closure.h

  1/* SPDX-License-Identifier: GPL-2.0 */
  2#ifndef _LINUX_CLOSURE_H
  3#define _LINUX_CLOSURE_H
  4
  5#include <linux/llist.h>
  6#include <linux/sched.h>
  7#include <linux/sched/task_stack.h>
  8#include <linux/workqueue.h>
  9
 10/*
 11 * Closure is perhaps the most overused and abused term in computer science, but
 12 * since I've been unable to come up with anything better you're stuck with it
 13 * again.
 14 *
 15 * What are closures?
 16 *
 17 * They embed a refcount. The basic idea is they count "things that are in
 18 * progress" - in flight bios, some other thread that's doing something else -
 19 * anything you might want to wait on.
 20 *
 21 * The refcount may be manipulated with closure_get() and closure_put().
 22 * closure_put() is where many of the interesting things happen, when it causes
 23 * the refcount to go to 0.
 24 *
 25 * Closures can be used to wait on things both synchronously and asynchronously,
 26 * and synchronous and asynchronous use can be mixed without restriction. To
 27 * wait synchronously, use closure_sync() - you will sleep until your closure's
 28 * refcount hits 1.
 29 *
 30 * To wait asynchronously, use
 31 *   continue_at(cl, next_function, workqueue);
 32 *
 33 * passing it, as you might expect, the function to run when nothing is pending
 34 * and the workqueue to run that function out of.
 35 *
 36 * continue_at() also, critically, requires a 'return' immediately following the
 37 * location where this macro is referenced, to return to the calling function.
 38 * There's good reason for this.
 39 *
 40 * To use safely closures asynchronously, they must always have a refcount while
 41 * they are running owned by the thread that is running them. Otherwise, suppose
 42 * you submit some bios and wish to have a function run when they all complete:
 43 *
 44 * foo_endio(struct bio *bio)
 45 * {
 46 *	closure_put(cl);
 47 * }
 48 *
 49 * closure_init(cl);
 50 *
 51 * do_stuff();
 52 * closure_get(cl);
 53 * bio1->bi_endio = foo_endio;
 54 * bio_submit(bio1);
 55 *
 56 * do_more_stuff();
 57 * closure_get(cl);
 58 * bio2->bi_endio = foo_endio;
 59 * bio_submit(bio2);
 60 *
 61 * continue_at(cl, complete_some_read, system_wq);
 62 *
 63 * If closure's refcount started at 0, complete_some_read() could run before the
 64 * second bio was submitted - which is almost always not what you want! More
 65 * importantly, it wouldn't be possible to say whether the original thread or
 66 * complete_some_read()'s thread owned the closure - and whatever state it was
 67 * associated with!
 68 *
 69 * So, closure_init() initializes a closure's refcount to 1 - and when a
 70 * closure_fn is run, the refcount will be reset to 1 first.
 71 *
 72 * Then, the rule is - if you got the refcount with closure_get(), release it
 73 * with closure_put() (i.e, in a bio->bi_endio function). If you have a refcount
 74 * on a closure because you called closure_init() or you were run out of a
 75 * closure - _always_ use continue_at(). Doing so consistently will help
 76 * eliminate an entire class of particularly pernicious races.
 77 *
 78 * Lastly, you might have a wait list dedicated to a specific event, and have no
 79 * need for specifying the condition - you just want to wait until someone runs
 80 * closure_wake_up() on the appropriate wait list. In that case, just use
 81 * closure_wait(). It will return either true or false, depending on whether the
 82 * closure was already on a wait list or not - a closure can only be on one wait
 83 * list at a time.
 84 *
 85 * Parents:
 86 *
 87 * closure_init() takes two arguments - it takes the closure to initialize, and
 88 * a (possibly null) parent.
 89 *
 90 * If parent is non null, the new closure will have a refcount for its lifetime;
 91 * a closure is considered to be "finished" when its refcount hits 0 and the
 92 * function to run is null. Hence
 93 *
 94 * continue_at(cl, NULL, NULL);
 95 *
 96 * returns up the (spaghetti) stack of closures, precisely like normal return
 97 * returns up the C stack. continue_at() with non null fn is better thought of
 98 * as doing a tail call.
 99 *
100 * All this implies that a closure should typically be embedded in a particular
101 * struct (which its refcount will normally control the lifetime of), and that
102 * struct can very much be thought of as a stack frame.
103 */
104
105struct closure;
106struct closure_syncer;
107typedef void (closure_fn) (struct work_struct *);
108extern struct dentry *bcache_debug;
109
110struct closure_waitlist {
111	struct llist_head	list;
112};
113
114enum closure_state {
115	/*
116	 * CLOSURE_WAITING: Set iff the closure is on a waitlist. Must be set by
117	 * the thread that owns the closure, and cleared by the thread that's
118	 * waking up the closure.
119	 *
120	 * The rest are for debugging and don't affect behaviour:
121	 *
122	 * CLOSURE_RUNNING: Set when a closure is running (i.e. by
123	 * closure_init() and when closure_put() runs then next function), and
124	 * must be cleared before remaining hits 0. Primarily to help guard
125	 * against incorrect usage and accidentally transferring references.
126	 * continue_at() and closure_return() clear it for you, if you're doing
127	 * something unusual you can use closure_set_dead() which also helps
128	 * annotate where references are being transferred.
129	 */
130
131	CLOSURE_BITS_START	= (1U << 26),
132	CLOSURE_DESTRUCTOR	= (1U << 26),
133	CLOSURE_WAITING		= (1U << 28),
134	CLOSURE_RUNNING		= (1U << 30),
135};
136
137#define CLOSURE_GUARD_MASK					\
138	((CLOSURE_DESTRUCTOR|CLOSURE_WAITING|CLOSURE_RUNNING) << 1)
139
140#define CLOSURE_REMAINING_MASK		(CLOSURE_BITS_START - 1)
141#define CLOSURE_REMAINING_INITIALIZER	(1|CLOSURE_RUNNING)
142
143struct closure {
144	union {
145		struct {
146			struct workqueue_struct *wq;
147			struct closure_syncer	*s;
148			struct llist_node	list;
149			closure_fn		*fn;
150		};
151		struct work_struct	work;
152	};
153
154	struct closure		*parent;
155
156	atomic_t		remaining;
157	bool			closure_get_happened;
158
159#ifdef CONFIG_DEBUG_CLOSURES
160#define CLOSURE_MAGIC_DEAD	0xc054dead
161#define CLOSURE_MAGIC_ALIVE	0xc054a11e
162#define CLOSURE_MAGIC_STACK	0xc05451cc
163
164	unsigned int		magic;
165	struct list_head	all;
166	unsigned long		ip;
167	unsigned long		waiting_on;
168#endif
169};
170
171void closure_sub(struct closure *cl, int v);
172void closure_put(struct closure *cl);
173void __closure_wake_up(struct closure_waitlist *list);
174bool closure_wait(struct closure_waitlist *list, struct closure *cl);
175void __closure_sync(struct closure *cl);
176
177static inline unsigned closure_nr_remaining(struct closure *cl)
178{
179	return atomic_read(&cl->remaining) & CLOSURE_REMAINING_MASK;
180}
181
182/**
183 * closure_sync - sleep until a closure a closure has nothing left to wait on
184 *
185 * Sleeps until the refcount hits 1 - the thread that's running the closure owns
186 * the last refcount.
187 */
188static inline void closure_sync(struct closure *cl)
189{
190#ifdef CONFIG_DEBUG_CLOSURES
191	BUG_ON(closure_nr_remaining(cl) != 1 && !cl->closure_get_happened);
192#endif
193
194	if (cl->closure_get_happened)
195		__closure_sync(cl);
196}
197
198int __closure_sync_timeout(struct closure *cl, unsigned long timeout);
199
200static inline int closure_sync_timeout(struct closure *cl, unsigned long timeout)
201{
202#ifdef CONFIG_DEBUG_CLOSURES
203	BUG_ON(closure_nr_remaining(cl) != 1 && !cl->closure_get_happened);
204#endif
205	return cl->closure_get_happened
206		? __closure_sync_timeout(cl, timeout)
207		: 0;
208}
209
210#ifdef CONFIG_DEBUG_CLOSURES
211
212void closure_debug_create(struct closure *cl);
213void closure_debug_destroy(struct closure *cl);
214
215#else
216
217static inline void closure_debug_create(struct closure *cl) {}
218static inline void closure_debug_destroy(struct closure *cl) {}
219
220#endif
221
222static inline void closure_set_ip(struct closure *cl)
223{
224#ifdef CONFIG_DEBUG_CLOSURES
225	cl->ip = _THIS_IP_;
226#endif
227}
228
229static inline void closure_set_ret_ip(struct closure *cl)
230{
231#ifdef CONFIG_DEBUG_CLOSURES
232	cl->ip = _RET_IP_;
233#endif
234}
235
236static inline void closure_set_waiting(struct closure *cl, unsigned long f)
237{
238#ifdef CONFIG_DEBUG_CLOSURES
239	cl->waiting_on = f;
240#endif
241}
242
243static inline void closure_set_stopped(struct closure *cl)
244{
245	atomic_sub(CLOSURE_RUNNING, &cl->remaining);
246}
247
248static inline void set_closure_fn(struct closure *cl, closure_fn *fn,
249				  struct workqueue_struct *wq)
250{
251	closure_set_ip(cl);
252	cl->fn = fn;
253	cl->wq = wq;
254}
255
256static inline void closure_queue(struct closure *cl)
257{
258	struct workqueue_struct *wq = cl->wq;
259	/**
260	 * Changes made to closure, work_struct, or a couple of other structs
261	 * may cause work.func not pointing to the right location.
262	 */
263	BUILD_BUG_ON(offsetof(struct closure, fn)
264		     != offsetof(struct work_struct, func));
265
266	if (wq) {
267		INIT_WORK(&cl->work, cl->work.func);
268		BUG_ON(!queue_work(wq, &cl->work));
269	} else
270		cl->fn(&cl->work);
271}
272
273/**
274 * closure_get - increment a closure's refcount
275 */
276static inline void closure_get(struct closure *cl)
277{
278	cl->closure_get_happened = true;
279
280#ifdef CONFIG_DEBUG_CLOSURES
281	BUG_ON((atomic_inc_return(&cl->remaining) &
282		CLOSURE_REMAINING_MASK) <= 1);
283#else
284	atomic_inc(&cl->remaining);
285#endif
286}
287
288/**
289 * closure_get_not_zero
290 */
291static inline bool closure_get_not_zero(struct closure *cl)
292{
293	unsigned old = atomic_read(&cl->remaining);
294	do {
295		if (!(old & CLOSURE_REMAINING_MASK))
296			return false;
297
298	} while (!atomic_try_cmpxchg_acquire(&cl->remaining, &old, old + 1));
299
300	return true;
301}
302
303/**
304 * closure_init - Initialize a closure, setting the refcount to 1
305 * @cl:		closure to initialize
306 * @parent:	parent of the new closure. cl will take a refcount on it for its
307 *		lifetime; may be NULL.
308 */
309static inline void closure_init(struct closure *cl, struct closure *parent)
310{
311	cl->fn = NULL;
312	cl->parent = parent;
313	if (parent)
314		closure_get(parent);
315
316	atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER);
317	cl->closure_get_happened = false;
318
319	closure_debug_create(cl);
320	closure_set_ip(cl);
321}
322
323static inline void closure_init_stack(struct closure *cl)
324{
325	memset(cl, 0, sizeof(struct closure));
326	atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER);
327#ifdef CONFIG_DEBUG_CLOSURES
328	cl->magic = CLOSURE_MAGIC_STACK;
329#endif
330}
331
332static inline void closure_init_stack_release(struct closure *cl)
333{
334	memset(cl, 0, sizeof(struct closure));
335	atomic_set_release(&cl->remaining, CLOSURE_REMAINING_INITIALIZER);
336#ifdef CONFIG_DEBUG_CLOSURES
337	cl->magic = CLOSURE_MAGIC_STACK;
338#endif
339}
340
341/**
342 * closure_wake_up - wake up all closures on a wait list,
343 *		     with memory barrier
344 */
345static inline void closure_wake_up(struct closure_waitlist *list)
346{
347	/* Memory barrier for the wait list */
348	smp_mb();
349	__closure_wake_up(list);
350}
351
352#define CLOSURE_CALLBACK(name)	void name(struct work_struct *ws)
353#define closure_type(name, type, member)				\
354	struct closure *cl = container_of(ws, struct closure, work);	\
355	type *name = container_of(cl, type, member)
356
357/**
358 * continue_at - jump to another function with barrier
359 *
360 * After @cl is no longer waiting on anything (i.e. all outstanding refs have
361 * been dropped with closure_put()), it will resume execution at @fn running out
362 * of @wq (or, if @wq is NULL, @fn will be called by closure_put() directly).
363 *
364 * This is because after calling continue_at() you no longer have a ref on @cl,
365 * and whatever @cl owns may be freed out from under you - a running closure fn
366 * has a ref on its own closure which continue_at() drops.
367 *
368 * Note you are expected to immediately return after using this macro.
369 */
370#define continue_at(_cl, _fn, _wq)					\
371do {									\
372	set_closure_fn(_cl, _fn, _wq);					\
373	closure_sub(_cl, CLOSURE_RUNNING + 1);				\
374} while (0)
375
376/**
377 * closure_return - finish execution of a closure
378 *
379 * This is used to indicate that @cl is finished: when all outstanding refs on
380 * @cl have been dropped @cl's ref on its parent closure (as passed to
381 * closure_init()) will be dropped, if one was specified - thus this can be
382 * thought of as returning to the parent closure.
383 */
384#define closure_return(_cl)	continue_at((_cl), NULL, NULL)
385
386void closure_return_sync(struct closure *cl);
387
388/**
389 * continue_at_nobarrier - jump to another function without barrier
390 *
391 * Causes @fn to be executed out of @cl, in @wq context (or called directly if
392 * @wq is NULL).
393 *
394 * The ref the caller of continue_at_nobarrier() had on @cl is now owned by @fn,
395 * thus it's not safe to touch anything protected by @cl after a
396 * continue_at_nobarrier().
397 */
398#define continue_at_nobarrier(_cl, _fn, _wq)				\
399do {									\
400	set_closure_fn(_cl, _fn, _wq);					\
401	closure_queue(_cl);						\
402} while (0)
403
404/**
405 * closure_return_with_destructor - finish execution of a closure,
406 *				    with destructor
407 *
408 * Works like closure_return(), except @destructor will be called when all
409 * outstanding refs on @cl have been dropped; @destructor may be used to safely
410 * free the memory occupied by @cl, and it is called with the ref on the parent
411 * closure still held - so @destructor could safely return an item to a
412 * freelist protected by @cl's parent.
413 */
414#define closure_return_with_destructor(_cl, _destructor)		\
415do {									\
416	set_closure_fn(_cl, _destructor, NULL);				\
417	closure_sub(_cl, CLOSURE_RUNNING - CLOSURE_DESTRUCTOR + 1);	\
418} while (0)
419
420/**
421 * closure_call - execute @fn out of a new, uninitialized closure
422 *
423 * Typically used when running out of one closure, and we want to run @fn
424 * asynchronously out of a new closure - @parent will then wait for @cl to
425 * finish.
426 */
427static inline void closure_call(struct closure *cl, closure_fn fn,
428				struct workqueue_struct *wq,
429				struct closure *parent)
430{
431	closure_init(cl, parent);
432	continue_at_nobarrier(cl, fn, wq);
433}
434
435#define __closure_wait_event(waitlist, _cond)				\
436do {									\
437	struct closure cl;						\
438									\
439	closure_init_stack(&cl);					\
440									\
441	while (1) {							\
442		closure_wait(waitlist, &cl);				\
443		if (_cond)						\
444			break;						\
445		closure_sync(&cl);					\
446	}								\
447	closure_wake_up(waitlist);					\
448	closure_sync(&cl);						\
449} while (0)
450
451#define closure_wait_event(waitlist, _cond)				\
452do {									\
453	if (!(_cond))							\
454		__closure_wait_event(waitlist, _cond);			\
455} while (0)
456
457#define __closure_wait_event_timeout(waitlist, _cond, _until)		\
458({									\
459	struct closure cl;						\
460	long _t;							\
461									\
462	closure_init_stack(&cl);					\
463									\
464	while (1) {							\
465		closure_wait(waitlist, &cl);				\
466		if (_cond) {						\
467			_t = max_t(long, 1L, _until - jiffies);		\
468			break;						\
469		}							\
470		_t = max_t(long, 0L, _until - jiffies);			\
471		if (!_t)						\
472			break;						\
473		closure_sync_timeout(&cl, _t);				\
474	}								\
475	closure_wake_up(waitlist);					\
476	closure_sync(&cl);						\
477	_t;								\
478})
479
480/*
481 * Returns 0 if timeout expired, remaining time in jiffies (at least 1) if
482 * condition became true
483 */
484#define closure_wait_event_timeout(waitlist, _cond, _timeout)		\
485({									\
486	unsigned long _until = jiffies + _timeout;			\
487	(_cond)								\
488		? max_t(long, 1L, _until - jiffies)			\
489		: __closure_wait_event_timeout(waitlist, _cond, _until);\
490})
491
492#endif /* _LINUX_CLOSURE_H */