Eric Dumazet implemented Big TCP that allowed bigger TSO/GRO packet sizes
for IPv6 traffic. See patch series:
'commit 89527be8d8d6 ("net: add IFLA_TSO_{MAX_SIZE|SEGS} attributes")'
This reduces the number of packets traversing the networking stack and
should usually improves performance. However, it also inserts a
temporary Hop-by-hop IPv6 extension header.
Using the HBH header removal method in the previous path, the extra header
be removed in bnxt drivers to allow it to send big TCP packets (bigger
TSO packets) as well.
Tested:
Compiled locally
To further test functional correctness, update the GSO/GRO limit on the
physical NIC:
ip link set eth0 gso_max_size 181000
ip link set eth0 gro_max_size 181000
Note that if there are bonding or ipvan devices on top of the physical
NIC, their GSO sizes need to be updated as well.
Then, IPv6/TCP packets with sizes larger than 64k can be observed.
Big TCP functionality is tested by Michael, feature checks not yet.
Tested by Michael:
I've confirmed with our hardware team that this is supported by our
chips, and I've tested it up to gso_max_size of 524280. Thanks.
Tested-by: Michael Chan <[email protected]>
Reviewed-by: Michael Chan <[email protected]>
Signed-off-by: Coco Li <[email protected]>
---
drivers/net/ethernet/broadcom/bnxt/bnxt.c | 26 ++++++++++++++++++++++-
1 file changed, 25 insertions(+), 1 deletion(-)
diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.c b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
index 0fe164b42c5d..c2713cb5debd 100644
--- a/drivers/net/ethernet/broadcom/bnxt/bnxt.c
+++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
@@ -389,6 +389,9 @@ static netdev_tx_t bnxt_start_xmit(struct sk_buff *skb, struct net_device *dev)
return NETDEV_TX_BUSY;
}
+ if (unlikely(ipv6_hopopt_jumbo_remove(skb)))
+ goto tx_free;
+
length = skb->len;
len = skb_headlen(skb);
last_frag = skb_shinfo(skb)->nr_frags;
@@ -11342,9 +11345,28 @@ static bool bnxt_exthdr_check(struct bnxt *bp, struct sk_buff *skb, int nw_off,
if (hdrlen > 64)
return false;
+
+ /* The ext header may be a hop-by-hop header inserted for
+ * big TCP purposes. This will be removed before sending
+ * from NIC, so do not count it.
+ */
+ if (*nexthdr == NEXTHDR_HOP) {
+ if (likely(skb->len <= GRO_LEGACY_MAX_SIZE))
+ goto increment_hdr;
+
+ struct hop_jumbo_hdr *jhdr = (struct hop_jumbo_hdr *)(nexthdr + hdrlen);
+
+ if (jhdr->tlv_type != IPV6_TLV_JUMBO || jhdr->hdrlen != 0 ||
+ (jhdr->nexthdr != IPPROTO_TCP && jhdr->nexthdr != IPPROTO_UDP))
+ goto increment_hdr;
+
+ goto next_hdr;
+ }
+increment_hdr:
+ hdr_count++;
+next_hdr:
nexthdr = &hp->nexthdr;
start += hdrlen;
- hdr_count++;
}
if (nextp) {
/* Caller will check inner protocol */
@@ -13657,6 +13679,8 @@ static int bnxt_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
dev->features &= ~NETIF_F_LRO;
dev->priv_flags |= IFF_UNICAST_FLT;
+ netif_set_tso_max_size(dev, GSO_MAX_SIZE);
+
#ifdef CONFIG_BNXT_SRIOV
init_waitqueue_head(&bp->sriov_cfg_wait);
#endif
--
2.39.0.rc0.267.gcb52ba06e7-goog
On Fri, Dec 2, 2022 at 11:12 PM Coco Li <[email protected]> wrote:
>
> Eric Dumazet implemented Big TCP that allowed bigger TSO/GRO packet sizes
> for IPv6 traffic. See patch series:
> 'commit 89527be8d8d6 ("net: add IFLA_TSO_{MAX_SIZE|SEGS} attributes")'
>
> This reduces the number of packets traversing the networking stack and
> should usually improves performance. However, it also inserts a
> temporary Hop-by-hop IPv6 extension header.
>
> Using the HBH header removal method in the previous path, the extra header
> be removed in bnxt drivers to allow it to send big TCP packets (bigger
> TSO packets) as well.
>
> Tested:
> Compiled locally
>
> To further test functional correctness, update the GSO/GRO limit on the
> physical NIC:
>
> ip link set eth0 gso_max_size 181000
> ip link set eth0 gro_max_size 181000
>
> Note that if there are bonding or ipvan devices on top of the physical
> NIC, their GSO sizes need to be updated as well.
>
> Then, IPv6/TCP packets with sizes larger than 64k can be observed.
>
> Big TCP functionality is tested by Michael, feature checks not yet.
>
> Tested by Michael:
> I've confirmed with our hardware team that this is supported by our
> chips, and I've tested it up to gso_max_size of 524280. Thanks.
>
> Tested-by: Michael Chan <[email protected]>
> Reviewed-by: Michael Chan <[email protected]>
> Signed-off-by: Coco Li <[email protected]>
> ---
> drivers/net/ethernet/broadcom/bnxt/bnxt.c | 26 ++++++++++++++++++++++-
> 1 file changed, 25 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.c b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
> index 0fe164b42c5d..c2713cb5debd 100644
> --- a/drivers/net/ethernet/broadcom/bnxt/bnxt.c
> +++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
> @@ -389,6 +389,9 @@ static netdev_tx_t bnxt_start_xmit(struct sk_buff *skb, struct net_device *dev)
> return NETDEV_TX_BUSY;
> }
>
> + if (unlikely(ipv6_hopopt_jumbo_remove(skb)))
> + goto tx_free;
> +
> length = skb->len;
> len = skb_headlen(skb);
> last_frag = skb_shinfo(skb)->nr_frags;
> @@ -11342,9 +11345,28 @@ static bool bnxt_exthdr_check(struct bnxt *bp, struct sk_buff *skb, int nw_off,
>
> if (hdrlen > 64)
> return false;
> +
> + /* The ext header may be a hop-by-hop header inserted for
> + * big TCP purposes. This will be removed before sending
> + * from NIC, so do not count it.
> + */
> + if (*nexthdr == NEXTHDR_HOP) {
> + if (likely(skb->len <= GRO_LEGACY_MAX_SIZE))
> + goto increment_hdr;
> +
> + struct hop_jumbo_hdr *jhdr = (struct hop_jumbo_hdr *)(nexthdr + hdrlen);
We discourage adding a variable declaration in the middle of code.
> +
> + if (jhdr->tlv_type != IPV6_TLV_JUMBO || jhdr->hdrlen != 0 ||
> + (jhdr->nexthdr != IPPROTO_TCP && jhdr->nexthdr != IPPROTO_UDP))
Why testing IPPROTO_UDP ? I do not think we support BIG UDP yet.
> + goto increment_hdr;
> +
> + goto next_hdr;
> + }
> +increment_hdr:
> + hdr_count++;
> +next_hdr:
> nexthdr = &hp->nexthdr;
> start += hdrlen;
> - hdr_count++;
> }
> if (nextp) {
> /* Caller will check inner protocol */
> @@ -13657,6 +13679,8 @@ static int bnxt_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
> dev->features &= ~NETIF_F_LRO;
> dev->priv_flags |= IFF_UNICAST_FLT;
>
> + netif_set_tso_max_size(dev, GSO_MAX_SIZE);
> +
> #ifdef CONFIG_BNXT_SRIOV
> init_waitqueue_head(&bp->sriov_cfg_wait);
> #endif
> --
> 2.39.0.rc0.267.gcb52ba06e7-goog
>
On Sun, Dec 4, 2022 at 9:14 PM Eric Dumazet <[email protected]> wrote:
>
> On Fri, Dec 2, 2022 at 11:12 PM Coco Li <[email protected]> wrote:
> >
> > Eric Dumazet implemented Big TCP that allowed bigger TSO/GRO packet sizes
> > for IPv6 traffic. See patch series:
> > 'commit 89527be8d8d6 ("net: add IFLA_TSO_{MAX_SIZE|SEGS} attributes")'
> >
> > This reduces the number of packets traversing the networking stack and
> > should usually improves performance. However, it also inserts a
> > temporary Hop-by-hop IPv6 extension header.
> >
> > Using the HBH header removal method in the previous path, the extra header
> > be removed in bnxt drivers to allow it to send big TCP packets (bigger
> > TSO packets) as well.
> >
> > Tested:
> > Compiled locally
> >
> > To further test functional correctness, update the GSO/GRO limit on the
> > physical NIC:
> >
> > ip link set eth0 gso_max_size 181000
> > ip link set eth0 gro_max_size 181000
> >
> > Note that if there are bonding or ipvan devices on top of the physical
> > NIC, their GSO sizes need to be updated as well.
> >
> > Then, IPv6/TCP packets with sizes larger than 64k can be observed.
> >
> > Big TCP functionality is tested by Michael, feature checks not yet.
> >
> > Tested by Michael:
> > I've confirmed with our hardware team that this is supported by our
> > chips, and I've tested it up to gso_max_size of 524280. Thanks.
> >
> > Tested-by: Michael Chan <[email protected]>
> > Reviewed-by: Michael Chan <[email protected]>
> > Signed-off-by: Coco Li <[email protected]>
> > ---
> > drivers/net/ethernet/broadcom/bnxt/bnxt.c | 26 ++++++++++++++++++++++-
> > 1 file changed, 25 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.c b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
> > index 0fe164b42c5d..c2713cb5debd 100644
> > --- a/drivers/net/ethernet/broadcom/bnxt/bnxt.c
> > +++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
> > @@ -389,6 +389,9 @@ static netdev_tx_t bnxt_start_xmit(struct sk_buff *skb, struct net_device *dev)
> > return NETDEV_TX_BUSY;
> > }
> >
> > + if (unlikely(ipv6_hopopt_jumbo_remove(skb)))
> > + goto tx_free;
> > +
> > length = skb->len;
> > len = skb_headlen(skb);
> > last_frag = skb_shinfo(skb)->nr_frags;
> > @@ -11342,9 +11345,28 @@ static bool bnxt_exthdr_check(struct bnxt *bp, struct sk_buff *skb, int nw_off,
> >
> > if (hdrlen > 64)
> > return false;
> > +
> > + /* The ext header may be a hop-by-hop header inserted for
> > + * big TCP purposes. This will be removed before sending
> > + * from NIC, so do not count it.
> > + */
> > + if (*nexthdr == NEXTHDR_HOP) {
> > + if (likely(skb->len <= GRO_LEGACY_MAX_SIZE))
> > + goto increment_hdr;
> > +
> > + struct hop_jumbo_hdr *jhdr = (struct hop_jumbo_hdr *)(nexthdr + hdrlen);
>
> We discourage adding a variable declaration in the middle of code.
>
This doesn't work. Initially nexthdr points to the nexthdr in the
ipv6 header. It should be coded like this:
struct hop_jumbo_hdr *jhdr = (struct hop_jumbo_hdr *)hp;
Thanks.