Re: [PATCH 0/2]: ARM: Enable THREAD_INFO_IN_TASK

[Kees Cook <keescook@chromium.org>]

On Thu, Sep 02, 2021 at 05:53:53PM +0100, Russell King (Oracle) wrote:
> On Thu, Sep 02, 2021 at 08:54:26AM -0700, Keith Packard wrote:
> > Placing thread_info in the kernel stack leaves it vulnerable to stack
> > overflow attacks. This short series addresses that by using the
> > existing THREAD_INFO_IN_TASK infrastructure.
> > 
> > As this is my first patch in this part of the kernel, I'm looking for
> > feedback about the general approach as well as specific comments on
> > places where I've missed something.
> > 
> > I've only run this on armhf running under qemu, so while I've tried to
> > make patches for other code paths, I haven't been able to test those.
> > 
> > (yes, I know checkpatch.pl complains about whitespace in asm-offsets.c, I
> > decided to leave the existing whitespace alone)
> > 
> > Signed-off-by: Keith Packard <keithpac@amazon.com>
> 
> I think you're introducing a circular dependency with this for
> certain kernel configurations:
> 
> E.g. Have you tried running this with CONFIG_CPU_V6 enabled?
> 
> +#define raw_smp_processor_id() this_cpu_read(cpu_number)
> +#define __smp_processor_id() __this_cpu_read(cpu_number)
> +
> +DECLARE_PER_CPU_READ_MOSTLY(unsigned int, cpu_number);
> 
> this_cpu_read() is defined as:
> 
> #define this_cpu_read(pcp) __pcpu_size_call_return(this_cpu_read_, pcp)
> (which will call this_cpu_read_4)
> 
> #define this_cpu_read_4(pcp)            this_cpu_generic_read(pcp)
> => __this_cpu_generic_read_nopreempt()
> =>         ___ret = READ_ONCE(*raw_cpu_ptr(&(pcp)));
> 
> #define raw_cpu_ptr(ptr)                                                \
> ({                                                                      \
>         __verify_pcpu_ptr(ptr);                                         \
>         arch_raw_cpu_ptr(ptr);                                          \
> })
> 
> #ifndef arch_raw_cpu_ptr
> #define arch_raw_cpu_ptr(ptr) SHIFT_PERCPU_PTR(ptr, __my_cpu_offset)
> #endif
> 
> #ifndef __my_cpu_offset
> #define __my_cpu_offset per_cpu_offset(raw_smp_processor_id())
> #endif
> 
> ... which then leads back to a use of raw_smp_processor_id(), thereby
> creating a circular loop of preprocessor definitions that the compiler
> can't resolve.

If this isn't easy to fix, perhaps this can be a V7-only feature?

-- 
Kees Cook