]>
Commit | Line | Data |
---|---|---|
1802d0be | 1 | // SPDX-License-Identifier: GPL-2.0-only |
fa60ce2c | 2 | /* |
d02f00cc MF |
3 | * reservations.c |
4 | * | |
5 | * Allocation reservations implementation | |
6 | * | |
7 | * Some code borrowed from fs/ext3/balloc.c and is: | |
8 | * | |
9 | * Copyright (C) 1992, 1993, 1994, 1995 | |
10 | * Remy Card (card@masi.ibp.fr) | |
11 | * Laboratoire MASI - Institut Blaise Pascal | |
12 | * Universite Pierre et Marie Curie (Paris VI) | |
13 | * | |
14 | * The rest is copyright (C) 2010 Novell. All rights reserved. | |
d02f00cc MF |
15 | */ |
16 | ||
17 | #include <linux/fs.h> | |
18 | #include <linux/types.h> | |
d02f00cc MF |
19 | #include <linux/highmem.h> |
20 | #include <linux/bitops.h> | |
21 | #include <linux/list.h> | |
22 | ||
d02f00cc MF |
23 | #include <cluster/masklog.h> |
24 | ||
25 | #include "ocfs2.h" | |
2661836e | 26 | #include "ocfs2_trace.h" |
d02f00cc MF |
27 | |
28 | #ifdef CONFIG_OCFS2_DEBUG_FS | |
29 | #define OCFS2_CHECK_RESERVATIONS | |
30 | #endif | |
31 | ||
8989b673 | 32 | static DEFINE_SPINLOCK(resv_lock); |
d02f00cc | 33 | |
83f92318 MF |
34 | int ocfs2_dir_resv_allowed(struct ocfs2_super *osb) |
35 | { | |
36 | return (osb->osb_resv_level && osb->osb_dir_resv_level); | |
37 | } | |
d02f00cc MF |
38 | |
39 | static unsigned int ocfs2_resv_window_bits(struct ocfs2_reservation_map *resmap, | |
40 | struct ocfs2_alloc_reservation *resv) | |
41 | { | |
42 | struct ocfs2_super *osb = resmap->m_osb; | |
43 | unsigned int bits; | |
44 | ||
e3b4a97d MF |
45 | if (!(resv->r_flags & OCFS2_RESV_FLAG_DIR)) { |
46 | /* 8, 16, 32, 64, 128, 256, 512, 1024 */ | |
47 | bits = 4 << osb->osb_resv_level; | |
b07f8f24 | 48 | } else { |
83f92318 | 49 | bits = 4 << osb->osb_dir_resv_level; |
b07f8f24 | 50 | } |
d02f00cc MF |
51 | return bits; |
52 | } | |
53 | ||
54 | static inline unsigned int ocfs2_resv_end(struct ocfs2_alloc_reservation *resv) | |
55 | { | |
56 | if (resv->r_len) | |
57 | return resv->r_start + resv->r_len - 1; | |
58 | return resv->r_start; | |
59 | } | |
60 | ||
61 | static inline int ocfs2_resv_empty(struct ocfs2_alloc_reservation *resv) | |
62 | { | |
63 | return !!(resv->r_len == 0); | |
64 | } | |
65 | ||
66 | static inline int ocfs2_resmap_disabled(struct ocfs2_reservation_map *resmap) | |
67 | { | |
68 | if (resmap->m_osb->osb_resv_level == 0) | |
69 | return 1; | |
70 | return 0; | |
71 | } | |
72 | ||
73 | static void ocfs2_dump_resv(struct ocfs2_reservation_map *resmap) | |
74 | { | |
75 | struct ocfs2_super *osb = resmap->m_osb; | |
76 | struct rb_node *node; | |
77 | struct ocfs2_alloc_reservation *resv; | |
78 | int i = 0; | |
79 | ||
80 | mlog(ML_NOTICE, "Dumping resmap for device %s. Bitmap length: %u\n", | |
81 | osb->dev_str, resmap->m_bitmap_len); | |
82 | ||
83 | node = rb_first(&resmap->m_reservations); | |
84 | while (node) { | |
85 | resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node); | |
86 | ||
87 | mlog(ML_NOTICE, "start: %u\tend: %u\tlen: %u\tlast_start: %u" | |
88 | "\tlast_len: %u\n", resv->r_start, | |
89 | ocfs2_resv_end(resv), resv->r_len, resv->r_last_start, | |
90 | resv->r_last_len); | |
91 | ||
92 | node = rb_next(node); | |
93 | i++; | |
94 | } | |
95 | ||
96 | mlog(ML_NOTICE, "%d reservations found. LRU follows\n", i); | |
97 | ||
98 | i = 0; | |
99 | list_for_each_entry(resv, &resmap->m_lru, r_lru) { | |
100 | mlog(ML_NOTICE, "LRU(%d) start: %u\tend: %u\tlen: %u\t" | |
101 | "last_start: %u\tlast_len: %u\n", i, resv->r_start, | |
102 | ocfs2_resv_end(resv), resv->r_len, resv->r_last_start, | |
103 | resv->r_last_len); | |
104 | ||
105 | i++; | |
106 | } | |
107 | } | |
108 | ||
109 | #ifdef OCFS2_CHECK_RESERVATIONS | |
110 | static int ocfs2_validate_resmap_bits(struct ocfs2_reservation_map *resmap, | |
111 | int i, | |
112 | struct ocfs2_alloc_reservation *resv) | |
113 | { | |
114 | char *disk_bitmap = resmap->m_disk_bitmap; | |
115 | unsigned int start = resv->r_start; | |
116 | unsigned int end = ocfs2_resv_end(resv); | |
117 | ||
118 | while (start <= end) { | |
119 | if (ocfs2_test_bit(start, disk_bitmap)) { | |
120 | mlog(ML_ERROR, | |
121 | "reservation %d covers an allocated area " | |
122 | "starting at bit %u!\n", i, start); | |
123 | return 1; | |
124 | } | |
125 | ||
126 | start++; | |
127 | } | |
128 | return 0; | |
129 | } | |
130 | ||
131 | static void ocfs2_check_resmap(struct ocfs2_reservation_map *resmap) | |
132 | { | |
133 | unsigned int off = 0; | |
134 | int i = 0; | |
135 | struct rb_node *node; | |
136 | struct ocfs2_alloc_reservation *resv; | |
137 | ||
138 | node = rb_first(&resmap->m_reservations); | |
139 | while (node) { | |
140 | resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node); | |
141 | ||
142 | if (i > 0 && resv->r_start <= off) { | |
143 | mlog(ML_ERROR, "reservation %d has bad start off!\n", | |
144 | i); | |
145 | goto bad; | |
146 | } | |
147 | ||
148 | if (resv->r_len == 0) { | |
149 | mlog(ML_ERROR, "reservation %d has no length!\n", | |
150 | i); | |
151 | goto bad; | |
152 | } | |
153 | ||
154 | if (resv->r_start > ocfs2_resv_end(resv)) { | |
155 | mlog(ML_ERROR, "reservation %d has invalid range!\n", | |
156 | i); | |
157 | goto bad; | |
158 | } | |
159 | ||
160 | if (ocfs2_resv_end(resv) >= resmap->m_bitmap_len) { | |
161 | mlog(ML_ERROR, "reservation %d extends past bitmap!\n", | |
162 | i); | |
163 | goto bad; | |
164 | } | |
165 | ||
166 | if (ocfs2_validate_resmap_bits(resmap, i, resv)) | |
167 | goto bad; | |
168 | ||
169 | off = ocfs2_resv_end(resv); | |
170 | node = rb_next(node); | |
171 | ||
172 | i++; | |
173 | } | |
174 | return; | |
175 | ||
176 | bad: | |
177 | ocfs2_dump_resv(resmap); | |
178 | BUG(); | |
179 | } | |
180 | #else | |
181 | static inline void ocfs2_check_resmap(struct ocfs2_reservation_map *resmap) | |
182 | { | |
183 | ||
184 | } | |
185 | #endif | |
186 | ||
187 | void ocfs2_resv_init_once(struct ocfs2_alloc_reservation *resv) | |
188 | { | |
189 | memset(resv, 0, sizeof(*resv)); | |
190 | INIT_LIST_HEAD(&resv->r_lru); | |
191 | } | |
192 | ||
193 | void ocfs2_resv_set_type(struct ocfs2_alloc_reservation *resv, | |
194 | unsigned int flags) | |
195 | { | |
196 | BUG_ON(flags & ~OCFS2_RESV_TYPES); | |
197 | ||
198 | resv->r_flags |= flags; | |
199 | } | |
200 | ||
201 | int ocfs2_resmap_init(struct ocfs2_super *osb, | |
202 | struct ocfs2_reservation_map *resmap) | |
203 | { | |
204 | memset(resmap, 0, sizeof(*resmap)); | |
205 | ||
206 | resmap->m_osb = osb; | |
207 | resmap->m_reservations = RB_ROOT; | |
208 | /* m_bitmap_len is initialized to zero by the above memset. */ | |
209 | INIT_LIST_HEAD(&resmap->m_lru); | |
210 | ||
211 | return 0; | |
212 | } | |
213 | ||
214 | static void ocfs2_resv_mark_lru(struct ocfs2_reservation_map *resmap, | |
215 | struct ocfs2_alloc_reservation *resv) | |
216 | { | |
217 | assert_spin_locked(&resv_lock); | |
218 | ||
219 | if (!list_empty(&resv->r_lru)) | |
220 | list_del_init(&resv->r_lru); | |
221 | ||
222 | list_add_tail(&resv->r_lru, &resmap->m_lru); | |
223 | } | |
224 | ||
225 | static void __ocfs2_resv_trunc(struct ocfs2_alloc_reservation *resv) | |
226 | { | |
227 | resv->r_len = 0; | |
228 | resv->r_start = 0; | |
229 | } | |
230 | ||
231 | static void ocfs2_resv_remove(struct ocfs2_reservation_map *resmap, | |
232 | struct ocfs2_alloc_reservation *resv) | |
233 | { | |
234 | if (resv->r_flags & OCFS2_RESV_FLAG_INUSE) { | |
235 | list_del_init(&resv->r_lru); | |
236 | rb_erase(&resv->r_node, &resmap->m_reservations); | |
237 | resv->r_flags &= ~OCFS2_RESV_FLAG_INUSE; | |
238 | } | |
239 | } | |
240 | ||
241 | static void __ocfs2_resv_discard(struct ocfs2_reservation_map *resmap, | |
242 | struct ocfs2_alloc_reservation *resv) | |
243 | { | |
244 | assert_spin_locked(&resv_lock); | |
245 | ||
246 | __ocfs2_resv_trunc(resv); | |
247 | /* | |
248 | * last_len and last_start no longer make sense if | |
249 | * we're changing the range of our allocations. | |
250 | */ | |
251 | resv->r_last_len = resv->r_last_start = 0; | |
252 | ||
253 | ocfs2_resv_remove(resmap, resv); | |
254 | } | |
255 | ||
256 | /* does nothing if 'resv' is null */ | |
257 | void ocfs2_resv_discard(struct ocfs2_reservation_map *resmap, | |
258 | struct ocfs2_alloc_reservation *resv) | |
259 | { | |
260 | if (resv) { | |
261 | spin_lock(&resv_lock); | |
262 | __ocfs2_resv_discard(resmap, resv); | |
263 | spin_unlock(&resv_lock); | |
264 | } | |
265 | } | |
266 | ||
267 | static void ocfs2_resmap_clear_all_resv(struct ocfs2_reservation_map *resmap) | |
268 | { | |
269 | struct rb_node *node; | |
270 | struct ocfs2_alloc_reservation *resv; | |
271 | ||
272 | assert_spin_locked(&resv_lock); | |
273 | ||
274 | while ((node = rb_last(&resmap->m_reservations)) != NULL) { | |
275 | resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node); | |
276 | ||
277 | __ocfs2_resv_discard(resmap, resv); | |
278 | } | |
279 | } | |
280 | ||
281 | void ocfs2_resmap_restart(struct ocfs2_reservation_map *resmap, | |
282 | unsigned int clen, char *disk_bitmap) | |
283 | { | |
284 | if (ocfs2_resmap_disabled(resmap)) | |
285 | return; | |
286 | ||
287 | spin_lock(&resv_lock); | |
288 | ||
289 | ocfs2_resmap_clear_all_resv(resmap); | |
290 | resmap->m_bitmap_len = clen; | |
291 | resmap->m_disk_bitmap = disk_bitmap; | |
292 | ||
293 | spin_unlock(&resv_lock); | |
294 | } | |
295 | ||
296 | void ocfs2_resmap_uninit(struct ocfs2_reservation_map *resmap) | |
297 | { | |
298 | /* Does nothing for now. Keep this around for API symmetry */ | |
299 | } | |
300 | ||
301 | static void ocfs2_resv_insert(struct ocfs2_reservation_map *resmap, | |
302 | struct ocfs2_alloc_reservation *new) | |
303 | { | |
304 | struct rb_root *root = &resmap->m_reservations; | |
305 | struct rb_node *parent = NULL; | |
306 | struct rb_node **p = &root->rb_node; | |
307 | struct ocfs2_alloc_reservation *tmp; | |
308 | ||
309 | assert_spin_locked(&resv_lock); | |
310 | ||
2661836e | 311 | trace_ocfs2_resv_insert(new->r_start, new->r_len); |
d02f00cc MF |
312 | |
313 | while (*p) { | |
314 | parent = *p; | |
315 | ||
316 | tmp = rb_entry(parent, struct ocfs2_alloc_reservation, r_node); | |
317 | ||
318 | if (new->r_start < tmp->r_start) { | |
319 | p = &(*p)->rb_left; | |
320 | ||
321 | /* | |
322 | * This is a good place to check for | |
323 | * overlapping reservations. | |
324 | */ | |
325 | BUG_ON(ocfs2_resv_end(new) >= tmp->r_start); | |
326 | } else if (new->r_start > ocfs2_resv_end(tmp)) { | |
327 | p = &(*p)->rb_right; | |
328 | } else { | |
329 | /* This should never happen! */ | |
330 | mlog(ML_ERROR, "Duplicate reservation window!\n"); | |
331 | BUG(); | |
332 | } | |
333 | } | |
334 | ||
335 | rb_link_node(&new->r_node, parent, p); | |
336 | rb_insert_color(&new->r_node, root); | |
337 | new->r_flags |= OCFS2_RESV_FLAG_INUSE; | |
338 | ||
339 | ocfs2_resv_mark_lru(resmap, new); | |
340 | ||
341 | ocfs2_check_resmap(resmap); | |
342 | } | |
343 | ||
344 | /** | |
345 | * ocfs2_find_resv_lhs() - find the window which contains goal | |
346 | * @resmap: reservation map to search | |
347 | * @goal: which bit to search for | |
348 | * | |
349 | * If a window containing that goal is not found, we return the window | |
350 | * which comes before goal. Returns NULL on empty rbtree or no window | |
351 | * before goal. | |
352 | */ | |
353 | static struct ocfs2_alloc_reservation * | |
354 | ocfs2_find_resv_lhs(struct ocfs2_reservation_map *resmap, unsigned int goal) | |
355 | { | |
356 | struct ocfs2_alloc_reservation *resv = NULL; | |
357 | struct ocfs2_alloc_reservation *prev_resv = NULL; | |
358 | struct rb_node *node = resmap->m_reservations.rb_node; | |
d02f00cc MF |
359 | |
360 | assert_spin_locked(&resv_lock); | |
361 | ||
362 | if (!node) | |
363 | return NULL; | |
364 | ||
365 | node = rb_first(&resmap->m_reservations); | |
366 | while (node) { | |
367 | resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node); | |
368 | ||
369 | if (resv->r_start <= goal && ocfs2_resv_end(resv) >= goal) | |
370 | break; | |
371 | ||
372 | /* Check if we overshot the reservation just before goal? */ | |
373 | if (resv->r_start > goal) { | |
374 | resv = prev_resv; | |
375 | break; | |
376 | } | |
377 | ||
378 | prev_resv = resv; | |
d02f00cc MF |
379 | node = rb_next(node); |
380 | } | |
381 | ||
382 | return resv; | |
383 | } | |
384 | ||
385 | /* | |
386 | * We are given a range within the bitmap, which corresponds to a gap | |
387 | * inside the reservations tree (search_start, search_len). The range | |
388 | * can be anything from the whole bitmap, to a gap between | |
389 | * reservations. | |
390 | * | |
391 | * The start value of *rstart is insignificant. | |
392 | * | |
393 | * This function searches the bitmap range starting at search_start | |
b065556a | 394 | * with length search_len for a set of contiguous free bits. We try |
d02f00cc MF |
395 | * to find up to 'wanted' bits, but can sometimes return less. |
396 | * | |
397 | * Returns the length of allocation, 0 if no free bits are found. | |
398 | * | |
399 | * *cstart and *clen will also be populated with the result. | |
400 | */ | |
401 | static int ocfs2_resmap_find_free_bits(struct ocfs2_reservation_map *resmap, | |
402 | unsigned int wanted, | |
403 | unsigned int search_start, | |
404 | unsigned int search_len, | |
405 | unsigned int *rstart, | |
406 | unsigned int *rlen) | |
407 | { | |
408 | void *bitmap = resmap->m_disk_bitmap; | |
409 | unsigned int best_start, best_len = 0; | |
410 | int offset, start, found; | |
411 | ||
2661836e TM |
412 | trace_ocfs2_resmap_find_free_bits_begin(search_start, search_len, |
413 | wanted, resmap->m_bitmap_len); | |
d02f00cc MF |
414 | |
415 | found = best_start = best_len = 0; | |
416 | ||
417 | start = search_start; | |
418 | while ((offset = ocfs2_find_next_zero_bit(bitmap, resmap->m_bitmap_len, | |
419 | start)) != -1) { | |
420 | /* Search reached end of the region */ | |
421 | if (offset >= (search_start + search_len)) | |
422 | break; | |
423 | ||
424 | if (offset == start) { | |
425 | /* we found a zero */ | |
426 | found++; | |
427 | /* move start to the next bit to test */ | |
428 | start++; | |
429 | } else { | |
430 | /* got a zero after some ones */ | |
431 | found = 1; | |
432 | start = offset + 1; | |
433 | } | |
434 | if (found > best_len) { | |
435 | best_len = found; | |
436 | best_start = start - found; | |
437 | } | |
438 | ||
439 | if (found >= wanted) | |
440 | break; | |
441 | } | |
442 | ||
443 | if (best_len == 0) | |
444 | return 0; | |
445 | ||
446 | if (best_len >= wanted) | |
447 | best_len = wanted; | |
448 | ||
449 | *rlen = best_len; | |
450 | *rstart = best_start; | |
451 | ||
2661836e | 452 | trace_ocfs2_resmap_find_free_bits_end(best_start, best_len); |
d02f00cc MF |
453 | |
454 | return *rlen; | |
455 | } | |
456 | ||
457 | static void __ocfs2_resv_find_window(struct ocfs2_reservation_map *resmap, | |
458 | struct ocfs2_alloc_reservation *resv, | |
459 | unsigned int goal, unsigned int wanted) | |
460 | { | |
461 | struct rb_root *root = &resmap->m_reservations; | |
462 | unsigned int gap_start, gap_end, gap_len; | |
463 | struct ocfs2_alloc_reservation *prev_resv, *next_resv; | |
464 | struct rb_node *prev, *next; | |
465 | unsigned int cstart, clen; | |
466 | unsigned int best_start = 0, best_len = 0; | |
467 | ||
468 | /* | |
469 | * Nasty cases to consider: | |
470 | * | |
471 | * - rbtree is empty | |
472 | * - our window should be first in all reservations | |
473 | * - our window should be last in all reservations | |
474 | * - need to make sure we don't go past end of bitmap | |
475 | */ | |
2661836e TM |
476 | trace_ocfs2_resv_find_window_begin(resv->r_start, ocfs2_resv_end(resv), |
477 | goal, wanted, RB_EMPTY_ROOT(root)); | |
d02f00cc MF |
478 | |
479 | assert_spin_locked(&resv_lock); | |
480 | ||
481 | if (RB_EMPTY_ROOT(root)) { | |
482 | /* | |
483 | * Easiest case - empty tree. We can just take | |
484 | * whatever window of free bits we want. | |
485 | */ | |
d02f00cc MF |
486 | clen = ocfs2_resmap_find_free_bits(resmap, wanted, goal, |
487 | resmap->m_bitmap_len - goal, | |
488 | &cstart, &clen); | |
489 | ||
490 | /* | |
491 | * This should never happen - the local alloc window | |
492 | * will always have free bits when we're called. | |
493 | */ | |
494 | BUG_ON(goal == 0 && clen == 0); | |
495 | ||
496 | if (clen == 0) | |
497 | return; | |
498 | ||
499 | resv->r_start = cstart; | |
500 | resv->r_len = clen; | |
501 | ||
502 | ocfs2_resv_insert(resmap, resv); | |
503 | return; | |
504 | } | |
505 | ||
506 | prev_resv = ocfs2_find_resv_lhs(resmap, goal); | |
507 | ||
508 | if (prev_resv == NULL) { | |
d02f00cc MF |
509 | /* |
510 | * A NULL here means that the search code couldn't | |
511 | * find a window that starts before goal. | |
512 | * | |
513 | * However, we can take the first window after goal, | |
514 | * which is also by definition, the leftmost window in | |
515 | * the entire tree. If we can find free bits in the | |
516 | * gap between goal and the LHS window, then the | |
517 | * reservation can safely be placed there. | |
518 | * | |
519 | * Otherwise we fall back to a linear search, checking | |
520 | * the gaps in between windows for a place to | |
521 | * allocate. | |
522 | */ | |
523 | ||
524 | next = rb_first(root); | |
525 | next_resv = rb_entry(next, struct ocfs2_alloc_reservation, | |
526 | r_node); | |
527 | ||
528 | /* | |
529 | * The search should never return such a window. (see | |
530 | * comment above | |
531 | */ | |
532 | if (next_resv->r_start <= goal) { | |
533 | mlog(ML_ERROR, "goal: %u next_resv: start %u len %u\n", | |
534 | goal, next_resv->r_start, next_resv->r_len); | |
535 | ocfs2_dump_resv(resmap); | |
536 | BUG(); | |
537 | } | |
538 | ||
539 | clen = ocfs2_resmap_find_free_bits(resmap, wanted, goal, | |
540 | next_resv->r_start - goal, | |
541 | &cstart, &clen); | |
542 | if (clen) { | |
543 | best_len = clen; | |
544 | best_start = cstart; | |
545 | if (best_len == wanted) | |
546 | goto out_insert; | |
547 | } | |
548 | ||
549 | prev_resv = next_resv; | |
550 | next_resv = NULL; | |
551 | } | |
552 | ||
2661836e TM |
553 | trace_ocfs2_resv_find_window_prev(prev_resv->r_start, |
554 | ocfs2_resv_end(prev_resv)); | |
555 | ||
d02f00cc MF |
556 | prev = &prev_resv->r_node; |
557 | ||
558 | /* Now we do a linear search for a window, starting at 'prev_rsv' */ | |
559 | while (1) { | |
560 | next = rb_next(prev); | |
561 | if (next) { | |
d02f00cc MF |
562 | next_resv = rb_entry(next, |
563 | struct ocfs2_alloc_reservation, | |
564 | r_node); | |
565 | ||
566 | gap_start = ocfs2_resv_end(prev_resv) + 1; | |
567 | gap_end = next_resv->r_start - 1; | |
568 | gap_len = gap_end - gap_start + 1; | |
569 | } else { | |
d02f00cc MF |
570 | /* |
571 | * We're at the rightmost edge of the | |
572 | * tree. See if a reservation between this | |
573 | * window and the end of the bitmap will work. | |
574 | */ | |
575 | gap_start = ocfs2_resv_end(prev_resv) + 1; | |
576 | gap_len = resmap->m_bitmap_len - gap_start; | |
577 | gap_end = resmap->m_bitmap_len - 1; | |
578 | } | |
579 | ||
2661836e TM |
580 | trace_ocfs2_resv_find_window_next(next ? next_resv->r_start: -1, |
581 | next ? ocfs2_resv_end(next_resv) : -1); | |
d02f00cc MF |
582 | /* |
583 | * No need to check this gap if we have already found | |
584 | * a larger region of free bits. | |
585 | */ | |
586 | if (gap_len <= best_len) | |
587 | goto next_resv; | |
588 | ||
589 | clen = ocfs2_resmap_find_free_bits(resmap, wanted, gap_start, | |
590 | gap_len, &cstart, &clen); | |
591 | if (clen == wanted) { | |
592 | best_len = clen; | |
593 | best_start = cstart; | |
594 | goto out_insert; | |
595 | } else if (clen > best_len) { | |
596 | best_len = clen; | |
597 | best_start = cstart; | |
598 | } | |
599 | ||
600 | next_resv: | |
601 | if (!next) | |
602 | break; | |
603 | ||
604 | prev = next; | |
605 | prev_resv = rb_entry(prev, struct ocfs2_alloc_reservation, | |
606 | r_node); | |
607 | } | |
608 | ||
609 | out_insert: | |
610 | if (best_len) { | |
611 | resv->r_start = best_start; | |
612 | resv->r_len = best_len; | |
613 | ocfs2_resv_insert(resmap, resv); | |
614 | } | |
615 | } | |
616 | ||
617 | static void ocfs2_cannibalize_resv(struct ocfs2_reservation_map *resmap, | |
618 | struct ocfs2_alloc_reservation *resv, | |
619 | unsigned int wanted) | |
620 | { | |
621 | struct ocfs2_alloc_reservation *lru_resv; | |
622 | int tmpwindow = !!(resv->r_flags & OCFS2_RESV_FLAG_TMP); | |
623 | unsigned int min_bits; | |
624 | ||
625 | if (!tmpwindow) | |
626 | min_bits = ocfs2_resv_window_bits(resmap, resv) >> 1; | |
627 | else | |
628 | min_bits = wanted; /* We at know the temp window will use all | |
629 | * of these bits */ | |
630 | ||
631 | /* | |
632 | * Take the first reservation off the LRU as our 'target'. We | |
633 | * don't try to be smart about it. There might be a case for | |
634 | * searching based on size but I don't have enough data to be | |
635 | * sure. --Mark (3/16/2010) | |
636 | */ | |
637 | lru_resv = list_first_entry(&resmap->m_lru, | |
638 | struct ocfs2_alloc_reservation, r_lru); | |
639 | ||
2661836e TM |
640 | trace_ocfs2_cannibalize_resv_begin(lru_resv->r_start, |
641 | lru_resv->r_len, | |
642 | ocfs2_resv_end(lru_resv)); | |
d02f00cc MF |
643 | |
644 | /* | |
645 | * Cannibalize (some or all) of the target reservation and | |
646 | * feed it to the current window. | |
647 | */ | |
648 | if (lru_resv->r_len <= min_bits) { | |
649 | /* | |
650 | * Discard completely if size is less than or equal to a | |
651 | * reasonable threshold - 50% of window bits for non temporary | |
652 | * windows. | |
653 | */ | |
654 | resv->r_start = lru_resv->r_start; | |
655 | resv->r_len = lru_resv->r_len; | |
656 | ||
657 | __ocfs2_resv_discard(resmap, lru_resv); | |
658 | } else { | |
659 | unsigned int shrink; | |
660 | if (tmpwindow) | |
661 | shrink = min_bits; | |
662 | else | |
663 | shrink = lru_resv->r_len / 2; | |
664 | ||
665 | lru_resv->r_len -= shrink; | |
666 | ||
667 | resv->r_start = ocfs2_resv_end(lru_resv) + 1; | |
668 | resv->r_len = shrink; | |
669 | } | |
670 | ||
2661836e TM |
671 | trace_ocfs2_cannibalize_resv_end(resv->r_start, ocfs2_resv_end(resv), |
672 | resv->r_len, resv->r_last_start, | |
673 | resv->r_last_len); | |
d02f00cc MF |
674 | |
675 | ocfs2_resv_insert(resmap, resv); | |
676 | } | |
677 | ||
678 | static void ocfs2_resv_find_window(struct ocfs2_reservation_map *resmap, | |
679 | struct ocfs2_alloc_reservation *resv, | |
680 | unsigned int wanted) | |
681 | { | |
682 | unsigned int goal = 0; | |
683 | ||
684 | BUG_ON(!ocfs2_resv_empty(resv)); | |
685 | ||
686 | /* | |
687 | * Begin by trying to get a window as close to the previous | |
688 | * one as possible. Using the most recent allocation as a | |
689 | * start goal makes sense. | |
690 | */ | |
691 | if (resv->r_last_len) { | |
692 | goal = resv->r_last_start + resv->r_last_len; | |
693 | if (goal >= resmap->m_bitmap_len) | |
694 | goal = 0; | |
695 | } | |
696 | ||
697 | __ocfs2_resv_find_window(resmap, resv, goal, wanted); | |
698 | ||
699 | /* Search from last alloc didn't work, try once more from beginning. */ | |
700 | if (ocfs2_resv_empty(resv) && goal != 0) | |
701 | __ocfs2_resv_find_window(resmap, resv, 0, wanted); | |
702 | ||
703 | if (ocfs2_resv_empty(resv)) { | |
704 | /* | |
705 | * Still empty? Pull oldest one off the LRU, remove it from | |
706 | * tree, put this one in it's place. | |
707 | */ | |
708 | ocfs2_cannibalize_resv(resmap, resv, wanted); | |
709 | } | |
710 | ||
711 | BUG_ON(ocfs2_resv_empty(resv)); | |
712 | } | |
713 | ||
714 | int ocfs2_resmap_resv_bits(struct ocfs2_reservation_map *resmap, | |
715 | struct ocfs2_alloc_reservation *resv, | |
716 | int *cstart, int *clen) | |
717 | { | |
d02f00cc MF |
718 | if (resv == NULL || ocfs2_resmap_disabled(resmap)) |
719 | return -ENOSPC; | |
720 | ||
721 | spin_lock(&resv_lock); | |
722 | ||
d02f00cc | 723 | if (ocfs2_resv_empty(resv)) { |
4a452de4 TM |
724 | /* |
725 | * We don't want to over-allocate for temporary | |
726 | * windows. Otherwise, we run the risk of fragmenting the | |
727 | * allocation space. | |
728 | */ | |
729 | unsigned int wanted = ocfs2_resv_window_bits(resmap, resv); | |
d02f00cc | 730 | |
4a452de4 TM |
731 | if ((resv->r_flags & OCFS2_RESV_FLAG_TMP) || wanted < *clen) |
732 | wanted = *clen; | |
733 | ||
d02f00cc MF |
734 | /* |
735 | * Try to get a window here. If it works, we must fall | |
736 | * through and test the bitmap . This avoids some | |
737 | * ping-ponging of windows due to non-reserved space | |
738 | * being allocation before we initialize a window for | |
739 | * that inode. | |
740 | */ | |
741 | ocfs2_resv_find_window(resmap, resv, wanted); | |
2661836e | 742 | trace_ocfs2_resmap_resv_bits(resv->r_start, resv->r_len); |
d02f00cc MF |
743 | } |
744 | ||
745 | BUG_ON(ocfs2_resv_empty(resv)); | |
746 | ||
747 | *cstart = resv->r_start; | |
748 | *clen = resv->r_len; | |
749 | ||
750 | spin_unlock(&resv_lock); | |
751 | return 0; | |
752 | } | |
753 | ||
754 | static void | |
755 | ocfs2_adjust_resv_from_alloc(struct ocfs2_reservation_map *resmap, | |
756 | struct ocfs2_alloc_reservation *resv, | |
757 | unsigned int start, unsigned int end) | |
758 | { | |
b065556a TM |
759 | unsigned int rhs = 0; |
760 | unsigned int old_end = ocfs2_resv_end(resv); | |
d02f00cc | 761 | |
b065556a | 762 | BUG_ON(start != resv->r_start || old_end < end); |
d02f00cc MF |
763 | |
764 | /* | |
765 | * Completely used? We can remove it then. | |
766 | */ | |
b065556a | 767 | if (old_end == end) { |
d02f00cc MF |
768 | __ocfs2_resv_discard(resmap, resv); |
769 | return; | |
770 | } | |
771 | ||
b065556a | 772 | rhs = old_end - end; |
d02f00cc MF |
773 | |
774 | /* | |
b065556a | 775 | * This should have been trapped above. |
d02f00cc | 776 | */ |
b065556a | 777 | BUG_ON(rhs == 0); |
d02f00cc | 778 | |
b065556a TM |
779 | resv->r_start = end + 1; |
780 | resv->r_len = old_end - resv->r_start + 1; | |
d02f00cc MF |
781 | } |
782 | ||
783 | void ocfs2_resmap_claimed_bits(struct ocfs2_reservation_map *resmap, | |
784 | struct ocfs2_alloc_reservation *resv, | |
785 | u32 cstart, u32 clen) | |
786 | { | |
787 | unsigned int cend = cstart + clen - 1; | |
788 | ||
789 | if (resmap == NULL || ocfs2_resmap_disabled(resmap)) | |
790 | return; | |
791 | ||
792 | if (resv == NULL) | |
793 | return; | |
794 | ||
b065556a TM |
795 | BUG_ON(cstart != resv->r_start); |
796 | ||
d02f00cc MF |
797 | spin_lock(&resv_lock); |
798 | ||
2661836e TM |
799 | trace_ocfs2_resmap_claimed_bits_begin(cstart, cend, clen, resv->r_start, |
800 | ocfs2_resv_end(resv), resv->r_len, | |
801 | resv->r_last_start, | |
802 | resv->r_last_len); | |
d02f00cc MF |
803 | |
804 | BUG_ON(cstart < resv->r_start); | |
805 | BUG_ON(cstart > ocfs2_resv_end(resv)); | |
806 | BUG_ON(cend > ocfs2_resv_end(resv)); | |
807 | ||
808 | ocfs2_adjust_resv_from_alloc(resmap, resv, cstart, cend); | |
809 | resv->r_last_start = cstart; | |
810 | resv->r_last_len = clen; | |
811 | ||
812 | /* | |
813 | * May have been discarded above from | |
814 | * ocfs2_adjust_resv_from_alloc(). | |
815 | */ | |
816 | if (!ocfs2_resv_empty(resv)) | |
817 | ocfs2_resv_mark_lru(resmap, resv); | |
818 | ||
2661836e TM |
819 | trace_ocfs2_resmap_claimed_bits_end(resv->r_start, ocfs2_resv_end(resv), |
820 | resv->r_len, resv->r_last_start, | |
821 | resv->r_last_len); | |
d02f00cc MF |
822 | |
823 | ocfs2_check_resmap(resmap); | |
824 | ||
825 | spin_unlock(&resv_lock); | |
826 | } |