skge killing off snd_via686 interrupts on Fedora Core 5

skge killing off snd_via686 interrupts on Fedora Core 5

[Sitsofe Wheler]

We have a few machines which have been upgraded to to D-Link System Inc
DGE-530T Gigabit Ethernet adaptors. The unfortunate news is that shortly
after this was done people found that programs were either going
into the D state or hanging while trying to access the soundcard using
ALSA. Looking in the logs turned up this message:

irq 11: nobody cared (try booting with the "irqpoll" option)
 [<c0139893>] __report_bad_irq+0x2b/0x69     [<c0139a54>] note_interrupt+0x183/0x1af
 [<c01393fa>] handle_IRQ_event+0x23/0x4c     [<c01394bd>] __do_IRQ+0x9a/0xcd
 [<c0104c2e>] do_IRQ+0x5c/0x77     =======================
 [<c01035b6>] common_interrupt+0x1a/0x20
 [<e08e6fd4>] skge_poll+0x20b/0x3fc [skge]     [<c028c1f7>] net_rx_action+0x78/0x13f
 [<c011f83a>] __do_softirq+0x35/0x7f     [<c0104c81>] do_softirq+0x38/0x3f
 =======================
 [<c011f8d7>] local_bh_enable+0x53/0x5e     [<c02ad51a>] tcp_recvmsg+0x317/0x6fb
 [<c02842c6>] sock_common_recvmsg+0x2f/0x45     [<c028250d>] sock_recvmsg+0xe9/0x104
 [<c012b10b>] autoremove_wake_function+0x0/0x2d     [<c02dd4b5>] _spin_unlock_irq+0x5/0x7
 [<c01c48b8>] vsnprintf+0x422/0x461     [<c0104c40>] do_IRQ+0x6e/0x77
 [<c02835a3>] sys_recvfrom+0xad/0x10c     [<c0151e25>] fd_install+0x24/0x50
 [<c02dd493>] _read_unlock_irq+0x5/0x7     [<c013a33a>] find_get_page+0x39/0x3f
 [<c0104c40>] do_IRQ+0x6e/0x77     [<c01035b6>] common_interrupt+0x1a/0x20
 [<c0104c40>] do_IRQ+0x6e/0x77     [<c01035b6>] common_interrupt+0x1a/0x20
 [<c028361b>] sys_recv+0x19/0x1d     [<c0283c47>] sys_socketcall+0x117/0x19e
 [<c0104c40>] do_IRQ+0x6e/0x77     [<c0102be9>] syscall_call+0x7/0xb
handlers:
[<e09dfca3>] (snd_via686_interrupt+0x0/0xcd [snd_via82xx])
Disabling IRQ #11

This has also been reported in the Red Hat bugzilla over here:
https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=192980
on a 2.6.16-1.2122_FC5 kernel. The problem is highly reproducible and
unloading/reloading the snd_via686 module temporarily fixes the problem
until the next big burst of network traffic.

Re: skge killing off snd_via686 interrupts on Fedora Core 5

[Alan Cox]

On Sad, 2006-05-27 at 15:43 +0100, Sitsofe Wheler wrote:
> We have a few machines which have been upgraded to to D-Link System Inc
> DGE-530T Gigabit Ethernet adaptors. The unfortunate news is that shortly
> after this was done people found that programs were either going
> into the D state or hanging while trying to access the soundcard using
> ALSA. Looking in the logs turned up this message:
> 
> irq 11: nobody cared (try booting with the "irqpoll" option)

If running with irqpoll fixes it then I'd suggest looking for BIOS
updates. The older VIA boards often had buggy IRQ router information.
Are you updating from a 100Mbit card in the same slot or onboard
ethernet ?

Re: skge killing off snd_via686 interrupts on Fedora Core 5

[Sitsofe Wheeler]

On Thu, 01 Jun 2006 18:03:37 +0100, Alan Cox wrote:

> On Sad, 2006-05-27 at 15:43 +0100, Sitsofe Wheler wrote:
>> We have a few machines which have been upgraded to to D-Link System Inc
>> DGE-530T Gigabit Ethernet adaptors. The unfortunate news is that shortly
>> after this was done people found that programs were either going
>> into the D state or hanging while trying to access the soundcard using
>> ALSA. Looking in the logs turned up this message:
>> 
>> irq 11: nobody cared (try booting with the "irqpoll" option)
> 
> If running with irqpoll fixes it then I'd suggest looking for BIOS
> updates. The older VIA boards often had buggy IRQ router information.

I have a feeling these machines are running the latest BIOS available to
them. I will look into this though.

> Are you updating from a 100Mbit card in the same slot or onboard
> ethernet ?

We are updating from the former (a removable PCI 100Mbit card being
replaced with a 1000Mbit one). If you think it's non BIOS related and
want to have a poke let me know and I'll add you to the sudoers list on
those machines.

-- 
Sitsofe | http://sucs.org/~sits/

Re: skge killing off snd_via686 interrupts on Fedora Core 5

[Sitsofe Wheeler]

(This accidentally fell off list so I'm going to see if I bodge it back)

On Sat, 03 Jun 2006 10:55:48 +0100, Alan Cox wrote:
> Ar Sad, 2006-06-03 am 21:31 +0100, ysgrifennodd Sitsofe Wheeler:
> > As mentioned in another reply the cards aren't onbord and are a PCI card
> > upgrade.
> 
> Same slot as the card you upgraded from ?

You've got me there. I have no idea - most of the machines had their
network card removed before a clean upgrade from Fedora Core 4 to Fedora
Core 5. I can go and swap an old card back on to the same slot though. On
the one machine that had its card swapped after the upgrade it looks like
ACPI allocated the old and new cards the same IRQ (16). Is there anything
in particular that I'm looking for on the old card (e.g. would showing
/proc/interrupts before and after help)? It might be worth noting that 
IRQ 11 doesn't seem to be disabled when there is anything else going
during the network transfer - e.g. if a video is being watched while the
netperf is being carried out the IRQ isn't disabled. Seemingly the problem
only shows when the card is hitting some sort of top speed...

[netperf showed that there was a performance hit to enabling irqpoll but
there was less of a hit using noapic. Either option cured the problem
and I added the netperf score to the mail]

> > irqpoll
> >  87380  16384  16384    10.00      29.63
> > noapic
> >  87380  16384  16384    10.00      53.21
> 
> That does strongly suggest IRQ routing problems are involved, but it may
> also be timing, hence the question about slots

-- 
Sitsofe | http://sucs.org/~sits/

Re: [SOLVED] skge killing off snd_via686 interrupts on Fedora Core 5

[Sitsofe Wheeler]

On Sun, 04 Jun 2006 10:45:46 +0100, Sitsofe Wheeler wrote:

> (This accidentally fell off list so I'm going to see if I bodge it back)
> 
> On Sat, 03 Jun 2006 10:55:48 +0100, Alan Cox wrote:
>> Ar Sad, 2006-06-03 am 21:31 +0100, ysgrifennodd Sitsofe Wheeler:
>> > As mentioned in another reply the cards aren't onbord and are a PCI
>> > card upgrade.
>> 
>> Same slot as the card you upgraded from ?

Alan correctly surmised this was due to a VIA motherboard bug and moving
the gigabit card to a different slot solved the problem without the need
for noapic or irqpoll. The issue turned up with other network cards (e.g.
a tuplip and an 8139) so it was fairly clear the issue wasn't driver
related.

Just for the record the motherboard in question is a Gigabyte GA-7DX+,
host bridge is a AMD-760 [IGD4-1P], ISA bridge is a VIA VT82C686 [Apollo
Super South] (rev 40)). At least two of the slots in the machine are
"cursed".

-- 
Sitsofe | http://sucs.org/~sits/