[PATCH v2 0/2] gro: optimise redundant parsing of packets

[PATCH v2 0/2] gro: optimise redundant parsing of packets

[Richard Gobert]

Currently the IPv6 extension headers are parsed twice: first in
ipv6_gro_receive, and then again in ipv6_gro_complete.

By using the new ->transport_proto field, and also storing the size of the
network header, we can avoid parsing extension headers a second time in
ipv6_gro_complete.

The first commit frees up space in the GRO CB. The second commit reduces the
redundant parsing during the complete phase, using the freed CB space.

In addition, the second commit contains a fix for a potential problem in BIG
TCP, which is detailed in the commit message itself.

Performance tests for TCP stream over IPv6 with extension headers demonstrate rx
improvement of ~0.7%.

For the benchmarks, I used 100Gbit NIC mlx5 single-core (power management off),
turboboost off.

Typical IPv6 traffic (zero extension headers):

    for i in {1..5}; do netperf -t TCP_STREAM -H 2001:db8:2:2::2 -l 90 | tail -1; done
    # before
    131072  16384  16384    90.00    16391.20
    131072  16384  16384    90.00    16403.50
    131072  16384  16384    90.00    16403.30
    131072  16384  16384    90.00    16397.84
    131072  16384  16384    90.00    16398.00

    # after
    131072  16384  16384    90.00    16399.85
    131072  16384  16384    90.00    16392.37
    131072  16384  16384    90.00    16403.06
    131072  16384  16384    90.00    16406.97
    131072  16384  16384    90.00    16406.09

IPv6 over IPv6 traffic:

    for i in {1..5}; do netperf -t TCP_STREAM -H 4001:db8:2:2::2 -l 90 | tail -1; done
    # before
    131072  16384  16384    90.00    14791.61
    131072  16384  16384    90.00    14791.66
    131072  16384  16384    90.00    14783.47
    131072  16384  16384    90.00    14810.17
    131072  16384  16384    90.00    14806.15

    # after
    131072  16384  16384    90.00    14793.49
    131072  16384  16384    90.00    14816.10
    131072  16384  16384    90.00    14818.41
    131072  16384  16384    90.00    14780.35
    131072  16384  16384    90.00    14800.48

IPv6 traffic with varying extension headers:

    for i in {1..5}; do netperf -t TCP_STREAM -H 2001:db8:2:2::2 -l 90 | tail -1; done
    # before
    131072  16384  16384    90.00    14812.37
    131072  16384  16384    90.00    14813.04
    131072  16384  16384    90.00    14802.54
    131072  16384  16384    90.00    14804.06
    131072  16384  16384    90.00    14819.08

    # after
    131072  16384  16384    90.00    14927.11
    131072  16384  16384    90.00    14910.45
    131072  16384  16384    90.00    14917.36
    131072  16384  16384    90.00    14916.53
    131072  16384  16384    90.00    14928.88


Richard Gobert (2):
  gro: decrease size of CB
  gro: optimise redundant parsing of packets

 include/net/gro.h      | 33 ++++++++++++++++++++++++---------
 net/core/gro.c         | 18 +++++++++++-------
 net/ethernet/eth.c     | 14 +++++++++++---
 net/ipv6/ip6_offload.c | 20 +++++++++++++++-----
 4 files changed, 61 insertions(+), 24 deletions(-)

-- 
2.36.1

[PATCH v2 1/2] gro: decrease size of CB

[Richard Gobert]

The GRO control block (NAPI_GRO_CB) is currently at its maximum size.  This
commit reduces its size by putting two groups of fields that are used only at
different times into a union.

Specifically, the fields frag0 and frag0_len are the fields that make up the
frag0 optimisation mechanism, which is used during the initial parsing of the
SKB.

The fields last and age are used after the initial parsing, while the SKB is
stored in the GRO list, waiting for other packets to arrive.

There was one location in dev_gro_receive that modified the frag0 fields after
setting last and age. I changed this accordingly without altering the code
behaviour.

Signed-off-by: Richard Gobert <richardbgobert@gmail.com>
---
 include/net/gro.h | 26 ++++++++++++++++----------
 net/core/gro.c    | 18 +++++++++++-------
 2 files changed, 27 insertions(+), 17 deletions(-)

diff --git a/include/net/gro.h b/include/net/gro.h
index a4fab706240d..7b47dd6ce94f 100644
--- a/include/net/gro.h
+++ b/include/net/gro.h
@@ -11,11 +11,23 @@
 #include <net/udp.h>
 
 struct napi_gro_cb {
-	/* Virtual address of skb_shinfo(skb)->frags[0].page + offset. */
-	void	*frag0;
+	union {
+		struct {
+			/* Virtual address of skb_shinfo(skb)->frags[0].page + offset. */
+			void	*frag0;
 
-	/* Length of frag0. */
-	unsigned int frag0_len;
+			/* Length of frag0. */
+			unsigned int frag0_len;
+		};
+
+		struct {
+			/* used in skb_gro_receive() slow path */
+			struct sk_buff *last;
+
+			/* jiffies when first packet was created/queued */
+			unsigned long age;
+		};
+	};
 
 	/* This indicates where we are processing relative to skb->data. */
 	int	data_offset;
@@ -32,9 +44,6 @@ struct napi_gro_cb {
 	/* Used in ipv6_gro_receive() and foo-over-udp */
 	u16	proto;
 
-	/* jiffies when first packet was created/queued */
-	unsigned long age;
-
 /* Used in napi_gro_cb::free */
 #define NAPI_GRO_FREE             1
 #define NAPI_GRO_FREE_STOLEN_HEAD 2
@@ -77,9 +86,6 @@ struct napi_gro_cb {
 
 	/* used to support CHECKSUM_COMPLETE for tunneling protocols */
 	__wsum	csum;
-
-	/* used in skb_gro_receive() slow path */
-	struct sk_buff *last;
 };
 
 #define NAPI_GRO_CB(skb) ((struct napi_gro_cb *)(skb)->cb)
diff --git a/net/core/gro.c b/net/core/gro.c
index a606705a0859..b1fdabd414a5 100644
--- a/net/core/gro.c
+++ b/net/core/gro.c
@@ -460,6 +460,14 @@ static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
 	}
 }
 
+static inline void gro_try_pull_from_frag0(struct sk_buff *skb)
+{
+	int grow = skb_gro_offset(skb) - skb_headlen(skb);
+
+	if (grow > 0)
+		gro_pull_from_frag0(skb, grow);
+}
+
 static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
 {
 	struct sk_buff *oldest;
@@ -489,7 +497,6 @@ static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff
 	struct sk_buff *pp = NULL;
 	enum gro_result ret;
 	int same_flow;
-	int grow;
 
 	if (netif_elide_gro(skb->dev))
 		goto normal;
@@ -564,17 +571,13 @@ static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff
 	else
 		gro_list->count++;
 
+	gro_try_pull_from_frag0(skb);
 	NAPI_GRO_CB(skb)->age = jiffies;
 	NAPI_GRO_CB(skb)->last = skb;
 	if (!skb_is_gso(skb))
 		skb_shinfo(skb)->gso_size = skb_gro_len(skb);
 	list_add(&skb->list, &gro_list->list);
 	ret = GRO_HELD;
-
-pull:
-	grow = skb_gro_offset(skb) - skb_headlen(skb);
-	if (grow > 0)
-		gro_pull_from_frag0(skb, grow);
 ok:
 	if (gro_list->count) {
 		if (!test_bit(bucket, &napi->gro_bitmask))
@@ -587,7 +590,8 @@ static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff
 
 normal:
 	ret = GRO_NORMAL;
-	goto pull;
+	gro_try_pull_from_frag0(skb);
+	goto ok;
 }
 
 struct packet_offload *gro_find_receive_by_type(__be16 type)
-- 
2.36.1

[PATCH v2 2/2] gro: optimise redundant parsing of packets

[Richard Gobert]

Currently the IPv6 extension headers are parsed twice: first in
ipv6_gro_receive, and then again in ipv6_gro_complete.

By using the new ->transport_proto field, and also storing the size of the
network header, we can avoid parsing extension headers a second time in
ipv6_gro_complete (which saves multiple memory dereferences and conditional
checks inside ipv6_exthdrs_len for a varying amount of extension headers in IPv6
packets).

The implementation had to handle both inner and outer layers in case of
encapsulation (as they can't use the same field).

Performance tests for TCP stream over IPv6 with a varying amount of extension
headers demonstrate throughput improvement of ~0.7%.

In addition, I fixed a potential existing problem:
 - The call to skb_set_inner_network_header at the beginning of
   ipv6_gro_complete calculates inner_network_header based on skb->data by
   calling skb_set_inner_network_header, and setting it to point to the beginning
   of the ip header.
 - If a packet is going to be handled by BIG TCP, the following code block is
   going to shift the packet header, and skb->data is going to be changed as
   well. 

When the two flows are combined, inner_network_header will point to the wrong
place.

The fix is to place the whole encapsulation branch after the BIG TCP code block.
This way, inner_network_header is calculated with a correct value of skb->data.
Also, by arranging the code that way, the optimisation does not add an additional
branch.

Signed-off-by: Richard Gobert <richardbgobert@gmail.com>
---
 include/net/gro.h      |  9 +++++++++
 net/ethernet/eth.c     | 14 +++++++++++---
 net/ipv6/ip6_offload.c | 20 +++++++++++++++-----
 3 files changed, 35 insertions(+), 8 deletions(-)

diff --git a/include/net/gro.h b/include/net/gro.h
index 7b47dd6ce94f..35f60ea99f6c 100644
--- a/include/net/gro.h
+++ b/include/net/gro.h
@@ -86,6 +86,15 @@ struct napi_gro_cb {
 
 	/* used to support CHECKSUM_COMPLETE for tunneling protocols */
 	__wsum	csum;
+
+	/* Used in ipv6_gro_receive() */
+	u16	network_len;
+
+	/* Used in eth_gro_receive() */
+	__be16	network_proto;
+
+	/* Used in ipv6_gro_receive() */
+	u8	transport_proto;
 };
 
 #define NAPI_GRO_CB(skb) ((struct napi_gro_cb *)(skb)->cb)
diff --git a/net/ethernet/eth.c b/net/ethernet/eth.c
index 2edc8b796a4e..c2b77d9401e4 100644
--- a/net/ethernet/eth.c
+++ b/net/ethernet/eth.c
@@ -439,6 +439,9 @@ struct sk_buff *eth_gro_receive(struct list_head *head, struct sk_buff *skb)
 		goto out;
 	}
 
+	if (!NAPI_GRO_CB(skb)->encap_mark)
+		NAPI_GRO_CB(skb)->network_proto = type;
+
 	skb_gro_pull(skb, sizeof(*eh));
 	skb_gro_postpull_rcsum(skb, eh, sizeof(*eh));
 
@@ -455,13 +458,18 @@ EXPORT_SYMBOL(eth_gro_receive);
 
 int eth_gro_complete(struct sk_buff *skb, int nhoff)
 {
-	struct ethhdr *eh = (struct ethhdr *)(skb->data + nhoff);
-	__be16 type = eh->h_proto;
 	struct packet_offload *ptype;
+	struct ethhdr *eh;
 	int err = -ENOSYS;
+	__be16 type;
 
-	if (skb->encapsulation)
+	if (skb->encapsulation) {
+		eh = (struct ethhdr *)(skb->data + nhoff);
 		skb_set_inner_mac_header(skb, nhoff);
+		type = eh->h_proto;
+	} else {
+		type = NAPI_GRO_CB(skb)->network_proto;
+	}
 
 	ptype = gro_find_complete_by_type(type);
 	if (ptype != NULL)
diff --git a/net/ipv6/ip6_offload.c b/net/ipv6/ip6_offload.c
index 00dc2e3b0184..6e3a923ad573 100644
--- a/net/ipv6/ip6_offload.c
+++ b/net/ipv6/ip6_offload.c
@@ -232,6 +232,11 @@ INDIRECT_CALLABLE_SCOPE struct sk_buff *ipv6_gro_receive(struct list_head *head,
 	flush--;
 	nlen = skb_network_header_len(skb);
 
+	if (!NAPI_GRO_CB(skb)->encap_mark) {
+		NAPI_GRO_CB(skb)->transport_proto = proto;
+		NAPI_GRO_CB(skb)->network_len = nlen;
+	}
+
 	list_for_each_entry(p, head, list) {
 		const struct ipv6hdr *iph2;
 		__be32 first_word; /* <Version:4><Traffic_Class:8><Flow_Label:20> */
@@ -324,10 +329,6 @@ INDIRECT_CALLABLE_SCOPE int ipv6_gro_complete(struct sk_buff *skb, int nhoff)
 	int err = -ENOSYS;
 	u32 payload_len;
 
-	if (skb->encapsulation) {
-		skb_set_inner_protocol(skb, cpu_to_be16(ETH_P_IPV6));
-		skb_set_inner_network_header(skb, nhoff);
-	}
 
 	payload_len = skb->len - nhoff - sizeof(*iph);
 	if (unlikely(payload_len > IPV6_MAXPLEN)) {
@@ -341,6 +342,7 @@ INDIRECT_CALLABLE_SCOPE int ipv6_gro_complete(struct sk_buff *skb, int nhoff)
 		skb->len += hoplen;
 		skb->mac_header -= hoplen;
 		skb->network_header -= hoplen;
+		NAPI_GRO_CB(skb)->network_len += hoplen;
 		iph = (struct ipv6hdr *)(skb->data + nhoff);
 		hop_jumbo = (struct hop_jumbo_hdr *)(iph + 1);
 
@@ -358,7 +360,15 @@ INDIRECT_CALLABLE_SCOPE int ipv6_gro_complete(struct sk_buff *skb, int nhoff)
 		iph->payload_len = htons(payload_len);
 	}
 
-	nhoff += sizeof(*iph) + ipv6_exthdrs_len(iph, &ops);
+	if (skb->encapsulation) {
+		skb_set_inner_protocol(skb, cpu_to_be16(ETH_P_IPV6));
+		skb_set_inner_network_header(skb, nhoff);
+		nhoff += sizeof(*iph) + ipv6_exthdrs_len(iph, &ops);
+	} else {
+		ops = rcu_dereference(inet6_offloads[NAPI_GRO_CB(skb)->transport_proto]);
+		nhoff += NAPI_GRO_CB(skb)->network_len;
+	}
+
 	if (WARN_ON(!ops || !ops->callbacks.gro_complete))
 		goto out;
 
-- 
2.36.1

RE: [PATCH v2 2/2] gro: optimise redundant parsing of packets

[Willem de Bruijn]

Richard Gobert wrote:
> Currently the IPv6 extension headers are parsed twice: first in
> ipv6_gro_receive, and then again in ipv6_gro_complete.
> 
> By using the new ->transport_proto field, and also storing the size of the
> network header, we can avoid parsing extension headers a second time in
> ipv6_gro_complete (which saves multiple memory dereferences and conditional
> checks inside ipv6_exthdrs_len for a varying amount of extension headers in IPv6
> packets).
> 
> The implementation had to handle both inner and outer layers in case of
> encapsulation (as they can't use the same field).
> 
> Performance tests for TCP stream over IPv6 with a varying amount of extension
> headers demonstrate throughput improvement of ~0.7%.
> 
> In addition, I fixed a potential existing problem:
>  - The call to skb_set_inner_network_header at the beginning of
>    ipv6_gro_complete calculates inner_network_header based on skb->data by
>    calling skb_set_inner_network_header, and setting it to point to the beginning
>    of the ip header.
>  - If a packet is going to be handled by BIG TCP, the following code block is
>    going to shift the packet header, and skb->data is going to be changed as
>    well. 
> 
> When the two flows are combined, inner_network_header will point to the wrong
> place.
> 
> The fix is to place the whole encapsulation branch after the BIG TCP code block.

This should be a separate fix patch?

> This way, inner_network_header is calculated with a correct value of skb->data.
> Also, by arranging the code that way, the optimisation does not add an additional
> branch.
> 
> Signed-off-by: Richard Gobert <richardbgobert@gmail.com>
> ---
>  include/net/gro.h      |  9 +++++++++
>  net/ethernet/eth.c     | 14 +++++++++++---
>  net/ipv6/ip6_offload.c | 20 +++++++++++++++-----
>  3 files changed, 35 insertions(+), 8 deletions(-)
> 
> diff --git a/include/net/gro.h b/include/net/gro.h
> index 7b47dd6ce94f..35f60ea99f6c 100644
> --- a/include/net/gro.h
> +++ b/include/net/gro.h
> @@ -86,6 +86,15 @@ struct napi_gro_cb {
>  
>  	/* used to support CHECKSUM_COMPLETE for tunneling protocols */
>  	__wsum	csum;
> +
> +	/* Used in ipv6_gro_receive() */
> +	u16	network_len;
> +
> +	/* Used in eth_gro_receive() */
> +	__be16	network_proto;
> +

Why also cache eth->h_proto? That is not mentioned in the commit message.

> +	/* Used in ipv6_gro_receive() */
> +	u8	transport_proto;

Re: [PATCH v2 2/2] gro: optimise redundant parsing of packets

[Eric Dumazet]

On Wed, Feb 22, 2023 at 4:13 PM Richard Gobert <richardbgobert@gmail.com> wrote:
>
> Currently the IPv6 extension headers are parsed twice: first in
> ipv6_gro_receive, and then again in ipv6_gro_complete.
>
> By using the new ->transport_proto field, and also storing the size of the
> network header, we can avoid parsing extension headers a second time in
> ipv6_gro_complete (which saves multiple memory dereferences and conditional
> checks inside ipv6_exthdrs_len for a varying amount of extension headers in IPv6
> packets).
>
> The implementation had to handle both inner and outer layers in case of
> encapsulation (as they can't use the same field).
>
> Performance tests for TCP stream over IPv6 with a varying amount of extension
> headers demonstrate throughput improvement of ~0.7%.
>
> In addition, I fixed a potential existing problem:
>  - The call to skb_set_inner_network_header at the beginning of
>    ipv6_gro_complete calculates inner_network_header based on skb->data by
>    calling skb_set_inner_network_header, and setting it to point to the beginning
>    of the ip header.
>  - If a packet is going to be handled by BIG TCP, the following code block is
>    going to shift the packet header, and skb->data is going to be changed as
>    well.
>
> When the two flows are combined, inner_network_header will point to the wrong
> place.

net-next is closed.

If you think a fix is needed, please send a stand-alone and minimal
patch so that we can discuss its merit.

Note :

BIG TCP only supports native IPv6, not encapsulated traffic,
so we should not bother with inner_network_header yet.

Re: [PATCH v2 2/2] gro: optimise redundant parsing of packets

[Richard Gobert]

> On Wed, Feb 22, 2023 at 4:13 PM Richard Gobert <richardbgobert@gmail.com> wrote:
> >
> > Currently the IPv6 extension headers are parsed twice: first in
> > ipv6_gro_receive, and then again in ipv6_gro_complete.
> >
> > By using the new ->transport_proto field, and also storing the size of the
> > network header, we can avoid parsing extension headers a second time in
> > ipv6_gro_complete (which saves multiple memory dereferences and conditional
> > checks inside ipv6_exthdrs_len for a varying amount of extension headers in IPv6
> > packets).
> >
> > The implementation had to handle both inner and outer layers in case of
> > encapsulation (as they can't use the same field).
> >
> > Performance tests for TCP stream over IPv6 with a varying amount of extension
> > headers demonstrate throughput improvement of ~0.7%.
> >
> > In addition, I fixed a potential existing problem:
> >  - The call to skb_set_inner_network_header at the beginning of
> >    ipv6_gro_complete calculates inner_network_header based on skb->data by
> >    calling skb_set_inner_network_header, and setting it to point to the beginning
> >    of the ip header.
> >  - If a packet is going to be handled by BIG TCP, the following code block is
> >    going to shift the packet header, and skb->data is going to be changed as
> >    well.
> >
> > When the two flows are combined, inner_network_header will point to the wrong
> > place.
> 
> net-next is closed.
> 
> If you think a fix is needed, please send a stand-alone and minimal
> patch so that we can discuss its merit.

I'll repost when net-next will be opened again.
Thanks.

> 
> Note :
> 
> BIG TCP only supports native IPv6, not encapsulated traffic,
> so we should not bother with inner_network_header yet.