Re: [Qemu-devel] [Nbd] [PATCH v2] doc: Add NBD_CMD_BLOCK_STATUS extension

[Wouter Verhelst]

On Thu, Apr 07, 2016 at 10:10:58AM -0600, Eric Blake wrote:
> On 04/07/2016 04:38 AM, Vladimir Sementsov-Ogievskiy wrote:
> > On 05.04.2016 16:43, Paolo Bonzini wrote:
> >>
> >> On 05/04/2016 06:05, Kevin Wolf wrote:
> >>> The options I can think of is adding a request field "max number of
> >>> descriptors" or a flag "only single descriptor" (with the assumption
> >>> that clients always want one or unlimited), but maybe you have a better
> >>> idea.
> >> I think a limit is better.  Even if the client is ultimately going to
> >> process the whole file, it may take a very long time and space to
> >> retrieve all the descriptors in one go.  Rather than query e.g. 16GB at
> >> a time, I think it's simpler to put a limit of 1024 descriptors or so.
> >>
> >> Paolo
> >>
> > 
> > I vote for the limit too. More over, I think, there should be two sides
> > limit:
> > 
> > 1. The client can specify the limit, so server should not return more
> > extents than requested. Of course, server should chose sequential
> > extents from the beginning of requested range.
> 
> For the client to request a limit would entail that we enhance the
> protocol to allow structured requests (where a wire-sniffer would know
> how many bytes to read for the client's additional data, even if it does
> not understand the extension's semantics).  Might not be a bad idea to
> have this in the long run, but so far I've been reluctant to bite the
> bullet.
> 
> > 2. Server side limit: if client asked too many extents or not specified
> > a limit at all, server should not return all extents, but only 1024 (for
> > ex.) from the beginning of the range.
> 
> Okay, I'm fairly convinced now that letting the server limit the reply
> is a good thing, and that one doesn't require a structured request from
> the client.  Since we just recently documented that strings should be no
> more than 4096 bytes, and my v2 proposal used 8 bytes per descriptor,
> maybe a good way to enforce a similar limit would be:
> 
> The server MAY choose to send fewer descriptors than what would describe
> the full extent of the client's request, but MUST send at least one
> descriptor unless an error is reported.  The server MUST NOT send more
> than 512 descriptors, even if that does not completely describe the
> client's requested length.
> 
> That way, a client in general should never expect more than ~4096 bytes
> + overhead on any server reply except a reply to NBD_CMD_READ, and can
> therefore utilize stack allocation for all other replies (if we do this,
> maybe we should make a hard rule that all future protocol extensions,
> other than NBD_CMD_READ, will guarantee that a reply has a bounded size)

The downside is more requests. A client-requested limit on the number of
replies sounds like a good idea, but a protocol-requested limited
doesn't, to me.

We could have a flag that turns the "length" field into "length in
number of described extents" rather than "length in number of bytes".
That way, the client could say "please give me metadata on X extents
starting at offset Y", rather than "please give me data on the extents
starting at offset X for Y bytes of length"

> I also think it may be okay to let the server reply with MORE data than
> the client requested, but only as long as it does not result in any
> extra descriptors (that is, only the last descriptor can result in a
> length beyond the client's request).  For example, if the client asks
> for block status of 1M of the file, but the server can conveniently
> learn via lseek(SEEK_HOLE) or other means that there are 2M of data
> before status changes, then there's no reason to force the server to
> throw away the information about the 1M beyond the client's read, and
> the client might even be able to be more efficient in later requests.

Yes, definitely. I had originally wanted to suggest that rather than my
above as a way to allow clients to do limited-extent requests (by just
declaring the above and stating that asking information about an offset
with zero-length is legal), but I think my other suggestion might be
more interesting.

> > 2.1 And/or, why not allow the server use the power of structured reply
> > and send several reply chunks? Why did you forbid this? (if I correctly
> > understand "This chunk type MUST appear at most once in a structured
> > reply.")
> 
> If we allow more than one chunk, then either every chunk has to include
> an offset (more traffic over the wire),

We can state that every chunk (as in, description of multiple extents
sent in one structured reply) has an offset, and that all extents
described in that chunk MUST be sequential.

> or the chunks have to be sent in
> a particular order (we aren't gaining any benefits that NBD_CMD_READ
> gains by allowing out-of-order transmission).  It's also more work for
> the client to reconstruct if it has to reassemble; with NBD_CMD_READ,
> the payload is dominated by the data being read, and you can pwrite()
> the data into its final location as the client; but with
> NBD_CMD_BLOCK_STATUS, the payload is dominated by the metadata and we
> want to keep it minimal; and there is no convenient command for the
> client to reassemble the information if received out of order.

There is a DF flag, which could be allowed here as well.

> Allowing for a short reply seems to be worth doing, but allowing for
> multiple reply chunks seems not worth the risk.

Disagree.

[...]

-- 
< ron> I mean, the main *practical* problem with C++, is there's like a dozen
       people in the world who think they really understand all of its rules,
       and pretty much all of them are just lying to themselves too.
 -- #debian-devel, OFTC, 2016-02-12