Re: [PATCH v3 2/3] mm/migrate_device.c: Copy pte dirty bit to page

[David Hildenbrand <david@redhat.com>]

On 26.08.22 23:37, Peter Xu wrote:
> On Fri, Aug 26, 2022 at 06:46:02PM +0200, David Hildenbrand wrote:
>> On 26.08.22 17:55, Peter Xu wrote:
>>> On Fri, Aug 26, 2022 at 04:47:22PM +0200, David Hildenbrand wrote:
>>>>> To me anon exclusive only shows this mm exclusively owns this page. I
>>>>> didn't quickly figure out why that requires different handling on tlb
>>>>> flushs.  Did I perhaps miss something?
>>>>
>>>> GUP-fast is the magic bit, we have to make sure that we won't see new
>>>> GUP pins, thus the TLB flush.
>>>>
>>>> include/linux/mm.h:gup_must_unshare() contains documentation.
>>>
>>> Hmm.. Shouldn't ptep_get_and_clear() (e.g., xchg() on x86_64) already
>>> guarantees that no other process/thread will see this pte anymore
>>> afterwards?
>>
>> You could have a GUP-fast thread that just looked up the PTE and is
>> going to pin the page afterwards, after the ptep_get_and_clear()
>> returned. You'll have to wait until that thread finished.
> 

Good that we're talking about it, very helpful! If that's actually not
required -- good.


What I learned how GUP-fast and TLB flushes interact is the following:

GUP-fast disables local interrupts. A TLB flush will send an IPI and
wait until it has been processed. This implies, that once the TLB flush
succeeded, that the interrupt was handled and GUP-fast cannot be running
anymore.

BUT, there is the new RCU variant nowadays, and the TLB flush itself
should not actually perform such a sync. They merely protect the page
tables from getting freed and the THP from getting split IIUC. And
you're correct that that wouldn't help.


> IIUC the early tlb flush won't protect concurrent fast-gup from happening,
> but I think it's safe because fast-gup will check pte after pinning, so
> either:
> 
>   (1) fast-gup runs before ptep_get_and_clear(), then
>       page_try_share_anon_rmap() will fail properly, or,
> 
>   (2) fast-gup runs during or after ptep_get_and_clear(), then fast-gup
>       will see that either the pte is none or changed, then it'll fail the
>       fast-gup itself.

I think you're right and I might have managed to confuse myself with the
write_protect_page() comment. I placed the gup_must_unshare() check
explicitly after the "pte changed" check for this reason. So once the
PTE was cleared, GUP-fast would undo any pin performed on a now-stale PTE.

> 
>>
>> Another user that relies on this interaction between GUP-fast and TLB
>> flushing is for example mm/ksm.c:write_protect_page()
>>
>> There is a comment in there explaining the interaction a bit more detailed.
>>
>> Maybe we'll be able to handle this differently in the future (maybe once
>> this turns out to be an actual performance problem). Unfortunately,
>> mm->write_protect_seq isn't easily usable because we'd need have to make
>> sure we're the exclusive writer.
>>
>>
>> For now, it's not too complicated. For PTEs:
>> * try_to_migrate_one() already uses ptep_clear_flush().
>> * try_to_unmap_one() already conditionally used ptep_clear_flush().
>> * migrate_vma_collect_pmd() was the one case that didn't use it already
>>  (and I wonder why it's different than try_to_migrate_one()).
> 
> I'm not sure whether I fully get the point, but here one major difference
> is all the rest handles one page, so a tlb flush alongside with the pte
> clear sounds reasonable.  Even if so try_to_unmap_one() was modified to use
> tlb batching, but then I see that anon exclusive made that batching
> conditional.  I also have question there on whether we can keep using the
> tlb batching even with anon exclusive pages there.
> 
> In general, I still don't see how stall tlb could affect anon exclusive
> pages on racing with fast-gup, because the only side effect of a stall tlb
> is unwanted page update iiuc, the problem is fast-gup doesn't even use tlb,
> afaict..

I have the gut feeling that the comment in write_protect_page() is
indeed stale, and that clearing PageAnonExclusive doesn't strictly need
the TLB flush.

I'll try to refresh my memory if there was any other case that I had to
handle over the weekend.

Thanks!

-- 
Thanks,

David / dhildenb