Re: [PATCH RFC] mm: account VMA before forced-COW via /proc/pid/mem

[Konstantin Khlebnikov]

Hugh Dickins wrote:
> On Thu, 5 Apr 2012, Konstantin Khlebnikov wrote:
> > Oleg Nesterov wrote:
> > > On 04/04, Konstantin Khlebnikov wrote:
> > > > Oleg Nesterov wrote:
> > > > > On 04/02, Konstantin Khlebnikov wrote:
> > > > > > Currently kernel does not account read-only private mappings into
> > > > > > memory commitment.
> > > > > > But these mappings can be force-COW-ed in get_user_pages().
> > > > >
> > > > > Heh. tail -n3 Documentation/vm/overcommit-accounting
> > > > > may be you should update it then.
> > > >
> > > > I just wonder how fragile this accounting...
> > >
> > > I meant, this patch could also remove this "TODO" from the docs.
> >
> > Actually I dug into this code for killing VM_ACCOUNT vma flag.
> > Currently we cannot do this only because asymmetry in mprotect_fixup():
> > it account vma on read-only ->  writable conversion, but keep on backward
> > operation.
> > Probably we can kill this asymmetry, and after that we can recognize
> > accountable vma
> > by its others flags state, so we don't need special VM_ACCOUNT for this.
>
> (I believe the VM_ACCOUNT flag will need to stay.)
>
> But this is just a quick note to say that I'm not ignoring you: I have
> a strong interest in this, but only now found time to look through the
> thread and ponder, and I'm not yet ready to decide.
>
> I've long detested that behaviour of GUP write,force, and my strong
> preference would be not to layer more strangeness upon strangeness,
> but limit the damage by making GUP write,force fail in that case,
> instead of inserting a PageAnon page into a VM_SHARED mapping.
>
> I think it's unlikely that it will cause a regression in real life
> (it already fails if you did not open the mmap'ed file for writing),
> but it would be a user-visible change in behaviour, and I've research
> to do before arriving at a conclusion.

Agree, but this stuff is very weak. Even if sysctl vm.overcommit_memory=2,
probably we should fixup accounting in /proc/pid/mem only for this case,
because vm.overcommit_memory=2 supposed to protect against overcommit, but it does not.

> Hugh

--
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/