GFP_ATOMIC vs GFP_KERNEL in netfilter module

GFP_ATOMIC vs GFP_KERNEL in netfilter module

[Roberto Jung Drebes]

Hello there,

I am testing a simple netfilter module which delays every 5th IP packet 
received. I am having problems allocating the timer structure. Here is 
the netfilter hook I register:

static unsigned int
comfirm_rx_hook(unsigned int hook,
                 struct sk_buff **pskb,
                 const struct net_device *in,
                 const struct net_device *out,
                 int (*okfn) (struct sk_buff *))
{
     static int counter;
     struct timer_list *tl;
     struct cf_delay_parms *dp;

     if ((*pskb)->nfmark != CF_MAGIC) {
         counter++;

         if (!(counter % 5)) {
             dp = kmalloc(sizeof(struct cf_delay_parms), GFP_KERNEL);
             tl = kmalloc(sizeof(struct timer_list), GFP_KERNEL);
             dp->timer = tl;
             (*pskb)->nfmark = CF_MAGIC;
             dp->skb = *pskb;
             tl->data = (unsigned long) dp;
             tl->function = (void *) cf_delay_rx;
             tl->expires = jiffies + 100;
             init_timer(tl);
             add_timer(tl);
             printk("packet delayed\n");
             return NF_STOLEN;
         } else
             return NF_ACCEPT;
     } else
         return NF_ACCEPT;
}

void cf_delay_rx(struct cf_delay_parms *parms)
{
     ip_rcv(parms->skb, parms->skb->dev, NULL);
     kfree(parms->timer);
     kfree(parms);
}

If I use GFP_KERNEL with the kmallocs, I get errors like

Debug: sleeping function called from invalid context at mm/slab.c:2000
in_atomic():1[expected: 0], irqs_disabled():0
  [<0211b765>] __might_sleep+0x82/0x8c
  [<02146dbc>] kmem_cache_alloc+0x1d/0x4c
  [<2204b047>] comfirm_rx_hook+0x47/0xd8 [mymodule]
  [<022a68eb>] nf_iterate+0x40/0x89
  [<022b5d63>] ip_rcv_finish+0x0/0x1d9
  [<022a6bac>] nf_hook_slow+0x47/0xb1
  [<022b5d63>] ip_rcv_finish+0x0/0x1d9
  [<022b5be7>] ip_rcv+0x36d/0x3a1
  [<022b5d63>] ip_rcv_finish+0x0/0x1d9
  [<0229eef3>] netif_receive_skb+0x1b0/0x1dd
  [<0229ef8c>] process_backlog+0x6c/0xd9
  [<0229f056>] net_rx_action+0x5d/0xcd
  [<02122fa1>] __do_softirq+0x35/0x73
  [<02108a7b>] do_softirq+0x46/0x4d
  =======================
  [<02107e67>] do_IRQ+0x2f7/0x303
  [<021fc4c1>] acpi_processor_idle+0xd3/0x1c5
  [<0210408c>] cpu_idle+0x1f/0x34
  [<0239b6ae>] start_kernel+0x221/0x224

If I use GFP_ATOMIC, I don't get the error, but I think timers are not 
being called after the delay. I have a similar code for transmition, 
which works OK with GFP_KERNEL (delays messages) but with GFP_ATOMIC it 
does also not delay.

I test delay with ping, and I am running kernel 2.6.8-1.521 from Fedora 
Core 2.

What am I doing wrong?

TIA,

ps: Please CC me any reply from the linux-kernel mailing list.

-- 
Roberto Jung Drebes <drebes@inf.ufrgs.br>
Porto Alegre, RS - Brasil
http://www.inf.ufrgs.br/~drebes/

Re: GFP_ATOMIC vs GFP_KERNEL in netfilter module

[Robert Love]

On Thu, 2004-12-02 at 21:12 -0200, Roberto Jung Drebes wrote:

> If I use GFP_KERNEL with the kmallocs, I get errors like

So don't use GFP_KERNEL; use GFP_ATOMIC.  You cannot sleep in netfilter
hooks, since they are run in interrupt context (when the packet is
received) by the net_rx_action softirq.  You can see this in your stack
trace.

> If I use GFP_ATOMIC, I don't get the error, but I think timers are not 
> being called after the delay. I have a similar code for transmition, 
> which works OK with GFP_KERNEL (delays messages) but with GFP_ATOMIC it 
> does also not delay.

If you have problems with GFP_ATOMIC, the problem is elsewhere and not
in the fact that your allocations are atomic versus not.

First thing, though, check the return value from kmalloc for failure (dp
and tl equal NULL) to ensure that the allocation is even working!

	Robert Love

Re: GFP_ATOMIC vs GFP_KERNEL in netfilter module

[Roberto Jung Drebes]

On 02/12/2004, at 21:12, Roberto Jung Drebes wrote:

> If I use GFP_ATOMIC, I don't get the error, but I think timers are not 
> being called after the delay. I have a similar code for transmition, 
> which works OK with GFP_KERNEL (delays messages) but with GFP_ATOMIC 
> it does also not delay.
>
> I test delay with ping, and I am running kernel 2.6.8-1.521 from 
> Fedora Core 2.
>
> What am I doing wrong?

Just for reference, everything was working alright with GFP_ATOMIC. I 
thought the timer was not installed because ping would not detect the 
delay. This happened because ping writes to the ICMP packet the 
timestamp when it is first received by the kernel, before netfilter 
hooks. Using ping with the old behavior (ping -U) worked as expected, 
printing the time the packet was kept in my netfilter hook.

Thanks,

-- 
Roberto Jung Drebes <drebes@inf.ufrgs.br>
Porto Alegre, RS - Brasil
http://www.inf.ufrgs.br/~drebes/