Re: igc problems with heavy traffic

From: Jim King <jim_at_jimking.net>
Date: Fri, 05 Aug 2022 15:41:42 UTC
On 8/5/2022 7:48 AM, mike tancsa wrote:
> On 8/5/2022 2:19 AM, Jim King wrote:
>> On 8/4/2022 12:29 PM, mike tancsa wrote:
>>> On 8/4/2022 1:12 PM, mike tancsa wrote:
>>>> On 7/31/2022 11:41 PM, Jim King wrote:
>>>>> I'm trying to use a Qnap dual port 2.5Gb NIC with FreeBSD 13.1. 
>>>>> FreeBSD recognizes the NIC and I'm able to connect to my LAN at 
>>>>> 1Gb.  However, trying to push heavy traffic through the NIC 
>>>>> results in a temporary loss of connectivity:  traffic stops and 
>>>>> the message "igc0: link state changed to DOWN" appears in the 
>>>>> system log. About 10 seconds later the message "igc0: link state 
>>>>> changed to UP" is logged, followed by some messages from dhclient 
>>>>> indicating that it has acquired a lease.  At that point network 
>>>>> connectivity works normally, until I try to push a lot of traffic 
>>>>> again.  I've been able to trigger this every time using iperf3 
>>>>> from another PC on my LAN, and by using Ookla's speedtest client 
>>>>> (my Internet service is good for about 900 Mbps).
>>>>>
>>>> How long does it take to wedge the nic ? I have a pair of Topton 
>>>> routers (4 IGC NICs) back to back and a simple iperf3 test seems to 
>>>> work ok between the 2 units.
>>>>
>>> Actually, I was able to reproduce this too running iperf3 -P4 for 
>>> about 60 seconds. Watching via sysutils/ifstat, it does indeed drop 
>>> traffic when the link drops
>>>
>>>
>> When I'm receiving traffic the wedge usually happens pretty quick. 
>> iperf while connected at 1Gb (I get similar results at 2.5Gb):
>>
> If I fiddle with the Flow Control settings it seems to make a 
> difference.  I can do it without bouncing the link.  I wonder what the 
> Linux driver has as default ?
>
> ie.
>
> sysctl -w dev.igc.1.fc=1
> sysctl -w dev.igc.0.fc=1
>
> Does it help in your case ?
>
Thanks for the suggestion.  Unfortunately it doesn't seem to help in my 
case.  All four values for flow control give pretty much the same 
results with my iperf test.

Jim