Re: [circular locking bug] Re: [patch 00/15] clocksource / timekeeping rework V4 (resend V3 + bug fix)

[Martin Schwidefsky <schwidefsky@de.ibm.com>]

On Thu, 20 Aug 2009 11:58:21 +0200
Ingo Molnar <mingo@elte.hu> wrote:

> 
> * Martin Schwidefsky <schwidefsky@de.ibm.com> wrote:
> 
> > On Wed, 19 Aug 2009 22:25:54 +0200
> > Ingo Molnar <mingo@elte.hu> wrote:
> > 
> > > 
> > > ok, with all the latest patches i re-added these bits to 
> > > -tip, and it triggered this lockdep assert on a testbox:
> > 
> > Another one :-(
> > 
> > > stack backtrace:
> > > Pid: 1, comm: swapper Not tainted 2.6.31-rc6-tip-01234-gcc9be0e-dirty #1054
> > > Call Trace:
> > >  [<c106f430>] print_usage_bug+0x130/0x180
> > >  [<c106f5eb>] mark_lock_irq+0x16b/0x260
> > >  [<c106f240>] ? check_usage_forwards+0x0/0xc0
> > >  [<c106f7fe>] mark_lock+0x11e/0x3a0
> > >  [<c106fbff>] mark_irqflags+0x17f/0x190
> > >  [<c107177a>] __lock_acquire+0x29a/0x520
> > >  [<c1071a6a>] lock_acquire+0x6a/0xc0
> > >  [<c10664d7>] ? clocksource_unregister+0x17/0x50
> > >  [<c175719b>] __mutex_lock_common+0x3b/0x340
> > >  [<c10664d7>] ? clocksource_unregister+0x17/0x50
> > >  [<c1757551>] mutex_lock_nested+0x31/0x40
> > >  [<c10664d7>] ? clocksource_unregister+0x17/0x50
> > >  [<c10664d7>] clocksource_unregister+0x17/0x50
> > >  [<c1008b3a>] pit_disable_clocksource+0x2a/0x40
> > >  [<c1008bb9>] init_pit_timer+0x29/0xb0
> > >  [<c106825a>] clockevents_set_mode+0x1a/0x50
> > >  [<c1069a96>] tick_switch_to_oneshot+0x96/0xc0
> > >  [<c1069ad2>] tick_init_highres+0x12/0x20
> > >  [<c105e32d>] hrtimer_switch_to_hres+0x4d/0x100
> > >  [<c105ebbd>] hrtimer_run_pending+0x4d/0x50
> > >  [<c104bb85>] run_timer_softirq+0x25/0x230
> > 
> > Ok, the cause is that the i8253 pit clocksource code 
> > tries to unregister a clocksource from a timer 
> > interrupt. Bad idea with the new code. Why does the pit 
> > clocksource have to >unregister< the clock if the 
> > set_mode callback is called with 
> > CLOCK_EVT_MODE_SHUTODWN, CLOCK_EVT_MODE_UNUSED, or 
> > CLOCK_EVT_MODE_ONESHOT? Very strange, I would argue 
> > that the clocksource should never unregister in the 
> > set_mode callback, the timekeeping code should not use 
> > the clocksource if it is unsuitable for e.g. the one 
> > shot mode.
> 
> i think this 'execute timer management functions right 
> from the deep bowels of time events' concept is 
> fundamentally flawed and one big layering violation. It 
> caused numerous problems (lockups, etc.) in the past.
> 
> There should be a time management kernel thread instead 
> (or workqueue), which does a proper state machine of all 
> these properties - without having to call this stuff from 
> within a timer handler.

We could use that time managment kernel thread for the watchdog
downgrade as well. Dunno if it is worth to create another kernel thread
that just sits there doing nothing for 99.9% of the time.

As for the fix: my brains starts to hurt looking at the pit clocksource
code. Why does it set CLOCK_EVT_FEAT_ONESHOT but then unregisters the
clocksource when the mode is set to CLOCK_EVT_MODE_ONESHOT?? That does
not make any sense to me. I would have expected that the pit does not
set CLOCK_EVT_MODE_ONESHOT. The timekeeping code wouldn't try use the
clock for one-shot if the bit is not set. And to unregister the clock
only because the mode is set to shutdown or unused doesn't seem to be
necessary either. My fix would be to remove the CLOCK_EVT_MODE_ONESHOT
bit from the features mask and to remove the clocksource_unregister
from the set_mode callback.

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.