| Message ID | 20200614071128.4ezfcyhjesot4vvr@SvensMacBookAir.sven.lan |
|---|---|
| State | Rejected |
| Delegated to: | David Miller |
| Headers | show |
| Series | [1/1] mvneta: fix prefetch location | expand |
On Sun, 14 Jun 2020 09:11:28 +0200 Sven Auhagen <sven.auhagen@voleatech.de> wrote: > The packet header prefetch is at an offset > now. Correct the prefetch address. > > Signed-off-by: Sven Auhagen <sven.auhagen@voleatech.de> > --- > drivers/net/ethernet/marvell/mvneta.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c > index 51889770958d..344fc5f649b4 100644 > --- a/drivers/net/ethernet/marvell/mvneta.c > +++ b/drivers/net/ethernet/marvell/mvneta.c > @@ -2229,7 +2229,7 @@ mvneta_swbm_rx_frame(struct mvneta_port *pp, > len, dma_dir); > > /* Prefetch header */ > - prefetch(data); > + prefetch(data + pp->rx_offset_correction + MVNETA_MH_SIZE); The comment does say "header", so I guess your change is correct if we are talking about the packet header. Currently this prefetch will help XDP-redirect, as it store xdp_frame in this area. > xdp->data_hard_start = data; > xdp->data = data + pp->rx_offset_correction + MVNETA_MH_SIZE; If you really need to prefetch the packet headers, it would be the same as calling prefetch(xdp->data), right? Have you benchmarked that this prefetch is a benefit?
On Sun, Jun 14, 2020 at 10:23:43AM +0200, Jesper Dangaard Brouer wrote: > On Sun, 14 Jun 2020 09:11:28 +0200 > Sven Auhagen <sven.auhagen@voleatech.de> wrote: > > > The packet header prefetch is at an offset > > now. Correct the prefetch address. > > > > Signed-off-by: Sven Auhagen <sven.auhagen@voleatech.de> > > --- > > drivers/net/ethernet/marvell/mvneta.c | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c > > index 51889770958d..344fc5f649b4 100644 > > --- a/drivers/net/ethernet/marvell/mvneta.c > > +++ b/drivers/net/ethernet/marvell/mvneta.c > > @@ -2229,7 +2229,7 @@ mvneta_swbm_rx_frame(struct mvneta_port *pp, > > len, dma_dir); > > > > /* Prefetch header */ > > - prefetch(data); > > + prefetch(data + pp->rx_offset_correction + MVNETA_MH_SIZE); > > The comment does say "header", so I guess your change is correct if we > are talking about the packet header. > > Currently this prefetch will help XDP-redirect, as it store xdp_frame > in this area. > > > xdp->data_hard_start = data; > > xdp->data = data + pp->rx_offset_correction + MVNETA_MH_SIZE; > > If you really need to prefetch the packet headers, it would be the same > as calling prefetch(xdp->data), right? Yes, that is correct. > > Have you benchmarked that this prefetch is a benefit? I just ran some forwarding tests and the difference is not measureable in normal forwarding. You are correct XDP Forwarding seems to be 1-2% slower. I guess you can disregard the patch. > > -- > Best regards, > Jesper Dangaard Brouer > MSc.CS, Principal Kernel Engineer at Red Hat > LinkedIn: https://eur03.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.linkedin.com%2Fin%2Fbrouer&data=02%7C01%7Csven.auhagen%40voleatech.de%7C4f3ae28c82514407dfd508d8103c4d1b%7Cb82a99f679814a7295344d35298f847b%7C0%7C0%7C637277198467217085&sdata=qIJ3tkU2mHFqZpArOGnQWSPEVtTBMXK8MpekSGcKavw%3D&reserved=0 >
diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c index 51889770958d..344fc5f649b4 100644 --- a/drivers/net/ethernet/marvell/mvneta.c +++ b/drivers/net/ethernet/marvell/mvneta.c @@ -2229,7 +2229,7 @@ mvneta_swbm_rx_frame(struct mvneta_port *pp, len, dma_dir); /* Prefetch header */ - prefetch(data); + prefetch(data + pp->rx_offset_correction + MVNETA_MH_SIZE); xdp->data_hard_start = data; xdp->data = data + pp->rx_offset_correction + MVNETA_MH_SIZE;
The packet header prefetch is at an offset now. Correct the prefetch address. Signed-off-by: Sven Auhagen <sven.auhagen@voleatech.de> --- drivers/net/ethernet/marvell/mvneta.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)