Re: [PATCH 2/2] mm/compaction: fix the range to pageblock_pfn_to_page()

[Michal Hocko <mhocko@suse.com>]

On Wed 08-10-25 12:17:29, David Hildenbrand wrote:
> On 02.10.25 05:31, Wei Yang wrote:
> > Current code may skip some part of the range to
> > pageblock_pfn_to_page() to check whether the range is in the same zone.
> > 
> > Function pageblock_pfn_to_page() is first introduced by commit 7d49d8868336
> > ("mm, compaction: reduce zone checking frequency in the migration scanner"),
> > in which it checks and isolates on the same range [pfn, block_end_pfn].
> > 
> > While after commit e1409c325fdc ("mm/compaction: pass only pageblock
> > aligned range to pageblock_pfn_to_page"), we operate on two different
> > ranges for check and isolation.
> > 
> >    * [block_start_pfn, block_end_pfn] to check it is in the same zone
> >    * [pfn, block_end_pfn] to do isolation
> > 
> > It miss some range to check when start_pfn and zone->zone_start_pfn is
> > in the same pageblock but (start_pfn < zone->zone_start_pfn). The range
> > before zone_start_pfn is missed.
> > 
> >           start_pfn     zone_start_pfn
> >      +----+-------------+-------------------+
> >      block_start_pfn                        block_end_pfn
> > 
> > This leads to the range check is passed, but it will isolate a range in
> > two different zones.
> 
> Can that actually happen? I recall that a zone always spans full pageblocks.

My recollection as well and we should rather focus on enforcing that to
be true than complicating already quite a complicated code.
-- 
Michal Hocko
SUSE Labs