On Thu, 19 Nov 2020 at 17:25, David Brazdil <dbrazdil@xxxxxxxxxx> wrote:
>
> KVM nVHE code runs under a different VA mapping than the kernel, hence
> so far it relied only on PC-relative addressing to avoid accidentally
> using a relocated kernel VA from a constant pool (see hyp_symbol_addr).
>
> So as to reduce the possibility of a programmer error, fixup the
> relocated addresses instead. Let the kernel relocate them to kernel VA
> first, but then iterate over them again, filter those that point to hyp
> code/data and convert the kernel VA to hyp VA.
>
> This is done after kvm_compute_layout and before apply_alternatives.
>
If this is significant enough to call out, please include the reason for it.
> Signed-off-by: David Brazdil <dbrazdil@xxxxxxxxxx>
> ---
> arch/arm64/include/asm/kvm_mmu.h | 1 +
> arch/arm64/kernel/smp.c | 4 +-
> arch/arm64/kvm/va_layout.c | 76 ++++++++++++++++++++++++++++++++
> 3 files changed, 80 insertions(+), 1 deletion(-)
>
> diff --git a/arch/arm64/include/asm/kvm_mmu.h b/arch/arm64/include/asm/kvm_mmu.h
> index 5168a0c516ae..e5226f7e4732 100644
> --- a/arch/arm64/include/asm/kvm_mmu.h
> +++ b/arch/arm64/include/asm/kvm_mmu.h
> @@ -105,6 +105,7 @@ alternative_cb_end
> void kvm_update_va_mask(struct alt_instr *alt,
> __le32 *origptr, __le32 *updptr, int nr_inst);
> void kvm_compute_layout(void);
> +void kvm_fixup_hyp_relocations(void);
>
> static __always_inline unsigned long __kern_hyp_va(unsigned long v)
> {
> diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
> index 18e9727d3f64..30241afc2c93 100644
> --- a/arch/arm64/kernel/smp.c
> +++ b/arch/arm64/kernel/smp.c
> @@ -434,8 +434,10 @@ static void __init hyp_mode_check(void)
> "CPU: CPUs started in inconsistent modes");
> else
> pr_info("CPU: All CPU(s) started at EL1\n");
> - if (IS_ENABLED(CONFIG_KVM))
> + if (IS_ENABLED(CONFIG_KVM)) {
> kvm_compute_layout();
> + kvm_fixup_hyp_relocations();
> + }
> }
>
> void __init smp_cpus_done(unsigned int max_cpus)
> diff --git a/arch/arm64/kvm/va_layout.c b/arch/arm64/kvm/va_layout.c
> index d8cc51bd60bf..b80fab974896 100644
> --- a/arch/arm64/kvm/va_layout.c
> +++ b/arch/arm64/kvm/va_layout.c
> @@ -10,6 +10,7 @@
> #include <asm/alternative.h>
> #include <asm/debug-monitors.h>
> #include <asm/insn.h>
> +#include <asm/kvm_asm.h>
> #include <asm/kvm_mmu.h>
> #include <asm/memory.h>
>
> @@ -82,6 +83,81 @@ __init void kvm_compute_layout(void)
> init_hyp_physvirt_offset();
> }
>
> +#define __load_elf_u64(s) \
> + ({ \
> + extern u64 s; \
> + u64 val; \
> + \
> + asm ("ldr %0, =%1" : "=r"(val) : "S"(&s)); \
> + val; \
> + })
> +
Do you need this to ensure that the reference is absolute? There may
be more elegant ways to achieve that, using weak references for
instance.
Also, in the relocation startup code, I deliberately used a 32-bit
quantity here, as it won't get confused for an absolute virtual
address that needs relocation.
> +static bool __is_within_bounds(u64 addr, char *start, char *end)
> +{
> + return start <= (char*)addr && (char*)addr < end;
> +}
> +
> +static bool __is_in_hyp_section(u64 addr)
> +{
> + return __is_within_bounds(addr, __hyp_text_start, __hyp_text_end) ||
> + __is_within_bounds(addr, __hyp_rodata_start, __hyp_rodata_end) ||
> + __is_within_bounds(addr,
> + CHOOSE_NVHE_SYM(__per_cpu_start),
> + CHOOSE_NVHE_SYM(__per_cpu_end));
> +}
> +
It is slightly disappointing that we need to filter these one by one
like this, but I don't think there are any guarantees about the order
in which the R_AARCH64_RELATIVE entries appear.
> +static void __fixup_hyp_rel(u64 addr)
__init ?
> +{
> + u64 *ptr, kern_va, hyp_va;
> +
> + /* Adjust the relocation address taken from ELF for KASLR. */
> + addr += kaslr_offset();
> +
> + /* Skip addresses not in any of the hyp sections. */
> + if (!__is_in_hyp_section(addr))
> + return;
> +
> + /* Get the LM alias of the relocation address. */
> + ptr = (u64*)kvm_ksym_ref((void*)addr);
> +
> + /*
> + * Read the value at the relocation address. It has already been
> + * relocated to the actual kernel kimg VA.
> + */
> + kern_va = (u64)kvm_ksym_ref((void*)*ptr);
> +
> + /* Convert to hyp VA. */
> + hyp_va = __early_kern_hyp_va(kern_va);
> +
> + /* Store hyp VA at the relocation address. */
> + *ptr = __early_kern_hyp_va(kern_va);
> +}
> +
> +static void __fixup_hyp_rela(void)
__init ?
> +{
> + Elf64_Rela *rel;
> + size_t i, n;
> +
> + rel = (Elf64_Rela*)(kimage_vaddr + __load_elf_u64(__rela_offset));
> + n = __load_elf_u64(__rela_size) / sizeof(*rel);
> +
> + for (i = 0; i < n; ++i)
> + __fixup_hyp_rel(rel[i].r_offset);
> +}
> +
> +/*
> + * The kernel relocated pointers to kernel VA. Iterate over relocations in
> + * the hypervisor ELF sections and convert them to hyp VA. This avoids the
> + * need to only use PC-relative addressing in hyp.
> + */
> +__init void kvm_fixup_hyp_relocations(void)
It is more idiomatic to put the __init after the 'void', and someone
is undoubtedly going to send a patch to 'fix' that if we merge it like
this.
> +{
> + if (!IS_ENABLED(CONFIG_RELOCATABLE) || has_vhe())
> + return;
> +
> + __fixup_hyp_rela();
> +}
> +
> static u32 compute_instruction(int n, u32 rd, u32 rn)
> {
> u32 insn = AARCH64_BREAK_FAULT;
> --
> 2.29.2.299.gdc1121823c-goog
>
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
