Re: [PATCH net-next] hv_netvsc: Use simple parser for IPv4 and v6 headers

Re: [PATCH net-next] hv_netvsc: Use simple parser for IPv4 and v6 headers

[Eric Dumazet]

On Wed, 2015-12-16 at 10:03 -0800, Haiyang Zhang wrote:
> To avoid performance overhead when using skb_flow_dissect_flow_keys(),
> we switch to the simple parsers to get the IP and port numbers.
> 
> Performance comparison: throughput (Gbps):
> Number of connections, before patch, after patch
> 1			8.56		10.18
> 4			11.17		14.07
> 16			12.21		21.78
> 64			18.71		32.08
> 256			15.92		26.32
> 1024			8.41		15.49
> 3000			7.82		11.58
> 
> Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
> Tested-by: Simon Xiao <sixiao@microsoft.com>
> Reviewed-by: K. Y. Srinivasan <kys@microsoft.com>
> ---
>  drivers/net/hyperv/netvsc_drv.c |   38 +++++++++++++++++++++++++++++---------
>  1 files changed, 29 insertions(+), 9 deletions(-)
> 
> diff --git a/drivers/net/hyperv/netvsc_drv.c b/drivers/net/hyperv/netvsc_drv.c
> index 1c8db9a..e28951f 100644
> --- a/drivers/net/hyperv/netvsc_drv.c
> +++ b/drivers/net/hyperv/netvsc_drv.c
> @@ -237,20 +237,40 @@ static u32 comp_hash(u8 *key, int klen, void *data, int dlen)
>  
>  static bool netvsc_set_hash(u32 *hash, struct sk_buff *skb)
>  {
> -	struct flow_keys flow;
> +	struct iphdr *iphdr;
> +	struct ipv6hdr *ipv6hdr;
> +	__be32 dbuf[9];
>  	int data_len;
>  
> -	if (!skb_flow_dissect_flow_keys(skb, &flow, 0) ||
> -	    !(flow.basic.n_proto == htons(ETH_P_IP) ||
> -	      flow.basic.n_proto == htons(ETH_P_IPV6)))
> +	if (eth_hdr(skb)->h_proto != htons(ETH_P_IP) &&
> +	    eth_hdr(skb)->h_proto != htons(ETH_P_IPV6))
>  		return false;
>  
> -	if (flow.basic.ip_proto == IPPROTO_TCP)
> -		data_len = 12;
> -	else
> -		data_len = 8;
> +	iphdr = ip_hdr(skb);
> +	ipv6hdr = ipv6_hdr(skb);
> +
> +	if (iphdr->version == 4) {
> +		dbuf[0] = iphdr->saddr;
> +		dbuf[1] = iphdr->daddr;
> +		if (iphdr->protocol == IPPROTO_TCP) {
> +			dbuf[2] = *(__be32 *)&tcp_hdr(skb)->source;
> +			data_len = 12;
> +		} else {
> +			data_len = 8;
> +		}
> +	} else if (ipv6hdr->version == 6) {
> +		memcpy(dbuf, &ipv6hdr->saddr, 32);
> +		if (ipv6hdr->nexthdr == IPPROTO_TCP) {
> +			dbuf[8] = *(__be32 *)&tcp_hdr(skb)->source;
> +			data_len = 36;
> +		} else {
> +			data_len = 32;
> +		}
> +	} else {
> +		return false;
> +	}
>  
> -	*hash = comp_hash(netvsc_hash_key, HASH_KEYLEN, &flow, data_len);
> +	*hash = comp_hash(netvsc_hash_key, HASH_KEYLEN, dbuf, data_len);
>  
>  	return true;
>  }


This looks very very wrong to me.

How many times this is called per second, for the 'one flow' case ?

Don't you use TSO in this driver ?

What about encapsulation ?

I suspect you have a quite different issue here.

You simply could use skb_get_hash() since local TCP flows will provide a
l4 skb->hash and you have no further flow dissection to do.

[PATCH net-next] hv_netvsc: Use simple parser for IPv4 and v6 headers

[Haiyang Zhang]

To avoid performance overhead when using skb_flow_dissect_flow_keys(),
we switch to the simple parsers to get the IP and port numbers.

Performance comparison: throughput (Gbps):
Number of connections, before patch, after patch
1			8.56		10.18
4			11.17		14.07
16			12.21		21.78
64			18.71		32.08
256			15.92		26.32
1024			8.41		15.49
3000			7.82		11.58

Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Tested-by: Simon Xiao <sixiao@microsoft.com>
Reviewed-by: K. Y. Srinivasan <kys@microsoft.com>
---
 drivers/net/hyperv/netvsc_drv.c |   38 +++++++++++++++++++++++++++++---------
 1 files changed, 29 insertions(+), 9 deletions(-)

diff --git a/drivers/net/hyperv/netvsc_drv.c b/drivers/net/hyperv/netvsc_drv.c
index 1c8db9a..e28951f 100644
--- a/drivers/net/hyperv/netvsc_drv.c
+++ b/drivers/net/hyperv/netvsc_drv.c
@@ -237,20 +237,40 @@ static u32 comp_hash(u8 *key, int klen, void *data, int dlen)
 
 static bool netvsc_set_hash(u32 *hash, struct sk_buff *skb)
 {
-	struct flow_keys flow;
+	struct iphdr *iphdr;
+	struct ipv6hdr *ipv6hdr;
+	__be32 dbuf[9];
 	int data_len;
 
-	if (!skb_flow_dissect_flow_keys(skb, &flow, 0) ||
-	    !(flow.basic.n_proto == htons(ETH_P_IP) ||
-	      flow.basic.n_proto == htons(ETH_P_IPV6)))
+	if (eth_hdr(skb)->h_proto != htons(ETH_P_IP) &&
+	    eth_hdr(skb)->h_proto != htons(ETH_P_IPV6))
 		return false;
 
-	if (flow.basic.ip_proto == IPPROTO_TCP)
-		data_len = 12;
-	else
-		data_len = 8;
+	iphdr = ip_hdr(skb);
+	ipv6hdr = ipv6_hdr(skb);
+
+	if (iphdr->version == 4) {
+		dbuf[0] = iphdr->saddr;
+		dbuf[1] = iphdr->daddr;
+		if (iphdr->protocol == IPPROTO_TCP) {
+			dbuf[2] = *(__be32 *)&tcp_hdr(skb)->source;
+			data_len = 12;
+		} else {
+			data_len = 8;
+		}
+	} else if (ipv6hdr->version == 6) {
+		memcpy(dbuf, &ipv6hdr->saddr, 32);
+		if (ipv6hdr->nexthdr == IPPROTO_TCP) {
+			dbuf[8] = *(__be32 *)&tcp_hdr(skb)->source;
+			data_len = 36;
+		} else {
+			data_len = 32;
+		}
+	} else {
+		return false;
+	}
 
-	*hash = comp_hash(netvsc_hash_key, HASH_KEYLEN, &flow, data_len);
+	*hash = comp_hash(netvsc_hash_key, HASH_KEYLEN, dbuf, data_len);
 
 	return true;
 }
-- 
1.7.4.1

Re: [PATCH net-next] hv_netvsc: Use simple parser for IPv4 and v6 headers

[Sergei Shtylyov]

Hello.

On 12/16/2015 09:03 PM, Haiyang Zhang wrote:

> To avoid performance overhead when using skb_flow_dissect_flow_keys(),
> we switch to the simple parsers to get the IP and port numbers.
>
> Performance comparison: throughput (Gbps):
> Number of connections, before patch, after patch
> 1			8.56		10.18
> 4			11.17		14.07
> 16			12.21		21.78
> 64			18.71		32.08
> 256			15.92		26.32
> 1024			8.41		15.49
> 3000			7.82		11.58
>
> Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
> Tested-by: Simon Xiao <sixiao@microsoft.com>
> Reviewed-by: K. Y. Srinivasan <kys@microsoft.com>
> ---
>   drivers/net/hyperv/netvsc_drv.c |   38 +++++++++++++++++++++++++++++---------
>   1 files changed, 29 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/net/hyperv/netvsc_drv.c b/drivers/net/hyperv/netvsc_drv.c
> index 1c8db9a..e28951f 100644
> --- a/drivers/net/hyperv/netvsc_drv.c
> +++ b/drivers/net/hyperv/netvsc_drv.c
> @@ -237,20 +237,40 @@ static u32 comp_hash(u8 *key, int klen, void *data, int dlen)
[...]
> +	if (iphdr->version == 4) {
> +		dbuf[0] = iphdr->saddr;
> +		dbuf[1] = iphdr->daddr;
> +		if (iphdr->protocol == IPPROTO_TCP) {
> +			dbuf[2] = *(__be32 *)&tcp_hdr(skb)->source;
> +			data_len = 12;
> +		} else {
> +			data_len = 8;
> +		}
> +	} else if (ipv6hdr->version == 6) {
> +		memcpy(dbuf, &ipv6hdr->saddr, 32);
> +		if (ipv6hdr->nexthdr == IPPROTO_TCP) {
> +			dbuf[8] = *(__be32 *)&tcp_hdr(skb)->source;
> +			data_len = 36;
> +		} else {
> +			data_len = 32;
> +		}
> +	} else {
> +		return false;
> +	}

    This is asking to be a *switch* statement.

[...]

MBR, Sergei

Re: [PATCH net-next] hv_netvsc: Use simple parser for IPv4 and v6 headers

[Sergei Shtylyov]

On 12/16/2015 09:34 PM, Sergei Shtylyov wrote:

>> To avoid performance overhead when using skb_flow_dissect_flow_keys(),
>> we switch to the simple parsers to get the IP and port numbers.
>>
>> Performance comparison: throughput (Gbps):
>> Number of connections, before patch, after patch
>> 1            8.56        10.18
>> 4            11.17        14.07
>> 16            12.21        21.78
>> 64            18.71        32.08
>> 256            15.92        26.32
>> 1024            8.41        15.49
>> 3000            7.82        11.58
>>
>> Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
>> Tested-by: Simon Xiao <sixiao@microsoft.com>
>> Reviewed-by: K. Y. Srinivasan <kys@microsoft.com>
>> ---
>>   drivers/net/hyperv/netvsc_drv.c |   38 +++++++++++++++++++++++++++++---------
>>   1 files changed, 29 insertions(+), 9 deletions(-)
>>
>> diff --git a/drivers/net/hyperv/netvsc_drv.c b/drivers/net/hyperv/netvsc_drv.c
>> index 1c8db9a..e28951f 100644
>> --- a/drivers/net/hyperv/netvsc_drv.c
>> +++ b/drivers/net/hyperv/netvsc_drv.c
>> @@ -237,20 +237,40 @@ static u32 comp_hash(u8 *key, int klen, void *data,
>> int dlen)
> [...]
>> +    if (iphdr->version == 4) {
>> +        dbuf[0] = iphdr->saddr;
>> +        dbuf[1] = iphdr->daddr;
>> +        if (iphdr->protocol == IPPROTO_TCP) {
>> +            dbuf[2] = *(__be32 *)&tcp_hdr(skb)->source;
>> +            data_len = 12;
>> +        } else {
>> +            data_len = 8;
>> +        }
>> +    } else if (ipv6hdr->version == 6) {
>> +        memcpy(dbuf, &ipv6hdr->saddr, 32);
>> +        if (ipv6hdr->nexthdr == IPPROTO_TCP) {
>> +            dbuf[8] = *(__be32 *)&tcp_hdr(skb)->source;
>> +            data_len = 36;
>> +        } else {
>> +            data_len = 32;
>> +        }
>> +    } else {
>> +        return false;
>> +    }
>
>     This is asking to be a *switch* statement.

    Oops, nevermind. I'd misread the code.

> [...]

MBR, Sergei

RE: [PATCH net-next] hv_netvsc: Use simple parser for IPv4 and v6 headers

[Haiyang Zhang]

> -----Original Message-----
> From: Eric Dumazet [mailto:eric.dumazet@gmail.com]
> Sent: Wednesday, December 16, 2015 12:08 PM
> 
> This looks very very wrong to me.
> 
> How many times this is called per second, for the 'one flow' case ?
> 
> Don't you use TSO in this driver ?
> 
> What about encapsulation ?
> 
> I suspect you have a quite different issue here.
> 
> You simply could use skb_get_hash() since local TCP flows will provide a
> l4 skb->hash and you have no further flow dissection to do.

In our test, we have bisected and found the following patch introduced big 
overhead into skb_flow_dissect_flow_keys(), and caused performance 
regression:
commit: d34af823
net: Add VLAN ID to flow_keys

This patch didn't add too many instructions, but we think the change to 
the size of struct flow_keys may cause different cache missing rate...

To avoid affecting other drivers using this function, our patch limits the 
change inside our driver to fix this performance regression.

Regarding your suggestion on skb_get_hash(), I looked at the code and ran 
some tests, and found the skb->l4_hash and skb->sw_hash bits are not set, 
so it calls __skb_get_hash() which eventually calls 
skb_flow_dissect_flow_keys(). So it still includes the performance 
overhead mentioned above.

static inline __u32 skb_get_hash(struct sk_buff *skb)
{
        if (!skb->l4_hash && !skb->sw_hash)
                __skb_get_hash(skb);

        return skb->hash;
}


void __skb_get_hash(struct sk_buff *skb)
{
        struct flow_keys keys;

        __flow_hash_secret_init();

        __skb_set_sw_hash(skb, ___skb_get_hash(skb, &keys, hashrnd),
                          flow_keys_have_l4(&keys));
}


static inline u32 ___skb_get_hash(const struct sk_buff *skb,
                                  struct flow_keys *keys, u32 keyval)
{
        skb_flow_dissect_flow_keys(skb, keys,
                                   FLOW_DISSECTOR_F_STOP_AT_FLOW_LABEL);

        return __flow_hash_from_keys(keys, keyval);
}


Thanks,
- Haiyang

Re: [PATCH net-next] hv_netvsc: Use simple parser for IPv4 and v6 headers

[Eric Dumazet]

On Wed, 2015-12-16 at 19:20 +0000, Haiyang Zhang wrote:
> > -----Original Message-----
> > From: Eric Dumazet [mailto:eric.dumazet@gmail.com]
> > Sent: Wednesday, December 16, 2015 12:08 PM
> > 
> > This looks very very wrong to me.
> > 
> > How many times this is called per second, for the 'one flow' case ?
> > 
> > Don't you use TSO in this driver ?
> > 
> > What about encapsulation ?
> > 
> > I suspect you have a quite different issue here.
> > 
> > You simply could use skb_get_hash() since local TCP flows will provide a
> > l4 skb->hash and you have no further flow dissection to do.
> 
> In our test, we have bisected and found the following patch introduced big 
> overhead into skb_flow_dissect_flow_keys(), and caused performance 
> regression:
> commit: d34af823
> net: Add VLAN ID to flow_keys

Adding Tom Herbert <tom@herbertland.com>

Your driver was assuming things about "struct flow_keys" layout.
This is not permitted.

Magic numbers like 12 and 8 are really bad...

static bool netvsc_set_hash(u32 *hash, struct sk_buff *skb)
{
        struct flow_keys flow;
        int data_len;

        if (!skb_flow_dissect_flow_keys(skb, &flow, 0) ||
            !(flow.basic.n_proto == htons(ETH_P_IP) ||
              flow.basic.n_proto == htons(ETH_P_IPV6)))
                return false;

        if (flow.basic.ip_proto == IPPROTO_TCP)
                data_len = 12;
        else
                data_len = 8;

        *hash = comp_hash(netvsc_hash_key, HASH_KEYLEN, &flow, data_len);

        return true;
}


> This patch didn't add too many instructions, but we think the change to 
> the size of struct flow_keys may cause different cache missing rate...
> 
> To avoid affecting other drivers using this function, our patch limits the 
> change inside our driver to fix this performance regression.
> 
> Regarding your suggestion on skb_get_hash(), I looked at the code and ran 
> some tests, and found the skb->l4_hash and skb->sw_hash bits are not set, 
> so it calls __skb_get_hash() which eventually calls 
> skb_flow_dissect_flow_keys(). So it still includes the performance 
> overhead mentioned above.

Okay, but have you tried this instead of just guessing ?

Are you forwarding traffic, or is the traffic locally generated ?

TCP stack does set skb->l4_hash for sure in current kernels.

Your 'basic flow dissection' is very buggy and a step backward.

Just call skb_get_hash() : Not only your perf problem will vanish, but
your driver will correctly work with all possible malformed packets
(like pretending to be TCP packets but too small to even contain one
byte of TCP header) and well formed ones, with all encapsulations.

Re: [PATCH net-next] hv_netvsc: Use simple parser for IPv4 and v6 headers

[David Miller]

From: Haiyang Zhang <haiyangz@microsoft.com>
Date: Wed, 16 Dec 2015 19:20:44 +0000

> In our test, we have bisected and found the following patch introduced big 
> overhead into skb_flow_dissect_flow_keys(), and caused performance 
> regression:
> commit: d34af823
> net: Add VLAN ID to flow_keys

NEVER _EVER_ work around this kind of problem by bypassing the code in
question in your driver.

ALWAYS work to fix the actual problem.