Am 06.11.2019 um 14:11 hat Max Reitz geschrieben:
> On 17.10.19 15:01, Kevin Wolf wrote:
> > Add a --export option to qemu-storage-daemon to export a block node. For
> > now, only NBD exports are implemented. Apart from the 'type' option
> > (which is the implied key), it maps the arguments for nbd-server-add to
> > the command line. Example:
> > 
> >     --export nbd,device=disk,name=test-export,writable=on
> > 
> > Signed-off-by: Kevin Wolf <kwolf@redhat.com>
> > ---
> >  qapi/block.json       | 27 +++++++++++++++++++++++++++
> >  qemu-storage-daemon.c | 31 +++++++++++++++++++++++++++++++
> >  2 files changed, 58 insertions(+)
> 
> Would it be better to collect the BlockExports in a list and work on it
> after all arguments have been parsed?  As it is, it’s important that
> users define block devices and things like NBD servers before --export.
>  Yes, I know, that’s exactly how it works with qemu, but is that really
> the best way?

It's actually not how QEMU works generally. QEMU collects things in
QemuOptsLists and then tries to interpret them in the right order. Of
course, we never get the order actually right, which results in constant
reshuffling of the order of initialisations in vl.c.

It also means that vl.c (!) has a list of -object types that need to be
created early so that other backends can make use of them, and of those
types that actually depend on a backend already being present (see
object_create_initial() for details).

I think it's much cleaner to simply use the order in the command line
instead of adding magic that tries to resolve (and fails at actually
resolving) all the dependencies. I seem to remember that this was in
fact one of the things Markus keeps mentioning he would change if he
were to rewrite the QEMU command line parser from scratch without
compatibility requirements.

Kevin