Re: [Qemu-devel] [PATCH] vl: Move cpu_synchronize_all_states() into qemu_system_reset()

[Laurent Vivier <lvivier@redhat.com>]

On 04/04/2016 06:49, David Gibson wrote:
> There are currently 3 calls to qemu_system_reset() in vl.c.  Two of them
> are immediately preceded by a cpu_synchronize_all_states9) and the
> remaining one should be.
> 
> The one which doesn't is the very first reset called directly from main().
> Without a cpu_synchronize_all_states(), kvm_vcpu_dirty is false at this
> point from the earlier cpu_synchronize_all_post_init().  That's incorrect
> because the reset path is quite likely to update the CPU state, and that
> updated state should be pushed back to KVM, not overwritten with stale
> data pushed to KVM immediately after init.
> 
> This patch moves the call to cpu_synchronize_all_states() into
> qemu_system_reset() for safety, so it is always called.  AFAICT this should
> be safe for the handful of callers outside vl.c - these all appear to be in
> places where the cpu state is already synchronized so the extra call
> will be a no-op.
> 
> Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
> ---
>  vl.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> This fixes a real but on ppc - the incorrect state of kvm_vcpu_dirty
> means that an extra cpu_synchronize_state() in revised code for
> updating the MSR is not a no-op as expected and loads a stale MSR
> value, resulting in an incorrect MSR value on entry to the guest.
> 
> Therefore, I'm hoping to apply this ASAP to 2.6.
> 
> Laurent, could you please verify that this does indeed address the
> problem with an incorrect MSR on entry.

This fixes the unset SF bit of MSR in my test case, so:

Tested-by: Laurent Vivier <lvivier@redhat.com>

Laurent