[Qemu-devel] [BUG] Missing feature, KVM on ARM, dirty page logging.

[Qemu-devel] [BUG] Missing feature, KVM on ARM, dirty page logging.

[Ian Molton]

Hi folks,

I'm looking at dirty page logging on KVM on ARM, which appears, at 
present to be non-existent. Is anyone working on this, or willing to 
lend a hand?

I'm running KVM on an OMAP5, and my guest is a vexpress-a15 with PL-111 
framebuffer.

running tcg mode, this works, however in KVM I was seeing (occasionally) 
one frame of video.

Some debugging led me to the dirty_log code, and I discovered that qemu 
makes a bad test for the return code from kvm_vm_ioctl(s, 
KVM_GET_DIRTY_LOG, &d) in kvm_physical_sync_dirty_bitmap(), so the 
problem was going unreported.

Furthermore, I think the vexpress emulation was missing a call to
memory_region_set_log(vram, true, DIRTY_MEMORY_VGA);
for the vram region, which was preventing the host kernel knowing that 
dirty page logging was required. (I think - Im not really very clear on 
the whole qemu memory API, as AFAICT, its barely documented...)

Anyhow, If I add the call to memory_region_set_log(), and if I ask the 
host kernel to mark absolutely everything as dirty, it does in fact, work.

Obviously this is a massive hack, but I'm not too sure where to go on 
from here, so pointers would be nice. Presumably, I'm going to have to 
do some prodding of the hosts page tables, but I'm not clear exactly 
what, as yet.

-Ian

Re: [Qemu-devel] [BUG] Missing feature, KVM on ARM, dirty page logging.

[Peter Maydell]

On 26 November 2012 17:58, Ian Molton <ian.molton@codethink.co.uk> wrote:
> I'm looking at dirty page logging on KVM on ARM, which appears, at present
> to be non-existent. Is anyone working on this, or willing to lend a hand?

I've cc'd the kvmarm mailing list to give this a slightly wider airing.

> I'm running KVM on an OMAP5, and my guest is a vexpress-a15 with PL-111
> framebuffer.
>
> running tcg mode, this works, however in KVM I was seeing (occasionally) one
> frame of video.

Yeah, not sure if anybody's tested video in a KVM guest.

> Some debugging led me to the dirty_log code, and I discovered that qemu
> makes a bad test for the return code from kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG,
> &d) in kvm_physical_sync_dirty_bitmap(), so the problem was going
> unreported.
>
> Furthermore, I think the vexpress emulation was missing a call to
> memory_region_set_log(vram, true, DIRTY_MEMORY_VGA);
> for the vram region, which was preventing the host kernel knowing that dirty
> page logging was required. (I think - Im not really very clear on the whole
> qemu memory API, as AFAICT, its barely documented...)
>
> Anyhow, If I add the call to memory_region_set_log(), and if I ask the host
> kernel to mark absolutely everything as dirty, it does in fact, work.
>
> Obviously this is a massive hack, but I'm not too sure where to go on from
> here, so pointers would be nice. Presumably, I'm going to have to do some
> prodding of the hosts page tables, but I'm not clear exactly what, as yet.

-- PMM