Re: [RFC][PATCH 0/5] VM_PINNED

[Peter Zijlstra]

On Tue, May 27, 2014 at 03:11:36PM +0400, Konstantin Khlebnikov wrote:
> On Tue, May 27, 2014 at 2:54 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > On Tue, May 27, 2014 at 12:29:09PM +0200, Peter Zijlstra wrote:
> >> On Tue, May 27, 2014 at 12:49:08AM +0400, Konstantin Khlebnikov wrote:
> >> > Another suggestion. VM_RESERVED is stronger than VM_LOCKED and extends
> >> > its functionality.
> >> > Maybe it's easier to add VM_DONTMIGRATE and use it together with VM_LOCKED.
> >> > This will make accounting easier. No?
> >>
> >> I prefer the PINNED name because the not being able to migrate is only
> >> one of the desired effects of it, not the primary effect. We're really
> >> looking to keep physical pages in place and preserve mappings.
> 
> Ah, I just mixed it up.
> 
> >>
> >> The -rt people for example really want to avoid faults (even minor
> >> faults), and DONTMIGRATE would still allow unmapping.
> >>
> >> Maybe always setting VM_PINNED and VM_LOCKED together is easier, I
> >> hadn't considered that. The first thing that came to mind is that that
> >> might make the fork() semantics difficult, but maybe it works out.
> >>
> >> And while we're on the subject, my patch preserves PINNED over fork()
> >> but maybe we don't actually need that either.
> >
> > So pinned_vm is userspace exposed, which means we have to maintain the
> > individual counts, and doing the fully orthogonal accounting is 'easier'
> > than trying to get the boundary cases right.
> >
> > That is, if we have a program that does mlockall() and then does the IB
> > ioctl() to 'pin' a region, we'd have to make mm_mpin() do munlock()
> > after it splits the vma, and then do the pinned accounting.
> >
> > Also, we'll have lost the LOCKED state and unless MCL_FUTURE was used,
> > we don't know what to restore the vma to on mm_munpin().
> >
> > So while the accounting looks tricky, it has simpler semantics.
> 
> What if VM_PINNED will require VM_LOCKED?
> I.e. user must mlock it before pining and cannot munlock vma while it's pinned.

So I don't like restrictions like that if its at all possible to avoid
-- and in this case, I already wrote the code and its not _that_
complicated.

But also; that would mean that we'd either have to make mm_mpin() do the
mlock unconditionally (which rather defeats the purpose) or break
userspace assumptions. I'm fairly sure the IB ioctl() don't require the
memory to be mlocked.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/