[PATCH v6 6/7] arm64: KVM: Set physical address size related factors in runtime
christoffer.dall at linaro.org
Tue May 27 07:02:41 PDT 2014
On Tue, May 27, 2014 at 03:53:49PM +0200, Christoffer Dall wrote:
> On Mon, May 12, 2014 at 06:40:54PM +0900, Jungseok Lee wrote:
> > This patch sets TCR_EL2.PS, VTCR_EL2.T0SZ and vttbr_baddr_mask in runtime,
> > not compile time.
> > In ARMv8, EL2 physical address size (TCR_EL2.PS) and stage2 input address
> > size (VTCR_EL2.T0SZE) cannot be determined in compile time since they
> > depends on hardware capability.
> > According to Table D4-23 and Table D4-25 in ARM DDI 0487A.b document,
> > vttbr_x is calculated using different hard-coded values with consideration
> super nit: I guess this is fixed values, and not hard-coded values
> > of T0SZ, granule size and the level of translation tables. Therefore,
> > vttbr_baddr_mask should be determined dynamically.
> so I think there's a deeper issue here, which is that we're not
> currently considering that for a given supported physical address size
> (run-time) and given page granularity (compile-time), we may have some
> flexibility in choosing the VTCR_EL2.SL0 field, and thereby the size of
> the initial stage2 pgd, by concatinating the initial level page tables.
> Additionally, the combinations of the givens may also force us to choose
> a specific SL0 value.
> Policy-wise, I would say we should concatenate as many initial level page
> tables as possible when using 4K pages, iow. always set VTCR_EL2.SL0 to
> the lowest possible value given the PARange and page size config we have
> at hand. That should always provide increased performance for VMs at
> the cost of maximum 16 concatenated tables, which is a 64K contiguous
> allocation and alignment, for 4K pages.
> For 64K pages, it becomes a 256K alignment and contiguous allocation
> requirement. One could argue that if this is not possible on your
> system, then you have no business runninng VMs on there, but I want to
> leave this open for comments...
Just had a brief chat with Marc, and he made me think of the fact that
we cannot decide this freely, because the code in kvm_mmu.c assumes that
the stage-2 page tables have the same number of levels etc. as the host
kernel (we re-use functions like pud_offset, pud_addr_end, etc. etc.).
I'm not sure this can always be aligned, so we may have to write our own
kvm_... versions of these to accomodate the best policy for KVM.
More information about the linux-arm-kernel