1 /* SPDX-License-Identifier: (GPL-2.0 WITH Linux-syscall-note) OR MIT */
3 * Header file for the io_uring interface.
5 * Copyright (C) 2019 Jens Axboe
6 * Copyright (C) 2019 Christoph Hellwig
8 #ifndef LINUX_IO_URING_H
9 #define LINUX_IO_URING_H
12 #include <linux/types.h>
13 #include <linux/time_types.h>
16 * IO submission data structure (Submission Queue Entry)
19 __u8 opcode; /* type of operation for this sqe */
20 __u8 flags; /* IOSQE_ flags */
21 __u16 ioprio; /* ioprio for the request */
22 __s32 fd; /* file descriptor to do IO on */
24 __u64 off; /* offset into file */
32 __u64 addr; /* pointer to buffer or iovecs */
35 __u32 len; /* buffer size or number of iovecs */
37 __kernel_rwf_t rw_flags;
39 __u16 poll_events; /* compatibility */
40 __u32 poll32_events; /* word-reversed for BE */
41 __u32 sync_range_flags;
56 __u64 user_data; /* data to be passed back at completion time */
57 /* pack this to avoid bogus arm OABI complaints */
59 /* index into fixed buffers, if used */
61 /* for grouped buffer selection */
63 } __attribute__((packed));
64 /* personality to use, if used */
70 __u16 notification_idx;
80 * If the ring is initialized with IORING_SETUP_SQE128, then
81 * this field is used for 80 bytes of arbitrary command data
88 * If sqe->file_index is set to this for opcodes that instantiate a new
89 * direct descriptor (like openat/openat2/accept), then io_uring will allocate
90 * an available direct descriptor instead of having the application pass one
91 * in. The picked direct descriptor will be returned in cqe->res, or -ENFILE
92 * if the space is full.
94 #define IORING_FILE_INDEX_ALLOC (~0U)
100 IOSQE_IO_HARDLINK_BIT,
102 IOSQE_BUFFER_SELECT_BIT,
103 IOSQE_CQE_SKIP_SUCCESS_BIT,
109 /* use fixed fileset */
110 #define IOSQE_FIXED_FILE (1U << IOSQE_FIXED_FILE_BIT)
111 /* issue after inflight IO */
112 #define IOSQE_IO_DRAIN (1U << IOSQE_IO_DRAIN_BIT)
114 #define IOSQE_IO_LINK (1U << IOSQE_IO_LINK_BIT)
115 /* like LINK, but stronger */
116 #define IOSQE_IO_HARDLINK (1U << IOSQE_IO_HARDLINK_BIT)
117 /* always go async */
118 #define IOSQE_ASYNC (1U << IOSQE_ASYNC_BIT)
119 /* select buffer from sqe->buf_group */
120 #define IOSQE_BUFFER_SELECT (1U << IOSQE_BUFFER_SELECT_BIT)
121 /* don't post CQE if request succeeded */
122 #define IOSQE_CQE_SKIP_SUCCESS (1U << IOSQE_CQE_SKIP_SUCCESS_BIT)
125 * io_uring_setup() flags
127 #define IORING_SETUP_IOPOLL (1U << 0) /* io_context is polled */
128 #define IORING_SETUP_SQPOLL (1U << 1) /* SQ poll thread */
129 #define IORING_SETUP_SQ_AFF (1U << 2) /* sq_thread_cpu is valid */
130 #define IORING_SETUP_CQSIZE (1U << 3) /* app defines CQ size */
131 #define IORING_SETUP_CLAMP (1U << 4) /* clamp SQ/CQ ring sizes */
132 #define IORING_SETUP_ATTACH_WQ (1U << 5) /* attach to existing wq */
133 #define IORING_SETUP_R_DISABLED (1U << 6) /* start with ring disabled */
134 #define IORING_SETUP_SUBMIT_ALL (1U << 7) /* continue submit on error */
136 * Cooperative task running. When requests complete, they often require
137 * forcing the submitter to transition to the kernel to complete. If this
138 * flag is set, work will be done when the task transitions anyway, rather
139 * than force an inter-processor interrupt reschedule. This avoids interrupting
140 * a task running in userspace, and saves an IPI.
142 #define IORING_SETUP_COOP_TASKRUN (1U << 8)
144 * If COOP_TASKRUN is set, get notified if task work is available for
145 * running and a kernel transition would be needed to run it. This sets
146 * IORING_SQ_TASKRUN in the sq ring flags. Not valid with COOP_TASKRUN.
148 #define IORING_SETUP_TASKRUN_FLAG (1U << 9)
149 #define IORING_SETUP_SQE128 (1U << 10) /* SQEs are 128 byte */
150 #define IORING_SETUP_CQE32 (1U << 11) /* CQEs are 32 byte */
152 * Only one task is allowed to submit requests
154 #define IORING_SETUP_SINGLE_ISSUER (1U << 12)
161 IORING_OP_READ_FIXED,
162 IORING_OP_WRITE_FIXED,
164 IORING_OP_POLL_REMOVE,
165 IORING_OP_SYNC_FILE_RANGE,
169 IORING_OP_TIMEOUT_REMOVE,
171 IORING_OP_ASYNC_CANCEL,
172 IORING_OP_LINK_TIMEOUT,
177 IORING_OP_RSRC_UPDATE,
178 IORING_OP_FILES_UPDATE = IORING_OP_RSRC_UPDATE,
189 IORING_OP_PROVIDE_BUFFERS,
190 IORING_OP_REMOVE_BUFFERS,
205 IORING_OP_SENDZC_NOTIF,
207 /* this goes last, obviously */
214 #define IORING_FSYNC_DATASYNC (1U << 0)
219 #define IORING_TIMEOUT_ABS (1U << 0)
220 #define IORING_TIMEOUT_UPDATE (1U << 1)
221 #define IORING_TIMEOUT_BOOTTIME (1U << 2)
222 #define IORING_TIMEOUT_REALTIME (1U << 3)
223 #define IORING_LINK_TIMEOUT_UPDATE (1U << 4)
224 #define IORING_TIMEOUT_ETIME_SUCCESS (1U << 5)
225 #define IORING_TIMEOUT_CLOCK_MASK (IORING_TIMEOUT_BOOTTIME | IORING_TIMEOUT_REALTIME)
226 #define IORING_TIMEOUT_UPDATE_MASK (IORING_TIMEOUT_UPDATE | IORING_LINK_TIMEOUT_UPDATE)
230 * extends splice(2) flags
232 #define SPLICE_F_FD_IN_FIXED (1U << 31) /* the last bit of __u32 */
235 * POLL_ADD flags. Note that since sqe->poll_events is the flag space, the
236 * command flags for POLL_ADD are stored in sqe->len.
238 * IORING_POLL_ADD_MULTI Multishot poll. Sets IORING_CQE_F_MORE if
239 * the poll handler will continue to report
240 * CQEs on behalf of the same SQE.
242 * IORING_POLL_UPDATE Update existing poll request, matching
243 * sqe->addr as the old user_data field.
245 * IORING_POLL_LEVEL Level triggered poll.
247 #define IORING_POLL_ADD_MULTI (1U << 0)
248 #define IORING_POLL_UPDATE_EVENTS (1U << 1)
249 #define IORING_POLL_UPDATE_USER_DATA (1U << 2)
250 #define IORING_POLL_ADD_LEVEL (1U << 3)
253 * ASYNC_CANCEL flags.
255 * IORING_ASYNC_CANCEL_ALL Cancel all requests that match the given key
256 * IORING_ASYNC_CANCEL_FD Key off 'fd' for cancelation rather than the
257 * request 'user_data'
258 * IORING_ASYNC_CANCEL_ANY Match any request
259 * IORING_ASYNC_CANCEL_FD_FIXED 'fd' passed in is a fixed descriptor
261 #define IORING_ASYNC_CANCEL_ALL (1U << 0)
262 #define IORING_ASYNC_CANCEL_FD (1U << 1)
263 #define IORING_ASYNC_CANCEL_ANY (1U << 2)
264 #define IORING_ASYNC_CANCEL_FD_FIXED (1U << 3)
267 * send/sendmsg and recv/recvmsg flags (sqe->ioprio)
269 * IORING_RECVSEND_POLL_FIRST If set, instead of first attempting to send
270 * or receive and arm poll if that yields an
271 * -EAGAIN result, arm poll upfront and skip
272 * the initial transfer attempt.
274 * IORING_RECV_MULTISHOT Multishot recv. Sets IORING_CQE_F_MORE if
275 * the handler will continue to report
276 * CQEs on behalf of the same SQE.
278 * IORING_RECVSEND_FIXED_BUF Use registered buffers, the index is stored in
279 * the buf_index field.
281 * IORING_RECVSEND_NOTIF_FLUSH Flush a notification after a successful
282 * successful. Only for zerocopy sends.
284 #define IORING_RECVSEND_POLL_FIRST (1U << 0)
285 #define IORING_RECV_MULTISHOT (1U << 1)
286 #define IORING_RECVSEND_FIXED_BUF (1U << 2)
287 #define IORING_RECVSEND_NOTIF_FLUSH (1U << 3)
290 * accept flags stored in sqe->ioprio
292 #define IORING_ACCEPT_MULTISHOT (1U << 0)
296 * IORING_OP_RSRC_UPDATE flags
299 IORING_RSRC_UPDATE_FILES,
300 IORING_RSRC_UPDATE_NOTIF,
304 * IORING_OP_MSG_RING command types, stored in sqe->addr
307 IORING_MSG_DATA, /* pass sqe->len as 'res' and off as user_data */
308 IORING_MSG_SEND_FD, /* send a registered fd to another ring */
312 * IORING_OP_MSG_RING flags (sqe->msg_ring_flags)
314 * IORING_MSG_RING_CQE_SKIP Don't post a CQE to the target ring. Not
315 * applicable for IORING_MSG_DATA, obviously.
317 #define IORING_MSG_RING_CQE_SKIP (1U << 0)
320 * IO completion data structure (Completion Queue Entry)
322 struct io_uring_cqe {
323 __u64 user_data; /* sqe->data submission passed back */
324 __s32 res; /* result code for this event */
328 * If the ring is initialized with IORING_SETUP_CQE32, then this field
329 * contains 16-bytes of padding, doubling the size of the CQE.
337 * IORING_CQE_F_BUFFER If set, the upper 16 bits are the buffer ID
338 * IORING_CQE_F_MORE If set, parent SQE will generate more CQE entries
339 * IORING_CQE_F_SOCK_NONEMPTY If set, more data to read after socket recv
341 #define IORING_CQE_F_BUFFER (1U << 0)
342 #define IORING_CQE_F_MORE (1U << 1)
343 #define IORING_CQE_F_SOCK_NONEMPTY (1U << 2)
346 IORING_CQE_BUFFER_SHIFT = 16,
350 * Magic offsets for the application to mmap the data it needs
352 #define IORING_OFF_SQ_RING 0ULL
353 #define IORING_OFF_CQ_RING 0x8000000ULL
354 #define IORING_OFF_SQES 0x10000000ULL
357 * Filled with the offset for mmap(2)
359 struct io_sqring_offsets {
374 #define IORING_SQ_NEED_WAKEUP (1U << 0) /* needs io_uring_enter wakeup */
375 #define IORING_SQ_CQ_OVERFLOW (1U << 1) /* CQ ring is overflown */
376 #define IORING_SQ_TASKRUN (1U << 2) /* task should enter the kernel */
378 struct io_cqring_offsets {
394 /* disable eventfd notifications */
395 #define IORING_CQ_EVENTFD_DISABLED (1U << 0)
398 * io_uring_enter(2) flags
400 #define IORING_ENTER_GETEVENTS (1U << 0)
401 #define IORING_ENTER_SQ_WAKEUP (1U << 1)
402 #define IORING_ENTER_SQ_WAIT (1U << 2)
403 #define IORING_ENTER_EXT_ARG (1U << 3)
404 #define IORING_ENTER_REGISTERED_RING (1U << 4)
407 * Passed in for io_uring_setup(2). Copied back with updated info on success
409 struct io_uring_params {
414 __u32 sq_thread_idle;
418 struct io_sqring_offsets sq_off;
419 struct io_cqring_offsets cq_off;
423 * io_uring_params->features flags
425 #define IORING_FEAT_SINGLE_MMAP (1U << 0)
426 #define IORING_FEAT_NODROP (1U << 1)
427 #define IORING_FEAT_SUBMIT_STABLE (1U << 2)
428 #define IORING_FEAT_RW_CUR_POS (1U << 3)
429 #define IORING_FEAT_CUR_PERSONALITY (1U << 4)
430 #define IORING_FEAT_FAST_POLL (1U << 5)
431 #define IORING_FEAT_POLL_32BITS (1U << 6)
432 #define IORING_FEAT_SQPOLL_NONFIXED (1U << 7)
433 #define IORING_FEAT_EXT_ARG (1U << 8)
434 #define IORING_FEAT_NATIVE_WORKERS (1U << 9)
435 #define IORING_FEAT_RSRC_TAGS (1U << 10)
436 #define IORING_FEAT_CQE_SKIP (1U << 11)
437 #define IORING_FEAT_LINKED_FILE (1U << 12)
440 * io_uring_register(2) opcodes and arguments
443 IORING_REGISTER_BUFFERS = 0,
444 IORING_UNREGISTER_BUFFERS = 1,
445 IORING_REGISTER_FILES = 2,
446 IORING_UNREGISTER_FILES = 3,
447 IORING_REGISTER_EVENTFD = 4,
448 IORING_UNREGISTER_EVENTFD = 5,
449 IORING_REGISTER_FILES_UPDATE = 6,
450 IORING_REGISTER_EVENTFD_ASYNC = 7,
451 IORING_REGISTER_PROBE = 8,
452 IORING_REGISTER_PERSONALITY = 9,
453 IORING_UNREGISTER_PERSONALITY = 10,
454 IORING_REGISTER_RESTRICTIONS = 11,
455 IORING_REGISTER_ENABLE_RINGS = 12,
457 /* extended with tagging */
458 IORING_REGISTER_FILES2 = 13,
459 IORING_REGISTER_FILES_UPDATE2 = 14,
460 IORING_REGISTER_BUFFERS2 = 15,
461 IORING_REGISTER_BUFFERS_UPDATE = 16,
463 /* set/clear io-wq thread affinities */
464 IORING_REGISTER_IOWQ_AFF = 17,
465 IORING_UNREGISTER_IOWQ_AFF = 18,
467 /* set/get max number of io-wq workers */
468 IORING_REGISTER_IOWQ_MAX_WORKERS = 19,
470 /* register/unregister io_uring fd with the ring */
471 IORING_REGISTER_RING_FDS = 20,
472 IORING_UNREGISTER_RING_FDS = 21,
474 /* register ring based provide buffer group */
475 IORING_REGISTER_PBUF_RING = 22,
476 IORING_UNREGISTER_PBUF_RING = 23,
478 /* sync cancelation API */
479 IORING_REGISTER_SYNC_CANCEL = 24,
481 /* register a range of fixed file slots for automatic slot allocation */
482 IORING_REGISTER_FILE_ALLOC_RANGE = 25,
484 /* zerocopy notification API */
485 IORING_REGISTER_NOTIFIERS = 26,
486 IORING_UNREGISTER_NOTIFIERS = 27,
492 /* io-wq worker categories */
498 /* deprecated, see struct io_uring_rsrc_update */
499 struct io_uring_files_update {
502 __aligned_u64 /* __s32 * */ fds;
506 * Register a fully sparse file space, rather than pass in an array of all
507 * -1 file descriptors.
509 #define IORING_RSRC_REGISTER_SPARSE (1U << 0)
511 struct io_uring_rsrc_register {
519 struct io_uring_rsrc_update {
525 struct io_uring_rsrc_update2 {
534 struct io_uring_notification_slot {
539 struct io_uring_notification_register {
547 /* Skip updating fd indexes set to this value in the fd table */
548 #define IORING_REGISTER_FILES_SKIP (-2)
550 #define IO_URING_OP_SUPPORTED (1U << 0)
552 struct io_uring_probe_op {
555 __u16 flags; /* IO_URING_OP_* flags */
559 struct io_uring_probe {
560 __u8 last_op; /* last opcode supported */
561 __u8 ops_len; /* length of ops[] array below */
564 struct io_uring_probe_op ops[];
567 struct io_uring_restriction {
570 __u8 register_op; /* IORING_RESTRICTION_REGISTER_OP */
571 __u8 sqe_op; /* IORING_RESTRICTION_SQE_OP */
572 __u8 sqe_flags; /* IORING_RESTRICTION_SQE_FLAGS_* */
578 struct io_uring_buf {
585 struct io_uring_buf_ring {
588 * To avoid spilling into more pages than we need to, the
589 * ring tail is overlaid with the io_uring_buf->resv field.
597 struct io_uring_buf bufs[0];
601 /* argument for IORING_(UN)REGISTER_PBUF_RING */
602 struct io_uring_buf_reg {
611 * io_uring_restriction->opcode values
614 /* Allow an io_uring_register(2) opcode */
615 IORING_RESTRICTION_REGISTER_OP = 0,
617 /* Allow an sqe opcode */
618 IORING_RESTRICTION_SQE_OP = 1,
620 /* Allow sqe flags */
621 IORING_RESTRICTION_SQE_FLAGS_ALLOWED = 2,
623 /* Require sqe flags (these flags must be set on each submission) */
624 IORING_RESTRICTION_SQE_FLAGS_REQUIRED = 3,
626 IORING_RESTRICTION_LAST
629 struct io_uring_getevents_arg {
637 * Argument for IORING_REGISTER_SYNC_CANCEL
639 struct io_uring_sync_cancel_reg {
643 struct __kernel_timespec timeout;
648 * Argument for IORING_REGISTER_FILE_ALLOC_RANGE
649 * The range is specified as [off, off + len)
651 struct io_uring_file_index_range {
657 struct io_uring_recvmsg_out {