On Friday 03 February 2012 03:01:35 KOSAKI Motohiro wrote:

quoted

Right now MAP_STACK does not mean anything since it is ignored. The
intention of this behaviour change is to make MAP_STACK mean that the
map is going to be used as a stack and hence, set it up like a stack
ought to be. I could not really think of a valid case for fixed size
stacks; it looks like a limitation in the pthread implementation in
glibc rather than a feature. So this patch will actually result in
uniform behaviour across threads when it comes to stacks.

This does change vm accounting since thread stacks were earlier
accounted as anon memory.

The fact is, now process stack and pthread stack clearly behave
different dance. libc don't expect pthread stack grow automatically.
So, your patch will break userland. Just only change display thing.

does it though ?  glibc doesn't keep track of the unused address space ... 
that's what the kernel is for.  pthread_attr_setstacksize explicitly operates 
on the *minimum* stack size, not the *exact* size.

where exactly do you think userland would break ?

http://pubs.opengroup.org/onlinepubs/9699919799/functions/pthread_attr_setstacksize.html
-mike