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

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

On 11/03/2024 17:49, Matthew Wilcox wrote:
> On Mon, Mar 11, 2024 at 04:14:06PM +0000, Ryan Roberts wrote:
>>> With this patch, it took about 700 seconds in my xfstests run to find
>>> the one in shrink_folio_list().  It took 3260 seconds to catch the one
>>> in move_folios_to_lru().
>>
>> 54 hours?? That's before I even reported that we still had a bug! Either you're
>> anticipating my every move, or you have a lot of stuff running in parallel :)
> 
> One hour is 3600 seconds ;-)  So more like 53 minutes.
> 
>>> That one is only probably OK.  We usually manage to split a folio before
>>> we get to this point, so we should remove it from the deferred list then.
>>> You'd need to buy a lottery ticket if you managed to get hwpoison in a
>>> folio that was deferred split ...
>>
>> OK, but "probably"? Given hwpoison is surely not a hot path, why not just be
>> safe and call folio_undo_large_rmappable()?
> 
> Yes, I've added it; just commenting on the likelihood of being able to
> hit it in testing, even with aggressive error injection.
> 
>>>> But taking a step back, I wonder whether the charge() and uncharge() functions
>>>> should really be checking to see if the folio is on a deferred split list and if
>>>> so, then move the folio to the corect list?
>>>
>>> I don't think that's the right thing to do.  All of these places which
>>> uncharge a folio are part of the freeing path, so we always want it
>>> removed, not moved to a different deferred list.
>>
>> Well I'm just thinking about trying to be robust. Clearly you would prefer that
>> folio_undo_large_rmappable() has been called before uncharge(), then uncharge()
>> notices that there is nothing on the deferred list (and doesn't take the lock).
>> But if it's not, is it better to randomly crash (costing best part of a week to
>> debug) or move the folio to the right list?
> 
> Neither ;-)  The right option is to include the assertion that the
> deferred list is empty.  That way we get to see the backtrace of whoever
> forgot to take the folio off the deferred list.
> 
>> Alternatively, can we refactor so that there aren't 9 separate uncharge() call
>> sites. Those sites are all trying to free the folio so is there a way to better
>> refactor that into a single place (I guess the argument for the current
>> arrangement is reducing the number of times that we have to iterate through the
>> batch?). Then we only have to get it right once.
> 
> I have been wondering about a better way to do it.  I've also been
> looking a bit askance at put_pages_list() which doesn't do memcg
> uncharging ...
> 
>>>
>>> But what about mem_cgroup_move_account()?  Looks like that's memcg v1
>>> only?  Should still be fixed though ...
>>
>> Right.
>>
>> And what about the first bug you found with the local list corruption? I'm not
>> running with that fix so its obviously not a problem here. But I still think its
>> a bug that we should fix? list_for_each_entry_safe() isn't safe against
>> *concurrent* list modification, right?
> 
> I've been thinking about that too.  I decided that the local list is
> actually protected by the lock after all.  It's a bit fiddly to prove,
> but:
> 
> 1. We have a reference on every folio ahead on the list (not behind us,
> but see below)
> 2. If split_folio succeeds, it takes the lock that would protect the
> list we are on.
> 3. If it doesn't, and folio_put() turns out to be the last reference,
> __folio_put_large -> destroy_large_folio -> folio_undo_large_rmappable
> takes the lock that protects the list we would be on.
> 
> So we can analyse this loop as:
> 
> 	list_for_each_entry_safe(folio, next, &list, _deferred_list) {
> 		if (random() & 1)
> 			continue;
> 		spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
> 		list_del_init(&folio->_deferred_list);
> 		spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
> 	}
> 
> We're guaranteed that 'next' is a valid folio because we hold a refcount
> on it.  Anything left on the list between &list and next may have been
> removed from the list, but we don't look at those entries until after
> we take the split_queue_lock again to do the list_splice_tail().
> 
> I'm too scared to write a loop like this, but I don't think it contains
> a bug.

OK, wow. Now that I'm looking at the implementation of
list_for_each_entry_safe() along with your reasoning, that is clear. But it's
certainly not obvious looking at deferred_split_scan().