xref: /openbmc/linux/include/uapi/linux/io_uring.h (revision cbb3a19f090d5a41b822caf9ff2058e1c6bc7ea3)
1 /* SPDX-License-Identifier: (GPL-2.0 WITH Linux-syscall-note) OR MIT */
2 /*
3  * Header file for the io_uring interface.
4  *
5  * Copyright (C) 2019 Jens Axboe
6  * Copyright (C) 2019 Christoph Hellwig
7  */
8 #ifndef LINUX_IO_URING_H
9 #define LINUX_IO_URING_H
10 
11 #include <linux/fs.h>
12 #include <linux/types.h>
13 
14 /*
15  * IO submission data structure (Submission Queue Entry)
16  */
17 struct io_uring_sqe {
18 	__u8	opcode;		/* type of operation for this sqe */
19 	__u8	flags;		/* IOSQE_ flags */
20 	__u16	ioprio;		/* ioprio for the request */
21 	__s32	fd;		/* file descriptor to do IO on */
22 	union {
23 		__u64	off;	/* offset into file */
24 		__u64	addr2;
25 		__u32	cmd_op;
26 	};
27 	union {
28 		__u64	addr;	/* pointer to buffer or iovecs */
29 		__u64	splice_off_in;
30 	};
31 	__u32	len;		/* buffer size or number of iovecs */
32 	union {
33 		__kernel_rwf_t	rw_flags;
34 		__u32		fsync_flags;
35 		__u16		poll_events;	/* compatibility */
36 		__u32		poll32_events;	/* word-reversed for BE */
37 		__u32		sync_range_flags;
38 		__u32		msg_flags;
39 		__u32		timeout_flags;
40 		__u32		accept_flags;
41 		__u32		cancel_flags;
42 		__u32		open_flags;
43 		__u32		statx_flags;
44 		__u32		fadvise_advice;
45 		__u32		splice_flags;
46 		__u32		rename_flags;
47 		__u32		unlink_flags;
48 		__u32		hardlink_flags;
49 		__u32		xattr_flags;
50 		__u32		close_flags;
51 	};
52 	__u64	user_data;	/* data to be passed back at completion time */
53 	/* pack this to avoid bogus arm OABI complaints */
54 	union {
55 		/* index into fixed buffers, if used */
56 		__u16	buf_index;
57 		/* for grouped buffer selection */
58 		__u16	buf_group;
59 	} __attribute__((packed));
60 	/* personality to use, if used */
61 	__u16	personality;
62 	union {
63 		__s32	splice_fd_in;
64 		__u32	file_index;
65 	};
66 	union {
67 		struct {
68 			__u64	addr3;
69 			__u64	__pad2[1];
70 		};
71 		/*
72 		 * If the ring is initialized with IORING_SETUP_SQE128, then
73 		 * this field is used for 80 bytes of arbitrary command data
74 		 */
75 		__u8	cmd[0];
76 	};
77 };
78 
79 /*
80  * If sqe->file_index is set to this for opcodes that instantiate a new
81  * direct descriptor (like openat/openat2/accept), then io_uring will allocate
82  * an available direct descriptor instead of having the application pass one
83  * in. The picked direct descriptor will be returned in cqe->res, or -ENFILE
84  * if the space is full.
85  */
86 #define IORING_FILE_INDEX_ALLOC		(~0U)
87 
88 enum {
89 	IOSQE_FIXED_FILE_BIT,
90 	IOSQE_IO_DRAIN_BIT,
91 	IOSQE_IO_LINK_BIT,
92 	IOSQE_IO_HARDLINK_BIT,
93 	IOSQE_ASYNC_BIT,
94 	IOSQE_BUFFER_SELECT_BIT,
95 	IOSQE_CQE_SKIP_SUCCESS_BIT,
96 };
97 
98 /*
99  * sqe->flags
100  */
101 /* use fixed fileset */
102 #define IOSQE_FIXED_FILE	(1U << IOSQE_FIXED_FILE_BIT)
103 /* issue after inflight IO */
104 #define IOSQE_IO_DRAIN		(1U << IOSQE_IO_DRAIN_BIT)
105 /* links next sqe */
106 #define IOSQE_IO_LINK		(1U << IOSQE_IO_LINK_BIT)
107 /* like LINK, but stronger */
108 #define IOSQE_IO_HARDLINK	(1U << IOSQE_IO_HARDLINK_BIT)
109 /* always go async */
110 #define IOSQE_ASYNC		(1U << IOSQE_ASYNC_BIT)
111 /* select buffer from sqe->buf_group */
112 #define IOSQE_BUFFER_SELECT	(1U << IOSQE_BUFFER_SELECT_BIT)
113 /* don't post CQE if request succeeded */
114 #define IOSQE_CQE_SKIP_SUCCESS	(1U << IOSQE_CQE_SKIP_SUCCESS_BIT)
115 
116 /*
117  * io_uring_setup() flags
118  */
119 #define IORING_SETUP_IOPOLL	(1U << 0)	/* io_context is polled */
120 #define IORING_SETUP_SQPOLL	(1U << 1)	/* SQ poll thread */
121 #define IORING_SETUP_SQ_AFF	(1U << 2)	/* sq_thread_cpu is valid */
122 #define IORING_SETUP_CQSIZE	(1U << 3)	/* app defines CQ size */
123 #define IORING_SETUP_CLAMP	(1U << 4)	/* clamp SQ/CQ ring sizes */
124 #define IORING_SETUP_ATTACH_WQ	(1U << 5)	/* attach to existing wq */
125 #define IORING_SETUP_R_DISABLED	(1U << 6)	/* start with ring disabled */
126 #define IORING_SETUP_SUBMIT_ALL	(1U << 7)	/* continue submit on error */
127 /*
128  * Cooperative task running. When requests complete, they often require
129  * forcing the submitter to transition to the kernel to complete. If this
130  * flag is set, work will be done when the task transitions anyway, rather
131  * than force an inter-processor interrupt reschedule. This avoids interrupting
132  * a task running in userspace, and saves an IPI.
133  */
134 #define IORING_SETUP_COOP_TASKRUN	(1U << 8)
135 /*
136  * If COOP_TASKRUN is set, get notified if task work is available for
137  * running and a kernel transition would be needed to run it. This sets
138  * IORING_SQ_TASKRUN in the sq ring flags. Not valid with COOP_TASKRUN.
139  */
140 #define IORING_SETUP_TASKRUN_FLAG	(1U << 9)
141 
142 #define IORING_SETUP_SQE128		(1U << 10) /* SQEs are 128 byte */
143 #define IORING_SETUP_CQE32		(1U << 11) /* CQEs are 32 byte */
144 
145 enum io_uring_op {
146 	IORING_OP_NOP,
147 	IORING_OP_READV,
148 	IORING_OP_WRITEV,
149 	IORING_OP_FSYNC,
150 	IORING_OP_READ_FIXED,
151 	IORING_OP_WRITE_FIXED,
152 	IORING_OP_POLL_ADD,
153 	IORING_OP_POLL_REMOVE,
154 	IORING_OP_SYNC_FILE_RANGE,
155 	IORING_OP_SENDMSG,
156 	IORING_OP_RECVMSG,
157 	IORING_OP_TIMEOUT,
158 	IORING_OP_TIMEOUT_REMOVE,
159 	IORING_OP_ACCEPT,
160 	IORING_OP_ASYNC_CANCEL,
161 	IORING_OP_LINK_TIMEOUT,
162 	IORING_OP_CONNECT,
163 	IORING_OP_FALLOCATE,
164 	IORING_OP_OPENAT,
165 	IORING_OP_CLOSE,
166 	IORING_OP_FILES_UPDATE,
167 	IORING_OP_STATX,
168 	IORING_OP_READ,
169 	IORING_OP_WRITE,
170 	IORING_OP_FADVISE,
171 	IORING_OP_MADVISE,
172 	IORING_OP_SEND,
173 	IORING_OP_RECV,
174 	IORING_OP_OPENAT2,
175 	IORING_OP_EPOLL_CTL,
176 	IORING_OP_SPLICE,
177 	IORING_OP_PROVIDE_BUFFERS,
178 	IORING_OP_REMOVE_BUFFERS,
179 	IORING_OP_TEE,
180 	IORING_OP_SHUTDOWN,
181 	IORING_OP_RENAMEAT,
182 	IORING_OP_UNLINKAT,
183 	IORING_OP_MKDIRAT,
184 	IORING_OP_SYMLINKAT,
185 	IORING_OP_LINKAT,
186 	IORING_OP_MSG_RING,
187 	IORING_OP_FSETXATTR,
188 	IORING_OP_SETXATTR,
189 	IORING_OP_FGETXATTR,
190 	IORING_OP_GETXATTR,
191 	IORING_OP_SOCKET,
192 	IORING_OP_URING_CMD,
193 
194 	/* this goes last, obviously */
195 	IORING_OP_LAST,
196 };
197 
198 /*
199  * sqe->fsync_flags
200  */
201 #define IORING_FSYNC_DATASYNC	(1U << 0)
202 
203 /*
204  * sqe->timeout_flags
205  */
206 #define IORING_TIMEOUT_ABS		(1U << 0)
207 #define IORING_TIMEOUT_UPDATE		(1U << 1)
208 #define IORING_TIMEOUT_BOOTTIME		(1U << 2)
209 #define IORING_TIMEOUT_REALTIME		(1U << 3)
210 #define IORING_LINK_TIMEOUT_UPDATE	(1U << 4)
211 #define IORING_TIMEOUT_ETIME_SUCCESS	(1U << 5)
212 #define IORING_TIMEOUT_CLOCK_MASK	(IORING_TIMEOUT_BOOTTIME | IORING_TIMEOUT_REALTIME)
213 #define IORING_TIMEOUT_UPDATE_MASK	(IORING_TIMEOUT_UPDATE | IORING_LINK_TIMEOUT_UPDATE)
214 /*
215  * sqe->splice_flags
216  * extends splice(2) flags
217  */
218 #define SPLICE_F_FD_IN_FIXED	(1U << 31) /* the last bit of __u32 */
219 
220 /*
221  * POLL_ADD flags. Note that since sqe->poll_events is the flag space, the
222  * command flags for POLL_ADD are stored in sqe->len.
223  *
224  * IORING_POLL_ADD_MULTI	Multishot poll. Sets IORING_CQE_F_MORE if
225  *				the poll handler will continue to report
226  *				CQEs on behalf of the same SQE.
227  *
228  * IORING_POLL_UPDATE		Update existing poll request, matching
229  *				sqe->addr as the old user_data field.
230  */
231 #define IORING_POLL_ADD_MULTI	(1U << 0)
232 #define IORING_POLL_UPDATE_EVENTS	(1U << 1)
233 #define IORING_POLL_UPDATE_USER_DATA	(1U << 2)
234 
235 /*
236  * ASYNC_CANCEL flags.
237  *
238  * IORING_ASYNC_CANCEL_ALL	Cancel all requests that match the given key
239  * IORING_ASYNC_CANCEL_FD	Key off 'fd' for cancelation rather than the
240  *				request 'user_data'
241  * IORING_ASYNC_CANCEL_ANY	Match any request
242  */
243 #define IORING_ASYNC_CANCEL_ALL	(1U << 0)
244 #define IORING_ASYNC_CANCEL_FD	(1U << 1)
245 #define IORING_ASYNC_CANCEL_ANY	(1U << 2)
246 
247 /*
248  * send/sendmsg and recv/recvmsg flags (sqe->addr2)
249  *
250  * IORING_RECVSEND_POLL_FIRST	If set, instead of first attempting to send
251  *				or receive and arm poll if that yields an
252  *				-EAGAIN result, arm poll upfront and skip
253  *				the initial transfer attempt.
254  */
255 #define IORING_RECVSEND_POLL_FIRST	(1U << 0)
256 
257 /*
258  * accept flags stored in sqe->ioprio
259  */
260 #define IORING_ACCEPT_MULTISHOT	(1U << 0)
261 
262 /*
263  * close flags, store in sqe->close_flags
264  */
265 #define IORING_CLOSE_FD_AND_FILE_SLOT	(1U << 0)
266 
267 /*
268  * IO completion data structure (Completion Queue Entry)
269  */
270 struct io_uring_cqe {
271 	__u64	user_data;	/* sqe->data submission passed back */
272 	__s32	res;		/* result code for this event */
273 	__u32	flags;
274 
275 	/*
276 	 * If the ring is initialized with IORING_SETUP_CQE32, then this field
277 	 * contains 16-bytes of padding, doubling the size of the CQE.
278 	 */
279 	__u64 big_cqe[];
280 };
281 
282 /*
283  * cqe->flags
284  *
285  * IORING_CQE_F_BUFFER	If set, the upper 16 bits are the buffer ID
286  * IORING_CQE_F_MORE	If set, parent SQE will generate more CQE entries
287  * IORING_CQE_F_SOCK_NONEMPTY	If set, more data to read after socket recv
288  */
289 #define IORING_CQE_F_BUFFER		(1U << 0)
290 #define IORING_CQE_F_MORE		(1U << 1)
291 #define IORING_CQE_F_SOCK_NONEMPTY	(1U << 2)
292 
293 enum {
294 	IORING_CQE_BUFFER_SHIFT		= 16,
295 };
296 
297 /*
298  * Magic offsets for the application to mmap the data it needs
299  */
300 #define IORING_OFF_SQ_RING		0ULL
301 #define IORING_OFF_CQ_RING		0x8000000ULL
302 #define IORING_OFF_SQES			0x10000000ULL
303 
304 /*
305  * Filled with the offset for mmap(2)
306  */
307 struct io_sqring_offsets {
308 	__u32 head;
309 	__u32 tail;
310 	__u32 ring_mask;
311 	__u32 ring_entries;
312 	__u32 flags;
313 	__u32 dropped;
314 	__u32 array;
315 	__u32 resv1;
316 	__u64 resv2;
317 };
318 
319 /*
320  * sq_ring->flags
321  */
322 #define IORING_SQ_NEED_WAKEUP	(1U << 0) /* needs io_uring_enter wakeup */
323 #define IORING_SQ_CQ_OVERFLOW	(1U << 1) /* CQ ring is overflown */
324 #define IORING_SQ_TASKRUN	(1U << 2) /* task should enter the kernel */
325 
326 struct io_cqring_offsets {
327 	__u32 head;
328 	__u32 tail;
329 	__u32 ring_mask;
330 	__u32 ring_entries;
331 	__u32 overflow;
332 	__u32 cqes;
333 	__u32 flags;
334 	__u32 resv1;
335 	__u64 resv2;
336 };
337 
338 /*
339  * cq_ring->flags
340  */
341 
342 /* disable eventfd notifications */
343 #define IORING_CQ_EVENTFD_DISABLED	(1U << 0)
344 
345 /*
346  * io_uring_enter(2) flags
347  */
348 #define IORING_ENTER_GETEVENTS		(1U << 0)
349 #define IORING_ENTER_SQ_WAKEUP		(1U << 1)
350 #define IORING_ENTER_SQ_WAIT		(1U << 2)
351 #define IORING_ENTER_EXT_ARG		(1U << 3)
352 #define IORING_ENTER_REGISTERED_RING	(1U << 4)
353 
354 /*
355  * Passed in for io_uring_setup(2). Copied back with updated info on success
356  */
357 struct io_uring_params {
358 	__u32 sq_entries;
359 	__u32 cq_entries;
360 	__u32 flags;
361 	__u32 sq_thread_cpu;
362 	__u32 sq_thread_idle;
363 	__u32 features;
364 	__u32 wq_fd;
365 	__u32 resv[3];
366 	struct io_sqring_offsets sq_off;
367 	struct io_cqring_offsets cq_off;
368 };
369 
370 /*
371  * io_uring_params->features flags
372  */
373 #define IORING_FEAT_SINGLE_MMAP		(1U << 0)
374 #define IORING_FEAT_NODROP		(1U << 1)
375 #define IORING_FEAT_SUBMIT_STABLE	(1U << 2)
376 #define IORING_FEAT_RW_CUR_POS		(1U << 3)
377 #define IORING_FEAT_CUR_PERSONALITY	(1U << 4)
378 #define IORING_FEAT_FAST_POLL		(1U << 5)
379 #define IORING_FEAT_POLL_32BITS 	(1U << 6)
380 #define IORING_FEAT_SQPOLL_NONFIXED	(1U << 7)
381 #define IORING_FEAT_EXT_ARG		(1U << 8)
382 #define IORING_FEAT_NATIVE_WORKERS	(1U << 9)
383 #define IORING_FEAT_RSRC_TAGS		(1U << 10)
384 #define IORING_FEAT_CQE_SKIP		(1U << 11)
385 #define IORING_FEAT_LINKED_FILE		(1U << 12)
386 
387 /*
388  * io_uring_register(2) opcodes and arguments
389  */
390 enum {
391 	IORING_REGISTER_BUFFERS			= 0,
392 	IORING_UNREGISTER_BUFFERS		= 1,
393 	IORING_REGISTER_FILES			= 2,
394 	IORING_UNREGISTER_FILES			= 3,
395 	IORING_REGISTER_EVENTFD			= 4,
396 	IORING_UNREGISTER_EVENTFD		= 5,
397 	IORING_REGISTER_FILES_UPDATE		= 6,
398 	IORING_REGISTER_EVENTFD_ASYNC		= 7,
399 	IORING_REGISTER_PROBE			= 8,
400 	IORING_REGISTER_PERSONALITY		= 9,
401 	IORING_UNREGISTER_PERSONALITY		= 10,
402 	IORING_REGISTER_RESTRICTIONS		= 11,
403 	IORING_REGISTER_ENABLE_RINGS		= 12,
404 
405 	/* extended with tagging */
406 	IORING_REGISTER_FILES2			= 13,
407 	IORING_REGISTER_FILES_UPDATE2		= 14,
408 	IORING_REGISTER_BUFFERS2		= 15,
409 	IORING_REGISTER_BUFFERS_UPDATE		= 16,
410 
411 	/* set/clear io-wq thread affinities */
412 	IORING_REGISTER_IOWQ_AFF		= 17,
413 	IORING_UNREGISTER_IOWQ_AFF		= 18,
414 
415 	/* set/get max number of io-wq workers */
416 	IORING_REGISTER_IOWQ_MAX_WORKERS	= 19,
417 
418 	/* register/unregister io_uring fd with the ring */
419 	IORING_REGISTER_RING_FDS		= 20,
420 	IORING_UNREGISTER_RING_FDS		= 21,
421 
422 	/* register ring based provide buffer group */
423 	IORING_REGISTER_PBUF_RING		= 22,
424 	IORING_UNREGISTER_PBUF_RING		= 23,
425 
426 	/* this goes last */
427 	IORING_REGISTER_LAST
428 };
429 
430 /* io-wq worker categories */
431 enum {
432 	IO_WQ_BOUND,
433 	IO_WQ_UNBOUND,
434 };
435 
436 /* deprecated, see struct io_uring_rsrc_update */
437 struct io_uring_files_update {
438 	__u32 offset;
439 	__u32 resv;
440 	__aligned_u64 /* __s32 * */ fds;
441 };
442 
443 /*
444  * Register a fully sparse file space, rather than pass in an array of all
445  * -1 file descriptors.
446  */
447 #define IORING_RSRC_REGISTER_SPARSE	(1U << 0)
448 
449 struct io_uring_rsrc_register {
450 	__u32 nr;
451 	__u32 flags;
452 	__u64 resv2;
453 	__aligned_u64 data;
454 	__aligned_u64 tags;
455 };
456 
457 struct io_uring_rsrc_update {
458 	__u32 offset;
459 	__u32 resv;
460 	__aligned_u64 data;
461 };
462 
463 struct io_uring_rsrc_update2 {
464 	__u32 offset;
465 	__u32 resv;
466 	__aligned_u64 data;
467 	__aligned_u64 tags;
468 	__u32 nr;
469 	__u32 resv2;
470 };
471 
472 /* Skip updating fd indexes set to this value in the fd table */
473 #define IORING_REGISTER_FILES_SKIP	(-2)
474 
475 #define IO_URING_OP_SUPPORTED	(1U << 0)
476 
477 struct io_uring_probe_op {
478 	__u8 op;
479 	__u8 resv;
480 	__u16 flags;	/* IO_URING_OP_* flags */
481 	__u32 resv2;
482 };
483 
484 struct io_uring_probe {
485 	__u8 last_op;	/* last opcode supported */
486 	__u8 ops_len;	/* length of ops[] array below */
487 	__u16 resv;
488 	__u32 resv2[3];
489 	struct io_uring_probe_op ops[0];
490 };
491 
492 struct io_uring_restriction {
493 	__u16 opcode;
494 	union {
495 		__u8 register_op; /* IORING_RESTRICTION_REGISTER_OP */
496 		__u8 sqe_op;      /* IORING_RESTRICTION_SQE_OP */
497 		__u8 sqe_flags;   /* IORING_RESTRICTION_SQE_FLAGS_* */
498 	};
499 	__u8 resv;
500 	__u32 resv2[3];
501 };
502 
503 struct io_uring_buf {
504 	__u64	addr;
505 	__u32	len;
506 	__u16	bid;
507 	__u16	resv;
508 };
509 
510 struct io_uring_buf_ring {
511 	union {
512 		/*
513 		 * To avoid spilling into more pages than we need to, the
514 		 * ring tail is overlaid with the io_uring_buf->resv field.
515 		 */
516 		struct {
517 			__u64	resv1;
518 			__u32	resv2;
519 			__u16	resv3;
520 			__u16	tail;
521 		};
522 		struct io_uring_buf	bufs[0];
523 	};
524 };
525 
526 /* argument for IORING_(UN)REGISTER_PBUF_RING */
527 struct io_uring_buf_reg {
528 	__u64	ring_addr;
529 	__u32	ring_entries;
530 	__u16	bgid;
531 	__u16	pad;
532 	__u64	resv[3];
533 };
534 
535 /*
536  * io_uring_restriction->opcode values
537  */
538 enum {
539 	/* Allow an io_uring_register(2) opcode */
540 	IORING_RESTRICTION_REGISTER_OP		= 0,
541 
542 	/* Allow an sqe opcode */
543 	IORING_RESTRICTION_SQE_OP		= 1,
544 
545 	/* Allow sqe flags */
546 	IORING_RESTRICTION_SQE_FLAGS_ALLOWED	= 2,
547 
548 	/* Require sqe flags (these flags must be set on each submission) */
549 	IORING_RESTRICTION_SQE_FLAGS_REQUIRED	= 3,
550 
551 	IORING_RESTRICTION_LAST
552 };
553 
554 struct io_uring_getevents_arg {
555 	__u64	sigmask;
556 	__u32	sigmask_sz;
557 	__u32	pad;
558 	__u64	ts;
559 };
560 
561 #endif
562