Re: [PATCH] mm: per-thread vma caching

[Davidlohr Bueso <davidlohr@hp.com>]

On Fri, 2014-02-21 at 20:55 -0800, Davidlohr Bueso wrote:
> On Fri, 2014-02-21 at 13:24 -0800, Linus Torvalds wrote:
> > On Fri, Feb 21, 2014 at 12:57 PM, Davidlohr Bueso <davidlohr@hp.com> wrote:
> > >
> > > Btw, one concern I had is regarding seqnum overflows... if such
> > > scenarios should happen we'd end up potentially returning bogus vmas and
> > > getting bus errors and other sorts of issues. So we'd have to flush the
> > > caches, but, do we care? I guess on 32bit systems it could be a bit more
> > > possible to trigger given enough forking.
> > 
> > I guess we should do something like
> > 
> >     if (unlikely(!++seqnum))
> >         flush_vma_cache()
> > 
> > just to not have to worry about it.
> > 
> > And we can either use a "#ifndef CONFIG_64BIT" to disable it for the
> > 64-bit case (because no, we really don't need to worry about overflow
> > in 64 bits ;), or just decide that a 32-bit sequence number actually
> > packs better in the structures, and make it be an "u32" even on 64-bit
> > architectures?
> > 
> > It looks like a 32-bit sequence number might pack nicely next to the
> > 
> >     unsigned brk_randomized:1;
> 
> And probably specially so for structures like task and mm. I hadn't
> considered the benefits of packing vs overflowing. So we can afford
> flushing all tasks's vmacache every 4 billion forks.

ah, not quite that much, I was just thinking of dup_mmap, of course we
also increment upon invalidations.