Re: [PATCH 3/4] kernel/fork: switch vmapped stack callation to __vmalloc_area()

[Konstantin Khlebnikov]

On 21.02.2018 03:16, Andrew Morton wrote:
> On Tue, 23 Jan 2018 16:57:21 +0300 Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx> wrote:
>
> > # stress-ng --clone 100 -t 10s --metrics-brief
> > at 32-core machine shows boost 35000 -> 36000 bogo ops
> >
> > Patch 4/4 is a kind of RFC.
> > Actually per-cpu cache of preallocated stacks works faster than buddy allocator thus
> > performance boots for it happens only at completely insane rate of clones.
>
> I'm not really sure what to make of this patchset.  Is it useful in any
> known real-world use cases?

Not yet. Feel free to ignore last patch.

> > +	  This option neutralize stack overflow protection but allows to
> > +	  achieve best performance for syscalls fork() and clone().
>
> That sounds problematic, but perhaps acceptable if the fallback only
> happens rarely.
>
> Can this code be folded into CONFIG_VMAP_STACk in some cleaner fashion?
> We now have options for non-vmapped stacks, vmapped stacks and a mix
> of both.
>
> And what about this comment in arch/Kconfig:VMAP_STACK:
>
>            This is presently incompatible with KASAN because KASAN expects
>            the stack to map directly to the KASAN shadow map using a formula
>            that is incorrect if the stack is in vmalloc space.
>
>
> So VMAP_STACK_AS_FALLBACK will intermittently break KASAN?

All of this (including CONFIG_VMAP_STACK) could be turned into boot option.
I think this would be a best solution.