xref: /linux/fs/xfs/xfs_mru_cache.c (revision 8e07e0e3964ca4e23ce7b68e2096fe660a888942)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2006-2007 Silicon Graphics, Inc.
4  * All Rights Reserved.
5  */
6 #include "xfs.h"
7 #include "xfs_mru_cache.h"
8 
9 /*
10  * The MRU Cache data structure consists of a data store, an array of lists and
11  * a lock to protect its internal state.  At initialisation time, the client
12  * supplies an element lifetime in milliseconds and a group count, as well as a
13  * function pointer to call when deleting elements.  A data structure for
14  * queueing up work in the form of timed callbacks is also included.
15  *
16  * The group count controls how many lists are created, and thereby how finely
17  * the elements are grouped in time.  When reaping occurs, all the elements in
18  * all the lists whose time has expired are deleted.
19  *
20  * To give an example of how this works in practice, consider a client that
21  * initialises an MRU Cache with a lifetime of ten seconds and a group count of
22  * five.  Five internal lists will be created, each representing a two second
23  * period in time.  When the first element is added, time zero for the data
24  * structure is initialised to the current time.
25  *
26  * All the elements added in the first two seconds are appended to the first
27  * list.  Elements added in the third second go into the second list, and so on.
28  * If an element is accessed at any point, it is removed from its list and
29  * inserted at the head of the current most-recently-used list.
30  *
31  * The reaper function will have nothing to do until at least twelve seconds
32  * have elapsed since the first element was added.  The reason for this is that
33  * if it were called at t=11s, there could be elements in the first list that
34  * have only been inactive for nine seconds, so it still does nothing.  If it is
35  * called anywhere between t=12 and t=14 seconds, it will delete all the
36  * elements that remain in the first list.  It's therefore possible for elements
37  * to remain in the data store even after they've been inactive for up to
38  * (t + t/g) seconds, where t is the inactive element lifetime and g is the
39  * number of groups.
40  *
41  * The above example assumes that the reaper function gets called at least once
42  * every (t/g) seconds.  If it is called less frequently, unused elements will
43  * accumulate in the reap list until the reaper function is eventually called.
44  * The current implementation uses work queue callbacks to carefully time the
45  * reaper function calls, so this should happen rarely, if at all.
46  *
47  * From a design perspective, the primary reason for the choice of a list array
48  * representing discrete time intervals is that it's only practical to reap
49  * expired elements in groups of some appreciable size.  This automatically
50  * introduces a granularity to element lifetimes, so there's no point storing an
51  * individual timeout with each element that specifies a more precise reap time.
52  * The bonus is a saving of sizeof(long) bytes of memory per element stored.
53  *
54  * The elements could have been stored in just one list, but an array of
55  * counters or pointers would need to be maintained to allow them to be divided
56  * up into discrete time groups.  More critically, the process of touching or
57  * removing an element would involve walking large portions of the entire list,
58  * which would have a detrimental effect on performance.  The additional memory
59  * requirement for the array of list heads is minimal.
60  *
61  * When an element is touched or deleted, it needs to be removed from its
62  * current list.  Doubly linked lists are used to make the list maintenance
63  * portion of these operations O(1).  Since reaper timing can be imprecise,
64  * inserts and lookups can occur when there are no free lists available.  When
65  * this happens, all the elements on the LRU list need to be migrated to the end
66  * of the reap list.  To keep the list maintenance portion of these operations
67  * O(1) also, list tails need to be accessible without walking the entire list.
68  * This is the reason why doubly linked list heads are used.
69  */
70 
71 /*
72  * An MRU Cache is a dynamic data structure that stores its elements in a way
73  * that allows efficient lookups, but also groups them into discrete time
74  * intervals based on insertion time.  This allows elements to be efficiently
75  * and automatically reaped after a fixed period of inactivity.
76  *
77  * When a client data pointer is stored in the MRU Cache it needs to be added to
78  * both the data store and to one of the lists.  It must also be possible to
79  * access each of these entries via the other, i.e. to:
80  *
81  *    a) Walk a list, removing the corresponding data store entry for each item.
82  *    b) Look up a data store entry, then access its list entry directly.
83  *
84  * To achieve both of these goals, each entry must contain both a list entry and
85  * a key, in addition to the user's data pointer.  Note that it's not a good
86  * idea to have the client embed one of these structures at the top of their own
87  * data structure, because inserting the same item more than once would most
88  * likely result in a loop in one of the lists.  That's a sure-fire recipe for
89  * an infinite loop in the code.
90  */
91 struct xfs_mru_cache {
92 	struct radix_tree_root	store;     /* Core storage data structure.  */
93 	struct list_head	*lists;    /* Array of lists, one per grp.  */
94 	struct list_head	reap_list; /* Elements overdue for reaping. */
95 	spinlock_t		lock;      /* Lock to protect this struct.  */
96 	unsigned int		grp_count; /* Number of discrete groups.    */
97 	unsigned int		grp_time;  /* Time period spanned by grps.  */
98 	unsigned int		lru_grp;   /* Group containing time zero.   */
99 	unsigned long		time_zero; /* Time first element was added. */
100 	xfs_mru_cache_free_func_t free_func; /* Function pointer for freeing. */
101 	struct delayed_work	work;      /* Workqueue data for reaping.   */
102 	unsigned int		queued;	   /* work has been queued */
103 	void			*data;
104 };
105 
106 static struct workqueue_struct	*xfs_mru_reap_wq;
107 
108 /*
109  * When inserting, destroying or reaping, it's first necessary to update the
110  * lists relative to a particular time.  In the case of destroying, that time
111  * will be well in the future to ensure that all items are moved to the reap
112  * list.  In all other cases though, the time will be the current time.
113  *
114  * This function enters a loop, moving the contents of the LRU list to the reap
115  * list again and again until either a) the lists are all empty, or b) time zero
116  * has been advanced sufficiently to be within the immediate element lifetime.
117  *
118  * Case a) above is detected by counting how many groups are migrated and
119  * stopping when they've all been moved.  Case b) is detected by monitoring the
120  * time_zero field, which is updated as each group is migrated.
121  *
122  * The return value is the earliest time that more migration could be needed, or
123  * zero if there's no need to schedule more work because the lists are empty.
124  */
125 STATIC unsigned long
126 _xfs_mru_cache_migrate(
127 	struct xfs_mru_cache	*mru,
128 	unsigned long		now)
129 {
130 	unsigned int		grp;
131 	unsigned int		migrated = 0;
132 	struct list_head	*lru_list;
133 
134 	/* Nothing to do if the data store is empty. */
135 	if (!mru->time_zero)
136 		return 0;
137 
138 	/* While time zero is older than the time spanned by all the lists. */
139 	while (mru->time_zero <= now - mru->grp_count * mru->grp_time) {
140 
141 		/*
142 		 * If the LRU list isn't empty, migrate its elements to the tail
143 		 * of the reap list.
144 		 */
145 		lru_list = mru->lists + mru->lru_grp;
146 		if (!list_empty(lru_list))
147 			list_splice_init(lru_list, mru->reap_list.prev);
148 
149 		/*
150 		 * Advance the LRU group number, freeing the old LRU list to
151 		 * become the new MRU list; advance time zero accordingly.
152 		 */
153 		mru->lru_grp = (mru->lru_grp + 1) % mru->grp_count;
154 		mru->time_zero += mru->grp_time;
155 
156 		/*
157 		 * If reaping is so far behind that all the elements on all the
158 		 * lists have been migrated to the reap list, it's now empty.
159 		 */
160 		if (++migrated == mru->grp_count) {
161 			mru->lru_grp = 0;
162 			mru->time_zero = 0;
163 			return 0;
164 		}
165 	}
166 
167 	/* Find the first non-empty list from the LRU end. */
168 	for (grp = 0; grp < mru->grp_count; grp++) {
169 
170 		/* Check the grp'th list from the LRU end. */
171 		lru_list = mru->lists + ((mru->lru_grp + grp) % mru->grp_count);
172 		if (!list_empty(lru_list))
173 			return mru->time_zero +
174 			       (mru->grp_count + grp) * mru->grp_time;
175 	}
176 
177 	/* All the lists must be empty. */
178 	mru->lru_grp = 0;
179 	mru->time_zero = 0;
180 	return 0;
181 }
182 
183 /*
184  * When inserting or doing a lookup, an element needs to be inserted into the
185  * MRU list.  The lists must be migrated first to ensure that they're
186  * up-to-date, otherwise the new element could be given a shorter lifetime in
187  * the cache than it should.
188  */
189 STATIC void
190 _xfs_mru_cache_list_insert(
191 	struct xfs_mru_cache	*mru,
192 	struct xfs_mru_cache_elem *elem)
193 {
194 	unsigned int		grp = 0;
195 	unsigned long		now = jiffies;
196 
197 	/*
198 	 * If the data store is empty, initialise time zero, leave grp set to
199 	 * zero and start the work queue timer if necessary.  Otherwise, set grp
200 	 * to the number of group times that have elapsed since time zero.
201 	 */
202 	if (!_xfs_mru_cache_migrate(mru, now)) {
203 		mru->time_zero = now;
204 		if (!mru->queued) {
205 			mru->queued = 1;
206 			queue_delayed_work(xfs_mru_reap_wq, &mru->work,
207 			                   mru->grp_count * mru->grp_time);
208 		}
209 	} else {
210 		grp = (now - mru->time_zero) / mru->grp_time;
211 		grp = (mru->lru_grp + grp) % mru->grp_count;
212 	}
213 
214 	/* Insert the element at the tail of the corresponding list. */
215 	list_add_tail(&elem->list_node, mru->lists + grp);
216 }
217 
218 /*
219  * When destroying or reaping, all the elements that were migrated to the reap
220  * list need to be deleted.  For each element this involves removing it from the
221  * data store, removing it from the reap list, calling the client's free
222  * function and deleting the element from the element cache.
223  *
224  * We get called holding the mru->lock, which we drop and then reacquire.
225  * Sparse need special help with this to tell it we know what we are doing.
226  */
227 STATIC void
228 _xfs_mru_cache_clear_reap_list(
229 	struct xfs_mru_cache	*mru)
230 		__releases(mru->lock) __acquires(mru->lock)
231 {
232 	struct xfs_mru_cache_elem *elem, *next;
233 	struct list_head	tmp;
234 
235 	INIT_LIST_HEAD(&tmp);
236 	list_for_each_entry_safe(elem, next, &mru->reap_list, list_node) {
237 
238 		/* Remove the element from the data store. */
239 		radix_tree_delete(&mru->store, elem->key);
240 
241 		/*
242 		 * remove to temp list so it can be freed without
243 		 * needing to hold the lock
244 		 */
245 		list_move(&elem->list_node, &tmp);
246 	}
247 	spin_unlock(&mru->lock);
248 
249 	list_for_each_entry_safe(elem, next, &tmp, list_node) {
250 		list_del_init(&elem->list_node);
251 		mru->free_func(mru->data, elem);
252 	}
253 
254 	spin_lock(&mru->lock);
255 }
256 
257 /*
258  * We fire the reap timer every group expiry interval so
259  * we always have a reaper ready to run. This makes shutdown
260  * and flushing of the reaper easy to do. Hence we need to
261  * keep when the next reap must occur so we can determine
262  * at each interval whether there is anything we need to do.
263  */
264 STATIC void
265 _xfs_mru_cache_reap(
266 	struct work_struct	*work)
267 {
268 	struct xfs_mru_cache	*mru =
269 		container_of(work, struct xfs_mru_cache, work.work);
270 	unsigned long		now, next;
271 
272 	ASSERT(mru && mru->lists);
273 	if (!mru || !mru->lists)
274 		return;
275 
276 	spin_lock(&mru->lock);
277 	next = _xfs_mru_cache_migrate(mru, jiffies);
278 	_xfs_mru_cache_clear_reap_list(mru);
279 
280 	mru->queued = next;
281 	if ((mru->queued > 0)) {
282 		now = jiffies;
283 		if (next <= now)
284 			next = 0;
285 		else
286 			next -= now;
287 		queue_delayed_work(xfs_mru_reap_wq, &mru->work, next);
288 	}
289 
290 	spin_unlock(&mru->lock);
291 }
292 
293 int
294 xfs_mru_cache_init(void)
295 {
296 	xfs_mru_reap_wq = alloc_workqueue("xfs_mru_cache",
297 			XFS_WQFLAGS(WQ_MEM_RECLAIM | WQ_FREEZABLE), 1);
298 	if (!xfs_mru_reap_wq)
299 		return -ENOMEM;
300 	return 0;
301 }
302 
303 void
304 xfs_mru_cache_uninit(void)
305 {
306 	destroy_workqueue(xfs_mru_reap_wq);
307 }
308 
309 /*
310  * To initialise a struct xfs_mru_cache pointer, call xfs_mru_cache_create()
311  * with the address of the pointer, a lifetime value in milliseconds, a group
312  * count and a free function to use when deleting elements.  This function
313  * returns 0 if the initialisation was successful.
314  */
315 int
316 xfs_mru_cache_create(
317 	struct xfs_mru_cache	**mrup,
318 	void			*data,
319 	unsigned int		lifetime_ms,
320 	unsigned int		grp_count,
321 	xfs_mru_cache_free_func_t free_func)
322 {
323 	struct xfs_mru_cache	*mru = NULL;
324 	int			err = 0, grp;
325 	unsigned int		grp_time;
326 
327 	if (mrup)
328 		*mrup = NULL;
329 
330 	if (!mrup || !grp_count || !lifetime_ms || !free_func)
331 		return -EINVAL;
332 
333 	if (!(grp_time = msecs_to_jiffies(lifetime_ms) / grp_count))
334 		return -EINVAL;
335 
336 	if (!(mru = kmem_zalloc(sizeof(*mru), 0)))
337 		return -ENOMEM;
338 
339 	/* An extra list is needed to avoid reaping up to a grp_time early. */
340 	mru->grp_count = grp_count + 1;
341 	mru->lists = kmem_zalloc(mru->grp_count * sizeof(*mru->lists), 0);
342 
343 	if (!mru->lists) {
344 		err = -ENOMEM;
345 		goto exit;
346 	}
347 
348 	for (grp = 0; grp < mru->grp_count; grp++)
349 		INIT_LIST_HEAD(mru->lists + grp);
350 
351 	/*
352 	 * We use GFP_KERNEL radix tree preload and do inserts under a
353 	 * spinlock so GFP_ATOMIC is appropriate for the radix tree itself.
354 	 */
355 	INIT_RADIX_TREE(&mru->store, GFP_ATOMIC);
356 	INIT_LIST_HEAD(&mru->reap_list);
357 	spin_lock_init(&mru->lock);
358 	INIT_DELAYED_WORK(&mru->work, _xfs_mru_cache_reap);
359 
360 	mru->grp_time  = grp_time;
361 	mru->free_func = free_func;
362 	mru->data = data;
363 	*mrup = mru;
364 
365 exit:
366 	if (err && mru && mru->lists)
367 		kmem_free(mru->lists);
368 	if (err && mru)
369 		kmem_free(mru);
370 
371 	return err;
372 }
373 
374 /*
375  * Call xfs_mru_cache_flush() to flush out all cached entries, calling their
376  * free functions as they're deleted.  When this function returns, the caller is
377  * guaranteed that all the free functions for all the elements have finished
378  * executing and the reaper is not running.
379  */
380 static void
381 xfs_mru_cache_flush(
382 	struct xfs_mru_cache	*mru)
383 {
384 	if (!mru || !mru->lists)
385 		return;
386 
387 	spin_lock(&mru->lock);
388 	if (mru->queued) {
389 		spin_unlock(&mru->lock);
390 		cancel_delayed_work_sync(&mru->work);
391 		spin_lock(&mru->lock);
392 	}
393 
394 	_xfs_mru_cache_migrate(mru, jiffies + mru->grp_count * mru->grp_time);
395 	_xfs_mru_cache_clear_reap_list(mru);
396 
397 	spin_unlock(&mru->lock);
398 }
399 
400 void
401 xfs_mru_cache_destroy(
402 	struct xfs_mru_cache	*mru)
403 {
404 	if (!mru || !mru->lists)
405 		return;
406 
407 	xfs_mru_cache_flush(mru);
408 
409 	kmem_free(mru->lists);
410 	kmem_free(mru);
411 }
412 
413 /*
414  * To insert an element, call xfs_mru_cache_insert() with the data store, the
415  * element's key and the client data pointer.  This function returns 0 on
416  * success or ENOMEM if memory for the data element couldn't be allocated.
417  */
418 int
419 xfs_mru_cache_insert(
420 	struct xfs_mru_cache	*mru,
421 	unsigned long		key,
422 	struct xfs_mru_cache_elem *elem)
423 {
424 	int			error;
425 
426 	ASSERT(mru && mru->lists);
427 	if (!mru || !mru->lists)
428 		return -EINVAL;
429 
430 	if (radix_tree_preload(GFP_NOFS))
431 		return -ENOMEM;
432 
433 	INIT_LIST_HEAD(&elem->list_node);
434 	elem->key = key;
435 
436 	spin_lock(&mru->lock);
437 	error = radix_tree_insert(&mru->store, key, elem);
438 	radix_tree_preload_end();
439 	if (!error)
440 		_xfs_mru_cache_list_insert(mru, elem);
441 	spin_unlock(&mru->lock);
442 
443 	return error;
444 }
445 
446 /*
447  * To remove an element without calling the free function, call
448  * xfs_mru_cache_remove() with the data store and the element's key.  On success
449  * the client data pointer for the removed element is returned, otherwise this
450  * function will return a NULL pointer.
451  */
452 struct xfs_mru_cache_elem *
453 xfs_mru_cache_remove(
454 	struct xfs_mru_cache	*mru,
455 	unsigned long		key)
456 {
457 	struct xfs_mru_cache_elem *elem;
458 
459 	ASSERT(mru && mru->lists);
460 	if (!mru || !mru->lists)
461 		return NULL;
462 
463 	spin_lock(&mru->lock);
464 	elem = radix_tree_delete(&mru->store, key);
465 	if (elem)
466 		list_del(&elem->list_node);
467 	spin_unlock(&mru->lock);
468 
469 	return elem;
470 }
471 
472 /*
473  * To remove and element and call the free function, call xfs_mru_cache_delete()
474  * with the data store and the element's key.
475  */
476 void
477 xfs_mru_cache_delete(
478 	struct xfs_mru_cache	*mru,
479 	unsigned long		key)
480 {
481 	struct xfs_mru_cache_elem *elem;
482 
483 	elem = xfs_mru_cache_remove(mru, key);
484 	if (elem)
485 		mru->free_func(mru->data, elem);
486 }
487 
488 /*
489  * To look up an element using its key, call xfs_mru_cache_lookup() with the
490  * data store and the element's key.  If found, the element will be moved to the
491  * head of the MRU list to indicate that it's been touched.
492  *
493  * The internal data structures are protected by a spinlock that is STILL HELD
494  * when this function returns.  Call xfs_mru_cache_done() to release it.  Note
495  * that it is not safe to call any function that might sleep in the interim.
496  *
497  * The implementation could have used reference counting to avoid this
498  * restriction, but since most clients simply want to get, set or test a member
499  * of the returned data structure, the extra per-element memory isn't warranted.
500  *
501  * If the element isn't found, this function returns NULL and the spinlock is
502  * released.  xfs_mru_cache_done() should NOT be called when this occurs.
503  *
504  * Because sparse isn't smart enough to know about conditional lock return
505  * status, we need to help it get it right by annotating the path that does
506  * not release the lock.
507  */
508 struct xfs_mru_cache_elem *
509 xfs_mru_cache_lookup(
510 	struct xfs_mru_cache	*mru,
511 	unsigned long		key)
512 {
513 	struct xfs_mru_cache_elem *elem;
514 
515 	ASSERT(mru && mru->lists);
516 	if (!mru || !mru->lists)
517 		return NULL;
518 
519 	spin_lock(&mru->lock);
520 	elem = radix_tree_lookup(&mru->store, key);
521 	if (elem) {
522 		list_del(&elem->list_node);
523 		_xfs_mru_cache_list_insert(mru, elem);
524 		__release(mru_lock); /* help sparse not be stupid */
525 	} else
526 		spin_unlock(&mru->lock);
527 
528 	return elem;
529 }
530 
531 /*
532  * To release the internal data structure spinlock after having performed an
533  * xfs_mru_cache_lookup() or an xfs_mru_cache_peek(), call xfs_mru_cache_done()
534  * with the data store pointer.
535  */
536 void
537 xfs_mru_cache_done(
538 	struct xfs_mru_cache	*mru)
539 		__releases(mru->lock)
540 {
541 	spin_unlock(&mru->lock);
542 }
543