Libguestfs - Jan 2021 - [PATCH 4/6] v2v: rhv-upload-plugin: Support multiple connections

On 1/23/21 12:38 AM, Richard W.M. Jones wrote:
>> Sending flush request is more tricky; on imageio side, we have one
>> qemu-nbd server, with multiple connections. I'm not sure if sending
one
>> flush command on one of the connections is good enough to flush all
>> commands, so we send flush command on all connections in the flush
>> callback.
> 
> I know the answer to this!  It depends if the NBD server advertises
> multi-conn or not.  With libnbd you can find out by querying
> nbd_can_multi_conn on any of the connections (if the server is
> behaving, the answer should be identical for any connection with the
> same exportname).  See:
> 
>   https://github.com/NetworkBlockDevice/nbd/blob/master/doc/proto.md
> 
>   "bit 8, NBD_FLAG_CAN_MULTI_CONN: Indicates that the server operates
>   entirely without cache, or that the cache it uses is shared among
>   all connections to the given device. In particular, if this flag is
>   present, then the effects of NBD_CMD_FLUSH and NBD_CMD_FLAG_FUA MUST
>   be visible across all connections when the server sends its reply to
>   that command to the client. In the absence of this flag, clients
>   SHOULD NOT multiplex their commands over more than one connection to
>   the export."
> 
> For unclear reasons qemu-nbd only advertises multi-conn for r/o
> connections, assuming my reading of the code is correct.  For nbdkit
> we went through the plugins a long time ago and made them advertise
> (or not) multi-conn as appropriate.
For qemu-nbd, supporting FLAG_CAN_MULTI_CONN on a writable connection
would require work to audit that a flush request on one connection
properly affects all pending writes on other connections, as visible by
all subsequent reads from any of the connections.  This is not a trivial
task.

However, that's for the general case (when multiple writers are sending
requests that can overlap one another, and when you have to worry about
read consistency).  But when copying an image, you have a special case:
each writer is only touching distinct subsets of the overall image, and
you are not reading from the image.  As long as those subsets do not
overlap, and you do not trigger read-modify-write actions, you really
don't care whether flushes are consistent between writers, because there
are no consistency issues in the first place.

So we _might_ be able to optimize nbdcopy for the case of parallel
writes even when the NBD server does not advertise CAN_MULTI_CONN; the
writers can either use FLAG_FUA for every write, or _each_ write without
flushes and perform a final NBD_CMD_FLUSH; as long as EACH writer
performs a final flush before disconnecting (rather than trying to
optimize to just one writer flushing), you will be guaranteed that the
entire copied image is now consistent, even though you could not
guarantee flush consistency during the parallel writes.

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org

On Tue, Jan 26, 2021 at 08:51:59AM -0600, Eric Blake
wrote:
> However, that's for the general case (when multiple writers are sending
> requests that can overlap one another, and when you have to worry about
> read consistency).  But when copying an image, you have a special case:
> each writer is only touching distinct subsets of the overall image, and
> you are not reading from the image.  As long as those subsets do not
> overlap, and you do not trigger read-modify-write actions, you really
> don't care whether flushes are consistent between writers, because
there
> are no consistency issues in the first place.
> 
> So we _might_ be able to optimize nbdcopy for the case of parallel
> writes even when the NBD server does not advertise CAN_MULTI_CONN; the
> writers can either use FLAG_FUA for every write, or _each_ write without
> flushes and perform a final NBD_CMD_FLUSH; as long as EACH writer
> performs a final flush before disconnecting (rather than trying to
> optimize to just one writer flushing), you will be guaranteed that the
> entire copied image is now consistent, even though you could not
> guarantee flush consistency during the parallel writes.
It happens that nbdcopy does call nbd_flush on each handle.  (A case
of being "right but for the wrong reasons").  However we don't do
multi-conn unless the server advertises it, and in light of what you
say above perhaps we should.
> >   https://github.com/NetworkBlockDevice/nbd/blob/master/doc/proto.md
> > 
> >   "bit 8, NBD_FLAG_CAN_MULTI_CONN: Indicates that the server
operates
> >   entirely without cache, or that the cache it uses is shared among
> >   all connections to the given device. In particular, if this flag is
> >   present, then the effects of NBD_CMD_FLUSH and NBD_CMD_FLAG_FUA MUST
> >   be visible across all connections when the server sends its reply to
> >   that command to the client. In the absence of this flag, clients
> >   SHOULD NOT multiplex their commands over more than one connection to
> >   the export."
Although the text only mentions flush & FUA, I wonder if there's
another case where multi-conn should not be advertised or used: That
is where the block or sector size is larger than the size of writes,
so writes are being turned into r/m/w cycles.  Multiple adjacent (even
non-overlapping) writes could then interfere with each other.

This is vanishingly unlikely to affect nbdcopy which uses a huge block
size, but would it be a problem in the NBD protocol itself?

Rich.

-- 
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
Read my programming and virtualization blog: http://rwmj.wordpress.com
Fedora Windows cross-compiler. Compile Windows programs, test, and
build Windows installers. Over 100 libraries supported.
http://fedoraproject.org/wiki/MinGW