[PATCH 1/2 v2] mv643xx_eth: use sw csum for big packets

[Saeed Bishara]

>-----Original Message-----
>From: Lennert Buytenhek [mailto:buytenh at wantstofly.org] 
>Sent: Saturday, June 12, 2010 11:32 AM
>To: Saeed Bishara
>Cc: linux-net at vger.kernel.org; linux-arm-kernel at lists.infradead.org
>Subject: Re: [PATCH 1/2 v2] mv643xx_eth: use sw csum for big packets
>
>On Thu, Jun 10, 2010 at 10:38:05AM +0300, Saeed Bishara wrote:
>
>> >> diff --git a/drivers/net/mv643xx_eth.c b/drivers/net/mv643xx_eth.c
>> >> index e345ec8..2d0e06b 100644
>> >> --- a/drivers/net/mv643xx_eth.c
>> >> +++ b/drivers/net/mv643xx_eth.c
>> >> @@ -289,6 +289,7 @@ struct mv643xx_eth_shared_private {
>> >>  	unsigned int t_clk;
>> >>  	int extended_rx_coal_limit;
>> >>  	int tx_bw_control;
>> >> +	int tx_csum_limit;
>> >>  };
>> >>  
>> >>  #define TX_BW_CONTROL_ABSENT		0
>> >> @@ -782,7 +783,8 @@ static int txq_submit_skb(struct 
>> >tx_queue *txq, struct sk_buff *skb)
>> >>  		       skb->protocol != htons(ETH_P_8021Q));
>> >>  
>> >>  		tag_bytes = (void *)ip_hdr(skb) - (void 
>> >*)skb->data - ETH_HLEN;
>> >> -		if (unlikely(tag_bytes & ~12)) {
>> >> +		if (unlikely(tag_bytes & ~12) ||
>> >> +			skb->len > mp->shared->tx_csum_limit) {
>> >
>> >Please line up skb->len with unlikely() on the line above it.
>> >
>> >
>> >> @@ -2666,6 +2668,9 @@ static int 
>> >mv643xx_eth_shared_probe(struct platform_device *pdev)
>> >>  	 * Detect hardware parameters.
>> >>  	 */
>> >>  	msp->t_clk = (pd != NULL && pd->t_clk != 0) ? pd->t_clk 
>> >: 133000000;
>> >> +	msp->tx_csum_limit = pd->tx_csum_limit ? 
>> >pd->tx_csum_limit : 9 * 1024;
>> >> +	/* add header count so we can compare against skb->len */
>> >> +	msp->tx_csum_limit += ETH_HLEN;
>> >>  	infer_hw_params(msp);
>> >
>> >Is the limit 9 * 1024 + 14 for the whole packet, or 9 * 
>1024 for the IP
>> >part?
>>
>> the limit is for the IP part, but I thought that adding the 
>header length, then comparing agains skb->len will be the 
>same. What do you suggest?
>
>Right, but for the header length you take 14, while if to include VLAN
>tags or DSA tags the header might actually be longer -- how does that
>affect the ability of the hardware to compute the checksum?
>
>I.e. is the restriction "total packet length must be < N + 14 bytes"
>or is it "the IP part must be < N bytes and it doesn't matter whether
>there are VLAN tags or not"?
It's the second option: "the IP part must be < N" regardless to the L2 header.
saeed
>
>I.e.:
>
>> >I.e. what happens if skb->len is 9 * 1024 + 15, but there 
>is one VLAN
>> >tag present -- will the hardware be able to do the checksum 
>offload or
>> >not?
>