Hi Marc, On 10/26/20 9:51 AM, Marc Zyngier wrote: > Registers x0/x1 get repeateadly pushed and poped during a host > HVC call. Instead, leave the registers on the stack, saving > a store instruction on the fast path for an add on the slow path. > > Signed-off-by: Marc Zyngier <maz@xxxxxxxxxx> > --- > arch/arm64/kvm/hyp/nvhe/host.S | 5 ++--- > 1 file changed, 2 insertions(+), 3 deletions(-) > > diff --git a/arch/arm64/kvm/hyp/nvhe/host.S b/arch/arm64/kvm/hyp/nvhe/host.S > index e2d316d13180..7b69f9ff8da0 100644 > --- a/arch/arm64/kvm/hyp/nvhe/host.S > +++ b/arch/arm64/kvm/hyp/nvhe/host.S > @@ -13,8 +13,6 @@ > .text > > SYM_FUNC_START(__host_exit) > - stp x0, x1, [sp, #-16]! > - > get_host_ctxt x0, x1 > > /* Store the host regs x2 and x3 */ > @@ -99,13 +97,14 @@ SYM_FUNC_END(__hyp_do_panic) > mrs x0, esr_el2 > lsr x0, x0, #ESR_ELx_EC_SHIFT > cmp x0, #ESR_ELx_EC_HVC64 > - ldp x0, x1, [sp], #16 > + ldp x0, x1, [sp] // Don't fixup the stack yet If I understand get_host_ctxt correctly, it will clobber x0 and x1, and this is the first thing that __host_exit does. I think that the values of x0 and x1 are only needed in host_el1_sync_vect: x0 to compare with HVC_STUB_HCALL_NR below, and x1 for the call to __kvm_handle_stub_hvc. I was thinking that we can restore x0 just before the comparison with HVC_STUB_HCALL_NR, after the first branch to __host_exit, to make it clear that it is not used by __host_exit. Not really important, but it might make the code a bit easier to understand (it looks a bit weird to me to have x0, x1 clobbered immediately after we restore them from the stack). Either way you prefer, the code looks correct to me: __host_exit assumes that x0 and x1 are at the top of the stack when it saves them, and the ADD in host_el1_sync_vect (when the code doesn't branch to __host_exit) makes sure the stack pointer is as expected: Reviewed-by: Alexandru Elisei <alexandru.elisei@xxxxxxx> Thanks, Alex > b.ne __host_exit > > /* Check for a stub HVC call */ > cmp x0, #HVC_STUB_HCALL_NR > b.hs __host_exit > > + add sp, sp, #16 > /* > * Compute the idmap address of __kvm_handle_stub_hvc and > * jump there. Since we use kimage_voffset, do not use the _______________________________________________ kvmarm mailing list kvmarm@xxxxxxxxxxxxxxxxxxxxx https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
