xref: /linux/fs/xfs/scrub/fscounters.c (revision f2d14bc4e437b8ed21e6890ae047a6ec47c030d9)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * Copyright (C) 2019-2023 Oracle.  All Rights Reserved.
4  * Author: Darrick J. Wong <djwong@kernel.org>
5  */
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_trans_resv.h"
11 #include "xfs_mount.h"
12 #include "xfs_alloc.h"
13 #include "xfs_ialloc.h"
14 #include "xfs_health.h"
15 #include "xfs_btree.h"
16 #include "xfs_ag.h"
17 #include "xfs_rtalloc.h"
18 #include "xfs_inode.h"
19 #include "scrub/scrub.h"
20 #include "scrub/common.h"
21 #include "scrub/trace.h"
22 
23 /*
24  * FS Summary Counters
25  * ===================
26  *
27  * The basics of filesystem summary counter checking are that we iterate the
28  * AGs counting the number of free blocks, free space btree blocks, per-AG
29  * reservations, inodes, delayed allocation reservations, and free inodes.
30  * Then we compare what we computed against the in-core counters.
31  *
32  * However, the reality is that summary counters are a tricky beast to check.
33  * While we /could/ freeze the filesystem and scramble around the AGs counting
34  * the free blocks, in practice we prefer not do that for a scan because
35  * freezing is costly.  To get around this, we added a per-cpu counter of the
36  * delalloc reservations so that we can rotor around the AGs relatively
37  * quickly, and we allow the counts to be slightly off because we're not taking
38  * any locks while we do this.
39  *
40  * So the first thing we do is warm up the buffer cache in the setup routine by
41  * walking all the AGs to make sure the incore per-AG structure has been
42  * initialized.  The expected value calculation then iterates the incore per-AG
43  * structures as quickly as it can.  We snapshot the percpu counters before and
44  * after this operation and use the difference in counter values to guess at
45  * our tolerance for mismatch between expected and actual counter values.
46  */
47 
48 struct xchk_fscounters {
49 	struct xfs_scrub	*sc;
50 	uint64_t		icount;
51 	uint64_t		ifree;
52 	uint64_t		fdblocks;
53 	uint64_t		frextents;
54 	unsigned long long	icount_min;
55 	unsigned long long	icount_max;
56 };
57 
58 /*
59  * Since the expected value computation is lockless but only browses incore
60  * values, the percpu counters should be fairly close to each other.  However,
61  * we'll allow ourselves to be off by at least this (arbitrary) amount.
62  */
63 #define XCHK_FSCOUNT_MIN_VARIANCE	(512)
64 
65 /*
66  * Make sure the per-AG structure has been initialized from the on-disk header
67  * contents and trust that the incore counters match the ondisk counters.  (The
68  * AGF and AGI scrubbers check them, and a normal xfs_scrub run checks the
69  * summary counters after checking all AG headers).  Do this from the setup
70  * function so that the inner AG aggregation loop runs as quickly as possible.
71  *
72  * This function runs during the setup phase /before/ we start checking any
73  * metadata.
74  */
75 STATIC int
76 xchk_fscount_warmup(
77 	struct xfs_scrub	*sc)
78 {
79 	struct xfs_mount	*mp = sc->mp;
80 	struct xfs_buf		*agi_bp = NULL;
81 	struct xfs_buf		*agf_bp = NULL;
82 	struct xfs_perag	*pag = NULL;
83 	xfs_agnumber_t		agno;
84 	int			error = 0;
85 
86 	for_each_perag(mp, agno, pag) {
87 		if (xchk_should_terminate(sc, &error))
88 			break;
89 		if (xfs_perag_initialised_agi(pag) &&
90 		    xfs_perag_initialised_agf(pag))
91 			continue;
92 
93 		/* Lock both AG headers. */
94 		error = xfs_ialloc_read_agi(pag, sc->tp, &agi_bp);
95 		if (error)
96 			break;
97 		error = xfs_alloc_read_agf(pag, sc->tp, 0, &agf_bp);
98 		if (error)
99 			break;
100 
101 		/*
102 		 * These are supposed to be initialized by the header read
103 		 * function.
104 		 */
105 		if (!xfs_perag_initialised_agi(pag) ||
106 		    !xfs_perag_initialised_agf(pag)) {
107 			error = -EFSCORRUPTED;
108 			break;
109 		}
110 
111 		xfs_buf_relse(agf_bp);
112 		agf_bp = NULL;
113 		xfs_buf_relse(agi_bp);
114 		agi_bp = NULL;
115 	}
116 
117 	if (agf_bp)
118 		xfs_buf_relse(agf_bp);
119 	if (agi_bp)
120 		xfs_buf_relse(agi_bp);
121 	if (pag)
122 		xfs_perag_rele(pag);
123 	return error;
124 }
125 
126 int
127 xchk_setup_fscounters(
128 	struct xfs_scrub	*sc)
129 {
130 	struct xchk_fscounters	*fsc;
131 	int			error;
132 
133 	/*
134 	 * If the AGF doesn't track btreeblks, we have to lock the AGF to count
135 	 * btree block usage by walking the actual btrees.
136 	 */
137 	if (!xfs_has_lazysbcount(sc->mp))
138 		xchk_fsgates_enable(sc, XCHK_FSGATES_DRAIN);
139 
140 	sc->buf = kzalloc(sizeof(struct xchk_fscounters), XCHK_GFP_FLAGS);
141 	if (!sc->buf)
142 		return -ENOMEM;
143 	fsc = sc->buf;
144 	fsc->sc = sc;
145 
146 	xfs_icount_range(sc->mp, &fsc->icount_min, &fsc->icount_max);
147 
148 	/* We must get the incore counters set up before we can proceed. */
149 	error = xchk_fscount_warmup(sc);
150 	if (error)
151 		return error;
152 
153 	return xchk_trans_alloc(sc, 0);
154 }
155 
156 /*
157  * Part 1: Collecting filesystem summary counts.  For each AG, we add its
158  * summary counts (total inodes, free inodes, free data blocks) to an incore
159  * copy of the overall filesystem summary counts.
160  *
161  * To avoid false corruption reports in part 2, any failure in this part must
162  * set the INCOMPLETE flag even when a negative errno is returned.  This care
163  * must be taken with certain errno values (i.e. EFSBADCRC, EFSCORRUPTED,
164  * ECANCELED) that are absorbed into a scrub state flag update by
165  * xchk_*_process_error.
166  */
167 
168 /* Count free space btree blocks manually for pre-lazysbcount filesystems. */
169 static int
170 xchk_fscount_btreeblks(
171 	struct xfs_scrub	*sc,
172 	struct xchk_fscounters	*fsc,
173 	xfs_agnumber_t		agno)
174 {
175 	xfs_extlen_t		blocks;
176 	int			error;
177 
178 	error = xchk_ag_init_existing(sc, agno, &sc->sa);
179 	if (error)
180 		goto out_free;
181 
182 	error = xfs_btree_count_blocks(sc->sa.bno_cur, &blocks);
183 	if (error)
184 		goto out_free;
185 	fsc->fdblocks += blocks - 1;
186 
187 	error = xfs_btree_count_blocks(sc->sa.cnt_cur, &blocks);
188 	if (error)
189 		goto out_free;
190 	fsc->fdblocks += blocks - 1;
191 
192 out_free:
193 	xchk_ag_free(sc, &sc->sa);
194 	return error;
195 }
196 
197 /*
198  * Calculate what the global in-core counters ought to be from the incore
199  * per-AG structure.  Callers can compare this to the actual in-core counters
200  * to estimate by how much both in-core and on-disk counters need to be
201  * adjusted.
202  */
203 STATIC int
204 xchk_fscount_aggregate_agcounts(
205 	struct xfs_scrub	*sc,
206 	struct xchk_fscounters	*fsc)
207 {
208 	struct xfs_mount	*mp = sc->mp;
209 	struct xfs_perag	*pag;
210 	uint64_t		delayed;
211 	xfs_agnumber_t		agno;
212 	int			tries = 8;
213 	int			error = 0;
214 
215 retry:
216 	fsc->icount = 0;
217 	fsc->ifree = 0;
218 	fsc->fdblocks = 0;
219 
220 	for_each_perag(mp, agno, pag) {
221 		if (xchk_should_terminate(sc, &error))
222 			break;
223 
224 		/* This somehow got unset since the warmup? */
225 		if (!xfs_perag_initialised_agi(pag) ||
226 		    !xfs_perag_initialised_agf(pag)) {
227 			error = -EFSCORRUPTED;
228 			break;
229 		}
230 
231 		/* Count all the inodes */
232 		fsc->icount += pag->pagi_count;
233 		fsc->ifree += pag->pagi_freecount;
234 
235 		/* Add up the free/freelist/bnobt/cntbt blocks */
236 		fsc->fdblocks += pag->pagf_freeblks;
237 		fsc->fdblocks += pag->pagf_flcount;
238 		if (xfs_has_lazysbcount(sc->mp)) {
239 			fsc->fdblocks += pag->pagf_btreeblks;
240 		} else {
241 			error = xchk_fscount_btreeblks(sc, fsc, agno);
242 			if (error)
243 				break;
244 		}
245 
246 		/*
247 		 * Per-AG reservations are taken out of the incore counters,
248 		 * so they must be left out of the free blocks computation.
249 		 */
250 		fsc->fdblocks -= pag->pag_meta_resv.ar_reserved;
251 		fsc->fdblocks -= pag->pag_rmapbt_resv.ar_orig_reserved;
252 
253 	}
254 	if (pag)
255 		xfs_perag_rele(pag);
256 	if (error) {
257 		xchk_set_incomplete(sc);
258 		return error;
259 	}
260 
261 	/*
262 	 * The global incore space reservation is taken from the incore
263 	 * counters, so leave that out of the computation.
264 	 */
265 	fsc->fdblocks -= mp->m_resblks_avail;
266 
267 	/*
268 	 * Delayed allocation reservations are taken out of the incore counters
269 	 * but not recorded on disk, so leave them and their indlen blocks out
270 	 * of the computation.
271 	 */
272 	delayed = percpu_counter_sum(&mp->m_delalloc_blks);
273 	fsc->fdblocks -= delayed;
274 
275 	trace_xchk_fscounters_calc(mp, fsc->icount, fsc->ifree, fsc->fdblocks,
276 			delayed);
277 
278 
279 	/* Bail out if the values we compute are totally nonsense. */
280 	if (fsc->icount < fsc->icount_min || fsc->icount > fsc->icount_max ||
281 	    fsc->fdblocks > mp->m_sb.sb_dblocks ||
282 	    fsc->ifree > fsc->icount_max)
283 		return -EFSCORRUPTED;
284 
285 	/*
286 	 * If ifree > icount then we probably had some perturbation in the
287 	 * counters while we were calculating things.  We'll try a few times
288 	 * to maintain ifree <= icount before giving up.
289 	 */
290 	if (fsc->ifree > fsc->icount) {
291 		if (tries--)
292 			goto retry;
293 		xchk_set_incomplete(sc);
294 		return 0;
295 	}
296 
297 	return 0;
298 }
299 
300 #ifdef CONFIG_XFS_RT
301 STATIC int
302 xchk_fscount_add_frextent(
303 	struct xfs_mount		*mp,
304 	struct xfs_trans		*tp,
305 	const struct xfs_rtalloc_rec	*rec,
306 	void				*priv)
307 {
308 	struct xchk_fscounters		*fsc = priv;
309 	int				error = 0;
310 
311 	fsc->frextents += rec->ar_extcount;
312 
313 	xchk_should_terminate(fsc->sc, &error);
314 	return error;
315 }
316 
317 /* Calculate the number of free realtime extents from the realtime bitmap. */
318 STATIC int
319 xchk_fscount_count_frextents(
320 	struct xfs_scrub	*sc,
321 	struct xchk_fscounters	*fsc)
322 {
323 	struct xfs_mount	*mp = sc->mp;
324 	int			error;
325 
326 	fsc->frextents = 0;
327 	if (!xfs_has_realtime(mp))
328 		return 0;
329 
330 	xfs_ilock(sc->mp->m_rbmip, XFS_ILOCK_SHARED | XFS_ILOCK_RTBITMAP);
331 	error = xfs_rtalloc_query_all(sc->mp, sc->tp,
332 			xchk_fscount_add_frextent, fsc);
333 	if (error) {
334 		xchk_set_incomplete(sc);
335 		goto out_unlock;
336 	}
337 
338 out_unlock:
339 	xfs_iunlock(sc->mp->m_rbmip, XFS_ILOCK_SHARED | XFS_ILOCK_RTBITMAP);
340 	return error;
341 }
342 #else
343 STATIC int
344 xchk_fscount_count_frextents(
345 	struct xfs_scrub	*sc,
346 	struct xchk_fscounters	*fsc)
347 {
348 	fsc->frextents = 0;
349 	return 0;
350 }
351 #endif /* CONFIG_XFS_RT */
352 
353 /*
354  * Part 2: Comparing filesystem summary counters.  All we have to do here is
355  * sum the percpu counters and compare them to what we've observed.
356  */
357 
358 /*
359  * Is the @counter reasonably close to the @expected value?
360  *
361  * We neither locked nor froze anything in the filesystem while aggregating the
362  * per-AG data to compute the @expected value, which means that the counter
363  * could have changed.  We know the @old_value of the summation of the counter
364  * before the aggregation, and we re-sum the counter now.  If the expected
365  * value falls between the two summations, we're ok.
366  *
367  * Otherwise, we /might/ have a problem.  If the change in the summations is
368  * more than we want to tolerate, the filesystem is probably busy and we should
369  * just send back INCOMPLETE and see if userspace will try again.
370  */
371 static inline bool
372 xchk_fscount_within_range(
373 	struct xfs_scrub	*sc,
374 	const int64_t		old_value,
375 	struct percpu_counter	*counter,
376 	uint64_t		expected)
377 {
378 	int64_t			min_value, max_value;
379 	int64_t			curr_value = percpu_counter_sum(counter);
380 
381 	trace_xchk_fscounters_within_range(sc->mp, expected, curr_value,
382 			old_value);
383 
384 	/* Negative values are always wrong. */
385 	if (curr_value < 0)
386 		return false;
387 
388 	/* Exact matches are always ok. */
389 	if (curr_value == expected)
390 		return true;
391 
392 	min_value = min(old_value, curr_value);
393 	max_value = max(old_value, curr_value);
394 
395 	/* Within the before-and-after range is ok. */
396 	if (expected >= min_value && expected <= max_value)
397 		return true;
398 
399 	/*
400 	 * If the difference between the two summations is too large, the fs
401 	 * might just be busy and so we'll mark the scrub incomplete.  Return
402 	 * true here so that we don't mark the counter corrupt.
403 	 *
404 	 * XXX: In the future when userspace can grant scrub permission to
405 	 * quiesce the filesystem to solve the outsized variance problem, this
406 	 * check should be moved up and the return code changed to signal to
407 	 * userspace that we need quiesce permission.
408 	 */
409 	if (max_value - min_value >= XCHK_FSCOUNT_MIN_VARIANCE) {
410 		xchk_set_incomplete(sc);
411 		return true;
412 	}
413 
414 	return false;
415 }
416 
417 /* Check the superblock counters. */
418 int
419 xchk_fscounters(
420 	struct xfs_scrub	*sc)
421 {
422 	struct xfs_mount	*mp = sc->mp;
423 	struct xchk_fscounters	*fsc = sc->buf;
424 	int64_t			icount, ifree, fdblocks, frextents;
425 	int			error;
426 
427 	/* Snapshot the percpu counters. */
428 	icount = percpu_counter_sum(&mp->m_icount);
429 	ifree = percpu_counter_sum(&mp->m_ifree);
430 	fdblocks = percpu_counter_sum(&mp->m_fdblocks);
431 	frextents = percpu_counter_sum(&mp->m_frextents);
432 
433 	/* No negative values, please! */
434 	if (icount < 0 || ifree < 0 || fdblocks < 0 || frextents < 0)
435 		xchk_set_corrupt(sc);
436 
437 	/* See if icount is obviously wrong. */
438 	if (icount < fsc->icount_min || icount > fsc->icount_max)
439 		xchk_set_corrupt(sc);
440 
441 	/* See if fdblocks is obviously wrong. */
442 	if (fdblocks > mp->m_sb.sb_dblocks)
443 		xchk_set_corrupt(sc);
444 
445 	/* See if frextents is obviously wrong. */
446 	if (frextents > mp->m_sb.sb_rextents)
447 		xchk_set_corrupt(sc);
448 
449 	/*
450 	 * XXX: We can't quiesce percpu counter updates, so exit early.
451 	 * This can be re-enabled when we gain exclusive freeze functionality.
452 	 */
453 	return 0;
454 
455 	/*
456 	 * If ifree exceeds icount by more than the minimum variance then
457 	 * something's probably wrong with the counters.
458 	 */
459 	if (ifree > icount && ifree - icount > XCHK_FSCOUNT_MIN_VARIANCE)
460 		xchk_set_corrupt(sc);
461 
462 	/* Walk the incore AG headers to calculate the expected counters. */
463 	error = xchk_fscount_aggregate_agcounts(sc, fsc);
464 	if (!xchk_process_error(sc, 0, XFS_SB_BLOCK(mp), &error))
465 		return error;
466 	if (sc->sm->sm_flags & XFS_SCRUB_OFLAG_INCOMPLETE)
467 		return 0;
468 
469 	/* Count the free extents counter for rt volumes. */
470 	error = xchk_fscount_count_frextents(sc, fsc);
471 	if (!xchk_process_error(sc, 0, XFS_SB_BLOCK(mp), &error))
472 		return error;
473 	if (sc->sm->sm_flags & XFS_SCRUB_OFLAG_INCOMPLETE)
474 		return 0;
475 
476 	/* Compare the in-core counters with whatever we counted. */
477 	if (!xchk_fscount_within_range(sc, icount, &mp->m_icount, fsc->icount))
478 		xchk_set_corrupt(sc);
479 
480 	if (!xchk_fscount_within_range(sc, ifree, &mp->m_ifree, fsc->ifree))
481 		xchk_set_corrupt(sc);
482 
483 	if (!xchk_fscount_within_range(sc, fdblocks, &mp->m_fdblocks,
484 			fsc->fdblocks))
485 		xchk_set_corrupt(sc);
486 
487 	if (!xchk_fscount_within_range(sc, frextents, &mp->m_frextents,
488 			fsc->frextents))
489 		xchk_set_corrupt(sc);
490 
491 	return 0;
492 }
493