Loading...
1/*
2 * Copyright (c) 2000-2002,2005 Silicon Graphics, Inc.
3 * Copyright (c) 2010 David Chinner.
4 * Copyright (c) 2011 Christoph Hellwig.
5 * All Rights Reserved.
6 *
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.
10 *
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.
15 *
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
19 */
20#include "xfs.h"
21#include "xfs_fs.h"
22#include "xfs_format.h"
23#include "xfs_log_format.h"
24#include "xfs_shared.h"
25#include "xfs_trans_resv.h"
26#include "xfs_sb.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"
32#include "xfs_log.h"
33
34void
35xfs_extent_busy_insert(
36 struct xfs_trans *tp,
37 xfs_agnumber_t agno,
38 xfs_agblock_t bno,
39 xfs_extlen_t len,
40 unsigned int flags)
41{
42 struct xfs_extent_busy *new;
43 struct xfs_extent_busy *busyp;
44 struct xfs_perag *pag;
45 struct rb_node **rbp;
46 struct rb_node *parent = NULL;
47
48 new = kmem_zalloc(sizeof(struct xfs_extent_busy), KM_SLEEP);
49 new->agno = agno;
50 new->bno = bno;
51 new->length = len;
52 INIT_LIST_HEAD(&new->list);
53 new->flags = flags;
54
55 /* trace before insert to be able to see failed inserts */
56 trace_xfs_extent_busy(tp->t_mountp, agno, bno, len);
57
58 pag = xfs_perag_get(tp->t_mountp, new->agno);
59 spin_lock(&pag->pagb_lock);
60 rbp = &pag->pagb_tree.rb_node;
61 while (*rbp) {
62 parent = *rbp;
63 busyp = rb_entry(parent, struct xfs_extent_busy, rb_node);
64
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);
71 } else {
72 ASSERT(0);
73 }
74 }
75
76 rb_link_node(&new->rb_node, parent, rbp);
77 rb_insert_color(&new->rb_node, &pag->pagb_tree);
78
79 list_add(&new->list, &tp->t_busy);
80 spin_unlock(&pag->pagb_lock);
81 xfs_perag_put(pag);
82}
83
84/*
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.
92 */
93int
94xfs_extent_busy_search(
95 struct xfs_mount *mp,
96 xfs_agnumber_t agno,
97 xfs_agblock_t bno,
98 xfs_extlen_t len)
99{
100 struct xfs_perag *pag;
101 struct rb_node *rbp;
102 struct xfs_extent_busy *busyp;
103 int match = 0;
104
105 pag = xfs_perag_get(mp, agno);
106 spin_lock(&pag->pagb_lock);
107
108 rbp = pag->pagb_tree.rb_node;
109
110 /* find closest start bno overlap */
111 while (rbp) {
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)
116 match = -1;
117 rbp = rbp->rb_left;
118 } else if (bno > busyp->bno) {
119 /* may overlap, but exact start block is higher */
120 if (bno < busyp->bno + busyp->length)
121 match = -1;
122 rbp = rbp->rb_right;
123 } else {
124 /* bno matches busyp, length determines exact match */
125 match = (busyp->length == len) ? 1 : -1;
126 break;
127 }
128 }
129 spin_unlock(&pag->pagb_lock);
130 xfs_perag_put(pag);
131 return match;
132}
133
134/*
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
140 * number or length.
141 *
142 * Returns true if the extent can safely be reused, or false if the search
143 * needs to be restarted.
144 */
145STATIC bool
146xfs_extent_busy_update_extent(
147 struct xfs_mount *mp,
148 struct xfs_perag *pag,
149 struct xfs_extent_busy *busyp,
150 xfs_agblock_t fbno,
151 xfs_extlen_t flen,
152 bool userdata) __releases(&pag->pagb_lock)
153 __acquires(&pag->pagb_lock)
154{
155 xfs_agblock_t fend = fbno + flen;
156 xfs_agblock_t bbno = busyp->bno;
157 xfs_agblock_t bend = bbno + busyp->length;
158
159 /*
160 * This extent is currently being discarded. Give the thread
161 * performing the discard a chance to mark the extent unbusy
162 * and retry.
163 */
164 if (busyp->flags & XFS_EXTENT_BUSY_DISCARDED) {
165 spin_unlock(&pag->pagb_lock);
166 delay(1);
167 spin_lock(&pag->pagb_lock);
168 return false;
169 }
170
171 /*
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.
174 *
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.
178 */
179 if (userdata)
180 goto out_force_log;
181
182 if (bbno < fbno && bend > fend) {
183 /*
184 * Case 1:
185 * bbno bend
186 * +BBBBBBBBBBBBBBBBB+
187 * +---------+
188 * fbno fend
189 */
190
191 /*
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.
196 *
197 * Force out the log to clear the busy extent and retry the
198 * search.
199 */
200 goto out_force_log;
201 } else if (bbno >= fbno && bend <= fend) {
202 /*
203 * Case 2:
204 * bbno bend
205 * +BBBBBBBBBBBBBBBBB+
206 * +-----------------+
207 * fbno fend
208 *
209 * Case 3:
210 * bbno bend
211 * +BBBBBBBBBBBBBBBBB+
212 * +--------------------------+
213 * fbno fend
214 *
215 * Case 4:
216 * bbno bend
217 * +BBBBBBBBBBBBBBBBB+
218 * +--------------------------+
219 * fbno fend
220 *
221 * Case 5:
222 * bbno bend
223 * +BBBBBBBBBBBBBBBBB+
224 * +-----------------------------------+
225 * fbno fend
226 *
227 */
228
229 /*
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.
235 *
236 * We also need to restart the busy extent search from the
237 * tree root, because erasing the node can rearrange the
238 * tree topology.
239 */
240 rb_erase(&busyp->rb_node, &pag->pagb_tree);
241 busyp->length = 0;
242 return false;
243 } else if (fend < bend) {
244 /*
245 * Case 6:
246 * bbno bend
247 * +BBBBBBBBBBBBBBBBB+
248 * +---------+
249 * fbno fend
250 *
251 * Case 7:
252 * bbno bend
253 * +BBBBBBBBBBBBBBBBB+
254 * +------------------+
255 * fbno fend
256 *
257 */
258 busyp->bno = fend;
259 } else if (bbno < fbno) {
260 /*
261 * Case 8:
262 * bbno bend
263 * +BBBBBBBBBBBBBBBBB+
264 * +-------------+
265 * fbno fend
266 *
267 * Case 9:
268 * bbno bend
269 * +BBBBBBBBBBBBBBBBB+
270 * +----------------------+
271 * fbno fend
272 */
273 busyp->length = fbno - busyp->bno;
274 } else {
275 ASSERT(0);
276 }
277
278 trace_xfs_extent_busy_reuse(mp, pag->pag_agno, fbno, flen);
279 return true;
280
281out_force_log:
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);
286 return false;
287}
288
289
290/*
291 * For a given extent [fbno, flen], make sure we can reuse it safely.
292 */
293void
294xfs_extent_busy_reuse(
295 struct xfs_mount *mp,
296 xfs_agnumber_t agno,
297 xfs_agblock_t fbno,
298 xfs_extlen_t flen,
299 bool userdata)
300{
301 struct xfs_perag *pag;
302 struct rb_node *rbp;
303
304 ASSERT(flen > 0);
305
306 pag = xfs_perag_get(mp, agno);
307 spin_lock(&pag->pagb_lock);
308restart:
309 rbp = pag->pagb_tree.rb_node;
310 while (rbp) {
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;
315
316 if (fbno + flen <= bbno) {
317 rbp = rbp->rb_left;
318 continue;
319 } else if (fbno >= bend) {
320 rbp = rbp->rb_right;
321 continue;
322 }
323
324 if (!xfs_extent_busy_update_extent(mp, pag, busyp, fbno, flen,
325 userdata))
326 goto restart;
327 }
328 spin_unlock(&pag->pagb_lock);
329 xfs_perag_put(pag);
330}
331
332/*
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.
337 *
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.
343 */
344bool
345xfs_extent_busy_trim(
346 struct xfs_alloc_arg *args,
347 xfs_agblock_t *bno,
348 xfs_extlen_t *len,
349 unsigned *busy_gen)
350{
351 xfs_agblock_t fbno;
352 xfs_extlen_t flen;
353 struct rb_node *rbp;
354 bool ret = false;
355
356 ASSERT(*len > 0);
357
358 spin_lock(&args->pag->pagb_lock);
359restart:
360 fbno = *bno;
361 flen = *len;
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;
369
370 if (fend <= bbno) {
371 rbp = rbp->rb_left;
372 continue;
373 } else if (fbno >= bend) {
374 rbp = rbp->rb_right;
375 continue;
376 }
377
378 /*
379 * If this is a metadata allocation, try to reuse the busy
380 * extent instead of trimming the allocation.
381 */
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,
385 busyp, fbno, flen,
386 false))
387 goto restart;
388 continue;
389 }
390
391 if (bbno <= fbno) {
392 /* start overlap */
393
394 /*
395 * Case 1:
396 * bbno bend
397 * +BBBBBBBBBBBBBBBBB+
398 * +---------+
399 * fbno fend
400 *
401 * Case 2:
402 * bbno bend
403 * +BBBBBBBBBBBBBBBBB+
404 * +-------------+
405 * fbno fend
406 *
407 * Case 3:
408 * bbno bend
409 * +BBBBBBBBBBBBBBBBB+
410 * +-------------+
411 * fbno fend
412 *
413 * Case 4:
414 * bbno bend
415 * +BBBBBBBBBBBBBBBBB+
416 * +-----------------+
417 * fbno fend
418 *
419 * No unbusy region in extent, return failure.
420 */
421 if (fend <= bend)
422 goto fail;
423
424 /*
425 * Case 5:
426 * bbno bend
427 * +BBBBBBBBBBBBBBBBB+
428 * +----------------------+
429 * fbno fend
430 *
431 * Case 6:
432 * bbno bend
433 * +BBBBBBBBBBBBBBBBB+
434 * +--------------------------+
435 * fbno fend
436 *
437 * Needs to be trimmed to:
438 * +-------+
439 * fbno fend
440 */
441 fbno = bend;
442 } else if (bend >= fend) {
443 /* end overlap */
444
445 /*
446 * Case 7:
447 * bbno bend
448 * +BBBBBBBBBBBBBBBBB+
449 * +------------------+
450 * fbno fend
451 *
452 * Case 8:
453 * bbno bend
454 * +BBBBBBBBBBBBBBBBB+
455 * +--------------------------+
456 * fbno fend
457 *
458 * Needs to be trimmed to:
459 * +-------+
460 * fbno fend
461 */
462 fend = bbno;
463 } else {
464 /* middle overlap */
465
466 /*
467 * Case 9:
468 * bbno bend
469 * +BBBBBBBBBBBBBBBBB+
470 * +-----------------------------------+
471 * fbno fend
472 *
473 * Can be trimmed to:
474 * +-------+ OR +-------+
475 * fbno fend fbno fend
476 *
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
481 * patterns.
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
487 * allocation.
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
495 * contiguous.
496 */
497 if (bbno - fbno >= args->maxlen) {
498 /* left candidate fits perfect */
499 fend = bbno;
500 } else if (fend - bend >= args->maxlen * 4) {
501 /* right candidate has enough free space */
502 fbno = bend;
503 } else if (bbno - fbno >= args->minlen) {
504 /* left candidate fits minimum requirement */
505 fend = bbno;
506 } else {
507 goto fail;
508 }
509 }
510
511 flen = fend - fbno;
512 }
513out:
514
515 if (fbno != *bno || flen != *len) {
516 trace_xfs_extent_busy_trim(args->mp, args->agno, *bno, *len,
517 fbno, flen);
518 *bno = fbno;
519 *len = flen;
520 *busy_gen = args->pag->pagb_gen;
521 ret = true;
522 }
523 spin_unlock(&args->pag->pagb_lock);
524 return ret;
525fail:
526 /*
527 * Return a zero extent length as failure indications. All callers
528 * re-check if the trimmed extent satisfies the minlen requirement.
529 */
530 flen = 0;
531 goto out;
532}
533
534STATIC void
535xfs_extent_busy_clear_one(
536 struct xfs_mount *mp,
537 struct xfs_perag *pag,
538 struct xfs_extent_busy *busyp)
539{
540 if (busyp->length) {
541 trace_xfs_extent_busy_clear(mp, busyp->agno, busyp->bno,
542 busyp->length);
543 rb_erase(&busyp->rb_node, &pag->pagb_tree);
544 }
545
546 list_del_init(&busyp->list);
547 kmem_free(busyp);
548}
549
550static void
551xfs_extent_busy_put_pag(
552 struct xfs_perag *pag,
553 bool wakeup)
554 __releases(pag->pagb_lock)
555{
556 if (wakeup) {
557 pag->pagb_gen++;
558 wake_up_all(&pag->pagb_wait);
559 }
560
561 spin_unlock(&pag->pagb_lock);
562 xfs_perag_put(pag);
563}
564
565/*
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.
569 */
570void
571xfs_extent_busy_clear(
572 struct xfs_mount *mp,
573 struct list_head *list,
574 bool do_discard)
575{
576 struct xfs_extent_busy *busyp, *n;
577 struct xfs_perag *pag = NULL;
578 xfs_agnumber_t agno = NULLAGNUMBER;
579 bool wakeup = false;
580
581 list_for_each_entry_safe(busyp, n, list, list) {
582 if (busyp->agno != agno) {
583 if (pag)
584 xfs_extent_busy_put_pag(pag, wakeup);
585 agno = busyp->agno;
586 pag = xfs_perag_get(mp, agno);
587 spin_lock(&pag->pagb_lock);
588 wakeup = false;
589 }
590
591 if (do_discard && busyp->length &&
592 !(busyp->flags & XFS_EXTENT_BUSY_SKIP_DISCARD)) {
593 busyp->flags = XFS_EXTENT_BUSY_DISCARDED;
594 } else {
595 xfs_extent_busy_clear_one(mp, pag, busyp);
596 wakeup = true;
597 }
598 }
599
600 if (pag)
601 xfs_extent_busy_put_pag(pag, wakeup);
602}
603
604/*
605 * Flush out all busy extents for this AG.
606 */
607void
608xfs_extent_busy_flush(
609 struct xfs_mount *mp,
610 struct xfs_perag *pag,
611 unsigned busy_gen)
612{
613 DEFINE_WAIT (wait);
614 int error;
615
616 error = xfs_log_force(mp, XFS_LOG_SYNC);
617 if (error)
618 return;
619
620 do {
621 prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
622 if (busy_gen != READ_ONCE(pag->pagb_gen))
623 break;
624 schedule();
625 } while (1);
626
627 finish_wait(&pag->pagb_wait, &wait);
628}
629
630void
631xfs_extent_busy_wait_all(
632 struct xfs_mount *mp)
633{
634 DEFINE_WAIT (wait);
635 xfs_agnumber_t agno;
636
637 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
638 struct xfs_perag *pag = xfs_perag_get(mp, agno);
639
640 do {
641 prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
642 if (RB_EMPTY_ROOT(&pag->pagb_tree))
643 break;
644 schedule();
645 } while (1);
646 finish_wait(&pag->pagb_wait, &wait);
647
648 xfs_perag_put(pag);
649 }
650}
651
652/*
653 * Callback for list_sort to sort busy extents by the AG they reside in.
654 */
655int
656xfs_extent_busy_ag_cmp(
657 void *priv,
658 struct list_head *l1,
659 struct list_head *l2)
660{
661 struct xfs_extent_busy *b1 =
662 container_of(l1, struct xfs_extent_busy, list);
663 struct xfs_extent_busy *b2 =
664 container_of(l2, struct xfs_extent_busy, list);
665 s32 diff;
666
667 diff = b1->agno - b2->agno;
668 if (!diff)
669 diff = b1->bno - b2->bno;
670 return diff;
671}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2002,2005 Silicon Graphics, Inc.
4 * Copyright (c) 2010 David Chinner.
5 * Copyright (c) 2011 Christoph Hellwig.
6 * All Rights Reserved.
7 */
8#include "xfs.h"
9#include "xfs_fs.h"
10#include "xfs_format.h"
11#include "xfs_log_format.h"
12#include "xfs_shared.h"
13#include "xfs_trans_resv.h"
14#include "xfs_mount.h"
15#include "xfs_alloc.h"
16#include "xfs_extent_busy.h"
17#include "xfs_trace.h"
18#include "xfs_trans.h"
19#include "xfs_log.h"
20#include "xfs_ag.h"
21
22void
23xfs_extent_busy_insert(
24 struct xfs_trans *tp,
25 struct xfs_perag *pag,
26 xfs_agblock_t bno,
27 xfs_extlen_t len,
28 unsigned int flags)
29{
30 struct xfs_extent_busy *new;
31 struct xfs_extent_busy *busyp;
32 struct rb_node **rbp;
33 struct rb_node *parent = NULL;
34
35 new = kmem_zalloc(sizeof(struct xfs_extent_busy), 0);
36 new->agno = pag->pag_agno;
37 new->bno = bno;
38 new->length = len;
39 INIT_LIST_HEAD(&new->list);
40 new->flags = flags;
41
42 /* trace before insert to be able to see failed inserts */
43 trace_xfs_extent_busy(tp->t_mountp, pag->pag_agno, bno, len);
44
45 spin_lock(&pag->pagb_lock);
46 rbp = &pag->pagb_tree.rb_node;
47 while (*rbp) {
48 parent = *rbp;
49 busyp = rb_entry(parent, struct xfs_extent_busy, rb_node);
50
51 if (new->bno < busyp->bno) {
52 rbp = &(*rbp)->rb_left;
53 ASSERT(new->bno + new->length <= busyp->bno);
54 } else if (new->bno > busyp->bno) {
55 rbp = &(*rbp)->rb_right;
56 ASSERT(bno >= busyp->bno + busyp->length);
57 } else {
58 ASSERT(0);
59 }
60 }
61
62 rb_link_node(&new->rb_node, parent, rbp);
63 rb_insert_color(&new->rb_node, &pag->pagb_tree);
64
65 list_add(&new->list, &tp->t_busy);
66 spin_unlock(&pag->pagb_lock);
67}
68
69/*
70 * Search for a busy extent within the range of the extent we are about to
71 * allocate. You need to be holding the busy extent tree lock when calling
72 * xfs_extent_busy_search(). This function returns 0 for no overlapping busy
73 * extent, -1 for an overlapping but not exact busy extent, and 1 for an exact
74 * match. This is done so that a non-zero return indicates an overlap that
75 * will require a synchronous transaction, but it can still be
76 * used to distinguish between a partial or exact match.
77 */
78int
79xfs_extent_busy_search(
80 struct xfs_mount *mp,
81 struct xfs_perag *pag,
82 xfs_agblock_t bno,
83 xfs_extlen_t len)
84{
85 struct rb_node *rbp;
86 struct xfs_extent_busy *busyp;
87 int match = 0;
88
89 /* find closest start bno overlap */
90 spin_lock(&pag->pagb_lock);
91 rbp = pag->pagb_tree.rb_node;
92 while (rbp) {
93 busyp = rb_entry(rbp, struct xfs_extent_busy, rb_node);
94 if (bno < busyp->bno) {
95 /* may overlap, but exact start block is lower */
96 if (bno + len > busyp->bno)
97 match = -1;
98 rbp = rbp->rb_left;
99 } else if (bno > busyp->bno) {
100 /* may overlap, but exact start block is higher */
101 if (bno < busyp->bno + busyp->length)
102 match = -1;
103 rbp = rbp->rb_right;
104 } else {
105 /* bno matches busyp, length determines exact match */
106 match = (busyp->length == len) ? 1 : -1;
107 break;
108 }
109 }
110 spin_unlock(&pag->pagb_lock);
111 return match;
112}
113
114/*
115 * The found free extent [fbno, fend] overlaps part or all of the given busy
116 * extent. If the overlap covers the beginning, the end, or all of the busy
117 * extent, the overlapping portion can be made unbusy and used for the
118 * allocation. We can't split a busy extent because we can't modify a
119 * transaction/CIL context busy list, but we can update an entry's block
120 * number or length.
121 *
122 * Returns true if the extent can safely be reused, or false if the search
123 * needs to be restarted.
124 */
125STATIC bool
126xfs_extent_busy_update_extent(
127 struct xfs_mount *mp,
128 struct xfs_perag *pag,
129 struct xfs_extent_busy *busyp,
130 xfs_agblock_t fbno,
131 xfs_extlen_t flen,
132 bool userdata) __releases(&pag->pagb_lock)
133 __acquires(&pag->pagb_lock)
134{
135 xfs_agblock_t fend = fbno + flen;
136 xfs_agblock_t bbno = busyp->bno;
137 xfs_agblock_t bend = bbno + busyp->length;
138
139 /*
140 * This extent is currently being discarded. Give the thread
141 * performing the discard a chance to mark the extent unbusy
142 * and retry.
143 */
144 if (busyp->flags & XFS_EXTENT_BUSY_DISCARDED) {
145 spin_unlock(&pag->pagb_lock);
146 delay(1);
147 spin_lock(&pag->pagb_lock);
148 return false;
149 }
150
151 /*
152 * If there is a busy extent overlapping a user allocation, we have
153 * no choice but to force the log and retry the search.
154 *
155 * Fortunately this does not happen during normal operation, but
156 * only if the filesystem is very low on space and has to dip into
157 * the AGFL for normal allocations.
158 */
159 if (userdata)
160 goto out_force_log;
161
162 if (bbno < fbno && bend > fend) {
163 /*
164 * Case 1:
165 * bbno bend
166 * +BBBBBBBBBBBBBBBBB+
167 * +---------+
168 * fbno fend
169 */
170
171 /*
172 * We would have to split the busy extent to be able to track
173 * it correct, which we cannot do because we would have to
174 * modify the list of busy extents attached to the transaction
175 * or CIL context, which is immutable.
176 *
177 * Force out the log to clear the busy extent and retry the
178 * search.
179 */
180 goto out_force_log;
181 } else if (bbno >= fbno && bend <= fend) {
182 /*
183 * Case 2:
184 * bbno bend
185 * +BBBBBBBBBBBBBBBBB+
186 * +-----------------+
187 * fbno fend
188 *
189 * Case 3:
190 * bbno bend
191 * +BBBBBBBBBBBBBBBBB+
192 * +--------------------------+
193 * fbno fend
194 *
195 * Case 4:
196 * bbno bend
197 * +BBBBBBBBBBBBBBBBB+
198 * +--------------------------+
199 * fbno fend
200 *
201 * Case 5:
202 * bbno bend
203 * +BBBBBBBBBBBBBBBBB+
204 * +-----------------------------------+
205 * fbno fend
206 *
207 */
208
209 /*
210 * The busy extent is fully covered by the extent we are
211 * allocating, and can simply be removed from the rbtree.
212 * However we cannot remove it from the immutable list
213 * tracking busy extents in the transaction or CIL context,
214 * so set the length to zero to mark it invalid.
215 *
216 * We also need to restart the busy extent search from the
217 * tree root, because erasing the node can rearrange the
218 * tree topology.
219 */
220 rb_erase(&busyp->rb_node, &pag->pagb_tree);
221 busyp->length = 0;
222 return false;
223 } else if (fend < bend) {
224 /*
225 * Case 6:
226 * bbno bend
227 * +BBBBBBBBBBBBBBBBB+
228 * +---------+
229 * fbno fend
230 *
231 * Case 7:
232 * bbno bend
233 * +BBBBBBBBBBBBBBBBB+
234 * +------------------+
235 * fbno fend
236 *
237 */
238 busyp->bno = fend;
239 busyp->length = bend - fend;
240 } else if (bbno < fbno) {
241 /*
242 * Case 8:
243 * bbno bend
244 * +BBBBBBBBBBBBBBBBB+
245 * +-------------+
246 * fbno fend
247 *
248 * Case 9:
249 * bbno bend
250 * +BBBBBBBBBBBBBBBBB+
251 * +----------------------+
252 * fbno fend
253 */
254 busyp->length = fbno - busyp->bno;
255 } else {
256 ASSERT(0);
257 }
258
259 trace_xfs_extent_busy_reuse(mp, pag->pag_agno, fbno, flen);
260 return true;
261
262out_force_log:
263 spin_unlock(&pag->pagb_lock);
264 xfs_log_force(mp, XFS_LOG_SYNC);
265 trace_xfs_extent_busy_force(mp, pag->pag_agno, fbno, flen);
266 spin_lock(&pag->pagb_lock);
267 return false;
268}
269
270
271/*
272 * For a given extent [fbno, flen], make sure we can reuse it safely.
273 */
274void
275xfs_extent_busy_reuse(
276 struct xfs_mount *mp,
277 struct xfs_perag *pag,
278 xfs_agblock_t fbno,
279 xfs_extlen_t flen,
280 bool userdata)
281{
282 struct rb_node *rbp;
283
284 ASSERT(flen > 0);
285 spin_lock(&pag->pagb_lock);
286restart:
287 rbp = pag->pagb_tree.rb_node;
288 while (rbp) {
289 struct xfs_extent_busy *busyp =
290 rb_entry(rbp, struct xfs_extent_busy, rb_node);
291 xfs_agblock_t bbno = busyp->bno;
292 xfs_agblock_t bend = bbno + busyp->length;
293
294 if (fbno + flen <= bbno) {
295 rbp = rbp->rb_left;
296 continue;
297 } else if (fbno >= bend) {
298 rbp = rbp->rb_right;
299 continue;
300 }
301
302 if (!xfs_extent_busy_update_extent(mp, pag, busyp, fbno, flen,
303 userdata))
304 goto restart;
305 }
306 spin_unlock(&pag->pagb_lock);
307}
308
309/*
310 * For a given extent [fbno, flen], search the busy extent list to find a
311 * subset of the extent that is not busy. If *rlen is smaller than
312 * args->minlen no suitable extent could be found, and the higher level
313 * code needs to force out the log and retry the allocation.
314 *
315 * Return the current busy generation for the AG if the extent is busy. This
316 * value can be used to wait for at least one of the currently busy extents
317 * to be cleared. Note that the busy list is not guaranteed to be empty after
318 * the gen is woken. The state of a specific extent must always be confirmed
319 * with another call to xfs_extent_busy_trim() before it can be used.
320 */
321bool
322xfs_extent_busy_trim(
323 struct xfs_alloc_arg *args,
324 xfs_agblock_t *bno,
325 xfs_extlen_t *len,
326 unsigned *busy_gen)
327{
328 xfs_agblock_t fbno;
329 xfs_extlen_t flen;
330 struct rb_node *rbp;
331 bool ret = false;
332
333 ASSERT(*len > 0);
334
335 spin_lock(&args->pag->pagb_lock);
336 fbno = *bno;
337 flen = *len;
338 rbp = args->pag->pagb_tree.rb_node;
339 while (rbp && flen >= args->minlen) {
340 struct xfs_extent_busy *busyp =
341 rb_entry(rbp, struct xfs_extent_busy, rb_node);
342 xfs_agblock_t fend = fbno + flen;
343 xfs_agblock_t bbno = busyp->bno;
344 xfs_agblock_t bend = bbno + busyp->length;
345
346 if (fend <= bbno) {
347 rbp = rbp->rb_left;
348 continue;
349 } else if (fbno >= bend) {
350 rbp = rbp->rb_right;
351 continue;
352 }
353
354 if (bbno <= fbno) {
355 /* start overlap */
356
357 /*
358 * Case 1:
359 * bbno bend
360 * +BBBBBBBBBBBBBBBBB+
361 * +---------+
362 * fbno fend
363 *
364 * Case 2:
365 * bbno bend
366 * +BBBBBBBBBBBBBBBBB+
367 * +-------------+
368 * fbno fend
369 *
370 * Case 3:
371 * bbno bend
372 * +BBBBBBBBBBBBBBBBB+
373 * +-------------+
374 * fbno fend
375 *
376 * Case 4:
377 * bbno bend
378 * +BBBBBBBBBBBBBBBBB+
379 * +-----------------+
380 * fbno fend
381 *
382 * No unbusy region in extent, return failure.
383 */
384 if (fend <= bend)
385 goto fail;
386
387 /*
388 * Case 5:
389 * bbno bend
390 * +BBBBBBBBBBBBBBBBB+
391 * +----------------------+
392 * fbno fend
393 *
394 * Case 6:
395 * bbno bend
396 * +BBBBBBBBBBBBBBBBB+
397 * +--------------------------+
398 * fbno fend
399 *
400 * Needs to be trimmed to:
401 * +-------+
402 * fbno fend
403 */
404 fbno = bend;
405 } else if (bend >= fend) {
406 /* end overlap */
407
408 /*
409 * Case 7:
410 * bbno bend
411 * +BBBBBBBBBBBBBBBBB+
412 * +------------------+
413 * fbno fend
414 *
415 * Case 8:
416 * bbno bend
417 * +BBBBBBBBBBBBBBBBB+
418 * +--------------------------+
419 * fbno fend
420 *
421 * Needs to be trimmed to:
422 * +-------+
423 * fbno fend
424 */
425 fend = bbno;
426 } else {
427 /* middle overlap */
428
429 /*
430 * Case 9:
431 * bbno bend
432 * +BBBBBBBBBBBBBBBBB+
433 * +-----------------------------------+
434 * fbno fend
435 *
436 * Can be trimmed to:
437 * +-------+ OR +-------+
438 * fbno fend fbno fend
439 *
440 * Backward allocation leads to significant
441 * fragmentation of directories, which degrades
442 * directory performance, therefore we always want to
443 * choose the option that produces forward allocation
444 * patterns.
445 * Preferring the lower bno extent will make the next
446 * request use "fend" as the start of the next
447 * allocation; if the segment is no longer busy at
448 * that point, we'll get a contiguous allocation, but
449 * even if it is still busy, we will get a forward
450 * allocation.
451 * We try to avoid choosing the segment at "bend",
452 * because that can lead to the next allocation
453 * taking the segment at "fbno", which would be a
454 * backward allocation. We only use the segment at
455 * "fbno" if it is much larger than the current
456 * requested size, because in that case there's a
457 * good chance subsequent allocations will be
458 * contiguous.
459 */
460 if (bbno - fbno >= args->maxlen) {
461 /* left candidate fits perfect */
462 fend = bbno;
463 } else if (fend - bend >= args->maxlen * 4) {
464 /* right candidate has enough free space */
465 fbno = bend;
466 } else if (bbno - fbno >= args->minlen) {
467 /* left candidate fits minimum requirement */
468 fend = bbno;
469 } else {
470 goto fail;
471 }
472 }
473
474 flen = fend - fbno;
475 }
476out:
477
478 if (fbno != *bno || flen != *len) {
479 trace_xfs_extent_busy_trim(args->mp, args->agno, *bno, *len,
480 fbno, flen);
481 *bno = fbno;
482 *len = flen;
483 *busy_gen = args->pag->pagb_gen;
484 ret = true;
485 }
486 spin_unlock(&args->pag->pagb_lock);
487 return ret;
488fail:
489 /*
490 * Return a zero extent length as failure indications. All callers
491 * re-check if the trimmed extent satisfies the minlen requirement.
492 */
493 flen = 0;
494 goto out;
495}
496
497STATIC void
498xfs_extent_busy_clear_one(
499 struct xfs_mount *mp,
500 struct xfs_perag *pag,
501 struct xfs_extent_busy *busyp)
502{
503 if (busyp->length) {
504 trace_xfs_extent_busy_clear(mp, busyp->agno, busyp->bno,
505 busyp->length);
506 rb_erase(&busyp->rb_node, &pag->pagb_tree);
507 }
508
509 list_del_init(&busyp->list);
510 kmem_free(busyp);
511}
512
513static void
514xfs_extent_busy_put_pag(
515 struct xfs_perag *pag,
516 bool wakeup)
517 __releases(pag->pagb_lock)
518{
519 if (wakeup) {
520 pag->pagb_gen++;
521 wake_up_all(&pag->pagb_wait);
522 }
523
524 spin_unlock(&pag->pagb_lock);
525 xfs_perag_put(pag);
526}
527
528/*
529 * Remove all extents on the passed in list from the busy extents tree.
530 * If do_discard is set skip extents that need to be discarded, and mark
531 * these as undergoing a discard operation instead.
532 */
533void
534xfs_extent_busy_clear(
535 struct xfs_mount *mp,
536 struct list_head *list,
537 bool do_discard)
538{
539 struct xfs_extent_busy *busyp, *n;
540 struct xfs_perag *pag = NULL;
541 xfs_agnumber_t agno = NULLAGNUMBER;
542 bool wakeup = false;
543
544 list_for_each_entry_safe(busyp, n, list, list) {
545 if (busyp->agno != agno) {
546 if (pag)
547 xfs_extent_busy_put_pag(pag, wakeup);
548 agno = busyp->agno;
549 pag = xfs_perag_get(mp, agno);
550 spin_lock(&pag->pagb_lock);
551 wakeup = false;
552 }
553
554 if (do_discard && busyp->length &&
555 !(busyp->flags & XFS_EXTENT_BUSY_SKIP_DISCARD)) {
556 busyp->flags = XFS_EXTENT_BUSY_DISCARDED;
557 } else {
558 xfs_extent_busy_clear_one(mp, pag, busyp);
559 wakeup = true;
560 }
561 }
562
563 if (pag)
564 xfs_extent_busy_put_pag(pag, wakeup);
565}
566
567/*
568 * Flush out all busy extents for this AG.
569 */
570void
571xfs_extent_busy_flush(
572 struct xfs_mount *mp,
573 struct xfs_perag *pag,
574 unsigned busy_gen)
575{
576 DEFINE_WAIT (wait);
577 int error;
578
579 error = xfs_log_force(mp, XFS_LOG_SYNC);
580 if (error)
581 return;
582
583 do {
584 prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
585 if (busy_gen != READ_ONCE(pag->pagb_gen))
586 break;
587 schedule();
588 } while (1);
589
590 finish_wait(&pag->pagb_wait, &wait);
591}
592
593void
594xfs_extent_busy_wait_all(
595 struct xfs_mount *mp)
596{
597 struct xfs_perag *pag;
598 DEFINE_WAIT (wait);
599 xfs_agnumber_t agno;
600
601 for_each_perag(mp, agno, pag) {
602 do {
603 prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
604 if (RB_EMPTY_ROOT(&pag->pagb_tree))
605 break;
606 schedule();
607 } while (1);
608 finish_wait(&pag->pagb_wait, &wait);
609 }
610}
611
612/*
613 * Callback for list_sort to sort busy extents by the AG they reside in.
614 */
615int
616xfs_extent_busy_ag_cmp(
617 void *priv,
618 const struct list_head *l1,
619 const struct list_head *l2)
620{
621 struct xfs_extent_busy *b1 =
622 container_of(l1, struct xfs_extent_busy, list);
623 struct xfs_extent_busy *b2 =
624 container_of(l2, struct xfs_extent_busy, list);
625 s32 diff;
626
627 diff = b1->agno - b2->agno;
628 if (!diff)
629 diff = b1->bno - b2->bno;
630 return diff;
631}