xref: /linux/include/linux/io_uring_types.h (revision 591beb0e3a03258ef9c01893a5209845799a7c33)
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 	/* executed inline from syscall */
30 	IO_URING_F_INLINE		= 16,
31 	/* int's last bit, sign checks are usually faster than a bit test */
32 	IO_URING_F_NONBLOCK		= INT_MIN,
33 
34 	/* ctx state flags, for URING_CMD */
35 	IO_URING_F_SQE128		= (1 << 8),
36 	IO_URING_F_CQE32		= (1 << 9),
37 	IO_URING_F_IOPOLL		= (1 << 10),
38 
39 	/* set when uring wants to cancel a previously issued command */
40 	IO_URING_F_CANCEL		= (1 << 11),
41 	IO_URING_F_COMPAT		= (1 << 12),
42 };
43 
44 struct io_wq_work_node {
45 	struct io_wq_work_node *next;
46 };
47 
48 struct io_wq_work_list {
49 	struct io_wq_work_node *first;
50 	struct io_wq_work_node *last;
51 };
52 
53 struct io_wq_work {
54 	struct io_wq_work_node list;
55 	atomic_t flags;
56 	/* place it here instead of io_kiocb as it fills padding and saves 4B */
57 	int cancel_seq;
58 };
59 
60 struct io_rsrc_data {
61 	unsigned int			nr;
62 	struct io_rsrc_node		**nodes;
63 };
64 
65 struct io_file_table {
66 	struct io_rsrc_data data;
67 	unsigned long *bitmap;
68 	unsigned int alloc_hint;
69 };
70 
71 struct io_hash_bucket {
72 	struct hlist_head	list;
73 } ____cacheline_aligned_in_smp;
74 
75 struct io_hash_table {
76 	struct io_hash_bucket	*hbs;
77 	unsigned		hash_bits;
78 };
79 
80 struct io_mapped_region {
81 	struct page		**pages;
82 	void			*ptr;
83 	unsigned		nr_pages;
84 	unsigned		flags;
85 };
86 
87 /*
88  * Return value from io_buffer_list selection, to avoid stashing it in
89  * struct io_kiocb. For legacy/classic provided buffers, keeping a reference
90  * across execution contexts are fine. But for ring provided buffers, the
91  * list may go away as soon as ->uring_lock is dropped. As the io_kiocb
92  * persists, it's better to just keep the buffer local for those cases.
93  */
94 struct io_br_sel {
95 	struct io_buffer_list *buf_list;
96 	/*
97 	 * Some selection parts return the user address, others return an error.
98 	 */
99 	union {
100 		void __user *addr;
101 		ssize_t val;
102 	};
103 };
104 
105 
106 /*
107  * Arbitrary limit, can be raised if need be
108  */
109 #define IO_RINGFD_REG_MAX 16
110 
111 struct io_uring_task {
112 	/* submission side */
113 	int				cached_refs;
114 	const struct io_ring_ctx 	*last;
115 	struct task_struct		*task;
116 	struct io_wq			*io_wq;
117 	struct file			*registered_rings[IO_RINGFD_REG_MAX];
118 
119 	struct xarray			xa;
120 	struct wait_queue_head		wait;
121 	atomic_t			in_cancel;
122 	atomic_t			inflight_tracked;
123 	struct percpu_counter		inflight;
124 
125 	struct { /* task_work */
126 		struct llist_head	task_list;
127 		struct callback_head	task_work;
128 	} ____cacheline_aligned_in_smp;
129 };
130 
131 struct iou_vec {
132 	union {
133 		struct iovec	*iovec;
134 		struct bio_vec	*bvec;
135 	};
136 	unsigned		nr; /* number of struct iovec it can hold */
137 };
138 
139 struct io_uring {
140 	u32 head;
141 	u32 tail;
142 };
143 
144 /*
145  * This data is shared with the application through the mmap at offsets
146  * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
147  *
148  * The offsets to the member fields are published through struct
149  * io_sqring_offsets when calling io_uring_setup.
150  */
151 struct io_rings {
152 	/*
153 	 * Head and tail offsets into the ring; the offsets need to be
154 	 * masked to get valid indices.
155 	 *
156 	 * The kernel controls head of the sq ring and the tail of the cq ring,
157 	 * and the application controls tail of the sq ring and the head of the
158 	 * cq ring.
159 	 */
160 	struct io_uring		sq, cq;
161 	/*
162 	 * Bitmasks to apply to head and tail offsets (constant, equals
163 	 * ring_entries - 1)
164 	 */
165 	u32			sq_ring_mask, cq_ring_mask;
166 	/* Ring sizes (constant, power of 2) */
167 	u32			sq_ring_entries, cq_ring_entries;
168 	/*
169 	 * Number of invalid entries dropped by the kernel due to
170 	 * invalid index stored in array
171 	 *
172 	 * Written by the kernel, shouldn't be modified by the
173 	 * application (i.e. get number of "new events" by comparing to
174 	 * cached value).
175 	 *
176 	 * After a new SQ head value was read by the application this
177 	 * counter includes all submissions that were dropped reaching
178 	 * the new SQ head (and possibly more).
179 	 */
180 	u32			sq_dropped;
181 	/*
182 	 * Runtime SQ flags
183 	 *
184 	 * Written by the kernel, shouldn't be modified by the
185 	 * application.
186 	 *
187 	 * The application needs a full memory barrier before checking
188 	 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
189 	 */
190 	atomic_t		sq_flags;
191 	/*
192 	 * Runtime CQ flags
193 	 *
194 	 * Written by the application, shouldn't be modified by the
195 	 * kernel.
196 	 */
197 	u32			cq_flags;
198 	/*
199 	 * Number of completion events lost because the queue was full;
200 	 * this should be avoided by the application by making sure
201 	 * there are not more requests pending than there is space in
202 	 * the completion queue.
203 	 *
204 	 * Written by the kernel, shouldn't be modified by the
205 	 * application (i.e. get number of "new events" by comparing to
206 	 * cached value).
207 	 *
208 	 * As completion events come in out of order this counter is not
209 	 * ordered with any other data.
210 	 */
211 	u32			cq_overflow;
212 	/*
213 	 * Ring buffer of completion events.
214 	 *
215 	 * The kernel writes completion events fresh every time they are
216 	 * produced, so the application is allowed to modify pending
217 	 * entries.
218 	 */
219 	struct io_uring_cqe	cqes[] ____cacheline_aligned_in_smp;
220 };
221 
222 struct io_bpf_filter;
223 struct io_bpf_filters {
224 	refcount_t refs;	/* ref for ->bpf_filters */
225 	spinlock_t lock;	/* protects ->bpf_filters modifications */
226 	struct io_bpf_filter __rcu **filters;
227 	struct rcu_head rcu_head;
228 };
229 
230 struct io_restriction {
231 	DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
232 	DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
233 	struct io_bpf_filters *bpf_filters;
234 	/* ->bpf_filters needs COW on modification */
235 	bool bpf_filters_cow;
236 	u8 sqe_flags_allowed;
237 	u8 sqe_flags_required;
238 	/* IORING_OP_* restrictions exist */
239 	bool op_registered;
240 	/* IORING_REGISTER_* restrictions exist */
241 	bool reg_registered;
242 };
243 
244 struct io_submit_link {
245 	struct io_kiocb		*head;
246 	struct io_kiocb		*last;
247 };
248 
249 struct io_submit_state {
250 	/* inline/task_work completion list, under ->uring_lock */
251 	struct io_wq_work_node	free_list;
252 	/* batch completion logic */
253 	struct io_wq_work_list	compl_reqs;
254 	struct io_submit_link	link;
255 
256 	bool			plug_started;
257 	bool			need_plug;
258 	bool			cq_flush;
259 	unsigned short		submit_nr;
260 	struct blk_plug		plug;
261 };
262 
263 struct io_alloc_cache {
264 	void			**entries;
265 	unsigned int		nr_cached;
266 	unsigned int		max_cached;
267 	unsigned int		elem_size;
268 	unsigned int		init_clear;
269 };
270 
271 struct io_ring_ctx {
272 	/* const or read-mostly hot data */
273 	struct {
274 		unsigned int		flags;
275 		unsigned int		drain_next: 1;
276 		unsigned int		op_restricted: 1;
277 		unsigned int		reg_restricted: 1;
278 		unsigned int		off_timeout_used: 1;
279 		unsigned int		drain_active: 1;
280 		unsigned int		has_evfd: 1;
281 		/* all CQEs should be posted only by the submitter task */
282 		unsigned int		task_complete: 1;
283 		unsigned int		lockless_cq: 1;
284 		unsigned int		syscall_iopoll: 1;
285 		unsigned int		poll_activated: 1;
286 		unsigned int		drain_disabled: 1;
287 		unsigned int		compat: 1;
288 		unsigned int		iowq_limits_set : 1;
289 
290 		struct task_struct	*submitter_task;
291 		struct io_rings		*rings;
292 		/* cache of ->restrictions.bpf_filters->filters */
293 		struct io_bpf_filter __rcu	**bpf_filters;
294 		struct percpu_ref	refs;
295 
296 		clockid_t		clockid;
297 		enum tk_offsets		clock_offset;
298 
299 		enum task_work_notify_mode	notify_method;
300 		unsigned			sq_thread_idle;
301 	} ____cacheline_aligned_in_smp;
302 
303 	/* submission data */
304 	struct {
305 		struct mutex		uring_lock;
306 
307 		/*
308 		 * Ring buffer of indices into array of io_uring_sqe, which is
309 		 * mmapped by the application using the IORING_OFF_SQES offset.
310 		 *
311 		 * This indirection could e.g. be used to assign fixed
312 		 * io_uring_sqe entries to operations and only submit them to
313 		 * the queue when needed.
314 		 *
315 		 * The kernel modifies neither the indices array nor the entries
316 		 * array.
317 		 */
318 		u32			*sq_array;
319 		struct io_uring_sqe	*sq_sqes;
320 		unsigned		cached_sq_head;
321 		unsigned		sq_entries;
322 
323 		/*
324 		 * Fixed resources fast path, should be accessed only under
325 		 * uring_lock, and updated through io_uring_register(2)
326 		 */
327 		atomic_t		cancel_seq;
328 
329 		/*
330 		 * ->iopoll_list is protected by the ctx->uring_lock for
331 		 * io_uring instances that don't use IORING_SETUP_SQPOLL.
332 		 * For SQPOLL, only the single threaded io_sq_thread() will
333 		 * manipulate the list, hence no extra locking is needed there.
334 		 */
335 		bool			poll_multi_queue;
336 		struct list_head	iopoll_list;
337 
338 		struct io_file_table	file_table;
339 		struct io_rsrc_data	buf_table;
340 		struct io_alloc_cache	node_cache;
341 		struct io_alloc_cache	imu_cache;
342 
343 		struct io_submit_state	submit_state;
344 
345 		/*
346 		 * Modifications are protected by ->uring_lock and ->mmap_lock.
347 		 * The buffer list's io mapped region should be stable once
348 		 * published.
349 		 */
350 		struct xarray		io_bl_xa;
351 
352 		struct io_hash_table	cancel_table;
353 		struct io_alloc_cache	apoll_cache;
354 		struct io_alloc_cache	netmsg_cache;
355 		struct io_alloc_cache	rw_cache;
356 		struct io_alloc_cache	cmd_cache;
357 
358 		/*
359 		 * Any cancelable uring_cmd is added to this list in
360 		 * ->uring_cmd() by io_uring_cmd_insert_cancelable()
361 		 */
362 		struct hlist_head	cancelable_uring_cmd;
363 		/*
364 		 * For Hybrid IOPOLL, runtime in hybrid polling, without
365 		 * scheduling time
366 		 */
367 		u64					hybrid_poll_time;
368 	} ____cacheline_aligned_in_smp;
369 
370 	struct {
371 		/*
372 		 * We cache a range of free CQEs we can use, once exhausted it
373 		 * should go through a slower range setup, see __io_get_cqe()
374 		 */
375 		struct io_uring_cqe	*cqe_cached;
376 		struct io_uring_cqe	*cqe_sentinel;
377 
378 		unsigned		cached_cq_tail;
379 		unsigned		cq_entries;
380 		struct io_ev_fd	__rcu	*io_ev_fd;
381 
382 		void			*cq_wait_arg;
383 		size_t			cq_wait_size;
384 	} ____cacheline_aligned_in_smp;
385 
386 	/*
387 	 * task_work and async notification delivery cacheline. Expected to
388 	 * regularly bounce b/w CPUs.
389 	 */
390 	struct {
391 		struct llist_head	work_llist;
392 		struct llist_head	retry_llist;
393 		unsigned long		check_cq;
394 		atomic_t		cq_wait_nr;
395 		atomic_t		cq_timeouts;
396 		struct wait_queue_head	cq_wait;
397 	} ____cacheline_aligned_in_smp;
398 
399 	/* timeouts */
400 	struct {
401 		raw_spinlock_t		timeout_lock;
402 		struct list_head	timeout_list;
403 		struct list_head	ltimeout_list;
404 		unsigned		cq_last_tm_flush;
405 	} ____cacheline_aligned_in_smp;
406 
407 	spinlock_t		completion_lock;
408 
409 	struct list_head	cq_overflow_list;
410 
411 	struct hlist_head	waitid_list;
412 
413 #ifdef CONFIG_FUTEX
414 	struct hlist_head	futex_list;
415 	struct io_alloc_cache	futex_cache;
416 #endif
417 
418 	const struct cred	*sq_creds;	/* cred used for __io_sq_thread() */
419 	struct io_sq_data	*sq_data;	/* if using sq thread polling */
420 
421 	struct wait_queue_head	sqo_sq_wait;
422 	struct list_head	sqd_list;
423 
424 	unsigned int		file_alloc_start;
425 	unsigned int		file_alloc_end;
426 
427 	/* Keep this last, we don't need it for the fast path */
428 	struct wait_queue_head		poll_wq;
429 	struct io_restriction		restrictions;
430 
431 	/* Stores zcrx object pointers of type struct io_zcrx_ifq */
432 	struct xarray			zcrx_ctxs;
433 
434 	u32			pers_next;
435 	struct xarray		personalities;
436 
437 	/* hashed buffered write serialization */
438 	struct io_wq_hash		*hash_map;
439 
440 	/* Only used for accounting purposes */
441 	struct user_struct		*user;
442 	struct mm_struct		*mm_account;
443 
444 	/*
445 	 * List of tctx nodes for this ctx, protected by tctx_lock. For
446 	 * cancelation purposes, nests under uring_lock.
447 	 */
448 	struct list_head		tctx_list;
449 	struct mutex			tctx_lock;
450 
451 	/* ctx exit and cancelation */
452 	struct llist_head		fallback_llist;
453 	struct delayed_work		fallback_work;
454 	struct work_struct		exit_work;
455 	struct completion		ref_comp;
456 
457 	/* io-wq management, e.g. thread count */
458 	u32				iowq_limits[2];
459 
460 	struct callback_head		poll_wq_task_work;
461 	struct list_head		defer_list;
462 	unsigned			nr_drained;
463 
464 	/* protected by ->completion_lock */
465 	unsigned			nr_req_allocated;
466 
467 #ifdef CONFIG_NET_RX_BUSY_POLL
468 	struct list_head	napi_list;	/* track busy poll napi_id */
469 	spinlock_t		napi_lock;	/* napi_list lock */
470 
471 	/* napi busy poll default timeout */
472 	ktime_t			napi_busy_poll_dt;
473 	bool			napi_prefer_busy_poll;
474 	u8			napi_track_mode;
475 
476 	DECLARE_HASHTABLE(napi_ht, 4);
477 #endif
478 
479 	/*
480 	 * Protection for resize vs mmap races - both the mmap and resize
481 	 * side will need to grab this lock, to prevent either side from
482 	 * being run concurrently with the other.
483 	 */
484 	struct mutex			mmap_lock;
485 
486 	struct io_mapped_region		sq_region;
487 	struct io_mapped_region		ring_region;
488 	/* used for optimised request parameter and wait argument passing  */
489 	struct io_mapped_region		param_region;
490 };
491 
492 /*
493  * Token indicating function is called in task work context:
494  * ctx->uring_lock is held and any completions generated will be flushed.
495  * ONLY core io_uring.c should instantiate this struct.
496  */
497 struct io_tw_state {
498 	bool cancel;
499 };
500 /* Alias to use in code that doesn't instantiate struct io_tw_state */
501 typedef struct io_tw_state io_tw_token_t;
502 
503 enum {
504 	REQ_F_FIXED_FILE_BIT	= IOSQE_FIXED_FILE_BIT,
505 	REQ_F_IO_DRAIN_BIT	= IOSQE_IO_DRAIN_BIT,
506 	REQ_F_LINK_BIT		= IOSQE_IO_LINK_BIT,
507 	REQ_F_HARDLINK_BIT	= IOSQE_IO_HARDLINK_BIT,
508 	REQ_F_FORCE_ASYNC_BIT	= IOSQE_ASYNC_BIT,
509 	REQ_F_BUFFER_SELECT_BIT	= IOSQE_BUFFER_SELECT_BIT,
510 	REQ_F_CQE_SKIP_BIT	= IOSQE_CQE_SKIP_SUCCESS_BIT,
511 
512 	/* first byte is taken by user flags, shift it to not overlap */
513 	REQ_F_FAIL_BIT		= 8,
514 	REQ_F_INFLIGHT_BIT,
515 	REQ_F_CUR_POS_BIT,
516 	REQ_F_NOWAIT_BIT,
517 	REQ_F_LINK_TIMEOUT_BIT,
518 	REQ_F_NEED_CLEANUP_BIT,
519 	REQ_F_POLLED_BIT,
520 	REQ_F_HYBRID_IOPOLL_STATE_BIT,
521 	REQ_F_BUFFER_SELECTED_BIT,
522 	REQ_F_BUFFER_RING_BIT,
523 	REQ_F_REISSUE_BIT,
524 	REQ_F_CREDS_BIT,
525 	REQ_F_REFCOUNT_BIT,
526 	REQ_F_ARM_LTIMEOUT_BIT,
527 	REQ_F_ASYNC_DATA_BIT,
528 	REQ_F_SKIP_LINK_CQES_BIT,
529 	REQ_F_SINGLE_POLL_BIT,
530 	REQ_F_DOUBLE_POLL_BIT,
531 	REQ_F_MULTISHOT_BIT,
532 	REQ_F_APOLL_MULTISHOT_BIT,
533 	REQ_F_CLEAR_POLLIN_BIT,
534 	/* keep async read/write and isreg together and in order */
535 	REQ_F_SUPPORT_NOWAIT_BIT,
536 	REQ_F_ISREG_BIT,
537 	REQ_F_POLL_NO_LAZY_BIT,
538 	REQ_F_CAN_POLL_BIT,
539 	REQ_F_BL_EMPTY_BIT,
540 	REQ_F_BL_NO_RECYCLE_BIT,
541 	REQ_F_BUFFERS_COMMIT_BIT,
542 	REQ_F_BUF_NODE_BIT,
543 	REQ_F_HAS_METADATA_BIT,
544 	REQ_F_IMPORT_BUFFER_BIT,
545 	REQ_F_SQE_COPIED_BIT,
546 
547 	/* not a real bit, just to check we're not overflowing the space */
548 	__REQ_F_LAST_BIT,
549 };
550 
551 typedef u64 __bitwise io_req_flags_t;
552 #define IO_REQ_FLAG(bitno)	((__force io_req_flags_t) BIT_ULL((bitno)))
553 
554 enum {
555 	/* ctx owns file */
556 	REQ_F_FIXED_FILE	= IO_REQ_FLAG(REQ_F_FIXED_FILE_BIT),
557 	/* drain existing IO first */
558 	REQ_F_IO_DRAIN		= IO_REQ_FLAG(REQ_F_IO_DRAIN_BIT),
559 	/* linked sqes */
560 	REQ_F_LINK		= IO_REQ_FLAG(REQ_F_LINK_BIT),
561 	/* doesn't sever on completion < 0 */
562 	REQ_F_HARDLINK		= IO_REQ_FLAG(REQ_F_HARDLINK_BIT),
563 	/* IOSQE_ASYNC */
564 	REQ_F_FORCE_ASYNC	= IO_REQ_FLAG(REQ_F_FORCE_ASYNC_BIT),
565 	/* IOSQE_BUFFER_SELECT */
566 	REQ_F_BUFFER_SELECT	= IO_REQ_FLAG(REQ_F_BUFFER_SELECT_BIT),
567 	/* IOSQE_CQE_SKIP_SUCCESS */
568 	REQ_F_CQE_SKIP		= IO_REQ_FLAG(REQ_F_CQE_SKIP_BIT),
569 
570 	/* fail rest of links */
571 	REQ_F_FAIL		= IO_REQ_FLAG(REQ_F_FAIL_BIT),
572 	/* on inflight list, should be cancelled and waited on exit reliably */
573 	REQ_F_INFLIGHT		= IO_REQ_FLAG(REQ_F_INFLIGHT_BIT),
574 	/* read/write uses file position */
575 	REQ_F_CUR_POS		= IO_REQ_FLAG(REQ_F_CUR_POS_BIT),
576 	/* must not punt to workers */
577 	REQ_F_NOWAIT		= IO_REQ_FLAG(REQ_F_NOWAIT_BIT),
578 	/* has or had linked timeout */
579 	REQ_F_LINK_TIMEOUT	= IO_REQ_FLAG(REQ_F_LINK_TIMEOUT_BIT),
580 	/* needs cleanup */
581 	REQ_F_NEED_CLEANUP	= IO_REQ_FLAG(REQ_F_NEED_CLEANUP_BIT),
582 	/* already went through poll handler */
583 	REQ_F_POLLED		= IO_REQ_FLAG(REQ_F_POLLED_BIT),
584 	/* every req only blocks once in hybrid poll */
585 	REQ_F_IOPOLL_STATE        = IO_REQ_FLAG(REQ_F_HYBRID_IOPOLL_STATE_BIT),
586 	/* buffer already selected */
587 	REQ_F_BUFFER_SELECTED	= IO_REQ_FLAG(REQ_F_BUFFER_SELECTED_BIT),
588 	/* buffer selected from ring, needs commit */
589 	REQ_F_BUFFER_RING	= IO_REQ_FLAG(REQ_F_BUFFER_RING_BIT),
590 	/* caller should reissue async */
591 	REQ_F_REISSUE		= IO_REQ_FLAG(REQ_F_REISSUE_BIT),
592 	/* supports async reads/writes */
593 	REQ_F_SUPPORT_NOWAIT	= IO_REQ_FLAG(REQ_F_SUPPORT_NOWAIT_BIT),
594 	/* regular file */
595 	REQ_F_ISREG		= IO_REQ_FLAG(REQ_F_ISREG_BIT),
596 	/* has creds assigned */
597 	REQ_F_CREDS		= IO_REQ_FLAG(REQ_F_CREDS_BIT),
598 	/* skip refcounting if not set */
599 	REQ_F_REFCOUNT		= IO_REQ_FLAG(REQ_F_REFCOUNT_BIT),
600 	/* there is a linked timeout that has to be armed */
601 	REQ_F_ARM_LTIMEOUT	= IO_REQ_FLAG(REQ_F_ARM_LTIMEOUT_BIT),
602 	/* ->async_data allocated */
603 	REQ_F_ASYNC_DATA	= IO_REQ_FLAG(REQ_F_ASYNC_DATA_BIT),
604 	/* don't post CQEs while failing linked requests */
605 	REQ_F_SKIP_LINK_CQES	= IO_REQ_FLAG(REQ_F_SKIP_LINK_CQES_BIT),
606 	/* single poll may be active */
607 	REQ_F_SINGLE_POLL	= IO_REQ_FLAG(REQ_F_SINGLE_POLL_BIT),
608 	/* double poll may active */
609 	REQ_F_DOUBLE_POLL	= IO_REQ_FLAG(REQ_F_DOUBLE_POLL_BIT),
610 	/* request posts multiple completions, should be set at prep time */
611 	REQ_F_MULTISHOT		= IO_REQ_FLAG(REQ_F_MULTISHOT_BIT),
612 	/* fast poll multishot mode */
613 	REQ_F_APOLL_MULTISHOT	= IO_REQ_FLAG(REQ_F_APOLL_MULTISHOT_BIT),
614 	/* recvmsg special flag, clear EPOLLIN */
615 	REQ_F_CLEAR_POLLIN	= IO_REQ_FLAG(REQ_F_CLEAR_POLLIN_BIT),
616 	/* don't use lazy poll wake for this request */
617 	REQ_F_POLL_NO_LAZY	= IO_REQ_FLAG(REQ_F_POLL_NO_LAZY_BIT),
618 	/* file is pollable */
619 	REQ_F_CAN_POLL		= IO_REQ_FLAG(REQ_F_CAN_POLL_BIT),
620 	/* buffer list was empty after selection of buffer */
621 	REQ_F_BL_EMPTY		= IO_REQ_FLAG(REQ_F_BL_EMPTY_BIT),
622 	/* don't recycle provided buffers for this request */
623 	REQ_F_BL_NO_RECYCLE	= IO_REQ_FLAG(REQ_F_BL_NO_RECYCLE_BIT),
624 	/* buffer ring head needs incrementing on put */
625 	REQ_F_BUFFERS_COMMIT	= IO_REQ_FLAG(REQ_F_BUFFERS_COMMIT_BIT),
626 	/* buf node is valid */
627 	REQ_F_BUF_NODE		= IO_REQ_FLAG(REQ_F_BUF_NODE_BIT),
628 	/* request has read/write metadata assigned */
629 	REQ_F_HAS_METADATA	= IO_REQ_FLAG(REQ_F_HAS_METADATA_BIT),
630 	/*
631 	 * For vectored fixed buffers, resolve iovec to registered buffers.
632 	 * For SEND_ZC, whether to import buffers (i.e. the first issue).
633 	 */
634 	REQ_F_IMPORT_BUFFER	= IO_REQ_FLAG(REQ_F_IMPORT_BUFFER_BIT),
635 	/* ->sqe_copy() has been called, if necessary */
636 	REQ_F_SQE_COPIED	= IO_REQ_FLAG(REQ_F_SQE_COPIED_BIT),
637 };
638 
639 struct io_tw_req {
640 	struct io_kiocb *req;
641 };
642 
643 typedef void (*io_req_tw_func_t)(struct io_tw_req tw_req, io_tw_token_t tw);
644 
645 struct io_task_work {
646 	struct llist_node		node;
647 	io_req_tw_func_t		func;
648 };
649 
650 struct io_cqe {
651 	__u64	user_data;
652 	__s32	res;
653 	/* fd initially, then cflags for completion */
654 	union {
655 		__u32	flags;
656 		int	fd;
657 	};
658 };
659 
660 /*
661  * Each request type overlays its private data structure on top of this one.
662  * They must not exceed this one in size.
663  */
664 struct io_cmd_data {
665 	struct file		*file;
666 	/* each command gets 56 bytes of data */
667 	__u8			data[56];
668 };
669 
670 static inline void io_kiocb_cmd_sz_check(size_t cmd_sz)
671 {
672 	BUILD_BUG_ON(cmd_sz > sizeof(struct io_cmd_data));
673 }
674 #define io_kiocb_to_cmd(req, cmd_type) ( \
675 	io_kiocb_cmd_sz_check(sizeof(cmd_type)) , \
676 	((cmd_type *)&(req)->cmd) \
677 )
678 
679 static inline struct io_kiocb *cmd_to_io_kiocb(void *ptr)
680 {
681 	return ptr;
682 }
683 
684 struct io_kiocb {
685 	union {
686 		/*
687 		 * NOTE! Each of the io_kiocb union members has the file pointer
688 		 * as the first entry in their struct definition. So you can
689 		 * access the file pointer through any of the sub-structs,
690 		 * or directly as just 'file' in this struct.
691 		 */
692 		struct file		*file;
693 		struct io_cmd_data	cmd;
694 	};
695 
696 	u8				opcode;
697 	/* polled IO has completed */
698 	u8				iopoll_completed;
699 	/*
700 	 * Can be either a fixed buffer index, or used with provided buffers.
701 	 * For the latter, it points to the selected buffer ID.
702 	 */
703 	u16				buf_index;
704 
705 	unsigned			nr_tw;
706 
707 	/* REQ_F_* flags */
708 	io_req_flags_t			flags;
709 
710 	struct io_cqe			cqe;
711 
712 	struct io_ring_ctx		*ctx;
713 	struct io_uring_task		*tctx;
714 
715 	union {
716 		/* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
717 		struct io_buffer	*kbuf;
718 
719 		struct io_rsrc_node	*buf_node;
720 	};
721 
722 	union {
723 		/* used by request caches, completion batching and iopoll */
724 		struct io_wq_work_node	comp_list;
725 		/* cache ->apoll->events */
726 		__poll_t apoll_events;
727 	};
728 
729 	struct io_rsrc_node		*file_node;
730 
731 	atomic_t			refs;
732 	bool				cancel_seq_set;
733 
734 	union {
735 		struct io_task_work	io_task_work;
736 		/* For IOPOLL setup queues, with hybrid polling */
737 		u64                     iopoll_start;
738 	};
739 
740 	union {
741 		/*
742 		 * for polled requests, i.e. IORING_OP_POLL_ADD and async armed
743 		 * poll
744 		 */
745 		struct hlist_node	hash_node;
746 		/* IOPOLL completion handling */
747 		struct list_head	iopoll_node;
748 		/* for private io_kiocb freeing */
749 		struct rcu_head		rcu_head;
750 	};
751 	/* internal polling, see IORING_FEAT_FAST_POLL */
752 	struct async_poll		*apoll;
753 	/* opcode allocated if it needs to store data for async defer */
754 	void				*async_data;
755 	/* linked requests, IFF REQ_F_HARDLINK or REQ_F_LINK are set */
756 	atomic_t			poll_refs;
757 	struct io_kiocb			*link;
758 	/* custom credentials, valid IFF REQ_F_CREDS is set */
759 	const struct cred		*creds;
760 	struct io_wq_work		work;
761 
762 	struct io_big_cqe {
763 		u64			extra1;
764 		u64			extra2;
765 	} big_cqe;
766 };
767 
768 struct io_overflow_cqe {
769 	struct list_head list;
770 	struct io_uring_cqe cqe;
771 };
772 #endif
773