Re: ZFS: data sometimes not getting correctly flushed to disk (possible mmap/write issue)?

On Fri, 2008-02-15 at 16:59 -0600, Richard Todd wrote:
> Bit of a long story behind this one: I've been playing with the "picard"
> program in ports to do re-tagging of audio (mp3/ogg) files, and was having
> problems with the tagging process creating corrupted .ogg files.  On
> further investigation trying to debug this, I found that the "picard"
> program seemed to be correctly writing the files if I checked them
> immediately, but the files were corrupt when I looked at them later.
> Further investigation and coming up with test cases brought me to the
> surprising conclusion that ZFS was to blame -- under certain conditions ZFS
> seems not to be writing all the data in the file out to disk correctly.
> The file appears to be good only so long as the file's data is in cache, but
> the data actually on disk doesn't match what's in cache.

I don't currently have a system that I can test this with, but is there
any chance you could make the following change to your script?

        #!/bin/sh
        NAME=$1
        LEN=$2
        rm $NAME
        ./a.out $NAME $LEN
        ./gen2.py $NAME
        md5 $NAME
+ cp $NAME /some/ufs/filesystem/$NAME.1
        ls -l $NAME
        sudo umount /u3
        sudo mount -t zfs u3 /u3
        md5 $NAME
+ cp $NAME /some/ufs/filesystem/$NAME.2
        ls -l $NAME
+ md5 /some/ufs/filesystem/$NAME.?

and when you find files that differ, compare the two copies stored on
the UFS filesystem with cmp(1)?  I'd be interested to know exactly what
the corruption looks like - is a single bit differen?  Or was an entire
page destroyed?

Gavin

Gavin