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