On 20 May 2014 23:05, Christoffer Dall [off-list ref] wrote:

On Tue, May 20, 2014 at 10:16:22PM +0800, Haojian Zhuang wrote:

quoted

On 20 May 2014 22:01, Christoffer Dall [off-list ref] wrote:
It's the implementation of gich_apr in arm32.

We needn't add or change anything in struct vgic_cpu. And both the
assembly code and the code could be much easier.

But we do end up with an extra memory access from EL2 in the critical
path, and I believe Marc's concern here is that if we cross a cache
line, this might really hurt performance.

Sorry. Do we may cross a cache line or a TLB entry?

I think that you're concerning to cross TLB entries. The reason is in
 below.

1. If the problem is on crossing cache line, it's caused by too much
instructions. Either the packing nr_lr or the gich_apr adds some
instructions. The packing nr_lr needs a little more instructions.

2. ldr instruction is a pseudo instruction. So it's parsed into operation
on PC register. Now I put gich_apr in interrupts_head.S, it results
in gich_apr variable before __kvm_hyp_code_start. It may cross the
TLB entries.
How about to declare gich_apr after __kvm_cpu_return in interrupts.S?
Since save_vgic_state & restore_vgic_state is only used once, declaring
gich_apr just after the code could avoid crossing TLB entry.

Regards
Haojian