Re: phenom, amd780g, tsc, hpet, kvm, kernel -- who's at fault?

[Ingo Molnar <mingo@elte.hu>]

* Michael Tokarev <mjt@tls.msk.ru> wrote:

> Ingo Molnar wrote:
>> * Michael Tokarev <mjt@tls.msk.ru> wrote:
>>
>>> Now, after quite some googling around, I tried to disable hpet,  
>>> booting with hpet=disable parameter.  And that one fixed all the  
>>> problems at once. 7 days uptime, I stress-tested it several times,  
>>> it works with TSC as timesource (still a problem within guests as  
>>> those shows unstable TSC anyway) since boot, no issues logged.  Even 
>>> cpufreq works as expected...
> []
>> It could again go bad like it did before - those messages are signs of 
>> HPET weirdnesses.
>>
>> Probably your box's hpet needs to be blacklisted, so that it gets  
>> disabled automatically on bootup.
>
> Well, I'm not convinced at all... at least not yet ;)
>
> The reason is simple: this box was rock solid a few months back. 
> With 2.6.25 and 2.6.26 kernels I think.  It had probs with kvm 
> (bugs), and lacked in general hardware support (both the chipset 
> and phenom cpu were still too new to be fully supported).  At that 
> time I installed the thing (was a test install with a random hdd, 
> so I added real drives and installed real distro), with quite a 
> lot of data copying back and forth (were rearranging partitions, 
> raid arrays, guests and so on, copying data to another disk, to 
> another machine and back).  There was no single issue, no single 
> mention of tsc or hpet instabilities, and system time was stable 
> too.  But since some time, -- unfortunately I don't know when 
> exactly, and sure thing it'd be very interesting to know, I'll try 
> to figure it out -- first it started showing system clock 
> weirdness, and finally come to this Friday the 13 incident.
>
> That all to say: it was stable with earlier kernel.  Now it's not. 
> Maybe, just maybe, at that time hpet wasn't supported, or maybe 
> wasn't used, or supported not in full to rely on it - I've no 
> idea.  If that's the case, I'll just shut up now because the whole 
> point becomes moot.

We added force-enabling of the hpets of certain boards over the past 
few kernel releases. Do you have kernel logs from earlier kernels, 
do you know it for sure that .28 was the first that enabled the 
hpet?

> Maybe it was due to somehow broken bios -- I did several bios 
> updates there, mostly because linux complained about something 
> scary (something akin "wasting so much megs memory due to bios not 
> set up something (GART? IOMMU?)") and I was hoping to fix that. 
> And it will be fixed someday in bios...
>
> (By the way: how bad the lack of hpet is?  It's used for 
> something, and having it malfunctioning and disabled does not 
> sound good, esp. on a machine which is running close to its 
> maximum...  Maybe I should return the mobo back? :)

a hpet isnt really important for server workloads. It's useful in 
terms of keeping dynticks timeouts long on the desktop - but on a 
busy server it has little relevance.

	Ingo