9.1 callout behavior

[Adrian Chadd]

Hi!

I know this is bringing up an old thread, but Dell has shipped us some
hardware (thanks Dell!) and I'm helping them take a look at it.


On 10 December 2013 07:23, Alexander Motin <mav at freebsd.org> wrote:
> On 10.12.2013 17:12, Bret Ketchum wrote:
>>
>>      Do either of you have a dual socket/package motherboard to play
>> with? I've tried a couple single socket motherboards and cannot
>> reproduce the issue. I'm wondering if this occurs on only multi-socket
>> mobos.
>
>
> My main test system is dual-socket (Supermicro X8DTU).

It's reasonably reproducable.

The tscdrift tool from jhb (in tools/tools/tscdrift) looks thus:

root at appollyon:~/src/tscdrift # ./tscdrift

CPU | TSC skew (min/avg/max/stddev)

----+------------------------------

  0 |     0     0     0    0.000
  1 |    34    79   380   39.005
  2 |   306   476  1326   120.472
  3 |   306   485  1426   123.487
  4 |   280   473  7500   248.965
  5 |   280   462  1320   126.691
  6 |   300   461  1934   129.362
  7 |   300   470  1354   122.654
  8 |   292   420  3640   152.029
  9 |   135   190   655   58.601
 10 |   112   188   620   56.490
 11 |   114   189   660   62.440
 12 |   129   204   566   53.731
 13 |   129   206   617   56.047
 14 |   126   211   620   54.450
 15 |   126   213   603   54.808
 16 |   440   590  1683   217.649
 17 |   440   612  1606   234.295
 18 |   468   642  4017   266.768
 19 |   463   653  8683   352.624
 20 |   480   671  1500   255.395
 21 |   480   689  8060   348.384
 22 |   468   707  3766   296.721
 23 |   466   703  1683   284.625
 24 |   480   767  8183   373.741
 25 |   486   782  7400   362.069
 26 |   480   664  1620   249.125
 27 |   477   686  3669   278.144
 28 |   469   621  1897   226.673
 29 |   469   649  8275   363.575
 30 |   457   636  1835   250.005
 31 |   451   641  4300   272.322

If I modify their supplied ticktock module to call
callout_reset_on(args, 0); to call on CPU #0, then it actually doesn't
seem to drift.

I'm going to look at mav's recent patch to the scheduler in case it
has an effect on things.


-a