Re: [PATCH v2 1/2] arm64/crash_core: Export KERNELPACMASK in vmcoreinfo

[Will Deacon]

On Mon, Apr 27, 2020 at 11:55:01AM +0530, Amit Daniel Kachhap wrote:
> Recently arm64 linux kernel added support for Armv8.3-A Pointer
> Authentication feature. If this feature is enabled in the kernel and the
> hardware supports address authentication then the return addresses are
> signed and stored in the stack to prevent ROP kind of attack. Kdump tool
> will now dump the kernel with signed lr values in the stack.
> 
> Any user analysis tool for this kernel dump may need the kernel pac mask
> information in vmcoreinfo to generate the correct return address for
> stacktrace purpose as well as to resolve the symbol name.
> 
> This patch is similar to commit ec6e822d1a22d0eef ("arm64: expose user PAC
> bit positions via ptrace") which exposes pac mask information via ptrace
> interfaces.
> 
> Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
> Cc: Will Deacon <will@xxxxxxxxxx>
> Cc: Mark Rutland <mark.rutland@xxxxxxx>
> Signed-off-by: Amit Daniel Kachhap <amit.kachhap@xxxxxxx>
> ---
> Changes since v1:
> * Rebased to kernel 5.7-rc3.
> * commit log change.
> 
> An implementation of this new KERNELPACMASK vmcoreinfo field used by crash
> tool can be found here[1]. This change is accepted by crash utility
> maintainer [2].
> 
> [1]: https://www.redhat.com/archives/crash-utility/2020-April/msg00095.html
> [2]: https://www.redhat.com/archives/crash-utility/2020-April/msg00099.html
> 
>  arch/arm64/include/asm/compiler.h | 3 +++
>  arch/arm64/kernel/crash_core.c    | 4 ++++
>  2 files changed, 7 insertions(+)
> 
> diff --git a/arch/arm64/include/asm/compiler.h b/arch/arm64/include/asm/compiler.h
> index eece20d..32d5900 100644
> --- a/arch/arm64/include/asm/compiler.h
> +++ b/arch/arm64/include/asm/compiler.h
> @@ -19,6 +19,9 @@
>  #define __builtin_return_address(val)					\
>  	(void *)(ptrauth_clear_pac((unsigned long)__builtin_return_address(val)))
>  
> +#else  /* !CONFIG_ARM64_PTR_AUTH */
> +#define	ptrauth_user_pac_mask()		0ULL
> +#define	ptrauth_kernel_pac_mask()	0ULL

This doesn't look quite right to me, since you still have to take into
account the case where CONFIG_ARM64_PTR_AUTH=y but the feature is not
available at runtime:

> @@ -16,4 +17,7 @@ void arch_crash_save_vmcoreinfo(void)
>  	vmcoreinfo_append_str("NUMBER(PHYS_OFFSET)=0x%llx\n",
>  						PHYS_OFFSET);
>  	vmcoreinfo_append_str("KERNELOFFSET=%lx\n", kaslr_offset());
> +	vmcoreinfo_append_str("NUMBER(KERNELPACMASK)=0x%llx\n",
> +						system_supports_address_auth() ?
> +						ptrauth_kernel_pac_mask() : 0);

In which case, would it make more sense to define
ptrauth_{kernel,user}_pac_mask() unconditionally? In fact, I'd probably
just remove the guards completely from asm/compiler.h because I think
they're misleading.

Will

--->8

diff --git a/arch/arm64/include/asm/compiler.h b/arch/arm64/include/asm/compiler.h
index eece20d2c55f..51a7ce87cdfe 100644
--- a/arch/arm64/include/asm/compiler.h
+++ b/arch/arm64/include/asm/compiler.h
@@ -2,8 +2,6 @@
 #ifndef __ASM_COMPILER_H
 #define __ASM_COMPILER_H
 
-#if defined(CONFIG_ARM64_PTR_AUTH)
-
 /*
  * The EL0/EL1 pointer bits used by a pointer authentication code.
  * This is dependent on TBI0/TBI1 being enabled, or bits 63:56 would also apply.
@@ -19,6 +17,4 @@
 #define __builtin_return_address(val)					\
 	(void *)(ptrauth_clear_pac((unsigned long)__builtin_return_address(val)))
 
-#endif /* CONFIG_ARM64_PTR_AUTH */
-
 #endif /* __ASM_COMPILER_H */