Re: [PATCH v2 08/23] mm/slab_common: make kmalloc_large_node() consistent with kmalloc_large()

[Hyeonggon Yoo]

On Tue, Apr 26, 2022 at 07:15:06PM +0200, Vlastimil Babka wrote:
> On 4/14/22 10:57, Hyeonggon Yoo wrote:
> > Move tracepoints into kmalloc_large_node() and add missing flag fix code.
> > 
> > Signed-off-by: Hyeonggon Yoo <42.hyeyoo@xxxxxxxxx>
> 

Hello Vlastimil, thanks for review! ;-)

> Hm so there's a problem with the tracepoint's caller.
> 
> kmalloc_large() is only called from kmalloc() which is an inline  thus the
> callsite of kmalloc() calls directly kmalloc_large().
> So when kmalloc_large() does "trace_kmalloc(_RET_IP_, ...)" the _RET_IP_ is the
> callsite of kmalloc(), which is what we want.

kmalloc_large() had the exact problem before my series when called from __kmalloc().

On top of current slab/for-next:
  [000] .....    43.172574: kmalloc: call_site=__kmalloc+0x2aa/0x300 ptr=ffff88e2183a0000 bytes_req=12368 bytes_alloc=16384 gfp_flags=GFP_KERNEL

Considering different usecases of kmalloc_large_node() (called from kmalloc_node() or __kmalloc_node()),
I think we need trace/notrace version of kmalloc_large_node().

> 
> But with kmalloc_large_node()...
> 
> > ---
> >  mm/slab_common.c |  6 ++++++
> >  mm/slub.c        | 22 ++++------------------
> >  2 files changed, 10 insertions(+), 18 deletions(-)
> > 
> > diff --git a/mm/slab_common.c b/mm/slab_common.c
> > index e72089515030..cf17be8cd9ad 100644
> > --- a/mm/slab_common.c
> > +++ b/mm/slab_common.c
> > @@ -955,6 +955,9 @@ void *kmalloc_large_node(size_t size, gfp_t flags, int node)
> >  	void *ptr = NULL;
> >  	unsigned int order = get_order(size);
> >  
> > +	if (unlikely(flags & GFP_SLAB_BUG_MASK))
> > +		flags = kmalloc_fix_flags(flags);
> > +
> >  	flags |= __GFP_COMP;
> >  	page = alloc_pages_node(node, flags, order);
> >  	if (page) {
> > @@ -966,6 +969,9 @@ void *kmalloc_large_node(size_t size, gfp_t flags, int node)
> >  	ptr = kasan_kmalloc_large(ptr, size, flags);
> >  	/* As ptr might get tagged, call kmemleak hook after KASAN. */
> >  	kmemleak_alloc(ptr, size, 1, flags);
> > +	trace_kmalloc_node(_RET_IP_, ptr,
> > +			   size, PAGE_SIZE << order,
> > +			   flags, node);
> 
> ... the _RET_IP_ here would be __kmalloc_node() which is not useful.
> 
> >  	return ptr;
> >  }
> > diff --git a/mm/slub.c b/mm/slub.c
> > index 640712706f2b..f10a892f1772 100644
> > --- a/mm/slub.c
> > +++ b/mm/slub.c
> > @@ -4396,15 +4396,8 @@ void *__kmalloc_node(size_t size, gfp_t flags, int node)
> >  	struct kmem_cache *s;
> >  	void *ret;
> >  
> > -	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
> > -		ret = kmalloc_large_node(size, flags, node);
> > -
> > -		trace_kmalloc_node(_RET_IP_, ret,
> > -				   size, PAGE_SIZE << get_order(size),
> > -				   flags, node);
> 
> Here it was OK because __kmalloc_node is expanded from something inline
> coming from slab.h.
> 
> > -
> > -		return ret;
> > -	}
> > +	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
> > +		return kmalloc_large_node(size, flags, node);
> >  
> >  	s = kmalloc_slab(size, flags);
> >  
> > @@ -4861,15 +4854,8 @@ void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
> >  	struct kmem_cache *s;
> >  	void *ret;
> >  
> > -	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
> > -		ret = kmalloc_large_node(size, gfpflags, node);
> > -
> > -		trace_kmalloc_node(caller, ret,
> > -				   size, PAGE_SIZE << get_order(size),
> > -				   gfpflags, node);
> > -
> > -		return ret;
> > -	}
> > +	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
> > +		return kmalloc_large_node(size, gfpflags, node);
> 
> And here it even forgets the 'caller'.
>

Thanks for catching this.
I think notrace version + tracepoint would fit here.

> >  
> >  	s = kmalloc_slab(size, gfpflags);
> >  
> 

-- 
Thanks,
Hyeonggon