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
47 #define OCFS2_RESV_DIR_WINDOW_BITS OCFS2_MIN_RESV_WINDOW_BITS
49 static unsigned int ocfs2_resv_window_bits(struct ocfs2_reservation_map *resmap,
50 struct ocfs2_alloc_reservation *resv)
52 struct ocfs2_super *osb = resmap->m_osb;
55 if (!(resv->r_flags & OCFS2_RESV_FLAG_DIR)) {
56 /* 8, 16, 32, 64, 128, 256, 512, 1024 */
57 bits = 4 << osb->osb_resv_level;
59 /* For now, treat directories the same as files. */
60 bits = 4 << osb->osb_resv_level;
65 static inline unsigned int ocfs2_resv_end(struct ocfs2_alloc_reservation *resv)
68 return resv->r_start + resv->r_len - 1;
72 static inline int ocfs2_resv_empty(struct ocfs2_alloc_reservation *resv)
74 return !!(resv->r_len == 0);
77 static inline int ocfs2_resmap_disabled(struct ocfs2_reservation_map *resmap)
79 if (resmap->m_osb->osb_resv_level == 0)
84 static void ocfs2_dump_resv(struct ocfs2_reservation_map *resmap)
86 struct ocfs2_super *osb = resmap->m_osb;
88 struct ocfs2_alloc_reservation *resv;
91 mlog(ML_NOTICE, "Dumping resmap for device %s. Bitmap length: %u\n",
92 osb->dev_str, resmap->m_bitmap_len);
94 node = rb_first(&resmap->m_reservations);
96 resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node);
98 mlog(ML_NOTICE, "start: %u\tend: %u\tlen: %u\tlast_start: %u"
99 "\tlast_len: %u\n", resv->r_start,
100 ocfs2_resv_end(resv), resv->r_len, resv->r_last_start,
103 node = rb_next(node);
107 mlog(ML_NOTICE, "%d reservations found. LRU follows\n", i);
110 list_for_each_entry(resv, &resmap->m_lru, r_lru) {
111 mlog(ML_NOTICE, "LRU(%d) start: %u\tend: %u\tlen: %u\t"
112 "last_start: %u\tlast_len: %u\n", i, resv->r_start,
113 ocfs2_resv_end(resv), resv->r_len, resv->r_last_start,
120 #ifdef OCFS2_CHECK_RESERVATIONS
121 static int ocfs2_validate_resmap_bits(struct ocfs2_reservation_map *resmap,
123 struct ocfs2_alloc_reservation *resv)
125 char *disk_bitmap = resmap->m_disk_bitmap;
126 unsigned int start = resv->r_start;
127 unsigned int end = ocfs2_resv_end(resv);
129 while (start <= end) {
130 if (ocfs2_test_bit(start, disk_bitmap)) {
132 "reservation %d covers an allocated area "
133 "starting at bit %u!\n", i, start);
142 static void ocfs2_check_resmap(struct ocfs2_reservation_map *resmap)
144 unsigned int off = 0;
146 struct rb_node *node;
147 struct ocfs2_alloc_reservation *resv;
149 node = rb_first(&resmap->m_reservations);
151 resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node);
153 if (i > 0 && resv->r_start <= off) {
154 mlog(ML_ERROR, "reservation %d has bad start off!\n",
159 if (resv->r_len == 0) {
160 mlog(ML_ERROR, "reservation %d has no length!\n",
165 if (resv->r_start > ocfs2_resv_end(resv)) {
166 mlog(ML_ERROR, "reservation %d has invalid range!\n",
171 if (ocfs2_resv_end(resv) >= resmap->m_bitmap_len) {
172 mlog(ML_ERROR, "reservation %d extends past bitmap!\n",
177 if (ocfs2_validate_resmap_bits(resmap, i, resv))
180 off = ocfs2_resv_end(resv);
181 node = rb_next(node);
188 ocfs2_dump_resv(resmap);
192 static inline void ocfs2_check_resmap(struct ocfs2_reservation_map *resmap)
198 void ocfs2_resv_init_once(struct ocfs2_alloc_reservation *resv)
200 memset(resv, 0, sizeof(*resv));
201 INIT_LIST_HEAD(&resv->r_lru);
204 void ocfs2_resv_set_type(struct ocfs2_alloc_reservation *resv,
207 BUG_ON(flags & ~OCFS2_RESV_TYPES);
209 resv->r_flags |= flags;
212 int ocfs2_resmap_init(struct ocfs2_super *osb,
213 struct ocfs2_reservation_map *resmap)
215 memset(resmap, 0, sizeof(*resmap));
218 resmap->m_reservations = RB_ROOT;
219 /* m_bitmap_len is initialized to zero by the above memset. */
220 INIT_LIST_HEAD(&resmap->m_lru);
225 static void ocfs2_resv_mark_lru(struct ocfs2_reservation_map *resmap,
226 struct ocfs2_alloc_reservation *resv)
228 assert_spin_locked(&resv_lock);
230 if (!list_empty(&resv->r_lru))
231 list_del_init(&resv->r_lru);
233 list_add_tail(&resv->r_lru, &resmap->m_lru);
236 static void __ocfs2_resv_trunc(struct ocfs2_alloc_reservation *resv)
242 static void ocfs2_resv_remove(struct ocfs2_reservation_map *resmap,
243 struct ocfs2_alloc_reservation *resv)
245 if (resv->r_flags & OCFS2_RESV_FLAG_INUSE) {
246 list_del_init(&resv->r_lru);
247 rb_erase(&resv->r_node, &resmap->m_reservations);
248 resv->r_flags &= ~OCFS2_RESV_FLAG_INUSE;
252 static void __ocfs2_resv_discard(struct ocfs2_reservation_map *resmap,
253 struct ocfs2_alloc_reservation *resv)
255 assert_spin_locked(&resv_lock);
257 __ocfs2_resv_trunc(resv);
259 * last_len and last_start no longer make sense if
260 * we're changing the range of our allocations.
262 resv->r_last_len = resv->r_last_start = 0;
264 ocfs2_resv_remove(resmap, resv);
267 /* does nothing if 'resv' is null */
268 void ocfs2_resv_discard(struct ocfs2_reservation_map *resmap,
269 struct ocfs2_alloc_reservation *resv)
272 spin_lock(&resv_lock);
273 __ocfs2_resv_discard(resmap, resv);
274 spin_unlock(&resv_lock);
278 static void ocfs2_resmap_clear_all_resv(struct ocfs2_reservation_map *resmap)
280 struct rb_node *node;
281 struct ocfs2_alloc_reservation *resv;
283 assert_spin_locked(&resv_lock);
285 while ((node = rb_last(&resmap->m_reservations)) != NULL) {
286 resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node);
288 __ocfs2_resv_discard(resmap, resv);
292 void ocfs2_resmap_restart(struct ocfs2_reservation_map *resmap,
293 unsigned int clen, char *disk_bitmap)
295 if (ocfs2_resmap_disabled(resmap))
298 spin_lock(&resv_lock);
300 ocfs2_resmap_clear_all_resv(resmap);
301 resmap->m_bitmap_len = clen;
302 resmap->m_disk_bitmap = disk_bitmap;
304 spin_unlock(&resv_lock);
307 void ocfs2_resmap_uninit(struct ocfs2_reservation_map *resmap)
309 /* Does nothing for now. Keep this around for API symmetry */
312 static void ocfs2_resv_insert(struct ocfs2_reservation_map *resmap,
313 struct ocfs2_alloc_reservation *new)
315 struct rb_root *root = &resmap->m_reservations;
316 struct rb_node *parent = NULL;
317 struct rb_node **p = &root->rb_node;
318 struct ocfs2_alloc_reservation *tmp;
320 assert_spin_locked(&resv_lock);
322 mlog(0, "Insert reservation start: %u len: %u\n", new->r_start,
328 tmp = rb_entry(parent, struct ocfs2_alloc_reservation, r_node);
330 if (new->r_start < tmp->r_start) {
334 * This is a good place to check for
335 * overlapping reservations.
337 BUG_ON(ocfs2_resv_end(new) >= tmp->r_start);
338 } else if (new->r_start > ocfs2_resv_end(tmp)) {
341 /* This should never happen! */
342 mlog(ML_ERROR, "Duplicate reservation window!\n");
347 rb_link_node(&new->r_node, parent, p);
348 rb_insert_color(&new->r_node, root);
349 new->r_flags |= OCFS2_RESV_FLAG_INUSE;
351 ocfs2_resv_mark_lru(resmap, new);
353 ocfs2_check_resmap(resmap);
357 * ocfs2_find_resv_lhs() - find the window which contains goal
358 * @resmap: reservation map to search
359 * @goal: which bit to search for
361 * If a window containing that goal is not found, we return the window
362 * which comes before goal. Returns NULL on empty rbtree or no window
365 static struct ocfs2_alloc_reservation *
366 ocfs2_find_resv_lhs(struct ocfs2_reservation_map *resmap, unsigned int goal)
368 struct ocfs2_alloc_reservation *resv = NULL;
369 struct ocfs2_alloc_reservation *prev_resv = NULL;
370 struct rb_node *node = resmap->m_reservations.rb_node;
371 struct rb_node *prev = NULL;
373 assert_spin_locked(&resv_lock);
378 node = rb_first(&resmap->m_reservations);
380 resv = rb_entry(node, struct ocfs2_alloc_reservation, r_node);
382 if (resv->r_start <= goal && ocfs2_resv_end(resv) >= goal)
385 /* Check if we overshot the reservation just before goal? */
386 if (resv->r_start > goal) {
393 node = rb_next(node);
400 * We are given a range within the bitmap, which corresponds to a gap
401 * inside the reservations tree (search_start, search_len). The range
402 * can be anything from the whole bitmap, to a gap between
405 * The start value of *rstart is insignificant.
407 * This function searches the bitmap range starting at search_start
408 * with length csearch_len for a set of contiguous free bits. We try
409 * to find up to 'wanted' bits, but can sometimes return less.
411 * Returns the length of allocation, 0 if no free bits are found.
413 * *cstart and *clen will also be populated with the result.
415 static int ocfs2_resmap_find_free_bits(struct ocfs2_reservation_map *resmap,
417 unsigned int search_start,
418 unsigned int search_len,
419 unsigned int *rstart,
422 void *bitmap = resmap->m_disk_bitmap;
423 unsigned int best_start, best_len = 0;
424 int offset, start, found;
426 mlog(0, "Find %u bits within range (%u, len %u) resmap len: %u\n",
427 wanted, search_start, search_len, resmap->m_bitmap_len);
429 found = best_start = best_len = 0;
431 start = search_start;
432 while ((offset = ocfs2_find_next_zero_bit(bitmap, resmap->m_bitmap_len,
434 /* Search reached end of the region */
435 if (offset >= (search_start + search_len))
438 if (offset == start) {
439 /* we found a zero */
441 /* move start to the next bit to test */
444 /* got a zero after some ones */
448 if (found > best_len) {
450 best_start = start - found;
460 if (best_len >= wanted)
464 *rstart = best_start;
466 mlog(0, "Found start: %u len: %u\n", best_start, best_len);
471 static void __ocfs2_resv_find_window(struct ocfs2_reservation_map *resmap,
472 struct ocfs2_alloc_reservation *resv,
473 unsigned int goal, unsigned int wanted)
475 struct rb_root *root = &resmap->m_reservations;
476 unsigned int gap_start, gap_end, gap_len;
477 struct ocfs2_alloc_reservation *prev_resv, *next_resv;
478 struct rb_node *prev, *next;
479 unsigned int cstart, clen;
480 unsigned int best_start = 0, best_len = 0;
483 * Nasty cases to consider:
486 * - our window should be first in all reservations
487 * - our window should be last in all reservations
488 * - need to make sure we don't go past end of bitmap
491 mlog(0, "resv start: %u resv end: %u goal: %u wanted: %u\n",
492 resv->r_start, ocfs2_resv_end(resv), goal, wanted);
494 assert_spin_locked(&resv_lock);
496 if (RB_EMPTY_ROOT(root)) {
498 * Easiest case - empty tree. We can just take
499 * whatever window of free bits we want.
502 mlog(0, "Empty root\n");
504 clen = ocfs2_resmap_find_free_bits(resmap, wanted, goal,
505 resmap->m_bitmap_len - goal,
509 * This should never happen - the local alloc window
510 * will always have free bits when we're called.
512 BUG_ON(goal == 0 && clen == 0);
517 resv->r_start = cstart;
520 ocfs2_resv_insert(resmap, resv);
524 prev_resv = ocfs2_find_resv_lhs(resmap, goal);
526 if (prev_resv == NULL) {
527 mlog(0, "Goal on LHS of leftmost window\n");
530 * A NULL here means that the search code couldn't
531 * find a window that starts before goal.
533 * However, we can take the first window after goal,
534 * which is also by definition, the leftmost window in
535 * the entire tree. If we can find free bits in the
536 * gap between goal and the LHS window, then the
537 * reservation can safely be placed there.
539 * Otherwise we fall back to a linear search, checking
540 * the gaps in between windows for a place to
544 next = rb_first(root);
545 next_resv = rb_entry(next, struct ocfs2_alloc_reservation,
549 * The search should never return such a window. (see
552 if (next_resv->r_start <= goal) {
553 mlog(ML_ERROR, "goal: %u next_resv: start %u len %u\n",
554 goal, next_resv->r_start, next_resv->r_len);
555 ocfs2_dump_resv(resmap);
559 clen = ocfs2_resmap_find_free_bits(resmap, wanted, goal,
560 next_resv->r_start - goal,
565 if (best_len == wanted)
569 prev_resv = next_resv;
573 prev = &prev_resv->r_node;
575 /* Now we do a linear search for a window, starting at 'prev_rsv' */
577 next = rb_next(prev);
579 mlog(0, "One more resv found in linear search\n");
580 next_resv = rb_entry(next,
581 struct ocfs2_alloc_reservation,
584 gap_start = ocfs2_resv_end(prev_resv) + 1;
585 gap_end = next_resv->r_start - 1;
586 gap_len = gap_end - gap_start + 1;
588 mlog(0, "No next node\n");
590 * We're at the rightmost edge of the
591 * tree. See if a reservation between this
592 * window and the end of the bitmap will work.
594 gap_start = ocfs2_resv_end(prev_resv) + 1;
595 gap_len = resmap->m_bitmap_len - gap_start;
596 gap_end = resmap->m_bitmap_len - 1;
600 * No need to check this gap if we have already found
601 * a larger region of free bits.
603 if (gap_len <= best_len)
606 clen = ocfs2_resmap_find_free_bits(resmap, wanted, gap_start,
607 gap_len, &cstart, &clen);
608 if (clen == wanted) {
612 } else if (clen > best_len) {
622 prev_resv = rb_entry(prev, struct ocfs2_alloc_reservation,
628 resv->r_start = best_start;
629 resv->r_len = best_len;
630 ocfs2_resv_insert(resmap, resv);
634 static void ocfs2_cannibalize_resv(struct ocfs2_reservation_map *resmap,
635 struct ocfs2_alloc_reservation *resv,
638 struct ocfs2_alloc_reservation *lru_resv;
639 int tmpwindow = !!(resv->r_flags & OCFS2_RESV_FLAG_TMP);
640 unsigned int min_bits;
643 min_bits = ocfs2_resv_window_bits(resmap, resv) >> 1;
645 min_bits = wanted; /* We at know the temp window will use all
649 * Take the first reservation off the LRU as our 'target'. We
650 * don't try to be smart about it. There might be a case for
651 * searching based on size but I don't have enough data to be
652 * sure. --Mark (3/16/2010)
654 lru_resv = list_first_entry(&resmap->m_lru,
655 struct ocfs2_alloc_reservation, r_lru);
657 mlog(0, "lru resv: start: %u len: %u end: %u\n", lru_resv->r_start,
658 lru_resv->r_len, ocfs2_resv_end(lru_resv));
661 * Cannibalize (some or all) of the target reservation and
662 * feed it to the current window.
664 if (lru_resv->r_len <= min_bits) {
666 * Discard completely if size is less than or equal to a
667 * reasonable threshold - 50% of window bits for non temporary
670 resv->r_start = lru_resv->r_start;
671 resv->r_len = lru_resv->r_len;
673 __ocfs2_resv_discard(resmap, lru_resv);
679 shrink = lru_resv->r_len / 2;
681 lru_resv->r_len -= shrink;
683 resv->r_start = ocfs2_resv_end(lru_resv) + 1;
684 resv->r_len = shrink;
687 mlog(0, "Reservation now looks like: r_start: %u r_end: %u "
688 "r_len: %u r_last_start: %u r_last_len: %u\n",
689 resv->r_start, ocfs2_resv_end(resv), resv->r_len,
690 resv->r_last_start, resv->r_last_len);
692 ocfs2_resv_insert(resmap, resv);
695 static void ocfs2_resv_find_window(struct ocfs2_reservation_map *resmap,
696 struct ocfs2_alloc_reservation *resv,
699 unsigned int goal = 0;
701 BUG_ON(!ocfs2_resv_empty(resv));
704 * Begin by trying to get a window as close to the previous
705 * one as possible. Using the most recent allocation as a
706 * start goal makes sense.
708 if (resv->r_last_len) {
709 goal = resv->r_last_start + resv->r_last_len;
710 if (goal >= resmap->m_bitmap_len)
714 __ocfs2_resv_find_window(resmap, resv, goal, wanted);
716 /* Search from last alloc didn't work, try once more from beginning. */
717 if (ocfs2_resv_empty(resv) && goal != 0)
718 __ocfs2_resv_find_window(resmap, resv, 0, wanted);
720 if (ocfs2_resv_empty(resv)) {
722 * Still empty? Pull oldest one off the LRU, remove it from
723 * tree, put this one in it's place.
725 ocfs2_cannibalize_resv(resmap, resv, wanted);
728 BUG_ON(ocfs2_resv_empty(resv));
731 int ocfs2_resmap_resv_bits(struct ocfs2_reservation_map *resmap,
732 struct ocfs2_alloc_reservation *resv,
733 int *cstart, int *clen)
735 unsigned int wanted = *clen;
737 if (resv == NULL || ocfs2_resmap_disabled(resmap))
740 spin_lock(&resv_lock);
743 * We don't want to over-allocate for temporary
744 * windows. Otherwise, we run the risk of fragmenting the
747 wanted = ocfs2_resv_window_bits(resmap, resv);
748 if ((resv->r_flags & OCFS2_RESV_FLAG_TMP) || wanted < *clen)
751 if (ocfs2_resv_empty(resv)) {
752 mlog(0, "empty reservation, find new window\n");
755 * Try to get a window here. If it works, we must fall
756 * through and test the bitmap . This avoids some
757 * ping-ponging of windows due to non-reserved space
758 * being allocation before we initialize a window for
761 ocfs2_resv_find_window(resmap, resv, wanted);
764 BUG_ON(ocfs2_resv_empty(resv));
766 *cstart = resv->r_start;
769 spin_unlock(&resv_lock);
774 ocfs2_adjust_resv_from_alloc(struct ocfs2_reservation_map *resmap,
775 struct ocfs2_alloc_reservation *resv,
776 unsigned int start, unsigned int end)
778 unsigned int lhs = 0, rhs = 0;
780 BUG_ON(start < resv->r_start);
783 * Completely used? We can remove it then.
785 if (ocfs2_resv_end(resv) <= end && resv->r_start >= start) {
786 __ocfs2_resv_discard(resmap, resv);
790 if (end < ocfs2_resv_end(resv))
791 rhs = end - ocfs2_resv_end(resv);
793 if (start > resv->r_start)
794 lhs = start - resv->r_start;
797 * This should have been trapped above. At the very least, rhs
798 * should be non zero.
800 BUG_ON(rhs == 0 && lhs == 0);
803 unsigned int old_end = ocfs2_resv_end(resv);
805 resv->r_start = end + 1;
806 resv->r_len = old_end - resv->r_start + 1;
808 resv->r_len = start - resv->r_start;
812 void ocfs2_resmap_claimed_bits(struct ocfs2_reservation_map *resmap,
813 struct ocfs2_alloc_reservation *resv,
814 u32 cstart, u32 clen)
816 unsigned int cend = cstart + clen - 1;
818 if (resmap == NULL || ocfs2_resmap_disabled(resmap))
824 spin_lock(&resv_lock);
826 mlog(0, "claim bits: cstart: %u cend: %u clen: %u r_start: %u "
827 "r_end: %u r_len: %u, r_last_start: %u r_last_len: %u\n",
828 cstart, cend, clen, resv->r_start, ocfs2_resv_end(resv),
829 resv->r_len, resv->r_last_start, resv->r_last_len);
831 BUG_ON(cstart < resv->r_start);
832 BUG_ON(cstart > ocfs2_resv_end(resv));
833 BUG_ON(cend > ocfs2_resv_end(resv));
835 ocfs2_adjust_resv_from_alloc(resmap, resv, cstart, cend);
836 resv->r_last_start = cstart;
837 resv->r_last_len = clen;
840 * May have been discarded above from
841 * ocfs2_adjust_resv_from_alloc().
843 if (!ocfs2_resv_empty(resv))
844 ocfs2_resv_mark_lru(resmap, resv);
846 mlog(0, "Reservation now looks like: r_start: %u r_end: %u "
847 "r_len: %u r_last_start: %u r_last_len: %u\n",
848 resv->r_start, ocfs2_resv_end(resv), resv->r_len,
849 resv->r_last_start, resv->r_last_len);
851 ocfs2_check_resmap(resmap);
853 spin_unlock(&resv_lock);