On Tue 16-02-21 13:01:54, Mike Rapoport wrote:
> On Tue, Feb 16, 2021 at 09:33:20AM +0100, Michal Hocko wrote:
> > On Mon 15-02-21 23:24:40, Mike Rapoport wrote:
> > > On Mon, Feb 15, 2021 at 10:00:31AM +0100, Michal Hocko wrote:
> > > > On Sun 14-02-21 20:00:16, Mike Rapoport wrote:
> > > > > On Fri, Feb 12, 2021 at 02:18:20PM +0100, Michal Hocko wrote:
> > > >
> > > > > We can correctly set the zone links for the reserved pages for holes
> > > > > in the
> > > > > middle of a zone based on the architecture constraints and with only
> > > > > the
> > > > > holes in the beginning/end of the memory will be not spanned by any
> > > > > node/zone which in practice does not seem to be a problem as the
> > > > > VM_BUG_ON
> > > > > in set_pfnblock_flags_mask() never triggered on pfn 0.
> > > >
> > > > I really fail to see what you mean by correct zone/node for a memory
> > > > range which is not associated with any real node.
> > >
> > > We know architectural zone constraints, so we can have always have 1:1
> > > match from pfn to zone. Node indeed will be a guess.
> >
> > That is true only for some zones.
>
> Hmm, and when is it not true?
ZONE_MOVABLE, ZONE_NORMAL and ZONE_DEVICE.
> > Also we do require those to be correct when the memory is managed by the
> > page allocator. I believe we can live with incorrect zones when they are
> > in holes.
>
> Note that the holes Andrea reported in the first place are not ranges that
> are not populated, but rather ranges that arch didn't report as usable,
> i.e. there is physical memory and it perfectly fits into an existing node
> and zone.
Except those are not usable so they have no clear meaning, right?
> > > > > > I am sorry, I haven't followed previous discussions. Has the
> > > > > > removal of
> > > > > > the VM_BUG_ON been considered as an immediate workaround?
> > > > >
> > > > > It was never discussed, but I'm not sure it's a good idea.
> > > > >
> > > > > Judging by the commit message that introduced the VM_BUG_ON (commit
> > > > > 86051ca5eaf5 ("mm: fix usemap initialization")) there was yet another
> > > > > inconsistency in the memory map that required a special care.
> > > >
> > > > Can we actually explore that path before adding yet additional
> > > > complexity and potentially a very involved fix for a subtle problem?
> > >
> > > This patch was intended as a fix for inconsistency of the memory map that
> > > is the root cause for triggering this VM_BUG_ON and other corner case
> > > problems.
> > >
> > > The previous version [1] is less involved as it does not extend node/zone
> > > spans.
> >
> > I do understand that. And I am not objecting to the patch. I have to
> > confess I haven't digested it yet. Any changes to early memory
> > intialization have turned out to be subtle and corner cases only pop up
> > later. This is almost impossible to review just by reading the code.
> > That's why I am asking whether we want to address the specific VM_BUG_ON
> > first with something much less tricky and actually reviewable. And
> > that's why I am asking whether dropping the bug_on itself is safe to do
> > and use as a hot fix which should be easier to backport.
>
> I can't say I'm familiar enough with migration and compaction code to say
> if it's ok to remove that bug_on. It does point to inconsistency in the
> memmap, but probably it's not important.
Yeah, I am not really clear on that myself TBH. On one hand this cannot
be really critical because it is a conditional assert on the debuging
mode. Most production users do not enable it so they wouldn't learn
about that inconsistency and maybe never notice any difference. This has
led me to think about this to be something mostly focused on debugging.
If that is an incorrect assumption then the VM_BUG_ON should be changed
to something else - either a graceful failure or a real BUG_ON if that
is really worth it.
--
Michal Hocko
SUSE Labs
