Re: [PATCH 3/3] mm: oom: show unreclaimable slab info when kernel panic

[Yang Shi]

On 9/15/17 5:00 AM, Tetsuo Handa wrote:
> On 2017/09/15 2:14, Yang Shi wrote:
> > @@ -1274,6 +1276,29 @@ static int slab_show(struct seq_file *m, void *p)
> >   	return 0;
> >   }
> >   
> > +void show_unreclaimable_slab()
> > +{
> > +	struct kmem_cache *s = NULL;
> > +	struct slabinfo sinfo;
> > +
> > +	memset(&sinfo, 0, sizeof(sinfo));
> > +
> > +	printk("Unreclaimable slabs:\n");
> > +	mutex_lock(&slab_mutex);
>
> Please avoid sleeping locks which potentially depend on memory allocation.
> There are
>
> 	mutex_lock(&slab_mutex);
> 	kmalloc(GFP_KERNEL);
> 	mutex_unlock(&slab_mutex);
>
> users which will fail to call panic() if they hit this path
Thanks for the heads up. Since this is just called by oom in panic path, 
so it sounds safe to just discard the mutex_lock()/mutex_unlock call 
since nobody can allocate memory without GFP_ATOMIC to change the 
statistics of slab.

Even though some GFP_ATOMIC callers allocate memory successfully, it 
should not have obvious impact to the slabinfo we need capture since 
typically GFP_ATOMIC allocation is small.

I will drop the mutext in v2 if no one has objection.

Thanks,
Yang

> > +	list_for_each_entry(s, &slab_caches, list) {
> > +		if (!is_root_cache(s))
> > +			continue;
> > +
> > +		get_slabinfo(s, &sinfo);
> > +
> > +		if (!is_reclaimable(s) && sinfo.num_objs > 0)
> > +			printk("%-17s %luKB\n", cache_name(s), K(sinfo.num_objs * s->size));
> > +	}
> > +	mutex_unlock(&slab_mutex);
> > +}
> > +EXPORT_SYMBOL(show_unreclaimable_slab);
> > +#undef K
> > +
> >   #if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
> >   void *memcg_slab_start(struct seq_file *m, loff_t *pos)
> >   {