Re: [net-next 8/8] stmmac: limit max_mtu in case of 4KiB and use __netdev_alloc_skb.

[Giuseppe CAVALLARO <peppe.cavallaro@st.com>]

On 10/18/2011 12:18 PM, Eric Dumazet wrote:
> Le mardi 18 octobre 2011 à 12:01 +0200, Giuseppe CAVALLARO a écrit :
>> Problem using big mtu around 4096 bytes is you end allocating (4096
>> +NET_SKB_PAD + NET_IP_ALIGN + sizeof(struct skb_shared_info) bytes ->
>> 8192 bytes : order-1 pages
>>
>> It's better to limit the mtu to SKB_MAX_HEAD(NET_SKB_PAD),
>> to have no more than one page per skb.
>>
>> Also the patch changes the netdev_alloc_skb_ip_align() done in
>> init_dma_desc_rings() and uses a variant allowing GFP_KERNEL allocations
>> allowing the driver to load even in case of memory pressure.
>>
>> Reported-by: Eric Dumazet <eric.dumazet@gmail.com>
>> Signed-off-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
>> ---
>>  drivers/net/ethernet/stmicro/stmmac/stmmac_main.c |   13 +++++++++----
>>  1 files changed, 9 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
>> index 1848a16..f5ca3be 100644
>> --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
>> +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
>> @@ -474,11 +474,13 @@ static void init_dma_desc_rings(struct net_device *dev)
>>  	for (i = 0; i < rxsize; i++) {
>>  		struct dma_desc *p = priv->dma_rx + i;
>>  
>> -		skb = netdev_alloc_skb_ip_align(dev, bfsize);
>> +		skb = __netdev_alloc_skb(dev, bfsize + NET_IP_ALIGN,
>> +					 GFP_KERNEL);
>>  		if (unlikely(skb == NULL)) {
>>  			pr_err("%s: Rx init fails; skb is NULL\n", __func__);
>>  			break;
>>  		}
>> +		skb_reserve(skb, NET_IP_ALIGN);
>>  		priv->rx_skbuff[i] = skb;
>>  		priv->rx_skbuff_dma[i] = dma_map_single(priv->device, skb->data,
>>  						bfsize, DMA_FROM_DEVICE);
>> @@ -1176,12 +1178,15 @@ static inline void stmmac_rx_refill(struct stmmac_priv *priv)
>>  
>>  			skb = __skb_dequeue(&priv->rx_recycle);
>>  			if (skb == NULL)
>> -				skb = netdev_alloc_skb_ip_align(priv->dev,
>> -								bfsize);
>> +				skb = __netdev_alloc_skb(priv->dev, bfsize +
>> +							 NET_IP_ALIGN,
>> +							 GFP_KERNEL);
>>  
> 
> No, you cant do that in softirq context. We cant sleep here and must use
> GFP_ATOMIC
> Only the init_dma_desc_rings() part is OK, we run in process context and
> are allowed to sleep in memory allocations (GFP_KERNEL)
> 
> 
>>  			if (unlikely(skb == NULL))
>>  				break;
>>  
>> +			skb_reserve(skb, NET_IP_ALIGN);
>> +
>>  			priv->rx_skbuff[entry] = skb;
>>  			priv->rx_skbuff_dma[entry] =
>>  			    dma_map_single(priv->device, skb->data, bfsize,
>> @@ -1401,7 +1406,7 @@ static int stmmac_change_mtu(struct net_device *dev, int new_mtu)
>>  	if (priv->plat->enh_desc)
>>  		max_mtu = JUMBO_LEN;
>>  	else
>> -		max_mtu = BUF_SIZE_4KiB;
>> +		max_mtu = SKB_MAX_HEAD(NET_SKB_PAD);
>>  
> 
> minor nit (since NET_IP_ALIGN is 0 anyway on modern x86)
> max_mtu = SKB_MAX_HEAD(NET_SKB_PAD + NET_IP_ALIGN);
> 
> 
>>  	if ((new_mtu < 46) || (new_mtu > max_mtu)) {
>>  		pr_err("%s: invalid MTU, max MTU is: %d\n", dev->name, max_mtu);
> 
> 
> 

Ok! I'm reworking and sending it again.

Thx
Peppe