Re: [PATCH 15/25] argo: implement the sendv op

["Jan Beulich" <JBeulich@suse.com>]

>>> On 04.01.19 at 09:13, <christopher.w.clark@gmail.com> wrote:
> ok, I'm at the point where I'm close to having a version three of the
> series to post that addresses all the feedback so far, plus some
> additional improvements, with the following two items remaining to
> discuss:
> 
> 1) the domain_cookie, with Jan's question about a) its exclusion of
> mismatches and b) its utility.
> 
> Given the expressed concern that the timer-based cookie initialization
> does not necessarily exclude mismatches, I've reimplemented it as a
> simple 128-bit counter protected by the L1 lock: this does now exclude
> mismatches.

... for all practical purposes, I assume you mean. In which case
I'd then immediately ask whether a 64-bit counter wouldn't do
as well.

> The utility of the cookie follows from this:
> 
> domid, despite its name, is not a unique domain identifier; it's a
> temporally unique id: Xen will ensure that no two domains that execute
> concurrently have the same domid. Domain authentication needs to take
> this into account.

Correct, at which point the question arises whether domain IDs
aren't too narrow. After all this isn't the first time we run into such
a restriction - see the opt_ibpb related code in context_switch().

> With Argo, it affects these points:
> 
> * ring registration: when the partner domain domid is specified, argo
> finds the currently executing domain with that domid, and needs to
> be able to confirm that it is the same domain later when a sendv is
> issued.
> 
> * sendv: needs to confirm that the domain sending a message is the same
> as the single domain authorized to transmit when the ring was first
> registered.
> 
> * notify: the querying domain asks about free space, and if there's not
> enough then a record is kept internal to the hypervisor, and a signal
> will be sent to the caller later when sufficient space becomes
> available.  Before sending the signal, Xen needs to confirm that the
> current domain with the domid it remembered is the same as the one that
> issued the query, otherwise Xen is sending spurious signals to domains
> that are not expecting it (and unless it checks, may not even be
> argo-enabled).
> 
> * domain teardown: in the absence of the domain cookie, or an
> alternative data structure that achieves the same ability to
> distinguish a reincarnated domain, all the rings that are registered
> that authorize the dying domid to send need to be torn down with
> suitable notification to their owners, and all the pending signals for
> that domain about available free space need to be nullified, to prevent
> a later domain inheriting these credentials and signals.
> 
> Doing so either entails a potentially-expensive walk of all rings of all
> domains, plus all the pending notifications on all rings the domain can
> access, or additional complexity with new data structures storing
> further metadata on the authorized domain on ring registration, etc.
> The domain cookie which enables identity confirmation on a domid is
> a reasonable alternative solution.

For all of these the question then is whether holding a reference
to the other domain (which has been looked up during ring
registration) wouldn't help. Furthermore this isn't a new problem,
see e.g. how event channel code deals with the ECS_INTERDOMAIN
case - without acquiring extra references, but instead with suitable
(and mutual) cleanup during domain destruction.

Jan



_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel