Re: Too many interrupts

Le Sun, 3 Apr 2005 20:54:34 -0700 (PDT), Donald Becker <[email protected]>=
 =20
a =E9crit:

> On Sat, 2 Apr 2005, infos chez fdrconseil wrote:
>
>> I am using an Intel eepro 100 dual port adapter and I have a problem.
>> I want to use only one port (eth0) and no cable is connected on the =20
>> second
>> one.
>> The board is using IRQ 6; Apparently, there is no IRQ sharing.
>
> The motherboard sets the IRQ assignment.  It tries to map one device pe=
r
> IRQ.  With the APIC, Advanced Programmable Interrupt Controller, the OS
> has the option of reconfiguing so that the interrupt arrives on a "high=
"
> IRQ.
>
>> The problem is that the network board is raising a huge amount of
>> interrupts and the CPUs
>> (bi pentium III motherboard) are busy processing these interrupts (abo=
ut
>> 10%).
>
> What is the packet count?  The eepro100 driver, before it was modified =
by
> others, should raise about one interrupt per Rx packet.  With normal
> traffic it's possible that all Tx packet will be handled without
> interrupts.  With heavy traffic there should be less than one interrupt
> per Rx packet.

The packet count seems normal :


  cat /proc/net/dev ; sleep 10 ; cat /proc/net/dev
Inter-|   Receive                                                |  =20
Transmit
  face |bytes    packets errs drop fifo frame compressed multicast|bytes =
   =20
packets errs drop fifo colls carrier compressed
     lo:85309643  107592    0    0    0     0          0         0 =20
85309643  107592    0    0    0     0       0          0
   eth0:41981809  476933    0    0    0     0          0         0 =20
177882307  149047    0    0    0     0       0          0
   eth1:       0       0    0    0    0     0          0         0       =
 =20
0       0    0    0    0     0       0          0


Inter-|   Receive                                                |  =20
Transmit
  face |bytes    packets errs drop fifo frame compressed multicast|bytes =
   =20
packets errs drop fifo colls carrier compressed
     lo:85314415  107598    0    0    0     0          0         0 =20
85314415  107598    0    0    0     0       0          0
   eth0:41985205  476985    0    0    0     0          0         0 =20
177883471  149055    0    0    0     0       0          0
   eth1:       0       0    0    0    0     0          0         0       =
 =20
0       0    0    0    0     0       0          0



I understand eth0 received 52 packets in 10 seconds.


> Sometime people don't notice that their machine is being attacked with
> packets that are being immediately discarded.

I don't notice the machine is being attacked...


>> Contents of /proc/interrupts :
>>    6:  501331320  501309654   IO-APIC-level  eth0
>> The same file, few seconds later :
>>    6:  502333959  502312308   IO-APIC-level  eth0
>
> About a million interrupts.  That's pretty intense.  I'm guessing that =
=20
> you
> don't have a million packets reported in /proc/net/dev?

You're right, I don't.


> Do you get the high interrupt rate as soon as you start the system?  Or
> only later?  If only later, there might be an interrupt source that can=
't
> be cleared.

I've got it since i start the system.


> Here is a quick check: when you start getting a high interrupt rate,
> disconnect the cable or use 'mii-diag -F 10baseT eth0' to force the
> interface into slow transfers.


I issued the command "mii-tool -F 10baseT-FD eth0"
but it did not change anything.
I went back with "mii-tool -F 100baseTx-FD eth0" as it was.


>> Here is the dmesg messages when starting the system :
>> Apr  2 15:52:47 localhost kernel: eepro100.c:v1.09j-t 9/29/99 Donald
>> Becker http://www.scyld.com/network/eepro100.html
>> Apr  2 15:52:47 localhost kernel: eepro100.c: $Revision: 1.36 =20
>> $ 2000/11/17
>> Modified by Andrey V. Savochkin <[email protected]> and others
>> Apr  2 15:52:47 localhost kernel: eth0: Intel Corp. 82557/8/9 [Etherne=
t
>> Pro 100], 00:02:B3:B6:2B:39, IRQ 6.
>>
>> The questions are :
>> - Am i wrong saying the problem comes from the network card ?
>
> Perhaps not.  There may be some other interrupt source on IRQ6.  The
> eepro100 driver interrupt handler is being run, but usually has nothing=
 =20
> to
> do.  If you didn't have a SMP your machine would appear to be hung.


I'm not sure I understand you fully : do you mean that with only
one processor, the system would be frozen ???


>> - What can I do ? (CPU power is wasted and CPU temperature increases)
>
> Run the above tests.

Ok. Done. Does it help ?

> Install the driver with debugging turned on to see if there are any
> interrupt sources.

Pretty complicated. Isn't it possible to activate some more information =20
without
recompiling the module ?


> Move the card to a different physical PCI slot to try to get another IR=
Q
> assignment.
>

As I said before on the list, this is a production server. It's a bit =20
difficult to stop it.
Anyway, I'll try to find a moment to do that.

Thank you for these answers,
FdR.