On Tue 26-09-17 12:43:44, Kirill A. Shutemov wrote:
> On Mon, Sep 25, 2017 at 03:53:05PM +0200, Michal Hocko wrote:
> > On Mon 25-09-17 16:07:15, Kirill A. Shutemov wrote:
> > > On Mon, Sep 25, 2017 at 01:54:30PM +0200, Michal Hocko wrote:
> > > > On Mon 25-09-17 10:39:13, Kirill A. Shutemov wrote:
> > > > > On machine with 5-level paging support a process can allocate
> > > > > significant amount of memory and stay unnoticed by oom-killer and
> > > > > memory cgroup. The trick is to allocate a lot of PUD page tables.
> > > > > We don't account PUD page tables, only PMD and PTE.
> > > > >
> > > > > We already addressed the same issue for PMD page tables, see
> > > > > dc6c9a35b66b ("mm: account pmd page tables to the process").
> > > > > Introduction 5-level paging bring the same issue for PUD page tables.
> > > > >
> > > > > The patch expands accounting to PUD level.
> > > >
> > > > OK, we definitely need this or something like that but I really do not
> > > > like how much code we actually need for each pte level for accounting.
> > > > Do we really need to distinguish each level? Do we have any arch that
> > > > would use a different number of pages to back pte/pmd/pud?
> > >
> > > Looks like we actually do. At least on mips. See PMD_ORDER/PUD_ORDER.
> >
> > Hmm, but then oom_badness does consider them a single page which is
> > wrong. I haven't checked other users. Anyway even if we've had different
> > sizes why cannot we deal with this in callers. They know which level of
> > page table they allocate/free, no?
>
> So do you want to see single counter for all page table levels?
I think it would make the code easier to maintain and follow. As a
follow up for this patch which I am willing to ack.
> Do we have anybody who relies on VmPTE/VmPMD now?
No idea. I would just account everything to the pte level. If somebody
complains we can revert that part. But the value has been exported since
dc6c9a35b66b ("mm: account pmd page tables to the process") and I
suspect you have done so just to keep it in sync the existing VmPTE
without an explicit usecase in mind, right?
--
Michal Hocko
SUSE Labs
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
