Re: [Qemu-devel] [PATCH v2 0/6] external backup api

From: Paolo Bonzini <pbonzini@redhat.com>
To: Markus Armbruster <armbru@redhat.com>, Fam Zheng <famz@redhat.com>
Cc: kwolf@redhat.com,
	Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>,
	"Denis V. Lunev" <den@virtuozzo.com>,
	Stefan Hajnoczi <stefanha@gmail.com>,
	qemu-devel@nongnu.org, jsnow@redhat.com
Subject: Re: [Qemu-devel] [PATCH v2 0/6] external backup api
Date: Fri, 26 Feb 2016 21:03:30 +0100	[thread overview]
Message-ID: <56D0AF92.7000509@redhat.com> (raw)
In-Reply-To: <56D0ADBD.20000@redhat.com>

On 26/02/2016 20:55, Paolo Bonzini wrote:
> 
> 
> On 19/02/2016 09:51, Markus Armbruster wrote:
>>> Is it an abuse to "Get LBA Status" to return dirty information? Because in SCSI
>>> the command reports "mapped", "allocated" and "anchored" statuses. Does that
>>> mean NBD will use a different status set?
>>
>> Perhaps some conceptual gymnastics can get us to standard semantics.
>>
>> Incremental backup wants to copy out an image's "dirty" blocks.
>>
>> We can view this as a bitmap telling us which of the image's blocks are
>> dirty.
>>
>> An alternative view would be base image + dirty delta image.  In the the
>> dirty delta, exactly the dirty blocks are allocated.  The delta image
>> may be conceptual.
> 
> I see a problem here. On one hand I agree that the "GET LBA STATUS" is
> a natural extension to the NBD protocol.
> 
> On the other hand, the Get LBA Status command in SCSI reflects the
> state over the whole chain, not only the top element.  It is the
> equivalent of bdrv_get_block_status_above(bs, NULL, ...), rather than
> bdrv_get_block_status(bs, ...).  My understanding is that the dirty
> block information would require the latter, especially in the
> "conceptual delta image" model that Markus describes above.
> 
> Having NBD implement subtly different semantics compared to SCSI is a
> bad idea in my opinion.
> 
> Of course if we call it "READ DIRTY BLOCKS" that would work, but I
> don't think such a command would be something that it makes sense to
> have in the general purpose NBD spec, so you would need a mechanism
> for vendor-specific extensions.

Trying to be constructive: shall we instead have a simple mini-protocol
with commands like "read bitmap slice", "clear bitmap slice", "query
next 0 bit", "query next 1 bit"?

Not necessarily QMP, just a little socket thing.  It could be JSON or
ASCII or binary.  It sucks to implement a new protocol, but perhaps it
could be something compatible with VMware or Parallels.

Sorry if this has already been proposed, I'm late to the game and I only
read this subthread.

Paolo