Re: [kernel-hardening] Re: [PATCH 4/4] refcount: Report failures through CHECK_DATA_CORRUPTION

[Kees Cook]

On Wed, Feb 8, 2017 at 1:12 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Tue, Feb 07, 2017 at 05:55:42PM +0000, Mark Rutland wrote:
>> On Tue, Feb 07, 2017 at 06:30:36PM +0100, Peter Zijlstra wrote:
>> > On Tue, Feb 07, 2017 at 04:03:01PM +0000, Mark Rutland wrote:
>> > > For x86 it's a little painful due to '%' in the register names, but it looks
>> > > possible. The below appears to do the mangling correctly (then screams due to
>> > > the mangled result being nonexistent).
>> >
>> > > asm(
>> > > " .macro  reg_to_offset   r\n"
>> > > " .irp rs,eax,ebx,ecx,edx\n"
>> > > " .ifc \\r, %\\rs\n"
>> > > " __offset_of_\\rs\n"
>> > > " .endif\n"
>> > > " .endr\n"
>> > > " .endm\n"
>> > > );
>> > >
>> > > #define asm_sym(var)              asm volatile("reg_to_offset %0\n" : : "r" (var))
>> >
>> > Oh gawd that's a most gnarly hack.
>>
>> :)
>>
>> > Do we want to go do that for all archs or somehow cook a generic
>> > fallback that ends up doing a full function call or something?
>>
>> Given the arch-specific reg->blah mapping is so "fun", I guess a generic
>> fallback would be a good start.
>>
>> I haven't figured out all the plumbing details. It'd be nice to reuse
>> the bug infrastructure so that arches don't have to implement another
>> trap and callback pair, but I guess the reg details need to live in
>> another data structure.
>
> On x86 have have __ex_table and __bug_table. The former is used for all
> sorts of things, including fixing up faults.
>
> Now, our struct exception_table_entry has a third field used to specify
> a handler, see commit:
>
>  548acf19234d ("x86/mm: Expand the exception table logic to allow new handling options")
>
> Also, given we trigger things with a known instruction at these sites,
> the ->to field is reusable and can be used to encode the register
> offset.
>
> Still, if we want to allow a generic implementation that does a function
> call, the handler prototype should probably look like:
>
>         void exception_value(unsigned long value);
>
> Which means the arch bits need a trampoline and we also need to encode
> that. The best I've come up with is having nr_regs trampolines and
> stuffing the trampoline function in the ->handler field and then using
> the ->to field to encode the actual handler.
>
> Something like:
>
> #define EX_REG_HANDLER(_reg)                                    \
> bool ex_handler_value_##_reg(const struct exception_table_entry *fixup, \
>                             struct pt_regs *regs, int trapnr)   \
> {                                                               \
>         void (*handler)(unsigned long) =                        \
>                 (void *)((unsigned long)&fixup->to + fixup->to); \
>                                                                 \
>         if (trapnr != X86_TRAP_UD)                              \
>                 return false;                                   \
>                                                                 \
>         regs->ip += 2; /* size of UD2 instruction */            \
>         handler(regs->_reg);                                    \
>         return true;                                            \
> }
>
> EX_REG_HANDLER(bx);
> EX_REG_HANDLER(cx);
> ...
> EX_REG_HANDLER(ss);
>
>
> asm (
> " .macro reg_to_handler r\n"
> " .irp rs,bx,cx,...,ss\n"
> " .ifc \\r, %\\rs\n"
> " ex_handler_value_\\rs\n"
> " .endif\n"
> " .endr\n"
> " .endm\n"
> );
>
> #define EXCEPTION_VALUE(val, handler)                   \
>         asm volatile ("1: ud2"                          \
>                       _ASM_EXTABLE_HANDLE(1b, handler,  \
>                                      reg_to_handler %0) \
>                       : : "r" (val))
>
>
> Where the generic version can simply be:
>
> #define EXCEPTION_VALUE(val, handler)   handler((unsigned long)val)
>
>
> Makes sense?

Ooooh, that is intense. And the trampolines (EX_REG_HANDLERs) are all
just there to catch whatever register gcc decides to stuff the value
into? *cover face* Sure, okay. :)

I wonder how many existing WARN callsites could be repurposed to use this?

-Kees

-- 
Kees Cook
Pixel Security