🚀 go-pugleaf

I am working on a program that creates its own lightweight threads (must manage
the stacks).  To detect stack overflow, I am using the mprotect() system
call to create a guard page at the end of the stack.  Instead of just dumping
core when the stack overflows, I want to exit the failed thread (the others
get to keep running).

My idea is to use setjmp() at the beginning of each thread to provide a "place"
to come back, and to call longjmp() from the bus error handler.  To do this,
I have to have a separate signal stack (using sigstack() and sigvec()).

The guard page works (I get my bus error), but I can't unwind the stack
using longjmp().  I'm doing this on a Sun4 (ELC), and I suspect that there
is some register window weirdness going on.

The function calling setjmp gets control after the longjmp but it gets a
bus error as soon as it returns to its caller.  The $i7 and $fp registers
are foobar when the function calling setjmp() gets control.  I tried the
#pragma unknown_control_flow for my function but it still doesn't get
the correct $i7 and $fp.  (If I set them to their correct values by hand
in the debugger, everything works ok).

After disassembling the setjmp/longjmp code in the library, I'm wondering
how they ever work with long stack depths with the register windows.

Does anyone know what I'm doing incorrectly?  Or am I faced with a
bug in SunOS (the rev. is 4.1.1)?

Thanks in advance for any help,

Brian Vetter
bjv@tivoli.com