Address Sanitizer vs. swapcontext

[Avi Kivity]

On 11/20/2015 03:29 AM, Ian Lance Taylor wrote:
> On Thu, Nov 19, 2015 at 2:24 AM, Avi Kivity <avi@scylladb.com> wrote:
>> I am using the swapcontext family to implement user-level threading.
>> Specifically, only getcontext(), setcontext(), and makecontext() are used,
>> during thread creation and teardown.  Beyond the initial switch into a
>> thread, I use setjmp()/longjmp() as they are significantly faster.
>>
>> This works well, except that in combination with Address Sanitizer I
>> stack-buffer-overflow errors accessing variables on a user-level-thread
>> stack, which, as far as I can tell, are false positives.
>>
>> See for example https://github.com/scylladb/scylla/issues/533.
>>
>> Is there any workaround for this?  I am willing to write an alternate code
>> path for debugging.  What would work here? sigaltstack()?
> If you #include <asan_interface.h>, you can use macros like
> ASAN_UNPOISON_MEMORY to tell asan that certain memory is OK to access.
>

I am not sure which memory I should unpoison.  The user-level-thread 
stack was allocated by the program, using malloc(), so it is visible to 
asan.