Re: [PATCH v5 00/21] Free some vmemmap pages of hugetlb page

[David Hildenbrand]

On 20.11.20 18:45, Mike Kravetz wrote:
> On 11/20/20 1:43 AM, David Hildenbrand wrote:
> > On 20.11.20 10:39, Michal Hocko wrote:
> > > On Fri 20-11-20 10:27:05, David Hildenbrand wrote:
> > > > On 20.11.20 09:42, Michal Hocko wrote:
> > > > > On Fri 20-11-20 14:43:04, Muchun Song wrote:
> > > > > [...]
> > > > >
> > > > > Thanks for improving the cover letter and providing some numbers. I have
> > > > > only glanced through the patchset because I didn't really have more time
> > > > > to dive depply into them.
> > > > >
> > > > > Overall it looks promissing. To summarize. I would prefer to not have
> > > > > the feature enablement controlled by compile time option and the kernel
> > > > > command line option should be opt-in. I also do not like that freeing
> > > > > the pool can trigger the oom killer or even shut the system down if no
> > > > > oom victim is eligible.
> > > > >
> > > > > One thing that I didn't really get to think hard about is what is the
> > > > > effect of vmemmap manipulation wrt pfn walkers. pfn_to_page can be
> > > > > invalid when racing with the split. How do we enforce that this won't
> > > > > blow up?
> > > >
> > > > I have the same concerns - the sections are online the whole time and
> > > > anybody with pfn_to_online_page() can grab them
> > > >
> > > > I think we have similar issues with memory offlining when removing the
> > > > vmemmap, it's just very hard to trigger and we can easily protect by
> > > > grabbing the memhotplug lock.
> > >
> > > I am not sure we can/want to span memory hotplug locking out to all pfn
> > > walkers. But you are right that the underlying problem is similar but
> > > much harder to trigger because vmemmaps are only removed when the
> > > physical memory is hotremoved and that happens very seldom. Maybe it
> > > will happen more with virtualization usecases. But this work makes it
> > > even more tricky. If a pfn walker races with a hotremove then it would
> > > just blow up when accessing the unmapped physical address space. For
> > > this feature a pfn walker would just grab a real struct page re-used for
> > > some unpredictable use under its feet. Any failure would be silent and
> > > hard to debug.
> >
> > Right, we don't want the memory hotplug locking, thus discussions regarding rcu. Luckily, for now I never saw a BUG report regarding this - maybe because the time between memory offlining (offline_pages()) and memory/vmemmap getting removed (try_remove_memory()) is just too long. Someone would have to sleep after pfn_to_online_page() for quite a while to trigger it.
> >
> > > [...]
> > > > To keep things easy, maybe simply never allow to free these hugetlb pages
> > > > again for now? If they were reserved during boot and the vmemmap condensed,
> > > > then just let them stick around for all eternity.
> > >
> > > Not sure I understand. Do you propose to only free those vmemmap pages
> > > when the pool is initialized during boot time and never allow to free
> > > them up? That would certainly make it safer and maybe even simpler wrt
> > > implementation.
> >
> > Exactly, let's keep it simple for now. I guess most use cases of this (virtualization, databases, ...) will allocate hugepages during boot and never free them.
>
> Not sure if I agree with that last statement.  Database and virtualization
> use cases from my employer allocate allocate hugetlb pages after boot.  It
> is shortly after boot, but still not from boot/kernel command line.

Right, but the ones that care about this optimization for now could be 
converted, I assume? I mean we are talking about "opt-in" from 
sysadmins, so requiring to specify a different cmdline parameter does 
not sound to weird to me. And it should simplify a first version quite a 
lot.

The more I think about this, the more I believe doing these vmemmap 
modifications after boot are very dangerous.

> Somewhat related, but not exactly addressing this issue ...
>
> One idea discussed in a previous patch set was to disable PMD/huge page
> mapping of vmemmap if this feature was enabled.  This would eliminate a bunch
> of the complex code doing page table manipulation.  It does not address
> the issue of struct page pages going away which is being discussed here,
> but it could be a way to simply the first version of this code.  If this
> is going to be an 'opt in' feature as previously suggested, then eliminating
> the  PMD/huge page vmemmap mapping may be acceptable.  My guess is that
> sysadmins would only 'opt in' if they expect most of system memory to be used
> by hugetlb pages.  We certainly have database and virtualization use cases
> where this is true.

It sounds like a hack to me, which does not fully solve the problem. But 
yeah, it's a simplification.

--
Thanks,

David / dhildenb