From: Jan Klemkow <jan@openbsd.org>
Subject: Re: vio: allow longer tx chains with tso
To: Stefan Fritsch <sf@openbsd.org>
Cc: tech@openbsd.org
Date: Wed, 18 Sep 2024 23:38:15 +0200

On Tue, Sep 17, 2024 at 11:12:07AM +0200, Stefan Fritsch wrote:
> most of the rest of my vio multi-queue work will wait until after the 
> release. But there are a few small parts that are not really related to 
> multi-queue that I would like to go into 7.6.
> 
> When TCP segmentation offload is supported, we may get larger packets with 
> more dma segments. Allocate more segments in the busdma_map in this case, 
> so that we need to defragment less often.
> 
> ok?

Looks good and works for me.

ok jan@

> diff --git a/sys/dev/pv/if_vio.c b/sys/dev/pv/if_vio.c
> index fc7f330c0ee..25f0316a36b 100644
> --- a/sys/dev/pv/if_vio.c
> +++ b/sys/dev/pv/if_vio.c
> @@ -269,7 +269,6 @@ struct vio_softc {
>  #define VIO_HAVE_MRG_RXBUF(sc)					\
>  	((sc)->sc_hdr_size == sizeof(struct virtio_net_hdr))
>  
> -#define VIRTIO_NET_TX_MAXNSEGS		16 /* for larger chains, defrag */
>  #define VIRTIO_NET_CTRL_MAC_MC_ENTRIES	64 /* for more entries, use ALLMULTI */
>  #define VIRTIO_NET_CTRL_MAC_UC_ENTRIES	 1 /* one entry for own unicast addr */
>  #define VIRTIO_NET_CTRL_TIMEOUT		(5*1000*1000*1000ULL) /* 5 seconds */
> @@ -321,7 +320,7 @@ int	vio_ctrl_start(struct vio_softc *, uint8_t, uint8_t, int, int *);
>  int	vio_ctrl_submit(struct vio_softc *, int);
>  void	vio_ctrl_finish(struct vio_softc *);
>  void	vio_ctrl_wakeup(struct vio_softc *, enum vio_ctrl_state);
> -int	vio_alloc_mem(struct vio_softc *);
> +int	vio_alloc_mem(struct vio_softc *, int);
>  int	vio_alloc_dmamem(struct vio_softc *);
>  void	vio_free_dmamem(struct vio_softc *);
>  
> @@ -416,7 +415,7 @@ vio_free_dmamem(struct vio_softc *sc)
>   *   viq_txmbufs[slot]:		mbuf pointer array for sent frames
>   */
>  int
> -vio_alloc_mem(struct vio_softc *sc)
> +vio_alloc_mem(struct vio_softc *sc, int tx_max_segments)
>  {
>  	struct virtio_softc	*vsc = sc->sc_virtio;
>  	struct ifnet		*ifp = &sc->sc_ac.ac_if;
> @@ -503,7 +502,7 @@ vio_alloc_mem(struct vio_softc *sc)
>  
>  		for (i = 0; i < txqsize; i++) {
>  			r = bus_dmamap_create(vsc->sc_dmat, txsize,
> -			    VIRTIO_NET_TX_MAXNSEGS, txsize, 0,
> +			    tx_max_segments, txsize, 0,
>  			    BUS_DMA_NOWAIT|BUS_DMA_ALLOCNOW,
>  			    &vioq->viq_txdmamaps[i]);
>  			if (r != 0)
> @@ -586,7 +585,7 @@ vio_attach(struct device *parent, struct device *self, void *aux)
>  {
>  	struct vio_softc *sc = (struct vio_softc *)self;
>  	struct virtio_softc *vsc = (struct virtio_softc *)parent;
> -	int i;
> +	int i, tx_max_segments;
>  	struct ifnet *ifp = &sc->sc_ac.ac_if;
>  
>  	if (vsc->sc_child != NULL) {
> @@ -650,6 +649,17 @@ vio_attach(struct device *parent, struct device *self, void *aux)
>  	else
>  		ifp->if_hardmtu = MCLBYTES - sc->sc_hdr_size - ETHER_HDR_LEN;
>  
> +	/* defrag for longer mbuf chains */
> +	tx_max_segments = 16;
> +	if (virtio_has_feature(vsc, VIRTIO_NET_F_HOST_TSO4) ||
> +	    virtio_has_feature(vsc, VIRTIO_NET_F_HOST_TSO6)) {
> +		/*
> +		 * With TSO, we may get 64K packets and want to be able to
> +		 * send longer chains without defragmenting
> +		 */
> +		tx_max_segments = 32;
> +	}
> +
>  	for (i = 0; i < sc->sc_nqueues; i++) {
>  		int vqidx = 2 * i;
>  		struct vio_queue *vioq = &sc->sc_q[i];
> @@ -664,7 +674,7 @@ vio_attach(struct device *parent, struct device *self, void *aux)
>  		vqidx++;
>  		vioq->viq_txvq = &vsc->sc_vqs[vqidx];
>  		if (virtio_alloc_vq(vsc, vioq->viq_txvq, vqidx,
> -		    VIRTIO_NET_TX_MAXNSEGS + 1, "tx") != 0) {
> +		    tx_max_segments + 1, "tx") != 0) {
>  			goto err;
>  		}
>  		vioq->viq_txvq->vq_done = vio_tx_intr;
> @@ -684,7 +694,7 @@ vio_attach(struct device *parent, struct device *self, void *aux)
>  		virtio_start_vq_intr(vsc, sc->sc_ctl_vq);
>  	}
>  
> -	if (vio_alloc_mem(sc) < 0)
> +	if (vio_alloc_mem(sc, tx_max_segments) < 0)
>  		goto err;
>  
>  	strlcpy(ifp->if_xname, self->dv_xname, IFNAMSIZ);
>