答复: [PATCH] kasan: avoid out-of-bounds in unwind_frame

On Wed, Oct 10, 2018 at 06:45:17AM +0000, Chunhui Li (???) wrote:

Hi Mark,

kasan detect out-of-bounds in stacktrace.c line 70, it's already over READ_ONCE_NOCHECK, but still crash
kernel-4.9/arch/arm64/kernel/stacktrace.c
69frame->sp = fp + 0x10;
70frame->fp = READ_ONCE_NOCHECK(*(unsigned long *)(fp));

we test on Android platform, kernel-4.9 build with clang 6.0.2, we will do experiment to clarify whether compiler related issue.

READ_ONCE_NOCHECK->__read_once_size_nocheck with __no_sanitize_address if enable CONFIG_KASAN

kernel-4.9/include/linux/compiler-gcc.h
#define __no_sanitize_address __attribute__((no_sanitize_address))

kernel-4.9/include/linux/compiler-clang.h
#define __no_sanitize_address __attribute__((no_sanitize("address")))

If READ_ONCE_NOCHECK() isn't working as intended, then this is a
compiler issue, and there are bigger problems regardless.

If we need to workaround some compiler issue, that should be done in the
implementation of READ_ONCE_NOCHECK().

Please note that there are other issues with using clang to compile the
kernel; at minimum 7.0.0 is required to build a mainline kernel (and
even then, the LSE atomics have to be explicitly disabled), so if this
only affects 6.0.x, I don't think we need to carry any workaround
upstream.

there is patch from internet, avoid kasan by wrapping with kasan_disable_current, it seems better.
https://lore.kernel.org/patchwork/patch/644463
such as:
+kasan_disable_current();
  // access fp
+kasan_enable_current();

Please note that the solution later in the thread [1] was to use
READ_ONCE_NOCHECK().

Thanks,
Mark.

[1] https://lore.kernel.org/patchwork/patch/644463/#829858