1 // SPDX-License-Identifier: GPL-2.0-only
2 /* -*- mode: c; c-basic-offset: 8; -*-
3 * vim: noexpandtab sw=8 ts=8 sts=0:
7 * Allocation reservations implementation
9 * Some code borrowed from fs/ext3/balloc.c and is:
11 * Copyright (C) 1992, 1993, 1994, 1995
12 * Remy Card (card@masi.ibp.fr)
13 * Laboratoire MASI - Institut Blaise Pascal
14 * Universite Pierre et Marie Curie (Paris VI)
16 * The rest is copyright (C) 2010 Novell. All rights reserved.
20 #include <linux/types.h>
21 #include <linux/highmem.h>
22 #include <linux/bitops.h>
23 #include <linux/list.h>
25 #include <cluster/masklog.h>
28 #include "ocfs2_trace.h"
30 #ifdef CONFIG_OCFS2_DEBUG_FS
31 #define OCFS2_CHECK_RESERVATIONS
34 static DEFINE_SPINLOCK(resv_lock
);
36 int ocfs2_dir_resv_allowed(struct ocfs2_super
*osb
)
38 return (osb
->osb_resv_level
&& osb
->osb_dir_resv_level
);
41 static unsigned int ocfs2_resv_window_bits(struct ocfs2_reservation_map
*resmap
,
42 struct ocfs2_alloc_reservation
*resv
)
44 struct ocfs2_super
*osb
= resmap
->m_osb
;
47 if (!(resv
->r_flags
& OCFS2_RESV_FLAG_DIR
)) {
48 /* 8, 16, 32, 64, 128, 256, 512, 1024 */
49 bits
= 4 << osb
->osb_resv_level
;
51 bits
= 4 << osb
->osb_dir_resv_level
;
56 static inline unsigned int ocfs2_resv_end(struct ocfs2_alloc_reservation
*resv
)
59 return resv
->r_start
+ resv
->r_len
- 1;
63 static inline int ocfs2_resv_empty(struct ocfs2_alloc_reservation
*resv
)
65 return !!(resv
->r_len
== 0);
68 static inline int ocfs2_resmap_disabled(struct ocfs2_reservation_map
*resmap
)
70 if (resmap
->m_osb
->osb_resv_level
== 0)
75 static void ocfs2_dump_resv(struct ocfs2_reservation_map
*resmap
)
77 struct ocfs2_super
*osb
= resmap
->m_osb
;
79 struct ocfs2_alloc_reservation
*resv
;
82 mlog(ML_NOTICE
, "Dumping resmap for device %s. Bitmap length: %u\n",
83 osb
->dev_str
, resmap
->m_bitmap_len
);
85 node
= rb_first(&resmap
->m_reservations
);
87 resv
= rb_entry(node
, struct ocfs2_alloc_reservation
, r_node
);
89 mlog(ML_NOTICE
, "start: %u\tend: %u\tlen: %u\tlast_start: %u"
90 "\tlast_len: %u\n", resv
->r_start
,
91 ocfs2_resv_end(resv
), resv
->r_len
, resv
->r_last_start
,
98 mlog(ML_NOTICE
, "%d reservations found. LRU follows\n", i
);
101 list_for_each_entry(resv
, &resmap
->m_lru
, r_lru
) {
102 mlog(ML_NOTICE
, "LRU(%d) start: %u\tend: %u\tlen: %u\t"
103 "last_start: %u\tlast_len: %u\n", i
, resv
->r_start
,
104 ocfs2_resv_end(resv
), resv
->r_len
, resv
->r_last_start
,
111 #ifdef OCFS2_CHECK_RESERVATIONS
112 static int ocfs2_validate_resmap_bits(struct ocfs2_reservation_map
*resmap
,
114 struct ocfs2_alloc_reservation
*resv
)
116 char *disk_bitmap
= resmap
->m_disk_bitmap
;
117 unsigned int start
= resv
->r_start
;
118 unsigned int end
= ocfs2_resv_end(resv
);
120 while (start
<= end
) {
121 if (ocfs2_test_bit(start
, disk_bitmap
)) {
123 "reservation %d covers an allocated area "
124 "starting at bit %u!\n", i
, start
);
133 static void ocfs2_check_resmap(struct ocfs2_reservation_map
*resmap
)
135 unsigned int off
= 0;
137 struct rb_node
*node
;
138 struct ocfs2_alloc_reservation
*resv
;
140 node
= rb_first(&resmap
->m_reservations
);
142 resv
= rb_entry(node
, struct ocfs2_alloc_reservation
, r_node
);
144 if (i
> 0 && resv
->r_start
<= off
) {
145 mlog(ML_ERROR
, "reservation %d has bad start off!\n",
150 if (resv
->r_len
== 0) {
151 mlog(ML_ERROR
, "reservation %d has no length!\n",
156 if (resv
->r_start
> ocfs2_resv_end(resv
)) {
157 mlog(ML_ERROR
, "reservation %d has invalid range!\n",
162 if (ocfs2_resv_end(resv
) >= resmap
->m_bitmap_len
) {
163 mlog(ML_ERROR
, "reservation %d extends past bitmap!\n",
168 if (ocfs2_validate_resmap_bits(resmap
, i
, resv
))
171 off
= ocfs2_resv_end(resv
);
172 node
= rb_next(node
);
179 ocfs2_dump_resv(resmap
);
183 static inline void ocfs2_check_resmap(struct ocfs2_reservation_map
*resmap
)
189 void ocfs2_resv_init_once(struct ocfs2_alloc_reservation
*resv
)
191 memset(resv
, 0, sizeof(*resv
));
192 INIT_LIST_HEAD(&resv
->r_lru
);
195 void ocfs2_resv_set_type(struct ocfs2_alloc_reservation
*resv
,
198 BUG_ON(flags
& ~OCFS2_RESV_TYPES
);
200 resv
->r_flags
|= flags
;
203 int ocfs2_resmap_init(struct ocfs2_super
*osb
,
204 struct ocfs2_reservation_map
*resmap
)
206 memset(resmap
, 0, sizeof(*resmap
));
209 resmap
->m_reservations
= RB_ROOT
;
210 /* m_bitmap_len is initialized to zero by the above memset. */
211 INIT_LIST_HEAD(&resmap
->m_lru
);
216 static void ocfs2_resv_mark_lru(struct ocfs2_reservation_map
*resmap
,
217 struct ocfs2_alloc_reservation
*resv
)
219 assert_spin_locked(&resv_lock
);
221 if (!list_empty(&resv
->r_lru
))
222 list_del_init(&resv
->r_lru
);
224 list_add_tail(&resv
->r_lru
, &resmap
->m_lru
);
227 static void __ocfs2_resv_trunc(struct ocfs2_alloc_reservation
*resv
)
233 static void ocfs2_resv_remove(struct ocfs2_reservation_map
*resmap
,
234 struct ocfs2_alloc_reservation
*resv
)
236 if (resv
->r_flags
& OCFS2_RESV_FLAG_INUSE
) {
237 list_del_init(&resv
->r_lru
);
238 rb_erase(&resv
->r_node
, &resmap
->m_reservations
);
239 resv
->r_flags
&= ~OCFS2_RESV_FLAG_INUSE
;
243 static void __ocfs2_resv_discard(struct ocfs2_reservation_map
*resmap
,
244 struct ocfs2_alloc_reservation
*resv
)
246 assert_spin_locked(&resv_lock
);
248 __ocfs2_resv_trunc(resv
);
250 * last_len and last_start no longer make sense if
251 * we're changing the range of our allocations.
253 resv
->r_last_len
= resv
->r_last_start
= 0;
255 ocfs2_resv_remove(resmap
, resv
);
258 /* does nothing if 'resv' is null */
259 void ocfs2_resv_discard(struct ocfs2_reservation_map
*resmap
,
260 struct ocfs2_alloc_reservation
*resv
)
263 spin_lock(&resv_lock
);
264 __ocfs2_resv_discard(resmap
, resv
);
265 spin_unlock(&resv_lock
);
269 static void ocfs2_resmap_clear_all_resv(struct ocfs2_reservation_map
*resmap
)
271 struct rb_node
*node
;
272 struct ocfs2_alloc_reservation
*resv
;
274 assert_spin_locked(&resv_lock
);
276 while ((node
= rb_last(&resmap
->m_reservations
)) != NULL
) {
277 resv
= rb_entry(node
, struct ocfs2_alloc_reservation
, r_node
);
279 __ocfs2_resv_discard(resmap
, resv
);
283 void ocfs2_resmap_restart(struct ocfs2_reservation_map
*resmap
,
284 unsigned int clen
, char *disk_bitmap
)
286 if (ocfs2_resmap_disabled(resmap
))
289 spin_lock(&resv_lock
);
291 ocfs2_resmap_clear_all_resv(resmap
);
292 resmap
->m_bitmap_len
= clen
;
293 resmap
->m_disk_bitmap
= disk_bitmap
;
295 spin_unlock(&resv_lock
);
298 void ocfs2_resmap_uninit(struct ocfs2_reservation_map
*resmap
)
300 /* Does nothing for now. Keep this around for API symmetry */
303 static void ocfs2_resv_insert(struct ocfs2_reservation_map
*resmap
,
304 struct ocfs2_alloc_reservation
*new)
306 struct rb_root
*root
= &resmap
->m_reservations
;
307 struct rb_node
*parent
= NULL
;
308 struct rb_node
**p
= &root
->rb_node
;
309 struct ocfs2_alloc_reservation
*tmp
;
311 assert_spin_locked(&resv_lock
);
313 trace_ocfs2_resv_insert(new->r_start
, new->r_len
);
318 tmp
= rb_entry(parent
, struct ocfs2_alloc_reservation
, r_node
);
320 if (new->r_start
< tmp
->r_start
) {
324 * This is a good place to check for
325 * overlapping reservations.
327 BUG_ON(ocfs2_resv_end(new) >= tmp
->r_start
);
328 } else if (new->r_start
> ocfs2_resv_end(tmp
)) {
331 /* This should never happen! */
332 mlog(ML_ERROR
, "Duplicate reservation window!\n");
337 rb_link_node(&new->r_node
, parent
, p
);
338 rb_insert_color(&new->r_node
, root
);
339 new->r_flags
|= OCFS2_RESV_FLAG_INUSE
;
341 ocfs2_resv_mark_lru(resmap
, new);
343 ocfs2_check_resmap(resmap
);
347 * ocfs2_find_resv_lhs() - find the window which contains goal
348 * @resmap: reservation map to search
349 * @goal: which bit to search for
351 * If a window containing that goal is not found, we return the window
352 * which comes before goal. Returns NULL on empty rbtree or no window
355 static struct ocfs2_alloc_reservation
*
356 ocfs2_find_resv_lhs(struct ocfs2_reservation_map
*resmap
, unsigned int goal
)
358 struct ocfs2_alloc_reservation
*resv
= NULL
;
359 struct ocfs2_alloc_reservation
*prev_resv
= NULL
;
360 struct rb_node
*node
= resmap
->m_reservations
.rb_node
;
362 assert_spin_locked(&resv_lock
);
367 node
= rb_first(&resmap
->m_reservations
);
369 resv
= rb_entry(node
, struct ocfs2_alloc_reservation
, r_node
);
371 if (resv
->r_start
<= goal
&& ocfs2_resv_end(resv
) >= goal
)
374 /* Check if we overshot the reservation just before goal? */
375 if (resv
->r_start
> goal
) {
381 node
= rb_next(node
);
388 * We are given a range within the bitmap, which corresponds to a gap
389 * inside the reservations tree (search_start, search_len). The range
390 * can be anything from the whole bitmap, to a gap between
393 * The start value of *rstart is insignificant.
395 * This function searches the bitmap range starting at search_start
396 * with length search_len for a set of contiguous free bits. We try
397 * to find up to 'wanted' bits, but can sometimes return less.
399 * Returns the length of allocation, 0 if no free bits are found.
401 * *cstart and *clen will also be populated with the result.
403 static int ocfs2_resmap_find_free_bits(struct ocfs2_reservation_map
*resmap
,
405 unsigned int search_start
,
406 unsigned int search_len
,
407 unsigned int *rstart
,
410 void *bitmap
= resmap
->m_disk_bitmap
;
411 unsigned int best_start
, best_len
= 0;
412 int offset
, start
, found
;
414 trace_ocfs2_resmap_find_free_bits_begin(search_start
, search_len
,
415 wanted
, resmap
->m_bitmap_len
);
417 found
= best_start
= best_len
= 0;
419 start
= search_start
;
420 while ((offset
= ocfs2_find_next_zero_bit(bitmap
, resmap
->m_bitmap_len
,
422 /* Search reached end of the region */
423 if (offset
>= (search_start
+ search_len
))
426 if (offset
== start
) {
427 /* we found a zero */
429 /* move start to the next bit to test */
432 /* got a zero after some ones */
436 if (found
> best_len
) {
438 best_start
= start
- found
;
448 if (best_len
>= wanted
)
452 *rstart
= best_start
;
454 trace_ocfs2_resmap_find_free_bits_end(best_start
, best_len
);
459 static void __ocfs2_resv_find_window(struct ocfs2_reservation_map
*resmap
,
460 struct ocfs2_alloc_reservation
*resv
,
461 unsigned int goal
, unsigned int wanted
)
463 struct rb_root
*root
= &resmap
->m_reservations
;
464 unsigned int gap_start
, gap_end
, gap_len
;
465 struct ocfs2_alloc_reservation
*prev_resv
, *next_resv
;
466 struct rb_node
*prev
, *next
;
467 unsigned int cstart
, clen
;
468 unsigned int best_start
= 0, best_len
= 0;
471 * Nasty cases to consider:
474 * - our window should be first in all reservations
475 * - our window should be last in all reservations
476 * - need to make sure we don't go past end of bitmap
478 trace_ocfs2_resv_find_window_begin(resv
->r_start
, ocfs2_resv_end(resv
),
479 goal
, wanted
, RB_EMPTY_ROOT(root
));
481 assert_spin_locked(&resv_lock
);
483 if (RB_EMPTY_ROOT(root
)) {
485 * Easiest case - empty tree. We can just take
486 * whatever window of free bits we want.
488 clen
= ocfs2_resmap_find_free_bits(resmap
, wanted
, goal
,
489 resmap
->m_bitmap_len
- goal
,
493 * This should never happen - the local alloc window
494 * will always have free bits when we're called.
496 BUG_ON(goal
== 0 && clen
== 0);
501 resv
->r_start
= cstart
;
504 ocfs2_resv_insert(resmap
, resv
);
508 prev_resv
= ocfs2_find_resv_lhs(resmap
, goal
);
510 if (prev_resv
== NULL
) {
512 * A NULL here means that the search code couldn't
513 * find a window that starts before goal.
515 * However, we can take the first window after goal,
516 * which is also by definition, the leftmost window in
517 * the entire tree. If we can find free bits in the
518 * gap between goal and the LHS window, then the
519 * reservation can safely be placed there.
521 * Otherwise we fall back to a linear search, checking
522 * the gaps in between windows for a place to
526 next
= rb_first(root
);
527 next_resv
= rb_entry(next
, struct ocfs2_alloc_reservation
,
531 * The search should never return such a window. (see
534 if (next_resv
->r_start
<= goal
) {
535 mlog(ML_ERROR
, "goal: %u next_resv: start %u len %u\n",
536 goal
, next_resv
->r_start
, next_resv
->r_len
);
537 ocfs2_dump_resv(resmap
);
541 clen
= ocfs2_resmap_find_free_bits(resmap
, wanted
, goal
,
542 next_resv
->r_start
- goal
,
547 if (best_len
== wanted
)
551 prev_resv
= next_resv
;
555 trace_ocfs2_resv_find_window_prev(prev_resv
->r_start
,
556 ocfs2_resv_end(prev_resv
));
558 prev
= &prev_resv
->r_node
;
560 /* Now we do a linear search for a window, starting at 'prev_rsv' */
562 next
= rb_next(prev
);
564 next_resv
= rb_entry(next
,
565 struct ocfs2_alloc_reservation
,
568 gap_start
= ocfs2_resv_end(prev_resv
) + 1;
569 gap_end
= next_resv
->r_start
- 1;
570 gap_len
= gap_end
- gap_start
+ 1;
573 * We're at the rightmost edge of the
574 * tree. See if a reservation between this
575 * window and the end of the bitmap will work.
577 gap_start
= ocfs2_resv_end(prev_resv
) + 1;
578 gap_len
= resmap
->m_bitmap_len
- gap_start
;
579 gap_end
= resmap
->m_bitmap_len
- 1;
582 trace_ocfs2_resv_find_window_next(next
? next_resv
->r_start
: -1,
583 next
? ocfs2_resv_end(next_resv
) : -1);
585 * No need to check this gap if we have already found
586 * a larger region of free bits.
588 if (gap_len
<= best_len
)
591 clen
= ocfs2_resmap_find_free_bits(resmap
, wanted
, gap_start
,
592 gap_len
, &cstart
, &clen
);
593 if (clen
== wanted
) {
597 } else if (clen
> best_len
) {
607 prev_resv
= rb_entry(prev
, struct ocfs2_alloc_reservation
,
613 resv
->r_start
= best_start
;
614 resv
->r_len
= best_len
;
615 ocfs2_resv_insert(resmap
, resv
);
619 static void ocfs2_cannibalize_resv(struct ocfs2_reservation_map
*resmap
,
620 struct ocfs2_alloc_reservation
*resv
,
623 struct ocfs2_alloc_reservation
*lru_resv
;
624 int tmpwindow
= !!(resv
->r_flags
& OCFS2_RESV_FLAG_TMP
);
625 unsigned int min_bits
;
628 min_bits
= ocfs2_resv_window_bits(resmap
, resv
) >> 1;
630 min_bits
= wanted
; /* We at know the temp window will use all
634 * Take the first reservation off the LRU as our 'target'. We
635 * don't try to be smart about it. There might be a case for
636 * searching based on size but I don't have enough data to be
637 * sure. --Mark (3/16/2010)
639 lru_resv
= list_first_entry(&resmap
->m_lru
,
640 struct ocfs2_alloc_reservation
, r_lru
);
642 trace_ocfs2_cannibalize_resv_begin(lru_resv
->r_start
,
644 ocfs2_resv_end(lru_resv
));
647 * Cannibalize (some or all) of the target reservation and
648 * feed it to the current window.
650 if (lru_resv
->r_len
<= min_bits
) {
652 * Discard completely if size is less than or equal to a
653 * reasonable threshold - 50% of window bits for non temporary
656 resv
->r_start
= lru_resv
->r_start
;
657 resv
->r_len
= lru_resv
->r_len
;
659 __ocfs2_resv_discard(resmap
, lru_resv
);
665 shrink
= lru_resv
->r_len
/ 2;
667 lru_resv
->r_len
-= shrink
;
669 resv
->r_start
= ocfs2_resv_end(lru_resv
) + 1;
670 resv
->r_len
= shrink
;
673 trace_ocfs2_cannibalize_resv_end(resv
->r_start
, ocfs2_resv_end(resv
),
674 resv
->r_len
, resv
->r_last_start
,
677 ocfs2_resv_insert(resmap
, resv
);
680 static void ocfs2_resv_find_window(struct ocfs2_reservation_map
*resmap
,
681 struct ocfs2_alloc_reservation
*resv
,
684 unsigned int goal
= 0;
686 BUG_ON(!ocfs2_resv_empty(resv
));
689 * Begin by trying to get a window as close to the previous
690 * one as possible. Using the most recent allocation as a
691 * start goal makes sense.
693 if (resv
->r_last_len
) {
694 goal
= resv
->r_last_start
+ resv
->r_last_len
;
695 if (goal
>= resmap
->m_bitmap_len
)
699 __ocfs2_resv_find_window(resmap
, resv
, goal
, wanted
);
701 /* Search from last alloc didn't work, try once more from beginning. */
702 if (ocfs2_resv_empty(resv
) && goal
!= 0)
703 __ocfs2_resv_find_window(resmap
, resv
, 0, wanted
);
705 if (ocfs2_resv_empty(resv
)) {
707 * Still empty? Pull oldest one off the LRU, remove it from
708 * tree, put this one in it's place.
710 ocfs2_cannibalize_resv(resmap
, resv
, wanted
);
713 BUG_ON(ocfs2_resv_empty(resv
));
716 int ocfs2_resmap_resv_bits(struct ocfs2_reservation_map
*resmap
,
717 struct ocfs2_alloc_reservation
*resv
,
718 int *cstart
, int *clen
)
720 if (resv
== NULL
|| ocfs2_resmap_disabled(resmap
))
723 spin_lock(&resv_lock
);
725 if (ocfs2_resv_empty(resv
)) {
727 * We don't want to over-allocate for temporary
728 * windows. Otherwise, we run the risk of fragmenting the
731 unsigned int wanted
= ocfs2_resv_window_bits(resmap
, resv
);
733 if ((resv
->r_flags
& OCFS2_RESV_FLAG_TMP
) || wanted
< *clen
)
737 * Try to get a window here. If it works, we must fall
738 * through and test the bitmap . This avoids some
739 * ping-ponging of windows due to non-reserved space
740 * being allocation before we initialize a window for
743 ocfs2_resv_find_window(resmap
, resv
, wanted
);
744 trace_ocfs2_resmap_resv_bits(resv
->r_start
, resv
->r_len
);
747 BUG_ON(ocfs2_resv_empty(resv
));
749 *cstart
= resv
->r_start
;
752 spin_unlock(&resv_lock
);
757 ocfs2_adjust_resv_from_alloc(struct ocfs2_reservation_map
*resmap
,
758 struct ocfs2_alloc_reservation
*resv
,
759 unsigned int start
, unsigned int end
)
761 unsigned int rhs
= 0;
762 unsigned int old_end
= ocfs2_resv_end(resv
);
764 BUG_ON(start
!= resv
->r_start
|| old_end
< end
);
767 * Completely used? We can remove it then.
769 if (old_end
== end
) {
770 __ocfs2_resv_discard(resmap
, resv
);
777 * This should have been trapped above.
781 resv
->r_start
= end
+ 1;
782 resv
->r_len
= old_end
- resv
->r_start
+ 1;
785 void ocfs2_resmap_claimed_bits(struct ocfs2_reservation_map
*resmap
,
786 struct ocfs2_alloc_reservation
*resv
,
787 u32 cstart
, u32 clen
)
789 unsigned int cend
= cstart
+ clen
- 1;
791 if (resmap
== NULL
|| ocfs2_resmap_disabled(resmap
))
797 BUG_ON(cstart
!= resv
->r_start
);
799 spin_lock(&resv_lock
);
801 trace_ocfs2_resmap_claimed_bits_begin(cstart
, cend
, clen
, resv
->r_start
,
802 ocfs2_resv_end(resv
), resv
->r_len
,
806 BUG_ON(cstart
< resv
->r_start
);
807 BUG_ON(cstart
> ocfs2_resv_end(resv
));
808 BUG_ON(cend
> ocfs2_resv_end(resv
));
810 ocfs2_adjust_resv_from_alloc(resmap
, resv
, cstart
, cend
);
811 resv
->r_last_start
= cstart
;
812 resv
->r_last_len
= clen
;
815 * May have been discarded above from
816 * ocfs2_adjust_resv_from_alloc().
818 if (!ocfs2_resv_empty(resv
))
819 ocfs2_resv_mark_lru(resmap
, resv
);
821 trace_ocfs2_resmap_claimed_bits_end(resv
->r_start
, ocfs2_resv_end(resv
),
822 resv
->r_len
, resv
->r_last_start
,
825 ocfs2_check_resmap(resmap
);
827 spin_unlock(&resv_lock
);