Re: [PATCH] clocksource, prevent overflow in clocksource_cyc2ns

[Prarit Bhargava <prarit@redhat.com>]

On 04/06/2012 07:29 PM, Thomas Gleixner wrote:
> On Thu, 5 Apr 2012, Prarit Bhargava wrote:
> 
>>>
>>> So what kernel version are you using?
>>
>> I retested using top of the linux.git tree, running
>>
>> echo 1 > /proc/sys/kernel/sysrq
>> for i in `seq 10000`; do sleep 1000 & done
>> echo t > /proc/sysrq-trigger
>>
>> and I no longer see a problem.  However, if I increase the number of threads to
>> 1000/cpu I get
>>
>> Clocksource %s unstable (delta = -429565427)
>> Clocksource switching to hpet
> 
> You are issuing a command which puts the kernel into a state where is
> dumps data for several seconds with interrupts disabled. And you expect that
> everything can cope with that?

Yes actually.  I do expect that everything "copes" with it.  I don't find it
unreasonable with system sizes increasing that functionality that has been
around for years works.

However, I also understand that no one expected or saw this problem -- I'm not
blaming anyone or screaming "Hey!  This is broken!!!".

>  
>> If I hack in (sorry for the cut-and-paste)
>> ....
>> +               cs_nsec = mult_frac(((csnow - cs->cs_last), cs->mult,
>> +                                   1UL << cs->shift);
>>
>> -               cs_nsec = clocksource_cyc2ns((csnow - cs->cs_last) &
>> -                                            cs->mask, cs->mult, cs->shift);
>> then I don't see unstable messages.
> 
> That does not make your approach more correct. The HPET wraparound
> time is ~3 seconds, so you screwed everything already, when your dump
> lasts longer than that. And there are clocksources which wrap way
> faster.
> 
> No, you can't fix that by hacking the timer code. A wraparound CANNOT
> be fixed by hacks.
> 
> So instead of fiddling in the victims, please fix the root cause,
> i.e. that stupid sysrq-t code which should not need to have interrupts
> disabled just to dump all that state. If that's not possible, send a
> patch to the sysrq documentation and warn about the consequences.
> 
> But stay away from code which is correct already. You CANNOT fix a
> problem which is caused by abnormal system state by hacking the code
> which is exposing the problem.
> 
> All you do is making hot pathes more expensive with a very dubious
> value. The time related calls are hotpath functions and optimized.
> 
> Aside of that you are breaking all architectures which do not have a
> native 64/32 instruction.
> 
> This mult_frac stuff is not going to happen, period.

Okay -- thanks for the info.  Much appreciated.

P.