Re: [Xen-devel] [PATCH] viridian: fix assertion failure

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Mon, Nov 26, 2018 at 03:08:50PM +0000, Paul Durrant wrote:
> > -----Original Message-----
> > From: Roger Pau Monne
> > Sent: 26 November 2018 15:04
> > To: Paul Durrant <Paul.Durrant@xxxxxxxxxx>
> > Cc: xen-devel@xxxxxxxxxxxxxxxxxxxx; Jan Beulich <jbeulich@xxxxxxxx>;
> > Andrew Cooper <Andrew.Cooper3@xxxxxxxxxx>; Wei Liu <wei.liu2@xxxxxxxxxx>
> > Subject: Re: [PATCH] viridian: fix assertion failure
> > 
> > On Mon, Nov 26, 2018 at 02:54:54PM +0000, Paul Durrant wrote:
> > > Whilst attempting to crash an apparently wedged Windows domain using
> > > 'xen-hvmcrash' I managed to trigger the following ASSERT:
> > >
> > > (XEN) Assertion '!vp->ptr' failed at viridian.c:607
> > >
> > > with stack:
> > >
> > > (XEN)    [<ffff82d08032c55d>] viridian_map_guest_page+0x1b4/0x1b6
> > > (XEN)    [<ffff82d08032b1db>] viridian_synic_load_vcpu_ctxt+0x39/0x3b
> > > (XEN)    [<ffff82d08032b90d>]
> > viridian.c#viridian_load_vcpu_ctxt+0x93/0xcc
> > > (XEN)    [<ffff82d0803096d6>] hvm_load+0x10e/0x19e
> > > (XEN)    [<ffff82d080274c6d>] arch_do_domctl+0xb74/0x25b4
> > > (XEN)    [<ffff82d0802068ab>] do_domctl+0x16f7/0x19d8
> > >
> > > This happened because viridian_map_guest_page() was not written to cope
> > > with being called multiple times, but this is unfortunately exactly what
> > > happens when xen-hvmcrash re-loads the domain context (having clobbered
> > > the values of RIP).
> > >
> > > This patch simply makes viridian_map_guest_page() return immediately if
> > it
> > > finds the page already mapped (i.e. vp->ptr != NULL).
> > >
> > > Signed-off-by: Paul Durrant <paul.durrant@xxxxxxxxxx>
> > > ---
> > > Cc: Jan Beulich <jbeulich@xxxxxxxx>
> > > Cc: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
> > > Cc: Wei Liu <wei.liu2@xxxxxxxxxx>
> > > Cc: "Roger Pau Monné" <roger.pau@xxxxxxxxxx>
> > > ---
> > >  xen/arch/x86/hvm/viridian/viridian.c | 3 ++-
> > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/xen/arch/x86/hvm/viridian/viridian.c
> > b/xen/arch/x86/hvm/viridian/viridian.c
> > > index 7d73f41de6..b99501eea2 100644
> > > --- a/xen/arch/x86/hvm/viridian/viridian.c
> > > +++ b/xen/arch/x86/hvm/viridian/viridian.c
> > > @@ -604,7 +604,8 @@ void viridian_map_guest_page(struct vcpu *v, struct
> > viridian_page *vp)
> > >      unsigned long gmfn = vp->msr.fields.pfn;
> > >      struct page_info *page = get_page_from_gfn(d, gmfn, NULL,
> > P2M_ALLOC);
> > >
> > > -    ASSERT(!vp->ptr);
> > > +    if ( vp->ptr )
> > > +        return;
> > 
> > Isn't this likely to get the page reference counting out of sync? You
> > will return without a put_page, and viridian_unmap_guest_page will set
> > vp->ptr = NULL and call put_page_and_type without taking into account
> > the reference count.
> > 
> 
> Damn. Missed that the side effect was in the initializer. I'll need to move 
> it below the check.
> 
> > If viridian_map_guest_page needs to be called multiple times you
> > likely need some reference counting AFAICT, so that the last call to
> > viridian_unmap_guest_page is the one that releases the page and sets
> > vp->ptr = NULL.
> 
> I shouldn't need to reference count. When the MSR is set by the guest there 
> is always an unmap then a map. It is just the context load case that should 
> only map on the first call (to handle resume correctly).

OK, I haven't looked at the usage, but if there's a paired number of
maps/unmaps adding a reference counter to viridian_page is fairly
trivial and would make the implementation of map/unmap always correct
regardless of it's usage.

Thanks, Roger.

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