Re: TSTILE lockups - not dead yet

[Don Lee <MacPPC2%c.icompute.com@localhost>]

On Dec 3, 2015, at 9:28 PM, John Klos <john%ziaspace.com@localhost> wrote:

>> I've seen this TSTILE problem. Never did get useful debug info. See:
>> https://mail-index.netbsd.org/port-macppc/2013/07/08/msg001958.html
> 
> It's not easy when there are hundreds of processes in TSTILE and people are bugging me because they can't check their email.

;->

>> I did a divide and conquer for the "analog bug", and understand that the fix is in post-6.1 code, but have not heard specifically that it's fixed in 6.1. I don't have a "scratch" macppc machine to test 7.0.
> 
> I think what you were seeing with Analog is what I was seeing with FLOPS:
> 
> http://mail-index.netbsd.org/port-macppc/2014/12/31/msg002138.html

Looks about the same. ISTR a small test case that would pop up FP errors on occasion if you ran it for a while.

The change that introduced the bug appeared to be a large mod written by Matt xxx (I don't remember the last name) that was marked as major change to CPU handling. The "divide" I did was to build kernels based on source from various dates. As I mention above, 

>>>>>> -D 2011-05-01 looks clean
>>>>>> -D 2011-05-03 looks bad.

This was on the pre-6.0 tree. It's apparently a May 2011 bug. Kernels from pre-5/3/2011 source were clean.

>> Now that I look at what I think is the commit, it looks like a small patch. (a bit keeping track of the FPU is getting lost)
>> 
>> If the kernel uses the FPU, this could be causing just about anything. I'm betting it does not.
> 
> I tried NetBSD 7 a year ago (well, end of 2014), and it still had FPU bugs. When I tried NetBSD 7 in October, it didn't. I never found any commits which I thought might've screwed up FP save/restore. Which commit do you think fixed the issue?
> 
> John

I think this looks promising, though I did not look at the code. I just saw this ref'd from the email archive.

http://mail-index.netbsd.org/source-changes/2015/07/06/msg067180.html

-dgl-