Re: [PATCH] mm/page_isolation: do not isolate the max order page

[David Hildenbrand <david@redhat.com>]

On 03.12.20 18:48, David Hildenbrand wrote:
> On 03.12.20 18:15, Vlastimil Babka wrote:
>> On 12/3/20 5:26 PM, David Hildenbrand wrote:
>>> On 03.12.20 01:03, Vlastimil Babka wrote:
>>>> On 12/2/20 1:21 PM, Muchun Song wrote:
>>>>> The max order page has no buddy page and never merge to other order.
>>>>> So isolating and then freeing it is pointless.
>>>>>
>>>>> Signed-off-by: Muchun Song <songmuchun@bytedance.com>
>>>>
>>>> Acked-by: Vlastimil Babka <vbabka@suse.cz>
>>>>
>>>>> ---
>>>>>  mm/page_isolation.c | 2 +-
>>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/mm/page_isolation.c b/mm/page_isolation.c
>>>>> index a254e1f370a3..bddf788f45bf 100644
>>>>> --- a/mm/page_isolation.c
>>>>> +++ b/mm/page_isolation.c
>>>>> @@ -88,7 +88,7 @@ static void unset_migratetype_isolate(struct page *page, unsigned migratetype)
>>>>>  	 */
>>>>>  	if (PageBuddy(page)) {
>>>>>  		order = buddy_order(page);
>>>>> -		if (order >= pageblock_order) {
>>>>> +		if (order >= pageblock_order && order < MAX_ORDER - 1) {
>>>>>  			pfn = page_to_pfn(page);
>>>>>  			buddy_pfn = __find_buddy_pfn(pfn, order);
>>>>>  			buddy = page + (buddy_pfn - pfn);
>>>>
>>>> Hm I wonder if order == MAX_ORDER - 1, then the buddy can actually be a
>>>> !pfn_valid() in some corner case? pfn_valid_within(buddy_pfn) that follows would
>>>> only catch it on archs with holes in zone. Then is_migrate_isolate_page(buddy)
>>>> might access an invalid buddy. So this might be actually a bug fix and not just
>>>> optimization, just the bug hasn't been observed in practice.
>>>
>>> I think we have no users that isolate/unisolate close to holes.
>>>
>>> CMA regions are properly aligned (to max of page_order /
>>> max_order_nr_pages) and don't contain holes.
>>
>> The problem as I see it, is that buddy_order(page) might be already MAX_ORDER -
>> 1 (e.g. two pageblocks on x86), and then finding buddy of that one is beyond the
>> guaranteed alignment (if they merged, which they can't, it would be four
> 
> Oh, I see. I would have assume that __find_buddy_pfn() would not hand
> out invalid buddies. But you're right, it's generic:
> 
> pfn = 1024 (4M)
> order = MAX_ORDER - 1 = 10
> buddy_pfn = __find_buddy_pfn(pfn, order)
> 
> -> pfn ^ (1 << order) = 0
> 
> 
> If that page has no struct page (!pfn_valid), we're doomed, I agree. It
> would be problematic if we have alloc_contig_range() users with ranges
> not aligned/multiples of to 8 MB (MAX_ORDER) I guess. virtio-mem and
> gigantic pages should be fine. CMA might be problematic, though? Do we
> have such small CMA ranges or with such alignment? COuld be I guess.
> 
> cma_init_reserved_mem() only checks
> 
> alignment = PAGE_SIZE << max_t(unsigned long, MAX_ORDER - 1,
> pageblock_order);
> 

Thinking again (SPARSE), we always end up in a single memory section.
Usually, all pfns within a single section are valid. The only exception
is with HAVE_ARCH_PFN_VALID - arm and arm6.

arm64 also has HOLES_IN_ZONE - so we always check for pfn_valid() in
this code.

arm only has HAVE_ARCH_PFN_VALID with SPARSE on ARCH_OMAP1. So only in
that combination, we might run into that issue if I am not wrong.


Not sure about !SPARSE and mips.

-- 
Thanks,

David / dhildenb