Re: [Xen-devel] [PATCH v10 01/32] ARM: vGIC: avoid rank lock when reading priority

[Julien Grall <julien.grall@xxxxxxx>]

On 05/31/2017 11:42 AM, Julien Grall wrote:
> Hi Stefano,
>
> On 30/05/17 22:39, Stefano Stabellini wrote:
> > On Tue, 30 May 2017, Julien Grall wrote:
> > > Hi Andre,
> > >
> > > On 26/05/17 18:35, Andre Przywara wrote:
> > > > When reading the priority value of a virtual interrupt, we were taking
> > > > the respective rank lock so far.
> > > > However for forwarded interrupts (Dom0 only so far) this may lead to a
> > > > deadlock with the following call chain:
> > > > - MMIO access to change the IRQ affinity, calling the ITARGETSR handler
> > > > - this handler takes the appropriate rank lock and calls
> > > > vgic_store_itargetsr()
> > > > - vgic_store_itargetsr() will eventually call vgic_migrate_irq()
> > > > - if this IRQ is already in-flight, it will remove it from the old
> > > >   VCPU and inject it into the new one, by calling 
> > > > vgic_vcpu_inject_irq()
> > > > - vgic_vcpu_inject_irq will call vgic_get_virq_priority()
> > > > - vgic_get_virq_priority() tries to take the rank lock - again!
> > > > It seems like this code path has never been exercised before.
> > > >
> > > > Fix this by avoiding taking the lock in vgic_get_virq_priority() 
> > > > (like we
> > > > do in vgic_get_target_vcpu()).
> > > > Actually we are just reading one byte, and priority changes while
> > > > interrupts are handled are a benign race that can happen on real 
> > > > hardware
> > > > too. So it looks safe to just use read_atomic() instead.
> > > >
> > > > Signed-off-by: Andre Przywara <andre.przywara@xxxxxxx>
> > > > ---
> > > >  xen/arch/arm/vgic.c | 8 +-------
> > > >  1 file changed, 1 insertion(+), 7 deletions(-)
> > > >
> > > > diff --git a/xen/arch/arm/vgic.c b/xen/arch/arm/vgic.c
> > > > index 83569b0..54b2aad 100644
> > > > --- a/xen/arch/arm/vgic.c
> > > > +++ b/xen/arch/arm/vgic.c
> > > > @@ -227,14 +227,8 @@ struct vcpu *vgic_get_target_vcpu(struct vcpu *v,
> > > > unsigned int virq)
> > > >  static int vgic_get_virq_priority(struct vcpu *v, unsigned int virq)
> > > >  {
> > > >      struct vgic_irq_rank *rank = vgic_rank_irq(v, virq);
> > > > -    unsigned long flags;
> > > > -    int priority;
> > > > -
> > > > -    vgic_lock_rank(v, rank, flags);
> > > > -    priority = rank->priority[virq & INTERRUPT_RANK_MASK];
> > > > -    vgic_unlock_rank(v, rank, flags);
> > > >
> > > > -    return priority;
> > > > +    return read_atomic(&rank->priority[virq & INTERRUPT_RANK_MASK]);
> > > The write in rank->priority will not be atomic (see vgic_reg_update
> > > implementation): the register is first masked, the the priority set.
> > >
> > > So you may end up to read 0 (which is the higher priority) by mistake.
> > >
> > > We should probably think to make vgic_reg_* helper atomic.
> > Right! That's why I wrote
> > alpine.DEB.2.10.1705231135130.18759@sstabellini-ThinkPad-X260
> It was not obvious from this e-mail why you wanted this. It looked more 
> a request for improvement rather than a potential bug. Hence my 
> suggestion to delay it.
FIY, I wrote down a couple of patch to make vgic_reg* atomic. It is not 
too invasive and would solve write atomically in the register. However, 
it will not prevent two write racing together. So the lock would still 
be need in some places.
I am still doing some testing. I will send a version hopefully Monday.

Cheers,

--
Julien Grall

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel