Re: [Xen-devel] [PATCH RFC] x86/altp2m: fix display frozen when switching to a new view early

[George Dunlap <george.dunlap@xxxxxxxxxx>]

On 09/03/2018 09:25 AM, Razvan Cojocaru wrote:
> When an new altp2m view is created very early on guest boot, the
> display will freeze (although the guest will run normally). This
> may also happen on resizing the display. The reason is the way
> Xen currently (mis)handles logdirty VGA: it intentionally
> misconfigures VGA pages so that they will fault.
> 
> The problem is that it only does this in the host p2m. Once we
> switch to a new altp2m, the misconfigured entries will no longer
> fault, so the display will not be updated.

Hey Razvan, thanks for doing this, and sorry it's taken so long to respond.

> This patch:
> 
> * updates ept_handle_misconfig() to use the active altp2m instead
>   of the hostp2m;

This is probably necessary.

> * has p2m_init_altp2m_ept() copy over max_mapped_pfn,
>   logdirty_ranges, global_logdirty, ept.ad and default_access
>   from the hostp2m (the latter more for completeness than for any
>   other reason).

I think this is probably the right approach.  These values change
rarely, but after a misconfig are read repeatedly.  So it's probably a
lot more efficient to propagate changes when they happen, rather than
trying to keep a single master copy.  However...

> We should discuss if just copying over
>   logdirty_ranges (which is a pointer) is sufficient, or if
>   this code requires more synchronization).

It's clearly not sufficient. :-)  The logdirty_ranges struct is
protected by the lock of the p2m structure that contains it; if you
point to it from a different p2m structure, then you'll have
inconsistent logging, and you'll have problems if one vcpu is reading
the structure while another is modifying it.

Another issue is that it doesn't look like you're propagating updates to
this shared state either -- if someone enables or disables
global_logdirty, or changes default_access, the altp2ms will still have
the old value.

I wonder if we should collect the various bits that need to be kept in
sync between hostp2m/altp2ms, put them all in a 'sync' sub-struct within
the p2m, and enforce using a function / macro to modify the values inside.

> Also, it's worth
>   clarifying if these variables (and which of them) should be
>   copied over from the hostp2m or the currently active p2m;
> * modifies p2m_change_entry_type_global() and
>   p2m_change_type_range() to propagate their changes to all
>   valid altp2ms.
> 
> Another aspect is that, while new modifications should work with
> these changes, _old_ modifications (written to the host2pm
> _before_ we've created the new altp2m) will, if I understand the
> code correctly be lost. That is to say, misconfigurations
> performed before p2m_init_altp2m_ept() in the hostp2m will
> presumably not trigger the necessary faults after switching to
> the new altp2m.

You're worried about the following sequence?

1. Misconfigure hostp2m
2. Enable altp2m
3. Switch to altp2m 1
4. Fault on a previously-misconfigured p2m entry

Actually, I'm pretty sure you don't have to worry about this (unless
I've completely forgotten how things work). Correct me if I'm wrong:

* The altp2ms start out as empty, and entries are copied from the host
p2m as needed, using hostp2m->get_entry() (at some level)

* There are a *lot* of users that call p2m->get_entry() without causing
a fault; these callers need to get the right values during a
misconfigure (the remnants of which may last indefinitely -- i.e.,
misconfigured entries may *never* be fixed up for regions of the p2m
which aren't touched)

So the fault in 4 should end up copying over the correct value (unless I
missed something).

Just one comment on the code...

>  static void ept_enable_pml(struct p2m_domain *p2m)
>  {
> +    unsigned int i;
> +    struct domain *d = p2m->domain;
> +
>      /* Domain must have been paused */
> -    ASSERT(atomic_read(&p2m->domain->pause_count));
> +    ASSERT(atomic_read(&d->pause_count));
>  
>      /*
>       * No need to return whether vmx_domain_enable_pml has succeeded, as
>       * ept_p2m_type_to_flags will do the check, and write protection will be
>       * used if PML is not enabled.
>       */
> -    if ( vmx_domain_enable_pml(p2m->domain) )
> +    if ( vmx_domain_enable_pml(d) )
>          return;
>  
>      /* Enable EPT A/D bit for PML */
>      p2m->ept.ad = 1;
> -    vmx_domain_update_eptp(p2m->domain);
> +
> +    if ( altp2m_active(d) )
> +        for ( i = 0; i < MAX_ALTP2M; i++ )
> +        {
> +            if ( d->arch.altp2m_eptp[i] == mfn_x(INVALID_MFN) )
> +                continue;
> +
> +            p2m = d->arch.altp2m_p2m[i];
> +            p2m->ept.ad = 1;
> +        }

You're not grabbing the respective p2m locks here -- I'm pretty sure
this will end up being three separate instructions (read, set ad bit,
write).

But there would something a bit funny here about grabbing the main p2m
lock in p2m.c, and then grabbing altp2m locks within the function.  But
on the other hand, you clearly only want to call this...

> +    vmx_domain_update_eptp(d);

...once.  Some refactoring might be wanted.

 -George

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