Customer explains: With heavy disk usage and net activity my machine crashes. I'm running 6.0 on a 686, 10 Gb EIDE, I forget the ethernet card (the machine is not local, can't check at the moment) but the driver is the epic100. This is definitely a serious 6.0 problem as 5.1 ran on the same hardware for around 8 months with no trouble whatsoever under extremely heavy usage conditions. He can reproduce this by running 'hdparm -t /dev/hda' while ping flooding the machine. I was not able to reproduce it. It seems similar to bug #2938, and customer tried steeping down to 386 optimization, though he did not remove Watchdog or APM. Same results.
I was unable to get this to fail on a test lab machine containing a Cyrix MII 300. I ping flooded the machine while running the hdparm -t /dev/hda command. It got noticable slower but never locked up. ------- Additional Comments From 06/12/99 01:43 ------- Some info got left off of my original report below. First, another user saw exactly the same behavior: http://x28.deja.com/getdoc.xp?AN=484168039.1&CONTEXT=928218503.648347748&hitnum=20 I don't get a reboot, the machine just locks up. Second, doing a hdparm -d 1/dev/hda (i.e. enabling DMA) "solves" the problem. Third, ifconfig shows that packets are being dropped prior to the failure. If I am ftp'ing a large file, the machine locks, but the halt can be avoided if I continually do network restarts during the transfer.
This is a real bug in the EPIC100 driver. It requires that interrupts are deferred for a long time while the card receives only small frames. The fix is in 2.2.7 or so and higher, I can provide it if need be.