Re: [PATCH 1/3] mm/slub: fix the race between validate_slab and slab_free

[Hyeonggon Yoo <42.hyeyoo@gmail.com>]

On Tue, May 31, 2022 at 11:47:22AM +0800, Muchun Song wrote:
> On Sun, May 29, 2022 at 11:37:06AM +0000, Hyeonggon Yoo wrote:
> > On Sun, May 29, 2022 at 04:15:33PM +0800, Rongwei Wang wrote:
> > > In use cases where allocating and freeing slab frequently, some
> > > error messages, such as "Left Redzone overwritten", "First byte
> > > 0xbb instead of 0xcc" would be printed when validating slabs.
> > > That's because an object has been filled with SLAB_RED_INACTIVE,
> > > but has not been added to slab's freelist. And between these
> > > two states, the behaviour of validating slab is likely to occur.
> > > 
> > > Actually, it doesn't mean the slab can not work stably. But, these
> > > confusing messages will disturb slab debugging more or less.
> > > 
> > > Signed-off-by: Rongwei Wang <rongwei.wang@linux.alibaba.com>
> > 
> > Have you observed it or it's from code inspection?
> > 
> > > ---
> > >  mm/slub.c | 40 +++++++++++++++++-----------------------
> > >  1 file changed, 17 insertions(+), 23 deletions(-)
> > > 
> > > diff --git a/mm/slub.c b/mm/slub.c
> > > index ed5c2c03a47a..310e56d99116 100644
> > > --- a/mm/slub.c
> > > +++ b/mm/slub.c
> > > @@ -1374,15 +1374,12 @@ static noinline int free_debug_processing(
> > >  	void *head, void *tail, int bulk_cnt,
> > >  	unsigned long addr)
> > >  {
> > > -	struct kmem_cache_node *n = get_node(s, slab_nid(slab));
> > >  	void *object = head;
> > >  	int cnt = 0;
> > > -	unsigned long flags, flags2;
> > > +	unsigned long flags;
> > >  	int ret = 0;
> > >  
> > > -	spin_lock_irqsave(&n->list_lock, flags);
> > > -	slab_lock(slab, &flags2);
> > > -
> > > +	slab_lock(slab, &flags);
> > >  	if (s->flags & SLAB_CONSISTENCY_CHECKS) {
> > >  		if (!check_slab(s, slab))
> > >  			goto out;
> > > @@ -1414,8 +1411,7 @@ static noinline int free_debug_processing(
> > >  		slab_err(s, slab, "Bulk freelist count(%d) invalid(%d)\n",
> > >  			 bulk_cnt, cnt);
> > >  
> > > -	slab_unlock(slab, &flags2);
> > > -	spin_unlock_irqrestore(&n->list_lock, flags);
> > > +	slab_unlock(slab, &flags);
> > >  	if (!ret)
> > >  		slab_fix(s, "Object at 0x%p not freed", object);
> > >  	return ret;
> > > @@ -3304,7 +3300,7 @@ static void __slab_free(struct kmem_cache *s, struct slab *slab,
> > >  
> > >  {
> > >  	void *prior;
> > > -	int was_frozen;
> > > +	int was_frozen, to_take_off = 0;
> > >  	struct slab new;
> > >  	unsigned long counters;
> > >  	struct kmem_cache_node *n = NULL;
> > > @@ -3315,15 +3311,19 @@ static void __slab_free(struct kmem_cache *s, struct slab *slab,
> > >  	if (kfence_free(head))
> > >  		return;
> > >  
> > > +	n = get_node(s, slab_nid(slab));
> > > +	spin_lock_irqsave(&n->list_lock, flags);
> > > +
> > 
> > Oh please don't do this.
> > 
> > SLUB free slowpath can be hit a lot depending on workload.
> > 
> > __slab_free() try its best not to take n->list_lock. currently takes n->list_lock
> > only when the slab need to be taken from list.
> > 
> > Unconditionally taking n->list_lock will degrade performance.
> >
> 
> I can confirm you are right. We have encountered this issue in practise.
> We have deployed somen HDFS instance on a 256-CPU machine.  When there
> are lots of IO requests from users, we can see lots of threads are contended
> on n->list_lock.  Lots of call traces are like following:
> 
>   ffffffff810dfbb4 native_queued_spin_lock_slowpath+0x1a4
>   ffffffff81780ffb _raw_spin_lock+0x1b
>   ffffffff8127327e get_partial_node.isra.81+0x5e
>   ffffffff812752d3 ___slab_alloc+0x2f3
>   ffffffff8127559c __slab_alloc+0x1c
>   ffffffff81275828 kmem_cache_alloc+0x278
>   ffffffff812e9e3d alloc_buffer_head+0x1d
>   ffffffff812e9f74 alloc_page_buffers+0xa4
>   ffffffff812eb0e9 create_empty_buffers+0x19
>   ffffffff812eb37d create_page_buffers+0x7d
>   ffffffff812ed78d __block_write_begin_int+0x9d
> 
> I thought it was because there are lots of threads which consume local
> CPU slab cache quickly and then both of them try to get a new slab
> from node partial list.  Since there are 256 CPUs, the contention
> is more competitive and easy to be visible.
> 
> Any thoughts on this issue (e.e. how to ease contention)? Comments
> are welcome.

How does increasing number of partial slabs affect your situation?
(increasing /sys/slab/<cache name>/cpu_partial)

> Thanks.
> 
>