svn commit: r274739 - head/sys/mips/conf

[Ian Lepore]

On Tue, 2014-11-25 at 08:59 -0800, Adrian Chadd wrote:
> On 25 November 2014 at 06:23, Ian Lepore <ian at freebsd.org> wrote:
> > On Tue, 2014-11-25 at 13:15 +0100, Dag-Erling Smørgrav wrote:
> >> Ian, please try the attached patch.
> >>
> >> The way this is intended to be used is that you set up a system with an
> >> /etc/rc that does nothing else than mount the root file system and
> >> append the output from "sysctl -b hw.attachtimes" to a log file, then
> >> reboot.  You let it run for as long as you wish, then copy the log file
> >> to another machine and run the attachtimes utility (source code below)
> >> on the log file to extract the data and display either the raw numbers
> >> or a histogram showing the distribution.
> >>
> >> See freefall:~des/software/attachtimes.tgz for an example of how to set
> >> this up.  All you need is the loader and kernel, a minimal /etc, and the
> >> tools required by /etc/rc (sh, fsck, mount, df, sysctl, halt, reboot).
> >> The example contains quite a bit more than that because I wanted to be
> >> able to boot it in single-user mode for debugging.
> >>
> >> The patch can easily be modified to record the actual timestamps instead
> >> of just the delta, but remember to modify the utility as well, or the
> >> output will be complete nonsense.
> >>
> >> DES
> >
> > Getting the results for this is going to take a while... I can't get the
> > system past mountroot right now.  I have no idea why.  It's been months
> > since I last successfully booted -current on this old hardware.  I'm
> > also very busy with $work.
> >
> > For the 3 numbers that are identical at the start, I think we're reading
> > garbage in binuptime() because the clock driver isn't working yet.  The
> > numbers first change with at91_st0, that's the clock driver.  atmelarm0
> > is the parent of everything listed between nexus and it, so the large
> > number there is probably the subtraction of the post-init clock value
> > from the pre-init clock garbage.
> >
> > In a more general sense, I'm going to repeat a couple things...
> >
> > The clock all this is being measured with runs at 32 KHz.  That's 'K'.
> >
> > This output is evidence that the system behaves exactly as I've been
> > saying it behaves for a long long while.  You seem convinced, I don't
> > understand why, that there must be some error here.  I don't understand
> > why you think a system like this would behave any differently each time
> > it is powered on.  The only actual entropy involved is whatever minor
> > thermal transients may exist in the crystal oscillator.  With 32KHz
> > resolution (or even a few MHz) that amounts to not a lot of measurable
> > variation at all.
> >
> > The repeatability of the boot sequence of hardware like this is nearly
> > perfect at the resolution we're measuring.  While that may be bad for
> > gathering entropy, it's a wonderful thing when you're debugging, because
> > problems that would be almost impossible to nail down on modern complex
> > hardware are 100% reproducible by just hitting the reset button.  That
> > reproducibility extends all the way into multiuser mode unless there is
> > a network connection where packet arrival times start adding
> > interrupt-based entropy.
> >
> 
> Can you bring up the clock first? Or at least earlier?
> 

First as in "before nexus"?  That sounds tricky.  Looking at the order
of things now, it's nexus, the main SoC bus, the interrupt controller,
power-management, clock driver.  That seems reasonable to me, you need
all that other stuff first to get the clock to the point where it can
register as a timecounter and provide timestamps.

> Should bintime() be returning garbage so early? I wonder if that'd
> have an effect on any other driver startup paths that may use it.
> 
> 

It seems odd that it returns garbage, but the alternative would be to
return zero, and while that makes more sense to me, it isn't helpful in
terms of providing useful data (other than as an indication that the
clock isn't running yet).

-- Ian