2.6.9 RCU breakage in dev_queue_xmit

2.6.9 RCU breakage in dev_queue_xmit

[Jeff V. Merkey]

Running dual gigabit interfaces at 196 MB/S (megabytes/second) on 
receive, 12 CLK interacket gap time, 1500 bytes payload
at 65000 packets per second per gigabit interface, and retransmitting 
received packets at 130 MB/S out of a third gigabit interface
with skb, RCU locks in dev_queue_xmit breaks and enters the following state:

Nov  8 15:38:08 ds kernel: Badness in local_bh_enable at 
kernel/softirq.c:141
Nov  8 15:38:08 ds kernel:  [<40107d1e>] dump_stack+0x1e/0x30
Nov  8 15:38:08 ds kernel:  [<401218b0>] local_bh_enable+0x70/0x80
Nov  8 15:38:08 ds kernel:  [<402c5bbb>] dev_queue_xmit+0x11b/0x250
Nov  8 15:38:08 ds kernel:  [<f8981cb7>] xmit_skb+0x17/0x20 [dsfs]
Nov  8 15:38:08 ds kernel:  [<f8981f8e>] xmit_packet+0x2e/0x80 [dsfs]
Nov  8 15:38:08 ds kernel:  [<f89820eb>] regen_data+0x10b/0x290 [dsfs]
Nov  8 15:38:08 ds kernel:  [<401052c5>] kernel_thread_helper+0x5/0x10
Nov  8 15:38:08 ds kernel: Badness in local_bh_enable at 
kernel/softirq.c:141
Nov  8 15:38:08 ds kernel:  [<40107d1e>] dump_stack+0x1e/0x30
Nov  8 15:38:08 ds kernel:  [<401218b0>] local_bh_enable+0x70/0x80
Nov  8 15:38:08 ds kernel:  [<402c5bbb>] dev_queue_xmit+0x11b/0x250
Nov  8 15:38:08 ds kernel:  [<f8981cb7>] xmit_skb+0x17/0x20 [dsfs]
Nov  8 15:38:08 ds kernel:  [<f8981f8e>] xmit_packet+0x2e/0x80 [dsfs]
Nov  8 15:38:08 ds kernel:  [<f89820eb>] regen_data+0x10b/0x290 [dsfs]
Nov  8 15:38:08 ds kernel:  [<401052c5>] kernel_thread_helper+0x5/0x10


And before any of you guys whine about "give me a test case" I just 
did.   Device driver is e1000 in the 2.6.9 kernel
tree.   System is a Xeon based system at 3 Ghz single processor with 4 
GB of ram and a 3GB/1GB kernel/user
split address space.

Jeff

Re: 2.6.9 RCU breakage in dev_queue_xmit

[Patrick McHardy]

Jeff V. Merkey wrote:

>
> Running dual gigabit interfaces at 196 MB/S (megabytes/second) on 
> receive, 12 CLK interacket gap time, 1500 bytes payload
> at 65000 packets per second per gigabit interface, and retransmitting 
> received packets at 130 MB/S out of a third gigabit interface
> with skb, RCU locks in dev_queue_xmit breaks and enters the following 
> state:
>
> Nov  8 15:38:08 ds kernel: Badness in local_bh_enable at 
> kernel/softirq.c:141
> Nov  8 15:38:08 ds kernel:  [<40107d1e>] dump_stack+0x1e/0x30
> Nov  8 15:38:08 ds kernel:  [<401218b0>] local_bh_enable+0x70/0x80
> Nov  8 15:38:08 ds kernel:  [<402c5bbb>] dev_queue_xmit+0x11b/0x250
> Nov  8 15:38:08 ds kernel:  [<f8981cb7>] xmit_skb+0x17/0x20 [dsfs]
> Nov  8 15:38:08 ds kernel:  [<f8981f8e>] xmit_packet+0x2e/0x80 [dsfs]
> Nov  8 15:38:08 ds kernel:  [<f89820eb>] regen_data+0x10b/0x290 [dsfs]

There is no such function in the 2.6.9 kernel.

Regards
Patrick

Re: 2.6.9 RCU breakage in dev_queue_xmit

[Stephen Hemminger]

On Tue, 09 Nov 2004 11:40:26 -0700
"Jeff V. Merkey" <jmerkey@drdos.com> wrote:

> Patrick McHardy wrote:
> 
> > Jeff V. Merkey wrote:
> >
> >>
> >> Running dual gigabit interfaces at 196 MB/S (megabytes/second) on 
> >> receive, 12 CLK interacket gap time, 1500 bytes payload
> >> at 65000 packets per second per gigabit interface, and retransmitting 
> >> received packets at 130 MB/S out of a third gigabit interface
> >> with skb, RCU locks in dev_queue_xmit breaks and enters the following 
> >> state:
> >>
> >> Nov  8 15:38:08 ds kernel: Badness in local_bh_enable at 
> >> kernel/softirq.c:141
> >> Nov  8 15:38:08 ds kernel:  [<40107d1e>] dump_stack+0x1e/0x30
> >> Nov  8 15:38:08 ds kernel:  [<401218b0>] local_bh_enable+0x70/0x80
> >> Nov  8 15:38:08 ds kernel:  [<402c5bbb>] dev_queue_xmit+0x11b/0x250
> >> Nov  8 15:38:08 ds kernel:  [<f8981cb7>] xmit_skb+0x17/0x20 [dsfs]
> >> Nov  8 15:38:08 ds kernel:  [<f8981f8e>] xmit_packet+0x2e/0x80 [dsfs]
> >> Nov  8 15:38:08 ds kernel:  [<f89820eb>] regen_data+0x10b/0x290 [dsfs]
> >
> >
> > There is no such function in the 2.6.9 kernel.
> 
> 
> Check /include/linux for local_bh_enable and /net/core/dev.c .
> 
> Jeff
> 
> >
> > Regards
> > Patrick

Patrick is asking about the function regen_data which doesn't exist in the
standard kernel.

Re: 2.6.9 RCU breakage in dev_queue_xmit

[Jeff V. Merkey]

Patrick McHardy wrote:

> Jeff V. Merkey wrote:
>
>>
>> Running dual gigabit interfaces at 196 MB/S (megabytes/second) on 
>> receive, 12 CLK interacket gap time, 1500 bytes payload
>> at 65000 packets per second per gigabit interface, and retransmitting 
>> received packets at 130 MB/S out of a third gigabit interface
>> with skb, RCU locks in dev_queue_xmit breaks and enters the following 
>> state:
>>
>> Nov  8 15:38:08 ds kernel: Badness in local_bh_enable at 
>> kernel/softirq.c:141
>> Nov  8 15:38:08 ds kernel:  [<40107d1e>] dump_stack+0x1e/0x30
>> Nov  8 15:38:08 ds kernel:  [<401218b0>] local_bh_enable+0x70/0x80
>> Nov  8 15:38:08 ds kernel:  [<402c5bbb>] dev_queue_xmit+0x11b/0x250
>> Nov  8 15:38:08 ds kernel:  [<f8981cb7>] xmit_skb+0x17/0x20 [dsfs]
>> Nov  8 15:38:08 ds kernel:  [<f8981f8e>] xmit_packet+0x2e/0x80 [dsfs]
>> Nov  8 15:38:08 ds kernel:  [<f89820eb>] regen_data+0x10b/0x290 [dsfs]
>
>
> There is no such function in the 2.6.9 kernel.


Check /include/linux for local_bh_enable and /net/core/dev.c .

Jeff

>
> Regards
> Patrick
>
> -
> To unsubscribe from this list: send the line "unsubscribe 
> linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
>

Re: 2.6.9 RCU breakage in dev_queue_xmit

[Jeff V. Merkey]

[-- Attachment #1: Type: text/plain, Size: 1612 bytes --]

Stephen Hemminger wrote:

>On Tue, 09 Nov 2004 11:40:26 -0700
>"Jeff V. Merkey" <jmerkey@drdos.com> wrote:
>
>  
>
>>Patrick McHardy wrote:
>>
>>    
>>
>>>Jeff V. Merkey wrote:
>>>
>>>      
>>>
>>>>Running dual gigabit interfaces at 196 MB/S (megabytes/second) on 
>>>>receive, 12 CLK interacket gap time, 1500 bytes payload
>>>>at 65000 packets per second per gigabit interface, and retransmitting 
>>>>received packets at 130 MB/S out of a third gigabit interface
>>>>with skb, RCU locks in dev_queue_xmit breaks and enters the following 
>>>>state:
>>>>
>>>>Nov  8 15:38:08 ds kernel: Badness in local_bh_enable at 
>>>>kernel/softirq.c:141
>>>>Nov  8 15:38:08 ds kernel:  [<40107d1e>] dump_stack+0x1e/0x30
>>>>Nov  8 15:38:08 ds kernel:  [<401218b0>] local_bh_enable+0x70/0x80
>>>>Nov  8 15:38:08 ds kernel:  [<402c5bbb>] dev_queue_xmit+0x11b/0x250
>>>>Nov  8 15:38:08 ds kernel:  [<f8981cb7>] xmit_skb+0x17/0x20 [dsfs]
>>>>Nov  8 15:38:08 ds kernel:  [<f8981f8e>] xmit_packet+0x2e/0x80 [dsfs]
>>>>Nov  8 15:38:08 ds kernel:  [<f89820eb>] regen_data+0x10b/0x290 [dsfs]
>>>>        
>>>>
>>>There is no such function in the 2.6.9 kernel.
>>>      
>>>
>>Check /include/linux for local_bh_enable and /net/core/dev.c .
>>
>>Jeff
>>
>>    
>>
>>>Regards
>>>Patrick
>>>      
>>>
>
>Patrick is asking about the function regen_data which doesn't exist in the
>standard kernel.
>-
>
>  
>

Code for regen_data cuntion attached. This code provides the example of 
the calls into linux that cause
the RCU locks to fail in dev_queue_xmit. The remainder of this module is 
not open source and
proprietary.

Jeff






[-- Attachment #2: regen.c --]
[-- Type: text/x-csrc, Size: 7442 bytes --]


int regen_active = 0;
int kill_regen = 0;
int regen_pid = 0;
unsigned long dsfs_skb_count = 0;

static void p_timeout(unsigned long data)
{
   SignalSemaphore((struct semaphore *)data);
}

int pm_sleep(int ticks)
{
   struct timer_list l_timer;
   register int retCode;
   P_InitSemaphore(semaphore);

   if (!seconds)
      return 0;

   init_timer(&l_timer);
   l_timer.expires = jiffies + ((HZ / 100) * ticks); 
   l_timer.data = (unsigned long)&semaphore;
   l_timer.function = p_timeout;

   add_timer(&l_timer);
   retCode = down_interruptible(&semaphore);
   del_timer(&l_timer);

   return ((retCode == -EINTR) ? -EINTR : 0);

}

int xmit_skb(struct sk_buff *skb)
{
   dsfs_skb_count--;
   return (dev_queue_xmit(skb));
}

void release_skb(struct sk_buff *skb)
{
   dsfs_skb_count--;
   kfree_skb(skb);
}

struct sk_buff *create_xmit_packet(int interface, int *err, int *skb_len)
{
	struct sk_buff *skb;
	struct net_device *dev;

        if (err)
           *err = 0;

        if (skb_len)
           *skb_len = 0;

	dev = dev_get_by_index(interface);
	if (!dev)
        {
           if (err)
              *err = -ENXIO;
           return NULL;
        }

	if (!(dev->flags & IFF_UP))
        {
	   dev_put(dev);
           if (err)
              *err = -ENETDOWN;
           return NULL;
        }

        skb = alloc_skb(dev->mtu + dev->hard_header_len + 15, GFP_ATOMIC);
        if (!skb)
        {
	   dev_put(dev);
           if (err)
              *err = -ENOBUFS;
           return NULL;
        }
        dsfs_skb_count++;

	skb_reserve(skb, (dev->hard_header_len + 15) & ~15);
	skb->nh.raw = skb->data;
        if (skb_len)
           *skb_len = (dev->mtu + dev->hard_header_len + 4);

	skb->dev = dev;

	dev_put(dev);
        return skb;

}

struct vlan_ethhdr 
{
   unsigned char	h_dest[6];     /* destination eth addr	*/
   unsigned char	h_source[6];   /* source ether addr	*/
   unsigned short       h_vlan_proto;  /* Should always be 0x8100 */
   unsigned short       h_vlan_TCI;    /* Encapsulates priority and VLAN ID */
   unsigned short	h_vlan_encapsulated_proto; /* packet type ID field (or len) */
};

void put_xmit_data(struct sk_buff *skb, unsigned char *data, int len, 
                   int proto, P_HANDLE *p_handle)
{
        struct vlan_ethhdr *vlan;

        if (skb->dev->hard_header) 
           skb->dev->hard_header(skb, skb->dev, ntohs(proto), 
                                 NULL, NULL, len);

        skb->tail = skb->data;
        skb->len = 0;

        // strip vlan headers if told to do so
        vlan = (struct vlan_ethhdr *)data;
        if ((p_handle->xmit_flags & STRIP_VLAN)
            && (ntohs(vlan->h_vlan_proto) == 0x8100))
        {
           memmove(skb->data, data, 12); 
           memmove(&skb->data[12], &data[16], len - 16); 
           skb_put(skb, len - 4);
        }
        else
        {
           memmove(skb->data, data, len); 
           skb_put(skb, len);
        }
        skb->protocol = proto;

        return;
}

int xmit_packet(struct sk_buff *skb)
{
	struct net_device *dev;
	int err;

	dev = dev_get_by_index(skb->dev->ifindex);
	if (!dev)
        {
           release_skb(skb);
	   return -ENXIO;
        }

	if (!(dev->flags & IFF_UP))
        {
	   dev_put(dev);
           release_skb(skb);
	   return -ENETDOWN;
        }

	skb->priority = 0;  // 0-6 set to 0

	err = xmit_skb(skb);
	if (err > 0 && (err = net_xmit_errno(err)) != 0)
        {
	   dev_put(dev);
           return err;
        }
	dev_put(dev);

	return 0;
}

int regen_data(void *arg)
{
    register ULONG pindex;
    struct sk_buff *skb;
    long long size;
    int err, skb_len;
    ULONG length = 0;
    VIRTUAL_SETUP *v = (VIRTUAL_SETUP *)arg;
    P_HANDLE *p_handle;

#if LINUX_26
    daemonize("if_regen%d", (int)v->pid);
#else
    sprintf(current->comm, "if_regen%d", (int)v->pid);
    daemonize();
#endif

    regen_active++;
    v->active++;
    while (v->ctl)
    {
retry:;
       if (v->interval)
       { 
#if LINUX_26
          v->currtime = CURRENT_TIME.tv_sec;
#else
          v->currtime = CURRENT_TIME;
#endif
          if (v->lasttime == v->currtime)
          {
             if (v->totalbytes >= (v->interval * (1000000 / 8)))
             {
                pi_sleep(1);
                goto retry; 
             }
          }
       }

       if (kill_regen)
          break;

       skb = create_xmit_packet(v->pindex, &err, &skb_len);
       if (!skb)
       {      
          switch (err)
          {
              case -ENXIO:
                 v->retry_errors++;
                 v->interface_errors++;
                 if (!pm_sleep(VIRTUAL_SLEEP))
                    goto retry;
                 goto exit_process;

              case -ENETDOWN:
                 v->interface_errors++;
                 v->retry_errors++;
                 if (!pm_sleep(VIRTUAL_SLEEP))
                    goto retry;
                 goto exit_process;

              case -EMSGSIZE:
                 v->size_errors++;
                 v->retry_errors++;
                 if (!pm_sleep(VIRTUAL_SLEEP))
                    goto retry;
                 goto exit_process;

              case -EINVAL:
                 v->fault_errors++;
                 v->retry_errors++;
                 if (!pm_sleep(VIRTUAL_SLEEP))
                    goto retry;
                 goto exit_process;

              case -ENOBUFS:
                 v->no_buffer_errors++;
                 v->retry_errors++;
                 if (!pm_sleep(VIRTUAL_SLEEP))
                    goto retry;
                 goto exit_process;

              default:
                 v->fault_errors++;
                 v->retry_errors++;
                 if (!pm_sleep(VIRTUAL_SLEEP))
                    goto retry;
                 goto exit_process;
          }
       }

read_again:;
       if ((kill_regen) || (!v->ctl))
       {
          release_skb(skb);
          goto exit_process;
       }

       p_handle = v->p_handle;
       if (!p_handle)
       {
          release_skb(skb);
          goto exit_process;
       }


       // this is the function that copies data into the skb
       // via copy_to_iovec
       pindex = regen_chain_packet(v->interface, skb, skb_len, p_handle, 
                                   &length, NULL, NULL, 
                                   &p_handle->start, &p_handle->end,
                                   v->d);

       if (pindex == -ENOENT)
       {
          release_skb(skb);
          goto exit_process;
       }

       if (pindex == 0xFFFFFFFF)
       {
          if (!pm_sleep(VIRTUAL_SLEEP))
             goto read_again;
          release_skb(skb);
          goto exit_process;
       }

       if (!length)
       {
          if (!pm_sleep(VIRTUAL_SLEEP))
             goto read_again;
          release_skb(skb);
          goto exit_process;
       }
   
       size = skb->len;

       err = xmit_packet(skb);
       if (err)
       {
          v->packets_aborted++;
       }
       else
       {
          v->bytes_xmit += size;
          v->packets_xmit++;
       }

       if (v->interval)
       {
#if LINUX_26
          v->currtime = CURRENT_TIME.tv_sec;
#else
          v->currtime = CURRENT_TIME;
#endif
          if (v->lasttime != v->currtime)
             v->totalbytes = 0;

          v->totalbytes += size;
          v->lasttime = v->currtime;
       }
    }

exit_process:;
    v->active--;
    regen_active--;

    return 0;

}

Re: 2.6.9 RCU breakage in dev_queue_xmit

[Herbert Xu]

Jeff V. Merkey <jmerkey@drdos.com> wrote:
> 
> Running dual gigabit interfaces at 196 MB/S (megabytes/second) on 
> receive, 12 CLK interacket gap time, 1500 bytes payload
> at 65000 packets per second per gigabit interface, and retransmitting 
> received packets at 130 MB/S out of a third gigabit interface
> with skb, RCU locks in dev_queue_xmit breaks and enters the following state:

This patch might help.
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
--
# This is a BitKeeper generated diff -Nru style patch.
#
# ChangeSet
#   2004/11/01 17:41:19-08:00 mingo@elte.hu 
#   [NET]: Fix unbalanced local_bh_enable() in dev_queue_xmit()
#   
#   Signed-off-by: Ingo Molnar <mingo@elte.hu>
#   Signed-off-by: David S. Miller <davem@davemloft.net>
# 
# net/core/dev.c
#   2004/11/01 17:40:59-08:00 mingo@elte.hu +5 -6
#   [NET]: Fix unbalanced local_bh_enable() in dev_queue_xmit()
#   
#   Signed-off-by: Ingo Molnar <mingo@elte.hu>
#   Signed-off-by: David S. Miller <davem@davemloft.net>
# 
diff -Nru a/net/core/dev.c b/net/core/dev.c
--- a/net/core/dev.c	2004-11-10 12:45:48 +11:00
+++ b/net/core/dev.c	2004-11-10 12:45:48 +11:00
@@ -1261,6 +1261,11 @@
 	struct Qdisc *q;
 	int rc = -ENOMEM;
 
+	/* Disable soft irqs for various locks below. Also 
+	 * stops preemption for RCU. 
+	 */
+	local_bh_disable(); 
+
 	if (skb_shinfo(skb)->frag_list &&
 	    !(dev->features & NETIF_F_FRAGLIST) &&
 	    __skb_linearize(skb, GFP_ATOMIC))
@@ -1284,12 +1289,6 @@
 	      skb->protocol != htons(ETH_P_IP))))
 	      	if (skb_checksum_help(skb, 0))
 	      		goto out_kfree_skb;
-
-
-	/* Disable soft irqs for various locks below. Also 
-	 * stops preemption for RCU. 
-	 */
-	local_bh_disable(); 
 
 	/* Updates of qdisc are serialized by queue_lock. 
 	 * The struct Qdisc which is pointed to by qdisc is now a 

Re: 2.6.9 RCU breakage in dev_queue_xmit

[Jeff V. Merkey]

Herbert Xu wrote:

>Jeff V. Merkey <jmerkey@drdos.com> wrote:
>  
>
>>Running dual gigabit interfaces at 196 MB/S (megabytes/second) on 
>>receive, 12 CLK interacket gap time, 1500 bytes payload
>>at 65000 packets per second per gigabit interface, and retransmitting 
>>received packets at 130 MB/S out of a third gigabit interface
>>with skb, RCU locks in dev_queue_xmit breaks and enters the following state:
>>    
>>
>
>This patch might help.
>  
>
Fixed.   Who an earth missed that?  dropping into an unlock case
by default without holding the lock?

Jeff

Re: 2.6.9 RCU breakage in dev_queue_xmit

[Jeff V. Merkey]

Herbert Xu wrote:

>Jeff V. Merkey <jmerkey@drdos.com> wrote:
>  
>
>>Running dual gigabit interfaces at 196 MB/S (megabytes/second) on 
>>receive, 12 CLK interacket gap time, 1500 bytes payload
>>at 65000 packets per second per gigabit interface, and retransmitting 
>>received packets at 130 MB/S out of a third gigabit interface
>>with skb, RCU locks in dev_queue_xmit breaks and enters the following state:
>>    
>>
>
>This patch might help.
>  
>

Herbert,

Even with this patch I still see RCU breakage at these data rates and 
the problem persists -- it just takes
longer for it to manifest (about 23 hours).  I am recoding 
dev_queue_xmit since the use of RCU primitives
is severely busted.   I looked over the code and the fact it breaks on 
uniprocessor is really a joke. 
No offense guys, but this is pretty bad.  How about something simple, 
like a spinlock or
multiple send queues per proc?

Jeff