Re: [PATCH] xen: reuse the same pirq allocated when driver load first time

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Mon, May 13, 2013 at 06:24:46PM +0100, Stefano Stabellini wrote:
> On Mon, 13 May 2013, Konrad Rzeszutek Wilk wrote:
> > On Mon, May 13, 2013 at 03:50:52PM +0100, Stefano Stabellini wrote:
> > > On Mon, 13 May 2013, Konrad Rzeszutek Wilk wrote:
> > > > On Mon, May 13, 2013 at 12:06:43PM +0100, Stefano Stabellini wrote:
> > > > > On Fri, 10 May 2013, Konrad Rzeszutek Wilk wrote:
> > > > > > On Wed, May 08, 2013 at 04:18:24PM +0800, Zhenzhong Duan wrote:
> > > > > > > When driver load and unload in a loop, pirq will exhaust finally.
> > > > > > > Try to use the same pirq which was already mapped and binded at first time
> > > > > > 
> > > > > > So what happens if I unload and reload two drivers in random order?
> > > > > > 
> > > > > > > when driver loaded.
> > > > > > > 
> > > > > > > Read pirq from msix entry and test if data is XEN_PIRQ_MSI_DATA
> > > > > > > xen_irq_from_pirq(pirq) < 0 checking is wrong as irq will be freed
> > > > > > > when driver unload, it's always true in second load.
> > > > > > 
> > > > > > If my understanding is right the issue at hand is that the caching
> > > > > > information about the pirq disappears once the driver has been
> > > > > > unloaded b/c the event's irq-info is removed (as the driver is
> > > > > > unloaded and free_irq is called).
> > > > > > 
> > > > > > Stefano,
> > > > > > Is there a specific write to the MSI structure that would cause the
> > > > > > hypervisor to drop the PIRQ? Or a nice hypercall to "free" an
> > > > > > PIRQ in usage?
> > > > > 
> > > > > We already have a "free PIRQ" hypercall, it's called
> > > > > PHYSDEVOP_unmap_pirq and should be called by QEMU.
> > > > 
> > > > Considering that we call function that allocates (PHYSDEVOP_get_free_pirq)
> > > > it in the Linux kernel (and not in QEMU), perhaps that should be done in the
> > > > Linux kernel as part of xen_destroy_irq()? Or would that confuse QEMU?
> > > 
> > > I think it would confuse QEMU. It is probably better to let the unmap
> > > being handled by it.
> > > 
> > > 
> > > > It looks like QEMU only does that hypercall (via xc_physdev_unmap_pirq)
> > > > unregister_real_device which is only called during pci unplug?
> > > 
> > > You are right! I would think that this behaviour is erroneous unless it
> > > was done on purpose to avoid allocating MSIs twice.
> > > If that is the case we would need to do something similar in Linux too.
> > > 
> > > I think that the issue is the mismatch between QEMU's and Linux's
> > > behaviours: either both should be allocating MSIs once, or they should
> > > both be allocating and deallocating MSIs every time the driver is loaded
> > > and unloaded.
> > 
> > Right. But we also have the scenario that QEMU and Linux are going to 
> > be out of sync. So we need fixes in both places - I think.
> 
> QEMU is the owner of the pirq, in fact it is the one that creates and
> destroys the mapping. I think that the right place to fix this problem
> is in QEMU, the ABI would be much cleaner as a result. As a side effect
> we don't need to make any changes in Linux.

You do. You need to remove the PHYSDEVOP_get_free_pirq call in that case.
> 
> > > 
> > > > > Linux should disable the MSI bit in the PCI config options of the
> > > > > device:
> > > > > 
> > > > > pci_disable_msi -> pci_msi_shutdown -> msi_set_enable(0)
> > > > 
> > > > Zhenzhong, does it do that? Looking at the driver it certainly seems that way.
> > > > > 
> > > > > That should cause QEMU to issue a xc_physdev_unmap_pirq that actually
> > > > > unmaps the pirq. If it doesn't, it's a bug :)
> > > > 
> > > > <sigh> It doesn't do that. So two bugs:
> > > >  - QEMU doing that
> > > >  - Linux kernel doing the hypercall as well.
> > > 
> > > At first sight I also thought that the Linux kernel was issuing
> > > PHYSDEVOP_unmap_pirq too but actually Linux is only doing it if
> > > xen_initial_domain().
> > 
> > That seems like an easy fix. Just do 'if (xen_initial_domain()
> > || xen_hvm_domain())'. I think the only one we cannot do it for
> > is 'xen_pv_domain()' (so PCI in PV guests) as the "owner" of the PIRQ is
> > actually dom0.
> 
> We could do that and it might be easier than changing QEMU, but I think
> it would be more clear and consistent if QEMU was the one to do the
> unmapping.

And allocating the PIRQ?