Network sort of stops working with the em (intel) driver under load.

[Michael Powell]

mark rowlands wrote:

> Newly built and cvsupped system,with GENERIC kernel,  when copying
> large amounts of data over a gigabit link via scp  the network  will
> hang after a couple of gig.  I can then no longer login via ssh. If I
> leave it be, after about 12-24 hours I can then login again. (A reboot
> of course fixes the issue immediately...) .
> 
> The system is very vanilla, only accf_http in loader.conf. No sysctl
> tuning done, no firewall hankypanky.
> 
> pciconf -lv | grep -A4 ^em
> em0 at pci0:3:4:0: class=0x020000 card=0x001e8086 chip=0x100e8086 rev=0x02
> hdr=0x00
>     vendor     = 'Intel Corporation'
>     device     = 'Gigabit Ethernet Controller (82540EM)'
>     class      = network
>     subclass   = ethernet
> 
> em0: <Intel(R) PRO/1000 Legacy Network Connection 1.0.1> port
> 0xec00-0xec3f mem 0xfeae0000-0xfeafffff,0xff irq 17 at device 4.0 on
> pci3
> 
> Suggestions as to where I could look for more information as to the
> precise nature of the problem gratefully received.  Current plan is to
> purchase another variety of gigabit card to see if it is specific to
> the intel card.
> 


If memory serves, I think there may have been some traffic about something 
like this on the -CURRENT list. You might look/search there and see if it 
sounds similar. If it seems like it might be the same thing, look for an MFC 
back to -STABLE. Sometimes the fix for very a specific item which has been 
addressed is to take a system to -STABLE in order to obtain the fixed bits.

Research and confirm first, before considering such an update. My policy on 
-STABLE in the past is I only think about going there for a very narrow and 
specific situation where I know I have a problem that the devs have seen, 
analyzed, and fixed, with subsequent MFC.

Something else too - if you can disable the vr and the USB chips completely, 
it might provide a data point. IRQ sharing is supposed to work well, and it 
is something that may be eliminated from the scenario easily if you do not 
need these things.

-Mike