Re: [PATCH 0/6] IRQFD without IRQ routing, enabled for XICS

[Alexander Graf <agraf@suse.de>]

On 30.06.14 12:51, Paul Mackerras wrote:
> This series of patches provides a way to implement IRQFD support
> without having to implement IRQ routing, and adds IRQFD support for
> the XICS interrupt controller emulation.  (XICS is the interrupt
> controller defined for the pSeries machine type, used on IBM POWER
> servers).
>
> The basic approach is to make it easy for code other than irqchip.c to
> provide a mapping from a global interrupt number (GSI) to an irq
> routing entry (struct kvm_kernel_irq_routing_entry).  To make the
> lifetime of this routing entry easier to manage, we change the IRQFD
> code to keep a copy of the routing entry (for the MSI fast-path case)
> rather than a pointer to the routing entry.  Since the copy can't be
> updated atomically, we add a seqcount_t to make sure that when reading
> it we get a copy that hasn't been half-way updated.
>
> Next we replace the hard-coded accesses outside irqchip.c to the
> fields of the kvm_irq_routing_table struct with calls to accessor
> functions in irqchip.c, namely kvm_irq_map_gsi() and
> kvm_irq_map_chip_pin().  That enables us to move all references to the
> kvm_irq_routing_table struct, and the definition of that struct, into
> irqchip.c.
>
> Then we move the irq notifier implementation from irqchip.c into
> eventfd.c and add a separate Kconfig option to enable IRQFD.  With
> that we can enable IRQFD without irq routing, which we achieve by
> compiling in eventfd.c but not irqchip.c, and providing an alternative
> implementation of kvm_irq_map_gsi() and kvm_irq_map_chip_pin().
>
> The last patch does that for XICS.  With this series I can use
> vhost-net with KVM guests, and I see the TCP bandwidth between guest
> and host on a POWER8 machine go from around 700MB/s to over 2GB/s.
>
> I would like to see this go into 3.17.

I think this approach is good enough for XICS, as it has a 100% flat 
number space. I'm not yet fully convinced how well this would work out 
with the GIC, but I know that one too little to make claims.

The only thing I disliked about this patch set is that it adds another 
piece of code that makes MPIC/XICS exclusive options. But I don't think 
that'd be hard to solve if we start caring.

Overall, nice work.


Alex