Skip to contentSkip to conversation

uli526x hangs on 2.6.29, not on rc8

DISCUSSION Santiago Garcia Mantinan 3 participants 6 messages Mar 26, 2009 – Mar 28, 2009

lore
Share & tools
lore
#0 lore raw

Hi!

After upgrading from rc8 to final 2.6.29 my machine's network stops working after some full speed (100Mb/s) downloads ifdown;ifup, or powering off the switch won't solve the problem, however it is solved if I unload the driver module and load it again. Tested with another NIC (e1000e based) and couldn't reproduce the problem.

This is happening on final 2.6.29 version but doesn't seem to happen on rc8 with the same kernel config. Machine is a AMD Athlon 64 running Debian 64 bits.

I don't see any changes on the git for uli526x from rc8 to final, but the tulip driver on which this one is based had some changes.

Don't know what else to add, just ask me for whatever data that may help on this.

Regards...

Signature · 1 line
Manty/BestiaTester -> http://manty.net
#1 lore raw
Quoted 12 lines · Santiago Garcia Mantinan

Hi!

After upgrading from rc8 to final 2.6.29 my machine's network stops working after some full speed (100Mb/s) downloads ifdown;ifup, or powering off the switch won't solve the problem, however it is solved if I unload the driver module and load it again. Tested with another NIC (e1000e based) and couldn't reproduce the problem.

This is happening on final 2.6.29 version but doesn't seem to happen on rc8 with the same kernel config. Machine is a AMD Athlon 64 running Debian 64 bits.

Quoted 7 lines

I don't see any changes on the git for uli526x from rc8 to final, but the tulip driver on which this one is based had some changes.

Don't know what else to add, just ask me for whatever data that may help on this.

Regards...

Hi

This is a known problem, and a fix should be available very soon for 2.6.29.1

http://www.spinics.net/lists/netdev/msg92829.html

You can check net/core/dev.c function process_backlog() should get this updated version :

static int process_backlog(struct napi_struct *napi, int quota)

{
	int work = 0;
	struct softnet_data *queue = &__get_cpu_var(softnet_data);
	unsigned long start_time = jiffies;

	napi->weight = weight_p;
	do {
		struct sk_buff *skb;

		local_irq_disable();
		skb = __skb_dequeue(&queue->input_pkt_queue);
		if (!skb) {
			list_del(&napi->poll_list);
			clear_bit(NAPI_STATE_SCHED, &napi->state);
			local_irq_enable();
			break;
		}
		local_irq_enable();

		napi_gro_receive(napi, skb);
	} while (++work < quota && jiffies == start_time);

	napi_gro_flush(napi);

	return work;
}
#2 lore raw
In reply to Santiago Garcia Mantinan

Hi Manty,

Quoted 5 lines · Santiago Garcia Mantinan

After upgrading from rc8 to final 2.6.29 my machine's network stops working after some full speed (100Mb/s) downloads ifdown;ifup, or powering off the switch won't solve the problem, however it is solved if I unload the driver module and load it again. Tested with another NIC (e1000e based) and couldn't reproduce the problem.

Known issue probably. Try: http://lkml.org/lkml/2009/3/25/217

Cheers,
FJP

#4 lore raw
Quoted 7 lines · Santiago Garcia Mantinan
Quoted 1 line

Known issue probably. Try: http://lkml.org/lkml/2009/3/25/217

Yes, seems related, I'm testing that patch right now on my machine, but I'm wondering if I should test that one or this one
(http://www.spinics.net/lists/netdev/msg92829.html) sent also by Herbert to the netdev list but at a later time. I'll try test the second one also just in case.

No, the plain revert is safer.

If you want to test another patch, you should probably try this one instead, which AFAIK is the best real fix available:
http://lkml.org/lkml/2009/3/26/79

Cheers,
FJP

#5 lore raw
Quoted 7 lines
Quoted 6 lines
Quoted 1 line

Known issue probably. Try: http://lkml.org/lkml/2009/3/25/217

Yes, seems related, I'm testing that patch right now on my machine, but I'm wondering if I should test that one or this one
(http://www.spinics.net/lists/netdev/msg92829.html) sent also by Herbert to the netdev list but at a later time. I'll try test the second one also just in case.

No, the plain revert is safer.

Ok, I have tested both during quite a long time (hours compared to the few minutes it took the system to hang before) and both patches seem to work ok, I'll stick with the first one following your recommendations.

Thanks!

Signature · 1 line
Manty/BestiaTester -> http://manty.net