Line data Source code
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 :
22 : void
23 40965154 : xfs_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 40965154 : struct xfs_extent_busy *new;
31 40965154 : struct xfs_extent_busy *busyp;
32 40965154 : struct rb_node **rbp;
33 40965154 : struct rb_node *parent = NULL;
34 :
35 40965154 : new = kmem_zalloc(sizeof(struct xfs_extent_busy), 0);
36 40965142 : new->agno = pag->pag_agno;
37 40965142 : new->bno = bno;
38 40965142 : new->length = len;
39 40965142 : INIT_LIST_HEAD(&new->list);
40 40965142 : new->flags = flags;
41 :
42 : /* trace before insert to be able to see failed inserts */
43 40965142 : trace_xfs_extent_busy(tp->t_mountp, pag->pag_agno, bno, len);
44 :
45 40965187 : spin_lock(&pag->pagb_lock);
46 40965348 : rbp = &pag->pagb_tree.rb_node;
47 392217852 : while (*rbp) {
48 351252661 : parent = *rbp;
49 351252661 : busyp = rb_entry(parent, struct xfs_extent_busy, rb_node);
50 :
51 351252661 : if (new->bno < busyp->bno) {
52 104122137 : rbp = &(*rbp)->rb_left;
53 104122137 : ASSERT(new->bno + new->length <= busyp->bno);
54 247130524 : } else if (new->bno > busyp->bno) {
55 247130524 : rbp = &(*rbp)->rb_right;
56 247130524 : ASSERT(bno >= busyp->bno + busyp->length);
57 : } else {
58 0 : ASSERT(0);
59 : }
60 : }
61 :
62 40965191 : rb_link_node(&new->rb_node, parent, rbp);
63 40965191 : rb_insert_color(&new->rb_node, &pag->pagb_tree);
64 :
65 40965262 : list_add(&new->list, &tp->t_busy);
66 40965309 : spin_unlock(&pag->pagb_lock);
67 40965228 : }
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 : */
78 : int
79 44924506 : xfs_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 44924506 : struct rb_node *rbp;
86 44924506 : struct xfs_extent_busy *busyp;
87 44924506 : int match = 0;
88 :
89 : /* find closest start bno overlap */
90 44924506 : spin_lock(&pag->pagb_lock);
91 44926338 : rbp = pag->pagb_tree.rb_node;
92 222599750 : while (rbp) {
93 177676333 : busyp = rb_entry(rbp, struct xfs_extent_busy, rb_node);
94 177676333 : if (bno < busyp->bno) {
95 : /* may overlap, but exact start block is lower */
96 71518693 : if (bno + len > busyp->bno)
97 2386 : match = -1;
98 71518693 : rbp = rbp->rb_left;
99 106157640 : } else if (bno > busyp->bno) {
100 : /* may overlap, but exact start block is higher */
101 106154719 : if (bno < busyp->bno + busyp->length)
102 0 : match = -1;
103 106154719 : rbp = rbp->rb_right;
104 : } else {
105 : /* bno matches busyp, length determines exact match */
106 2921 : match = (busyp->length == len) ? 1 : -1;
107 : break;
108 : }
109 : }
110 44926338 : spin_unlock(&pag->pagb_lock);
111 44925876 : 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 : */
125 : STATIC bool
126 4034 : xfs_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 4034 : xfs_agblock_t fend = fbno + flen;
136 4034 : xfs_agblock_t bbno = busyp->bno;
137 4034 : 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 4034 : if (busyp->flags & XFS_EXTENT_BUSY_DISCARDED) {
145 0 : spin_unlock(&pag->pagb_lock);
146 0 : delay(1);
147 0 : spin_lock(&pag->pagb_lock);
148 0 : 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 4034 : if (userdata)
160 0 : goto out_force_log;
161 :
162 4034 : 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 0 : goto out_force_log;
181 4034 : } 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 4034 : rb_erase(&busyp->rb_node, &pag->pagb_tree);
221 4034 : busyp->length = 0;
222 4034 : return false;
223 0 : } 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 0 : busyp->bno = fend;
239 0 : busyp->length = bend - fend;
240 0 : } 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 0 : busyp->length = fbno - busyp->bno;
255 : } else {
256 0 : ASSERT(0);
257 : }
258 :
259 0 : trace_xfs_extent_busy_reuse(mp, pag->pag_agno, fbno, flen);
260 0 : return true;
261 :
262 0 : out_force_log:
263 0 : spin_unlock(&pag->pagb_lock);
264 0 : xfs_log_force(mp, XFS_LOG_SYNC);
265 0 : trace_xfs_extent_busy_force(mp, pag->pag_agno, fbno, flen);
266 0 : spin_lock(&pag->pagb_lock);
267 0 : return false;
268 : }
269 :
270 :
271 : /*
272 : * For a given extent [fbno, flen], make sure we can reuse it safely.
273 : */
274 : void
275 338197 : xfs_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 338197 : struct rb_node *rbp;
283 :
284 338197 : ASSERT(flen > 0);
285 338197 : spin_lock(&pag->pagb_lock);
286 342231 : restart:
287 342231 : rbp = pag->pagb_tree.rb_node;
288 1520926 : while (rbp) {
289 1182729 : struct xfs_extent_busy *busyp =
290 1182729 : rb_entry(rbp, struct xfs_extent_busy, rb_node);
291 1182729 : xfs_agblock_t bbno = busyp->bno;
292 1182729 : xfs_agblock_t bend = bbno + busyp->length;
293 :
294 1182729 : if (fbno + flen <= bbno) {
295 452909 : rbp = rbp->rb_left;
296 452909 : continue;
297 729820 : } else if (fbno >= bend) {
298 725786 : rbp = rbp->rb_right;
299 725786 : continue;
300 : }
301 :
302 4034 : if (!xfs_extent_busy_update_extent(mp, pag, busyp, fbno, flen,
303 : userdata))
304 4034 : goto restart;
305 : }
306 338197 : spin_unlock(&pag->pagb_lock);
307 338197 : }
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 : */
321 : bool
322 3304834309 : xfs_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 3304834309 : xfs_agblock_t fbno;
329 3304834309 : xfs_extlen_t flen;
330 3304834309 : struct rb_node *rbp;
331 3304834309 : bool ret = false;
332 :
333 3304834309 : ASSERT(*len > 0);
334 :
335 3304834309 : spin_lock(&args->pag->pagb_lock);
336 3313453636 : fbno = *bno;
337 3313453636 : flen = *len;
338 3313453636 : rbp = args->pag->pagb_tree.rb_node;
339 26711293397 : while (rbp && flen >= args->minlen) {
340 24898669291 : struct xfs_extent_busy *busyp =
341 24898669291 : rb_entry(rbp, struct xfs_extent_busy, rb_node);
342 24898669291 : xfs_agblock_t fend = fbno + flen;
343 24898669291 : xfs_agblock_t bbno = busyp->bno;
344 24898669291 : xfs_agblock_t bend = bbno + busyp->length;
345 :
346 24898669291 : if (fend <= bbno) {
347 11455257153 : rbp = rbp->rb_left;
348 11455257153 : continue;
349 13443412138 : } else if (fbno >= bend) {
350 11045423929 : rbp = rbp->rb_right;
351 11045423929 : continue;
352 : }
353 :
354 2397988209 : 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 1814556353 : if (fend <= bend)
385 1497282126 : 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 583431856 : } 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 347622667 : if (bbno - fbno >= args->maxlen) {
461 : /* left candidate fits perfect */
462 : fend = bbno;
463 4155424 : } else if (fend - bend >= args->maxlen * 4) {
464 : /* right candidate has enough free space */
465 : fbno = bend;
466 3555050 : } else if (bbno - fbno >= args->minlen) {
467 : /* left candidate fits minimum requirement */
468 : fend = bbno;
469 : } else {
470 3547404 : goto fail;
471 : }
472 : }
473 :
474 897158679 : flen = fend - fbno;
475 : }
476 1812624106 : out:
477 :
478 3313453636 : if (fbno != *bno || flen != *len) {
479 1635376088 : trace_xfs_extent_busy_trim(args->mp, args->agno, *bno, *len,
480 : fbno, flen);
481 1634514127 : *bno = fbno;
482 1634514127 : *len = flen;
483 1634514127 : *busy_gen = args->pag->pagb_gen;
484 1634514127 : ret = true;
485 : }
486 3312591675 : spin_unlock(&args->pag->pagb_lock);
487 3312075377 : return ret;
488 1500829530 : fail:
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 1500829530 : flen = 0;
494 1500829530 : goto out;
495 : }
496 :
497 : STATIC void
498 40965494 : xfs_extent_busy_clear_one(
499 : struct xfs_mount *mp,
500 : struct xfs_perag *pag,
501 : struct xfs_extent_busy *busyp)
502 : {
503 40965494 : if (busyp->length) {
504 40961460 : trace_xfs_extent_busy_clear(mp, busyp->agno, busyp->bno,
505 : busyp->length);
506 40961460 : rb_erase(&busyp->rb_node, &pag->pagb_tree);
507 : }
508 :
509 40965494 : list_del_init(&busyp->list);
510 40965494 : kmem_free(busyp);
511 40965494 : }
512 :
513 : static void
514 2791888 : xfs_extent_busy_put_pag(
515 : struct xfs_perag *pag,
516 : bool wakeup)
517 : __releases(pag->pagb_lock)
518 : {
519 2791888 : if (wakeup) {
520 2791849 : pag->pagb_gen++;
521 2791849 : wake_up_all(&pag->pagb_wait);
522 : }
523 :
524 2791888 : spin_unlock(&pag->pagb_lock);
525 2791888 : xfs_perag_put(pag);
526 2791888 : }
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 : */
533 : void
534 3226500214 : xfs_extent_busy_clear(
535 : struct xfs_mount *mp,
536 : struct list_head *list,
537 : bool do_discard)
538 : {
539 3226500214 : struct xfs_extent_busy *busyp, *n;
540 3226500214 : struct xfs_perag *pag = NULL;
541 3226500214 : xfs_agnumber_t agno = NULLAGNUMBER;
542 3226500214 : bool wakeup = false;
543 :
544 3267703879 : list_for_each_entry_safe(busyp, n, list, list) {
545 40965746 : if (busyp->agno != agno) {
546 2791888 : if (pag)
547 1356658 : xfs_extent_busy_put_pag(pag, wakeup);
548 2791888 : agno = busyp->agno;
549 2791888 : pag = xfs_perag_get(mp, agno);
550 2791888 : spin_lock(&pag->pagb_lock);
551 2791888 : wakeup = false;
552 : }
553 :
554 40965746 : if (do_discard && busyp->length &&
555 357 : !(busyp->flags & XFS_EXTENT_BUSY_SKIP_DISCARD)) {
556 252 : busyp->flags = XFS_EXTENT_BUSY_DISCARDED;
557 : } else {
558 40965494 : xfs_extent_busy_clear_one(mp, pag, busyp);
559 40965494 : wakeup = true;
560 : }
561 : }
562 :
563 3226738133 : if (pag)
564 1435230 : xfs_extent_busy_put_pag(pag, wakeup);
565 3226738133 : }
566 :
567 : /*
568 : * Flush out all busy extents for this AG.
569 : *
570 : * If the current transaction is holding busy extents, the caller may not want
571 : * to wait for committed busy extents to resolve. If we are being told just to
572 : * try a flush or progress has been made since we last skipped a busy extent,
573 : * return immediately to allow the caller to try again.
574 : *
575 : * If we are freeing extents, we might actually be holding the only free extents
576 : * in the transaction busy list and the log force won't resolve that situation.
577 : * In this case, we must return -EAGAIN to avoid a deadlock by informing the
578 : * caller it needs to commit the busy extents it holds before retrying the
579 : * extent free operation.
580 : */
581 : int
582 45995 : xfs_extent_busy_flush(
583 : struct xfs_trans *tp,
584 : struct xfs_perag *pag,
585 : unsigned busy_gen,
586 : uint32_t alloc_flags)
587 : {
588 45995 : DEFINE_WAIT (wait);
589 45995 : int error;
590 :
591 45995 : error = xfs_log_force(tp->t_mountp, XFS_LOG_SYNC);
592 45994 : if (error)
593 : return error;
594 :
595 : /* Avoid deadlocks on uncommitted busy extents. */
596 45992 : if (!list_empty(&tp->t_busy)) {
597 0 : if (alloc_flags & XFS_ALLOC_FLAG_TRYFLUSH)
598 : return 0;
599 :
600 0 : if (busy_gen != READ_ONCE(pag->pagb_gen))
601 : return 0;
602 :
603 0 : if (alloc_flags & XFS_ALLOC_FLAG_FREEING)
604 : return -EAGAIN;
605 : }
606 :
607 : /* Wait for committed busy extents to resolve. */
608 45994 : do {
609 45993 : prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
610 45994 : if (busy_gen != READ_ONCE(pag->pagb_gen))
611 : break;
612 1 : schedule();
613 : } while (1);
614 :
615 45993 : finish_wait(&pag->pagb_wait, &wait);
616 45993 : return 0;
617 : }
618 :
619 : void
620 22482 : xfs_extent_busy_wait_all(
621 : struct xfs_mount *mp)
622 : {
623 22482 : struct xfs_perag *pag;
624 22482 : DEFINE_WAIT (wait);
625 22482 : xfs_agnumber_t agno;
626 :
627 149348 : for_each_perag(mp, agno, pag) {
628 126870 : do {
629 126868 : prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
630 126868 : if (RB_EMPTY_ROOT(&pag->pagb_tree))
631 : break;
632 2 : schedule();
633 : } while (1);
634 126866 : finish_wait(&pag->pagb_wait, &wait);
635 : }
636 22482 : }
637 :
638 : /*
639 : * Callback for list_sort to sort busy extents by the AG they reside in.
640 : */
641 : int
642 272237425 : xfs_extent_busy_ag_cmp(
643 : void *priv,
644 : const struct list_head *l1,
645 : const struct list_head *l2)
646 : {
647 272237425 : struct xfs_extent_busy *b1 =
648 272237425 : container_of(l1, struct xfs_extent_busy, list);
649 272237425 : struct xfs_extent_busy *b2 =
650 272237425 : container_of(l2, struct xfs_extent_busy, list);
651 272237425 : s32 diff;
652 :
653 272237425 : diff = b1->agno - b2->agno;
654 272237425 : if (!diff)
655 238736866 : diff = b1->bno - b2->bno;
656 272237425 : return diff;
657 : }
|
