Re: [patch] flexible-mmap-2.6.7-D5

[Ingo Molnar <mingo@elte.hu>]

* William Lee Irwin III <wli@holomorphy.com> wrote:

> Also, I suspect some more graceful fallback would make sense
> particularly for the case of RLIM_INFINITY, which would leave users
> that run with, say, all rlimits at RLIM_INFINITY in the interest of
> having full access to system resources with a mere 512MB of
> virtualspace for the heap, which IIRC glibc is intelligent enough to
> circumvent for malloc(), but not for mmap(NULL, ...). [...]

well, the 5/6=stack 1/6=malloc rule in the RLIM_INFINITY can be changed. 
What would make the most sense - 1/2 for both?

> If it's been in production that long, I find it hard to believe that's
> never been tripped over. [...]

it's been tripped over and the 5/6 rule was a fix for such a bugreport. 
What happens more in practice frequently is that someone needs a big
stack and sets the stack ulimit to RLIM_INFINITY.

> [...] (also, that 128MB is currently wasted); [...]

the 128MB is 'wasted' to give some flexibility to the stack rlimits
changing runtime. But in practice it's far more important to have the
mmap()/malloc() space maximized and flexible than to give the stack
automatic flexibility.

	Ingo