bpo-21131: don't use SIGSTKSZ for stack size in sigaltstack by peadar · Pull Request #13649 · python/cpython (original) (raw)
Following up discussion on http://bugs.python.org/issue21131
vstinner suggested using the logic in thread_pthread.h to work out the "good" stack size for sigaltstack. I've integrated the logic in there, but it doesn't actually provide an accessible way of finding the default thread stack size - it just nominally requests pthread_create to create a stack of a "default" size by not poking at the stack size attribute.
So, instead, I've rejigged the macros in that file to provide PLATFORM_STACK_SIZE and PLATFORM_STACK_PREFERRED that are available regardless of the ability of pthreads to set the thread stack size attribute (i.e., regardless of the presence of _POSIX_THREAD_ATTR_STACK_SIZE) (and simplified some of the conditionals in the code using THREAD_STACK_SIZE in the process). The eventual sizes past to pthread_attr_setstacksize are unaffected.
This is then used as a fallback in the new function, size_t _Pthread_preferred_stacksize() , which will first attempt to use a default-constructed pthread_attr_t to work out at runtime what size stack the system prefers for new threads. This value is far more likely to be usable than the value defined in SIGSTKSZ.
If that call fails, or the thread size attribute is not available, we default to the maximum of PTHREAD_STACK_MIN, whatever we've defined for PLATFORM_STACK_SIZE for those platforms with an inadequate default stacksize, or finally the arbitrary 32k setting previously used as a floor.
All feedback gratefully appreciated.