Re: 2.6.14-mm1

[Andrew Morton]

Valdis.Kletnieks@xxxxxx wrote:
>
> On Mon, 07 Nov 2005 11:52:10 PST, Andrew Morton said:
> 
> > > 2) The PS/2 keyboard death on ppp traffic is still not fixed. 
> > > Reproducible even on slow GPRS if there's something else (e.g. glxgears) 
> > > that eats some CPU time. When keyboard is dead, events/0 consomes 100% 
> > > of CPU. Nothing in dmesg. If you outline some suspicious pieces of code, 
> > > I will insert printks there in order to debug this.
> > 
> > input guys cc'ed.
> 
> Getting myself on the cc: list, as I've seen this one on 2.6.14-rc5-mm1 (haven't
> nailed it on 14-mm1 *yet*, but only been up for 12 hours).  Also, some
> additional info:
> 
> The keyboard is dead, but other stuff still works - I've been able to issue
> commands by laborious cut-n-paste into an xterm window.  X is still up and
> responding, as are all the clients, so it's *not* a hard loop in events/0.
> 
> Also, I've had gkrellm running when it hits, and it will show incoming data
> rates on the modem of 3.5Mbytes/sec (as opposed to the 5K/sec you'd expect from
> a 56k modem).  A few times, I've had it go into auto-ambush on an iptables rule,
> with the same rule tripping several tens of thousands of times in a row,
> which makes me think it's got to do with a short packet (such as an inbound
> SYN packet) going into replicator mode and just being handed up from the
> device driver over and over, thousands of times....
> 
> alt-sysrq still works - I can sysrq-T to get traces, -S to sync, -B to reboot
> and so on, and the output gets through klogd and syslogd and into /var/adm/messages.

That sounds like softirq starvation.

Or maybe the input layer uses keventd services and keventd is stuck.

> I'm able to often trigger the bug by opening a new tab in Firefox, as that (a)
> involves small SYN+ACK packets coming back and (b) a Firefox bug causes it to
> chew CPU when displaying a page in a tab....
> 
> I'm willing to test-drive any debugging/patches needed, as this is probably the
> single biggest stability hit I have in -mm at the moment.
> 

I'd try hitting sysrq-p ten times, then take a look at the logs.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/