Re: [PATCH v3 bpf-next] net: veth: alloc skb in bulk for ndo_xdp_xmit

[Jesper Dangaard Brouer <brouer@redhat.com>]

On Thu, 4 Feb 2021 01:14:56 +0100
Daniel Borkmann <daniel@iogearbox.net> wrote:

> On 1/29/21 11:04 PM, Lorenzo Bianconi wrote:
> > Split ndo_xdp_xmit and ndo_start_xmit use cases in veth_xdp_rcv routine
> > in order to alloc skbs in bulk for XDP_PASS verdict.
> > Introduce xdp_alloc_skb_bulk utility routine to alloc skb bulk list.
> > The proposed approach has been tested in the following scenario:  
> [...]
> > diff --git a/net/core/xdp.c b/net/core/xdp.c
> > index 0d2630a35c3e..05354976c1fc 100644
> > --- a/net/core/xdp.c
> > +++ b/net/core/xdp.c
> > @@ -514,6 +514,17 @@ void xdp_warn(const char *msg, const char *func, const int line)
> >   };
> >   EXPORT_SYMBOL_GPL(xdp_warn);
> >   
> > +int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
> > +{
> > +	n_skb = kmem_cache_alloc_bulk(skbuff_head_cache, gfp,
> > +				      n_skb, skbs);  
> 
> Applied, but one question I was wondering about when reading the kmem_cache_alloc_bulk()
> code was whether it would be safer to simply test for kmem_cache_alloc_bulk() != n_skb
> given it could potentially in future also alloc less objs than requested, but I presume
> if such extension would get implemented then call-sites might need to indicate 'best
> effort' somehow via flag instead (to handle < n_skb case). Either way all current callers
> assume for != 0 that everything went well, so lgtm.

It was Andrew (AKPM) that wanted the API to either return the requested
number of objects or fail. I respected the MM-maintainers request at
that point, even-though I wanted the other API as there is a small
performance advantage (not crossing page boundary in SLUB).

At that time we discussed it on MM-list, and I see his/the point:
If API can allocate less objs than requested, then think about how this
complicated the surrounding code. E.g. in this specific code we already
have VETH_XDP_BATCH(16) xdp_frame objects, which we need to get 16 SKB
objects for.  What should the code do if it cannot get 16 SKBs(?).

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer