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

From: Willy Tarreau
Date: Tue Jul 04 2017 - 14:40:02 EST

On Tue, Jul 04, 2017 at 11:37:15AM -0700, Linus Torvalds wrote:
> On Tue, Jul 4, 2017 at 10:22 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> >
> > Well, I've been thinking about this some more and the more I think about
> > it the less I am convinced we should try to be clever here. Why? Because
> > as soon as somebody tries to manage stacks explicitly you cannot simply
> > assume anything about the previous mapping. Say some interpret uses
> > [ mngmnt data][red zone] <--[- MAP_GROWSDOWN ]
> >
> > Now if we consider the red zone's (PROT_NONE) prev mapping we would fail
> > the expansion even though we haven't hit the red zone and that is
> > essentially what the Java and rust bugs are about. So we just risk yet
> > another regression.
> Ack.
> Let's make the initial version at least only check the first vma.
> The long-term fix for this is to have the binaries do proper stack
> expansion probing anyway, and it's quite possible that people who do
> their own stack redzoning by adding a PROT_NONE thing already do that
> proper fix (eg the Java stack may simply not *have* those big crazy
> structures on it in the first place).

But what is wrong with stopping the loop as soon as the distance gets
larger than the stack_guard_gap ?