xref: /linux/fs/ocfs2/uptodate.c (revision 63307d015b91e626c97bb82e88054af3d0b74643)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* -*- mode: c; c-basic-offset: 8; -*-
3  * vim: noexpandtab sw=8 ts=8 sts=0:
4  *
5  * uptodate.c
6  *
7  * Tracking the up-to-date-ness of a local buffer_head with respect to
8  * the cluster.
9  *
10  * Copyright (C) 2002, 2004, 2005 Oracle.  All rights reserved.
11  *
12  * Standard buffer head caching flags (uptodate, etc) are insufficient
13  * in a clustered environment - a buffer may be marked up to date on
14  * our local node but could have been modified by another cluster
15  * member. As a result an additional (and performant) caching scheme
16  * is required. A further requirement is that we consume as little
17  * memory as possible - we never pin buffer_head structures in order
18  * to cache them.
19  *
20  * We track the existence of up to date buffers on the inodes which
21  * are associated with them. Because we don't want to pin
22  * buffer_heads, this is only a (strong) hint and several other checks
23  * are made in the I/O path to ensure that we don't use a stale or
24  * invalid buffer without going to disk:
25  *	- buffer_jbd is used liberally - if a bh is in the journal on
26  *	  this node then it *must* be up to date.
27  *	- the standard buffer_uptodate() macro is used to detect buffers
28  *	  which may be invalid (even if we have an up to date tracking
29  * 	  item for them)
30  *
31  * For a full understanding of how this code works together, one
32  * should read the callers in dlmglue.c, the I/O functions in
33  * buffer_head_io.c and ocfs2_journal_access in journal.c
34  */
35 
36 #include <linux/fs.h>
37 #include <linux/types.h>
38 #include <linux/slab.h>
39 #include <linux/highmem.h>
40 #include <linux/buffer_head.h>
41 #include <linux/rbtree.h>
42 
43 #include <cluster/masklog.h>
44 
45 #include "ocfs2.h"
46 
47 #include "inode.h"
48 #include "uptodate.h"
49 #include "ocfs2_trace.h"
50 
51 struct ocfs2_meta_cache_item {
52 	struct rb_node	c_node;
53 	sector_t	c_block;
54 };
55 
56 static struct kmem_cache *ocfs2_uptodate_cachep;
57 
58 u64 ocfs2_metadata_cache_owner(struct ocfs2_caching_info *ci)
59 {
60 	BUG_ON(!ci || !ci->ci_ops);
61 
62 	return ci->ci_ops->co_owner(ci);
63 }
64 
65 struct super_block *ocfs2_metadata_cache_get_super(struct ocfs2_caching_info *ci)
66 {
67 	BUG_ON(!ci || !ci->ci_ops);
68 
69 	return ci->ci_ops->co_get_super(ci);
70 }
71 
72 static void ocfs2_metadata_cache_lock(struct ocfs2_caching_info *ci)
73 {
74 	BUG_ON(!ci || !ci->ci_ops);
75 
76 	ci->ci_ops->co_cache_lock(ci);
77 }
78 
79 static void ocfs2_metadata_cache_unlock(struct ocfs2_caching_info *ci)
80 {
81 	BUG_ON(!ci || !ci->ci_ops);
82 
83 	ci->ci_ops->co_cache_unlock(ci);
84 }
85 
86 void ocfs2_metadata_cache_io_lock(struct ocfs2_caching_info *ci)
87 {
88 	BUG_ON(!ci || !ci->ci_ops);
89 
90 	ci->ci_ops->co_io_lock(ci);
91 }
92 
93 void ocfs2_metadata_cache_io_unlock(struct ocfs2_caching_info *ci)
94 {
95 	BUG_ON(!ci || !ci->ci_ops);
96 
97 	ci->ci_ops->co_io_unlock(ci);
98 }
99 
100 
101 static void ocfs2_metadata_cache_reset(struct ocfs2_caching_info *ci,
102 				       int clear)
103 {
104 	ci->ci_flags |= OCFS2_CACHE_FL_INLINE;
105 	ci->ci_num_cached = 0;
106 
107 	if (clear) {
108 		ci->ci_created_trans = 0;
109 		ci->ci_last_trans = 0;
110 	}
111 }
112 
113 void ocfs2_metadata_cache_init(struct ocfs2_caching_info *ci,
114 			       const struct ocfs2_caching_operations *ops)
115 {
116 	BUG_ON(!ops);
117 
118 	ci->ci_ops = ops;
119 	ocfs2_metadata_cache_reset(ci, 1);
120 }
121 
122 void ocfs2_metadata_cache_exit(struct ocfs2_caching_info *ci)
123 {
124 	ocfs2_metadata_cache_purge(ci);
125 	ocfs2_metadata_cache_reset(ci, 1);
126 }
127 
128 
129 /* No lock taken here as 'root' is not expected to be visible to other
130  * processes. */
131 static unsigned int ocfs2_purge_copied_metadata_tree(struct rb_root *root)
132 {
133 	unsigned int purged = 0;
134 	struct rb_node *node;
135 	struct ocfs2_meta_cache_item *item;
136 
137 	while ((node = rb_last(root)) != NULL) {
138 		item = rb_entry(node, struct ocfs2_meta_cache_item, c_node);
139 
140 		trace_ocfs2_purge_copied_metadata_tree(
141 					(unsigned long long) item->c_block);
142 
143 		rb_erase(&item->c_node, root);
144 		kmem_cache_free(ocfs2_uptodate_cachep, item);
145 
146 		purged++;
147 	}
148 	return purged;
149 }
150 
151 /* Called from locking and called from ocfs2_clear_inode. Dump the
152  * cache for a given inode.
153  *
154  * This function is a few more lines longer than necessary due to some
155  * accounting done here, but I think it's worth tracking down those
156  * bugs sooner -- Mark */
157 void ocfs2_metadata_cache_purge(struct ocfs2_caching_info *ci)
158 {
159 	unsigned int tree, to_purge, purged;
160 	struct rb_root root = RB_ROOT;
161 
162 	BUG_ON(!ci || !ci->ci_ops);
163 
164 	ocfs2_metadata_cache_lock(ci);
165 	tree = !(ci->ci_flags & OCFS2_CACHE_FL_INLINE);
166 	to_purge = ci->ci_num_cached;
167 
168 	trace_ocfs2_metadata_cache_purge(
169 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
170 		to_purge, tree);
171 
172 	/* If we're a tree, save off the root so that we can safely
173 	 * initialize the cache. We do the work to free tree members
174 	 * without the spinlock. */
175 	if (tree)
176 		root = ci->ci_cache.ci_tree;
177 
178 	ocfs2_metadata_cache_reset(ci, 0);
179 	ocfs2_metadata_cache_unlock(ci);
180 
181 	purged = ocfs2_purge_copied_metadata_tree(&root);
182 	/* If possible, track the number wiped so that we can more
183 	 * easily detect counting errors. Unfortunately, this is only
184 	 * meaningful for trees. */
185 	if (tree && purged != to_purge)
186 		mlog(ML_ERROR, "Owner %llu, count = %u, purged = %u\n",
187 		     (unsigned long long)ocfs2_metadata_cache_owner(ci),
188 		     to_purge, purged);
189 }
190 
191 /* Returns the index in the cache array, -1 if not found.
192  * Requires ip_lock. */
193 static int ocfs2_search_cache_array(struct ocfs2_caching_info *ci,
194 				    sector_t item)
195 {
196 	int i;
197 
198 	for (i = 0; i < ci->ci_num_cached; i++) {
199 		if (item == ci->ci_cache.ci_array[i])
200 			return i;
201 	}
202 
203 	return -1;
204 }
205 
206 /* Returns the cache item if found, otherwise NULL.
207  * Requires ip_lock. */
208 static struct ocfs2_meta_cache_item *
209 ocfs2_search_cache_tree(struct ocfs2_caching_info *ci,
210 			sector_t block)
211 {
212 	struct rb_node * n = ci->ci_cache.ci_tree.rb_node;
213 	struct ocfs2_meta_cache_item *item = NULL;
214 
215 	while (n) {
216 		item = rb_entry(n, struct ocfs2_meta_cache_item, c_node);
217 
218 		if (block < item->c_block)
219 			n = n->rb_left;
220 		else if (block > item->c_block)
221 			n = n->rb_right;
222 		else
223 			return item;
224 	}
225 
226 	return NULL;
227 }
228 
229 static int ocfs2_buffer_cached(struct ocfs2_caching_info *ci,
230 			       struct buffer_head *bh)
231 {
232 	int index = -1;
233 	struct ocfs2_meta_cache_item *item = NULL;
234 
235 	ocfs2_metadata_cache_lock(ci);
236 
237 	trace_ocfs2_buffer_cached_begin(
238 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
239 		(unsigned long long) bh->b_blocknr,
240 		!!(ci->ci_flags & OCFS2_CACHE_FL_INLINE));
241 
242 	if (ci->ci_flags & OCFS2_CACHE_FL_INLINE)
243 		index = ocfs2_search_cache_array(ci, bh->b_blocknr);
244 	else
245 		item = ocfs2_search_cache_tree(ci, bh->b_blocknr);
246 
247 	ocfs2_metadata_cache_unlock(ci);
248 
249 	trace_ocfs2_buffer_cached_end(index, item);
250 
251 	return (index != -1) || (item != NULL);
252 }
253 
254 /* Warning: even if it returns true, this does *not* guarantee that
255  * the block is stored in our inode metadata cache.
256  *
257  * This can be called under lock_buffer()
258  */
259 int ocfs2_buffer_uptodate(struct ocfs2_caching_info *ci,
260 			  struct buffer_head *bh)
261 {
262 	/* Doesn't matter if the bh is in our cache or not -- if it's
263 	 * not marked uptodate then we know it can't have correct
264 	 * data. */
265 	if (!buffer_uptodate(bh))
266 		return 0;
267 
268 	/* OCFS2 does not allow multiple nodes to be changing the same
269 	 * block at the same time. */
270 	if (buffer_jbd(bh))
271 		return 1;
272 
273 	/* Ok, locally the buffer is marked as up to date, now search
274 	 * our cache to see if we can trust that. */
275 	return ocfs2_buffer_cached(ci, bh);
276 }
277 
278 /*
279  * Determine whether a buffer is currently out on a read-ahead request.
280  * ci_io_sem should be held to serialize submitters with the logic here.
281  */
282 int ocfs2_buffer_read_ahead(struct ocfs2_caching_info *ci,
283 			    struct buffer_head *bh)
284 {
285 	return buffer_locked(bh) && ocfs2_buffer_cached(ci, bh);
286 }
287 
288 /* Requires ip_lock */
289 static void ocfs2_append_cache_array(struct ocfs2_caching_info *ci,
290 				     sector_t block)
291 {
292 	BUG_ON(ci->ci_num_cached >= OCFS2_CACHE_INFO_MAX_ARRAY);
293 
294 	trace_ocfs2_append_cache_array(
295 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
296 		(unsigned long long)block, ci->ci_num_cached);
297 
298 	ci->ci_cache.ci_array[ci->ci_num_cached] = block;
299 	ci->ci_num_cached++;
300 }
301 
302 /* By now the caller should have checked that the item does *not*
303  * exist in the tree.
304  * Requires ip_lock. */
305 static void __ocfs2_insert_cache_tree(struct ocfs2_caching_info *ci,
306 				      struct ocfs2_meta_cache_item *new)
307 {
308 	sector_t block = new->c_block;
309 	struct rb_node *parent = NULL;
310 	struct rb_node **p = &ci->ci_cache.ci_tree.rb_node;
311 	struct ocfs2_meta_cache_item *tmp;
312 
313 	trace_ocfs2_insert_cache_tree(
314 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
315 		(unsigned long long)block, ci->ci_num_cached);
316 
317 	while(*p) {
318 		parent = *p;
319 
320 		tmp = rb_entry(parent, struct ocfs2_meta_cache_item, c_node);
321 
322 		if (block < tmp->c_block)
323 			p = &(*p)->rb_left;
324 		else if (block > tmp->c_block)
325 			p = &(*p)->rb_right;
326 		else {
327 			/* This should never happen! */
328 			mlog(ML_ERROR, "Duplicate block %llu cached!\n",
329 			     (unsigned long long) block);
330 			BUG();
331 		}
332 	}
333 
334 	rb_link_node(&new->c_node, parent, p);
335 	rb_insert_color(&new->c_node, &ci->ci_cache.ci_tree);
336 	ci->ci_num_cached++;
337 }
338 
339 /* co_cache_lock() must be held */
340 static inline int ocfs2_insert_can_use_array(struct ocfs2_caching_info *ci)
341 {
342 	return (ci->ci_flags & OCFS2_CACHE_FL_INLINE) &&
343 		(ci->ci_num_cached < OCFS2_CACHE_INFO_MAX_ARRAY);
344 }
345 
346 /* tree should be exactly OCFS2_CACHE_INFO_MAX_ARRAY wide. NULL the
347  * pointers in tree after we use them - this allows caller to detect
348  * when to free in case of error.
349  *
350  * The co_cache_lock() must be held. */
351 static void ocfs2_expand_cache(struct ocfs2_caching_info *ci,
352 			       struct ocfs2_meta_cache_item **tree)
353 {
354 	int i;
355 
356 	mlog_bug_on_msg(ci->ci_num_cached != OCFS2_CACHE_INFO_MAX_ARRAY,
357 			"Owner %llu, num cached = %u, should be %u\n",
358 			(unsigned long long)ocfs2_metadata_cache_owner(ci),
359 			ci->ci_num_cached, OCFS2_CACHE_INFO_MAX_ARRAY);
360 	mlog_bug_on_msg(!(ci->ci_flags & OCFS2_CACHE_FL_INLINE),
361 			"Owner %llu not marked as inline anymore!\n",
362 			(unsigned long long)ocfs2_metadata_cache_owner(ci));
363 
364 	/* Be careful to initialize the tree members *first* because
365 	 * once the ci_tree is used, the array is junk... */
366 	for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++)
367 		tree[i]->c_block = ci->ci_cache.ci_array[i];
368 
369 	ci->ci_flags &= ~OCFS2_CACHE_FL_INLINE;
370 	ci->ci_cache.ci_tree = RB_ROOT;
371 	/* this will be set again by __ocfs2_insert_cache_tree */
372 	ci->ci_num_cached = 0;
373 
374 	for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) {
375 		__ocfs2_insert_cache_tree(ci, tree[i]);
376 		tree[i] = NULL;
377 	}
378 
379 	trace_ocfs2_expand_cache(
380 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
381 		ci->ci_flags, ci->ci_num_cached);
382 }
383 
384 /* Slow path function - memory allocation is necessary. See the
385  * comment above ocfs2_set_buffer_uptodate for more information. */
386 static void __ocfs2_set_buffer_uptodate(struct ocfs2_caching_info *ci,
387 					sector_t block,
388 					int expand_tree)
389 {
390 	int i;
391 	struct ocfs2_meta_cache_item *new = NULL;
392 	struct ocfs2_meta_cache_item *tree[OCFS2_CACHE_INFO_MAX_ARRAY] =
393 		{ NULL, };
394 
395 	trace_ocfs2_set_buffer_uptodate(
396 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
397 		(unsigned long long)block, expand_tree);
398 
399 	new = kmem_cache_alloc(ocfs2_uptodate_cachep, GFP_NOFS);
400 	if (!new) {
401 		mlog_errno(-ENOMEM);
402 		return;
403 	}
404 	new->c_block = block;
405 
406 	if (expand_tree) {
407 		/* Do *not* allocate an array here - the removal code
408 		 * has no way of tracking that. */
409 		for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) {
410 			tree[i] = kmem_cache_alloc(ocfs2_uptodate_cachep,
411 						   GFP_NOFS);
412 			if (!tree[i]) {
413 				mlog_errno(-ENOMEM);
414 				goto out_free;
415 			}
416 
417 			/* These are initialized in ocfs2_expand_cache! */
418 		}
419 	}
420 
421 	ocfs2_metadata_cache_lock(ci);
422 	if (ocfs2_insert_can_use_array(ci)) {
423 		/* Ok, items were removed from the cache in between
424 		 * locks. Detect this and revert back to the fast path */
425 		ocfs2_append_cache_array(ci, block);
426 		ocfs2_metadata_cache_unlock(ci);
427 		goto out_free;
428 	}
429 
430 	if (expand_tree)
431 		ocfs2_expand_cache(ci, tree);
432 
433 	__ocfs2_insert_cache_tree(ci, new);
434 	ocfs2_metadata_cache_unlock(ci);
435 
436 	new = NULL;
437 out_free:
438 	if (new)
439 		kmem_cache_free(ocfs2_uptodate_cachep, new);
440 
441 	/* If these were used, then ocfs2_expand_cache re-set them to
442 	 * NULL for us. */
443 	if (tree[0]) {
444 		for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++)
445 			if (tree[i])
446 				kmem_cache_free(ocfs2_uptodate_cachep,
447 						tree[i]);
448 	}
449 }
450 
451 /* Item insertion is guarded by co_io_lock(), so the insertion path takes
452  * advantage of this by not rechecking for a duplicate insert during
453  * the slow case. Additionally, if the cache needs to be bumped up to
454  * a tree, the code will not recheck after acquiring the lock --
455  * multiple paths cannot be expanding to a tree at the same time.
456  *
457  * The slow path takes into account that items can be removed
458  * (including the whole tree wiped and reset) when this process it out
459  * allocating memory. In those cases, it reverts back to the fast
460  * path.
461  *
462  * Note that this function may actually fail to insert the block if
463  * memory cannot be allocated. This is not fatal however (but may
464  * result in a performance penalty)
465  *
466  * Readahead buffers can be passed in here before the I/O request is
467  * completed.
468  */
469 void ocfs2_set_buffer_uptodate(struct ocfs2_caching_info *ci,
470 			       struct buffer_head *bh)
471 {
472 	int expand;
473 
474 	/* The block may very well exist in our cache already, so avoid
475 	 * doing any more work in that case. */
476 	if (ocfs2_buffer_cached(ci, bh))
477 		return;
478 
479 	trace_ocfs2_set_buffer_uptodate_begin(
480 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
481 		(unsigned long long)bh->b_blocknr);
482 
483 	/* No need to recheck under spinlock - insertion is guarded by
484 	 * co_io_lock() */
485 	ocfs2_metadata_cache_lock(ci);
486 	if (ocfs2_insert_can_use_array(ci)) {
487 		/* Fast case - it's an array and there's a free
488 		 * spot. */
489 		ocfs2_append_cache_array(ci, bh->b_blocknr);
490 		ocfs2_metadata_cache_unlock(ci);
491 		return;
492 	}
493 
494 	expand = 0;
495 	if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) {
496 		/* We need to bump things up to a tree. */
497 		expand = 1;
498 	}
499 	ocfs2_metadata_cache_unlock(ci);
500 
501 	__ocfs2_set_buffer_uptodate(ci, bh->b_blocknr, expand);
502 }
503 
504 /* Called against a newly allocated buffer. Most likely nobody should
505  * be able to read this sort of metadata while it's still being
506  * allocated, but this is careful to take co_io_lock() anyway. */
507 void ocfs2_set_new_buffer_uptodate(struct ocfs2_caching_info *ci,
508 				   struct buffer_head *bh)
509 {
510 	/* This should definitely *not* exist in our cache */
511 	BUG_ON(ocfs2_buffer_cached(ci, bh));
512 
513 	set_buffer_uptodate(bh);
514 
515 	ocfs2_metadata_cache_io_lock(ci);
516 	ocfs2_set_buffer_uptodate(ci, bh);
517 	ocfs2_metadata_cache_io_unlock(ci);
518 }
519 
520 /* Requires ip_lock. */
521 static void ocfs2_remove_metadata_array(struct ocfs2_caching_info *ci,
522 					int index)
523 {
524 	sector_t *array = ci->ci_cache.ci_array;
525 	int bytes;
526 
527 	BUG_ON(index < 0 || index >= OCFS2_CACHE_INFO_MAX_ARRAY);
528 	BUG_ON(index >= ci->ci_num_cached);
529 	BUG_ON(!ci->ci_num_cached);
530 
531 	trace_ocfs2_remove_metadata_array(
532 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
533 		index, ci->ci_num_cached);
534 
535 	ci->ci_num_cached--;
536 
537 	/* don't need to copy if the array is now empty, or if we
538 	 * removed at the tail */
539 	if (ci->ci_num_cached && index < ci->ci_num_cached) {
540 		bytes = sizeof(sector_t) * (ci->ci_num_cached - index);
541 		memmove(&array[index], &array[index + 1], bytes);
542 	}
543 }
544 
545 /* Requires ip_lock. */
546 static void ocfs2_remove_metadata_tree(struct ocfs2_caching_info *ci,
547 				       struct ocfs2_meta_cache_item *item)
548 {
549 	trace_ocfs2_remove_metadata_tree(
550 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
551 		(unsigned long long)item->c_block);
552 
553 	rb_erase(&item->c_node, &ci->ci_cache.ci_tree);
554 	ci->ci_num_cached--;
555 }
556 
557 static void ocfs2_remove_block_from_cache(struct ocfs2_caching_info *ci,
558 					  sector_t block)
559 {
560 	int index;
561 	struct ocfs2_meta_cache_item *item = NULL;
562 
563 	ocfs2_metadata_cache_lock(ci);
564 	trace_ocfs2_remove_block_from_cache(
565 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
566 		(unsigned long long) block, ci->ci_num_cached,
567 		ci->ci_flags);
568 
569 	if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) {
570 		index = ocfs2_search_cache_array(ci, block);
571 		if (index != -1)
572 			ocfs2_remove_metadata_array(ci, index);
573 	} else {
574 		item = ocfs2_search_cache_tree(ci, block);
575 		if (item)
576 			ocfs2_remove_metadata_tree(ci, item);
577 	}
578 	ocfs2_metadata_cache_unlock(ci);
579 
580 	if (item)
581 		kmem_cache_free(ocfs2_uptodate_cachep, item);
582 }
583 
584 /*
585  * Called when we remove a chunk of metadata from an inode. We don't
586  * bother reverting things to an inlined array in the case of a remove
587  * which moves us back under the limit.
588  */
589 void ocfs2_remove_from_cache(struct ocfs2_caching_info *ci,
590 			     struct buffer_head *bh)
591 {
592 	sector_t block = bh->b_blocknr;
593 
594 	ocfs2_remove_block_from_cache(ci, block);
595 }
596 
597 /* Called when we remove xattr clusters from an inode. */
598 void ocfs2_remove_xattr_clusters_from_cache(struct ocfs2_caching_info *ci,
599 					    sector_t block,
600 					    u32 c_len)
601 {
602 	struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
603 	unsigned int i, b_len = ocfs2_clusters_to_blocks(sb, 1) * c_len;
604 
605 	for (i = 0; i < b_len; i++, block++)
606 		ocfs2_remove_block_from_cache(ci, block);
607 }
608 
609 int __init init_ocfs2_uptodate_cache(void)
610 {
611 	ocfs2_uptodate_cachep = kmem_cache_create("ocfs2_uptodate",
612 				  sizeof(struct ocfs2_meta_cache_item),
613 				  0, SLAB_HWCACHE_ALIGN, NULL);
614 	if (!ocfs2_uptodate_cachep)
615 		return -ENOMEM;
616 
617 	return 0;
618 }
619 
620 void exit_ocfs2_uptodate_cache(void)
621 {
622 	kmem_cache_destroy(ocfs2_uptodate_cachep);
623 }
624