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_format.h"
23 #include "xfs_log_format.h"
24 #include "xfs_shared.h"
25 #include "xfs_trans_resv.h"
27 #include "xfs_mount.h"
28 #include "xfs_alloc.h"
29 #include "xfs_extent_busy.h"
30 #include "xfs_trace.h"
31 #include "xfs_trans.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_SLEEP);
52 INIT_LIST_HEAD(&new->list);
55 /* trace before insert to be able to see failed inserts */
56 trace_xfs_extent_busy(tp->t_mountp, agno, bno, len);
58 pag = xfs_perag_get(tp->t_mountp, new->agno);
59 spin_lock(&pag->pagb_lock);
60 rbp = &pag->pagb_tree.rb_node;
63 busyp = rb_entry(parent, struct xfs_extent_busy, rb_node);
65 if (new->bno < busyp->bno) {
66 rbp = &(*rbp)->rb_left;
67 ASSERT(new->bno + new->length <= busyp->bno);
68 } else if (new->bno > busyp->bno) {
69 rbp = &(*rbp)->rb_right;
70 ASSERT(bno >= busyp->bno + busyp->length);
76 rb_link_node(&new->rb_node, parent, rbp);
77 rb_insert_color(&new->rb_node, &pag->pagb_tree);
79 list_add(&new->list, &tp->t_busy);
80 spin_unlock(&pag->pagb_lock);
85 * Search for a busy extent within the range of the extent we are about to
86 * allocate. You need to be holding the busy extent tree lock when calling
87 * xfs_extent_busy_search(). This function returns 0 for no overlapping busy
88 * extent, -1 for an overlapping but not exact busy extent, and 1 for an exact
89 * match. This is done so that a non-zero return indicates an overlap that
90 * will require a synchronous transaction, but it can still be
91 * used to distinguish between a partial or exact match.
94 xfs_extent_busy_search(
100 struct xfs_perag *pag;
102 struct xfs_extent_busy *busyp;
105 pag = xfs_perag_get(mp, agno);
106 spin_lock(&pag->pagb_lock);
108 rbp = pag->pagb_tree.rb_node;
110 /* find closest start bno overlap */
112 busyp = rb_entry(rbp, struct xfs_extent_busy, rb_node);
113 if (bno < busyp->bno) {
114 /* may overlap, but exact start block is lower */
115 if (bno + len > busyp->bno)
118 } else if (bno > busyp->bno) {
119 /* may overlap, but exact start block is higher */
120 if (bno < busyp->bno + busyp->length)
124 /* bno matches busyp, length determines exact match */
125 match = (busyp->length == len) ? 1 : -1;
129 spin_unlock(&pag->pagb_lock);
135 * The found free extent [fbno, fend] overlaps part or all of the given busy
136 * extent. If the overlap covers the beginning, the end, or all of the busy
137 * extent, the overlapping portion can be made unbusy and used for the
138 * allocation. We can't split a busy extent because we can't modify a
139 * transaction/CIL context busy list, but we can update an entry's block
142 * Returns true if the extent can safely be reused, or false if the search
143 * needs to be restarted.
146 xfs_extent_busy_update_extent(
147 struct xfs_mount *mp,
148 struct xfs_perag *pag,
149 struct xfs_extent_busy *busyp,
152 bool userdata) __releases(&pag->pagb_lock)
153 __acquires(&pag->pagb_lock)
155 xfs_agblock_t fend = fbno + flen;
156 xfs_agblock_t bbno = busyp->bno;
157 xfs_agblock_t bend = bbno + busyp->length;
160 * This extent is currently being discarded. Give the thread
161 * performing the discard a chance to mark the extent unbusy
164 if (busyp->flags & XFS_EXTENT_BUSY_DISCARDED) {
165 spin_unlock(&pag->pagb_lock);
167 spin_lock(&pag->pagb_lock);
172 * If there is a busy extent overlapping a user allocation, we have
173 * no choice but to force the log and retry the search.
175 * Fortunately this does not happen during normal operation, but
176 * only if the filesystem is very low on space and has to dip into
177 * the AGFL for normal allocations.
182 if (bbno < fbno && bend > fend) {
186 * +BBBBBBBBBBBBBBBBB+
192 * We would have to split the busy extent to be able to track
193 * it correct, which we cannot do because we would have to
194 * modify the list of busy extents attached to the transaction
195 * or CIL context, which is immutable.
197 * Force out the log to clear the busy extent and retry the
201 } else if (bbno >= fbno && bend <= fend) {
205 * +BBBBBBBBBBBBBBBBB+
206 * +-----------------+
211 * +BBBBBBBBBBBBBBBBB+
212 * +--------------------------+
217 * +BBBBBBBBBBBBBBBBB+
218 * +--------------------------+
223 * +BBBBBBBBBBBBBBBBB+
224 * +-----------------------------------+
230 * The busy extent is fully covered by the extent we are
231 * allocating, and can simply be removed from the rbtree.
232 * However we cannot remove it from the immutable list
233 * tracking busy extents in the transaction or CIL context,
234 * so set the length to zero to mark it invalid.
236 * We also need to restart the busy extent search from the
237 * tree root, because erasing the node can rearrange the
240 rb_erase(&busyp->rb_node, &pag->pagb_tree);
243 } else if (fend < bend) {
247 * +BBBBBBBBBBBBBBBBB+
253 * +BBBBBBBBBBBBBBBBB+
254 * +------------------+
259 } else if (bbno < fbno) {
263 * +BBBBBBBBBBBBBBBBB+
269 * +BBBBBBBBBBBBBBBBB+
270 * +----------------------+
273 busyp->length = fbno - busyp->bno;
278 trace_xfs_extent_busy_reuse(mp, pag->pag_agno, fbno, flen);
282 spin_unlock(&pag->pagb_lock);
283 xfs_log_force(mp, XFS_LOG_SYNC);
284 trace_xfs_extent_busy_force(mp, pag->pag_agno, fbno, flen);
285 spin_lock(&pag->pagb_lock);
291 * For a given extent [fbno, flen], make sure we can reuse it safely.
294 xfs_extent_busy_reuse(
295 struct xfs_mount *mp,
301 struct xfs_perag *pag;
306 pag = xfs_perag_get(mp, agno);
307 spin_lock(&pag->pagb_lock);
309 rbp = pag->pagb_tree.rb_node;
311 struct xfs_extent_busy *busyp =
312 rb_entry(rbp, struct xfs_extent_busy, rb_node);
313 xfs_agblock_t bbno = busyp->bno;
314 xfs_agblock_t bend = bbno + busyp->length;
316 if (fbno + flen <= bbno) {
319 } else if (fbno >= bend) {
324 if (!xfs_extent_busy_update_extent(mp, pag, busyp, fbno, flen,
328 spin_unlock(&pag->pagb_lock);
333 * For a given extent [fbno, flen], search the busy extent list to find a
334 * subset of the extent that is not busy. If *rlen is smaller than
335 * args->minlen no suitable extent could be found, and the higher level
336 * code needs to force out the log and retry the allocation.
338 * Return the current busy generation for the AG if the extent is busy. This
339 * value can be used to wait for at least one of the currently busy extents
340 * to be cleared. Note that the busy list is not guaranteed to be empty after
341 * the gen is woken. The state of a specific extent must always be confirmed
342 * with another call to xfs_extent_busy_trim() before it can be used.
345 xfs_extent_busy_trim(
346 struct xfs_alloc_arg *args,
358 spin_lock(&args->pag->pagb_lock);
362 rbp = args->pag->pagb_tree.rb_node;
363 while (rbp && flen >= args->minlen) {
364 struct xfs_extent_busy *busyp =
365 rb_entry(rbp, struct xfs_extent_busy, rb_node);
366 xfs_agblock_t fend = fbno + flen;
367 xfs_agblock_t bbno = busyp->bno;
368 xfs_agblock_t bend = bbno + busyp->length;
373 } else if (fbno >= bend) {
379 * If this is a metadata allocation, try to reuse the busy
380 * extent instead of trimming the allocation.
382 if (!xfs_alloc_is_userdata(args->datatype) &&
383 !(busyp->flags & XFS_EXTENT_BUSY_DISCARDED)) {
384 if (!xfs_extent_busy_update_extent(args->mp, args->pag,
397 * +BBBBBBBBBBBBBBBBB+
403 * +BBBBBBBBBBBBBBBBB+
409 * +BBBBBBBBBBBBBBBBB+
415 * +BBBBBBBBBBBBBBBBB+
416 * +-----------------+
419 * No unbusy region in extent, return failure.
427 * +BBBBBBBBBBBBBBBBB+
428 * +----------------------+
433 * +BBBBBBBBBBBBBBBBB+
434 * +--------------------------+
437 * Needs to be trimmed to:
442 } else if (bend >= fend) {
448 * +BBBBBBBBBBBBBBBBB+
449 * +------------------+
454 * +BBBBBBBBBBBBBBBBB+
455 * +--------------------------+
458 * Needs to be trimmed to:
469 * +BBBBBBBBBBBBBBBBB+
470 * +-----------------------------------+
474 * +-------+ OR +-------+
475 * fbno fend fbno fend
477 * Backward allocation leads to significant
478 * fragmentation of directories, which degrades
479 * directory performance, therefore we always want to
480 * choose the option that produces forward allocation
482 * Preferring the lower bno extent will make the next
483 * request use "fend" as the start of the next
484 * allocation; if the segment is no longer busy at
485 * that point, we'll get a contiguous allocation, but
486 * even if it is still busy, we will get a forward
488 * We try to avoid choosing the segment at "bend",
489 * because that can lead to the next allocation
490 * taking the segment at "fbno", which would be a
491 * backward allocation. We only use the segment at
492 * "fbno" if it is much larger than the current
493 * requested size, because in that case there's a
494 * good chance subsequent allocations will be
497 if (bbno - fbno >= args->maxlen) {
498 /* left candidate fits perfect */
500 } else if (fend - bend >= args->maxlen * 4) {
501 /* right candidate has enough free space */
503 } else if (bbno - fbno >= args->minlen) {
504 /* left candidate fits minimum requirement */
515 if (fbno != *bno || flen != *len) {
516 trace_xfs_extent_busy_trim(args->mp, args->agno, *bno, *len,
520 *busy_gen = args->pag->pagb_gen;
523 spin_unlock(&args->pag->pagb_lock);
527 * Return a zero extent length as failure indications. All callers
528 * re-check if the trimmed extent satisfies the minlen requirement.
535 xfs_extent_busy_clear_one(
536 struct xfs_mount *mp,
537 struct xfs_perag *pag,
538 struct xfs_extent_busy *busyp)
541 trace_xfs_extent_busy_clear(mp, busyp->agno, busyp->bno,
543 rb_erase(&busyp->rb_node, &pag->pagb_tree);
546 list_del_init(&busyp->list);
551 xfs_extent_busy_put_pag(
552 struct xfs_perag *pag,
554 __releases(pag->pagb_lock)
558 wake_up_all(&pag->pagb_wait);
561 spin_unlock(&pag->pagb_lock);
566 * Remove all extents on the passed in list from the busy extents tree.
567 * If do_discard is set skip extents that need to be discarded, and mark
568 * these as undergoing a discard operation instead.
571 xfs_extent_busy_clear(
572 struct xfs_mount *mp,
573 struct list_head *list,
576 struct xfs_extent_busy *busyp, *n;
577 struct xfs_perag *pag = NULL;
578 xfs_agnumber_t agno = NULLAGNUMBER;
581 list_for_each_entry_safe(busyp, n, list, list) {
582 if (busyp->agno != agno) {
584 xfs_extent_busy_put_pag(pag, wakeup);
586 pag = xfs_perag_get(mp, agno);
587 spin_lock(&pag->pagb_lock);
591 if (do_discard && busyp->length &&
592 !(busyp->flags & XFS_EXTENT_BUSY_SKIP_DISCARD)) {
593 busyp->flags = XFS_EXTENT_BUSY_DISCARDED;
595 xfs_extent_busy_clear_one(mp, pag, busyp);
601 xfs_extent_busy_put_pag(pag, wakeup);
605 * Flush out all busy extents for this AG.
608 xfs_extent_busy_flush(
609 struct xfs_mount *mp,
610 struct xfs_perag *pag,
614 int log_flushed = 0, error;
616 trace_xfs_log_force(mp, 0, _THIS_IP_);
617 error = _xfs_log_force(mp, XFS_LOG_SYNC, &log_flushed);
622 prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
623 if (busy_gen != READ_ONCE(pag->pagb_gen))
628 finish_wait(&pag->pagb_wait, &wait);
632 xfs_extent_busy_wait_all(
633 struct xfs_mount *mp)
638 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
639 struct xfs_perag *pag = xfs_perag_get(mp, agno);
642 prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
643 if (RB_EMPTY_ROOT(&pag->pagb_tree))
647 finish_wait(&pag->pagb_wait, &wait);
654 * Callback for list_sort to sort busy extents by the AG they reside in.
657 xfs_extent_busy_ag_cmp(
659 struct list_head *l1,
660 struct list_head *l2)
662 struct xfs_extent_busy *b1 =
663 container_of(l1, struct xfs_extent_busy, list);
664 struct xfs_extent_busy *b2 =
665 container_of(l2, struct xfs_extent_busy, list);
668 diff = b1->agno - b2->agno;
670 diff = b1->bno - b2->bno;