On Wed, Aug 13, 2025 at 12:40 AM Maciej Fijalkowski
<maciej.fijalkowski@xxxxxxxxx> wrote:
>
> On Mon, Aug 11, 2025 at 09:12:35PM +0800, Jason Xing wrote:
> > From: Jason Xing <kernelxing@xxxxxxxxxxx>
> >
> > This patch is to prepare for later batch xmit in generic path. Add a new
> > socket option to provide an alternative to achieve a higher overall
> > throughput.
> >
> > skb_batch will be used to store newly allocated skb at one time in the
> > xmit path.
>
> I don't think we need yet another setsockopt. You previously added a knob
> for manipulating max tx budget on generic xmit and that should be enough.
> I think that we should strive for making the batching approach a default
> path in xsk generic xmit.
You're right, it‘s the right direction that we should take. But I
considered this as well before cooking the series and then gave up, my
experiments show that in some real cases (not xdpsock) the batch
process might increase latency. It's a side effect. At that time I
thought many years ago the invention of GRO didn't become the default.
Thanks,
Jason
>
> >
> > Signed-off-by: Jason Xing <kernelxing@xxxxxxxxxxx>
> > ---
> > Documentation/networking/af_xdp.rst | 9 ++++++++
> > include/net/xdp_sock.h | 2 ++
> > include/uapi/linux/if_xdp.h | 1 +
> > net/xdp/xsk.c | 32 +++++++++++++++++++++++++++++
> > tools/include/uapi/linux/if_xdp.h | 1 +
> > 5 files changed, 45 insertions(+)
> >
> > diff --git a/Documentation/networking/af_xdp.rst b/Documentation/networking/af_xdp.rst
> > index 50d92084a49c..1194bdfaf61e 100644
> > --- a/Documentation/networking/af_xdp.rst
> > +++ b/Documentation/networking/af_xdp.rst
> > @@ -447,6 +447,15 @@ mode to allow application to tune the per-socket maximum iteration for
> > better throughput and less frequency of send syscall.
> > Allowed range is [32, xs->tx->nentries].
> >
> > +XDP_GENERIC_XMIT_BATCH
> > +----------------------
> > +
> > +It provides an option that allows application to use batch xmit in the copy
> > +mode. Batch process minimizes the number of grabbing/releasing queue lock
> > +without redundant actions compared to before to gain the overall performance
> > +improvement whereas it might increase the latency of per packet. The maximum
> > +value shouldn't be larger than xs->max_tx_budget.
> > +
> > XDP_STATISTICS getsockopt
> > -------------------------
> >
> > diff --git a/include/net/xdp_sock.h b/include/net/xdp_sock.h
> > index ce587a225661..b5a3e37da8db 100644
> > --- a/include/net/xdp_sock.h
> > +++ b/include/net/xdp_sock.h
> > @@ -61,6 +61,7 @@ struct xdp_sock {
> > XSK_BOUND,
> > XSK_UNBOUND,
> > } state;
> > + struct sk_buff **skb_batch;
> >
> > struct xsk_queue *tx ____cacheline_aligned_in_smp;
> > struct list_head tx_list;
> > @@ -70,6 +71,7 @@ struct xdp_sock {
> > * preventing other XSKs from being starved.
> > */
> > u32 tx_budget_spent;
> > + u32 generic_xmit_batch;
> >
> > /* Statistics */
> > u64 rx_dropped;
> > diff --git a/include/uapi/linux/if_xdp.h b/include/uapi/linux/if_xdp.h
> > index 23a062781468..44cb72cd328e 100644
> > --- a/include/uapi/linux/if_xdp.h
> > +++ b/include/uapi/linux/if_xdp.h
> > @@ -80,6 +80,7 @@ struct xdp_mmap_offsets {
> > #define XDP_STATISTICS 7
> > #define XDP_OPTIONS 8
> > #define XDP_MAX_TX_SKB_BUDGET 9
> > +#define XDP_GENERIC_XMIT_BATCH 10
> >
> > struct xdp_umem_reg {
> > __u64 addr; /* Start of packet data area */
> > diff --git a/net/xdp/xsk.c b/net/xdp/xsk.c
> > index 9c3acecc14b1..7a149f4ac273 100644
> > --- a/net/xdp/xsk.c
> > +++ b/net/xdp/xsk.c
> > @@ -1122,6 +1122,7 @@ static int xsk_release(struct socket *sock)
> > xskq_destroy(xs->tx);
> > xskq_destroy(xs->fq_tmp);
> > xskq_destroy(xs->cq_tmp);
> > + kfree(xs->skb_batch);
> >
> > sock_orphan(sk);
> > sock->sk = NULL;
> > @@ -1456,6 +1457,37 @@ static int xsk_setsockopt(struct socket *sock, int level, int optname,
> > WRITE_ONCE(xs->max_tx_budget, budget);
> > return 0;
> > }
> > + case XDP_GENERIC_XMIT_BATCH:
> > + {
> > + unsigned int batch, batch_alloc_len;
> > + struct sk_buff **new;
> > +
> > + if (optlen != sizeof(batch))
> > + return -EINVAL;
> > + if (copy_from_sockptr(&batch, optval, sizeof(batch)))
> > + return -EFAULT;
> > + if (batch > xs->max_tx_budget)
> > + return -EACCES;
> > +
> > + mutex_lock(&xs->mutex);
> > + if (!batch) {
> > + kfree(xs->skb_batch);
> > + xs->generic_xmit_batch = 0;
> > + goto out;
> > + }
> > + batch_alloc_len = sizeof(struct sk_buff *) * batch;
> > + new = kmalloc(batch_alloc_len, GFP_KERNEL);
> > + if (!new)
> > + return -ENOMEM;
> > + if (xs->skb_batch)
> > + kfree(xs->skb_batch);
> > +
> > + xs->skb_batch = new;
> > + xs->generic_xmit_batch = batch;
> > +out:
> > + mutex_unlock(&xs->mutex);
> > + return 0;
> > + }
> > default:
> > break;
> > }
> > diff --git a/tools/include/uapi/linux/if_xdp.h b/tools/include/uapi/linux/if_xdp.h
> > index 23a062781468..44cb72cd328e 100644
> > --- a/tools/include/uapi/linux/if_xdp.h
> > +++ b/tools/include/uapi/linux/if_xdp.h
> > @@ -80,6 +80,7 @@ struct xdp_mmap_offsets {
> > #define XDP_STATISTICS 7
> > #define XDP_OPTIONS 8
> > #define XDP_MAX_TX_SKB_BUDGET 9
> > +#define XDP_GENERIC_XMIT_BATCH 10
> >
> > struct xdp_umem_reg {
> > __u64 addr; /* Start of packet data area */
> > --
> > 2.41.3
> >
