On Mon, Mar 03, 2025 at 05:36:41PM +0530, Malladi, Meghana wrote:
> > > +static int emac_run_xdp(struct prueth_emac *emac, struct xdp_buff *xdp,
> > > + struct page *page)
> > > +{
> > > + struct net_device *ndev = emac->ndev;
> > > + int err, result = ICSSG_XDP_PASS;
> > > + struct bpf_prog *xdp_prog;
> > > + struct xdp_frame *xdpf;
> > > + int q_idx;
> > > + u32 act;
> > > +
> > > + xdp_prog = READ_ONCE(emac->xdp_prog);
> > > + act = bpf_prog_run_xdp(xdp_prog, xdp);
> > > + switch (act) {
> > > + case XDP_PASS:
> > > + break;
> > > + case XDP_TX:
> > > + /* Send packet to TX ring for immediate transmission */
> > > + xdpf = xdp_convert_buff_to_frame(xdp);
> > > + if (unlikely(!xdpf))
> >
> > This is the second unlikely() macro which is added in this patchset.
> > The rule with likely/unlikely() is that it should only be added if it
> > likely makes a difference in benchmarking. Quite often the compiler
> > is able to predict that valid pointers are more likely than NULL
> > pointers so often these types of annotations don't make any difference
> > at all to the compiled code. But it depends on the compiler and the -O2
> > options.
> >
>
> Do correct me if I am wrong, but from my understanding, XDP feature depends
> alot of performance and benchmarking and having unlikely does make a
> difference. Atleast in all the other drivers I see this being used for XDP.
>
Which compiler are you on when you say that "having unlikely does make a
difference"?
I'm on gcc version 14.2.0 (Debian 14.2.0-16) and it doesn't make a
difference to the compiled code. This matches what one would expect from
a compiler. Valid pointers are fast path and NULL pointers are slow path.
Adding an unlikely() is a micro optimization. There are so many other
things you can do to speed up the code. I wouldn't start with that.
regards,
dan
