[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH for-2.9 4/5] rbd: Peel off redundant RbdAuthMeth

From: Jeff Cody
Subject: Re: [Qemu-devel] [PATCH for-2.9 4/5] rbd: Peel off redundant RbdAuthMethod wrapper struct
Date: Thu, 23 Mar 2017 23:55:37 -0400
User-agent: Mutt/1.5.24 (2015-08-30)

On Thu, Mar 23, 2017 at 04:56:30PM -0500, Eric Blake wrote:
> On 03/23/2017 04:43 PM, Eric Blake wrote:
> > We'd still have to allow libvirt's use of
> > ":key=...:auth_supported=cephx\\;none" on the command line, but from the
> > QMP side, we would support exactly one auth method, and libvirt will be
> > updated to match when it starts targetting blockdev-add instead of
> > legacy command line.
> > 
> > If librados really needs 'cephx;none' any time cephx authorization is
> > requested, then even though the QMP only requests 'cephx', we can still
> > map it to 'cephx;none' in qemu - but I'm hoping that setting
> > auth_supported=cephx rather than auth_supported=cephx;none on the
> > librados side gives us what we (and libvirt) really want in the first place.
> Or, if it becomes something that libvirt wants to allow users to tune in
> their XML (right now, users don't get a choice, they get either 'none'
> or 'cephx;none'), a forward-compatible solution under my QMP proposal
> would be to have qemu add a third enum state, "none", "cephx", and
> "cephx-no-fallback".
> On the other hand, if supporting multiple methods at once makes sense
> (say librados adds a 'cephy' method, and that users could legitimately
> use both 'cephx;cephy' and 'cephy;cephx' with different behaviors), then
> keeping auth as an array of instances of a simple flat union scales
> nicer than having to add a combinatorial explosion of branches to the
> flat union to cover every useful combination of auth_supported methods.
> Maybe I'm overthinking it.

I spoke over email with Jason Dillaman (cc'ed), and this is what he told me
with regards to the authentication methods, and what cephx;none means:

On Thu, Mar 23, 2017 at 06:04:30PM -0400, Jason Dillaman wrote:
> It's saying that the client is willing to connect to a server that
> uses cephx auth or a server that uses no auth. Looking at the code,
> the "auth_supported" configuration key is actually deprecated and
> "auth_client_required" should be used instead (which defaults to
> 'cephx, none' already). Since it's been deprecated since v0.55 and if
> you are cleaning things up, feel free to switch to the new one if
> possible.

So from what Jason is saying, it seems like the conclusion we reached over
IRC is correct: it will attempt cephx but also fallback to no auth.

Also, since the preferred auth option may be named different depending on
ceph versions, I know think we probably should not tie the QAPI parameter to
the name of the older deprecated option.

I suggest that the 'auth_supported' parameter for QAPI be renamed to
'auth_method'.  If you don't like the array and the flexibility it provides
at the cost of complexity, I think a flat enum consisting of 3 values like
you mentioned is reasonable.  Since the QAPI does not need to map to the
legacy commandline used by libvirt, I would suggest maybe naming them
slightly different, though: any, none, strict

For 2.9, they could each map to 'auth_supported' like so:
    any:     "cephx;none"
    none:    "none"
    strict:  "cephx"

For 2.10, we could try to discover if 'auth_client_required' is supported or
not, and use either it or 'auth_supported' as appropriate (with the same
mappings as above).

The reason I like "any" and "strict", is it gives consistent meanings to
options even if new auth methods are introduced.  For a hypothetical "cephy"
method example:

    any:     "cephy;cephx;none"
    none:    "none"
    strict:  "cephy;cephx"


reply via email to

[Prev in Thread] Current Thread [Next in Thread]