On Tue, 11 Nov 2014, Dave Hansen wrote:
> On 11/11/2014 10:27 AM, Thomas Gleixner wrote:
> > On Thu, 6 Nov 2014, Dave Hansen wrote:
> >> Instead of all of these games with dropping and reacquiring mmap_sem and
> >> adding other locks, or deferring the work, why don't we just do a
> >> get_user_pages()? Something along the lines of:
> >>
> >> while (1) {
> >> ret = cmpxchg(addr)
> >> if (!ret)
> >> break;
> >> if (ret == -EFAULT)
> >> get_user_pages(addr);
> >> }
> >>
> >> Does anybody see a problem with that?
> >
> > You want to do that under mmap_sem write held, right? Not a problem per
> > se, except that you block normal faults for a possibly long time when
> > the page(s) need to be swapped in.
>
> Yeah, it might hold mmap_sem for write while doing this in the unmap
> path. But, that's only if the bounds directory entry has been swapped
> out. There's only 1 pointer of bounds directory entries there for every
> 1MB of data, so it _should_ be relatively rare. It would mean that
> nobody's been accessing a 512MB swath of data controlled by the same
> page of the bounds directory.
>
> If it gets to be an issue, we can always add some code to fault it in
> before mmap_sem is acquired.
I don't think it's a real issue.
> FWIW, I believe we have a fairly long road ahead of us to optimize MPX
> in practice. I have a list of things I want to go investigate, but I
> have not looked in to it in detail at all.
:)
> > But yes, this might solve most of the issues at hand. Did not think
> > about GUP at all :(
>
> Whew. Fixing it was getting nasty and complicated. :)
Indeed. Though I think that distangling specific parts of MPX from
mmap_sem is still a worthwhile exercise. So not all of the complex
ideas we came up with during the discussion are lost in the pointless
complexity universe :)
Thanks,
tglx
