[zfs-discuss] ZFS on RAID-5 array

Gionatan Danti g.danti at assyoma.it
Fri Apr 13 10:42:03 EDT 2018


On 13/04/2018 01:09, Gordan Bobic via zfs-discuss wrote:
> SATA disks also aren't supposed to feed back duff data instead of 
> returning an error, because all sectors have ECC on them, but it still 
> nevertheless happens.

Both SATA and SAS disks have ECC protected sectors. The difference is 
that the latter can use an *additiontal* 8/20 bytes code to permit 
end-to-end checksumming. Some high level profile use of the DIF field 
also enable *application* checksum to be stored on the sector itself.

> Write-read-verify has existed on SATA drives since forever. Not all 
> implement it but most do. I'm the one that wrote the hdparm patch to add 
> a feature to toggle this using hdparm. Unfortunately, sectors rot - just 
> because the data was sound when it was written, and was verified as 
> such, does not mean it will still be sound when you read it months 
> later. Enabling the feature also halves the performance.

While real, my main concern is not bitroot, rather a corruption due to 
bad disk's RAM chip and/or low quality power supply. I identified at 
least two machines with this problem. In this case, a system which can 
checksum *and* heal your data is extremely important. This is the very 
reason why the DIF field exist: to guarantee additional protection (and 
healing capability) using classical filesystem as XFS and the likes.


> Bottom line, it's not about the cost, all disks have sector ECC, from 
> the cheapest to the most expensive, and many cheap SATA disks have WRV. 
> The problem is that this doesn't magically help mitigate the error 
> conditions that ZFS was specifically designed to address.

DIF *specifically* address the very same concern: it enables RAID 
controllers to detect corruption and re-try reading from another disk 
(mirror) and stipe reconstruct (RAID5/6), rather than returning fast 
garbage.


However, it is restricted to SAS controller/backplanes/disks only. For 
this reason, there is also an on-going effort to emulate it via device 
mapper. Give a look here: 
https://github.com/torvalds/linux/blob/master/Documentation/device-mapper/dm-integrity.txt

That said, ZFS offer comparable protection now, and with consumer disks, 
which is awesome.

-- 
Danti Gionatan
Supporto Tecnico
Assyoma S.r.l. - www.assyoma.it
email: g.danti at assyoma.it - info at assyoma.it
GPG public key ID: FF5F32A8


More information about the zfs-discuss mailing list