On 07/28/2015 10:01 PM, Paul Eggert wrote:
> Eli Zaretskii wrote:
>> The implementation of stack_overflow on sysdep.c was recently changed
>> so as not to use sys/resource.h and getrlimit, but configure.ac still
>> insists on these two features in order to include the recovery code,
>> which I think should be fixed.
> 
> Thanks, fixed with the attached patch.
> 
>> More importantly, the recovery simply longjmps to command_loop,
>> whereas similar features like Fthrow and Fsignal do much more in
>> unwind_to_catch.  Shouldn't stack overflow recovery do that as well?
>> Otherwise, the specpdl stack, byte_stack_list, lisp_eval_depth
>> etc. all stay at their values they had at stack overflow time, no?
> 
> No, they are cleared back to top-level values, because when
> command_loop's call to sigsetjmp returns nonzero, it then calls
> init_eval and this resets them.
> 
> There is a problem that the speclpdl stack's unwind-protect and
> dynamic-let-bindings are unceremoniously discarded on stack overflow.  I
> suppose that should get fixed, though it may be a bit tricky to avoid
> looping.

What's wrong with just mprotecting a guard page at the end of the stack,
and on overflow, giving that region normal protection, unwinding as
normal, then, at top level, restoring the guard page?