Re: [cobalt-developers] 2.2.14 kernel

[Dennis <dennis@xxxxxxxxx>]

At 11:42 AM 6/1/00 -0400, Dennis wrote:
>At 04:30 PM 5/31/00 -0700, Tim Hockin wrote:
>>Dennis wrote:
>>
>>> the 1.06 driver locks up under heavy load, so it is unusable in a heavily
>>
>>We haven't seen that.  I don't know what you mean by heavy load -
>>memory, CPU, or network?  We can put out almost wire-speed with our
>>RaQ3s.
>
>when the system runs out of buffers (ie cannot replenish the receive ring),
>the 1.06 driver never properly replenishes them when the buffers become
>available. The problem occurs when  packets are coming in faster than the
>box can process them (easily duplicatable by putting one ethernet on
>100Mb/s and the other on 10 and loading the 100Mb/s end)...the receive
>queue grows until there are no buffers, and the ethernet controller shuts
>down and has to be restarted manually.
>
>>
>>> I've also been trying to tell you that compiling eepro100 drivers does NOT
>>> WORK "out of the box" on the RaQ3. I get continuous mdio_read() errors and
>>> the driver doesnt function properly.
>>
>>Again, strange - I just dropped the 2.2.15 pre something driver into my
>>2.2.14 build tree and recompiled.  This was before 2.2.15 was 'done'. 
>>Also, you will find on shipping RaQ3s that the new driver, which uses
>>memory mapped IO, will probably BREAK on without newer firmware (PCI
>>initialization issue).  The firmware patch is being fixed, due to a bug
>>found by several developers.
>
>Well I'm running 2.2.12C5 (2.2.14 isnt out yet, remember?) We are also
>using modules. The modules load ok but then the console gets continuous
>mdio_read() errors.

Thinking about it....if you weren't properly setting 100Mb/s mode that
would explain your "10-fold" decrease in performance. Are you certain that
this wasnt the case?

dennis