Re: Static TLS size and stack size

[Siddhesh Poyarekar <siddhesh at redhat dot com>]

On Tue, 05 Jun 2012 06:28:00 -0700, Ian wrote:

> If I set the desired stack size via pthread_attr_setstacksize in a
> pthread_attr_t passed to pthread_create, I don't get a stack of that
> size.  The guard page is subtracted from the stack size.  More
> importantly, the static TLS size is subtracted.  If the result is
> negative, pthread_create returns an EINVAL error.  If the result is
> non-negative, I can wind up a stack much smaller than requested.
> 
> This is a problem for the Go library.  The Go library uses GCC's
> -fsplit-stack support, so when creating a thread it asks for a stack
> size of PTHREAD_STACK_MIN.  If the Go library is used by a program
> with a large number of TLS variables, or a large TLS array, I can
> wind up with an unexpected failure, or with a segmentation violation
> as the small stack is overrun.
> 
> I tried working around this by calling the glibc-private function
> _dl_get_static_tls_info, but, besides being highly nonportable, Jakub
> informs me that that will cause rpm to refuse to build a gcc RPM.
> 
> I don't see a good way to handle this.  To me it simply seems like a
> bug that glibc takes the static TLS size out of the requested stack
> size. Why not increase the requested stack size instead?
> 
> Any suggestions for what I can do here?

This is a known issue in glibc that is being worked on. See:

http://sourceware.org/bugzilla/show_bug.cgi?id=11787

Regards,
Siddhesh