Re: [RFC PATCH v4 8/8] hugetlb: use new vma_lock for pmd sharing synchronization

[Miaohe Lin]

On 2022/7/30 1:41, Mike Kravetz wrote:
> On 07/29/22 09:41, Miaohe Lin wrote:
>> On 2022/7/29 1:47, Mike Kravetz wrote:
>>> On 07/28/22 14:51, Miaohe Lin wrote:
>> snip
>>>>
>>>> Do we need to check &mm->mm_users == 0 here in case the address_space of corresponding process
>>>> has exited? In this case, mmdrop will drop the last reference and free the skipped_mm. So we will
>>>> use skipped_mm below after it's freed?
>>>>
>>>
>>> Good point!
>>> I think we need to wait to drop since we want to hold the read lock.
>>> Will update.
>>>
>>>>> +		vma = find_vma(skipped_mm, skipped_vm_start);
>>>>> +		if (!vma || vma->vm_file->f_mapping != mapping ||
>>>>
>>>> If skipped_vm_start is unmapped and remapped as a anon vma before we taking the mmap_read_lock,
>>>> vma->vm_file will be NULL?
>>>>
>>>
>>> IIUC, vma->vm_file will always be set even for an anon vma.  The fault
>>> code depends on this.  See beginning of hugetlb_fault() where we
>>> unconditionally do:
>>>
>>> mapping = vma->vm_file->f_mapping;
>>
>> What if vma is non-hugetlb anon vma?
>>
> 
> Right.  The first check after !vma should be for hugetlb as we do not
> care if non-hugetlb vmas and want to quit.  Something like,
> 
> 	vma = find_vma(skipped_mm, skipped_vm_start);
> 	if (!vma || !is_vm_hugetlb_page(vma) ||
> 

This looks good to me. Thanks.