1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Allocation reservations implementation
8 * Some code borrowed from fs/ext3/balloc.c and is:
10 * Copyright (C) 1992, 1993, 1994, 1995
11 * Remy Card (card@masi.ibp.fr)
12 * Laboratoire MASI - Institut Blaise Pascal
13 * Universite Pierre et Marie Curie (Paris VI)
15 * The rest is copyright (C) 2010 Novell. All rights reserved.
17 * This program is free software; you can redistribute it and/or
18 * modify it under the terms of the GNU General Public
19 * License version 2 as published by the Free Software Foundation.
21 * This program is distributed in the hope that it will be useful,
22 * but WITHOUT ANY WARRANTY; without even the implied warranty of
23 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
24 * General Public License for more details.
28 #include <linux/types.h>
29 #include <linux/slab.h>
30 #include <linux/highmem.h>
31 #include <linux/bitops.h>
32 #include <linux/list.h>
34 #define MLOG_MASK_PREFIX ML_RESERVATIONS
35 #include <cluster/masklog.h>
39 #ifdef CONFIG_OCFS2_DEBUG_FS
40 #define OCFS2_CHECK_RESERVATIONS
43 DEFINE_SPINLOCK(resv_lock);
45 #define OCFS2_MIN_RESV_WINDOW_BITS 8
46 #define OCFS2_MAX_RESV_WINDOW_BITS 1024
48 int ocfs2_dir_resv_allowed(struct ocfs2_super *osb)
50 return (osb->osb_resv_level && osb->osb_dir_resv_level);
53 static unsigned int ocfs2_resv_window_bits(struct ocfs2_reservation_map *resmap,
54 struct ocfs2_alloc_reservation *resv)
56 struct ocfs2_super *osb = resmap->m_osb;
59 if (!(resv->r_flags & OCFS2_RESV_FLAG_DIR)) {
60 /* 8, 16, 32, 64, 128, 256, 512, 1024 */
61 bits = 4 << osb->osb_resv_level;
63 bits = 4 << osb->osb_dir_resv_level;
68 static inline unsigned int ocfs2_resv_end(struct ocfs2_alloc_reservation *resv)
71 return resv->r_start + resv->r_len - 1;
75 static inline int ocfs2_resv_empty(struct ocfs2_alloc_reservation *resv)
77 return !!(resv->r_len == 0);
80 static inline int ocfs2_resmap_disabled(struct ocfs2_reservation_map *resmap)
82 if (resmap->m_osb->osb_resv_level == 0)
87 static void ocfs2_dump_resv(struct ocfs2_reservation_map *resmap)
89 struct ocfs2_super *osb = resmap->m_osb;
91 struct ocfs2_alloc_reservation *resv;
94 mlog(ML_NOTICE, "Dumping resmap for device %s. Bitmap length: %u\n",
95 osb->dev_str, resmap->m_bitmap_len);
97 node = rb_first(&resmap->m_reservations);
99 resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node);
101 mlog(ML_NOTICE, "start: %u\tend: %u\tlen: %u\tlast_start: %u"
102 "\tlast_len: %u\n", resv->r_start,
103 ocfs2_resv_end(resv), resv->r_len, resv->r_last_start,
106 node = rb_next(node);
110 mlog(ML_NOTICE, "%d reservations found. LRU follows\n", i);
113 list_for_each_entry(resv, &resmap->m_lru, r_lru) {
114 mlog(ML_NOTICE, "LRU(%d) start: %u\tend: %u\tlen: %u\t"
115 "last_start: %u\tlast_len: %u\n", i, resv->r_start,
116 ocfs2_resv_end(resv), resv->r_len, resv->r_last_start,
123 #ifdef OCFS2_CHECK_RESERVATIONS
124 static int ocfs2_validate_resmap_bits(struct ocfs2_reservation_map *resmap,
126 struct ocfs2_alloc_reservation *resv)
128 char *disk_bitmap = resmap->m_disk_bitmap;
129 unsigned int start = resv->r_start;
130 unsigned int end = ocfs2_resv_end(resv);
132 while (start <= end) {
133 if (ocfs2_test_bit(start, disk_bitmap)) {
135 "reservation %d covers an allocated area "
136 "starting at bit %u!\n", i, start);
145 static void ocfs2_check_resmap(struct ocfs2_reservation_map *resmap)
147 unsigned int off = 0;
149 struct rb_node *node;
150 struct ocfs2_alloc_reservation *resv;
152 node = rb_first(&resmap->m_reservations);
154 resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node);
156 if (i > 0 && resv->r_start <= off) {
157 mlog(ML_ERROR, "reservation %d has bad start off!\n",
162 if (resv->r_len == 0) {
163 mlog(ML_ERROR, "reservation %d has no length!\n",
168 if (resv->r_start > ocfs2_resv_end(resv)) {
169 mlog(ML_ERROR, "reservation %d has invalid range!\n",
174 if (ocfs2_resv_end(resv) >= resmap->m_bitmap_len) {
175 mlog(ML_ERROR, "reservation %d extends past bitmap!\n",
180 if (ocfs2_validate_resmap_bits(resmap, i, resv))
183 off = ocfs2_resv_end(resv);
184 node = rb_next(node);
191 ocfs2_dump_resv(resmap);
195 static inline void ocfs2_check_resmap(struct ocfs2_reservation_map *resmap)
201 void ocfs2_resv_init_once(struct ocfs2_alloc_reservation *resv)
203 memset(resv, 0, sizeof(*resv));
204 INIT_LIST_HEAD(&resv->r_lru);
207 void ocfs2_resv_set_type(struct ocfs2_alloc_reservation *resv,
210 BUG_ON(flags & ~OCFS2_RESV_TYPES);
212 resv->r_flags |= flags;
215 int ocfs2_resmap_init(struct ocfs2_super *osb,
216 struct ocfs2_reservation_map *resmap)
218 memset(resmap, 0, sizeof(*resmap));
221 resmap->m_reservations = RB_ROOT;
222 /* m_bitmap_len is initialized to zero by the above memset. */
223 INIT_LIST_HEAD(&resmap->m_lru);
228 static void ocfs2_resv_mark_lru(struct ocfs2_reservation_map *resmap,
229 struct ocfs2_alloc_reservation *resv)
231 assert_spin_locked(&resv_lock);
233 if (!list_empty(&resv->r_lru))
234 list_del_init(&resv->r_lru);
236 list_add_tail(&resv->r_lru, &resmap->m_lru);
239 static void __ocfs2_resv_trunc(struct ocfs2_alloc_reservation *resv)
245 static void ocfs2_resv_remove(struct ocfs2_reservation_map *resmap,
246 struct ocfs2_alloc_reservation *resv)
248 if (resv->r_flags & OCFS2_RESV_FLAG_INUSE) {
249 list_del_init(&resv->r_lru);
250 rb_erase(&resv->r_node, &resmap->m_reservations);
251 resv->r_flags &= ~OCFS2_RESV_FLAG_INUSE;
255 static void __ocfs2_resv_discard(struct ocfs2_reservation_map *resmap,
256 struct ocfs2_alloc_reservation *resv)
258 assert_spin_locked(&resv_lock);
260 __ocfs2_resv_trunc(resv);
262 * last_len and last_start no longer make sense if
263 * we're changing the range of our allocations.
265 resv->r_last_len = resv->r_last_start = 0;
267 ocfs2_resv_remove(resmap, resv);
270 /* does nothing if 'resv' is null */
271 void ocfs2_resv_discard(struct ocfs2_reservation_map *resmap,
272 struct ocfs2_alloc_reservation *resv)
275 spin_lock(&resv_lock);
276 __ocfs2_resv_discard(resmap, resv);
277 spin_unlock(&resv_lock);
281 static void ocfs2_resmap_clear_all_resv(struct ocfs2_reservation_map *resmap)
283 struct rb_node *node;
284 struct ocfs2_alloc_reservation *resv;
286 assert_spin_locked(&resv_lock);
288 while ((node = rb_last(&resmap->m_reservations)) != NULL) {
289 resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node);
291 __ocfs2_resv_discard(resmap, resv);
295 void ocfs2_resmap_restart(struct ocfs2_reservation_map *resmap,
296 unsigned int clen, char *disk_bitmap)
298 if (ocfs2_resmap_disabled(resmap))
301 spin_lock(&resv_lock);
303 ocfs2_resmap_clear_all_resv(resmap);
304 resmap->m_bitmap_len = clen;
305 resmap->m_disk_bitmap = disk_bitmap;
307 spin_unlock(&resv_lock);
310 void ocfs2_resmap_uninit(struct ocfs2_reservation_map *resmap)
312 /* Does nothing for now. Keep this around for API symmetry */
315 static void ocfs2_resv_insert(struct ocfs2_reservation_map *resmap,
316 struct ocfs2_alloc_reservation *new)
318 struct rb_root *root = &resmap->m_reservations;
319 struct rb_node *parent = NULL;
320 struct rb_node **p = &root->rb_node;
321 struct ocfs2_alloc_reservation *tmp;
323 assert_spin_locked(&resv_lock);
325 mlog(0, "Insert reservation start: %u len: %u\n", new->r_start,
331 tmp = rb_entry(parent, struct ocfs2_alloc_reservation, r_node);
333 if (new->r_start < tmp->r_start) {
337 * This is a good place to check for
338 * overlapping reservations.
340 BUG_ON(ocfs2_resv_end(new) >= tmp->r_start);
341 } else if (new->r_start > ocfs2_resv_end(tmp)) {
344 /* This should never happen! */
345 mlog(ML_ERROR, "Duplicate reservation window!\n");
350 rb_link_node(&new->r_node, parent, p);
351 rb_insert_color(&new->r_node, root);
352 new->r_flags |= OCFS2_RESV_FLAG_INUSE;
354 ocfs2_resv_mark_lru(resmap, new);
356 ocfs2_check_resmap(resmap);
360 * ocfs2_find_resv_lhs() - find the window which contains goal
361 * @resmap: reservation map to search
362 * @goal: which bit to search for
364 * If a window containing that goal is not found, we return the window
365 * which comes before goal. Returns NULL on empty rbtree or no window
368 static struct ocfs2_alloc_reservation *
369 ocfs2_find_resv_lhs(struct ocfs2_reservation_map *resmap, unsigned int goal)
371 struct ocfs2_alloc_reservation *resv = NULL;
372 struct ocfs2_alloc_reservation *prev_resv = NULL;
373 struct rb_node *node = resmap->m_reservations.rb_node;
374 struct rb_node *prev = NULL;
376 assert_spin_locked(&resv_lock);
381 node = rb_first(&resmap->m_reservations);
383 resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node);
385 if (resv->r_start <= goal && ocfs2_resv_end(resv) >= goal)
388 /* Check if we overshot the reservation just before goal? */
389 if (resv->r_start > goal) {
396 node = rb_next(node);
403 * We are given a range within the bitmap, which corresponds to a gap
404 * inside the reservations tree (search_start, search_len). The range
405 * can be anything from the whole bitmap, to a gap between
408 * The start value of *rstart is insignificant.
410 * This function searches the bitmap range starting at search_start
411 * with length search_len for a set of contiguous free bits. We try
412 * to find up to 'wanted' bits, but can sometimes return less.
414 * Returns the length of allocation, 0 if no free bits are found.
416 * *cstart and *clen will also be populated with the result.
418 static int ocfs2_resmap_find_free_bits(struct ocfs2_reservation_map *resmap,
420 unsigned int search_start,
421 unsigned int search_len,
422 unsigned int *rstart,
425 void *bitmap = resmap->m_disk_bitmap;
426 unsigned int best_start, best_len = 0;
427 int offset, start, found;
429 mlog(0, "Find %u bits within range (%u, len %u) resmap len: %u\n",
430 wanted, search_start, search_len, resmap->m_bitmap_len);
432 found = best_start = best_len = 0;
434 start = search_start;
435 while ((offset = ocfs2_find_next_zero_bit(bitmap, resmap->m_bitmap_len,
437 /* Search reached end of the region */
438 if (offset >= (search_start + search_len))
441 if (offset == start) {
442 /* we found a zero */
444 /* move start to the next bit to test */
447 /* got a zero after some ones */
451 if (found > best_len) {
453 best_start = start - found;
463 if (best_len >= wanted)
467 *rstart = best_start;
469 mlog(0, "Found start: %u len: %u\n", best_start, best_len);
474 static void __ocfs2_resv_find_window(struct ocfs2_reservation_map *resmap,
475 struct ocfs2_alloc_reservation *resv,
476 unsigned int goal, unsigned int wanted)
478 struct rb_root *root = &resmap->m_reservations;
479 unsigned int gap_start, gap_end, gap_len;
480 struct ocfs2_alloc_reservation *prev_resv, *next_resv;
481 struct rb_node *prev, *next;
482 unsigned int cstart, clen;
483 unsigned int best_start = 0, best_len = 0;
486 * Nasty cases to consider:
489 * - our window should be first in all reservations
490 * - our window should be last in all reservations
491 * - need to make sure we don't go past end of bitmap
494 mlog(0, "resv start: %u resv end: %u goal: %u wanted: %u\n",
495 resv->r_start, ocfs2_resv_end(resv), goal, wanted);
497 assert_spin_locked(&resv_lock);
499 if (RB_EMPTY_ROOT(root)) {
501 * Easiest case - empty tree. We can just take
502 * whatever window of free bits we want.
505 mlog(0, "Empty root\n");
507 clen = ocfs2_resmap_find_free_bits(resmap, wanted, goal,
508 resmap->m_bitmap_len - goal,
512 * This should never happen - the local alloc window
513 * will always have free bits when we're called.
515 BUG_ON(goal == 0 && clen == 0);
520 resv->r_start = cstart;
523 ocfs2_resv_insert(resmap, resv);
527 prev_resv = ocfs2_find_resv_lhs(resmap, goal);
529 if (prev_resv == NULL) {
530 mlog(0, "Goal on LHS of leftmost window\n");
533 * A NULL here means that the search code couldn't
534 * find a window that starts before goal.
536 * However, we can take the first window after goal,
537 * which is also by definition, the leftmost window in
538 * the entire tree. If we can find free bits in the
539 * gap between goal and the LHS window, then the
540 * reservation can safely be placed there.
542 * Otherwise we fall back to a linear search, checking
543 * the gaps in between windows for a place to
547 next = rb_first(root);
548 next_resv = rb_entry(next, struct ocfs2_alloc_reservation,
552 * The search should never return such a window. (see
555 if (next_resv->r_start <= goal) {
556 mlog(ML_ERROR, "goal: %u next_resv: start %u len %u\n",
557 goal, next_resv->r_start, next_resv->r_len);
558 ocfs2_dump_resv(resmap);
562 clen = ocfs2_resmap_find_free_bits(resmap, wanted, goal,
563 next_resv->r_start - goal,
568 if (best_len == wanted)
572 prev_resv = next_resv;
576 prev = &prev_resv->r_node;
578 /* Now we do a linear search for a window, starting at 'prev_rsv' */
580 next = rb_next(prev);
582 mlog(0, "One more resv found in linear search\n");
583 next_resv = rb_entry(next,
584 struct ocfs2_alloc_reservation,
587 gap_start = ocfs2_resv_end(prev_resv) + 1;
588 gap_end = next_resv->r_start - 1;
589 gap_len = gap_end - gap_start + 1;
591 mlog(0, "No next node\n");
593 * We're at the rightmost edge of the
594 * tree. See if a reservation between this
595 * window and the end of the bitmap will work.
597 gap_start = ocfs2_resv_end(prev_resv) + 1;
598 gap_len = resmap->m_bitmap_len - gap_start;
599 gap_end = resmap->m_bitmap_len - 1;
603 * No need to check this gap if we have already found
604 * a larger region of free bits.
606 if (gap_len <= best_len)
609 clen = ocfs2_resmap_find_free_bits(resmap, wanted, gap_start,
610 gap_len, &cstart, &clen);
611 if (clen == wanted) {
615 } else if (clen > best_len) {
625 prev_resv = rb_entry(prev, struct ocfs2_alloc_reservation,
631 resv->r_start = best_start;
632 resv->r_len = best_len;
633 ocfs2_resv_insert(resmap, resv);
637 static void ocfs2_cannibalize_resv(struct ocfs2_reservation_map *resmap,
638 struct ocfs2_alloc_reservation *resv,
641 struct ocfs2_alloc_reservation *lru_resv;
642 int tmpwindow = !!(resv->r_flags & OCFS2_RESV_FLAG_TMP);
643 unsigned int min_bits;
646 min_bits = ocfs2_resv_window_bits(resmap, resv) >> 1;
648 min_bits = wanted; /* We at know the temp window will use all
652 * Take the first reservation off the LRU as our 'target'. We
653 * don't try to be smart about it. There might be a case for
654 * searching based on size but I don't have enough data to be
655 * sure. --Mark (3/16/2010)
657 lru_resv = list_first_entry(&resmap->m_lru,
658 struct ocfs2_alloc_reservation, r_lru);
660 mlog(0, "lru resv: start: %u len: %u end: %u\n", lru_resv->r_start,
661 lru_resv->r_len, ocfs2_resv_end(lru_resv));
664 * Cannibalize (some or all) of the target reservation and
665 * feed it to the current window.
667 if (lru_resv->r_len <= min_bits) {
669 * Discard completely if size is less than or equal to a
670 * reasonable threshold - 50% of window bits for non temporary
673 resv->r_start = lru_resv->r_start;
674 resv->r_len = lru_resv->r_len;
676 __ocfs2_resv_discard(resmap, lru_resv);
682 shrink = lru_resv->r_len / 2;
684 lru_resv->r_len -= shrink;
686 resv->r_start = ocfs2_resv_end(lru_resv) + 1;
687 resv->r_len = shrink;
690 mlog(0, "Reservation now looks like: r_start: %u r_end: %u "
691 "r_len: %u r_last_start: %u r_last_len: %u\n",
692 resv->r_start, ocfs2_resv_end(resv), resv->r_len,
693 resv->r_last_start, resv->r_last_len);
695 ocfs2_resv_insert(resmap, resv);
698 static void ocfs2_resv_find_window(struct ocfs2_reservation_map *resmap,
699 struct ocfs2_alloc_reservation *resv,
702 unsigned int goal = 0;
704 BUG_ON(!ocfs2_resv_empty(resv));
707 * Begin by trying to get a window as close to the previous
708 * one as possible. Using the most recent allocation as a
709 * start goal makes sense.
711 if (resv->r_last_len) {
712 goal = resv->r_last_start + resv->r_last_len;
713 if (goal >= resmap->m_bitmap_len)
717 __ocfs2_resv_find_window(resmap, resv, goal, wanted);
719 /* Search from last alloc didn't work, try once more from beginning. */
720 if (ocfs2_resv_empty(resv) && goal != 0)
721 __ocfs2_resv_find_window(resmap, resv, 0, wanted);
723 if (ocfs2_resv_empty(resv)) {
725 * Still empty? Pull oldest one off the LRU, remove it from
726 * tree, put this one in it's place.
728 ocfs2_cannibalize_resv(resmap, resv, wanted);
731 BUG_ON(ocfs2_resv_empty(resv));
734 int ocfs2_resmap_resv_bits(struct ocfs2_reservation_map *resmap,
735 struct ocfs2_alloc_reservation *resv,
736 int *cstart, int *clen)
738 unsigned int wanted = *clen;
740 if (resv == NULL || ocfs2_resmap_disabled(resmap))
743 spin_lock(&resv_lock);
746 * We don't want to over-allocate for temporary
747 * windows. Otherwise, we run the risk of fragmenting the
750 wanted = ocfs2_resv_window_bits(resmap, resv);
751 if ((resv->r_flags & OCFS2_RESV_FLAG_TMP) || wanted < *clen)
754 if (ocfs2_resv_empty(resv)) {
755 mlog(0, "empty reservation, find new window\n");
758 * Try to get a window here. If it works, we must fall
759 * through and test the bitmap . This avoids some
760 * ping-ponging of windows due to non-reserved space
761 * being allocation before we initialize a window for
764 ocfs2_resv_find_window(resmap, resv, wanted);
767 BUG_ON(ocfs2_resv_empty(resv));
769 *cstart = resv->r_start;
772 spin_unlock(&resv_lock);
777 ocfs2_adjust_resv_from_alloc(struct ocfs2_reservation_map *resmap,
778 struct ocfs2_alloc_reservation *resv,
779 unsigned int start, unsigned int end)
781 unsigned int rhs = 0;
782 unsigned int old_end = ocfs2_resv_end(resv);
784 BUG_ON(start != resv->r_start || old_end < end);
787 * Completely used? We can remove it then.
789 if (old_end == end) {
790 __ocfs2_resv_discard(resmap, resv);
797 * This should have been trapped above.
801 resv->r_start = end + 1;
802 resv->r_len = old_end - resv->r_start + 1;
805 void ocfs2_resmap_claimed_bits(struct ocfs2_reservation_map *resmap,
806 struct ocfs2_alloc_reservation *resv,
807 u32 cstart, u32 clen)
809 unsigned int cend = cstart + clen - 1;
811 if (resmap == NULL || ocfs2_resmap_disabled(resmap))
817 BUG_ON(cstart != resv->r_start);
819 spin_lock(&resv_lock);
821 mlog(0, "claim bits: cstart: %u cend: %u clen: %u r_start: %u "
822 "r_end: %u r_len: %u, r_last_start: %u r_last_len: %u\n",
823 cstart, cend, clen, resv->r_start, ocfs2_resv_end(resv),
824 resv->r_len, resv->r_last_start, resv->r_last_len);
826 BUG_ON(cstart < resv->r_start);
827 BUG_ON(cstart > ocfs2_resv_end(resv));
828 BUG_ON(cend > ocfs2_resv_end(resv));
830 ocfs2_adjust_resv_from_alloc(resmap, resv, cstart, cend);
831 resv->r_last_start = cstart;
832 resv->r_last_len = clen;
835 * May have been discarded above from
836 * ocfs2_adjust_resv_from_alloc().
838 if (!ocfs2_resv_empty(resv))
839 ocfs2_resv_mark_lru(resmap, resv);
841 mlog(0, "Reservation now looks like: r_start: %u r_end: %u "
842 "r_len: %u r_last_start: %u r_last_len: %u\n",
843 resv->r_start, ocfs2_resv_end(resv), resv->r_len,
844 resv->r_last_start, resv->r_last_len);
846 ocfs2_check_resmap(resmap);
848 spin_unlock(&resv_lock);