[Qemu-devel] Re: [PATCH 6/6] QEMU support for the Kernel Virtual Machine interface

[Fabrice Bellard <fabrice@bellard.org>]

Anthony Liguori wrote:
> This patch actually enables KVM support for QEMU.  I apologize that it is so
> large but this was the only sane way to preserve bisectability.
> 
> The goal of this patch is to add KVM support, but not to impact users when
> KVM isn't being used.  It achieves this by using a kvm_enabled() macro that
> evaluates to (0) when KVM support is not enabled.  An if (kvm_enabled()) is
> just as good as using an #ifdef since GCC will eliminate the dead code.
> 
> This patches touches a lot of areas.  For performance reasons, the guest CPU
> state is not kept in sync with CPUState.  This requires an explicit
> synchronization whenever CPUState is required.  KVM also uses it's own main
> loop as it runs each VCPU in it's own thread.
> 
> Trapping VGA updates via MMIO is far too slow when running KVM so there is
> additional logic to allow VGA memory to be accessed as RAM.  We use KVM's
> shadow page tables to keep track of which portions of RAM have been dirtied.
> 
> KVM also supports an in-kernel APIC implementation as a performance
> enhancement.  Finally, KVM supports APIC TPR patching.  This allows TPR
> accesses (which are very frequently for Windows) to be patches into CALL
> instructions to the BIOS (for 32-bit guests).  This results in a very
> sigificant performance improvement for Windows guests.
> 
> While this patch is very large, the new files are only included when KVM
> support is compiled in.  Every change to QEMU is wrapped in an
> if (kvm_enabled()) so the code disappears when KVM support is not compiled in.
> This is done to ensure no regressions are introduced to normal QEMU.

Some questions:

- QEMU already maintains modified page status for VGA memory (and kqemu 
for example fully supports that), so I don't see why KVM needs a new method.

- Why is kvm_cpu_register_physical_memory() needed ? kqemu can work 
without it because there is a remapping between physical memory and RAM 
address. I suggest to add that feature in KVM or to modify 
cpu_register_physical_memory() to hide it.

- If KVM implements its own CPU loop, why are there patches in libqemu.a 
(CPU core) ?

Regards,

Fabrice.