Re: [PATCH] mm: larger stack guard gap, between vmas

[Ben Hutchings <ben@decadent.org.uk>]

[-- Attachment #1: Type: text/plain, Size: 2718 bytes --]

On Wed, 2017-07-05 at 10:15 -0700, Linus Torvalds wrote:
> > On Wed, Jul 5, 2017 at 9:58 AM, Ben Hutchings <ben@decadent.org.uk> wrote:
> > 
> > I ended up with the following two patches, which seem to deal with
> > both the Java and Rust regressions.  These don't touch the
> > stack-grows-up paths at all because Rust doesn't run on those
> > architectures and the Java weirdness is i386-specific.
> > 
> > They definitely need longer commit messages and comments, but aside
> > from that do these look reasonable?
> 
> I thin kthey both look reasonable, but I think we might still want to
> massage things a bit (cutting down the quoting to a minimum, hopefully
> leaving enough context to still make sense):
> 
> > Subject: [1/2] mmap: Skip a single VM_NONE mapping when checking the stack gap
> > 
> >         prev = vma->vm_prev;
> > +       if (prev && !(prev->vm_flags & (VM_READ | VM_WRITE | VM_EXEC)))
> > +               prev = prev->vm_prev;
> >         if (prev && prev->vm_end > gap_addr) {
> 
> Do we just want to ignore the user-supplied guard mapping, or do we
> want to say "if the user does a guard mapping, we use that *instead*
> of our stack gap"?
> 
> IOW, instead of "prev = prev->vm_prev;" and continuing, maybe we want
> to just return "ok".

Rust effectively added a second guard page to the main thread stack.  
But it does not (yet) implement stack probing
(https://github.com/rust-lang/rust/issues/16012) so I think it will
benefit from the kernel's larger stack guard gap.

> > Subject: [2/2] mmap: Avoid mapping anywhere within the full stack extent  if finite
> 
> This is good thinking, but no, I don't think the "if finite" is right.
> 
> I've seen people use "really big values" as replacement for
> RLIM_INIFITY, for various reasons.
> 
> We've had huge confusion about RLIM_INFINITY over the years - look for
> things like COMPAT_RLIM_OLD_INFINITY to see the kinds of confusions
> we've had.

That sounds familiar...

> Some people just use MAX_LONG etc, which is *not* the same as
> RLIM_INFINITY, but in practice ends up doing the same thing. Yadda
> yadda.
> 
> So I'm personally leery of checking and depending on "exactly
> RLIM_INIFITY", because I've seen it go wrong so many times.
> 
> And I think your second patch breaks that "use a really large value to
> approximate infinity" case that definitely has existed as a pattern.

Right.  Well that seems to leave us with remembering the MAP_FIXED flag
and using that as the condition to ignore the previous mapping.

Ben.

-- 
Ben Hutchings
Man invented language to satisfy his deep need to complain. - Lily
Tomlin

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 833 bytes --]