xref: /linux/fs/xfs/xfs_extent_busy.c (revision c0c914eca7f251c70facc37dfebeaf176601918d)
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 
34 void
35 xfs_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_MAYFAIL);
49 	if (!new) {
50 		/*
51 		 * No Memory!  Since it is now not possible to track the free
52 		 * block, make this a synchronous transaction to insure that
53 		 * the block is not reused before this transaction commits.
54 		 */
55 		trace_xfs_extent_busy_enomem(tp->t_mountp, agno, bno, len);
56 		xfs_trans_set_sync(tp);
57 		return;
58 	}
59 
60 	new->agno = agno;
61 	new->bno = bno;
62 	new->length = len;
63 	INIT_LIST_HEAD(&new->list);
64 	new->flags = flags;
65 
66 	/* trace before insert to be able to see failed inserts */
67 	trace_xfs_extent_busy(tp->t_mountp, agno, bno, len);
68 
69 	pag = xfs_perag_get(tp->t_mountp, new->agno);
70 	spin_lock(&pag->pagb_lock);
71 	rbp = &pag->pagb_tree.rb_node;
72 	while (*rbp) {
73 		parent = *rbp;
74 		busyp = rb_entry(parent, struct xfs_extent_busy, rb_node);
75 
76 		if (new->bno < busyp->bno) {
77 			rbp = &(*rbp)->rb_left;
78 			ASSERT(new->bno + new->length <= busyp->bno);
79 		} else if (new->bno > busyp->bno) {
80 			rbp = &(*rbp)->rb_right;
81 			ASSERT(bno >= busyp->bno + busyp->length);
82 		} else {
83 			ASSERT(0);
84 		}
85 	}
86 
87 	rb_link_node(&new->rb_node, parent, rbp);
88 	rb_insert_color(&new->rb_node, &pag->pagb_tree);
89 
90 	list_add(&new->list, &tp->t_busy);
91 	spin_unlock(&pag->pagb_lock);
92 	xfs_perag_put(pag);
93 }
94 
95 /*
96  * Search for a busy extent within the range of the extent we are about to
97  * allocate.  You need to be holding the busy extent tree lock when calling
98  * xfs_extent_busy_search(). This function returns 0 for no overlapping busy
99  * extent, -1 for an overlapping but not exact busy extent, and 1 for an exact
100  * match. This is done so that a non-zero return indicates an overlap that
101  * will require a synchronous transaction, but it can still be
102  * used to distinguish between a partial or exact match.
103  */
104 int
105 xfs_extent_busy_search(
106 	struct xfs_mount	*mp,
107 	xfs_agnumber_t		agno,
108 	xfs_agblock_t		bno,
109 	xfs_extlen_t		len)
110 {
111 	struct xfs_perag	*pag;
112 	struct rb_node		*rbp;
113 	struct xfs_extent_busy	*busyp;
114 	int			match = 0;
115 
116 	pag = xfs_perag_get(mp, agno);
117 	spin_lock(&pag->pagb_lock);
118 
119 	rbp = pag->pagb_tree.rb_node;
120 
121 	/* find closest start bno overlap */
122 	while (rbp) {
123 		busyp = rb_entry(rbp, struct xfs_extent_busy, rb_node);
124 		if (bno < busyp->bno) {
125 			/* may overlap, but exact start block is lower */
126 			if (bno + len > busyp->bno)
127 				match = -1;
128 			rbp = rbp->rb_left;
129 		} else if (bno > busyp->bno) {
130 			/* may overlap, but exact start block is higher */
131 			if (bno < busyp->bno + busyp->length)
132 				match = -1;
133 			rbp = rbp->rb_right;
134 		} else {
135 			/* bno matches busyp, length determines exact match */
136 			match = (busyp->length == len) ? 1 : -1;
137 			break;
138 		}
139 	}
140 	spin_unlock(&pag->pagb_lock);
141 	xfs_perag_put(pag);
142 	return match;
143 }
144 
145 /*
146  * The found free extent [fbno, fend] overlaps part or all of the given busy
147  * extent.  If the overlap covers the beginning, the end, or all of the busy
148  * extent, the overlapping portion can be made unbusy and used for the
149  * allocation.  We can't split a busy extent because we can't modify a
150  * transaction/CIL context busy list, but we can update an entry's block
151  * number or length.
152  *
153  * Returns true if the extent can safely be reused, or false if the search
154  * needs to be restarted.
155  */
156 STATIC bool
157 xfs_extent_busy_update_extent(
158 	struct xfs_mount	*mp,
159 	struct xfs_perag	*pag,
160 	struct xfs_extent_busy	*busyp,
161 	xfs_agblock_t		fbno,
162 	xfs_extlen_t		flen,
163 	bool			userdata) __releases(&pag->pagb_lock)
164 					  __acquires(&pag->pagb_lock)
165 {
166 	xfs_agblock_t		fend = fbno + flen;
167 	xfs_agblock_t		bbno = busyp->bno;
168 	xfs_agblock_t		bend = bbno + busyp->length;
169 
170 	/*
171 	 * This extent is currently being discarded.  Give the thread
172 	 * performing the discard a chance to mark the extent unbusy
173 	 * and retry.
174 	 */
175 	if (busyp->flags & XFS_EXTENT_BUSY_DISCARDED) {
176 		spin_unlock(&pag->pagb_lock);
177 		delay(1);
178 		spin_lock(&pag->pagb_lock);
179 		return false;
180 	}
181 
182 	/*
183 	 * If there is a busy extent overlapping a user allocation, we have
184 	 * no choice but to force the log and retry the search.
185 	 *
186 	 * Fortunately this does not happen during normal operation, but
187 	 * only if the filesystem is very low on space and has to dip into
188 	 * the AGFL for normal allocations.
189 	 */
190 	if (userdata)
191 		goto out_force_log;
192 
193 	if (bbno < fbno && bend > fend) {
194 		/*
195 		 * Case 1:
196 		 *    bbno           bend
197 		 *    +BBBBBBBBBBBBBBBBB+
198 		 *        +---------+
199 		 *        fbno   fend
200 		 */
201 
202 		/*
203 		 * We would have to split the busy extent to be able to track
204 		 * it correct, which we cannot do because we would have to
205 		 * modify the list of busy extents attached to the transaction
206 		 * or CIL context, which is immutable.
207 		 *
208 		 * Force out the log to clear the busy extent and retry the
209 		 * search.
210 		 */
211 		goto out_force_log;
212 	} else if (bbno >= fbno && bend <= fend) {
213 		/*
214 		 * Case 2:
215 		 *    bbno           bend
216 		 *    +BBBBBBBBBBBBBBBBB+
217 		 *    +-----------------+
218 		 *    fbno           fend
219 		 *
220 		 * Case 3:
221 		 *    bbno           bend
222 		 *    +BBBBBBBBBBBBBBBBB+
223 		 *    +--------------------------+
224 		 *    fbno                    fend
225 		 *
226 		 * Case 4:
227 		 *             bbno           bend
228 		 *             +BBBBBBBBBBBBBBBBB+
229 		 *    +--------------------------+
230 		 *    fbno                    fend
231 		 *
232 		 * Case 5:
233 		 *             bbno           bend
234 		 *             +BBBBBBBBBBBBBBBBB+
235 		 *    +-----------------------------------+
236 		 *    fbno                             fend
237 		 *
238 		 */
239 
240 		/*
241 		 * The busy extent is fully covered by the extent we are
242 		 * allocating, and can simply be removed from the rbtree.
243 		 * However we cannot remove it from the immutable list
244 		 * tracking busy extents in the transaction or CIL context,
245 		 * so set the length to zero to mark it invalid.
246 		 *
247 		 * We also need to restart the busy extent search from the
248 		 * tree root, because erasing the node can rearrange the
249 		 * tree topology.
250 		 */
251 		rb_erase(&busyp->rb_node, &pag->pagb_tree);
252 		busyp->length = 0;
253 		return false;
254 	} else if (fend < bend) {
255 		/*
256 		 * Case 6:
257 		 *              bbno           bend
258 		 *             +BBBBBBBBBBBBBBBBB+
259 		 *             +---------+
260 		 *             fbno   fend
261 		 *
262 		 * Case 7:
263 		 *             bbno           bend
264 		 *             +BBBBBBBBBBBBBBBBB+
265 		 *    +------------------+
266 		 *    fbno            fend
267 		 *
268 		 */
269 		busyp->bno = fend;
270 	} else if (bbno < fbno) {
271 		/*
272 		 * Case 8:
273 		 *    bbno           bend
274 		 *    +BBBBBBBBBBBBBBBBB+
275 		 *        +-------------+
276 		 *        fbno       fend
277 		 *
278 		 * Case 9:
279 		 *    bbno           bend
280 		 *    +BBBBBBBBBBBBBBBBB+
281 		 *        +----------------------+
282 		 *        fbno                fend
283 		 */
284 		busyp->length = fbno - busyp->bno;
285 	} else {
286 		ASSERT(0);
287 	}
288 
289 	trace_xfs_extent_busy_reuse(mp, pag->pag_agno, fbno, flen);
290 	return true;
291 
292 out_force_log:
293 	spin_unlock(&pag->pagb_lock);
294 	xfs_log_force(mp, XFS_LOG_SYNC);
295 	trace_xfs_extent_busy_force(mp, pag->pag_agno, fbno, flen);
296 	spin_lock(&pag->pagb_lock);
297 	return false;
298 }
299 
300 
301 /*
302  * For a given extent [fbno, flen], make sure we can reuse it safely.
303  */
304 void
305 xfs_extent_busy_reuse(
306 	struct xfs_mount	*mp,
307 	xfs_agnumber_t		agno,
308 	xfs_agblock_t		fbno,
309 	xfs_extlen_t		flen,
310 	bool			userdata)
311 {
312 	struct xfs_perag	*pag;
313 	struct rb_node		*rbp;
314 
315 	ASSERT(flen > 0);
316 
317 	pag = xfs_perag_get(mp, agno);
318 	spin_lock(&pag->pagb_lock);
319 restart:
320 	rbp = pag->pagb_tree.rb_node;
321 	while (rbp) {
322 		struct xfs_extent_busy *busyp =
323 			rb_entry(rbp, struct xfs_extent_busy, rb_node);
324 		xfs_agblock_t	bbno = busyp->bno;
325 		xfs_agblock_t	bend = bbno + busyp->length;
326 
327 		if (fbno + flen <= bbno) {
328 			rbp = rbp->rb_left;
329 			continue;
330 		} else if (fbno >= bend) {
331 			rbp = rbp->rb_right;
332 			continue;
333 		}
334 
335 		if (!xfs_extent_busy_update_extent(mp, pag, busyp, fbno, flen,
336 						  userdata))
337 			goto restart;
338 	}
339 	spin_unlock(&pag->pagb_lock);
340 	xfs_perag_put(pag);
341 }
342 
343 /*
344  * For a given extent [fbno, flen], search the busy extent list to find a
345  * subset of the extent that is not busy.  If *rlen is smaller than
346  * args->minlen no suitable extent could be found, and the higher level
347  * code needs to force out the log and retry the allocation.
348  */
349 void
350 xfs_extent_busy_trim(
351 	struct xfs_alloc_arg	*args,
352 	xfs_agblock_t		bno,
353 	xfs_extlen_t		len,
354 	xfs_agblock_t		*rbno,
355 	xfs_extlen_t		*rlen)
356 {
357 	xfs_agblock_t		fbno;
358 	xfs_extlen_t		flen;
359 	struct rb_node		*rbp;
360 
361 	ASSERT(len > 0);
362 
363 	spin_lock(&args->pag->pagb_lock);
364 restart:
365 	fbno = bno;
366 	flen = len;
367 	rbp = args->pag->pagb_tree.rb_node;
368 	while (rbp && flen >= args->minlen) {
369 		struct xfs_extent_busy *busyp =
370 			rb_entry(rbp, struct xfs_extent_busy, rb_node);
371 		xfs_agblock_t	fend = fbno + flen;
372 		xfs_agblock_t	bbno = busyp->bno;
373 		xfs_agblock_t	bend = bbno + busyp->length;
374 
375 		if (fend <= bbno) {
376 			rbp = rbp->rb_left;
377 			continue;
378 		} else if (fbno >= bend) {
379 			rbp = rbp->rb_right;
380 			continue;
381 		}
382 
383 		/*
384 		 * If this is a metadata allocation, try to reuse the busy
385 		 * extent instead of trimming the allocation.
386 		 */
387 		if (!args->userdata &&
388 		    !(busyp->flags & XFS_EXTENT_BUSY_DISCARDED)) {
389 			if (!xfs_extent_busy_update_extent(args->mp, args->pag,
390 							  busyp, fbno, flen,
391 							  false))
392 				goto restart;
393 			continue;
394 		}
395 
396 		if (bbno <= fbno) {
397 			/* start overlap */
398 
399 			/*
400 			 * Case 1:
401 			 *    bbno           bend
402 			 *    +BBBBBBBBBBBBBBBBB+
403 			 *        +---------+
404 			 *        fbno   fend
405 			 *
406 			 * Case 2:
407 			 *    bbno           bend
408 			 *    +BBBBBBBBBBBBBBBBB+
409 			 *    +-------------+
410 			 *    fbno       fend
411 			 *
412 			 * Case 3:
413 			 *    bbno           bend
414 			 *    +BBBBBBBBBBBBBBBBB+
415 			 *        +-------------+
416 			 *        fbno       fend
417 			 *
418 			 * Case 4:
419 			 *    bbno           bend
420 			 *    +BBBBBBBBBBBBBBBBB+
421 			 *    +-----------------+
422 			 *    fbno           fend
423 			 *
424 			 * No unbusy region in extent, return failure.
425 			 */
426 			if (fend <= bend)
427 				goto fail;
428 
429 			/*
430 			 * Case 5:
431 			 *    bbno           bend
432 			 *    +BBBBBBBBBBBBBBBBB+
433 			 *        +----------------------+
434 			 *        fbno                fend
435 			 *
436 			 * Case 6:
437 			 *    bbno           bend
438 			 *    +BBBBBBBBBBBBBBBBB+
439 			 *    +--------------------------+
440 			 *    fbno                    fend
441 			 *
442 			 * Needs to be trimmed to:
443 			 *                       +-------+
444 			 *                       fbno fend
445 			 */
446 			fbno = bend;
447 		} else if (bend >= fend) {
448 			/* end overlap */
449 
450 			/*
451 			 * Case 7:
452 			 *             bbno           bend
453 			 *             +BBBBBBBBBBBBBBBBB+
454 			 *    +------------------+
455 			 *    fbno            fend
456 			 *
457 			 * Case 8:
458 			 *             bbno           bend
459 			 *             +BBBBBBBBBBBBBBBBB+
460 			 *    +--------------------------+
461 			 *    fbno                    fend
462 			 *
463 			 * Needs to be trimmed to:
464 			 *    +-------+
465 			 *    fbno fend
466 			 */
467 			fend = bbno;
468 		} else {
469 			/* middle overlap */
470 
471 			/*
472 			 * Case 9:
473 			 *             bbno           bend
474 			 *             +BBBBBBBBBBBBBBBBB+
475 			 *    +-----------------------------------+
476 			 *    fbno                             fend
477 			 *
478 			 * Can be trimmed to:
479 			 *    +-------+        OR         +-------+
480 			 *    fbno fend                   fbno fend
481 			 *
482 			 * Backward allocation leads to significant
483 			 * fragmentation of directories, which degrades
484 			 * directory performance, therefore we always want to
485 			 * choose the option that produces forward allocation
486 			 * patterns.
487 			 * Preferring the lower bno extent will make the next
488 			 * request use "fend" as the start of the next
489 			 * allocation;  if the segment is no longer busy at
490 			 * that point, we'll get a contiguous allocation, but
491 			 * even if it is still busy, we will get a forward
492 			 * allocation.
493 			 * We try to avoid choosing the segment at "bend",
494 			 * because that can lead to the next allocation
495 			 * taking the segment at "fbno", which would be a
496 			 * backward allocation.  We only use the segment at
497 			 * "fbno" if it is much larger than the current
498 			 * requested size, because in that case there's a
499 			 * good chance subsequent allocations will be
500 			 * contiguous.
501 			 */
502 			if (bbno - fbno >= args->maxlen) {
503 				/* left candidate fits perfect */
504 				fend = bbno;
505 			} else if (fend - bend >= args->maxlen * 4) {
506 				/* right candidate has enough free space */
507 				fbno = bend;
508 			} else if (bbno - fbno >= args->minlen) {
509 				/* left candidate fits minimum requirement */
510 				fend = bbno;
511 			} else {
512 				goto fail;
513 			}
514 		}
515 
516 		flen = fend - fbno;
517 	}
518 	spin_unlock(&args->pag->pagb_lock);
519 
520 	if (fbno != bno || flen != len) {
521 		trace_xfs_extent_busy_trim(args->mp, args->agno, bno, len,
522 					  fbno, flen);
523 	}
524 	*rbno = fbno;
525 	*rlen = flen;
526 	return;
527 fail:
528 	/*
529 	 * Return a zero extent length as failure indications.  All callers
530 	 * re-check if the trimmed extent satisfies the minlen requirement.
531 	 */
532 	spin_unlock(&args->pag->pagb_lock);
533 	trace_xfs_extent_busy_trim(args->mp, args->agno, bno, len, fbno, 0);
534 	*rbno = fbno;
535 	*rlen = 0;
536 }
537 
538 STATIC void
539 xfs_extent_busy_clear_one(
540 	struct xfs_mount	*mp,
541 	struct xfs_perag	*pag,
542 	struct xfs_extent_busy	*busyp)
543 {
544 	if (busyp->length) {
545 		trace_xfs_extent_busy_clear(mp, busyp->agno, busyp->bno,
546 						busyp->length);
547 		rb_erase(&busyp->rb_node, &pag->pagb_tree);
548 	}
549 
550 	list_del_init(&busyp->list);
551 	kmem_free(busyp);
552 }
553 
554 /*
555  * Remove all extents on the passed in list from the busy extents tree.
556  * If do_discard is set skip extents that need to be discarded, and mark
557  * these as undergoing a discard operation instead.
558  */
559 void
560 xfs_extent_busy_clear(
561 	struct xfs_mount	*mp,
562 	struct list_head	*list,
563 	bool			do_discard)
564 {
565 	struct xfs_extent_busy	*busyp, *n;
566 	struct xfs_perag	*pag = NULL;
567 	xfs_agnumber_t		agno = NULLAGNUMBER;
568 
569 	list_for_each_entry_safe(busyp, n, list, list) {
570 		if (busyp->agno != agno) {
571 			if (pag) {
572 				spin_unlock(&pag->pagb_lock);
573 				xfs_perag_put(pag);
574 			}
575 			pag = xfs_perag_get(mp, busyp->agno);
576 			spin_lock(&pag->pagb_lock);
577 			agno = busyp->agno;
578 		}
579 
580 		if (do_discard && busyp->length &&
581 		    !(busyp->flags & XFS_EXTENT_BUSY_SKIP_DISCARD))
582 			busyp->flags = XFS_EXTENT_BUSY_DISCARDED;
583 		else
584 			xfs_extent_busy_clear_one(mp, pag, busyp);
585 	}
586 
587 	if (pag) {
588 		spin_unlock(&pag->pagb_lock);
589 		xfs_perag_put(pag);
590 	}
591 }
592 
593 /*
594  * Callback for list_sort to sort busy extents by the AG they reside in.
595  */
596 int
597 xfs_extent_busy_ag_cmp(
598 	void			*priv,
599 	struct list_head	*a,
600 	struct list_head	*b)
601 {
602 	return container_of(a, struct xfs_extent_busy, list)->agno -
603 		container_of(b, struct xfs_extent_busy, list)->agno;
604 }
605