Re: MD/RAID time out writing superblock

[Mark Lord <liml@rtr.ca>]

Tejun Heo wrote:
> Hello,
> 
> Mark Lord wrote:
>> Tejun.. do we do a FLUSH CACHE before issuing a non-NCQ command ?
> 
> Nope.
> 
>> If not, then I think we may need to add code to do it.
> 
> Hmm... can you explain a bit more?  That seems rather extreme to me.
..

You may recall that I first raised this issue about a year ago,
when my own RAID0 array (MythTV box) started showing errors very
similar to what Chris is reporting.

These were easily triggered by running hddtemp once every few seconds
to log drive temperatures during Myth recording sessions.

hddtemp uses SMART commands.

The actual errors in the logs were command timeouts,
but at this point I no longer remember which opcode was
actually timing out.  Disabling the onboard write cache
immediately "cured" the problem, at the expense of MUCH
slower I/O times.

My theory at the time, was that some non-NCQ commands might be triggering
an internal FLUSH CACHE within the (Hitachi) drive firmware, which then
caused the original command to timeout in libata (due to the large amounts
of data present in the onboard write-caches).

Now that more people are playing the game, we're seeing more and more
reports of strange interactions with smartd running in the background.

I suspect more and more now that this is an (avoidable) interaction
between the write-cache and the SMART opcode, and it could perhaps be
avoided by doing a FLUSH CACHE before any SMART (or non-data command) opcode.

Cheers