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>
20 * IO submission data structure (Submission Queue Entry)
23 __u8 opcode; /* type of operation for this sqe */
24 __u8 flags; /* IOSQE_ flags */
25 __u16 ioprio; /* ioprio for the request */
26 __s32 fd; /* file descriptor to do IO on */
28 __u64 off; /* offset into file */
36 __u64 addr; /* pointer to buffer or iovecs */
39 __u32 len; /* buffer size or number of iovecs */
41 __kernel_rwf_t rw_flags;
43 __u16 poll_events; /* compatibility */
44 __u32 poll32_events; /* word-reversed for BE */
45 __u32 sync_range_flags;
59 __u32 uring_cmd_flags;
61 __u64 user_data; /* data to be passed back at completion time */
62 /* pack this to avoid bogus arm OABI complaints */
64 /* index into fixed buffers, if used */
66 /* for grouped buffer selection */
68 } __attribute__((packed));
69 /* personality to use, if used */
85 * If the ring is initialized with IORING_SETUP_SQE128, then
86 * this field is used for 80 bytes of arbitrary command data
93 * If sqe->file_index is set to this for opcodes that instantiate a new
94 * direct descriptor (like openat/openat2/accept), then io_uring will allocate
95 * an available direct descriptor instead of having the application pass one
96 * in. The picked direct descriptor will be returned in cqe->res, or -ENFILE
97 * if the space is full.
99 #define IORING_FILE_INDEX_ALLOC (~0U)
102 IOSQE_FIXED_FILE_BIT,
105 IOSQE_IO_HARDLINK_BIT,
107 IOSQE_BUFFER_SELECT_BIT,
108 IOSQE_CQE_SKIP_SUCCESS_BIT,
114 /* use fixed fileset */
115 #define IOSQE_FIXED_FILE (1U << IOSQE_FIXED_FILE_BIT)
116 /* issue after inflight IO */
117 #define IOSQE_IO_DRAIN (1U << IOSQE_IO_DRAIN_BIT)
119 #define IOSQE_IO_LINK (1U << IOSQE_IO_LINK_BIT)
120 /* like LINK, but stronger */
121 #define IOSQE_IO_HARDLINK (1U << IOSQE_IO_HARDLINK_BIT)
122 /* always go async */
123 #define IOSQE_ASYNC (1U << IOSQE_ASYNC_BIT)
124 /* select buffer from sqe->buf_group */
125 #define IOSQE_BUFFER_SELECT (1U << IOSQE_BUFFER_SELECT_BIT)
126 /* don't post CQE if request succeeded */
127 #define IOSQE_CQE_SKIP_SUCCESS (1U << IOSQE_CQE_SKIP_SUCCESS_BIT)
130 * io_uring_setup() flags
132 #define IORING_SETUP_IOPOLL (1U << 0) /* io_context is polled */
133 #define IORING_SETUP_SQPOLL (1U << 1) /* SQ poll thread */
134 #define IORING_SETUP_SQ_AFF (1U << 2) /* sq_thread_cpu is valid */
135 #define IORING_SETUP_CQSIZE (1U << 3) /* app defines CQ size */
136 #define IORING_SETUP_CLAMP (1U << 4) /* clamp SQ/CQ ring sizes */
137 #define IORING_SETUP_ATTACH_WQ (1U << 5) /* attach to existing wq */
138 #define IORING_SETUP_R_DISABLED (1U << 6) /* start with ring disabled */
139 #define IORING_SETUP_SUBMIT_ALL (1U << 7) /* continue submit on error */
141 * Cooperative task running. When requests complete, they often require
142 * forcing the submitter to transition to the kernel to complete. If this
143 * flag is set, work will be done when the task transitions anyway, rather
144 * than force an inter-processor interrupt reschedule. This avoids interrupting
145 * a task running in userspace, and saves an IPI.
147 #define IORING_SETUP_COOP_TASKRUN (1U << 8)
149 * If COOP_TASKRUN is set, get notified if task work is available for
150 * running and a kernel transition would be needed to run it. This sets
151 * IORING_SQ_TASKRUN in the sq ring flags. Not valid with COOP_TASKRUN.
153 #define IORING_SETUP_TASKRUN_FLAG (1U << 9)
154 #define IORING_SETUP_SQE128 (1U << 10) /* SQEs are 128 byte */
155 #define IORING_SETUP_CQE32 (1U << 11) /* CQEs are 32 byte */
157 * Only one task is allowed to submit requests
159 #define IORING_SETUP_SINGLE_ISSUER (1U << 12)
162 * Defer running task work to get events.
163 * Rather than running bits of task work whenever the task transitions
164 * try to do it just before it is needed.
166 #define IORING_SETUP_DEFER_TASKRUN (1U << 13)
173 IORING_OP_READ_FIXED,
174 IORING_OP_WRITE_FIXED,
176 IORING_OP_POLL_REMOVE,
177 IORING_OP_SYNC_FILE_RANGE,
181 IORING_OP_TIMEOUT_REMOVE,
183 IORING_OP_ASYNC_CANCEL,
184 IORING_OP_LINK_TIMEOUT,
189 IORING_OP_FILES_UPDATE,
200 IORING_OP_PROVIDE_BUFFERS,
201 IORING_OP_REMOVE_BUFFERS,
217 IORING_OP_SENDMSG_ZC,
219 /* this goes last, obviously */
224 * sqe->uring_cmd_flags
225 * IORING_URING_CMD_FIXED use registered buffer; pass this flag
226 * along with setting sqe->buf_index.
228 #define IORING_URING_CMD_FIXED (1U << 0)
234 #define IORING_FSYNC_DATASYNC (1U << 0)
239 #define IORING_TIMEOUT_ABS (1U << 0)
240 #define IORING_TIMEOUT_UPDATE (1U << 1)
241 #define IORING_TIMEOUT_BOOTTIME (1U << 2)
242 #define IORING_TIMEOUT_REALTIME (1U << 3)
243 #define IORING_LINK_TIMEOUT_UPDATE (1U << 4)
244 #define IORING_TIMEOUT_ETIME_SUCCESS (1U << 5)
245 #define IORING_TIMEOUT_CLOCK_MASK (IORING_TIMEOUT_BOOTTIME | IORING_TIMEOUT_REALTIME)
246 #define IORING_TIMEOUT_UPDATE_MASK (IORING_TIMEOUT_UPDATE | IORING_LINK_TIMEOUT_UPDATE)
249 * extends splice(2) flags
251 #define SPLICE_F_FD_IN_FIXED (1U << 31) /* the last bit of __u32 */
254 * POLL_ADD flags. Note that since sqe->poll_events is the flag space, the
255 * command flags for POLL_ADD are stored in sqe->len.
257 * IORING_POLL_ADD_MULTI Multishot poll. Sets IORING_CQE_F_MORE if
258 * the poll handler will continue to report
259 * CQEs on behalf of the same SQE.
261 * IORING_POLL_UPDATE Update existing poll request, matching
262 * sqe->addr as the old user_data field.
264 * IORING_POLL_LEVEL Level triggered poll.
266 #define IORING_POLL_ADD_MULTI (1U << 0)
267 #define IORING_POLL_UPDATE_EVENTS (1U << 1)
268 #define IORING_POLL_UPDATE_USER_DATA (1U << 2)
269 #define IORING_POLL_ADD_LEVEL (1U << 3)
272 * ASYNC_CANCEL flags.
274 * IORING_ASYNC_CANCEL_ALL Cancel all requests that match the given key
275 * IORING_ASYNC_CANCEL_FD Key off 'fd' for cancelation rather than the
276 * request 'user_data'
277 * IORING_ASYNC_CANCEL_ANY Match any request
278 * IORING_ASYNC_CANCEL_FD_FIXED 'fd' passed in is a fixed descriptor
280 #define IORING_ASYNC_CANCEL_ALL (1U << 0)
281 #define IORING_ASYNC_CANCEL_FD (1U << 1)
282 #define IORING_ASYNC_CANCEL_ANY (1U << 2)
283 #define IORING_ASYNC_CANCEL_FD_FIXED (1U << 3)
286 * send/sendmsg and recv/recvmsg flags (sqe->ioprio)
288 * IORING_RECVSEND_POLL_FIRST If set, instead of first attempting to send
289 * or receive and arm poll if that yields an
290 * -EAGAIN result, arm poll upfront and skip
291 * the initial transfer attempt.
293 * IORING_RECV_MULTISHOT Multishot recv. Sets IORING_CQE_F_MORE if
294 * the handler will continue to report
295 * CQEs on behalf of the same SQE.
297 * IORING_RECVSEND_FIXED_BUF Use registered buffers, the index is stored in
298 * the buf_index field.
300 * IORING_SEND_ZC_REPORT_USAGE
301 * If set, SEND[MSG]_ZC should report
302 * the zerocopy usage in cqe.res
303 * for the IORING_CQE_F_NOTIF cqe.
304 * 0 is reported if zerocopy was actually possible.
305 * IORING_NOTIF_USAGE_ZC_COPIED if data was copied
306 * (at least partially).
308 #define IORING_RECVSEND_POLL_FIRST (1U << 0)
309 #define IORING_RECV_MULTISHOT (1U << 1)
310 #define IORING_RECVSEND_FIXED_BUF (1U << 2)
311 #define IORING_SEND_ZC_REPORT_USAGE (1U << 3)
314 * cqe.res for IORING_CQE_F_NOTIF if
315 * IORING_SEND_ZC_REPORT_USAGE was requested
317 * It should be treated as a flag, all other
318 * bits of cqe.res should be treated as reserved!
320 #define IORING_NOTIF_USAGE_ZC_COPIED (1U << 31)
323 * accept flags stored in sqe->ioprio
325 #define IORING_ACCEPT_MULTISHOT (1U << 0)
328 * IORING_OP_MSG_RING command types, stored in sqe->addr
331 IORING_MSG_DATA, /* pass sqe->len as 'res' and off as user_data */
332 IORING_MSG_SEND_FD, /* send a registered fd to another ring */
336 * IORING_OP_MSG_RING flags (sqe->msg_ring_flags)
338 * IORING_MSG_RING_CQE_SKIP Don't post a CQE to the target ring. Not
339 * applicable for IORING_MSG_DATA, obviously.
341 #define IORING_MSG_RING_CQE_SKIP (1U << 0)
344 * IO completion data structure (Completion Queue Entry)
346 struct io_uring_cqe {
347 __u64 user_data; /* sqe->data submission passed back */
348 __s32 res; /* result code for this event */
352 * If the ring is initialized with IORING_SETUP_CQE32, then this field
353 * contains 16-bytes of padding, doubling the size of the CQE.
361 * IORING_CQE_F_BUFFER If set, the upper 16 bits are the buffer ID
362 * IORING_CQE_F_MORE If set, parent SQE will generate more CQE entries
363 * IORING_CQE_F_SOCK_NONEMPTY If set, more data to read after socket recv
364 * IORING_CQE_F_NOTIF Set for notification CQEs. Can be used to distinct
367 #define IORING_CQE_F_BUFFER (1U << 0)
368 #define IORING_CQE_F_MORE (1U << 1)
369 #define IORING_CQE_F_SOCK_NONEMPTY (1U << 2)
370 #define IORING_CQE_F_NOTIF (1U << 3)
373 IORING_CQE_BUFFER_SHIFT = 16,
377 * Magic offsets for the application to mmap the data it needs
379 #define IORING_OFF_SQ_RING 0ULL
380 #define IORING_OFF_CQ_RING 0x8000000ULL
381 #define IORING_OFF_SQES 0x10000000ULL
384 * Filled with the offset for mmap(2)
386 struct io_sqring_offsets {
401 #define IORING_SQ_NEED_WAKEUP (1U << 0) /* needs io_uring_enter wakeup */
402 #define IORING_SQ_CQ_OVERFLOW (1U << 1) /* CQ ring is overflown */
403 #define IORING_SQ_TASKRUN (1U << 2) /* task should enter the kernel */
405 struct io_cqring_offsets {
421 /* disable eventfd notifications */
422 #define IORING_CQ_EVENTFD_DISABLED (1U << 0)
425 * io_uring_enter(2) flags
427 #define IORING_ENTER_GETEVENTS (1U << 0)
428 #define IORING_ENTER_SQ_WAKEUP (1U << 1)
429 #define IORING_ENTER_SQ_WAIT (1U << 2)
430 #define IORING_ENTER_EXT_ARG (1U << 3)
431 #define IORING_ENTER_REGISTERED_RING (1U << 4)
434 * Passed in for io_uring_setup(2). Copied back with updated info on success
436 struct io_uring_params {
441 __u32 sq_thread_idle;
445 struct io_sqring_offsets sq_off;
446 struct io_cqring_offsets cq_off;
450 * io_uring_params->features flags
452 #define IORING_FEAT_SINGLE_MMAP (1U << 0)
453 #define IORING_FEAT_NODROP (1U << 1)
454 #define IORING_FEAT_SUBMIT_STABLE (1U << 2)
455 #define IORING_FEAT_RW_CUR_POS (1U << 3)
456 #define IORING_FEAT_CUR_PERSONALITY (1U << 4)
457 #define IORING_FEAT_FAST_POLL (1U << 5)
458 #define IORING_FEAT_POLL_32BITS (1U << 6)
459 #define IORING_FEAT_SQPOLL_NONFIXED (1U << 7)
460 #define IORING_FEAT_EXT_ARG (1U << 8)
461 #define IORING_FEAT_NATIVE_WORKERS (1U << 9)
462 #define IORING_FEAT_RSRC_TAGS (1U << 10)
463 #define IORING_FEAT_CQE_SKIP (1U << 11)
464 #define IORING_FEAT_LINKED_FILE (1U << 12)
467 * io_uring_register(2) opcodes and arguments
470 IORING_REGISTER_BUFFERS = 0,
471 IORING_UNREGISTER_BUFFERS = 1,
472 IORING_REGISTER_FILES = 2,
473 IORING_UNREGISTER_FILES = 3,
474 IORING_REGISTER_EVENTFD = 4,
475 IORING_UNREGISTER_EVENTFD = 5,
476 IORING_REGISTER_FILES_UPDATE = 6,
477 IORING_REGISTER_EVENTFD_ASYNC = 7,
478 IORING_REGISTER_PROBE = 8,
479 IORING_REGISTER_PERSONALITY = 9,
480 IORING_UNREGISTER_PERSONALITY = 10,
481 IORING_REGISTER_RESTRICTIONS = 11,
482 IORING_REGISTER_ENABLE_RINGS = 12,
484 /* extended with tagging */
485 IORING_REGISTER_FILES2 = 13,
486 IORING_REGISTER_FILES_UPDATE2 = 14,
487 IORING_REGISTER_BUFFERS2 = 15,
488 IORING_REGISTER_BUFFERS_UPDATE = 16,
490 /* set/clear io-wq thread affinities */
491 IORING_REGISTER_IOWQ_AFF = 17,
492 IORING_UNREGISTER_IOWQ_AFF = 18,
494 /* set/get max number of io-wq workers */
495 IORING_REGISTER_IOWQ_MAX_WORKERS = 19,
497 /* register/unregister io_uring fd with the ring */
498 IORING_REGISTER_RING_FDS = 20,
499 IORING_UNREGISTER_RING_FDS = 21,
501 /* register ring based provide buffer group */
502 IORING_REGISTER_PBUF_RING = 22,
503 IORING_UNREGISTER_PBUF_RING = 23,
505 /* sync cancelation API */
506 IORING_REGISTER_SYNC_CANCEL = 24,
508 /* register a range of fixed file slots for automatic slot allocation */
509 IORING_REGISTER_FILE_ALLOC_RANGE = 25,
515 /* io-wq worker categories */
521 /* deprecated, see struct io_uring_rsrc_update */
522 struct io_uring_files_update {
525 __aligned_u64 /* __s32 * */ fds;
529 * Register a fully sparse file space, rather than pass in an array of all
530 * -1 file descriptors.
532 #define IORING_RSRC_REGISTER_SPARSE (1U << 0)
534 struct io_uring_rsrc_register {
542 struct io_uring_rsrc_update {
548 struct io_uring_rsrc_update2 {
557 struct io_uring_notification_slot {
562 struct io_uring_notification_register {
570 /* Skip updating fd indexes set to this value in the fd table */
571 #define IORING_REGISTER_FILES_SKIP (-2)
573 #define IO_URING_OP_SUPPORTED (1U << 0)
575 struct io_uring_probe_op {
578 __u16 flags; /* IO_URING_OP_* flags */
582 struct io_uring_probe {
583 __u8 last_op; /* last opcode supported */
584 __u8 ops_len; /* length of ops[] array below */
587 struct io_uring_probe_op ops[];
590 struct io_uring_restriction {
593 __u8 register_op; /* IORING_RESTRICTION_REGISTER_OP */
594 __u8 sqe_op; /* IORING_RESTRICTION_SQE_OP */
595 __u8 sqe_flags; /* IORING_RESTRICTION_SQE_FLAGS_* */
601 struct io_uring_buf {
608 struct io_uring_buf_ring {
611 * To avoid spilling into more pages than we need to, the
612 * ring tail is overlaid with the io_uring_buf->resv field.
620 __DECLARE_FLEX_ARRAY(struct io_uring_buf, bufs);
624 /* argument for IORING_(UN)REGISTER_PBUF_RING */
625 struct io_uring_buf_reg {
634 * io_uring_restriction->opcode values
637 /* Allow an io_uring_register(2) opcode */
638 IORING_RESTRICTION_REGISTER_OP = 0,
640 /* Allow an sqe opcode */
641 IORING_RESTRICTION_SQE_OP = 1,
643 /* Allow sqe flags */
644 IORING_RESTRICTION_SQE_FLAGS_ALLOWED = 2,
646 /* Require sqe flags (these flags must be set on each submission) */
647 IORING_RESTRICTION_SQE_FLAGS_REQUIRED = 3,
649 IORING_RESTRICTION_LAST
652 struct io_uring_getevents_arg {
660 * Argument for IORING_REGISTER_SYNC_CANCEL
662 struct io_uring_sync_cancel_reg {
666 struct __kernel_timespec timeout;
671 * Argument for IORING_REGISTER_FILE_ALLOC_RANGE
672 * The range is specified as [off, off + len)
674 struct io_uring_file_index_range {
680 struct io_uring_recvmsg_out {