[net-next 0/3] ip_gre: a bunch of fixes for mtu

[net-next 0/3] ip_gre: a bunch of fixes for mtu

[William Tu]

The first two patches are to fix some issues for mtu and needed_headroom length
calculation from the gre and erspan tunnel header.  The last path tries to avoid
arithmetic operation for every packet when checking for erspan truncate.

William Tu (3):
  ip_gre: fix mtu and headroom size
  ip_gre: fix erspan tunnel mtu calculation
  ip_gre: cache the device mtu hard_header_len calc

 include/net/erspan.h     |  1 +
 include/net/ip_tunnels.h |  1 +
 net/ipv4/ip_gre.c        | 23 +++++++++++------------
 net/ipv4/ip_tunnel.c     |  3 +++
 4 files changed, 16 insertions(+), 12 deletions(-)

-- 
2.7.4

[net-next 1/3] ip_gre: fix mtu and headroom size

[William Tu]

The ip_gre_calc_hlen() already counts gre's base header len (4-byte)
plus optional header's len, so tunnel->hlen has the entire gre
headeri + options len. Thus, remove the -4 and +4 when calculating
the needed_headroom and mtu.

Fixes: 4565e9919cda ("gre: Setup and TX path for gre/UDP foo-over-udp encapsulation")
Signed-off-by: William Tu <u9012063@gmail.com>
Cc: Tom Herbert <tom@quantonium.net>
---
 net/ipv4/ip_gre.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/net/ipv4/ip_gre.c b/net/ipv4/ip_gre.c
index c105a315b1a3..286065c35959 100644
--- a/net/ipv4/ip_gre.c
+++ b/net/ipv4/ip_gre.c
@@ -947,8 +947,8 @@ static void __gre_tunnel_init(struct net_device *dev)
 
 	t_hlen = tunnel->hlen + sizeof(struct iphdr);
 
-	dev->needed_headroom	= LL_MAX_HEADER + t_hlen + 4;
-	dev->mtu		= ETH_DATA_LEN - t_hlen - 4;
+	dev->needed_headroom	= LL_MAX_HEADER + t_hlen;
+	dev->mtu		= ETH_DATA_LEN - t_hlen;
 
 	dev->features		|= GRE_FEATURES;
 	dev->hw_features	|= GRE_FEATURES;
-- 
2.7.4

[net-next 2/3] ip_gre: fix erspan tunnel mtu calculation

[William Tu]

Remove the unnecessary -4 and +4 bytes at mtu and headroom calculation.
In addition, erspan uses fixed 8-byte gre header, so add ERSPAN_GREHDR_LEN
macro for better readability.

Now tunnel->hlen = grehdr(8) + erspanhdr(8) = 16 byte.
The mtu should be ETH_DATA_LEN - 16 - iph(20) = 1464.
After the ip_tunnel_bind_dev(), the mtu is adjusted to
1464 - 14 (dev->hard_header_len) = 1450.
The maximum skb->len the erspan tunnel can carry without
being truncated is 1450 + 14 = 1464 byte.

Signed-off-by: William Tu <u9012063@gmail.com>
Cc: Xin Long <lucien.xin@gmail.com>
---
 include/net/erspan.h |  1 +
 net/ipv4/ip_gre.c    | 11 +++++------
 2 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/include/net/erspan.h b/include/net/erspan.h
index ca94fc86865e..e28294e248d0 100644
--- a/include/net/erspan.h
+++ b/include/net/erspan.h
@@ -28,6 +28,7 @@
  */
 
 #define ERSPAN_VERSION	0x1
+#define ERSPAN_GREHDR_LEN 8	/* ERSPAN has fixed 8-byte GRE header */
 
 #define VER_MASK	0xf000
 #define VLAN_MASK	0x0fff
diff --git a/net/ipv4/ip_gre.c b/net/ipv4/ip_gre.c
index 286065c35959..6e6e4c4811cc 100644
--- a/net/ipv4/ip_gre.c
+++ b/net/ipv4/ip_gre.c
@@ -569,8 +569,7 @@ static void erspan_fb_xmit(struct sk_buff *skb, struct net_device *dev,
 
 	key = &tun_info->key;
 
-	/* ERSPAN has fixed 8 byte GRE header */
-	tunnel_hlen = 8 + sizeof(struct erspanhdr);
+	tunnel_hlen = ERSPAN_GREHDR_LEN + sizeof(struct erspanhdr);
 
 	rt = prepare_fb_xmit(skb, dev, &fl, tunnel_hlen);
 	if (!rt)
@@ -591,7 +590,7 @@ static void erspan_fb_xmit(struct sk_buff *skb, struct net_device *dev,
 	erspan_build_header(skb, tunnel_id_to_key32(key->tun_id),
 			    ntohl(md->index), truncate);
 
-	gre_build_header(skb, 8, TUNNEL_SEQ,
+	gre_build_header(skb, ERSPAN_GREHDR_LEN, TUNNEL_SEQ,
 			 htons(ETH_P_ERSPAN), 0, htonl(tunnel->o_seqno++));
 
 	df = key->tun_flags & TUNNEL_DONT_FRAGMENT ?  htons(IP_DF) : 0;
@@ -1242,14 +1241,14 @@ static int erspan_tunnel_init(struct net_device *dev)
 	struct ip_tunnel *tunnel = netdev_priv(dev);
 	int t_hlen;
 
-	tunnel->tun_hlen = 8;
+	tunnel->tun_hlen = ERSPAN_GREHDR_LEN;
 	tunnel->parms.iph.protocol = IPPROTO_GRE;
 	tunnel->hlen = tunnel->tun_hlen + tunnel->encap_hlen +
 		       sizeof(struct erspanhdr);
 	t_hlen = tunnel->hlen + sizeof(struct iphdr);
 
-	dev->needed_headroom = LL_MAX_HEADER + t_hlen + 4;
-	dev->mtu = ETH_DATA_LEN - t_hlen - 4;
+	dev->needed_headroom = LL_MAX_HEADER + t_hlen;
+	dev->mtu = ETH_DATA_LEN - t_hlen;
 	dev->features		|= GRE_FEATURES;
 	dev->hw_features	|= GRE_FEATURES;
 	dev->priv_flags		|= IFF_LIVE_ADDR_CHANGE;
-- 
2.7.4

Re: [net-next 2/3] ip_gre: fix erspan tunnel mtu calculation

[Xin Long]

On Tue, Oct 10, 2017 at 4:47 AM, William Tu <u9012063@gmail.com> wrote:
> Remove the unnecessary -4 and +4 bytes at mtu and headroom calculation.
> In addition, erspan uses fixed 8-byte gre header, so add ERSPAN_GREHDR_LEN
> macro for better readability.
>
> Now tunnel->hlen = grehdr(8) + erspanhdr(8) = 16 byte.
> The mtu should be ETH_DATA_LEN - 16 - iph(20) = 1464.
> After the ip_tunnel_bind_dev(), the mtu is adjusted to
> 1464 - 14 (dev->hard_header_len) = 1450.
> The maximum skb->len the erspan tunnel can carry without
> being truncated is 1450 + 14 = 1464 byte.
>
> Signed-off-by: William Tu <u9012063@gmail.com>
> Cc: Xin Long <lucien.xin@gmail.com>
> ---
>  include/net/erspan.h |  1 +
>  net/ipv4/ip_gre.c    | 11 +++++------
>  2 files changed, 6 insertions(+), 6 deletions(-)
>
> diff --git a/include/net/erspan.h b/include/net/erspan.h
> index ca94fc86865e..e28294e248d0 100644
> --- a/include/net/erspan.h
> +++ b/include/net/erspan.h
> @@ -28,6 +28,7 @@
>   */
>
>  #define ERSPAN_VERSION 0x1
> +#define ERSPAN_GREHDR_LEN 8    /* ERSPAN has fixed 8-byte GRE header */
>
>  #define VER_MASK       0xf000
>  #define VLAN_MASK      0x0fff
> diff --git a/net/ipv4/ip_gre.c b/net/ipv4/ip_gre.c
> index 286065c35959..6e6e4c4811cc 100644
> --- a/net/ipv4/ip_gre.c
> +++ b/net/ipv4/ip_gre.c
> @@ -569,8 +569,7 @@ static void erspan_fb_xmit(struct sk_buff *skb, struct net_device *dev,
>
>         key = &tun_info->key;
>
> -       /* ERSPAN has fixed 8 byte GRE header */
> -       tunnel_hlen = 8 + sizeof(struct erspanhdr);
> +       tunnel_hlen = ERSPAN_GREHDR_LEN + sizeof(struct erspanhdr);
>
>         rt = prepare_fb_xmit(skb, dev, &fl, tunnel_hlen);
>         if (!rt)
> @@ -591,7 +590,7 @@ static void erspan_fb_xmit(struct sk_buff *skb, struct net_device *dev,
>         erspan_build_header(skb, tunnel_id_to_key32(key->tun_id),
>                             ntohl(md->index), truncate);
>
> -       gre_build_header(skb, 8, TUNNEL_SEQ,
> +       gre_build_header(skb, ERSPAN_GREHDR_LEN, TUNNEL_SEQ,
>                          htons(ETH_P_ERSPAN), 0, htonl(tunnel->o_seqno++));
>
>         df = key->tun_flags & TUNNEL_DONT_FRAGMENT ?  htons(IP_DF) : 0;
> @@ -1242,14 +1241,14 @@ static int erspan_tunnel_init(struct net_device *dev)
>         struct ip_tunnel *tunnel = netdev_priv(dev);
>         int t_hlen;
>
> -       tunnel->tun_hlen = 8;
> +       tunnel->tun_hlen = ERSPAN_GREHDR_LEN;
>         tunnel->parms.iph.protocol = IPPROTO_GRE;
>         tunnel->hlen = tunnel->tun_hlen + tunnel->encap_hlen +
>                        sizeof(struct erspanhdr);
>         t_hlen = tunnel->hlen + sizeof(struct iphdr);
>
> -       dev->needed_headroom = LL_MAX_HEADER + t_hlen + 4;
> -       dev->mtu = ETH_DATA_LEN - t_hlen - 4;
> +       dev->needed_headroom = LL_MAX_HEADER + t_hlen;
> +       dev->mtu = ETH_DATA_LEN - t_hlen;
1. I guess '+4-4' stuff was copied from __gre_tunnel_init(), I'm thinking
it may be there for some reason.

2. 'dev->needed_headroom =' and 'dev->mtu =' are really needed ?
As I've seen both will be updated in .newlink:
ipgre_newlink() -> ip_tunnel_newlink() -> ip_tunnel_bind_dev()

>         dev->features           |= GRE_FEATURES;
>         dev->hw_features        |= GRE_FEATURES;
>         dev->priv_flags         |= IFF_LIVE_ADDR_CHANGE;
> --
> 2.7.4
>

Re: [net-next 2/3] ip_gre: fix erspan tunnel mtu calculation

[William Tu]

>> @@ -1242,14 +1241,14 @@ static int erspan_tunnel_init(struct net_device *dev)
>>         struct ip_tunnel *tunnel = netdev_priv(dev);
>>         int t_hlen;
>>
>> -       tunnel->tun_hlen = 8;
>> +       tunnel->tun_hlen = ERSPAN_GREHDR_LEN;
>>         tunnel->parms.iph.protocol = IPPROTO_GRE;
>>         tunnel->hlen = tunnel->tun_hlen + tunnel->encap_hlen +
>>                        sizeof(struct erspanhdr);
>>         t_hlen = tunnel->hlen + sizeof(struct iphdr);
>>
>> -       dev->needed_headroom = LL_MAX_HEADER + t_hlen + 4;
>> -       dev->mtu = ETH_DATA_LEN - t_hlen - 4;
>> +       dev->needed_headroom = LL_MAX_HEADER + t_hlen;
>> +       dev->mtu = ETH_DATA_LEN - t_hlen;
> 1. I guess '+4-4' stuff was copied from __gre_tunnel_init(), I'm thinking
> it may be there for some reason.
>
I traced back to
4565e9919cda ("gre: Setup and TX path for gre/UDP foo-over-udp encapsulation")
and I think '+4-4' is there for GRE base header length.

Since now we do
    dev->mtu = ETH_DATA_LEN - t_hlen;
and t_hlen already counts the the gre base header + optional header
len, I think it's not needed.

> 2. 'dev->needed_headroom =' and 'dev->mtu =' are really needed ?
> As I've seen both will be updated in .newlink:
> ipgre_newlink() -> ip_tunnel_newlink() -> ip_tunnel_bind_dev()
>
right, I also find both values gets overwritten by
ip_tunnel_bind_dev() using my test cases. Maybe we can remove them?

Thanks
William

Re: [net-next 2/3] ip_gre: fix erspan tunnel mtu calculation

[Xin Long]

On Tue, Oct 10, 2017 at 8:59 PM, William Tu <u9012063@gmail.com> wrote:
>>> @@ -1242,14 +1241,14 @@ static int erspan_tunnel_init(struct net_device *dev)
>>>         struct ip_tunnel *tunnel = netdev_priv(dev);
>>>         int t_hlen;
>>>
>>> -       tunnel->tun_hlen = 8;
>>> +       tunnel->tun_hlen = ERSPAN_GREHDR_LEN;
>>>         tunnel->parms.iph.protocol = IPPROTO_GRE;
>>>         tunnel->hlen = tunnel->tun_hlen + tunnel->encap_hlen +
>>>                        sizeof(struct erspanhdr);
>>>         t_hlen = tunnel->hlen + sizeof(struct iphdr);
>>>
>>> -       dev->needed_headroom = LL_MAX_HEADER + t_hlen + 4;
>>> -       dev->mtu = ETH_DATA_LEN - t_hlen - 4;
>>> +       dev->needed_headroom = LL_MAX_HEADER + t_hlen;
>>> +       dev->mtu = ETH_DATA_LEN - t_hlen;
>> 1. I guess '+4-4' stuff was copied from __gre_tunnel_init(), I'm thinking
>> it may be there for some reason.
>>
> I traced back to
> 4565e9919cda ("gre: Setup and TX path for gre/UDP foo-over-udp encapsulation")
> and I think '+4-4' is there for GRE base header length.
>
> Since now we do
>     dev->mtu = ETH_DATA_LEN - t_hlen;
> and t_hlen already counts the the gre base header + optional header
> len, I think it's not needed.

okay. thanks.

>
>> 2. 'dev->needed_headroom =' and 'dev->mtu =' are really needed ?
>> As I've seen both will be updated in .newlink:
>> ipgre_newlink() -> ip_tunnel_newlink() -> ip_tunnel_bind_dev()
>>
> right, I also find both values gets overwritten by
> ip_tunnel_bind_dev() using my test cases. Maybe we can remove them?

It's there just in case that there is no lower dev found, but
ip_tunnel_newlink/create always updates dev->mtu even if
there is no lower dev found.

let's leave as it is for now, ipgre may just not be sure ip_tunnel_xxx would
do it when no lower dev found.

[net-next 3/3] ip_gre: cache the device mtu hard_header_len calc

[William Tu]

The patch introduces ip_tunnel->ether_mtu fields to cache the value of
dev->mtu + dev->hard_header_len.  This avoids the arithmetic operation
on every packet.

Signed-off-by: William Tu <u9012063@gmail.com>
Cc: David Laight <David.Laight@aculab.com>
---
 include/net/ip_tunnels.h | 1 +
 net/ipv4/ip_gre.c        | 8 ++++----
 net/ipv4/ip_tunnel.c     | 3 +++
 3 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/include/net/ip_tunnels.h b/include/net/ip_tunnels.h
index b41a1e057fce..19565be26e13 100644
--- a/include/net/ip_tunnels.h
+++ b/include/net/ip_tunnels.h
@@ -117,6 +117,7 @@ struct ip_tunnel {
 
 	/* This field used only by ERSPAN */
 	u32		index;		/* ERSPAN type II index */
+	unsigned int	ether_mtu;	/* The mtu including the ether hdr */
 
 	struct dst_cache dst_cache;
 
diff --git a/net/ipv4/ip_gre.c b/net/ipv4/ip_gre.c
index 6e6e4c4811cc..994b8ddea0b1 100644
--- a/net/ipv4/ip_gre.c
+++ b/net/ipv4/ip_gre.c
@@ -578,8 +578,8 @@ static void erspan_fb_xmit(struct sk_buff *skb, struct net_device *dev,
 	if (gre_handle_offloads(skb, false))
 		goto err_free_rt;
 
-	if (skb->len > dev->mtu + dev->hard_header_len) {
-		pskb_trim(skb, dev->mtu + dev->hard_header_len);
+	if (skb->len > tunnel->ether_mtu) {
+		pskb_trim(skb, tunnel->ether_mtu);
 		truncate = true;
 	}
 
@@ -730,8 +730,8 @@ static netdev_tx_t erspan_xmit(struct sk_buff *skb,
 	if (skb_cow_head(skb, dev->needed_headroom))
 		goto free_skb;
 
-	if (skb->len > dev->mtu + dev->hard_header_len) {
-		pskb_trim(skb, dev->mtu + dev->hard_header_len);
+	if (skb->len > tunnel->ether_mtu) {
+		pskb_trim(skb, tunnel->ether_mtu);
 		truncate = true;
 	}
 
diff --git a/net/ipv4/ip_tunnel.c b/net/ipv4/ip_tunnel.c
index fe6fee728ce4..859af5b86802 100644
--- a/net/ipv4/ip_tunnel.c
+++ b/net/ipv4/ip_tunnel.c
@@ -348,6 +348,7 @@ static int ip_tunnel_bind_dev(struct net_device *dev)
 
 	dev->needed_headroom = t_hlen + hlen;
 	mtu -= (dev->hard_header_len + t_hlen);
+	tunnel->ether_mtu = mtu + dev->hard_header_len;
 
 	if (mtu < 68)
 		mtu = 68;
@@ -952,6 +953,7 @@ int __ip_tunnel_change_mtu(struct net_device *dev, int new_mtu, bool strict)
 	}
 
 	dev->mtu = new_mtu;
+	tunnel->ether_mtu = new_mtu + dev->hard_header_len;
 	return 0;
 }
 EXPORT_SYMBOL_GPL(__ip_tunnel_change_mtu);
@@ -1183,6 +1185,7 @@ int ip_tunnel_init(struct net_device *dev)
 
 	tunnel->dev = dev;
 	tunnel->net = dev_net(dev);
+	tunnel->ether_mtu = dev->mtu + dev->hard_header_len;
 	strcpy(tunnel->parms.name, dev->name);
 	iph->version		= 4;
 	iph->ihl		= 5;
-- 
2.7.4

Re: [net-next 3/3] ip_gre: cache the device mtu hard_header_len calc

[Xin Long]

On Tue, Oct 10, 2017 at 4:47 AM, William Tu <u9012063@gmail.com> wrote:
> The patch introduces ip_tunnel->ether_mtu fields to cache the value of
> dev->mtu + dev->hard_header_len.  This avoids the arithmetic operation
> on every packet.
>
> Signed-off-by: William Tu <u9012063@gmail.com>
> Cc: David Laight <David.Laight@aculab.com>
> ---
>  include/net/ip_tunnels.h | 1 +
>  net/ipv4/ip_gre.c        | 8 ++++----
>  net/ipv4/ip_tunnel.c     | 3 +++
>  3 files changed, 8 insertions(+), 4 deletions(-)
>
> diff --git a/include/net/ip_tunnels.h b/include/net/ip_tunnels.h
> index b41a1e057fce..19565be26e13 100644
> --- a/include/net/ip_tunnels.h
> +++ b/include/net/ip_tunnels.h
> @@ -117,6 +117,7 @@ struct ip_tunnel {
>
>         /* This field used only by ERSPAN */
>         u32             index;          /* ERSPAN type II index */
> +       unsigned int    ether_mtu;      /* The mtu including the ether hdr */
>
>         struct dst_cache dst_cache;
ip_tunnel is a very common structure for various tunnels.
Adding ether_mtu into it to avoid  ONLY an addition operation ONLY for
erspan, I think it's not worth it.

>
> diff --git a/net/ipv4/ip_gre.c b/net/ipv4/ip_gre.c
> index 6e6e4c4811cc..994b8ddea0b1 100644
> --- a/net/ipv4/ip_gre.c
> +++ b/net/ipv4/ip_gre.c
> @@ -578,8 +578,8 @@ static void erspan_fb_xmit(struct sk_buff *skb, struct net_device *dev,
>         if (gre_handle_offloads(skb, false))
>                 goto err_free_rt;
>
> -       if (skb->len > dev->mtu + dev->hard_header_len) {
> -               pskb_trim(skb, dev->mtu + dev->hard_header_len);
> +       if (skb->len > tunnel->ether_mtu) {
> +               pskb_trim(skb, tunnel->ether_mtu);
>                 truncate = true;
>         }
>
> @@ -730,8 +730,8 @@ static netdev_tx_t erspan_xmit(struct sk_buff *skb,
>         if (skb_cow_head(skb, dev->needed_headroom))
>                 goto free_skb;
>
> -       if (skb->len > dev->mtu + dev->hard_header_len) {
> -               pskb_trim(skb, dev->mtu + dev->hard_header_len);
> +       if (skb->len > tunnel->ether_mtu) {
> +               pskb_trim(skb, tunnel->ether_mtu);
>                 truncate = true;
>         }
>
> diff --git a/net/ipv4/ip_tunnel.c b/net/ipv4/ip_tunnel.c
> index fe6fee728ce4..859af5b86802 100644
> --- a/net/ipv4/ip_tunnel.c
> +++ b/net/ipv4/ip_tunnel.c
> @@ -348,6 +348,7 @@ static int ip_tunnel_bind_dev(struct net_device *dev)
>
>         dev->needed_headroom = t_hlen + hlen;
>         mtu -= (dev->hard_header_len + t_hlen);
> +       tunnel->ether_mtu = mtu + dev->hard_header_len;
>
>         if (mtu < 68)
>                 mtu = 68;
> @@ -952,6 +953,7 @@ int __ip_tunnel_change_mtu(struct net_device *dev, int new_mtu, bool strict)
>         }
>
>         dev->mtu = new_mtu;
> +       tunnel->ether_mtu = new_mtu + dev->hard_header_len;
>         return 0;
>  }
>  EXPORT_SYMBOL_GPL(__ip_tunnel_change_mtu);
> @@ -1183,6 +1185,7 @@ int ip_tunnel_init(struct net_device *dev)
>
>         tunnel->dev = dev;
>         tunnel->net = dev_net(dev);
> +       tunnel->ether_mtu = dev->mtu + dev->hard_header_len;
>         strcpy(tunnel->parms.name, dev->name);
>         iph->version            = 4;
>         iph->ihl                = 5;
> --
> 2.7.4
>