zfs checksums on non-raidz

[Peter Maloney]

Am 08.02.2012 19:47, schrieb Tom Vier:
> In mirror and single dev zpools, are data and metadata still checksummed, or
> only when using raidz? 
>
> I've only been able to determine so far that the uberblocks at the zpool level
> are always checksummed (according to Sun's on-disk format pdf).
>
Everything, even a single disk, is checksummed.

Test it with:


zpool create testpool gpt/testslice
cp /some/big/test/file /testpool/

# next command is dangerous because it means you can now directly write
to disks and damage the data
sysctl kern.geom.debugflags=0x10

# next command is more dangerous, because it intentionally destroys data
on the slice (be careful about what you put in of=... )
dd if=/dev/zero of=/dev/gpt/testslice bs=1M count=5000

# after this command, you should have some checksum errors
md5 /testpool/file

# after this command and waiting, you should have all the errors there
are (all files, including not read by the md5 above)
zpool scrub testpool

...wait...

zpool status testpool

Now you probably (depending on whether or not the part we wrote with dd
actually went on top of where data was placed) have checksum errors, and
a message about an unrecoverable file. If it was raidz or mirror, it
would have automatically fixed it for you, but with no redundancy, all
the checksum can do is tell you the file is broken.