On Wed, May 29, 2024 at 01:12:08PM +0100, Pierre-Clément Tosi wrote:
> Fix the mismatch between the (incorrect) C signature, C call site, and
> asm implementation by aligning all three on an API passing the
> parameters (pgd and SP) separately, instead of as a bundled struct.
>
> Remove the now unnecessary memory accesses while the MMU is off from the
> asm, which simplifies the C caller (as it does not need to convert a VA
> struct pointer to PA) and makes the code slightly more robust by
> offsetting the struct fields from C and properly expressing the call to
> the C compiler (e.g. type checker and kCFI).
>
> Fixes: f320bc742bc2 ("KVM: arm64: Prepare the creation of s1 mappings at EL2")
> Signed-off-by: Pierre-Clément Tosi <ptosi@xxxxxxxxxx>
> ---
> arch/arm64/include/asm/kvm_hyp.h | 3 +--
> arch/arm64/kvm/hyp/nvhe/hyp-init.S | 17 +++++++++--------
> arch/arm64/kvm/hyp/nvhe/setup.c | 4 ++--
> 3 files changed, 12 insertions(+), 12 deletions(-)
>
> diff --git a/arch/arm64/include/asm/kvm_hyp.h b/arch/arm64/include/asm/kvm_hyp.h
> index 3e80464f8953..58b5a2b14d88 100644
> --- a/arch/arm64/include/asm/kvm_hyp.h
> +++ b/arch/arm64/include/asm/kvm_hyp.h
> @@ -123,8 +123,7 @@ void __noreturn __hyp_do_panic(struct kvm_cpu_context *host_ctxt, u64 spsr,
> #endif
>
> #ifdef __KVM_NVHE_HYPERVISOR__
> -void __pkvm_init_switch_pgd(phys_addr_t phys, unsigned long size,
> - phys_addr_t pgd, void *sp, void *cont_fn);
> +void __pkvm_init_switch_pgd(phys_addr_t pgd, void *sp, void (*fn)(void));
> int __pkvm_init(phys_addr_t phys, unsigned long size, unsigned long nr_cpus,
> unsigned long *per_cpu_base, u32 hyp_va_bits);
> void __noreturn __host_enter(struct kvm_cpu_context *host_ctxt);
> diff --git a/arch/arm64/kvm/hyp/nvhe/hyp-init.S b/arch/arm64/kvm/hyp/nvhe/hyp-init.S
> index 2994878d68ea..d859c4de06b6 100644
> --- a/arch/arm64/kvm/hyp/nvhe/hyp-init.S
> +++ b/arch/arm64/kvm/hyp/nvhe/hyp-init.S
> @@ -265,33 +265,34 @@ alternative_else_nop_endif
>
> SYM_CODE_END(__kvm_handle_stub_hvc)
>
> +/*
> + * void __pkvm_init_switch_pgd(phys_addr_t pgd, void *sp, void (*fn)(void));
> + */
> SYM_FUNC_START(__pkvm_init_switch_pgd)
> /* Turn the MMU off */
> pre_disable_mmu_workaround
> - mrs x2, sctlr_el2
> - bic x3, x2, #SCTLR_ELx_M
> + mrs x9, sctlr_el2
> + bic x3, x9, #SCTLR_ELx_M
This is fine, but there's no need to jump all the way to x9 for the
register allocation. I think it would be neatest to re-jig the function
so it uses x4 here for the sctlr and then uses x5 later for the ttbr.
> diff --git a/arch/arm64/kvm/hyp/nvhe/setup.c b/arch/arm64/kvm/hyp/nvhe/setup.c
> index 859f22f754d3..1cbd2c78f7a1 100644
> --- a/arch/arm64/kvm/hyp/nvhe/setup.c
> +++ b/arch/arm64/kvm/hyp/nvhe/setup.c
> @@ -316,7 +316,7 @@ int __pkvm_init(phys_addr_t phys, unsigned long size, unsigned long nr_cpus,
> {
> struct kvm_nvhe_init_params *params;
> void *virt = hyp_phys_to_virt(phys);
> - void (*fn)(phys_addr_t params_pa, void *finalize_fn_va);
> + typeof(__pkvm_init_switch_pgd) *fn;
> int ret;
>
> BUG_ON(kvm_check_pvm_sysreg_table());
> @@ -340,7 +340,7 @@ int __pkvm_init(phys_addr_t phys, unsigned long size, unsigned long nr_cpus,
> /* Jump in the idmap page to switch to the new page-tables */
> params = this_cpu_ptr(&kvm_init_params);
> fn = (typeof(fn))__hyp_pa(__pkvm_init_switch_pgd);
> - fn(__hyp_pa(params), __pkvm_init_finalise);
> + fn(params->pgd_pa, (void *)params->stack_hyp_va, __pkvm_init_finalise);
Why not have the prototype of __pkvm_init_switch_pgd() take the SP as
an 'unsigned long' so that you can avoid this cast altogether?
Will
