Re: [Qemu-devel] [Nbd] [PATCH 2/2] NBD proto: add GET_LBA_STATUS extension

[Kevin Wolf]

Am 23.03.2016 um 18:58 hat Wouter Verhelst geschrieben:
> On Wed, Mar 23, 2016 at 05:16:02PM +0300, Denis V. Lunev wrote:
> > +    The type of information required by the client is passed to server in 
> > the
> > +    command flags field. If the server does not implement requested type or
> > +    have no means to express it, it MUST NOT return an error, but instead 
> > MUST
> > +    return a single LBA status descriptor with *offset* and *length* equal 
> > to
> > +    the *offset* and *length* from request, and *status* set to `0`.
> > +
> > +    The following request types are currently defined for the command:
> > +
> > +    1. Block provisioning state
> > +
> > +    Upon receiving an `NBD_CMD_GET_LBA_STATUS` command with command flags
> > +    field set to `NBD_FLAG_GET_ALLOCATED` (0x0), the server MUST return
> 
> I prefer to have a non-zero flag value.
> 
> > +    the provisioning state of the device. The following provisionnig states
> > +    are defined for the command:
> > +
> > +      - `NBD_STATE_ALLOCATED` (0x0), LBA extent is present on the block 
> > device;
> > +      - `NBD_STATE_ZEROED` (0x1), LBA extent is present on the block device
> > +        and contains zeroes;
> 
> Presumably this should be "contains only zeroes"?
> 
> Also, this may end up being a fairly expensive call for the server to
> process. Is it really useful?

I think we need to make clear that this is meant as an optimisation and
it's always a valid option for a server to return NBD_STATE_ALLOCATED
even if the contents is zeroed.

It is definitely useful if the server has a means to efficiently find
out the allocation status (e.g. SEEK_HOLE). In that case the client may
be able to avoid reading the block and sending it over the network, or
when making a copy, it could use it to keep the target file sparse. If
the client can't take advantage, we didn't have much overhead, so it's
fine.

It's less clear in a case where the server needs to read in the block
and scan its contents. It could still be a net win if the next thing the
client does is retrieving the block: The server would still have the
cost of reading the block, but it wouldn't be transferred. But when the
client doesn't follow up with a read, it's quite a bit of overhead that
we had for no benefit. Returning NBD_STATE_ALLOCATED might be more
appropriate in this case than scanning for zeros.

Kevin

pgpR_qmddlASq.pgp
Description: PGP signature