Re: [PATCH v2 1/2] mm/migrate_device.c: Copy pte dirty bit to page

[Alistair Popple <apopple@nvidia.com>]

"Huang, Ying" <ying.huang@intel.com> writes:

> Peter Xu <peterx@redhat.com> writes:
>
>> On Thu, Aug 18, 2022 at 02:34:45PM +0800, Huang, Ying wrote:
>>> > In this specific case, the only way to do safe tlb batching in my mind is:
>>> >
>>> > 	pte_offset_map_lock();
>>> > 	arch_enter_lazy_mmu_mode();
>>> >         // If any pending tlb, do it now
>>> >         if (mm_tlb_flush_pending())
>>> > 		flush_tlb_range(vma, start, end);
>>> >         else
>>> >                 flush_tlb_batched_pending();
>>>
>>> I don't think we need the above 4 lines.  Because we will flush TLB
>>> before we access the pages.

I agree. For migration the TLB flush is only important if the PTE is
present, and in that case we do a TLB flush anyway.

>> Could you elaborate?
>
> As you have said below, we don't use non-present PTEs and flush present
> PTEs before we access the pages.
>
>>> Can you find any issue if we don't use the above 4 lines?
>>
>> It seems okay to me to leave stall tlb at least within the scope of this
>> function. It only collects present ptes and flush propoerly for them.  I
>> don't quickly see any other implications to other not touched ptes - unlike
>> e.g. mprotect(), there's a strong barrier of not allowing further write
>> after mprotect() returns.
>
> Yes.  I think so too.
>
>> Still I don't know whether there'll be any side effect of having stall tlbs
>> in !present ptes because I'm not familiar enough with the private dev swap
>> migration code.  But I think having them will be safe, even if redundant.

What side-effect were you thinking of? I don't see any issue with not
TLB flushing stale device-private TLBs prior to the migration because
they're not accessible anyway and shouldn't be in any TLB.

> I don't think it's a good idea to be redundant.  That may hide the real
> issue.
>
> Best Regards,
> Huang, Ying

Thanks all for the discussion. Having done some more reading I agree
that it's safe to assume HW dirty bits are write-through, so will remove
the ptep_clear_flush() and use ptep_get_and_clear() instead. Will split
out the TLB flushing fix into a separate patch in this series.

 - Alistair