Re: upstream regression (IO-APIC?)

[Bartlomiej Zolnierkiewicz <bzolnier@gmail.com>]

On Monday 03 November 2008, Alok Kataria wrote:
> On Mon, 2008-11-03 at 09:28 -0800, Bartlomiej Zolnierkiewicz wrote:
> > On Sunday 02 November 2008, Bartlomiej Zolnierkiewicz wrote:
> > > On Sunday 02 November 2008, Bartlomiej Zolnierkiewicz wrote:
> > > > On Thursday 30 October 2008, Robert Hancock wrote:
> > > > > Bartlomiej Zolnierkiewicz wrote:
> > > > > > The current Linus tree as of commit e946217e4fdaa67681bbabfa8e6b18641921f750
> > > > > > is broken for me.  I get either the following panic (see log from qemu below)
> > > > > > or lost IRQs on ATA init...  Is this a known issue?
> > > > > > 
> > > > > > PS The tree that I used before and was supposedly good (sorry, I'm too tired
> > > > > > to verify it now) had commit 57f8f7b60db6f1ed2c6918ab9230c4623a9dbe37 at head.
> > > > 
> > > > Unfortunately 57f8f7b60db6f1ed2c6918ab9230c4623a9dbe37 (v2.6.28-rc1)
> > > > is also bad.  Bisecting it further was a real pain (i.e. I hit broken
> > > > build with x86 irqbalance changes, broken build with netfilter nat
> > > > changes and jbd journal problem).  In the end it turned out that 2.6.27
> > > > is bad too!  However with 2.6.27 the panic occurs only once per several
> > > > attempts and if there is no panic kernel boots normally (no lost IRQs).
> > > > 
> > > > [...]
> > > > 
> > > > I finally managed to narrow it down to change making x86 use tsc_khz
> > > > for loops_per_jiffy -- commit 3da757daf86e498872855f0b5e101f763ba79499
> > > > ("x86: use cpu_khz for loops_per_jiffy calculation").  This approach
> > > > seems too simplistic (as I see now Arjan & Pavel expressed concerns
> > > > about it back when the patch was posted initially [1][2]).  Also it
> > > > would probably be preferred to re-use existing preset_lpj variable
> > > > (just like KVM does it for similar purpose [3]) instead of adding a
> > > > lpj_tsc one and increasing complexity.
> > > 
> > > It turned out that I can boot a kernel with different config with
> > > HZ == 250 just fine and switching to HZ == 1000 makes it fail.
> > > 
> > > 
> > > Looking into it some more:
> > > 
> > > HZ == 250 kernel (good):
> > > 
> > > Calibrating delay loop (skipped), value calculated using timer frequency.. 2986.79 BogoMIPS (lpj=5973580)
> > > 
> > > HZ == 1000 kernel (bad):
> > > 
> > > Calibrating delay loop (skipped), using tsc calculated value.. 2990.35 BogoMIPS (lpj=1495176)
> > > 
> > > HZ == 1000 kernel with hackyfix (good):
> > > 
> > > Calibrating delay using timer specific routine.. 3016.68 BogoMIPS (lpj=6033376)
> > > 
> > > 
> > > Argggh... lpj is used for udelay() & friends so this bug is quite
> > > dangerous (since udelay() & friends are used for hardware delays)...
> > 
> > It may be not as severe as I initially thought,
> > (obviously) the real hardware works fine:
> > 
> > calibrate_delay_direct(): lpj=1495884
> > Calibrating delay loop (skipped), value calculated using timer frequency.. 2990.36 BogoMIPS (lpj=1495183)
> > 
> > So the issue only affects qemu ATM.
> Oh so its on a emulator, something wrong in the timer emulation logic in
> qemu ?

Probably.  I now noticed that the problem happens only with HZ == 250 host
and HZ == 1000 guest.  When host and guest are using the same HZ setting
everything works fine.

Thanks,
Bart