Re: [PATCH] MIPS: Do not flush tlb when setting pmd entry

[maobibo]

On 06/22/2020 11:48 PM, Thomas Bogendoerfer wrote:
> On Sat, Jun 20, 2020 at 11:47:35AM +0800, maobibo wrote:
>>
>>
>> On 06/17/2020 07:14 PM, Thomas Bogendoerfer wrote:
>>> On Tue, Jun 16, 2020 at 06:34:21PM +0800, maobibo wrote:
>>>>
>>>>
>>>> On 06/15/2020 06:14 PM, Thomas Bogendoerfer wrote:
>>>>> On Wed, Jun 03, 2020 at 05:42:13PM +0800, Bibo Mao wrote:
>>>>>> Function set_pmd_at is to set pmd entry, if tlb entry need to
>>>>>> be flushed, there exists pmdp_huge_clear_flush alike function
>>>>>> before set_pmd_at is called. So it is not necessary to call
>>>>>> flush_tlb_all in this function.
>>>>>
>>>>> have you checked all set_pmd_at() calls ? I found a few case where
>>>>> it's not clear to me, if tlb flushing is done... If you think this
>>>>> is still the right thing to do, please change arch/mips/mm/pgtable-32.c
>>>>> as well.
>>>> well, I will double check this and do more testing about thp and hugepage.
>>>
>>> I was more concerned about
>>>
>>> fs/dax.c
>>> fs/proc/task_mmu.c
>>> mm/rmap.c
>>
>> I think that flush_tlb_all should not be called in function set_pmd_at
>> on mips platform. However update_mmu_cache_pmd() should be called __after__
>> set_pmd_at() function to update tlb entry at some places, it is another issue.
>> Here is my analysis in the three files where set_pmd_at is called.
>> [..]
> 
> thank you for confirming that we are good with removing flush_tlb_all().
Sorry, there is something wrong if remove flush_tlb_all(). If pmd_none is true,
pmd points to invalid_pte_table, maybe there exists pte entry with normal page size
for fault address. And we need invalidate this pte entry like it is done in function build_huge_handler_tail in file arch/mips/mm/tlbex.c

I will send another patch.
> 
> Thomas.
>