Re: [PATCH v2 1/3] mm/shuffle: don't move pages between zones and don't read garbage memmaps

[David Hildenbrand <david@redhat.com>]

On 22.06.20 15:10, Wei Yang wrote:
> On Mon, Jun 22, 2020 at 11:51:34AM +0200, David Hildenbrand wrote:
>> On 22.06.20 11:22, Wei Yang wrote:
>>> On Mon, Jun 22, 2020 at 10:43:11AM +0200, David Hildenbrand wrote:
>>>> On 22.06.20 10:26, Wei Yang wrote:
>>>>> On Fri, Jun 19, 2020 at 02:59:20PM +0200, David Hildenbrand wrote:
>>>>>> Especially with memory hotplug, we can have offline sections (with a
>>>>>> garbage memmap) and overlapping zones. We have to make sure to only
>>>>>> touch initialized memmaps (online sections managed by the buddy) and that
>>>>>> the zone matches, to not move pages between zones.
>>>>>>
>>>>>> To test if this can actually happen, I added a simple
>>>>>> 	BUG_ON(page_zone(page_i) != page_zone(page_j));
>>>>>> right before the swap. When hotplugging a 256M DIMM to a 4G x86-64 VM and
>>>>>> onlining the first memory block "online_movable" and the second memory
>>>>>> block "online_kernel", it will trigger the BUG, as both zones (NORMAL
>>>>>> and MOVABLE) overlap.
>>>>>>
>>>>>> This might result in all kinds of weird situations (e.g., double
>>>>>> allocations, list corruptions, unmovable allocations ending up in the
>>>>>> movable zone).
>>>>>>
>>>>>> Fixes: e900a918b098 ("mm: shuffle initial free memory to improve memory-side-cache utilization")
>>>>>> Acked-by: Michal Hocko <mhocko@suse.com>
>>>>>> Cc: stable@vger.kernel.org # v5.2+
>>>>>> Cc: Andrew Morton <akpm@linux-foundation.org>
>>>>>> Cc: Johannes Weiner <hannes@cmpxchg.org>
>>>>>> Cc: Michal Hocko <mhocko@suse.com>
>>>>>> Cc: Minchan Kim <minchan@kernel.org>
>>>>>> Cc: Huang Ying <ying.huang@intel.com>
>>>>>> Cc: Wei Yang <richard.weiyang@gmail.com>
>>>>>> Cc: Mel Gorman <mgorman@techsingularity.net>
>>>>>> Signed-off-by: David Hildenbrand <david@redhat.com>
>>>>>> ---
>>>>>> mm/shuffle.c | 18 +++++++++---------
>>>>>> 1 file changed, 9 insertions(+), 9 deletions(-)
>>>>>>
>>>>>> diff --git a/mm/shuffle.c b/mm/shuffle.c
>>>>>> index 44406d9977c77..dd13ab851b3ee 100644
>>>>>> --- a/mm/shuffle.c
>>>>>> +++ b/mm/shuffle.c
>>>>>> @@ -58,25 +58,25 @@ module_param_call(shuffle, shuffle_store, shuffle_show, &shuffle_param, 0400);
>>>>>>  * For two pages to be swapped in the shuffle, they must be free (on a
>>>>>>  * 'free_area' lru), have the same order, and have the same migratetype.
>>>>>>  */
>>>>>> -static struct page * __meminit shuffle_valid_page(unsigned long pfn, int order)
>>>>>> +static struct page * __meminit shuffle_valid_page(struct zone *zone,
>>>>>> +						  unsigned long pfn, int order)
>>>>>> {
>>>>>> -	struct page *page;
>>>>>> +	struct page *page = pfn_to_online_page(pfn);
>>>>>
>>>>> Hi, David and Dan,
>>>>>
>>>>> One thing I want to confirm here is we won't have partially online section,
>>>>> right? We can add a sub-section to system, but we won't manage it by buddy.
>>>>
>>>> Hi,
>>>>
>>>> there is still a BUG with sub-section hot-add (devmem), which broke
>>>> pfn_to_online_page() in corner cases (especially, see the description in
>>>> include/linux/mmzone.h). We can have a boot-memory section partially
>>>> populated and marked online. Then, we can hot-add devmem, marking the
>>>> remaining pfns valid - and as the section is maked online, also as online.
>>>
>>> Oh, yes, I see this description.
>>>
>>> This means we could have section marked as online, but with a sub-section even
>>> not added.
>>>
>>> While the good news is even the sub-section is not added, but its memmap is
>>> populated for an early section. So the page returned from pfn_to_online_page()
>>> is a valid one.
>>>
>>> But what would happen, if the sub-section is removed after added? Would
>>> section_deactivate() release related memmap to this "struct page"?
>>
>> If devmem is removed, the memmap will be freed and the sub-sections are
>> marked as non-present. So this works as expected.
>>
> 
> Sorry, I may not catch your point. If my understanding is correct, the
> above behavior happens in function section_deactivate().
> 
> Let me draw my understanding of function section_deactivate():
> 
>     section_deactivate(pfn, nr_pages)
>         clear_subsection_map(pfn, nr_pages)
> 	depopulate_section_memmap(pfn, nr_pages)
> 
> Since we just remove a sub-section, I skipped some un-related codes. These two
> functions would:
> 
>   * clear bitmap in ms->usage->subsection_map
>   * free memmap for the sub-section
> 
> While since the section is not empty, ms->section_mem_map is not set no null.

Let me clarify, sub-section hotremove works differently when overlying
with (online) boot memory within a section.

Early sections (IOW, boot memory) are never partially removed. See
mm/sparse.c:section_deactivate(). We only free a early memmap when the
section is completely empty. Also see how
include/linux/mmzone.h:pfn_valid() handles early sections.

So when we have a partially present section with boot memory, we
a) marked the whole section present and online (there is only a single
   bit)
b) allocated the memmap for the whole section
c) Only exposed the relevant pages to the buddy. The memmap of non-
   present parts in a section were initialized and are reserved.

pfn_valid() will return for all non-present pfns valid, because there is
a memmap. pfn_to_online_page() will return for all pfns "true", because
we only have a single bit for the whole section. This has been the case
before sub-section hotplug and is still the case. It simply looks like
just another memory hole for which we have a memmap.

Now, with devmem it is possible to suddenly change these sub-section
holes (memmaps) to become ZONE_DEVICE memory. pfn_to_online_page() would
have to detect that and report a "false". Possible fixes were already
discussed (e.g., sub-section online map instead of a single bit).

Again, the zone check safes us from the worst, just as in the case of
all other pfn walkers that use (as documented) pfn_to_online_page(). It
still needs a fix as dicussed, but it seems to work reasonably fine like
that for now.

-- 
Thanks,

David / dhildenb