2 * Copyright (c) 2000-2002,2005 Silicon Graphics, Inc.
3 * Copyright (c) 2010 David Chinner.
4 * Copyright (c) 2011 Christoph Hellwig.
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License as
9 * published by the Free Software Foundation.
11 * This program is distributed in the hope that it would be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
22 #include "xfs_types.h"
24 #include "xfs_trans.h"
27 #include "xfs_mount.h"
28 #include "xfs_bmap_btree.h"
29 #include "xfs_alloc.h"
30 #include "xfs_inode.h"
31 #include "xfs_extent_busy.h"
32 #include "xfs_trace.h"
35 xfs_extent_busy_insert(
42 struct xfs_extent_busy
*new;
43 struct xfs_extent_busy
*busyp
;
44 struct xfs_perag
*pag
;
46 struct rb_node
*parent
= NULL
;
48 new = kmem_zalloc(sizeof(struct xfs_extent_busy
), KM_MAYFAIL
);
51 * No Memory! Since it is now not possible to track the free
52 * block, make this a synchronous transaction to insure that
53 * the block is not reused before this transaction commits.
55 trace_xfs_extent_busy_enomem(tp
->t_mountp
, agno
, bno
, len
);
56 xfs_trans_set_sync(tp
);
63 INIT_LIST_HEAD(&new->list
);
66 /* trace before insert to be able to see failed inserts */
67 trace_xfs_extent_busy(tp
->t_mountp
, agno
, bno
, len
);
69 pag
= xfs_perag_get(tp
->t_mountp
, new->agno
);
70 spin_lock(&pag
->pagb_lock
);
71 rbp
= &pag
->pagb_tree
.rb_node
;
74 busyp
= rb_entry(parent
, struct xfs_extent_busy
, rb_node
);
76 if (new->bno
< busyp
->bno
) {
77 rbp
= &(*rbp
)->rb_left
;
78 ASSERT(new->bno
+ new->length
<= busyp
->bno
);
79 } else if (new->bno
> busyp
->bno
) {
80 rbp
= &(*rbp
)->rb_right
;
81 ASSERT(bno
>= busyp
->bno
+ busyp
->length
);
87 rb_link_node(&new->rb_node
, parent
, rbp
);
88 rb_insert_color(&new->rb_node
, &pag
->pagb_tree
);
90 list_add(&new->list
, &tp
->t_busy
);
91 spin_unlock(&pag
->pagb_lock
);
96 * Search for a busy extent within the range of the extent we are about to
97 * allocate. You need to be holding the busy extent tree lock when calling
98 * xfs_extent_busy_search(). This function returns 0 for no overlapping busy
99 * extent, -1 for an overlapping but not exact busy extent, and 1 for an exact
100 * match. This is done so that a non-zero return indicates an overlap that
101 * will require a synchronous transaction, but it can still be
102 * used to distinguish between a partial or exact match.
105 xfs_extent_busy_search(
106 struct xfs_mount
*mp
,
111 struct xfs_perag
*pag
;
113 struct xfs_extent_busy
*busyp
;
116 pag
= xfs_perag_get(mp
, agno
);
117 spin_lock(&pag
->pagb_lock
);
119 rbp
= pag
->pagb_tree
.rb_node
;
121 /* find closest start bno overlap */
123 busyp
= rb_entry(rbp
, struct xfs_extent_busy
, rb_node
);
124 if (bno
< busyp
->bno
) {
125 /* may overlap, but exact start block is lower */
126 if (bno
+ len
> busyp
->bno
)
129 } else if (bno
> busyp
->bno
) {
130 /* may overlap, but exact start block is higher */
131 if (bno
< busyp
->bno
+ busyp
->length
)
135 /* bno matches busyp, length determines exact match */
136 match
= (busyp
->length
== len
) ? 1 : -1;
140 spin_unlock(&pag
->pagb_lock
);
146 * The found free extent [fbno, fend] overlaps part or all of the given busy
147 * extent. If the overlap covers the beginning, the end, or all of the busy
148 * extent, the overlapping portion can be made unbusy and used for the
149 * allocation. We can't split a busy extent because we can't modify a
150 * transaction/CIL context busy list, but we can update an entries block
153 * Returns true if the extent can safely be reused, or false if the search
154 * needs to be restarted.
157 xfs_extent_busy_update_extent(
158 struct xfs_mount
*mp
,
159 struct xfs_perag
*pag
,
160 struct xfs_extent_busy
*busyp
,
165 xfs_agblock_t fend
= fbno
+ flen
;
166 xfs_agblock_t bbno
= busyp
->bno
;
167 xfs_agblock_t bend
= bbno
+ busyp
->length
;
170 * This extent is currently being discarded. Give the thread
171 * performing the discard a chance to mark the extent unbusy
174 if (busyp
->flags
& XFS_EXTENT_BUSY_DISCARDED
) {
175 spin_unlock(&pag
->pagb_lock
);
177 spin_lock(&pag
->pagb_lock
);
182 * If there is a busy extent overlapping a user allocation, we have
183 * no choice but to force the log and retry the search.
185 * Fortunately this does not happen during normal operation, but
186 * only if the filesystem is very low on space and has to dip into
187 * the AGFL for normal allocations.
192 if (bbno
< fbno
&& bend
> fend
) {
196 * +BBBBBBBBBBBBBBBBB+
202 * We would have to split the busy extent to be able to track
203 * it correct, which we cannot do because we would have to
204 * modify the list of busy extents attached to the transaction
205 * or CIL context, which is immutable.
207 * Force out the log to clear the busy extent and retry the
211 } else if (bbno
>= fbno
&& bend
<= fend
) {
215 * +BBBBBBBBBBBBBBBBB+
216 * +-----------------+
221 * +BBBBBBBBBBBBBBBBB+
222 * +--------------------------+
227 * +BBBBBBBBBBBBBBBBB+
228 * +--------------------------+
233 * +BBBBBBBBBBBBBBBBB+
234 * +-----------------------------------+
240 * The busy extent is fully covered by the extent we are
241 * allocating, and can simply be removed from the rbtree.
242 * However we cannot remove it from the immutable list
243 * tracking busy extents in the transaction or CIL context,
244 * so set the length to zero to mark it invalid.
246 * We also need to restart the busy extent search from the
247 * tree root, because erasing the node can rearrange the
250 rb_erase(&busyp
->rb_node
, &pag
->pagb_tree
);
253 } else if (fend
< bend
) {
257 * +BBBBBBBBBBBBBBBBB+
263 * +BBBBBBBBBBBBBBBBB+
264 * +------------------+
269 } else if (bbno
< fbno
) {
273 * +BBBBBBBBBBBBBBBBB+
279 * +BBBBBBBBBBBBBBBBB+
280 * +----------------------+
283 busyp
->length
= fbno
- busyp
->bno
;
288 trace_xfs_extent_busy_reuse(mp
, pag
->pag_agno
, fbno
, flen
);
292 spin_unlock(&pag
->pagb_lock
);
293 xfs_log_force(mp
, XFS_LOG_SYNC
);
294 trace_xfs_extent_busy_force(mp
, pag
->pag_agno
, fbno
, flen
);
295 spin_lock(&pag
->pagb_lock
);
301 * For a given extent [fbno, flen], make sure we can reuse it safely.
304 xfs_extent_busy_reuse(
305 struct xfs_mount
*mp
,
311 struct xfs_perag
*pag
;
316 pag
= xfs_perag_get(mp
, agno
);
317 spin_lock(&pag
->pagb_lock
);
319 rbp
= pag
->pagb_tree
.rb_node
;
321 struct xfs_extent_busy
*busyp
=
322 rb_entry(rbp
, struct xfs_extent_busy
, rb_node
);
323 xfs_agblock_t bbno
= busyp
->bno
;
324 xfs_agblock_t bend
= bbno
+ busyp
->length
;
326 if (fbno
+ flen
<= bbno
) {
329 } else if (fbno
>= bend
) {
334 if (!xfs_extent_busy_update_extent(mp
, pag
, busyp
, fbno
, flen
,
338 spin_unlock(&pag
->pagb_lock
);
343 * For a given extent [fbno, flen], search the busy extent list to find a
344 * subset of the extent that is not busy. If *rlen is smaller than
345 * args->minlen no suitable extent could be found, and the higher level
346 * code needs to force out the log and retry the allocation.
349 xfs_extent_busy_trim(
350 struct xfs_alloc_arg
*args
,
362 spin_lock(&args
->pag
->pagb_lock
);
366 rbp
= args
->pag
->pagb_tree
.rb_node
;
367 while (rbp
&& flen
>= args
->minlen
) {
368 struct xfs_extent_busy
*busyp
=
369 rb_entry(rbp
, struct xfs_extent_busy
, rb_node
);
370 xfs_agblock_t fend
= fbno
+ flen
;
371 xfs_agblock_t bbno
= busyp
->bno
;
372 xfs_agblock_t bend
= bbno
+ busyp
->length
;
377 } else if (fbno
>= bend
) {
383 * If this is a metadata allocation, try to reuse the busy
384 * extent instead of trimming the allocation.
386 if (!args
->userdata
&&
387 !(busyp
->flags
& XFS_EXTENT_BUSY_DISCARDED
)) {
388 if (!xfs_extent_busy_update_extent(args
->mp
, args
->pag
,
401 * +BBBBBBBBBBBBBBBBB+
407 * +BBBBBBBBBBBBBBBBB+
413 * +BBBBBBBBBBBBBBBBB+
419 * +BBBBBBBBBBBBBBBBB+
420 * +-----------------+
423 * No unbusy region in extent, return failure.
431 * +BBBBBBBBBBBBBBBBB+
432 * +----------------------+
437 * +BBBBBBBBBBBBBBBBB+
438 * +--------------------------+
441 * Needs to be trimmed to:
446 } else if (bend
>= fend
) {
452 * +BBBBBBBBBBBBBBBBB+
453 * +------------------+
458 * +BBBBBBBBBBBBBBBBB+
459 * +--------------------------+
462 * Needs to be trimmed to:
473 * +BBBBBBBBBBBBBBBBB+
474 * +-----------------------------------+
478 * +-------+ OR +-------+
479 * fbno fend fbno fend
481 * Backward allocation leads to significant
482 * fragmentation of directories, which degrades
483 * directory performance, therefore we always want to
484 * choose the option that produces forward allocation
486 * Preferring the lower bno extent will make the next
487 * request use "fend" as the start of the next
488 * allocation; if the segment is no longer busy at
489 * that point, we'll get a contiguous allocation, but
490 * even if it is still busy, we will get a forward
492 * We try to avoid choosing the segment at "bend",
493 * because that can lead to the next allocation
494 * taking the segment at "fbno", which would be a
495 * backward allocation. We only use the segment at
496 * "fbno" if it is much larger than the current
497 * requested size, because in that case there's a
498 * good chance subsequent allocations will be
501 if (bbno
- fbno
>= args
->maxlen
) {
502 /* left candidate fits perfect */
504 } else if (fend
- bend
>= args
->maxlen
* 4) {
505 /* right candidate has enough free space */
507 } else if (bbno
- fbno
>= args
->minlen
) {
508 /* left candidate fits minimum requirement */
517 spin_unlock(&args
->pag
->pagb_lock
);
519 if (fbno
!= bno
|| flen
!= len
) {
520 trace_xfs_extent_busy_trim(args
->mp
, args
->agno
, bno
, len
,
528 * Return a zero extent length as failure indications. All callers
529 * re-check if the trimmed extent satisfies the minlen requirement.
531 spin_unlock(&args
->pag
->pagb_lock
);
532 trace_xfs_extent_busy_trim(args
->mp
, args
->agno
, bno
, len
, fbno
, 0);
538 xfs_extent_busy_clear_one(
539 struct xfs_mount
*mp
,
540 struct xfs_perag
*pag
,
541 struct xfs_extent_busy
*busyp
)
544 trace_xfs_extent_busy_clear(mp
, busyp
->agno
, busyp
->bno
,
546 rb_erase(&busyp
->rb_node
, &pag
->pagb_tree
);
549 list_del_init(&busyp
->list
);
554 * Remove all extents on the passed in list from the busy extents tree.
555 * If do_discard is set skip extents that need to be discarded, and mark
556 * these as undergoing a discard operation instead.
559 xfs_extent_busy_clear(
560 struct xfs_mount
*mp
,
561 struct list_head
*list
,
564 struct xfs_extent_busy
*busyp
, *n
;
565 struct xfs_perag
*pag
= NULL
;
566 xfs_agnumber_t agno
= NULLAGNUMBER
;
568 list_for_each_entry_safe(busyp
, n
, list
, list
) {
569 if (busyp
->agno
!= agno
) {
571 spin_unlock(&pag
->pagb_lock
);
574 pag
= xfs_perag_get(mp
, busyp
->agno
);
575 spin_lock(&pag
->pagb_lock
);
579 if (do_discard
&& busyp
->length
&&
580 !(busyp
->flags
& XFS_EXTENT_BUSY_SKIP_DISCARD
))
581 busyp
->flags
= XFS_EXTENT_BUSY_DISCARDED
;
583 xfs_extent_busy_clear_one(mp
, pag
, busyp
);
587 spin_unlock(&pag
->pagb_lock
);
593 * Callback for list_sort to sort busy extents by the AG they reside in.
596 xfs_extent_busy_ag_cmp(
601 return container_of(a
, struct xfs_extent_busy
, list
)->agno
-
602 container_of(b
, struct xfs_extent_busy
, list
)->agno
;