fs/xfs/scrub/fscounters.c

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