xref: /openbmc/linux/fs/xfs/scrub/fscounters.c (revision 07c7c6bf)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * Copyright (C) 2019 Oracle.  All Rights Reserved.
4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
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_defer.h"
13 #include "xfs_btree.h"
14 #include "xfs_bit.h"
15 #include "xfs_log_format.h"
16 #include "xfs_trans.h"
17 #include "xfs_sb.h"
18 #include "xfs_inode.h"
19 #include "xfs_alloc.h"
20 #include "xfs_ialloc.h"
21 #include "xfs_rmap.h"
22 #include "xfs_error.h"
23 #include "xfs_errortag.h"
24 #include "xfs_icache.h"
25 #include "xfs_health.h"
26 #include "xfs_bmap.h"
27 #include "scrub/xfs_scrub.h"
28 #include "scrub/scrub.h"
29 #include "scrub/common.h"
30 #include "scrub/trace.h"
31 
32 /*
33  * FS Summary Counters
34  * ===================
35  *
36  * The basics of filesystem summary counter checking are that we iterate the
37  * AGs counting the number of free blocks, free space btree blocks, per-AG
38  * reservations, inodes, delayed allocation reservations, and free inodes.
39  * Then we compare what we computed against the in-core counters.
40  *
41  * However, the reality is that summary counters are a tricky beast to check.
42  * While we /could/ freeze the filesystem and scramble around the AGs counting
43  * the free blocks, in practice we prefer not do that for a scan because
44  * freezing is costly.  To get around this, we added a per-cpu counter of the
45  * delalloc reservations so that we can rotor around the AGs relatively
46  * quickly, and we allow the counts to be slightly off because we're not taking
47  * any locks while we do this.
48  *
49  * So the first thing we do is warm up the buffer cache in the setup routine by
50  * walking all the AGs to make sure the incore per-AG structure has been
51  * initialized.  The expected value calculation then iterates the incore per-AG
52  * structures as quickly as it can.  We snapshot the percpu counters before and
53  * after this operation and use the difference in counter values to guess at
54  * our tolerance for mismatch between expected and actual counter values.
55  */
56 
57 /*
58  * Since the expected value computation is lockless but only browses incore
59  * values, the percpu counters should be fairly close to each other.  However,
60  * we'll allow ourselves to be off by at least this (arbitrary) amount.
61  */
62 #define XCHK_FSCOUNT_MIN_VARIANCE	(512)
63 
64 /*
65  * Make sure the per-AG structure has been initialized from the on-disk header
66  * contents and trust that the incore counters match the ondisk counters.  (The
67  * AGF and AGI scrubbers check them, and a normal xfs_scrub run checks the
68  * summary counters after checking all AG headers).  Do this from the setup
69  * function so that the inner AG aggregation loop runs as quickly as possible.
70  *
71  * This function runs during the setup phase /before/ we start checking any
72  * metadata.
73  */
74 STATIC int
75 xchk_fscount_warmup(
76 	struct xfs_scrub	*sc)
77 {
78 	struct xfs_mount	*mp = sc->mp;
79 	struct xfs_buf		*agi_bp = NULL;
80 	struct xfs_buf		*agf_bp = NULL;
81 	struct xfs_perag	*pag = NULL;
82 	xfs_agnumber_t		agno;
83 	int			error = 0;
84 
85 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
86 		pag = xfs_perag_get(mp, agno);
87 
88 		if (pag->pagi_init && pag->pagf_init)
89 			goto next_loop_perag;
90 
91 		/* Lock both AG headers. */
92 		error = xfs_ialloc_read_agi(mp, sc->tp, agno, &agi_bp);
93 		if (error)
94 			break;
95 		error = xfs_alloc_read_agf(mp, sc->tp, agno, 0, &agf_bp);
96 		if (error)
97 			break;
98 		error = -ENOMEM;
99 		if (!agf_bp || !agi_bp)
100 			break;
101 
102 		/*
103 		 * These are supposed to be initialized by the header read
104 		 * function.
105 		 */
106 		error = -EFSCORRUPTED;
107 		if (!pag->pagi_init || !pag->pagf_init)
108 			break;
109 
110 		xfs_buf_relse(agf_bp);
111 		agf_bp = NULL;
112 		xfs_buf_relse(agi_bp);
113 		agi_bp = NULL;
114 next_loop_perag:
115 		xfs_perag_put(pag);
116 		pag = NULL;
117 		error = 0;
118 
119 		if (fatal_signal_pending(current))
120 			break;
121 	}
122 
123 	if (agf_bp)
124 		xfs_buf_relse(agf_bp);
125 	if (agi_bp)
126 		xfs_buf_relse(agi_bp);
127 	if (pag)
128 		xfs_perag_put(pag);
129 	return error;
130 }
131 
132 int
133 xchk_setup_fscounters(
134 	struct xfs_scrub	*sc,
135 	struct xfs_inode	*ip)
136 {
137 	struct xchk_fscounters	*fsc;
138 	int			error;
139 
140 	sc->buf = kmem_zalloc(sizeof(struct xchk_fscounters), KM_SLEEP);
141 	if (!sc->buf)
142 		return -ENOMEM;
143 	fsc = sc->buf;
144 
145 	xfs_icount_range(sc->mp, &fsc->icount_min, &fsc->icount_max);
146 
147 	/* We must get the incore counters set up before we can proceed. */
148 	error = xchk_fscount_warmup(sc);
149 	if (error)
150 		return error;
151 
152 	/*
153 	 * Pause background reclaim while we're scrubbing to reduce the
154 	 * likelihood of background perturbations to the counters throwing off
155 	 * our calculations.
156 	 */
157 	xchk_stop_reaping(sc);
158 
159 	return xchk_trans_alloc(sc, 0);
160 }
161 
162 /*
163  * Calculate what the global in-core counters ought to be from the incore
164  * per-AG structure.  Callers can compare this to the actual in-core counters
165  * to estimate by how much both in-core and on-disk counters need to be
166  * adjusted.
167  */
168 STATIC int
169 xchk_fscount_aggregate_agcounts(
170 	struct xfs_scrub	*sc,
171 	struct xchk_fscounters	*fsc)
172 {
173 	struct xfs_mount	*mp = sc->mp;
174 	struct xfs_perag	*pag;
175 	uint64_t		delayed;
176 	xfs_agnumber_t		agno;
177 	int			tries = 8;
178 
179 retry:
180 	fsc->icount = 0;
181 	fsc->ifree = 0;
182 	fsc->fdblocks = 0;
183 
184 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
185 		pag = xfs_perag_get(mp, agno);
186 
187 		/* This somehow got unset since the warmup? */
188 		if (!pag->pagi_init || !pag->pagf_init) {
189 			xfs_perag_put(pag);
190 			return -EFSCORRUPTED;
191 		}
192 
193 		/* Count all the inodes */
194 		fsc->icount += pag->pagi_count;
195 		fsc->ifree += pag->pagi_freecount;
196 
197 		/* Add up the free/freelist/bnobt/cntbt blocks */
198 		fsc->fdblocks += pag->pagf_freeblks;
199 		fsc->fdblocks += pag->pagf_flcount;
200 		fsc->fdblocks += pag->pagf_btreeblks;
201 
202 		/*
203 		 * Per-AG reservations are taken out of the incore counters,
204 		 * so they must be left out of the free blocks computation.
205 		 */
206 		fsc->fdblocks -= pag->pag_meta_resv.ar_reserved;
207 		fsc->fdblocks -= pag->pag_rmapbt_resv.ar_orig_reserved;
208 
209 		xfs_perag_put(pag);
210 
211 		if (fatal_signal_pending(current))
212 			break;
213 	}
214 
215 	/*
216 	 * The global incore space reservation is taken from the incore
217 	 * counters, so leave that out of the computation.
218 	 */
219 	fsc->fdblocks -= mp->m_resblks_avail;
220 
221 	/*
222 	 * Delayed allocation reservations are taken out of the incore counters
223 	 * but not recorded on disk, so leave them and their indlen blocks out
224 	 * of the computation.
225 	 */
226 	delayed = percpu_counter_sum(&mp->m_delalloc_blks);
227 	fsc->fdblocks -= delayed;
228 
229 	trace_xchk_fscounters_calc(mp, fsc->icount, fsc->ifree, fsc->fdblocks,
230 			delayed);
231 
232 
233 	/* Bail out if the values we compute are totally nonsense. */
234 	if (fsc->icount < fsc->icount_min || fsc->icount > fsc->icount_max ||
235 	    fsc->fdblocks > mp->m_sb.sb_dblocks ||
236 	    fsc->ifree > fsc->icount_max)
237 		return -EFSCORRUPTED;
238 
239 	/*
240 	 * If ifree > icount then we probably had some perturbation in the
241 	 * counters while we were calculating things.  We'll try a few times
242 	 * to maintain ifree <= icount before giving up.
243 	 */
244 	if (fsc->ifree > fsc->icount) {
245 		if (tries--)
246 			goto retry;
247 		xchk_set_incomplete(sc);
248 		return 0;
249 	}
250 
251 	return 0;
252 }
253 
254 /*
255  * Is the @counter reasonably close to the @expected value?
256  *
257  * We neither locked nor froze anything in the filesystem while aggregating the
258  * per-AG data to compute the @expected value, which means that the counter
259  * could have changed.  We know the @old_value of the summation of the counter
260  * before the aggregation, and we re-sum the counter now.  If the expected
261  * value falls between the two summations, we're ok.
262  *
263  * Otherwise, we /might/ have a problem.  If the change in the summations is
264  * more than we want to tolerate, the filesystem is probably busy and we should
265  * just send back INCOMPLETE and see if userspace will try again.
266  */
267 static inline bool
268 xchk_fscount_within_range(
269 	struct xfs_scrub	*sc,
270 	const int64_t		old_value,
271 	struct percpu_counter	*counter,
272 	uint64_t		expected)
273 {
274 	int64_t			min_value, max_value;
275 	int64_t			curr_value = percpu_counter_sum(counter);
276 
277 	trace_xchk_fscounters_within_range(sc->mp, expected, curr_value,
278 			old_value);
279 
280 	/* Negative values are always wrong. */
281 	if (curr_value < 0)
282 		return false;
283 
284 	/* Exact matches are always ok. */
285 	if (curr_value == expected)
286 		return true;
287 
288 	min_value = min(old_value, curr_value);
289 	max_value = max(old_value, curr_value);
290 
291 	/* Within the before-and-after range is ok. */
292 	if (expected >= min_value && expected <= max_value)
293 		return true;
294 
295 	/*
296 	 * If the difference between the two summations is too large, the fs
297 	 * might just be busy and so we'll mark the scrub incomplete.  Return
298 	 * true here so that we don't mark the counter corrupt.
299 	 *
300 	 * XXX: In the future when userspace can grant scrub permission to
301 	 * quiesce the filesystem to solve the outsized variance problem, this
302 	 * check should be moved up and the return code changed to signal to
303 	 * userspace that we need quiesce permission.
304 	 */
305 	if (max_value - min_value >= XCHK_FSCOUNT_MIN_VARIANCE) {
306 		xchk_set_incomplete(sc);
307 		return true;
308 	}
309 
310 	return false;
311 }
312 
313 /* Check the superblock counters. */
314 int
315 xchk_fscounters(
316 	struct xfs_scrub	*sc)
317 {
318 	struct xfs_mount	*mp = sc->mp;
319 	struct xchk_fscounters	*fsc = sc->buf;
320 	int64_t			icount, ifree, fdblocks;
321 	int			error;
322 
323 	/* Snapshot the percpu counters. */
324 	icount = percpu_counter_sum(&mp->m_icount);
325 	ifree = percpu_counter_sum(&mp->m_ifree);
326 	fdblocks = percpu_counter_sum(&mp->m_fdblocks);
327 
328 	/* No negative values, please! */
329 	if (icount < 0 || ifree < 0 || fdblocks < 0)
330 		xchk_set_corrupt(sc);
331 
332 	/* See if icount is obviously wrong. */
333 	if (icount < fsc->icount_min || icount > fsc->icount_max)
334 		xchk_set_corrupt(sc);
335 
336 	/* See if fdblocks is obviously wrong. */
337 	if (fdblocks > mp->m_sb.sb_dblocks)
338 		xchk_set_corrupt(sc);
339 
340 	/*
341 	 * If ifree exceeds icount by more than the minimum variance then
342 	 * something's probably wrong with the counters.
343 	 */
344 	if (ifree > icount && ifree - icount > XCHK_FSCOUNT_MIN_VARIANCE)
345 		xchk_set_corrupt(sc);
346 
347 	/* Walk the incore AG headers to calculate the expected counters. */
348 	error = xchk_fscount_aggregate_agcounts(sc, fsc);
349 	if (!xchk_process_error(sc, 0, XFS_SB_BLOCK(mp), &error))
350 		return error;
351 	if (sc->sm->sm_flags & XFS_SCRUB_OFLAG_INCOMPLETE)
352 		return 0;
353 
354 	/* Compare the in-core counters with whatever we counted. */
355 	if (!xchk_fscount_within_range(sc, icount, &mp->m_icount, fsc->icount))
356 		xchk_set_corrupt(sc);
357 
358 	if (!xchk_fscount_within_range(sc, ifree, &mp->m_ifree, fsc->ifree))
359 		xchk_set_corrupt(sc);
360 
361 	if (!xchk_fscount_within_range(sc, fdblocks, &mp->m_fdblocks,
362 			fsc->fdblocks))
363 		xchk_set_corrupt(sc);
364 
365 	return 0;
366 }
367