Re: [XEN PATCH v2 2/5] xen: export get_free_port

[Julien Grall <julien@xxxxxxx>]

Hi Stefano,

On 25/01/2022 22:49, Stefano Stabellini wrote:
> On Tue, 25 Jan 2022, Jan Beulich wrote:
> > On 25.01.2022 02:10, Stefano Stabellini wrote:
> > > On Sun, 23 Jan 2022, Julien Grall wrote:
> > > > > diff --git a/xen/common/event_channel.c b/xen/common/event_channel.c
> > > > > index da88ad141a..5b0bcaaad4 100644
> > > > > --- a/xen/common/event_channel.c
> > > > > +++ b/xen/common/event_channel.c
> > > > > @@ -232,7 +232,7 @@ int evtchn_allocate_port(struct domain *d, evtchn_port_t
> > > > > port)
> > > > >        return 0;
> > > > >    }
> > > > >    -static int get_free_port(struct domain *d)
> > > > > +int get_free_port(struct domain *d)
> > > > I dislike the idea to expose get_free_port() (or whichever name we decide)
> > > > because this can be easily misused.
> > > >
> > > > In fact looking at your next patch (#3), you are misusing it as it is meant to
> > > > be called with d->event_lock. I know this doesn't much matter
> > > > in your situation because this is done at boot with no other domains running
> > > > (or potentially any event channel allocation). However, I still think we
> > > > should get the API right.
> > > >
> > > > I am also not entirely happy of open-coding the allocation in domain_build.c.
> > > > Instead, I would prefer if we provide a new helper to allocate an unbound
> > > > event channel. This would be similar to your v1 (I still need to review the
> > > > patch though).
> > > I am happy to go back to v1 and address feedback on that patch. However,
> > > I am having difficulties with the implementation. Jan pointed out:
> > >
> > >
> > > > > -
> > > > > -    chn->state = ECS_UNBOUND;
> > > > This cannot be pulled ahead of the XSM check (or in general anything
> > > > potentially resulting in an error), as check_free_port() relies on
> > > > ->state remaining ECS_FREE until it is known that the calling function
> > > > can't fail anymore.
> > > This makes it difficult to reuse _evtchn_alloc_unbound for the
> > > implementation of evtchn_alloc_unbound. In fact, I couldn't find a way
> > > to do it.
> > >
> > > Instead, I just create a new public function called
> > > "evtchn_alloc_unbound" and renamed the existing funtion to
> > > "_evtchn_alloc_unbound" (this to addresses Jan's feedback that the
> > > static function should be the one starting with "_"). So the function
> > > names are inverted compared to v1.
> > >
> > > Please let me know if you have any better suggestions.
> > >
> > >
> > > diff --git a/xen/common/event_channel.c b/xen/common/event_channel.c
> > > index da88ad141a..c6b7dd7fbd 100644
> > > --- a/xen/common/event_channel.c
> > > +++ b/xen/common/event_channel.c
> > > @@ -18,6 +18,7 @@
> > >   
> > >   #include <xen/init.h>
> > >   #include <xen/lib.h>
> > > +#include <xen/err.h>
> > >   #include <xen/errno.h>
> > >   #include <xen/sched.h>
> > >   #include <xen/irq.h>
> > > @@ -284,7 +285,27 @@ void evtchn_free(struct domain *d, struct evtchn *chn)
> > >       xsm_evtchn_close_post(chn);
> > >   }
> > >   
> > > -static int evtchn_alloc_unbound(evtchn_alloc_unbound_t *alloc)
> > > +struct evtchn *evtchn_alloc_unbound(struct domain *d, domid_t remote_dom)
> > > +{
> > > +    struct evtchn *chn;
> > > +    int port;
> > > +
> > > +    if ( (port = get_free_port(d)) < 0 )
> > > +        return ERR_PTR(port);
> > > +    chn = evtchn_from_port(d, port);
> > > +
> > > +    evtchn_write_lock(chn);
> > > +
> > > +    chn->state = ECS_UNBOUND;
> > > +    chn->u.unbound.remote_domid = remote_dom;
> > > +    evtchn_port_init(d, chn);
> > > +
> > > +    evtchn_write_unlock(chn);
> > > +
> > > +    return chn;
> > > +}
> > > +
> > > +static int _evtchn_alloc_unbound(evtchn_alloc_unbound_t *alloc)
> > >   {
> > >       struct evtchn *chn;
> > >       struct domain *d;
> > Instead of introducing a clone of this function (with, btw, still
> > insufficient locking), did you consider simply using the existing
> > evtchn_alloc_unbound() as-is, i.e. with the caller passing
> > evtchn_alloc_unbound_t *?
> Yes, we tried that first. Unfortunately the (dummy) XSM check cannot
> work. This is how we would want to call the function:
>
>
>      alloc.dom = d->domain_id;
>      alloc.remote_dom = hardware_domain->domain_id;
>      rc = evtchn_alloc_unbound(&alloc);
>
>
> This is the implementation of the XSM check:
>
> static XSM_INLINE int xsm_evtchn_unbound(
>      XSM_DEFAULT_ARG struct domain *d, struct evtchn *chn, domid_t id2)
> {
>      XSM_ASSERT_ACTION(XSM_TARGET);
>      return xsm_default_action(action, current->domain, d);
> }
>
>
> Note the usage of current->domain. If you have any suggestions on how to
> fix it please let me know.
If I am not mistaken, current should still point to a domain (in this 
case idle).
So one alternative would be to ignore XSM if current->domain == idle and 
the system is booting (this could be part of xsm_default_action())
Another alternative would be to switch current to another domain. 'dom0' 
wouldn't be a solution because it doesn't exist for "true" dom0less. So 
a possibility would be to use dom_xen or create a fake build domain to 
be used for XSM check during boot.
Cheers,

--
Julien Grall