Re: [External] Re: [PATCH v2 07/19] mm/hugetlb: Free the vmemmap pages associated with each hugetlb page

[Mike Kravetz]

On 10/28/20 11:13 PM, Muchun Song wrote:
> On Thu, Oct 29, 2020 at 7:42 AM Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote:
>>
>> On 10/26/20 7:51 AM, Muchun Song wrote:
>>> +
>>> +static inline spinlock_t *vmemmap_pmd_lockptr(pmd_t *pmd)
>>> +{
>>> +     static DEFINE_SPINLOCK(pgtable_lock);
>>> +
>>> +     return &pgtable_lock;
>>> +}
>>
>> This is just a global lock.  Correct?  And hugetlb specific?
> 
> Yes, it is a global lock. Originally, I wanted to use the pmd lock(e.g.
> pmd_lockptr()). But we need to allocate memory for the spinlock and
> initialize it when ALLOC_SPLIT_PTLOCKS. It may increase the
> complexity.
> 
> And I think that here alloc/free hugetlb pages is not a frequent operation.
> So I finally use a global lock. Maybe it is enough.
> 
>>
>> It should be OK as the page table entries for huegtlb pages will not
>> overlap with other entries.
> 
> Does "hugetlb specific" mean the pmd lock? or per hugetlb lock?
> If it is pmd lock, this is fine to me. If not, it may not be enough.
> Because the lock also guards the splitting of pmd pgtable.

By "hugetlb specific", I was trying to say that only hugetlb code would
use this lock.  It is not a concern now.  However, there has been talk
about other code doing something similar to remove struct pages.  If that
ever happens then we will need a different locking scheme.

Disregard my statement about there being no overlap.  I was confusing
page tables for huge pages with page tables for mappings mmap entries
of huge pages.
-- 
Mike Kravetz