xref: /openbmc/linux/fs/xfs/xfs_extent_busy.c (revision 58d47044)
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 static void
23 xfs_extent_busy_insert_list(
24 	struct xfs_perag	*pag,
25 	xfs_agblock_t		bno,
26 	xfs_extlen_t		len,
27 	unsigned int		flags,
28 	struct list_head	*busy_list)
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(pag->pag_mount, 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 	/* always process discard lists in fifo order */
66 	list_add_tail(&new->list, busy_list);
67 	spin_unlock(&pag->pagb_lock);
68 }
69 
70 void
71 xfs_extent_busy_insert(
72 	struct xfs_trans	*tp,
73 	struct xfs_perag	*pag,
74 	xfs_agblock_t		bno,
75 	xfs_extlen_t		len,
76 	unsigned int		flags)
77 {
78 	xfs_extent_busy_insert_list(pag, bno, len, flags, &tp->t_busy);
79 }
80 
81 void
82 xfs_extent_busy_insert_discard(
83 	struct xfs_perag	*pag,
84 	xfs_agblock_t		bno,
85 	xfs_extlen_t		len,
86 	struct list_head	*busy_list)
87 {
88 	xfs_extent_busy_insert_list(pag, bno, len, XFS_EXTENT_BUSY_DISCARDED,
89 			busy_list);
90 }
91 
92 /*
93  * Search for a busy extent within the range of the extent we are about to
94  * allocate.  You need to be holding the busy extent tree lock when calling
95  * xfs_extent_busy_search(). This function returns 0 for no overlapping busy
96  * extent, -1 for an overlapping but not exact busy extent, and 1 for an exact
97  * match. This is done so that a non-zero return indicates an overlap that
98  * will require a synchronous transaction, but it can still be
99  * used to distinguish between a partial or exact match.
100  */
101 int
102 xfs_extent_busy_search(
103 	struct xfs_mount	*mp,
104 	struct xfs_perag	*pag,
105 	xfs_agblock_t		bno,
106 	xfs_extlen_t		len)
107 {
108 	struct rb_node		*rbp;
109 	struct xfs_extent_busy	*busyp;
110 	int			match = 0;
111 
112 	/* find closest start bno overlap */
113 	spin_lock(&pag->pagb_lock);
114 	rbp = pag->pagb_tree.rb_node;
115 	while (rbp) {
116 		busyp = rb_entry(rbp, struct xfs_extent_busy, rb_node);
117 		if (bno < busyp->bno) {
118 			/* may overlap, but exact start block is lower */
119 			if (bno + len > busyp->bno)
120 				match = -1;
121 			rbp = rbp->rb_left;
122 		} else if (bno > busyp->bno) {
123 			/* may overlap, but exact start block is higher */
124 			if (bno < busyp->bno + busyp->length)
125 				match = -1;
126 			rbp = rbp->rb_right;
127 		} else {
128 			/* bno matches busyp, length determines exact match */
129 			match = (busyp->length == len) ? 1 : -1;
130 			break;
131 		}
132 	}
133 	spin_unlock(&pag->pagb_lock);
134 	return match;
135 }
136 
137 /*
138  * The found free extent [fbno, fend] overlaps part or all of the given busy
139  * extent.  If the overlap covers the beginning, the end, or all of the busy
140  * extent, the overlapping portion can be made unbusy and used for the
141  * allocation.  We can't split a busy extent because we can't modify a
142  * transaction/CIL context busy list, but we can update an entry's block
143  * number or length.
144  *
145  * Returns true if the extent can safely be reused, or false if the search
146  * needs to be restarted.
147  */
148 STATIC bool
149 xfs_extent_busy_update_extent(
150 	struct xfs_mount	*mp,
151 	struct xfs_perag	*pag,
152 	struct xfs_extent_busy	*busyp,
153 	xfs_agblock_t		fbno,
154 	xfs_extlen_t		flen,
155 	bool			userdata) __releases(&pag->pagb_lock)
156 					  __acquires(&pag->pagb_lock)
157 {
158 	xfs_agblock_t		fend = fbno + flen;
159 	xfs_agblock_t		bbno = busyp->bno;
160 	xfs_agblock_t		bend = bbno + busyp->length;
161 
162 	/*
163 	 * This extent is currently being discarded.  Give the thread
164 	 * performing the discard a chance to mark the extent unbusy
165 	 * and retry.
166 	 */
167 	if (busyp->flags & XFS_EXTENT_BUSY_DISCARDED) {
168 		spin_unlock(&pag->pagb_lock);
169 		delay(1);
170 		spin_lock(&pag->pagb_lock);
171 		return false;
172 	}
173 
174 	/*
175 	 * If there is a busy extent overlapping a user allocation, we have
176 	 * no choice but to force the log and retry the search.
177 	 *
178 	 * Fortunately this does not happen during normal operation, but
179 	 * only if the filesystem is very low on space and has to dip into
180 	 * the AGFL for normal allocations.
181 	 */
182 	if (userdata)
183 		goto out_force_log;
184 
185 	if (bbno < fbno && bend > fend) {
186 		/*
187 		 * Case 1:
188 		 *    bbno           bend
189 		 *    +BBBBBBBBBBBBBBBBB+
190 		 *        +---------+
191 		 *        fbno   fend
192 		 */
193 
194 		/*
195 		 * We would have to split the busy extent to be able to track
196 		 * it correct, which we cannot do because we would have to
197 		 * modify the list of busy extents attached to the transaction
198 		 * or CIL context, which is immutable.
199 		 *
200 		 * Force out the log to clear the busy extent and retry the
201 		 * search.
202 		 */
203 		goto out_force_log;
204 	} else if (bbno >= fbno && bend <= fend) {
205 		/*
206 		 * Case 2:
207 		 *    bbno           bend
208 		 *    +BBBBBBBBBBBBBBBBB+
209 		 *    +-----------------+
210 		 *    fbno           fend
211 		 *
212 		 * Case 3:
213 		 *    bbno           bend
214 		 *    +BBBBBBBBBBBBBBBBB+
215 		 *    +--------------------------+
216 		 *    fbno                    fend
217 		 *
218 		 * Case 4:
219 		 *             bbno           bend
220 		 *             +BBBBBBBBBBBBBBBBB+
221 		 *    +--------------------------+
222 		 *    fbno                    fend
223 		 *
224 		 * Case 5:
225 		 *             bbno           bend
226 		 *             +BBBBBBBBBBBBBBBBB+
227 		 *    +-----------------------------------+
228 		 *    fbno                             fend
229 		 *
230 		 */
231 
232 		/*
233 		 * The busy extent is fully covered by the extent we are
234 		 * allocating, and can simply be removed from the rbtree.
235 		 * However we cannot remove it from the immutable list
236 		 * tracking busy extents in the transaction or CIL context,
237 		 * so set the length to zero to mark it invalid.
238 		 *
239 		 * We also need to restart the busy extent search from the
240 		 * tree root, because erasing the node can rearrange the
241 		 * tree topology.
242 		 */
243 		rb_erase(&busyp->rb_node, &pag->pagb_tree);
244 		busyp->length = 0;
245 		return false;
246 	} else if (fend < bend) {
247 		/*
248 		 * Case 6:
249 		 *              bbno           bend
250 		 *             +BBBBBBBBBBBBBBBBB+
251 		 *             +---------+
252 		 *             fbno   fend
253 		 *
254 		 * Case 7:
255 		 *             bbno           bend
256 		 *             +BBBBBBBBBBBBBBBBB+
257 		 *    +------------------+
258 		 *    fbno            fend
259 		 *
260 		 */
261 		busyp->bno = fend;
262 		busyp->length = bend - fend;
263 	} else if (bbno < fbno) {
264 		/*
265 		 * Case 8:
266 		 *    bbno           bend
267 		 *    +BBBBBBBBBBBBBBBBB+
268 		 *        +-------------+
269 		 *        fbno       fend
270 		 *
271 		 * Case 9:
272 		 *    bbno           bend
273 		 *    +BBBBBBBBBBBBBBBBB+
274 		 *        +----------------------+
275 		 *        fbno                fend
276 		 */
277 		busyp->length = fbno - busyp->bno;
278 	} else {
279 		ASSERT(0);
280 	}
281 
282 	trace_xfs_extent_busy_reuse(mp, pag->pag_agno, fbno, flen);
283 	return true;
284 
285 out_force_log:
286 	spin_unlock(&pag->pagb_lock);
287 	xfs_log_force(mp, XFS_LOG_SYNC);
288 	trace_xfs_extent_busy_force(mp, pag->pag_agno, fbno, flen);
289 	spin_lock(&pag->pagb_lock);
290 	return false;
291 }
292 
293 
294 /*
295  * For a given extent [fbno, flen], make sure we can reuse it safely.
296  */
297 void
298 xfs_extent_busy_reuse(
299 	struct xfs_mount	*mp,
300 	struct xfs_perag	*pag,
301 	xfs_agblock_t		fbno,
302 	xfs_extlen_t		flen,
303 	bool			userdata)
304 {
305 	struct rb_node		*rbp;
306 
307 	ASSERT(flen > 0);
308 	spin_lock(&pag->pagb_lock);
309 restart:
310 	rbp = pag->pagb_tree.rb_node;
311 	while (rbp) {
312 		struct xfs_extent_busy *busyp =
313 			rb_entry(rbp, struct xfs_extent_busy, rb_node);
314 		xfs_agblock_t	bbno = busyp->bno;
315 		xfs_agblock_t	bend = bbno + busyp->length;
316 
317 		if (fbno + flen <= bbno) {
318 			rbp = rbp->rb_left;
319 			continue;
320 		} else if (fbno >= bend) {
321 			rbp = rbp->rb_right;
322 			continue;
323 		}
324 
325 		if (!xfs_extent_busy_update_extent(mp, pag, busyp, fbno, flen,
326 						  userdata))
327 			goto restart;
328 	}
329 	spin_unlock(&pag->pagb_lock);
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  */
344 bool
345 xfs_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);
359 	fbno = *bno;
360 	flen = *len;
361 	rbp = args->pag->pagb_tree.rb_node;
362 	while (rbp && flen >= args->minlen) {
363 		struct xfs_extent_busy *busyp =
364 			rb_entry(rbp, struct xfs_extent_busy, rb_node);
365 		xfs_agblock_t	fend = fbno + flen;
366 		xfs_agblock_t	bbno = busyp->bno;
367 		xfs_agblock_t	bend = bbno + busyp->length;
368 
369 		if (fend <= bbno) {
370 			rbp = rbp->rb_left;
371 			continue;
372 		} else if (fbno >= bend) {
373 			rbp = rbp->rb_right;
374 			continue;
375 		}
376 
377 		if (bbno <= fbno) {
378 			/* start overlap */
379 
380 			/*
381 			 * Case 1:
382 			 *    bbno           bend
383 			 *    +BBBBBBBBBBBBBBBBB+
384 			 *        +---------+
385 			 *        fbno   fend
386 			 *
387 			 * Case 2:
388 			 *    bbno           bend
389 			 *    +BBBBBBBBBBBBBBBBB+
390 			 *    +-------------+
391 			 *    fbno       fend
392 			 *
393 			 * Case 3:
394 			 *    bbno           bend
395 			 *    +BBBBBBBBBBBBBBBBB+
396 			 *        +-------------+
397 			 *        fbno       fend
398 			 *
399 			 * Case 4:
400 			 *    bbno           bend
401 			 *    +BBBBBBBBBBBBBBBBB+
402 			 *    +-----------------+
403 			 *    fbno           fend
404 			 *
405 			 * No unbusy region in extent, return failure.
406 			 */
407 			if (fend <= bend)
408 				goto fail;
409 
410 			/*
411 			 * Case 5:
412 			 *    bbno           bend
413 			 *    +BBBBBBBBBBBBBBBBB+
414 			 *        +----------------------+
415 			 *        fbno                fend
416 			 *
417 			 * Case 6:
418 			 *    bbno           bend
419 			 *    +BBBBBBBBBBBBBBBBB+
420 			 *    +--------------------------+
421 			 *    fbno                    fend
422 			 *
423 			 * Needs to be trimmed to:
424 			 *                       +-------+
425 			 *                       fbno fend
426 			 */
427 			fbno = bend;
428 		} else if (bend >= fend) {
429 			/* end overlap */
430 
431 			/*
432 			 * Case 7:
433 			 *             bbno           bend
434 			 *             +BBBBBBBBBBBBBBBBB+
435 			 *    +------------------+
436 			 *    fbno            fend
437 			 *
438 			 * Case 8:
439 			 *             bbno           bend
440 			 *             +BBBBBBBBBBBBBBBBB+
441 			 *    +--------------------------+
442 			 *    fbno                    fend
443 			 *
444 			 * Needs to be trimmed to:
445 			 *    +-------+
446 			 *    fbno fend
447 			 */
448 			fend = bbno;
449 		} else {
450 			/* middle overlap */
451 
452 			/*
453 			 * Case 9:
454 			 *             bbno           bend
455 			 *             +BBBBBBBBBBBBBBBBB+
456 			 *    +-----------------------------------+
457 			 *    fbno                             fend
458 			 *
459 			 * Can be trimmed to:
460 			 *    +-------+        OR         +-------+
461 			 *    fbno fend                   fbno fend
462 			 *
463 			 * Backward allocation leads to significant
464 			 * fragmentation of directories, which degrades
465 			 * directory performance, therefore we always want to
466 			 * choose the option that produces forward allocation
467 			 * patterns.
468 			 * Preferring the lower bno extent will make the next
469 			 * request use "fend" as the start of the next
470 			 * allocation;  if the segment is no longer busy at
471 			 * that point, we'll get a contiguous allocation, but
472 			 * even if it is still busy, we will get a forward
473 			 * allocation.
474 			 * We try to avoid choosing the segment at "bend",
475 			 * because that can lead to the next allocation
476 			 * taking the segment at "fbno", which would be a
477 			 * backward allocation.  We only use the segment at
478 			 * "fbno" if it is much larger than the current
479 			 * requested size, because in that case there's a
480 			 * good chance subsequent allocations will be
481 			 * contiguous.
482 			 */
483 			if (bbno - fbno >= args->maxlen) {
484 				/* left candidate fits perfect */
485 				fend = bbno;
486 			} else if (fend - bend >= args->maxlen * 4) {
487 				/* right candidate has enough free space */
488 				fbno = bend;
489 			} else if (bbno - fbno >= args->minlen) {
490 				/* left candidate fits minimum requirement */
491 				fend = bbno;
492 			} else {
493 				goto fail;
494 			}
495 		}
496 
497 		flen = fend - fbno;
498 	}
499 out:
500 
501 	if (fbno != *bno || flen != *len) {
502 		trace_xfs_extent_busy_trim(args->mp, args->agno, *bno, *len,
503 					  fbno, flen);
504 		*bno = fbno;
505 		*len = flen;
506 		*busy_gen = args->pag->pagb_gen;
507 		ret = true;
508 	}
509 	spin_unlock(&args->pag->pagb_lock);
510 	return ret;
511 fail:
512 	/*
513 	 * Return a zero extent length as failure indications.  All callers
514 	 * re-check if the trimmed extent satisfies the minlen requirement.
515 	 */
516 	flen = 0;
517 	goto out;
518 }
519 
520 STATIC void
521 xfs_extent_busy_clear_one(
522 	struct xfs_mount	*mp,
523 	struct xfs_perag	*pag,
524 	struct xfs_extent_busy	*busyp)
525 {
526 	if (busyp->length) {
527 		trace_xfs_extent_busy_clear(mp, busyp->agno, busyp->bno,
528 						busyp->length);
529 		rb_erase(&busyp->rb_node, &pag->pagb_tree);
530 	}
531 
532 	list_del_init(&busyp->list);
533 	kmem_free(busyp);
534 }
535 
536 static void
537 xfs_extent_busy_put_pag(
538 	struct xfs_perag	*pag,
539 	bool			wakeup)
540 		__releases(pag->pagb_lock)
541 {
542 	if (wakeup) {
543 		pag->pagb_gen++;
544 		wake_up_all(&pag->pagb_wait);
545 	}
546 
547 	spin_unlock(&pag->pagb_lock);
548 	xfs_perag_put(pag);
549 }
550 
551 /*
552  * Remove all extents on the passed in list from the busy extents tree.
553  * If do_discard is set skip extents that need to be discarded, and mark
554  * these as undergoing a discard operation instead.
555  */
556 void
557 xfs_extent_busy_clear(
558 	struct xfs_mount	*mp,
559 	struct list_head	*list,
560 	bool			do_discard)
561 {
562 	struct xfs_extent_busy	*busyp, *n;
563 	struct xfs_perag	*pag = NULL;
564 	xfs_agnumber_t		agno = NULLAGNUMBER;
565 	bool			wakeup = false;
566 
567 	list_for_each_entry_safe(busyp, n, list, list) {
568 		if (busyp->agno != agno) {
569 			if (pag)
570 				xfs_extent_busy_put_pag(pag, wakeup);
571 			agno = busyp->agno;
572 			pag = xfs_perag_get(mp, agno);
573 			spin_lock(&pag->pagb_lock);
574 			wakeup = false;
575 		}
576 
577 		if (do_discard && busyp->length &&
578 		    !(busyp->flags & XFS_EXTENT_BUSY_SKIP_DISCARD)) {
579 			busyp->flags = XFS_EXTENT_BUSY_DISCARDED;
580 		} else {
581 			xfs_extent_busy_clear_one(mp, pag, busyp);
582 			wakeup = true;
583 		}
584 	}
585 
586 	if (pag)
587 		xfs_extent_busy_put_pag(pag, wakeup);
588 }
589 
590 /*
591  * Flush out all busy extents for this AG.
592  *
593  * If the current transaction is holding busy extents, the caller may not want
594  * to wait for committed busy extents to resolve. If we are being told just to
595  * try a flush or progress has been made since we last skipped a busy extent,
596  * return immediately to allow the caller to try again.
597  *
598  * If we are freeing extents, we might actually be holding the only free extents
599  * in the transaction busy list and the log force won't resolve that situation.
600  * In this case, we must return -EAGAIN to avoid a deadlock by informing the
601  * caller it needs to commit the busy extents it holds before retrying the
602  * extent free operation.
603  */
604 int
605 xfs_extent_busy_flush(
606 	struct xfs_trans	*tp,
607 	struct xfs_perag	*pag,
608 	unsigned		busy_gen,
609 	uint32_t		alloc_flags)
610 {
611 	DEFINE_WAIT		(wait);
612 	int			error;
613 
614 	error = xfs_log_force(tp->t_mountp, XFS_LOG_SYNC);
615 	if (error)
616 		return error;
617 
618 	/* Avoid deadlocks on uncommitted busy extents. */
619 	if (!list_empty(&tp->t_busy)) {
620 		if (alloc_flags & XFS_ALLOC_FLAG_TRYFLUSH)
621 			return 0;
622 
623 		if (busy_gen != READ_ONCE(pag->pagb_gen))
624 			return 0;
625 
626 		if (alloc_flags & XFS_ALLOC_FLAG_FREEING)
627 			return -EAGAIN;
628 	}
629 
630 	/* Wait for committed busy extents to resolve. */
631 	do {
632 		prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
633 		if  (busy_gen != READ_ONCE(pag->pagb_gen))
634 			break;
635 		schedule();
636 	} while (1);
637 
638 	finish_wait(&pag->pagb_wait, &wait);
639 	return 0;
640 }
641 
642 void
643 xfs_extent_busy_wait_all(
644 	struct xfs_mount	*mp)
645 {
646 	struct xfs_perag	*pag;
647 	DEFINE_WAIT		(wait);
648 	xfs_agnumber_t		agno;
649 
650 	for_each_perag(mp, agno, pag) {
651 		do {
652 			prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
653 			if  (RB_EMPTY_ROOT(&pag->pagb_tree))
654 				break;
655 			schedule();
656 		} while (1);
657 		finish_wait(&pag->pagb_wait, &wait);
658 	}
659 }
660 
661 /*
662  * Callback for list_sort to sort busy extents by the AG they reside in.
663  */
664 int
665 xfs_extent_busy_ag_cmp(
666 	void			*priv,
667 	const struct list_head	*l1,
668 	const struct list_head	*l2)
669 {
670 	struct xfs_extent_busy	*b1 =
671 		container_of(l1, struct xfs_extent_busy, list);
672 	struct xfs_extent_busy	*b2 =
673 		container_of(l2, struct xfs_extent_busy, list);
674 	s32 diff;
675 
676 	diff = b1->agno - b2->agno;
677 	if (!diff)
678 		diff = b1->bno - b2->bno;
679 	return diff;
680 }
681