Re: Debug: sleeping function called from invalid context at mm/slab.c:2126

[Tony <tony.uestc@gmail.com>]

Parag Warudkar wrote:
>>I'm writing a net_device driver. I want to send a packet when the timer 
>>is out. I get the following warning. It seems that I should not call 
>>alloc_skb. Can anyone tell me how to get rid of this? Thanks in advance.
>>
> 
> 
> You are calling alloc_skb which in turn calls kmem_cache_alloc in interrupt context where things can't sleep and kmem_cache_alloc can sleep.  The reason for this is that you are passing GFP_KERNEL to alloc_skb. Try passing GFP_ATOMIC instead.
> 
> Other alternative is to may be use a precreated pool of skbs - may be this can be done in driver init function or any other safe context. But I don't know how much feasible that is in your situation.
> 
> HTH
> Parag
> 
> 
> 
Thanks a lot. Another question.

My interface is a virtual interface which represent a radio connected to 
the host using ethernet NIC. I designed my own L2 protocol on top of 
802.3, which must be used, since the radio and the host are connected by 
ethernet.

Now, my radio_hard_header will only add my L2 header, and my 
radio_hard_start_xmit will do (simplified):
1) ajust the headroom space
     hh_len = LL_RESERVED_SPACE(bdev);
     if (unlikely(skb_headroom(skb) < hh_len && bdev->hard_header)) {
		struct sk_buff *skb2;

		skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
		if (skb2 == NULL) {
             stats->tx_dropped++;
			dev_kfree_skb(skb);
			return 0;
		}
		if (skb->sk)
			skb_set_owner_w(skb2, skb->sk);
		dev_kfree_skb(skb);
		skb = skb2;
	}

2) call eth0->hard_header
3) skb->dev = eth0
    return dev_queue_xmit()

The problem is when system try to retransmit the packet, I add another 
ethernet header mistakenly.

I have two question:
1) I do not modify the skb passed to hard_start_xmit if 
skb_realloc_headroom is executed. only in this case the retransmission 
runs well. Is my understanding right?
2) Should I do this way or add the ethernet header in my 
radio_hard_header? If I choose the later, the problem will be how should 
I handle it when eth_hard_header return a negative number, when ARP is 
needed.

Thx