1#ifndef _LINUX_CLOSURE_H
  2#define _LINUX_CLOSURE_H
  3
  4#include <linux/llist.h>
  5#include <linux/sched.h>
  6#include <linux/workqueue.h>
  7
  8/*
  9 * Closure is perhaps the most overused and abused term in computer science, but
 10 * since I've been unable to come up with anything better you're stuck with it
 11 * again.
 12 *
 13 * What are closures?
 14 *
 15 * They embed a refcount. The basic idea is they count "things that are in
 16 * progress" - in flight bios, some other thread that's doing something else -
 17 * anything you might want to wait on.
 18 *
 19 * The refcount may be manipulated with closure_get() and closure_put().
 20 * closure_put() is where many of the interesting things happen, when it causes
 21 * the refcount to go to 0.
 22 *
 23 * Closures can be used to wait on things both synchronously and asynchronously,
 24 * and synchronous and asynchronous use can be mixed without restriction. To
 25 * wait synchronously, use closure_sync() - you will sleep until your closure's
 26 * refcount hits 1.
 27 *
 28 * To wait asynchronously, use
 29 *   continue_at(cl, next_function, workqueue);
 30 *
 31 * passing it, as you might expect, the function to run when nothing is pending
 32 * and the workqueue to run that function out of.
 33 *
 34 * continue_at() also, critically, requires a 'return' immediately following the
 35 * location where this macro is referenced, to return to the calling function.
 36 * There's good reason for this.
 37 *
 38 * To use safely closures asynchronously, they must always have a refcount while
 39 * they are running owned by the thread that is running them. Otherwise, suppose
 40 * you submit some bios and wish to have a function run when they all complete:
 41 *
 42 * foo_endio(struct bio *bio)
 43 * {
 44 *	closure_put(cl);
 45 * }
 46 *
 47 * closure_init(cl);
 48 *
 49 * do_stuff();
 50 * closure_get(cl);
 51 * bio1->bi_endio = foo_endio;
 52 * bio_submit(bio1);
 53 *
 54 * do_more_stuff();
 55 * closure_get(cl);
 56 * bio2->bi_endio = foo_endio;
 57 * bio_submit(bio2);
 58 *
 59 * continue_at(cl, complete_some_read, system_wq);
 60 *
 61 * If closure's refcount started at 0, complete_some_read() could run before the
 62 * second bio was submitted - which is almost always not what you want! More
 63 * importantly, it wouldn't be possible to say whether the original thread or
 64 * complete_some_read()'s thread owned the closure - and whatever state it was
 65 * associated with!
 66 *
 67 * So, closure_init() initializes a closure's refcount to 1 - and when a
 68 * closure_fn is run, the refcount will be reset to 1 first.
 69 *
 70 * Then, the rule is - if you got the refcount with closure_get(), release it
 71 * with closure_put() (i.e, in a bio->bi_endio function). If you have a refcount
 72 * on a closure because you called closure_init() or you were run out of a
 73 * closure - _always_ use continue_at(). Doing so consistently will help
 74 * eliminate an entire class of particularly pernicious races.
 75 *
 76 * Lastly, you might have a wait list dedicated to a specific event, and have no
 77 * need for specifying the condition - you just want to wait until someone runs
 78 * closure_wake_up() on the appropriate wait list. In that case, just use
 79 * closure_wait(). It will return either true or false, depending on whether the
 80 * closure was already on a wait list or not - a closure can only be on one wait
 81 * list at a time.
 82 *
 83 * Parents:
 84 *
 85 * closure_init() takes two arguments - it takes the closure to initialize, and
 86 * a (possibly null) parent.
 87 *
 88 * If parent is non null, the new closure will have a refcount for its lifetime;
 89 * a closure is considered to be "finished" when its refcount hits 0 and the
 90 * function to run is null. Hence
 91 *
 92 * continue_at(cl, NULL, NULL);
 93 *
 94 * returns up the (spaghetti) stack of closures, precisely like normal return
 95 * returns up the C stack. continue_at() with non null fn is better thought of
 96 * as doing a tail call.
 97 *
 98 * All this implies that a closure should typically be embedded in a particular
 99 * struct (which its refcount will normally control the lifetime of), and that
100 * struct can very much be thought of as a stack frame.
101 */
102
103struct closure;
104typedef void (closure_fn) (struct closure *);
105
106struct closure_waitlist {
107	struct llist_head	list;
108};
109
110enum closure_state {
111	/*
112	 * CLOSURE_WAITING: Set iff the closure is on a waitlist. Must be set by
113	 * the thread that owns the closure, and cleared by the thread that's
114	 * waking up the closure.
115	 *
116	 * CLOSURE_SLEEPING: Must be set before a thread uses a closure to sleep
117	 * - indicates that cl->task is valid and closure_put() may wake it up.
118	 * Only set or cleared by the thread that owns 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	 * CLOSURE_STACK: Sanity check - remaining should never hit 0 on a
131	 * closure with this flag set
132	 */
133
134	CLOSURE_BITS_START	= (1 << 23),
135	CLOSURE_DESTRUCTOR	= (1 << 23),
136	CLOSURE_WAITING		= (1 << 25),
137	CLOSURE_SLEEPING	= (1 << 27),
138	CLOSURE_RUNNING		= (1 << 29),
139	CLOSURE_STACK		= (1 << 31),
140};
141
142#define CLOSURE_GUARD_MASK					\
143	((CLOSURE_DESTRUCTOR|CLOSURE_WAITING|CLOSURE_SLEEPING|	\
144	  CLOSURE_RUNNING|CLOSURE_STACK) << 1)
145
146#define CLOSURE_REMAINING_MASK		(CLOSURE_BITS_START - 1)
147#define CLOSURE_REMAINING_INITIALIZER	(1|CLOSURE_RUNNING)
148
149struct closure {
150	union {
151		struct {
152			struct workqueue_struct *wq;
153			struct task_struct	*task;
154			struct llist_node	list;
155			closure_fn		*fn;
156		};
157		struct work_struct	work;
158	};
159
160	struct closure		*parent;
161
162	atomic_t		remaining;
163
164#ifdef CONFIG_BCACHE_CLOSURES_DEBUG
165#define CLOSURE_MAGIC_DEAD	0xc054dead
166#define CLOSURE_MAGIC_ALIVE	0xc054a11e
167
168	unsigned		magic;
169	struct list_head	all;
170	unsigned long		ip;
171	unsigned long		waiting_on;
172#endif
173};
174
175void closure_sub(struct closure *cl, int v);
176void closure_put(struct closure *cl);
177void __closure_wake_up(struct closure_waitlist *list);
178bool closure_wait(struct closure_waitlist *list, struct closure *cl);
179void closure_sync(struct closure *cl);
180
181#ifdef CONFIG_BCACHE_CLOSURES_DEBUG
182
183void closure_debug_init(void);
184void closure_debug_create(struct closure *cl);
185void closure_debug_destroy(struct closure *cl);
186
187#else
188
189static inline void closure_debug_init(void) {}
190static inline void closure_debug_create(struct closure *cl) {}
191static inline void closure_debug_destroy(struct closure *cl) {}
192
193#endif
194
195static inline void closure_set_ip(struct closure *cl)
196{
197#ifdef CONFIG_BCACHE_CLOSURES_DEBUG
198	cl->ip = _THIS_IP_;
199#endif
200}
201
202static inline void closure_set_ret_ip(struct closure *cl)
203{
204#ifdef CONFIG_BCACHE_CLOSURES_DEBUG
205	cl->ip = _RET_IP_;
206#endif
207}
208
209static inline void closure_set_waiting(struct closure *cl, unsigned long f)
210{
211#ifdef CONFIG_BCACHE_CLOSURES_DEBUG
212	cl->waiting_on = f;
213#endif
214}
215
216static inline void __closure_end_sleep(struct closure *cl)
217{
218	__set_current_state(TASK_RUNNING);
219
220	if (atomic_read(&cl->remaining) & CLOSURE_SLEEPING)
221		atomic_sub(CLOSURE_SLEEPING, &cl->remaining);
222}
223
224static inline void __closure_start_sleep(struct closure *cl)
225{
226	closure_set_ip(cl);
227	cl->task = current;
228	set_current_state(TASK_UNINTERRUPTIBLE);
229
230	if (!(atomic_read(&cl->remaining) & CLOSURE_SLEEPING))
231		atomic_add(CLOSURE_SLEEPING, &cl->remaining);
232}
233
234static inline void closure_set_stopped(struct closure *cl)
235{
236	atomic_sub(CLOSURE_RUNNING, &cl->remaining);
237}
238
239static inline void set_closure_fn(struct closure *cl, closure_fn *fn,
240				  struct workqueue_struct *wq)
241{
242	BUG_ON(object_is_on_stack(cl));
243	closure_set_ip(cl);
244	cl->fn = fn;
245	cl->wq = wq;
246	/* between atomic_dec() in closure_put() */
247	smp_mb__before_atomic();
248}
249
250static inline void closure_queue(struct closure *cl)
251{
252	struct workqueue_struct *wq = cl->wq;
253	if (wq) {
254		INIT_WORK(&cl->work, cl->work.func);
255		BUG_ON(!queue_work(wq, &cl->work));
256	} else
257		cl->fn(cl);
258}
259
260/**
261 * closure_get - increment a closure's refcount
262 */
263static inline void closure_get(struct closure *cl)
264{
265#ifdef CONFIG_BCACHE_CLOSURES_DEBUG
266	BUG_ON((atomic_inc_return(&cl->remaining) &
267		CLOSURE_REMAINING_MASK) <= 1);
268#else
269	atomic_inc(&cl->remaining);
270#endif
271}
272
273/**
274 * closure_init - Initialize a closure, setting the refcount to 1
275 * @cl:		closure to initialize
276 * @parent:	parent of the new closure. cl will take a refcount on it for its
277 *		lifetime; may be NULL.
278 */
279static inline void closure_init(struct closure *cl, struct closure *parent)
280{
281	memset(cl, 0, sizeof(struct closure));
282	cl->parent = parent;
283	if (parent)
284		closure_get(parent);
285
286	atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER);
287
288	closure_debug_create(cl);
289	closure_set_ip(cl);
290}
291
292static inline void closure_init_stack(struct closure *cl)
293{
294	memset(cl, 0, sizeof(struct closure));
295	atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER|CLOSURE_STACK);
296}
297
298/**
299 * closure_wake_up - wake up all closures on a wait list.
300 */
301static inline void closure_wake_up(struct closure_waitlist *list)
302{
303	smp_mb();
304	__closure_wake_up(list);
305}
306
307/**
308 * continue_at - jump to another function with barrier
309 *
310 * After @cl is no longer waiting on anything (i.e. all outstanding refs have
311 * been dropped with closure_put()), it will resume execution at @fn running out
312 * of @wq (or, if @wq is NULL, @fn will be called by closure_put() directly).
313 *
314 * NOTE: This macro expands to a return in the calling function!
315 *
316 * This is because after calling continue_at() you no longer have a ref on @cl,
317 * and whatever @cl owns may be freed out from under you - a running closure fn
318 * has a ref on its own closure which continue_at() drops.
319 */
320#define continue_at(_cl, _fn, _wq)					\
321do {									\
322	set_closure_fn(_cl, _fn, _wq);					\
323	closure_sub(_cl, CLOSURE_RUNNING + 1);				\
324} while (0)
325
326/**
327 * closure_return - finish execution of a closure
328 *
329 * This is used to indicate that @cl is finished: when all outstanding refs on
330 * @cl have been dropped @cl's ref on its parent closure (as passed to
331 * closure_init()) will be dropped, if one was specified - thus this can be
332 * thought of as returning to the parent closure.
333 */
334#define closure_return(_cl)	continue_at((_cl), NULL, NULL)
335
336/**
337 * continue_at_nobarrier - jump to another function without barrier
338 *
339 * Causes @fn to be executed out of @cl, in @wq context (or called directly if
340 * @wq is NULL).
341 *
342 * NOTE: like continue_at(), this macro expands to a return in the caller!
343 *
344 * The ref the caller of continue_at_nobarrier() had on @cl is now owned by @fn,
345 * thus it's not safe to touch anything protected by @cl after a
346 * continue_at_nobarrier().
347 */
348#define continue_at_nobarrier(_cl, _fn, _wq)				\
349do {									\
350	set_closure_fn(_cl, _fn, _wq);					\
351	closure_queue(_cl);						\
352} while (0)
353
354/**
355 * closure_return - finish execution of a closure, with destructor
356 *
357 * Works like closure_return(), except @destructor will be called when all
358 * outstanding refs on @cl have been dropped; @destructor may be used to safely
359 * free the memory occupied by @cl, and it is called with the ref on the parent
360 * closure still held - so @destructor could safely return an item to a
361 * freelist protected by @cl's parent.
362 */
363#define closure_return_with_destructor(_cl, _destructor)		\
364do {									\
365	set_closure_fn(_cl, _destructor, NULL);				\
366	closure_sub(_cl, CLOSURE_RUNNING - CLOSURE_DESTRUCTOR + 1);	\
367} while (0)
368
369/**
370 * closure_call - execute @fn out of a new, uninitialized closure
371 *
372 * Typically used when running out of one closure, and we want to run @fn
373 * asynchronously out of a new closure - @parent will then wait for @cl to
374 * finish.
375 */
376static inline void closure_call(struct closure *cl, closure_fn fn,
377				struct workqueue_struct *wq,
378				struct closure *parent)
379{
380	closure_init(cl, parent);
381	continue_at_nobarrier(cl, fn, wq);
382}
383
384#endif /* _LINUX_CLOSURE_H */