Re: [Qemu-devel] [PATCH v5 17/38] block: Add BlockBackendRootState

[Kevin Wolf <kwolf@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 2382 bytes --]

Am 29.09.2015 um 14:30 hat Eric Blake geschrieben:
> On 09/29/2015 05:17 AM, Alberto Garcia wrote:
> > On Tue 22 Sep 2015 05:00:05 PM CEST, Max Reitz wrote:
> > 
> >>> The correct way to solve this seems to be that each BB has its own
> >>> I/O throttling filter. Actually, if we could lift the throttling code
> >>> to BlockBackend, that might solve the problem.
> >>
> >> So yes, as long as we have throttling on the BDS level, something may
> >> always break when having multiple BBs per BDS, because you'd probably
> >> want throttling to be on the BB level. But lifting that doesn't seem a
> >> trivial task, and I don't really know whether I want this in this
> >> series.
> >>
> >> Especially considering that right now you cannot have multiple BBs per
> >> BDS.
> >>
> >> All in all: Yes, before we allow multiple BBs per BDS we want
> >> throttling to be on the BB level. But this is nothing that should be
> >> done in or for this series, since right now we do not allow multiple
> >> BBs per BDS.
> > 
> > I agree that it makes sense to move throttling to the BB level. It's
> > probably not trivial but I don't think it's too complex either. I can
> > give it a look once this series has been merged.
> 
> Ultimately, I think we want throttling at both levels.  I can make the
> argument that in a cloud, a guest should not be able to consume more
> than X resources (throttling at the BB, regardless of what BDS tree it
> is associated with); but I can also argue that I may want to throttle
> specific BDS (a backing chain with network backing file and local active
> file: throttle the backing file to limit network traffic, and the guest
> gets faster as it moves more local; or conversely, a common backing file
> and one-off guest: throttle the active layer to do performance analysis
> of the bottlenecks where the guest differs from the common backing layer).

Yes, but the throttling that we currently have is the one that belongs
into BlockBackend. This becomes very clear when you consider that the
throttling moves to the new top when you create live snapshots.

Throttling for individual BDSes should be done by a filter block driver.
In fact, I also suspect that the BB-level filtering should eventually be
done by a filter block driver, but that's something for which we don't
have a clear design yet.

Kevin

[-- Attachment #2: Type: application/pgp-signature, Size: 836 bytes --]