Re: [PATCH v3 10/18] mm: Allow non-hugetlb large folios to be batch processed

[Ryan Roberts <ryan.roberts@arm.com>]

On 10/03/2024 11:01, Ryan Roberts wrote:
> On 06/03/2024 16:09, Matthew Wilcox wrote:
>> On Wed, Mar 06, 2024 at 01:42:06PM +0000, Ryan Roberts wrote:
>>> When running some swap tests with this change (which is in mm-stable)
>>> present, I see BadThings(TM). Usually I see a "bad page state"
>>> followed by a delay of a few seconds, followed by an oops or NULL
>>> pointer deref. Bisect points to this change, and if I revert it,
>>> the problem goes away.
>>
>> That oops is really messed up ;-(  We're clearly got two CPUs oopsing at
>> the same time and it's all interleaved.  That said, I can pick some
>> nuggets out of it.
>>
>>> [   76.239466] BUG: Bad page state in process usemem  pfn:2554a0
>>> [   76.240196] kernel BUG at include/linux/mm.h:1120!
>>
>> These are the two different BUGs being called simultaneously ...
>>
>> The first one is bad_page() in page_alloc.c and the second is
>> put_page_testzero()
>>         VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
>>
>> I'm sure it's significant that both of these are the same page (pfn
>> 2554a0).  Feels like we have two CPUs calling put_folio() at the same
>> time, and one of them underflows.  It probably doesn't matter which call
>> trace ends up in bad_page() and which in put_page_testzero().
>>
>> One of them is coming from deferred_split_scan(), which is weird because
>> we can see the folio_try_get() earlier in the function.  So whatever
>> this folio was, we found it on the deferred split list, got its refcount,
>> moved it to the local list, either failed to get the lock, or
>> successfully got the lock, split it, unlocked it and put it.
>>
>> (I can see this was invoked from page fault -> memcg shrinking.  That's
>> probably irrelevant but explains some of the functions in the backtrace)
>>
>> The other call trace comes from migrate_folio_done() where we're putting
>> the _source_ folio.  That was called from migrate_pages_batch() which
>> was called from kcompactd.
>>
>> Um.  Where do we handle the deferred list in the migration code?
>>
>>
>> I've also tried looking at this from a different angle -- what is it
>> about this commit that produces this problem?  It's a fairly small
>> commit:
>>
>> -               if (folio_test_large(folio)) {
>> +               /* hugetlb has its own memcg */
>> +               if (folio_test_hugetlb(folio)) {
>>                         if (lruvec) {
>>                                 unlock_page_lruvec_irqrestore(lruvec, flags);
>>                                 lruvec = NULL;
>>                         }
>> -                       __folio_put_large(folio);
>> +                       free_huge_folio(folio);
>>
>> So all that's changed is that large non-hugetlb folios do not call
>> __folio_put_large().  As a reminder, that function does:
>>
>>         if (!folio_test_hugetlb(folio))
>>                 page_cache_release(folio);
>>         destroy_large_folio(folio);
>>
>> and destroy_large_folio() does:
>>         if (folio_test_large_rmappable(folio))
>>                 folio_undo_large_rmappable(folio);
>>
>>         mem_cgroup_uncharge(folio);
>>         free_the_page(&folio->page, folio_order(folio));
>>
>> So after my patch, instead of calling (in order):
>>
>> 	page_cache_release(folio);
>> 	folio_undo_large_rmappable(folio);
>> 	mem_cgroup_uncharge(folio);
>> 	free_unref_page()
>>
>> it calls:
>>
>> 	__page_cache_release(folio, &lruvec, &flags);
>> 	mem_cgroup_uncharge_folios()
>> 	folio_undo_large_rmappable(folio);
> 
> I was just looking at this again, and something pops out...
> 
> You have swapped the order of folio_undo_large_rmappable() and
> mem_cgroup_uncharge(). But folio_undo_large_rmappable() calls
> get_deferred_split_queue() which tries to get the split queue from
> folio_memcg(folio) first and falls back to pgdat otherwise. If you are now
> calling mem_cgroup_uncharge_folios() first, will that remove the folio from the
> cgroup? Then we are operating on the wrong list? (just a guess based on the name
> of the function...)

Infact, looking at mem_cgroup_uncharge_folios() that's exactly what it does - it
calls uncharge_folio(), which zeros memcg_data. And this is completely
consistent with the behaviour I've seen, including the original bisection
result. And it explains why the "workaround to re-narrow" the window is 100%
successful - its reverting the ordering to be correct again.

Assuming you agree, I'll leave you to work up the patch(s).

> 
> 
> 
>>
>> So have I simply widened the window for this race, whatever it is
>> exactly?  Something involving mis-handling of the deferred list?
>>
>