tree 237086ec658a3adccc6a6da9275692547afa5ebb
parent e276ac5cbd22b928d2b486c588f91f075c3bca1b
author Darrick J. Wong <djwong@kernel.org> 1720041674 -0700
committer Darrick J. Wong <djwong@kernel.org> 1723064179 -0700

xfs_scrub: tune fstrim minlen parameter based on free space histograms

Currently, phase 8 runs very slowly on filesystems with a lot of small
free space extents.  To reduce the amount of time spent on fstrim
activities during phase 8, we want to balance estimated runtime against
completeness of the trim.  In short, the goal is to reduce runtime by
avoiding small trim requests.

At the start of phase 8, a CDF is computed in decreasing order of extent
length from the histogram buckets created during the fsmap scan in phase
7.  A point corresponding to the fstrim percentage target is chosen from
the CDF and mapped back to a histogram bucket, and free space extents
smaller than that amount are ommitted from fstrim.

On my aging /home filesystem, the free space histogram reported by
xfs_spaceman looks like this:

   from      to extents    blocks    pct blkcdf extcdf
      1       1  121953    121953   0.04 100.00 100.00
      2       3  124741    299694   0.09  99.96  81.16
      4       7  113492    593763   0.18  99.87  61.89
      8      15  109215   1179524   0.36  99.69  44.36
     16      31   76972   1695455   0.52  99.33  27.48
     32      63   48655   2219667   0.68  98.82  15.59
     64     127   31398   2876898   0.88  98.14   8.08
    128     255    8014   1447920   0.44  97.27   3.23
    256     511    4142   1501758   0.46  96.82   1.99
    512    1023    2433   1768732   0.54  96.37   1.35
   1024    2047    1795   2648460   0.81  95.83   0.97
   2048    4095    1429   4206103   1.28  95.02   0.69
   4096    8191    1045   6162111   1.88  93.74   0.47
   8192   16383     791   9242745   2.81  91.87   0.31
  16384   32767     473  10883977   3.31  89.06   0.19
  32768   65535     272  12385566   3.77  85.74   0.12
  65536  131071     192  18098739   5.51  81.98   0.07
 131072  262143     108  20675199   6.29  76.47   0.04
 262144  524287      80  29061285   8.84  70.18   0.03
 524288 1048575      39  29002829   8.83  61.33   0.02
1048576 2097151      25  36824985  11.21  52.51   0.01
2097152 4194303      32 101727192  30.95  41.30   0.01
4194304 8388607       7  34007410  10.35  10.35   0.00

From this table, we see that free space extents that are 16 blocks or
longer constitute 99.3% of the free space in the filesystem but only
27.5% of the extents.  If we set the fstrim minlen parameter to 16
blocks, that means that we can trim over 99% of the space in one third
of the time it would take to trim everything.

Add a new -o fstrim_pct= option to xfs_scrub just in case there are
users out there who want a different percentage.  For example, accepting
a 95% trim would net us a speed increase of nearly two orders of
magnitude, ignoring system call overhead.  Setting it to 100% will trim
everything, just like fstrim(8).

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>