Re: NFS problem after 2.4.19-pre3, not solved

[Trond Myklebust]

>>>>> " " == Andrea Arcangeli <[e-mail]> writes:

     > the thing that makes the difference is the backout-cto patch
     > according those numbers and I doubt it is influencing the page
     > replacement in any way (is kernel-side memory pressure going to
     > increase significantly with cto?).

No. The only possibility I can see is if the extra checks are causing
extra cache invalidations. I don't why that should be the case, but
then again I'm not able to reproduce those numbers...

     > 2.4.19-pre3 vanilla first time after boot 5m15s, then 1m56s
     > 1m57s 1m56s 1m56s 1m57s

     > 2.4.19-pre4 vanilla 5m20s, then 4m00s 4m01s 4m00s 4m01s 4m01a

     > 2.4.19-pre4-nfs-backout-cto, from pr8aa2, 2 Hunks 5m13a, then
     > 1m57s 1m58s 1m57s 1m58s 1m59s

     > nfs-backout-cto is appended. Now if the previous kernel was
     > buggy and it was not invalidating "invalid" cache then cto is
     > right, otherwise it sounds like the cto patch is invalidating
     > more cache than necessary.

Check the patch: it doesn't invalidate the cache when the mtime stays
the same. A tcpdump would show whether this is the case or not.

I would be interested to see if this is something that is related to
nfs-server only. (I.e. whether or not Mario can see the same problem
with knfsd.)

Cheers,
  Trond