Re: [Xen-devel] [PATCH] xen: reuse the same pirq allocated whendriver load first time

[Konrad Rzeszutek Wilk]

On Tue, May 21, 2013 at 10:50:09PM +0100, Stefano Stabellini wrote:
> On Tue, 21 May 2013, Konrad Rzeszutek Wilk wrote:
> > On Tue, May 21, 2013 at 05:51:02PM +0100, Stefano Stabellini wrote:
> > > On Tue, 21 May 2013, Konrad Rzeszutek Wilk wrote:
> > > > > Looking at the hypervisor code I couldn't see anything obviously wrong.
> > > > 
> > > > I think the culprit is "physdev_unmap_pirq":
> > > > 
> > > >    if ( is_hvm_domain(d) )                                                     
> > > >     {                                                                           
> > > >         spin_lock(&d->event_lock);                                              
> > > >         gdprintk(XENLOG_WARNING,"d%d, pirq: %d is %x %s, irq: %d\n",            
> > > >             d->domain_id, pirq, domain_pirq_to_emuirq(d, pirq),                 
> > > >             domain_pirq_to_emuirq(d, pirq) == IRQ_UNBOUND ? "unbound" : "",        
> > > >             domain_pirq_to_irq(d, pirq));                                       
> > > >                                                                                 
> > > >         if ( domain_pirq_to_emuirq(d, pirq) != IRQ_UNBOUND )                    
> > > >             ret = unmap_domain_pirq_emuirq(d, pirq);                            
> > > >         spin_unlock(&d->event_lock);                                            
> > > >         if ( domid == DOMID_SELF || ret )                                       
> > > >             goto free_domain;                                             
> > > > 
> > > > It always tells me unbound:
> > > > 
> > > > (XEN) physdev.c:237:d14 14, pirq: 54 is ffffffff
> > > > (XEN) irq.c:1873:d14 14, nr_pirqs: 56
> > > > (XEN) physdev.c:237:d14 14, pirq: 53 is ffffffff
> > > > (XEN) irq.c:1873:d14 14, nr_pirqs: 56
> > > > (XEN) physdev.c:237:d14 14, pirq: 52 is ffffffff
> > > > (XEN) irq.c:1873:d14 14, nr_pirqs: 56
> > > > (XEN) physdev.c:237:d14 14, pirq: 51 is ffffffff
> > > > (XEN) irq.c:1873:d14 14, nr_pirqs: 56
> > > > (XEN) physdev.c:237:d14 14, pirq: 50 is ffffffff
> > > > (XEN) irq.c:1873:d14 14, nr_pirqs: 56
> > > > (a bit older debug code, so the 'unbound' does not show up here).
> > > > 
> > > > Which means that the call to unmap_domain_pirq_emuirq does not happen.
> > > > The checks in unmap_domain_pirq_emuirq also look to be depend
> > > > on the code being IRQ_UNBOUND.
> > > > 
> > > > In other words, all of that code looks to only clear things when
> > > > they are !IRQ_UNBOUND.
> > > > 
> > > > But the other logic (IRQ_UNBOUND) looks to be missing a removal
> > > > in the radix tree:
> > > > 
> > > >   if ( emuirq != IRQ_PT )                                                     
> > > >         radix_tree_delete(&d->arch.hvm_domain.emuirq_pirq, emuirq);             
> > > >                                                                         
> > > > And I think that is what is causing the leak - the radix tree
> > > > needs to be pruned? Or perhaps the allocate_pirq should check
> > > > the radix tree for IRQ_UNBOUND ones and re-use them?
> > > 
> > > I think that you are looking in the wrong place.
> > > The issue is that QEMU doesn't call pt_msi_disable in
> > > pt_msgctrl_reg_write if (!val & PCI_MSI_FLAGS_ENABLE).
> > 
> > In my test-case I am not even calling QEMU. I am just doing two hypercalls 
> > hypercall - get_free_pirq and unmap.
> > > 
> > > The code above is correct as is because it is trying to handle emulated
> > > IRQs and MSIs, not real passthrough MSIs. They latter are not added to
> > > that radix tree, see physdev_hvm_map_pirq and physdev_map_pirq.
> > 
> > The bug is in the hypervisor. This little patch solves the test-case
> > (I hadn't tried to do the PCI passthrough yet)
> > 
> > 
> > diff --git a/xen/arch/x86/irq.c b/xen/arch/x86/irq.c
> > index b0b0c65..b78717a 100644
> > --- a/xen/arch/x86/irq.c
> > +++ b/xen/arch/x86/irq.c
> > @@ -1851,8 +1851,8 @@ static int pirq_guest_force_unbind(struct domain *d, struct pirq *pirq)
> >  static inline bool_t is_free_pirq(const struct domain *d,
> >                                    const struct pirq *pirq)
> >  {
> > -    return !pirq || (!pirq->arch.irq && (!is_hvm_domain(d) ||
> > -        pirq->arch.hvm.emuirq == IRQ_UNBOUND));
> > +    return !pirq || ((pirq->arch.irq == 0 || (pirq->arch.irq == PIRQ_ALLOCATED)) &&
> > +           (!is_hvm_domain(d) || pirq->arch.hvm.emuirq == IRQ_UNBOUND));
> >  }
> >  
> >  int get_free_pirq(struct domain *d, int type)

To be fair, this diff is just to demonstrate that the pirq->arch.irq is
the one that seems to gate things. I am not suggesting that this is the
final patch - just the 'aha, this is what is happening!'.
> > 
> > 
> > The reason is that pirq->arch.irq in PHYSDEVOP_get_free_pirq is set to
> > from the value of zero to -1 (PIRQ_ALLOCATED). Then in map_domain_pirq
> > we check it first:
> > 
> > 904     old_irq = domain_pirq_to_irq(d, pirq);                  
> > .. snip..                    
> > 1907     if ( (old_irq > 0 && (old_irq != irq) ) ||                                  
> > 
> > and since the 'old_irq' is -1 (or zero), and the irq passed in
> > is different, then all checks pass and the value is over-written:
> > 
> >  1988         set_domain_irq_pirq(d, irq, info);                                      
> > 
> > And that is it.
> 
> 
> We have to be careful about this: the point of PHYSDEVOP_get_free_pirq is
> that Linux can know for sure the pirq that is going to be used to map the
> MSI by QEMU. If you modify is_free_pirq that way, suddenly the pirq
> could be allocated for something else after Linux called
> PHYSDEVOP_get_free_pirq and before QEMU called xc_physdev_map_pirq_msi.

Yes. And I think the 'is_free_pirq' modification above is incorrect.

I think the fix should be in the unmap_pirq code (hypervisor) to check
if the arch.irq is either zero or PIRQ_ALLOCATED. Right now it only
checks for zero.

Then as you say there is also xc_physdev_map_pirq_msi, but the call
chain looks to be INTx-MSIx. The pt_msi_update is the call that is
used when guest writes the PIRQ (xc_domain_update_msi_irq -> XEN_DOMCTL_bind_pt_irq)

And that looks to be it. The pt_msi_update can be called multiple times if
the guest decides to use a different PIRQ.

> 
> Right now the unmap is supposed to be done by QEMU, not Linux. So I
> think that it is "normal" (although counterintuitive) that your little
> test works that way.

Yes, the test-case is flawed.
> 
> pirq allocated via PHYSDEVOP_get_free_pirq should be passed to QEMU,
> mapped by QEMU, unmapped by QEMU and eventually freed by QEMU.
> 
> This is not the bestest interface ever written of course but that's how
> it works now.

That is kindly said :-)
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/