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

[Huang, Ying]

Nadav Amit <nadav.amit at gmail.com> writes:

> On Aug 17, 2022, at 12:17 AM, Huang, Ying <ying.huang at intel.com> wrote:
>
>> Alistair Popple <apopple at nvidia.com> writes:
>> 
>>> Peter Xu <peterx at redhat.com> writes:
>>> 
>>>> On Wed, Aug 17, 2022 at 11:49:03AM +1000, Alistair Popple wrote:
>>>>> Peter Xu <peterx at redhat.com> writes:
>>>>> 
>>>>>> On Tue, Aug 16, 2022 at 04:10:29PM +0800, huang ying wrote:
>>>>>>>> @@ -193,11 +194,10 @@ static int migrate_vma_collect_pmd(pmd_t *pmdp,
>>>>>>>>                        bool anon_exclusive;
>>>>>>>>                        pte_t swp_pte;
>>>>>>>> 
>>>>>>>> +                       flush_cache_page(vma, addr, pte_pfn(*ptep));
>>>>>>>> +                       pte = ptep_clear_flush(vma, addr, ptep);
>>>>>>> 
>>>>>>> Although I think it's possible to batch the TLB flushing just before
>>>>>>> unlocking PTL.  The current code looks correct.
>>>>>> 
>>>>>> If we're with unconditionally ptep_clear_flush(), does it mean we should
>>>>>> probably drop the "unmapped" and the last flush_tlb_range() already since
>>>>>> they'll be redundant?
>>>>> 
>>>>> This patch does that, unless I missed something?
>>>> 
>>>> Yes it does.  Somehow I didn't read into the real v2 patch, sorry!
>>>> 
>>>>>> If that'll need to be dropped, it looks indeed better to still keep the
>>>>>> batch to me but just move it earlier (before unlock iiuc then it'll be
>>>>>> safe), then we can keep using ptep_get_and_clear() afaiu but keep "pte"
>>>>>> updated.
>>>>> 
>>>>> I think we would also need to check should_defer_flush(). Looking at
>>>>> try_to_unmap_one() there is this comment:
>>>>> 
>>>>> 			if (should_defer_flush(mm, flags) && !anon_exclusive) {
>>>>> 				/*
>>>>> 				 * We clear the PTE but do not flush so potentially
>>>>> 				 * a remote CPU could still be writing to the folio.
>>>>> 				 * If the entry was previously clean then the
>>>>> 				 * architecture must guarantee that a clear->dirty
>>>>> 				 * transition on a cached TLB entry is written through
>>>>> 				 * and traps if the PTE is unmapped.
>>>>> 				 */
>>>>> 
>>>>> And as I understand it we'd need the same guarantee here. Given
>>>>> try_to_migrate_one() doesn't do batched TLB flushes either I'd rather
>>>>> keep the code as consistent as possible between
>>>>> migrate_vma_collect_pmd() and try_to_migrate_one(). I could look at
>>>>> introducing TLB flushing for both in some future patch series.
>>>> 
>>>> should_defer_flush() is TTU-specific code?
>>> 
>>> I'm not sure, but I think we need the same guarantee here as mentioned
>>> in the comment otherwise we wouldn't see a subsequent CPU write that
>>> could dirty the PTE after we have cleared it but before the TLB flush.
>>> 
>>> My assumption was should_defer_flush() would ensure we have that
>>> guarantee from the architecture, but maybe there are alternate/better
>>> ways of enforcing that?
>>>> IIUC the caller sets TTU_BATCH_FLUSH showing that tlb can be omitted since
>>>> the caller will be responsible for doing it.  In migrate_vma_collect_pmd()
>>>> iiuc we don't need that hint because it'll be flushed within the same
>>>> function but just only after the loop of modifying the ptes.  Also it'll be
>>>> with the pgtable lock held.
>>> 
>>> Right, but the pgtable lock doesn't protect against HW PTE changes such
>>> as setting the dirty bit so we need to ensure the HW does the right
>>> thing here and I don't know if all HW does.
>> 
>> This sounds sensible.  But I take a look at zap_pte_range(), and find
>> that it appears that the implementation requires the PTE dirty bit to be
>> write-through.  Do I miss something?
>> 
>> Hi, Nadav, Can you help?
>
> Sorry for joining the discussion late. I read most ofthis thread and I hope
> I understand what you ask me. So at the risk of rehashing or repeating what
> you already know - here are my 2 cents. Feel free to ask me again if I did
> not understand your questions:
>
> 1. ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH is currently x86 specific. There is a
> recent patch that want to use it for arm64 as well [1]. The assumption that
> Alistair cited from the code (regarding should_defer_flush()) might not be
> applicable to certain architectures (although most likely it is). I tried
> to encapsulate the logic on whether an unflushed RO entry can become dirty
> in an arch specific function [2].
>
> 2. Having said all of that, using the logic of “flush if there are pending
> TLB flushes for this mm” as done by UNMAP_TLB_FLUSH makes sense IMHO
> (although I would have considered doing it in finer granularity of
> VMA/page-table as I proposed before and got somewhat lukewarm response [3]).
>
> 3. There is no question that flushing after dropping the ptl is wrong. But
> reading the thread, I think that you only focus on whether a dirty
> indication might get lost. The problem, I think, is bigger, as it might also
> cause correction problems after concurrently removing mappings. What happens
> if we get for a clean PTE something like:
>
>   CPU0					CPU1
>   ----					----
>
>   migrate_vma_collect_pmd()
>   [ defer flush, release ptl ]
> 					madvise(MADV_DONTNEED)
> 					-> zap_pte_range()
> 					[ PTE not present; mmu_gather
> 					  not updated ]
> 					
> 					[ no flush; stale PTE in TLB ]
>
> 					[ page is still accessible ]
>
> [ might apply to munmap(); I usually regard MADV_DONTNEED since it does
>   not call mmap_write_lock() ]

Yes.  You are right.  Flushing after PTL unlocking can cause more
serious problems.

I also want to check whether the dirty bit can be lost in
zap_pte_range(), where the TLB flush will be delayed if the PTE is
clean.  Per my understanding, PTE dirty bit must be write-through to
make this work correctly.  And I cannot imagine how CPU can do except
page fault if

- PTE is non-present
- TLB entry is clean
- CPU writes the page

Then, can we assume PTE dirty bit are always write-through on any
architecture?

> 4. Having multiple TLB flushing infrastructures makes all of these
> discussions very complicated and unmaintainable. I need to convince myself
> in every occasion (including this one) whether calls to
> flush_tlb_batched_pending() and tlb_flush_pending() are needed or not.
>
> What I would like to have [3] is a single infrastructure that gets a
> “ticket” (generation when the batching started), the old PTE and the new PTE
> and checks whether a TLB flush is needed based on the arch behavior and the
> current TLB generation. If needed, it would update the “ticket” to the new
> generation. Andy wanted a ring for pending TLB flushes, but I think it is an
> overkill with more overhead and complexity than needed.
>
> But the current situation in which every TLB flush is a basis for long
> discussions and prone to bugs is impossible.
>
> I hope it helps. Let me know if you want me to revive the patch-set or other
> feedback.
>
> [1] https://lore.kernel.org/all/20220711034615.482895-5-21cnbao@gmail.com/
> [2] https://lore.kernel.org/all/20220718120212.3180-13-namit@vmware.com/
> [3] https://lore.kernel.org/all/20210131001132.3368247-16-namit@vmware.com/

Haven't looked at this in depth yet.  Will do that.

Best Regards,
Huang, Ying