xref: /linux/fs/xfs/xfs_extent_busy.c (revision 37744feebc086908fd89760650f458ab19071750)
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
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
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
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
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
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
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
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
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
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
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
644 xfs_extent_busy_ag_cmp(
645 	void			*priv,
646 	struct list_head	*l1,
647 	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