx->timeout_lock);
7465 /* drop submission reference */
7469 static void io_queue_sqe_arm_apoll(struct io_kiocb *req)
7470 __must_hold(&req->ctx->uring_lock)
7472 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
7474 switch (io_arm_poll_handler(req, 0)) {
7475 case IO_APOLL_READY:
7476 io_req_task_queue(req);
7478 case IO_APOLL_ABORTED:
7480 * Queued up for async execution, worker will release
7481 * submit reference when the iocb is actually submitted.
7483 io_queue_async_work(req, NULL);
7490 io_queue_linked_timeout(linked_timeout);
7493 static inline void __io_queue_sqe(struct io_kiocb *req)
7494 __must_hold(&req->ctx->uring_lock)
7496 struct io_kiocb *linked_timeout;
7499 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
7501 if (req->flags & REQ_F_COMPLETE_INLINE) {
7502 io_req_add_compl_list(req);
7506 * We async punt it if the file wasn't marked NOWAIT, or if the file
7507 * doesn't support non-blocking read/write attempts
7510 linked_timeout = io_prep_linked_timeout(req);
7512 io_queue_linked_timeout(linked_timeout);
7513 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
7514 io_queue_sqe_arm_apoll(req);
7516 io_req_complete_failed(req, ret);
7520 static void io_queue_sqe_fallback(struct io_kiocb *req)
7521 __must_hold(&req->ctx->uring_lock)
7523 if (req->flags & REQ_F_FAIL) {
7524 io_req_complete_fail_submit(req);
7525 } else if (unlikely(req->ctx->drain_active)) {
7528 int ret = io_req_prep_async(req);
7531 io_req_complete_failed(req, ret);
7533 io_queue_async_work(req, NULL);
7537 static inline void io_queue_sqe(struct io_kiocb *req)
7538 __must_hold(&req->ctx->uring_lock)
7540 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))))
7541 __io_queue_sqe(req);
7543 io_queue_sqe_fallback(req);
7547 * Check SQE restrictions (opcode and flags).
7549 * Returns 'true' if SQE is allowed, 'false' otherwise.
7551 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7552 struct io_kiocb *req,
7553 unsigned int sqe_flags)
7555 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7558 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7559 ctx->restrictions.sqe_flags_required)
7562 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7563 ctx->restrictions.sqe_flags_required))
7569 static void io_init_req_drain(struct io_kiocb *req)
7571 struct io_ring_ctx *ctx = req->ctx;
7572 struct io_kiocb *head = ctx->submit_state.link.head;
7574 ctx->drain_active = true;
7577 * If we need to drain a request in the middle of a link, drain
7578 * the head request and the next request/link after the current
7579 * link. Considering sequential execution of links,
7580 * REQ_F_IO_DRAIN will be maintained for every request of our
7583 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
7584 ctx->drain_next = true;
7588 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7589 const struct io_uring_sqe *sqe)
7590 __must_hold(&ctx->uring_lock)
7592 unsigned int sqe_flags;
7596 /* req is partially pre-initialised, see io_preinit_req() */
7597 req->opcode = opcode = READ_ONCE(sqe->opcode);
7598 /* same numerical values with corresponding REQ_F_*, safe to copy */
7599 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7600 req->user_data = READ_ONCE(sqe->user_data);
7602 req->fixed_rsrc_refs = NULL;
7603 req->task = current;
7605 if (unlikely(opcode >= IORING_OP_LAST)) {
7609 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
7610 /* enforce forwards compatibility on users */
7611 if (sqe_flags & ~SQE_VALID_FLAGS)
7613 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7614 !io_op_defs[opcode].buffer_select)
7616 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
7617 ctx->drain_disabled = true;
7618 if (sqe_flags & IOSQE_IO_DRAIN) {
7619 if (ctx->drain_disabled)
7621 io_init_req_drain(req);
7624 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
7625 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
7627 /* knock it to the slow queue path, will be drained there */
7628 if (ctx->drain_active)
7629 req->flags |= REQ_F_FORCE_ASYNC;
7630 /* if there is no link, we're at "next" request and need to drain */
7631 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
7632 ctx->drain_next = false;
7633 ctx->drain_active = true;
7634 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
7638 if (io_op_defs[opcode].needs_file) {
7639 struct io_submit_state *state = &ctx->submit_state;
7641 req->fd = READ_ONCE(sqe->fd);
7644 * Plug now if we have more than 2 IO left after this, and the
7645 * target is potentially a read/write to block based storage.
7647 if (state->need_plug && io_op_defs[opcode].plug) {
7648 state->plug_started = true;
7649 state->need_plug = false;
7650 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
7654 personality = READ_ONCE(sqe->personality);
7658 req->creds = xa_load(&ctx->personalities, personality);
7661 get_cred(req->creds);
7662 ret = security_uring_override_creds(req->creds);
7664 put_cred(req->creds);
7667 req->flags |= REQ_F_CREDS;
7670 return io_req_prep(req, sqe);
7673 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7674 const struct io_uring_sqe *sqe)
7675 __must_hold(&ctx->uring_lock)
7677 struct io_submit_link *link = &ctx->submit_state.link;
7680 ret = io_init_req(ctx, req, sqe);
7681 if (unlikely(ret)) {
7682 trace_io_uring_req_failed(sqe, ctx, req, ret);
7684 /* fail even hard links since we don't submit */
7687 * we can judge a link req is failed or cancelled by if
7688 * REQ_F_FAIL is set, but the head is an exception since
7689 * it may be set REQ_F_FAIL because of other req's failure
7690 * so let's leverage req->result to distinguish if a head
7691 * is set REQ_F_FAIL because of its failure or other req's
7692 * failure so that we can set the correct ret code for it.
7693 * init result here to avoid affecting the normal path.
7695 if (!(link->head->flags & REQ_F_FAIL))
7696 req_fail_link_node(link->head, -ECANCELED);
7697 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7699 * the current req is a normal req, we should return
7700 * error and thus break the submittion loop.
7702 io_req_complete_failed(req, ret);
7705 req_fail_link_node(req, ret);
7708 /* don't need @sqe from now on */
7709 trace_io_uring_submit_sqe(ctx, req, req->user_data, req->opcode,
7711 ctx->flags & IORING_SETUP_SQPOLL);
7714 * If we already have a head request, queue this one for async
7715 * submittal once the head completes. If we don't have a head but
7716 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7717 * submitted sync once the chain is complete. If none of those
7718 * conditions are true (normal request), then just queue it.
7721 struct io_kiocb *head = link->head;
7723 if (!(req->flags & REQ_F_FAIL)) {
7724 ret = io_req_prep_async(req);
7725 if (unlikely(ret)) {
7726 req_fail_link_node(req, ret);
7727 if (!(head->flags & REQ_F_FAIL))
7728 req_fail_link_node(head, -ECANCELED);
7731 trace_io_uring_link(ctx, req, head);
7732 link->last->link = req;
7735 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
7737 /* last request of a link, enqueue the link */
7740 } else if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7751 * Batched submission is done, ensure local IO is flushed out.
7753 static void io_submit_state_end(struct io_ring_ctx *ctx)
7755 struct io_submit_state *state = &ctx->submit_state;
7757 if (state->link.head)
7758 io_queue_sqe(state->link.head);
7759 /* flush only after queuing links as they can generate completions */
7760 io_submit_flush_completions(ctx);
7761 if (state->plug_started)
7762 blk_finish_plug(&state->plug);
7766 * Start submission side cache.
7768 static void io_submit_state_start(struct io_submit_state *state,
7769 unsigned int max_ios)
7771 state->plug_started = false;
7772 state->need_plug = max_ios > 2;
7773 state->submit_nr = max_ios;
7774 /* set only head, no need to init link_last in advance */
7775 state->link.head = NULL;
7778 static void io_commit_sqring(struct io_ring_ctx *ctx)
7780 struct io_rings *rings = ctx->rings;
7783 * Ensure any loads from the SQEs are done at this point,
7784 * since once we write the new head, the application could
7785 * write new data to them.
7787 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7791 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7792 * that is mapped by userspace. This means that care needs to be taken to
7793 * ensure that reads are stable, as we cannot rely on userspace always
7794 * being a good citizen. If members of the sqe are validated and then later
7795 * used, it's important that those reads are done through READ_ONCE() to
7796 * prevent a re-load down the line.
7798 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7800 unsigned head, mask = ctx->sq_entries - 1;
7801 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7804 * The cached sq head (or cq tail) serves two purposes:
7806 * 1) allows us to batch the cost of updating the user visible
7808 * 2) allows the kernel side to track the head on its own, even
7809 * though the application is the one updating it.
7811 head = READ_ONCE(ctx->sq_array[sq_idx]);
7812 if (likely(head < ctx->sq_entries))
7813 return &ctx->sq_sqes[head];
7815 /* drop invalid entries */
7817 WRITE_ONCE(ctx->rings->sq_dropped,
7818 READ_ONCE(ctx->rings->sq_dropped) + 1);
7822 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7823 __must_hold(&ctx->uring_lock)
7825 unsigned int entries = io_sqring_entries(ctx);
7828 if (unlikely(!entries))
7830 /* make sure SQ entry isn't read before tail */
7831 nr = min3(nr, ctx->sq_entries, entries);
7832 io_get_task_refs(nr);
7834 io_submit_state_start(&ctx->submit_state, nr);
7836 const struct io_uring_sqe *sqe;
7837 struct io_kiocb *req;
7839 if (unlikely(!io_alloc_req_refill(ctx))) {
7841 submitted = -EAGAIN;
7844 req = io_alloc_req(ctx);
7845 sqe = io_get_sqe(ctx);
7846 if (unlikely(!sqe)) {
7847 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
7850 /* will complete beyond this point, count as submitted */
7852 if (io_submit_sqe(ctx, req, sqe)) {
7854 * Continue submitting even for sqe failure if the
7855 * ring was setup with IORING_SETUP_SUBMIT_ALL
7857 if (!(ctx->flags & IORING_SETUP_SUBMIT_ALL))
7860 } while (submitted < nr);
7862 if (unlikely(submitted != nr)) {
7863 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7864 int unused = nr - ref_used;
7866 current->io_uring->cached_refs += unused;
7869 io_submit_state_end(ctx);
7870 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7871 io_commit_sqring(ctx);
7876 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7878 return READ_ONCE(sqd->state);
7881 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7883 /* Tell userspace we may need a wakeup call */
7884 spin_lock(&ctx->completion_lock);
7885 WRITE_ONCE(ctx->rings->sq_flags,
7886 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7887 spin_unlock(&ctx->completion_lock);
7890 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7892 spin_lock(&ctx->completion_lock);
7893 WRITE_ONCE(ctx->rings->sq_flags,
7894 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7895 spin_unlock(&ctx->completion_lock);
7898 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7900 unsigned int to_submit;
7903 to_submit = io_sqring_entries(ctx);
7904 /* if we're handling multiple rings, cap submit size for fairness */
7905 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7906 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7908 if (!wq_list_empty(&ctx->iopoll_list) || to_submit) {
7909 const struct cred *creds = NULL;
7911 if (ctx->sq_creds != current_cred())
7912 creds = override_creds(ctx->sq_creds);
7914 mutex_lock(&ctx->uring_lock);
7915 if (!wq_list_empty(&ctx->iopoll_list))
7916 io_do_iopoll(ctx, true);
7919 * Don't submit if refs are dying, good for io_uring_register(),
7920 * but also it is relied upon by io_ring_exit_work()
7922 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7923 !(ctx->flags & IORING_SETUP_R_DISABLED))
7924 ret = io_submit_sqes(ctx, to_submit);
7925 mutex_unlock(&ctx->uring_lock);
7927 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7928 wake_up(&ctx->sqo_sq_wait);
7930 revert_creds(creds);
7936 static __cold void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7938 struct io_ring_ctx *ctx;
7939 unsigned sq_thread_idle = 0;
7941 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7942 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7943 sqd->sq_thread_idle = sq_thread_idle;
7946 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7948 bool did_sig = false;
7949 struct ksignal ksig;
7951 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7952 signal_pending(current)) {
7953 mutex_unlock(&sqd->lock);
7954 if (signal_pending(current))
7955 did_sig = get_signal(&ksig);
7957 mutex_lock(&sqd->lock);
7959 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7962 static int io_sq_thread(void *data)
7964 struct io_sq_data *sqd = data;
7965 struct io_ring_ctx *ctx;
7966 unsigned long timeout = 0;
7967 char buf[TASK_COMM_LEN];
7970 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7971 set_task_comm(current, buf);
7973 if (sqd->sq_cpu != -1)
7974 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7976 set_cpus_allowed_ptr(current, cpu_online_mask);
7977 current->flags |= PF_NO_SETAFFINITY;
7979 audit_alloc_kernel(current);
7981 mutex_lock(&sqd->lock);
7983 bool cap_entries, sqt_spin = false;
7985 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7986 if (io_sqd_handle_event(sqd))
7988 timeout = jiffies + sqd->sq_thread_idle;
7991 cap_entries = !list_is_singular(&sqd->ctx_list);
7992 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7993 int ret = __io_sq_thread(ctx, cap_entries);
7995 if (!sqt_spin && (ret > 0 || !wq_list_empty(&ctx->iopoll_list)))
7998 if (io_run_task_work())
8001 if (sqt_spin || !time_after(jiffies, timeout)) {
8004 timeout = jiffies + sqd->sq_thread_idle;
8008 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
8009 if (!io_sqd_events_pending(sqd) && !task_work_pending(current)) {
8010 bool needs_sched = true;
8012 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
8013 io_ring_set_wakeup_flag(ctx);
8015 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
8016 !wq_list_empty(&ctx->iopoll_list)) {
8017 needs_sched = false;
8022 * Ensure the store of the wakeup flag is not
8023 * reordered with the load of the SQ tail
8027 if (io_sqring_entries(ctx)) {
8028 needs_sched = false;
8034 mutex_unlock(&sqd->lock);
8036 mutex_lock(&sqd->lock);
8038 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
8039 io_ring_clear_wakeup_flag(ctx);
8042 finish_wait(&sqd->wait, &wait);
8043 timeout = jiffies + sqd->sq_thread_idle;
8046 io_uring_cancel_generic(true, sqd);
8048 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
8049 io_ring_set_wakeup_flag(ctx);
8051 mutex_unlock(&sqd->lock);
8053 audit_free(current);
8055 complete(&sqd->exited);
8059 struct io_wait_queue {
8060 struct wait_queue_entry wq;
8061 struct io_ring_ctx *ctx;
8063 unsigned nr_timeouts;
8066 static inline bool io_should_wake(struct io_wait_queue *iowq)
8068 struct io_ring_ctx *ctx = iowq->ctx;
8069 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
8072 * Wake up if we have enough events, or if a timeout occurred since we
8073 * started waiting. For timeouts, we always want to return to userspace,
8074 * regardless of event count.
8076 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
8079 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
8080 int wake_flags, void *key)
8082 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
8086 * Cannot safely flush overflowed CQEs from here, ensure we wake up
8087 * the task, and the next invocation will do it.
8089 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
8090 return autoremove_wake_function(curr, mode, wake_flags, key);
8094 static int io_run_task_work_sig(void)
8096 if (io_run_task_work())
8098 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
8099 return -ERESTARTSYS;
8100 if (task_sigpending(current))
8105 /* when returns >0, the caller should retry */
8106 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
8107 struct io_wait_queue *iowq,
8112 /* make sure we run task_work before checking for signals */
8113 ret = io_run_task_work_sig();
8114 if (ret || io_should_wake(iowq))
8116 /* let the caller flush overflows, retry */
8117 if (test_bit(0, &ctx->check_cq_overflow))
8120 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
8126 * Wait until events become available, if we don't already have some. The
8127 * application must reap them itself, as they reside on the shared cq ring.
8129 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
8130 const sigset_t __user *sig, size_t sigsz,
8131 struct __kernel_timespec __user *uts)
8133 struct io_wait_queue iowq;
8134 struct io_rings *rings = ctx->rings;
8135 ktime_t timeout = KTIME_MAX;
8139 io_cqring_overflow_flush(ctx);
8140 if (io_cqring_events(ctx) >= min_events)
8142 if (!io_run_task_work())
8147 #ifdef CONFIG_COMPAT
8148 if (in_compat_syscall())
8149 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
8153 ret = set_user_sigmask(sig, sigsz);
8160 struct timespec64 ts;
8162 if (get_timespec64(&ts, uts))
8164 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
8167 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
8168 iowq.wq.private = current;
8169 INIT_LIST_HEAD(&iowq.wq.entry);
8171 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
8172 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
8174 trace_io_uring_cqring_wait(ctx, min_events);
8176 /* if we can't even flush overflow, don't wait for more */
8177 if (!io_cqring_overflow_flush(ctx)) {
8181 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
8182 TASK_INTERRUPTIBLE);
8183 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
8184 finish_wait(&ctx->cq_wait, &iowq.wq);
8188 restore_saved_sigmask_unless(ret == -EINTR);
8190 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
8193 static void io_free_page_table(void **table, size_t size)
8195 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
8197 for (i = 0; i < nr_tables; i++)
8202 static __cold void **io_alloc_page_table(size_t size)
8204 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
8205 size_t init_size = size;
8208 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
8212 for (i = 0; i < nr_tables; i++) {
8213 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
8215 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
8217 io_free_page_table(table, init_size);
8225 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
8227 percpu_ref_exit(&ref_node->refs);
8231 static __cold void io_rsrc_node_ref_zero(struct percpu_ref *ref)
8233 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
8234 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
8235 unsigned long flags;
8236 bool first_add = false;
8237 unsigned long delay = HZ;
8239 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
8242 /* if we are mid-quiesce then do not delay */
8243 if (node->rsrc_data->quiesce)
8246 while (!list_empty(&ctx->rsrc_ref_list)) {
8247 node = list_first_entry(&ctx->rsrc_ref_list,
8248 struct io_rsrc_node, node);
8249 /* recycle ref nodes in order */
8252 list_del(&node->node);
8253 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
8255 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
8258 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
8261 static struct io_rsrc_node *io_rsrc_node_alloc(void)
8263 struct io_rsrc_node *ref_node;
8265 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
8269 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
8274 INIT_LIST_HEAD(&ref_node->node);
8275 INIT_LIST_HEAD(&ref_node->rsrc_list);
8276 ref_node->done = false;
8280 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
8281 struct io_rsrc_data *data_to_kill)
8282 __must_hold(&ctx->uring_lock)
8284 WARN_ON_ONCE(!ctx->rsrc_backup_node);
8285 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
8287 io_rsrc_refs_drop(ctx);
8290 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
8292 rsrc_node->rsrc_data = data_to_kill;
8293 spin_lock_irq(&ctx->rsrc_ref_lock);
8294 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
8295 spin_unlock_irq(&ctx->rsrc_ref_lock);
8297 atomic_inc(&data_to_kill->refs);
8298 percpu_ref_kill(&rsrc_node->refs);
8299 ctx->rsrc_node = NULL;
8302 if (!ctx->rsrc_node) {
8303 ctx->rsrc_node = ctx->rsrc_backup_node;
8304 ctx->rsrc_backup_node = NULL;
8308 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
8310 if (ctx->rsrc_backup_node)
8312 ctx->rsrc_backup_node = io_rsrc_node_alloc();
8313 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
8316 static __cold int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
8317 struct io_ring_ctx *ctx)
8321 /* As we may drop ->uring_lock, other task may have started quiesce */
8325 data->quiesce = true;
8327 ret = io_rsrc_node_switch_start(ctx);
8330 io_rsrc_node_switch(ctx, data);
8332 /* kill initial ref, already quiesced if zero */
8333 if (atomic_dec_and_test(&data->refs))
8335 mutex_unlock(&ctx->uring_lock);
8336 flush_delayed_work(&ctx->rsrc_put_work);
8337 ret = wait_for_completion_interruptible(&data->done);
8339 mutex_lock(&ctx->uring_lock);
8340 if (atomic_read(&data->refs) > 0) {
8342 * it has been revived by another thread while
8345 mutex_unlock(&ctx->uring_lock);
8351 atomic_inc(&data->refs);
8352 /* wait for all works potentially completing data->done */
8353 flush_delayed_work(&ctx->rsrc_put_work);
8354 reinit_completion(&data->done);
8356 ret = io_run_task_work_sig();
8357 mutex_lock(&ctx->uring_lock);
8359 data->quiesce = false;
8364 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
8366 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
8367 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
8369 return &data->tags[table_idx][off];
8372 static void io_rsrc_data_free(struct io_rsrc_data *data)
8374 size_t size = data->nr * sizeof(data->tags[0][0]);
8377 io_free_page_table((void **)data->tags, size);
8381 static __cold int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
8382 u64 __user *utags, unsigned nr,
8383 struct io_rsrc_data **pdata)
8385 struct io_rsrc_data *data;
8389 data = kzalloc(sizeof(*data), GFP_KERNEL);
8392 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
8400 data->do_put = do_put;
8403 for (i = 0; i < nr; i++) {
8404 u64 *tag_slot = io_get_tag_slot(data, i);
8406 if (copy_from_user(tag_slot, &utags[i],
8412 atomic_set(&data->refs, 1);
8413 init_completion(&data->done);
8417 io_rsrc_data_free(data);
8421 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
8423 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
8424 GFP_KERNEL_ACCOUNT);
8425 return !!table->files;
8428 static void io_free_file_tables(struct io_file_table *table)
8430 kvfree(table->files);
8431 table->files = NULL;
8434 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
8436 #if defined(CONFIG_UNIX)
8437 if (ctx->ring_sock) {
8438 struct sock *sock = ctx->ring_sock->sk;
8439 struct sk_buff *skb;
8441 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
8447 for (i = 0; i < ctx->nr_user_files; i++) {
8450 file = io_file_from_index(ctx, i);
8455 io_free_file_tables(&ctx->file_table);
8456 io_rsrc_data_free(ctx->file_data);
8457 ctx->file_data = NULL;
8458 ctx->nr_user_files = 0;
8461 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
8465 if (!ctx->file_data)
8467 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
8469 __io_sqe_files_unregister(ctx);
8473 static void io_sq_thread_unpark(struct io_sq_data *sqd)
8474 __releases(&sqd->lock)
8476 WARN_ON_ONCE(sqd->thread == current);
8479 * Do the dance but not conditional clear_bit() because it'd race with
8480 * other threads incrementing park_pending and setting the bit.
8482 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
8483 if (atomic_dec_return(&sqd->park_pending))
8484 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
8485 mutex_unlock(&sqd->lock);
8488 static void io_sq_thread_park(struct io_sq_data *sqd)
8489 __acquires(&sqd->lock)
8491 WARN_ON_ONCE(sqd->thread == current);
8493 atomic_inc(&sqd->park_pending);
8494 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
8495 mutex_lock(&sqd->lock);
8497 wake_up_process(sqd->thread);
8500 static void io_sq_thread_stop(struct io_sq_data *sqd)
8502 WARN_ON_ONCE(sqd->thread == current);
8503 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
8505 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
8506 mutex_lock(&sqd->lock);
8508 wake_up_process(sqd->thread);
8509 mutex_unlock(&sqd->lock);
8510 wait_for_completion(&sqd->exited);
8513 static void io_put_sq_data(struct io_sq_data *sqd)
8515 if (refcount_dec_and_test(&sqd->refs)) {
8516 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
8518 io_sq_thread_stop(sqd);
8523 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
8525 struct io_sq_data *sqd = ctx->sq_data;
8528 io_sq_thread_park(sqd);
8529 list_del_init(&ctx->sqd_list);
8530 io_sqd_update_thread_idle(sqd);
8531 io_sq_thread_unpark(sqd);
8533 io_put_sq_data(sqd);
8534 ctx->sq_data = NULL;
8538 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
8540 struct io_ring_ctx *ctx_attach;
8541 struct io_sq_data *sqd;
8544 f = fdget(p->wq_fd);
8546 return ERR_PTR(-ENXIO);
8547 if (f.file->f_op != &io_uring_fops) {
8549 return ERR_PTR(-EINVAL);
8552 ctx_attach = f.file->private_data;
8553 sqd = ctx_attach->sq_data;
8556 return ERR_PTR(-EINVAL);
8558 if (sqd->task_tgid != current->tgid) {
8560 return ERR_PTR(-EPERM);
8563 refcount_inc(&sqd->refs);
8568 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
8571 struct io_sq_data *sqd;
8574 if (p->flags & IORING_SETUP_ATTACH_WQ) {
8575 sqd = io_attach_sq_data(p);
8580 /* fall through for EPERM case, setup new sqd/task */
8581 if (PTR_ERR(sqd) != -EPERM)
8585 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
8587 return ERR_PTR(-ENOMEM);
8589 atomic_set(&sqd->park_pending, 0);
8590 refcount_set(&sqd->refs, 1);
8591 INIT_LIST_HEAD(&sqd->ctx_list);
8592 mutex_init(&sqd->lock);
8593 init_waitqueue_head(&sqd->wait);
8594 init_completion(&sqd->exited);
8598 #if defined(CONFIG_UNIX)
8600 * Ensure the UNIX gc is aware of our file set, so we are certain that
8601 * the io_uring can be safely unregistered on process exit, even if we have
8602 * loops in the file referencing.
8604 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8606 struct sock *sk = ctx->ring_sock->sk;
8607 struct scm_fp_list *fpl;
8608 struct sk_buff *skb;
8611 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8615 skb = alloc_skb(0, GFP_KERNEL);
8624 fpl->user = get_uid(current_user());
8625 for (i = 0; i < nr; i++) {
8626 struct file *file = io_file_from_index(ctx, i + offset);
8630 fpl->fp[nr_files] = get_file(file);
8631 unix_inflight(fpl->user, fpl->fp[nr_files]);
8636 fpl->max = SCM_MAX_FD;
8637 fpl->count = nr_files;
8638 UNIXCB(skb).fp = fpl;
8639 skb->destructor = unix_destruct_scm;
8640 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8641 skb_queue_head(&sk->sk_receive_queue, skb);
8643 for (i = 0; i < nr; i++) {
8644 struct file *file = io_file_from_index(ctx, i + offset);
8651 free_uid(fpl->user);
8659 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8660 * causes regular reference counting to break down. We rely on the UNIX
8661 * garbage collection to take care of this problem for us.
8663 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8665 unsigned left, total;
8669 left = ctx->nr_user_files;
8671 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8673 ret = __io_sqe_files_scm(ctx, this_files, total);
8677 total += this_files;
8683 while (total < ctx->nr_user_files) {
8684 struct file *file = io_file_from_index(ctx, total);
8694 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8700 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8702 struct file *file = prsrc->file;
8703 #if defined(CONFIG_UNIX)
8704 struct sock *sock = ctx->ring_sock->sk;
8705 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8706 struct sk_buff *skb;
8709 __skb_queue_head_init(&list);
8712 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8713 * remove this entry and rearrange the file array.
8715 skb = skb_dequeue(head);
8717 struct scm_fp_list *fp;
8719 fp = UNIXCB(skb).fp;
8720 for (i = 0; i < fp->count; i++) {
8723 if (fp->fp[i] != file)
8726 unix_notinflight(fp->user, fp->fp[i]);
8727 left = fp->count - 1 - i;
8729 memmove(&fp->fp[i], &fp->fp[i + 1],
8730 left * sizeof(struct file *));
8737 __skb_queue_tail(&list, skb);
8747 __skb_queue_tail(&list, skb);
8749 skb = skb_dequeue(head);
8752 if (skb_peek(&list)) {
8753 spin_lock_irq(&head->lock);
8754 while ((skb = __skb_dequeue(&list)) != NULL)
8755 __skb_queue_tail(head, skb);
8756 spin_unlock_irq(&head->lock);
8763 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8765 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8766 struct io_ring_ctx *ctx = rsrc_data->ctx;
8767 struct io_rsrc_put *prsrc, *tmp;
8769 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8770 list_del(&prsrc->list);
8773 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8775 io_ring_submit_lock(ctx, lock_ring);
8776 spin_lock(&ctx->completion_lock);
8777 io_fill_cqe_aux(ctx, prsrc->tag, 0, 0);
8778 io_commit_cqring(ctx);
8779 spin_unlock(&ctx->completion_lock);
8780 io_cqring_ev_posted(ctx);
8781 io_ring_submit_unlock(ctx, lock_ring);
8784 rsrc_data->do_put(ctx, prsrc);
8788 io_rsrc_node_destroy(ref_node);
8789 if (atomic_dec_and_test(&rsrc_data->refs))
8790 complete(&rsrc_data->done);
8793 static void io_rsrc_put_work(struct work_struct *work)
8795 struct io_ring_ctx *ctx;
8796 struct llist_node *node;
8798 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8799 node = llist_del_all(&ctx->rsrc_put_llist);
8802 struct io_rsrc_node *ref_node;
8803 struct llist_node *next = node->next;
8805 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8806 __io_rsrc_put_work(ref_node);
8811 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8812 unsigned nr_args, u64 __user *tags)
8814 __s32 __user *fds = (__s32 __user *) arg;
8823 if (nr_args > IORING_MAX_FIXED_FILES)
8825 if (nr_args > rlimit(RLIMIT_NOFILE))
8827 ret = io_rsrc_node_switch_start(ctx);
8830 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8836 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8839 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8840 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8844 /* allow sparse sets */
8847 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8854 if (unlikely(!file))
8858 * Don't allow io_uring instances to be registered. If UNIX
8859 * isn't enabled, then this causes a reference cycle and this
8860 * instance can never get freed. If UNIX is enabled we'll
8861 * handle it just fine, but there's still no point in allowing
8862 * a ring fd as it doesn't support regular read/write anyway.
8864 if (file->f_op == &io_uring_fops) {
8868 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8871 ret = io_sqe_files_scm(ctx);
8873 __io_sqe_files_unregister(ctx);
8877 io_rsrc_node_switch(ctx, NULL);
8880 for (i = 0; i < ctx->nr_user_files; i++) {
8881 file = io_file_from_index(ctx, i);
8885 io_free_file_tables(&ctx->file_table);
8886 ctx->nr_user_files = 0;
8888 io_rsrc_data_free(ctx->file_data);
8889 ctx->file_data = NULL;
8893 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8896 #if defined(CONFIG_UNIX)
8897 struct sock *sock = ctx->ring_sock->sk;
8898 struct sk_buff_head *head = &sock->sk_receive_queue;
8899 struct sk_buff *skb;
8902 * See if we can merge this file into an existing skb SCM_RIGHTS
8903 * file set. If there's no room, fall back to allocating a new skb
8904 * and filling it in.
8906 spin_lock_irq(&head->lock);
8907 skb = skb_peek(head);
8909 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8911 if (fpl->count < SCM_MAX_FD) {
8912 __skb_unlink(skb, head);
8913 spin_unlock_irq(&head->lock);
8914 fpl->fp[fpl->count] = get_file(file);
8915 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8917 spin_lock_irq(&head->lock);
8918 __skb_queue_head(head, skb);
8923 spin_unlock_irq(&head->lock);
8930 return __io_sqe_files_scm(ctx, 1, index);
8936 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8937 struct io_rsrc_node *node, void *rsrc)
8939 u64 *tag_slot = io_get_tag_slot(data, idx);
8940 struct io_rsrc_put *prsrc;
8942 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8946 prsrc->tag = *tag_slot;
8949 list_add(&prsrc->list, &node->rsrc_list);
8953 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8954 unsigned int issue_flags, u32 slot_index)
8956 struct io_ring_ctx *ctx = req->ctx;
8957 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
8958 bool needs_switch = false;
8959 struct io_fixed_file *file_slot;
8962 io_ring_submit_lock(ctx, needs_lock);
8963 if (file->f_op == &io_uring_fops)
8966 if (!ctx->file_data)
8969 if (slot_index >= ctx->nr_user_files)
8972 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8973 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8975 if (file_slot->file_ptr) {
8976 struct file *old_file;
8978 ret = io_rsrc_node_switch_start(ctx);
8982 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8983 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8984 ctx->rsrc_node, old_file);
8987 file_slot->file_ptr = 0;
8988 needs_switch = true;
8991 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8992 io_fixed_file_set(file_slot, file);
8993 ret = io_sqe_file_register(ctx, file, slot_index);
8995 file_slot->file_ptr = 0;
9002 io_rsrc_node_switch(ctx, ctx->file_data);
9003 io_ring_submit_unlock(ctx, needs_lock);
9009 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
9011 unsigned int offset = req->close.file_slot - 1;
9012 struct io_ring_ctx *ctx = req->ctx;
9013 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
9014 struct io_fixed_file *file_slot;
9018 io_ring_submit_lock(ctx, needs_lock);
9020 if (unlikely(!ctx->file_data))
9023 if (offset >= ctx->nr_user_files)
9025 ret = io_rsrc_node_switch_start(ctx);
9029 offset = array_index_nospec(offset, ctx->nr_user_files);
9030 file_slot = io_fixed_file_slot(&ctx->file_table, offset);
9032 if (!file_slot->file_ptr)
9035 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
9036 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
9040 file_slot->file_ptr = 0;
9041 io_rsrc_node_switch(ctx, ctx->file_data);
9044 io_ring_submit_unlock(ctx, needs_lock);
9048 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
9049 struct io_uring_rsrc_update2 *up,
9052 u64 __user *tags = u64_to_user_ptr(up->tags);
9053 __s32 __user *fds = u64_to_user_ptr(up->data);
9054 struct io_rsrc_data *data = ctx->file_data;
9055 struct io_fixed_file *file_slot;
9059 bool needs_switch = false;
9061 if (!ctx->file_data)
9063 if (up->offset + nr_args > ctx->nr_user_files)
9066 for (done = 0; done < nr_args; done++) {
9069 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
9070 copy_from_user(&fd, &fds[done], sizeof(fd))) {
9074 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
9078 if (fd == IORING_REGISTER_FILES_SKIP)
9081 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
9082 file_slot = io_fixed_file_slot(&ctx->file_table, i);
9084 if (file_slot->file_ptr) {
9085 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
9086 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
9089 file_slot->file_ptr = 0;
9090 needs_switch = true;
9099 * Don't allow io_uring instances to be registered. If
9100 * UNIX isn't enabled, then this causes a reference
9101 * cycle and this instance can never get freed. If UNIX
9102 * is enabled we'll handle it just fine, but there's
9103 * still no point in allowing a ring fd as it doesn't
9104 * support regular read/write anyway.
9106 if (file->f_op == &io_uring_fops) {
9111 *io_get_tag_slot(data, i) = tag;
9112 io_fixed_file_set(file_slot, file);
9113 err = io_sqe_file_register(ctx, file, i);
9115 file_slot->file_ptr = 0;
9123 io_rsrc_node_switch(ctx, data);
9124 return done ? done : err;
9127 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
9128 struct task_struct *task)
9130 struct io_wq_hash *hash;
9131 struct io_wq_data data;
9132 unsigned int concurrency;
9134 mutex_lock(&ctx->uring_lock);
9135 hash = ctx->hash_map;
9137 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
9139 mutex_unlock(&ctx->uring_lock);
9140 return ERR_PTR(-ENOMEM);
9142 refcount_set(&hash->refs, 1);
9143 init_waitqueue_head(&hash->wait);
9144 ctx->hash_map = hash;
9146 mutex_unlock(&ctx->uring_lock);
9150 data.free_work = io_wq_free_work;
9151 data.do_work = io_wq_submit_work;
9153 /* Do QD, or 4 * CPUS, whatever is smallest */
9154 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
9156 return io_wq_create(concurrency, &data);
9159 static __cold int io_uring_alloc_task_context(struct task_struct *task,
9160 struct io_ring_ctx *ctx)
9162 struct io_uring_task *tctx;
9165 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
9166 if (unlikely(!tctx))
9169 tctx->registered_rings = kcalloc(IO_RINGFD_REG_MAX,
9170 sizeof(struct file *), GFP_KERNEL);
9171 if (unlikely(!tctx->registered_rings)) {
9176 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
9177 if (unlikely(ret)) {
9178 kfree(tctx->registered_rings);
9183 tctx->io_wq = io_init_wq_offload(ctx, task);
9184 if (IS_ERR(tctx->io_wq)) {
9185 ret = PTR_ERR(tctx->io_wq);
9186 percpu_counter_destroy(&tctx->inflight);
9187 kfree(tctx->registered_rings);
9193 init_waitqueue_head(&tctx->wait);
9194 atomic_set(&tctx->in_idle, 0);
9195 task->io_uring = tctx;
9196 spin_lock_init(&tctx->task_lock);
9197 INIT_WQ_LIST(&tctx->task_list);
9198 INIT_WQ_LIST(&tctx->prior_task_list);
9199 init_task_work(&tctx->task_work, tctx_task_work);
9203 void __io_uring_free(struct task_struct *tsk)
9205 struct io_uring_task *tctx = tsk->io_uring;
9207 WARN_ON_ONCE(!xa_empty(&tctx->xa));
9208 WARN_ON_ONCE(tctx->io_wq);
9209 WARN_ON_ONCE(tctx->cached_refs);
9211 kfree(tctx->registered_rings);
9212 percpu_counter_destroy(&tctx->inflight);
9214 tsk->io_uring = NULL;
9217 static __cold int io_sq_offload_create(struct io_ring_ctx *ctx,
9218 struct io_uring_params *p)
9222 /* Retain compatibility with failing for an invalid attach attempt */
9223 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
9224 IORING_SETUP_ATTACH_WQ) {
9227 f = fdget(p->wq_fd);
9230 if (f.file->f_op != &io_uring_fops) {
9236 if (ctx->flags & IORING_SETUP_SQPOLL) {
9237 struct task_struct *tsk;
9238 struct io_sq_data *sqd;
9241 ret = security_uring_sqpoll();
9245 sqd = io_get_sq_data(p, &attached);
9251 ctx->sq_creds = get_current_cred();
9253 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
9254 if (!ctx->sq_thread_idle)
9255 ctx->sq_thread_idle = HZ;
9257 io_sq_thread_park(sqd);
9258 list_add(&ctx->sqd_list, &sqd->ctx_list);
9259 io_sqd_update_thread_idle(sqd);
9260 /* don't attach to a dying SQPOLL thread, would be racy */
9261 ret = (attached && !sqd->thread) ? -ENXIO : 0;
9262 io_sq_thread_unpark(sqd);
9269 if (p->flags & IORING_SETUP_SQ_AFF) {
9270 int cpu = p->sq_thread_cpu;
9273 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
9280 sqd->task_pid = current->pid;
9281 sqd->task_tgid = current->tgid;
9282 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
9289 ret = io_uring_alloc_task_context(tsk, ctx);
9290 wake_up_new_task(tsk);
9293 } else if (p->flags & IORING_SETUP_SQ_AFF) {
9294 /* Can't have SQ_AFF without SQPOLL */
9301 complete(&ctx->sq_data->exited);
9303 io_sq_thread_finish(ctx);
9307 static inline void __io_unaccount_mem(struct user_struct *user,
9308 unsigned long nr_pages)
9310 atomic_long_sub(nr_pages, &user->locked_vm);
9313 static inline int __io_account_mem(struct user_struct *user,
9314 unsigned long nr_pages)
9316 unsigned long page_limit, cur_pages, new_pages;
9318 /* Don't allow more pages than we can safely lock */
9319 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
9322 cur_pages = atomic_long_read(&user->locked_vm);
9323 new_pages = cur_pages + nr_pages;
9324 if (new_pages > page_limit)
9326 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
9327 new_pages) != cur_pages);
9332 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
9335 __io_unaccount_mem(ctx->user, nr_pages);
9337 if (ctx->mm_account)
9338 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
9341 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
9346 ret = __io_account_mem(ctx->user, nr_pages);
9351 if (ctx->mm_account)
9352 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
9357 static void io_mem_free(void *ptr)
9364 page = virt_to_head_page(ptr);
9365 if (put_page_testzero(page))
9366 free_compound_page(page);
9369 static void *io_mem_alloc(size_t size)
9371 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
9373 return (void *) __get_free_pages(gfp, get_order(size));
9376 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
9379 struct io_rings *rings;
9380 size_t off, sq_array_size;
9382 off = struct_size(rings, cqes, cq_entries);
9383 if (off == SIZE_MAX)
9387 off = ALIGN(off, SMP_CACHE_BYTES);
9395 sq_array_size = array_size(sizeof(u32), sq_entries);
9396 if (sq_array_size == SIZE_MAX)
9399 if (check_add_overflow(off, sq_array_size, &off))
9405 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
9407 struct io_mapped_ubuf *imu = *slot;
9410 if (imu != ctx->dummy_ubuf) {
9411 for (i = 0; i < imu->nr_bvecs; i++)
9412 unpin_user_page(imu->bvec[i].bv_page);
9413 if (imu->acct_pages)
9414 io_unaccount_mem(ctx, imu->acct_pages);
9420 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
9422 io_buffer_unmap(ctx, &prsrc->buf);
9426 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
9430 for (i = 0; i < ctx->nr_user_bufs; i++)
9431 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
9432 kfree(ctx->user_bufs);
9433 io_rsrc_data_free(ctx->buf_data);
9434 ctx->user_bufs = NULL;
9435 ctx->buf_data = NULL;
9436 ctx->nr_user_bufs = 0;
9439 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
9446 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
9448 __io_sqe_buffers_unregister(ctx);
9452 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
9453 void __user *arg, unsigned index)
9455 struct iovec __user *src;
9457 #ifdef CONFIG_COMPAT
9459 struct compat_iovec __user *ciovs;
9460 struct compat_iovec ciov;
9462 ciovs = (struct compat_iovec __user *) arg;
9463 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
9466 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
9467 dst->iov_len = ciov.iov_len;
9471 src = (struct iovec __user *) arg;
9472 if (copy_from_user(dst, &src[index], sizeof(*dst)))
9478 * Not super efficient, but this is just a registration time. And we do cache
9479 * the last compound head, so generally we'll only do a full search if we don't
9482 * We check if the given compound head page has already been accounted, to
9483 * avoid double accounting it. This allows us to account the full size of the
9484 * page, not just the constituent pages of a huge page.
9486 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
9487 int nr_pages, struct page *hpage)
9491 /* check current page array */
9492 for (i = 0; i < nr_pages; i++) {
9493 if (!PageCompound(pages[i]))
9495 if (compound_head(pages[i]) == hpage)
9499 /* check previously registered pages */
9500 for (i = 0; i < ctx->nr_user_bufs; i++) {
9501 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
9503 for (j = 0; j < imu->nr_bvecs; j++) {
9504 if (!PageCompound(imu->bvec[j].bv_page))
9506 if (compound_head(imu->bvec[j].bv_page) == hpage)
9514 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
9515 int nr_pages, struct io_mapped_ubuf *imu,
9516 struct page **last_hpage)
9520 imu->acct_pages = 0;
9521 for (i = 0; i < nr_pages; i++) {
9522 if (!PageCompound(pages[i])) {
9527 hpage = compound_head(pages[i]);
9528 if (hpage == *last_hpage)
9530 *last_hpage = hpage;
9531 if (headpage_already_acct(ctx, pages, i, hpage))
9533 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
9537 if (!imu->acct_pages)
9540 ret = io_account_mem(ctx, imu->acct_pages);
9542 imu->acct_pages = 0;
9546 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
9547 struct io_mapped_ubuf **pimu,
9548 struct page **last_hpage)
9550 struct io_mapped_ubuf *imu = NULL;
9551 struct vm_area_struct **vmas = NULL;
9552 struct page **pages = NULL;
9553 unsigned long off, start, end, ubuf;
9555 int ret, pret, nr_pages, i;
9557 if (!iov->iov_base) {
9558 *pimu = ctx->dummy_ubuf;
9562 ubuf = (unsigned long) iov->iov_base;
9563 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
9564 start = ubuf >> PAGE_SHIFT;
9565 nr_pages = end - start;
9570 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
9574 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
9579 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
9584 mmap_read_lock(current->mm);
9585 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
9587 if (pret == nr_pages) {
9588 /* don't support file backed memory */
9589 for (i = 0; i < nr_pages; i++) {
9590 struct vm_area_struct *vma = vmas[i];
9592 if (vma_is_shmem(vma))
9595 !is_file_hugepages(vma->vm_file)) {
9601 ret = pret < 0 ? pret : -EFAULT;
9603 mmap_read_unlock(current->mm);
9606 * if we did partial map, or found file backed vmas,
9607 * release any pages we did get
9610 unpin_user_pages(pages, pret);
9614 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9616 unpin_user_pages(pages, pret);
9620 off = ubuf & ~PAGE_MASK;
9621 size = iov->iov_len;
9622 for (i = 0; i < nr_pages; i++) {
9625 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9626 imu->bvec[i].bv_page = pages[i];
9627 imu->bvec[i].bv_len = vec_len;
9628 imu->bvec[i].bv_offset = off;
9632 /* store original address for later verification */
9634 imu->ubuf_end = ubuf + iov->iov_len;
9635 imu->nr_bvecs = nr_pages;
9646 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9648 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9649 return ctx->user_bufs ? 0 : -ENOMEM;
9652 static int io_buffer_validate(struct iovec *iov)
9654 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9657 * Don't impose further limits on the size and buffer
9658 * constraints here, we'll -EINVAL later when IO is
9659 * submitted if they are wrong.
9662 return iov->iov_len ? -EFAULT : 0;
9666 /* arbitrary limit, but we need something */
9667 if (iov->iov_len > SZ_1G)
9670 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9676 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9677 unsigned int nr_args, u64 __user *tags)
9679 struct page *last_hpage = NULL;
9680 struct io_rsrc_data *data;
9686 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9688 ret = io_rsrc_node_switch_start(ctx);
9691 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9694 ret = io_buffers_map_alloc(ctx, nr_args);
9696 io_rsrc_data_free(data);
9700 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9701 ret = io_copy_iov(ctx, &iov, arg, i);
9704 ret = io_buffer_validate(&iov);
9707 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9712 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9718 WARN_ON_ONCE(ctx->buf_data);
9720 ctx->buf_data = data;
9722 __io_sqe_buffers_unregister(ctx);
9724 io_rsrc_node_switch(ctx, NULL);
9728 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9729 struct io_uring_rsrc_update2 *up,
9730 unsigned int nr_args)
9732 u64 __user *tags = u64_to_user_ptr(up->tags);
9733 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9734 struct page *last_hpage = NULL;
9735 bool needs_switch = false;
9741 if (up->offset + nr_args > ctx->nr_user_bufs)
9744 for (done = 0; done < nr_args; done++) {
9745 struct io_mapped_ubuf *imu;
9746 int offset = up->offset + done;
9749 err = io_copy_iov(ctx, &iov, iovs, done);
9752 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9756 err = io_buffer_validate(&iov);
9759 if (!iov.iov_base && tag) {
9763 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9767 i = array_index_nospec(offset, ctx->nr_user_bufs);
9768 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9769 err = io_queue_rsrc_removal(ctx->buf_data, i,
9770 ctx->rsrc_node, ctx->user_bufs[i]);
9771 if (unlikely(err)) {
9772 io_buffer_unmap(ctx, &imu);
9775 ctx->user_bufs[i] = NULL;
9776 needs_switch = true;
9779 ctx->user_bufs[i] = imu;
9780 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9784 io_rsrc_node_switch(ctx, ctx->buf_data);
9785 return done ? done : err;
9788 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
9789 unsigned int eventfd_async)
9791 struct io_ev_fd *ev_fd;
9792 __s32 __user *fds = arg;
9795 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
9796 lockdep_is_held(&ctx->uring_lock));
9800 if (copy_from_user(&fd, fds, sizeof(*fds)))
9803 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
9807 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
9808 if (IS_ERR(ev_fd->cq_ev_fd)) {
9809 int ret = PTR_ERR(ev_fd->cq_ev_fd);
9813 ev_fd->eventfd_async = eventfd_async;
9814 ctx->has_evfd = true;
9815 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
9819 static void io_eventfd_put(struct rcu_head *rcu)
9821 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
9823 eventfd_ctx_put(ev_fd->cq_ev_fd);
9827 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9829 struct io_ev_fd *ev_fd;
9831 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
9832 lockdep_is_held(&ctx->uring_lock));
9834 ctx->has_evfd = false;
9835 rcu_assign_pointer(ctx->io_ev_fd, NULL);
9836 call_rcu(&ev_fd->rcu, io_eventfd_put);
9843 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9847 for (i = 0; i < (1U << IO_BUFFERS_HASH_BITS); i++) {
9848 struct list_head *list = &ctx->io_buffers[i];
9850 while (!list_empty(list)) {
9851 struct io_buffer_list *bl;
9853 bl = list_first_entry(list, struct io_buffer_list, list);
9854 __io_remove_buffers(ctx, bl, -1U);
9855 list_del(&bl->list);
9860 while (!list_empty(&ctx->io_buffers_pages)) {
9863 page = list_first_entry(&ctx->io_buffers_pages, struct page, lru);
9864 list_del_init(&page->lru);
9869 static void io_req_caches_free(struct io_ring_ctx *ctx)
9871 struct io_submit_state *state = &ctx->submit_state;
9874 mutex_lock(&ctx->uring_lock);
9875 io_flush_cached_locked_reqs(ctx, state);
9877 while (state->free_list.next) {
9878 struct io_wq_work_node *node;
9879 struct io_kiocb *req;
9881 node = wq_stack_extract(&state->free_list);
9882 req = container_of(node, struct io_kiocb, comp_list);
9883 kmem_cache_free(req_cachep, req);
9887 percpu_ref_put_many(&ctx->refs, nr);
9888 mutex_unlock(&ctx->uring_lock);
9891 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9893 if (data && !atomic_dec_and_test(&data->refs))
9894 wait_for_completion(&data->done);
9897 static void io_flush_apoll_cache(struct io_ring_ctx *ctx)
9899 struct async_poll *apoll;
9901 while (!list_empty(&ctx->apoll_cache)) {
9902 apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
9904 list_del(&apoll->poll.wait.entry);
9909 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
9911 io_sq_thread_finish(ctx);
9913 if (ctx->mm_account) {
9914 mmdrop(ctx->mm_account);
9915 ctx->mm_account = NULL;
9918 io_rsrc_refs_drop(ctx);
9919 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9920 io_wait_rsrc_data(ctx->buf_data);
9921 io_wait_rsrc_data(ctx->file_data);
9923 mutex_lock(&ctx->uring_lock);
9925 __io_sqe_buffers_unregister(ctx);
9927 __io_sqe_files_unregister(ctx);
9929 __io_cqring_overflow_flush(ctx, true);
9930 io_eventfd_unregister(ctx);
9931 io_flush_apoll_cache(ctx);
9932 mutex_unlock(&ctx->uring_lock);
9933 io_destroy_buffers(ctx);
9935 put_cred(ctx->sq_creds);
9937 /* there are no registered resources left, nobody uses it */
9939 io_rsrc_node_destroy(ctx->rsrc_node);
9940 if (ctx->rsrc_backup_node)
9941 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9942 flush_delayed_work(&ctx->rsrc_put_work);
9943 flush_delayed_work(&ctx->fallback_work);
9945 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9946 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9948 #if defined(CONFIG_UNIX)
9949 if (ctx->ring_sock) {
9950 ctx->ring_sock->file = NULL; /* so that iput() is called */
9951 sock_release(ctx->ring_sock);
9954 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9956 io_mem_free(ctx->rings);
9957 io_mem_free(ctx->sq_sqes);
9959 percpu_ref_exit(&ctx->refs);
9960 free_uid(ctx->user);
9961 io_req_caches_free(ctx);
9963 io_wq_put_hash(ctx->hash_map);
9964 kfree(ctx->cancel_hash);
9965 kfree(ctx->dummy_ubuf);
9966 kfree(ctx->io_buffers);
9970 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9972 struct io_ring_ctx *ctx = file->private_data;
9975 poll_wait(file, &ctx->cq_wait, wait);
9977 * synchronizes with barrier from wq_has_sleeper call in
9981 if (!io_sqring_full(ctx))
9982 mask |= EPOLLOUT | EPOLLWRNORM;
9985 * Don't flush cqring overflow list here, just do a simple check.
9986 * Otherwise there could possible be ABBA deadlock:
9989 * lock(&ctx->uring_lock);
9991 * lock(&ctx->uring_lock);
9994 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9995 * pushs them to do the flush.
9997 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9998 mask |= EPOLLIN | EPOLLRDNORM;
10003 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
10005 const struct cred *creds;
10007 creds = xa_erase(&ctx->personalities, id);
10016 struct io_tctx_exit {
10017 struct callback_head task_work;
10018 struct completion completion;
10019 struct io_ring_ctx *ctx;
10022 static __cold void io_tctx_exit_cb(struct callback_head *cb)
10024 struct io_uring_task *tctx = current->io_uring;
10025 struct io_tctx_exit *work;
10027 work = container_of(cb, struct io_tctx_exit, task_work);
10029 * When @in_idle, we're in cancellation and it's racy to remove the
10030 * node. It'll be removed by the end of cancellation, just ignore it.
10032 if (!atomic_read(&tctx->in_idle))
10033 io_uring_del_tctx_node((unsigned long)work->ctx);
10034 complete(&work->completion);
10037 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
10039 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
10041 return req->ctx == data;
10044 static __cold void io_ring_exit_work(struct work_struct *work)
10046 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
10047 unsigned long timeout = jiffies + HZ * 60 * 5;
10048 unsigned long interval = HZ / 20;
10049 struct io_tctx_exit exit;
10050 struct io_tctx_node *node;
10054 * If we're doing polled IO and end up having requests being
10055 * submitted async (out-of-line), then completions can come in while
10056 * we're waiting for refs to drop. We need to reap these manually,
10057 * as nobody else will be looking for them.
10060 io_uring_try_cancel_requests(ctx, NULL, true);
10061 if (ctx->sq_data) {
10062 struct io_sq_data *sqd = ctx->sq_data;
10063 struct task_struct *tsk;
10065 io_sq_thread_park(sqd);
10067 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
10068 io_wq_cancel_cb(tsk->io_uring->io_wq,
10069 io_cancel_ctx_cb, ctx, true);
10070 io_sq_thread_unpark(sqd);
10073 io_req_caches_free(ctx);
10075 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
10076 /* there is little hope left, don't run it too often */
10077 interval = HZ * 60;
10079 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
10081 init_completion(&exit.completion);
10082 init_task_work(&exit.task_work, io_tctx_exit_cb);
10085 * Some may use context even when all refs and requests have been put,
10086 * and they are free to do so while still holding uring_lock or
10087 * completion_lock, see io_req_task_submit(). Apart from other work,
10088 * this lock/unlock section also waits them to finish.
10090 mutex_lock(&ctx->uring_lock);
10091 while (!list_empty(&ctx->tctx_list)) {
10092 WARN_ON_ONCE(time_after(jiffies, timeout));
10094 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
10096 /* don't spin on a single task if cancellation failed */
10097 list_rotate_left(&ctx->tctx_list);
10098 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
10099 if (WARN_ON_ONCE(ret))
10102 mutex_unlock(&ctx->uring_lock);
10103 wait_for_completion(&exit.completion);
10104 mutex_lock(&ctx->uring_lock);
10106 mutex_unlock(&ctx->uring_lock);
10107 spin_lock(&ctx->completion_lock);
10108 spin_unlock(&ctx->completion_lock);
10110 io_ring_ctx_free(ctx);
10113 /* Returns true if we found and killed one or more timeouts */
10114 static __cold bool io_kill_timeouts(struct io_ring_ctx *ctx,
10115 struct task_struct *tsk, bool cancel_all)
10117 struct io_kiocb *req, *tmp;
10120 spin_lock(&ctx->completion_lock);
10121 spin_lock_irq(&ctx->timeout_lock);
10122 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
10123 if (io_match_task(req, tsk, cancel_all)) {
10124 io_kill_timeout(req, -ECANCELED);
10128 spin_unlock_irq(&ctx->timeout_lock);
10130 io_commit_cqring(ctx);
10131 spin_unlock(&ctx->completion_lock);
10133 io_cqring_ev_posted(ctx);
10134 return canceled != 0;
10137 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
10139 unsigned long index;
10140 struct creds *creds;
10142 mutex_lock(&ctx->uring_lock);
10143 percpu_ref_kill(&ctx->refs);
10145 __io_cqring_overflow_flush(ctx, true);
10146 xa_for_each(&ctx->personalities, index, creds)
10147 io_unregister_personality(ctx, index);
10148 mutex_unlock(&ctx->uring_lock);
10150 io_kill_timeouts(ctx, NULL, true);
10151 io_poll_remove_all(ctx, NULL, true);
10153 /* if we failed setting up the ctx, we might not have any rings */
10154 io_iopoll_try_reap_events(ctx);
10156 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
10158 * Use system_unbound_wq to avoid spawning tons of event kworkers
10159 * if we're exiting a ton of rings at the same time. It just adds
10160 * noise and overhead, there's no discernable change in runtime
10161 * over using system_wq.
10163 queue_work(system_unbound_wq, &ctx->exit_work);
10166 static int io_uring_release(struct inode *inode, struct file *file)
10168 struct io_ring_ctx *ctx = file->private_data;
10170 file->private_data = NULL;
10171 io_ring_ctx_wait_and_kill(ctx);
10175 struct io_task_cancel {
10176 struct task_struct *task;
10180 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
10182 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
10183 struct io_task_cancel *cancel = data;
10185 return io_match_task_safe(req, cancel->task, cancel->all);
10188 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
10189 struct task_struct *task,
10192 struct io_defer_entry *de;
10195 spin_lock(&ctx->completion_lock);
10196 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
10197 if (io_match_task_safe(de->req, task, cancel_all)) {
10198 list_cut_position(&list, &ctx->defer_list, &de->list);
10202 spin_unlock(&ctx->completion_lock);
10203 if (list_empty(&list))
10206 while (!list_empty(&list)) {
10207 de = list_first_entry(&list, struct io_defer_entry, list);
10208 list_del_init(&de->list);
10209 io_req_complete_failed(de->req, -ECANCELED);
10215 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
10217 struct io_tctx_node *node;
10218 enum io_wq_cancel cret;
10221 mutex_lock(&ctx->uring_lock);
10222 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
10223 struct io_uring_task *tctx = node->task->io_uring;
10226 * io_wq will stay alive while we hold uring_lock, because it's
10227 * killed after ctx nodes, which requires to take the lock.
10229 if (!tctx || !tctx->io_wq)
10231 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
10232 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
10234 mutex_unlock(&ctx->uring_lock);
10239 static __cold void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
10240 struct task_struct *task,
10243 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
10244 struct io_uring_task *tctx = task ? task->io_uring : NULL;
10247 enum io_wq_cancel cret;
10251 ret |= io_uring_try_cancel_iowq(ctx);
10252 } else if (tctx && tctx->io_wq) {
10254 * Cancels requests of all rings, not only @ctx, but
10255 * it's fine as the task is in exit/exec.
10257 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
10259 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
10262 /* SQPOLL thread does its own polling */
10263 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
10264 (ctx->sq_data && ctx->sq_data->thread == current)) {
10265 while (!wq_list_empty(&ctx->iopoll_list)) {
10266 io_iopoll_try_reap_events(ctx);
10271 ret |= io_cancel_defer_files(ctx, task, cancel_all);
10272 ret |= io_poll_remove_all(ctx, task, cancel_all);
10273 ret |= io_kill_timeouts(ctx, task, cancel_all);
10275 ret |= io_run_task_work();
10282 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
10284 struct io_uring_task *tctx = current->io_uring;
10285 struct io_tctx_node *node;
10288 if (unlikely(!tctx)) {
10289 ret = io_uring_alloc_task_context(current, ctx);
10293 tctx = current->io_uring;
10294 if (ctx->iowq_limits_set) {
10295 unsigned int limits[2] = { ctx->iowq_limits[0],
10296 ctx->iowq_limits[1], };
10298 ret = io_wq_max_workers(tctx->io_wq, limits);
10303 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
10304 node = kmalloc(sizeof(*node), GFP_KERNEL);
10308 node->task = current;
10310 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
10311 node, GFP_KERNEL));
10317 mutex_lock(&ctx->uring_lock);
10318 list_add(&node->ctx_node, &ctx->tctx_list);
10319 mutex_unlock(&ctx->uring_lock);
10326 * Note that this task has used io_uring. We use it for cancelation purposes.
10328 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
10330 struct io_uring_task *tctx = current->io_uring;
10332 if (likely(tctx && tctx->last == ctx))
10334 return __io_uring_add_tctx_node(ctx);
10338 * Remove this io_uring_file -> task mapping.
10340 static __cold void io_uring_del_tctx_node(unsigned long index)
10342 struct io_uring_task *tctx = current->io_uring;
10343 struct io_tctx_node *node;
10347 node = xa_erase(&tctx->xa, index);
10351 WARN_ON_ONCE(current != node->task);
10352 WARN_ON_ONCE(list_empty(&node->ctx_node));
10354 mutex_lock(&node->ctx->uring_lock);
10355 list_del(&node->ctx_node);
10356 mutex_unlock(&node->ctx->uring_lock);
10358 if (tctx->last == node->ctx)
10363 static __cold void io_uring_clean_tctx(struct io_uring_task *tctx)
10365 struct io_wq *wq = tctx->io_wq;
10366 struct io_tctx_node *node;
10367 unsigned long index;
10369 xa_for_each(&tctx->xa, index, node) {
10370 io_uring_del_tctx_node(index);
10375 * Must be after io_uring_del_tctx_node() (removes nodes under
10376 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
10378 io_wq_put_and_exit(wq);
10379 tctx->io_wq = NULL;
10383 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
10387 return percpu_counter_sum(&tctx->inflight);
10391 * Find any io_uring ctx that this task has registered or done IO on, and cancel
10392 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
10394 static __cold void io_uring_cancel_generic(bool cancel_all,
10395 struct io_sq_data *sqd)
10397 struct io_uring_task *tctx = current->io_uring;
10398 struct io_ring_ctx *ctx;
10402 WARN_ON_ONCE(sqd && sqd->thread != current);
10404 if (!current->io_uring)
10407 io_wq_exit_start(tctx->io_wq);
10409 atomic_inc(&tctx->in_idle);
10411 io_uring_drop_tctx_refs(current);
10412 /* read completions before cancelations */
10413 inflight = tctx_inflight(tctx, !cancel_all);
10418 struct io_tctx_node *node;
10419 unsigned long index;
10421 xa_for_each(&tctx->xa, index, node) {
10422 /* sqpoll task will cancel all its requests */
10423 if (node->ctx->sq_data)
10425 io_uring_try_cancel_requests(node->ctx, current,
10429 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
10430 io_uring_try_cancel_requests(ctx, current,
10434 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
10435 io_run_task_work();
10436 io_uring_drop_tctx_refs(current);
10439 * If we've seen completions, retry without waiting. This
10440 * avoids a race where a completion comes in before we did
10441 * prepare_to_wait().
10443 if (inflight == tctx_inflight(tctx, !cancel_all))
10445 finish_wait(&tctx->wait, &wait);
10448 io_uring_clean_tctx(tctx);
10451 * We shouldn't run task_works after cancel, so just leave
10452 * ->in_idle set for normal exit.
10454 atomic_dec(&tctx->in_idle);
10455 /* for exec all current's requests should be gone, kill tctx */
10456 __io_uring_free(current);
10460 void __io_uring_cancel(bool cancel_all)
10462 io_uring_cancel_generic(cancel_all, NULL);
10465 void io_uring_unreg_ringfd(void)
10467 struct io_uring_task *tctx = current->io_uring;
10470 for (i = 0; i < IO_RINGFD_REG_MAX; i++) {
10471 if (tctx->registered_rings[i]) {
10472 fput(tctx->registered_rings[i]);
10473 tctx->registered_rings[i] = NULL;
10478 static int io_ring_add_registered_fd(struct io_uring_task *tctx, int fd,
10479 int start, int end)
10484 for (offset = start; offset < end; offset++) {
10485 offset = array_index_nospec(offset, IO_RINGFD_REG_MAX);
10486 if (tctx->registered_rings[offset])
10492 } else if (file->f_op != &io_uring_fops) {
10494 return -EOPNOTSUPP;
10496 tctx->registered_rings[offset] = file;
10504 * Register a ring fd to avoid fdget/fdput for each io_uring_enter()
10505 * invocation. User passes in an array of struct io_uring_rsrc_update
10506 * with ->data set to the ring_fd, and ->offset given for the desired
10507 * index. If no index is desired, application may set ->offset == -1U
10508 * and we'll find an available index. Returns number of entries
10509 * successfully processed, or < 0 on error if none were processed.
10511 static int io_ringfd_register(struct io_ring_ctx *ctx, void __user *__arg,
10514 struct io_uring_rsrc_update __user *arg = __arg;
10515 struct io_uring_rsrc_update reg;
10516 struct io_uring_task *tctx;
10519 if (!nr_args || nr_args > IO_RINGFD_REG_MAX)
10522 mutex_unlock(&ctx->uring_lock);
10523 ret = io_uring_add_tctx_node(ctx);
10524 mutex_lock(&ctx->uring_lock);
10528 tctx = current->io_uring;
10529 for (i = 0; i < nr_args; i++) {
10532 if (copy_from_user(®, &arg[i], sizeof(reg))) {
10542 if (reg.offset == -1U) {
10544 end = IO_RINGFD_REG_MAX;
10546 if (reg.offset >= IO_RINGFD_REG_MAX) {
10550 start = reg.offset;
10554 ret = io_ring_add_registered_fd(tctx, reg.data, start, end);
10559 if (copy_to_user(&arg[i], ®, sizeof(reg))) {
10560 fput(tctx->registered_rings[reg.offset]);
10561 tctx->registered_rings[reg.offset] = NULL;
10567 return i ? i : ret;
10570 static int io_ringfd_unregister(struct io_ring_ctx *ctx, void __user *__arg,
10573 struct io_uring_rsrc_update __user *arg = __arg;
10574 struct io_uring_task *tctx = current->io_uring;
10575 struct io_uring_rsrc_update reg;
10578 if (!nr_args || nr_args > IO_RINGFD_REG_MAX)
10583 for (i = 0; i < nr_args; i++) {
10584 if (copy_from_user(®, &arg[i], sizeof(reg))) {
10588 if (reg.resv || reg.offset >= IO_RINGFD_REG_MAX) {
10593 reg.offset = array_index_nospec(reg.offset, IO_RINGFD_REG_MAX);
10594 if (tctx->registered_rings[reg.offset]) {
10595 fput(tctx->registered_rings[reg.offset]);
10596 tctx->registered_rings[reg.offset] = NULL;
10600 return i ? i : ret;
10603 static void *io_uring_validate_mmap_request(struct file *file,
10604 loff_t pgoff, size_t sz)
10606 struct io_ring_ctx *ctx = file->private_data;
10607 loff_t offset = pgoff << PAGE_SHIFT;
10612 case IORING_OFF_SQ_RING:
10613 case IORING_OFF_CQ_RING:
10616 case IORING_OFF_SQES:
10617 ptr = ctx->sq_sqes;
10620 return ERR_PTR(-EINVAL);
10623 page = virt_to_head_page(ptr);
10624 if (sz > page_size(page))
10625 return ERR_PTR(-EINVAL);
10632 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
10634 size_t sz = vma->vm_end - vma->vm_start;
10638 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
10640 return PTR_ERR(ptr);
10642 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
10643 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
10646 #else /* !CONFIG_MMU */
10648 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
10650 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
10653 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
10655 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
10658 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
10659 unsigned long addr, unsigned long len,
10660 unsigned long pgoff, unsigned long flags)
10664 ptr = io_uring_validate_mmap_request(file, pgoff, len);
10666 return PTR_ERR(ptr);
10668 return (unsigned long) ptr;
10671 #endif /* !CONFIG_MMU */
10673 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
10678 if (!io_sqring_full(ctx))
10680 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
10682 if (!io_sqring_full(ctx))
10685 } while (!signal_pending(current));
10687 finish_wait(&ctx->sqo_sq_wait, &wait);
10691 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
10692 struct __kernel_timespec __user **ts,
10693 const sigset_t __user **sig)
10695 struct io_uring_getevents_arg arg;
10698 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
10699 * is just a pointer to the sigset_t.
10701 if (!(flags & IORING_ENTER_EXT_ARG)) {
10702 *sig = (const sigset_t __user *) argp;
10708 * EXT_ARG is set - ensure we agree on the size of it and copy in our
10709 * timespec and sigset_t pointers if good.
10711 if (*argsz != sizeof(arg))
10713 if (copy_from_user(&arg, argp, sizeof(arg)))
10717 *sig = u64_to_user_ptr(arg.sigmask);
10718 *argsz = arg.sigmask_sz;
10719 *ts = u64_to_user_ptr(arg.ts);
10723 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
10724 u32, min_complete, u32, flags, const void __user *, argp,
10727 struct io_ring_ctx *ctx;
10732 io_run_task_work();
10734 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
10735 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
10736 IORING_ENTER_REGISTERED_RING)))
10740 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
10741 * need only dereference our task private array to find it.
10743 if (flags & IORING_ENTER_REGISTERED_RING) {
10744 struct io_uring_task *tctx = current->io_uring;
10746 if (!tctx || fd >= IO_RINGFD_REG_MAX)
10748 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
10749 f.file = tctx->registered_rings[fd];
10750 if (unlikely(!f.file))
10754 if (unlikely(!f.file))
10759 if (unlikely(f.file->f_op != &io_uring_fops))
10763 ctx = f.file->private_data;
10764 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
10768 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
10772 * For SQ polling, the thread will do all submissions and completions.
10773 * Just return the requested submit count, and wake the thread if
10774 * we were asked to.
10777 if (ctx->flags & IORING_SETUP_SQPOLL) {
10778 io_cqring_overflow_flush(ctx);
10780 if (unlikely(ctx->sq_data->thread == NULL)) {
10784 if (flags & IORING_ENTER_SQ_WAKEUP)
10785 wake_up(&ctx->sq_data->wait);
10786 if (flags & IORING_ENTER_SQ_WAIT) {
10787 ret = io_sqpoll_wait_sq(ctx);
10791 submitted = to_submit;
10792 } else if (to_submit) {
10793 ret = io_uring_add_tctx_node(ctx);
10796 mutex_lock(&ctx->uring_lock);
10797 submitted = io_submit_sqes(ctx, to_submit);
10798 mutex_unlock(&ctx->uring_lock);
10800 if (submitted != to_submit)
10803 if (flags & IORING_ENTER_GETEVENTS) {
10804 const sigset_t __user *sig;
10805 struct __kernel_timespec __user *ts;
10807 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10811 min_complete = min(min_complete, ctx->cq_entries);
10814 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10815 * space applications don't need to do io completion events
10816 * polling again, they can rely on io_sq_thread to do polling
10817 * work, which can reduce cpu usage and uring_lock contention.
10819 if (ctx->flags & IORING_SETUP_IOPOLL &&
10820 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10821 ret = io_iopoll_check(ctx, min_complete);
10823 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10828 percpu_ref_put(&ctx->refs);
10830 if (!(flags & IORING_ENTER_REGISTERED_RING))
10832 return submitted ? submitted : ret;
10835 #ifdef CONFIG_PROC_FS
10836 static __cold int io_uring_show_cred(struct seq_file *m, unsigned int id,
10837 const struct cred *cred)
10839 struct user_namespace *uns = seq_user_ns(m);
10840 struct group_info *gi;
10845 seq_printf(m, "%5d\n", id);
10846 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10847 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10848 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10849 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10850 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10851 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10852 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10853 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10854 seq_puts(m, "\n\tGroups:\t");
10855 gi = cred->group_info;
10856 for (g = 0; g < gi->ngroups; g++) {
10857 seq_put_decimal_ull(m, g ? " " : "",
10858 from_kgid_munged(uns, gi->gid[g]));
10860 seq_puts(m, "\n\tCapEff:\t");
10861 cap = cred->cap_effective;
10862 CAP_FOR_EACH_U32(__capi)
10863 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10868 static __cold void __io_uring_show_fdinfo(struct io_ring_ctx *ctx,
10869 struct seq_file *m)
10871 struct io_sq_data *sq = NULL;
10872 struct io_overflow_cqe *ocqe;
10873 struct io_rings *r = ctx->rings;
10874 unsigned int sq_mask = ctx->sq_entries - 1, cq_mask = ctx->cq_entries - 1;
10875 unsigned int sq_head = READ_ONCE(r->sq.head);
10876 unsigned int sq_tail = READ_ONCE(r->sq.tail);
10877 unsigned int cq_head = READ_ONCE(r->cq.head);
10878 unsigned int cq_tail = READ_ONCE(r->cq.tail);
10879 unsigned int sq_entries, cq_entries;
10884 * we may get imprecise sqe and cqe info if uring is actively running
10885 * since we get cached_sq_head and cached_cq_tail without uring_lock
10886 * and sq_tail and cq_head are changed by userspace. But it's ok since
10887 * we usually use these info when it is stuck.
10889 seq_printf(m, "SqMask:\t0x%x\n", sq_mask);
10890 seq_printf(m, "SqHead:\t%u\n", sq_head);
10891 seq_printf(m, "SqTail:\t%u\n", sq_tail);
10892 seq_printf(m, "CachedSqHead:\t%u\n", ctx->cached_sq_head);
10893 seq_printf(m, "CqMask:\t0x%x\n", cq_mask);
10894 seq_printf(m, "CqHead:\t%u\n", cq_head);
10895 seq_printf(m, "CqTail:\t%u\n", cq_tail);
10896 seq_printf(m, "CachedCqTail:\t%u\n", ctx->cached_cq_tail);
10897 seq_printf(m, "SQEs:\t%u\n", sq_tail - ctx->cached_sq_head);
10898 sq_entries = min(sq_tail - sq_head, ctx->sq_entries);
10899 for (i = 0; i < sq_entries; i++) {
10900 unsigned int entry = i + sq_head;
10901 unsigned int sq_idx = READ_ONCE(ctx->sq_array[entry & sq_mask]);
10902 struct io_uring_sqe *sqe;
10904 if (sq_idx > sq_mask)
10906 sqe = &ctx->sq_sqes[sq_idx];
10907 seq_printf(m, "%5u: opcode:%d, fd:%d, flags:%x, user_data:%llu\n",
10908 sq_idx, sqe->opcode, sqe->fd, sqe->flags,
10911 seq_printf(m, "CQEs:\t%u\n", cq_tail - cq_head);
10912 cq_entries = min(cq_tail - cq_head, ctx->cq_entries);
10913 for (i = 0; i < cq_entries; i++) {
10914 unsigned int entry = i + cq_head;
10915 struct io_uring_cqe *cqe = &r->cqes[entry & cq_mask];
10917 seq_printf(m, "%5u: user_data:%llu, res:%d, flag:%x\n",
10918 entry & cq_mask, cqe->user_data, cqe->res,
10923 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10924 * since fdinfo case grabs it in the opposite direction of normal use
10925 * cases. If we fail to get the lock, we just don't iterate any
10926 * structures that could be going away outside the io_uring mutex.
10928 has_lock = mutex_trylock(&ctx->uring_lock);
10930 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10936 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10937 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10938 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10939 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10940 struct file *f = io_file_from_index(ctx, i);
10943 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10945 seq_printf(m, "%5u: <none>\n", i);
10947 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10948 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10949 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10950 unsigned int len = buf->ubuf_end - buf->ubuf;
10952 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10954 if (has_lock && !xa_empty(&ctx->personalities)) {
10955 unsigned long index;
10956 const struct cred *cred;
10958 seq_printf(m, "Personalities:\n");
10959 xa_for_each(&ctx->personalities, index, cred)
10960 io_uring_show_cred(m, index, cred);
10963 mutex_unlock(&ctx->uring_lock);
10965 seq_puts(m, "PollList:\n");
10966 spin_lock(&ctx->completion_lock);
10967 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10968 struct hlist_head *list = &ctx->cancel_hash[i];
10969 struct io_kiocb *req;
10971 hlist_for_each_entry(req, list, hash_node)
10972 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10973 task_work_pending(req->task));
10976 seq_puts(m, "CqOverflowList:\n");
10977 list_for_each_entry(ocqe, &ctx->cq_overflow_list, list) {
10978 struct io_uring_cqe *cqe = &ocqe->cqe;
10980 seq_printf(m, " user_data=%llu, res=%d, flags=%x\n",
10981 cqe->user_data, cqe->res, cqe->flags);
10985 spin_unlock(&ctx->completion_lock);
10988 static __cold void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10990 struct io_ring_ctx *ctx = f->private_data;
10992 if (percpu_ref_tryget(&ctx->refs)) {
10993 __io_uring_show_fdinfo(ctx, m);
10994 percpu_ref_put(&ctx->refs);
10999 static const struct file_operations io_uring_fops = {
11000 .release = io_uring_release,
11001 .mmap = io_uring_mmap,
11003 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
11004 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
11006 .poll = io_uring_poll,
11007 #ifdef CONFIG_PROC_FS
11008 .show_fdinfo = io_uring_show_fdinfo,
11012 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
11013 struct io_uring_params *p)
11015 struct io_rings *rings;
11016 size_t size, sq_array_offset;
11018 /* make sure these are sane, as we already accounted them */
11019 ctx->sq_entries = p->sq_entries;
11020 ctx->cq_entries = p->cq_entries;
11022 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
11023 if (size == SIZE_MAX)
11026 rings = io_mem_alloc(size);
11030 ctx->rings = rings;
11031 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
11032 rings->sq_ring_mask = p->sq_entries - 1;
11033 rings->cq_ring_mask = p->cq_entries - 1;
11034 rings->sq_ring_entries = p->sq_entries;
11035 rings->cq_ring_entries = p->cq_entries;
11037 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
11038 if (size == SIZE_MAX) {
11039 io_mem_free(ctx->rings);
11044 ctx->sq_sqes = io_mem_alloc(size);
11045 if (!ctx->sq_sqes) {
11046 io_mem_free(ctx->rings);
11054 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
11058 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
11062 ret = io_uring_add_tctx_node(ctx);
11067 fd_install(fd, file);
11072 * Allocate an anonymous fd, this is what constitutes the application
11073 * visible backing of an io_uring instance. The application mmaps this
11074 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
11075 * we have to tie this fd to a socket for file garbage collection purposes.
11077 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
11080 #if defined(CONFIG_UNIX)
11083 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
11086 return ERR_PTR(ret);
11089 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
11090 O_RDWR | O_CLOEXEC, NULL);
11091 #if defined(CONFIG_UNIX)
11092 if (IS_ERR(file)) {
11093 sock_release(ctx->ring_sock);
11094 ctx->ring_sock = NULL;
11096 ctx->ring_sock->file = file;
11102 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
11103 struct io_uring_params __user *params)
11105 struct io_ring_ctx *ctx;
11111 if (entries > IORING_MAX_ENTRIES) {
11112 if (!(p->flags & IORING_SETUP_CLAMP))
11114 entries = IORING_MAX_ENTRIES;
11118 * Use twice as many entries for the CQ ring. It's possible for the
11119 * application to drive a higher depth than the size of the SQ ring,
11120 * since the sqes are only used at submission time. This allows for
11121 * some flexibility in overcommitting a bit. If the application has
11122 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
11123 * of CQ ring entries manually.
11125 p->sq_entries = roundup_pow_of_two(entries);
11126 if (p->flags & IORING_SETUP_CQSIZE) {
11128 * If IORING_SETUP_CQSIZE is set, we do the same roundup
11129 * to a power-of-two, if it isn't already. We do NOT impose
11130 * any cq vs sq ring sizing.
11132 if (!p->cq_entries)
11134 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
11135 if (!(p->flags & IORING_SETUP_CLAMP))
11137 p->cq_entries = IORING_MAX_CQ_ENTRIES;
11139 p->cq_entries = roundup_pow_of_two(p->cq_entries);
11140 if (p->cq_entries < p->sq_entries)
11143 p->cq_entries = 2 * p->sq_entries;
11146 ctx = io_ring_ctx_alloc(p);
11149 ctx->compat = in_compat_syscall();
11150 if (!capable(CAP_IPC_LOCK))
11151 ctx->user = get_uid(current_user());
11154 * This is just grabbed for accounting purposes. When a process exits,
11155 * the mm is exited and dropped before the files, hence we need to hang
11156 * on to this mm purely for the purposes of being able to unaccount
11157 * memory (locked/pinned vm). It's not used for anything else.
11159 mmgrab(current->mm);
11160 ctx->mm_account = current->mm;
11162 ret = io_allocate_scq_urings(ctx, p);
11166 ret = io_sq_offload_create(ctx, p);
11169 /* always set a rsrc node */
11170 ret = io_rsrc_node_switch_start(ctx);
11173 io_rsrc_node_switch(ctx, NULL);
11175 memset(&p->sq_off, 0, sizeof(p->sq_off));
11176 p->sq_off.head = offsetof(struct io_rings, sq.head);
11177 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
11178 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
11179 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
11180 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
11181 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
11182 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
11184 memset(&p->cq_off, 0, sizeof(p->cq_off));
11185 p->cq_off.head = offsetof(struct io_rings, cq.head);
11186 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
11187 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
11188 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
11189 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
11190 p->cq_off.cqes = offsetof(struct io_rings, cqes);
11191 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
11193 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
11194 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
11195 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
11196 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
11197 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
11198 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
11199 IORING_FEAT_LINKED_FILE;
11201 if (copy_to_user(params, p, sizeof(*p))) {
11206 file = io_uring_get_file(ctx);
11207 if (IS_ERR(file)) {
11208 ret = PTR_ERR(file);
11213 * Install ring fd as the very last thing, so we don't risk someone
11214 * having closed it before we finish setup
11216 ret = io_uring_install_fd(ctx, file);
11218 /* fput will clean it up */
11223 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
11226 io_ring_ctx_wait_and_kill(ctx);
11231 * Sets up an aio uring context, and returns the fd. Applications asks for a
11232 * ring size, we return the actual sq/cq ring sizes (among other things) in the
11233 * params structure passed in.
11235 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
11237 struct io_uring_params p;
11240 if (copy_from_user(&p, params, sizeof(p)))
11242 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
11247 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
11248 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
11249 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
11250 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL))
11253 return io_uring_create(entries, &p, params);
11256 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
11257 struct io_uring_params __user *, params)
11259 return io_uring_setup(entries, params);
11262 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
11265 struct io_uring_probe *p;
11269 size = struct_size(p, ops, nr_args);
11270 if (size == SIZE_MAX)
11272 p = kzalloc(size, GFP_KERNEL);
11277 if (copy_from_user(p, arg, size))
11280 if (memchr_inv(p, 0, size))
11283 p->last_op = IORING_OP_LAST - 1;
11284 if (nr_args > IORING_OP_LAST)
11285 nr_args = IORING_OP_LAST;
11287 for (i = 0; i < nr_args; i++) {
11289 if (!io_op_defs[i].not_supported)
11290 p->ops[i].flags = IO_URING_OP_SUPPORTED;
11295 if (copy_to_user(arg, p, size))
11302 static int io_register_personality(struct io_ring_ctx *ctx)
11304 const struct cred *creds;
11308 creds = get_current_cred();
11310 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
11311 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
11319 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
11320 void __user *arg, unsigned int nr_args)
11322 struct io_uring_restriction *res;
11326 /* Restrictions allowed only if rings started disabled */
11327 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
11330 /* We allow only a single restrictions registration */
11331 if (ctx->restrictions.registered)
11334 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
11337 size = array_size(nr_args, sizeof(*res));
11338 if (size == SIZE_MAX)
11341 res = memdup_user(arg, size);
11343 return PTR_ERR(res);
11347 for (i = 0; i < nr_args; i++) {
11348 switch (res[i].opcode) {
11349 case IORING_RESTRICTION_REGISTER_OP:
11350 if (res[i].register_op >= IORING_REGISTER_LAST) {
11355 __set_bit(res[i].register_op,
11356 ctx->restrictions.register_op);
11358 case IORING_RESTRICTION_SQE_OP:
11359 if (res[i].sqe_op >= IORING_OP_LAST) {
11364 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
11366 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
11367 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
11369 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
11370 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
11379 /* Reset all restrictions if an error happened */
11381 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
11383 ctx->restrictions.registered = true;
11389 static int io_register_enable_rings(struct io_ring_ctx *ctx)
11391 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
11394 if (ctx->restrictions.registered)
11395 ctx->restricted = 1;
11397 ctx->flags &= ~IORING_SETUP_R_DISABLED;
11398 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
11399 wake_up(&ctx->sq_data->wait);
11403 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
11404 struct io_uring_rsrc_update2 *up,
11410 if (check_add_overflow(up->offset, nr_args, &tmp))
11412 err = io_rsrc_node_switch_start(ctx);
11417 case IORING_RSRC_FILE:
11418 return __io_sqe_files_update(ctx, up, nr_args);
11419 case IORING_RSRC_BUFFER:
11420 return __io_sqe_buffers_update(ctx, up, nr_args);
11425 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
11428 struct io_uring_rsrc_update2 up;
11432 memset(&up, 0, sizeof(up));
11433 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
11435 if (up.resv || up.resv2)
11437 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
11440 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
11441 unsigned size, unsigned type)
11443 struct io_uring_rsrc_update2 up;
11445 if (size != sizeof(up))
11447 if (copy_from_user(&up, arg, sizeof(up)))
11449 if (!up.nr || up.resv || up.resv2)
11451 return __io_register_rsrc_update(ctx, type, &up, up.nr);
11454 static __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
11455 unsigned int size, unsigned int type)
11457 struct io_uring_rsrc_register rr;
11459 /* keep it extendible */
11460 if (size != sizeof(rr))
11463 memset(&rr, 0, sizeof(rr));
11464 if (copy_from_user(&rr, arg, size))
11466 if (!rr.nr || rr.resv || rr.resv2)
11470 case IORING_RSRC_FILE:
11471 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
11472 rr.nr, u64_to_user_ptr(rr.tags));
11473 case IORING_RSRC_BUFFER:
11474 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
11475 rr.nr, u64_to_user_ptr(rr.tags));
11480 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
11481 void __user *arg, unsigned len)
11483 struct io_uring_task *tctx = current->io_uring;
11484 cpumask_var_t new_mask;
11487 if (!tctx || !tctx->io_wq)
11490 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
11493 cpumask_clear(new_mask);
11494 if (len > cpumask_size())
11495 len = cpumask_size();
11497 if (in_compat_syscall()) {
11498 ret = compat_get_bitmap(cpumask_bits(new_mask),
11499 (const compat_ulong_t __user *)arg,
11500 len * 8 /* CHAR_BIT */);
11502 ret = copy_from_user(new_mask, arg, len);
11506 free_cpumask_var(new_mask);
11510 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
11511 free_cpumask_var(new_mask);
11515 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
11517 struct io_uring_task *tctx = current->io_uring;
11519 if (!tctx || !tctx->io_wq)
11522 return io_wq_cpu_affinity(tctx->io_wq, NULL);
11525 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
11527 __must_hold(&ctx->uring_lock)
11529 struct io_tctx_node *node;
11530 struct io_uring_task *tctx = NULL;
11531 struct io_sq_data *sqd = NULL;
11532 __u32 new_count[2];
11535 if (copy_from_user(new_count, arg, sizeof(new_count)))
11537 for (i = 0; i < ARRAY_SIZE(new_count); i++)
11538 if (new_count[i] > INT_MAX)
11541 if (ctx->flags & IORING_SETUP_SQPOLL) {
11542 sqd = ctx->sq_data;
11545 * Observe the correct sqd->lock -> ctx->uring_lock
11546 * ordering. Fine to drop uring_lock here, we hold
11547 * a ref to the ctx.
11549 refcount_inc(&sqd->refs);
11550 mutex_unlock(&ctx->uring_lock);
11551 mutex_lock(&sqd->lock);
11552 mutex_lock(&ctx->uring_lock);
11554 tctx = sqd->thread->io_uring;
11557 tctx = current->io_uring;
11560 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
11562 for (i = 0; i < ARRAY_SIZE(new_count); i++)
11564 ctx->iowq_limits[i] = new_count[i];
11565 ctx->iowq_limits_set = true;
11567 if (tctx && tctx->io_wq) {
11568 ret = io_wq_max_workers(tctx->io_wq, new_count);
11572 memset(new_count, 0, sizeof(new_count));
11576 mutex_unlock(&sqd->lock);
11577 io_put_sq_data(sqd);
11580 if (copy_to_user(arg, new_count, sizeof(new_count)))
11583 /* that's it for SQPOLL, only the SQPOLL task creates requests */
11587 /* now propagate the restriction to all registered users */
11588 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
11589 struct io_uring_task *tctx = node->task->io_uring;
11591 if (WARN_ON_ONCE(!tctx->io_wq))
11594 for (i = 0; i < ARRAY_SIZE(new_count); i++)
11595 new_count[i] = ctx->iowq_limits[i];
11596 /* ignore errors, it always returns zero anyway */
11597 (void)io_wq_max_workers(tctx->io_wq, new_count);
11602 mutex_unlock(&sqd->lock);
11603 io_put_sq_data(sqd);
11608 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
11609 void __user *arg, unsigned nr_args)
11610 __releases(ctx->uring_lock)
11611 __acquires(ctx->uring_lock)
11616 * We're inside the ring mutex, if the ref is already dying, then
11617 * someone else killed the ctx or is already going through
11618 * io_uring_register().
11620 if (percpu_ref_is_dying(&ctx->refs))
11623 if (ctx->restricted) {
11624 if (opcode >= IORING_REGISTER_LAST)
11626 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
11627 if (!test_bit(opcode, ctx->restrictions.register_op))
11632 case IORING_REGISTER_BUFFERS:
11633 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
11635 case IORING_UNREGISTER_BUFFERS:
11637 if (arg || nr_args)
11639 ret = io_sqe_buffers_unregister(ctx);
11641 case IORING_REGISTER_FILES:
11642 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
11644 case IORING_UNREGISTER_FILES:
11646 if (arg || nr_args)
11648 ret = io_sqe_files_unregister(ctx);
11650 case IORING_REGISTER_FILES_UPDATE:
11651 ret = io_register_files_update(ctx, arg, nr_args);
11653 case IORING_REGISTER_EVENTFD:
11657 ret = io_eventfd_register(ctx, arg, 0);
11659 case IORING_REGISTER_EVENTFD_ASYNC:
11663 ret = io_eventfd_register(ctx, arg, 1);
11665 case IORING_UNREGISTER_EVENTFD:
11667 if (arg || nr_args)
11669 ret = io_eventfd_unregister(ctx);
11671 case IORING_REGISTER_PROBE:
11673 if (!arg || nr_args > 256)
11675 ret = io_probe(ctx, arg, nr_args);
11677 case IORING_REGISTER_PERSONALITY:
11679 if (arg || nr_args)
11681 ret = io_register_personality(ctx);
11683 case IORING_UNREGISTER_PERSONALITY:
11687 ret = io_unregister_personality(ctx, nr_args);
11689 case IORING_REGISTER_ENABLE_RINGS:
11691 if (arg || nr_args)
11693 ret = io_register_enable_rings(ctx);
11695 case IORING_REGISTER_RESTRICTIONS:
11696 ret = io_register_restrictions(ctx, arg, nr_args);
11698 case IORING_REGISTER_FILES2:
11699 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
11701 case IORING_REGISTER_FILES_UPDATE2:
11702 ret = io_register_rsrc_update(ctx, arg, nr_args,
11705 case IORING_REGISTER_BUFFERS2:
11706 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
11708 case IORING_REGISTER_BUFFERS_UPDATE:
11709 ret = io_register_rsrc_update(ctx, arg, nr_args,
11710 IORING_RSRC_BUFFER);
11712 case IORING_REGISTER_IOWQ_AFF:
11714 if (!arg || !nr_args)
11716 ret = io_register_iowq_aff(ctx, arg, nr_args);
11718 case IORING_UNREGISTER_IOWQ_AFF:
11720 if (arg || nr_args)
11722 ret = io_unregister_iowq_aff(ctx);
11724 case IORING_REGISTER_IOWQ_MAX_WORKERS:
11726 if (!arg || nr_args != 2)
11728 ret = io_register_iowq_max_workers(ctx, arg);
11730 case IORING_REGISTER_RING_FDS:
11731 ret = io_ringfd_register(ctx, arg, nr_args);
11733 case IORING_UNREGISTER_RING_FDS:
11734 ret = io_ringfd_unregister(ctx, arg, nr_args);
11744 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
11745 void __user *, arg, unsigned int, nr_args)
11747 struct io_ring_ctx *ctx;
11756 if (f.file->f_op != &io_uring_fops)
11759 ctx = f.file->private_data;
11761 io_run_task_work();
11763 mutex_lock(&ctx->uring_lock);
11764 ret = __io_uring_register(ctx, opcode, arg, nr_args);
11765 mutex_unlock(&ctx->uring_lock);
11766 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
11772 static int __init io_uring_init(void)
11774 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
11775 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
11776 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
11779 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
11780 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
11781 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
11782 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
11783 BUILD_BUG_SQE_ELEM(1, __u8, flags);
11784 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
11785 BUILD_BUG_SQE_ELEM(4, __s32, fd);
11786 BUILD_BUG_SQE_ELEM(8, __u64, off);
11787 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
11788 BUILD_BUG_SQE_ELEM(16, __u64, addr);
11789 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
11790 BUILD_BUG_SQE_ELEM(24, __u32, len);
11791 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
11792 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
11793 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
11794 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
11795 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
11796 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
11797 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
11798 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
11799 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
11800 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
11801 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
11802 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
11803 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
11804 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
11805 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
11806 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
11807 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
11808 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
11809 BUILD_BUG_SQE_ELEM(42, __u16, personality);
11810 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
11811 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
11813 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
11814 sizeof(struct io_uring_rsrc_update));
11815 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
11816 sizeof(struct io_uring_rsrc_update2));
11818 /* ->buf_index is u16 */
11819 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
11821 /* should fit into one byte */
11822 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
11823 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
11824 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
11826 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
11827 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
11829 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11833 __initcall(io_uring_init);