On 14/12/20 23:29, Andy Lutomirski wrote:
One downside to that is that we'd need to do the VMSAVE on every
iteration of vcpu_run(), as opposed to just once when we enter
from userspace via KVM_RUN. It ends up being a similar situation to
Andy's earlier suggestion of moving VMLOAD just after vmexit, but
in that case we were able to remove an MSR write to MSR_GS_BASE,
which cancelled out the overhead, but in this case I think it could
only cost us extra.
If you want to micro-optimize, there is a trick you could play: use
WRGSBASE if available. If X86_FEATURE_GSBASE is available, you could
use WRGSBASE to restore GSBASE and defer VMLOAD to vcpu_put(). This
would need benchmarking on Zen 3 to see if it’s worthwhile.
If the improvement is big (100 cycles or so) it would be worthwhile.
Otherwise, for the sake of code clarity doing VMLOAD in the assembly
code is the simplest.
Paolo
