Re: [Qemu-devel] [PATCH 0/2] block: Write out internal caches even with cache=unsafe

[Kevin Wolf]

Am 24.10.2011 09:53, schrieb Paolo Bonzini:
> On 10/24/2011 09:37 AM, Kevin Wolf wrote:
>>> Why? cache=unsafe is explicitly allowing to s/data/manure/ on
>>> crash.
>>
>> It's surely expected on a host crash, but is it for a qemu crash?
>> cache=unsafe was introduced to avoid fsync() costs, which it still
>> does after this patch.
> 
> I think it's not about "why is it there", but rather about "what is it 
> useful for".  My interpretation of it is "I do not need the image 
> anymore unless the command exits cleanly": VM installations, qemu-img 
> conversions, BDRV_O_SNAPSHOT (doesn't do it yet, but it could).  Even 
> SIGINT and SIGTERM would be excluded from this definition, but they cost 
> nothing so it's nice to include them.

I think another common interpretation is: "I don't run this VM in
production but for development. I want the VM to go faster and I can
recreate the image in the unlikely event that power fails during my
work. But it certainly would be nasty."

>>> If you do this for raw-posix, you need to do it for all protocols.
>>
>> rbd could use it, too, right. Any other protocol I missed?
> 
> NBD could, but it doesn't support flush yet.
> 
> In general, even if it were useful to implement this, I'm not sure this 
> is the best way to implement it.  For example you could simply clear 
> BDRV_O_NO_FLUSH in qcow2_open.

That could work, too. On the other hand I don't like block drivers to
modify their flags. For example, would query-block still provide the
correct cache mode then?

But I think that starting to make exceptions for single block drivers
isn't a good idea anyway. If we want bdrv_flush() to write out all
metadata internal to qemu, I think the approach with checking the flag
in drivers calling things like fsync() is better. The common thing is to
do the flush.

Kevin