Commit | Line | Data |
---|---|---|
ccd979bd MF |
1 | /* -*- mode: c; c-basic-offset: 8; -*- |
2 | * vim: noexpandtab sw=8 ts=8 sts=0: | |
3 | * | |
4 | * uptodate.c | |
5 | * | |
6 | * Tracking the up-to-date-ness of a local buffer_head with respect to | |
7 | * the cluster. | |
8 | * | |
9 | * Copyright (C) 2002, 2004, 2005 Oracle. All rights reserved. | |
10 | * | |
11 | * This program is free software; you can redistribute it and/or | |
12 | * modify it under the terms of the GNU General Public | |
13 | * License as published by the Free Software Foundation; either | |
14 | * version 2 of the License, or (at your option) any later version. | |
15 | * | |
16 | * This program is distributed in the hope that it will be useful, | |
17 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
18 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
19 | * General Public License for more details. | |
20 | * | |
21 | * You should have received a copy of the GNU General Public | |
22 | * License along with this program; if not, write to the | |
23 | * Free Software Foundation, Inc., 59 Temple Place - Suite 330, | |
24 | * Boston, MA 021110-1307, USA. | |
25 | * | |
26 | * Standard buffer head caching flags (uptodate, etc) are insufficient | |
27 | * in a clustered environment - a buffer may be marked up to date on | |
28 | * our local node but could have been modified by another cluster | |
29 | * member. As a result an additional (and performant) caching scheme | |
30 | * is required. A further requirement is that we consume as little | |
31 | * memory as possible - we never pin buffer_head structures in order | |
32 | * to cache them. | |
33 | * | |
34 | * We track the existence of up to date buffers on the inodes which | |
35 | * are associated with them. Because we don't want to pin | |
36 | * buffer_heads, this is only a (strong) hint and several other checks | |
37 | * are made in the I/O path to ensure that we don't use a stale or | |
38 | * invalid buffer without going to disk: | |
39 | * - buffer_jbd is used liberally - if a bh is in the journal on | |
40 | * this node then it *must* be up to date. | |
41 | * - the standard buffer_uptodate() macro is used to detect buffers | |
42 | * which may be invalid (even if we have an up to date tracking | |
43 | * item for them) | |
44 | * | |
45 | * For a full understanding of how this code works together, one | |
46 | * should read the callers in dlmglue.c, the I/O functions in | |
47 | * buffer_head_io.c and ocfs2_journal_access in journal.c | |
48 | */ | |
49 | ||
50 | #include <linux/fs.h> | |
51 | #include <linux/types.h> | |
52 | #include <linux/slab.h> | |
53 | #include <linux/highmem.h> | |
54 | #include <linux/buffer_head.h> | |
55 | #include <linux/rbtree.h> | |
2b4e30fb JB |
56 | #ifndef CONFIG_OCFS2_COMPAT_JBD |
57 | # include <linux/jbd2.h> | |
58 | #else | |
59 | # include <linux/jbd.h> | |
60 | #endif | |
ccd979bd MF |
61 | |
62 | #define MLOG_MASK_PREFIX ML_UPTODATE | |
63 | ||
64 | #include <cluster/masklog.h> | |
65 | ||
66 | #include "ocfs2.h" | |
67 | ||
68 | #include "inode.h" | |
69 | #include "uptodate.h" | |
70 | ||
71 | struct ocfs2_meta_cache_item { | |
72 | struct rb_node c_node; | |
73 | sector_t c_block; | |
74 | }; | |
75 | ||
e18b890b | 76 | static struct kmem_cache *ocfs2_uptodate_cachep = NULL; |
ccd979bd | 77 | |
8cb471e8 | 78 | u64 ocfs2_metadata_cache_owner(struct ocfs2_caching_info *ci) |
6e5a3d75 JB |
79 | { |
80 | BUG_ON(!ci || !ci->ci_ops); | |
81 | ||
82 | return ci->ci_ops->co_owner(ci); | |
83 | } | |
84 | ||
8cb471e8 JB |
85 | struct super_block *ocfs2_metadata_cache_get_super(struct ocfs2_caching_info *ci) |
86 | { | |
87 | BUG_ON(!ci || !ci->ci_ops); | |
88 | ||
89 | return ci->ci_ops->co_get_super(ci); | |
90 | } | |
91 | ||
6e5a3d75 JB |
92 | static void ocfs2_metadata_cache_lock(struct ocfs2_caching_info *ci) |
93 | { | |
94 | BUG_ON(!ci || !ci->ci_ops); | |
95 | ||
96 | ci->ci_ops->co_cache_lock(ci); | |
97 | } | |
98 | ||
99 | static void ocfs2_metadata_cache_unlock(struct ocfs2_caching_info *ci) | |
100 | { | |
101 | BUG_ON(!ci || !ci->ci_ops); | |
102 | ||
103 | ci->ci_ops->co_cache_unlock(ci); | |
104 | } | |
105 | ||
8cb471e8 | 106 | void ocfs2_metadata_cache_io_lock(struct ocfs2_caching_info *ci) |
6e5a3d75 JB |
107 | { |
108 | BUG_ON(!ci || !ci->ci_ops); | |
109 | ||
110 | ci->ci_ops->co_io_lock(ci); | |
111 | } | |
112 | ||
8cb471e8 | 113 | void ocfs2_metadata_cache_io_unlock(struct ocfs2_caching_info *ci) |
6e5a3d75 JB |
114 | { |
115 | BUG_ON(!ci || !ci->ci_ops); | |
116 | ||
117 | ci->ci_ops->co_io_unlock(ci); | |
118 | } | |
119 | ||
120 | ||
66fb345d JB |
121 | static void ocfs2_metadata_cache_reset(struct ocfs2_caching_info *ci, |
122 | int clear) | |
123 | { | |
124 | ci->ci_flags |= OCFS2_CACHE_FL_INLINE; | |
125 | ci->ci_num_cached = 0; | |
126 | ||
292dd27e JB |
127 | if (clear) { |
128 | ci->ci_created_trans = 0; | |
66fb345d | 129 | ci->ci_last_trans = 0; |
292dd27e | 130 | } |
66fb345d JB |
131 | } |
132 | ||
47460d65 | 133 | void ocfs2_metadata_cache_init(struct ocfs2_caching_info *ci, |
6e5a3d75 | 134 | const struct ocfs2_caching_operations *ops) |
ccd979bd | 135 | { |
6e5a3d75 JB |
136 | BUG_ON(!ops); |
137 | ||
138 | ci->ci_ops = ops; | |
66fb345d JB |
139 | ocfs2_metadata_cache_reset(ci, 1); |
140 | } | |
141 | ||
142 | void ocfs2_metadata_cache_exit(struct ocfs2_caching_info *ci) | |
143 | { | |
144 | ocfs2_metadata_cache_purge(ci); | |
145 | ocfs2_metadata_cache_reset(ci, 1); | |
ccd979bd MF |
146 | } |
147 | ||
66fb345d | 148 | |
ccd979bd MF |
149 | /* No lock taken here as 'root' is not expected to be visible to other |
150 | * processes. */ | |
151 | static unsigned int ocfs2_purge_copied_metadata_tree(struct rb_root *root) | |
152 | { | |
153 | unsigned int purged = 0; | |
154 | struct rb_node *node; | |
155 | struct ocfs2_meta_cache_item *item; | |
156 | ||
157 | while ((node = rb_last(root)) != NULL) { | |
158 | item = rb_entry(node, struct ocfs2_meta_cache_item, c_node); | |
159 | ||
160 | mlog(0, "Purge item %llu\n", | |
161 | (unsigned long long) item->c_block); | |
162 | ||
163 | rb_erase(&item->c_node, root); | |
164 | kmem_cache_free(ocfs2_uptodate_cachep, item); | |
165 | ||
166 | purged++; | |
167 | } | |
168 | return purged; | |
169 | } | |
170 | ||
171 | /* Called from locking and called from ocfs2_clear_inode. Dump the | |
172 | * cache for a given inode. | |
173 | * | |
174 | * This function is a few more lines longer than necessary due to some | |
175 | * accounting done here, but I think it's worth tracking down those | |
176 | * bugs sooner -- Mark */ | |
8cb471e8 | 177 | void ocfs2_metadata_cache_purge(struct ocfs2_caching_info *ci) |
ccd979bd | 178 | { |
ccd979bd | 179 | unsigned int tree, to_purge, purged; |
ccd979bd MF |
180 | struct rb_root root = RB_ROOT; |
181 | ||
6e5a3d75 JB |
182 | BUG_ON(!ci || !ci->ci_ops); |
183 | ||
184 | ocfs2_metadata_cache_lock(ci); | |
47460d65 | 185 | tree = !(ci->ci_flags & OCFS2_CACHE_FL_INLINE); |
ccd979bd MF |
186 | to_purge = ci->ci_num_cached; |
187 | ||
6e5a3d75 JB |
188 | mlog(0, "Purge %u %s items from Owner %llu\n", to_purge, |
189 | tree ? "array" : "tree", | |
190 | (unsigned long long)ocfs2_metadata_cache_owner(ci)); | |
ccd979bd MF |
191 | |
192 | /* If we're a tree, save off the root so that we can safely | |
193 | * initialize the cache. We do the work to free tree members | |
194 | * without the spinlock. */ | |
195 | if (tree) | |
196 | root = ci->ci_cache.ci_tree; | |
197 | ||
66fb345d | 198 | ocfs2_metadata_cache_reset(ci, 0); |
6e5a3d75 | 199 | ocfs2_metadata_cache_unlock(ci); |
ccd979bd MF |
200 | |
201 | purged = ocfs2_purge_copied_metadata_tree(&root); | |
202 | /* If possible, track the number wiped so that we can more | |
203 | * easily detect counting errors. Unfortunately, this is only | |
204 | * meaningful for trees. */ | |
205 | if (tree && purged != to_purge) | |
6e5a3d75 JB |
206 | mlog(ML_ERROR, "Owner %llu, count = %u, purged = %u\n", |
207 | (unsigned long long)ocfs2_metadata_cache_owner(ci), | |
208 | to_purge, purged); | |
ccd979bd MF |
209 | } |
210 | ||
211 | /* Returns the index in the cache array, -1 if not found. | |
212 | * Requires ip_lock. */ | |
213 | static int ocfs2_search_cache_array(struct ocfs2_caching_info *ci, | |
214 | sector_t item) | |
215 | { | |
216 | int i; | |
217 | ||
218 | for (i = 0; i < ci->ci_num_cached; i++) { | |
219 | if (item == ci->ci_cache.ci_array[i]) | |
220 | return i; | |
221 | } | |
222 | ||
223 | return -1; | |
224 | } | |
225 | ||
226 | /* Returns the cache item if found, otherwise NULL. | |
227 | * Requires ip_lock. */ | |
228 | static struct ocfs2_meta_cache_item * | |
229 | ocfs2_search_cache_tree(struct ocfs2_caching_info *ci, | |
230 | sector_t block) | |
231 | { | |
232 | struct rb_node * n = ci->ci_cache.ci_tree.rb_node; | |
233 | struct ocfs2_meta_cache_item *item = NULL; | |
234 | ||
235 | while (n) { | |
236 | item = rb_entry(n, struct ocfs2_meta_cache_item, c_node); | |
237 | ||
238 | if (block < item->c_block) | |
239 | n = n->rb_left; | |
240 | else if (block > item->c_block) | |
241 | n = n->rb_right; | |
242 | else | |
243 | return item; | |
244 | } | |
245 | ||
246 | return NULL; | |
247 | } | |
248 | ||
8cb471e8 | 249 | static int ocfs2_buffer_cached(struct ocfs2_caching_info *ci, |
ccd979bd MF |
250 | struct buffer_head *bh) |
251 | { | |
252 | int index = -1; | |
253 | struct ocfs2_meta_cache_item *item = NULL; | |
254 | ||
6e5a3d75 | 255 | ocfs2_metadata_cache_lock(ci); |
ccd979bd | 256 | |
6e5a3d75 JB |
257 | mlog(0, "Owner %llu, query block %llu (inline = %u)\n", |
258 | (unsigned long long)ocfs2_metadata_cache_owner(ci), | |
b0697053 | 259 | (unsigned long long) bh->b_blocknr, |
47460d65 | 260 | !!(ci->ci_flags & OCFS2_CACHE_FL_INLINE)); |
ccd979bd | 261 | |
47460d65 | 262 | if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) |
8cb471e8 | 263 | index = ocfs2_search_cache_array(ci, bh->b_blocknr); |
ccd979bd | 264 | else |
8cb471e8 | 265 | item = ocfs2_search_cache_tree(ci, bh->b_blocknr); |
ccd979bd | 266 | |
6e5a3d75 | 267 | ocfs2_metadata_cache_unlock(ci); |
ccd979bd MF |
268 | |
269 | mlog(0, "index = %d, item = %p\n", index, item); | |
270 | ||
271 | return (index != -1) || (item != NULL); | |
272 | } | |
273 | ||
274 | /* Warning: even if it returns true, this does *not* guarantee that | |
aa958874 MF |
275 | * the block is stored in our inode metadata cache. |
276 | * | |
277 | * This can be called under lock_buffer() | |
278 | */ | |
8cb471e8 | 279 | int ocfs2_buffer_uptodate(struct ocfs2_caching_info *ci, |
ccd979bd MF |
280 | struct buffer_head *bh) |
281 | { | |
282 | /* Doesn't matter if the bh is in our cache or not -- if it's | |
283 | * not marked uptodate then we know it can't have correct | |
284 | * data. */ | |
285 | if (!buffer_uptodate(bh)) | |
286 | return 0; | |
287 | ||
288 | /* OCFS2 does not allow multiple nodes to be changing the same | |
289 | * block at the same time. */ | |
290 | if (buffer_jbd(bh)) | |
291 | return 1; | |
292 | ||
293 | /* Ok, locally the buffer is marked as up to date, now search | |
294 | * our cache to see if we can trust that. */ | |
8cb471e8 | 295 | return ocfs2_buffer_cached(ci, bh); |
ccd979bd MF |
296 | } |
297 | ||
8cb471e8 | 298 | /* |
aa958874 | 299 | * Determine whether a buffer is currently out on a read-ahead request. |
47460d65 | 300 | * ci_io_sem should be held to serialize submitters with the logic here. |
aa958874 | 301 | */ |
8cb471e8 | 302 | int ocfs2_buffer_read_ahead(struct ocfs2_caching_info *ci, |
aa958874 MF |
303 | struct buffer_head *bh) |
304 | { | |
8cb471e8 | 305 | return buffer_locked(bh) && ocfs2_buffer_cached(ci, bh); |
aa958874 MF |
306 | } |
307 | ||
ccd979bd MF |
308 | /* Requires ip_lock */ |
309 | static void ocfs2_append_cache_array(struct ocfs2_caching_info *ci, | |
310 | sector_t block) | |
311 | { | |
47460d65 | 312 | BUG_ON(ci->ci_num_cached >= OCFS2_CACHE_INFO_MAX_ARRAY); |
ccd979bd MF |
313 | |
314 | mlog(0, "block %llu takes position %u\n", (unsigned long long) block, | |
315 | ci->ci_num_cached); | |
316 | ||
317 | ci->ci_cache.ci_array[ci->ci_num_cached] = block; | |
318 | ci->ci_num_cached++; | |
319 | } | |
320 | ||
321 | /* By now the caller should have checked that the item does *not* | |
322 | * exist in the tree. | |
323 | * Requires ip_lock. */ | |
324 | static void __ocfs2_insert_cache_tree(struct ocfs2_caching_info *ci, | |
325 | struct ocfs2_meta_cache_item *new) | |
326 | { | |
327 | sector_t block = new->c_block; | |
328 | struct rb_node *parent = NULL; | |
329 | struct rb_node **p = &ci->ci_cache.ci_tree.rb_node; | |
330 | struct ocfs2_meta_cache_item *tmp; | |
331 | ||
332 | mlog(0, "Insert block %llu num = %u\n", (unsigned long long) block, | |
333 | ci->ci_num_cached); | |
334 | ||
335 | while(*p) { | |
336 | parent = *p; | |
337 | ||
338 | tmp = rb_entry(parent, struct ocfs2_meta_cache_item, c_node); | |
339 | ||
340 | if (block < tmp->c_block) | |
341 | p = &(*p)->rb_left; | |
342 | else if (block > tmp->c_block) | |
343 | p = &(*p)->rb_right; | |
344 | else { | |
345 | /* This should never happen! */ | |
346 | mlog(ML_ERROR, "Duplicate block %llu cached!\n", | |
347 | (unsigned long long) block); | |
348 | BUG(); | |
349 | } | |
350 | } | |
351 | ||
352 | rb_link_node(&new->c_node, parent, p); | |
353 | rb_insert_color(&new->c_node, &ci->ci_cache.ci_tree); | |
354 | ci->ci_num_cached++; | |
355 | } | |
356 | ||
6e5a3d75 | 357 | /* co_cache_lock() must be held */ |
8cb471e8 | 358 | static inline int ocfs2_insert_can_use_array(struct ocfs2_caching_info *ci) |
ccd979bd | 359 | { |
47460d65 JB |
360 | return (ci->ci_flags & OCFS2_CACHE_FL_INLINE) && |
361 | (ci->ci_num_cached < OCFS2_CACHE_INFO_MAX_ARRAY); | |
ccd979bd MF |
362 | } |
363 | ||
47460d65 | 364 | /* tree should be exactly OCFS2_CACHE_INFO_MAX_ARRAY wide. NULL the |
ccd979bd | 365 | * pointers in tree after we use them - this allows caller to detect |
6e5a3d75 JB |
366 | * when to free in case of error. |
367 | * | |
368 | * The co_cache_lock() must be held. */ | |
8cb471e8 | 369 | static void ocfs2_expand_cache(struct ocfs2_caching_info *ci, |
ccd979bd MF |
370 | struct ocfs2_meta_cache_item **tree) |
371 | { | |
372 | int i; | |
ccd979bd | 373 | |
47460d65 | 374 | mlog_bug_on_msg(ci->ci_num_cached != OCFS2_CACHE_INFO_MAX_ARRAY, |
6e5a3d75 JB |
375 | "Owner %llu, num cached = %u, should be %u\n", |
376 | (unsigned long long)ocfs2_metadata_cache_owner(ci), | |
377 | ci->ci_num_cached, OCFS2_CACHE_INFO_MAX_ARRAY); | |
47460d65 | 378 | mlog_bug_on_msg(!(ci->ci_flags & OCFS2_CACHE_FL_INLINE), |
6e5a3d75 JB |
379 | "Owner %llu not marked as inline anymore!\n", |
380 | (unsigned long long)ocfs2_metadata_cache_owner(ci)); | |
ccd979bd MF |
381 | |
382 | /* Be careful to initialize the tree members *first* because | |
383 | * once the ci_tree is used, the array is junk... */ | |
47460d65 | 384 | for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) |
ccd979bd MF |
385 | tree[i]->c_block = ci->ci_cache.ci_array[i]; |
386 | ||
47460d65 | 387 | ci->ci_flags &= ~OCFS2_CACHE_FL_INLINE; |
ccd979bd MF |
388 | ci->ci_cache.ci_tree = RB_ROOT; |
389 | /* this will be set again by __ocfs2_insert_cache_tree */ | |
390 | ci->ci_num_cached = 0; | |
391 | ||
47460d65 | 392 | for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) { |
ccd979bd MF |
393 | __ocfs2_insert_cache_tree(ci, tree[i]); |
394 | tree[i] = NULL; | |
395 | } | |
396 | ||
b0697053 | 397 | mlog(0, "Expanded %llu to a tree cache: flags 0x%x, num = %u\n", |
6e5a3d75 JB |
398 | (unsigned long long)ocfs2_metadata_cache_owner(ci), |
399 | ci->ci_flags, ci->ci_num_cached); | |
ccd979bd MF |
400 | } |
401 | ||
402 | /* Slow path function - memory allocation is necessary. See the | |
403 | * comment above ocfs2_set_buffer_uptodate for more information. */ | |
8cb471e8 | 404 | static void __ocfs2_set_buffer_uptodate(struct ocfs2_caching_info *ci, |
ccd979bd MF |
405 | sector_t block, |
406 | int expand_tree) | |
407 | { | |
408 | int i; | |
ccd979bd | 409 | struct ocfs2_meta_cache_item *new = NULL; |
47460d65 | 410 | struct ocfs2_meta_cache_item *tree[OCFS2_CACHE_INFO_MAX_ARRAY] = |
ccd979bd MF |
411 | { NULL, }; |
412 | ||
6e5a3d75 JB |
413 | mlog(0, "Owner %llu, block %llu, expand = %d\n", |
414 | (unsigned long long)ocfs2_metadata_cache_owner(ci), | |
b0697053 | 415 | (unsigned long long)block, expand_tree); |
ccd979bd | 416 | |
afae00ab | 417 | new = kmem_cache_alloc(ocfs2_uptodate_cachep, GFP_NOFS); |
ccd979bd MF |
418 | if (!new) { |
419 | mlog_errno(-ENOMEM); | |
420 | return; | |
421 | } | |
422 | new->c_block = block; | |
423 | ||
424 | if (expand_tree) { | |
425 | /* Do *not* allocate an array here - the removal code | |
426 | * has no way of tracking that. */ | |
47460d65 | 427 | for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) { |
ccd979bd | 428 | tree[i] = kmem_cache_alloc(ocfs2_uptodate_cachep, |
afae00ab | 429 | GFP_NOFS); |
ccd979bd MF |
430 | if (!tree[i]) { |
431 | mlog_errno(-ENOMEM); | |
432 | goto out_free; | |
433 | } | |
434 | ||
435 | /* These are initialized in ocfs2_expand_cache! */ | |
436 | } | |
437 | } | |
438 | ||
6e5a3d75 | 439 | ocfs2_metadata_cache_lock(ci); |
8cb471e8 | 440 | if (ocfs2_insert_can_use_array(ci)) { |
ccd979bd MF |
441 | mlog(0, "Someone cleared the tree underneath us\n"); |
442 | /* Ok, items were removed from the cache in between | |
443 | * locks. Detect this and revert back to the fast path */ | |
444 | ocfs2_append_cache_array(ci, block); | |
6e5a3d75 | 445 | ocfs2_metadata_cache_unlock(ci); |
ccd979bd MF |
446 | goto out_free; |
447 | } | |
448 | ||
449 | if (expand_tree) | |
8cb471e8 | 450 | ocfs2_expand_cache(ci, tree); |
ccd979bd MF |
451 | |
452 | __ocfs2_insert_cache_tree(ci, new); | |
6e5a3d75 | 453 | ocfs2_metadata_cache_unlock(ci); |
ccd979bd MF |
454 | |
455 | new = NULL; | |
456 | out_free: | |
457 | if (new) | |
458 | kmem_cache_free(ocfs2_uptodate_cachep, new); | |
459 | ||
460 | /* If these were used, then ocfs2_expand_cache re-set them to | |
461 | * NULL for us. */ | |
462 | if (tree[0]) { | |
47460d65 | 463 | for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) |
ccd979bd MF |
464 | if (tree[i]) |
465 | kmem_cache_free(ocfs2_uptodate_cachep, | |
466 | tree[i]); | |
467 | } | |
468 | } | |
469 | ||
6e5a3d75 | 470 | /* Item insertion is guarded by co_io_lock(), so the insertion path takes |
ccd979bd MF |
471 | * advantage of this by not rechecking for a duplicate insert during |
472 | * the slow case. Additionally, if the cache needs to be bumped up to | |
473 | * a tree, the code will not recheck after acquiring the lock -- | |
474 | * multiple paths cannot be expanding to a tree at the same time. | |
475 | * | |
476 | * The slow path takes into account that items can be removed | |
477 | * (including the whole tree wiped and reset) when this process it out | |
478 | * allocating memory. In those cases, it reverts back to the fast | |
479 | * path. | |
480 | * | |
481 | * Note that this function may actually fail to insert the block if | |
482 | * memory cannot be allocated. This is not fatal however (but may | |
aa958874 MF |
483 | * result in a performance penalty) |
484 | * | |
485 | * Readahead buffers can be passed in here before the I/O request is | |
486 | * completed. | |
487 | */ | |
8cb471e8 | 488 | void ocfs2_set_buffer_uptodate(struct ocfs2_caching_info *ci, |
ccd979bd MF |
489 | struct buffer_head *bh) |
490 | { | |
491 | int expand; | |
ccd979bd MF |
492 | |
493 | /* The block may very well exist in our cache already, so avoid | |
494 | * doing any more work in that case. */ | |
8cb471e8 | 495 | if (ocfs2_buffer_cached(ci, bh)) |
ccd979bd MF |
496 | return; |
497 | ||
6e5a3d75 JB |
498 | mlog(0, "Owner %llu, inserting block %llu\n", |
499 | (unsigned long long)ocfs2_metadata_cache_owner(ci), | |
b0697053 | 500 | (unsigned long long)bh->b_blocknr); |
ccd979bd MF |
501 | |
502 | /* No need to recheck under spinlock - insertion is guarded by | |
6e5a3d75 JB |
503 | * co_io_lock() */ |
504 | ocfs2_metadata_cache_lock(ci); | |
8cb471e8 | 505 | if (ocfs2_insert_can_use_array(ci)) { |
ccd979bd MF |
506 | /* Fast case - it's an array and there's a free |
507 | * spot. */ | |
508 | ocfs2_append_cache_array(ci, bh->b_blocknr); | |
6e5a3d75 | 509 | ocfs2_metadata_cache_unlock(ci); |
ccd979bd MF |
510 | return; |
511 | } | |
512 | ||
513 | expand = 0; | |
47460d65 | 514 | if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) { |
ccd979bd MF |
515 | /* We need to bump things up to a tree. */ |
516 | expand = 1; | |
517 | } | |
6e5a3d75 | 518 | ocfs2_metadata_cache_unlock(ci); |
ccd979bd | 519 | |
8cb471e8 | 520 | __ocfs2_set_buffer_uptodate(ci, bh->b_blocknr, expand); |
ccd979bd MF |
521 | } |
522 | ||
523 | /* Called against a newly allocated buffer. Most likely nobody should | |
524 | * be able to read this sort of metadata while it's still being | |
6e5a3d75 | 525 | * allocated, but this is careful to take co_io_lock() anyway. */ |
8cb471e8 | 526 | void ocfs2_set_new_buffer_uptodate(struct ocfs2_caching_info *ci, |
ccd979bd MF |
527 | struct buffer_head *bh) |
528 | { | |
ccd979bd | 529 | /* This should definitely *not* exist in our cache */ |
8cb471e8 | 530 | BUG_ON(ocfs2_buffer_cached(ci, bh)); |
ccd979bd MF |
531 | |
532 | set_buffer_uptodate(bh); | |
533 | ||
6e5a3d75 | 534 | ocfs2_metadata_cache_io_lock(ci); |
8cb471e8 | 535 | ocfs2_set_buffer_uptodate(ci, bh); |
6e5a3d75 | 536 | ocfs2_metadata_cache_io_unlock(ci); |
ccd979bd MF |
537 | } |
538 | ||
539 | /* Requires ip_lock. */ | |
540 | static void ocfs2_remove_metadata_array(struct ocfs2_caching_info *ci, | |
541 | int index) | |
542 | { | |
543 | sector_t *array = ci->ci_cache.ci_array; | |
544 | int bytes; | |
545 | ||
47460d65 | 546 | BUG_ON(index < 0 || index >= OCFS2_CACHE_INFO_MAX_ARRAY); |
ccd979bd MF |
547 | BUG_ON(index >= ci->ci_num_cached); |
548 | BUG_ON(!ci->ci_num_cached); | |
549 | ||
550 | mlog(0, "remove index %d (num_cached = %u\n", index, | |
551 | ci->ci_num_cached); | |
552 | ||
553 | ci->ci_num_cached--; | |
554 | ||
555 | /* don't need to copy if the array is now empty, or if we | |
556 | * removed at the tail */ | |
557 | if (ci->ci_num_cached && index < ci->ci_num_cached) { | |
558 | bytes = sizeof(sector_t) * (ci->ci_num_cached - index); | |
559 | memmove(&array[index], &array[index + 1], bytes); | |
560 | } | |
561 | } | |
562 | ||
563 | /* Requires ip_lock. */ | |
564 | static void ocfs2_remove_metadata_tree(struct ocfs2_caching_info *ci, | |
565 | struct ocfs2_meta_cache_item *item) | |
566 | { | |
567 | mlog(0, "remove block %llu from tree\n", | |
568 | (unsigned long long) item->c_block); | |
569 | ||
570 | rb_erase(&item->c_node, &ci->ci_cache.ci_tree); | |
571 | ci->ci_num_cached--; | |
572 | } | |
573 | ||
8cb471e8 | 574 | static void ocfs2_remove_block_from_cache(struct ocfs2_caching_info *ci, |
ac11c827 | 575 | sector_t block) |
ccd979bd MF |
576 | { |
577 | int index; | |
ccd979bd | 578 | struct ocfs2_meta_cache_item *item = NULL; |
ccd979bd | 579 | |
6e5a3d75 JB |
580 | ocfs2_metadata_cache_lock(ci); |
581 | mlog(0, "Owner %llu, remove %llu, items = %u, array = %u\n", | |
582 | (unsigned long long)ocfs2_metadata_cache_owner(ci), | |
b0697053 | 583 | (unsigned long long) block, ci->ci_num_cached, |
47460d65 | 584 | ci->ci_flags & OCFS2_CACHE_FL_INLINE); |
ccd979bd | 585 | |
47460d65 | 586 | if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) { |
ccd979bd MF |
587 | index = ocfs2_search_cache_array(ci, block); |
588 | if (index != -1) | |
589 | ocfs2_remove_metadata_array(ci, index); | |
590 | } else { | |
591 | item = ocfs2_search_cache_tree(ci, block); | |
592 | if (item) | |
593 | ocfs2_remove_metadata_tree(ci, item); | |
594 | } | |
6e5a3d75 | 595 | ocfs2_metadata_cache_unlock(ci); |
ccd979bd MF |
596 | |
597 | if (item) | |
598 | kmem_cache_free(ocfs2_uptodate_cachep, item); | |
599 | } | |
600 | ||
ac11c827 TM |
601 | /* |
602 | * Called when we remove a chunk of metadata from an inode. We don't | |
603 | * bother reverting things to an inlined array in the case of a remove | |
604 | * which moves us back under the limit. | |
605 | */ | |
8cb471e8 | 606 | void ocfs2_remove_from_cache(struct ocfs2_caching_info *ci, |
ac11c827 TM |
607 | struct buffer_head *bh) |
608 | { | |
609 | sector_t block = bh->b_blocknr; | |
610 | ||
8cb471e8 | 611 | ocfs2_remove_block_from_cache(ci, block); |
ac11c827 TM |
612 | } |
613 | ||
614 | /* Called when we remove xattr clusters from an inode. */ | |
8cb471e8 | 615 | void ocfs2_remove_xattr_clusters_from_cache(struct ocfs2_caching_info *ci, |
ac11c827 TM |
616 | sector_t block, |
617 | u32 c_len) | |
618 | { | |
8cb471e8 JB |
619 | struct super_block *sb = ocfs2_metadata_cache_get_super(ci); |
620 | unsigned int i, b_len = ocfs2_clusters_to_blocks(sb, 1) * c_len; | |
ac11c827 TM |
621 | |
622 | for (i = 0; i < b_len; i++, block++) | |
8cb471e8 | 623 | ocfs2_remove_block_from_cache(ci, block); |
ac11c827 TM |
624 | } |
625 | ||
ccd979bd MF |
626 | int __init init_ocfs2_uptodate_cache(void) |
627 | { | |
628 | ocfs2_uptodate_cachep = kmem_cache_create("ocfs2_uptodate", | |
629 | sizeof(struct ocfs2_meta_cache_item), | |
20c2df83 | 630 | 0, SLAB_HWCACHE_ALIGN, NULL); |
ccd979bd MF |
631 | if (!ocfs2_uptodate_cachep) |
632 | return -ENOMEM; | |
633 | ||
634 | mlog(0, "%u inlined cache items per inode.\n", | |
47460d65 | 635 | OCFS2_CACHE_INFO_MAX_ARRAY); |
ccd979bd MF |
636 | |
637 | return 0; | |
638 | } | |
639 | ||
0c6c98fb | 640 | void exit_ocfs2_uptodate_cache(void) |
ccd979bd MF |
641 | { |
642 | if (ocfs2_uptodate_cachep) | |
643 | kmem_cache_destroy(ocfs2_uptodate_cachep); | |
644 | } |