Re: [PATCH 5/5] mm/slub: Refactor deactivate_slab()

From: Vlastimil Babka <vbabka@suse.cz>
To: Hyeonggon Yoo <42.hyeyoo@gmail.com>
Cc: linux-mm@kvack.org, Roman Gushchin <guro@fb.com>,
	Andrew Morton <akpm@linux-foundation.org>,
	linux-kernel@vger.kernel.org,
	Joonsoo Kim <iamjoonsoo.kim@lge.com>,
	David Rientjes <rientjes@google.com>,
	Christoph Lameter <cl@linux.com>,
	Pekka Enberg <penberg@kernel.org>
Subject: Re: [PATCH 5/5] mm/slub: Refactor deactivate_slab()
Date: Fri, 25 Feb 2022 11:07:45 +0100	[thread overview]
Message-ID: <21c9fa1a-a003-3325-dd92-982ae3102336@suse.cz> (raw)
In-Reply-To: <YhimVo7oKmnMSkYS@ip-172-31-19-208.ap-northeast-1.compute.internal>

On 2/25/22 10:50, Hyeonggon Yoo wrote:
> On Fri, Feb 25, 2022 at 09:34:09AM +0000, Hyeonggon Yoo wrote:
>> On Thu, Feb 24, 2022 at 07:16:11PM +0100, Vlastimil Babka wrote:
>> > On 2/21/22 11:53, Hyeonggon Yoo wrote:
>> > > Simply deactivate_slab() by removing variable 'lock' and replacing
>> > > 'l' and 'm' with 'mode'. Instead, remove slab from list and unlock
>> > > n->list_lock when cmpxchg_double() fails, and then retry.
>> > > 
>> > > One slight functional change is releasing and taking n->list_lock again
>> > > when cmpxchg_double() fails. This is not harmful because SLUB avoids
>> > > deactivating slabs as much as possible.
>> > > 
>> > > Signed-off-by: Hyeonggon Yoo <42.hyeyoo@gmail.com>
>> > 
>> > Hm I wonder if we could simplify even a bit more. Do we have to actually
>> > place the slab on a partial (full) list before the cmpxchg, only to remove
>> > it when cmpxchg fails? Seems it's to avoid anyone else seeing the slab
>> > un-frozen, but not on the list, which would be unexpected. However if anyone
>> > sees such slab, they have to take the list_lock first to start working with
>> > the slab... so this should be safe, because we hold the list_lock here, and
>> > will place the slab on the list before we release it. But it thus shouldn't
>> > matter if the placement happens before or after a successful cmpxchg, no? So
>> > we can only do it once after a successful cmpxchg and need no undo's?
>> >
>> 
>> My thought was similar. But after testing I noticed that &n->list_lock prevents
>> race between __slab_free() and deactivate_slab().
>> 
>> > Specifically AFAIK the only possible race should be with a __slab_free()
>> > which might observe !was_frozen after we succeed an unfreezing cmpxchg and
>> > go through the
>> > "} else { /* Needs to be taken off a list */"
>> > branch but then it takes the list_lock as the first thing, so will be able
>> > to proceed only after the slab is actually on the list.
>> > 
>> > Do I miss anything or would you agree?
>> >
>> 
>> It's so tricky.
>> 
>> I tried to simplify more as you said. Seeing frozen slab on list was not
>> problem. But the problem was that something might interfere between
>> cmpxchg_double() and taking spinlock.
>> 
>> This is what I faced:
>> 
>> 	CPU A				CPU B
>> deactivate_slab() {			__slab_free() {
>> 	/* slab is full */
>> 	slab.frozen = 0;
>> 	cmpxchg_double();
>> 						/* Hmm... 
>> 						slab->frozen == 0 &&
>> 						slab->freelist != NULL?
>> 						Oh This must be on the list.. */
> 						Oh this is wrong.
> 						slab->freelist must be
> 						NULL because it's full
> 						slab.
> 
> 						It's more complex
> 						than I thought...
> 
> 
>> 						spin_lock_irqsave();
>> 						cmpxchg_double();
>> 						/* Corruption: slab
>> 						 * was not yet inserted to
>> 						 * list but try removing */
>> 						remove_full();
>> 						spin_unlock_irqrestore();
>> 					}
>> 	spin_lock_irqsave();
>> 	add_full();
>> 	spin_unlock_irqrestore();
>> }
> 
> So it was...
> 
>  	CPU A				CPU B
>  deactivate_slab() {			__slab_free() {
>  	/* slab is full */
>  	slab.frozen = 0;
>  	cmpxchg_double();
>  						/*
> 							Hmm... 
> 							!was_frozen &&
> 							prior == NULL?
> 							Let's freeze this!
> 						*/
> 						put_cpu_partial();
>  					}
>  	spin_lock_irqsave();

Yeah in my proposal I didn't intend to only take spin_lock_irqsave() here.
My idea for CPU A would be something like:

spin_lock_irqsave();
slab.frozen = 0;
if (cmpxchg_double()); {
	/* success */
	add_partial(); // (or add_full())
	spin_unlock_irqrestore();
} else {
	/* fail */
	spin_unlock_irqrestore();
	goto redo;
}

So we would still have the list_lock protection around cmpxchg as in the
current code. We just wouldn't do e.g. add_partial() before cmpxchg, only to
remove_partial() when cmpxchg failed.

>  	add_full();
> 	/* It's now frozen by CPU B and at the same time on full list */
>  	spin_unlock_irqrestore();
> 
> And &n->list_lock prevents such a race.