Re: Fwd: [BISECTED] "sparc64: Fix numa distance values" breakage

[Nitin Gupta <nitin.m.gupta@oracle.com>]

Mike,

I believe this is due to the firmware exporting wrong/incomplete
information about memory latency groups in the machine descriptor (MD).
Before this patch, this information was not used at all and kernel
always used default values for numa node distance values. With incorrect
values, scheduler can have a skewed view of the machine causing this
non optimal usage. My testing on T7, T5, T4 with recent firmwares never
showed such issues.

Can you please provide output of 'numactl --hardware' on your machine?
Ideally, I would also require dump of the MD but I don't have a script
handy for this which I can share externally.

Dave: would you have a script to dump MD which you can share?

Thanks,
Nitin

>>
>> From: Mikael Pettersson <mikpelinux@gmail.com>
>> Subject: [BISECTED] "sparc64: Fix numa distance values" breakage (was: 4.4-rc kernels only use one of two CPUs on Sun Blade 2500)
>> Date: December 30, 2015 at 9:18:57 AM MST
>> To: Mikael Pettersson <mikpelinux@gmail.com>
>> Cc: Linux SPARC Kernel Mailing List <sparclinux@vger.kernel.org>
>>
>> Mikael Pettersson writes:
>>> Something is causing the 4.4-rc kernels to only use half the CPU
>>> capacity of my Sun Blade 2500 (dual USIIIi).  The kernel does detect
>>> both CPUs, but it doesn't seem to want to schedule processes on
>>> both of them.  During CPU-intensive jobs like GCC bootstraps, 'top'
>>> indicates the machine is 50% idle and aggregate CPU usage is 100%
>>> (should be 200%).  This is completely deterministic.
>>>
>>> Going back to 4.3.0 resolves the problems.
>>
>> A git bisect identified the commit below as the culprit.
>> I've confirmed that reverting it from 4.4-rc7 solves the problem.
>>
>> commit 52708d690b8be132ba9d294464625dbbdb9fa5df
>> Author: Nitin Gupta <nitin.m.gupta@oracle.com>
>> Date:   Mon Nov 2 16:30:24 2015 -0500
>>
>>     sparc64: Fix numa distance values
>>
>>     Orabug: 21896119
>>
>>     Use machine descriptor (MD) to get node latency
>>     values instead of just using default values.
>>
>>     Testing:
>>     On an T5-8 system with:
>>      - total nodes = 8
>>      - self latencies = 0x26d18
>>      - latency to other nodes = 0x3a598
>>        => latency ratio = ~1.5
>>
>>     output of numactl --hardware
>>
>>      - before fix:
>>
>>     node distances:
>>     node   0   1   2   3   4   5   6   7
>>       0:  10  20  20  20  20  20  20  20
>>       1:  20  10  20  20  20  20  20  20
>>       2:  20  20  10  20  20  20  20  20
>>       3:  20  20  20  10  20  20  20  20
>>       4:  20  20  20  20  10  20  20  20
>>       5:  20  20  20  20  20  10  20  20
>>       6:  20  20  20  20  20  20  10  20
>>       7:  20  20  20  20  20  20  20  10
>>
>>      - after fix:
>>
>>     node distances:
>>     node   0   1   2   3   4   5   6   7
>>       0:  10  15  15  15  15  15  15  15
>>       1:  15  10  15  15  15  15  15  15
>>       2:  15  15  10  15  15  15  15  15
>>       3:  15  15  15  10  15  15  15  15
>>       4:  15  15  15  15  10  15  15  15
>>       5:  15  15  15  15  15  10  15  15
>>       6:  15  15  15  15  15  15  10  15
>>       7:  15  15  15  15  15  15  15  10
>>
>>     Signed-off-by: Nitin Gupta <nitin.m.gupta@oracle.com>
>>     Reviewed-by: Chris Hyser <chris.hyser@oracle.com>
>>     Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
>>     Signed-off-by: David S. Miller <davem@davemloft.net>
>> --