[PATCH net-next 0/2] net: add support for ip generic checksum offload for gre

This patchset it to add ip generic csum processing first in
skb_csum_hwoffload_help() in Patch 1/2 and then add csum
offload support for GRE header in Patch 2/2.

v1->v2:
  - See each patch's changelog.

Xin Long (2):
  net: support ip generic csum processing in skb_csum_hwoffload_help
  ip_gre: add csum offload support for gre header

 include/net/gre.h      | 19 +++++++------------
 net/core/dev.c         | 13 ++++++++++++-
 net/ipv4/gre_offload.c | 15 +++++++++++++--
 3 files changed, 32 insertions(+), 15 deletions(-)

-- 
2.1.0

NETIF_F_IP|IPV6_CSUM feature flag indicates UDP and TCP csum offload
while NETIF_F_HW_CSUM feature flag indicates ip generic csum offload
for HW, which includes not only for TCP/UDP csum, but also for other
protocols' csum like GRE's.

However, in skb_csum_hwoffload_help() it only checks features against
NETIF_F_CSUM_MASK(NETIF_F_HW|IP|IPV6_CSUM). So if it's a non TCP/UDP
packet and the features doesn't support NETIF_F_HW_CSUM, but supports
NETIF_F_IP|IPV6_CSUM only, it would still return 0 and leave the HW
to do csum.

This patch is to support ip generic csum processing by checking
NETIF_F_HW_CSUM for all protocols, and check (NETIF_F_IP_CSUM |
NETIF_F_IPV6_CSUM) only for TCP and UDP.

v1->v2:
  - not extend skb->csum_not_inet, but use skb->csum_offset to tell
    if it's an UDP/TCP csum packet.

Suggested-by: Alexander Duyck <redacted>
Signed-off-by: Xin Long <lucien.xin@gmail.com>
---
 net/core/dev.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 6df3f1b..aae116d 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -3621,7 +3621,18 @@ int skb_csum_hwoffload_help(struct sk_buff *skb,
 		return !!(features & NETIF_F_SCTP_CRC) ? 0 :
 			skb_crc32c_csum_help(skb);
 
-	return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
+	if (features & NETIF_F_HW_CSUM)
+		return 0;
+
+	if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
+		switch (skb->csum_offset) {
+		case offsetof(struct tcphdr, check):
+		case offsetof(struct udphdr, check):
+			return 0;
+		}
+	}
+
+	return skb_checksum_help(skb);
 }
 EXPORT_SYMBOL(skb_csum_hwoffload_help);
 

-- 
2.1.0

This patch is to add csum offload support for gre header:

On the TX path in gre_build_header(), when CHECKSUM_PARTIAL's set
for inner proto, it will calculate the csum for outer proto, and
inner csum will be offloaded later. Otherwise, CHECKSUM_PARTIAL
and csum_start/offset will be set for outer proto, and the outer
csum will be offloaded later.

On the GSO path in gre_gso_segment(), when CHECKSUM_PARTIAL is
not set for inner proto and the hardware supports csum offload,
CHECKSUM_PARTIAL and csum_start/offset will be set for outer
proto, and outer csum will be offloaded later. Otherwise, it
will do csum for outer proto by calling gso_make_checksum().

Note that SCTP has to do the csum by itself for non GSO path in
sctp_packet_pack(), as gre_build_header() can't handle the csum
with CHECKSUM_PARTIAL set for SCTP CRC csum offload.

v1->v2:
  - remove the SCTP part, as GRE dev doesn't support SCTP CRC CSUM
    and it will always do checksum for SCTP in sctp_packet_pack()
    when it's not a GSO packet.

Signed-off-by: Xin Long <lucien.xin@gmail.com>
---
 include/net/gre.h      | 19 +++++++------------
 net/ipv4/gre_offload.c | 15 +++++++++++++--
 2 files changed, 20 insertions(+), 14 deletions(-)

diff --git a/include/net/gre.h b/include/net/gre.h
index b60f212..4e20970 100644
--- a/include/net/gre.h
+++ b/include/net/gre.h

@@ -106,17 +106,6 @@ static inline __be16 gre_tnl_flags_to_gre_flags(__be16 tflags)
 	return flags;
 }
 
-static inline __sum16 gre_checksum(struct sk_buff *skb)
-{
-	__wsum csum;
-
-	if (skb->ip_summed == CHECKSUM_PARTIAL)
-		csum = lco_csum(skb);
-	else
-		csum = skb_checksum(skb, 0, skb->len, 0);
-	return csum_fold(csum);
-}
-
 static inline void gre_build_header(struct sk_buff *skb, int hdr_len,
 				    __be16 flags, __be16 proto,
 				    __be32 key, __be32 seq)

@@ -146,7 +135,13 @@ static inline void gre_build_header(struct sk_buff *skb, int hdr_len,
 		    !(skb_shinfo(skb)->gso_type &
 		      (SKB_GSO_GRE | SKB_GSO_GRE_CSUM))) {
 			*ptr = 0;
-			*(__sum16 *)ptr = gre_checksum(skb);
+			if (skb->ip_summed == CHECKSUM_PARTIAL) {
+				*(__sum16 *)ptr = csum_fold(lco_csum(skb));
+			} else {
+				skb->ip_summed = CHECKSUM_PARTIAL;
+				skb->csum_start = skb_transport_header(skb) - skb->head;
+				skb->csum_offset = sizeof(*greh);
+			}
 		}
 	}
 }

diff --git a/net/ipv4/gre_offload.c b/net/ipv4/gre_offload.c
index 10bc49b..1121a9d 100644
--- a/net/ipv4/gre_offload.c
+++ b/net/ipv4/gre_offload.c

@@ -15,10 +15,10 @@ static struct sk_buff *gre_gso_segment(struct sk_buff *skb,
 				       netdev_features_t features)
 {
 	int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
+	bool need_csum, offload_csum, gso_partial, need_ipsec;
 	struct sk_buff *segs = ERR_PTR(-EINVAL);
 	u16 mac_offset = skb->mac_header;
 	__be16 protocol = skb->protocol;
-	bool need_csum, gso_partial;
 	u16 mac_len = skb->mac_len;
 	int gre_offset, outer_hlen;
 

@@ -47,6 +47,11 @@ static struct sk_buff *gre_gso_segment(struct sk_buff *skb,
 	if (need_csum)
 		features &= ~NETIF_F_SCTP_CRC;
 
+	need_ipsec = skb_dst(skb) && dst_xfrm(skb_dst(skb));
+	/* Try to offload checksum if possible */
+	offload_csum = !!(need_csum && !need_ipsec &&
+			  (skb->dev->features & NETIF_F_HW_CSUM));
+
 	/* segment inner packet. */
 	segs = skb_mac_gso_segment(skb, features);
 	if (IS_ERR_OR_NULL(segs)) {

@@ -100,7 +105,13 @@ static struct sk_buff *gre_gso_segment(struct sk_buff *skb,
 		}
 
 		*(pcsum + 1) = 0;
-		*pcsum = gso_make_checksum(skb, 0);
+		if (skb->encapsulation || !offload_csum) {
+			*pcsum = gso_make_checksum(skb, 0);
+		} else {
+			skb->ip_summed = CHECKSUM_PARTIAL;
+			skb->csum_start = skb_transport_header(skb) - skb->head;
+			skb->csum_offset = sizeof(*greh);
+		}
 	} while ((skb = skb->next));
 out:
 	return segs;

-- 
2.1.0

On Sun, Jan 24, 2021 at 3:47 AM Xin Long [off-list ref] wrote:

This relies on no other protocols requesting CHECKSUM_PARTIAL
with these csum_offset values.

That is a fragile assumption. It may well be correct, and Alex argues
that point in v1 of the patch. I think that argumentation at the least
should be captured as a comment or in the commit message.

Or perhaps limit this optimization over s/w checksumming to

  skb->sk &&
  (skb->sk->sk_family == AF_INET  || .. ) &&
  (skb->sk->sk_type == SOCK_STREAM || ..)

?

On Tue, Jan 26, 2021 at 9:59 AM Willem de Bruijn
[off-list ref] wrote:

On Sun, Jan 24, 2021 at 3:47 AM Xin Long [off-list ref] wrote:

quoted

NETIF_F_IP|IPV6_CSUM feature flag indicates UDP and TCP csum offload
while NETIF_F_HW_CSUM feature flag indicates ip generic csum offload
for HW, which includes not only for TCP/UDP csum, but also for other
protocols' csum like GRE's.

However, in skb_csum_hwoffload_help() it only checks features against
NETIF_F_CSUM_MASK(NETIF_F_HW|IP|IPV6_CSUM). So if it's a non TCP/UDP
packet and the features doesn't support NETIF_F_HW_CSUM, but supports
NETIF_F_IP|IPV6_CSUM only, it would still return 0 and leave the HW
to do csum.

This patch is to support ip generic csum processing by checking
NETIF_F_HW_CSUM for all protocols, and check (NETIF_F_IP_CSUM |
NETIF_F_IPV6_CSUM) only for TCP and UDP.

v1->v2:
  - not extend skb->csum_not_inet, but use skb->csum_offset to tell
    if it's an UDP/TCP csum packet.

Suggested-by: Alexander Duyck <redacted>
Signed-off-by: Xin Long <lucien.xin@gmail.com>
---
 net/core/dev.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 6df3f1b..aae116d 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -3621,7 +3621,18 @@ int skb_csum_hwoffload_help(struct sk_buff *skb,
                return !!(features & NETIF_F_SCTP_CRC) ? 0 :
                        skb_crc32c_csum_help(skb);

-       return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
+       if (features & NETIF_F_HW_CSUM)
+               return 0;
+
+       if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
+               switch (skb->csum_offset) {
+               case offsetof(struct tcphdr, check):
+               case offsetof(struct udphdr, check):

This relies on no other protocols requesting CHECKSUM_PARTIAL
with these csum_offset values.

That is a fragile assumption. It may well be correct, and Alex argues
that point in v1 of the patch. I think that argumentation at the least
should be captured as a comment or in the commit message.

will add a note in changelog and repost, thanks!

Or perhaps limit this optimization over s/w checksumming to

  skb->sk &&
  (skb->sk->sk_family == AF_INET  || .. ) &&
  (skb->sk->sk_type == SOCK_STREAM || ..)

?