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

From: Muchun Song <songmuchun@bytedance.com>
To: Mike Kravetz <mike.kravetz@oracle.com>
Cc: David Hildenbrand <david@redhat.com>,
	Michal Hocko <mhocko@suse.com>, Jonathan Corbet <corbet@lwn.net>,
	Thomas Gleixner <tglx@linutronix.de>,
	mingo@redhat.com, bp@alien8.de, x86@kernel.org, hpa@zytor.com,
	dave.hansen@linux.intel.com, luto@kernel.org,
	Peter Zijlstra <peterz@infradead.org>,
	viro@zeniv.linux.org.uk,
	Andrew Morton <akpm@linux-foundation.org>,
	paulmck@kernel.org, mchehab+huawei@kernel.org,
	pawan.kumar.gupta@linux.intel.com,
	Randy Dunlap <rdunlap@infradead.org>,
	oneukum@suse.com, anshuman.khandual@arm.com, jroedel@suse.de,
	Mina Almasry <almasrymina@google.com>,
	David Rientjes <rientjes@google.com>,
	Matthew Wilcox <willy@infradead.org>,
	Oscar Salvador <osalvador@suse.de>,
	"Song Bao Hua (Barry Song)" <song.bao.hua@hisilicon.com>,
	Xiongchun duan <duanxiongchun@bytedance.com>,
	linux-doc@vger.kernel.org, LKML <linux-kernel@vger.kernel.org>,
	Linux Memory Management List <linux-mm@kvack.org>,
	linux-fsdevel <linux-fsdevel@vger.kernel.org>
Subject: Re: [External] Re: [PATCH v5 00/21] Free some vmemmap pages of hugetlb page
Date: Sun, 22 Nov 2020 15:29:40 +0800	[thread overview]
Message-ID: <CAMZfGtUSc6QNTy34U5GQgFMzJ6_pcYJUFZwy4hNexrGN0f0hKA@mail.gmail.com> (raw)
In-Reply-To: <55e53264-a07a-a3ec-4253-e72c718b4ee6@oracle.com>

On Sat, Nov 21, 2020 at 1:47 AM Mike Kravetz <mike.kravetz@oracle.com> 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.
>
> 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.

Hi Mike,

Yeah, I agree with you that the first version of this feature should be
simply. I can do that (disable PMD/huge page mapping of vmemmap)
in the next version patch. But I have another question: what the
problem is when struct page pages go away? I have not understood
the issues discussed here, hope you can answer for me. Thanks.

> --
> Mike Kravetz

-- 
Yours,
Muchun