Re: ksoftirqd uses 99% CPU triggered by network traffic (maybe RLT-8139 related)

[OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>]

Francois Romieu <romieu@fr.zoreil.com> writes:

> Hector Martin <hector@marcansoft.com> :
> [...]
> > using a stabler TCP/IP stack. This one works OK. No problem so far... 
> > anyway, even though it stopped and I had to restart the PS2 some times, 
> > the PC has been receiving packets with no reboot whatsoever. I think 
> > it's fixed :)
> 
> Ok, I'll send the final version of the patches for inclusion in -netdev
> and/or -mm this evening. It will provide a broader testing.

Bit interesting.

On the final analysis, what was the cause of this problem?
I found the following in progguide-8100(100).pdf. Does this help something?

2.5 Software Issues

This section covers the handling of various data reception topics.

    1. Handling a Receive Buffer Overflow:
       The Rx DMA (FIFO to buffer) is stopped. The CAPR must be
       updated first to dismiss the ISR (RxBufferOverflow) event. The
       correct actions to process RxBufOvw are:
           a. Update CAPR.
           b. Write a 1 to ISR (ROK).
       The Rx DMA resumes after step b.

    2. Handling RxFIFOOvw:
       When RxFIFOOvw occurs, all incoming packets are discarded.
       Clearing ISR (RxFIFOOvw) doesn t dismiss the RxFIFOOvw event.
       To dismiss the RxFIFOOvw event, the ISR (RxBufOvw) must be
       written with a 1.

-- 
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>