Re: [PATCH 3/3 net-next] net_device: Add minimal padding to allow for net_device pointer alignment

[Thomas Graf <tgraf@suug.ch>]

On 04/09/15 at 05:34pm, Eric Dumazet wrote:
> On Fri, 2015-04-10 at 01:43 +0200, Thomas Graf wrote:
> > diff --git a/net/core/dev.c b/net/core/dev.c
> > index b2775f0..2b43aba 100644
> > --- a/net/core/dev.c
> > +++ b/net/core/dev.c
> > @@ -6794,8 +6794,7 @@ struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
> >  		alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
> >  		alloc_size += sizeof_priv;
> >  	}
> > -	/* ensure 32-byte alignment of whole construct */
> > -	alloc_size += NETDEV_ALIGN - 1;
> > +	alloc_size += PTR_ALIGN_PAD(struct net_device, NETDEV_ALIGN);
> >  
> >  	p = kzalloc(alloc_size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
> >  	if (!p)
> 
> 
> I believe code intent was to get an alignment even if kmalloc() returns
> say 0xxxxx0008.
> 
> AFAIK, SLAB in debug mode was able to do this.

I assumed kmalloc would guarantee alignment of the return pointer as
indicated by __alignof__ but I now realize that this can't possibly be the
case. kmalloc only knows about the size. We'll have to drop patch 3.
Fortunately it's still below 2K.