On Mon, Jun 16, 2025 at 11:52 AM Daniel P. Berrangé <berra...@redhat.com> wrote:
>
> On Mon, Jun 16, 2025 at 11:25:54AM +0200, Ilya Dryomov wrote:
> > On Thu, May 15, 2025 at 1:29 PM Fiona Ebner <f.eb...@proxmox.com> wrote:
> > >
> > > Currently, most Ceph configuration options are not exposed via QAPI.
> > > While it is possible to specify a dedicated Ceph configuration file,
> > > specialized options are often only required for a selection of images
> > > on the RBD storage, not all of them. To avoid the need to generate a
> > > dedicated Ceph configuration file for each image (or for each required
> > > combination of options), support a selection of key-value pairs via
> > > QAPI.
> > >
> > > Initially, this is just 'rbd_cache_policy'. For example, this is
> > > useful with small images used as a pflash for EFI variables. Setting
> > > the 'rbd_cache_policy' to 'writeback' yields a substantial improvement
> > > there [0].
> > >
> > > The function qemu_rbd_extract_key_value_pairs() was copied/adapted
> > > from the existing qemu_rbd_extract_encryption_create_options().
> > >
> > > [0]: https://bugzilla.proxmox.com/show_bug.cgi?id=3329#c9
> > >
> > > Signed-off-by: Fiona Ebner <f.eb...@proxmox.com>
>
> snip
>
> > > ##
> > > # @BlockdevOptionsRbd:
> > > #
> > > @@ -4327,6 +4360,9 @@
> > > # authentication. This maps to Ceph configuration option "key".
> > > # (Since 3.0)
> > > #
> > > +# @key-value-pairs: Key-value pairs for additional Ceph configuraton.
> > > +# (Since 10.1)
> > > +#
> > > # @server: Monitor host address and port. This maps to the "mon_host"
> > > # Ceph option.
> > > #
> > > @@ -4342,6 +4378,7 @@
> > > '*user': 'str',
> > > '*auth-client-required': ['RbdAuthMode'],
> > > '*key-secret': 'str',
> > > + '*key-value-pairs' : 'RbdKeyValuePairs',
> >
> > To side-step all of the above, have you considered implementing
> > a straightforward passthrough to Ceph instead? Something like
> >
> > '*key-value-pairs': ['RbdKeyValuePair']
> >
> > where RbdKeyValuePair is just a pair arbitrary strings (and
> > key-value-pairs is thus an optional list of those). rados_conf_set()
> > would be called just the same but the user would be able to override
> > any Ceph option they wish, not just a few that we thought of here.
>
> Passing through arbitrary key/value pairs as strings is essentially
> abdicating our design responsibility in QAPI. enums would no longer
> be introspectable. Integers / booleans would require abnormal formatting
> by clients. API stability / deprecation promises can no longer be made.
> and more besides.
>
> Given that limitation, if we did go the string pairs route, I would
> expect it to be marked as "unstable" in the QAPI schema, so apps have
> a suitable warning NOT to rely on this.
This sounds sensible to me. We can continue exposing the most common
Ceph options through a proper QAPI schema but add key-value-pairs as an
alternative low-level route for those who want to avoid dealing with
physical configuration files.
Thanks,
Ilya
