xref: /linux/include/linux/io_uring_types.h (revision 3a4d319a8fb5a9bbdf5b31ef32841eb286b1dcc2)
1 #ifndef IO_URING_TYPES_H
2 #define IO_URING_TYPES_H
3 
4 #include <linux/blkdev.h>
5 #include <linux/hashtable.h>
6 #include <linux/task_work.h>
7 #include <linux/bitmap.h>
8 #include <linux/llist.h>
9 #include <uapi/linux/io_uring.h>
10 
11 enum {
12 	/*
13 	 * A hint to not wake right away but delay until there are enough of
14 	 * tw's queued to match the number of CQEs the task is waiting for.
15 	 *
16 	 * Must not be used with requests generating more than one CQE.
17 	 * It's also ignored unless IORING_SETUP_DEFER_TASKRUN is set.
18 	 */
19 	IOU_F_TWQ_LAZY_WAKE			= 1,
20 };
21 
22 enum io_uring_cmd_flags {
23 	IO_URING_F_COMPLETE_DEFER	= 1,
24 	IO_URING_F_UNLOCKED		= 2,
25 	/* the request is executed from poll, it should not be freed */
26 	IO_URING_F_MULTISHOT		= 4,
27 	/* executed by io-wq */
28 	IO_URING_F_IOWQ			= 8,
29 	/* int's last bit, sign checks are usually faster than a bit test */
30 	IO_URING_F_NONBLOCK		= INT_MIN,
31 
32 	/* ctx state flags, for URING_CMD */
33 	IO_URING_F_SQE128		= (1 << 8),
34 	IO_URING_F_CQE32		= (1 << 9),
35 	IO_URING_F_IOPOLL		= (1 << 10),
36 
37 	/* set when uring wants to cancel a previously issued command */
38 	IO_URING_F_CANCEL		= (1 << 11),
39 	IO_URING_F_COMPAT		= (1 << 12),
40 };
41 
42 struct io_wq_work_node {
43 	struct io_wq_work_node *next;
44 };
45 
46 struct io_wq_work_list {
47 	struct io_wq_work_node *first;
48 	struct io_wq_work_node *last;
49 };
50 
51 struct io_wq_work {
52 	struct io_wq_work_node list;
53 	atomic_t flags;
54 	/* place it here instead of io_kiocb as it fills padding and saves 4B */
55 	int cancel_seq;
56 };
57 
58 struct io_fixed_file {
59 	/* file * with additional FFS_* flags */
60 	unsigned long file_ptr;
61 };
62 
63 struct io_file_table {
64 	struct io_fixed_file *files;
65 	unsigned long *bitmap;
66 	unsigned int alloc_hint;
67 };
68 
69 struct io_hash_bucket {
70 	spinlock_t		lock;
71 	struct hlist_head	list;
72 } ____cacheline_aligned_in_smp;
73 
74 struct io_hash_table {
75 	struct io_hash_bucket	*hbs;
76 	unsigned		hash_bits;
77 };
78 
79 /*
80  * Arbitrary limit, can be raised if need be
81  */
82 #define IO_RINGFD_REG_MAX 16
83 
84 struct io_uring_task {
85 	/* submission side */
86 	int				cached_refs;
87 	const struct io_ring_ctx 	*last;
88 	struct io_wq			*io_wq;
89 	struct file			*registered_rings[IO_RINGFD_REG_MAX];
90 
91 	struct xarray			xa;
92 	struct wait_queue_head		wait;
93 	atomic_t			in_cancel;
94 	atomic_t			inflight_tracked;
95 	struct percpu_counter		inflight;
96 
97 	struct { /* task_work */
98 		struct llist_head	task_list;
99 		struct callback_head	task_work;
100 	} ____cacheline_aligned_in_smp;
101 };
102 
103 struct io_uring {
104 	u32 head;
105 	u32 tail;
106 };
107 
108 /*
109  * This data is shared with the application through the mmap at offsets
110  * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111  *
112  * The offsets to the member fields are published through struct
113  * io_sqring_offsets when calling io_uring_setup.
114  */
115 struct io_rings {
116 	/*
117 	 * Head and tail offsets into the ring; the offsets need to be
118 	 * masked to get valid indices.
119 	 *
120 	 * The kernel controls head of the sq ring and the tail of the cq ring,
121 	 * and the application controls tail of the sq ring and the head of the
122 	 * cq ring.
123 	 */
124 	struct io_uring		sq, cq;
125 	/*
126 	 * Bitmasks to apply to head and tail offsets (constant, equals
127 	 * ring_entries - 1)
128 	 */
129 	u32			sq_ring_mask, cq_ring_mask;
130 	/* Ring sizes (constant, power of 2) */
131 	u32			sq_ring_entries, cq_ring_entries;
132 	/*
133 	 * Number of invalid entries dropped by the kernel due to
134 	 * invalid index stored in array
135 	 *
136 	 * Written by the kernel, shouldn't be modified by the
137 	 * application (i.e. get number of "new events" by comparing to
138 	 * cached value).
139 	 *
140 	 * After a new SQ head value was read by the application this
141 	 * counter includes all submissions that were dropped reaching
142 	 * the new SQ head (and possibly more).
143 	 */
144 	u32			sq_dropped;
145 	/*
146 	 * Runtime SQ flags
147 	 *
148 	 * Written by the kernel, shouldn't be modified by the
149 	 * application.
150 	 *
151 	 * The application needs a full memory barrier before checking
152 	 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
153 	 */
154 	atomic_t		sq_flags;
155 	/*
156 	 * Runtime CQ flags
157 	 *
158 	 * Written by the application, shouldn't be modified by the
159 	 * kernel.
160 	 */
161 	u32			cq_flags;
162 	/*
163 	 * Number of completion events lost because the queue was full;
164 	 * this should be avoided by the application by making sure
165 	 * there are not more requests pending than there is space in
166 	 * the completion queue.
167 	 *
168 	 * Written by the kernel, shouldn't be modified by the
169 	 * application (i.e. get number of "new events" by comparing to
170 	 * cached value).
171 	 *
172 	 * As completion events come in out of order this counter is not
173 	 * ordered with any other data.
174 	 */
175 	u32			cq_overflow;
176 	/*
177 	 * Ring buffer of completion events.
178 	 *
179 	 * The kernel writes completion events fresh every time they are
180 	 * produced, so the application is allowed to modify pending
181 	 * entries.
182 	 */
183 	struct io_uring_cqe	cqes[] ____cacheline_aligned_in_smp;
184 };
185 
186 struct io_restriction {
187 	DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
188 	DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
189 	u8 sqe_flags_allowed;
190 	u8 sqe_flags_required;
191 	bool registered;
192 };
193 
194 struct io_submit_link {
195 	struct io_kiocb		*head;
196 	struct io_kiocb		*last;
197 };
198 
199 struct io_submit_state {
200 	/* inline/task_work completion list, under ->uring_lock */
201 	struct io_wq_work_node	free_list;
202 	/* batch completion logic */
203 	struct io_wq_work_list	compl_reqs;
204 	struct io_submit_link	link;
205 
206 	bool			plug_started;
207 	bool			need_plug;
208 	bool			cq_flush;
209 	unsigned short		submit_nr;
210 	struct blk_plug		plug;
211 };
212 
213 struct io_alloc_cache {
214 	void			**entries;
215 	unsigned int		nr_cached;
216 	unsigned int		max_cached;
217 	size_t			elem_size;
218 };
219 
220 struct io_ring_ctx {
221 	/* const or read-mostly hot data */
222 	struct {
223 		unsigned int		flags;
224 		unsigned int		drain_next: 1;
225 		unsigned int		restricted: 1;
226 		unsigned int		off_timeout_used: 1;
227 		unsigned int		drain_active: 1;
228 		unsigned int		has_evfd: 1;
229 		/* all CQEs should be posted only by the submitter task */
230 		unsigned int		task_complete: 1;
231 		unsigned int		lockless_cq: 1;
232 		unsigned int		syscall_iopoll: 1;
233 		unsigned int		poll_activated: 1;
234 		unsigned int		drain_disabled: 1;
235 		unsigned int		compat: 1;
236 		unsigned int		iowq_limits_set : 1;
237 
238 		struct task_struct	*submitter_task;
239 		struct io_rings		*rings;
240 		struct percpu_ref	refs;
241 
242 		clockid_t		clockid;
243 		enum tk_offsets		clock_offset;
244 
245 		enum task_work_notify_mode	notify_method;
246 		unsigned			sq_thread_idle;
247 	} ____cacheline_aligned_in_smp;
248 
249 	/* submission data */
250 	struct {
251 		struct mutex		uring_lock;
252 
253 		/*
254 		 * Ring buffer of indices into array of io_uring_sqe, which is
255 		 * mmapped by the application using the IORING_OFF_SQES offset.
256 		 *
257 		 * This indirection could e.g. be used to assign fixed
258 		 * io_uring_sqe entries to operations and only submit them to
259 		 * the queue when needed.
260 		 *
261 		 * The kernel modifies neither the indices array nor the entries
262 		 * array.
263 		 */
264 		u32			*sq_array;
265 		struct io_uring_sqe	*sq_sqes;
266 		unsigned		cached_sq_head;
267 		unsigned		sq_entries;
268 
269 		/*
270 		 * Fixed resources fast path, should be accessed only under
271 		 * uring_lock, and updated through io_uring_register(2)
272 		 */
273 		struct io_rsrc_node	*rsrc_node;
274 		atomic_t		cancel_seq;
275 
276 		/*
277 		 * ->iopoll_list is protected by the ctx->uring_lock for
278 		 * io_uring instances that don't use IORING_SETUP_SQPOLL.
279 		 * For SQPOLL, only the single threaded io_sq_thread() will
280 		 * manipulate the list, hence no extra locking is needed there.
281 		 */
282 		bool			poll_multi_queue;
283 		struct io_wq_work_list	iopoll_list;
284 
285 		struct io_file_table	file_table;
286 		struct io_mapped_ubuf	**user_bufs;
287 		unsigned		nr_user_files;
288 		unsigned		nr_user_bufs;
289 
290 		struct io_submit_state	submit_state;
291 
292 		struct xarray		io_bl_xa;
293 
294 		struct io_hash_table	cancel_table_locked;
295 		struct io_alloc_cache	apoll_cache;
296 		struct io_alloc_cache	netmsg_cache;
297 		struct io_alloc_cache	rw_cache;
298 		struct io_alloc_cache	uring_cache;
299 
300 		/*
301 		 * Any cancelable uring_cmd is added to this list in
302 		 * ->uring_cmd() by io_uring_cmd_insert_cancelable()
303 		 */
304 		struct hlist_head	cancelable_uring_cmd;
305 	} ____cacheline_aligned_in_smp;
306 
307 	struct {
308 		/*
309 		 * We cache a range of free CQEs we can use, once exhausted it
310 		 * should go through a slower range setup, see __io_get_cqe()
311 		 */
312 		struct io_uring_cqe	*cqe_cached;
313 		struct io_uring_cqe	*cqe_sentinel;
314 
315 		unsigned		cached_cq_tail;
316 		unsigned		cq_entries;
317 		struct io_ev_fd	__rcu	*io_ev_fd;
318 		unsigned		cq_extra;
319 	} ____cacheline_aligned_in_smp;
320 
321 	/*
322 	 * task_work and async notification delivery cacheline. Expected to
323 	 * regularly bounce b/w CPUs.
324 	 */
325 	struct {
326 		struct llist_head	work_llist;
327 		unsigned long		check_cq;
328 		atomic_t		cq_wait_nr;
329 		atomic_t		cq_timeouts;
330 		struct wait_queue_head	cq_wait;
331 	} ____cacheline_aligned_in_smp;
332 
333 	/* timeouts */
334 	struct {
335 		spinlock_t		timeout_lock;
336 		struct list_head	timeout_list;
337 		struct list_head	ltimeout_list;
338 		unsigned		cq_last_tm_flush;
339 	} ____cacheline_aligned_in_smp;
340 
341 	spinlock_t		completion_lock;
342 
343 	struct list_head	io_buffers_comp;
344 	struct list_head	cq_overflow_list;
345 	struct io_hash_table	cancel_table;
346 
347 	struct hlist_head	waitid_list;
348 
349 #ifdef CONFIG_FUTEX
350 	struct hlist_head	futex_list;
351 	struct io_alloc_cache	futex_cache;
352 #endif
353 
354 	const struct cred	*sq_creds;	/* cred used for __io_sq_thread() */
355 	struct io_sq_data	*sq_data;	/* if using sq thread polling */
356 
357 	struct wait_queue_head	sqo_sq_wait;
358 	struct list_head	sqd_list;
359 
360 	unsigned int		file_alloc_start;
361 	unsigned int		file_alloc_end;
362 
363 	struct list_head	io_buffers_cache;
364 
365 	/* Keep this last, we don't need it for the fast path */
366 	struct wait_queue_head		poll_wq;
367 	struct io_restriction		restrictions;
368 
369 	/* slow path rsrc auxilary data, used by update/register */
370 	struct io_rsrc_data		*file_data;
371 	struct io_rsrc_data		*buf_data;
372 
373 	/* protected by ->uring_lock */
374 	struct list_head		rsrc_ref_list;
375 	struct io_alloc_cache		rsrc_node_cache;
376 	struct wait_queue_head		rsrc_quiesce_wq;
377 	unsigned			rsrc_quiesce;
378 
379 	u32			pers_next;
380 	struct xarray		personalities;
381 
382 	/* hashed buffered write serialization */
383 	struct io_wq_hash		*hash_map;
384 
385 	/* Only used for accounting purposes */
386 	struct user_struct		*user;
387 	struct mm_struct		*mm_account;
388 
389 	/* ctx exit and cancelation */
390 	struct llist_head		fallback_llist;
391 	struct delayed_work		fallback_work;
392 	struct work_struct		exit_work;
393 	struct list_head		tctx_list;
394 	struct completion		ref_comp;
395 
396 	/* io-wq management, e.g. thread count */
397 	u32				iowq_limits[2];
398 
399 	struct callback_head		poll_wq_task_work;
400 	struct list_head		defer_list;
401 
402 	struct io_alloc_cache		msg_cache;
403 	spinlock_t			msg_lock;
404 
405 #ifdef CONFIG_NET_RX_BUSY_POLL
406 	struct list_head	napi_list;	/* track busy poll napi_id */
407 	spinlock_t		napi_lock;	/* napi_list lock */
408 
409 	/* napi busy poll default timeout */
410 	ktime_t			napi_busy_poll_dt;
411 	bool			napi_prefer_busy_poll;
412 	bool			napi_enabled;
413 
414 	DECLARE_HASHTABLE(napi_ht, 4);
415 #endif
416 
417 	/* protected by ->completion_lock */
418 	unsigned			evfd_last_cq_tail;
419 
420 	/*
421 	 * If IORING_SETUP_NO_MMAP is used, then the below holds
422 	 * the gup'ed pages for the two rings, and the sqes.
423 	 */
424 	unsigned short			n_ring_pages;
425 	unsigned short			n_sqe_pages;
426 	struct page			**ring_pages;
427 	struct page			**sqe_pages;
428 };
429 
430 struct io_tw_state {
431 };
432 
433 enum {
434 	REQ_F_FIXED_FILE_BIT	= IOSQE_FIXED_FILE_BIT,
435 	REQ_F_IO_DRAIN_BIT	= IOSQE_IO_DRAIN_BIT,
436 	REQ_F_LINK_BIT		= IOSQE_IO_LINK_BIT,
437 	REQ_F_HARDLINK_BIT	= IOSQE_IO_HARDLINK_BIT,
438 	REQ_F_FORCE_ASYNC_BIT	= IOSQE_ASYNC_BIT,
439 	REQ_F_BUFFER_SELECT_BIT	= IOSQE_BUFFER_SELECT_BIT,
440 	REQ_F_CQE_SKIP_BIT	= IOSQE_CQE_SKIP_SUCCESS_BIT,
441 
442 	/* first byte is taken by user flags, shift it to not overlap */
443 	REQ_F_FAIL_BIT		= 8,
444 	REQ_F_INFLIGHT_BIT,
445 	REQ_F_CUR_POS_BIT,
446 	REQ_F_NOWAIT_BIT,
447 	REQ_F_LINK_TIMEOUT_BIT,
448 	REQ_F_NEED_CLEANUP_BIT,
449 	REQ_F_POLLED_BIT,
450 	REQ_F_BUFFER_SELECTED_BIT,
451 	REQ_F_BUFFER_RING_BIT,
452 	REQ_F_REISSUE_BIT,
453 	REQ_F_CREDS_BIT,
454 	REQ_F_REFCOUNT_BIT,
455 	REQ_F_ARM_LTIMEOUT_BIT,
456 	REQ_F_ASYNC_DATA_BIT,
457 	REQ_F_SKIP_LINK_CQES_BIT,
458 	REQ_F_SINGLE_POLL_BIT,
459 	REQ_F_DOUBLE_POLL_BIT,
460 	REQ_F_APOLL_MULTISHOT_BIT,
461 	REQ_F_CLEAR_POLLIN_BIT,
462 	REQ_F_HASH_LOCKED_BIT,
463 	/* keep async read/write and isreg together and in order */
464 	REQ_F_SUPPORT_NOWAIT_BIT,
465 	REQ_F_ISREG_BIT,
466 	REQ_F_POLL_NO_LAZY_BIT,
467 	REQ_F_CAN_POLL_BIT,
468 	REQ_F_BL_EMPTY_BIT,
469 	REQ_F_BL_NO_RECYCLE_BIT,
470 	REQ_F_BUFFERS_COMMIT_BIT,
471 
472 	/* not a real bit, just to check we're not overflowing the space */
473 	__REQ_F_LAST_BIT,
474 };
475 
476 typedef u64 __bitwise io_req_flags_t;
477 #define IO_REQ_FLAG(bitno)	((__force io_req_flags_t) BIT_ULL((bitno)))
478 
479 enum {
480 	/* ctx owns file */
481 	REQ_F_FIXED_FILE	= IO_REQ_FLAG(REQ_F_FIXED_FILE_BIT),
482 	/* drain existing IO first */
483 	REQ_F_IO_DRAIN		= IO_REQ_FLAG(REQ_F_IO_DRAIN_BIT),
484 	/* linked sqes */
485 	REQ_F_LINK		= IO_REQ_FLAG(REQ_F_LINK_BIT),
486 	/* doesn't sever on completion < 0 */
487 	REQ_F_HARDLINK		= IO_REQ_FLAG(REQ_F_HARDLINK_BIT),
488 	/* IOSQE_ASYNC */
489 	REQ_F_FORCE_ASYNC	= IO_REQ_FLAG(REQ_F_FORCE_ASYNC_BIT),
490 	/* IOSQE_BUFFER_SELECT */
491 	REQ_F_BUFFER_SELECT	= IO_REQ_FLAG(REQ_F_BUFFER_SELECT_BIT),
492 	/* IOSQE_CQE_SKIP_SUCCESS */
493 	REQ_F_CQE_SKIP		= IO_REQ_FLAG(REQ_F_CQE_SKIP_BIT),
494 
495 	/* fail rest of links */
496 	REQ_F_FAIL		= IO_REQ_FLAG(REQ_F_FAIL_BIT),
497 	/* on inflight list, should be cancelled and waited on exit reliably */
498 	REQ_F_INFLIGHT		= IO_REQ_FLAG(REQ_F_INFLIGHT_BIT),
499 	/* read/write uses file position */
500 	REQ_F_CUR_POS		= IO_REQ_FLAG(REQ_F_CUR_POS_BIT),
501 	/* must not punt to workers */
502 	REQ_F_NOWAIT		= IO_REQ_FLAG(REQ_F_NOWAIT_BIT),
503 	/* has or had linked timeout */
504 	REQ_F_LINK_TIMEOUT	= IO_REQ_FLAG(REQ_F_LINK_TIMEOUT_BIT),
505 	/* needs cleanup */
506 	REQ_F_NEED_CLEANUP	= IO_REQ_FLAG(REQ_F_NEED_CLEANUP_BIT),
507 	/* already went through poll handler */
508 	REQ_F_POLLED		= IO_REQ_FLAG(REQ_F_POLLED_BIT),
509 	/* buffer already selected */
510 	REQ_F_BUFFER_SELECTED	= IO_REQ_FLAG(REQ_F_BUFFER_SELECTED_BIT),
511 	/* buffer selected from ring, needs commit */
512 	REQ_F_BUFFER_RING	= IO_REQ_FLAG(REQ_F_BUFFER_RING_BIT),
513 	/* caller should reissue async */
514 	REQ_F_REISSUE		= IO_REQ_FLAG(REQ_F_REISSUE_BIT),
515 	/* supports async reads/writes */
516 	REQ_F_SUPPORT_NOWAIT	= IO_REQ_FLAG(REQ_F_SUPPORT_NOWAIT_BIT),
517 	/* regular file */
518 	REQ_F_ISREG		= IO_REQ_FLAG(REQ_F_ISREG_BIT),
519 	/* has creds assigned */
520 	REQ_F_CREDS		= IO_REQ_FLAG(REQ_F_CREDS_BIT),
521 	/* skip refcounting if not set */
522 	REQ_F_REFCOUNT		= IO_REQ_FLAG(REQ_F_REFCOUNT_BIT),
523 	/* there is a linked timeout that has to be armed */
524 	REQ_F_ARM_LTIMEOUT	= IO_REQ_FLAG(REQ_F_ARM_LTIMEOUT_BIT),
525 	/* ->async_data allocated */
526 	REQ_F_ASYNC_DATA	= IO_REQ_FLAG(REQ_F_ASYNC_DATA_BIT),
527 	/* don't post CQEs while failing linked requests */
528 	REQ_F_SKIP_LINK_CQES	= IO_REQ_FLAG(REQ_F_SKIP_LINK_CQES_BIT),
529 	/* single poll may be active */
530 	REQ_F_SINGLE_POLL	= IO_REQ_FLAG(REQ_F_SINGLE_POLL_BIT),
531 	/* double poll may active */
532 	REQ_F_DOUBLE_POLL	= IO_REQ_FLAG(REQ_F_DOUBLE_POLL_BIT),
533 	/* fast poll multishot mode */
534 	REQ_F_APOLL_MULTISHOT	= IO_REQ_FLAG(REQ_F_APOLL_MULTISHOT_BIT),
535 	/* recvmsg special flag, clear EPOLLIN */
536 	REQ_F_CLEAR_POLLIN	= IO_REQ_FLAG(REQ_F_CLEAR_POLLIN_BIT),
537 	/* hashed into ->cancel_hash_locked, protected by ->uring_lock */
538 	REQ_F_HASH_LOCKED	= IO_REQ_FLAG(REQ_F_HASH_LOCKED_BIT),
539 	/* don't use lazy poll wake for this request */
540 	REQ_F_POLL_NO_LAZY	= IO_REQ_FLAG(REQ_F_POLL_NO_LAZY_BIT),
541 	/* file is pollable */
542 	REQ_F_CAN_POLL		= IO_REQ_FLAG(REQ_F_CAN_POLL_BIT),
543 	/* buffer list was empty after selection of buffer */
544 	REQ_F_BL_EMPTY		= IO_REQ_FLAG(REQ_F_BL_EMPTY_BIT),
545 	/* don't recycle provided buffers for this request */
546 	REQ_F_BL_NO_RECYCLE	= IO_REQ_FLAG(REQ_F_BL_NO_RECYCLE_BIT),
547 	/* buffer ring head needs incrementing on put */
548 	REQ_F_BUFFERS_COMMIT	= IO_REQ_FLAG(REQ_F_BUFFERS_COMMIT_BIT),
549 };
550 
551 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, struct io_tw_state *ts);
552 
553 struct io_task_work {
554 	struct llist_node		node;
555 	io_req_tw_func_t		func;
556 };
557 
558 struct io_cqe {
559 	__u64	user_data;
560 	__s32	res;
561 	/* fd initially, then cflags for completion */
562 	union {
563 		__u32	flags;
564 		int	fd;
565 	};
566 };
567 
568 /*
569  * Each request type overlays its private data structure on top of this one.
570  * They must not exceed this one in size.
571  */
572 struct io_cmd_data {
573 	struct file		*file;
574 	/* each command gets 56 bytes of data */
575 	__u8			data[56];
576 };
577 
io_kiocb_cmd_sz_check(size_t cmd_sz)578 static inline void io_kiocb_cmd_sz_check(size_t cmd_sz)
579 {
580 	BUILD_BUG_ON(cmd_sz > sizeof(struct io_cmd_data));
581 }
582 #define io_kiocb_to_cmd(req, cmd_type) ( \
583 	io_kiocb_cmd_sz_check(sizeof(cmd_type)) , \
584 	((cmd_type *)&(req)->cmd) \
585 )
586 #define cmd_to_io_kiocb(ptr)	((struct io_kiocb *) ptr)
587 
588 struct io_kiocb {
589 	union {
590 		/*
591 		 * NOTE! Each of the io_kiocb union members has the file pointer
592 		 * as the first entry in their struct definition. So you can
593 		 * access the file pointer through any of the sub-structs,
594 		 * or directly as just 'file' in this struct.
595 		 */
596 		struct file		*file;
597 		struct io_cmd_data	cmd;
598 	};
599 
600 	u8				opcode;
601 	/* polled IO has completed */
602 	u8				iopoll_completed;
603 	/*
604 	 * Can be either a fixed buffer index, or used with provided buffers.
605 	 * For the latter, before issue it points to the buffer group ID,
606 	 * and after selection it points to the buffer ID itself.
607 	 */
608 	u16				buf_index;
609 
610 	unsigned			nr_tw;
611 
612 	/* REQ_F_* flags */
613 	io_req_flags_t			flags;
614 
615 	struct io_cqe			cqe;
616 
617 	struct io_ring_ctx		*ctx;
618 	struct task_struct		*task;
619 
620 	union {
621 		/* store used ubuf, so we can prevent reloading */
622 		struct io_mapped_ubuf	*imu;
623 
624 		/* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
625 		struct io_buffer	*kbuf;
626 
627 		/*
628 		 * stores buffer ID for ring provided buffers, valid IFF
629 		 * REQ_F_BUFFER_RING is set.
630 		 */
631 		struct io_buffer_list	*buf_list;
632 	};
633 
634 	union {
635 		/* used by request caches, completion batching and iopoll */
636 		struct io_wq_work_node	comp_list;
637 		/* cache ->apoll->events */
638 		__poll_t apoll_events;
639 	};
640 
641 	struct io_rsrc_node		*rsrc_node;
642 
643 	atomic_t			refs;
644 	bool				cancel_seq_set;
645 	struct io_task_work		io_task_work;
646 	/* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
647 	struct hlist_node		hash_node;
648 	/* internal polling, see IORING_FEAT_FAST_POLL */
649 	struct async_poll		*apoll;
650 	/* opcode allocated if it needs to store data for async defer */
651 	void				*async_data;
652 	/* linked requests, IFF REQ_F_HARDLINK or REQ_F_LINK are set */
653 	atomic_t			poll_refs;
654 	struct io_kiocb			*link;
655 	/* custom credentials, valid IFF REQ_F_CREDS is set */
656 	const struct cred		*creds;
657 	struct io_wq_work		work;
658 
659 	struct {
660 		u64			extra1;
661 		u64			extra2;
662 	} big_cqe;
663 };
664 
665 struct io_overflow_cqe {
666 	struct list_head list;
667 	struct io_uring_cqe cqe;
668 };
669 
670 #endif
671