On 09.06.2020 21:46, Kees Cook wrote:
> On Thu, Jun 04, 2020 at 04:49:54PM +0300, Alexander Popov wrote:
>> Let's improve the instrumentation to avoid this:
>>
>> 1. Make stackleak_track_stack() save all register that it works with.
>> Use no_caller_saved_registers attribute for that function. This attribute
>> is available for x86_64 and i386 starting from gcc-7.
>>
>> 2. Insert calling stackleak_track_stack() in asm:
>> asm volatile("call stackleak_track_stack" :: "r" (current_stack_pointer))
>> Here we use ASM_CALL_CONSTRAINT trick from arch/x86/include/asm/asm.h.
>> The input constraint is taken into account during gcc shrink-wrapping
>> optimization. It is needed to be sure that stackleak_track_stack() call is
>> inserted after the prologue of the containing function, when the stack
>> frame is prepared.
>
> Very cool; nice work!
>
>> +static void add_stack_tracking(gimple_stmt_iterator *gsi)
>> +{
>> + /*
>> + * The 'no_caller_saved_registers' attribute is used for
>> + * stackleak_track_stack(). If the compiler supports this attribute for
>> + * the target arch, we can add calling stackleak_track_stack() in asm.
>> + * That improves performance: we avoid useless operations with the
>> + * caller-saved registers in the functions from which we will remove
>> + * stackleak_track_stack() call during the stackleak_cleanup pass.
>> + */
>> + if (lookup_attribute_spec(get_identifier("no_caller_saved_registers")))
>> + add_stack_tracking_gasm(gsi);
>> + else
>> + add_stack_tracking_gcall(gsi);
>> +}
>
> The build_for_x86 flag is only ever used as an assert() test against
> no_caller_saved_registers, but we're able to test for that separately.
> Why does the architecture need to be tested? (i.e. when this flag
> becomes supported o other architectures, why must it still be x86-only?)
The inline asm statement that is used for instrumentation is arch-specific.
Trying to add
asm volatile("call stackleak_track_stack")
in gcc plugin on aarch64 makes gcc break spectacularly.
I pass the target arch name to the plugin and check it explicitly to avoid that.
Moreover, I'm going to create a gcc enhancement request for supporting
no_caller_saved_registers attribute on aarch64.
Best regards,
Alexander
